2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/delay.h>
24 #include <linux/crash_dump.h>
25 #include <linux/prefetch.h>
27 #include <trace/events/block.h>
29 #include <linux/blk-mq.h>
32 #include "blk-mq-tag.h"
35 #include "blk-mq-sched.h"
37 static DEFINE_MUTEX(all_q_mutex
);
38 static LIST_HEAD(all_q_list
);
41 * Check if any of the ctx's have pending work in this hardware queue
43 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
45 return sbitmap_any_bit_set(&hctx
->ctx_map
) ||
46 !list_empty_careful(&hctx
->dispatch
) ||
47 blk_mq_sched_has_work(hctx
);
51 * Mark this ctx as having pending work in this hardware queue
53 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
54 struct blk_mq_ctx
*ctx
)
56 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
57 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
60 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
61 struct blk_mq_ctx
*ctx
)
63 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
66 void blk_mq_freeze_queue_start(struct request_queue
*q
)
70 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
71 if (freeze_depth
== 1) {
72 percpu_ref_kill(&q
->q_usage_counter
);
73 blk_mq_run_hw_queues(q
, false);
76 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start
);
78 static void blk_mq_freeze_queue_wait(struct request_queue
*q
)
80 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
84 * Guarantee no request is in use, so we can change any data structure of
85 * the queue afterward.
87 void blk_freeze_queue(struct request_queue
*q
)
90 * In the !blk_mq case we are only calling this to kill the
91 * q_usage_counter, otherwise this increases the freeze depth
92 * and waits for it to return to zero. For this reason there is
93 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
94 * exported to drivers as the only user for unfreeze is blk_mq.
96 blk_mq_freeze_queue_start(q
);
97 blk_mq_freeze_queue_wait(q
);
100 void blk_mq_freeze_queue(struct request_queue
*q
)
103 * ...just an alias to keep freeze and unfreeze actions balanced
104 * in the blk_mq_* namespace
108 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
110 void blk_mq_unfreeze_queue(struct request_queue
*q
)
114 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
115 WARN_ON_ONCE(freeze_depth
< 0);
117 percpu_ref_reinit(&q
->q_usage_counter
);
118 wake_up_all(&q
->mq_freeze_wq
);
121 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
124 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
127 * Note: this function does not prevent that the struct request end_io()
128 * callback function is invoked. Additionally, it is not prevented that
129 * new queue_rq() calls occur unless the queue has been stopped first.
131 void blk_mq_quiesce_queue(struct request_queue
*q
)
133 struct blk_mq_hw_ctx
*hctx
;
137 blk_mq_stop_hw_queues(q
);
139 queue_for_each_hw_ctx(q
, hctx
, i
) {
140 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
141 synchronize_srcu(&hctx
->queue_rq_srcu
);
148 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
150 void blk_mq_wake_waiters(struct request_queue
*q
)
152 struct blk_mq_hw_ctx
*hctx
;
155 queue_for_each_hw_ctx(q
, hctx
, i
)
156 if (blk_mq_hw_queue_mapped(hctx
))
157 blk_mq_tag_wakeup_all(hctx
->tags
, true);
160 * If we are called because the queue has now been marked as
161 * dying, we need to ensure that processes currently waiting on
162 * the queue are notified as well.
