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"
34 static DEFINE_MUTEX(all_q_mutex
);
35 static LIST_HEAD(all_q_list
);
38 * Check if any of the ctx's have pending work in this hardware queue
40 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
42 return sbitmap_any_bit_set(&hctx
->ctx_map
);
46 * Mark this ctx as having pending work in this hardware queue
48 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
49 struct blk_mq_ctx
*ctx
)
51 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
52 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
55 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
56 struct blk_mq_ctx
*ctx
)
58 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
61 void blk_mq_freeze_queue_start(struct request_queue
*q
)
65 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
66 if (freeze_depth
== 1) {
67 percpu_ref_kill(&q
->q_usage_counter
);
68 blk_mq_run_hw_queues(q
, false);
71 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start
);
73 static void blk_mq_freeze_queue_wait(struct request_queue
*q
)
75 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
79 * Guarantee no request is in use, so we can change any data structure of
80 * the queue afterward.
82 void blk_freeze_queue(struct request_queue
*q
)
85 * In the !blk_mq case we are only calling this to kill the
86 * q_usage_counter, otherwise this increases the freeze depth
87 * and waits for it to return to zero. For this reason there is
88 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
89 * exported to drivers as the only user for unfreeze is blk_mq.
91 blk_mq_freeze_queue_start(q
);
92 blk_mq_freeze_queue_wait(q
);
95 void blk_mq_freeze_queue(struct request_queue
*q
)
98 * ...just an alias to keep freeze and unfreeze actions balanced
99 * in the blk_mq_* namespace
103 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
105 void blk_mq_unfreeze_queue(struct request_queue
*q
)
109 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
110 WARN_ON_ONCE(freeze_depth
< 0);
112 percpu_ref_reinit(&q
->q_usage_counter
);
113 wake_up_all(&q
->mq_freeze_wq
);
116 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
118 void blk_mq_wake_waiters(struct request_queue
*q
)
120 struct blk_mq_hw_ctx
*hctx
;
123 queue_for_each_hw_ctx(q
, hctx
, i
)
124 if (blk_mq_hw_queue_mapped(hctx
))
125 blk_mq_tag_wakeup_all(hctx
->tags
, true);
128 * If we are called because the queue has now been marked as
129 * dying, we need to ensure that processes currently waiting on
130 * the queue are notified as well.
132 wake_up_all(&q
->mq_freeze_wq
);
135 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
137 return blk_mq_has_free_tags(hctx
->tags
);
139 EXPORT_SYMBOL(blk_mq_can_queue
);
141 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
142 struct request
*rq
, int op
,
143 unsigned int op_flags
)
145 INIT_LIST_HEAD(&rq
->queuelist
);
146 /* csd/requeue_work/fifo_time is initialized before use */
149 req_set_op_attrs(rq
, op
, op_flags
);
150 if (blk_queue_io_stat(q
))
151 rq
->rq_flags
|= RQF_IO_STAT
;
152 /* do not touch atomic flags, it needs atomic ops against the timer */
154 INIT_HLIST_NODE(&rq
->hash
);
155 RB_CLEAR_NODE(&rq
->rb_node
);
158 rq
->start_time
= jiffies
;
159 #ifdef CONFIG_BLK_CGROUP
161 set_start_time_ns(rq
);
162 rq
->io_start_time_ns
= 0;
164 rq
->nr_phys_segments
= 0;
165 #if defined(CONFIG_BLK_DEV_INTEGRITY)
166 rq
->nr_integrity_segments
= 0;
169 /* tag was already set */
179 INIT_LIST_HEAD(&rq
->timeout_list
);
183 rq
->end_io_data
= NULL
;
186 ctx
->rq_dispatched
[rw_is_sync(op
, op_flags
)]++;
189 static struct request
*
190 __blk_mq_alloc_request(struct blk_mq_alloc_data
*data
, int op
, int op_flags
)
195 tag
= blk_mq_get_tag(data
);
196 if (tag
!= BLK_MQ_TAG_FAIL
) {
197 rq
= data
->hctx
->tags
->rqs
[tag
];
199 if (blk_mq_tag_busy(data
->hctx
)) {
200 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
201 atomic_inc(&data
->hctx
->nr_active
);
205 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
, op_flags
);
212 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
215 struct blk_mq_ctx
*ctx
;
216 struct blk_mq_hw_ctx
*hctx
;
218 struct blk_mq_alloc_data alloc_data
;
221 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
225 ctx
= blk_mq_get_ctx(q
);
226 hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
227 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
228 rq
= __blk_mq_alloc_request(&alloc_data
, rw
, 0);
233 return ERR_PTR(-EWOULDBLOCK
);
237 rq
->__sector
= (sector_t
) -1;
238 rq
->bio
= rq
->biotail
= NULL
;
241 EXPORT_SYMBOL(blk_mq_alloc_request
);
243 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
244 unsigned int flags
, unsigned int hctx_idx
)
246 struct blk_mq_hw_ctx
*hctx
;
247 struct blk_mq_ctx
*ctx
;
249 struct blk_mq_alloc_data alloc_data
;
253 * If the tag allocator sleeps we could get an allocation for a
254 * different hardware context. No need to complicate the low level
255 * allocator for this for the rare use case of a command tied to
258 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
259 return ERR_PTR(-EINVAL
);
261 if (hctx_idx
>= q
->nr_hw_queues
)
262 return ERR_PTR(-EIO
);
264 ret
= blk_queue_enter(q
, true);
269 * Check if the hardware context is actually mapped to anything.
270 * If not tell the caller that it should skip this queue.
