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);
798 trace_block_rq_error(req, error, nr_bytes);
801 blk_account_io_completion(req, nr_bytes);
805 struct bio *bio = req->bio;
806 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
808 if (bio_bytes == bio->bi_iter.bi_size)
809 req->bio = bio->bi_next;
811 /* Completion has already been traced */
812 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
813 req_bio_endio(req, bio, bio_bytes, error);
815 total_bytes += bio_bytes;
816 nr_bytes -= bio_bytes;
827 * Reset counters so that the request stacking driver
828 * can find how many bytes remain in the request
835 req->__data_len -= total_bytes;
837 /* update sector only for requests with clear definition of sector */
838 if (!blk_rq_is_passthrough(req))
839 req->__sector += total_bytes >> 9;
841 /* mixed attributes always follow the first bio */
842 if (req->rq_flags & RQF_MIXED_MERGE) {
843 req->cmd_flags &= ~REQ_FAILFAST_MASK;
844 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
847 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
849 * If total number of sectors is less than the first segment
850 * size, something has gone terribly wrong.
852 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
853 blk_dump_rq_flags(req, "request botched");
854 req->__data_len = blk_rq_cur_bytes(req);
857 /* recalculate the number of segments */
858 req->nr_phys_segments = blk_recalc_rq_segments(req);
863 EXPORT_SYMBOL_GPL(blk_update_request);
865 static void __blk_account_io_done(struct request *req, u64 now)
867 const int sgrp = op_stat_group(req_op(req));
870 update_io_ticks(req->part, jiffies, true);
871 part_stat_inc(req->part, ios[sgrp]);
872 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
876 static inline void blk_account_io_done(struct request *req, u64 now)
879 * Account IO completion. flush_rq isn't accounted as a
880 * normal IO on queueing nor completion. Accounting the
881 * containing request is enough.
883 if (blk_do_io_stat(req) && req->part &&
884 !(req->rq_flags & RQF_FLUSH_SEQ))
885 __blk_account_io_done(req, now);
888 static void __blk_account_io_start(struct request *rq)
890 /* passthrough requests can hold bios that do not have ->bi_bdev set */
891 if (rq->bio && rq->bio->bi_bdev)
892 rq->part = rq->bio->bi_bdev;
893 else if (rq->q->disk)
894 rq->part = rq->q->disk->part0;
897 update_io_ticks(rq->part, jiffies, false);
901 static inline void blk_account_io_start(struct request *req)
903 if (blk_do_io_stat(req))
904 __blk_account_io_start(req);
907 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
909 if (rq->rq_flags & RQF_STATS) {
910 blk_mq_poll_stats_start(rq->q);
911 blk_stat_add(rq, now);
914 blk_mq_sched_completed_request(rq, now);
915 blk_account_io_done(rq, now);
918 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
920 if (blk_mq_need_time_stamp(rq))
921 __blk_mq_end_request_acct(rq, ktime_get_ns());
924 rq_qos_done(rq->q, rq);
925 rq->end_io(rq, error);
927 blk_mq_free_request(rq);
930 EXPORT_SYMBOL(__blk_mq_end_request);
932 void blk_mq_end_request(struct request *rq, blk_status_t error)
934 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
936 __blk_mq_end_request(rq, error);
938 EXPORT_SYMBOL(blk_mq_end_request);
940 #define TAG_COMP_BATCH 32
942 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
943 int *tag_array, int nr_tags)
945 struct request_queue *q = hctx->queue;
948 * All requests should have been marked as RQF_MQ_INFLIGHT, so
949 * update hctx->nr_active in batch
951 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
952 __blk_mq_sub_active_requests(hctx, nr_tags);
954 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
955 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
958 void blk_mq_end_request_batch(struct io_comp_batch *iob)
960 int tags[TAG_COMP_BATCH], nr_tags = 0;
961 struct blk_mq_hw_ctx *cur_hctx = NULL;
966 now = ktime_get_ns();
968 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
970 prefetch(rq->rq_next);
972 blk_complete_request(rq);
974 __blk_mq_end_request_acct(rq, now);
976 rq_qos_done(rq->q, rq);
978 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
979 if (!req_ref_put_and_test(rq))
982 blk_crypto_free_request(rq);
983 blk_pm_mark_last_busy(rq);
985 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
987 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
989 cur_hctx = rq->mq_hctx;
991 tags[nr_tags++] = rq->tag;
995 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
997 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
999 static void blk_complete_reqs(struct llist_head *list)
1001 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1002 struct request *rq, *next;
1004 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1005 rq->q->mq_ops->complete(rq);
1008 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1010 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1013 static int blk_softirq_cpu_dead(unsigned int cpu)
1015 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1019 static void __blk_mq_complete_request_remote(void *data)
1021 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1024 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1026 int cpu = raw_smp_processor_id();
1028 if (!IS_ENABLED(CONFIG_SMP) ||
1029 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1032 * With force threaded interrupts enabled, raising softirq from an SMP
1033 * function call will always result in waking the ksoftirqd thread.
1034 * This is probably worse than completing the request on a different
1037 if (force_irqthreads())
1040 /* same CPU or cache domain? Complete locally */
1041 if (cpu == rq->mq_ctx->cpu ||
1042 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1043 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1046 /* don't try to IPI to an offline CPU */
1047 return cpu_online(rq->mq_ctx->cpu);
1050 static void blk_mq_complete_send_ipi(struct request *rq)
1052 struct llist_head *list;
1055 cpu = rq->mq_ctx->cpu;
1056 list = &per_cpu(blk_cpu_done, cpu);
1057 if (llist_add(&rq->ipi_list, list)) {
1058 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1059 smp_call_function_single_async(cpu, &rq->csd);
1063 static void blk_mq_raise_softirq(struct request *rq)
1065 struct llist_head *list;
1068 list = this_cpu_ptr(&blk_cpu_done);
1069 if (llist_add(&rq->ipi_list, list))
1070 raise_softirq(BLOCK_SOFTIRQ);
1074 bool blk_mq_complete_request_remote(struct request *rq)
1076 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1079 * For a polled request, always complete locallly, it's pointless
1080 * to redirect the completion.
1082 if (rq->cmd_flags & REQ_POLLED)
1085 if (blk_mq_complete_need_ipi(rq)) {
1086 blk_mq_complete_send_ipi(rq);
1090 if (rq->q->nr_hw_queues == 1) {
1091 blk_mq_raise_softirq(rq);
1096 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1099 * blk_mq_complete_request - end I/O on a request
1100 * @rq: the request being processed
1103 * Complete a request by scheduling the ->complete_rq operation.
1105 void blk_mq_complete_request(struct request *rq)
1107 if (!blk_mq_complete_request_remote(rq))
1108 rq->q->mq_ops->complete(rq);
1110 EXPORT_SYMBOL(blk_mq_complete_request);
1113 * blk_mq_start_request - Start processing a request
1114 * @rq: Pointer to request to be started
1116 * Function used by device drivers to notify the block layer that a request
1117 * is going to be processed now, so blk layer can do proper initializations
1118 * such as starting the timeout timer.
1120 void blk_mq_start_request(struct request *rq)
1122 struct request_queue *q = rq->q;
1124 trace_block_rq_issue(rq);
1126 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1128 #ifdef CONFIG_BLK_CGROUP
1130 start_time = bio_issue_time(&rq->bio->bi_issue);
1133 start_time = ktime_get_ns();
1134 rq->io_start_time_ns = start_time;
1135 rq->stats_sectors = blk_rq_sectors(rq);
1136 rq->rq_flags |= RQF_STATS;
1137 rq_qos_issue(q, rq);
1140 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1143 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1145 #ifdef CONFIG_BLK_DEV_INTEGRITY
1146 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1147 q->integrity.profile->prepare_fn(rq);
1149 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1150 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1152 EXPORT_SYMBOL(blk_mq_start_request);
1155 * blk_end_sync_rq - executes a completion event on a request
1156 * @rq: request to complete
1157 * @error: end I/O status of the request
1159 static void blk_end_sync_rq(struct request *rq, blk_status_t error)
1161 struct completion *waiting = rq->end_io_data;
1163 rq->end_io_data = (void *)(uintptr_t)error;
1166 * complete last, if this is a stack request the process (and thus
1167 * the rq pointer) could be invalid right after this complete()
1173 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1174 * @rq: request to insert
1175 * @at_head: insert request at head or tail of queue
1176 * @done: I/O completion handler
1179 * Insert a fully prepared request at the back of the I/O scheduler queue
1180 * for execution. Don't wait for completion.
1183 * This function will invoke @done directly if the queue is dead.
1185 void blk_execute_rq_nowait(struct request *rq, bool at_head, rq_end_io_fn *done)
1187 WARN_ON(irqs_disabled());
1188 WARN_ON(!blk_rq_is_passthrough(rq));
1192 blk_account_io_start(rq);
1195 * don't check dying flag for MQ because the request won't
1196 * be reused after dying flag is set
1198 blk_mq_sched_insert_request(rq, at_head, true, false);
1200 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1202 static bool blk_rq_is_poll(struct request *rq)
1206 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1208 if (WARN_ON_ONCE(!rq->bio))
1213 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1216 bio_poll(rq->bio, NULL, 0);
1218 } while (!completion_done(wait));
1222 * blk_execute_rq - insert a request into queue for execution
1223 * @rq: request to insert
1224 * @at_head: insert request at head or tail of queue
1227 * Insert a fully prepared request at the back of the I/O scheduler queue
1228 * for execution and wait for completion.
1229 * Return: The blk_status_t result provided to blk_mq_end_request().
