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"
35 #include "blk-mq-sched.h"
37 static DEFINE_MUTEX(all_q_mutex
);
38 static LIST_HEAD(all_q_list
);
41 * Check if any of the ctx's have pending work in this hardware queue
43 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
45 return sbitmap_any_bit_set(&hctx
->ctx_map
) ||
46 !list_empty_careful(&hctx
->dispatch
) ||
47 blk_mq_sched_has_work(hctx
);
51 * Mark this ctx as having pending work in this hardware queue
53 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
54 struct blk_mq_ctx
*ctx
)
56 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
57 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
60 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
61 struct blk_mq_ctx
*ctx
)
63 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
66 void blk_mq_freeze_queue_start(struct request_queue
*q
)
70 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
71 if (freeze_depth
== 1) {
72 percpu_ref_kill(&q
->q_usage_counter
);
73 blk_mq_run_hw_queues(q
, false);
76 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start
);
78 static void blk_mq_freeze_queue_wait(struct request_queue
*q
)
80 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
84 * Guarantee no request is in use, so we can change any data structure of
85 * the queue afterward.
87 void blk_freeze_queue(struct request_queue
*q
)
90 * In the !blk_mq case we are only calling this to kill the
91 * q_usage_counter, otherwise this increases the freeze depth
92 * and waits for it to return to zero. For this reason there is
93 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
94 * exported to drivers as the only user for unfreeze is blk_mq.
96 blk_mq_freeze_queue_start(q
);
97 blk_mq_freeze_queue_wait(q
);
100 void blk_mq_freeze_queue(struct request_queue
*q
)
103 * ...just an alias to keep freeze and unfreeze actions balanced
104 * in the blk_mq_* namespace
108 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
110 void blk_mq_unfreeze_queue(struct request_queue
*q
)
114 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
115 WARN_ON_ONCE(freeze_depth
< 0);
117 percpu_ref_reinit(&q
->q_usage_counter
);
118 wake_up_all(&q
->mq_freeze_wq
);
121 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
124 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
127 * Note: this function does not prevent that the struct request end_io()
128 * callback function is invoked. Additionally, it is not prevented that
129 * new queue_rq() calls occur unless the queue has been stopped first.
131 void blk_mq_quiesce_queue(struct request_queue
*q
)
133 struct blk_mq_hw_ctx
*hctx
;
137 blk_mq_stop_hw_queues(q
);
139 queue_for_each_hw_ctx(q
, hctx
, i
) {
140 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
141 synchronize_srcu(&hctx
->queue_rq_srcu
);
148 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
150 void blk_mq_wake_waiters(struct request_queue
*q
)
152 struct blk_mq_hw_ctx
*hctx
;
155 queue_for_each_hw_ctx(q
, hctx
, i
)
156 if (blk_mq_hw_queue_mapped(hctx
))
157 blk_mq_tag_wakeup_all(hctx
->tags
, true);
160 * If we are called because the queue has now been marked as
161 * dying, we need to ensure that processes currently waiting on
162 * the queue are notified as well.
164 wake_up_all(&q
->mq_freeze_wq
);
167 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
169 return blk_mq_has_free_tags(hctx
->tags
);
171 EXPORT_SYMBOL(blk_mq_can_queue
);
173 void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
174 struct request
*rq
, unsigned int op
)
176 INIT_LIST_HEAD(&rq
->queuelist
);
177 /* csd/requeue_work/fifo_time is initialized before use */
181 if (blk_queue_io_stat(q
))
182 rq
->rq_flags
|= RQF_IO_STAT
;
183 /* do not touch atomic flags, it needs atomic ops against the timer */
185 INIT_HLIST_NODE(&rq
->hash
);
186 RB_CLEAR_NODE(&rq
->rb_node
);
189 rq
->start_time
= jiffies
;
190 #ifdef CONFIG_BLK_CGROUP
192 set_start_time_ns(rq
);
193 rq
->io_start_time_ns
= 0;
195 rq
->nr_phys_segments
= 0;
196 #if defined(CONFIG_BLK_DEV_INTEGRITY)
197 rq
->nr_integrity_segments
= 0;
200 /* tag was already set */
210 INIT_LIST_HEAD(&rq
->timeout_list
);
214 rq
->end_io_data
= NULL
;
217 ctx
->rq_dispatched
[op_is_sync(op
)]++;
219 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init
);
221 struct request
*__blk_mq_alloc_request(struct blk_mq_alloc_data
*data
,
227 tag
= blk_mq_get_tag(data
);
228 if (tag
!= BLK_MQ_TAG_FAIL
) {
229 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
231 rq
= tags
->static_rqs
[tag
];
233 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
235 rq
->internal_tag
= tag
;
237 if (blk_mq_tag_busy(data
->hctx
)) {
238 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
239 atomic_inc(&data
->hctx
->nr_active
);
242 rq
->internal_tag
= -1;
245 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
);
251 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request
);
253 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
256 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
260 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
264 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
266 blk_mq_put_ctx(alloc_data
.ctx
);
270 return ERR_PTR(-EWOULDBLOCK
);
273 rq
->__sector
= (sector_t
) -1;
274 rq
->bio
= rq
->biotail
= NULL
;
277 EXPORT_SYMBOL(blk_mq_alloc_request
);
279 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
280 unsigned int flags
, unsigned int hctx_idx
)
282 struct blk_mq_hw_ctx
*hctx
;
283 struct blk_mq_ctx
*ctx
;
285 struct blk_mq_alloc_data alloc_data
;
289 * If the tag allocator sleeps we could get an allocation for a
290 * different hardware context. No need to complicate the low level
291 * allocator for this for the rare use case of a command tied to
294 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
295 return ERR_PTR(-EINVAL
);
297 if (hctx_idx
>= q
->nr_hw_queues
)
298 return ERR_PTR(-EIO
);
300 ret
= blk_queue_enter(q
, true);
305 * Check if the hardware context is actually mapped to anything.
306 * If not tell the caller that it should skip this queue.
308 hctx
= q
->queue_hw_ctx
[hctx_idx
];
309 if (!blk_mq_hw_queue_mapped(hctx
)) {
313 ctx
= __blk_mq_get_ctx(q
, cpumask_first(hctx
->cpumask
));
315 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
316 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
328 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
330 void __blk_mq_finish_request(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
333 const int sched_tag
= rq
->internal_tag
;
334 struct request_queue
*q
= rq
->q
;
336 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
337 atomic_dec(&hctx
->nr_active
);
339 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
342 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
343 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
345 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
347 blk_mq_sched_completed_request(hctx
, rq
);
351 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx
*hctx
,
354 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
356 ctx
->rq_completed
[rq_is_sync(rq
)]++;
357 __blk_mq_finish_request(hctx
, ctx
, rq
);
360 void blk_mq_finish_request(struct request
*rq
)
362 blk_mq_finish_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
365 void blk_mq_free_request(struct request
*rq
)
367 blk_mq_sched_put_request(rq
);
369 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
371 inline void __blk_mq_end_request(struct request
*rq
, int error
)
373 blk_account_io_done(rq
);
376 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
377 rq
->end_io(rq
, error
);
379 if (unlikely(blk_bidi_rq(rq
)))
380 blk_mq_free_request(rq
->next_rq
);
381 blk_mq_free_request(rq
);
384 EXPORT_SYMBOL(__blk_mq_end_request
);
386 void blk_mq_end_request(struct request
*rq
, int error
)
388 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
390 __blk_mq_end_request(rq
, error
);
392 EXPORT_SYMBOL(blk_mq_end_request
);
394 static void __blk_mq_complete_request_remote(void *data
)
396 struct request
*rq
= data
;
398 rq
->q
->softirq_done_fn(rq
);
401 static void blk_mq_ipi_complete_request(struct request
*rq
)
403 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
407 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
408 rq
->q
->softirq_done_fn(rq
);
413 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
414 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
416 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
417 rq
->csd
.func
= __blk_mq_complete_request_remote
;
420 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
422 rq
->q
->softirq_done_fn(rq
);
427 static void blk_mq_stat_add(struct request
*rq
)
429 if (rq
->rq_flags
& RQF_STATS
) {
431 * We could rq->mq_ctx here, but there's less of a risk
432 * of races if we have the completion event add the stats
433 * to the local software queue.
