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 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
));
82 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
84 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
85 unsigned long timeout
)
87 return wait_event_timeout(q
->mq_freeze_wq
,
88 percpu_ref_is_zero(&q
->q_usage_counter
),
91 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
94 * Guarantee no request is in use, so we can change any data structure of
95 * the queue afterward.
97 void blk_freeze_queue(struct request_queue
*q
)
100 * In the !blk_mq case we are only calling this to kill the
101 * q_usage_counter, otherwise this increases the freeze depth
102 * and waits for it to return to zero. For this reason there is
103 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
104 * exported to drivers as the only user for unfreeze is blk_mq.
106 blk_mq_freeze_queue_start(q
);
107 blk_mq_freeze_queue_wait(q
);
110 void blk_mq_freeze_queue(struct request_queue
*q
)
113 * ...just an alias to keep freeze and unfreeze actions balanced
114 * in the blk_mq_* namespace
118 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
120 void blk_mq_unfreeze_queue(struct request_queue
*q
)
124 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
125 WARN_ON_ONCE(freeze_depth
< 0);
127 percpu_ref_reinit(&q
->q_usage_counter
);
128 wake_up_all(&q
->mq_freeze_wq
);
131 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
134 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
137 * Note: this function does not prevent that the struct request end_io()
138 * callback function is invoked. Additionally, it is not prevented that
139 * new queue_rq() calls occur unless the queue has been stopped first.
141 void blk_mq_quiesce_queue(struct request_queue
*q
)
143 struct blk_mq_hw_ctx
*hctx
;
147 blk_mq_stop_hw_queues(q
);
149 queue_for_each_hw_ctx(q
, hctx
, i
) {
150 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
151 synchronize_srcu(&hctx
->queue_rq_srcu
);
158 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
160 void blk_mq_wake_waiters(struct request_queue
*q
)
162 struct blk_mq_hw_ctx
*hctx
;
165 queue_for_each_hw_ctx(q
, hctx
, i
)
166 if (blk_mq_hw_queue_mapped(hctx
))
167 blk_mq_tag_wakeup_all(hctx
->tags
, true);
170 * If we are called because the queue has now been marked as
171 * dying, we need to ensure that processes currently waiting on
172 * the queue are notified as well.
174 wake_up_all(&q
->mq_freeze_wq
);
177 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
179 return blk_mq_has_free_tags(hctx
->tags
);
181 EXPORT_SYMBOL(blk_mq_can_queue
);
183 void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
184 struct request
*rq
, unsigned int op
)
186 INIT_LIST_HEAD(&rq
->queuelist
);
187 /* csd/requeue_work/fifo_time is initialized before use */
191 if (blk_queue_io_stat(q
))
192 rq
->rq_flags
|= RQF_IO_STAT
;
193 /* do not touch atomic flags, it needs atomic ops against the timer */
195 INIT_HLIST_NODE(&rq
->hash
);
196 RB_CLEAR_NODE(&rq
->rb_node
);
199 rq
->start_time
= jiffies
;
200 #ifdef CONFIG_BLK_CGROUP
202 set_start_time_ns(rq
);
203 rq
->io_start_time_ns
= 0;
205 rq
->nr_phys_segments
= 0;
206 #if defined(CONFIG_BLK_DEV_INTEGRITY)
207 rq
->nr_integrity_segments
= 0;
210 /* tag was already set */
214 INIT_LIST_HEAD(&rq
->timeout_list
);
218 rq
->end_io_data
= NULL
;
221 ctx
->rq_dispatched
[op_is_sync(op
)]++;
223 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init
);
225 struct request
*__blk_mq_alloc_request(struct blk_mq_alloc_data
*data
,
231 tag
= blk_mq_get_tag(data
);
232 if (tag
!= BLK_MQ_TAG_FAIL
) {
233 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
235 rq
= tags
->static_rqs
[tag
];
237 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
239 rq
->internal_tag
= tag
;
241 if (blk_mq_tag_busy(data
->hctx
)) {
242 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
243 atomic_inc(&data
->hctx
->nr_active
);
246 rq
->internal_tag
= -1;
247 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
250 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
);
256 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request
);
258 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
261 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
265 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
269 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
271 blk_mq_put_ctx(alloc_data
.ctx
);
275 return ERR_PTR(-EWOULDBLOCK
);
278 rq
->__sector
= (sector_t
) -1;
279 rq
->bio
= rq
->biotail
= NULL
;
282 EXPORT_SYMBOL(blk_mq_alloc_request
);
284 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
285 unsigned int flags
, unsigned int hctx_idx
)
287 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
293 * If the tag allocator sleeps we could get an allocation for a
294 * different hardware context. No need to complicate the low level
295 * allocator for this for the rare use case of a command tied to
298 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
299 return ERR_PTR(-EINVAL
);
301 if (hctx_idx
>= q
->nr_hw_queues
)
302 return ERR_PTR(-EIO
);
304 ret
= blk_queue_enter(q
, true);
309 * Check if the hardware context is actually mapped to anything.
310 * If not tell the caller that it should skip this queue.
312 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
313 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
315 return ERR_PTR(-EXDEV
);
317 cpu
= cpumask_first(alloc_data
.hctx
->cpumask
);
318 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
320 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
322 blk_mq_put_ctx(alloc_data
.ctx
);
326 return ERR_PTR(-EWOULDBLOCK
);
330 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
332 void __blk_mq_finish_request(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
335 const int sched_tag
= rq
->internal_tag
;
336 struct request_queue
*q
= rq
->q
;
338 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
339 atomic_dec(&hctx
->nr_active
);
341 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
344 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
345 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
347 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
349 blk_mq_sched_completed_request(hctx
, rq
);
350 blk_mq_sched_restart_queues(hctx
);
354 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx
*hctx
,
357 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
359 ctx
->rq_completed
[rq_is_sync(rq
)]++;
360 __blk_mq_finish_request(hctx
, ctx
, rq
);
363 void blk_mq_finish_request(struct request
*rq
)
365 blk_mq_finish_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
368 void blk_mq_free_request(struct request
*rq
)
370 blk_mq_sched_put_request(rq
);
372 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
374 inline void __blk_mq_end_request(struct request
*rq
, int error
)
376 blk_account_io_done(rq
);
379 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
380 rq
->end_io(rq
, error
);
382 if (unlikely(blk_bidi_rq(rq
)))
383 blk_mq_free_request(rq
->next_rq
);
384 blk_mq_free_request(rq
);
387 EXPORT_SYMBOL(__blk_mq_end_request
);
389 void blk_mq_end_request(struct request
*rq
, int error
)
391 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
393 __blk_mq_end_request(rq
, error
);
395 EXPORT_SYMBOL(blk_mq_end_request
);
397 static void __blk_mq_complete_request_remote(void *data
)
399 struct request
*rq
= data
;
401 rq
->q
->softirq_done_fn(rq
);
404 static void blk_mq_ipi_complete_request(struct request
*rq
)
406 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
410 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
411 rq
->q
->softirq_done_fn(rq
);
416 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
417 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
419 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
420 rq
->csd
.func
= __blk_mq_complete_request_remote
;
423 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
425 rq
->q
->softirq_done_fn(rq
);
430 static void blk_mq_stat_add(struct request
*rq
)
432 if (rq
->rq_flags
& RQF_STATS
) {
434 * We could rq->mq_ctx here, but there's less of a risk
435 * of races if we have the completion event add the stats
436 * to the local software queue.
