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/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
29 #include <trace/events/block.h>
31 #include <linux/blk-mq.h>
34 #include "blk-mq-tag.h"
37 #include "blk-mq-sched.h"
39 static DEFINE_MUTEX(all_q_mutex
);
40 static LIST_HEAD(all_q_list
);
43 * Check if any of the ctx's have pending work in this hardware queue
45 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
47 return sbitmap_any_bit_set(&hctx
->ctx_map
) ||
48 !list_empty_careful(&hctx
->dispatch
) ||
49 blk_mq_sched_has_work(hctx
);
53 * Mark this ctx as having pending work in this hardware queue
55 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
56 struct blk_mq_ctx
*ctx
)
58 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
59 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
62 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
63 struct blk_mq_ctx
*ctx
)
65 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
68 void blk_mq_freeze_queue_start(struct request_queue
*q
)
72 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
73 if (freeze_depth
== 1) {
74 percpu_ref_kill(&q
->q_usage_counter
);
75 blk_mq_run_hw_queues(q
, false);
78 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start
);
80 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
82 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
84 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
86 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
87 unsigned long timeout
)
89 return wait_event_timeout(q
->mq_freeze_wq
,
90 percpu_ref_is_zero(&q
->q_usage_counter
),
93 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
96 * Guarantee no request is in use, so we can change any data structure of
97 * the queue afterward.
99 void blk_freeze_queue(struct request_queue
*q
)
102 * In the !blk_mq case we are only calling this to kill the
103 * q_usage_counter, otherwise this increases the freeze depth
104 * and waits for it to return to zero. For this reason there is
105 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
106 * exported to drivers as the only user for unfreeze is blk_mq.
108 blk_mq_freeze_queue_start(q
);
109 blk_mq_freeze_queue_wait(q
);
112 void blk_mq_freeze_queue(struct request_queue
*q
)
115 * ...just an alias to keep freeze and unfreeze actions balanced
116 * in the blk_mq_* namespace
120 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
122 void blk_mq_unfreeze_queue(struct request_queue
*q
)
126 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
127 WARN_ON_ONCE(freeze_depth
< 0);
129 percpu_ref_reinit(&q
->q_usage_counter
);
130 wake_up_all(&q
->mq_freeze_wq
);
133 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
136 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
139 * Note: this function does not prevent that the struct request end_io()
140 * callback function is invoked. Additionally, it is not prevented that
141 * new queue_rq() calls occur unless the queue has been stopped first.
143 void blk_mq_quiesce_queue(struct request_queue
*q
)
145 struct blk_mq_hw_ctx
*hctx
;
149 blk_mq_stop_hw_queues(q
);
151 queue_for_each_hw_ctx(q
, hctx
, i
) {
152 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
153 synchronize_srcu(&hctx
->queue_rq_srcu
);
160 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
162 void blk_mq_wake_waiters(struct request_queue
*q
)
164 struct blk_mq_hw_ctx
*hctx
;
167 queue_for_each_hw_ctx(q
, hctx
, i
)
168 if (blk_mq_hw_queue_mapped(hctx
))
169 blk_mq_tag_wakeup_all(hctx
->tags
, true);
172 * If we are called because the queue has now been marked as
173 * dying, we need to ensure that processes currently waiting on
174 * the queue are notified as well.
176 wake_up_all(&q
->mq_freeze_wq
);
179 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
181 return blk_mq_has_free_tags(hctx
->tags
);
183 EXPORT_SYMBOL(blk_mq_can_queue
);
185 void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
186 struct request
*rq
, unsigned int op
)
188 INIT_LIST_HEAD(&rq
->queuelist
);
189 /* csd/requeue_work/fifo_time is initialized before use */
193 if (blk_queue_io_stat(q
))
194 rq
->rq_flags
|= RQF_IO_STAT
;
195 /* do not touch atomic flags, it needs atomic ops against the timer */
197 INIT_HLIST_NODE(&rq
->hash
);
198 RB_CLEAR_NODE(&rq
->rb_node
);
201 rq
->start_time
= jiffies
;
202 #ifdef CONFIG_BLK_CGROUP
204 set_start_time_ns(rq
);
205 rq
->io_start_time_ns
= 0;
207 rq
->nr_phys_segments
= 0;
208 #if defined(CONFIG_BLK_DEV_INTEGRITY)
209 rq
->nr_integrity_segments
= 0;
212 /* tag was already set */
216 INIT_LIST_HEAD(&rq
->timeout_list
);
220 rq
->end_io_data
= NULL
;
223 ctx
->rq_dispatched
[op_is_sync(op
)]++;
225 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init
);
227 struct request
*__blk_mq_alloc_request(struct blk_mq_alloc_data
*data
,
233 tag
= blk_mq_get_tag(data
);
234 if (tag
!= BLK_MQ_TAG_FAIL
) {
235 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
237 rq
= tags
->static_rqs
[tag
];
239 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
241 rq
->internal_tag
= tag
;
243 if (blk_mq_tag_busy(data
->hctx
)) {
244 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
245 atomic_inc(&data
->hctx
->nr_active
);
248 rq
->internal_tag
= -1;
249 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
252 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
);
258 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request
);
260 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
263 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
267 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
271 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
273 blk_mq_put_ctx(alloc_data
.ctx
);
277 return ERR_PTR(-EWOULDBLOCK
);
280 rq
->__sector
= (sector_t
) -1;
281 rq
->bio
= rq
->biotail
= NULL
;
284 EXPORT_SYMBOL(blk_mq_alloc_request
);
286 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
287 unsigned int flags
, unsigned int hctx_idx
)
289 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
295 * If the tag allocator sleeps we could get an allocation for a
296 * different hardware context. No need to complicate the low level
297 * allocator for this for the rare use case of a command tied to
300 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
301 return ERR_PTR(-EINVAL
);
303 if (hctx_idx
>= q
->nr_hw_queues
)
304 return ERR_PTR(-EIO
);
306 ret
= blk_queue_enter(q
, true);
311 * Check if the hardware context is actually mapped to anything.
312 * If not tell the caller that it should skip this queue.
314 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
315 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
317 return ERR_PTR(-EXDEV
);
319 cpu
= cpumask_first(alloc_data
.hctx
->cpumask
);
320 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
322 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
324 blk_mq_put_ctx(alloc_data
.ctx
);
328 return ERR_PTR(-EWOULDBLOCK
);
332 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
334 void __blk_mq_finish_request(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
337 const int sched_tag
= rq
->internal_tag
;
338 struct request_queue
*q
= rq
->q
;
340 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
341 atomic_dec(&hctx
->nr_active
);
343 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
346 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
347 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
349 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
351 blk_mq_sched_completed_request(hctx
, rq
);
352 blk_mq_sched_restart_queues(hctx
);
356 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx
*hctx
,
359 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
361 ctx
->rq_completed
[rq_is_sync(rq
)]++;
362 __blk_mq_finish_request(hctx
, ctx
, rq
);
365 void blk_mq_finish_request(struct request
*rq
)
367 blk_mq_finish_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
370 void blk_mq_free_request(struct request
*rq
)
372 blk_mq_sched_put_request(rq
);
374 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
376 inline void __blk_mq_end_request(struct request
*rq
, int error
)
378 blk_account_io_done(rq
);
381 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
382 rq
->end_io(rq
, error
);
384 if (unlikely(blk_bidi_rq(rq
)))
385 blk_mq_free_request(rq
->next_rq
);
386 blk_mq_free_request(rq
);
389 EXPORT_SYMBOL(__blk_mq_end_request
);
391 void blk_mq_end_request(struct request
*rq
, int error
)
393 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
395 __blk_mq_end_request(rq
, error
);
397 EXPORT_SYMBOL(blk_mq_end_request
);
399 static void __blk_mq_complete_request_remote(void *data
)
401 struct request
*rq
= data
;
403 rq
->q
->softirq_done_fn(rq
);
406 static void blk_mq_ipi_complete_request(struct request
*rq
)
408 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
412 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
413 rq
->q
->softirq_done_fn(rq
);
418 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
419 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
421 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
422 rq
->csd
.func
= __blk_mq_complete_request_remote
;
425 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
427 rq
->q
->softirq_done_fn(rq
);
432 static void blk_mq_stat_add(struct request
*rq
)
434 if (rq
->rq_flags
& RQF_STATS
) {
436 * We could rq->mq_ctx here, but there's less of a risk
437 * of races if we have the completion event add the stats
438 * to the local software queue.