272 hctx
= q
->queue_hw_ctx
[hctx_idx
];
273 if (!blk_mq_hw_queue_mapped(hctx
)) {
277 ctx
= __blk_mq_get_ctx(q
, cpumask_first(hctx
->cpumask
));
279 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
280 rq
= __blk_mq_alloc_request(&alloc_data
, rw
, 0);
292 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
294 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
295 struct blk_mq_ctx
*ctx
, struct request
*rq
)
297 const int tag
= rq
->tag
;
298 struct request_queue
*q
= rq
->q
;
300 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
301 atomic_dec(&hctx
->nr_active
);
304 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
305 blk_mq_put_tag(hctx
, ctx
, tag
);
309 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
311 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
313 ctx
->rq_completed
[rq_is_sync(rq
)]++;
314 __blk_mq_free_request(hctx
, ctx
, rq
);
317 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request
);
319 void blk_mq_free_request(struct request
*rq
)
321 blk_mq_free_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
323 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
325 inline void __blk_mq_end_request(struct request
*rq
, int error
)
327 blk_account_io_done(rq
);
330 rq
->end_io(rq
, error
);
332 if (unlikely(blk_bidi_rq(rq
)))
333 blk_mq_free_request(rq
->next_rq
);
334 blk_mq_free_request(rq
);
337 EXPORT_SYMBOL(__blk_mq_end_request
);
339 void blk_mq_end_request(struct request
*rq
, int error
)
341 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
343 __blk_mq_end_request(rq
, error
);
345 EXPORT_SYMBOL(blk_mq_end_request
);
347 static void __blk_mq_complete_request_remote(void *data
)
349 struct request
*rq
= data
;
351 rq
->q
->softirq_done_fn(rq
);
354 static void blk_mq_ipi_complete_request(struct request
*rq
)
356 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
360 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
361 rq
->q
->softirq_done_fn(rq
);
366 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
367 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
369 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
370 rq
->csd
.func
= __blk_mq_complete_request_remote
;
373 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
375 rq
->q
->softirq_done_fn(rq
);
380 static void __blk_mq_complete_request(struct request
*rq
)
382 struct request_queue
*q
= rq
->q
;
384 if (!q
->softirq_done_fn
)
385 blk_mq_end_request(rq
, rq
->errors
);
387 blk_mq_ipi_complete_request(rq
);
391 * blk_mq_complete_request - end I/O on a request
392 * @rq: the request being processed
395 * Ends all I/O on a request. It does not handle partial completions.
396 * The actual completion happens out-of-order, through a IPI handler.
398 void blk_mq_complete_request(struct request
*rq
, int error
)
400 struct request_queue
*q
= rq
->q
;
402 if (unlikely(blk_should_fake_timeout(q
)))
404 if (!blk_mark_rq_complete(rq
)) {
406 __blk_mq_complete_request(rq
);
409 EXPORT_SYMBOL(blk_mq_complete_request
);
411 int blk_mq_request_started(struct request
*rq
)
413 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
415 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
417 void blk_mq_start_request(struct request
*rq
)
419 struct request_queue
*q
= rq
->q
;
421 trace_block_rq_issue(q
, rq
);
423 rq
->resid_len
= blk_rq_bytes(rq
);
424 if (unlikely(blk_bidi_rq(rq
)))
425 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
430 * Ensure that ->deadline is visible before set the started
431 * flag and clear the completed flag.
433 smp_mb__before_atomic();
436 * Mark us as started and clear complete. Complete might have been
437 * set if requeue raced with timeout, which then marked it as
438 * complete. So be sure to clear complete again when we start
439 * the request, otherwise we'll ignore the completion event.
441 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
442 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
443 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
444 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
446 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
448 * Make sure space for the drain appears. We know we can do
449 * this because max_hw_segments has been adjusted to be one
450 * fewer than the device can handle.
452 rq
->nr_phys_segments
++;
455 EXPORT_SYMBOL(blk_mq_start_request
);
457 static void __blk_mq_requeue_request(struct request
*rq
)
459 struct request_queue
*q
= rq
->q
;
461 trace_block_rq_requeue(q
, rq
);
463 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
464 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
465 rq
->nr_phys_segments
--;
469 void blk_mq_requeue_request(struct request
*rq
)
471 __blk_mq_requeue_request(rq
);
473 BUG_ON(blk_queued_rq(rq
));
474 blk_mq_add_to_requeue_list(rq
, true);
476 EXPORT_SYMBOL(blk_mq_requeue_request
);
478 static void blk_mq_requeue_work(struct work_struct
*work
)
480 struct request_queue
*q
=
481 container_of(work
, struct request_queue
, requeue_work
.work
);
483 struct request
*rq
, *next
;
486 spin_lock_irqsave(&q
->requeue_lock
, flags
);
487 list_splice_init(&q
->requeue_list
, &rq_list
);
488 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
490 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
491 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
494 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
495 list_del_init(&rq
->queuelist
);
496 blk_mq_insert_request(rq
, true, false, false);
499 while (!list_empty(&rq_list
)) {
500 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
501 list_del_init(&rq
->queuelist
);
502 blk_mq_insert_request(rq
, false, false, false);
506 * Use the start variant of queue running here, so that running
507 * the requeue work will kick stopped queues.
509 blk_mq_start_hw_queues(q
);
512 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
)
514 struct request_queue
*q
= rq
->q
;
518 * We abuse this flag that is otherwise used by the I/O scheduler to
519 * request head insertation from the workqueue.
521 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
523 spin_lock_irqsave(&q
->requeue_lock
, flags
);
525 rq
->rq_flags
|= RQF_SOFTBARRIER
;
526 list_add(&rq
->queuelist
, &q
->requeue_list
);
528 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
530 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
532 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
534 void blk_mq_cancel_requeue_work(struct request_queue
*q
)
536 cancel_delayed_work_sync(&q
->requeue_work
);
538 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work
);
540 void blk_mq_kick_requeue_list(struct request_queue
*q
)
542 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
544 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
546 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
549 kblockd_schedule_delayed_work(&q
->requeue_work
,
550 msecs_to_jiffies(msecs
));
552 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
554 void blk_mq_abort_requeue_list(struct request_queue
*q
)
559 spin_lock_irqsave(&q
->requeue_lock
, flags
);
560 list_splice_init(&q
->requeue_list
, &rq_list
);
561 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
563 while (!list_empty(&rq_list
)) {
566 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
567 list_del_init(&rq
->queuelist
);
569 blk_mq_end_request(rq
, rq
->errors
);
572 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
574 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
576 if (tag
< tags
->nr_tags
) {
577 prefetch(tags
->rqs
[tag
]);
578 return tags
->rqs
[tag
];
583 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
585 struct blk_mq_timeout_data
{
587 unsigned int next_set
;
590 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
592 struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
593 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
596 * We know that complete is set at this point. If STARTED isn't set
597 * anymore, then the request isn't active and the "timeout" should
598 * just be ignored. This can happen due to the bitflag ordering.
