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>
26 #include <trace/events/block.h>
28 #include <linux/blk-mq.h>
31 #include "blk-mq-tag.h"
33 static DEFINE_MUTEX(all_q_mutex
);
34 static LIST_HEAD(all_q_list
);
36 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
);
39 * Check if any of the ctx's have pending work in this hardware queue
41 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
45 for (i
= 0; i
< hctx
->ctx_map
.size
; i
++)
46 if (hctx
->ctx_map
.map
[i
].word
)
52 static inline struct blk_align_bitmap
*get_bm(struct blk_mq_hw_ctx
*hctx
,
53 struct blk_mq_ctx
*ctx
)
55 return &hctx
->ctx_map
.map
[ctx
->index_hw
/ hctx
->ctx_map
.bits_per_word
];
58 #define CTX_TO_BIT(hctx, ctx) \
59 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
62 * Mark this ctx as having pending work in this hardware queue
64 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
65 struct blk_mq_ctx
*ctx
)
67 struct blk_align_bitmap
*bm
= get_bm(hctx
, ctx
);
69 if (!test_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
))
70 set_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
);
73 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
74 struct blk_mq_ctx
*ctx
)
76 struct blk_align_bitmap
*bm
= get_bm(hctx
, ctx
);
78 clear_bit(CTX_TO_BIT(hctx
, ctx
), &bm
->word
);
81 void blk_mq_freeze_queue_start(struct request_queue
*q
)
85 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
86 if (freeze_depth
== 1) {
87 percpu_ref_kill(&q
->q_usage_counter
);
88 blk_mq_run_hw_queues(q
, false);
91 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start
);
93 static void blk_mq_freeze_queue_wait(struct request_queue
*q
)
95 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
99 * Guarantee no request is in use, so we can change any data structure of
100 * the queue afterward.
102 void blk_freeze_queue(struct request_queue
*q
)
105 * In the !blk_mq case we are only calling this to kill the
106 * q_usage_counter, otherwise this increases the freeze depth
107 * and waits for it to return to zero. For this reason there is
108 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
109 * exported to drivers as the only user for unfreeze is blk_mq.
111 blk_mq_freeze_queue_start(q
);
112 blk_mq_freeze_queue_wait(q
);
115 void blk_mq_freeze_queue(struct request_queue
*q
)
118 * ...just an alias to keep freeze and unfreeze actions balanced
119 * in the blk_mq_* namespace
123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
125 void blk_mq_unfreeze_queue(struct request_queue
*q
)
129 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
130 WARN_ON_ONCE(freeze_depth
< 0);
132 percpu_ref_reinit(&q
->q_usage_counter
);
133 wake_up_all(&q
->mq_freeze_wq
);
136 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
138 void blk_mq_wake_waiters(struct request_queue
*q
)
140 struct blk_mq_hw_ctx
*hctx
;
143 queue_for_each_hw_ctx(q
, hctx
, i
)
144 if (blk_mq_hw_queue_mapped(hctx
))
145 blk_mq_tag_wakeup_all(hctx
->tags
, true);
148 * If we are called because the queue has now been marked as
149 * dying, we need to ensure that processes currently waiting on
150 * the queue are notified as well.
152 wake_up_all(&q
->mq_freeze_wq
);
155 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
157 return blk_mq_has_free_tags(hctx
->tags
);
159 EXPORT_SYMBOL(blk_mq_can_queue
);
161 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
162 struct request
*rq
, unsigned int rw_flags
)
164 if (blk_queue_io_stat(q
))
165 rw_flags
|= REQ_IO_STAT
;
167 INIT_LIST_HEAD(&rq
->queuelist
);
168 /* csd/requeue_work/fifo_time is initialized before use */
171 rq
->cmd_flags
|= rw_flags
;
172 /* do not touch atomic flags, it needs atomic ops against the timer */
174 INIT_HLIST_NODE(&rq
->hash
);
175 RB_CLEAR_NODE(&rq
->rb_node
);
178 rq
->start_time
= jiffies
;
179 #ifdef CONFIG_BLK_CGROUP
181 set_start_time_ns(rq
);
182 rq
->io_start_time_ns
= 0;
184 rq
->nr_phys_segments
= 0;
185 #if defined(CONFIG_BLK_DEV_INTEGRITY)
186 rq
->nr_integrity_segments
= 0;
189 /* tag was already set */
199 INIT_LIST_HEAD(&rq
->timeout_list
);
203 rq
->end_io_data
= NULL
;
206 ctx
->rq_dispatched
[rw_is_sync(rw_flags
)]++;
209 static struct request
*
210 __blk_mq_alloc_request(struct blk_mq_alloc_data
*data
, int rw
)
215 tag
= blk_mq_get_tag(data
);
216 if (tag
!= BLK_MQ_TAG_FAIL
) {
217 rq
= data
->hctx
->tags
->rqs
[tag
];
219 if (blk_mq_tag_busy(data
->hctx
)) {
220 rq
->cmd_flags
= REQ_MQ_INFLIGHT
;
221 atomic_inc(&data
->hctx
->nr_active
);
225 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, rw
);
232 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
235 struct blk_mq_ctx
*ctx
;
236 struct blk_mq_hw_ctx
*hctx
;
238 struct blk_mq_alloc_data alloc_data
;
241 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
245 ctx
= blk_mq_get_ctx(q
);
246 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
247 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
249 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
250 if (!rq
&& !(flags
& BLK_MQ_REQ_NOWAIT
)) {
251 __blk_mq_run_hw_queue(hctx
);
254 ctx
= blk_mq_get_ctx(q
);
255 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
256 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
257 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
258 ctx
= alloc_data
.ctx
;
263 return ERR_PTR(-EWOULDBLOCK
);
267 EXPORT_SYMBOL(blk_mq_alloc_request
);
269 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
270 unsigned int flags
, unsigned int hctx_idx
)
272 struct blk_mq_hw_ctx
*hctx
;
273 struct blk_mq_ctx
*ctx
;
275 struct blk_mq_alloc_data alloc_data
;
279 * If the tag allocator sleeps we could get an allocation for a
280 * different hardware context. No need to complicate the low level
281 * allocator for this for the rare use case of a command tied to
284 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
285 return ERR_PTR(-EINVAL
);
287 if (hctx_idx
>= q
->nr_hw_queues
)
288 return ERR_PTR(-EIO
);
290 ret
= blk_queue_enter(q
, true);
294 hctx
= q
->queue_hw_ctx
[hctx_idx
];
295 ctx
= __blk_mq_get_ctx(q
, cpumask_first(hctx
->cpumask
));
297 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
298 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
301 return ERR_PTR(-EWOULDBLOCK
);
306 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
308 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
309 struct blk_mq_ctx
*ctx
, struct request
*rq
)
311 const int tag
= rq
->tag
;
312 struct request_queue
*q
= rq
->q
;
314 if (rq
->cmd_flags
& REQ_MQ_INFLIGHT
)
315 atomic_dec(&hctx
->nr_active
);
318 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
319 blk_mq_put_tag(hctx
, tag
, &ctx
->last_tag
);
323 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
325 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
327 ctx
->rq_completed
[rq_is_sync(rq
)]++;
328 __blk_mq_free_request(hctx
, ctx
, rq
);
331 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request