440 struct blk_mq_ctx
*ctx
;
442 ctx
= __blk_mq_get_ctx(rq
->q
, raw_smp_processor_id());
443 blk_stat_add(&ctx
->stat
[rq_data_dir(rq
)], rq
);
447 static void __blk_mq_complete_request(struct request
*rq
)
449 struct request_queue
*q
= rq
->q
;
453 if (!q
->softirq_done_fn
)
454 blk_mq_end_request(rq
, rq
->errors
);
456 blk_mq_ipi_complete_request(rq
);
460 * blk_mq_complete_request - end I/O on a request
461 * @rq: the request being processed
464 * Ends all I/O on a request. It does not handle partial completions.
465 * The actual completion happens out-of-order, through a IPI handler.
467 void blk_mq_complete_request(struct request
*rq
, int error
)
469 struct request_queue
*q
= rq
->q
;
471 if (unlikely(blk_should_fake_timeout(q
)))
473 if (!blk_mark_rq_complete(rq
)) {
475 __blk_mq_complete_request(rq
);
478 EXPORT_SYMBOL(blk_mq_complete_request
);
480 int blk_mq_request_started(struct request
*rq
)
482 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
484 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
486 void blk_mq_start_request(struct request
*rq
)
488 struct request_queue
*q
= rq
->q
;
490 blk_mq_sched_started_request(rq
);
492 trace_block_rq_issue(q
, rq
);
494 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
495 blk_stat_set_issue_time(&rq
->issue_stat
);
496 rq
->rq_flags
|= RQF_STATS
;
497 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
503 * Ensure that ->deadline is visible before set the started
504 * flag and clear the completed flag.
506 smp_mb__before_atomic();
509 * Mark us as started and clear complete. Complete might have been
510 * set if requeue raced with timeout, which then marked it as
511 * complete. So be sure to clear complete again when we start
512 * the request, otherwise we'll ignore the completion event.
514 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
515 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
516 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
517 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
519 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
521 * Make sure space for the drain appears. We know we can do
522 * this because max_hw_segments has been adjusted to be one
523 * fewer than the device can handle.
525 rq
->nr_phys_segments
++;
528 EXPORT_SYMBOL(blk_mq_start_request
);
530 static void __blk_mq_requeue_request(struct request
*rq
)
532 struct request_queue
*q
= rq
->q
;
534 trace_block_rq_requeue(q
, rq
);
535 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
536 blk_mq_sched_requeue_request(rq
);
538 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
539 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
540 rq
->nr_phys_segments
--;
544 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
546 __blk_mq_requeue_request(rq
);
548 BUG_ON(blk_queued_rq(rq
));
549 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
551 EXPORT_SYMBOL(blk_mq_requeue_request
);
553 static void blk_mq_requeue_work(struct work_struct
*work
)
555 struct request_queue
*q
=
556 container_of(work
, struct request_queue
, requeue_work
.work
);
558 struct request
*rq
, *next
;
561 spin_lock_irqsave(&q
->requeue_lock
, flags
);
562 list_splice_init(&q
->requeue_list
, &rq_list
);
563 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
565 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
566 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
569 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
570 list_del_init(&rq
->queuelist
);
571 blk_mq_sched_insert_request(rq
, true, false, false, true);
574 while (!list_empty(&rq_list
)) {
575 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
576 list_del_init(&rq
->queuelist
);
577 blk_mq_sched_insert_request(rq
, false, false, false, true);
580 blk_mq_run_hw_queues(q
, false);
583 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
584 bool kick_requeue_list
)
586 struct request_queue
*q
= rq
->q
;
590 * We abuse this flag that is otherwise used by the I/O scheduler to
591 * request head insertation from the workqueue.
593 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
595 spin_lock_irqsave(&q
->requeue_lock
, flags
);
597 rq
->rq_flags
|= RQF_SOFTBARRIER
;
598 list_add(&rq
->queuelist
, &q
->requeue_list
);
600 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
602 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
604 if (kick_requeue_list
)
605 blk_mq_kick_requeue_list(q
);
607 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
609 void blk_mq_kick_requeue_list(struct request_queue
*q
)
611 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
613 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
615 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
618 kblockd_schedule_delayed_work(&q
->requeue_work
,
619 msecs_to_jiffies(msecs
));
621 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
623 void blk_mq_abort_requeue_list(struct request_queue
*q
)
628 spin_lock_irqsave(&q
->requeue_lock
, flags
);
629 list_splice_init(&q
->requeue_list
, &rq_list
);
630 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
632 while (!list_empty(&rq_list
)) {
635 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
636 list_del_init(&rq
->queuelist
);
638 blk_mq_end_request(rq
, rq
->errors
);
641 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
643 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
645 if (tag
< tags
->nr_tags
) {
646 prefetch(tags
->rqs
[tag
]);
647 return tags
->rqs
[tag
];
652 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
654 struct blk_mq_timeout_data
{
656 unsigned int next_set
;
659 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
661 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
662 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
665 * We know that complete is set at this point. If STARTED isn't set
666 * anymore, then the request isn't active and the "timeout" should
667 * just be ignored. This can happen due to the bitflag ordering.
668 * Timeout first checks if STARTED is set, and if it is, assumes
669 * the request is active. But if we race with completion, then
670 * we both flags will get cleared. So check here again, and ignore
671 * a timeout event with a request that isn't active.
673 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
677 ret
= ops
->timeout(req
, reserved
);
681 __blk_mq_complete_request(req
);
683 case BLK_EH_RESET_TIMER
:
685 blk_clear_rq_complete(req
);
687 case BLK_EH_NOT_HANDLED
:
690 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
695 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
696 struct request
*rq
, void *priv
, bool reserved
)
698 struct blk_mq_timeout_data
*data
= priv
;
700 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
702 * If a request wasn't started before the queue was
703 * marked dying, kill it here or it'll go unnoticed.
705 if (unlikely(blk_queue_dying(rq
->q
))) {
707 blk_mq_end_request(rq
, rq
->errors
);
712 if (time_after_eq(jiffies
, rq
->deadline
)) {
713 if (!blk_mark_rq_complete(rq
))
714 blk_mq_rq_timed_out(rq
, reserved
);
715 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
716 data
->next
= rq
->deadline
;
721 static void blk_mq_timeout_work(struct work_struct
*work
)
723 struct request_queue
*q
=
724 container_of(work
, struct request_queue
, timeout_work
);
725 struct blk_mq_timeout_data data
= {
731 /* A deadlock might occur if a request is stuck requiring a
732 * timeout at the same time a queue freeze is waiting
733 * completion, since the timeout code would not be able to
734 * acquire the queue reference here.