1231 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1233 DECLARE_COMPLETION_ONSTACK(wait);
1234 unsigned long hang_check;
1236 rq->end_io_data = &wait;
1237 blk_execute_rq_nowait(rq, at_head, blk_end_sync_rq);
1239 /* Prevent hang_check timer from firing at us during very long I/O */
1240 hang_check = sysctl_hung_task_timeout_secs;
1242 if (blk_rq_is_poll(rq))
1243 blk_rq_poll_completion(rq, &wait);
1244 else if (hang_check)
1245 while (!wait_for_completion_io_timeout(&wait,
1246 hang_check * (HZ/2)))
1249 wait_for_completion_io(&wait);
1251 return (blk_status_t)(uintptr_t)rq->end_io_data;
1253 EXPORT_SYMBOL(blk_execute_rq);
1255 static void __blk_mq_requeue_request(struct request *rq)
1257 struct request_queue *q = rq->q;
1259 blk_mq_put_driver_tag(rq);
1261 trace_block_rq_requeue(rq);
1262 rq_qos_requeue(q, rq);
1264 if (blk_mq_request_started(rq)) {
1265 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1266 rq->rq_flags &= ~RQF_TIMED_OUT;
1270 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1272 __blk_mq_requeue_request(rq);
1274 /* this request will be re-inserted to io scheduler queue */
1275 blk_mq_sched_requeue_request(rq);
1277 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1279 EXPORT_SYMBOL(blk_mq_requeue_request);
1281 static void blk_mq_requeue_work(struct work_struct *work)
1283 struct request_queue *q =
1284 container_of(work, struct request_queue, requeue_work.work);
1286 struct request *rq, *next;
1288 spin_lock_irq(&q->requeue_lock);
1289 list_splice_init(&q->requeue_list, &rq_list);
1290 spin_unlock_irq(&q->requeue_lock);
1292 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1293 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1296 rq->rq_flags &= ~RQF_SOFTBARRIER;
1297 list_del_init(&rq->queuelist);
1299 * If RQF_DONTPREP, rq has contained some driver specific
1300 * data, so insert it to hctx dispatch list to avoid any
1303 if (rq->rq_flags & RQF_DONTPREP)
1304 blk_mq_request_bypass_insert(rq, false, false);
1306 blk_mq_sched_insert_request(rq, true, false, false);
1309 while (!list_empty(&rq_list)) {
1310 rq = list_entry(rq_list.next, struct request, queuelist);
1311 list_del_init(&rq->queuelist);
1312 blk_mq_sched_insert_request(rq, false, false, false);
1315 blk_mq_run_hw_queues(q, false);
1318 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1319 bool kick_requeue_list)
1321 struct request_queue *q = rq->q;
1322 unsigned long flags;
1325 * We abuse this flag that is otherwise used by the I/O scheduler to
1326 * request head insertion from the workqueue.
1328 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1330 spin_lock_irqsave(&q->requeue_lock, flags);
1332 rq->rq_flags |= RQF_SOFTBARRIER;
1333 list_add(&rq->queuelist, &q->requeue_list);
1335 list_add_tail(&rq->queuelist, &q->requeue_list);
1337 spin_unlock_irqrestore(&q->requeue_lock, flags);
1339 if (kick_requeue_list)
1340 blk_mq_kick_requeue_list(q);
1343 void blk_mq_kick_requeue_list(struct request_queue *q)
1345 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1347 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1349 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1350 unsigned long msecs)
1352 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1353 msecs_to_jiffies(msecs));
1355 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1357 static bool blk_mq_rq_inflight(struct request *rq, void *priv,
1361 * If we find a request that isn't idle we know the queue is busy
1362 * as it's checked in the iter.
1363 * Return false to stop the iteration.
1365 if (blk_mq_request_started(rq)) {
1375 bool blk_mq_queue_inflight(struct request_queue *q)
1379 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1382 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1384 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
1386 req->rq_flags |= RQF_TIMED_OUT;
1387 if (req->q->mq_ops->timeout) {
1388 enum blk_eh_timer_return ret;
1390 ret = req->q->mq_ops->timeout(req, reserved);
1391 if (ret == BLK_EH_DONE)
1393 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1399 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
1401 unsigned long deadline;
1403 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1405 if (rq->rq_flags & RQF_TIMED_OUT)
1408 deadline = READ_ONCE(rq->deadline);
1409 if (time_after_eq(jiffies, deadline))
1414 else if (time_after(*next, deadline))
1419 void blk_mq_put_rq_ref(struct request *rq)
1421 if (is_flush_rq(rq))
1423 else if (req_ref_put_and_test(rq))
1424 __blk_mq_free_request(rq);
1427 static bool blk_mq_check_expired(struct request *rq, void *priv, bool reserved)
1429 unsigned long *next = priv;
1432 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1433 * be reallocated underneath the timeout handler's processing, then
1434 * the expire check is reliable. If the request is not expired, then
1435 * it was completed and reallocated as a new request after returning
1436 * from blk_mq_check_expired().
1438 if (blk_mq_req_expired(rq, next))
1439 blk_mq_rq_timed_out(rq, reserved);
1443 static void blk_mq_timeout_work(struct work_struct *work)
1445 struct request_queue *q =
1446 container_of(work, struct request_queue, timeout_work);
1447 unsigned long next = 0;
1448 struct blk_mq_hw_ctx *hctx;
1451 /* A deadlock might occur if a request is stuck requiring a
1452 * timeout at the same time a queue freeze is waiting
1453 * completion, since the timeout code would not be able to
1454 * acquire the queue reference here.
1456 * That's why we don't use blk_queue_enter here; instead, we use
1457 * percpu_ref_tryget directly, because we need to be able to
1458 * obtain a reference even in the short window between the queue
1459 * starting to freeze, by dropping the first reference in
1460 * blk_freeze_queue_start, and the moment the last request is
1461 * consumed, marked by the instant q_usage_counter reaches
1464 if (!percpu_ref_tryget(&q->q_usage_counter))
1467 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1470 mod_timer(&q->timeout, next);
1473 * Request timeouts are handled as a forward rolling timer. If
1474 * we end up here it means that no requests are pending and
1475 * also that no request has been pending for a while. Mark
1476 * each hctx as idle.
1478 queue_for_each_hw_ctx(q, hctx, i) {
1479 /* the hctx may be unmapped, so check it here */
1480 if (blk_mq_hw_queue_mapped(hctx))
1481 blk_mq_tag_idle(hctx);
1487 struct flush_busy_ctx_data {
1488 struct blk_mq_hw_ctx *hctx;
1489 struct list_head *list;
1492 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1494 struct flush_busy_ctx_data *flush_data = data;
1495 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1496 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1497 enum hctx_type type = hctx->type;
1499 spin_lock(&ctx->lock);
1500 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1501 sbitmap_clear_bit(sb, bitnr);
1502 spin_unlock(&ctx->lock);
1507 * Process software queues that have been marked busy, splicing them
1508 * to the for-dispatch
1510 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1512 struct flush_busy_ctx_data data = {
1517 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1519 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1521 struct dispatch_rq_data {
1522 struct blk_mq_hw_ctx *hctx;
1526 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1529 struct dispatch_rq_data *dispatch_data = data;
1530 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1531 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1532 enum hctx_type type = hctx->type;
1534 spin_lock(&ctx->lock);
1535 if (!list_empty(&ctx->rq_lists[type])) {
1536 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1537 list_del_init(&dispatch_data->rq->queuelist);
1538 if (list_empty(&ctx->rq_lists[type]))
1539 sbitmap_clear_bit(sb, bitnr);
1541 spin_unlock(&ctx->lock);
1543 return !dispatch_data->rq;
1546 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1547 struct blk_mq_ctx *start)
1549 unsigned off = start ? start->index_hw[hctx->type] : 0;
1550 struct dispatch_rq_data data = {
1555 __sbitmap_for_each_set(&hctx->ctx_map, off,
1556 dispatch_rq_from_ctx, &data);
1561 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1563 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1564 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1567 blk_mq_tag_busy(rq->mq_hctx);
1569 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1570 bt = &rq->mq_hctx->tags->breserved_tags;
1573 if (!hctx_may_queue(rq->mq_hctx, bt))
1577 tag = __sbitmap_queue_get(bt);
1578 if (tag == BLK_MQ_NO_TAG)
1581 rq->tag = tag + tag_offset;
1585 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1587 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1590 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1591 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1592 rq->rq_flags |= RQF_MQ_INFLIGHT;
1593 __blk_mq_inc_active_requests(hctx);
1595 hctx->tags->rqs[rq->tag] = rq;
1599 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1600 int flags, void *key)
1602 struct blk_mq_hw_ctx *hctx;
1604 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1606 spin_lock(&hctx->dispatch_wait_lock);
1607 if (!list_empty(&wait->entry)) {
1608 struct sbitmap_queue *sbq;
1610 list_del_init(&wait->entry);
1611 sbq = &hctx->tags->bitmap_tags;
1612 atomic_dec(&sbq->ws_active);
1614 spin_unlock(&hctx->dispatch_wait_lock);
1616 blk_mq_run_hw_queue(hctx, true);
1621 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1622 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1623 * restart. For both cases, take care to check the condition again after
1624 * marking us as waiting.
1626 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1629 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1630 struct wait_queue_head *wq;
1631 wait_queue_entry_t *wait;
1634 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1635 blk_mq_sched_mark_restart_hctx(hctx);
1638 * It's possible that a tag was freed in the window between the
1639 * allocation failure and adding the hardware queue to the wait
1642 * Don't clear RESTART here, someone else could have set it.
1643 * At most this will cost an extra queue run.
1645 return blk_mq_get_driver_tag(rq);
1648 wait = &hctx->dispatch_wait;
1649 if (!list_empty_careful(&wait->entry))
1652 wq = &bt_wait_ptr(sbq, hctx)->wait;
1654 spin_lock_irq(&wq->lock);
1655 spin_lock(&hctx->dispatch_wait_lock);
1656 if (!list_empty(&wait->entry)) {
1657 spin_unlock(&hctx->dispatch_wait_lock);
1658 spin_unlock_irq(&wq->lock);
1662 atomic_inc(&sbq->ws_active);
1663 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1664 __add_wait_queue(wq, wait);
1667 * It's possible that a tag was freed in the window between the
1668 * allocation failure and adding the hardware queue to the wait
1671 ret = blk_mq_get_driver_tag(rq);
1673 spin_unlock(&hctx->dispatch_wait_lock);
1674 spin_unlock_irq(&wq->lock);
1679 * We got a tag, remove ourselves from the wait queue to ensure
1680 * someone else gets the wakeup.
1682 list_del_init(&wait->entry);
1683 atomic_dec(&sbq->ws_active);
1684 spin_unlock(&hctx->dispatch_wait_lock);
1685 spin_unlock_irq(&wq->lock);
1690 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1691 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1693 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1694 * - EWMA is one simple way to compute running average value
1695 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1696 * - take 4 as factor for avoiding to get too small(0) result, and this
1697 * factor doesn't matter because EWMA decreases exponentially
1699 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1703 ewma = hctx->dispatch_busy;
1708 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1710 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1711 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1713 hctx->dispatch_busy = ewma;
1716 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1718 static void blk_mq_handle_dev_resource(struct request *rq,
1719 struct list_head *list)
1721 struct request *next =
1722 list_first_entry_or_null(list, struct request, queuelist);
1725 * If an I/O scheduler has been configured and we got a driver tag for
1726 * the next request already, free it.
1729 blk_mq_put_driver_tag(next);
1731 list_add(&rq->queuelist, list);
1732 __blk_mq_requeue_request(rq);
1735 static void blk_mq_handle_zone_resource(struct request *rq,
1736 struct list_head *zone_list)
1739 * If we end up here it is because we cannot dispatch a request to a
1740 * specific zone due to LLD level zone-write locking or other zone
1741 * related resource not being available. In this case, set the request
1742 * aside in zone_list for retrying it later.