);
333 void blk_mq_free_request(struct request
*rq
)
335 struct blk_mq_hw_ctx
*hctx
;
336 struct request_queue
*q
= rq
->q
;
338 hctx
= q
->mq_ops
->map_queue(q
, rq
->mq_ctx
->cpu
);
339 blk_mq_free_hctx_request(hctx
, rq
);
341 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
343 inline void __blk_mq_end_request(struct request
*rq
, int error
)
345 blk_account_io_done(rq
);
348 rq
->end_io(rq
, error
);
350 if (unlikely(blk_bidi_rq(rq
)))
351 blk_mq_free_request(rq
->next_rq
);
352 blk_mq_free_request(rq
);
355 EXPORT_SYMBOL(__blk_mq_end_request
);
357 void blk_mq_end_request(struct request
*rq
, int error
)
359 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
361 __blk_mq_end_request(rq
, error
);
363 EXPORT_SYMBOL(blk_mq_end_request
);
365 static void __blk_mq_complete_request_remote(void *data
)
367 struct request
*rq
= data
;
369 rq
->q
->softirq_done_fn(rq
);
372 static void blk_mq_ipi_complete_request(struct request
*rq
)
374 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
378 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
379 rq
->q
->softirq_done_fn(rq
);
384 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
385 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
387 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
388 rq
->csd
.func
= __blk_mq_complete_request_remote
;
391 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
393 rq
->q
->softirq_done_fn(rq
);
398 static void __blk_mq_complete_request(struct request
*rq
)
400 struct request_queue
*q
= rq
->q
;
402 if (!q
->softirq_done_fn
)
403 blk_mq_end_request(rq
, rq
->errors
);
405 blk_mq_ipi_complete_request(rq
);
409 * blk_mq_complete_request - end I/O on a request
410 * @rq: the request being processed
413 * Ends all I/O on a request. It does not handle partial completions.
414 * The actual completion happens out-of-order, through a IPI handler.
416 void blk_mq_complete_request(struct request
*rq
, int error
)
418 struct request_queue
*q
= rq
->q
;
420 if (unlikely(blk_should_fake_timeout(q
)))
422 if (!blk_mark_rq_complete(rq
)) {
424 __blk_mq_complete_request(rq
);
427 EXPORT_SYMBOL(blk_mq_complete_request
);
429 int blk_mq_request_started(struct request
*rq
)
431 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
433 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
435 void blk_mq_start_request(struct request
*rq
)
437 struct request_queue
*q
= rq
->q
;
439 trace_block_rq_issue(q
, rq
);
441 rq
->resid_len
= blk_rq_bytes(rq
);
442 if (unlikely(blk_bidi_rq(rq
)))
443 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
448 * Ensure that ->deadline is visible before set the started
449 * flag and clear the completed flag.
451 smp_mb__before_atomic();
454 * Mark us as started and clear complete. Complete might have been
455 * set if requeue raced with timeout, which then marked it as
456 * complete. So be sure to clear complete again when we start
457 * the request, otherwise we'll ignore the completion event.
459 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
460 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
461 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
462 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
464 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
466 * Make sure space for the drain appears. We know we can do
467 * this because max_hw_segments has been adjusted to be one
468 * fewer than the device can handle.
470 rq
->nr_phys_segments
++;
473 EXPORT_SYMBOL(blk_mq_start_request
);
475 static void __blk_mq_requeue_request(struct request
*rq
)
477 struct request_queue
*q
= rq
->q
;
479 trace_block_rq_requeue(q
, rq
);
481 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
482 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
483 rq
->nr_phys_segments
--;
487 void blk_mq_requeue_request(struct request
*rq
)
489 __blk_mq_requeue_request(rq
);
491 BUG_ON(blk_queued_rq(rq
));
492 blk_mq_add_to_requeue_list(rq
, true);
494 EXPORT_SYMBOL(blk_mq_requeue_request
);
496 static void blk_mq_requeue_work(struct work_struct
*work
)
498 struct request_queue
*q
=
499 container_of(work
, struct request_queue
, requeue_work
);
501 struct request
*rq
, *next
;
504 spin_lock_irqsave(&q
->requeue_lock
, flags
);
505 list_splice_init(&q
->requeue_list
, &rq_list
);
506 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
508 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
509 if (!(rq
->cmd_flags
& REQ_SOFTBARRIER
))
512 rq
->cmd_flags
&= ~REQ_SOFTBARRIER
;
513 list_del_init(&rq
->queuelist
);
514 blk_mq_insert_request(rq
, true, false, false);
517 while (!list_empty(&rq_list
)) {
518 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
519 list_del_init(&rq
->queuelist
);
520 blk_mq_insert_request(rq
, false, false, false);
524 * Use the start variant of queue running here, so that running
525 * the requeue work will kick stopped queues.
527 blk_mq_start_hw_queues(q
);
530 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
)
532 struct request_queue
*q
= rq
->q
;
536 * We abuse this flag that is otherwise used by the I/O scheduler to
537 * request head insertation from the workqueue.
539 BUG_ON(rq
->cmd_flags
& REQ_SOFTBARRIER
);
541 spin_lock_irqsave(&q
->requeue_lock
, flags
);
543 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
544 list_add(&rq
->queuelist
, &q
->requeue_list
);
546 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
548 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
550 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
552 void blk_mq_cancel_requeue_work(struct request_queue
*q
)
554 cancel_work_sync(&q
->requeue_work
);
556 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work
);
558 void blk_mq_kick_requeue_list(struct request_queue
*q
)
560 kblockd_schedule_work(&q
->requeue_work
);
562 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
564 void blk_mq_abort_requeue_list(struct request_queue
*q
)
569 spin_lock_irqsave(&q
->requeue_lock
, flags
);
570 list_splice_init(&q
->requeue_list
, &rq_list
);
571 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
573 while (!list_empty(&rq_list
)) {
576 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
577 list_del_init(&rq
->queuelist
);
579 blk_mq_end_request(rq
, rq
->errors
);
582 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
584 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
586 return tags
->rqs
[tag
];
588 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
590 struct blk_mq_timeout_data
{
592 unsigned int next_set
;
595 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
597 struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
598 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
601 * We know that complete is set at this point. If STARTED isn't set
602 * anymore, then the request isn't active and the "timeout" should
603 * just be ignored. This can happen due to the bitflag ordering.
