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 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 */
204 INIT_LIST_HEAD(&rq
->timeout_list
);
208 rq
->end_io_data
= NULL
;
211 ctx
->rq_dispatched
[op_is_sync(op
)]++;
213 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init
);
215 struct request
*__blk_mq_alloc_request(struct blk_mq_alloc_data
*data
,
221 tag
= blk_mq_get_tag(data
);
222 if (tag
!= BLK_MQ_TAG_FAIL
) {
223 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
225 rq
= tags
->static_rqs
[tag
];
227 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
229 rq
->internal_tag
= tag
;
231 if (blk_mq_tag_busy(data
->hctx
)) {
232 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
233 atomic_inc(&data
->hctx
->nr_active
);
236 rq
->internal_tag
= -1;
239 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
);
245 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request
);
247 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
250 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
254 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
258 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
260 blk_mq_put_ctx(alloc_data
.ctx
);
264 return ERR_PTR(-EWOULDBLOCK
);
267 rq
->__sector
= (sector_t
) -1;
268 rq
->bio
= rq
->biotail
= NULL
;
271 EXPORT_SYMBOL(blk_mq_alloc_request
);
273 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
274 unsigned int flags
, unsigned int hctx_idx
)
276 struct blk_mq_hw_ctx
*hctx
;
277 struct blk_mq_ctx
*ctx
;
279 struct blk_mq_alloc_data alloc_data
;
283 * If the tag allocator sleeps we could get an allocation for a
284 * different hardware context. No need to complicate the low level
285 * allocator for this for the rare use case of a command tied to
288 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
289 return ERR_PTR(-EINVAL
);
291 if (hctx_idx
>= q
->nr_hw_queues
)
292 return ERR_PTR(-EIO
);
294 ret
= blk_queue_enter(q
, true);
299 * Check if the hardware context is actually mapped to anything.
300 * If not tell the caller that it should skip this queue.
302 hctx
= q
->queue_hw_ctx
[hctx_idx
];
303 if (!blk_mq_hw_queue_mapped(hctx
)) {
307 ctx
= __blk_mq_get_ctx(q
, cpumask_first(hctx
->cpumask
));
309 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
310 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
322 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
324 void __blk_mq_finish_request(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
327 const int sched_tag
= rq
->internal_tag
;
328 struct request_queue
*q
= rq
->q
;
330 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
331 atomic_dec(&hctx
->nr_active
);
333 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
336 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
337 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
339 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
341 blk_mq_sched_completed_request(hctx
, rq
);
342 blk_mq_sched_restart_queues(hctx
);
346 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx
*hctx
,
349 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
351 ctx
->rq_completed
[rq_is_sync(rq
)]++;
352 __blk_mq_finish_request(hctx
, ctx
, rq
);
355 void blk_mq_finish_request(struct request
*rq
)
357 blk_mq_finish_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
360 void blk_mq_free_request(struct request
*rq
)
362 blk_mq_sched_put_request(rq
);
364 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
366 inline void __blk_mq_end_request(struct request
*rq
, int error
)
368 blk_account_io_done(rq
);
371 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
372 rq
->end_io(rq
, error
);
374 if (unlikely(blk_bidi_rq(rq
)))
375 blk_mq_free_request(rq
->next_rq
);
376 blk_mq_free_request(rq
);
379 EXPORT_SYMBOL(__blk_mq_end_request
);
381 void blk_mq_end_request(struct request
*rq
, int error
)
383 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
385 __blk_mq_end_request(rq
, error
);
387 EXPORT_SYMBOL(blk_mq_end_request
);
389 static void __blk_mq_complete_request_remote(void *data
)
391 struct request
*rq
= data
;
393 rq
->q
->softirq_done_fn(rq
);
396 static void blk_mq_ipi_complete_request(struct request
*rq
)
398 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
402 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
403 rq
->q
->softirq_done_fn(rq
);
408 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
409 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
411 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
412 rq
->csd
.func
= __blk_mq_complete_request_remote
;
415 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
417 rq
->q
->softirq_done_fn(rq
);
422 static void blk_mq_stat_add(struct request
*rq
)
424 if (rq
->rq_flags
& RQF_STATS
) {
426 * We could rq->mq_ctx here, but there's less of a risk
427 * of races if we have the completion event add the stats
428 * to the local software queue.
430 struct blk_mq_ctx
*ctx
;
432 ctx
= __blk_mq_get_ctx(rq
->q
, raw_smp_processor_id());
433 blk_stat_add(&ctx
->stat
[rq_data_dir(rq
)], rq
);
437 static void __blk_mq_complete_request(struct request
*rq
)
439 struct request_queue
*q
= rq
->q
;
443 if (!q
->softirq_done_fn
)
444 blk_mq_end_request(rq
, rq
->errors
);
446 blk_mq_ipi_complete_request(rq
);
450 * blk_mq_complete_request - end I/O on a request
451 * @rq: the request being processed
454 * Ends all I/O on a request. It does not handle partial completions.
455 * The actual completion happens out-of-order, through a IPI handler.
457 void blk_mq_complete_request(struct request
*rq
, int error
)
459 struct request_queue
*q
= rq
->q
;
461 if (unlikely(blk_should_fake_timeout(q
)))
463 if (!blk_mark_rq_complete(rq
)) {
465 __blk_mq_complete_request(rq
);
468 EXPORT_SYMBOL(blk_mq_complete_request
);
470 int blk_mq_request_started(struct request
*rq
)
472 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
474 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
476 void blk_mq_start_request(struct request
*rq
)
478 struct request_queue
*q
= rq
->q
;
480 blk_mq_sched_started_request(rq
);
482 trace_block_rq_issue(q
, rq
);
484 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
485 blk_stat_set_issue_time(&rq
->issue_stat
);
486 rq
->rq_flags
|= RQF_STATS
;
487 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
493 * Ensure that ->deadline is visible before set the started
494 * flag and clear the completed flag.
496 smp_mb__before_atomic();
499 * Mark us as started and clear complete. Complete might have been
500 * set if requeue raced with timeout, which then marked it as
501 * complete. So be sure to clear complete again when we start
502 * the request, otherwise we'll ignore the completion event.
504 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
505 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
506 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
507 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
509 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
511 * Make sure space for the drain appears. We know we can do
512 * this because max_hw_segments has been adjusted to be one
513 * fewer than the device can handle.