1744 list_add(&rq->queuelist, zone_list);
1745 __blk_mq_requeue_request(rq);
1748 enum prep_dispatch {
1750 PREP_DISPATCH_NO_TAG,
1751 PREP_DISPATCH_NO_BUDGET,
1754 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1757 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1758 int budget_token = -1;
1761 budget_token = blk_mq_get_dispatch_budget(rq->q);
1762 if (budget_token < 0) {
1763 blk_mq_put_driver_tag(rq);
1764 return PREP_DISPATCH_NO_BUDGET;
1766 blk_mq_set_rq_budget_token(rq, budget_token);
1769 if (!blk_mq_get_driver_tag(rq)) {
1771 * The initial allocation attempt failed, so we need to
1772 * rerun the hardware queue when a tag is freed. The
1773 * waitqueue takes care of that. If the queue is run
1774 * before we add this entry back on the dispatch list,
1775 * we'll re-run it below.
1777 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1779 * All budgets not got from this function will be put
1780 * together during handling partial dispatch
1783 blk_mq_put_dispatch_budget(rq->q, budget_token);
1784 return PREP_DISPATCH_NO_TAG;
1788 return PREP_DISPATCH_OK;
1791 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1792 static void blk_mq_release_budgets(struct request_queue *q,
1793 struct list_head *list)
1797 list_for_each_entry(rq, list, queuelist) {
1798 int budget_token = blk_mq_get_rq_budget_token(rq);
1800 if (budget_token >= 0)
1801 blk_mq_put_dispatch_budget(q, budget_token);
1806 * Returns true if we did some work AND can potentially do more.
1808 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1809 unsigned int nr_budgets)
1811 enum prep_dispatch prep;
1812 struct request_queue *q = hctx->queue;
1813 struct request *rq, *nxt;
1815 blk_status_t ret = BLK_STS_OK;
1816 LIST_HEAD(zone_list);
1817 bool needs_resource = false;
1819 if (list_empty(list))
1823 * Now process all the entries, sending them to the driver.
1825 errors = queued = 0;
1827 struct blk_mq_queue_data bd;
1829 rq = list_first_entry(list, struct request, queuelist);
1831 WARN_ON_ONCE(hctx != rq->mq_hctx);
1832 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1833 if (prep != PREP_DISPATCH_OK)
1836 list_del_init(&rq->queuelist);
1841 * Flag last if we have no more requests, or if we have more
1842 * but can't assign a driver tag to it.
1844 if (list_empty(list))
1847 nxt = list_first_entry(list, struct request, queuelist);
1848 bd.last = !blk_mq_get_driver_tag(nxt);
1852 * once the request is queued to lld, no need to cover the
1857 ret = q->mq_ops->queue_rq(hctx, &bd);
1862 case BLK_STS_RESOURCE:
1863 needs_resource = true;
1865 case BLK_STS_DEV_RESOURCE:
1866 blk_mq_handle_dev_resource(rq, list);
1868 case BLK_STS_ZONE_RESOURCE:
1870 * Move the request to zone_list and keep going through
1871 * the dispatch list to find more requests the drive can
1874 blk_mq_handle_zone_resource(rq, &zone_list);
1875 needs_resource = true;
1879 blk_mq_end_request(rq, ret);
1881 } while (!list_empty(list));
1883 if (!list_empty(&zone_list))
1884 list_splice_tail_init(&zone_list, list);
1886 /* If we didn't flush the entire list, we could have told the driver
1887 * there was more coming, but that turned out to be a lie.
1889 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1890 q->mq_ops->commit_rqs(hctx);
1892 * Any items that need requeuing? Stuff them into hctx->dispatch,
1893 * that is where we will continue on next queue run.
1895 if (!list_empty(list)) {
1897 /* For non-shared tags, the RESTART check will suffice */
1898 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1899 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1902 blk_mq_release_budgets(q, list);
1904 spin_lock(&hctx->lock);
1905 list_splice_tail_init(list, &hctx->dispatch);
1906 spin_unlock(&hctx->lock);
1909 * Order adding requests to hctx->dispatch and checking
1910 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1911 * in blk_mq_sched_restart(). Avoid restart code path to
1912 * miss the new added requests to hctx->dispatch, meantime
1913 * SCHED_RESTART is observed here.
1918 * If SCHED_RESTART was set by the caller of this function and
1919 * it is no longer set that means that it was cleared by another
1920 * thread and hence that a queue rerun is needed.
1922 * If 'no_tag' is set, that means that we failed getting
1923 * a driver tag with an I/O scheduler attached. If our dispatch
1924 * waitqueue is no longer active, ensure that we run the queue
1925 * AFTER adding our entries back to the list.
1927 * If no I/O scheduler has been configured it is possible that
1928 * the hardware queue got stopped and restarted before requests
1929 * were pushed back onto the dispatch list. Rerun the queue to
1930 * avoid starvation. Notes:
1931 * - blk_mq_run_hw_queue() checks whether or not a queue has
1932 * been stopped before rerunning a queue.
1933 * - Some but not all block drivers stop a queue before
1934 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1937 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1938 * bit is set, run queue after a delay to avoid IO stalls
1939 * that could otherwise occur if the queue is idle. We'll do
1940 * similar if we couldn't get budget or couldn't lock a zone
1941 * and SCHED_RESTART is set.
1943 needs_restart = blk_mq_sched_needs_restart(hctx);
1944 if (prep == PREP_DISPATCH_NO_BUDGET)
1945 needs_resource = true;
1946 if (!needs_restart ||
1947 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1948 blk_mq_run_hw_queue(hctx, true);
1949 else if (needs_restart && needs_resource)
1950 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1952 blk_mq_update_dispatch_busy(hctx, true);
1955 blk_mq_update_dispatch_busy(hctx, false);
1957 return (queued + errors) != 0;
1961 * __blk_mq_run_hw_queue - Run a hardware queue.
1962 * @hctx: Pointer to the hardware queue to run.
1964 * Send pending requests to the hardware.
1966 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1969 * We can't run the queue inline with ints disabled. Ensure that
1970 * we catch bad users of this early.
1972 WARN_ON_ONCE(in_interrupt());
1974 blk_mq_run_dispatch_ops(hctx->queue,
1975 blk_mq_sched_dispatch_requests(hctx));
1978 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1980 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1982 if (cpu >= nr_cpu_ids)
1983 cpu = cpumask_first(hctx->cpumask);
1988 * It'd be great if the workqueue API had a way to pass
1989 * in a mask and had some smarts for more clever placement.
1990 * For now we just round-robin here, switching for every
1991 * BLK_MQ_CPU_WORK_BATCH queued items.
1993 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1996 int next_cpu = hctx->next_cpu;
1998 if (hctx->queue->nr_hw_queues == 1)
1999 return WORK_CPU_UNBOUND;
2001 if (--hctx->next_cpu_batch <= 0) {
2003 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2005 if (next_cpu >= nr_cpu_ids)
2006 next_cpu = blk_mq_first_mapped_cpu(hctx);
2007 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2011 * Do unbound schedule if we can't find a online CPU for this hctx,
2012 * and it should only happen in the path of handling CPU DEAD.
2014 if (!cpu_online(next_cpu)) {
2021 * Make sure to re-select CPU next time once after CPUs
2022 * in hctx->cpumask become online again.
2024 hctx->next_cpu = next_cpu;
2025 hctx->next_cpu_batch = 1;
2026 return WORK_CPU_UNBOUND;
2029 hctx->next_cpu = next_cpu;
2034 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2035 * @hctx: Pointer to the hardware queue to run.
2036 * @async: If we want to run the queue asynchronously.
2037 * @msecs: Milliseconds of delay to wait before running the queue.
2039 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2040 * with a delay of @msecs.
2042 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2043 unsigned long msecs)
2045 if (unlikely(blk_mq_hctx_stopped(hctx)))
2048 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2049 int cpu = get_cpu();
2050 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
2051 __blk_mq_run_hw_queue(hctx);
2059 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2060 msecs_to_jiffies(msecs));
2064 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2065 * @hctx: Pointer to the hardware queue to run.
2066 * @msecs: Milliseconds of delay to wait before running the queue.
2068 * Run a hardware queue asynchronously with a delay of @msecs.
2070 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2072 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2074 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2077 * blk_mq_run_hw_queue - Start to run a hardware queue.
2078 * @hctx: Pointer to the hardware queue to run.
2079 * @async: If we want to run the queue asynchronously.
2081 * Check if the request queue is not in a quiesced state and if there are
2082 * pending requests to be sent. If this is true, run the queue to send requests
2085 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2090 * When queue is quiesced, we may be switching io scheduler, or
2091 * updating nr_hw_queues, or other things, and we can't run queue
2092 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2094 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2097 __blk_mq_run_dispatch_ops(hctx->queue, false,
2098 need_run = !blk_queue_quiesced(hctx->queue) &&
2099 blk_mq_hctx_has_pending(hctx));
2102 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2104 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2107 * Is the request queue handled by an IO scheduler that does not respect
2108 * hardware queues when dispatching?
2110 static bool blk_mq_has_sqsched(struct request_queue *q)
2112 struct elevator_queue *e = q->elevator;
2114 if (e && e->type->ops.dispatch_request &&
2115 !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE))
2121 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2124 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2126 struct blk_mq_hw_ctx *hctx;
2129 * If the IO scheduler does not respect hardware queues when
2130 * dispatching, we just don't bother with multiple HW queues and
2131 * dispatch from hctx for the current CPU since running multiple queues
2132 * just causes lock contention inside the scheduler and pointless cache
2135 hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT,
2136 raw_smp_processor_id());
2137 if (!blk_mq_hctx_stopped(hctx))
2143 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2144 * @q: Pointer to the request queue to run.
2145 * @async: If we want to run the queue asynchronously.
2147 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2149 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2153 if (blk_mq_has_sqsched(q))
2154 sq_hctx = blk_mq_get_sq_hctx(q);
2155 queue_for_each_hw_ctx(q, hctx, i) {
2156 if (blk_mq_hctx_stopped(hctx))
2159 * Dispatch from this hctx either if there's no hctx preferred
2160 * by IO scheduler or if it has requests that bypass the
2163 if (!sq_hctx || sq_hctx == hctx ||
2164 !list_empty_careful(&hctx->dispatch))
2165 blk_mq_run_hw_queue(hctx, async);
2168 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2171 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2172 * @q: Pointer to the request queue to run.
2173 * @msecs: Milliseconds of delay to wait before running the queues.
2175 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2177 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2181 if (blk_mq_has_sqsched(q))
2182 sq_hctx = blk_mq_get_sq_hctx(q);
2183 queue_for_each_hw_ctx(q, hctx, i) {
2184 if (blk_mq_hctx_stopped(hctx))
2187 * If there is already a run_work pending, leave the
2188 * pending delay untouched. Otherwise, a hctx can stall
2189 * if another hctx is re-delaying the other's work
2190 * before the work executes.
2192 if (delayed_work_pending(&hctx->run_work))
2195 * Dispatch from this hctx either if there's no hctx preferred
2196 * by IO scheduler or if it has requests that bypass the
2199 if (!sq_hctx || sq_hctx == hctx ||
2200 !list_empty_careful(&hctx->dispatch))
2201 blk_mq_delay_run_hw_queue(hctx, msecs);
2204 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2207 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
2208 * @q: request queue.