604 * Timeout first checks if STARTED is set, and if it is, assumes
605 * the request is active. But if we race with completion, then
606 * we both flags will get cleared. So check here again, and ignore
607 * a timeout event with a request that isn't active.
609 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
613 ret
= ops
->timeout(req
, reserved
);
617 __blk_mq_complete_request(req
);
619 case BLK_EH_RESET_TIMER
:
621 blk_clear_rq_complete(req
);
623 case BLK_EH_NOT_HANDLED
:
626 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
631 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
632 struct request
*rq
, void *priv
, bool reserved
)
634 struct blk_mq_timeout_data
*data
= priv
;
636 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
638 * If a request wasn't started before the queue was
639 * marked dying, kill it here or it'll go unnoticed.
641 if (unlikely(blk_queue_dying(rq
->q
))) {
643 blk_mq_end_request(rq
, rq
->errors
);
647 if (rq
->cmd_flags
& REQ_NO_TIMEOUT
)
650 if (time_after_eq(jiffies
, rq
->deadline
)) {
651 if (!blk_mark_rq_complete(rq
))
652 blk_mq_rq_timed_out(rq
, reserved
);
653 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
654 data
->next
= rq
->deadline
;
659 static void blk_mq_timeout_work(struct work_struct
*work
)
661 struct request_queue
*q
=
662 container_of(work
, struct request_queue
, timeout_work
);
663 struct blk_mq_timeout_data data
= {
669 /* A deadlock might occur if a request is stuck requiring a
670 * timeout at the same time a queue freeze is waiting
671 * completion, since the timeout code would not be able to
672 * acquire the queue reference here.
674 * That's why we don't use blk_queue_enter here; instead, we use
675 * percpu_ref_tryget directly, because we need to be able to
676 * obtain a reference even in the short window between the queue
677 * starting to freeze, by dropping the first reference in
678 * blk_mq_freeze_queue_start, and the moment the last request is
679 * consumed, marked by the instant q_usage_counter reaches
682 if (!percpu_ref_tryget(&q
->q_usage_counter
))
685 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
688 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
689 mod_timer(&q
->timeout
, data
.next
);
691 struct blk_mq_hw_ctx
*hctx
;
693 queue_for_each_hw_ctx(q
, hctx
, i
) {
694 /* the hctx may be unmapped, so check it here */
695 if (blk_mq_hw_queue_mapped(hctx
))
696 blk_mq_tag_idle(hctx
);
703 * Reverse check our software queue for entries that we could potentially
704 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
705 * too much time checking for merges.
707 static bool blk_mq_attempt_merge(struct request_queue
*q
,
708 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
713 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
719 if (!blk_rq_merge_ok(rq
, bio
))
722 el_ret
= blk_try_merge(rq
, bio
);
723 if (el_ret
== ELEVATOR_BACK_MERGE
) {
724 if (bio_attempt_back_merge(q
, rq
, bio
)) {
729 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
730 if (bio_attempt_front_merge(q
, rq
, bio
)) {
742 * Process software queues that have been marked busy, splicing them
743 * to the for-dispatch
745 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
747 struct blk_mq_ctx
*ctx
;
750 for (i
= 0; i
< hctx
->ctx_map
.size
; i
++) {
751 struct blk_align_bitmap
*bm
= &hctx
->ctx_map
.map
[i
];
752 unsigned int off
, bit
;
758 off
= i
* hctx
->ctx_map
.bits_per_word
;
760 bit
= find_next_bit(&bm
->word
, bm
->depth
, bit
);
761 if (bit
>= bm
->depth
)
764 ctx
= hctx
->ctxs
[bit
+ off
];
765 clear_bit(bit
, &bm
->word
);
766 spin_lock(&ctx
->lock
);
767 list_splice_tail_init(&ctx
->rq_list
, list
);
768 spin_unlock(&ctx
->lock
);
776 * Run this hardware queue, pulling any software queues mapped to it in.
777 * Note that this function currently has various problems around ordering
778 * of IO. In particular, we'd like FIFO behaviour on handling existing
779 * items on the hctx->dispatch list. Ignore that for now.
781 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
783 struct request_queue
*q
= hctx
->queue
;
786 LIST_HEAD(driver_list
);
787 struct list_head
*dptr
;
790 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
793 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
794 cpu_online(hctx
->next_cpu
));
799 * Touch any software queue that has pending entries.
801 flush_busy_ctxs(hctx
, &rq_list
);
804 * If we have previous entries on our dispatch list, grab them
805 * and stuff them at the front for more fair dispatch.
807 if (!list_empty_careful(&hctx
->dispatch
)) {
808 spin_lock(&hctx
->lock
);
809 if (!list_empty(&hctx
->dispatch
))
810 list_splice_init(&hctx
->dispatch
, &rq_list
);
811 spin_unlock(&hctx
->lock
);
815 * Start off with dptr being NULL, so we start the first request
816 * immediately, even if we have more pending.
821 * Now process all the entries, sending them to the driver.
824 while (!list_empty(&rq_list
)) {
825 struct blk_mq_queue_data bd
;
828 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
829 list_del_init(&rq
->queuelist
);
833 bd
.last
= list_empty(&rq_list
);
835 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
837 case BLK_MQ_RQ_QUEUE_OK
:
840 case BLK_MQ_RQ_QUEUE_BUSY
:
841 list_add(&rq
->queuelist
, &rq_list
);
842 __blk_mq_requeue_request(rq
);
845 pr_err("blk-mq: bad return on queue: %d\n", ret
);
846 case BLK_MQ_RQ_QUEUE_ERROR
:
848 blk_mq_end_request(rq
, rq
->errors
);
852 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
856 * We've done the first request. If we have more than 1
857 * left in the list, set dptr to defer issue.
859 if (!dptr
&& rq_list
.next
!= rq_list
.prev
)
864 hctx
->dispatched
[0]++;
865 else if (queued
< (1 << (BLK_MQ_MAX_DISPATCH_ORDER
- 1)))
866 hctx
->dispatched
[ilog2(queued
) + 1]++;
869 * Any items that need requeuing? Stuff them into hctx->dispatch,
870 * that is where we will continue on next queue run.
872 if (!list_empty(&rq_list
)) {
873 spin_lock(&hctx
->lock
);
874 list_splice(&rq_list
, &hctx
->dispatch
);
875 spin_unlock(&hctx
->lock
);
877 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
878 * it's possible the queue is stopped and restarted again
879 * before this. Queue restart will dispatch requests. And since
880 * requests in rq_list aren't added into hctx->dispatch yet,
881 * the requests in rq_list might get lost.