438 struct blk_mq_ctx
*ctx
;
440 ctx
= __blk_mq_get_ctx(rq
->q
, raw_smp_processor_id());
441 blk_stat_add(&ctx
->stat
[rq_data_dir(rq
)], rq
);
445 static void __blk_mq_complete_request(struct request
*rq
)
447 struct request_queue
*q
= rq
->q
;
451 if (!q
->softirq_done_fn
)
452 blk_mq_end_request(rq
, rq
->errors
);
454 blk_mq_ipi_complete_request(rq
);
458 * blk_mq_complete_request - end I/O on a request
459 * @rq: the request being processed
462 * Ends all I/O on a request. It does not handle partial completions.
463 * The actual completion happens out-of-order, through a IPI handler.
465 void blk_mq_complete_request(struct request
*rq
, int error
)
467 struct request_queue
*q
= rq
->q
;
469 if (unlikely(blk_should_fake_timeout(q
)))
471 if (!blk_mark_rq_complete(rq
)) {
473 __blk_mq_complete_request(rq
);
476 EXPORT_SYMBOL(blk_mq_complete_request
);
478 int blk_mq_request_started(struct request
*rq
)
480 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
482 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
484 void blk_mq_start_request(struct request
*rq
)
486 struct request_queue
*q
= rq
->q
;
488 blk_mq_sched_started_request(rq
);
490 trace_block_rq_issue(q
, rq
);
492 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
493 blk_stat_set_issue_time(&rq
->issue_stat
);
494 rq
->rq_flags
|= RQF_STATS
;
495 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
501 * Ensure that ->deadline is visible before set the started
502 * flag and clear the completed flag.
504 smp_mb__before_atomic();
507 * Mark us as started and clear complete. Complete might have been
508 * set if requeue raced with timeout, which then marked it as
509 * complete. So be sure to clear complete again when we start
510 * the request, otherwise we'll ignore the completion event.
512 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
513 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
514 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
515 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
517 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
519 * Make sure space for the drain appears. We know we can do
520 * this because max_hw_segments has been adjusted to be one
521 * fewer than the device can handle.
523 rq
->nr_phys_segments
++;
526 EXPORT_SYMBOL(blk_mq_start_request
);
528 static void __blk_mq_requeue_request(struct request
*rq
)
530 struct request_queue
*q
= rq
->q
;
532 trace_block_rq_requeue(q
, rq
);
533 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
534 blk_mq_sched_requeue_request(rq
);
536 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
537 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
538 rq
->nr_phys_segments
--;
542 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
544 __blk_mq_requeue_request(rq
);
546 BUG_ON(blk_queued_rq(rq
));
547 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
549 EXPORT_SYMBOL(blk_mq_requeue_request
);
551 static void blk_mq_requeue_work(struct work_struct
*work
)
553 struct request_queue
*q
=
554 container_of(work
, struct request_queue
, requeue_work
.work
);
556 struct request
*rq
, *next
;
559 spin_lock_irqsave(&q
->requeue_lock
, flags
);
560 list_splice_init(&q
->requeue_list
, &rq_list
);
561 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
563 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
564 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
567 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
568 list_del_init(&rq
->queuelist
);
569 blk_mq_sched_insert_request(rq
, true, false, false, true);
572 while (!list_empty(&rq_list
)) {
573 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
574 list_del_init(&rq
->queuelist
);
575 blk_mq_sched_insert_request(rq
, false, false, false, true);
578 blk_mq_run_hw_queues(q
, false);
581 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
582 bool kick_requeue_list
)
584 struct request_queue
*q
= rq
->q
;
588 * We abuse this flag that is otherwise used by the I/O scheduler to
589 * request head insertation from the workqueue.
591 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
593 spin_lock_irqsave(&q
->requeue_lock
, flags
);
595 rq
->rq_flags
|= RQF_SOFTBARRIER
;
596 list_add(&rq
->queuelist
, &q
->requeue_list
);
598 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
600 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
602 if (kick_requeue_list
)
603 blk_mq_kick_requeue_list(q
);
605 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
607 void blk_mq_kick_requeue_list(struct request_queue
*q
)
609 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
611 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
613 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
616 kblockd_schedule_delayed_work(&q
->requeue_work
,
617 msecs_to_jiffies(msecs
));
619 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
621 void blk_mq_abort_requeue_list(struct request_queue
*q
)
626 spin_lock_irqsave(&q
->requeue_lock
, flags
);
627 list_splice_init(&q
->requeue_list
, &rq_list
);
628 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
630 while (!list_empty(&rq_list
)) {
633 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
634 list_del_init(&rq
->queuelist
);
636 blk_mq_end_request(rq
, rq
->errors
);
639 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
641 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
643 if (tag
< tags
->nr_tags
) {
644 prefetch(tags
->rqs
[tag
]);
645 return tags
->rqs
[tag
];
650 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
652 struct blk_mq_timeout_data
{
654 unsigned int next_set
;
657 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
659 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
660 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
663 * We know that complete is set at this point. If STARTED isn't set
664 * anymore, then the request isn't active and the "timeout" should
665 * just be ignored. This can happen due to the bitflag ordering.
666 * Timeout first checks if STARTED is set, and if it is, assumes
667 * the request is active. But if we race with completion, then
668 * we both flags will get cleared. So check here again, and ignore
669 * a timeout event with a request that isn't active.
671 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
675 ret
= ops
->timeout(req
, reserved
);
679 __blk_mq_complete_request(req
);
681 case BLK_EH_RESET_TIMER
:
683 blk_clear_rq_complete(req
);
685 case BLK_EH_NOT_HANDLED
:
688 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
693 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
694 struct request
*rq
, void *priv
, bool reserved
)
696 struct blk_mq_timeout_data
*data
= priv
;
698 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
700 * If a request wasn't started before the queue was
701 * marked dying, kill it here or it'll go unnoticed.
703 if (unlikely(blk_queue_dying(rq
->q
))) {
705 blk_mq_end_request(rq
, rq
->errors
);
710 if (time_after_eq(jiffies
, rq
->deadline
)) {
711 if (!blk_mark_rq_complete(rq
))
712 blk_mq_rq_timed_out(rq
, reserved
);
713 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
714 data
->next
= rq
->deadline
;
719 static void blk_mq_timeout_work(struct work_struct
*work
)
721 struct request_queue
*q
=
722 container_of(work
, struct request_queue
, timeout_work
);
723 struct blk_mq_timeout_data data
= {
729 /* A deadlock might occur if a request is stuck requiring a
730 * timeout at the same time a queue freeze is waiting
731 * completion, since the timeout code would not be able to
732 * acquire the queue reference here.
734 * That's why we don't use blk_queue_enter here; instead, we use
735 * percpu_ref_tryget directly, because we need to be able to
736 * obtain a reference even in the short window between the queue
737 * starting to freeze, by dropping the first reference in
738 * blk_mq_freeze_queue_start, and the moment the last request is
739 * consumed, marked by the instant q_usage_counter reaches
742 if (!percpu_ref_tryget(&q
->q_usage_counter
))
745 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
748 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
749 mod_timer(&q
->timeout
, data
.next
);
751 struct blk_mq_hw_ctx
*hctx
;
753 queue_for_each_hw_ctx(q
, hctx
, i
) {
754 /* the hctx may be unmapped, so check it here */
755 if (blk_mq_hw_queue_mapped(hctx
))
756 blk_mq_tag_idle(hctx
);
763 * Reverse check our software queue for entries that we could potentially
764 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
765 * too much time checking for merges.