736 * That's why we don't use blk_queue_enter here; instead, we use
737 * percpu_ref_tryget directly, because we need to be able to
738 * obtain a reference even in the short window between the queue
739 * starting to freeze, by dropping the first reference in
740 * blk_mq_freeze_queue_start, and the moment the last request is
741 * consumed, marked by the instant q_usage_counter reaches
744 if (!percpu_ref_tryget(&q
->q_usage_counter
))
747 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
750 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
751 mod_timer(&q
->timeout
, data
.next
);
753 struct blk_mq_hw_ctx
*hctx
;
755 queue_for_each_hw_ctx(q
, hctx
, i
) {
756 /* the hctx may be unmapped, so check it here */
757 if (blk_mq_hw_queue_mapped(hctx
))
758 blk_mq_tag_idle(hctx
);
765 * Reverse check our software queue for entries that we could potentially
766 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
767 * too much time checking for merges.
769 static bool blk_mq_attempt_merge(struct request_queue
*q
,
770 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
775 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
781 if (!blk_rq_merge_ok(rq
, bio
))
784 switch (blk_try_merge(rq
, bio
)) {
785 case ELEVATOR_BACK_MERGE
:
786 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
787 merged
= bio_attempt_back_merge(q
, rq
, bio
);
789 case ELEVATOR_FRONT_MERGE
:
790 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
791 merged
= bio_attempt_front_merge(q
, rq
, bio
);
793 case ELEVATOR_DISCARD_MERGE
:
794 merged
= bio_attempt_discard_merge(q
, rq
, bio
);
808 struct flush_busy_ctx_data
{
809 struct blk_mq_hw_ctx
*hctx
;
810 struct list_head
*list
;
813 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
815 struct flush_busy_ctx_data
*flush_data
= data
;
816 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
817 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
819 sbitmap_clear_bit(sb
, bitnr
);
820 spin_lock(&ctx
->lock
);
821 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
822 spin_unlock(&ctx
->lock
);
827 * Process software queues that have been marked busy, splicing them
828 * to the for-dispatch
830 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
832 struct flush_busy_ctx_data data
= {
837 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
839 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
841 static inline unsigned int queued_to_index(unsigned int queued
)
846 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
849 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
852 struct blk_mq_alloc_data data
= {
854 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
855 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
865 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
866 data
.flags
|= BLK_MQ_REQ_RESERVED
;
868 rq
->tag
= blk_mq_get_tag(&data
);
870 if (blk_mq_tag_busy(data
.hctx
)) {
871 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
872 atomic_inc(&data
.hctx
->nr_active
);
874 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
881 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx
*hctx
,
884 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
887 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
888 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
889 atomic_dec(&hctx
->nr_active
);
893 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx
*hctx
,
896 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
899 __blk_mq_put_driver_tag(hctx
, rq
);
902 static void blk_mq_put_driver_tag(struct request
*rq
)
904 struct blk_mq_hw_ctx
*hctx
;
906 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
909 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
910 __blk_mq_put_driver_tag(hctx
, rq
);
914 * If we fail getting a driver tag because all the driver tags are already
915 * assigned and on the dispatch list, BUT the first entry does not have a
916 * tag, then we could deadlock. For that case, move entries with assigned
917 * driver tags to the front, leaving the set of tagged requests in the
918 * same order, and the untagged set in the same order.
920 static bool reorder_tags_to_front(struct list_head
*list
)
922 struct request
*rq
, *tmp
, *first
= NULL
;
924 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
928 list_move(&rq
->queuelist
, list
);
934 return first
!= NULL
;
937 static int blk_mq_dispatch_wake(wait_queue_t
*wait
, unsigned mode
, int flags
,
940 struct blk_mq_hw_ctx
*hctx
;
942 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
944 list_del(&wait
->task_list
);
945 clear_bit_unlock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
);
946 blk_mq_run_hw_queue(hctx
, true);
950 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx
*hctx
)
952 struct sbq_wait_state
*ws
;
955 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
956 * The thread which wins the race to grab this bit adds the hardware
957 * queue to the wait queue.
959 if (test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
) ||
960 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
963 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
964 ws
= bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
);
967 * As soon as this returns, it's no longer safe to fiddle with
968 * hctx->dispatch_wait, since a completion can wake up the wait queue
969 * and unlock the bit.
971 add_wait_queue(&ws
->wait
, &hctx
->dispatch_wait
);
975 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
977 struct request_queue
*q
= hctx
->queue
;
979 LIST_HEAD(driver_list
);
980 struct list_head
*dptr
;
981 int queued
, ret
= BLK_MQ_RQ_QUEUE_OK
;
984 * Start off with dptr being NULL, so we start the first request
985 * immediately, even if we have more pending.
990 * Now process all the entries, sending them to the driver.
993 while (!list_empty(list
)) {
994 struct blk_mq_queue_data bd
;
996 rq
= list_first_entry(list
, struct request
, queuelist
);
997 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
998 if (!queued
&& reorder_tags_to_front(list
))
1002 * The initial allocation attempt failed, so we need to
1003 * rerun the hardware queue when a tag is freed.
1005 if (blk_mq_dispatch_wait_add(hctx
)) {
1007 * It's possible that a tag was freed in the
1008 * window between the allocation failure and
1009 * adding the hardware queue to the wait queue.
1011 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1018 list_del_init(&rq
->queuelist
);
1024 * Flag last if we have no more requests, or if we have more
1025 * but can't assign a driver tag to it.
1027 if (list_empty(list
))
1030 struct request
*nxt
;
1032 nxt
= list_first_entry(list
, struct request
, queuelist
);
1033 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1036 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1038 case BLK_MQ_RQ_QUEUE_OK
:
1041 case BLK_MQ_RQ_QUEUE_BUSY
:
1042 blk_mq_put_driver_tag_hctx(hctx
, rq
);
1043 list_add(&rq
->queuelist
, list
);
1044 __blk_mq_requeue_request(rq
);
1047 pr_err("blk-mq: bad return on queue: %d\n", ret
);
1048 case BLK_MQ_RQ_QUEUE_ERROR
:
1050 blk_mq_end_request(rq
, rq
->errors
);
1054 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
1058 * We've done the first request. If we have more than 1
1059 * left in the list, set dptr to defer issue.
1061 if (!dptr
&& list
->next
!= list
->prev
)
1062 dptr
= &driver_list
;
1065 hctx
->dispatched
[queued_to_index(queued
)]++;
1068 * Any items that need requeuing? Stuff them into hctx->dispatch,
1069 * that is where we will continue on next queue run.
1071 if (!list_empty(list
)) {
1073 * If we got a driver tag for the next request already,
1076 rq
= list_first_entry(list
, struct request
, queuelist
);
1077 blk_mq_put_driver_tag(rq
);
1079 spin_lock(&hctx
->lock
);
1080 list_splice_init(list
, &hctx
->dispatch
);
1081 spin_unlock(&hctx
->lock
);
1084 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
1085 * it's possible the queue is stopped and restarted again
1086 * before this. Queue restart will dispatch requests. And since
1087 * requests in rq_list aren't added into hctx->dispatch yet,
1088 * the requests in rq_list might get lost.
1090 * blk_mq_run_hw_queue() already checks the STOPPED bit
1092 * If RESTART or TAG_WAITING is set, then let completion restart
1093 * the queue instead of potentially looping here.