883 * blk_mq_run_hw_queue() already checks the STOPPED bit
885 blk_mq_run_hw_queue(hctx
, true);
890 * It'd be great if the workqueue API had a way to pass
891 * in a mask and had some smarts for more clever placement.
892 * For now we just round-robin here, switching for every
893 * BLK_MQ_CPU_WORK_BATCH queued items.
895 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
897 if (hctx
->queue
->nr_hw_queues
== 1)
898 return WORK_CPU_UNBOUND
;
900 if (--hctx
->next_cpu_batch
<= 0) {
903 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
904 if (next_cpu
>= nr_cpu_ids
)
905 next_cpu
= cpumask_first(hctx
->cpumask
);
907 hctx
->next_cpu
= next_cpu
;
908 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
911 return hctx
->next_cpu
;
914 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
916 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
) ||
917 !blk_mq_hw_queue_mapped(hctx
)))
922 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
923 __blk_mq_run_hw_queue(hctx
);
931 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
935 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
937 struct blk_mq_hw_ctx
*hctx
;
940 queue_for_each_hw_ctx(q
, hctx
, i
) {
941 if ((!blk_mq_hctx_has_pending(hctx
) &&
942 list_empty_careful(&hctx
->dispatch
)) ||
943 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
946 blk_mq_run_hw_queue(hctx
, async
);
949 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
951 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
953 cancel_delayed_work(&hctx
->run_work
);
954 cancel_delayed_work(&hctx
->delay_work
);
955 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
957 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
959 void blk_mq_stop_hw_queues(struct request_queue
*q
)
961 struct blk_mq_hw_ctx
*hctx
;
964 queue_for_each_hw_ctx(q
, hctx
, i
)
965 blk_mq_stop_hw_queue(hctx
);
967 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
969 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
971 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
973 blk_mq_run_hw_queue(hctx
, false);
975 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
977 void blk_mq_start_hw_queues(struct request_queue
*q
)
979 struct blk_mq_hw_ctx
*hctx
;
982 queue_for_each_hw_ctx(q
, hctx
, i
)
983 blk_mq_start_hw_queue(hctx
);
985 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
987 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
989 struct blk_mq_hw_ctx
*hctx
;
992 queue_for_each_hw_ctx(q
, hctx
, i
) {
993 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
996 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
997 blk_mq_run_hw_queue(hctx
, async
);
1000 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1002 static void blk_mq_run_work_fn(struct work_struct
*work
)
1004 struct blk_mq_hw_ctx
*hctx
;
1006 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1008 __blk_mq_run_hw_queue(hctx
);
1011 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1013 struct blk_mq_hw_ctx
*hctx
;
1015 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1017 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1018 __blk_mq_run_hw_queue(hctx
);
1021 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1023 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1026 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1027 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1029 EXPORT_SYMBOL(blk_mq_delay_queue
);
1031 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1035 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1037 trace_block_rq_insert(hctx
->queue
, rq
);
1040 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1042 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1045 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
1046 struct request
*rq
, bool at_head
)
1048 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1050 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1051 blk_mq_hctx_mark_pending(hctx
, ctx
);
1054 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
1057 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1058 struct request_queue
*q
= rq
->q
;
1059 struct blk_mq_hw_ctx
*hctx
;
1061 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1063 spin_lock(&ctx
->lock
);
1064 __blk_mq_insert_request(hctx
, rq
, at_head
);
1065 spin_unlock(&ctx
->lock
);
1068 blk_mq_run_hw_queue(hctx
, async
);
1071 static void blk_mq_insert_requests(struct request_queue
*q
,
1072 struct blk_mq_ctx
*ctx
,
1073 struct list_head
*list
,
1078 struct blk_mq_hw_ctx
*hctx
;
1080 trace_block_unplug(q
, depth
, !from_schedule
);
1082 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1085 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1088 spin_lock(&ctx
->lock
);
1089 while (!list_empty(list
)) {
1092 rq
= list_first_entry(list
, struct request
, queuelist
);
1093 BUG_ON(rq
->mq_ctx
!= ctx
);
1094 list_del_init(&rq
->queuelist
);
1095 __blk_mq_insert_req_list(hctx
, rq
, false);
1097 blk_mq_hctx_mark_pending(hctx
, ctx
);
1098 spin_unlock(&ctx
->lock
);
1100 blk_mq_run_hw_queue(hctx
, from_schedule
);
1103 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1105 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1106 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1108 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1109 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1110 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1113 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1115 struct blk_mq_ctx
*this_ctx
;
1116 struct request_queue
*this_q
;
1119 LIST_HEAD(ctx_list
);
1122 list_splice_init(&plug
->mq_list
, &list
);
1124 list_sort(NULL
, &list
, plug_ctx_cmp
);
1130 while (!list_empty(&list
)) {
1131 rq
= list_entry_rq(list
.next
);
1132 list_del_init(&rq
->queuelist
);
1134 if (rq
->mq_ctx
!= this_ctx
) {
1136 blk_mq_insert_requests(this_q
, this_ctx
,
1141 this_ctx
= rq
->mq_ctx
;
1147 list_add_tail(&rq
->queuelist
, &ctx_list
);
1151 * If 'this_ctx' is set, we know we have entries to complete
1152 * on 'ctx_list'. Do those.