767 static bool blk_mq_attempt_merge(struct request_queue
*q
,
768 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
773 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
779 if (!blk_rq_merge_ok(rq
, bio
))
782 switch (blk_try_merge(rq
, bio
)) {
783 case ELEVATOR_BACK_MERGE
:
784 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
785 merged
= bio_attempt_back_merge(q
, rq
, bio
);
787 case ELEVATOR_FRONT_MERGE
:
788 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
789 merged
= bio_attempt_front_merge(q
, rq
, bio
);
791 case ELEVATOR_DISCARD_MERGE
:
792 merged
= bio_attempt_discard_merge(q
, rq
, bio
);
806 struct flush_busy_ctx_data
{
807 struct blk_mq_hw_ctx
*hctx
;
808 struct list_head
*list
;
811 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
813 struct flush_busy_ctx_data
*flush_data
= data
;
814 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
815 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
817 sbitmap_clear_bit(sb
, bitnr
);
818 spin_lock(&ctx
->lock
);
819 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
820 spin_unlock(&ctx
->lock
);
825 * Process software queues that have been marked busy, splicing them
826 * to the for-dispatch
828 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
830 struct flush_busy_ctx_data data
= {
835 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
837 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
839 static inline unsigned int queued_to_index(unsigned int queued
)
844 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
847 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
850 struct blk_mq_alloc_data data
= {
852 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
853 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
863 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
864 data
.flags
|= BLK_MQ_REQ_RESERVED
;
866 rq
->tag
= blk_mq_get_tag(&data
);
868 if (blk_mq_tag_busy(data
.hctx
)) {
869 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
870 atomic_inc(&data
.hctx
->nr_active
);
872 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
879 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx
*hctx
,
882 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
885 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
886 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
887 atomic_dec(&hctx
->nr_active
);
891 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx
*hctx
,
894 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
897 __blk_mq_put_driver_tag(hctx
, rq
);
900 static void blk_mq_put_driver_tag(struct request
*rq
)
902 struct blk_mq_hw_ctx
*hctx
;
904 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
907 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
908 __blk_mq_put_driver_tag(hctx
, rq
);
912 * If we fail getting a driver tag because all the driver tags are already
913 * assigned and on the dispatch list, BUT the first entry does not have a
914 * tag, then we could deadlock. For that case, move entries with assigned
915 * driver tags to the front, leaving the set of tagged requests in the
916 * same order, and the untagged set in the same order.
918 static bool reorder_tags_to_front(struct list_head
*list
)
920 struct request
*rq
, *tmp
, *first
= NULL
;
922 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
926 list_move(&rq
->queuelist
, list
);
932 return first
!= NULL
;
935 static int blk_mq_dispatch_wake(wait_queue_t
*wait
, unsigned mode
, int flags
,
938 struct blk_mq_hw_ctx
*hctx
;
940 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
942 list_del(&wait
->task_list
);
943 clear_bit_unlock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
);
944 blk_mq_run_hw_queue(hctx
, true);
948 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx
*hctx
)
950 struct sbq_wait_state
*ws
;
953 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
954 * The thread which wins the race to grab this bit adds the hardware
955 * queue to the wait queue.
957 if (test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
) ||
958 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
961 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
962 ws
= bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
);
965 * As soon as this returns, it's no longer safe to fiddle with
966 * hctx->dispatch_wait, since a completion can wake up the wait queue
967 * and unlock the bit.
969 add_wait_queue(&ws
->wait
, &hctx
->dispatch_wait
);
973 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
975 struct request_queue
*q
= hctx
->queue
;
977 LIST_HEAD(driver_list
);
978 struct list_head
*dptr
;
979 int queued
, ret
= BLK_MQ_RQ_QUEUE_OK
;
982 * Start off with dptr being NULL, so we start the first request
983 * immediately, even if we have more pending.
988 * Now process all the entries, sending them to the driver.
991 while (!list_empty(list
)) {
992 struct blk_mq_queue_data bd
;
994 rq
= list_first_entry(list
, struct request
, queuelist
);
995 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
996 if (!queued
&& reorder_tags_to_front(list
))
1000 * The initial allocation attempt failed, so we need to
1001 * rerun the hardware queue when a tag is freed.
1003 if (blk_mq_dispatch_wait_add(hctx
)) {
1005 * It's possible that a tag was freed in the
1006 * window between the allocation failure and
1007 * adding the hardware queue to the wait queue.
1009 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1016 list_del_init(&rq
->queuelist
);
1022 * Flag last if we have no more requests, or if we have more
1023 * but can't assign a driver tag to it.
1025 if (list_empty(list
))
1028 struct request
*nxt
;
1030 nxt
= list_first_entry(list
, struct request
, queuelist
);
1031 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1034 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1036 case BLK_MQ_RQ_QUEUE_OK
:
1039 case BLK_MQ_RQ_QUEUE_BUSY
:
1040 blk_mq_put_driver_tag_hctx(hctx
, rq
);
1041 list_add(&rq
->queuelist
, list
);
1042 __blk_mq_requeue_request(rq
);
1045 pr_err("blk-mq: bad return on queue: %d\n", ret
);
1046 case BLK_MQ_RQ_QUEUE_ERROR
:
1048 blk_mq_end_request(rq
, rq
->errors
);
1052 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
1056 * We've done the first request. If we have more than 1
1057 * left in the list, set dptr to defer issue.
1059 if (!dptr
&& list
->next
!= list
->prev
)
1060 dptr
= &driver_list
;
1063 hctx
->dispatched
[queued_to_index(queued
)]++;
1066 * Any items that need requeuing? Stuff them into hctx->dispatch,
1067 * that is where we will continue on next queue run.
1069 if (!list_empty(list
)) {
1071 * If we got a driver tag for the next request already,
1074 rq
= list_first_entry(list
, struct request
, queuelist
);
1075 blk_mq_put_driver_tag(rq
);
1077 spin_lock(&hctx
->lock
);
1078 list_splice_init(list
, &hctx
->dispatch
);
1079 spin_unlock(&hctx
->lock
);
1082 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
1083 * it's possible the queue is stopped and restarted again
1084 * before this. Queue restart will dispatch requests. And since
1085 * requests in rq_list aren't added into hctx->dispatch yet,
1086 * the requests in rq_list might get lost.
1088 * blk_mq_run_hw_queue() already checks the STOPPED bit
1090 * If RESTART or TAG_WAITING is set, then let completion restart
1091 * the queue instead of potentially looping here.
1093 if (!blk_mq_sched_needs_restart(hctx
) &&
1094 !test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1095 blk_mq_run_hw_queue(hctx
, true);
1101 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1105 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1106 cpu_online(hctx
->next_cpu
));
1108 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1110 blk_mq_sched_dispatch_requests(hctx
);
1113 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1114 blk_mq_sched_dispatch_requests(hctx
);
1115 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1120 * It'd be great if the workqueue API had a way to pass
1121 * in a mask and had some smarts for more clever placement.
1122 * For now we just round-robin here, switching for every
1123 * BLK_MQ_CPU_WORK_BATCH queued items.