435 struct blk_mq_ctx
*ctx
;
437 ctx
= __blk_mq_get_ctx(rq
->q
, raw_smp_processor_id());
438 blk_stat_add(&ctx
->stat
[rq_data_dir(rq
)], rq
);
442 static void __blk_mq_complete_request(struct request
*rq
)
444 struct request_queue
*q
= rq
->q
;
448 if (!q
->softirq_done_fn
)
449 blk_mq_end_request(rq
, rq
->errors
);
451 blk_mq_ipi_complete_request(rq
);
455 * blk_mq_complete_request - end I/O on a request
456 * @rq: the request being processed
459 * Ends all I/O on a request. It does not handle partial completions.
460 * The actual completion happens out-of-order, through a IPI handler.
462 void blk_mq_complete_request(struct request
*rq
, int error
)
464 struct request_queue
*q
= rq
->q
;
466 if (unlikely(blk_should_fake_timeout(q
)))
468 if (!blk_mark_rq_complete(rq
)) {
470 __blk_mq_complete_request(rq
);
473 EXPORT_SYMBOL(blk_mq_complete_request
);
475 int blk_mq_request_started(struct request
*rq
)
477 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
479 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
481 void blk_mq_start_request(struct request
*rq
)
483 struct request_queue
*q
= rq
->q
;
485 blk_mq_sched_started_request(rq
);
487 trace_block_rq_issue(q
, rq
);
489 rq
->resid_len
= blk_rq_bytes(rq
);
490 if (unlikely(blk_bidi_rq(rq
)))
491 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
493 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
494 blk_stat_set_issue_time(&rq
->issue_stat
);
495 rq
->rq_flags
|= RQF_STATS
;
496 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
502 * Ensure that ->deadline is visible before set the started
503 * flag and clear the completed flag.
505 smp_mb__before_atomic();
508 * Mark us as started and clear complete. Complete might have been
509 * set if requeue raced with timeout, which then marked it as
510 * complete. So be sure to clear complete again when we start
511 * the request, otherwise we'll ignore the completion event.
513 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
514 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
515 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
516 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
518 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
520 * Make sure space for the drain appears. We know we can do
521 * this because max_hw_segments has been adjusted to be one
522 * fewer than the device can handle.
524 rq
->nr_phys_segments
++;
527 EXPORT_SYMBOL(blk_mq_start_request
);
529 static void __blk_mq_requeue_request(struct request
*rq
)
531 struct request_queue
*q
= rq
->q
;
533 trace_block_rq_requeue(q
, rq
);
534 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
535 blk_mq_sched_requeue_request(rq
);
537 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
538 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
539 rq
->nr_phys_segments
--;
543 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
545 __blk_mq_requeue_request(rq
);
547 BUG_ON(blk_queued_rq(rq
));
548 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
550 EXPORT_SYMBOL(blk_mq_requeue_request
);
552 static void blk_mq_requeue_work(struct work_struct
*work
)
554 struct request_queue
*q
=
555 container_of(work
, struct request_queue
, requeue_work
.work
);
557 struct request
*rq
, *next
;
560 spin_lock_irqsave(&q
->requeue_lock
, flags
);
561 list_splice_init(&q
->requeue_list
, &rq_list
);
562 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
564 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
565 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
568 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
569 list_del_init(&rq
->queuelist
);
570 blk_mq_sched_insert_request(rq
, true, false, false);
573 while (!list_empty(&rq_list
)) {
574 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
575 list_del_init(&rq
->queuelist
);
576 blk_mq_sched_insert_request(rq
, false, false, false);
579 blk_mq_run_hw_queues(q
, false);
582 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
583 bool kick_requeue_list
)
585 struct request_queue
*q
= rq
->q
;
589 * We abuse this flag that is otherwise used by the I/O scheduler to
590 * request head insertation from the workqueue.
592 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
594 spin_lock_irqsave(&q
->requeue_lock
, flags
);
596 rq
->rq_flags
|= RQF_SOFTBARRIER
;
597 list_add(&rq
->queuelist
, &q
->requeue_list
);
599 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
601 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
603 if (kick_requeue_list
)
604 blk_mq_kick_requeue_list(q
);
606 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
608 void blk_mq_kick_requeue_list(struct request_queue
*q
)
610 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
612 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
614 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
617 kblockd_schedule_delayed_work(&q
->requeue_work
,
618 msecs_to_jiffies(msecs
));
620 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
622 void blk_mq_abort_requeue_list(struct request_queue
*q
)
627 spin_lock_irqsave(&q
->requeue_lock
, flags
);
628 list_splice_init(&q
->requeue_list
, &rq_list
);
629 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
631 while (!list_empty(&rq_list
)) {
634 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
635 list_del_init(&rq
->queuelist
);
637 blk_mq_end_request(rq
, rq
->errors
);
640 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
642 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
644 if (tag
< tags
->nr_tags
) {
645 prefetch(tags
->rqs
[tag
]);
646 return tags
->rqs
[tag
];
651 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
653 struct blk_mq_timeout_data
{
655 unsigned int next_set
;
658 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
660 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
661 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
664 * We know that complete is set at this point. If STARTED isn't set
665 * anymore, then the request isn't active and the "timeout" should
666 * just be ignored. This can happen due to the bitflag ordering.
667 * Timeout first checks if STARTED is set, and if it is, assumes
668 * the request is active. But if we race with completion, then
669 * we both flags will get cleared. So check here again, and ignore
670 * a timeout event with a request that isn't active.
672 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
676 ret
= ops
->timeout(req
, reserved
);
680 __blk_mq_complete_request(req
);
682 case BLK_EH_RESET_TIMER
:
684 blk_clear_rq_complete(req
);
686 case BLK_EH_NOT_HANDLED
:
689 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
694 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
695 struct request
*rq
, void *priv
, bool reserved
)
697 struct blk_mq_timeout_data
*data
= priv
;
699 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
701 * If a request wasn't started before the queue was
702 * marked dying, kill it here or it'll go unnoticed.