599 * Timeout first checks if STARTED is set, and if it is, assumes
600 * the request is active. But if we race with completion, then
601 * we both flags will get cleared. So check here again, and ignore
602 * a timeout event with a request that isn't active.
604 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
608 ret
= ops
->timeout(req
, reserved
);
612 __blk_mq_complete_request(req
);
614 case BLK_EH_RESET_TIMER
:
616 blk_clear_rq_complete(req
);
618 case BLK_EH_NOT_HANDLED
:
621 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
626 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
627 struct request
*rq
, void *priv
, bool reserved
)
629 struct blk_mq_timeout_data
*data
= priv
;
631 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
633 * If a request wasn't started before the queue was
634 * marked dying, kill it here or it'll go unnoticed.
636 if (unlikely(blk_queue_dying(rq
->q
))) {
638 blk_mq_end_request(rq
, rq
->errors
);
643 if (time_after_eq(jiffies
, rq
->deadline
)) {
644 if (!blk_mark_rq_complete(rq
))
645 blk_mq_rq_timed_out(rq
, reserved
);
646 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
647 data
->next
= rq
->deadline
;
652 static void blk_mq_timeout_work(struct work_struct
*work
)
654 struct request_queue
*q
=
655 container_of(work
, struct request_queue
, timeout_work
);
656 struct blk_mq_timeout_data data
= {
662 /* A deadlock might occur if a request is stuck requiring a
663 * timeout at the same time a queue freeze is waiting
664 * completion, since the timeout code would not be able to
665 * acquire the queue reference here.
667 * That's why we don't use blk_queue_enter here; instead, we use
668 * percpu_ref_tryget directly, because we need to be able to
669 * obtain a reference even in the short window between the queue
670 * starting to freeze, by dropping the first reference in
671 * blk_mq_freeze_queue_start, and the moment the last request is
672 * consumed, marked by the instant q_usage_counter reaches
675 if (!percpu_ref_tryget(&q
->q_usage_counter
))
678 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
681 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
682 mod_timer(&q
->timeout
, data
.next
);
684 struct blk_mq_hw_ctx
*hctx
;
686 queue_for_each_hw_ctx(q
, hctx
, i
) {
687 /* the hctx may be unmapped, so check it here */
688 if (blk_mq_hw_queue_mapped(hctx
))
689 blk_mq_tag_idle(hctx
);
696 * Reverse check our software queue for entries that we could potentially
697 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
698 * too much time checking for merges.
700 static bool blk_mq_attempt_merge(struct request_queue
*q
,
701 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
706 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
712 if (!blk_rq_merge_ok(rq
, bio
))
715 el_ret
= blk_try_merge(rq
, bio
);
716 if (el_ret
== ELEVATOR_BACK_MERGE
) {
717 if (bio_attempt_back_merge(q
, rq
, bio
)) {
722 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
723 if (bio_attempt_front_merge(q
, rq
, bio
)) {
734 struct flush_busy_ctx_data
{
735 struct blk_mq_hw_ctx
*hctx
;
736 struct list_head
*list
;
739 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
741 struct flush_busy_ctx_data
*flush_data
= data
;
742 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
743 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
745 sbitmap_clear_bit(sb
, bitnr
);
746 spin_lock(&ctx
->lock
);
747 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
748 spin_unlock(&ctx
->lock
);
753 * Process software queues that have been marked busy, splicing them
754 * to the for-dispatch
756 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
758 struct flush_busy_ctx_data data
= {
763 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
766 static inline unsigned int queued_to_index(unsigned int queued
)
771 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
775 * Run this hardware queue, pulling any software queues mapped to it in.
776 * Note that this function currently has various problems around ordering
777 * of IO. In particular, we'd like FIFO behaviour on handling existing
778 * items on the hctx->dispatch list. Ignore that for now.
780 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
782 struct request_queue
*q
= hctx
->queue
;
785 LIST_HEAD(driver_list
);
786 struct list_head
*dptr
;
789 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)))
792 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
793 cpu_online(hctx
->next_cpu
));
798 * Touch any software queue that has pending entries.
800 flush_busy_ctxs(hctx
, &rq_list
);
803 * If we have previous entries on our dispatch list, grab them
804 * and stuff them at the front for more fair dispatch.
806 if (!list_empty_careful(&hctx
->dispatch
)) {
807 spin_lock(&hctx
->lock
);
808 if (!list_empty(&hctx
->dispatch
))
809 list_splice_init(&hctx
->dispatch
, &rq_list
);
810 spin_unlock(&hctx
->lock
);
814 * Start off with dptr being NULL, so we start the first request
815 * immediately, even if we have more pending.
820 * Now process all the entries, sending them to the driver.
823 while (!list_empty(&rq_list
)) {
824 struct blk_mq_queue_data bd
;
827 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
828 list_del_init(&rq
->queuelist
);
832 bd
.last
= list_empty(&rq_list
);
834 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
836 case BLK_MQ_RQ_QUEUE_OK
:
839 case BLK_MQ_RQ_QUEUE_BUSY
:
840 list_add(&rq
->queuelist
, &rq_list
);
841 __blk_mq_requeue_request(rq
);
844 pr_err("blk-mq: bad return on queue: %d\n", ret
);
845 case BLK_MQ_RQ_QUEUE_ERROR
:
847 blk_mq_end_request(rq
, rq
->errors
);
851 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
855 * We've done the first request. If we have more than 1
856 * left in the list, set dptr to defer issue.
858 if (!dptr
&& rq_list
.next
!= rq_list
.prev
)
862 hctx
->dispatched
[queued_to_index(queued
)]++;
865 * Any items that need requeuing? Stuff them into hctx->dispatch,
866 * that is where we will continue on next queue run.
868 if (!list_empty(&rq_list
)) {
869 spin_lock(&hctx
->lock
);
870 list_splice(&rq_list
, &hctx
->dispatch
);
871 spin_unlock(&hctx
->lock
);
873 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
874 * it's possible the queue is stopped and restarted again
875 * before this. Queue restart will dispatch requests. And since
876 * requests in rq_list aren't added into hctx->dispatch yet,
877 * the requests in rq_list might get lost.