2210 * The caller is responsible for serializing this function against
2211 * blk_mq_{start,stop}_hw_queue().
2213 bool blk_mq_queue_stopped(struct request_queue *q)
2215 struct blk_mq_hw_ctx *hctx;
2218 queue_for_each_hw_ctx(q, hctx, i)
2219 if (blk_mq_hctx_stopped(hctx))
2224 EXPORT_SYMBOL(blk_mq_queue_stopped);
2227 * This function is often used for pausing .queue_rq() by driver when
2228 * there isn't enough resource or some conditions aren't satisfied, and
2229 * BLK_STS_RESOURCE is usually returned.
2231 * We do not guarantee that dispatch can be drained or blocked
2232 * after blk_mq_stop_hw_queue() returns. Please use
2233 * blk_mq_quiesce_queue() for that requirement.
2235 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2237 cancel_delayed_work(&hctx->run_work);
2239 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2241 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2244 * This function is often used for pausing .queue_rq() by driver when
2245 * there isn't enough resource or some conditions aren't satisfied, and
2246 * BLK_STS_RESOURCE is usually returned.
2248 * We do not guarantee that dispatch can be drained or blocked
2249 * after blk_mq_stop_hw_queues() returns. Please use
2250 * blk_mq_quiesce_queue() for that requirement.
2252 void blk_mq_stop_hw_queues(struct request_queue *q)
2254 struct blk_mq_hw_ctx *hctx;
2257 queue_for_each_hw_ctx(q, hctx, i)
2258 blk_mq_stop_hw_queue(hctx);
2260 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2262 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2264 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2266 blk_mq_run_hw_queue(hctx, false);
2268 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2270 void blk_mq_start_hw_queues(struct request_queue *q)
2272 struct blk_mq_hw_ctx *hctx;
2275 queue_for_each_hw_ctx(q, hctx, i)
2276 blk_mq_start_hw_queue(hctx);
2278 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2280 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2282 if (!blk_mq_hctx_stopped(hctx))
2285 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2286 blk_mq_run_hw_queue(hctx, async);
2288 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2290 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2292 struct blk_mq_hw_ctx *hctx;
2295 queue_for_each_hw_ctx(q, hctx, i)
2296 blk_mq_start_stopped_hw_queue(hctx, async);
2298 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2300 static void blk_mq_run_work_fn(struct work_struct *work)
2302 struct blk_mq_hw_ctx *hctx;
2304 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2307 * If we are stopped, don't run the queue.
2309 if (blk_mq_hctx_stopped(hctx))
2312 __blk_mq_run_hw_queue(hctx);
2315 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2319 struct blk_mq_ctx *ctx = rq->mq_ctx;
2320 enum hctx_type type = hctx->type;
2322 lockdep_assert_held(&ctx->lock);
2324 trace_block_rq_insert(rq);
2327 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2329 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2332 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2335 struct blk_mq_ctx *ctx = rq->mq_ctx;
2337 lockdep_assert_held(&ctx->lock);
2339 __blk_mq_insert_req_list(hctx, rq, at_head);
2340 blk_mq_hctx_mark_pending(hctx, ctx);
2344 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2345 * @rq: Pointer to request to be inserted.
2346 * @at_head: true if the request should be inserted at the head of the list.
2347 * @run_queue: If we should run the hardware queue after inserting the request.
2349 * Should only be used carefully, when the caller knows we want to
2350 * bypass a potential IO scheduler on the target device.
2352 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2355 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2357 spin_lock(&hctx->lock);
2359 list_add(&rq->queuelist, &hctx->dispatch);
2361 list_add_tail(&rq->queuelist, &hctx->dispatch);
2362 spin_unlock(&hctx->lock);
2365 blk_mq_run_hw_queue(hctx, false);
2368 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2369 struct list_head *list)
2373 enum hctx_type type = hctx->type;
2376 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2379 list_for_each_entry(rq, list, queuelist) {
2380 BUG_ON(rq->mq_ctx != ctx);
2381 trace_block_rq_insert(rq);
2384 spin_lock(&ctx->lock);
2385 list_splice_tail_init(list, &ctx->rq_lists[type]);
2386 blk_mq_hctx_mark_pending(hctx, ctx);
2387 spin_unlock(&ctx->lock);
2390 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2393 if (hctx->queue->mq_ops->commit_rqs) {
2394 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2395 hctx->queue->mq_ops->commit_rqs(hctx);
2400 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2401 unsigned int nr_segs)
2405 if (bio->bi_opf & REQ_RAHEAD)
2406 rq->cmd_flags |= REQ_FAILFAST_MASK;
2408 rq->__sector = bio->bi_iter.bi_sector;
2409 blk_rq_bio_prep(rq, bio, nr_segs);
2411 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2412 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2415 blk_account_io_start(rq);
2418 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2419 struct request *rq, bool last)
2421 struct request_queue *q = rq->q;
2422 struct blk_mq_queue_data bd = {
2429 * For OK queue, we are done. For error, caller may kill it.
2430 * Any other error (busy), just add it to our list as we
2431 * previously would have done.
2433 ret = q->mq_ops->queue_rq(hctx, &bd);
2436 blk_mq_update_dispatch_busy(hctx, false);
2438 case BLK_STS_RESOURCE:
2439 case BLK_STS_DEV_RESOURCE:
2440 blk_mq_update_dispatch_busy(hctx, true);
2441 __blk_mq_requeue_request(rq);
2444 blk_mq_update_dispatch_busy(hctx, false);
2451 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2453 bool bypass_insert, bool last)
2455 struct request_queue *q = rq->q;
2456 bool run_queue = true;
2460 * RCU or SRCU read lock is needed before checking quiesced flag.
2462 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2463 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2464 * and avoid driver to try to dispatch again.
2466 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2468 bypass_insert = false;
2472 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2475 budget_token = blk_mq_get_dispatch_budget(q);
2476 if (budget_token < 0)
2479 blk_mq_set_rq_budget_token(rq, budget_token);
2481 if (!blk_mq_get_driver_tag(rq)) {
2482 blk_mq_put_dispatch_budget(q, budget_token);
2486 return __blk_mq_issue_directly(hctx, rq, last);
2489 return BLK_STS_RESOURCE;
2491 blk_mq_sched_insert_request(rq, false, run_queue, false);
2497 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2498 * @hctx: Pointer of the associated hardware queue.
2499 * @rq: Pointer to request to be sent.
2501 * If the device has enough resources to accept a new request now, send the
2502 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2503 * we can try send it another time in the future. Requests inserted at this
2504 * queue have higher priority.
2506 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2510 __blk_mq_try_issue_directly(hctx, rq, false, true);
2512 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2513 blk_mq_request_bypass_insert(rq, false, true);
2514 else if (ret != BLK_STS_OK)
2515 blk_mq_end_request(rq, ret);
2518 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2520 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2523 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2525 struct blk_mq_hw_ctx *hctx = NULL;
2530 while ((rq = rq_list_pop(&plug->mq_list))) {
2531 bool last = rq_list_empty(plug->mq_list);
2534 if (hctx != rq->mq_hctx) {
2536 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2540 ret = blk_mq_request_issue_directly(rq, last);
2545 case BLK_STS_RESOURCE:
2546 case BLK_STS_DEV_RESOURCE:
2547 blk_mq_request_bypass_insert(rq, false, last);
2548 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2551 blk_mq_end_request(rq, ret);
2558 * If we didn't flush the entire list, we could have told the driver
2559 * there was more coming, but that turned out to be a lie.
2562 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2565 static void __blk_mq_flush_plug_list(struct request_queue *q,
2566 struct blk_plug *plug)
2568 if (blk_queue_quiesced(q))
2570 q->mq_ops->queue_rqs(&plug->mq_list);
2573 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2575 struct blk_mq_hw_ctx *this_hctx;
2576 struct blk_mq_ctx *this_ctx;
2581 if (rq_list_empty(plug->mq_list))
2585 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2586 struct request_queue *q;
2588 rq = rq_list_peek(&plug->mq_list);
2592 * Peek first request and see if we have a ->queue_rqs() hook.
2593 * If we do, we can dispatch the whole plug list in one go. We
2594 * already know at this point that all requests belong to the
2595 * same queue, caller must ensure that's the case.
2597 * Since we pass off the full list to the driver at this point,
2598 * we do not increment the active request count for the queue.
2599 * Bypass shared tags for now because of that.
2601 if (q->mq_ops->queue_rqs &&
2602 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2603 blk_mq_run_dispatch_ops(q,
2604 __blk_mq_flush_plug_list(q, plug));
2605 if (rq_list_empty(plug->mq_list))
2609 blk_mq_run_dispatch_ops(q,
2610 blk_mq_plug_issue_direct(plug, false));
2611 if (rq_list_empty(plug->mq_list))
2619 rq = rq_list_pop(&plug->mq_list);
2622 this_hctx = rq->mq_hctx;
2623 this_ctx = rq->mq_ctx;
2624 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2625 trace_block_unplug(this_hctx->queue, depth,
2627 blk_mq_sched_insert_requests(this_hctx, this_ctx,
2628 &list, from_schedule);
2630 this_hctx = rq->mq_hctx;
2631 this_ctx = rq->mq_ctx;
2635 list_add(&rq->queuelist, &list);
2637 } while (!rq_list_empty(plug->mq_list));
2639 if (!list_empty(&list)) {
2640 trace_block_unplug(this_hctx->queue, depth, !from_schedule);
2641 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list,
2646 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2647 struct list_head *list)
2652 while (!list_empty(list)) {
2654 struct request *rq = list_first_entry(list, struct request,
2657 list_del_init(&rq->queuelist);
2658 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2659 if (ret != BLK_STS_OK) {
2660 if (ret == BLK_STS_RESOURCE ||
2661 ret == BLK_STS_DEV_RESOURCE) {
2662 blk_mq_request_bypass_insert(rq, false,
2666 blk_mq_end_request(rq, ret);
2673 * If we didn't flush the entire list, we could have told
2674 * the driver there was more coming, but that turned out to
2677 if ((!list_empty(list) || errors) &&
2678 hctx->queue->mq_ops->commit_rqs && queued)
2679 hctx->queue->mq_ops->commit_rqs(hctx);
2683 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
2684 * queues. This is important for md arrays to benefit from merging
2687 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
2689 if (plug->multiple_queues)
2690 return BLK_MAX_REQUEST_COUNT * 2;
2691 return BLK_MAX_REQUEST_COUNT;
2694 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2696 struct request *last = rq_list_peek(&plug->mq_list);
2698 if (!plug->rq_count) {
2699 trace_block_plug(rq->q);
2700 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
2701 (!blk_queue_nomerges(rq->q) &&
2702 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2703 blk_mq_flush_plug_list(plug, false);
2704 trace_block_plug(rq->q);
2707 if (!plug->multiple_queues && last && last->q != rq->q)
2708 plug->multiple_queues = true;
2709 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
2710 plug->has_elevator = true;
2712 rq_list_add(&plug->mq_list, rq);
2716 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2717 struct bio *bio, unsigned int nr_segs)
2719 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2720 if (blk_attempt_plug_merge(q, bio, nr_segs))
2722 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2728 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2729 struct blk_plug *plug,
2732 struct blk_mq_alloc_data data = {
2735 .cmd_flags = bio->bi_opf,
2739 if (unlikely(bio_queue_enter(bio)))
2743 data.nr_tags = plug->nr_ios;
2745 data.cached_rq = &plug->cached_rq;
2748 rq = __blk_mq_alloc_requests(&data);
2751 rq_qos_cleanup(q, bio);
2752 if (bio->bi_opf & REQ_NOWAIT)
2753 bio_wouldblock_error(bio);
2758 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2759 struct blk_plug *plug, struct bio *bio)
2765 rq = rq_list_peek(&plug->cached_rq);
2766 if (!rq || rq->q != q)
2769 if (blk_mq_get_hctx_type(bio->bi_opf) != rq->mq_hctx->type)
2771 if (op_is_flush(rq->cmd_flags) != op_is_flush(bio->bi_opf))
2774 rq->cmd_flags = bio->bi_opf;
2775 plug->cached_rq = rq_list_next(rq);
2776 INIT_LIST_HEAD(&rq->queuelist);
2781 * blk_mq_submit_bio - Create and send a request to block device.