704 if (unlikely(blk_queue_dying(rq
->q
))) {
706 blk_mq_end_request(rq
, rq
->errors
);
711 if (time_after_eq(jiffies
, rq
->deadline
)) {
712 if (!blk_mark_rq_complete(rq
))
713 blk_mq_rq_timed_out(rq
, reserved
);
714 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
715 data
->next
= rq
->deadline
;
720 static void blk_mq_timeout_work(struct work_struct
*work
)
722 struct request_queue
*q
=
723 container_of(work
, struct request_queue
, timeout_work
);
724 struct blk_mq_timeout_data data
= {
730 /* A deadlock might occur if a request is stuck requiring a
731 * timeout at the same time a queue freeze is waiting
732 * completion, since the timeout code would not be able to
733 * acquire the queue reference here.
735 * That's why we don't use blk_queue_enter here; instead, we use
736 * percpu_ref_tryget directly, because we need to be able to
737 * obtain a reference even in the short window between the queue
738 * starting to freeze, by dropping the first reference in
739 * blk_mq_freeze_queue_start, and the moment the last request is
740 * consumed, marked by the instant q_usage_counter reaches
743 if (!percpu_ref_tryget(&q
->q_usage_counter
))
746 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
749 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
750 mod_timer(&q
->timeout
, data
.next
);
752 struct blk_mq_hw_ctx
*hctx
;
754 queue_for_each_hw_ctx(q
, hctx
, i
) {
755 /* the hctx may be unmapped, so check it here */
756 if (blk_mq_hw_queue_mapped(hctx
))
757 blk_mq_tag_idle(hctx
);
764 * Reverse check our software queue for entries that we could potentially
765 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
766 * too much time checking for merges.
768 static bool blk_mq_attempt_merge(struct request_queue
*q
,
769 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
774 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
780 if (!blk_rq_merge_ok(rq
, bio
))
783 el_ret
= blk_try_merge(rq
, bio
);
784 if (el_ret
== ELEVATOR_NO_MERGE
)
787 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
790 if (el_ret
== ELEVATOR_BACK_MERGE
) {
791 if (bio_attempt_back_merge(q
, rq
, bio
)) {
796 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
797 if (bio_attempt_front_merge(q
, rq
, bio
)) {
808 struct flush_busy_ctx_data
{
809 struct blk_mq_hw_ctx
*hctx
;
810 struct list_head
*list
;
813 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
815 struct flush_busy_ctx_data
*flush_data
= data
;
816 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
817 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
819 sbitmap_clear_bit(sb
, bitnr
);
820 spin_lock(&ctx
->lock
);
821 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
822 spin_unlock(&ctx
->lock
);
827 * Process software queues that have been marked busy, splicing them
828 * to the for-dispatch
830 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
832 struct flush_busy_ctx_data data
= {
837 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
839 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
841 static inline unsigned int queued_to_index(unsigned int queued
)
846 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
849 static bool blk_mq_get_driver_tag(struct request
*rq
,
850 struct blk_mq_hw_ctx
**hctx
, bool wait
)
852 struct blk_mq_alloc_data data
= {
855 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
856 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
859 if (blk_mq_hctx_stopped(data
.hctx
))
869 rq
->tag
= blk_mq_get_tag(&data
);
871 if (blk_mq_tag_busy(data
.hctx
)) {
872 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
873 atomic_inc(&data
.hctx
->nr_active
);
875 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
883 * If we fail getting a driver tag because all the driver tags are already
884 * assigned and on the dispatch list, BUT the first entry does not have a
885 * tag, then we could deadlock. For that case, move entries with assigned
886 * driver tags to the front, leaving the set of tagged requests in the
887 * same order, and the untagged set in the same order.
889 static bool reorder_tags_to_front(struct list_head
*list
)
891 struct request
*rq
, *tmp
, *first
= NULL
;
893 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
897 list_move(&rq
->queuelist
, list
);
903 return first
!= NULL
;
906 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
908 struct request_queue
*q
= hctx
->queue
;
910 LIST_HEAD(driver_list
);
911 struct list_head
*dptr
;
912 int queued
, ret
= BLK_MQ_RQ_QUEUE_OK
;
915 * Start off with dptr being NULL, so we start the first request
916 * immediately, even if we have more pending.
921 * Now process all the entries, sending them to the driver.
924 while (!list_empty(list
)) {
925 struct blk_mq_queue_data bd
;
927 rq
= list_first_entry(list
, struct request
, queuelist
);
928 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
929 if (!queued
&& reorder_tags_to_front(list
))
931 blk_mq_sched_mark_restart(hctx
);
934 list_del_init(&rq
->queuelist
);
938 bd
.last
= list_empty(list
);
940 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
942 case BLK_MQ_RQ_QUEUE_OK
:
945 case BLK_MQ_RQ_QUEUE_BUSY
:
946 list_add(&rq
->queuelist
, list
);
947 __blk_mq_requeue_request(rq
);
950 pr_err("blk-mq: bad return on queue: %d\n", ret
);
951 case BLK_MQ_RQ_QUEUE_ERROR
:
953 blk_mq_end_request(rq
, rq
->errors
);
957 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
961 * We've done the first request. If we have more than 1
962 * left in the list, set dptr to defer issue.
964 if (!dptr
&& list
->next
!= list
->prev
)
968 hctx
->dispatched
[queued_to_index(queued
)]++;
971 * Any items that need requeuing? Stuff them into hctx->dispatch,
972 * that is where we will continue on next queue run.
974 if (!list_empty(list
)) {
975 spin_lock(&hctx
->lock
);
976 list_splice(list
, &hctx
->dispatch
);
977 spin_unlock(&hctx
->lock
);
980 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
981 * it's possible the queue is stopped and restarted again
982 * before this. Queue restart will dispatch requests. And since
983 * requests in rq_list aren't added into hctx->dispatch yet,
984 * the requests in rq_list might get lost.
986 * blk_mq_run_hw_queue() already checks the STOPPED bit
988 * If RESTART is set, then let completion restart the queue
989 * instead of potentially looping here.
991 if (!blk_mq_sched_needs_restart(hctx
))
992 blk_mq_run_hw_queue(hctx
, true);
995 return ret
!= BLK_MQ_RQ_QUEUE_BUSY
;
998 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1002 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1003 cpu_online(hctx
->next_cpu
));
1005 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1007 blk_mq_sched_dispatch_requests(hctx
);
1010 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1011 blk_mq_sched_dispatch_requests(hctx
);
1012 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1017 * It'd be great if the workqueue API had a way to pass
1018 * in a mask and had some smarts for more clever placement.
1019 * For now we just round-robin here, switching for every
1020 * BLK_MQ_CPU_WORK_BATCH queued items.
1022 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1024 if (hctx
->queue
->nr_hw_queues
== 1)
1025 return WORK_CPU_UNBOUND
;
1027 if (--hctx
->next_cpu_batch
<= 0) {
1030 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1031 if (next_cpu
>= nr_cpu_ids
)
1032 next_cpu
= cpumask_first(hctx
->cpumask
);
1034 hctx
->next_cpu
= next_cpu
;
1035 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1038 return hctx
->next_cpu
;
1041 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1043 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
1044 !blk_mq_hw_queue_mapped(hctx
)))
1047 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1048 int cpu
= get_cpu();
1049 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1050 __blk_mq_run_hw_queue(hctx
);
1058 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
);
1061 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1063 struct blk_mq_hw_ctx
*hctx
;
1066 queue_for_each_hw_ctx(q
, hctx
, i
) {
1067 if (!blk_mq_hctx_has_pending(hctx
) ||
1068 blk_mq_hctx_stopped(hctx
))
1071 blk_mq_run_hw_queue(hctx
, async
);
1074 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1077 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1078 * @q: request queue.