879 * blk_mq_run_hw_queue() already checks the STOPPED bit
881 blk_mq_run_hw_queue(hctx
, true);
886 * It'd be great if the workqueue API had a way to pass
887 * in a mask and had some smarts for more clever placement.
888 * For now we just round-robin here, switching for every
889 * BLK_MQ_CPU_WORK_BATCH queued items.
891 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
893 if (hctx
->queue
->nr_hw_queues
== 1)
894 return WORK_CPU_UNBOUND
;
896 if (--hctx
->next_cpu_batch
<= 0) {
897 int cpu
= hctx
->next_cpu
, next_cpu
;
899 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
900 if (next_cpu
>= nr_cpu_ids
)
901 next_cpu
= cpumask_first(hctx
->cpumask
);
903 hctx
->next_cpu
= next_cpu
;
904 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
909 return hctx
->next_cpu
;
912 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
914 if (unlikely(test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
) ||
915 !blk_mq_hw_queue_mapped(hctx
)))
918 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
920 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
921 __blk_mq_run_hw_queue(hctx
);
929 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
);
932 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
934 struct blk_mq_hw_ctx
*hctx
;
937 queue_for_each_hw_ctx(q
, hctx
, i
) {
938 if ((!blk_mq_hctx_has_pending(hctx
) &&
939 list_empty_careful(&hctx
->dispatch
)) ||
940 test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
943 blk_mq_run_hw_queue(hctx
, async
);
946 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
948 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
950 cancel_work(&hctx
->run_work
);
951 cancel_delayed_work(&hctx
->delay_work
);
952 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
954 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
956 void blk_mq_stop_hw_queues(struct request_queue
*q
)
958 struct blk_mq_hw_ctx
*hctx
;
961 queue_for_each_hw_ctx(q
, hctx
, i
)
962 blk_mq_stop_hw_queue(hctx
);
964 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
966 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
968 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
970 blk_mq_run_hw_queue(hctx
, false);
972 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
974 void blk_mq_start_hw_queues(struct request_queue
*q
)
976 struct blk_mq_hw_ctx
*hctx
;
979 queue_for_each_hw_ctx(q
, hctx
, i
)
980 blk_mq_start_hw_queue(hctx
);
982 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
984 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
986 struct blk_mq_hw_ctx
*hctx
;
989 queue_for_each_hw_ctx(q
, hctx
, i
) {
990 if (!test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
993 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
994 blk_mq_run_hw_queue(hctx
, async
);
997 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
999 static void blk_mq_run_work_fn(struct work_struct
*work
)
1001 struct blk_mq_hw_ctx
*hctx
;
1003 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1005 __blk_mq_run_hw_queue(hctx
);
1008 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1010 struct blk_mq_hw_ctx
*hctx
;
1012 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1014 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1015 __blk_mq_run_hw_queue(hctx
);
1018 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1020 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1023 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1024 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1026 EXPORT_SYMBOL(blk_mq_delay_queue
);
1028 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1032 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1034 trace_block_rq_insert(hctx
->queue
, rq
);
1037 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1039 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1042 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
1043 struct request
*rq
, bool at_head
)
1045 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1047 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1048 blk_mq_hctx_mark_pending(hctx
, ctx
);
1051 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
1054 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1055 struct request_queue
*q
= rq
->q
;
1056 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1058 spin_lock(&ctx
->lock
);
1059 __blk_mq_insert_request(hctx
, rq
, at_head
);
1060 spin_unlock(&ctx
->lock
);
1063 blk_mq_run_hw_queue(hctx
, async
);
1066 static void blk_mq_insert_requests(struct request_queue
*q
,
1067 struct blk_mq_ctx
*ctx
,
1068 struct list_head
*list
,
1073 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1075 trace_block_unplug(q
, depth
, !from_schedule
);
1078 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1081 spin_lock(&ctx
->lock
);
1082 while (!list_empty(list
)) {
1085 rq
= list_first_entry(list
, struct request
, queuelist
);
1086 BUG_ON(rq
->mq_ctx
!= ctx
);
1087 list_del_init(&rq
->queuelist
);
1088 __blk_mq_insert_req_list(hctx
, rq
, false);
1090 blk_mq_hctx_mark_pending(hctx
, ctx
);
1091 spin_unlock(&ctx
->lock
);
1093 blk_mq_run_hw_queue(hctx
, from_schedule
);
1096 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1098 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1099 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1101 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1102 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1103 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1106 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1108 struct blk_mq_ctx
*this_ctx
;
1109 struct request_queue
*this_q
;
1112 LIST_HEAD(ctx_list
);
1115 list_splice_init(&plug
->mq_list
, &list
);
1117 list_sort(NULL
, &list
, plug_ctx_cmp
);
1123 while (!list_empty(&list
)) {
1124 rq
= list_entry_rq(list
.next
);
1125 list_del_init(&rq
->queuelist
);
1127 if (rq
->mq_ctx
!= this_ctx
) {
1129 blk_mq_insert_requests(this_q
, this_ctx
,
1134 this_ctx
= rq
->mq_ctx
;
1140 list_add_tail(&rq
->queuelist
, &ctx_list
);
1144 * If 'this_ctx' is set, we know we have entries to complete
1145 * on 'ctx_list'. Do those.
