2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/delay.h>
24 #include <linux/crash_dump.h>
25 #include <linux/prefetch.h>
27 #include <trace/events/block.h>
29 #include <linux/blk-mq.h>
32 #include "blk-mq-tag.h"
36 static DEFINE_MUTEX(all_q_mutex
);
37 static LIST_HEAD(all_q_list
);
40 * Check if any of the ctx's have pending work in this hardware queue
42 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
44 return sbitmap_any_bit_set(&hctx
->ctx_map
);
48 * Mark this ctx as having pending work in this hardware queue
50 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
51 struct blk_mq_ctx
*ctx
)
53 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
54 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
57 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
58 struct blk_mq_ctx
*ctx
)
60 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
63 void blk_mq_freeze_queue_start(struct request_queue
*q
)
67 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
68 if (freeze_depth
== 1) {
69 percpu_ref_kill(&q
->q_usage_counter
);
70 blk_mq_run_hw_queues(q
, false);
73 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start
);
75 static void blk_mq_freeze_queue_wait(struct request_queue
*q
)
77 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
81 * Guarantee no request is in use, so we can change any data structure of
82 * the queue afterward.
84 void blk_freeze_queue(struct request_queue
*q
)
87 * In the !blk_mq case we are only calling this to kill the
88 * q_usage_counter, otherwise this increases the freeze depth
89 * and waits for it to return to zero. For this reason there is
90 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
91 * exported to drivers as the only user for unfreeze is blk_mq.
93 blk_mq_freeze_queue_start(q
);
94 blk_mq_freeze_queue_wait(q
);
97 void blk_mq_freeze_queue(struct request_queue
*q
)
100 * ...just an alias to keep freeze and unfreeze actions balanced
101 * in the blk_mq_* namespace
105 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
107 void blk_mq_unfreeze_queue(struct request_queue
*q
)
111 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
112 WARN_ON_ONCE(freeze_depth
< 0);
114 percpu_ref_reinit(&q
->q_usage_counter
);
115 wake_up_all(&q
->mq_freeze_wq
);
118 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
121 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
124 * Note: this function does not prevent that the struct request end_io()
125 * callback function is invoked. Additionally, it is not prevented that
126 * new queue_rq() calls occur unless the queue has been stopped first.
128 void blk_mq_quiesce_queue(struct request_queue
*q
)
130 struct blk_mq_hw_ctx
*hctx
;
134 blk_mq_stop_hw_queues(q
);
136 queue_for_each_hw_ctx(q
, hctx
, i
) {
137 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
138 synchronize_srcu(&hctx
->queue_rq_srcu
);
145 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
147 void blk_mq_wake_waiters(struct request_queue
*q
)
149 struct blk_mq_hw_ctx
*hctx
;
152 queue_for_each_hw_ctx(q
, hctx
, i
)
153 if (blk_mq_hw_queue_mapped(hctx
))
154 blk_mq_tag_wakeup_all(hctx
->tags
, true);
157 * If we are called because the queue has now been marked as
158 * dying, we need to ensure that processes currently waiting on
159 * the queue are notified as well.
161 wake_up_all(&q
->mq_freeze_wq
);
164 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
166 return blk_mq_has_free_tags(hctx
->tags
);
168 EXPORT_SYMBOL(blk_mq_can_queue
);
170 void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
171 struct request
*rq
, unsigned int op
)
173 INIT_LIST_HEAD(&rq
->queuelist
);
174 /* csd/requeue_work/fifo_time is initialized before use */
178 if (blk_queue_io_stat(q
))
179 rq
->rq_flags
|= RQF_IO_STAT
;
180 /* do not touch atomic flags, it needs atomic ops against the timer */
182 INIT_HLIST_NODE(&rq
->hash
);
183 RB_CLEAR_NODE(&rq
->rb_node
);
186 rq
->start_time
= jiffies
;
187 #ifdef CONFIG_BLK_CGROUP
189 set_start_time_ns(rq
);
190 rq
->io_start_time_ns
= 0;
192 rq
->nr_phys_segments
= 0;
193 #if defined(CONFIG_BLK_DEV_INTEGRITY)
194 rq
->nr_integrity_segments
= 0;
197 /* tag was already set */
207 INIT_LIST_HEAD(&rq
->timeout_list
);
211 rq
->end_io_data
= NULL
;
214 ctx
->rq_dispatched
[op_is_sync(op
)]++;
216 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init
);
218 struct request
*__blk_mq_alloc_request(struct blk_mq_alloc_data
*data
,
224 tag
= blk_mq_get_tag(data
);
225 if (tag
!= BLK_MQ_TAG_FAIL
) {
226 rq
= data
->hctx
->tags
->rqs
[tag
];
228 if (blk_mq_tag_busy(data
->hctx
)) {
229 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
230 atomic_inc(&data
->hctx
->nr_active
);
234 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
);
240 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request
);
242 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
245 struct blk_mq_ctx
*ctx
;
246 struct blk_mq_hw_ctx
*hctx
;
248 struct blk_mq_alloc_data alloc_data
;
251 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
255 ctx
= blk_mq_get_ctx(q
);
256 hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
257 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
258 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
263 return ERR_PTR(-EWOULDBLOCK
);
267 rq
->__sector
= (sector_t
) -1;
268 rq
->bio
= rq
->biotail
= NULL
;
271 EXPORT_SYMBOL(blk_mq_alloc_request
);
273 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
274 unsigned int flags
, unsigned int hctx_idx
)
276 struct blk_mq_hw_ctx
*hctx
;
277 struct blk_mq_ctx
*ctx
;
279 struct blk_mq_alloc_data alloc_data
;
283 * If the tag allocator sleeps we could get an allocation for a
284 * different hardware context. No need to complicate the low level
285 * allocator for this for the rare use case of a command tied to
288 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
289 return ERR_PTR(-EINVAL
);
291 if (hctx_idx
>= q
->nr_hw_queues
)
292 return ERR_PTR(-EIO
);
294 ret
= blk_queue_enter(q
, true);
299 * Check if the hardware context is actually mapped to anything.
300 * If not tell the caller that it should skip this queue.