515 rq
->nr_phys_segments
++;
518 EXPORT_SYMBOL(blk_mq_start_request
);
520 static void __blk_mq_requeue_request(struct request
*rq
)
522 struct request_queue
*q
= rq
->q
;
524 trace_block_rq_requeue(q
, rq
);
525 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
526 blk_mq_sched_requeue_request(rq
);
528 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
529 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
530 rq
->nr_phys_segments
--;
534 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
536 __blk_mq_requeue_request(rq
);
538 BUG_ON(blk_queued_rq(rq
));
539 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
541 EXPORT_SYMBOL(blk_mq_requeue_request
);
543 static void blk_mq_requeue_work(struct work_struct
*work
)
545 struct request_queue
*q
=
546 container_of(work
, struct request_queue
, requeue_work
.work
);
548 struct request
*rq
, *next
;
551 spin_lock_irqsave(&q
->requeue_lock
, flags
);
552 list_splice_init(&q
->requeue_list
, &rq_list
);
553 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
555 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
556 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
559 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
560 list_del_init(&rq
->queuelist
);
561 blk_mq_sched_insert_request(rq
, true, false, false, true);
564 while (!list_empty(&rq_list
)) {
565 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
566 list_del_init(&rq
->queuelist
);
567 blk_mq_sched_insert_request(rq
, false, false, false, true);
570 blk_mq_run_hw_queues(q
, false);
573 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
574 bool kick_requeue_list
)
576 struct request_queue
*q
= rq
->q
;
580 * We abuse this flag that is otherwise used by the I/O scheduler to
581 * request head insertation from the workqueue.
583 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
585 spin_lock_irqsave(&q
->requeue_lock
, flags
);
587 rq
->rq_flags
|= RQF_SOFTBARRIER
;
588 list_add(&rq
->queuelist
, &q
->requeue_list
);
590 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
592 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
594 if (kick_requeue_list
)
595 blk_mq_kick_requeue_list(q
);
597 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
599 void blk_mq_kick_requeue_list(struct request_queue
*q
)
601 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
603 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
605 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
608 kblockd_schedule_delayed_work(&q
->requeue_work
,
609 msecs_to_jiffies(msecs
));
611 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
613 void blk_mq_abort_requeue_list(struct request_queue
*q
)
618 spin_lock_irqsave(&q
->requeue_lock
, flags
);
619 list_splice_init(&q
->requeue_list
, &rq_list
);
620 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
622 while (!list_empty(&rq_list
)) {
625 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
626 list_del_init(&rq
->queuelist
);
628 blk_mq_end_request(rq
, rq
->errors
);
631 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
633 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
635 if (tag
< tags
->nr_tags
) {
636 prefetch(tags
->rqs
[tag
]);
637 return tags
->rqs
[tag
];
642 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
644 struct blk_mq_timeout_data
{
646 unsigned int next_set
;
649 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
651 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
652 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
655 * We know that complete is set at this point. If STARTED isn't set
656 * anymore, then the request isn't active and the "timeout" should
657 * just be ignored. This can happen due to the bitflag ordering.
658 * Timeout first checks if STARTED is set, and if it is, assumes
659 * the request is active. But if we race with completion, then
660 * we both flags will get cleared. So check here again, and ignore
661 * a timeout event with a request that isn't active.
663 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
667 ret
= ops
->timeout(req
, reserved
);
671 __blk_mq_complete_request(req
);
673 case BLK_EH_RESET_TIMER
:
675 blk_clear_rq_complete(req
);
677 case BLK_EH_NOT_HANDLED
:
680 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
685 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
686 struct request
*rq
, void *priv
, bool reserved
)
688 struct blk_mq_timeout_data
*data
= priv
;
690 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
692 * If a request wasn't started before the queue was
693 * marked dying, kill it here or it'll go unnoticed.
695 if (unlikely(blk_queue_dying(rq
->q
))) {
697 blk_mq_end_request(rq
, rq
->errors
);
702 if (time_after_eq(jiffies
, rq
->deadline
)) {
703 if (!blk_mark_rq_complete(rq
))
704 blk_mq_rq_timed_out(rq
, reserved
);
705 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
706 data
->next
= rq
->deadline
;
711 static void blk_mq_timeout_work(struct work_struct
*work
)
713 struct request_queue
*q
=
714 container_of(work
, struct request_queue
, timeout_work
);
715 struct blk_mq_timeout_data data
= {
721 /* A deadlock might occur if a request is stuck requiring a
722 * timeout at the same time a queue freeze is waiting
723 * completion, since the timeout code would not be able to
724 * acquire the queue reference here.
726 * That's why we don't use blk_queue_enter here; instead, we use
727 * percpu_ref_tryget directly, because we need to be able to
728 * obtain a reference even in the short window between the queue
729 * starting to freeze, by dropping the first reference in
730 * blk_mq_freeze_queue_start, and the moment the last request is
731 * consumed, marked by the instant q_usage_counter reaches
734 if (!percpu_ref_tryget(&q
->q_usage_counter
))
737 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
740 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
741 mod_timer(&q
->timeout
, data
.next
);
743 struct blk_mq_hw_ctx
*hctx
;
745 queue_for_each_hw_ctx(q
, hctx
, i
) {
746 /* the hctx may be unmapped, so check it here */
747 if (blk_mq_hw_queue_mapped(hctx
))
748 blk_mq_tag_idle(hctx
);
755 * Reverse check our software queue for entries that we could potentially
756 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
757 * too much time checking for merges.
759 static bool blk_mq_attempt_merge(struct request_queue
*q
,
760 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
765 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
771 if (!blk_rq_merge_ok(rq
, bio
))
774 switch (blk_try_merge(rq
, bio
)) {
775 case ELEVATOR_BACK_MERGE
:
776 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
777 merged
= bio_attempt_back_merge(q
, rq
, bio
);
779 case ELEVATOR_FRONT_MERGE
:
780 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
781 merged
= bio_attempt_front_merge(q
, rq
, bio
);
783 case ELEVATOR_DISCARD_MERGE
:
784 merged
= bio_attempt_discard_merge(q
, rq
, bio
);
798 struct flush_busy_ctx_data
{
799 struct blk_mq_hw_ctx
*hctx
;
800 struct list_head
*list
;
803 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
805 struct flush_busy_ctx_data
*flush_data
= data
;
806 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
807 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
809 sbitmap_clear_bit(sb
, bitnr
);
810 spin_lock(&ctx
->lock
);
811 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
812 spin_unlock(&ctx
->lock
);
817 * Process software queues that have been marked busy, splicing them
818 * to the for-dispatch
820 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
822 struct flush_busy_ctx_data data
= {
827 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
829 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
831 static inline unsigned int queued_to_index(unsigned int queued
)
836 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
839 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
842 struct blk_mq_alloc_data data
= {
844 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
845 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
855 rq
->tag
= blk_mq_get_tag(&data
);
857 if (blk_mq_tag_busy(data
.hctx
)) {
858 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
859 atomic_inc(&data
.hctx
->nr_active
);
861 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
868 static void blk_mq_put_driver_tag(struct blk_mq_hw_ctx
*hctx
,
871 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
874 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
877 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
878 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
879 atomic_dec(&hctx
->nr_active
);
884 * If we fail getting a driver tag because all the driver tags are already
885 * assigned and on the dispatch list, BUT the first entry does not have a
886 * tag, then we could deadlock. For that case, move entries with assigned
887 * driver tags to the front, leaving the set of tagged requests in the
888 * same order, and the untagged set in the same order.