1095 if (!blk_mq_sched_needs_restart(hctx
) &&
1096 !test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1097 blk_mq_run_hw_queue(hctx
, true);
1103 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1107 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1108 cpu_online(hctx
->next_cpu
));
1110 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1112 blk_mq_sched_dispatch_requests(hctx
);
1115 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1116 blk_mq_sched_dispatch_requests(hctx
);
1117 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1122 * It'd be great if the workqueue API had a way to pass
1123 * in a mask and had some smarts for more clever placement.
1124 * For now we just round-robin here, switching for every
1125 * BLK_MQ_CPU_WORK_BATCH queued items.
1127 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1129 if (hctx
->queue
->nr_hw_queues
== 1)
1130 return WORK_CPU_UNBOUND
;
1132 if (--hctx
->next_cpu_batch
<= 0) {
1135 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1136 if (next_cpu
>= nr_cpu_ids
)
1137 next_cpu
= cpumask_first(hctx
->cpumask
);
1139 hctx
->next_cpu
= next_cpu
;
1140 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1143 return hctx
->next_cpu
;
1146 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1148 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
1149 !blk_mq_hw_queue_mapped(hctx
)))
1152 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1153 int cpu
= get_cpu();
1154 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1155 __blk_mq_run_hw_queue(hctx
);
1163 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
);
1166 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1168 struct blk_mq_hw_ctx
*hctx
;
1171 queue_for_each_hw_ctx(q
, hctx
, i
) {
1172 if (!blk_mq_hctx_has_pending(hctx
) ||
1173 blk_mq_hctx_stopped(hctx
))
1176 blk_mq_run_hw_queue(hctx
, async
);
1179 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1182 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1183 * @q: request queue.
1185 * The caller is responsible for serializing this function against
1186 * blk_mq_{start,stop}_hw_queue().
1188 bool blk_mq_queue_stopped(struct request_queue
*q
)
1190 struct blk_mq_hw_ctx
*hctx
;
1193 queue_for_each_hw_ctx(q
, hctx
, i
)
1194 if (blk_mq_hctx_stopped(hctx
))
1199 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1201 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1203 cancel_work(&hctx
->run_work
);
1204 cancel_delayed_work(&hctx
->delay_work
);
1205 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1207 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1209 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1211 struct blk_mq_hw_ctx
*hctx
;
1214 queue_for_each_hw_ctx(q
, hctx
, i
)
1215 blk_mq_stop_hw_queue(hctx
);
1217 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1219 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1221 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1223 blk_mq_run_hw_queue(hctx
, false);
1225 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1227 void blk_mq_start_hw_queues(struct request_queue
*q
)
1229 struct blk_mq_hw_ctx
*hctx
;
1232 queue_for_each_hw_ctx(q
, hctx
, i
)
1233 blk_mq_start_hw_queue(hctx
);
1235 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1237 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1239 if (!blk_mq_hctx_stopped(hctx
))
1242 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1243 blk_mq_run_hw_queue(hctx
, async
);
1245 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1247 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1249 struct blk_mq_hw_ctx
*hctx
;
1252 queue_for_each_hw_ctx(q
, hctx
, i
)
1253 blk_mq_start_stopped_hw_queue(hctx
, async
);
1255 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1257 static void blk_mq_run_work_fn(struct work_struct
*work
)
1259 struct blk_mq_hw_ctx
*hctx
;
1261 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1263 __blk_mq_run_hw_queue(hctx
);
1266 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1268 struct blk_mq_hw_ctx
*hctx
;
1270 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1272 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1273 __blk_mq_run_hw_queue(hctx
);
1276 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1278 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1281 blk_mq_stop_hw_queue(hctx
);
1282 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1283 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1285 EXPORT_SYMBOL(blk_mq_delay_queue
);
1287 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1291 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1293 trace_block_rq_insert(hctx
->queue
, rq
);
1296 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1298 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1301 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1304 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1306 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1307 blk_mq_hctx_mark_pending(hctx
, ctx
);
1310 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1311 struct list_head
*list
)
1315 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1318 spin_lock(&ctx
->lock
);
1319 while (!list_empty(list
)) {
1322 rq
= list_first_entry(list
, struct request
, queuelist
);
1323 BUG_ON(rq
->mq_ctx
!= ctx
);
1324 list_del_init(&rq
->queuelist
);
1325 __blk_mq_insert_req_list(hctx
, rq
, false);
1327 blk_mq_hctx_mark_pending(hctx
, ctx
);
1328 spin_unlock(&ctx
->lock
);
1331 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1333 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1334 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1336 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1337 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1338 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1341 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1343 struct blk_mq_ctx
*this_ctx
;
1344 struct request_queue
*this_q
;
1347 LIST_HEAD(ctx_list
);
1350 list_splice_init(&plug
->mq_list
, &list
);
1352 list_sort(NULL
, &list
, plug_ctx_cmp
);
1358 while (!list_empty(&list
)) {
1359 rq
= list_entry_rq(list
.next
);
1360 list_del_init(&rq
->queuelist
);
1362 if (rq
->mq_ctx
!= this_ctx
) {
1364 trace_block_unplug(this_q
, depth
, from_schedule
);
1365 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1370 this_ctx
= rq
->mq_ctx
;
1376 list_add_tail(&rq
->queuelist
, &ctx_list
);
1380 * If 'this_ctx' is set, we know we have entries to complete
1381 * on 'ctx_list'. Do those.
1384 trace_block_unplug(this_q
, depth
, from_schedule
);
1385 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1390 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1392 init_request_from_bio(rq
, bio
);
1394 blk_account_io_start(rq
, true);
1397 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1399 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1400 !blk_queue_nomerges(hctx
->queue
);
1403 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1404 struct blk_mq_ctx
*ctx
,
1405 struct request
*rq
, struct bio
*bio
)
1407 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1408 blk_mq_bio_to_request(rq
, bio
);
1409 spin_lock(&ctx
->lock
);
1411 __blk_mq_insert_request(hctx
, rq
, false);
1412 spin_unlock(&ctx
->lock
);
1415 struct request_queue
*q
= hctx
->queue
;
1417 spin_lock(&ctx
->lock
);
1418 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1419 blk_mq_bio_to_request(rq
, bio
);
1423 spin_unlock(&ctx
->lock
);
1424 __blk_mq_finish_request(hctx
, ctx
, rq
);
1429 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1432 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1434 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1437 static void blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
,
1440 struct request_queue
*q
= rq
->q
;
1441 struct blk_mq_queue_data bd
= {
1446 struct blk_mq_hw_ctx
*hctx
;
1447 blk_qc_t new_cookie
;
1453 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1456 new_cookie
= request_to_qc_t(hctx
, rq
);
1459 * For OK queue, we are done. For error, kill it. Any other
1460 * error (busy), just add it to our list as we previously
1463 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1464 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1465 *cookie
= new_cookie
;
1469 __blk_mq_requeue_request(rq
);
1471 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1472 *cookie
= BLK_QC_T_NONE
;
1474 blk_mq_end_request(rq
, rq
->errors
);
1479 blk_mq_sched_insert_request(rq
, false, true, false, may_sleep
);
1483 * Multiple hardware queue variant. This will not use per-process plugs,
1484 * but will attempt to bypass the hctx queueing if we can go straight to
1485 * hardware for SYNC IO.