1080 * The caller is responsible for serializing this function against
1081 * blk_mq_{start,stop}_hw_queue().
1083 bool blk_mq_queue_stopped(struct request_queue
*q
)
1085 struct blk_mq_hw_ctx
*hctx
;
1088 queue_for_each_hw_ctx(q
, hctx
, i
)
1089 if (blk_mq_hctx_stopped(hctx
))
1094 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1096 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1098 cancel_work(&hctx
->run_work
);
1099 cancel_delayed_work(&hctx
->delay_work
);
1100 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1102 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1104 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1106 struct blk_mq_hw_ctx
*hctx
;
1109 queue_for_each_hw_ctx(q
, hctx
, i
)
1110 blk_mq_stop_hw_queue(hctx
);
1112 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1114 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1116 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1118 blk_mq_run_hw_queue(hctx
, false);
1120 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1122 void blk_mq_start_hw_queues(struct request_queue
*q
)
1124 struct blk_mq_hw_ctx
*hctx
;
1127 queue_for_each_hw_ctx(q
, hctx
, i
)
1128 blk_mq_start_hw_queue(hctx
);
1130 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1132 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1134 if (!blk_mq_hctx_stopped(hctx
))
1137 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1138 blk_mq_run_hw_queue(hctx
, async
);
1140 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1142 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1144 struct blk_mq_hw_ctx
*hctx
;
1147 queue_for_each_hw_ctx(q
, hctx
, i
)
1148 blk_mq_start_stopped_hw_queue(hctx
, async
);
1150 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1152 static void blk_mq_run_work_fn(struct work_struct
*work
)
1154 struct blk_mq_hw_ctx
*hctx
;
1156 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1158 __blk_mq_run_hw_queue(hctx
);
1161 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1163 struct blk_mq_hw_ctx
*hctx
;
1165 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1167 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1168 __blk_mq_run_hw_queue(hctx
);
1171 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1173 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1176 blk_mq_stop_hw_queue(hctx
);
1177 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1178 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1180 EXPORT_SYMBOL(blk_mq_delay_queue
);
1182 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1186 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1188 trace_block_rq_insert(hctx
->queue
, rq
);
1191 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1193 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1196 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1199 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1201 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1202 blk_mq_hctx_mark_pending(hctx
, ctx
);
1205 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1206 struct list_head
*list
)
1210 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1213 spin_lock(&ctx
->lock
);
1214 while (!list_empty(list
)) {
1217 rq
= list_first_entry(list
, struct request
, queuelist
);
1218 BUG_ON(rq
->mq_ctx
!= ctx
);
1219 list_del_init(&rq
->queuelist
);
1220 __blk_mq_insert_req_list(hctx
, rq
, false);
1222 blk_mq_hctx_mark_pending(hctx
, ctx
);
1223 spin_unlock(&ctx
->lock
);
1226 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1228 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1229 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1231 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1232 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1233 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1236 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1238 struct blk_mq_ctx
*this_ctx
;
1239 struct request_queue
*this_q
;
1242 LIST_HEAD(ctx_list
);
1245 list_splice_init(&plug
->mq_list
, &list
);
1247 list_sort(NULL
, &list
, plug_ctx_cmp
);
1253 while (!list_empty(&list
)) {
1254 rq
= list_entry_rq(list
.next
);
1255 list_del_init(&rq
->queuelist
);
1257 if (rq
->mq_ctx
!= this_ctx
) {
1259 trace_block_unplug(this_q
, depth
, from_schedule
);
1260 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1265 this_ctx
= rq
->mq_ctx
;
1271 list_add_tail(&rq
->queuelist
, &ctx_list
);
1275 * If 'this_ctx' is set, we know we have entries to complete
1276 * on 'ctx_list'. Do those.
1279 trace_block_unplug(this_q
, depth
, from_schedule
);
1280 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1285 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1287 init_request_from_bio(rq
, bio
);
1289 blk_account_io_start(rq
, true);
1292 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1294 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1295 !blk_queue_nomerges(hctx
->queue
);
1298 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1299 struct blk_mq_ctx
*ctx
,
1300 struct request
*rq
, struct bio
*bio
)
1302 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1303 blk_mq_bio_to_request(rq
, bio
);
1304 spin_lock(&ctx
->lock
);
1306 __blk_mq_insert_request(hctx
, rq
, false);
1307 spin_unlock(&ctx
->lock
);
1310 struct request_queue
*q
= hctx
->queue
;
1312 spin_lock(&ctx
->lock
);
1313 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1314 blk_mq_bio_to_request(rq
, bio
);
1318 spin_unlock(&ctx
->lock
);
1319 __blk_mq_finish_request(hctx
, ctx
, rq
);
1324 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1327 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1329 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1332 static void blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
)
1334 struct request_queue
*q
= rq
->q
;
1335 struct blk_mq_queue_data bd
= {
1340 struct blk_mq_hw_ctx
*hctx
;
1341 blk_qc_t new_cookie
;
1347 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1350 new_cookie
= request_to_qc_t(hctx
, rq
);
1353 * For OK queue, we are done. For error, kill it. Any other
1354 * error (busy), just add it to our list as we previously
1357 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1358 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1359 *cookie
= new_cookie
;
1363 __blk_mq_requeue_request(rq
);
1365 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1366 *cookie
= BLK_QC_T_NONE
;
1368 blk_mq_end_request(rq
, rq
->errors
);
1373 blk_mq_sched_insert_request(rq
, false, true, true);
1377 * Multiple hardware queue variant. This will not use per-process plugs,
1378 * but will attempt to bypass the hctx queueing if we can go straight to
1379 * hardware for SYNC IO.
1381 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1383 const int is_sync
= op_is_sync(bio
->bi_opf
);
1384 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1385 struct blk_mq_alloc_data data
= { .flags
= 0 };
1387 unsigned int request_count
= 0, srcu_idx
;
1388 struct blk_plug
*plug
;
1389 struct request
*same_queue_rq
= NULL
;
1391 unsigned int wb_acct
;
1393 blk_queue_bounce(q
, &bio
);
1395 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1397 return BLK_QC_T_NONE
;
1400 blk_queue_split(q
, &bio
, q
->bio_split
);
1402 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1403 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1404 return BLK_QC_T_NONE
;
1406 if (blk_mq_sched_bio_merge(q
, bio
))
1407 return BLK_QC_T_NONE
;
1409 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1411 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1413 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1414 if (unlikely(!rq
)) {
1415 __wbt_done(q
->rq_wb
, wb_acct
);
1416 return BLK_QC_T_NONE
;
1419 wbt_track(&rq
->issue_stat
, wb_acct
);
1421 cookie
= request_to_qc_t(data
.hctx
, rq
);
1423 if (unlikely(is_flush_fua
)) {
1424 blk_mq_bio_to_request(rq
, bio
);
1425 blk_mq_get_driver_tag(rq
, NULL
, true);
1426 blk_insert_flush(rq
);
1430 plug
= current
->plug
;
1432 * If the driver supports defer issued based on 'last', then
1433 * queue it up like normal since we can potentially save some
1436 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1437 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1438 struct request
*old_rq
= NULL
;
1440 blk_mq_bio_to_request(rq
, bio
);
1443 * We do limited plugging. If the bio can be merged, do that.