890 static bool reorder_tags_to_front(struct list_head
*list
)
892 struct request
*rq
, *tmp
, *first
= NULL
;
894 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
898 list_move(&rq
->queuelist
, list
);
904 return first
!= NULL
;
907 static int blk_mq_dispatch_wake(wait_queue_t
*wait
, unsigned mode
, int flags
,
910 struct blk_mq_hw_ctx
*hctx
;
912 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
914 list_del(&wait
->task_list
);
915 clear_bit_unlock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
);
916 blk_mq_run_hw_queue(hctx
, true);
920 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx
*hctx
)
922 struct sbq_wait_state
*ws
;
925 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
926 * The thread which wins the race to grab this bit adds the hardware
927 * queue to the wait queue.
929 if (test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
) ||
930 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
933 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
934 ws
= bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
);
937 * As soon as this returns, it's no longer safe to fiddle with
938 * hctx->dispatch_wait, since a completion can wake up the wait queue
939 * and unlock the bit.
941 add_wait_queue(&ws
->wait
, &hctx
->dispatch_wait
);
945 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
947 struct request_queue
*q
= hctx
->queue
;
949 LIST_HEAD(driver_list
);
950 struct list_head
*dptr
;
951 int queued
, ret
= BLK_MQ_RQ_QUEUE_OK
;
954 * Start off with dptr being NULL, so we start the first request
955 * immediately, even if we have more pending.
960 * Now process all the entries, sending them to the driver.
963 while (!list_empty(list
)) {
964 struct blk_mq_queue_data bd
;
966 rq
= list_first_entry(list
, struct request
, queuelist
);
967 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
968 if (!queued
&& reorder_tags_to_front(list
))
972 * The initial allocation attempt failed, so we need to
973 * rerun the hardware queue when a tag is freed.
975 if (blk_mq_dispatch_wait_add(hctx
)) {
977 * It's possible that a tag was freed in the
978 * window between the allocation failure and
979 * adding the hardware queue to the wait queue.
981 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
988 list_del_init(&rq
->queuelist
);
992 bd
.last
= list_empty(list
);
994 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
996 case BLK_MQ_RQ_QUEUE_OK
:
999 case BLK_MQ_RQ_QUEUE_BUSY
:
1000 blk_mq_put_driver_tag(hctx
, rq
);
1001 list_add(&rq
->queuelist
, list
);
1002 __blk_mq_requeue_request(rq
);
1005 pr_err("blk-mq: bad return on queue: %d\n", ret
);
1006 case BLK_MQ_RQ_QUEUE_ERROR
:
1008 blk_mq_end_request(rq
, rq
->errors
);
1012 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
1016 * We've done the first request. If we have more than 1
1017 * left in the list, set dptr to defer issue.
1019 if (!dptr
&& list
->next
!= list
->prev
)
1020 dptr
= &driver_list
;
1023 hctx
->dispatched
[queued_to_index(queued
)]++;
1026 * Any items that need requeuing? Stuff them into hctx->dispatch,
1027 * that is where we will continue on next queue run.
1029 if (!list_empty(list
)) {
1030 spin_lock(&hctx
->lock
);
1031 list_splice_init(list
, &hctx
->dispatch
);
1032 spin_unlock(&hctx
->lock
);
1035 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
1036 * it's possible the queue is stopped and restarted again
1037 * before this. Queue restart will dispatch requests. And since
1038 * requests in rq_list aren't added into hctx->dispatch yet,
1039 * the requests in rq_list might get lost.
1041 * blk_mq_run_hw_queue() already checks the STOPPED bit
1043 * If RESTART or TAG_WAITING is set, then let completion restart
1044 * the queue instead of potentially looping here.
1046 if (!blk_mq_sched_needs_restart(hctx
) &&
1047 !test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1048 blk_mq_run_hw_queue(hctx
, true);
1054 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1058 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1059 cpu_online(hctx
->next_cpu
));
1061 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1063 blk_mq_sched_dispatch_requests(hctx
);
1066 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1067 blk_mq_sched_dispatch_requests(hctx
);
1068 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1073 * It'd be great if the workqueue API had a way to pass
1074 * in a mask and had some smarts for more clever placement.
1075 * For now we just round-robin here, switching for every
1076 * BLK_MQ_CPU_WORK_BATCH queued items.
1078 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1080 if (hctx
->queue
->nr_hw_queues
== 1)
1081 return WORK_CPU_UNBOUND
;
1083 if (--hctx
->next_cpu_batch
<= 0) {
1086 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1087 if (next_cpu
>= nr_cpu_ids
)
1088 next_cpu
= cpumask_first(hctx
->cpumask
);
1090 hctx
->next_cpu
= next_cpu
;
1091 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1094 return hctx
->next_cpu
;
1097 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1099 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
1100 !blk_mq_hw_queue_mapped(hctx
)))
1103 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1104 int cpu
= get_cpu();
1105 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1106 __blk_mq_run_hw_queue(hctx
);
1114 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
);
1117 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1119 struct blk_mq_hw_ctx
*hctx
;
1122 queue_for_each_hw_ctx(q
, hctx
, i
) {
1123 if (!blk_mq_hctx_has_pending(hctx
) ||
1124 blk_mq_hctx_stopped(hctx
))
1127 blk_mq_run_hw_queue(hctx
, async
);
1130 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1133 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1134 * @q: request queue.
1136 * The caller is responsible for serializing this function against
1137 * blk_mq_{start,stop}_hw_queue().