2782 * @bio: Bio pointer.
2784 * Builds up a request structure from @q and @bio and send to the device. The
2785 * request may not be queued directly to hardware if:
2786 * * This request can be merged with another one
2787 * * We want to place request at plug queue for possible future merging
2788 * * There is an IO scheduler active at this queue
2790 * It will not queue the request if there is an error with the bio, or at the
2793 void blk_mq_submit_bio(struct bio *bio)
2795 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2796 struct blk_plug *plug = blk_mq_plug(q, bio);
2797 const int is_sync = op_is_sync(bio->bi_opf);
2799 unsigned int nr_segs = 1;
2802 blk_queue_bounce(q, &bio);
2803 if (blk_may_split(q, bio))
2804 __blk_queue_split(q, &bio, &nr_segs);
2806 if (!bio_integrity_prep(bio))
2809 if (blk_mq_attempt_bio_merge(q, bio, nr_segs))
2812 rq_qos_throttle(q, bio);
2814 rq = blk_mq_get_cached_request(q, plug, bio);
2816 rq = blk_mq_get_new_requests(q, plug, bio);
2821 trace_block_getrq(bio);
2823 rq_qos_track(q, rq, bio);
2825 blk_mq_bio_to_request(rq, bio, nr_segs);
2827 ret = blk_crypto_init_request(rq);
2828 if (ret != BLK_STS_OK) {
2829 bio->bi_status = ret;
2831 blk_mq_free_request(rq);
2835 if (op_is_flush(bio->bi_opf)) {
2836 blk_insert_flush(rq);
2841 blk_add_rq_to_plug(plug, rq);
2842 else if ((rq->rq_flags & RQF_ELV) ||
2843 (rq->mq_hctx->dispatch_busy &&
2844 (q->nr_hw_queues == 1 || !is_sync)))
2845 blk_mq_sched_insert_request(rq, false, true, true);
2847 blk_mq_run_dispatch_ops(rq->q,
2848 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2851 #ifdef CONFIG_BLK_MQ_STACKING
2853 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2854 * @rq: the request being queued
2856 blk_status_t blk_insert_cloned_request(struct request *rq)
2858 struct request_queue *q = rq->q;
2859 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
2862 if (blk_rq_sectors(rq) > max_sectors) {
2864 * SCSI device does not have a good way to return if
2865 * Write Same/Zero is actually supported. If a device rejects
2866 * a non-read/write command (discard, write same,etc.) the
2867 * low-level device driver will set the relevant queue limit to
2868 * 0 to prevent blk-lib from issuing more of the offending
2869 * operations. Commands queued prior to the queue limit being
2870 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
2871 * errors being propagated to upper layers.
2873 if (max_sectors == 0)
2874 return BLK_STS_NOTSUPP;
2876 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
2877 __func__, blk_rq_sectors(rq), max_sectors);
2878 return BLK_STS_IOERR;
2882 * The queue settings related to segment counting may differ from the
2885 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
2886 if (rq->nr_phys_segments > queue_max_segments(q)) {
2887 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
2888 __func__, rq->nr_phys_segments, queue_max_segments(q));
2889 return BLK_STS_IOERR;
2892 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
2893 return BLK_STS_IOERR;
2895 if (blk_crypto_insert_cloned_request(rq))
2896 return BLK_STS_IOERR;
2898 blk_account_io_start(rq);
2901 * Since we have a scheduler attached on the top device,
2902 * bypass a potential scheduler on the bottom device for
2905 blk_mq_run_dispatch_ops(q,
2906 ret = blk_mq_request_issue_directly(rq, true));
2908 blk_account_io_done(rq, ktime_get_ns());
2911 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
2914 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
2915 * @rq: the clone request to be cleaned up
2918 * Free all bios in @rq for a cloned request.
2920 void blk_rq_unprep_clone(struct request *rq)
2924 while ((bio = rq->bio) != NULL) {
2925 rq->bio = bio->bi_next;
2930 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
2933 * blk_rq_prep_clone - Helper function to setup clone request
2934 * @rq: the request to be setup
2935 * @rq_src: original request to be cloned
2936 * @bs: bio_set that bios for clone are allocated from
2937 * @gfp_mask: memory allocation mask for bio
2938 * @bio_ctr: setup function to be called for each clone bio.
2939 * Returns %0 for success, non %0 for failure.
2940 * @data: private data to be passed to @bio_ctr
2943 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
2944 * Also, pages which the original bios are pointing to are not copied
2945 * and the cloned bios just point same pages.
2946 * So cloned bios must be completed before original bios, which means
2947 * the caller must complete @rq before @rq_src.
2949 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
2950 struct bio_set *bs, gfp_t gfp_mask,
2951 int (*bio_ctr)(struct bio *, struct bio *, void *),
2954 struct bio *bio, *bio_src;
2959 __rq_for_each_bio(bio_src, rq_src) {
2960 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
2965 if (bio_ctr && bio_ctr(bio, bio_src, data))
2969 rq->biotail->bi_next = bio;
2972 rq->bio = rq->biotail = bio;
2977 /* Copy attributes of the original request to the clone request. */
2978 rq->__sector = blk_rq_pos(rq_src);
2979 rq->__data_len = blk_rq_bytes(rq_src);
2980 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
2981 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
2982 rq->special_vec = rq_src->special_vec;
2984 rq->nr_phys_segments = rq_src->nr_phys_segments;
2985 rq->ioprio = rq_src->ioprio;
2987 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
2995 blk_rq_unprep_clone(rq);
2999 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3000 #endif /* CONFIG_BLK_MQ_STACKING */
3003 * Steal bios from a request and add them to a bio list.
3004 * The request must not have been partially completed before.
3006 void blk_steal_bios(struct bio_list *list, struct request *rq)
3010 list->tail->bi_next = rq->bio;
3012 list->head = rq->bio;
3013 list->tail = rq->biotail;
3021 EXPORT_SYMBOL_GPL(blk_steal_bios);
3023 static size_t order_to_size(unsigned int order)
3025 return (size_t)PAGE_SIZE << order;
3028 /* called before freeing request pool in @tags */
3029 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3030 struct blk_mq_tags *tags)
3033 unsigned long flags;
3035 /* There is no need to clear a driver tags own mapping */
3036 if (drv_tags == tags)
3039 list_for_each_entry(page, &tags->page_list, lru) {
3040 unsigned long start = (unsigned long)page_address(page);
3041 unsigned long end = start + order_to_size(page->private);
3044 for (i = 0; i < drv_tags->nr_tags; i++) {
3045 struct request *rq = drv_tags->rqs[i];
3046 unsigned long rq_addr = (unsigned long)rq;
3048 if (rq_addr >= start && rq_addr < end) {
3049 WARN_ON_ONCE(req_ref_read(rq) != 0);
3050 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3056 * Wait until all pending iteration is done.
3058 * Request reference is cleared and it is guaranteed to be observed
3059 * after the ->lock is released.
3061 spin_lock_irqsave(&drv_tags->lock, flags);
3062 spin_unlock_irqrestore(&drv_tags->lock, flags);
3065 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3066 unsigned int hctx_idx)
3068 struct blk_mq_tags *drv_tags;
3071 if (blk_mq_is_shared_tags(set->flags))
3072 drv_tags = set->shared_tags;
3074 drv_tags = set->tags[hctx_idx];
3076 if (tags->static_rqs && set->ops->exit_request) {
3079 for (i = 0; i < tags->nr_tags; i++) {
3080 struct request *rq = tags->static_rqs[i];
3084 set->ops->exit_request(set, rq, hctx_idx);
3085 tags->static_rqs[i] = NULL;
3089 blk_mq_clear_rq_mapping(drv_tags, tags);
3091 while (!list_empty(&tags->page_list)) {
3092 page = list_first_entry(&tags->page_list, struct page, lru);
3093 list_del_init(&page->lru);
3095 * Remove kmemleak object previously allocated in
3096 * blk_mq_alloc_rqs().
3098 kmemleak_free(page_address(page));
3099 __free_pages(page, page->private);
3103 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3107 kfree(tags->static_rqs);
3108 tags->static_rqs = NULL;
3110 blk_mq_free_tags(tags);
3113 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3114 unsigned int hctx_idx,
3115 unsigned int nr_tags,
3116 unsigned int reserved_tags)
3118 struct blk_mq_tags *tags;
3121 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
3122 if (node == NUMA_NO_NODE)
3123 node = set->numa_node;
3125 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3126 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3130 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3131 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3134 blk_mq_free_tags(tags);
3138 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3139 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3141 if (!tags->static_rqs) {
3143 blk_mq_free_tags(tags);
3150 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3151 unsigned int hctx_idx, int node)
3155 if (set->ops->init_request) {
3156 ret = set->ops->init_request(set, rq, hctx_idx, node);
3161 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3165 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3166 struct blk_mq_tags *tags,
3167 unsigned int hctx_idx, unsigned int depth)
3169 unsigned int i, j, entries_per_page, max_order = 4;
3170 size_t rq_size, left;
3173 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
3174 if (node == NUMA_NO_NODE)
3175 node = set->numa_node;
3177 INIT_LIST_HEAD(&tags->page_list);
3180 * rq_size is the size of the request plus driver payload, rounded
3181 * to the cacheline size
3183 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3185 left = rq_size * depth;
3187 for (i = 0; i < depth; ) {
3188 int this_order = max_order;
3193 while (this_order && left < order_to_size(this_order - 1))
3197 page = alloc_pages_node(node,
3198 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3204 if (order_to_size(this_order) < rq_size)
3211 page->private = this_order;
3212 list_add_tail(&page->lru, &tags->page_list);
3214 p = page_address(page);
3216 * Allow kmemleak to scan these pages as they contain pointers
3217 * to additional allocations like via ops->init_request().