1148 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1153 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1155 init_request_from_bio(rq
, bio
);
1157 blk_account_io_start(rq
, 1);
1160 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1162 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1163 !blk_queue_nomerges(hctx
->queue
);
1166 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1167 struct blk_mq_ctx
*ctx
,
1168 struct request
*rq
, struct bio
*bio
)
1170 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1171 blk_mq_bio_to_request(rq
, bio
);
1172 spin_lock(&ctx
->lock
);
1174 __blk_mq_insert_request(hctx
, rq
, false);
1175 spin_unlock(&ctx
->lock
);
1178 struct request_queue
*q
= hctx
->queue
;
1180 spin_lock(&ctx
->lock
);
1181 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1182 blk_mq_bio_to_request(rq
, bio
);
1186 spin_unlock(&ctx
->lock
);
1187 __blk_mq_free_request(hctx
, ctx
, rq
);
1192 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1194 struct blk_mq_alloc_data
*data
)
1196 struct blk_mq_hw_ctx
*hctx
;
1197 struct blk_mq_ctx
*ctx
;
1199 int op
= bio_data_dir(bio
);
1202 blk_queue_enter_live(q
);
1203 ctx
= blk_mq_get_ctx(q
);
1204 hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1206 if (rw_is_sync(bio_op(bio
), bio
->bi_opf
))
1207 op_flags
|= REQ_SYNC
;
1209 trace_block_getrq(q
, bio
, op
);
1210 blk_mq_set_alloc_data(data
, q
, 0, ctx
, hctx
);
1211 rq
= __blk_mq_alloc_request(data
, op
, op_flags
);
1213 data
->hctx
->queued
++;
1217 static int blk_mq_direct_issue_request(struct request
*rq
, blk_qc_t
*cookie
)
1220 struct request_queue
*q
= rq
->q
;
1221 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, rq
->mq_ctx
->cpu
);
1222 struct blk_mq_queue_data bd
= {
1227 blk_qc_t new_cookie
= blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
);
1230 * For OK queue, we are done. For error, kill it. Any other
1231 * error (busy), just add it to our list as we previously
1234 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1235 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1236 *cookie
= new_cookie
;
1240 __blk_mq_requeue_request(rq
);
1242 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1243 *cookie
= BLK_QC_T_NONE
;
1245 blk_mq_end_request(rq
, rq
->errors
);
1253 * Multiple hardware queue variant. This will not use per-process plugs,
1254 * but will attempt to bypass the hctx queueing if we can go straight to
1255 * hardware for SYNC IO.
1257 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1259 const int is_sync
= rw_is_sync(bio_op(bio
), bio
->bi_opf
);
1260 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1261 struct blk_mq_alloc_data data
;
1263 unsigned int request_count
= 0;
1264 struct blk_plug
*plug
;
1265 struct request
*same_queue_rq
= NULL
;
1268 blk_queue_bounce(q
, &bio
);
1270 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1272 return BLK_QC_T_NONE
;
1275 blk_queue_split(q
, &bio
, q
->bio_split
);
1277 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1278 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1279 return BLK_QC_T_NONE
;
1281 rq
= blk_mq_map_request(q
, bio
, &data
);
1283 return BLK_QC_T_NONE
;
1285 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1287 if (unlikely(is_flush_fua
)) {
1288 blk_mq_bio_to_request(rq
, bio
);
1289 blk_insert_flush(rq
);
1293 plug
= current
->plug
;
1295 * If the driver supports defer issued based on 'last', then
1296 * queue it up like normal since we can potentially save some
1299 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1300 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1301 struct request
*old_rq
= NULL
;
1303 blk_mq_bio_to_request(rq
, bio
);
1306 * We do limited pluging. If the bio can be merged, do that.
1307 * Otherwise the existing request in the plug list will be
1308 * issued. So the plug list will have one request at most
1312 * The plug list might get flushed before this. If that
1313 * happens, same_queue_rq is invalid and plug list is
1316 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1317 old_rq
= same_queue_rq
;
1318 list_del_init(&old_rq
->queuelist
);
1320 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1321 } else /* is_sync */
1323 blk_mq_put_ctx(data
.ctx
);
1326 if (!blk_mq_direct_issue_request(old_rq
, &cookie
))
1328 blk_mq_insert_request(old_rq
, false, true, true);
1332 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1334 * For a SYNC request, send it to the hardware immediately. For
1335 * an ASYNC request, just ensure that we run it later on. The
1336 * latter allows for merging opportunities and more efficient
1340 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1342 blk_mq_put_ctx(data
.ctx
);
1348 * Single hardware queue variant. This will attempt to use any per-process
1349 * plug for merging and IO deferral.
1351 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1353 const int is_sync
= rw_is_sync(bio_op(bio
), bio
->bi_opf
);
1354 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1355 struct blk_plug
*plug
;
1356 unsigned int request_count
= 0;
1357 struct blk_mq_alloc_data data
;
1361 blk_queue_bounce(q
, &bio
);
1363 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1365 return BLK_QC_T_NONE
;
1368 blk_queue_split(q
, &bio
, q
->bio_split
);
1370 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1371 if (blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1372 return BLK_QC_T_NONE
;
1374 request_count
= blk_plug_queued_count(q
);
1376 rq
= blk_mq_map_request(q
, bio
, &data
);
1378 return BLK_QC_T_NONE
;
1380 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1382 if (unlikely(is_flush_fua
)) {
1383 blk_mq_bio_to_request(rq
, bio
);
1384 blk_insert_flush(rq
);
1389 * A task plug currently exists. Since this is completely lockless,
1390 * utilize that to temporarily store requests until the task is
1391 * either done or scheduled away.
1393 plug
= current
->plug
;
1395 blk_mq_bio_to_request(rq
, bio
);
1397 trace_block_plug(q
);
1399 blk_mq_put_ctx(data
.ctx
);
1401 if (request_count
>= BLK_MAX_REQUEST_COUNT
) {
1402 blk_flush_plug_list(plug
, false);
1403 trace_block_plug(q
);
1406 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1410 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1412 * For a SYNC request, send it to the hardware immediately. For
1413 * an ASYNC request, just ensure that we run it later on. The
1414 * latter allows for merging opportunities and more efficient
1418 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1421 blk_mq_put_ctx(data
.ctx
);
1425 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1426 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1430 if (tags
->rqs
&& set
->ops
->exit_request
) {
1433 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1436 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1438 tags
->rqs
[i
] = NULL
;
1442 while (!list_empty(&tags
->page_list
)) {
1443 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1444 list_del_init(&page
->lru
);
1446 * Remove kmemleak object previously allocated in
1447 * blk_mq_init_rq_map().