302 hctx
= q
->queue_hw_ctx
[hctx_idx
];
303 if (!blk_mq_hw_queue_mapped(hctx
)) {
307 ctx
= __blk_mq_get_ctx(q
, cpumask_first(hctx
->cpumask
));
309 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
310 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
322 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
324 void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
327 const int tag
= rq
->tag
;
328 struct request_queue
*q
= rq
->q
;
330 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
331 atomic_dec(&hctx
->nr_active
);
333 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
336 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
337 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
338 blk_mq_put_tag(hctx
, ctx
, tag
);
342 static void blk_mq_free_hctx_request(struct blk_mq_hw_ctx
*hctx
,
345 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
347 ctx
->rq_completed
[rq_is_sync(rq
)]++;
348 __blk_mq_free_request(hctx
, ctx
, rq
);
351 void blk_mq_free_request(struct request
*rq
)
353 blk_mq_free_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
355 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
357 inline void __blk_mq_end_request(struct request
*rq
, int error
)
359 blk_account_io_done(rq
);
362 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
363 rq
->end_io(rq
, error
);
365 if (unlikely(blk_bidi_rq(rq
)))
366 blk_mq_free_request(rq
->next_rq
);
367 blk_mq_free_request(rq
);
370 EXPORT_SYMBOL(__blk_mq_end_request
);
372 void blk_mq_end_request(struct request
*rq
, int error
)
374 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
376 __blk_mq_end_request(rq
, error
);
378 EXPORT_SYMBOL(blk_mq_end_request
);
380 static void __blk_mq_complete_request_remote(void *data
)
382 struct request
*rq
= data
;
384 rq
->q
->softirq_done_fn(rq
);
387 static void blk_mq_ipi_complete_request(struct request
*rq
)
389 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
393 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
394 rq
->q
->softirq_done_fn(rq
);
399 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
400 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
402 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
403 rq
->csd
.func
= __blk_mq_complete_request_remote
;
406 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
408 rq
->q
->softirq_done_fn(rq
);
413 static void blk_mq_stat_add(struct request
*rq
)
415 if (rq
->rq_flags
& RQF_STATS
) {
417 * We could rq->mq_ctx here, but there's less of a risk
418 * of races if we have the completion event add the stats
419 * to the local software queue.
421 struct blk_mq_ctx
*ctx
;
423 ctx
= __blk_mq_get_ctx(rq
->q
, raw_smp_processor_id());
424 blk_stat_add(&ctx
->stat
[rq_data_dir(rq
)], rq
);
428 static void __blk_mq_complete_request(struct request
*rq
)
430 struct request_queue
*q
= rq
->q
;
434 if (!q
->softirq_done_fn
)
435 blk_mq_end_request(rq
, rq
->errors
);
437 blk_mq_ipi_complete_request(rq
);
441 * blk_mq_complete_request - end I/O on a request
442 * @rq: the request being processed
445 * Ends all I/O on a request. It does not handle partial completions.
446 * The actual completion happens out-of-order, through a IPI handler.
448 void blk_mq_complete_request(struct request
*rq
, int error
)
450 struct request_queue
*q
= rq
->q
;
452 if (unlikely(blk_should_fake_timeout(q
)))
454 if (!blk_mark_rq_complete(rq
)) {
456 __blk_mq_complete_request(rq
);
459 EXPORT_SYMBOL(blk_mq_complete_request
);
461 int blk_mq_request_started(struct request
*rq
)
463 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
465 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
467 void blk_mq_start_request(struct request
*rq
)
469 struct request_queue
*q
= rq
->q
;
471 trace_block_rq_issue(q
, rq
);
473 rq
->resid_len
= blk_rq_bytes(rq
);
474 if (unlikely(blk_bidi_rq(rq
)))
475 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
477 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
478 blk_stat_set_issue_time(&rq
->issue_stat
);
479 rq
->rq_flags
|= RQF_STATS
;
480 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
486 * Ensure that ->deadline is visible before set the started
487 * flag and clear the completed flag.
489 smp_mb__before_atomic();
492 * Mark us as started and clear complete. Complete might have been
493 * set if requeue raced with timeout, which then marked it as
494 * complete. So be sure to clear complete again when we start
495 * the request, otherwise we'll ignore the completion event.
497 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
498 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
499 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
500 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
502 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
504 * Make sure space for the drain appears. We know we can do
505 * this because max_hw_segments has been adjusted to be one
506 * fewer than the device can handle.
508 rq
->nr_phys_segments
++;
511 EXPORT_SYMBOL(blk_mq_start_request
);
513 static void __blk_mq_requeue_request(struct request
*rq
)
515 struct request_queue
*q
= rq
->q
;
517 trace_block_rq_requeue(q
, rq
);
518 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
520 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
521 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
522 rq
->nr_phys_segments
--;
526 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
528 __blk_mq_requeue_request(rq
);
530 BUG_ON(blk_queued_rq(rq
));
531 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
533 EXPORT_SYMBOL(blk_mq_requeue_request
);
535 static void blk_mq_requeue_work(struct work_struct
*work
)
537 struct request_queue
*q
=
538 container_of(work
, struct request_queue
, requeue_work
.work
);
540 struct request
*rq
, *next
;
543 spin_lock_irqsave(&q
->requeue_lock
, flags
);
544 list_splice_init(&q
->requeue_list
, &rq_list
);
545 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
547 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
548 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
551 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
552 list_del_init(&rq
->queuelist
);
553 blk_mq_insert_request(rq
, true, false, false);
556 while (!list_empty(&rq_list
)) {
557 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
558 list_del_init(&rq
->queuelist
);
559 blk_mq_insert_request(rq
, false, false, false);
562 blk_mq_run_hw_queues(q
, false);
565 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
566 bool kick_requeue_list
)
568 struct request_queue
*q
= rq
->q
;
572 * We abuse this flag that is otherwise used by the I/O scheduler to
573 * request head insertation from the workqueue.
575 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
577 spin_lock_irqsave(&q
->requeue_lock
, flags
);
579 rq
->rq_flags
|= RQF_SOFTBARRIER
;
580 list_add(&rq
->queuelist
, &q
->requeue_list
);
582 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
584 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
586 if (kick_requeue_list
)
587 blk_mq_kick_requeue_list(q
);
589 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
591 void blk_mq_kick_requeue_list(struct request_queue
*q
)
593 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
595 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
597 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
600 kblockd_schedule_delayed_work(&q
->requeue_work
,
601 msecs_to_jiffies(msecs
));
603 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
605 void blk_mq_abort_requeue_list(struct request_queue
*q
)
610 spin_lock_irqsave(&q
->requeue_lock
, flags
);
611 list_splice_init(&q
->requeue_list
, &rq_list
);
612 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
614 while (!list_empty(&rq_list
)) {
617 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
618 list_del_init(&rq
->queuelist
);
620 blk_mq_end_request(rq
, rq
->errors
);
623 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
625 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
627 if (tag
< tags
->nr_tags
) {
628 prefetch(tags
->rqs
[tag
]);
629 return tags
->rqs
[tag
];
634 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
636 struct blk_mq_timeout_data
{
638 unsigned int next_set
;
641 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
643 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
644 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
647 * We know that complete is set at this point. If STARTED isn't set
648 * anymore, then the request isn't active and the "timeout" should
649 * just be ignored. This can happen due to the bitflag ordering.
650 * Timeout first checks if STARTED is set, and if it is, assumes
651 * the request is active. But if we race with completion, then
652 * we both flags will get cleared. So check here again, and ignore
653 * a timeout event with a request that isn't active.
655 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
659 ret
= ops
->timeout(req
, reserved
);
663 __blk_mq_complete_request(req
);
665 case BLK_EH_RESET_TIMER
:
667 blk_clear_rq_complete(req
);
669 case BLK_EH_NOT_HANDLED
:
672 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
677 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
678 struct request
*rq
, void *priv
, bool reserved
)
680 struct blk_mq_timeout_data
*data
= priv
;
682 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
684 * If a request wasn't started before the queue was
685 * marked dying, kill it here or it'll go unnoticed.