3219 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3220 entries_per_page = order_to_size(this_order) / rq_size;
3221 to_do = min(entries_per_page, depth - i);
3222 left -= to_do * rq_size;
3223 for (j = 0; j < to_do; j++) {
3224 struct request *rq = p;
3226 tags->static_rqs[i] = rq;
3227 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3228 tags->static_rqs[i] = NULL;
3239 blk_mq_free_rqs(set, tags, hctx_idx);
3243 struct rq_iter_data {
3244 struct blk_mq_hw_ctx *hctx;
3248 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
3250 struct rq_iter_data *iter_data = data;
3252 if (rq->mq_hctx != iter_data->hctx)
3254 iter_data->has_rq = true;
3258 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3260 struct blk_mq_tags *tags = hctx->sched_tags ?
3261 hctx->sched_tags : hctx->tags;
3262 struct rq_iter_data data = {
3266 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3270 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3271 struct blk_mq_hw_ctx *hctx)
3273 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3275 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3280 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3282 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3283 struct blk_mq_hw_ctx, cpuhp_online);
3285 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3286 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3290 * Prevent new request from being allocated on the current hctx.
3292 * The smp_mb__after_atomic() Pairs with the implied barrier in
3293 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3294 * seen once we return from the tag allocator.
3296 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3297 smp_mb__after_atomic();
3300 * Try to grab a reference to the queue and wait for any outstanding
3301 * requests. If we could not grab a reference the queue has been
3302 * frozen and there are no requests.
3304 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3305 while (blk_mq_hctx_has_requests(hctx))
3307 percpu_ref_put(&hctx->queue->q_usage_counter);
3313 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3315 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3316 struct blk_mq_hw_ctx, cpuhp_online);
3318 if (cpumask_test_cpu(cpu, hctx->cpumask))
3319 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3324 * 'cpu' is going away. splice any existing rq_list entries from this
3325 * software queue to the hw queue dispatch list, and ensure that it
3328 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3330 struct blk_mq_hw_ctx *hctx;
3331 struct blk_mq_ctx *ctx;
3333 enum hctx_type type;
3335 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3336 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3339 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3342 spin_lock(&ctx->lock);
3343 if (!list_empty(&ctx->rq_lists[type])) {
3344 list_splice_init(&ctx->rq_lists[type], &tmp);
3345 blk_mq_hctx_clear_pending(hctx, ctx);
3347 spin_unlock(&ctx->lock);
3349 if (list_empty(&tmp))
3352 spin_lock(&hctx->lock);
3353 list_splice_tail_init(&tmp, &hctx->dispatch);
3354 spin_unlock(&hctx->lock);
3356 blk_mq_run_hw_queue(hctx, true);
3360 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3362 if (!(hctx->flags & BLK_MQ_F_STACKING))
3363 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3364 &hctx->cpuhp_online);
3365 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3370 * Before freeing hw queue, clearing the flush request reference in
3371 * tags->rqs[] for avoiding potential UAF.
3373 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3374 unsigned int queue_depth, struct request *flush_rq)
3377 unsigned long flags;
3379 /* The hw queue may not be mapped yet */
3383 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3385 for (i = 0; i < queue_depth; i++)
3386 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3389 * Wait until all pending iteration is done.
3391 * Request reference is cleared and it is guaranteed to be observed
3392 * after the ->lock is released.
3394 spin_lock_irqsave(&tags->lock, flags);
3395 spin_unlock_irqrestore(&tags->lock, flags);
3398 /* hctx->ctxs will be freed in queue's release handler */
3399 static void blk_mq_exit_hctx(struct request_queue *q,
3400 struct blk_mq_tag_set *set,
3401 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3403 struct request *flush_rq = hctx->fq->flush_rq;
3405 if (blk_mq_hw_queue_mapped(hctx))
3406 blk_mq_tag_idle(hctx);
3408 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3409 set->queue_depth, flush_rq);
3410 if (set->ops->exit_request)
3411 set->ops->exit_request(set, flush_rq, hctx_idx);
3413 if (set->ops->exit_hctx)
3414 set->ops->exit_hctx(hctx, hctx_idx);
3416 blk_mq_remove_cpuhp(hctx);
3418 spin_lock(&q->unused_hctx_lock);
3419 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3420 spin_unlock(&q->unused_hctx_lock);
3423 static void blk_mq_exit_hw_queues(struct request_queue *q,
3424 struct blk_mq_tag_set *set, int nr_queue)
3426 struct blk_mq_hw_ctx *hctx;
3429 queue_for_each_hw_ctx(q, hctx, i) {
3432 blk_mq_debugfs_unregister_hctx(hctx);
3433 blk_mq_exit_hctx(q, set, hctx, i);
3437 static int blk_mq_init_hctx(struct request_queue *q,
3438 struct blk_mq_tag_set *set,
3439 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3441 hctx->queue_num = hctx_idx;
3443 if (!(hctx->flags & BLK_MQ_F_STACKING))
3444 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3445 &hctx->cpuhp_online);
3446 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3448 hctx->tags = set->tags[hctx_idx];
3450 if (set->ops->init_hctx &&
3451 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3452 goto unregister_cpu_notifier;
3454 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3460 if (set->ops->exit_hctx)
3461 set->ops->exit_hctx(hctx, hctx_idx);
3462 unregister_cpu_notifier:
3463 blk_mq_remove_cpuhp(hctx);
3467 static struct blk_mq_hw_ctx *
3468 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3471 struct blk_mq_hw_ctx *hctx;
3472 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3474 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3476 goto fail_alloc_hctx;
3478 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3481 atomic_set(&hctx->nr_active, 0);
3482 if (node == NUMA_NO_NODE)
3483 node = set->numa_node;
3484 hctx->numa_node = node;
3486 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3487 spin_lock_init(&hctx->lock);
3488 INIT_LIST_HEAD(&hctx->dispatch);
3490 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3492 INIT_LIST_HEAD(&hctx->hctx_list);
3495 * Allocate space for all possible cpus to avoid allocation at
3498 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3503 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3504 gfp, node, false, false))
3508 spin_lock_init(&hctx->dispatch_wait_lock);
3509 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3510 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3512 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3516 blk_mq_hctx_kobj_init(hctx);
3521 sbitmap_free(&hctx->ctx_map);
3525 free_cpumask_var(hctx->cpumask);
3532 static void blk_mq_init_cpu_queues(struct request_queue *q,
3533 unsigned int nr_hw_queues)
3535 struct blk_mq_tag_set *set = q->tag_set;
3538 for_each_possible_cpu(i) {
3539 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3540 struct blk_mq_hw_ctx *hctx;
3544 spin_lock_init(&__ctx->lock);
3545 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3546 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3551 * Set local node, IFF we have more than one hw queue. If
3552 * not, we remain on the home node of the device
3554 for (j = 0; j < set->nr_maps; j++) {
3555 hctx = blk_mq_map_queue_type(q, j, i);
3556 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3557 hctx->numa_node = cpu_to_node(i);
3562 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3563 unsigned int hctx_idx,
3566 struct blk_mq_tags *tags;
3569 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3573 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3575 blk_mq_free_rq_map(tags);
3582 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3585 if (blk_mq_is_shared_tags(set->flags)) {
3586 set->tags[hctx_idx] = set->shared_tags;
3591 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3594 return set->tags[hctx_idx];
3597 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3598 struct blk_mq_tags *tags,
3599 unsigned int hctx_idx)
3602 blk_mq_free_rqs(set, tags, hctx_idx);
3603 blk_mq_free_rq_map(tags);
3607 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3608 unsigned int hctx_idx)
3610 if (!blk_mq_is_shared_tags(set->flags))
3611 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3613 set->tags[hctx_idx] = NULL;
3616 static void blk_mq_map_swqueue(struct request_queue *q)
3618 unsigned int i, j, hctx_idx;
3619 struct blk_mq_hw_ctx *hctx;
3620 struct blk_mq_ctx *ctx;
3621 struct blk_mq_tag_set *set = q->tag_set;
3623 queue_for_each_hw_ctx(q, hctx, i) {
3624 cpumask_clear(hctx->cpumask);
3626 hctx->dispatch_from = NULL;
3630 * Map software to hardware queues.
3632 * If the cpu isn't present, the cpu is mapped to first hctx.
3634 for_each_possible_cpu(i) {
3636 ctx = per_cpu_ptr(q->queue_ctx, i);
3637 for (j = 0; j < set->nr_maps; j++) {
3638 if (!set->map[j].nr_queues) {
3639 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3640 HCTX_TYPE_DEFAULT, i);
3643 hctx_idx = set->map[j].mq_map[i];
3644 /* unmapped hw queue can be remapped after CPU topo changed */
3645 if (!set->tags[hctx_idx] &&
3646 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3648 * If tags initialization fail for some hctx,
3649 * that hctx won't be brought online. In this
3650 * case, remap the current ctx to hctx[0] which
3651 * is guaranteed to always have tags allocated
3653 set->map[j].mq_map[i] = 0;
3656 hctx = blk_mq_map_queue_type(q, j, i);
3657 ctx->hctxs[j] = hctx;
3659 * If the CPU is already set in the mask, then we've
3660 * mapped this one already. This can happen if
3661 * devices share queues across queue maps.
3663 if (cpumask_test_cpu(i, hctx->cpumask))
3666 cpumask_set_cpu(i, hctx->cpumask);
3668 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3669 hctx->ctxs[hctx->nr_ctx++] = ctx;
3672 * If the nr_ctx type overflows, we have exceeded the
3673 * amount of sw queues we can support.
3675 BUG_ON(!hctx->nr_ctx);
3678 for (; j < HCTX_MAX_TYPES; j++)
3679 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3680 HCTX_TYPE_DEFAULT, i);
3683 queue_for_each_hw_ctx(q, hctx, i) {
3685 * If no software queues are mapped to this hardware queue,
3686 * disable it and free the request entries.
3688 if (!hctx->nr_ctx) {
3689 /* Never unmap queue 0. We need it as a
3690 * fallback in case of a new remap fails
3694 __blk_mq_free_map_and_rqs(set, i);
3700 hctx->tags = set->tags[i];
3701 WARN_ON(!hctx->tags);
3704 * Set the map size to the number of mapped software queues.
3705 * This is more accurate and more efficient than looping
3706 * over all possibly mapped software queues.
3708 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3711 * Initialize batch roundrobin counts
3713 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3714 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3719 * Caller needs to ensure that we're either frozen/quiesced, or that
3720 * the queue isn't live yet.