1125 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1127 if (hctx
->queue
->nr_hw_queues
== 1)
1128 return WORK_CPU_UNBOUND
;
1130 if (--hctx
->next_cpu_batch
<= 0) {
1133 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1134 if (next_cpu
>= nr_cpu_ids
)
1135 next_cpu
= cpumask_first(hctx
->cpumask
);
1137 hctx
->next_cpu
= next_cpu
;
1138 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1141 return hctx
->next_cpu
;
1144 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1146 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
1147 !blk_mq_hw_queue_mapped(hctx
)))
1150 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1151 int cpu
= get_cpu();
1152 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1153 __blk_mq_run_hw_queue(hctx
);
1161 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
);
1164 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1166 struct blk_mq_hw_ctx
*hctx
;
1169 queue_for_each_hw_ctx(q
, hctx
, i
) {
1170 if (!blk_mq_hctx_has_pending(hctx
) ||
1171 blk_mq_hctx_stopped(hctx
))
1174 blk_mq_run_hw_queue(hctx
, async
);
1177 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1180 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1181 * @q: request queue.
1183 * The caller is responsible for serializing this function against
1184 * blk_mq_{start,stop}_hw_queue().
1186 bool blk_mq_queue_stopped(struct request_queue
*q
)
1188 struct blk_mq_hw_ctx
*hctx
;
1191 queue_for_each_hw_ctx(q
, hctx
, i
)
1192 if (blk_mq_hctx_stopped(hctx
))
1197 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1199 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1201 cancel_work(&hctx
->run_work
);
1202 cancel_delayed_work(&hctx
->delay_work
);
1203 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1205 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1207 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1209 struct blk_mq_hw_ctx
*hctx
;
1212 queue_for_each_hw_ctx(q
, hctx
, i
)
1213 blk_mq_stop_hw_queue(hctx
);
1215 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1217 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1219 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1221 blk_mq_run_hw_queue(hctx
, false);
1223 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1225 void blk_mq_start_hw_queues(struct request_queue
*q
)
1227 struct blk_mq_hw_ctx
*hctx
;
1230 queue_for_each_hw_ctx(q
, hctx
, i
)
1231 blk_mq_start_hw_queue(hctx
);
1233 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1235 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1237 if (!blk_mq_hctx_stopped(hctx
))
1240 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1241 blk_mq_run_hw_queue(hctx
, async
);
1243 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1245 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1247 struct blk_mq_hw_ctx
*hctx
;
1250 queue_for_each_hw_ctx(q
, hctx
, i
)
1251 blk_mq_start_stopped_hw_queue(hctx
, async
);
1253 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1255 static void blk_mq_run_work_fn(struct work_struct
*work
)
1257 struct blk_mq_hw_ctx
*hctx
;
1259 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1261 __blk_mq_run_hw_queue(hctx
);
1264 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1266 struct blk_mq_hw_ctx
*hctx
;
1268 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1270 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1271 __blk_mq_run_hw_queue(hctx
);
1274 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1276 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1279 blk_mq_stop_hw_queue(hctx
);
1280 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1281 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1283 EXPORT_SYMBOL(blk_mq_delay_queue
);
1285 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1289 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1291 trace_block_rq_insert(hctx
->queue
, rq
);
1294 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1296 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1299 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1302 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1304 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1305 blk_mq_hctx_mark_pending(hctx
, ctx
);
1308 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1309 struct list_head
*list
)
1313 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1316 spin_lock(&ctx
->lock
);
1317 while (!list_empty(list
)) {
1320 rq
= list_first_entry(list
, struct request
, queuelist
);
1321 BUG_ON(rq
->mq_ctx
!= ctx
);
1322 list_del_init(&rq
->queuelist
);
1323 __blk_mq_insert_req_list(hctx
, rq
, false);
1325 blk_mq_hctx_mark_pending(hctx
, ctx
);
1326 spin_unlock(&ctx
->lock
);
1329 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1331 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1332 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1334 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1335 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1336 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1339 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1341 struct blk_mq_ctx
*this_ctx
;
1342 struct request_queue
*this_q
;
1345 LIST_HEAD(ctx_list
);
1348 list_splice_init(&plug
->mq_list
, &list
);
1350 list_sort(NULL
, &list
, plug_ctx_cmp
);
1356 while (!list_empty(&list
)) {
1357 rq
= list_entry_rq(list
.next
);
1358 list_del_init(&rq
->queuelist
);
1360 if (rq
->mq_ctx
!= this_ctx
) {
1362 trace_block_unplug(this_q
, depth
, from_schedule
);
1363 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1368 this_ctx
= rq
->mq_ctx
;
1374 list_add_tail(&rq
->queuelist
, &ctx_list
);
1378 * If 'this_ctx' is set, we know we have entries to complete
1379 * on 'ctx_list'. Do those.
1382 trace_block_unplug(this_q
, depth
, from_schedule
);
1383 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1388 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1390 init_request_from_bio(rq
, bio
);
1392 blk_account_io_start(rq
, true);
1395 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1397 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1398 !blk_queue_nomerges(hctx
->queue
);
1401 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1402 struct blk_mq_ctx
*ctx
,
1403 struct request
*rq
, struct bio
*bio
)
1405 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1406 blk_mq_bio_to_request(rq
, bio
);
1407 spin_lock(&ctx
->lock
);
1409 __blk_mq_insert_request(hctx
, rq
, false);
1410 spin_unlock(&ctx
->lock
);
1413 struct request_queue
*q
= hctx
->queue
;
1415 spin_lock(&ctx
->lock
);
1416 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1417 blk_mq_bio_to_request(rq
, bio
);
1421 spin_unlock(&ctx
->lock
);
1422 __blk_mq_finish_request(hctx
, ctx
, rq
);
1427 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1430 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1432 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1435 static void blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
)
1437 struct request_queue
*q
= rq
->q
;
1438 struct blk_mq_queue_data bd
= {
1443 struct blk_mq_hw_ctx
*hctx
;
1444 blk_qc_t new_cookie
;
1450 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1453 new_cookie
= request_to_qc_t(hctx
, rq
);
1456 * For OK queue, we are done. For error, kill it. Any other
1457 * error (busy), just add it to our list as we previously
1460 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1461 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1462 *cookie
= new_cookie
;
1466 __blk_mq_requeue_request(rq
);
1468 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1469 *cookie
= BLK_QC_T_NONE
;
1471 blk_mq_end_request(rq
, rq
->errors
);
1476 blk_mq_sched_insert_request(rq
, false, true, true, false);
1480 * Multiple hardware queue variant. This will not use per-process plugs,
1481 * but will attempt to bypass the hctx queueing if we can go straight to
1482 * hardware for SYNC IO.
1484 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1486 const int is_sync
= op_is_sync(bio
->bi_opf
);
1487 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1488 struct blk_mq_alloc_data data
= { .flags
= 0 };
1490 unsigned int request_count
= 0, srcu_idx
;
1491 struct blk_plug
*plug
;
1492 struct request
*same_queue_rq
= NULL
;
1494 unsigned int wb_acct
;
1496 blk_queue_bounce(q
, &bio
);
1498 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1500 return BLK_QC_T_NONE
;
1503 blk_queue_split(q
, &bio
, q
->bio_split
);
1505 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1506 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1507 return BLK_QC_T_NONE
;
1509 if (blk_mq_sched_bio_merge(q
, bio
))
1510 return BLK_QC_T_NONE
;
1512 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1514 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1516 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1517 if (unlikely(!rq
)) {
1518 __wbt_done(q
->rq_wb
, wb_acct
);
1519 return BLK_QC_T_NONE
;
1522 wbt_track(&rq
->issue_stat
, wb_acct
);
1524 cookie
= request_to_qc_t(data
.hctx
, rq
);
1526 if (unlikely(is_flush_fua
)) {
1529 blk_mq_bio_to_request(rq
, bio
);
1530 blk_insert_flush(rq
);
1534 plug
= current
->plug
;
1536 * If the driver supports defer issued based on 'last', then
1537 * queue it up like normal since we can potentially save some
1540 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1541 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1542 struct request
*old_rq
= NULL
;
1544 blk_mq_bio_to_request(rq
, bio
);
1547 * We do limited plugging. If the bio can be merged, do that.