1449 kmemleak_free(page_address(page
));
1450 __free_pages(page
, page
->private);
1455 blk_mq_free_tags(tags
);
1458 static size_t order_to_size(unsigned int order
)
1460 return (size_t)PAGE_SIZE
<< order
;
1463 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1464 unsigned int hctx_idx
)
1466 struct blk_mq_tags
*tags
;
1467 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1468 size_t rq_size
, left
;
1470 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1472 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1476 INIT_LIST_HEAD(&tags
->page_list
);
1478 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1479 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1482 blk_mq_free_tags(tags
);
1487 * rq_size is the size of the request plus driver payload, rounded
1488 * to the cacheline size
1490 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1492 left
= rq_size
* set
->queue_depth
;
1494 for (i
= 0; i
< set
->queue_depth
; ) {
1495 int this_order
= max_order
;
1500 while (this_order
&& left
< order_to_size(this_order
- 1))
1504 page
= alloc_pages_node(set
->numa_node
,
1505 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1511 if (order_to_size(this_order
) < rq_size
)
1518 page
->private = this_order
;
1519 list_add_tail(&page
->lru
, &tags
->page_list
);
1521 p
= page_address(page
);
1523 * Allow kmemleak to scan these pages as they contain pointers
1524 * to additional allocations like via ops->init_request().
1526 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_KERNEL
);
1527 entries_per_page
= order_to_size(this_order
) / rq_size
;
1528 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1529 left
-= to_do
* rq_size
;
1530 for (j
= 0; j
< to_do
; j
++) {
1532 if (set
->ops
->init_request
) {
1533 if (set
->ops
->init_request(set
->driver_data
,
1534 tags
->rqs
[i
], hctx_idx
, i
,
1536 tags
->rqs
[i
] = NULL
;
1548 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1553 * 'cpu' is going away. splice any existing rq_list entries from this
1554 * software queue to the hw queue dispatch list, and ensure that it
1557 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1559 struct blk_mq_hw_ctx
*hctx
;
1560 struct blk_mq_ctx
*ctx
;
1563 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1564 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1566 spin_lock(&ctx
->lock
);
1567 if (!list_empty(&ctx
->rq_list
)) {
1568 list_splice_init(&ctx
->rq_list
, &tmp
);
1569 blk_mq_hctx_clear_pending(hctx
, ctx
);
1571 spin_unlock(&ctx
->lock
);
1573 if (list_empty(&tmp
))
1576 spin_lock(&hctx
->lock
);
1577 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1578 spin_unlock(&hctx
->lock
);
1580 blk_mq_run_hw_queue(hctx
, true);
1584 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1586 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1590 /* hctx->ctxs will be freed in queue's release handler */
1591 static void blk_mq_exit_hctx(struct request_queue
*q
,
1592 struct blk_mq_tag_set
*set
,
1593 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1595 unsigned flush_start_tag
= set
->queue_depth
;
1597 blk_mq_tag_idle(hctx
);
1599 if (set
->ops
->exit_request
)
1600 set
->ops
->exit_request(set
->driver_data
,
1601 hctx
->fq
->flush_rq
, hctx_idx
,
1602 flush_start_tag
+ hctx_idx
);
1604 if (set
->ops
->exit_hctx
)
1605 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1607 blk_mq_remove_cpuhp(hctx
);
1608 blk_free_flush_queue(hctx
->fq
);
1609 sbitmap_free(&hctx
->ctx_map
);
1612 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1613 struct blk_mq_tag_set
*set
, int nr_queue
)
1615 struct blk_mq_hw_ctx
*hctx
;
1618 queue_for_each_hw_ctx(q
, hctx
, i
) {
1621 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1625 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1626 struct blk_mq_tag_set
*set
)
1628 struct blk_mq_hw_ctx
*hctx
;
1631 queue_for_each_hw_ctx(q
, hctx
, i
)
1632 free_cpumask_var(hctx
->cpumask
);
1635 static int blk_mq_init_hctx(struct request_queue
*q
,
1636 struct blk_mq_tag_set
*set
,
1637 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1640 unsigned flush_start_tag
= set
->queue_depth
;
1642 node
= hctx
->numa_node
;
1643 if (node
== NUMA_NO_NODE
)
1644 node
= hctx
->numa_node
= set
->numa_node
;
1646 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1647 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1648 spin_lock_init(&hctx
->lock
);
1649 INIT_LIST_HEAD(&hctx
->dispatch
);
1651 hctx
->queue_num
= hctx_idx
;
1652 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1654 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1656 hctx
->tags
= set
->tags
[hctx_idx
];
1659 * Allocate space for all possible cpus to avoid allocation at
1662 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1665 goto unregister_cpu_notifier
;
1667 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1673 if (set
->ops
->init_hctx
&&
1674 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1677 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1681 if (set
->ops
->init_request
&&
1682 set
->ops
->init_request(set
->driver_data
,
1683 hctx
->fq
->flush_rq
, hctx_idx
,
1684 flush_start_tag
+ hctx_idx
, node
))
1692 if (set
->ops
->exit_hctx
)
1693 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1695 sbitmap_free(&hctx
->ctx_map
);
1698 unregister_cpu_notifier
:
1699 blk_mq_remove_cpuhp(hctx
);
1703 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1704 unsigned int nr_hw_queues
)
1708 for_each_possible_cpu(i
) {
1709 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1710 struct blk_mq_hw_ctx
*hctx
;
1712 memset(__ctx
, 0, sizeof(*__ctx
));
1714 spin_lock_init(&__ctx
->lock
);
1715 INIT_LIST_HEAD(&__ctx
->rq_list
);
1718 /* If the cpu isn't online, the cpu is mapped to first hctx */
1722 hctx
= blk_mq_map_queue(q
, i
);
1725 * Set local node, IFF we have more than one hw queue. If
1726 * not, we remain on the home node of the device
1728 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1729 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1733 static void blk_mq_map_swqueue(struct request_queue
*q
,
1734 const struct cpumask
*online_mask
)
1737 struct blk_mq_hw_ctx
*hctx
;
1738 struct blk_mq_ctx
*ctx
;
1739 struct blk_mq_tag_set
*set
= q
->tag_set
;
1742 * Avoid others reading imcomplete hctx->cpumask through sysfs
1744 mutex_lock(&q
->sysfs_lock
);
1746 queue_for_each_hw_ctx(q
, hctx
, i
) {
1747 cpumask_clear(hctx
->cpumask
);
1752 * Map software to hardware queues
1754 for_each_possible_cpu(i
) {
1755 /* If the cpu isn't online, the cpu is mapped to first hctx */
1756 if (!cpumask_test_cpu(i
, online_mask
))
1759 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1760 hctx
= blk_mq_map_queue(q
, i
);
1762 cpumask_set_cpu(i
, hctx
->cpumask
);
1763 ctx
->index_hw
= hctx
->nr_ctx
;
1764 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1767 mutex_unlock(&q
->sysfs_lock
);
1769 queue_for_each_hw_ctx(q
, hctx
, i
) {
1771 * If no software queues are mapped to this hardware queue,
1772 * disable it and free the request entries.