1444 * Otherwise the existing request in the plug list will be
1445 * issued. So the plug list will have one request at most
1449 * The plug list might get flushed before this. If that
1450 * happens, same_queue_rq is invalid and plug list is
1453 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1454 old_rq
= same_queue_rq
;
1455 list_del_init(&old_rq
->queuelist
);
1457 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1458 } else /* is_sync */
1460 blk_mq_put_ctx(data
.ctx
);
1464 if (!(data
.hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1466 blk_mq_try_issue_directly(old_rq
, &cookie
);
1469 srcu_idx
= srcu_read_lock(&data
.hctx
->queue_rq_srcu
);
1470 blk_mq_try_issue_directly(old_rq
, &cookie
);
1471 srcu_read_unlock(&data
.hctx
->queue_rq_srcu
, srcu_idx
);
1477 blk_mq_put_ctx(data
.ctx
);
1478 blk_mq_bio_to_request(rq
, bio
);
1479 blk_mq_sched_insert_request(rq
, false, true, true);
1482 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1484 * For a SYNC request, send it to the hardware immediately. For
1485 * an ASYNC request, just ensure that we run it later on. The
1486 * latter allows for merging opportunities and more efficient
1490 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1492 blk_mq_put_ctx(data
.ctx
);
1498 * Single hardware queue variant. This will attempt to use any per-process
1499 * plug for merging and IO deferral.
1501 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1503 const int is_sync
= op_is_sync(bio
->bi_opf
);
1504 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1505 struct blk_plug
*plug
;
1506 unsigned int request_count
= 0;
1507 struct blk_mq_alloc_data data
= { .flags
= 0 };
1510 unsigned int wb_acct
;
1512 blk_queue_bounce(q
, &bio
);
1514 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1516 return BLK_QC_T_NONE
;
1519 blk_queue_split(q
, &bio
, q
->bio_split
);
1521 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1522 if (blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1523 return BLK_QC_T_NONE
;
1525 request_count
= blk_plug_queued_count(q
);
1527 if (blk_mq_sched_bio_merge(q
, bio
))
1528 return BLK_QC_T_NONE
;
1530 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1532 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1534 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1535 if (unlikely(!rq
)) {
1536 __wbt_done(q
->rq_wb
, wb_acct
);
1537 return BLK_QC_T_NONE
;
1540 wbt_track(&rq
->issue_stat
, wb_acct
);
1542 cookie
= request_to_qc_t(data
.hctx
, rq
);
1544 if (unlikely(is_flush_fua
)) {
1545 blk_mq_bio_to_request(rq
, bio
);
1546 blk_mq_get_driver_tag(rq
, NULL
, true);
1547 blk_insert_flush(rq
);
1552 * A task plug currently exists. Since this is completely lockless,
1553 * utilize that to temporarily store requests until the task is
1554 * either done or scheduled away.
1556 plug
= current
->plug
;
1558 struct request
*last
= NULL
;
1560 blk_mq_bio_to_request(rq
, bio
);
1563 * @request_count may become stale because of schedule
1564 * out, so check the list again.
1566 if (list_empty(&plug
->mq_list
))
1569 trace_block_plug(q
);
1571 last
= list_entry_rq(plug
->mq_list
.prev
);
1573 blk_mq_put_ctx(data
.ctx
);
1575 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1576 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1577 blk_flush_plug_list(plug
, false);
1578 trace_block_plug(q
);
1581 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1586 blk_mq_put_ctx(data
.ctx
);
1587 blk_mq_bio_to_request(rq
, bio
);
1588 blk_mq_sched_insert_request(rq
, false, true, true);
1591 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1593 * For a SYNC request, send it to the hardware immediately. For
1594 * an ASYNC request, just ensure that we run it later on. The
1595 * latter allows for merging opportunities and more efficient
1599 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1602 blk_mq_put_ctx(data
.ctx
);
1607 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1608 unsigned int hctx_idx
)
1612 if (tags
->rqs
&& set
->ops
->exit_request
) {
1615 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1616 struct request
*rq
= tags
->static_rqs
[i
];
1620 set
->ops
->exit_request(set
->driver_data
, rq
,
1622 tags
->static_rqs
[i
] = NULL
;
1626 while (!list_empty(&tags
->page_list
)) {
1627 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1628 list_del_init(&page
->lru
);
1630 * Remove kmemleak object previously allocated in
1631 * blk_mq_init_rq_map().
1633 kmemleak_free(page_address(page
));
1634 __free_pages(page
, page
->private);
1638 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1642 kfree(tags
->static_rqs
);
1643 tags
->static_rqs
= NULL
;
1645 blk_mq_free_tags(tags
);
1648 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1649 unsigned int hctx_idx
,
1650 unsigned int nr_tags
,
1651 unsigned int reserved_tags
)
1653 struct blk_mq_tags
*tags
;
1655 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
,
1657 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1661 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1662 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1665 blk_mq_free_tags(tags
);
1669 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1670 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1672 if (!tags
->static_rqs
) {
1674 blk_mq_free_tags(tags
);
1681 static size_t order_to_size(unsigned int order
)
1683 return (size_t)PAGE_SIZE
<< order
;
1686 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1687 unsigned int hctx_idx
, unsigned int depth
)
1689 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1690 size_t rq_size
, left
;
1692 INIT_LIST_HEAD(&tags
->page_list
);
1695 * rq_size is the size of the request plus driver payload, rounded
1696 * to the cacheline size
1698 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1700 left
= rq_size
* depth
;
1702 for (i
= 0; i
< depth
; ) {
1703 int this_order
= max_order
;
1708 while (this_order
&& left
< order_to_size(this_order
- 1))
1712 page
= alloc_pages_node(set
->numa_node
,
1713 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1719 if (order_to_size(this_order
) < rq_size
)
1726 page
->private = this_order
;
1727 list_add_tail(&page
->lru
, &tags
->page_list
);
1729 p
= page_address(page
);
1731 * Allow kmemleak to scan these pages as they contain pointers
1732 * to additional allocations like via ops->init_request().