164 wake_up_all(&q
->mq_freeze_wq
);
167 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
169 return blk_mq_has_free_tags(hctx
->tags
);
171 EXPORT_SYMBOL(blk_mq_can_queue
);
173 void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
174 struct request
*rq
, unsigned int op
)
176 INIT_LIST_HEAD(&rq
->queuelist
);
177 /* csd/requeue_work/fifo_time is initialized before use */
181 if (blk_queue_io_stat(q
))
182 rq
->rq_flags
|= RQF_IO_STAT
;
183 /* do not touch atomic flags, it needs atomic ops against the timer */
185 INIT_HLIST_NODE(&rq
->hash
);
186 RB_CLEAR_NODE(&rq
->rb_node
);
189 rq
->start_time
= jiffies
;
190 #ifdef CONFIG_BLK_CGROUP
192 set_start_time_ns(rq
);
193 rq
->io_start_time_ns
= 0;
195 rq
->nr_phys_segments
= 0;
196 #if defined(CONFIG_BLK_DEV_INTEGRITY)
197 rq
->nr_integrity_segments
= 0;
200 /* tag was already set */
210 INIT_LIST_HEAD(&rq
->timeout_list
);
214 rq
->end_io_data
= NULL
;
217 ctx
->rq_dispatched
[op_is_sync(op
)]++;
219 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init
);
221 struct request
*__blk_mq_alloc_request(struct blk_mq_alloc_data
*data
,
227 tag
= blk_mq_get_tag(data
);
228 if (tag
!= BLK_MQ_TAG_FAIL
) {
229 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
231 rq
= tags
->static_rqs
[tag
];
233 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
235 rq
->internal_tag
= tag
;
237 if (blk_mq_tag_busy(data
->hctx
)) {
238 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
239 atomic_inc(&data
->hctx
->nr_active
);
242 rq
->internal_tag
= -1;
245 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
);
251 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request
);
253 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
256 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
260 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
264 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
266 blk_mq_put_ctx(alloc_data
.ctx
);
270 return ERR_PTR(-EWOULDBLOCK
);
273 rq
->__sector
= (sector_t
) -1;
274 rq
->bio
= rq
->biotail
= NULL
;
277 EXPORT_SYMBOL(blk_mq_alloc_request
);
279 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
280 unsigned int flags
, unsigned int hctx_idx
)
282 struct blk_mq_hw_ctx
*hctx
;
283 struct blk_mq_ctx
*ctx
;
285 struct blk_mq_alloc_data alloc_data
;
289 * If the tag allocator sleeps we could get an allocation for a
290 * different hardware context. No need to complicate the low level
291 * allocator for this for the rare use case of a command tied to
294 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
295 return ERR_PTR(-EINVAL
);
297 if (hctx_idx
>= q
->nr_hw_queues
)
298 return ERR_PTR(-EIO
);
300 ret
= blk_queue_enter(q
, true);
305 * Check if the hardware context is actually mapped to anything.
306 * If not tell the caller that it should skip this queue.
308 hctx
= q
->queue_hw_ctx
[hctx_idx
];
309 if (!blk_mq_hw_queue_mapped(hctx
)) {
313 ctx
= __blk_mq_get_ctx(q
, cpumask_first(hctx
->cpumask
));
315 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
316 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
328 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
330 void __blk_mq_finish_request(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
333 const int sched_tag
= rq
->internal_tag
;
334 struct request_queue
*q
= rq
->q
;
336 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
337 atomic_dec(&hctx
->nr_active
);
339 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
342 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
343 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
345 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
347 blk_mq_sched_completed_request(hctx
, rq
);
348 blk_mq_sched_restart_queues(hctx
);
352 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx
*hctx
,
355 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
357 ctx
->rq_completed
[rq_is_sync(rq
)]++;
358 __blk_mq_finish_request(hctx
, ctx
, rq
);
361 void blk_mq_finish_request(struct request
*rq
)
363 blk_mq_finish_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
366 void blk_mq_free_request(struct request
*rq
)
368 blk_mq_sched_put_request(rq
);
370 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
372 inline void __blk_mq_end_request(struct request
*rq
, int error
)
374 blk_account_io_done(rq
);
377 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
378 rq
->end_io(rq
, error
);
380 if (unlikely(blk_bidi_rq(rq
)))
381 blk_mq_free_request(rq
->next_rq
);
382 blk_mq_free_request(rq
);
385 EXPORT_SYMBOL(__blk_mq_end_request
);
387 void blk_mq_end_request(struct request
*rq
, int error
)
389 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
391 __blk_mq_end_request(rq
, error
);
393 EXPORT_SYMBOL(blk_mq_end_request
);
395 static void __blk_mq_complete_request_remote(void *data
)
397 struct request
*rq
= data
;
399 rq
->q
->softirq_done_fn(rq
);
402 static void blk_mq_ipi_complete_request(struct request
*rq
)
404 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
408 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
409 rq
->q
->softirq_done_fn(rq
);
414 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
415 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
417 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
418 rq
->csd
.func
= __blk_mq_complete_request_remote
;
421 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
423 rq
->q
->softirq_done_fn(rq
);
428 static void blk_mq_stat_add(struct request
*rq
)
430 if (rq
->rq_flags
& RQF_STATS
) {
432 * We could rq->mq_ctx here, but there's less of a risk
433 * of races if we have the completion event add the stats
434 * to the local software queue.