1155 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1160 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1162 init_request_from_bio(rq
, bio
);
1164 if (blk_do_io_stat(rq
))
1165 blk_account_io_start(rq
, 1);
1168 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1170 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1171 !blk_queue_nomerges(hctx
->queue
);
1174 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1175 struct blk_mq_ctx
*ctx
,
1176 struct request
*rq
, struct bio
*bio
)
1178 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1179 blk_mq_bio_to_request(rq
, bio
);
1180 spin_lock(&ctx
->lock
);
1182 __blk_mq_insert_request(hctx
, rq
, false);
1183 spin_unlock(&ctx
->lock
);
1186 struct request_queue
*q
= hctx
->queue
;
1188 spin_lock(&ctx
->lock
);
1189 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1190 blk_mq_bio_to_request(rq
, bio
);
1194 spin_unlock(&ctx
->lock
);
1195 __blk_mq_free_request(hctx
, ctx
, rq
);
1200 struct blk_map_ctx
{
1201 struct blk_mq_hw_ctx
*hctx
;
1202 struct blk_mq_ctx
*ctx
;
1205 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1207 struct blk_map_ctx
*data
)
1209 struct blk_mq_hw_ctx
*hctx
;
1210 struct blk_mq_ctx
*ctx
;
1212 int rw
= bio_data_dir(bio
);
1213 struct blk_mq_alloc_data alloc_data
;
1215 blk_queue_enter_live(q
);
1216 ctx
= blk_mq_get_ctx(q
);
1217 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1219 if (rw_is_sync(bio
->bi_rw
))
1222 trace_block_getrq(q
, bio
, rw
);
1223 blk_mq_set_alloc_data(&alloc_data
, q
, BLK_MQ_REQ_NOWAIT
, ctx
, hctx
);
1224 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1225 if (unlikely(!rq
)) {
1226 __blk_mq_run_hw_queue(hctx
);
1227 blk_mq_put_ctx(ctx
);
1228 trace_block_sleeprq(q
, bio
, rw
);
1230 ctx
= blk_mq_get_ctx(q
);
1231 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
1232 blk_mq_set_alloc_data(&alloc_data
, q
, 0, ctx
, hctx
);
1233 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
1234 ctx
= alloc_data
.ctx
;
1235 hctx
= alloc_data
.hctx
;
1244 static int blk_mq_direct_issue_request(struct request
*rq
, blk_qc_t
*cookie
)
1247 struct request_queue
*q
= rq
->q
;
1248 struct blk_mq_hw_ctx
*hctx
= q
->mq_ops
->map_queue(q
,
1250 struct blk_mq_queue_data bd
= {
1255 blk_qc_t new_cookie
= blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
);
1258 * For OK queue, we are done. For error, kill it. Any other
1259 * error (busy), just add it to our list as we previously
1262 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1263 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1264 *cookie
= new_cookie
;
1268 __blk_mq_requeue_request(rq
);
1270 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1271 *cookie
= BLK_QC_T_NONE
;
1273 blk_mq_end_request(rq
, rq
->errors
);
1281 * Multiple hardware queue variant. This will not use per-process plugs,
1282 * but will attempt to bypass the hctx queueing if we can go straight to
1283 * hardware for SYNC IO.
1285 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1287 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1288 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1289 struct blk_map_ctx data
;
1291 unsigned int request_count
= 0;
1292 struct blk_plug
*plug
;
1293 struct request
*same_queue_rq
= NULL
;
1296 blk_queue_bounce(q
, &bio
);
1298 blk_queue_split(q
, &bio
, q
->bio_split
);
1300 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1302 return BLK_QC_T_NONE
;
1305 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1306 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1307 return BLK_QC_T_NONE
;
1309 rq
= blk_mq_map_request(q
, bio
, &data
);
1311 return BLK_QC_T_NONE
;
1313 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1315 if (unlikely(is_flush_fua
)) {
1316 blk_mq_bio_to_request(rq
, bio
);
1317 blk_insert_flush(rq
);
1321 plug
= current
->plug
;
1323 * If the driver supports defer issued based on 'last', then
1324 * queue it up like normal since we can potentially save some
1327 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1328 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1329 struct request
*old_rq
= NULL
;
1331 blk_mq_bio_to_request(rq
, bio
);
1334 * We do limited pluging. If the bio can be merged, do that.
1335 * Otherwise the existing request in the plug list will be
1336 * issued. So the plug list will have one request at most
1340 * The plug list might get flushed before this. If that
1341 * happens, same_queue_rq is invalid and plug list is
1344 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1345 old_rq
= same_queue_rq
;
1346 list_del_init(&old_rq
->queuelist
);
1348 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1349 } else /* is_sync */
1351 blk_mq_put_ctx(data
.ctx
);
1354 if (test_bit(BLK_MQ_S_STOPPED
, &data
.hctx
->state
) ||
1355 blk_mq_direct_issue_request(old_rq
, &cookie
) != 0)
1356 blk_mq_insert_request(old_rq
, false, true, true);
1360 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1362 * For a SYNC request, send it to the hardware immediately. For
1363 * an ASYNC request, just ensure that we run it later on. The
1364 * latter allows for merging opportunities and more efficient
1368 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1370 blk_mq_put_ctx(data
.ctx
);
1376 * Single hardware queue variant. This will attempt to use any per-process
1377 * plug for merging and IO deferral.
1379 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1381 const int is_sync
= rw_is_sync(bio
->bi_rw
);
1382 const int is_flush_fua
= bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
1383 struct blk_plug
*plug
;
1384 unsigned int request_count
= 0;
1385 struct blk_map_ctx data
;
1389 blk_queue_bounce(q
, &bio
);
1391 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1393 return BLK_QC_T_NONE
;
1396 blk_queue_split(q
, &bio
, q
->bio_split
);
1398 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1399 if (blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1400 return BLK_QC_T_NONE
;
1402 request_count
= blk_plug_queued_count(q
);
1404 rq
= blk_mq_map_request(q
, bio
, &data
);
1406 return BLK_QC_T_NONE
;
1408 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1410 if (unlikely(is_flush_fua
)) {
1411 blk_mq_bio_to_request(rq
, bio
);
1412 blk_insert_flush(rq
);
1417 * A task plug currently exists. Since this is completely lockless,
1418 * utilize that to temporarily store requests until the task is
1419 * either done or scheduled away.
1421 plug
= current
->plug
;
1423 blk_mq_bio_to_request(rq
, bio
);
1425 trace_block_plug(q
);
1427 blk_mq_put_ctx(data
.ctx
);
1429 if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1430 blk_flush_plug_list(plug
, false);
1431 trace_block_plug(q
);
1434 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1438 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1440 * For a SYNC request, send it to the hardware immediately. For
1441 * an ASYNC request, just ensure that we run it later on. The
1442 * latter allows for merging opportunities and more efficient
1446 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1449 blk_mq_put_ctx(data
.ctx
);
1454 * Default mapping to a software queue, since we use one per CPU.
1456 struct blk_mq_hw_ctx
*blk_mq_map_queue(struct request_queue
*q
, const int cpu
)
1458 return q
->queue_hw_ctx
[q
->mq_map
[cpu
]];
1460 EXPORT_SYMBOL(blk_mq_map_queue
);
1462 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1463 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1467 if (tags
->rqs
&& set
->ops
->exit_request
) {
1470 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1473 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1475 tags
->rqs
[i
] = NULL
;
1479 while (!list_empty(&tags
->page_list
)) {
1480 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1481 list_del_init(&page
->lru
);
1483 * Remove kmemleak object previously allocated in
1484 * blk_mq_init_rq_map().
1486 kmemleak_free(page_address(page
));
1487 __free_pages(page
, page
->private);
1492 blk_mq_free_tags(tags
);
1495 static size_t order_to_size(unsigned int order
)
1497 return (size_t)PAGE_SIZE
<< order
;
1500 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1501 unsigned int hctx_idx
)
1503 struct blk_mq_tags
*tags
;
1504 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1505 size_t rq_size
, left
;
1507 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1509 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1513 INIT_LIST_HEAD(&tags
->page_list
);
1515 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1516 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1519 blk_mq_free_tags(tags
);
1524 * rq_size is the size of the request plus driver payload, rounded
1525 * to the cacheline size
1527 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1529 left
= rq_size
* set
->queue_depth
;
1531 for (i
= 0; i
< set
->queue_depth
; ) {
1532 int this_order
= max_order
;
1537 while (left
< order_to_size(this_order
- 1) && this_order
)
1541 page
= alloc_pages_node(set
->numa_node
,
1542 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1548 if (order_to_size(this_order
) < rq_size
)
1555 page
->private = this_order
;
1556 list_add_tail(&page
->lru
, &tags
->page_list
);
1558 p
= page_address(page
);
1560 * Allow kmemleak to scan these pages as they contain pointers
1561 * to additional allocations like via ops->init_request().