687 if (unlikely(blk_queue_dying(rq
->q
))) {
689 blk_mq_end_request(rq
, rq
->errors
);
694 if (time_after_eq(jiffies
, rq
->deadline
)) {
695 if (!blk_mark_rq_complete(rq
))
696 blk_mq_rq_timed_out(rq
, reserved
);
697 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
698 data
->next
= rq
->deadline
;
703 static void blk_mq_timeout_work(struct work_struct
*work
)
705 struct request_queue
*q
=
706 container_of(work
, struct request_queue
, timeout_work
);
707 struct blk_mq_timeout_data data
= {
713 /* A deadlock might occur if a request is stuck requiring a
714 * timeout at the same time a queue freeze is waiting
715 * completion, since the timeout code would not be able to
716 * acquire the queue reference here.
718 * That's why we don't use blk_queue_enter here; instead, we use
719 * percpu_ref_tryget directly, because we need to be able to
720 * obtain a reference even in the short window between the queue
721 * starting to freeze, by dropping the first reference in
722 * blk_mq_freeze_queue_start, and the moment the last request is
723 * consumed, marked by the instant q_usage_counter reaches
726 if (!percpu_ref_tryget(&q
->q_usage_counter
))
729 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
732 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
733 mod_timer(&q
->timeout
, data
.next
);
735 struct blk_mq_hw_ctx
*hctx
;
737 queue_for_each_hw_ctx(q
, hctx
, i
) {
738 /* the hctx may be unmapped, so check it here */
739 if (blk_mq_hw_queue_mapped(hctx
))
740 blk_mq_tag_idle(hctx
);
747 * Reverse check our software queue for entries that we could potentially
748 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
749 * too much time checking for merges.
751 static bool blk_mq_attempt_merge(struct request_queue
*q
,
752 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
757 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
763 if (!blk_rq_merge_ok(rq
, bio
))
766 el_ret
= blk_try_merge(rq
, bio
);
767 if (el_ret
== ELEVATOR_BACK_MERGE
) {
768 if (bio_attempt_back_merge(q
, rq
, bio
)) {
773 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
774 if (bio_attempt_front_merge(q
, rq
, bio
)) {
785 struct flush_busy_ctx_data
{
786 struct blk_mq_hw_ctx
*hctx
;
787 struct list_head
*list
;
790 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
792 struct flush_busy_ctx_data
*flush_data
= data
;
793 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
794 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
796 sbitmap_clear_bit(sb
, bitnr
);
797 spin_lock(&ctx
->lock
);
798 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
799 spin_unlock(&ctx
->lock
);
804 * Process software queues that have been marked busy, splicing them
805 * to the for-dispatch
807 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
809 struct flush_busy_ctx_data data
= {
814 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
816 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
818 static inline unsigned int queued_to_index(unsigned int queued
)
823 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
826 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
828 struct request_queue
*q
= hctx
->queue
;
830 LIST_HEAD(driver_list
);
831 struct list_head
*dptr
;
832 int queued
, ret
= BLK_MQ_RQ_QUEUE_OK
;
835 * Start off with dptr being NULL, so we start the first request
836 * immediately, even if we have more pending.
841 * Now process all the entries, sending them to the driver.
844 while (!list_empty(list
)) {
845 struct blk_mq_queue_data bd
;
847 rq
= list_first_entry(list
, struct request
, queuelist
);
848 list_del_init(&rq
->queuelist
);
852 bd
.last
= list_empty(list
);
854 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
856 case BLK_MQ_RQ_QUEUE_OK
:
859 case BLK_MQ_RQ_QUEUE_BUSY
:
860 list_add(&rq
->queuelist
, list
);
861 __blk_mq_requeue_request(rq
);
864 pr_err("blk-mq: bad return on queue: %d\n", ret
);
865 case BLK_MQ_RQ_QUEUE_ERROR
:
867 blk_mq_end_request(rq
, rq
->errors
);
871 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
875 * We've done the first request. If we have more than 1
876 * left in the list, set dptr to defer issue.
878 if (!dptr
&& list
->next
!= list
->prev
)
882 hctx
->dispatched
[queued_to_index(queued
)]++;
885 * Any items that need requeuing? Stuff them into hctx->dispatch,
886 * that is where we will continue on next queue run.
888 if (!list_empty(list
)) {
889 spin_lock(&hctx
->lock
);
890 list_splice(list
, &hctx
->dispatch
);
891 spin_unlock(&hctx
->lock
);
894 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
895 * it's possible the queue is stopped and restarted again
896 * before this. Queue restart will dispatch requests. And since
897 * requests in rq_list aren't added into hctx->dispatch yet,
898 * the requests in rq_list might get lost.
900 * blk_mq_run_hw_queue() already checks the STOPPED bit
902 blk_mq_run_hw_queue(hctx
, true);
905 return ret
!= BLK_MQ_RQ_QUEUE_BUSY
;
909 * Run this hardware queue, pulling any software queues mapped to it in.
910 * Note that this function currently has various problems around ordering
911 * of IO. In particular, we'd like FIFO behaviour on handling existing
912 * items on the hctx->dispatch list. Ignore that for now.
914 static void blk_mq_process_rq_list(struct blk_mq_hw_ctx
*hctx
)
917 LIST_HEAD(driver_list
);
919 if (unlikely(blk_mq_hctx_stopped(hctx
)))
925 * Touch any software queue that has pending entries.
927 blk_mq_flush_busy_ctxs(hctx
, &rq_list
);
930 * If we have previous entries on our dispatch list, grab them
931 * and stuff them at the front for more fair dispatch.
933 if (!list_empty_careful(&hctx
->dispatch
)) {
934 spin_lock(&hctx
->lock
);
935 if (!list_empty(&hctx
->dispatch
))
936 list_splice_init(&hctx
->dispatch
, &rq_list
);
937 spin_unlock(&hctx
->lock
);
940 blk_mq_dispatch_rq_list(hctx
, &rq_list
);
943 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
947 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
948 cpu_online(hctx
->next_cpu
));
950 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
952 blk_mq_process_rq_list(hctx
);
955 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
956 blk_mq_process_rq_list(hctx
);
957 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
962 * It'd be great if the workqueue API had a way to pass
963 * in a mask and had some smarts for more clever placement.
964 * For now we just round-robin here, switching for every
965 * BLK_MQ_CPU_WORK_BATCH queued items.