3722 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3724 struct blk_mq_hw_ctx *hctx;
3727 queue_for_each_hw_ctx(q, hctx, i) {
3729 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3731 blk_mq_tag_idle(hctx);
3732 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3737 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3740 struct request_queue *q;
3742 lockdep_assert_held(&set->tag_list_lock);
3744 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3745 blk_mq_freeze_queue(q);
3746 queue_set_hctx_shared(q, shared);
3747 blk_mq_unfreeze_queue(q);
3751 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3753 struct blk_mq_tag_set *set = q->tag_set;
3755 mutex_lock(&set->tag_list_lock);
3756 list_del(&q->tag_set_list);
3757 if (list_is_singular(&set->tag_list)) {
3758 /* just transitioned to unshared */
3759 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3760 /* update existing queue */
3761 blk_mq_update_tag_set_shared(set, false);
3763 mutex_unlock(&set->tag_list_lock);
3764 INIT_LIST_HEAD(&q->tag_set_list);
3767 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3768 struct request_queue *q)
3770 mutex_lock(&set->tag_list_lock);
3773 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3775 if (!list_empty(&set->tag_list) &&
3776 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3777 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3778 /* update existing queue */
3779 blk_mq_update_tag_set_shared(set, true);
3781 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3782 queue_set_hctx_shared(q, true);
3783 list_add_tail(&q->tag_set_list, &set->tag_list);
3785 mutex_unlock(&set->tag_list_lock);
3788 /* All allocations will be freed in release handler of q->mq_kobj */
3789 static int blk_mq_alloc_ctxs(struct request_queue *q)
3791 struct blk_mq_ctxs *ctxs;
3794 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3798 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3799 if (!ctxs->queue_ctx)
3802 for_each_possible_cpu(cpu) {
3803 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3807 q->mq_kobj = &ctxs->kobj;
3808 q->queue_ctx = ctxs->queue_ctx;
3817 * It is the actual release handler for mq, but we do it from
3818 * request queue's release handler for avoiding use-after-free
3819 * and headache because q->mq_kobj shouldn't have been introduced,
3820 * but we can't group ctx/kctx kobj without it.
3822 void blk_mq_release(struct request_queue *q)
3824 struct blk_mq_hw_ctx *hctx, *next;
3827 queue_for_each_hw_ctx(q, hctx, i)
3828 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3830 /* all hctx are in .unused_hctx_list now */
3831 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3832 list_del_init(&hctx->hctx_list);
3833 kobject_put(&hctx->kobj);
3836 kfree(q->queue_hw_ctx);
3839 * release .mq_kobj and sw queue's kobject now because
3840 * both share lifetime with request queue.
3842 blk_mq_sysfs_deinit(q);
3845 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3848 struct request_queue *q;
3851 q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING);
3853 return ERR_PTR(-ENOMEM);
3854 q->queuedata = queuedata;
3855 ret = blk_mq_init_allocated_queue(set, q);
3857 blk_cleanup_queue(q);
3858 return ERR_PTR(ret);
3863 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3865 return blk_mq_init_queue_data(set, NULL);
3867 EXPORT_SYMBOL(blk_mq_init_queue);
3869 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
3870 struct lock_class_key *lkclass)
3872 struct request_queue *q;
3873 struct gendisk *disk;
3875 q = blk_mq_init_queue_data(set, queuedata);
3879 disk = __alloc_disk_node(q, set->numa_node, lkclass);
3881 blk_cleanup_queue(q);
3882 return ERR_PTR(-ENOMEM);
3886 EXPORT_SYMBOL(__blk_mq_alloc_disk);
3888 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3889 struct blk_mq_tag_set *set, struct request_queue *q,
3890 int hctx_idx, int node)
3892 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3894 /* reuse dead hctx first */
3895 spin_lock(&q->unused_hctx_lock);
3896 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3897 if (tmp->numa_node == node) {
3903 list_del_init(&hctx->hctx_list);
3904 spin_unlock(&q->unused_hctx_lock);
3907 hctx = blk_mq_alloc_hctx(q, set, node);
3911 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3917 kobject_put(&hctx->kobj);
3922 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3923 struct request_queue *q)
3926 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3928 if (q->nr_hw_queues < set->nr_hw_queues) {
3929 struct blk_mq_hw_ctx **new_hctxs;
3931 new_hctxs = kcalloc_node(set->nr_hw_queues,
3932 sizeof(*new_hctxs), GFP_KERNEL,
3937 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3939 q->queue_hw_ctx = new_hctxs;
3944 /* protect against switching io scheduler */
3945 mutex_lock(&q->sysfs_lock);
3946 for (i = 0; i < set->nr_hw_queues; i++) {
3948 struct blk_mq_hw_ctx *hctx;
3950 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3952 * If the hw queue has been mapped to another numa node,
3953 * we need to realloc the hctx. If allocation fails, fallback
3954 * to use the previous one.
3956 if (hctxs[i] && (hctxs[i]->numa_node == node))
3959 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3962 blk_mq_exit_hctx(q, set, hctxs[i], i);
3966 pr_warn("Allocate new hctx on node %d fails,\
3967 fallback to previous one on node %d\n",
3968 node, hctxs[i]->numa_node);
3974 * Increasing nr_hw_queues fails. Free the newly allocated
3975 * hctxs and keep the previous q->nr_hw_queues.
3977 if (i != set->nr_hw_queues) {
3978 j = q->nr_hw_queues;
3982 end = q->nr_hw_queues;
3983 q->nr_hw_queues = set->nr_hw_queues;
3986 for (; j < end; j++) {
3987 struct blk_mq_hw_ctx *hctx = hctxs[j];
3990 blk_mq_exit_hctx(q, set, hctx, j);
3994 mutex_unlock(&q->sysfs_lock);
3997 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3998 struct request_queue *q)
4000 WARN_ON_ONCE(blk_queue_has_srcu(q) !=
4001 !!(set->flags & BLK_MQ_F_BLOCKING));
4003 /* mark the queue as mq asap */
4004 q->mq_ops = set->ops;
4006 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4007 blk_mq_poll_stats_bkt,
4008 BLK_MQ_POLL_STATS_BKTS, q);
4012 if (blk_mq_alloc_ctxs(q))
4015 /* init q->mq_kobj and sw queues' kobjects */
4016 blk_mq_sysfs_init(q);
4018 INIT_LIST_HEAD(&q->unused_hctx_list);
4019 spin_lock_init(&q->unused_hctx_lock);
4021 blk_mq_realloc_hw_ctxs(set, q);
4022 if (!q->nr_hw_queues)
4025 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4026 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4030 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4031 if (set->nr_maps > HCTX_TYPE_POLL &&
4032 set->map[HCTX_TYPE_POLL].nr_queues)
4033 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4035 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4036 INIT_LIST_HEAD(&q->requeue_list);
4037 spin_lock_init(&q->requeue_lock);
4039 q->nr_requests = set->queue_depth;
4042 * Default to classic polling
4044 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4046 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4047 blk_mq_add_queue_tag_set(set, q);
4048 blk_mq_map_swqueue(q);
4052 kfree(q->queue_hw_ctx);
4053 q->nr_hw_queues = 0;
4054 blk_mq_sysfs_deinit(q);
4056 blk_stat_free_callback(q->poll_cb);
4062 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4064 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4065 void blk_mq_exit_queue(struct request_queue *q)
4067 struct blk_mq_tag_set *set = q->tag_set;
4069 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4070 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4071 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4072 blk_mq_del_queue_tag_set(q);
4075 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4079 if (blk_mq_is_shared_tags(set->flags)) {
4080 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4083 if (!set->shared_tags)
4087 for (i = 0; i < set->nr_hw_queues; i++) {
4088 if (!__blk_mq_alloc_map_and_rqs(set, i))
4097 __blk_mq_free_map_and_rqs(set, i);
4099 if (blk_mq_is_shared_tags(set->flags)) {
4100 blk_mq_free_map_and_rqs(set, set->shared_tags,
4101 BLK_MQ_NO_HCTX_IDX);
4108 * Allocate the request maps associated with this tag_set. Note that this
4109 * may reduce the depth asked for, if memory is tight. set->queue_depth
4110 * will be updated to reflect the allocated depth.
4112 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4117 depth = set->queue_depth;
4119 err = __blk_mq_alloc_rq_maps(set);
4123 set->queue_depth >>= 1;
4124 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4128 } while (set->queue_depth);
4130 if (!set->queue_depth || err) {
4131 pr_err("blk-mq: failed to allocate request map\n");
4135 if (depth != set->queue_depth)
4136 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4137 depth, set->queue_depth);
4142 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4145 * blk_mq_map_queues() and multiple .map_queues() implementations
4146 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4147 * number of hardware queues.
4149 if (set->nr_maps == 1)
4150 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4152 if (set->ops->map_queues && !is_kdump_kernel()) {
4156 * transport .map_queues is usually done in the following
4159 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4160 * mask = get_cpu_mask(queue)
4161 * for_each_cpu(cpu, mask)
4162 * set->map[x].mq_map[cpu] = queue;
4165 * When we need to remap, the table has to be cleared for
4166 * killing stale mapping since one CPU may not be mapped
4169 for (i = 0; i < set->nr_maps; i++)
4170 blk_mq_clear_mq_map(&set->map[i]);
4172 return set->ops->map_queues(set);
4174 BUG_ON(set->nr_maps > 1);
4175 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4179 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4180 int cur_nr_hw_queues, int new_nr_hw_queues)
4182 struct blk_mq_tags **new_tags;
4184 if (cur_nr_hw_queues >= new_nr_hw_queues)
4187 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4188 GFP_KERNEL, set->numa_node);
4193 memcpy(new_tags, set->tags, cur_nr_hw_queues *
4194 sizeof(*set->tags));
4196 set->tags = new_tags;
4197 set->nr_hw_queues = new_nr_hw_queues;
4202 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
4203 int new_nr_hw_queues)
4205 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
4209 * Alloc a tag set to be associated with one or more request queues.
4210 * May fail with EINVAL for various error conditions. May adjust the
4211 * requested depth down, if it's too large. In that case, the set
4212 * value will be stored in set->queue_depth.
4214 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4218 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4220 if (!set->nr_hw_queues)
4222 if (!set->queue_depth)
4224 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4227 if (!set->ops->queue_rq)
4230 if (!set->ops->get_budget ^ !set->ops->put_budget)
4233 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4234 pr_info("blk-mq: reduced tag depth to %u\n",
4236 set->queue_depth = BLK_MQ_MAX_DEPTH;
4241 else if (set->nr_maps > HCTX_MAX_TYPES)
4245 * If a crashdump is active, then we are potentially in a very
4246 * memory constrained environment. Limit us to 1 queue and
4247 * 64 tags to prevent using too much memory.