1774 if (!hctx
->nr_ctx
) {
1776 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1777 set
->tags
[i
] = NULL
;
1783 /* unmapped hw queue can be remapped after CPU topo changed */
1785 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1786 hctx
->tags
= set
->tags
[i
];
1787 WARN_ON(!hctx
->tags
);
1790 * Set the map size to the number of mapped software queues.
1791 * This is more accurate and more efficient than looping
1792 * over all possibly mapped software queues.
1794 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
1797 * Initialize batch roundrobin counts
1799 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1800 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1804 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
1806 struct blk_mq_hw_ctx
*hctx
;
1809 queue_for_each_hw_ctx(q
, hctx
, i
) {
1811 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1813 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1817 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
1819 struct request_queue
*q
;
1821 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1822 blk_mq_freeze_queue(q
);
1823 queue_set_hctx_shared(q
, shared
);
1824 blk_mq_unfreeze_queue(q
);
1828 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1830 struct blk_mq_tag_set
*set
= q
->tag_set
;
1832 mutex_lock(&set
->tag_list_lock
);
1833 list_del_init(&q
->tag_set_list
);
1834 if (list_is_singular(&set
->tag_list
)) {
1835 /* just transitioned to unshared */
1836 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1837 /* update existing queue */
1838 blk_mq_update_tag_set_depth(set
, false);
1840 mutex_unlock(&set
->tag_list_lock
);
1843 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1844 struct request_queue
*q
)
1848 mutex_lock(&set
->tag_list_lock
);
1850 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1851 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1852 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
1853 /* update existing queue */
1854 blk_mq_update_tag_set_depth(set
, true);
1856 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
1857 queue_set_hctx_shared(q
, true);
1858 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1860 mutex_unlock(&set
->tag_list_lock
);
1864 * It is the actual release handler for mq, but we do it from
1865 * request queue's release handler for avoiding use-after-free
1866 * and headache because q->mq_kobj shouldn't have been introduced,
1867 * but we can't group ctx/kctx kobj without it.
1869 void blk_mq_release(struct request_queue
*q
)
1871 struct blk_mq_hw_ctx
*hctx
;
1874 /* hctx kobj stays in hctx */
1875 queue_for_each_hw_ctx(q
, hctx
, i
) {
1884 kfree(q
->queue_hw_ctx
);
1886 /* ctx kobj stays in queue_ctx */
1887 free_percpu(q
->queue_ctx
);
1890 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
1892 struct request_queue
*uninit_q
, *q
;
1894 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
1896 return ERR_PTR(-ENOMEM
);
1898 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
1900 blk_cleanup_queue(uninit_q
);
1904 EXPORT_SYMBOL(blk_mq_init_queue
);
1906 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
1907 struct request_queue
*q
)
1910 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
1912 blk_mq_sysfs_unregister(q
);
1913 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
1919 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
1920 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
1925 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
1932 atomic_set(&hctxs
[i
]->nr_active
, 0);
1933 hctxs
[i
]->numa_node
= node
;
1934 hctxs
[i
]->queue_num
= i
;
1936 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
1937 free_cpumask_var(hctxs
[i
]->cpumask
);
1942 blk_mq_hctx_kobj_init(hctxs
[i
]);
1944 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
1945 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
1949 blk_mq_free_rq_map(set
, hctx
->tags
, j
);
1950 set
->tags
[j
] = NULL
;
1952 blk_mq_exit_hctx(q
, set
, hctx
, j
);
1953 free_cpumask_var(hctx
->cpumask
);
1954 kobject_put(&hctx
->kobj
);
1961 q
->nr_hw_queues
= i
;
1962 blk_mq_sysfs_register(q
);
1965 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
1966 struct request_queue
*q
)
1968 /* mark the queue as mq asap */
1969 q
->mq_ops
= set
->ops
;
1971 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
1975 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
1976 GFP_KERNEL
, set
->numa_node
);
1977 if (!q
->queue_hw_ctx
)
1980 q
->mq_map
= set
->mq_map
;
1982 blk_mq_realloc_hw_ctxs(set
, q
);
1983 if (!q
->nr_hw_queues
)
1986 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
1987 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
1989 q
->nr_queues
= nr_cpu_ids
;
1991 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
1993 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
1994 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
1996 q
->sg_reserved_size
= INT_MAX
;
1998 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
1999 INIT_LIST_HEAD(&q
->requeue_list
);
2000 spin_lock_init(&q
->requeue_lock
);
2002 if (q
->nr_hw_queues
> 1)
2003 blk_queue_make_request(q
, blk_mq_make_request
);
2005 blk_queue_make_request(q
, blk_sq_make_request
);
2008 * Do this after blk_queue_make_request() overrides it...
2010 q
->nr_requests
= set
->queue_depth
;
2012 if (set
->ops
->complete
)
2013 blk_queue_softirq_done(q
, set
->ops
->complete
);
2015 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2018 mutex_lock(&all_q_mutex
);
2020 list_add_tail(&q
->all_q_node
, &all_q_list
);
2021 blk_mq_add_queue_tag_set(set
, q
);
2022 blk_mq_map_swqueue(q
, cpu_online_mask
);
2024 mutex_unlock(&all_q_mutex
);
2030 kfree(q
->queue_hw_ctx
);
2032 free_percpu(q
->queue_ctx
);
2035 return ERR_PTR(-ENOMEM
);
2037 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2039 void blk_mq_free_queue(struct request_queue
*q
)
2041 struct blk_mq_tag_set
*set
= q
->tag_set
;
2043 mutex_lock(&all_q_mutex
);
2044 list_del_init(&q
->all_q_node
);
2045 mutex_unlock(&all_q_mutex
);
2047 blk_mq_del_queue_tag_set(q
);
2049 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2050 blk_mq_free_hw_queues(q
, set
);
2053 /* Basically redo blk_mq_init_queue with queue frozen */
2054 static void blk_mq_queue_reinit(struct request_queue
*q
,
2055 const struct cpumask
*online_mask
)
2057 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2059 blk_mq_sysfs_unregister(q
);
2062 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2063 * we should change hctx numa_node according to new topology (this
2064 * involves free and re-allocate memory, worthy doing?)