436 struct blk_mq_ctx
*ctx
;
438 ctx
= __blk_mq_get_ctx(rq
->q
, raw_smp_processor_id());
439 blk_stat_add(&ctx
->stat
[rq_data_dir(rq
)], rq
);
443 static void __blk_mq_complete_request(struct request
*rq
)
445 struct request_queue
*q
= rq
->q
;
449 if (!q
->softirq_done_fn
)
450 blk_mq_end_request(rq
, rq
->errors
);
452 blk_mq_ipi_complete_request(rq
);
456 * blk_mq_complete_request - end I/O on a request
457 * @rq: the request being processed
460 * Ends all I/O on a request. It does not handle partial completions.
461 * The actual completion happens out-of-order, through a IPI handler.
463 void blk_mq_complete_request(struct request
*rq
, int error
)
465 struct request_queue
*q
= rq
->q
;
467 if (unlikely(blk_should_fake_timeout(q
)))
469 if (!blk_mark_rq_complete(rq
)) {
471 __blk_mq_complete_request(rq
);
474 EXPORT_SYMBOL(blk_mq_complete_request
);
476 int blk_mq_request_started(struct request
*rq
)
478 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
480 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
482 void blk_mq_start_request(struct request
*rq
)
484 struct request_queue
*q
= rq
->q
;
486 blk_mq_sched_started_request(rq
);
488 trace_block_rq_issue(q
, rq
);
490 rq
->resid_len
= blk_rq_bytes(rq
);
491 if (unlikely(blk_bidi_rq(rq
)))
492 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_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 el_ret
= blk_try_merge(rq
, bio
);
785 if (el_ret
== ELEVATOR_NO_MERGE
)
788 if (!blk_mq_sched_allow_merge(q
, rq
, bio
))
791 if (el_ret
== ELEVATOR_BACK_MERGE
) {
792 if (bio_attempt_back_merge(q
, rq
, bio
)) {
797 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
798 if (bio_attempt_front_merge(q
, rq
, bio
)) {
809 struct flush_busy_ctx_data
{
810 struct blk_mq_hw_ctx
*hctx
;
811 struct list_head
*list
;
814 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
816 struct flush_busy_ctx_data
*flush_data
= data
;
817 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
818 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
820 sbitmap_clear_bit(sb
, bitnr
);
821 spin_lock(&ctx
->lock
);
822 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
823 spin_unlock(&ctx
->lock
);
828 * Process software queues that have been marked busy, splicing them
829 * to the for-dispatch
831 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
833 struct flush_busy_ctx_data data
= {
838 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
840 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
842 static inline unsigned int queued_to_index(unsigned int queued
)
847 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
850 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
853 struct blk_mq_alloc_data data
= {
855 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
856 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
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 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
885 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
888 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
889 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
890 atomic_dec(&hctx
->nr_active
);
895 * If we fail getting a driver tag because all the driver tags are already
896 * assigned and on the dispatch list, BUT the first entry does not have a
897 * tag, then we could deadlock. For that case, move entries with assigned
898 * driver tags to the front, leaving the set of tagged requests in the
899 * same order, and the untagged set in the same order.
