2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/delay.h>
24 #include <linux/crash_dump.h>
25 #include <linux/prefetch.h>
27 #include <trace/events/block.h>
29 #include <linux/blk-mq.h>
32 #include "blk-mq-tag.h"
36 static DEFINE_MUTEX(all_q_mutex
);
37 static LIST_HEAD(all_q_list
);
40 * Check if any of the ctx's have pending work in this hardware queue
42 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
44 return sbitmap_any_bit_set(&hctx
->ctx_map
);
48 * Mark this ctx as having pending work in this hardware queue
50 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
51 struct blk_mq_ctx
*ctx
)
53 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
54 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
57 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
58 struct blk_mq_ctx
*ctx
)
60 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
63 void blk_mq_freeze_queue_start(struct request_queue
*q
)
67 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
68 if (freeze_depth
== 1) {
69 percpu_ref_kill(&q
->q_usage_counter
);
70 blk_mq_run_hw_queues(q
, false);
73 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start
);
75 static void blk_mq_freeze_queue_wait(struct request_queue
*q
)
77 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
81 * Guarantee no request is in use, so we can change any data structure of
82 * the queue afterward.
84 void blk_freeze_queue(struct request_queue
*q
)
87 * In the !blk_mq case we are only calling this to kill the
88 * q_usage_counter, otherwise this increases the freeze depth
89 * and waits for it to return to zero. For this reason there is
90 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
91 * exported to drivers as the only user for unfreeze is blk_mq.
93 blk_mq_freeze_queue_start(q
);
94 blk_mq_freeze_queue_wait(q
);
97 void blk_mq_freeze_queue(struct request_queue
*q
)
100 * ...just an alias to keep freeze and unfreeze actions balanced
101 * in the blk_mq_* namespace
105 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
107 void blk_mq_unfreeze_queue(struct request_queue
*q
)
111 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
112 WARN_ON_ONCE(freeze_depth
< 0);
114 percpu_ref_reinit(&q
->q_usage_counter
);
115 wake_up_all(&q
->mq_freeze_wq
);
118 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
121 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
124 * Note: this function does not prevent that the struct request end_io()
125 * callback function is invoked. Additionally, it is not prevented that
126 * new queue_rq() calls occur unless the queue has been stopped first.
128 void blk_mq_quiesce_queue(struct request_queue
*q
)
130 struct blk_mq_hw_ctx
*hctx
;
134 blk_mq_stop_hw_queues(q
);
136 queue_for_each_hw_ctx(q
, hctx
, i
) {
137 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
138 synchronize_srcu(&hctx
->queue_rq_srcu
);
145 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
147 void blk_mq_wake_waiters(struct request_queue
*q
)
149 struct blk_mq_hw_ctx
*hctx
;
152 queue_for_each_hw_ctx(q
, hctx
, i
)
153 if (blk_mq_hw_queue_mapped(hctx
))
154 blk_mq_tag_wakeup_all(hctx
->tags
, true);
157 * If we are called because the queue has now been marked as
158 * dying, we need to ensure that processes currently waiting on
159 * the queue are notified as well.
161 wake_up_all(&q
->mq_freeze_wq
);
164 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
166 return blk_mq_has_free_tags(hctx
->tags
);
168 EXPORT_SYMBOL(blk_mq_can_queue
);
170 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
171 struct request
*rq
, unsigned int op
)
173 INIT_LIST_HEAD(&rq
->queuelist
);
174 /* csd/requeue_work/fifo_time is initialized before use */
178 if (blk_queue_io_stat(q
))
179 rq
->rq_flags
|= RQF_IO_STAT
;
180 /* do not touch atomic flags, it needs atomic ops against the timer */
182 INIT_HLIST_NODE(&rq
->hash
);
183 RB_CLEAR_NODE(&rq
->rb_node
);
186 rq
->start_time
= jiffies
;
187 #ifdef CONFIG_BLK_CGROUP
189 set_start_time_ns(rq
);
190 rq
->io_start_time_ns
= 0;
192 rq
->nr_phys_segments
= 0;
193 #if defined(CONFIG_BLK_DEV_INTEGRITY)
194 rq
->nr_integrity_segments
= 0;
197 /* tag was already set */
207 INIT_LIST_HEAD(&rq
->timeout_list
);
211 rq
->end_io_data
= NULL
;
214 ctx
->rq_dispatched
[op_is_sync(op
)]++;
217 static struct request
*
218 __blk_mq_alloc_request(struct blk_mq_alloc_data
*data
, unsigned int op
)
223 tag
= blk_mq_get_tag(data
);
224 if (tag
!= BLK_MQ_TAG_FAIL
) {
225 rq
= data
->hctx
->tags
->rqs
[tag
];
227 if (blk_mq_tag_busy(data
->hctx
)) {
228 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
229 atomic_inc(&data
->hctx
->nr_active
);
233 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
);
240 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
243 struct blk_mq_ctx
*ctx
;
244 struct blk_mq_hw_ctx
*hctx
;
246 struct blk_mq_alloc_data alloc_data
;
249 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
253 ctx
= blk_mq_get_ctx(q
);
254 hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
255 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
256 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
261 return ERR_PTR(-EWOULDBLOCK
);
265 rq
->__sector
= (sector_t
) -1;
266 rq
->bio
= rq
->biotail
= NULL
;
269 EXPORT_SYMBOL(blk_mq_alloc_request
);
271 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
272 unsigned int flags
, unsigned int hctx_idx
)
274 struct blk_mq_hw_ctx
*hctx
;
275 struct blk_mq_ctx
*ctx
;
277 struct blk_mq_alloc_data alloc_data
;
281 * If the tag allocator sleeps we could get an allocation for a
282 * different hardware context. No need to complicate the low level
283 * allocator for this for the rare use case of a command tied to
286 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
287 return ERR_PTR(-EINVAL
);
289 if (hctx_idx
>= q
->nr_hw_queues
)
290 return ERR_PTR(-EIO
);
292 ret
= blk_queue_enter(q
, true);
297 * Check if the hardware context is actually mapped to anything.
298 * If not tell the caller that it should skip this queue.