1563 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1564 entries_per_page
= order_to_size(this_order
) / rq_size
;
1565 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1566 left
-= to_do
* rq_size
;
1567 for (j
= 0; j
< to_do
; j
++) {
1569 if (set
->ops
->init_request
) {
1570 if (set
->ops
->init_request(set
->driver_data
,
1571 tags
->rqs
[i
], hctx_idx
, i
,
1573 tags
->rqs
[i
] = NULL
;
1585 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1589 static void blk_mq_free_bitmap(struct blk_mq_ctxmap
*bitmap
)
1594 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap
*bitmap
, int node
)
1596 unsigned int bpw
= 8, total
, num_maps
, i
;
1598 bitmap
->bits_per_word
= bpw
;
1600 num_maps
= ALIGN(nr_cpu_ids
, bpw
) / bpw
;
1601 bitmap
->map
= kzalloc_node(num_maps
* sizeof(struct blk_align_bitmap
),
1607 for (i
= 0; i
< num_maps
; i
++) {
1608 bitmap
->map
[i
].depth
= min(total
, bitmap
->bits_per_word
);
1609 total
-= bitmap
->map
[i
].depth
;
1616 * 'cpu' is going away. splice any existing rq_list entries from this
1617 * software queue to the hw queue dispatch list, and ensure that it
1620 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx
*hctx
, int cpu
)
1622 struct blk_mq_ctx
*ctx
;
1625 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1627 spin_lock(&ctx
->lock
);
1628 if (!list_empty(&ctx
->rq_list
)) {
1629 list_splice_init(&ctx
->rq_list
, &tmp
);
1630 blk_mq_hctx_clear_pending(hctx
, ctx
);
1632 spin_unlock(&ctx
->lock
);
1634 if (list_empty(&tmp
))
1637 spin_lock(&hctx
->lock
);
1638 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1639 spin_unlock(&hctx
->lock
);
1641 blk_mq_run_hw_queue(hctx
, true);
1645 static int blk_mq_hctx_notify(void *data
, unsigned long action
,
1648 struct blk_mq_hw_ctx
*hctx
= data
;
1650 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
1651 return blk_mq_hctx_cpu_offline(hctx
, cpu
);
1654 * In case of CPU online, tags may be reallocated
1655 * in blk_mq_map_swqueue() after mapping is updated.
1661 /* hctx->ctxs will be freed in queue's release handler */
1662 static void blk_mq_exit_hctx(struct request_queue
*q
,
1663 struct blk_mq_tag_set
*set
,
1664 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1666 unsigned flush_start_tag
= set
->queue_depth
;
1668 blk_mq_tag_idle(hctx
);
1670 if (set
->ops
->exit_request
)
1671 set
->ops
->exit_request(set
->driver_data
,
1672 hctx
->fq
->flush_rq
, hctx_idx
,
1673 flush_start_tag
+ hctx_idx
);
1675 if (set
->ops
->exit_hctx
)
1676 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1678 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1679 blk_free_flush_queue(hctx
->fq
);
1680 blk_mq_free_bitmap(&hctx
->ctx_map
);
1683 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1684 struct blk_mq_tag_set
*set
, int nr_queue
)
1686 struct blk_mq_hw_ctx
*hctx
;
1689 queue_for_each_hw_ctx(q
, hctx
, i
) {
1692 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1696 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1697 struct blk_mq_tag_set
*set
)
1699 struct blk_mq_hw_ctx
*hctx
;
1702 queue_for_each_hw_ctx(q
, hctx
, i
)
1703 free_cpumask_var(hctx
->cpumask
);
1706 static int blk_mq_init_hctx(struct request_queue
*q
,
1707 struct blk_mq_tag_set
*set
,
1708 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1711 unsigned flush_start_tag
= set
->queue_depth
;
1713 node
= hctx
->numa_node
;
1714 if (node
== NUMA_NO_NODE
)
1715 node
= hctx
->numa_node
= set
->numa_node
;
1717 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1718 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1719 spin_lock_init(&hctx
->lock
);
1720 INIT_LIST_HEAD(&hctx
->dispatch
);
1722 hctx
->queue_num
= hctx_idx
;
1723 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1725 blk_mq_init_cpu_notifier(&hctx
->cpu_notifier
,
1726 blk_mq_hctx_notify
, hctx
);
1727 blk_mq_register_cpu_notifier(&hctx
->cpu_notifier
);
1729 hctx
->tags
= set
->tags
[hctx_idx
];
1732 * Allocate space for all possible cpus to avoid allocation at
1735 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1738 goto unregister_cpu_notifier
;
1740 if (blk_mq_alloc_bitmap(&hctx
->ctx_map
, node
))
1745 if (set
->ops
->init_hctx
&&
1746 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1749 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1753 if (set
->ops
->init_request
&&
1754 set
->ops
->init_request(set
->driver_data
,
1755 hctx
->fq
->flush_rq
, hctx_idx
,
1756 flush_start_tag
+ hctx_idx
, node
))
1764 if (set
->ops
->exit_hctx
)
1765 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1767 blk_mq_free_bitmap(&hctx
->ctx_map
);
1770 unregister_cpu_notifier
:
1771 blk_mq_unregister_cpu_notifier(&hctx
->cpu_notifier
);
1776 static int blk_mq_init_hw_queues(struct request_queue
*q
,
1777 struct blk_mq_tag_set
*set
)
1779 struct blk_mq_hw_ctx
*hctx
;
1783 * Initialize hardware queues
1785 queue_for_each_hw_ctx(q
, hctx
, i
) {
1786 if (blk_mq_init_hctx(q
, set
, hctx
, i
))
1790 if (i
== q
->nr_hw_queues
)
1796 blk_mq_exit_hw_queues(q
, set
, i
);
1801 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1802 unsigned int nr_hw_queues
)
1806 for_each_possible_cpu(i
) {
1807 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1808 struct blk_mq_hw_ctx
*hctx
;
1810 memset(__ctx
, 0, sizeof(*__ctx
));
1812 spin_lock_init(&__ctx
->lock
);
1813 INIT_LIST_HEAD(&__ctx
->rq_list
);
1816 /* If the cpu isn't online, the cpu is mapped to first hctx */
1820 hctx
= q
->mq_ops
->map_queue(q
, i
);
1823 * Set local node, IFF we have more than one hw queue. If
1824 * not, we remain on the home node of the device
1826 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1827 hctx
->numa_node
= cpu_to_node(i
);
1831 static void blk_mq_map_swqueue(struct request_queue
*q
,
1832 const struct cpumask
*online_mask
)
1834 unsigned int i
, hctx_idx
;
1835 struct blk_mq_hw_ctx
*hctx
;
1836 struct blk_mq_ctx
*ctx
;
1837 struct blk_mq_tag_set
*set
= q
->tag_set
;