901 static bool reorder_tags_to_front(struct list_head
*list
)
903 struct request
*rq
, *tmp
, *first
= NULL
;
905 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
909 list_move(&rq
->queuelist
, list
);
915 return first
!= NULL
;
918 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
920 struct request_queue
*q
= hctx
->queue
;
922 LIST_HEAD(driver_list
);
923 struct list_head
*dptr
;
924 int queued
, ret
= BLK_MQ_RQ_QUEUE_OK
;
927 * Start off with dptr being NULL, so we start the first request
928 * immediately, even if we have more pending.
933 * Now process all the entries, sending them to the driver.
936 while (!list_empty(list
)) {
937 struct blk_mq_queue_data bd
;
939 rq
= list_first_entry(list
, struct request
, queuelist
);
940 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
941 if (!queued
&& reorder_tags_to_front(list
))
945 * We failed getting a driver tag. Mark the queue(s)
946 * as needing a restart. Retry getting a tag again,
947 * in case the needed IO completed right before we
948 * marked the queue as needing a restart.
950 blk_mq_sched_mark_restart(hctx
);
951 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
954 list_del_init(&rq
->queuelist
);
958 bd
.last
= list_empty(list
);
960 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
962 case BLK_MQ_RQ_QUEUE_OK
:
965 case BLK_MQ_RQ_QUEUE_BUSY
:
966 blk_mq_put_driver_tag(hctx
, rq
);
967 list_add(&rq
->queuelist
, list
);
968 __blk_mq_requeue_request(rq
);
971 pr_err("blk-mq: bad return on queue: %d\n", ret
);
972 case BLK_MQ_RQ_QUEUE_ERROR
:
974 blk_mq_end_request(rq
, rq
->errors
);
978 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
982 * We've done the first request. If we have more than 1
983 * left in the list, set dptr to defer issue.
985 if (!dptr
&& list
->next
!= list
->prev
)
989 hctx
->dispatched
[queued_to_index(queued
)]++;
992 * Any items that need requeuing? Stuff them into hctx->dispatch,
993 * that is where we will continue on next queue run.
995 if (!list_empty(list
)) {
996 spin_lock(&hctx
->lock
);
997 list_splice_init(list
, &hctx
->dispatch
);
998 spin_unlock(&hctx
->lock
);
1001 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
1002 * it's possible the queue is stopped and restarted again
1003 * before this. Queue restart will dispatch requests. And since
1004 * requests in rq_list aren't added into hctx->dispatch yet,
1005 * the requests in rq_list might get lost.
1007 * blk_mq_run_hw_queue() already checks the STOPPED bit
1009 * If RESTART is set, then let completion restart the queue
1010 * instead of potentially looping here.
1012 if (!blk_mq_sched_needs_restart(hctx
))
1013 blk_mq_run_hw_queue(hctx
, true);
1016 return ret
!= BLK_MQ_RQ_QUEUE_BUSY
;
1019 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1023 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1024 cpu_online(hctx
->next_cpu
));
1026 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1028 blk_mq_sched_dispatch_requests(hctx
);
1031 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1032 blk_mq_sched_dispatch_requests(hctx
);
1033 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1038 * It'd be great if the workqueue API had a way to pass
1039 * in a mask and had some smarts for more clever placement.
1040 * For now we just round-robin here, switching for every
1041 * BLK_MQ_CPU_WORK_BATCH queued items.
1043 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1045 if (hctx
->queue
->nr_hw_queues
== 1)
1046 return WORK_CPU_UNBOUND
;
1048 if (--hctx
->next_cpu_batch
<= 0) {
1051 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1052 if (next_cpu
>= nr_cpu_ids
)
1053 next_cpu
= cpumask_first(hctx
->cpumask
);
1055 hctx
->next_cpu
= next_cpu
;
1056 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1059 return hctx
->next_cpu
;
1062 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1064 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
1065 !blk_mq_hw_queue_mapped(hctx
)))
1068 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1069 int cpu
= get_cpu();
1070 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1071 __blk_mq_run_hw_queue(hctx
);
1079 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
);
1082 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1084 struct blk_mq_hw_ctx
*hctx
;
1087 queue_for_each_hw_ctx(q
, hctx
, i
) {
1088 if (!blk_mq_hctx_has_pending(hctx
) ||
1089 blk_mq_hctx_stopped(hctx
))
1092 blk_mq_run_hw_queue(hctx
, async
);
1095 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1098 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1099 * @q: request queue.