1139 bool blk_mq_queue_stopped(struct request_queue
*q
)
1141 struct blk_mq_hw_ctx
*hctx
;
1144 queue_for_each_hw_ctx(q
, hctx
, i
)
1145 if (blk_mq_hctx_stopped(hctx
))
1150 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1152 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1154 cancel_work(&hctx
->run_work
);
1155 cancel_delayed_work(&hctx
->delay_work
);
1156 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1158 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1160 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1162 struct blk_mq_hw_ctx
*hctx
;
1165 queue_for_each_hw_ctx(q
, hctx
, i
)
1166 blk_mq_stop_hw_queue(hctx
);
1168 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1170 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1172 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1174 blk_mq_run_hw_queue(hctx
, false);
1176 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1178 void blk_mq_start_hw_queues(struct request_queue
*q
)
1180 struct blk_mq_hw_ctx
*hctx
;
1183 queue_for_each_hw_ctx(q
, hctx
, i
)
1184 blk_mq_start_hw_queue(hctx
);
1186 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1188 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1190 if (!blk_mq_hctx_stopped(hctx
))
1193 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1194 blk_mq_run_hw_queue(hctx
, async
);
1196 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1198 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1200 struct blk_mq_hw_ctx
*hctx
;
1203 queue_for_each_hw_ctx(q
, hctx
, i
)
1204 blk_mq_start_stopped_hw_queue(hctx
, async
);
1206 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1208 static void blk_mq_run_work_fn(struct work_struct
*work
)
1210 struct blk_mq_hw_ctx
*hctx
;
1212 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1214 __blk_mq_run_hw_queue(hctx
);
1217 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1219 struct blk_mq_hw_ctx
*hctx
;
1221 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1223 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1224 __blk_mq_run_hw_queue(hctx
);
1227 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1229 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1232 blk_mq_stop_hw_queue(hctx
);
1233 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1234 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1236 EXPORT_SYMBOL(blk_mq_delay_queue
);
1238 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1242 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1244 trace_block_rq_insert(hctx
->queue
, rq
);
1247 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1249 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1252 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1255 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1257 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1258 blk_mq_hctx_mark_pending(hctx
, ctx
);
1261 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1262 struct list_head
*list
)
1266 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1269 spin_lock(&ctx
->lock
);
1270 while (!list_empty(list
)) {
1273 rq
= list_first_entry(list
, struct request
, queuelist
);
1274 BUG_ON(rq
->mq_ctx
!= ctx
);
1275 list_del_init(&rq
->queuelist
);
1276 __blk_mq_insert_req_list(hctx
, rq
, false);
1278 blk_mq_hctx_mark_pending(hctx
, ctx
);
1279 spin_unlock(&ctx
->lock
);
1282 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1284 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1285 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1287 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1288 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1289 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1292 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1294 struct blk_mq_ctx
*this_ctx
;
1295 struct request_queue
*this_q
;
1298 LIST_HEAD(ctx_list
);
1301 list_splice_init(&plug
->mq_list
, &list
);
1303 list_sort(NULL
, &list
, plug_ctx_cmp
);
1309 while (!list_empty(&list
)) {
1310 rq
= list_entry_rq(list
.next
);
1311 list_del_init(&rq
->queuelist
);
1313 if (rq
->mq_ctx
!= this_ctx
) {
1315 trace_block_unplug(this_q
, depth
, from_schedule
);
1316 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1321 this_ctx
= rq
->mq_ctx
;
1327 list_add_tail(&rq
->queuelist
, &ctx_list
);
1331 * If 'this_ctx' is set, we know we have entries to complete
1332 * on 'ctx_list'. Do those.
1335 trace_block_unplug(this_q
, depth
, from_schedule
);
1336 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1341 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1343 init_request_from_bio(rq
, bio
);
1345 blk_account_io_start(rq
, true);
1348 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1350 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1351 !blk_queue_nomerges(hctx
->queue
);
1354 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1355 struct blk_mq_ctx
*ctx
,
1356 struct request
*rq
, struct bio
*bio
)
1358 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1359 blk_mq_bio_to_request(rq
, bio
);
1360 spin_lock(&ctx
->lock
);
1362 __blk_mq_insert_request(hctx
, rq
, false);
1363 spin_unlock(&ctx
->lock
);
1366 struct request_queue
*q
= hctx
->queue
;
1368 spin_lock(&ctx
->lock
);
1369 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1370 blk_mq_bio_to_request(rq
, bio
);
1374 spin_unlock(&ctx
->lock
);
1375 __blk_mq_finish_request(hctx
, ctx
, rq
);
1380 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1383 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1385 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1388 static void blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
)
1390 struct request_queue
*q
= rq
->q
;
1391 struct blk_mq_queue_data bd
= {
1396 struct blk_mq_hw_ctx
*hctx
;
1397 blk_qc_t new_cookie
;
1403 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1406 new_cookie
= request_to_qc_t(hctx
, rq
);
1409 * For OK queue, we are done. For error, kill it. Any other
1410 * error (busy), just add it to our list as we previously
1413 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1414 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1415 *cookie
= new_cookie
;
1419 __blk_mq_requeue_request(rq
);
1421 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1422 *cookie
= BLK_QC_T_NONE
;
1424 blk_mq_end_request(rq
, rq
->errors
);
1429 blk_mq_sched_insert_request(rq
, false, true, true, false);
1433 * Multiple hardware queue variant. This will not use per-process plugs,
1434 * but will attempt to bypass the hctx queueing if we can go straight to
1435 * hardware for SYNC IO.
1437 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1439 const int is_sync
= op_is_sync(bio
->bi_opf
);
1440 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1441 struct blk_mq_alloc_data data
= { .flags
= 0 };
1443 unsigned int request_count
= 0, srcu_idx
;
1444 struct blk_plug
*plug
;
1445 struct request
*same_queue_rq
= NULL
;
1447 unsigned int wb_acct
;
1449 blk_queue_bounce(q
, &bio
);
1451 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1453 return BLK_QC_T_NONE
;
1456 blk_queue_split(q
, &bio
, q
->bio_split
);
1458 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1459 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1460 return BLK_QC_T_NONE
;
1462 if (blk_mq_sched_bio_merge(q
, bio
))
1463 return BLK_QC_T_NONE
;
1465 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1467 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1469 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1470 if (unlikely(!rq
)) {
1471 __wbt_done(q
->rq_wb
, wb_acct
);
1472 return BLK_QC_T_NONE
;
1475 wbt_track(&rq
->issue_stat
, wb_acct
);
1477 cookie
= request_to_qc_t(data
.hctx
, rq
);
1479 if (unlikely(is_flush_fua
)) {
1482 blk_mq_bio_to_request(rq
, bio
);
1483 blk_insert_flush(rq
);
1487 plug
= current
->plug
;
1489 * If the driver supports defer issued based on 'last', then
1490 * queue it up like normal since we can potentially save some
1493 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1494 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1495 struct request
*old_rq
= NULL
;
1497 blk_mq_bio_to_request(rq
, bio
);
1500 * We do limited plugging. If the bio can be merged, do that.
1501 * Otherwise the existing request in the plug list will be
1502 * issued. So the plug list will have one request at most
1506 * The plug list might get flushed before this. If that
1507 * happens, same_queue_rq is invalid and plug list is
1510 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1511 old_rq
= same_queue_rq
;
1512 list_del_init(&old_rq
->queuelist
);
1514 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1515 } else /* is_sync */
1517 blk_mq_put_ctx(data
.ctx
);
1521 if (!(data
.hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1523 blk_mq_try_issue_directly(old_rq
, &cookie
);
1526 srcu_idx
= srcu_read_lock(&data
.hctx
->queue_rq_srcu
);
1527 blk_mq_try_issue_directly(old_rq
, &cookie
);
1528 srcu_read_unlock(&data
.hctx
->queue_rq_srcu
, srcu_idx
);
1535 blk_mq_put_ctx(data
.ctx
);
1536 blk_mq_bio_to_request(rq
, bio
);
1537 blk_mq_sched_insert_request(rq
, false, true,
1538 !is_sync
|| is_flush_fua
, true);
1541 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1543 * For a SYNC request, send it to the hardware immediately. For
1544 * an ASYNC request, just ensure that we run it later on. The
1545 * latter allows for merging opportunities and more efficient
1549 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1551 blk_mq_put_ctx(data
.ctx
);
1557 * Single hardware queue variant. This will attempt to use any per-process
1558 * plug for merging and IO deferral.