300 hctx
= q
->queue_hw_ctx
[hctx_idx
];
301 if (!blk_mq_hw_queue_mapped(hctx
)) {
305 ctx
= __blk_mq_get_ctx(q
, cpumask_first(hctx
->cpumask
));
307 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
308 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
320 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
322 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
323 struct blk_mq_ctx
*ctx
, struct request
*rq
)
325 const int tag
= rq
->tag
;
326 struct request_queue
*q
= rq
->q
;
328 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
329 atomic_dec(&hctx
->nr_active
);
331 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
334 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
335 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
336 blk_mq_put_tag(hctx
, ctx
, tag
);
340 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
342 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
344 ctx
->rq_completed
[rq_is_sync(rq
)]++;
345 __blk_mq_free_request(hctx
, ctx
, rq
);
348 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request
);
350 void blk_mq_free_request(struct request
*rq
)
352 blk_mq_free_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
354 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
356 inline void __blk_mq_end_request(struct request
*rq
, int error
)
358 blk_account_io_done(rq
);
361 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
362 rq
->end_io(rq
, error
);
364 if (unlikely(blk_bidi_rq(rq
)))
365 blk_mq_free_request(rq
->next_rq
);
366 blk_mq_free_request(rq
);
369 EXPORT_SYMBOL(__blk_mq_end_request
);
371 void blk_mq_end_request(struct request
*rq
, int error
)
373 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
375 __blk_mq_end_request(rq
, error
);
377 EXPORT_SYMBOL(blk_mq_end_request
);
379 static void __blk_mq_complete_request_remote(void *data
)
381 struct request
*rq
= data
;
383 rq
->q
->softirq_done_fn(rq
);
386 static void blk_mq_ipi_complete_request(struct request
*rq
)
388 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
392 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
393 rq
->q
->softirq_done_fn(rq
);
398 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
399 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
401 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
402 rq
->csd
.func
= __blk_mq_complete_request_remote
;
405 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
407 rq
->q
->softirq_done_fn(rq
);
412 static void blk_mq_stat_add(struct request
*rq
)
414 if (rq
->rq_flags
& RQF_STATS
) {
416 * We could rq->mq_ctx here, but there's less of a risk
417 * of races if we have the completion event add the stats
418 * to the local software queue.
420 struct blk_mq_ctx
*ctx
;
422 ctx
= __blk_mq_get_ctx(rq
->q
, raw_smp_processor_id());
423 blk_stat_add(&ctx
->stat
[rq_data_dir(rq
)], rq
);
427 static void __blk_mq_complete_request(struct request
*rq
)
429 struct request_queue
*q
= rq
->q
;
433 if (!q
->softirq_done_fn
)
434 blk_mq_end_request(rq
, rq
->errors
);
436 blk_mq_ipi_complete_request(rq
);
440 * blk_mq_complete_request - end I/O on a request
441 * @rq: the request being processed
444 * Ends all I/O on a request. It does not handle partial completions.
445 * The actual completion happens out-of-order, through a IPI handler.
447 void blk_mq_complete_request(struct request
*rq
, int error
)
449 struct request_queue
*q
= rq
->q
;
451 if (unlikely(blk_should_fake_timeout(q
)))
453 if (!blk_mark_rq_complete(rq
)) {
455 __blk_mq_complete_request(rq
);
458 EXPORT_SYMBOL(blk_mq_complete_request
);
460 int blk_mq_request_started(struct request
*rq
)
462 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
464 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
466 void blk_mq_start_request(struct request
*rq
)
468 struct request_queue
*q
= rq
->q
;
470 trace_block_rq_issue(q
, rq
);
472 rq
->resid_len
= blk_rq_bytes(rq
);
473 if (unlikely(blk_bidi_rq(rq
)))
474 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
476 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
477 blk_stat_set_issue_time(&rq
->issue_stat
);
478 rq
->rq_flags
|= RQF_STATS
;
479 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
485 * Ensure that ->deadline is visible before set the started
486 * flag and clear the completed flag.
488 smp_mb__before_atomic();
491 * Mark us as started and clear complete. Complete might have been
492 * set if requeue raced with timeout, which then marked it as
493 * complete. So be sure to clear complete again when we start
494 * the request, otherwise we'll ignore the completion event.
496 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
497 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
498 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
499 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
501 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
503 * Make sure space for the drain appears. We know we can do
504 * this because max_hw_segments has been adjusted to be one
505 * fewer than the device can handle.
507 rq
->nr_phys_segments
++;
510 EXPORT_SYMBOL(blk_mq_start_request
);
512 static void __blk_mq_requeue_request(struct request
*rq
)
514 struct request_queue
*q
= rq
->q
;
516 trace_block_rq_requeue(q
, rq
);
517 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
519 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
520 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
521 rq
->nr_phys_segments
--;
525 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
527 __blk_mq_requeue_request(rq
);
529 BUG_ON(blk_queued_rq(rq
));
530 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
532 EXPORT_SYMBOL(blk_mq_requeue_request
);
534 static void blk_mq_requeue_work(struct work_struct
*work
)
536 struct request_queue
*q
=
537 container_of(work
, struct request_queue
, requeue_work
.work
);
539 struct request
*rq
, *next
;
542 spin_lock_irqsave(&q
->requeue_lock
, flags
);
543 list_splice_init(&q
->requeue_list
, &rq_list
);
544 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
546 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
547 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
550 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
551 list_del_init(&rq
->queuelist
);
552 blk_mq_insert_request(rq
, true, false, false);
555 while (!list_empty(&rq_list
)) {
556 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
557 list_del_init(&rq
->queuelist
);
558 blk_mq_insert_request(rq
, false, false, false);
561 blk_mq_run_hw_queues(q
, false);
564 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
565 bool kick_requeue_list
)
567 struct request_queue
*q
= rq
->q
;
571 * We abuse this flag that is otherwise used by the I/O scheduler to
572 * request head insertation from the workqueue.
574 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
576 spin_lock_irqsave(&q
->requeue_lock
, flags
);
578 rq
->rq_flags
|= RQF_SOFTBARRIER
;
579 list_add(&rq
->queuelist
, &q
->requeue_list
);
581 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
583 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
585 if (kick_requeue_list
)
586 blk_mq_kick_requeue_list(q
);
588 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
590 void blk_mq_kick_requeue_list(struct request_queue
*q
)
592 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
594 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
596 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
599 kblockd_schedule_delayed_work(&q
->requeue_work
,
600 msecs_to_jiffies(msecs
));
602 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
604 void blk_mq_abort_requeue_list(struct request_queue
*q
)
609 spin_lock_irqsave(&q
->requeue_lock
, flags
);
610 list_splice_init(&q
->requeue_list
, &rq_list
);
611 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
613 while (!list_empty(&rq_list
)) {
616 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
617 list_del_init(&rq
->queuelist
);
619 blk_mq_end_request(rq
, rq
->errors
);
622 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
624 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
626 if (tag
< tags
->nr_tags
) {
627 prefetch(tags
->rqs
[tag
]);
628 return tags
->rqs
[tag
];
633 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
635 struct blk_mq_timeout_data
{
637 unsigned int next_set
;
640 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
642 struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
643 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
646 * We know that complete is set at this point. If STARTED isn't set
647 * anymore, then the request isn't active and the "timeout" should
648 * just be ignored. This can happen due to the bitflag ordering.
649 * Timeout first checks if STARTED is set, and if it is, assumes
650 * the request is active. But if we race with completion, then
651 * we both flags will get cleared. So check here again, and ignore
652 * a timeout event with a request that isn't active.