1101 * The caller is responsible for serializing this function against
1102 * blk_mq_{start,stop}_hw_queue().
1104 bool blk_mq_queue_stopped(struct request_queue
*q
)
1106 struct blk_mq_hw_ctx
*hctx
;
1109 queue_for_each_hw_ctx(q
, hctx
, i
)
1110 if (blk_mq_hctx_stopped(hctx
))
1115 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1117 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1119 cancel_work(&hctx
->run_work
);
1120 cancel_delayed_work(&hctx
->delay_work
);
1121 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1123 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1125 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1127 struct blk_mq_hw_ctx
*hctx
;
1130 queue_for_each_hw_ctx(q
, hctx
, i
)
1131 blk_mq_stop_hw_queue(hctx
);
1133 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1135 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1137 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1139 blk_mq_run_hw_queue(hctx
, false);
1141 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1143 void blk_mq_start_hw_queues(struct request_queue
*q
)
1145 struct blk_mq_hw_ctx
*hctx
;
1148 queue_for_each_hw_ctx(q
, hctx
, i
)
1149 blk_mq_start_hw_queue(hctx
);
1151 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1153 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1155 if (!blk_mq_hctx_stopped(hctx
))
1158 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1159 blk_mq_run_hw_queue(hctx
, async
);
1161 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1163 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1165 struct blk_mq_hw_ctx
*hctx
;
1168 queue_for_each_hw_ctx(q
, hctx
, i
)
1169 blk_mq_start_stopped_hw_queue(hctx
, async
);
1171 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1173 static void blk_mq_run_work_fn(struct work_struct
*work
)
1175 struct blk_mq_hw_ctx
*hctx
;
1177 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1179 __blk_mq_run_hw_queue(hctx
);
1182 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1184 struct blk_mq_hw_ctx
*hctx
;
1186 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1188 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1189 __blk_mq_run_hw_queue(hctx
);
1192 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1194 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1197 blk_mq_stop_hw_queue(hctx
);
1198 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1199 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1201 EXPORT_SYMBOL(blk_mq_delay_queue
);
1203 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1207 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1209 trace_block_rq_insert(hctx
->queue
, rq
);
1212 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1214 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1217 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1220 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1222 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1223 blk_mq_hctx_mark_pending(hctx
, ctx
);
1226 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1227 struct list_head
*list
)
1231 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1234 spin_lock(&ctx
->lock
);
1235 while (!list_empty(list
)) {
1238 rq
= list_first_entry(list
, struct request
, queuelist
);
1239 BUG_ON(rq
->mq_ctx
!= ctx
);
1240 list_del_init(&rq
->queuelist
);
1241 __blk_mq_insert_req_list(hctx
, rq
, false);
1243 blk_mq_hctx_mark_pending(hctx
, ctx
);
1244 spin_unlock(&ctx
->lock
);
1247 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1249 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1250 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1252 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1253 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1254 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1257 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1259 struct blk_mq_ctx
*this_ctx
;
1260 struct request_queue
*this_q
;
1263 LIST_HEAD(ctx_list
);
1266 list_splice_init(&plug
->mq_list
, &list
);
1268 list_sort(NULL
, &list
, plug_ctx_cmp
);
1274 while (!list_empty(&list
)) {
1275 rq
= list_entry_rq(list
.next
);
1276 list_del_init(&rq
->queuelist
);
1278 if (rq
->mq_ctx
!= this_ctx
) {
1280 trace_block_unplug(this_q
, depth
, from_schedule
);
1281 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1286 this_ctx
= rq
->mq_ctx
;
1292 list_add_tail(&rq
->queuelist
, &ctx_list
);
1296 * If 'this_ctx' is set, we know we have entries to complete
1297 * on 'ctx_list'. Do those.