967 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
969 if (hctx
->queue
->nr_hw_queues
== 1)
970 return WORK_CPU_UNBOUND
;
972 if (--hctx
->next_cpu_batch
<= 0) {
975 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
976 if (next_cpu
>= nr_cpu_ids
)
977 next_cpu
= cpumask_first(hctx
->cpumask
);
979 hctx
->next_cpu
= next_cpu
;
980 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
983 return hctx
->next_cpu
;
986 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
988 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
989 !blk_mq_hw_queue_mapped(hctx
)))
992 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
994 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
995 __blk_mq_run_hw_queue(hctx
);
1003 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
);
1006 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1008 struct blk_mq_hw_ctx
*hctx
;
1011 queue_for_each_hw_ctx(q
, hctx
, i
) {
1012 if ((!blk_mq_hctx_has_pending(hctx
) &&
1013 list_empty_careful(&hctx
->dispatch
)) ||
1014 blk_mq_hctx_stopped(hctx
))
1017 blk_mq_run_hw_queue(hctx
, async
);
1020 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1023 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1024 * @q: request queue.
1026 * The caller is responsible for serializing this function against
1027 * blk_mq_{start,stop}_hw_queue().
1029 bool blk_mq_queue_stopped(struct request_queue
*q
)
1031 struct blk_mq_hw_ctx
*hctx
;
1034 queue_for_each_hw_ctx(q
, hctx
, i
)
1035 if (blk_mq_hctx_stopped(hctx
))
1040 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1042 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1044 cancel_work(&hctx
->run_work
);
1045 cancel_delayed_work(&hctx
->delay_work
);
1046 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1048 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1050 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1052 struct blk_mq_hw_ctx
*hctx
;
1055 queue_for_each_hw_ctx(q
, hctx
, i
)
1056 blk_mq_stop_hw_queue(hctx
);
1058 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1060 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1062 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1064 blk_mq_run_hw_queue(hctx
, false);
1066 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1068 void blk_mq_start_hw_queues(struct request_queue
*q
)
1070 struct blk_mq_hw_ctx
*hctx
;
1073 queue_for_each_hw_ctx(q
, hctx
, i
)
1074 blk_mq_start_hw_queue(hctx
);
1076 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1078 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1080 if (!blk_mq_hctx_stopped(hctx
))
1083 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1084 blk_mq_run_hw_queue(hctx
, async
);
1086 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1088 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1090 struct blk_mq_hw_ctx
*hctx
;
1093 queue_for_each_hw_ctx(q
, hctx
, i
)
1094 blk_mq_start_stopped_hw_queue(hctx
, async
);
1096 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1098 static void blk_mq_run_work_fn(struct work_struct
*work
)
1100 struct blk_mq_hw_ctx
*hctx
;
1102 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1104 __blk_mq_run_hw_queue(hctx
);
1107 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1109 struct blk_mq_hw_ctx
*hctx
;
1111 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1113 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1114 __blk_mq_run_hw_queue(hctx
);
1117 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1119 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1122 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1123 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1125 EXPORT_SYMBOL(blk_mq_delay_queue
);
1127 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1131 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1133 trace_block_rq_insert(hctx
->queue
, rq
);
1136 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1138 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1141 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1144 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1146 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1147 blk_mq_hctx_mark_pending(hctx
, ctx
);
1150 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
1153 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1154 struct request_queue
*q
= rq
->q
;
1155 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1157 spin_lock(&ctx
->lock
);
1158 __blk_mq_insert_request(hctx
, rq
, at_head
);
1159 spin_unlock(&ctx
->lock
);
1162 blk_mq_run_hw_queue(hctx
, async
);
1165 static void blk_mq_insert_requests(struct request_queue
*q
,
1166 struct blk_mq_ctx
*ctx
,
1167 struct list_head
*list
,
1172 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1174 trace_block_unplug(q
, depth
, !from_schedule
);
1177 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1180 spin_lock(&ctx
->lock
);
1181 while (!list_empty(list
)) {
1184 rq
= list_first_entry(list
, struct request
, queuelist
);
1185 BUG_ON(rq
->mq_ctx
!= ctx
);
1186 list_del_init(&rq
->queuelist
);
1187 __blk_mq_insert_req_list(hctx
, rq
, false);
1189 blk_mq_hctx_mark_pending(hctx
, ctx
);
1190 spin_unlock(&ctx
->lock
);
1192 blk_mq_run_hw_queue(hctx
, from_schedule
);
1195 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1197 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1198 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1200 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1201 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1202 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1205 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1207 struct blk_mq_ctx
*this_ctx
;
1208 struct request_queue
*this_q
;
1211 LIST_HEAD(ctx_list
);
1214 list_splice_init(&plug
->mq_list
, &list
);
1216 list_sort(NULL
, &list
, plug_ctx_cmp
);
1222 while (!list_empty(&list
)) {
1223 rq
= list_entry_rq(list
.next
);
1224 list_del_init(&rq
->queuelist
);
1226 if (rq
->mq_ctx
!= this_ctx
) {
1228 blk_mq_insert_requests(this_q
, this_ctx
,
1233 this_ctx
= rq
->mq_ctx
;
1239 list_add_tail(&rq
->queuelist
, &ctx_list
);
1243 * If 'this_ctx' is set, we know we have entries to complete
1244 * on 'ctx_list'. Do those.
1247 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1252 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1254 init_request_from_bio(rq
, bio
);
1256 blk_account_io_start(rq
, true);
1259 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1261 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1262 !blk_queue_nomerges(hctx
->queue
);
1265 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1266 struct blk_mq_ctx
*ctx
,
1267 struct request
*rq
, struct bio
*bio
)
1269 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1270 blk_mq_bio_to_request(rq
, bio
);
1271 spin_lock(&ctx
->lock
);
1273 __blk_mq_insert_request(hctx
, rq
, false);
1274 spin_unlock(&ctx
->lock
);
1277 struct request_queue
*q
= hctx
->queue
;
1279 spin_lock(&ctx
->lock
);
1280 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1281 blk_mq_bio_to_request(rq
, bio
);
1285 spin_unlock(&ctx
->lock
);
1286 __blk_mq_free_request(hctx
, ctx
, rq
);
1291 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1293 struct blk_mq_alloc_data
*data
)
1295 struct blk_mq_hw_ctx
*hctx
;
1296 struct blk_mq_ctx
*ctx
;
1299 blk_queue_enter_live(q
);
1300 ctx
= blk_mq_get_ctx(q
);
1301 hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1303 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1304 blk_mq_set_alloc_data(data
, q
, 0, ctx
, hctx
);
1305 rq
= __blk_mq_alloc_request(data
, bio
->bi_opf
);
1307 data
->hctx
->queued
++;
1311 static void blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
)
1314 struct request_queue
*q
= rq
->q
;
1315 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, rq
->mq_ctx
->cpu
);
1316 struct blk_mq_queue_data bd
= {
1321 blk_qc_t new_cookie
= blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
);
1323 if (blk_mq_hctx_stopped(hctx
))
1327 * For OK queue, we are done. For error, kill it. Any other
1328 * error (busy), just add it to our list as we previously
1331 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1332 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1333 *cookie
= new_cookie
;
1337 __blk_mq_requeue_request(rq
);
1339 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1340 *cookie
= BLK_QC_T_NONE
;
1342 blk_mq_end_request(rq
, rq
->errors
);
1347 blk_mq_insert_request(rq
, false, true, true);
1351 * Multiple hardware queue variant. This will not use per-process plugs,
1352 * but will attempt to bypass the hctx queueing if we can go straight to
1353 * hardware for SYNC IO.