1734 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1735 entries_per_page
= order_to_size(this_order
) / rq_size
;
1736 to_do
= min(entries_per_page
, depth
- i
);
1737 left
-= to_do
* rq_size
;
1738 for (j
= 0; j
< to_do
; j
++) {
1739 struct request
*rq
= p
;
1741 tags
->static_rqs
[i
] = rq
;
1742 if (set
->ops
->init_request
) {
1743 if (set
->ops
->init_request(set
->driver_data
,
1746 tags
->static_rqs
[i
] = NULL
;
1758 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1763 * 'cpu' is going away. splice any existing rq_list entries from this
1764 * software queue to the hw queue dispatch list, and ensure that it
1767 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1769 struct blk_mq_hw_ctx
*hctx
;
1770 struct blk_mq_ctx
*ctx
;
1773 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1774 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1776 spin_lock(&ctx
->lock
);
1777 if (!list_empty(&ctx
->rq_list
)) {
1778 list_splice_init(&ctx
->rq_list
, &tmp
);
1779 blk_mq_hctx_clear_pending(hctx
, ctx
);
1781 spin_unlock(&ctx
->lock
);
1783 if (list_empty(&tmp
))
1786 spin_lock(&hctx
->lock
);
1787 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1788 spin_unlock(&hctx
->lock
);
1790 blk_mq_run_hw_queue(hctx
, true);
1794 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1796 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1800 /* hctx->ctxs will be freed in queue's release handler */
1801 static void blk_mq_exit_hctx(struct request_queue
*q
,
1802 struct blk_mq_tag_set
*set
,
1803 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1805 unsigned flush_start_tag
= set
->queue_depth
;
1807 blk_mq_tag_idle(hctx
);
1809 if (set
->ops
->exit_request
)
1810 set
->ops
->exit_request(set
->driver_data
,
1811 hctx
->fq
->flush_rq
, hctx_idx
,
1812 flush_start_tag
+ hctx_idx
);
1814 if (set
->ops
->exit_hctx
)
1815 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1817 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1818 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1820 blk_mq_remove_cpuhp(hctx
);
1821 blk_free_flush_queue(hctx
->fq
);
1822 sbitmap_free(&hctx
->ctx_map
);
1825 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1826 struct blk_mq_tag_set
*set
, int nr_queue
)
1828 struct blk_mq_hw_ctx
*hctx
;
1831 queue_for_each_hw_ctx(q
, hctx
, i
) {
1834 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1838 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1839 struct blk_mq_tag_set
*set
)
1841 struct blk_mq_hw_ctx
*hctx
;
1844 queue_for_each_hw_ctx(q
, hctx
, i
)
1845 free_cpumask_var(hctx
->cpumask
);
1848 static int blk_mq_init_hctx(struct request_queue
*q
,
1849 struct blk_mq_tag_set
*set
,
1850 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1853 unsigned flush_start_tag
= set
->queue_depth
;
1855 node
= hctx
->numa_node
;
1856 if (node
== NUMA_NO_NODE
)
1857 node
= hctx
->numa_node
= set
->numa_node
;
1859 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1860 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1861 spin_lock_init(&hctx
->lock
);
1862 INIT_LIST_HEAD(&hctx
->dispatch
);
1864 hctx
->queue_num
= hctx_idx
;
1865 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1867 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1869 hctx
->tags
= set
->tags
[hctx_idx
];
1872 * Allocate space for all possible cpus to avoid allocation at
1875 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1878 goto unregister_cpu_notifier
;
1880 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1886 if (set
->ops
->init_hctx
&&
1887 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1890 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1894 if (set
->ops
->init_request
&&
1895 set
->ops
->init_request(set
->driver_data
,
1896 hctx
->fq
->flush_rq
, hctx_idx
,
1897 flush_start_tag
+ hctx_idx
, node
))
1900 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1901 init_srcu_struct(&hctx
->queue_rq_srcu
);
1908 if (set
->ops
->exit_hctx
)
1909 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1911 sbitmap_free(&hctx
->ctx_map
);
1914 unregister_cpu_notifier
:
1915 blk_mq_remove_cpuhp(hctx
);
1919 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1920 unsigned int nr_hw_queues
)
1924 for_each_possible_cpu(i
) {
1925 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1926 struct blk_mq_hw_ctx
*hctx
;
1928 memset(__ctx
, 0, sizeof(*__ctx
));
1930 spin_lock_init(&__ctx
->lock
);
1931 INIT_LIST_HEAD(&__ctx
->rq_list
);
1933 blk_stat_init(&__ctx
->stat
[BLK_STAT_READ
]);
1934 blk_stat_init(&__ctx
->stat
[BLK_STAT_WRITE
]);
1936 /* If the cpu isn't online, the cpu is mapped to first hctx */
1940 hctx
= blk_mq_map_queue(q
, i
);
1943 * Set local node, IFF we have more than one hw queue. If
1944 * not, we remain on the home node of the device
1946 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1947 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1951 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
1955 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
1956 set
->queue_depth
, set
->reserved_tags
);
1957 if (!set
->tags
[hctx_idx
])
1960 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
1965 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
1966 set
->tags
[hctx_idx
] = NULL
;
1970 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
1971 unsigned int hctx_idx
)
1973 if (set
->tags
[hctx_idx
]) {
1974 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
1975 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
1976 set
->tags
[hctx_idx
] = NULL
;
1980 static void blk_mq_map_swqueue(struct request_queue
*q
,
1981 const struct cpumask
*online_mask
)
1983 unsigned int i
, hctx_idx
;
1984 struct blk_mq_hw_ctx
*hctx
;
1985 struct blk_mq_ctx
*ctx
;
1986 struct blk_mq_tag_set
*set
= q
->tag_set
;
1989 * Avoid others reading imcomplete hctx->cpumask through sysfs
1991 mutex_lock(&q
->sysfs_lock
);
1993 queue_for_each_hw_ctx(q
, hctx
, i
) {
1994 cpumask_clear(hctx
->cpumask
);
1999 * Map software to hardware queues
2001 for_each_possible_cpu(i
) {
2002 /* If the cpu isn't online, the cpu is mapped to first hctx */
2003 if (!cpumask_test_cpu(i
, online_mask
))
2006 hctx_idx
= q
->mq_map
[i
];
2007 /* unmapped hw queue can be remapped after CPU topo changed */
2008 if (!set
->tags
[hctx_idx
] &&
2009 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2011 * If tags initialization fail for some hctx,
2012 * that hctx won't be brought online. In this
2013 * case, remap the current ctx to hctx[0] which
2014 * is guaranteed to always have tags allocated
2019 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2020 hctx
= blk_mq_map_queue(q
, i
);
2022 cpumask_set_cpu(i
, hctx
->cpumask
);
2023 ctx
->index_hw
= hctx
->nr_ctx
;
2024 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2027 mutex_unlock(&q
->sysfs_lock
);
2029 queue_for_each_hw_ctx(q
, hctx
, i
) {
2031 * If no software queues are mapped to this hardware queue,
2032 * disable it and free the request entries.
2034 if (!hctx
->nr_ctx
) {
2035 /* Never unmap queue 0. We need it as a
2036 * fallback in case of a new remap fails
2039 if (i
&& set
->tags
[i
])
2040 blk_mq_free_map_and_requests(set
, i
);
2046 hctx
->tags
= set
->tags
[i
];
2047 WARN_ON(!hctx
->tags
);
2050 * Set the map size to the number of mapped software queues.
2051 * This is more accurate and more efficient than looping
2052 * over all possibly mapped software queues.