1840 * Avoid others reading imcomplete hctx->cpumask through sysfs
1842 mutex_lock(&q
->sysfs_lock
);
1844 queue_for_each_hw_ctx(q
, hctx
, i
) {
1845 cpumask_clear(hctx
->cpumask
);
1850 * Map software to hardware queues
1852 for_each_possible_cpu(i
) {
1853 /* If the cpu isn't online, the cpu is mapped to first hctx */
1854 if (!cpumask_test_cpu(i
, online_mask
))
1857 hctx_idx
= q
->mq_map
[i
];
1858 /* unmapped hw queue can be remapped after CPU topo changed */
1859 if (!set
->tags
[hctx_idx
]) {
1860 set
->tags
[hctx_idx
] = blk_mq_init_rq_map(set
, hctx_idx
);
1863 * If tags initialization fail for some hctx,
1864 * that hctx won't be brought online. In this
1865 * case, remap the current ctx to hctx[0] which
1866 * is guaranteed to always have tags allocated
1868 if (!set
->tags
[hctx_idx
])
1872 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1873 hctx
= q
->mq_ops
->map_queue(q
, i
);
1874 cpumask_set_cpu(i
, hctx
->cpumask
);
1875 ctx
->index_hw
= hctx
->nr_ctx
;
1876 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1879 mutex_unlock(&q
->sysfs_lock
);
1881 queue_for_each_hw_ctx(q
, hctx
, i
) {
1882 struct blk_mq_ctxmap
*map
= &hctx
->ctx_map
;
1885 * If no software queues are mapped to this hardware queue,
1886 * disable it and free the request entries.
1888 if (!hctx
->nr_ctx
) {
1889 /* Never unmap queue 0. We need it as a
1890 * fallback in case of a new remap fails
1893 if (i
&& set
->tags
[i
]) {
1894 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1895 set
->tags
[i
] = NULL
;
1901 hctx
->tags
= set
->tags
[i
];
1902 WARN_ON(!hctx
->tags
);
1904 cpumask_copy(hctx
->tags
->cpumask
, hctx
->cpumask
);
1906 * Set the map size to the number of mapped software queues.
1907 * This is more accurate and more efficient than looping
1908 * over all possibly mapped software queues.
1910 map
->size
= DIV_ROUND_UP(hctx
->nr_ctx
, map
->bits_per_word
);
1913 * Initialize batch roundrobin counts
1915 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1916 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1920 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
1922 struct blk_mq_hw_ctx
*hctx
;
1925 queue_for_each_hw_ctx(q
, hctx
, i
) {
1927 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1929 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1933 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
1935 struct request_queue
*q
;
1937 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1938 blk_mq_freeze_queue(q
);
1939 queue_set_hctx_shared(q
, shared
);
1940 blk_mq_unfreeze_queue(q
);
1944 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1946 struct blk_mq_tag_set
*set
= q
->tag_set
;
1948 mutex_lock(&set
->tag_list_lock
);
1949 list_del_init(&q
->tag_set_list
);
1950 if (list_is_singular(&set
->tag_list
)) {
1951 /* just transitioned to unshared */
1952 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1953 /* update existing queue */
1954 blk_mq_update_tag_set_depth(set
, false);
1956 mutex_unlock(&set
->tag_list_lock
);
1959 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1960 struct request_queue
*q
)
1964 mutex_lock(&set
->tag_list_lock
);
1966 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1967 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1968 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
1969 /* update existing queue */
1970 blk_mq_update_tag_set_depth(set
, true);
1972 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
1973 queue_set_hctx_shared(q
, true);
1974 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1976 mutex_unlock(&set
->tag_list_lock
);
1980 * It is the actual release handler for mq, but we do it from
1981 * request queue's release handler for avoiding use-after-free
1982 * and headache because q->mq_kobj shouldn't have been introduced,
1983 * but we can't group ctx/kctx kobj without it.
1985 void blk_mq_release(struct request_queue
*q
)
1987 struct blk_mq_hw_ctx
*hctx
;
1990 /* hctx kobj stays in hctx */
1991 queue_for_each_hw_ctx(q
, hctx
, i
) {
2001 kfree(q
->queue_hw_ctx
);
2003 /* ctx kobj stays in queue_ctx */
2004 free_percpu(q
->queue_ctx
);
2007 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2009 struct request_queue
*uninit_q
, *q
;
2011 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2013 return ERR_PTR(-ENOMEM
);
2015 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2017 blk_cleanup_queue(uninit_q
);
2021 EXPORT_SYMBOL(blk_mq_init_queue
);
2023 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2024 struct request_queue
*q
)
2026 struct blk_mq_hw_ctx
**hctxs
;
2027 struct blk_mq_ctx __percpu
*ctx
;
2031 ctx
= alloc_percpu(struct blk_mq_ctx
);
2033 return ERR_PTR(-ENOMEM
);
2035 hctxs
= kmalloc_node(set
->nr_hw_queues
* sizeof(*hctxs
), GFP_KERNEL
,
2041 map
= blk_mq_make_queue_map(set
);
2045 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2046 int node
= blk_mq_hw_queue_to_node(map
, i
);
2048 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2053 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2057 atomic_set(&hctxs
[i
]->nr_active
, 0);
2058 hctxs
[i
]->numa_node
= node
;
2059 hctxs
[i
]->queue_num
= i
;
2062 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2063 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2065 q
->nr_queues
= nr_cpu_ids
;
2066 q
->nr_hw_queues
= set
->nr_hw_queues
;
2070 q
->queue_hw_ctx
= hctxs
;
2072 q
->mq_ops
= set
->ops
;
2073 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2075 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2076 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2078 q
->sg_reserved_size
= INT_MAX
;
2080 INIT_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2081 INIT_LIST_HEAD(&q
->requeue_list
);
2082 spin_lock_init(&q
->requeue_lock
);
2084 if (q
->nr_hw_queues
> 1)
2085 blk_queue_make_request(q
, blk_mq_make_request
);
2087 blk_queue_make_request(q
, blk_sq_make_request
);
2090 * Do this after blk_queue_make_request() overrides it...