1487 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1489 const int is_sync
= op_is_sync(bio
->bi_opf
);
1490 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1491 struct blk_mq_alloc_data data
= { .flags
= 0 };
1493 unsigned int request_count
= 0, srcu_idx
;
1494 struct blk_plug
*plug
;
1495 struct request
*same_queue_rq
= NULL
;
1497 unsigned int wb_acct
;
1499 blk_queue_bounce(q
, &bio
);
1501 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1503 return BLK_QC_T_NONE
;
1506 blk_queue_split(q
, &bio
, q
->bio_split
);
1508 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1509 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1510 return BLK_QC_T_NONE
;
1512 if (blk_mq_sched_bio_merge(q
, bio
))
1513 return BLK_QC_T_NONE
;
1515 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1517 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1519 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1520 if (unlikely(!rq
)) {
1521 __wbt_done(q
->rq_wb
, wb_acct
);
1522 return BLK_QC_T_NONE
;
1525 wbt_track(&rq
->issue_stat
, wb_acct
);
1527 cookie
= request_to_qc_t(data
.hctx
, rq
);
1529 if (unlikely(is_flush_fua
)) {
1532 blk_mq_bio_to_request(rq
, bio
);
1533 blk_insert_flush(rq
);
1537 plug
= current
->plug
;
1539 * If the driver supports defer issued based on 'last', then
1540 * queue it up like normal since we can potentially save some
1543 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1544 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1545 struct request
*old_rq
= NULL
;
1547 blk_mq_bio_to_request(rq
, bio
);
1550 * We do limited plugging. If the bio can be merged, do that.
1551 * Otherwise the existing request in the plug list will be
1552 * issued. So the plug list will have one request at most
1556 * The plug list might get flushed before this. If that
1557 * happens, same_queue_rq is invalid and plug list is
1560 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1561 old_rq
= same_queue_rq
;
1562 list_del_init(&old_rq
->queuelist
);
1564 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1565 } else /* is_sync */
1567 blk_mq_put_ctx(data
.ctx
);
1571 if (!(data
.hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1573 blk_mq_try_issue_directly(old_rq
, &cookie
, false);
1576 srcu_idx
= srcu_read_lock(&data
.hctx
->queue_rq_srcu
);
1577 blk_mq_try_issue_directly(old_rq
, &cookie
, true);
1578 srcu_read_unlock(&data
.hctx
->queue_rq_srcu
, srcu_idx
);
1585 blk_mq_put_ctx(data
.ctx
);
1586 blk_mq_bio_to_request(rq
, bio
);
1587 blk_mq_sched_insert_request(rq
, false, true,
1588 !is_sync
|| is_flush_fua
, true);
1591 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1593 * For a SYNC request, send it to the hardware immediately. For
1594 * an ASYNC request, just ensure that we run it later on. The
1595 * latter allows for merging opportunities and more efficient
1599 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1601 blk_mq_put_ctx(data
.ctx
);
1607 * Single hardware queue variant. This will attempt to use any per-process
1608 * plug for merging and IO deferral.
1610 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1612 const int is_sync
= op_is_sync(bio
->bi_opf
);
1613 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1614 struct blk_plug
*plug
;
1615 unsigned int request_count
= 0;
1616 struct blk_mq_alloc_data data
= { .flags
= 0 };
1619 unsigned int wb_acct
;
1621 blk_queue_bounce(q
, &bio
);
1623 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1625 return BLK_QC_T_NONE
;
1628 blk_queue_split(q
, &bio
, q
->bio_split
);
1630 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1631 if (blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1632 return BLK_QC_T_NONE
;
1634 request_count
= blk_plug_queued_count(q
);
1636 if (blk_mq_sched_bio_merge(q
, bio
))
1637 return BLK_QC_T_NONE
;
1639 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1641 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1643 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1644 if (unlikely(!rq
)) {
1645 __wbt_done(q
->rq_wb
, wb_acct
);
1646 return BLK_QC_T_NONE
;
1649 wbt_track(&rq
->issue_stat
, wb_acct
);
1651 cookie
= request_to_qc_t(data
.hctx
, rq
);
1653 if (unlikely(is_flush_fua
)) {
1656 blk_mq_bio_to_request(rq
, bio
);
1657 blk_insert_flush(rq
);
1662 * A task plug currently exists. Since this is completely lockless,
1663 * utilize that to temporarily store requests until the task is
1664 * either done or scheduled away.
1666 plug
= current
->plug
;
1668 struct request
*last
= NULL
;
1670 blk_mq_bio_to_request(rq
, bio
);
1673 * @request_count may become stale because of schedule
1674 * out, so check the list again.
1676 if (list_empty(&plug
->mq_list
))
1679 trace_block_plug(q
);
1681 last
= list_entry_rq(plug
->mq_list
.prev
);
1683 blk_mq_put_ctx(data
.ctx
);
1685 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1686 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1687 blk_flush_plug_list(plug
, false);
1688 trace_block_plug(q
);
1691 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1697 blk_mq_put_ctx(data
.ctx
);
1698 blk_mq_bio_to_request(rq
, bio
);
1699 blk_mq_sched_insert_request(rq
, false, true,
1700 !is_sync
|| is_flush_fua
, true);
1703 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1705 * For a SYNC request, send it to the hardware immediately. For
1706 * an ASYNC request, just ensure that we run it later on. The
1707 * latter allows for merging opportunities and more efficient
1711 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1714 blk_mq_put_ctx(data
.ctx
);
1719 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1720 unsigned int hctx_idx
)
1724 if (tags
->rqs
&& set
->ops
->exit_request
) {
1727 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1728 struct request
*rq
= tags
->static_rqs
[i
];
1732 set
->ops
->exit_request(set
->driver_data
, rq
,
1734 tags
->static_rqs
[i
] = NULL
;
1738 while (!list_empty(&tags
->page_list
)) {
1739 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1740 list_del_init(&page
->lru
);
1742 * Remove kmemleak object previously allocated in
1743 * blk_mq_init_rq_map().
1745 kmemleak_free(page_address(page
));
1746 __free_pages(page
, page
->private);
1750 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1754 kfree(tags
->static_rqs
);
1755 tags
->static_rqs
= NULL
;
1757 blk_mq_free_tags(tags
);
1760 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1761 unsigned int hctx_idx
,
1762 unsigned int nr_tags
,
1763 unsigned int reserved_tags
)
1765 struct blk_mq_tags
*tags
;
1768 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1769 if (node
== NUMA_NO_NODE
)
1770 node
= set
->numa_node
;
1772 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1773 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1777 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1778 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1781 blk_mq_free_tags(tags
);
1785 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1786 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1788 if (!tags
->static_rqs
) {
1790 blk_mq_free_tags(tags
);
1797 static size_t order_to_size(unsigned int order
)
1799 return (size_t)PAGE_SIZE
<< order
;
1802 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1803 unsigned int hctx_idx
, unsigned int depth
)
1805 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1806 size_t rq_size
, left
;
1809 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1810 if (node
== NUMA_NO_NODE
)
1811 node
= set
->numa_node
;
1813 INIT_LIST_HEAD(&tags
->page_list
);
1816 * rq_size is the size of the request plus driver payload, rounded
1817 * to the cacheline size
1819 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1821 left
= rq_size
* depth
;
1823 for (i
= 0; i
< depth
; ) {
1824 int this_order
= max_order
;
1829 while (this_order
&& left
< order_to_size(this_order
- 1))
1833 page
= alloc_pages_node(node
,
1834 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1840 if (order_to_size(this_order
) < rq_size
)
1847 page
->private = this_order
;
1848 list_add_tail(&page
->lru
, &tags
->page_list
);
1850 p
= page_address(page
);
1852 * Allow kmemleak to scan these pages as they contain pointers
1853 * to additional allocations like via ops->init_request().