1355 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1357 const int is_sync
= op_is_sync(bio
->bi_opf
);
1358 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1359 struct blk_mq_alloc_data data
;
1361 unsigned int request_count
= 0, srcu_idx
;
1362 struct blk_plug
*plug
;
1363 struct request
*same_queue_rq
= NULL
;
1365 unsigned int wb_acct
;
1367 blk_queue_bounce(q
, &bio
);
1369 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1371 return BLK_QC_T_NONE
;
1374 blk_queue_split(q
, &bio
, q
->bio_split
);
1376 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1377 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1378 return BLK_QC_T_NONE
;
1380 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1382 rq
= blk_mq_map_request(q
, bio
, &data
);
1383 if (unlikely(!rq
)) {
1384 __wbt_done(q
->rq_wb
, wb_acct
);
1385 return BLK_QC_T_NONE
;
1388 wbt_track(&rq
->issue_stat
, wb_acct
);
1390 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1392 if (unlikely(is_flush_fua
)) {
1393 blk_mq_bio_to_request(rq
, bio
);
1394 blk_insert_flush(rq
);
1398 plug
= current
->plug
;
1400 * If the driver supports defer issued based on 'last', then
1401 * queue it up like normal since we can potentially save some
1404 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1405 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1406 struct request
*old_rq
= NULL
;
1408 blk_mq_bio_to_request(rq
, bio
);
1411 * We do limited plugging. If the bio can be merged, do that.
1412 * Otherwise the existing request in the plug list will be
1413 * issued. So the plug list will have one request at most
1417 * The plug list might get flushed before this. If that
1418 * happens, same_queue_rq is invalid and plug list is
1421 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1422 old_rq
= same_queue_rq
;
1423 list_del_init(&old_rq
->queuelist
);
1425 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1426 } else /* is_sync */
1428 blk_mq_put_ctx(data
.ctx
);
1432 if (!(data
.hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1434 blk_mq_try_issue_directly(old_rq
, &cookie
);
1437 srcu_idx
= srcu_read_lock(&data
.hctx
->queue_rq_srcu
);
1438 blk_mq_try_issue_directly(old_rq
, &cookie
);
1439 srcu_read_unlock(&data
.hctx
->queue_rq_srcu
, srcu_idx
);
1444 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1446 * For a SYNC request, send it to the hardware immediately. For
1447 * an ASYNC request, just ensure that we run it later on. The
1448 * latter allows for merging opportunities and more efficient
1452 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1454 blk_mq_put_ctx(data
.ctx
);
1460 * Single hardware queue variant. This will attempt to use any per-process
1461 * plug for merging and IO deferral.
1463 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1465 const int is_sync
= op_is_sync(bio
->bi_opf
);
1466 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1467 struct blk_plug
*plug
;
1468 unsigned int request_count
= 0;
1469 struct blk_mq_alloc_data data
;
1472 unsigned int wb_acct
;
1474 blk_queue_bounce(q
, &bio
);
1476 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1478 return BLK_QC_T_NONE
;
1481 blk_queue_split(q
, &bio
, q
->bio_split
);
1483 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1484 if (blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1485 return BLK_QC_T_NONE
;
1487 request_count
= blk_plug_queued_count(q
);
1489 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1491 rq
= blk_mq_map_request(q
, bio
, &data
);
1492 if (unlikely(!rq
)) {
1493 __wbt_done(q
->rq_wb
, wb_acct
);
1494 return BLK_QC_T_NONE
;
1497 wbt_track(&rq
->issue_stat
, wb_acct
);
1499 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1501 if (unlikely(is_flush_fua
)) {
1502 blk_mq_bio_to_request(rq
, bio
);
1503 blk_insert_flush(rq
);
1508 * A task plug currently exists. Since this is completely lockless,
1509 * utilize that to temporarily store requests until the task is
1510 * either done or scheduled away.
1512 plug
= current
->plug
;
1514 struct request
*last
= NULL
;
1516 blk_mq_bio_to_request(rq
, bio
);
1519 * @request_count may become stale because of schedule
1520 * out, so check the list again.
1522 if (list_empty(&plug
->mq_list
))
1525 trace_block_plug(q
);
1527 last
= list_entry_rq(plug
->mq_list
.prev
);
1529 blk_mq_put_ctx(data
.ctx
);
1531 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1532 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1533 blk_flush_plug_list(plug
, false);
1534 trace_block_plug(q
);
1537 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1541 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1543 * For a SYNC request, send it to the hardware immediately. For
1544 * an ASYNC request, just ensure that we run it later on. The
1545 * latter allows for merging opportunities and more efficient
1549 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1552 blk_mq_put_ctx(data
.ctx
);
1556 void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1557 unsigned int hctx_idx
)
1561 if (tags
->rqs
&& set
->ops
->exit_request
) {
1564 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1567 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1569 tags
->rqs
[i
] = NULL
;
1573 while (!list_empty(&tags
->page_list
)) {
1574 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1575 list_del_init(&page
->lru
);
1577 * Remove kmemleak object previously allocated in
1578 * blk_mq_init_rq_map().
1580 kmemleak_free(page_address(page
));
1581 __free_pages(page
, page
->private);
1586 blk_mq_free_tags(tags
);
1589 static size_t order_to_size(unsigned int order
)
1591 return (size_t)PAGE_SIZE
<< order
;
1594 struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1595 unsigned int hctx_idx
)
1597 struct blk_mq_tags
*tags
;
1598 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1599 size_t rq_size
, left
;
1601 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1603 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1607 INIT_LIST_HEAD(&tags
->page_list
);
1609 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1610 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1613 blk_mq_free_tags(tags
);
1618 * rq_size is the size of the request plus driver payload, rounded
1619 * to the cacheline size
1621 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1623 left
= rq_size
* set
->queue_depth
;
1625 for (i
= 0; i
< set
->queue_depth
; ) {
1626 int this_order
= max_order
;
1631 while (this_order
&& left
< order_to_size(this_order
- 1))
1635 page
= alloc_pages_node(set
->numa_node
,
1636 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1642 if (order_to_size(this_order
) < rq_size
)
1649 page
->private = this_order
;
1650 list_add_tail(&page
->lru
, &tags
->page_list
);
1652 p
= page_address(page
);
1654 * Allow kmemleak to scan these pages as they contain pointers
1655 * to additional allocations like via ops->init_request().