654 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
658 ret
= ops
->timeout(req
, reserved
);
662 __blk_mq_complete_request(req
);
664 case BLK_EH_RESET_TIMER
:
666 blk_clear_rq_complete(req
);
668 case BLK_EH_NOT_HANDLED
:
671 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
676 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
677 struct request
*rq
, void *priv
, bool reserved
)
679 struct blk_mq_timeout_data
*data
= priv
;
681 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
683 * If a request wasn't started before the queue was
684 * marked dying, kill it here or it'll go unnoticed.
686 if (unlikely(blk_queue_dying(rq
->q
))) {
688 blk_mq_end_request(rq
, rq
->errors
);
693 if (time_after_eq(jiffies
, rq
->deadline
)) {
694 if (!blk_mark_rq_complete(rq
))
695 blk_mq_rq_timed_out(rq
, reserved
);
696 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
697 data
->next
= rq
->deadline
;
702 static void blk_mq_timeout_work(struct work_struct
*work
)
704 struct request_queue
*q
=
705 container_of(work
, struct request_queue
, timeout_work
);
706 struct blk_mq_timeout_data data
= {
712 /* A deadlock might occur if a request is stuck requiring a
713 * timeout at the same time a queue freeze is waiting
714 * completion, since the timeout code would not be able to
715 * acquire the queue reference here.
717 * That's why we don't use blk_queue_enter here; instead, we use
718 * percpu_ref_tryget directly, because we need to be able to
719 * obtain a reference even in the short window between the queue
720 * starting to freeze, by dropping the first reference in
721 * blk_mq_freeze_queue_start, and the moment the last request is
722 * consumed, marked by the instant q_usage_counter reaches
725 if (!percpu_ref_tryget(&q
->q_usage_counter
))
728 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
731 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
732 mod_timer(&q
->timeout
, data
.next
);
734 struct blk_mq_hw_ctx
*hctx
;
736 queue_for_each_hw_ctx(q
, hctx
, i
) {
737 /* the hctx may be unmapped, so check it here */
738 if (blk_mq_hw_queue_mapped(hctx
))
739 blk_mq_tag_idle(hctx
);
746 * Reverse check our software queue for entries that we could potentially
747 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
748 * too much time checking for merges.
750 static bool blk_mq_attempt_merge(struct request_queue
*q
,
751 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
756 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
762 if (!blk_rq_merge_ok(rq
, bio
))
765 el_ret
= blk_try_merge(rq
, bio
);
766 if (el_ret
== ELEVATOR_BACK_MERGE
) {
767 if (bio_attempt_back_merge(q
, rq
, bio
)) {
772 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
773 if (bio_attempt_front_merge(q
, rq
, bio
)) {
784 struct flush_busy_ctx_data
{
785 struct blk_mq_hw_ctx
*hctx
;
786 struct list_head
*list
;
789 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
791 struct flush_busy_ctx_data
*flush_data
= data
;
792 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
793 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
795 sbitmap_clear_bit(sb
, bitnr
);
796 spin_lock(&ctx
->lock
);
797 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
798 spin_unlock(&ctx
->lock
);
803 * Process software queues that have been marked busy, splicing them
804 * to the for-dispatch
806 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
808 struct flush_busy_ctx_data data
= {
813 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
816 static inline unsigned int queued_to_index(unsigned int queued
)
821 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
825 * Run this hardware queue, pulling any software queues mapped to it in.
826 * Note that this function currently has various problems around ordering
827 * of IO. In particular, we'd like FIFO behaviour on handling existing
828 * items on the hctx->dispatch list. Ignore that for now.
830 static void blk_mq_process_rq_list(struct blk_mq_hw_ctx
*hctx
)
832 struct request_queue
*q
= hctx
->queue
;
835 LIST_HEAD(driver_list
);
836 struct list_head
*dptr
;
839 if (unlikely(blk_mq_hctx_stopped(hctx
)))
845 * Touch any software queue that has pending entries.
847 flush_busy_ctxs(hctx
, &rq_list
);
850 * If we have previous entries on our dispatch list, grab them
851 * and stuff them at the front for more fair dispatch.
853 if (!list_empty_careful(&hctx
->dispatch
)) {
854 spin_lock(&hctx
->lock
);
855 if (!list_empty(&hctx
->dispatch
))
856 list_splice_init(&hctx
->dispatch
, &rq_list
);
857 spin_unlock(&hctx
->lock
);
861 * Start off with dptr being NULL, so we start the first request
862 * immediately, even if we have more pending.
867 * Now process all the entries, sending them to the driver.
870 while (!list_empty(&rq_list
)) {
871 struct blk_mq_queue_data bd
;
874 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
875 list_del_init(&rq
->queuelist
);
879 bd
.last
= list_empty(&rq_list
);
881 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
883 case BLK_MQ_RQ_QUEUE_OK
:
886 case BLK_MQ_RQ_QUEUE_BUSY
:
887 list_add(&rq
->queuelist
, &rq_list
);
888 __blk_mq_requeue_request(rq
);
891 pr_err("blk-mq: bad return on queue: %d\n", ret
);
892 case BLK_MQ_RQ_QUEUE_ERROR
:
894 blk_mq_end_request(rq
, rq
->errors
);
898 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
902 * We've done the first request. If we have more than 1
903 * left in the list, set dptr to defer issue.
905 if (!dptr
&& rq_list
.next
!= rq_list
.prev
)
909 hctx
->dispatched
[queued_to_index(queued
)]++;
912 * Any items that need requeuing? Stuff them into hctx->dispatch,
913 * that is where we will continue on next queue run.
915 if (!list_empty(&rq_list
)) {
916 spin_lock(&hctx
->lock
);
917 list_splice(&rq_list
, &hctx
->dispatch
);
918 spin_unlock(&hctx
->lock
);
920 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
921 * it's possible the queue is stopped and restarted again
922 * before this. Queue restart will dispatch requests. And since
923 * requests in rq_list aren't added into hctx->dispatch yet,
924 * the requests in rq_list might get lost.
926 * blk_mq_run_hw_queue() already checks the STOPPED bit
928 blk_mq_run_hw_queue(hctx
, true);
932 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
936 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
937 cpu_online(hctx
->next_cpu
));
939 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
941 blk_mq_process_rq_list(hctx
);
944 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
945 blk_mq_process_rq_list(hctx
);
946 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
951 * It'd be great if the workqueue API had a way to pass
952 * in a mask and had some smarts for more clever placement.
953 * For now we just round-robin here, switching for every
954 * BLK_MQ_CPU_WORK_BATCH queued items.
956 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
958 if (hctx
->queue
->nr_hw_queues
== 1)
959 return WORK_CPU_UNBOUND
;
961 if (--hctx
->next_cpu_batch
<= 0) {
964 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
965 if (next_cpu
>= nr_cpu_ids
)
966 next_cpu
= cpumask_first(hctx
->cpumask
);
968 hctx
->next_cpu
= next_cpu
;
969 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
972 return hctx
->next_cpu
;
975 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
977 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
978 !blk_mq_hw_queue_mapped(hctx
)))
981 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
983 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
984 __blk_mq_run_hw_queue(hctx
);
992 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
);
995 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
997 struct blk_mq_hw_ctx
*hctx
;
1000 queue_for_each_hw_ctx(q
, hctx
, i
) {
1001 if ((!blk_mq_hctx_has_pending(hctx
) &&
1002 list_empty_careful(&hctx
->dispatch
)) ||
1003 blk_mq_hctx_stopped(hctx
))
1006 blk_mq_run_hw_queue(hctx
, async
);
1009 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1012 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1013 * @q: request queue.