1300 trace_block_unplug(this_q
, depth
, from_schedule
);
1301 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1306 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1308 init_request_from_bio(rq
, bio
);
1310 blk_account_io_start(rq
, true);
1313 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1315 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1316 !blk_queue_nomerges(hctx
->queue
);
1319 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1320 struct blk_mq_ctx
*ctx
,
1321 struct request
*rq
, struct bio
*bio
)
1323 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1324 blk_mq_bio_to_request(rq
, bio
);
1325 spin_lock(&ctx
->lock
);
1327 __blk_mq_insert_request(hctx
, rq
, false);
1328 spin_unlock(&ctx
->lock
);
1331 struct request_queue
*q
= hctx
->queue
;
1333 spin_lock(&ctx
->lock
);
1334 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1335 blk_mq_bio_to_request(rq
, bio
);
1339 spin_unlock(&ctx
->lock
);
1340 __blk_mq_finish_request(hctx
, ctx
, rq
);
1345 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1348 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1350 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1353 static void blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
)
1355 struct request_queue
*q
= rq
->q
;
1356 struct blk_mq_queue_data bd
= {
1361 struct blk_mq_hw_ctx
*hctx
;
1362 blk_qc_t new_cookie
;
1368 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1371 new_cookie
= request_to_qc_t(hctx
, rq
);
1374 * For OK queue, we are done. For error, kill it. Any other
1375 * error (busy), just add it to our list as we previously
1378 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1379 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1380 *cookie
= new_cookie
;
1384 __blk_mq_requeue_request(rq
);
1386 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1387 *cookie
= BLK_QC_T_NONE
;
1389 blk_mq_end_request(rq
, rq
->errors
);
1394 blk_mq_sched_insert_request(rq
, false, true, true, false);
1398 * Multiple hardware queue variant. This will not use per-process plugs,
1399 * but will attempt to bypass the hctx queueing if we can go straight to
1400 * hardware for SYNC IO.
1402 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1404 const int is_sync
= op_is_sync(bio
->bi_opf
);
1405 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1406 struct blk_mq_alloc_data data
= { .flags
= 0 };
1408 unsigned int request_count
= 0, srcu_idx
;
1409 struct blk_plug
*plug
;
1410 struct request
*same_queue_rq
= NULL
;
1412 unsigned int wb_acct
;
1414 blk_queue_bounce(q
, &bio
);
1416 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1418 return BLK_QC_T_NONE
;
1421 blk_queue_split(q
, &bio
, q
->bio_split
);
1423 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1424 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1425 return BLK_QC_T_NONE
;
1427 if (blk_mq_sched_bio_merge(q
, bio
))
1428 return BLK_QC_T_NONE
;
1430 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1432 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1434 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1435 if (unlikely(!rq
)) {
1436 __wbt_done(q
->rq_wb
, wb_acct
);
1437 return BLK_QC_T_NONE
;
1440 wbt_track(&rq
->issue_stat
, wb_acct
);
1442 cookie
= request_to_qc_t(data
.hctx
, rq
);
1444 if (unlikely(is_flush_fua
)) {
1445 blk_mq_put_ctx(data
.ctx
);
1446 blk_mq_bio_to_request(rq
, bio
);
1447 blk_mq_get_driver_tag(rq
, NULL
, true);
1448 blk_insert_flush(rq
);
1449 blk_mq_run_hw_queue(data
.hctx
, true);
1453 plug
= current
->plug
;
1455 * If the driver supports defer issued based on 'last', then
1456 * queue it up like normal since we can potentially save some
1459 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1460 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1461 struct request
*old_rq
= NULL
;
1463 blk_mq_bio_to_request(rq
, bio
);
1466 * We do limited plugging. If the bio can be merged, do that.