1657 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1658 entries_per_page
= order_to_size(this_order
) / rq_size
;
1659 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1660 left
-= to_do
* rq_size
;
1661 for (j
= 0; j
< to_do
; j
++) {
1663 if (set
->ops
->init_request
) {
1664 if (set
->ops
->init_request(set
->driver_data
,
1665 tags
->rqs
[i
], hctx_idx
, i
,
1667 tags
->rqs
[i
] = NULL
;
1679 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1684 * 'cpu' is going away. splice any existing rq_list entries from this
1685 * software queue to the hw queue dispatch list, and ensure that it
1688 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1690 struct blk_mq_hw_ctx
*hctx
;
1691 struct blk_mq_ctx
*ctx
;
1694 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1695 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1697 spin_lock(&ctx
->lock
);
1698 if (!list_empty(&ctx
->rq_list
)) {
1699 list_splice_init(&ctx
->rq_list
, &tmp
);
1700 blk_mq_hctx_clear_pending(hctx
, ctx
);
1702 spin_unlock(&ctx
->lock
);
1704 if (list_empty(&tmp
))
1707 spin_lock(&hctx
->lock
);
1708 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1709 spin_unlock(&hctx
->lock
);
1711 blk_mq_run_hw_queue(hctx
, true);
1715 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1717 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1721 /* hctx->ctxs will be freed in queue's release handler */
1722 static void blk_mq_exit_hctx(struct request_queue
*q
,
1723 struct blk_mq_tag_set
*set
,
1724 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1726 unsigned flush_start_tag
= set
->queue_depth
;
1728 blk_mq_tag_idle(hctx
);
1730 if (set
->ops
->exit_request
)
1731 set
->ops
->exit_request(set
->driver_data
,
1732 hctx
->fq
->flush_rq
, hctx_idx
,
1733 flush_start_tag
+ hctx_idx
);
1735 if (set
->ops
->exit_hctx
)
1736 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1738 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1739 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1741 blk_mq_remove_cpuhp(hctx
);
1742 blk_free_flush_queue(hctx
->fq
);
1743 sbitmap_free(&hctx
->ctx_map
);
1746 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1747 struct blk_mq_tag_set
*set
, int nr_queue
)
1749 struct blk_mq_hw_ctx
*hctx
;
1752 queue_for_each_hw_ctx(q
, hctx
, i
) {
1755 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1759 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1760 struct blk_mq_tag_set
*set
)
1762 struct blk_mq_hw_ctx
*hctx
;
1765 queue_for_each_hw_ctx(q
, hctx
, i
)
1766 free_cpumask_var(hctx
->cpumask
);
1769 static int blk_mq_init_hctx(struct request_queue
*q
,
1770 struct blk_mq_tag_set
*set
,
1771 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1774 unsigned flush_start_tag
= set
->queue_depth
;
1776 node
= hctx
->numa_node
;
1777 if (node
== NUMA_NO_NODE
)
1778 node
= hctx
->numa_node
= set
->numa_node
;
1780 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1781 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1782 spin_lock_init(&hctx
->lock
);
1783 INIT_LIST_HEAD(&hctx
->dispatch
);
1785 hctx
->queue_num
= hctx_idx
;
1786 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1788 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1790 hctx
->tags
= set
->tags
[hctx_idx
];
1793 * Allocate space for all possible cpus to avoid allocation at
1796 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1799 goto unregister_cpu_notifier
;
1801 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1807 if (set
->ops
->init_hctx
&&
1808 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1811 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1815 if (set
->ops
->init_request
&&
1816 set
->ops
->init_request(set
->driver_data
,
1817 hctx
->fq
->flush_rq
, hctx_idx
,
1818 flush_start_tag
+ hctx_idx
, node
))
1821 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1822 init_srcu_struct(&hctx
->queue_rq_srcu
);
1829 if (set
->ops
->exit_hctx
)
1830 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1832 sbitmap_free(&hctx
->ctx_map
);
1835 unregister_cpu_notifier
:
1836 blk_mq_remove_cpuhp(hctx
);
1840 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1841 unsigned int nr_hw_queues
)
1845 for_each_possible_cpu(i
) {
1846 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1847 struct blk_mq_hw_ctx
*hctx
;
1849 memset(__ctx
, 0, sizeof(*__ctx
));
1851 spin_lock_init(&__ctx
->lock
);
1852 INIT_LIST_HEAD(&__ctx
->rq_list
);
1854 blk_stat_init(&__ctx
->stat
[BLK_STAT_READ
]);
1855 blk_stat_init(&__ctx
->stat
[BLK_STAT_WRITE
]);
1857 /* If the cpu isn't online, the cpu is mapped to first hctx */
1861 hctx
= blk_mq_map_queue(q
, i
);
1864 * Set local node, IFF we have more than one hw queue. If
1865 * not, we remain on the home node of the device
1867 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1868 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1872 static void blk_mq_map_swqueue(struct request_queue
*q
,
1873 const struct cpumask
*online_mask
)
1875 unsigned int i
, hctx_idx
;
1876 struct blk_mq_hw_ctx
*hctx
;
1877 struct blk_mq_ctx
*ctx
;
1878 struct blk_mq_tag_set
*set
= q
->tag_set
;
1881 * Avoid others reading imcomplete hctx->cpumask through sysfs
1883 mutex_lock(&q
->sysfs_lock
);
1885 queue_for_each_hw_ctx(q
, hctx
, i
) {
1886 cpumask_clear(hctx
->cpumask
);
1891 * Map software to hardware queues
1893 for_each_possible_cpu(i
) {
1894 /* If the cpu isn't online, the cpu is mapped to first hctx */
1895 if (!cpumask_test_cpu(i
, online_mask
))
1898 hctx_idx
= q
->mq_map
[i
];
1899 /* unmapped hw queue can be remapped after CPU topo changed */
1900 if (!set
->tags
[hctx_idx
]) {
1901 set
->tags
[hctx_idx
] = blk_mq_init_rq_map(set
, hctx_idx
);
1904 * If tags initialization fail for some hctx,
1905 * that hctx won't be brought online. In this
1906 * case, remap the current ctx to hctx[0] which
1907 * is guaranteed to always have tags allocated
1909 if (!set
->tags
[hctx_idx
])
1913 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1914 hctx
= blk_mq_map_queue(q
, i
);
1916 cpumask_set_cpu(i
, hctx
->cpumask
);
1917 ctx
->index_hw
= hctx
->nr_ctx
;
1918 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1921 mutex_unlock(&q
->sysfs_lock
);
1923 queue_for_each_hw_ctx(q
, hctx
, i
) {
1925 * If no software queues are mapped to this hardware queue,
1926 * disable it and free the request entries.
1928 if (!hctx
->nr_ctx
) {
1929 /* Never unmap queue 0. We need it as a
1930 * fallback in case of a new remap fails
1933 if (i
&& set
->tags
[i
]) {
1934 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1935 set
->tags
[i
] = NULL
;
1941 hctx
->tags
= set
->tags
[i
];
1942 WARN_ON(!hctx
->tags
);
1945 * Set the map size to the number of mapped software queues.
1946 * This is more accurate and more efficient than looping
1947 * over all possibly mapped software queues.