1015 * The caller is responsible for serializing this function against
1016 * blk_mq_{start,stop}_hw_queue().
1018 bool blk_mq_queue_stopped(struct request_queue
*q
)
1020 struct blk_mq_hw_ctx
*hctx
;
1023 queue_for_each_hw_ctx(q
, hctx
, i
)
1024 if (blk_mq_hctx_stopped(hctx
))
1029 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1031 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1033 cancel_work(&hctx
->run_work
);
1034 cancel_delayed_work(&hctx
->delay_work
);
1035 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1037 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1039 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1041 struct blk_mq_hw_ctx
*hctx
;
1044 queue_for_each_hw_ctx(q
, hctx
, i
)
1045 blk_mq_stop_hw_queue(hctx
);
1047 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1049 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1051 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1053 blk_mq_run_hw_queue(hctx
, false);
1055 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1057 void blk_mq_start_hw_queues(struct request_queue
*q
)
1059 struct blk_mq_hw_ctx
*hctx
;
1062 queue_for_each_hw_ctx(q
, hctx
, i
)
1063 blk_mq_start_hw_queue(hctx
);
1065 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1067 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1069 if (!blk_mq_hctx_stopped(hctx
))
1072 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1073 blk_mq_run_hw_queue(hctx
, async
);
1075 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1077 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1079 struct blk_mq_hw_ctx
*hctx
;
1082 queue_for_each_hw_ctx(q
, hctx
, i
)
1083 blk_mq_start_stopped_hw_queue(hctx
, async
);
1085 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1087 static void blk_mq_run_work_fn(struct work_struct
*work
)
1089 struct blk_mq_hw_ctx
*hctx
;
1091 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1093 __blk_mq_run_hw_queue(hctx
);
1096 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1098 struct blk_mq_hw_ctx
*hctx
;
1100 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1102 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1103 __blk_mq_run_hw_queue(hctx
);
1106 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1108 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1111 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1112 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1114 EXPORT_SYMBOL(blk_mq_delay_queue
);
1116 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1120 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1122 trace_block_rq_insert(hctx
->queue
, rq
);
1125 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1127 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1130 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
1131 struct request
*rq
, bool at_head
)
1133 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1135 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1136 blk_mq_hctx_mark_pending(hctx
, ctx
);
1139 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
1142 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1143 struct request_queue
*q
= rq
->q
;
1144 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1146 spin_lock(&ctx
->lock
);
1147 __blk_mq_insert_request(hctx
, rq
, at_head
);
1148 spin_unlock(&ctx
->lock
);
1151 blk_mq_run_hw_queue(hctx
, async
);
1154 static void blk_mq_insert_requests(struct request_queue
*q
,
1155 struct blk_mq_ctx
*ctx
,
1156 struct list_head
*list
,
1161 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1163 trace_block_unplug(q
, depth
, !from_schedule
);
1166 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1169 spin_lock(&ctx
->lock
);
1170 while (!list_empty(list
)) {
1173 rq
= list_first_entry(list
, struct request
, queuelist
);
1174 BUG_ON(rq
->mq_ctx
!= ctx
);
1175 list_del_init(&rq
->queuelist
);
1176 __blk_mq_insert_req_list(hctx
, rq
, false);
1178 blk_mq_hctx_mark_pending(hctx
, ctx
);
1179 spin_unlock(&ctx
->lock
);
1181 blk_mq_run_hw_queue(hctx
, from_schedule
);
1184 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1186 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1187 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1189 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1190 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1191 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1194 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1196 struct blk_mq_ctx
*this_ctx
;
1197 struct request_queue
*this_q
;
1200 LIST_HEAD(ctx_list
);
1203 list_splice_init(&plug
->mq_list
, &list
);
1205 list_sort(NULL
, &list
, plug_ctx_cmp
);
1211 while (!list_empty(&list
)) {
1212 rq
= list_entry_rq(list
.next
);
1213 list_del_init(&rq
->queuelist
);
1215 if (rq
->mq_ctx
!= this_ctx
) {
1217 blk_mq_insert_requests(this_q
, this_ctx
,
1222 this_ctx
= rq
->mq_ctx
;
1228 list_add_tail(&rq
->queuelist
, &ctx_list
);
1232 * If 'this_ctx' is set, we know we have entries to complete
1233 * on 'ctx_list'. Do those.
1236 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1241 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1243 init_request_from_bio(rq
, bio
);
1245 blk_account_io_start(rq
, true);
1248 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1250 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1251 !blk_queue_nomerges(hctx
->queue
);
1254 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1255 struct blk_mq_ctx
*ctx
,
1256 struct request
*rq
, struct bio
*bio
)
1258 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1259 blk_mq_bio_to_request(rq
, bio
);
1260 spin_lock(&ctx
->lock
);
1262 __blk_mq_insert_request(hctx
, rq
, false);
1263 spin_unlock(&ctx
->lock
);
1266 struct request_queue
*q
= hctx
->queue
;
1268 spin_lock(&ctx
->lock
);
1269 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1270 blk_mq_bio_to_request(rq
, bio
);
1274 spin_unlock(&ctx
->lock
);
1275 __blk_mq_free_request(hctx
, ctx
, rq
);
1280 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1282 struct blk_mq_alloc_data
*data
)
1284 struct blk_mq_hw_ctx
*hctx
;
1285 struct blk_mq_ctx
*ctx
;
1288 blk_queue_enter_live(q
);
1289 ctx
= blk_mq_get_ctx(q
);
1290 hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1292 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1293 blk_mq_set_alloc_data(data
, q
, 0, ctx
, hctx
);
1294 rq
= __blk_mq_alloc_request(data
, bio
->bi_opf
);
1296 data
->hctx
->queued
++;
1300 static void blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
)
1303 struct request_queue
*q
= rq
->q
;
1304 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, rq
->mq_ctx
->cpu
);
1305 struct blk_mq_queue_data bd
= {
1310 blk_qc_t new_cookie
= blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
);
1312 if (blk_mq_hctx_stopped(hctx
))
1316 * For OK queue, we are done. For error, kill it. Any other
1317 * error (busy), just add it to our list as we previously
1320 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1321 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1322 *cookie
= new_cookie
;
1326 __blk_mq_requeue_request(rq
);
1328 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1329 *cookie
= BLK_QC_T_NONE
;
1331 blk_mq_end_request(rq
, rq
->errors
);
1336 blk_mq_insert_request(rq
, false, true, true);
1340 * Multiple hardware queue variant. This will not use per-process plugs,
1341 * but will attempt to bypass the hctx queueing if we can go straight to
1342 * hardware for SYNC IO.