1467 * Otherwise the existing request in the plug list will be
1468 * issued. So the plug list will have one request at most
1472 * The plug list might get flushed before this. If that
1473 * happens, same_queue_rq is invalid and plug list is
1476 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1477 old_rq
= same_queue_rq
;
1478 list_del_init(&old_rq
->queuelist
);
1480 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1481 } else /* is_sync */
1483 blk_mq_put_ctx(data
.ctx
);
1487 if (!(data
.hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1489 blk_mq_try_issue_directly(old_rq
, &cookie
);
1492 srcu_idx
= srcu_read_lock(&data
.hctx
->queue_rq_srcu
);
1493 blk_mq_try_issue_directly(old_rq
, &cookie
);
1494 srcu_read_unlock(&data
.hctx
->queue_rq_srcu
, srcu_idx
);
1500 blk_mq_put_ctx(data
.ctx
);
1501 blk_mq_bio_to_request(rq
, bio
);
1502 blk_mq_sched_insert_request(rq
, false, true,
1503 !is_sync
|| is_flush_fua
, true);
1506 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1508 * For a SYNC request, send it to the hardware immediately. For
1509 * an ASYNC request, just ensure that we run it later on. The
1510 * latter allows for merging opportunities and more efficient
1513 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1515 blk_mq_put_ctx(data
.ctx
);
1521 * Single hardware queue variant. This will attempt to use any per-process
1522 * plug for merging and IO deferral.
1524 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1526 const int is_sync
= op_is_sync(bio
->bi_opf
);
1527 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1528 struct blk_plug
*plug
;
1529 unsigned int request_count
= 0;
1530 struct blk_mq_alloc_data data
= { .flags
= 0 };
1533 unsigned int wb_acct
;
1535 blk_queue_bounce(q
, &bio
);
1537 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1539 return BLK_QC_T_NONE
;
1542 blk_queue_split(q
, &bio
, q
->bio_split
);
1544 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1545 if (blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1546 return BLK_QC_T_NONE
;
1548 request_count
= blk_plug_queued_count(q
);
1550 if (blk_mq_sched_bio_merge(q
, bio
))
1551 return BLK_QC_T_NONE
;
1553 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1555 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1557 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1558 if (unlikely(!rq
)) {
1559 __wbt_done(q
->rq_wb
, wb_acct
);
1560 return BLK_QC_T_NONE
;
1563 wbt_track(&rq
->issue_stat
, wb_acct
);
1565 cookie
= request_to_qc_t(data
.hctx
, rq
);
1567 if (unlikely(is_flush_fua
)) {
1568 blk_mq_put_ctx(data
.ctx
);
1569 blk_mq_bio_to_request(rq
, bio
);
1570 blk_mq_get_driver_tag(rq
, NULL
, true);
1571 blk_insert_flush(rq
);
1572 blk_mq_run_hw_queue(data
.hctx
, true);
1577 * A task plug currently exists. Since this is completely lockless,
1578 * utilize that to temporarily store requests until the task is
1579 * either done or scheduled away.
1581 plug
= current
->plug
;
1583 struct request
*last
= NULL
;
1585 blk_mq_bio_to_request(rq
, bio
);
1588 * @request_count may become stale because of schedule
1589 * out, so check the list again.
1591 if (list_empty(&plug
->mq_list
))
1594 trace_block_plug(q
);
1596 last
= list_entry_rq(plug
->mq_list
.prev
);
1598 blk_mq_put_ctx(data
.ctx
);
1600 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1601 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1602 blk_flush_plug_list(plug
, false);
1603 trace_block_plug(q
);
1606 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1611 blk_mq_put_ctx(data
.ctx
);
1612 blk_mq_bio_to_request(rq
, bio
);
1613 blk_mq_sched_insert_request(rq
, false, true,
1614 !is_sync
|| is_flush_fua
, true);
1617 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1619 * For a SYNC request, send it to the hardware immediately. For
1620 * an ASYNC request, just ensure that we run it later on. The
1621 * latter allows for merging opportunities and more efficient
1624 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1627 blk_mq_put_ctx(data
.ctx
);
1632 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1633 unsigned int hctx_idx
)
1637 if (tags
->rqs
&& set
->ops
->exit_request
) {
1640 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1641 struct request
*rq
= tags
->static_rqs
[i
];
1645 set
->ops
->exit_request(set
->driver_data
, rq
,
1647 tags
->static_rqs
[i
] = NULL
;
1651 while (!list_empty(&tags
->page_list
)) {
1652 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1653 list_del_init(&page
->lru
);
1655 * Remove kmemleak object previously allocated in
1656 * blk_mq_init_rq_map().