1855 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1856 entries_per_page
= order_to_size(this_order
) / rq_size
;
1857 to_do
= min(entries_per_page
, depth
- i
);
1858 left
-= to_do
* rq_size
;
1859 for (j
= 0; j
< to_do
; j
++) {
1860 struct request
*rq
= p
;
1862 tags
->static_rqs
[i
] = rq
;
1863 if (set
->ops
->init_request
) {
1864 if (set
->ops
->init_request(set
->driver_data
,
1867 tags
->static_rqs
[i
] = NULL
;
1879 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1884 * 'cpu' is going away. splice any existing rq_list entries from this
1885 * software queue to the hw queue dispatch list, and ensure that it
1888 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1890 struct blk_mq_hw_ctx
*hctx
;
1891 struct blk_mq_ctx
*ctx
;
1894 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1895 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1897 spin_lock(&ctx
->lock
);
1898 if (!list_empty(&ctx
->rq_list
)) {
1899 list_splice_init(&ctx
->rq_list
, &tmp
);
1900 blk_mq_hctx_clear_pending(hctx
, ctx
);
1902 spin_unlock(&ctx
->lock
);
1904 if (list_empty(&tmp
))
1907 spin_lock(&hctx
->lock
);
1908 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1909 spin_unlock(&hctx
->lock
);
1911 blk_mq_run_hw_queue(hctx
, true);
1915 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1917 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1921 /* hctx->ctxs will be freed in queue's release handler */
1922 static void blk_mq_exit_hctx(struct request_queue
*q
,
1923 struct blk_mq_tag_set
*set
,
1924 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1926 unsigned flush_start_tag
= set
->queue_depth
;
1928 blk_mq_tag_idle(hctx
);
1930 if (set
->ops
->exit_request
)
1931 set
->ops
->exit_request(set
->driver_data
,
1932 hctx
->fq
->flush_rq
, hctx_idx
,
1933 flush_start_tag
+ hctx_idx
);
1935 if (set
->ops
->exit_hctx
)
1936 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1938 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1939 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1941 blk_mq_remove_cpuhp(hctx
);
1942 blk_free_flush_queue(hctx
->fq
);
1943 sbitmap_free(&hctx
->ctx_map
);
1946 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1947 struct blk_mq_tag_set
*set
, int nr_queue
)
1949 struct blk_mq_hw_ctx
*hctx
;
1952 queue_for_each_hw_ctx(q
, hctx
, i
) {
1955 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1959 static int blk_mq_init_hctx(struct request_queue
*q
,
1960 struct blk_mq_tag_set
*set
,
1961 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1964 unsigned flush_start_tag
= set
->queue_depth
;
1966 node
= hctx
->numa_node
;
1967 if (node
== NUMA_NO_NODE
)
1968 node
= hctx
->numa_node
= set
->numa_node
;
1970 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1971 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1972 spin_lock_init(&hctx
->lock
);
1973 INIT_LIST_HEAD(&hctx
->dispatch
);
1975 hctx
->queue_num
= hctx_idx
;
1976 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1978 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1980 hctx
->tags
= set
->tags
[hctx_idx
];
1983 * Allocate space for all possible cpus to avoid allocation at
1986 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1989 goto unregister_cpu_notifier
;
1991 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1997 if (set
->ops
->init_hctx
&&
1998 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2001 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
2005 if (set
->ops
->init_request
&&
2006 set
->ops
->init_request(set
->driver_data
,
2007 hctx
->fq
->flush_rq
, hctx_idx
,
2008 flush_start_tag
+ hctx_idx
, node
))
2011 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2012 init_srcu_struct(&hctx
->queue_rq_srcu
);
2019 if (set
->ops
->exit_hctx
)
2020 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2022 sbitmap_free(&hctx
->ctx_map
);
2025 unregister_cpu_notifier
:
2026 blk_mq_remove_cpuhp(hctx
);
2030 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2031 unsigned int nr_hw_queues
)
2035 for_each_possible_cpu(i
) {
2036 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2037 struct blk_mq_hw_ctx
*hctx
;
2040 spin_lock_init(&__ctx
->lock
);
2041 INIT_LIST_HEAD(&__ctx
->rq_list
);
2043 blk_stat_init(&__ctx
->stat
[BLK_STAT_READ
]);
2044 blk_stat_init(&__ctx
->stat
[BLK_STAT_WRITE
]);
2046 /* If the cpu isn't online, the cpu is mapped to first hctx */
2050 hctx
= blk_mq_map_queue(q
, i
);
2053 * Set local node, IFF we have more than one hw queue. If
2054 * not, we remain on the home node of the device
2056 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2057 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2061 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2065 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2066 set
->queue_depth
, set
->reserved_tags
);
2067 if (!set
->tags
[hctx_idx
])
2070 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2075 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2076 set
->tags
[hctx_idx
] = NULL
;
2080 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2081 unsigned int hctx_idx
)
2083 if (set
->tags
[hctx_idx
]) {
2084 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2085 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2086 set
->tags
[hctx_idx
] = NULL
;
2090 static void blk_mq_map_swqueue(struct request_queue
*q
,
2091 const struct cpumask
*online_mask
)
2093 unsigned int i
, hctx_idx
;
2094 struct blk_mq_hw_ctx
*hctx
;
2095 struct blk_mq_ctx
*ctx
;
2096 struct blk_mq_tag_set
*set
= q
->tag_set
;
2099 * Avoid others reading imcomplete hctx->cpumask through sysfs
2101 mutex_lock(&q
->sysfs_lock
);
2103 queue_for_each_hw_ctx(q
, hctx
, i
) {
2104 cpumask_clear(hctx
->cpumask
);
2109 * Map software to hardware queues
2111 for_each_possible_cpu(i
) {
2112 /* If the cpu isn't online, the cpu is mapped to first hctx */
2113 if (!cpumask_test_cpu(i
, online_mask
))
2116 hctx_idx
= q
->mq_map
[i
];
2117 /* unmapped hw queue can be remapped after CPU topo changed */
2118 if (!set
->tags
[hctx_idx
] &&
2119 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2121 * If tags initialization fail for some hctx,
2122 * that hctx won't be brought online. In this
2123 * case, remap the current ctx to hctx[0] which
2124 * is guaranteed to always have tags allocated
2129 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2130 hctx
= blk_mq_map_queue(q
, i
);
2132 cpumask_set_cpu(i
, hctx
->cpumask
);
2133 ctx
->index_hw
= hctx
->nr_ctx
;
2134 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2137 mutex_unlock(&q
->sysfs_lock
);
2139 queue_for_each_hw_ctx(q
, hctx
, i
) {
2141 * If no software queues are mapped to this hardware queue,
2142 * disable it and free the request entries.
2144 if (!hctx
->nr_ctx
) {
2145 /* Never unmap queue 0. We need it as a
2146 * fallback in case of a new remap fails
2149 if (i
&& set
->tags
[i
])
2150 blk_mq_free_map_and_requests(set
, i
);
2156 hctx
->tags
= set
->tags
[i
];
2157 WARN_ON(!hctx
->tags
);
2160 * Set the map size to the number of mapped software queues.
2161 * This is more accurate and more efficient than looping
2162 * over all possibly mapped software queues.