1344 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1346 const int is_sync
= op_is_sync(bio
->bi_opf
);
1347 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1348 struct blk_mq_alloc_data data
;
1350 unsigned int request_count
= 0, srcu_idx
;
1351 struct blk_plug
*plug
;
1352 struct request
*same_queue_rq
= NULL
;
1354 unsigned int wb_acct
;
1356 blk_queue_bounce(q
, &bio
);
1358 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1360 return BLK_QC_T_NONE
;
1363 blk_queue_split(q
, &bio
, q
->bio_split
);
1365 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1366 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1367 return BLK_QC_T_NONE
;
1369 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1371 rq
= blk_mq_map_request(q
, bio
, &data
);
1372 if (unlikely(!rq
)) {
1373 __wbt_done(q
->rq_wb
, wb_acct
);
1374 return BLK_QC_T_NONE
;
1377 wbt_track(&rq
->issue_stat
, wb_acct
);
1379 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1381 if (unlikely(is_flush_fua
)) {
1382 blk_mq_bio_to_request(rq
, bio
);
1383 blk_insert_flush(rq
);
1387 plug
= current
->plug
;
1389 * If the driver supports defer issued based on 'last', then
1390 * queue it up like normal since we can potentially save some
1393 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1394 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1395 struct request
*old_rq
= NULL
;
1397 blk_mq_bio_to_request(rq
, bio
);
1400 * We do limited plugging. If the bio can be merged, do that.
1401 * Otherwise the existing request in the plug list will be
1402 * issued. So the plug list will have one request at most
1406 * The plug list might get flushed before this. If that
1407 * happens, same_queue_rq is invalid and plug list is
1410 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1411 old_rq
= same_queue_rq
;
1412 list_del_init(&old_rq
->queuelist
);
1414 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1415 } else /* is_sync */
1417 blk_mq_put_ctx(data
.ctx
);
1421 if (!(data
.hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1423 blk_mq_try_issue_directly(old_rq
, &cookie
);
1426 srcu_idx
= srcu_read_lock(&data
.hctx
->queue_rq_srcu
);
1427 blk_mq_try_issue_directly(old_rq
, &cookie
);
1428 srcu_read_unlock(&data
.hctx
->queue_rq_srcu
, srcu_idx
);
1433 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1435 * For a SYNC request, send it to the hardware immediately. For
1436 * an ASYNC request, just ensure that we run it later on. The
1437 * latter allows for merging opportunities and more efficient
1441 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1443 blk_mq_put_ctx(data
.ctx
);
1449 * Single hardware queue variant. This will attempt to use any per-process
1450 * plug for merging and IO deferral.
1452 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1454 const int is_sync
= op_is_sync(bio
->bi_opf
);
1455 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1456 struct blk_plug
*plug
;
1457 unsigned int request_count
= 0;
1458 struct blk_mq_alloc_data data
;
1461 unsigned int wb_acct
;
1463 blk_queue_bounce(q
, &bio
);
1465 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1467 return BLK_QC_T_NONE
;
1470 blk_queue_split(q
, &bio
, q
->bio_split
);
1472 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1473 if (blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1474 return BLK_QC_T_NONE
;
1476 request_count
= blk_plug_queued_count(q
);
1478 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1480 rq
= blk_mq_map_request(q
, bio
, &data
);
1481 if (unlikely(!rq
)) {
1482 __wbt_done(q
->rq_wb
, wb_acct
);
1483 return BLK_QC_T_NONE
;
1486 wbt_track(&rq
->issue_stat
, wb_acct
);
1488 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1490 if (unlikely(is_flush_fua
)) {
1491 blk_mq_bio_to_request(rq
, bio
);
1492 blk_insert_flush(rq
);
1497 * A task plug currently exists. Since this is completely lockless,
1498 * utilize that to temporarily store requests until the task is
1499 * either done or scheduled away.
1501 plug
= current
->plug
;
1503 struct request
*last
= NULL
;
1505 blk_mq_bio_to_request(rq
, bio
);
1508 * @request_count may become stale because of schedule
1509 * out, so check the list again.
1511 if (list_empty(&plug
->mq_list
))
1514 trace_block_plug(q
);
1516 last
= list_entry_rq(plug
->mq_list
.prev
);
1518 blk_mq_put_ctx(data
.ctx
);
1520 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1521 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1522 blk_flush_plug_list(plug
, false);
1523 trace_block_plug(q
);
1526 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1530 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1532 * For a SYNC request, send it to the hardware immediately. For
1533 * an ASYNC request, just ensure that we run it later on. The
1534 * latter allows for merging opportunities and more efficient
1538 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1541 blk_mq_put_ctx(data
.ctx
);
1545 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1546 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1550 if (tags
->rqs
&& set
->ops
->exit_request
) {
1553 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1556 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1558 tags
->rqs
[i
] = NULL
;
1562 while (!list_empty(&tags
->page_list
)) {
1563 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1564 list_del_init(&page
->lru
);
1566 * Remove kmemleak object previously allocated in
1567 * blk_mq_init_rq_map().
1569 kmemleak_free(page_address(page
));
1570 __free_pages(page
, page
->private);
1575 blk_mq_free_tags(tags
);
1578 static size_t order_to_size(unsigned int order
)
1580 return (size_t)PAGE_SIZE
<< order
;
1583 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1584 unsigned int hctx_idx
)
1586 struct blk_mq_tags
*tags
;
1587 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1588 size_t rq_size
, left
;
1590 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1592 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1596 INIT_LIST_HEAD(&tags
->page_list
);
1598 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1599 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1602 blk_mq_free_tags(tags
);
1607 * rq_size is the size of the request plus driver payload, rounded
1608 * to the cacheline size
1610 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1612 left
= rq_size
* set
->queue_depth
;
1614 for (i
= 0; i
< set
->queue_depth
; ) {
1615 int this_order
= max_order
;
1620 while (this_order
&& left
< order_to_size(this_order
- 1))
1624 page
= alloc_pages_node(set
->numa_node
,
1625 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1631 if (order_to_size(this_order
) < rq_size
)
1638 page
->private = this_order
;
1639 list_add_tail(&page
->lru
, &tags
->page_list
);
1641 p
= page_address(page
);
1643 * Allow kmemleak to scan these pages as they contain pointers
1644 * to additional allocations like via ops->init_request().