1560 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1562 const int is_sync
= op_is_sync(bio
->bi_opf
);
1563 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1564 struct blk_plug
*plug
;
1565 unsigned int request_count
= 0;
1566 struct blk_mq_alloc_data data
= { .flags
= 0 };
1569 unsigned int wb_acct
;
1571 blk_queue_bounce(q
, &bio
);
1573 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1575 return BLK_QC_T_NONE
;
1578 blk_queue_split(q
, &bio
, q
->bio_split
);
1580 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1581 if (blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1582 return BLK_QC_T_NONE
;
1584 request_count
= blk_plug_queued_count(q
);
1586 if (blk_mq_sched_bio_merge(q
, bio
))
1587 return BLK_QC_T_NONE
;
1589 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1591 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1593 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1594 if (unlikely(!rq
)) {
1595 __wbt_done(q
->rq_wb
, wb_acct
);
1596 return BLK_QC_T_NONE
;
1599 wbt_track(&rq
->issue_stat
, wb_acct
);
1601 cookie
= request_to_qc_t(data
.hctx
, rq
);
1603 if (unlikely(is_flush_fua
)) {
1606 blk_mq_bio_to_request(rq
, bio
);
1607 blk_insert_flush(rq
);
1612 * A task plug currently exists. Since this is completely lockless,
1613 * utilize that to temporarily store requests until the task is
1614 * either done or scheduled away.
1616 plug
= current
->plug
;
1618 struct request
*last
= NULL
;
1620 blk_mq_bio_to_request(rq
, bio
);
1623 * @request_count may become stale because of schedule
1624 * out, so check the list again.
1626 if (list_empty(&plug
->mq_list
))
1629 trace_block_plug(q
);
1631 last
= list_entry_rq(plug
->mq_list
.prev
);
1633 blk_mq_put_ctx(data
.ctx
);
1635 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1636 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1637 blk_flush_plug_list(plug
, false);
1638 trace_block_plug(q
);
1641 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1647 blk_mq_put_ctx(data
.ctx
);
1648 blk_mq_bio_to_request(rq
, bio
);
1649 blk_mq_sched_insert_request(rq
, false, true,
1650 !is_sync
|| is_flush_fua
, true);
1653 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1655 * For a SYNC request, send it to the hardware immediately. For
1656 * an ASYNC request, just ensure that we run it later on. The
1657 * latter allows for merging opportunities and more efficient
1661 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1664 blk_mq_put_ctx(data
.ctx
);
1669 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1670 unsigned int hctx_idx
)
1674 if (tags
->rqs
&& set
->ops
->exit_request
) {
1677 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1678 struct request
*rq
= tags
->static_rqs
[i
];
1682 set
->ops
->exit_request(set
->driver_data
, rq
,
1684 tags
->static_rqs
[i
] = NULL
;
1688 while (!list_empty(&tags
->page_list
)) {
1689 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1690 list_del_init(&page
->lru
);
1692 * Remove kmemleak object previously allocated in
1693 * blk_mq_init_rq_map().
1695 kmemleak_free(page_address(page
));
1696 __free_pages(page
, page
->private);
1700 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1704 kfree(tags
->static_rqs
);
1705 tags
->static_rqs
= NULL
;
1707 blk_mq_free_tags(tags
);
1710 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1711 unsigned int hctx_idx
,
1712 unsigned int nr_tags
,
1713 unsigned int reserved_tags
)
1715 struct blk_mq_tags
*tags
;
1717 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
,
1719 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1723 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1724 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1727 blk_mq_free_tags(tags
);
1731 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1732 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1734 if (!tags
->static_rqs
) {
1736 blk_mq_free_tags(tags
);
1743 static size_t order_to_size(unsigned int order
)
1745 return (size_t)PAGE_SIZE
<< order
;
1748 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1749 unsigned int hctx_idx
, unsigned int depth
)
1751 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1752 size_t rq_size
, left
;
1754 INIT_LIST_HEAD(&tags
->page_list
);
1757 * rq_size is the size of the request plus driver payload, rounded
1758 * to the cacheline size
1760 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1762 left
= rq_size
* depth
;
1764 for (i
= 0; i
< depth
; ) {
1765 int this_order
= max_order
;
1770 while (this_order
&& left
< order_to_size(this_order
- 1))
1774 page
= alloc_pages_node(set
->numa_node
,
1775 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1781 if (order_to_size(this_order
) < rq_size
)
1788 page
->private = this_order
;
1789 list_add_tail(&page
->lru
, &tags
->page_list
);
1791 p
= page_address(page
);
1793 * Allow kmemleak to scan these pages as they contain pointers
1794 * to additional allocations like via ops->init_request().
1796 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1797 entries_per_page
= order_to_size(this_order
) / rq_size
;
1798 to_do
= min(entries_per_page
, depth
- i
);
1799 left
-= to_do
* rq_size
;
1800 for (j
= 0; j
< to_do
; j
++) {
1801 struct request
*rq
= p
;
1803 tags
->static_rqs
[i
] = rq
;
1804 if (set
->ops
->init_request
) {
1805 if (set
->ops
->init_request(set
->driver_data
,
1808 tags
->static_rqs
[i
] = NULL
;
1820 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1825 * 'cpu' is going away. splice any existing rq_list entries from this
1826 * software queue to the hw queue dispatch list, and ensure that it
1829 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1831 struct blk_mq_hw_ctx
*hctx
;
1832 struct blk_mq_ctx
*ctx
;
1835 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1836 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1838 spin_lock(&ctx
->lock
);
1839 if (!list_empty(&ctx
->rq_list
)) {
1840 list_splice_init(&ctx
->rq_list
, &tmp
);
1841 blk_mq_hctx_clear_pending(hctx
, ctx
);
1843 spin_unlock(&ctx
->lock
);
1845 if (list_empty(&tmp
))
1848 spin_lock(&hctx
->lock
);
1849 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1850 spin_unlock(&hctx
->lock
);
1852 blk_mq_run_hw_queue(hctx
, true);
1856 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1858 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1862 /* hctx->ctxs will be freed in queue's release handler */
1863 static void blk_mq_exit_hctx(struct request_queue
*q
,
1864 struct blk_mq_tag_set
*set
,
1865 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1867 unsigned flush_start_tag
= set
->queue_depth
;
1869 blk_mq_tag_idle(hctx
);
1871 if (set
->ops
->exit_request
)
1872 set
->ops
->exit_request(set
->driver_data
,
1873 hctx
->fq
->flush_rq
, hctx_idx
,
1874 flush_start_tag
+ hctx_idx
);
1876 if (set
->ops
->exit_hctx
)
1877 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1879 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1880 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1882 blk_mq_remove_cpuhp(hctx
);
1883 blk_free_flush_queue(hctx
->fq
);
1884 sbitmap_free(&hctx
->ctx_map
);
1887 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1888 struct blk_mq_tag_set