1949 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
1952 * Initialize batch roundrobin counts
1954 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1955 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1959 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
1961 struct blk_mq_hw_ctx
*hctx
;
1964 queue_for_each_hw_ctx(q
, hctx
, i
) {
1966 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1968 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1972 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
1974 struct request_queue
*q
;
1976 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1977 blk_mq_freeze_queue(q
);
1978 queue_set_hctx_shared(q
, shared
);
1979 blk_mq_unfreeze_queue(q
);
1983 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1985 struct blk_mq_tag_set
*set
= q
->tag_set
;
1987 mutex_lock(&set
->tag_list_lock
);
1988 list_del_init(&q
->tag_set_list
);
1989 if (list_is_singular(&set
->tag_list
)) {
1990 /* just transitioned to unshared */
1991 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1992 /* update existing queue */
1993 blk_mq_update_tag_set_depth(set
, false);
1995 mutex_unlock(&set
->tag_list_lock
);
1998 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1999 struct request_queue
*q
)
2003 mutex_lock(&set
->tag_list_lock
);
2005 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2006 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2007 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2008 /* update existing queue */
2009 blk_mq_update_tag_set_depth(set
, true);
2011 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2012 queue_set_hctx_shared(q
, true);
2013 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
2015 mutex_unlock(&set
->tag_list_lock
);
2019 * It is the actual release handler for mq, but we do it from
2020 * request queue's release handler for avoiding use-after-free
2021 * and headache because q->mq_kobj shouldn't have been introduced,
2022 * but we can't group ctx/kctx kobj without it.
2024 void blk_mq_release(struct request_queue
*q
)
2026 struct blk_mq_hw_ctx
*hctx
;
2029 /* hctx kobj stays in hctx */
2030 queue_for_each_hw_ctx(q
, hctx
, i
) {
2039 kfree(q
->queue_hw_ctx
);
2041 /* ctx kobj stays in queue_ctx */
2042 free_percpu(q
->queue_ctx
);
2045 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2047 struct request_queue
*uninit_q
, *q
;
2049 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2051 return ERR_PTR(-ENOMEM
);
2053 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2055 blk_cleanup_queue(uninit_q
);
2059 EXPORT_SYMBOL(blk_mq_init_queue
);
2061 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2062 struct request_queue
*q
)
2065 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2067 blk_mq_sysfs_unregister(q
);
2068 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2074 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2075 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2080 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2087 atomic_set(&hctxs
[i
]->nr_active
, 0);
2088 hctxs
[i
]->numa_node
= node
;
2089 hctxs
[i
]->queue_num
= i
;
2091 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2092 free_cpumask_var(hctxs
[i
]->cpumask
);
2097 blk_mq_hctx_kobj_init(hctxs
[i
]);
2099 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2100 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2104 blk_mq_free_rq_map(set
, hctx
->tags
, j
);
2105 set
->tags
[j
] = NULL
;
2107 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2108 free_cpumask_var(hctx
->cpumask
);
2109 kobject_put(&hctx
->kobj
);
2116 q
->nr_hw_queues
= i
;
2117 blk_mq_sysfs_register(q
);
2120 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2121 struct request_queue
*q
)
2123 /* mark the queue as mq asap */
2124 q
->mq_ops
= set
->ops
;
2126 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2130 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2131 GFP_KERNEL
, set
->numa_node
);
2132 if (!q
->queue_hw_ctx
)
2135 q
->mq_map
= set
->mq_map
;
2137 blk_mq_realloc_hw_ctxs(set
, q
);
2138 if (!q
->nr_hw_queues
)
2141 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2142 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2144 q
->nr_queues
= nr_cpu_ids
;
2146 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2148 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2149 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2151 q
->sg_reserved_size
= INT_MAX
;
2153 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2154 INIT_LIST_HEAD(&q
->requeue_list
);
2155 spin_lock_init(&q
->requeue_lock
);
2157 if (q
->nr_hw_queues
> 1)
2158 blk_queue_make_request(q
, blk_mq_make_request
);
2160 blk_queue_make_request(q
, blk_sq_make_request
);
2163 * Do this after blk_queue_make_request() overrides it...
2165 q
->nr_requests
= set
->queue_depth
;
2168 * Default to classic polling
2172 if (set
->ops
->complete
)
2173 blk_queue_softirq_done(q
, set
->ops
->complete
);
2175 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2178 mutex_lock(&all_q_mutex
);
2180 list_add_tail(&q
->all_q_node
, &all_q_list
);
2181 blk_mq_add_queue_tag_set(set
, q
);
2182 blk_mq_map_swqueue(q
, cpu_online_mask
);
2184 mutex_unlock(&all_q_mutex
);
2190 kfree(q
->queue_hw_ctx
);
2192 free_percpu(q
->queue_ctx
);
2195 return ERR_PTR(-ENOMEM
);
2197 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2199 void blk_mq_free_queue(struct request_queue
*q
)
2201 struct blk_mq_tag_set
*set
= q
->tag_set
;
2203 mutex_lock(&all_q_mutex
);
2204 list_del_init(&q
->all_q_node
);
2205 mutex_unlock(&all_q_mutex
);
2209 blk_mq_del_queue_tag_set(q
);
2211 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2212 blk_mq_free_hw_queues(q
, set
);
2215 /* Basically redo blk_mq_init_queue with queue frozen */
2216 static void blk_mq_queue_reinit(struct request_queue
*q
,
2217 const struct cpumask
*online_mask
)
2219 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2221 blk_mq_sysfs_unregister(q
);
2224 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2225 * we should change hctx numa_node according to new topology (this
2226 * involves free and re-allocate memory, worthy doing?)
2229 blk_mq_map_swqueue(q
, online_mask
);
2231 blk_mq_sysfs_register(q
);
2235 * New online cpumask which is going to be set in this hotplug event.
2236 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2237 * one-by-one and dynamically allocating this could result in a failure.
2239 static struct cpumask cpuhp_online_new
;
2241 static void blk_mq_queue_reinit_work(void)
2243 struct request_queue
*q
;
2245 mutex_lock(&all_q_mutex
);
2247 * We need to freeze and reinit all existing queues. Freezing
2248 * involves synchronous wait for an RCU grace period and doing it
2249 * one by one may take a long time. Start freezing all queues in
2250 * one swoop and then wait for the completions so that freezing can
2251 * take place in parallel.
2253 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2254 blk_mq_freeze_queue_start(q
);
2255 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2256 blk_mq_freeze_queue_wait(q
);
2258 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2259 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2261 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2262 blk_mq_unfreeze_queue(q
);
2264 mutex_unlock(&all_q_mutex
);
2267 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2269 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2270 blk_mq_queue_reinit_work();
2275 * Before hotadded cpu starts handling requests, new mappings must be
2276 * established. Otherwise, these requests in hw queue might never be
2279 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2280 * for CPU0, and ctx1 for CPU1).
2282 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2283 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2285 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2286 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2287 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2290 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2292 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2293 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2294 blk_mq_queue_reinit_work();
2298 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2302 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2303 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2312 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2318 * Allocate the request maps associated with this tag_set. Note that this
2319 * may reduce the depth asked for, if memory is tight. set->queue_depth
2320 * will be updated to reflect the allocated depth.