1548 * Otherwise the existing request in the plug list will be
1549 * issued. So the plug list will have one request at most
1553 * The plug list might get flushed before this. If that
1554 * happens, same_queue_rq is invalid and plug list is
1557 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1558 old_rq
= same_queue_rq
;
1559 list_del_init(&old_rq
->queuelist
);
1561 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1562 } else /* is_sync */
1564 blk_mq_put_ctx(data
.ctx
);
1568 if (!(data
.hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1570 blk_mq_try_issue_directly(old_rq
, &cookie
);
1573 srcu_idx
= srcu_read_lock(&data
.hctx
->queue_rq_srcu
);
1574 blk_mq_try_issue_directly(old_rq
, &cookie
);
1575 srcu_read_unlock(&data
.hctx
->queue_rq_srcu
, srcu_idx
);
1582 blk_mq_put_ctx(data
.ctx
);
1583 blk_mq_bio_to_request(rq
, bio
);
1584 blk_mq_sched_insert_request(rq
, false, true,
1585 !is_sync
|| is_flush_fua
, true);
1588 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1590 * For a SYNC request, send it to the hardware immediately. For
1591 * an ASYNC request, just ensure that we run it later on. The
1592 * latter allows for merging opportunities and more efficient
1596 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1598 blk_mq_put_ctx(data
.ctx
);
1604 * Single hardware queue variant. This will attempt to use any per-process
1605 * plug for merging and IO deferral.
1607 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1609 const int is_sync
= op_is_sync(bio
->bi_opf
);
1610 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1611 struct blk_plug
*plug
;
1612 unsigned int request_count
= 0;
1613 struct blk_mq_alloc_data data
= { .flags
= 0 };
1616 unsigned int wb_acct
;
1618 blk_queue_bounce(q
, &bio
);
1620 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1622 return BLK_QC_T_NONE
;
1625 blk_queue_split(q
, &bio
, q
->bio_split
);
1627 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1628 if (blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1629 return BLK_QC_T_NONE
;
1631 request_count
= blk_plug_queued_count(q
);
1633 if (blk_mq_sched_bio_merge(q
, bio
))
1634 return BLK_QC_T_NONE
;
1636 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1638 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1640 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1641 if (unlikely(!rq
)) {
1642 __wbt_done(q
->rq_wb
, wb_acct
);
1643 return BLK_QC_T_NONE
;
1646 wbt_track(&rq
->issue_stat
, wb_acct
);
1648 cookie
= request_to_qc_t(data
.hctx
, rq
);
1650 if (unlikely(is_flush_fua
)) {
1653 blk_mq_bio_to_request(rq
, bio
);
1654 blk_insert_flush(rq
);
1659 * A task plug currently exists. Since this is completely lockless,
1660 * utilize that to temporarily store requests until the task is
1661 * either done or scheduled away.
1663 plug
= current
->plug
;
1665 struct request
*last
= NULL
;
1667 blk_mq_bio_to_request(rq
, bio
);
1670 * @request_count may become stale because of schedule
1671 * out, so check the list again.
1673 if (list_empty(&plug
->mq_list
))
1676 trace_block_plug(q
);
1678 last
= list_entry_rq(plug
->mq_list
.prev
);
1680 blk_mq_put_ctx(data
.ctx
);
1682 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1683 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1684 blk_flush_plug_list(plug
, false);
1685 trace_block_plug(q
);
1688 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1694 blk_mq_put_ctx(data
.ctx
);
1695 blk_mq_bio_to_request(rq
, bio
);
1696 blk_mq_sched_insert_request(rq
, false, true,
1697 !is_sync
|| is_flush_fua
, true);
1700 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1702 * For a SYNC request, send it to the hardware immediately. For
1703 * an ASYNC request, just ensure that we run it later on. The
1704 * latter allows for merging opportunities and more efficient
1708 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1711 blk_mq_put_ctx(data
.ctx
);
1716 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1717 unsigned int hctx_idx
)
1721 if (tags
->rqs
&& set
->ops
->exit_request
) {
1724 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1725 struct request
*rq
= tags
->static_rqs
[i
];
1729 set
->ops
->exit_request(set
->driver_data
, rq
,
1731 tags
->static_rqs
[i
] = NULL
;
1735 while (!list_empty(&tags
->page_list
)) {
1736 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1737 list_del_init(&page
->lru
);
1739 * Remove kmemleak object previously allocated in
1740 * blk_mq_init_rq_map().
1742 kmemleak_free(page_address(page
));
1743 __free_pages(page
, page
->private);
1747 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1751 kfree(tags
->static_rqs
);
1752 tags
->static_rqs
= NULL
;
1754 blk_mq_free_tags(tags
);
1757 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1758 unsigned int hctx_idx
,
1759 unsigned int nr_tags
,
1760 unsigned int reserved_tags
)
1762 struct blk_mq_tags
*tags
;
1765 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1766 if (node
== NUMA_NO_NODE
)
1767 node
= set
->numa_node
;
1769 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1770 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1774 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1775 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1778 blk_mq_free_tags(tags
);
1782 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1783 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1785 if (!tags
->static_rqs
) {
1787 blk_mq_free_tags(tags
);
1794 static size_t order_to_size(unsigned int order
)
1796 return (size_t)PAGE_SIZE
<< order
;
1799 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1800 unsigned int hctx_idx
, unsigned int depth
)
1802 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1803 size_t rq_size
, left
;
1806 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1807 if (node
== NUMA_NO_NODE
)
1808 node
= set
->numa_node
;
1810 INIT_LIST_HEAD(&tags
->page_list
);
1813 * rq_size is the size of the request plus driver payload, rounded
1814 * to the cacheline size
1816 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1818 left
= rq_size
* depth
;
1820 for (i
= 0; i
< depth
; ) {
1821 int this_order
= max_order
;
1826 while (this_order
&& left
< order_to_size(this_order
- 1))
1830 page
= alloc_pages_node(node
,
1831 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1837 if (order_to_size(this_order
) < rq_size
)
1844 page
->private = this_order
;
1845 list_add_tail(&page
->lru
, &tags
->page_list
);
1847 p
= page_address(page
);
1849 * Allow kmemleak to scan these pages as they contain pointers
1850 * to additional allocations like via ops->init_request().