*set
, int nr_queue
)
1890 struct blk_mq_hw_ctx
*hctx
;
1893 queue_for_each_hw_ctx(q
, hctx
, i
) {
1896 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1900 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1901 struct blk_mq_tag_set
*set
)
1903 struct blk_mq_hw_ctx
*hctx
;
1906 queue_for_each_hw_ctx(q
, hctx
, i
)
1907 free_cpumask_var(hctx
->cpumask
);
1910 static int blk_mq_init_hctx(struct request_queue
*q
,
1911 struct blk_mq_tag_set
*set
,
1912 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1915 unsigned flush_start_tag
= set
->queue_depth
;
1917 node
= hctx
->numa_node
;
1918 if (node
== NUMA_NO_NODE
)
1919 node
= hctx
->numa_node
= set
->numa_node
;
1921 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1922 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1923 spin_lock_init(&hctx
->lock
);
1924 INIT_LIST_HEAD(&hctx
->dispatch
);
1926 hctx
->queue_num
= hctx_idx
;
1927 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1929 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1931 hctx
->tags
= set
->tags
[hctx_idx
];
1934 * Allocate space for all possible cpus to avoid allocation at
1937 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1940 goto unregister_cpu_notifier
;
1942 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1948 if (set
->ops
->init_hctx
&&
1949 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1952 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1956 if (set
->ops
->init_request
&&
1957 set
->ops
->init_request(set
->driver_data
,
1958 hctx
->fq
->flush_rq
, hctx_idx
,
1959 flush_start_tag
+ hctx_idx
, node
))
1962 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1963 init_srcu_struct(&hctx
->queue_rq_srcu
);
1970 if (set
->ops
->exit_hctx
)
1971 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1973 sbitmap_free(&hctx
->ctx_map
);
1976 unregister_cpu_notifier
:
1977 blk_mq_remove_cpuhp(hctx
);
1981 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1982 unsigned int nr_hw_queues
)
1986 for_each_possible_cpu(i
) {
1987 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1988 struct blk_mq_hw_ctx
*hctx
;
1990 memset(__ctx
, 0, sizeof(*__ctx
));
1992 spin_lock_init(&__ctx
->lock
);
1993 INIT_LIST_HEAD(&__ctx
->rq_list
);
1995 blk_stat_init(&__ctx
->stat
[BLK_STAT_READ
]);
1996 blk_stat_init(&__ctx
->stat
[BLK_STAT_WRITE
]);
1998 /* If the cpu isn't online, the cpu is mapped to first hctx */
2002 hctx
= blk_mq_map_queue(q
, i
);
2005 * Set local node, IFF we have more than one hw queue. If
2006 * not, we remain on the home node of the device
2008 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2009 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2013 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2017 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2018 set
->queue_depth
, set
->reserved_tags
);
2019 if (!set
->tags
[hctx_idx
])
2022 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2027 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2028 set
->tags
[hctx_idx
] = NULL
;
2032 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2033 unsigned int hctx_idx
)
2035 if (set
->tags
[hctx_idx
]) {
2036 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2037 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2038 set
->tags
[hctx_idx
] = NULL
;
2042 static void blk_mq_map_swqueue(struct request_queue
*q
,
2043 const struct cpumask
*online_mask
)
2045 unsigned int i
, hctx_idx
;
2046 struct blk_mq_hw_ctx
*hctx
;
2047 struct blk_mq_ctx
*ctx
;
2048 struct blk_mq_tag_set
*set
= q
->tag_set
;
2051 * Avoid others reading imcomplete hctx->cpumask through sysfs
2053 mutex_lock(&q
->sysfs_lock
);
2055 queue_for_each_hw_ctx(q
, hctx
, i
) {
2056 cpumask_clear(hctx
->cpumask
);
2061 * Map software to hardware queues
2063 for_each_possible_cpu(i
) {
2064 /* If the cpu isn't online, the cpu is mapped to first hctx */
2065 if (!cpumask_test_cpu(i
, online_mask
))
2068 hctx_idx
= q
->mq_map
[i
];
2069 /* unmapped hw queue can be remapped after CPU topo changed */
2070 if (!set
->tags
[hctx_idx
] &&
2071 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2073 * If tags initialization fail for some hctx,
2074 * that hctx won't be brought online. In this
2075 * case, remap the current ctx to hctx[0] which
2076 * is guaranteed to always have tags allocated
2081 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2082 hctx
= blk_mq_map_queue(q
, i
);
2084 cpumask_set_cpu(i
, hctx
->cpumask
);
2085 ctx
->index_hw
= hctx
->nr_ctx
;
2086 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2089 mutex_unlock(&q
->sysfs_lock
);
2091 queue_for_each_hw_ctx(q
, hctx
, i
) {
2093 * If no software queues are mapped to this hardware queue,
2094 * disable it and free the request entries.
2096 if (!hctx
->nr_ctx
) {
2097 /* Never unmap queue 0. We need it as a
2098 * fallback in case of a new remap fails
2101 if (i
&& set
->tags
[i
])
2102 blk_mq_free_map_and_requests(set
, i
);
2108 hctx
->tags
= set
->tags
[i
];
2109 WARN_ON(!hctx
->tags
);
2112 * Set the map size to the number of mapped software queues.
2113 * This is more accurate and more efficient than looping
2114 * over all possibly mapped software queues.
2116 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2119 * Initialize batch roundrobin counts
2121 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2122 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2126 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2128 struct blk_mq_hw_ctx
*hctx
;
2131 queue_for_each_hw_ctx(q
, hctx
, i
) {
2133 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2135 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2139 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
2141 struct request_queue
*q
;
2143 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2144 blk_mq_freeze_queue(q
);
2145 queue_set_hctx_shared(q
, shared
);
2146 blk_mq_unfreeze_queue(q
);
2150 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2152 struct blk_mq_tag_set
*set
= q
->tag_set
;
2154 mutex_lock(&set
->tag_list_lock
);
2155 list_del_init(&q
->tag_set_list
);
2156 if (list_is_singular(&set
->tag_list
)) {
2157 /* just transitioned to unshared */
2158 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2159 /* update existing queue */
2160 blk_mq_update_tag_set_depth(set
, false);
2162 mutex_unlock(&set
->tag_list_lock
);
2165 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2166 struct request_queue
*q
)
2170 mutex_lock(&set
->tag_list_lock
);
2172 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2173 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2174 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2175 /* update existing queue */
2176 blk_mq_update_tag_set_depth(set
, true);
2178 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2179 queue_set_hctx_shared(q
, true);
2180 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
2182 mutex_unlock(&set
->tag_list_lock
);
2186 * It is the actual release handler for mq, but we do it from
2187 * request queue's release handler for avoiding use-after-free
2188 * and headache because q->mq_kobj shouldn't have been introduced,
2189 * but we can't group ctx/kctx kobj without it.