2322 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2327 depth
= set
->queue_depth
;
2329 err
= __blk_mq_alloc_rq_maps(set
);
2333 set
->queue_depth
>>= 1;
2334 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2338 } while (set
->queue_depth
);
2340 if (!set
->queue_depth
|| err
) {
2341 pr_err("blk-mq: failed to allocate request map\n");
2345 if (depth
!= set
->queue_depth
)
2346 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2347 depth
, set
->queue_depth
);
2353 * Alloc a tag set to be associated with one or more request queues.
2354 * May fail with EINVAL for various error conditions. May adjust the
2355 * requested depth down, if if it too large. In that case, the set
2356 * value will be stored in set->queue_depth.
2358 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2362 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2364 if (!set
->nr_hw_queues
)
2366 if (!set
->queue_depth
)
2368 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2371 if (!set
->ops
->queue_rq
)
2374 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2375 pr_info("blk-mq: reduced tag depth to %u\n",
2377 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2381 * If a crashdump is active, then we are potentially in a very
2382 * memory constrained environment. Limit us to 1 queue and
2383 * 64 tags to prevent using too much memory.
2385 if (is_kdump_kernel()) {
2386 set
->nr_hw_queues
= 1;
2387 set
->queue_depth
= min(64U, set
->queue_depth
);
2390 * There is no use for more h/w queues than cpus.
2392 if (set
->nr_hw_queues
> nr_cpu_ids
)
2393 set
->nr_hw_queues
= nr_cpu_ids
;
2395 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2396 GFP_KERNEL
, set
->numa_node
);
2401 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2402 GFP_KERNEL
, set
->numa_node
);
2406 if (set
->ops
->map_queues
)
2407 ret
= set
->ops
->map_queues(set
);
2409 ret
= blk_mq_map_queues(set
);
2411 goto out_free_mq_map
;
2413 ret
= blk_mq_alloc_rq_maps(set
);
2415 goto out_free_mq_map
;
2417 mutex_init(&set
->tag_list_lock
);
2418 INIT_LIST_HEAD(&set
->tag_list
);
2430 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2432 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2436 for (i
= 0; i
< nr_cpu_ids
; i
++) {
2438 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2447 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2449 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2451 struct blk_mq_tag_set
*set
= q
->tag_set
;
2452 struct blk_mq_hw_ctx
*hctx
;
2455 if (!set
|| nr
> set
->queue_depth
)
2459 queue_for_each_hw_ctx(q
, hctx
, i
) {
2462 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2468 q
->nr_requests
= nr
;
2473 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2475 struct request_queue
*q
;
2477 if (nr_hw_queues
> nr_cpu_ids
)
2478 nr_hw_queues
= nr_cpu_ids
;
2479 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2482 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2483 blk_mq_freeze_queue(q
);
2485 set
->nr_hw_queues
= nr_hw_queues
;
2486 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2487 blk_mq_realloc_hw_ctxs(set
, q
);
2489 if (q
->nr_hw_queues
> 1)
2490 blk_queue_make_request(q
, blk_mq_make_request
);
2492 blk_queue_make_request(q
, blk_sq_make_request
);
2494 blk_mq_queue_reinit(q
, cpu_online_mask
);
2497 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2498 blk_mq_unfreeze_queue(q
);
2500 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2502 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2503 struct blk_mq_hw_ctx
*hctx
,
2506 struct blk_rq_stat stat
[2];
2507 unsigned long ret
= 0;
2510 * If stats collection isn't on, don't sleep but turn it on for
2513 if (!blk_stat_enable(q
))
2517 * We don't have to do this once per IO, should optimize this
2518 * to just use the current window of stats until it changes
2520 memset(&stat
, 0, sizeof(stat
));
2521 blk_hctx_stat_get(hctx
, stat
);
2524 * As an optimistic guess, use half of the mean service time
2525 * for this type of request. We can (and should) make this smarter.
2526 * For instance, if the completion latencies are tight, we can
2527 * get closer than just half the mean. This is especially
2528 * important on devices where the completion latencies are longer
2531 if (req_op(rq
) == REQ_OP_READ
&& stat
[BLK_STAT_READ
].nr_samples
)
2532 ret
= (stat
[BLK_STAT_READ
].mean
+ 1) / 2;
2533 else if (req_op(rq
) == REQ_OP_WRITE
&& stat
[BLK_STAT_WRITE
].nr_samples
)
2534 ret
= (stat
[BLK_STAT_WRITE
].mean
+ 1) / 2;
2539 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2540 struct blk_mq_hw_ctx
*hctx
,
2543 struct hrtimer_sleeper hs
;
2544 enum hrtimer_mode mode
;
2548 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2554 * -1: don't ever hybrid sleep
2555 * 0: use half of prev avg
2556 * >0: use this specific value
2558 if (q
->poll_nsec
== -1)
2560 else if (q
->poll_nsec
> 0)
2561 nsecs
= q
->poll_nsec
;
2563 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2568 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2571 * This will be replaced with the stats tracking code, using
2572 * 'avg_completion_time / 2' as the pre-sleep target.
2576 mode
= HRTIMER_MODE_REL
;
2577 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2578 hrtimer_set_expires(&hs
.timer
, kt
);
2580 hrtimer_init_sleeper(&hs
, current
);
2582 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2584 set_current_state(TASK_UNINTERRUPTIBLE
);
2585 hrtimer_start_expires(&hs
.timer
, mode
);
2588 hrtimer_cancel(&hs
.timer
);
2589 mode
= HRTIMER_MODE_ABS
;
2590 } while (hs
.task
&& !signal_pending(current
));
2592 __set_current_state(TASK_RUNNING
);
2593 destroy_hrtimer_on_stack(&hs
.timer
);
2597 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2599 struct request_queue
*q
= hctx
->queue
;
2603 * If we sleep, have the caller restart the poll loop to reset
2604 * the state. Like for the other success return cases, the
2605 * caller is responsible for checking if the IO completed. If
2606 * the IO isn't complete, we'll get called again and will go
2607 * straight to the busy poll loop.
2609 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2612 hctx
->poll_considered
++;
2614 state
= current
->state
;
2615 while (!need_resched()) {
2618 hctx
->poll_invoked
++;
2620 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2622 hctx
->poll_success
++;
2623 set_current_state(TASK_RUNNING
);
2627 if (signal_pending_state(state
, current
))
2628 set_current_state(TASK_RUNNING
);
2630 if (current
->state
== TASK_RUNNING
)
2640 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2642 struct blk_mq_hw_ctx
*hctx
;
2643 struct blk_plug
*plug
;
2646 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2647 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2650 plug
= current
->plug
;
2652 blk_flush_plug_list(plug
, false);
2654 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2655 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2657 return __blk_mq_poll(hctx
, rq
);
2659 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2661 void blk_mq_disable_hotplug(void)
2663 mutex_lock(&all_q_mutex
);
2666 void blk_mq_enable_hotplug(void)
2668 mutex_unlock(&all_q_mutex
);
2671 static int __init
blk_mq_init(void)
2673 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2674 blk_mq_hctx_notify_dead
);
2676 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2677 blk_mq_queue_reinit_prepare
,
2678 blk_mq_queue_reinit_dead
);
2681 subsys_initcall(blk_mq_init
);