1852 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1853 entries_per_page
= order_to_size(this_order
) / rq_size
;
1854 to_do
= min(entries_per_page
, depth
- i
);
1855 left
-= to_do
* rq_size
;
1856 for (j
= 0; j
< to_do
; j
++) {
1857 struct request
*rq
= p
;
1859 tags
->static_rqs
[i
] = rq
;
1860 if (set
->ops
->init_request
) {
1861 if (set
->ops
->init_request(set
->driver_data
,
1864 tags
->static_rqs
[i
] = NULL
;
1876 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1881 * 'cpu' is going away. splice any existing rq_list entries from this
1882 * software queue to the hw queue dispatch list, and ensure that it
1885 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1887 struct blk_mq_hw_ctx
*hctx
;
1888 struct blk_mq_ctx
*ctx
;
1891 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1892 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1894 spin_lock(&ctx
->lock
);
1895 if (!list_empty(&ctx
->rq_list
)) {
1896 list_splice_init(&ctx
->rq_list
, &tmp
);
1897 blk_mq_hctx_clear_pending(hctx
, ctx
);
1899 spin_unlock(&ctx
->lock
);
1901 if (list_empty(&tmp
))
1904 spin_lock(&hctx
->lock
);
1905 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1906 spin_unlock(&hctx
->lock
);
1908 blk_mq_run_hw_queue(hctx
, true);
1912 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1914 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1918 /* hctx->ctxs will be freed in queue's release handler */
1919 static void blk_mq_exit_hctx(struct request_queue
*q
,
1920 struct blk_mq_tag_set
*set
,
1921 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1923 unsigned flush_start_tag
= set
->queue_depth
;
1925 blk_mq_tag_idle(hctx
);
1927 if (set
->ops
->exit_request
)
1928 set
->ops
->exit_request(set
->driver_data
,
1929 hctx
->fq
->flush_rq
, hctx_idx
,
1930 flush_start_tag
+ hctx_idx
);
1932 if (set
->ops
->exit_hctx
)
1933 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1935 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1936 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1938 blk_mq_remove_cpuhp(hctx
);
1939 blk_free_flush_queue(hctx
->fq
);
1940 sbitmap_free(&hctx
->ctx_map
);
1943 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1944 struct blk_mq_tag_set
*set
, int nr_queue
)
1946 struct blk_mq_hw_ctx
*hctx
;
1949 queue_for_each_hw_ctx(q
, hctx
, i
) {
1952 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1956 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1957 struct blk_mq_tag_set
*set
)
1959 struct blk_mq_hw_ctx
*hctx
;
1962 queue_for_each_hw_ctx(q
, hctx
, i
)
1963 free_cpumask_var(hctx
->cpumask
);
1966 static int blk_mq_init_hctx(struct request_queue
*q
,
1967 struct blk_mq_tag_set
*set
,
1968 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1971 unsigned flush_start_tag
= set
->queue_depth
;
1973 node
= hctx
->numa_node
;
1974 if (node
== NUMA_NO_NODE
)
1975 node
= hctx
->numa_node
= set
->numa_node
;
1977 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1978 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1979 spin_lock_init(&hctx
->lock
);
1980 INIT_LIST_HEAD(&hctx
->dispatch
);
1982 hctx
->queue_num
= hctx_idx
;
1983 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1985 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1987 hctx
->tags
= set
->tags
[hctx_idx
];
1990 * Allocate space for all possible cpus to avoid allocation at
1993 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1996 goto unregister_cpu_notifier
;
1998 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
2004 if (set
->ops
->init_hctx
&&
2005 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2008 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
2012 if (set
->ops
->init_request
&&
2013 set
->ops
->init_request(set
->driver_data
,
2014 hctx
->fq
->flush_rq
, hctx_idx
,
2015 flush_start_tag
+ hctx_idx
, node
))
2018 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2019 init_srcu_struct(&hctx
->queue_rq_srcu
);
2026 if (set
->ops
->exit_hctx
)
2027 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2029 sbitmap_free(&hctx
->ctx_map
);
2032 unregister_cpu_notifier
:
2033 blk_mq_remove_cpuhp(hctx
);
2037 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2038 unsigned int nr_hw_queues
)
2042 for_each_possible_cpu(i
) {
2043 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2044 struct blk_mq_hw_ctx
*hctx
;
2046 memset(__ctx
, 0, sizeof(*__ctx
));
2048 spin_lock_init(&__ctx
->lock
);
2049 INIT_LIST_HEAD(&__ctx
->rq_list
);
2051 blk_stat_init(&__ctx
->stat
[BLK_STAT_READ
]);
2052 blk_stat_init(&__ctx
->stat
[BLK_STAT_WRITE
]);
2054 /* If the cpu isn't online, the cpu is mapped to first hctx */
2058 hctx
= blk_mq_map_queue(q
, i
);
2061 * Set local node, IFF we have more than one hw queue. If
2062 * not, we remain on the home node of the device
2064 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2065 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2069 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2073 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2074 set
->queue_depth
, set
->reserved_tags
);
2075 if (!set
->tags
[hctx_idx
])
2078 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2083 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2084 set
->tags
[hctx_idx
] = NULL
;
2088 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2089 unsigned int hctx_idx
)
2091 if (set
->tags
[hctx_idx
]) {
2092 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2093 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2094 set
->tags
[hctx_idx
] = NULL
;
2098 static void blk_mq_map_swqueue(struct request_queue
*q
,
2099 const struct cpumask
*online_mask
)
2101 unsigned int i
, hctx_idx
;
2102 struct blk_mq_hw_ctx
*hctx
;
2103 struct blk_mq_ctx
*ctx
;
2104 struct blk_mq_tag_set
*set
= q
->tag_set
;
2107 * Avoid others reading imcomplete hctx->cpumask through sysfs
2109 mutex_lock(&q
->sysfs_lock
);
2111 queue_for_each_hw_ctx(q
, hctx
, i
) {
2112 cpumask_clear(hctx
->cpumask
);
2117 * Map software to hardware queues
2119 for_each_possible_cpu(i
) {
2120 /* If the cpu isn't online, the cpu is mapped to first hctx */
2121 if (!cpumask_test_cpu(i
, online_mask
))
2124 hctx_idx
= q
->mq_map
[i
];
2125 /* unmapped hw queue can be remapped after CPU topo changed */
2126 if (!set
->tags
[hctx_idx
] &&
2127 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2129 * If tags initialization fail for some hctx,
2130 * that hctx won't be brought online. In this
2131 * case, remap the current ctx to hctx[0] which
2132 * is guaranteed to always have tags allocated
2137 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2138 hctx
= blk_mq_map_queue(q
, i
);
2140 cpumask_set_cpu(i
, hctx
->cpumask
);
2141 ctx
->index_hw
= hctx
->nr_ctx
;
2142 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2145 mutex_unlock(&q
->sysfs_lock
);
2147 queue_for_each_hw_ctx(q
, hctx
, i
) {
2149 * If no software queues are mapped to this hardware queue,
2150 * disable it and free the request entries.
2152 if (!hctx
->nr_ctx
) {
2153 /* Never unmap queue 0. We need it as a
2154 * fallback in case of a new remap fails
2157 if (i
&& set
->tags
[i
])
2158 blk_mq_free_map_and_requests(set
, i
);
2164 hctx
->tags
= set
->tags
[i
];
2165 WARN_ON(!hctx
->tags
);
2168 * Set the map size to the number of mapped software queues.
2169 * This is more accurate and more efficient than looping
2170 * over all possibly mapped software queues.