1646 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_KERNEL
);
1647 entries_per_page
= order_to_size(this_order
) / rq_size
;
1648 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1649 left
-= to_do
* rq_size
;
1650 for (j
= 0; j
< to_do
; j
++) {
1652 if (set
->ops
->init_request
) {
1653 if (set
->ops
->init_request(set
->driver_data
,
1654 tags
->rqs
[i
], hctx_idx
, i
,
1656 tags
->rqs
[i
] = NULL
;
1668 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1673 * 'cpu' is going away. splice any existing rq_list entries from this
1674 * software queue to the hw queue dispatch list, and ensure that it
1677 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1679 struct blk_mq_hw_ctx
*hctx
;
1680 struct blk_mq_ctx
*ctx
;
1683 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1684 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1686 spin_lock(&ctx
->lock
);
1687 if (!list_empty(&ctx
->rq_list
)) {
1688 list_splice_init(&ctx
->rq_list
, &tmp
);
1689 blk_mq_hctx_clear_pending(hctx
, ctx
);
1691 spin_unlock(&ctx
->lock
);
1693 if (list_empty(&tmp
))
1696 spin_lock(&hctx
->lock
);
1697 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1698 spin_unlock(&hctx
->lock
);
1700 blk_mq_run_hw_queue(hctx
, true);
1704 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1706 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1710 /* hctx->ctxs will be freed in queue's release handler */
1711 static void blk_mq_exit_hctx(struct request_queue
*q
,
1712 struct blk_mq_tag_set
*set
,
1713 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1715 unsigned flush_start_tag
= set
->queue_depth
;
1717 blk_mq_tag_idle(hctx
);
1719 if (set
->ops
->exit_request
)
1720 set
->ops
->exit_request(set
->driver_data
,
1721 hctx
->fq
->flush_rq
, hctx_idx
,
1722 flush_start_tag
+ hctx_idx
);
1724 if (set
->ops
->exit_hctx
)
1725 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1727 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1728 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1730 blk_mq_remove_cpuhp(hctx
);
1731 blk_free_flush_queue(hctx
->fq
);
1732 sbitmap_free(&hctx
->ctx_map
);
1735 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1736 struct blk_mq_tag_set
*set
, int nr_queue
)
1738 struct blk_mq_hw_ctx
*hctx
;
1741 queue_for_each_hw_ctx(q
, hctx
, i
) {
1744 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1748 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1749 struct blk_mq_tag_set
*set
)
1751 struct blk_mq_hw_ctx
*hctx
;
1754 queue_for_each_hw_ctx(q
, hctx
, i
)
1755 free_cpumask_var(hctx
->cpumask
);
1758 static int blk_mq_init_hctx(struct request_queue
*q
,
1759 struct blk_mq_tag_set
*set
,
1760 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1763 unsigned flush_start_tag
= set
->queue_depth
;
1765 node
= hctx
->numa_node
;
1766 if (node
== NUMA_NO_NODE
)
1767 node
= hctx
->numa_node
= set
->numa_node
;
1769 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1770 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1771 spin_lock_init(&hctx
->lock
);
1772 INIT_LIST_HEAD(&hctx
->dispatch
);
1774 hctx
->queue_num
= hctx_idx
;
1775 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1777 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1779 hctx
->tags
= set
->tags
[hctx_idx
];
1782 * Allocate space for all possible cpus to avoid allocation at
1785 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1788 goto unregister_cpu_notifier
;
1790 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1796 if (set
->ops
->init_hctx
&&
1797 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1800 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1804 if (set
->ops
->init_request
&&
1805 set
->ops
->init_request(set
->driver_data
,
1806 hctx
->fq
->flush_rq
, hctx_idx
,
1807 flush_start_tag
+ hctx_idx
, node
))
1810 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1811 init_srcu_struct(&hctx
->queue_rq_srcu
);
1818 if (set
->ops
->exit_hctx
)
1819 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1821 sbitmap_free(&hctx
->ctx_map
);
1824 unregister_cpu_notifier
:
1825 blk_mq_remove_cpuhp(hctx
);
1829 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1830 unsigned int nr_hw_queues
)
1834 for_each_possible_cpu(i
) {
1835 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1836 struct blk_mq_hw_ctx
*hctx
;
1838 memset(__ctx
, 0, sizeof(*__ctx
));
1840 spin_lock_init(&__ctx
->lock
);
1841 INIT_LIST_HEAD(&__ctx
->rq_list
);
1843 blk_stat_init(&__ctx
->stat
[BLK_STAT_READ
]);
1844 blk_stat_init(&__ctx
->stat
[BLK_STAT_WRITE
]);
1846 /* If the cpu isn't online, the cpu is mapped to first hctx */
1850 hctx
= blk_mq_map_queue(q
, i
);
1853 * Set local node, IFF we have more than one hw queue. If
1854 * not, we remain on the home node of the device
1856 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1857 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1861 static void blk_mq_map_swqueue(struct request_queue
*q
,
1862 const struct cpumask
*online_mask
)
1865 struct blk_mq_hw_ctx
*hctx
;
1866 struct blk_mq_ctx
*ctx
;
1867 struct blk_mq_tag_set
*set
= q
->tag_set
;
1870 * Avoid others reading imcomplete hctx->cpumask through sysfs
1872 mutex_lock(&q
->sysfs_lock
);
1874 queue_for_each_hw_ctx(q
, hctx
, i
) {
1875 cpumask_clear(hctx
->cpumask
);
1880 * Map software to hardware queues
1882 for_each_possible_cpu(i
) {
1883 /* If the cpu isn't online, the cpu is mapped to first hctx */
1884 if (!cpumask_test_cpu(i
, online_mask
))
1887 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1888 hctx
= blk_mq_map_queue(q
, i
);
1890 cpumask_set_cpu(i
, hctx
->cpumask
);
1891 ctx
->index_hw
= hctx
->nr_ctx
;
1892 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1895 mutex_unlock(&q
->sysfs_lock
);
1897 queue_for_each_hw_ctx(q
, hctx
, i
) {
1899 * If no software queues are mapped to this hardware queue,
1900 * disable it and free the request entries.
1902 if (!hctx
->nr_ctx
) {
1904 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1905 set
->tags
[i
] = NULL
;
1911 /* unmapped hw queue can be remapped after CPU topo changed */
1913 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1914 hctx
->tags
= set
->tags
[i
];
1915 WARN_ON(!hctx
->tags
);
1918 * Set the map size to the number of mapped software queues.
1919 * This is more accurate and more efficient than looping
1920 * over all possibly mapped software queues.
1922 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
1925 * Initialize batch roundrobin counts
1927 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1928 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1932 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
1934 struct blk_mq_hw_ctx
*hctx
;
1937 queue_for_each_hw_ctx(q
, hctx
, i
) {
1939 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1941 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1945 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
1947 struct request_queue
*q
;
1949 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1950 blk_mq_freeze_queue(q
);
1951 queue_set_hctx_shared(q
, shared
);
1952 blk_mq_unfreeze_queue(q
);
1956 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1958 struct blk_mq_tag_set
*set
= q
->tag_set
;
1960 mutex_lock(&set
->tag_list_lock
);
1961 list_del_init(&q
->tag_set_list
);
1962 if (list_is_singular(&set
->tag_list
)) {
1963 /* just transitioned to unshared */
1964 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1965 /* update existing queue */
1966 blk_mq_update_tag_set_depth(set
, false);
1968 mutex_unlock(&set
->tag_list_lock
);
1971 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1972 struct request_queue
*q
)
1976 mutex_lock(&set
->tag_list_lock
);
1978 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1979 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1980 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
1981 /* update existing queue */
1982 blk_mq_update_tag_set_depth(set
, true);
1984 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
1985 queue_set_hctx_shared(q
, true);
1986 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1988 mutex_unlock(&set
->tag_list_lock
);
1992 * It is the actual release handler for mq, but we do it from
1993 * request queue's release handler for avoiding use-after-free
1994 * and headache because q->mq_kobj shouldn't have been introduced,
1995 * but we can't group ctx/kctx kobj without it.