2191 void blk_mq_release(struct request_queue
*q
)
2193 struct blk_mq_hw_ctx
*hctx
;
2196 blk_mq_sched_teardown(q
);
2198 /* hctx kobj stays in hctx */
2199 queue_for_each_hw_ctx(q
, hctx
, i
) {
2208 kfree(q
->queue_hw_ctx
);
2210 /* ctx kobj stays in queue_ctx */
2211 free_percpu(q
->queue_ctx
);
2214 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2216 struct request_queue
*uninit_q
, *q
;
2218 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2220 return ERR_PTR(-ENOMEM
);
2222 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2224 blk_cleanup_queue(uninit_q
);
2228 EXPORT_SYMBOL(blk_mq_init_queue
);
2230 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2231 struct request_queue
*q
)
2234 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2236 blk_mq_sysfs_unregister(q
);
2237 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2243 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2244 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2249 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2256 atomic_set(&hctxs
[i
]->nr_active
, 0);
2257 hctxs
[i
]->numa_node
= node
;
2258 hctxs
[i
]->queue_num
= i
;
2260 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2261 free_cpumask_var(hctxs
[i
]->cpumask
);
2266 blk_mq_hctx_kobj_init(hctxs
[i
]);
2268 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2269 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2273 blk_mq_free_map_and_requests(set
, j
);
2274 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2275 free_cpumask_var(hctx
->cpumask
);
2276 kobject_put(&hctx
->kobj
);
2283 q
->nr_hw_queues
= i
;
2284 blk_mq_sysfs_register(q
);
2287 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2288 struct request_queue
*q
)
2290 /* mark the queue as mq asap */
2291 q
->mq_ops
= set
->ops
;
2293 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2297 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2298 GFP_KERNEL
, set
->numa_node
);
2299 if (!q
->queue_hw_ctx
)
2302 q
->mq_map
= set
->mq_map
;
2304 blk_mq_realloc_hw_ctxs(set
, q
);
2305 if (!q
->nr_hw_queues
)
2308 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2309 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2311 q
->nr_queues
= nr_cpu_ids
;
2313 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2315 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2316 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2318 q
->sg_reserved_size
= INT_MAX
;
2320 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2321 INIT_LIST_HEAD(&q
->requeue_list
);
2322 spin_lock_init(&q
->requeue_lock
);
2324 if (q
->nr_hw_queues
> 1)
2325 blk_queue_make_request(q
, blk_mq_make_request
);
2327 blk_queue_make_request(q
, blk_sq_make_request
);
2330 * Do this after blk_queue_make_request() overrides it...
2332 q
->nr_requests
= set
->queue_depth
;
2335 * Default to classic polling
2339 if (set
->ops
->complete
)
2340 blk_queue_softirq_done(q
, set
->ops
->complete
);
2342 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2345 mutex_lock(&all_q_mutex
);
2347 list_add_tail(&q
->all_q_node
, &all_q_list
);
2348 blk_mq_add_queue_tag_set(set
, q
);
2349 blk_mq_map_swqueue(q
, cpu_online_mask
);
2351 mutex_unlock(&all_q_mutex
);
2354 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2357 ret
= blk_mq_sched_init(q
);
2359 return ERR_PTR(ret
);
2365 kfree(q
->queue_hw_ctx
);
2367 free_percpu(q
->queue_ctx
);
2370 return ERR_PTR(-ENOMEM
);
2372 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2374 void blk_mq_free_queue(struct request_queue
*q
)
2376 struct blk_mq_tag_set
*set
= q
->tag_set
;
2378 mutex_lock(&all_q_mutex
);
2379 list_del_init(&q
->all_q_node
);
2380 mutex_unlock(&all_q_mutex
);
2384 blk_mq_del_queue_tag_set(q
);
2386 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2387 blk_mq_free_hw_queues(q
, set
);
2390 /* Basically redo blk_mq_init_queue with queue frozen */
2391 static void blk_mq_queue_reinit(struct request_queue
*q
,
2392 const struct cpumask
*online_mask
)
2394 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2396 blk_mq_sysfs_unregister(q
);
2399 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2400 * we should change hctx numa_node according to new topology (this
2401 * involves free and re-allocate memory, worthy doing?)
2404 blk_mq_map_swqueue(q
, online_mask
);
2406 blk_mq_sysfs_register(q
);
2410 * New online cpumask which is going to be set in this hotplug event.
2411 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2412 * one-by-one and dynamically allocating this could result in a failure.
2414 static struct cpumask cpuhp_online_new
;
2416 static void blk_mq_queue_reinit_work(void)
2418 struct request_queue
*q
;
2420 mutex_lock(&all_q_mutex
);
2422 * We need to freeze and reinit all existing queues. Freezing
2423 * involves synchronous wait for an RCU grace period and doing it
2424 * one by one may take a long time. Start freezing all queues in
2425 * one swoop and then wait for the completions so that freezing can
2426 * take place in parallel.
2428 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2429 blk_mq_freeze_queue_start(q
);
2430 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2431 blk_mq_freeze_queue_wait(q
);
2433 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2434 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2436 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2437 blk_mq_unfreeze_queue(q
);
2439 mutex_unlock(&all_q_mutex
);
2442 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2444 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2445 blk_mq_queue_reinit_work();
2450 * Before hotadded cpu starts handling requests, new mappings must be
2451 * established. Otherwise, these requests in hw queue might never be
2454 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2455 * for CPU0, and ctx1 for CPU1).
2457 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2458 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2460 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2461 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2462 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2465 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2467 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2468 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2469 blk_mq_queue_reinit_work();
2473 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2477 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2478 if (!__blk_mq_alloc_rq_map(set
, i
))
2485 blk_mq_free_rq_map(set
->tags
[i
]);
2491 * Allocate the request maps associated with this tag_set. Note that this
2492 * may reduce the depth asked for, if memory is tight. set->queue_depth
2493 * will be updated to reflect the allocated depth.
2495 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2500 depth
= set
->queue_depth
;
2502 err
= __blk_mq_alloc_rq_maps(set
);
2506 set
->queue_depth
>>= 1;
2507 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2511 } while (set
->queue_depth
);
2513 if (!set
->queue_depth
|| err
) {
2514 pr_err("blk-mq: failed to allocate request map\n");
2518 if (depth
!= set
->queue_depth
)
2519 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2520 depth
, set
->queue_depth
);
2526 * Alloc a tag set to be associated with one or more request queues.
2527 * May fail with EINVAL for various error conditions. May adjust the
2528 * requested depth down, if if it too large. In that case, the set
2529 * value will be stored in set->queue_depth.
2531 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2535 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2537 if (!set
->nr_hw_queues
)
2539 if (!set
->queue_depth
)
2541 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2544 if (!set
->ops
->queue_rq
)
2547 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2548 pr_info("blk-mq: reduced tag depth to %u\n",
2550 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2554 * If a crashdump is active, then we are potentially in a very
2555 * memory constrained environment. Limit us to 1 queue and
2556 * 64 tags to prevent using too much memory.