2164 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2167 * Initialize batch roundrobin counts
2169 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2170 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2174 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2176 struct blk_mq_hw_ctx
*hctx
;
2179 queue_for_each_hw_ctx(q
, hctx
, i
) {
2181 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2183 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2187 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
2189 struct request_queue
*q
;
2191 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2192 blk_mq_freeze_queue(q
);
2193 queue_set_hctx_shared(q
, shared
);
2194 blk_mq_unfreeze_queue(q
);
2198 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2200 struct blk_mq_tag_set
*set
= q
->tag_set
;
2202 mutex_lock(&set
->tag_list_lock
);
2203 list_del_init(&q
->tag_set_list
);
2204 if (list_is_singular(&set
->tag_list
)) {
2205 /* just transitioned to unshared */
2206 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2207 /* update existing queue */
2208 blk_mq_update_tag_set_depth(set
, false);
2210 mutex_unlock(&set
->tag_list_lock
);
2213 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2214 struct request_queue
*q
)
2218 mutex_lock(&set
->tag_list_lock
);
2220 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2221 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2222 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2223 /* update existing queue */
2224 blk_mq_update_tag_set_depth(set
, true);
2226 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2227 queue_set_hctx_shared(q
, true);
2228 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
2230 mutex_unlock(&set
->tag_list_lock
);
2234 * It is the actual release handler for mq, but we do it from
2235 * request queue's release handler for avoiding use-after-free
2236 * and headache because q->mq_kobj shouldn't have been introduced,
2237 * but we can't group ctx/kctx kobj without it.
2239 void blk_mq_release(struct request_queue
*q
)
2241 struct blk_mq_hw_ctx
*hctx
;
2244 blk_mq_sched_teardown(q
);
2246 /* hctx kobj stays in hctx */
2247 queue_for_each_hw_ctx(q
, hctx
, i
) {
2250 kobject_put(&hctx
->kobj
);
2255 kfree(q
->queue_hw_ctx
);
2258 * release .mq_kobj and sw queue's kobject now because
2259 * both share lifetime with request queue.
2261 blk_mq_sysfs_deinit(q
);
2263 free_percpu(q
->queue_ctx
);
2266 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2268 struct request_queue
*uninit_q
, *q
;
2270 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2272 return ERR_PTR(-ENOMEM
);
2274 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2276 blk_cleanup_queue(uninit_q
);
2280 EXPORT_SYMBOL(blk_mq_init_queue
);
2282 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2283 struct request_queue
*q
)
2286 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2288 blk_mq_sysfs_unregister(q
);
2289 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2295 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2296 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2301 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2308 atomic_set(&hctxs
[i
]->nr_active
, 0);
2309 hctxs
[i
]->numa_node
= node
;
2310 hctxs
[i
]->queue_num
= i
;
2312 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2313 free_cpumask_var(hctxs
[i
]->cpumask
);
2318 blk_mq_hctx_kobj_init(hctxs
[i
]);
2320 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2321 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2325 blk_mq_free_map_and_requests(set
, j
);
2326 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2327 kobject_put(&hctx
->kobj
);
2332 q
->nr_hw_queues
= i
;
2333 blk_mq_sysfs_register(q
);
2336 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2337 struct request_queue
*q
)
2339 /* mark the queue as mq asap */
2340 q
->mq_ops
= set
->ops
;
2342 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2346 /* init q->mq_kobj and sw queues' kobjects */
2347 blk_mq_sysfs_init(q
);
2349 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2350 GFP_KERNEL
, set
->numa_node
);
2351 if (!q
->queue_hw_ctx
)
2354 q
->mq_map
= set
->mq_map
;
2356 blk_mq_realloc_hw_ctxs(set
, q
);
2357 if (!q
->nr_hw_queues
)
2360 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2361 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2363 q
->nr_queues
= nr_cpu_ids
;
2365 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2367 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2368 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2370 q
->sg_reserved_size
= INT_MAX
;
2372 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2373 INIT_LIST_HEAD(&q
->requeue_list
);
2374 spin_lock_init(&q
->requeue_lock
);
2376 if (q
->nr_hw_queues
> 1)
2377 blk_queue_make_request(q
, blk_mq_make_request
);
2379 blk_queue_make_request(q
, blk_sq_make_request
);
2382 * Do this after blk_queue_make_request() overrides it...
2384 q
->nr_requests
= set
->queue_depth
;
2387 * Default to classic polling
2391 if (set
->ops
->complete
)
2392 blk_queue_softirq_done(q
, set
->ops
->complete
);
2394 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2397 mutex_lock(&all_q_mutex
);
2399 list_add_tail(&q
->all_q_node
, &all_q_list
);
2400 blk_mq_add_queue_tag_set(set
, q
);
2401 blk_mq_map_swqueue(q
, cpu_online_mask
);
2403 mutex_unlock(&all_q_mutex
);
2406 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2409 ret
= blk_mq_sched_init(q
);
2411 return ERR_PTR(ret
);
2417 kfree(q
->queue_hw_ctx
);
2419 free_percpu(q
->queue_ctx
);
2422 return ERR_PTR(-ENOMEM
);
2424 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2426 void blk_mq_free_queue(struct request_queue
*q
)
2428 struct blk_mq_tag_set
*set
= q
->tag_set
;
2430 mutex_lock(&all_q_mutex
);
2431 list_del_init(&q
->all_q_node
);
2432 mutex_unlock(&all_q_mutex
);
2436 blk_mq_del_queue_tag_set(q
);
2438 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2441 /* Basically redo blk_mq_init_queue with queue frozen */
2442 static void blk_mq_queue_reinit(struct request_queue
*q
,
2443 const struct cpumask
*online_mask
)
2445 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2447 blk_mq_sysfs_unregister(q
);
2450 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2451 * we should change hctx numa_node according to new topology (this
2452 * involves free and re-allocate memory, worthy doing?)
2455 blk_mq_map_swqueue(q
, online_mask
);
2457 blk_mq_sysfs_register(q
);
2461 * New online cpumask which is going to be set in this hotplug event.
2462 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2463 * one-by-one and dynamically allocating this could result in a failure.
2465 static struct cpumask cpuhp_online_new
;
2467 static void blk_mq_queue_reinit_work(void)
2469 struct request_queue
*q
;
2471 mutex_lock(&all_q_mutex
);
2473 * We need to freeze and reinit all existing queues. Freezing
2474 * involves synchronous wait for an RCU grace period and doing it
2475 * one by one may take a long time. Start freezing all queues in
2476 * one swoop and then wait for the completions so that freezing can
2477 * take place in parallel.
2479 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2480 blk_mq_freeze_queue_start(q
);
2481 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2482 blk_mq_freeze_queue_wait(q
);
2484 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2485 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2487 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2488 blk_mq_unfreeze_queue(q
);
2490 mutex_unlock(&all_q_mutex
);
2493 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2495 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2496 blk_mq_queue_reinit_work();
2501 * Before hotadded cpu starts handling requests, new mappings must be
2502 * established. Otherwise, these requests in hw queue might never be
2505 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2506 * for CPU0, and ctx1 for CPU1).
2508 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2509 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2511 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2512 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2513 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2516 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2518 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2519 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2520 blk_mq_queue_reinit_work();
2524 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2528 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2529 if (!__blk_mq_alloc_rq_map(set
, i
))
2536 blk_mq_free_rq_map(set
->tags
[i
]);
2542 * Allocate the request maps associated with this tag_set. Note that this
2543 * may reduce the depth asked for, if memory is tight. set->queue_depth
2544 * will be updated to reflect the allocated depth.
2546 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2551 depth
= set
->queue_depth
;
2553 err
= __blk_mq_alloc_rq_maps(set
);
2557 set
->queue_depth
>>= 1;
2558 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2562 } while (set
->queue_depth
);
2564 if (!set
->queue_depth
|| err
) {
2565 pr_err("blk-mq: failed to allocate request map\n");
2569 if (depth
!= set
->queue_depth
)
2570 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2571 depth
, set
->queue_depth
);
2577 * Alloc a tag set to be associated with one or more request queues.