1997 void blk_mq_release(struct request_queue
*q
)
1999 struct blk_mq_hw_ctx
*hctx
;
2002 /* hctx kobj stays in hctx */
2003 queue_for_each_hw_ctx(q
, hctx
, i
) {
2012 kfree(q
->queue_hw_ctx
);
2014 /* ctx kobj stays in queue_ctx */
2015 free_percpu(q
->queue_ctx
);
2018 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2020 struct request_queue
*uninit_q
, *q
;
2022 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2024 return ERR_PTR(-ENOMEM
);
2026 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2028 blk_cleanup_queue(uninit_q
);
2032 EXPORT_SYMBOL(blk_mq_init_queue
);
2034 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2035 struct request_queue
*q
)
2038 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2040 blk_mq_sysfs_unregister(q
);
2041 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2047 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2048 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2053 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2060 atomic_set(&hctxs
[i
]->nr_active
, 0);
2061 hctxs
[i
]->numa_node
= node
;
2062 hctxs
[i
]->queue_num
= i
;
2064 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2065 free_cpumask_var(hctxs
[i
]->cpumask
);
2070 blk_mq_hctx_kobj_init(hctxs
[i
]);
2072 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2073 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2077 blk_mq_free_rq_map(set
, hctx
->tags
, j
);
2078 set
->tags
[j
] = NULL
;
2080 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2081 free_cpumask_var(hctx
->cpumask
);
2082 kobject_put(&hctx
->kobj
);
2089 q
->nr_hw_queues
= i
;
2090 blk_mq_sysfs_register(q
);
2093 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2094 struct request_queue
*q
)
2096 /* mark the queue as mq asap */
2097 q
->mq_ops
= set
->ops
;
2099 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2103 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2104 GFP_KERNEL
, set
->numa_node
);
2105 if (!q
->queue_hw_ctx
)
2108 q
->mq_map
= set
->mq_map
;
2110 blk_mq_realloc_hw_ctxs(set
, q
);
2111 if (!q
->nr_hw_queues
)
2114 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2115 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2117 q
->nr_queues
= nr_cpu_ids
;
2119 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2121 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2122 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2124 q
->sg_reserved_size
= INT_MAX
;
2126 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2127 INIT_LIST_HEAD(&q
->requeue_list
);
2128 spin_lock_init(&q
->requeue_lock
);
2130 if (q
->nr_hw_queues
> 1)
2131 blk_queue_make_request(q
, blk_mq_make_request
);
2133 blk_queue_make_request(q
, blk_sq_make_request
);
2136 * Do this after blk_queue_make_request() overrides it...
2138 q
->nr_requests
= set
->queue_depth
;
2141 * Default to classic polling
2145 if (set
->ops
->complete
)
2146 blk_queue_softirq_done(q
, set
->ops
->complete
);
2148 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2151 mutex_lock(&all_q_mutex
);
2153 list_add_tail(&q
->all_q_node
, &all_q_list
);
2154 blk_mq_add_queue_tag_set(set
, q
);
2155 blk_mq_map_swqueue(q
, cpu_online_mask
);
2157 mutex_unlock(&all_q_mutex
);
2163 kfree(q
->queue_hw_ctx
);
2165 free_percpu(q
->queue_ctx
);
2168 return ERR_PTR(-ENOMEM
);
2170 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2172 void blk_mq_free_queue(struct request_queue
*q
)
2174 struct blk_mq_tag_set
*set
= q
->tag_set
;
2176 mutex_lock(&all_q_mutex
);
2177 list_del_init(&q
->all_q_node
);
2178 mutex_unlock(&all_q_mutex
);
2182 blk_mq_del_queue_tag_set(q
);
2184 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2185 blk_mq_free_hw_queues(q
, set
);
2188 /* Basically redo blk_mq_init_queue with queue frozen */
2189 static void blk_mq_queue_reinit(struct request_queue
*q
,
2190 const struct cpumask
*online_mask
)
2192 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2194 blk_mq_sysfs_unregister(q
);
2197 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2198 * we should change hctx numa_node according to new topology (this
2199 * involves free and re-allocate memory, worthy doing?)
2202 blk_mq_map_swqueue(q
, online_mask
);
2204 blk_mq_sysfs_register(q
);
2208 * New online cpumask which is going to be set in this hotplug event.
2209 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2210 * one-by-one and dynamically allocating this could result in a failure.
2212 static struct cpumask cpuhp_online_new
;
2214 static void blk_mq_queue_reinit_work(void)
2216 struct request_queue
*q
;
2218 mutex_lock(&all_q_mutex
);
2220 * We need to freeze and reinit all existing queues. Freezing
2221 * involves synchronous wait for an RCU grace period and doing it
2222 * one by one may take a long time. Start freezing all queues in
2223 * one swoop and then wait for the completions so that freezing can
2224 * take place in parallel.
2226 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2227 blk_mq_freeze_queue_start(q
);
2228 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2229 blk_mq_freeze_queue_wait(q
);
2231 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2232 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2234 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2235 blk_mq_unfreeze_queue(q
);
2237 mutex_unlock(&all_q_mutex
);
2240 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2242 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2243 blk_mq_queue_reinit_work();
2248 * Before hotadded cpu starts handling requests, new mappings must be
2249 * established. Otherwise, these requests in hw queue might never be
2252 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2253 * for CPU0, and ctx1 for CPU1).
2255 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2256 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2258 * And then while running hw queue, flush_busy_ctxs() finds bit0 is set in
2259 * pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2260 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list
2263 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2265 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2266 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2267 blk_mq_queue_reinit_work();
2271 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2275 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2276 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2285 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2291 * Allocate the request maps associated with this tag_set. Note that this
2292 * may reduce the depth asked for, if memory is tight. set->queue_depth
2293 * will be updated to reflect the allocated depth.
2295 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2300 depth
= set
->queue_depth
;
2302 err
= __blk_mq_alloc_rq_maps(set
);
2306 set
->queue_depth
>>= 1;
2307 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2311 } while (set
->queue_depth
);
2313 if (!set
->queue_depth
|| err
) {
2314 pr_err("blk-mq: failed to allocate request map\n");
2318 if (depth
!= set
->queue_depth
)
2319 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2320 depth
, set
->queue_depth
);
2326 * Alloc a tag set to be associated with one or more request queues.