2054 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2057 * Initialize batch roundrobin counts
2059 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2060 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2064 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2066 struct blk_mq_hw_ctx
*hctx
;
2069 queue_for_each_hw_ctx(q
, hctx
, i
) {
2071 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2073 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2077 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
2079 struct request_queue
*q
;
2081 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2082 blk_mq_freeze_queue(q
);
2083 queue_set_hctx_shared(q
, shared
);
2084 blk_mq_unfreeze_queue(q
);
2088 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2090 struct blk_mq_tag_set
*set
= q
->tag_set
;
2092 mutex_lock(&set
->tag_list_lock
);
2093 list_del_init(&q
->tag_set_list
);
2094 if (list_is_singular(&set
->tag_list
)) {
2095 /* just transitioned to unshared */
2096 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2097 /* update existing queue */
2098 blk_mq_update_tag_set_depth(set
, false);
2100 mutex_unlock(&set
->tag_list_lock
);
2103 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2104 struct request_queue
*q
)
2108 mutex_lock(&set
->tag_list_lock
);
2110 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2111 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2112 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2113 /* update existing queue */
2114 blk_mq_update_tag_set_depth(set
, true);
2116 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2117 queue_set_hctx_shared(q
, true);
2118 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
2120 mutex_unlock(&set
->tag_list_lock
);
2124 * It is the actual release handler for mq, but we do it from
2125 * request queue's release handler for avoiding use-after-free
2126 * and headache because q->mq_kobj shouldn't have been introduced,
2127 * but we can't group ctx/kctx kobj without it.
2129 void blk_mq_release(struct request_queue
*q
)
2131 struct blk_mq_hw_ctx
*hctx
;
2134 blk_mq_sched_teardown(q
);
2136 /* hctx kobj stays in hctx */
2137 queue_for_each_hw_ctx(q
, hctx
, i
) {
2146 kfree(q
->queue_hw_ctx
);
2148 /* ctx kobj stays in queue_ctx */
2149 free_percpu(q
->queue_ctx
);
2152 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2154 struct request_queue
*uninit_q
, *q
;
2156 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2158 return ERR_PTR(-ENOMEM
);
2160 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2162 blk_cleanup_queue(uninit_q
);
2166 EXPORT_SYMBOL(blk_mq_init_queue
);
2168 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2169 struct request_queue
*q
)
2172 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2174 blk_mq_sysfs_unregister(q
);
2175 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2181 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2182 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2187 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2194 atomic_set(&hctxs
[i
]->nr_active
, 0);
2195 hctxs
[i
]->numa_node
= node
;
2196 hctxs
[i
]->queue_num
= i
;
2198 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2199 free_cpumask_var(hctxs
[i
]->cpumask
);
2204 blk_mq_hctx_kobj_init(hctxs
[i
]);
2206 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2207 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2211 blk_mq_free_map_and_requests(set
, j
);
2212 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2213 free_cpumask_var(hctx
->cpumask
);
2214 kobject_put(&hctx
->kobj
);
2221 q
->nr_hw_queues
= i
;
2222 blk_mq_sysfs_register(q
);
2225 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2226 struct request_queue
*q
)
2228 /* mark the queue as mq asap */
2229 q
->mq_ops
= set
->ops
;
2231 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2235 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2236 GFP_KERNEL
, set
->numa_node
);
2237 if (!q
->queue_hw_ctx
)
2240 q
->mq_map
= set
->mq_map
;
2242 blk_mq_realloc_hw_ctxs(set
, q
);
2243 if (!q
->nr_hw_queues
)
2246 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2247 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2249 q
->nr_queues
= nr_cpu_ids
;
2251 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2253 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2254 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2256 q
->sg_reserved_size
= INT_MAX
;
2258 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2259 INIT_LIST_HEAD(&q
->requeue_list
);
2260 spin_lock_init(&q
->requeue_lock
);
2262 if (q
->nr_hw_queues
> 1)
2263 blk_queue_make_request(q
, blk_mq_make_request
);
2265 blk_queue_make_request(q
, blk_sq_make_request
);
2268 * Do this after blk_queue_make_request() overrides it...
2270 q
->nr_requests
= set
->queue_depth
;
2273 * Default to classic polling
2277 if (set
->ops
->complete
)
2278 blk_queue_softirq_done(q
, set
->ops
->complete
);
2280 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2283 mutex_lock(&all_q_mutex
);
2285 list_add_tail(&q
->all_q_node
, &all_q_list
);
2286 blk_mq_add_queue_tag_set(set
, q
);
2287 blk_mq_map_swqueue(q
, cpu_online_mask
);
2289 mutex_unlock(&all_q_mutex
);
2292 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2295 ret
= blk_mq_sched_init(q
);
2297 return ERR_PTR(ret
);
2303 kfree(q
->queue_hw_ctx
);
2305 free_percpu(q
->queue_ctx
);
2308 return ERR_PTR(-ENOMEM
);
2310 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2312 void blk_mq_free_queue(struct request_queue
*q
)
2314 struct blk_mq_tag_set
*set
= q
->tag_set
;
2316 mutex_lock(&all_q_mutex
);
2317 list_del_init(&q
->all_q_node
);
2318 mutex_unlock(&all_q_mutex
);
2322 blk_mq_del_queue_tag_set(q
);
2324 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2325 blk_mq_free_hw_queues(q
, set
);
2328 /* Basically redo blk_mq_init_queue with queue frozen */
2329 static void blk_mq_queue_reinit(struct request_queue
*q
,
2330 const struct cpumask
*online_mask
)
2332 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2334 blk_mq_sysfs_unregister(q
);
2337 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2338 * we should change hctx numa_node according to new topology (this
2339 * involves free and re-allocate memory, worthy doing?)
2342 blk_mq_map_swqueue(q
, online_mask
);
2344 blk_mq_sysfs_register(q
);
2348 * New online cpumask which is going to be set in this hotplug event.
2349 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2350 * one-by-one and dynamically allocating this could result in a failure.
2352 static struct cpumask cpuhp_online_new
;
2354 static void blk_mq_queue_reinit_work(void)
2356 struct request_queue
*q
;
2358 mutex_lock(&all_q_mutex
);
2360 * We need to freeze and reinit all existing queues. Freezing
2361 * involves synchronous wait for an RCU grace period and doing it
2362 * one by one may take a long time. Start freezing all queues in
2363 * one swoop and then wait for the completions so that freezing can
2364 * take place in parallel.
2366 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2367 blk_mq_freeze_queue_start(q
);
2368 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2369 blk_mq_freeze_queue_wait(q
);
2371 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2372 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2374 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2375 blk_mq_unfreeze_queue(q
);
2377 mutex_unlock(&all_q_mutex
);
2380 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2382 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2383 blk_mq_queue_reinit_work();
2388 * Before hotadded cpu starts handling requests, new mappings must be
2389 * established. Otherwise, these requests in hw queue might never be
2392 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2393 * for CPU0, and ctx1 for CPU1).
2395 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2396 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2398 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2399 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2400 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2403 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2405 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2406 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2407 blk_mq_queue_reinit_work();
2411 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2415 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2416 if (!__blk_mq_alloc_rq_map(set
, i
))
2423 blk_mq_free_rq_map(set
->tags
[i
]);
2429 * Allocate the request maps associated with this tag_set. Note that this
2430 * may reduce the depth asked for, if memory is tight. set->queue_depth
2431 * will be updated to reflect the allocated depth.
2433 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2438 depth
= set
->queue_depth
;
2440 err
= __blk_mq_alloc_rq_maps(set
);
2444 set
->queue_depth
>>= 1;
2445 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2449 } while (set
->queue_depth
);
2451 if (!set
->queue_depth
|| err
) {
2452 pr_err("blk-mq: failed to allocate request map\n");
2456 if (depth
!= set
->queue_depth
)
2457 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2458 depth
, set
->queue_depth
);
2464 * Alloc a tag set to be associated with one or more request queues.