4249 if (is_kdump_kernel()) {
4250 set->nr_hw_queues = 1;
4252 set->queue_depth = min(64U, set->queue_depth);
4255 * There is no use for more h/w queues than cpus if we just have
4258 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4259 set->nr_hw_queues = nr_cpu_ids;
4261 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
4265 for (i = 0; i < set->nr_maps; i++) {
4266 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4267 sizeof(set->map[i].mq_map[0]),
4268 GFP_KERNEL, set->numa_node);
4269 if (!set->map[i].mq_map)
4270 goto out_free_mq_map;
4271 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4274 ret = blk_mq_update_queue_map(set);
4276 goto out_free_mq_map;
4278 ret = blk_mq_alloc_set_map_and_rqs(set);
4280 goto out_free_mq_map;
4282 mutex_init(&set->tag_list_lock);
4283 INIT_LIST_HEAD(&set->tag_list);
4288 for (i = 0; i < set->nr_maps; i++) {
4289 kfree(set->map[i].mq_map);
4290 set->map[i].mq_map = NULL;
4296 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4298 /* allocate and initialize a tagset for a simple single-queue device */
4299 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4300 const struct blk_mq_ops *ops, unsigned int queue_depth,
4301 unsigned int set_flags)
4303 memset(set, 0, sizeof(*set));
4305 set->nr_hw_queues = 1;
4307 set->queue_depth = queue_depth;
4308 set->numa_node = NUMA_NO_NODE;
4309 set->flags = set_flags;
4310 return blk_mq_alloc_tag_set(set);
4312 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4314 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4318 for (i = 0; i < set->nr_hw_queues; i++)
4319 __blk_mq_free_map_and_rqs(set, i);
4321 if (blk_mq_is_shared_tags(set->flags)) {
4322 blk_mq_free_map_and_rqs(set, set->shared_tags,
4323 BLK_MQ_NO_HCTX_IDX);
4326 for (j = 0; j < set->nr_maps; j++) {
4327 kfree(set->map[j].mq_map);
4328 set->map[j].mq_map = NULL;
4334 EXPORT_SYMBOL(blk_mq_free_tag_set);
4336 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4338 struct blk_mq_tag_set *set = q->tag_set;
4339 struct blk_mq_hw_ctx *hctx;
4345 if (q->nr_requests == nr)
4348 blk_mq_freeze_queue(q);
4349 blk_mq_quiesce_queue(q);
4352 queue_for_each_hw_ctx(q, hctx, i) {
4356 * If we're using an MQ scheduler, just update the scheduler
4357 * queue depth. This is similar to what the old code would do.
4359 if (hctx->sched_tags) {
4360 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4363 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4368 if (q->elevator && q->elevator->type->ops.depth_updated)
4369 q->elevator->type->ops.depth_updated(hctx);
4372 q->nr_requests = nr;
4373 if (blk_mq_is_shared_tags(set->flags)) {
4375 blk_mq_tag_update_sched_shared_tags(q);
4377 blk_mq_tag_resize_shared_tags(set, nr);
4381 blk_mq_unquiesce_queue(q);
4382 blk_mq_unfreeze_queue(q);
4388 * request_queue and elevator_type pair.
4389 * It is just used by __blk_mq_update_nr_hw_queues to cache
4390 * the elevator_type associated with a request_queue.
4392 struct blk_mq_qe_pair {
4393 struct list_head node;
4394 struct request_queue *q;
4395 struct elevator_type *type;
4399 * Cache the elevator_type in qe pair list and switch the
4400 * io scheduler to 'none'
4402 static bool blk_mq_elv_switch_none(struct list_head *head,
4403 struct request_queue *q)
4405 struct blk_mq_qe_pair *qe;
4410 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4414 INIT_LIST_HEAD(&qe->node);
4416 qe->type = q->elevator->type;
4417 list_add(&qe->node, head);
4419 mutex_lock(&q->sysfs_lock);
4421 * After elevator_switch_mq, the previous elevator_queue will be
4422 * released by elevator_release. The reference of the io scheduler
4423 * module get by elevator_get will also be put. So we need to get
4424 * a reference of the io scheduler module here to prevent it to be
4427 __module_get(qe->type->elevator_owner);
4428 elevator_switch_mq(q, NULL);
4429 mutex_unlock(&q->sysfs_lock);
4434 static void blk_mq_elv_switch_back(struct list_head *head,
4435 struct request_queue *q)
4437 struct blk_mq_qe_pair *qe;
4438 struct elevator_type *t = NULL;
4440 list_for_each_entry(qe, head, node)
4449 list_del(&qe->node);
4452 mutex_lock(&q->sysfs_lock);
4453 elevator_switch_mq(q, t);
4454 mutex_unlock(&q->sysfs_lock);
4457 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4460 struct request_queue *q;
4462 int prev_nr_hw_queues;
4464 lockdep_assert_held(&set->tag_list_lock);
4466 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4467 nr_hw_queues = nr_cpu_ids;
4468 if (nr_hw_queues < 1)
4470 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4473 list_for_each_entry(q, &set->tag_list, tag_set_list)
4474 blk_mq_freeze_queue(q);
4476 * Switch IO scheduler to 'none', cleaning up the data associated
4477 * with the previous scheduler. We will switch back once we are done
4478 * updating the new sw to hw queue mappings.
4480 list_for_each_entry(q, &set->tag_list, tag_set_list)
4481 if (!blk_mq_elv_switch_none(&head, q))
4484 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4485 blk_mq_debugfs_unregister_hctxs(q);
4486 blk_mq_sysfs_unregister(q);
4489 prev_nr_hw_queues = set->nr_hw_queues;
4490 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4494 set->nr_hw_queues = nr_hw_queues;
4496 blk_mq_update_queue_map(set);
4497 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4498 blk_mq_realloc_hw_ctxs(set, q);
4499 if (q->nr_hw_queues != set->nr_hw_queues) {
4500 int i = prev_nr_hw_queues;
4502 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4503 nr_hw_queues, prev_nr_hw_queues);
4504 for (; i < set->nr_hw_queues; i++)
4505 __blk_mq_free_map_and_rqs(set, i);
4507 set->nr_hw_queues = prev_nr_hw_queues;
4508 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4511 blk_mq_map_swqueue(q);
4515 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4516 blk_mq_sysfs_register(q);
4517 blk_mq_debugfs_register_hctxs(q);
4521 list_for_each_entry(q, &set->tag_list, tag_set_list)
4522 blk_mq_elv_switch_back(&head, q);
4524 list_for_each_entry(q, &set->tag_list, tag_set_list)
4525 blk_mq_unfreeze_queue(q);
4528 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4530 mutex_lock(&set->tag_list_lock);
4531 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4532 mutex_unlock(&set->tag_list_lock);
4534 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4536 /* Enable polling stats and return whether they were already enabled. */
4537 static bool blk_poll_stats_enable(struct request_queue *q)
4542 return blk_stats_alloc_enable(q);
4545 static void blk_mq_poll_stats_start(struct request_queue *q)
4548 * We don't arm the callback if polling stats are not enabled or the
4549 * callback is already active.
4551 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4554 blk_stat_activate_msecs(q->poll_cb, 100);
4557 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4559 struct request_queue *q = cb->data;
4562 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4563 if (cb->stat[bucket].nr_samples)
4564 q->poll_stat[bucket] = cb->stat[bucket];
4568 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4571 unsigned long ret = 0;
4575 * If stats collection isn't on, don't sleep but turn it on for
4578 if (!blk_poll_stats_enable(q))
4582 * As an optimistic guess, use half of the mean service time
4583 * for this type of request. We can (and should) make this smarter.
4584 * For instance, if the completion latencies are tight, we can
4585 * get closer than just half the mean. This is especially
4586 * important on devices where the completion latencies are longer
4587 * than ~10 usec. We do use the stats for the relevant IO size
4588 * if available which does lead to better estimates.
4590 bucket = blk_mq_poll_stats_bkt(rq);
4594 if (q->poll_stat[bucket].nr_samples)
4595 ret = (q->poll_stat[bucket].mean + 1) / 2;
4600 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4602 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4603 struct request *rq = blk_qc_to_rq(hctx, qc);
4604 struct hrtimer_sleeper hs;
4605 enum hrtimer_mode mode;
4610 * If a request has completed on queue that uses an I/O scheduler, we
4611 * won't get back a request from blk_qc_to_rq.
4613 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4617 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4619 * 0: use half of prev avg
4620 * >0: use this specific value
4622 if (q->poll_nsec > 0)
4623 nsecs = q->poll_nsec;
4625 nsecs = blk_mq_poll_nsecs(q, rq);
4630 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4633 * This will be replaced with the stats tracking code, using
4634 * 'avg_completion_time / 2' as the pre-sleep target.
4638 mode = HRTIMER_MODE_REL;
4639 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4640 hrtimer_set_expires(&hs.timer, kt);
4643 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4645 set_current_state(TASK_UNINTERRUPTIBLE);
4646 hrtimer_sleeper_start_expires(&hs, mode);
4649 hrtimer_cancel(&hs.timer);
4650 mode = HRTIMER_MODE_ABS;
4651 } while (hs.task && !signal_pending(current));
4653 __set_current_state(TASK_RUNNING);
4654 destroy_hrtimer_on_stack(&hs.timer);
4657 * If we sleep, have the caller restart the poll loop to reset the
4658 * state. Like for the other success return cases, the caller is
4659 * responsible for checking if the IO completed. If the IO isn't
4660 * complete, we'll get called again and will go straight to the busy
4666 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4667 struct io_comp_batch *iob, unsigned int flags)
4669 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4670 long state = get_current_state();
4674 ret = q->mq_ops->poll(hctx, iob);
4676 __set_current_state(TASK_RUNNING);
4680 if (signal_pending_state(state, current))
4681 __set_current_state(TASK_RUNNING);
4682 if (task_is_running(current))
4685 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4688 } while (!need_resched());
4690 __set_current_state(TASK_RUNNING);
4694 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4697 if (!(flags & BLK_POLL_NOSLEEP) &&
4698 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4699 if (blk_mq_poll_hybrid(q, cookie))
4702 return blk_mq_poll_classic(q, cookie, iob, flags);
4705 unsigned int blk_mq_rq_cpu(struct request *rq)
4707 return rq->mq_ctx->cpu;
4709 EXPORT_SYMBOL(blk_mq_rq_cpu);
4711 void blk_mq_cancel_work_sync(struct request_queue *q)
4713 if (queue_is_mq(q)) {
4714 struct blk_mq_hw_ctx *hctx;
4717 cancel_delayed_work_sync(&q->requeue_work);
4719 queue_for_each_hw_ctx(q, hctx, i)
4720 cancel_delayed_work_sync(&hctx->run_work);
4724 static int __init blk_mq_init(void)
4728 for_each_possible_cpu(i)
4729 init_llist_head(&per_cpu(blk_cpu_done, i));
4730 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4732 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4733 "block/softirq:dead", NULL,
4734 blk_softirq_cpu_dead);
4735 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4736 blk_mq_hctx_notify_dead);
4737 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4738 blk_mq_hctx_notify_online,
4739 blk_mq_hctx_notify_offline);
4742 subsys_initcall(blk_mq_init);