2067 blk_mq_map_swqueue(q
, online_mask
);
2069 blk_mq_sysfs_register(q
);
2073 * New online cpumask which is going to be set in this hotplug event.
2074 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2075 * one-by-one and dynamically allocating this could result in a failure.
2077 static struct cpumask cpuhp_online_new
;
2079 static void blk_mq_queue_reinit_work(void)
2081 struct request_queue
*q
;
2083 mutex_lock(&all_q_mutex
);
2085 * We need to freeze and reinit all existing queues. Freezing
2086 * involves synchronous wait for an RCU grace period and doing it
2087 * one by one may take a long time. Start freezing all queues in
2088 * one swoop and then wait for the completions so that freezing can
2089 * take place in parallel.
2091 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2092 blk_mq_freeze_queue_start(q
);
2093 list_for_each_entry(q
, &all_q_list
, all_q_node
) {
2094 blk_mq_freeze_queue_wait(q
);
2097 * timeout handler can't touch hw queue during the
2100 del_timer_sync(&q
->timeout
);
2103 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2104 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2106 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2107 blk_mq_unfreeze_queue(q
);
2109 mutex_unlock(&all_q_mutex
);
2112 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2114 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2115 blk_mq_queue_reinit_work();
2120 * Before hotadded cpu starts handling requests, new mappings must be
2121 * established. Otherwise, these requests in hw queue might never be
2124 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2125 * for CPU0, and ctx1 for CPU1).
2127 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2128 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2130 * And then while running hw queue, flush_busy_ctxs() finds bit0 is set in
2131 * pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2132 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list
2135 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2137 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2138 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2139 blk_mq_queue_reinit_work();
2143 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2147 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2148 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2157 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2163 * Allocate the request maps associated with this tag_set. Note that this
2164 * may reduce the depth asked for, if memory is tight. set->queue_depth
2165 * will be updated to reflect the allocated depth.
2167 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2172 depth
= set
->queue_depth
;
2174 err
= __blk_mq_alloc_rq_maps(set
);
2178 set
->queue_depth
>>= 1;
2179 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2183 } while (set
->queue_depth
);
2185 if (!set
->queue_depth
|| err
) {
2186 pr_err("blk-mq: failed to allocate request map\n");
2190 if (depth
!= set
->queue_depth
)
2191 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2192 depth
, set
->queue_depth
);
2198 * Alloc a tag set to be associated with one or more request queues.
2199 * May fail with EINVAL for various error conditions. May adjust the
2200 * requested depth down, if if it too large. In that case, the set
2201 * value will be stored in set->queue_depth.
2203 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2207 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2209 if (!set
->nr_hw_queues
)
2211 if (!set
->queue_depth
)
2213 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2216 if (!set
->ops
->queue_rq
)
2219 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2220 pr_info("blk-mq: reduced tag depth to %u\n",
2222 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2226 * If a crashdump is active, then we are potentially in a very
2227 * memory constrained environment. Limit us to 1 queue and
2228 * 64 tags to prevent using too much memory.
2230 if (is_kdump_kernel()) {
2231 set
->nr_hw_queues
= 1;
2232 set
->queue_depth
= min(64U, set
->queue_depth
);
2235 * There is no use for more h/w queues than cpus.
2237 if (set
->nr_hw_queues
> nr_cpu_ids
)
2238 set
->nr_hw_queues
= nr_cpu_ids
;
2240 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2241 GFP_KERNEL
, set
->numa_node
);
2246 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2247 GFP_KERNEL
, set
->numa_node
);
2251 if (set
->ops
->map_queues
)
2252 ret
= set
->ops
->map_queues(set
);
2254 ret
= blk_mq_map_queues(set
);
2256 goto out_free_mq_map
;
2258 ret
= blk_mq_alloc_rq_maps(set
);
2260 goto out_free_mq_map
;
2262 mutex_init(&set
->tag_list_lock
);
2263 INIT_LIST_HEAD(&set
->tag_list
);
2275 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2277 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2281 for (i
= 0; i
< nr_cpu_ids
; i
++) {
2283 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2292 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2294 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2296 struct blk_mq_tag_set
*set
= q
->tag_set
;
2297 struct blk_mq_hw_ctx
*hctx
;
2300 if (!set
|| nr
> set
->queue_depth
)
2304 queue_for_each_hw_ctx(q
, hctx
, i
) {
2307 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2313 q
->nr_requests
= nr
;
2318 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2320 struct request_queue
*q
;
2322 if (nr_hw_queues
> nr_cpu_ids
)
2323 nr_hw_queues
= nr_cpu_ids
;
2324 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2327 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2328 blk_mq_freeze_queue(q
);
2330 set
->nr_hw_queues
= nr_hw_queues
;
2331 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2332 blk_mq_realloc_hw_ctxs(set
, q
);
2334 if (q
->nr_hw_queues
> 1)
2335 blk_queue_make_request(q
, blk_mq_make_request
);
2337 blk_queue_make_request(q
, blk_sq_make_request
);
2339 blk_mq_queue_reinit(q
, cpu_online_mask
);
2342 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2343 blk_mq_unfreeze_queue(q
);
2345 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2347 void blk_mq_disable_hotplug(void)
2349 mutex_lock(&all_q_mutex
);
2352 void blk_mq_enable_hotplug(void)
2354 mutex_unlock(&all_q_mutex
);
2357 static int __init
blk_mq_init(void)
2359 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2360 blk_mq_hctx_notify_dead
);
2362 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2363 blk_mq_queue_reinit_prepare
,
2364 blk_mq_queue_reinit_dead
);
2367 subsys_initcall(blk_mq_init
);