2092 q
->nr_requests
= set
->queue_depth
;
2094 if (set
->ops
->complete
)
2095 blk_queue_softirq_done(q
, set
->ops
->complete
);
2097 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2099 if (blk_mq_init_hw_queues(q
, set
))
2103 mutex_lock(&all_q_mutex
);
2105 list_add_tail(&q
->all_q_node
, &all_q_list
);
2106 blk_mq_add_queue_tag_set(set
, q
);
2107 blk_mq_map_swqueue(q
, cpu_online_mask
);
2109 mutex_unlock(&all_q_mutex
);
2116 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2119 free_cpumask_var(hctxs
[i
]->cpumask
);
2126 return ERR_PTR(-ENOMEM
);
2128 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2130 void blk_mq_free_queue(struct request_queue
*q
)
2132 struct blk_mq_tag_set
*set
= q
->tag_set
;
2134 mutex_lock(&all_q_mutex
);
2135 list_del_init(&q
->all_q_node
);
2136 mutex_unlock(&all_q_mutex
);
2138 blk_mq_del_queue_tag_set(q
);
2140 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2141 blk_mq_free_hw_queues(q
, set
);
2144 /* Basically redo blk_mq_init_queue with queue frozen */
2145 static void blk_mq_queue_reinit(struct request_queue
*q
,
2146 const struct cpumask
*online_mask
)
2148 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2150 blk_mq_sysfs_unregister(q
);
2152 blk_mq_update_queue_map(q
->mq_map
, q
->nr_hw_queues
, online_mask
);
2155 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2156 * we should change hctx numa_node according to new topology (this
2157 * involves free and re-allocate memory, worthy doing?)
2160 blk_mq_map_swqueue(q
, online_mask
);
2162 blk_mq_sysfs_register(q
);
2165 static int blk_mq_queue_reinit_notify(struct notifier_block
*nb
,
2166 unsigned long action
, void *hcpu
)
2168 struct request_queue
*q
;
2169 int cpu
= (unsigned long)hcpu
;
2171 * New online cpumask which is going to be set in this hotplug event.
2172 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2173 * one-by-one and dynamically allocating this could result in a failure.
2175 static struct cpumask online_new
;
2178 * Before hotadded cpu starts handling requests, new mappings must
2179 * be established. Otherwise, these requests in hw queue might
2180 * never be dispatched.
2182 * For example, there is a single hw queue (hctx) and two CPU queues
2183 * (ctx0 for CPU0, and ctx1 for CPU1).
2185 * Now CPU1 is just onlined and a request is inserted into
2186 * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2189 * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2190 * set in pending bitmap and tries to retrieve requests in
2191 * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0,
2192 * so the request in ctx1->rq_list is ignored.
2194 switch (action
& ~CPU_TASKS_FROZEN
) {
2196 case CPU_UP_CANCELED
:
2197 cpumask_copy(&online_new
, cpu_online_mask
);
2199 case CPU_UP_PREPARE
:
2200 cpumask_copy(&online_new
, cpu_online_mask
);
2201 cpumask_set_cpu(cpu
, &online_new
);
2207 mutex_lock(&all_q_mutex
);
2210 * We need to freeze and reinit all existing queues. Freezing
2211 * involves synchronous wait for an RCU grace period and doing it
2212 * one by one may take a long time. Start freezing all queues in
2213 * one swoop and then wait for the completions so that freezing can
2214 * take place in parallel.
2216 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2217 blk_mq_freeze_queue_start(q
);
2218 list_for_each_entry(q
, &all_q_list
, all_q_node
) {
2219 blk_mq_freeze_queue_wait(q
);
2222 * timeout handler can't touch hw queue during the
2225 del_timer_sync(&q
->timeout
);
2228 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2229 blk_mq_queue_reinit(q
, &online_new
);
2231 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2232 blk_mq_unfreeze_queue(q
);
2234 mutex_unlock(&all_q_mutex
);
2238 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2242 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2243 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2252 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2258 * Allocate the request maps associated with this tag_set. Note that this
2259 * may reduce the depth asked for, if memory is tight. set->queue_depth
2260 * will be updated to reflect the allocated depth.
2262 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2267 depth
= set
->queue_depth
;
2269 err
= __blk_mq_alloc_rq_maps(set
);
2273 set
->queue_depth
>>= 1;
2274 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2278 } while (set
->queue_depth
);
2280 if (!set
->queue_depth
|| err
) {
2281 pr_err("blk-mq: failed to allocate request map\n");
2285 if (depth
!= set
->queue_depth
)
2286 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2287 depth
, set
->queue_depth
);
2292 struct cpumask
*blk_mq_tags_cpumask(struct blk_mq_tags
*tags
)
2294 return tags
->cpumask
;
2296 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask
);
2299 * Alloc a tag set to be associated with one or more request queues.
2300 * May fail with EINVAL for various error conditions. May adjust the
2301 * requested depth down, if if it too large. In that case, the set
2302 * value will be stored in set->queue_depth.
2304 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2306 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2308 if (!set
->nr_hw_queues
)
2310 if (!set
->queue_depth
)
2312 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2315 if (!set
->ops
->queue_rq
|| !set
->ops
->map_queue
)
2318 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2319 pr_info("blk-mq: reduced tag depth to %u\n",
2321 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2325 * If a crashdump is active, then we are potentially in a very
2326 * memory constrained environment. Limit us to 1 queue and
2327 * 64 tags to prevent using too much memory.
2329 if (is_kdump_kernel()) {
2330 set
->nr_hw_queues
= 1;
2331 set
->queue_depth
= min(64U, set
->queue_depth
);
2334 set
->tags
= kmalloc_node(set
->nr_hw_queues
*
2335 sizeof(struct blk_mq_tags
*),
2336 GFP_KERNEL
, set
->numa_node
);
2340 if (blk_mq_alloc_rq_maps(set
))
2343 mutex_init(&set
->tag_list_lock
);
2344 INIT_LIST_HEAD(&set
->tag_list
);
2352 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2354 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2358 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2360 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2366 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2368 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2370 struct blk_mq_tag_set
*set
= q
->tag_set
;
2371 struct blk_mq_hw_ctx
*hctx
;
2374 if (!set
|| nr
> set
->queue_depth
)
2378 queue_for_each_hw_ctx(q
, hctx
, i
) {
2379 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2385 q
->nr_requests
= nr
;
2390 void blk_mq_disable_hotplug(void)
2392 mutex_lock(&all_q_mutex
);
2395 void blk_mq_enable_hotplug(void)
2397 mutex_unlock(&all_q_mutex
);
2400 static int __init
blk_mq_init(void)
2404 hotcpu_notifier(blk_mq_queue_reinit_notify
, 0);
2408 subsys_initcall(blk_mq_init
);