2558 if (is_kdump_kernel()) {
2559 set
->nr_hw_queues
= 1;
2560 set
->queue_depth
= min(64U, set
->queue_depth
);
2563 * There is no use for more h/w queues than cpus.
2565 if (set
->nr_hw_queues
> nr_cpu_ids
)
2566 set
->nr_hw_queues
= nr_cpu_ids
;
2568 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2569 GFP_KERNEL
, set
->numa_node
);
2574 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2575 GFP_KERNEL
, set
->numa_node
);
2579 if (set
->ops
->map_queues
)
2580 ret
= set
->ops
->map_queues(set
);
2582 ret
= blk_mq_map_queues(set
);
2584 goto out_free_mq_map
;
2586 ret
= blk_mq_alloc_rq_maps(set
);
2588 goto out_free_mq_map
;
2590 mutex_init(&set
->tag_list_lock
);
2591 INIT_LIST_HEAD(&set
->tag_list
);
2603 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2605 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2609 for (i
= 0; i
< nr_cpu_ids
; i
++)
2610 blk_mq_free_map_and_requests(set
, i
);
2618 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2620 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2622 struct blk_mq_tag_set
*set
= q
->tag_set
;
2623 struct blk_mq_hw_ctx
*hctx
;
2629 blk_mq_freeze_queue(q
);
2630 blk_mq_quiesce_queue(q
);
2633 queue_for_each_hw_ctx(q
, hctx
, i
) {
2637 * If we're using an MQ scheduler, just update the scheduler
2638 * queue depth. This is similar to what the old code would do.
2640 if (!hctx
->sched_tags
) {
2641 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2642 min(nr
, set
->queue_depth
),
2645 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2653 q
->nr_requests
= nr
;
2655 blk_mq_unfreeze_queue(q
);
2656 blk_mq_start_stopped_hw_queues(q
, true);
2661 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2663 struct request_queue
*q
;
2665 if (nr_hw_queues
> nr_cpu_ids
)
2666 nr_hw_queues
= nr_cpu_ids
;
2667 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2670 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2671 blk_mq_freeze_queue(q
);
2673 set
->nr_hw_queues
= nr_hw_queues
;
2674 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2675 blk_mq_realloc_hw_ctxs(set
, q
);
2678 * Manually set the make_request_fn as blk_queue_make_request
2679 * resets a lot of the queue settings.
2681 if (q
->nr_hw_queues
> 1)
2682 q
->make_request_fn
= blk_mq_make_request
;
2684 q
->make_request_fn
= blk_sq_make_request
;
2686 blk_mq_queue_reinit(q
, cpu_online_mask
);
2689 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2690 blk_mq_unfreeze_queue(q
);
2692 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2694 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2695 struct blk_mq_hw_ctx
*hctx
,
2698 struct blk_rq_stat stat
[2];
2699 unsigned long ret
= 0;
2702 * If stats collection isn't on, don't sleep but turn it on for
2705 if (!blk_stat_enable(q
))
2709 * We don't have to do this once per IO, should optimize this
2710 * to just use the current window of stats until it changes
2712 memset(&stat
, 0, sizeof(stat
));
2713 blk_hctx_stat_get(hctx
, stat
);
2716 * As an optimistic guess, use half of the mean service time
2717 * for this type of request. We can (and should) make this smarter.
2718 * For instance, if the completion latencies are tight, we can
2719 * get closer than just half the mean. This is especially
2720 * important on devices where the completion latencies are longer
2723 if (req_op(rq
) == REQ_OP_READ
&& stat
[BLK_STAT_READ
].nr_samples
)
2724 ret
= (stat
[BLK_STAT_READ
].mean
+ 1) / 2;
2725 else if (req_op(rq
) == REQ_OP_WRITE
&& stat
[BLK_STAT_WRITE
].nr_samples
)
2726 ret
= (stat
[BLK_STAT_WRITE
].mean
+ 1) / 2;
2731 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2732 struct blk_mq_hw_ctx
*hctx
,
2735 struct hrtimer_sleeper hs
;
2736 enum hrtimer_mode mode
;
2740 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2746 * -1: don't ever hybrid sleep
2747 * 0: use half of prev avg
2748 * >0: use this specific value
2750 if (q
->poll_nsec
== -1)
2752 else if (q
->poll_nsec
> 0)
2753 nsecs
= q
->poll_nsec
;
2755 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2760 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2763 * This will be replaced with the stats tracking code, using
2764 * 'avg_completion_time / 2' as the pre-sleep target.
2768 mode
= HRTIMER_MODE_REL
;
2769 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2770 hrtimer_set_expires(&hs
.timer
, kt
);
2772 hrtimer_init_sleeper(&hs
, current
);
2774 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2776 set_current_state(TASK_UNINTERRUPTIBLE
);
2777 hrtimer_start_expires(&hs
.timer
, mode
);
2780 hrtimer_cancel(&hs
.timer
);
2781 mode
= HRTIMER_MODE_ABS
;
2782 } while (hs
.task
&& !signal_pending(current
));
2784 __set_current_state(TASK_RUNNING
);
2785 destroy_hrtimer_on_stack(&hs
.timer
);
2789 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2791 struct request_queue
*q
= hctx
->queue
;
2795 * If we sleep, have the caller restart the poll loop to reset
2796 * the state. Like for the other success return cases, the
2797 * caller is responsible for checking if the IO completed. If
2798 * the IO isn't complete, we'll get called again and will go
2799 * straight to the busy poll loop.
2801 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2804 hctx
->poll_considered
++;
2806 state
= current
->state
;
2807 while (!need_resched()) {
2810 hctx
->poll_invoked
++;
2812 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2814 hctx
->poll_success
++;
2815 set_current_state(TASK_RUNNING
);
2819 if (signal_pending_state(state
, current
))
2820 set_current_state(TASK_RUNNING
);
2822 if (current
->state
== TASK_RUNNING
)
2832 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2834 struct blk_mq_hw_ctx
*hctx
;
2835 struct blk_plug
*plug
;
2838 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2839 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2842 plug
= current
->plug
;
2844 blk_flush_plug_list(plug
, false);
2846 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2847 if (!blk_qc_t_is_internal(cookie
))
2848 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2850 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2852 return __blk_mq_poll(hctx
, rq
);
2854 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2856 void blk_mq_disable_hotplug(void)
2858 mutex_lock(&all_q_mutex
);
2861 void blk_mq_enable_hotplug(void)
2863 mutex_unlock(&all_q_mutex
);
2866 static int __init
blk_mq_init(void)
2868 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2869 blk_mq_hctx_notify_dead
);
2871 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2872 blk_mq_queue_reinit_prepare
,
2873 blk_mq_queue_reinit_dead
);
2876 subsys_initcall(blk_mq_init
);