2327 * May fail with EINVAL for various error conditions. May adjust the
2328 * requested depth down, if if it too large. In that case, the set
2329 * value will be stored in set->queue_depth.
2331 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2335 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2337 if (!set
->nr_hw_queues
)
2339 if (!set
->queue_depth
)
2341 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2344 if (!set
->ops
->queue_rq
)
2347 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2348 pr_info("blk-mq: reduced tag depth to %u\n",
2350 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2354 * If a crashdump is active, then we are potentially in a very
2355 * memory constrained environment. Limit us to 1 queue and
2356 * 64 tags to prevent using too much memory.
2358 if (is_kdump_kernel()) {
2359 set
->nr_hw_queues
= 1;
2360 set
->queue_depth
= min(64U, set
->queue_depth
);
2363 * There is no use for more h/w queues than cpus.
2365 if (set
->nr_hw_queues
> nr_cpu_ids
)
2366 set
->nr_hw_queues
= nr_cpu_ids
;
2368 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2369 GFP_KERNEL
, set
->numa_node
);
2374 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2375 GFP_KERNEL
, set
->numa_node
);
2379 if (set
->ops
->map_queues
)
2380 ret
= set
->ops
->map_queues(set
);
2382 ret
= blk_mq_map_queues(set
);
2384 goto out_free_mq_map
;
2386 ret
= blk_mq_alloc_rq_maps(set
);
2388 goto out_free_mq_map
;
2390 mutex_init(&set
->tag_list_lock
);
2391 INIT_LIST_HEAD(&set
->tag_list
);
2403 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2405 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2409 for (i
= 0; i
< nr_cpu_ids
; i
++) {
2411 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2420 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2422 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2424 struct blk_mq_tag_set
*set
= q
->tag_set
;
2425 struct blk_mq_hw_ctx
*hctx
;
2428 if (!set
|| nr
> set
->queue_depth
)
2432 queue_for_each_hw_ctx(q
, hctx
, i
) {
2435 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2441 q
->nr_requests
= nr
;
2446 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2448 struct request_queue
*q
;
2450 if (nr_hw_queues
> nr_cpu_ids
)
2451 nr_hw_queues
= nr_cpu_ids
;
2452 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2455 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2456 blk_mq_freeze_queue(q
);
2458 set
->nr_hw_queues
= nr_hw_queues
;
2459 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2460 blk_mq_realloc_hw_ctxs(set
, q
);
2462 if (q
->nr_hw_queues
> 1)
2463 blk_queue_make_request(q
, blk_mq_make_request
);
2465 blk_queue_make_request(q
, blk_sq_make_request
);
2467 blk_mq_queue_reinit(q
, cpu_online_mask
);
2470 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2471 blk_mq_unfreeze_queue(q
);
2473 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2475 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2476 struct blk_mq_hw_ctx
*hctx
,
2479 struct blk_rq_stat stat
[2];
2480 unsigned long ret
= 0;
2483 * If stats collection isn't on, don't sleep but turn it on for
2486 if (!blk_stat_enable(q
))
2490 * We don't have to do this once per IO, should optimize this
2491 * to just use the current window of stats until it changes
2493 memset(&stat
, 0, sizeof(stat
));
2494 blk_hctx_stat_get(hctx
, stat
);
2497 * As an optimistic guess, use half of the mean service time
2498 * for this type of request. We can (and should) make this smarter.
2499 * For instance, if the completion latencies are tight, we can
2500 * get closer than just half the mean. This is especially
2501 * important on devices where the completion latencies are longer
2504 if (req_op(rq
) == REQ_OP_READ
&& stat
[BLK_STAT_READ
].nr_samples
)
2505 ret
= (stat
[BLK_STAT_READ
].mean
+ 1) / 2;
2506 else if (req_op(rq
) == REQ_OP_WRITE
&& stat
[BLK_STAT_WRITE
].nr_samples
)
2507 ret
= (stat
[BLK_STAT_WRITE
].mean
+ 1) / 2;
2512 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2513 struct blk_mq_hw_ctx
*hctx
,
2516 struct hrtimer_sleeper hs
;
2517 enum hrtimer_mode mode
;
2521 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2527 * -1: don't ever hybrid sleep
2528 * 0: use half of prev avg
2529 * >0: use this specific value
2531 if (q
->poll_nsec
== -1)
2533 else if (q
->poll_nsec
> 0)
2534 nsecs
= q
->poll_nsec
;
2536 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2541 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2544 * This will be replaced with the stats tracking code, using
2545 * 'avg_completion_time / 2' as the pre-sleep target.
2547 kt
= ktime_set(0, nsecs
);
2549 mode
= HRTIMER_MODE_REL
;
2550 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2551 hrtimer_set_expires(&hs
.timer
, kt
);
2553 hrtimer_init_sleeper(&hs
, current
);
2555 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2557 set_current_state(TASK_UNINTERRUPTIBLE
);
2558 hrtimer_start_expires(&hs
.timer
, mode
);
2561 hrtimer_cancel(&hs
.timer
);
2562 mode
= HRTIMER_MODE_ABS
;
2563 } while (hs
.task
&& !signal_pending(current
));
2565 __set_current_state(TASK_RUNNING
);
2566 destroy_hrtimer_on_stack(&hs
.timer
);
2570 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2572 struct request_queue
*q
= hctx
->queue
;
2576 * If we sleep, have the caller restart the poll loop to reset
2577 * the state. Like for the other success return cases, the
2578 * caller is responsible for checking if the IO completed. If
2579 * the IO isn't complete, we'll get called again and will go
2580 * straight to the busy poll loop.
2582 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2585 hctx
->poll_considered
++;
2587 state
= current
->state
;
2588 while (!need_resched()) {
2591 hctx
->poll_invoked
++;
2593 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2595 hctx
->poll_success
++;
2596 set_current_state(TASK_RUNNING
);
2600 if (signal_pending_state(state
, current
))
2601 set_current_state(TASK_RUNNING
);
2603 if (current
->state
== TASK_RUNNING
)
2613 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2615 struct blk_mq_hw_ctx
*hctx
;
2616 struct blk_plug
*plug
;
2619 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2620 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2623 plug
= current
->plug
;
2625 blk_flush_plug_list(plug
, false);
2627 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2628 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2630 return __blk_mq_poll(hctx
, rq
);
2632 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2634 void blk_mq_disable_hotplug(void)
2636 mutex_lock(&all_q_mutex
);
2639 void blk_mq_enable_hotplug(void)
2641 mutex_unlock(&all_q_mutex
);
2644 static int __init
blk_mq_init(void)
2646 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2647 blk_mq_hctx_notify_dead
);
2649 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2650 blk_mq_queue_reinit_prepare
,
2651 blk_mq_queue_reinit_dead
);
2654 subsys_initcall(blk_mq_init
);