2465 * May fail with EINVAL for various error conditions. May adjust the
2466 * requested depth down, if if it too large. In that case, the set
2467 * value will be stored in set->queue_depth.
2469 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2473 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2475 if (!set
->nr_hw_queues
)
2477 if (!set
->queue_depth
)
2479 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2482 if (!set
->ops
->queue_rq
)
2485 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2486 pr_info("blk-mq: reduced tag depth to %u\n",
2488 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2492 * If a crashdump is active, then we are potentially in a very
2493 * memory constrained environment. Limit us to 1 queue and
2494 * 64 tags to prevent using too much memory.
2496 if (is_kdump_kernel()) {
2497 set
->nr_hw_queues
= 1;
2498 set
->queue_depth
= min(64U, set
->queue_depth
);
2501 * There is no use for more h/w queues than cpus.
2503 if (set
->nr_hw_queues
> nr_cpu_ids
)
2504 set
->nr_hw_queues
= nr_cpu_ids
;
2506 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2507 GFP_KERNEL
, set
->numa_node
);
2512 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2513 GFP_KERNEL
, set
->numa_node
);
2517 if (set
->ops
->map_queues
)
2518 ret
= set
->ops
->map_queues(set
);
2520 ret
= blk_mq_map_queues(set
);
2522 goto out_free_mq_map
;
2524 ret
= blk_mq_alloc_rq_maps(set
);
2526 goto out_free_mq_map
;
2528 mutex_init(&set
->tag_list_lock
);
2529 INIT_LIST_HEAD(&set
->tag_list
);
2541 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2543 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2547 for (i
= 0; i
< nr_cpu_ids
; i
++)
2548 blk_mq_free_map_and_requests(set
, i
);
2556 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2558 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2560 struct blk_mq_tag_set
*set
= q
->tag_set
;
2561 struct blk_mq_hw_ctx
*hctx
;
2567 blk_mq_freeze_queue(q
);
2568 blk_mq_quiesce_queue(q
);
2571 queue_for_each_hw_ctx(q
, hctx
, i
) {
2575 * If we're using an MQ scheduler, just update the scheduler
2576 * queue depth. This is similar to what the old code would do.
2578 if (!hctx
->sched_tags
) {
2579 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2580 min(nr
, set
->queue_depth
),
2583 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2591 q
->nr_requests
= nr
;
2593 blk_mq_unfreeze_queue(q
);
2594 blk_mq_start_stopped_hw_queues(q
, true);
2599 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2601 struct request_queue
*q
;
2603 if (nr_hw_queues
> nr_cpu_ids
)
2604 nr_hw_queues
= nr_cpu_ids
;
2605 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2608 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2609 blk_mq_freeze_queue(q
);
2611 set
->nr_hw_queues
= nr_hw_queues
;
2612 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2613 blk_mq_realloc_hw_ctxs(set
, q
);
2615 if (q
->nr_hw_queues
> 1)
2616 blk_queue_make_request(q
, blk_mq_make_request
);
2618 blk_queue_make_request(q
, blk_sq_make_request
);
2620 blk_mq_queue_reinit(q
, cpu_online_mask
);
2623 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2624 blk_mq_unfreeze_queue(q
);
2626 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2628 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2629 struct blk_mq_hw_ctx
*hctx
,
2632 struct blk_rq_stat stat
[2];
2633 unsigned long ret
= 0;
2636 * If stats collection isn't on, don't sleep but turn it on for
2639 if (!blk_stat_enable(q
))
2643 * We don't have to do this once per IO, should optimize this
2644 * to just use the current window of stats until it changes
2646 memset(&stat
, 0, sizeof(stat
));
2647 blk_hctx_stat_get(hctx
, stat
);
2650 * As an optimistic guess, use half of the mean service time
2651 * for this type of request. We can (and should) make this smarter.
2652 * For instance, if the completion latencies are tight, we can
2653 * get closer than just half the mean. This is especially
2654 * important on devices where the completion latencies are longer
2657 if (req_op(rq
) == REQ_OP_READ
&& stat
[BLK_STAT_READ
].nr_samples
)
2658 ret
= (stat
[BLK_STAT_READ
].mean
+ 1) / 2;
2659 else if (req_op(rq
) == REQ_OP_WRITE
&& stat
[BLK_STAT_WRITE
].nr_samples
)
2660 ret
= (stat
[BLK_STAT_WRITE
].mean
+ 1) / 2;
2665 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2666 struct blk_mq_hw_ctx
*hctx
,
2669 struct hrtimer_sleeper hs
;
2670 enum hrtimer_mode mode
;
2674 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2680 * -1: don't ever hybrid sleep
2681 * 0: use half of prev avg
2682 * >0: use this specific value
2684 if (q
->poll_nsec
== -1)
2686 else if (q
->poll_nsec
> 0)
2687 nsecs
= q
->poll_nsec
;
2689 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2694 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2697 * This will be replaced with the stats tracking code, using
2698 * 'avg_completion_time / 2' as the pre-sleep target.
2702 mode
= HRTIMER_MODE_REL
;
2703 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2704 hrtimer_set_expires(&hs
.timer
, kt
);
2706 hrtimer_init_sleeper(&hs
, current
);
2708 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2710 set_current_state(TASK_UNINTERRUPTIBLE
);
2711 hrtimer_start_expires(&hs
.timer
, mode
);
2714 hrtimer_cancel(&hs
.timer
);
2715 mode
= HRTIMER_MODE_ABS
;
2716 } while (hs
.task
&& !signal_pending(current
));
2718 __set_current_state(TASK_RUNNING
);
2719 destroy_hrtimer_on_stack(&hs
.timer
);
2723 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2725 struct request_queue
*q
= hctx
->queue
;
2729 * If we sleep, have the caller restart the poll loop to reset
2730 * the state. Like for the other success return cases, the
2731 * caller is responsible for checking if the IO completed. If
2732 * the IO isn't complete, we'll get called again and will go
2733 * straight to the busy poll loop.
2735 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2738 hctx
->poll_considered
++;
2740 state
= current
->state
;
2741 while (!need_resched()) {
2744 hctx
->poll_invoked
++;
2746 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2748 hctx
->poll_success
++;
2749 set_current_state(TASK_RUNNING
);
2753 if (signal_pending_state(state
, current
))
2754 set_current_state(TASK_RUNNING
);
2756 if (current
->state
== TASK_RUNNING
)
2766 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2768 struct blk_mq_hw_ctx
*hctx
;
2769 struct blk_plug
*plug
;
2772 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2773 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2776 plug
= current
->plug
;
2778 blk_flush_plug_list(plug
, false);
2780 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2781 if (!blk_qc_t_is_internal(cookie
))
2782 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2784 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2786 return __blk_mq_poll(hctx
, rq
);
2788 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2790 void blk_mq_disable_hotplug(void)
2792 mutex_lock(&all_q_mutex
);
2795 void blk_mq_enable_hotplug(void)
2797 mutex_unlock(&all_q_mutex
);
2800 static int __init
blk_mq_init(void)
2802 blk_mq_debugfs_init();
2804 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2805 blk_mq_hctx_notify_dead
);
2807 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2808 blk_mq_queue_reinit_prepare
,
2809 blk_mq_queue_reinit_dead
);
2812 subsys_initcall(blk_mq_init
);