2578 * May fail with EINVAL for various error conditions. May adjust the
2579 * requested depth down, if if it too large. In that case, the set
2580 * value will be stored in set->queue_depth.
2582 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2586 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2588 if (!set
->nr_hw_queues
)
2590 if (!set
->queue_depth
)
2592 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2595 if (!set
->ops
->queue_rq
)
2598 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2599 pr_info("blk-mq: reduced tag depth to %u\n",
2601 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2605 * If a crashdump is active, then we are potentially in a very
2606 * memory constrained environment. Limit us to 1 queue and
2607 * 64 tags to prevent using too much memory.
2609 if (is_kdump_kernel()) {
2610 set
->nr_hw_queues
= 1;
2611 set
->queue_depth
= min(64U, set
->queue_depth
);
2614 * There is no use for more h/w queues than cpus.
2616 if (set
->nr_hw_queues
> nr_cpu_ids
)
2617 set
->nr_hw_queues
= nr_cpu_ids
;
2619 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2620 GFP_KERNEL
, set
->numa_node
);
2625 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2626 GFP_KERNEL
, set
->numa_node
);
2630 if (set
->ops
->map_queues
)
2631 ret
= set
->ops
->map_queues(set
);
2633 ret
= blk_mq_map_queues(set
);
2635 goto out_free_mq_map
;
2637 ret
= blk_mq_alloc_rq_maps(set
);
2639 goto out_free_mq_map
;
2641 mutex_init(&set
->tag_list_lock
);
2642 INIT_LIST_HEAD(&set
->tag_list
);
2654 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2656 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2660 for (i
= 0; i
< nr_cpu_ids
; i
++)
2661 blk_mq_free_map_and_requests(set
, i
);
2669 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2671 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2673 struct blk_mq_tag_set
*set
= q
->tag_set
;
2674 struct blk_mq_hw_ctx
*hctx
;
2680 blk_mq_freeze_queue(q
);
2681 blk_mq_quiesce_queue(q
);
2684 queue_for_each_hw_ctx(q
, hctx
, i
) {
2688 * If we're using an MQ scheduler, just update the scheduler
2689 * queue depth. This is similar to what the old code would do.
2691 if (!hctx
->sched_tags
) {
2692 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2693 min(nr
, set
->queue_depth
),
2696 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2704 q
->nr_requests
= nr
;
2706 blk_mq_unfreeze_queue(q
);
2707 blk_mq_start_stopped_hw_queues(q
, true);
2712 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2714 struct request_queue
*q
;
2716 if (nr_hw_queues
> nr_cpu_ids
)
2717 nr_hw_queues
= nr_cpu_ids
;
2718 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2721 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2722 blk_mq_freeze_queue(q
);
2724 set
->nr_hw_queues
= nr_hw_queues
;
2725 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2726 blk_mq_realloc_hw_ctxs(set
, q
);
2729 * Manually set the make_request_fn as blk_queue_make_request
2730 * resets a lot of the queue settings.
2732 if (q
->nr_hw_queues
> 1)
2733 q
->make_request_fn
= blk_mq_make_request
;
2735 q
->make_request_fn
= blk_sq_make_request
;
2737 blk_mq_queue_reinit(q
, cpu_online_mask
);
2740 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2741 blk_mq_unfreeze_queue(q
);
2743 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2745 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2746 struct blk_mq_hw_ctx
*hctx
,
2749 struct blk_rq_stat stat
[2];
2750 unsigned long ret
= 0;
2753 * If stats collection isn't on, don't sleep but turn it on for
2756 if (!blk_stat_enable(q
))
2760 * We don't have to do this once per IO, should optimize this
2761 * to just use the current window of stats until it changes
2763 memset(&stat
, 0, sizeof(stat
));
2764 blk_hctx_stat_get(hctx
, stat
);
2767 * As an optimistic guess, use half of the mean service time
2768 * for this type of request. We can (and should) make this smarter.
2769 * For instance, if the completion latencies are tight, we can
2770 * get closer than just half the mean. This is especially
2771 * important on devices where the completion latencies are longer
2774 if (req_op(rq
) == REQ_OP_READ
&& stat
[BLK_STAT_READ
].nr_samples
)
2775 ret
= (stat
[BLK_STAT_READ
].mean
+ 1) / 2;
2776 else if (req_op(rq
) == REQ_OP_WRITE
&& stat
[BLK_STAT_WRITE
].nr_samples
)
2777 ret
= (stat
[BLK_STAT_WRITE
].mean
+ 1) / 2;
2782 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2783 struct blk_mq_hw_ctx
*hctx
,
2786 struct hrtimer_sleeper hs
;
2787 enum hrtimer_mode mode
;
2791 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2797 * -1: don't ever hybrid sleep
2798 * 0: use half of prev avg
2799 * >0: use this specific value
2801 if (q
->poll_nsec
== -1)
2803 else if (q
->poll_nsec
> 0)
2804 nsecs
= q
->poll_nsec
;
2806 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2811 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2814 * This will be replaced with the stats tracking code, using
2815 * 'avg_completion_time / 2' as the pre-sleep target.
2819 mode
= HRTIMER_MODE_REL
;
2820 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2821 hrtimer_set_expires(&hs
.timer
, kt
);
2823 hrtimer_init_sleeper(&hs
, current
);
2825 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2827 set_current_state(TASK_UNINTERRUPTIBLE
);
2828 hrtimer_start_expires(&hs
.timer
, mode
);
2831 hrtimer_cancel(&hs
.timer
);
2832 mode
= HRTIMER_MODE_ABS
;
2833 } while (hs
.task
&& !signal_pending(current
));
2835 __set_current_state(TASK_RUNNING
);
2836 destroy_hrtimer_on_stack(&hs
.timer
);
2840 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2842 struct request_queue
*q
= hctx
->queue
;
2846 * If we sleep, have the caller restart the poll loop to reset
2847 * the state. Like for the other success return cases, the
2848 * caller is responsible for checking if the IO completed. If
2849 * the IO isn't complete, we'll get called again and will go
2850 * straight to the busy poll loop.
2852 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2855 hctx
->poll_considered
++;
2857 state
= current
->state
;
2858 while (!need_resched()) {
2861 hctx
->poll_invoked
++;
2863 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2865 hctx
->poll_success
++;
2866 set_current_state(TASK_RUNNING
);
2870 if (signal_pending_state(state
, current
))
2871 set_current_state(TASK_RUNNING
);
2873 if (current
->state
== TASK_RUNNING
)
2883 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2885 struct blk_mq_hw_ctx
*hctx
;
2886 struct blk_plug
*plug
;
2889 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2890 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2893 plug
= current
->plug
;
2895 blk_flush_plug_list(plug
, false);
2897 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2898 if (!blk_qc_t_is_internal(cookie
))
2899 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2901 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2903 return __blk_mq_poll(hctx
, rq
);
2905 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2907 void blk_mq_disable_hotplug(void)
2909 mutex_lock(&all_q_mutex
);
2912 void blk_mq_enable_hotplug(void)
2914 mutex_unlock(&all_q_mutex
);
2917 static int __init
blk_mq_init(void)
2919 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2920 blk_mq_hctx_notify_dead
);
2922 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2923 blk_mq_queue_reinit_prepare
,
2924 blk_mq_queue_reinit_dead
);
2927 subsys_initcall(blk_mq_init
);