2172 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2175 * Initialize batch roundrobin counts
2177 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2178 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2182 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2184 struct blk_mq_hw_ctx
*hctx
;
2187 queue_for_each_hw_ctx(q
, hctx
, i
) {
2189 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2191 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2195 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
2197 struct request_queue
*q
;
2199 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2200 blk_mq_freeze_queue(q
);
2201 queue_set_hctx_shared(q
, shared
);
2202 blk_mq_unfreeze_queue(q
);
2206 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2208 struct blk_mq_tag_set
*set
= q
->tag_set
;
2210 mutex_lock(&set
->tag_list_lock
);
2211 list_del_init(&q
->tag_set_list
);
2212 if (list_is_singular(&set
->tag_list
)) {
2213 /* just transitioned to unshared */
2214 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2215 /* update existing queue */
2216 blk_mq_update_tag_set_depth(set
, false);
2218 mutex_unlock(&set
->tag_list_lock
);
2221 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2222 struct request_queue
*q
)
2226 mutex_lock(&set
->tag_list_lock
);
2228 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2229 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2230 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2231 /* update existing queue */
2232 blk_mq_update_tag_set_depth(set
, true);
2234 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2235 queue_set_hctx_shared(q
, true);
2236 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
2238 mutex_unlock(&set
->tag_list_lock
);
2242 * It is the actual release handler for mq, but we do it from
2243 * request queue's release handler for avoiding use-after-free
2244 * and headache because q->mq_kobj shouldn't have been introduced,
2245 * but we can't group ctx/kctx kobj without it.
2247 void blk_mq_release(struct request_queue
*q
)
2249 struct blk_mq_hw_ctx
*hctx
;
2252 blk_mq_sched_teardown(q
);
2254 /* hctx kobj stays in hctx */
2255 queue_for_each_hw_ctx(q
, hctx
, i
) {
2264 kfree(q
->queue_hw_ctx
);
2266 /* ctx kobj stays in queue_ctx */
2267 free_percpu(q
->queue_ctx
);
2270 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2272 struct request_queue
*uninit_q
, *q
;
2274 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2276 return ERR_PTR(-ENOMEM
);
2278 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2280 blk_cleanup_queue(uninit_q
);
2284 EXPORT_SYMBOL(blk_mq_init_queue
);
2286 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2287 struct request_queue
*q
)
2290 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2292 blk_mq_sysfs_unregister(q
);
2293 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2299 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2300 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2305 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2312 atomic_set(&hctxs
[i
]->nr_active
, 0);
2313 hctxs
[i
]->numa_node
= node
;
2314 hctxs
[i
]->queue_num
= i
;
2316 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2317 free_cpumask_var(hctxs
[i
]->cpumask
);
2322 blk_mq_hctx_kobj_init(hctxs
[i
]);
2324 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2325 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2329 blk_mq_free_map_and_requests(set
, j
);
2330 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2331 free_cpumask_var(hctx
->cpumask
);
2332 kobject_put(&hctx
->kobj
);
2339 q
->nr_hw_queues
= i
;
2340 blk_mq_sysfs_register(q
);
2343 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2344 struct request_queue
*q
)
2346 /* mark the queue as mq asap */
2347 q
->mq_ops
= set
->ops
;
2349 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2353 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2354 GFP_KERNEL
, set
->numa_node
);
2355 if (!q
->queue_hw_ctx
)
2358 q
->mq_map
= set
->mq_map
;
2360 blk_mq_realloc_hw_ctxs(set
, q
);
2361 if (!q
->nr_hw_queues
)
2364 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2365 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2367 q
->nr_queues
= nr_cpu_ids
;
2369 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2371 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2372 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2374 q
->sg_reserved_size
= INT_MAX
;
2376 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2377 INIT_LIST_HEAD(&q
->requeue_list
);
2378 spin_lock_init(&q
->requeue_lock
);
2380 if (q
->nr_hw_queues
> 1)
2381 blk_queue_make_request(q
, blk_mq_make_request
);
2383 blk_queue_make_request(q
, blk_sq_make_request
);
2386 * Do this after blk_queue_make_request() overrides it...
2388 q
->nr_requests
= set
->queue_depth
;
2391 * Default to classic polling
2395 if (set
->ops
->complete
)
2396 blk_queue_softirq_done(q
, set
->ops
->complete
);
2398 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2401 mutex_lock(&all_q_mutex
);
2403 list_add_tail(&q
->all_q_node
, &all_q_list
);
2404 blk_mq_add_queue_tag_set(set
, q
);
2405 blk_mq_map_swqueue(q
, cpu_online_mask
);
2407 mutex_unlock(&all_q_mutex
);
2410 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2413 ret
= blk_mq_sched_init(q
);
2415 return ERR_PTR(ret
);
2421 kfree(q
->queue_hw_ctx
);
2423 free_percpu(q
->queue_ctx
);
2426 return ERR_PTR(-ENOMEM
);
2428 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2430 void blk_mq_free_queue(struct request_queue
*q
)
2432 struct blk_mq_tag_set
*set
= q
->tag_set
;
2434 mutex_lock(&all_q_mutex
);
2435 list_del_init(&q
->all_q_node
);
2436 mutex_unlock(&all_q_mutex
);
2440 blk_mq_del_queue_tag_set(q
);
2442 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2443 blk_mq_free_hw_queues(q
, set
);
2446 /* Basically redo blk_mq_init_queue with queue frozen */
2447 static void blk_mq_queue_reinit(struct request_queue
*q
,
2448 const struct cpumask
*online_mask
)
2450 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2452 blk_mq_sysfs_unregister(q
);
2455 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2456 * we should change hctx numa_node according to new topology (this
2457 * involves free and re-allocate memory, worthy doing?)
2460 blk_mq_map_swqueue(q
, online_mask
);
2462 blk_mq_sysfs_register(q
);
2466 * New online cpumask which is going to be set in this hotplug event.
2467 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2468 * one-by-one and dynamically allocating this could result in a failure.
2470 static struct cpumask cpuhp_online_new
;
2472 static void blk_mq_queue_reinit_work(void)
2474 struct request_queue
*q
;
2476 mutex_lock(&all_q_mutex
);
2478 * We need to freeze and reinit all existing queues. Freezing
2479 * involves synchronous wait for an RCU grace period and doing it
2480 * one by one may take a long time. Start freezing all queues in
2481 * one swoop and then wait for the completions so that freezing can
2482 * take place in parallel.
2484 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2485 blk_mq_freeze_queue_start(q
);
2486 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2487 blk_mq_freeze_queue_wait(q
);
2489 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2490 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2492 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2493 blk_mq_unfreeze_queue(q
);
2495 mutex_unlock(&all_q_mutex
);
2498 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2500 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2501 blk_mq_queue_reinit_work();
2506 * Before hotadded cpu starts handling requests, new mappings must be
2507 * established. Otherwise, these requests in hw queue might never be
2510 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2511 * for CPU0, and ctx1 for CPU1).
2513 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2514 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2516 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2517 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2518 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2521 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2523 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2524 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2525 blk_mq_queue_reinit_work();
2529 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2533 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2534 if (!__blk_mq_alloc_rq_map(set
, i
))
2541 blk_mq_free_rq_map(set
->tags
[i
]);
2547 * Allocate the request maps associated with this tag_set. Note that this
2548 * may reduce the depth asked for, if memory is tight. set->queue_depth
2549 * will be updated to reflect the allocated depth.
2551 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2556 depth
= set
->queue_depth
;
2558 err
= __blk_mq_alloc_rq_maps(set
);
2562 set
->queue_depth
>>= 1;
2563 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2567 } while (set
->queue_depth
);
2569 if (!set
->queue_depth
|| err
) {
2570 pr_err("blk-mq: failed to allocate request map\n");
2574 if (depth
!= set
->queue_depth
)
2575 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2576 depth
, set
->queue_depth
);
2582 * Alloc a tag set to be associated with one or more request queues.
2583 * May fail with EINVAL for various error conditions. May adjust the
2584 * requested depth down, if if it too large. In that case, the set
2585 * value will be stored in set->queue_depth.
2587 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2591 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2593 if (!set
->nr_hw_queues
)
2595 if (!set
->queue_depth
)
2597 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2600 if (!set
->ops
->queue_rq
)
2603 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2604 pr_info("blk-mq: reduced tag depth to %u\n",
2606 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2610 * If a crashdump is active, then we are potentially in a very
2611 * memory constrained environment. Limit us to 1 queue and
2612 * 64 tags to prevent using too much memory.
2614 if (is_kdump_kernel()) {
2615 set
->nr_hw_queues
= 1;
2616 set
->queue_depth
= min(64U, set
->queue_depth
);
2619 * There is no use for more h/w queues than cpus.
2621 if (set
->nr_hw_queues
> nr_cpu_ids
)
2622 set
->nr_hw_queues
= nr_cpu_ids
;
2624 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2625 GFP_KERNEL
, set
->numa_node
);
2630 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2631 GFP_KERNEL
, set
->numa_node
);
2635 if (set
->ops
->map_queues
)
2636 ret
= set
->ops
->map_queues(set
);
2638 ret
= blk_mq_map_queues(set
);
2640 goto out_free_mq_map
;
2642 ret
= blk_mq_alloc_rq_maps(set
);
2644 goto out_free_mq_map
;
2646 mutex_init(&set
->tag_list_lock
);
2647 INIT_LIST_HEAD(&set
->tag_list
);
2659 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2661 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2665 for (i
= 0; i
< nr_cpu_ids
; i
++)
2666 blk_mq_free_map_and_requests(set
, i
);
2674 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2676 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2678 struct blk_mq_tag_set
*set
= q
->tag_set
;
2679 struct blk_mq_hw_ctx
*hctx
;
2685 blk_mq_freeze_queue(q
);
2686 blk_mq_quiesce_queue(q
);
2689 queue_for_each_hw_ctx(q
, hctx
, i
) {
2693 * If we're using an MQ scheduler, just update the scheduler
2694 * queue depth. This is similar to what the old code would do.
2696 if (!hctx
->sched_tags
) {
2697 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2698 min(nr
, set
->queue_depth
),
2701 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2709 q
->nr_requests
= nr
;
2711 blk_mq_unfreeze_queue(q
);
2712 blk_mq_start_stopped_hw_queues(q
, true);
2717 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2719 struct request_queue
*q
;
2721 if (nr_hw_queues
> nr_cpu_ids
)
2722 nr_hw_queues
= nr_cpu_ids
;
2723 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2726 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2727 blk_mq_freeze_queue(q
);
2729 set
->nr_hw_queues
= nr_hw_queues
;
2730 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2731 blk_mq_realloc_hw_ctxs(set
, q
);
2734 * Manually set the make_request_fn as blk_queue_make_request
2735 * resets a lot of the queue settings.
2737 if (q
->nr_hw_queues
> 1)
2738 q
->make_request_fn
= blk_mq_make_request
;
2740 q
->make_request_fn
= blk_sq_make_request
;
2742 blk_mq_queue_reinit(q
, cpu_online_mask
);
2745 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2746 blk_mq_unfreeze_queue(q
);
2748 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2750 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2751 struct blk_mq_hw_ctx
*hctx
,
2754 struct blk_rq_stat stat
[2];
2755 unsigned long ret
= 0;
2758 * If stats collection isn't on, don't sleep but turn it on for
2761 if (!blk_stat_enable(q
))
2765 * We don't have to do this once per IO, should optimize this
2766 * to just use the current window of stats until it changes
2768 memset(&stat
, 0, sizeof(stat
));
2769 blk_hctx_stat_get(hctx
, stat
);
2772 * As an optimistic guess, use half of the mean service time
2773 * for this type of request. We can (and should) make this smarter.
2774 * For instance, if the completion latencies are tight, we can
2775 * get closer than just half the mean. This is especially
2776 * important on devices where the completion latencies are longer
2779 if (req_op(rq
) == REQ_OP_READ
&& stat
[BLK_STAT_READ
].nr_samples
)
2780 ret
= (stat
[BLK_STAT_READ
].mean
+ 1) / 2;
2781 else if (req_op(rq
) == REQ_OP_WRITE
&& stat
[BLK_STAT_WRITE
].nr_samples
)
2782 ret
= (stat
[BLK_STAT_WRITE
].mean
+ 1) / 2;
2787 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2788 struct blk_mq_hw_ctx
*hctx
,
2791 struct hrtimer_sleeper hs
;
2792 enum hrtimer_mode mode
;
2796 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2802 * -1: don't ever hybrid sleep
2803 * 0: use half of prev avg
2804 * >0: use this specific value
2806 if (q
->poll_nsec
== -1)
2808 else if (q
->poll_nsec
> 0)
2809 nsecs
= q
->poll_nsec
;
2811 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2816 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2819 * This will be replaced with the stats tracking code, using
2820 * 'avg_completion_time / 2' as the pre-sleep target.
2824 mode
= HRTIMER_MODE_REL
;
2825 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2826 hrtimer_set_expires(&hs
.timer
, kt
);
2828 hrtimer_init_sleeper(&hs
, current
);
2830 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2832 set_current_state(TASK_UNINTERRUPTIBLE
);
2833 hrtimer_start_expires(&hs
.timer
, mode
);
2836 hrtimer_cancel(&hs
.timer
);
2837 mode
= HRTIMER_MODE_ABS
;
2838 } while (hs
.task
&& !signal_pending(current
));
2840 __set_current_state(TASK_RUNNING
);
2841 destroy_hrtimer_on_stack(&hs
.timer
);
2845 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2847 struct request_queue
*q
= hctx
->queue
;
2851 * If we sleep, have the caller restart the poll loop to reset
2852 * the state. Like for the other success return cases, the
2853 * caller is responsible for checking if the IO completed. If
2854 * the IO isn't complete, we'll get called again and will go
2855 * straight to the busy poll loop.
2857 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2860 hctx
->poll_considered
++;
2862 state
= current
->state
;
2863 while (!need_resched()) {
2866 hctx
->poll_invoked
++;
2868 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2870 hctx
->poll_success
++;
2871 set_current_state(TASK_RUNNING
);
2875 if (signal_pending_state(state
, current
))
2876 set_current_state(TASK_RUNNING
);
2878 if (current
->state
== TASK_RUNNING
)
2888 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2890 struct blk_mq_hw_ctx
*hctx
;
2891 struct blk_plug
*plug
;
2894 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2895 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2898 plug
= current
->plug
;
2900 blk_flush_plug_list(plug
, false);
2902 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2903 if (!blk_qc_t_is_internal(cookie
))
2904 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2906 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2908 return __blk_mq_poll(hctx
, rq
);
2910 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2912 void blk_mq_disable_hotplug(void)
2914 mutex_lock(&all_q_mutex
);
2917 void blk_mq_enable_hotplug(void)
2919 mutex_unlock(&all_q_mutex
);
2922 static int __init
blk_mq_init(void)
2924 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2925 blk_mq_hctx_notify_dead
);
2927 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2928 blk_mq_queue_reinit_prepare
,
2929 blk_mq_queue_reinit_dead
);
2932 subsys_initcall(blk_mq_init
);