1658 kmemleak_free(page_address(page
));
1659 __free_pages(page
, page
->private);
1663 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1667 kfree(tags
->static_rqs
);
1668 tags
->static_rqs
= NULL
;
1670 blk_mq_free_tags(tags
);
1673 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1674 unsigned int hctx_idx
,
1675 unsigned int nr_tags
,
1676 unsigned int reserved_tags
)
1678 struct blk_mq_tags
*tags
;
1680 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
,
1682 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1686 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1687 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1690 blk_mq_free_tags(tags
);
1694 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1695 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1697 if (!tags
->static_rqs
) {
1699 blk_mq_free_tags(tags
);
1706 static size_t order_to_size(unsigned int order
)
1708 return (size_t)PAGE_SIZE
<< order
;
1711 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1712 unsigned int hctx_idx
, unsigned int depth
)
1714 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1715 size_t rq_size
, left
;
1717 INIT_LIST_HEAD(&tags
->page_list
);
1720 * rq_size is the size of the request plus driver payload, rounded
1721 * to the cacheline size
1723 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1725 left
= rq_size
* depth
;
1727 for (i
= 0; i
< depth
; ) {
1728 int this_order
= max_order
;
1733 while (this_order
&& left
< order_to_size(this_order
- 1))
1737 page
= alloc_pages_node(set
->numa_node
,
1738 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1744 if (order_to_size(this_order
) < rq_size
)
1751 page
->private = this_order
;
1752 list_add_tail(&page
->lru
, &tags
->page_list
);
1754 p
= page_address(page
);
1756 * Allow kmemleak to scan these pages as they contain pointers
1757 * to additional allocations like via ops->init_request().
1759 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1760 entries_per_page
= order_to_size(this_order
) / rq_size
;
1761 to_do
= min(entries_per_page
, depth
- i
);
1762 left
-= to_do
* rq_size
;
1763 for (j
= 0; j
< to_do
; j
++) {
1764 struct request
*rq
= p
;
1766 tags
->static_rqs
[i
] = rq
;
1767 if (set
->ops
->init_request
) {
1768 if (set
->ops
->init_request(set
->driver_data
,
1771 tags
->static_rqs
[i
] = NULL
;
1783 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1788 * 'cpu' is going away. splice any existing rq_list entries from this
1789 * software queue to the hw queue dispatch list, and ensure that it
1792 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1794 struct blk_mq_hw_ctx
*hctx
;
1795 struct blk_mq_ctx
*ctx
;
1798 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1799 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1801 spin_lock(&ctx
->lock
);
1802 if (!list_empty(&ctx
->rq_list
)) {
1803 list_splice_init(&ctx
->rq_list
, &tmp
);
1804 blk_mq_hctx_clear_pending(hctx
, ctx
);
1806 spin_unlock(&ctx
->lock
);
1808 if (list_empty(&tmp
))
1811 spin_lock(&hctx
->lock
);
1812 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1813 spin_unlock(&hctx
->lock
);
1815 blk_mq_run_hw_queue(hctx
, true);
1819 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1821 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1825 /* hctx->ctxs will be freed in queue's release handler */
1826 static void blk_mq_exit_hctx(struct request_queue
*q
,
1827 struct blk_mq_tag_set
*set
,
1828 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1830 unsigned flush_start_tag
= set
->queue_depth
;
1832 blk_mq_tag_idle(hctx
);
1834 if (set
->ops
->exit_request
)
1835 set
->ops
->exit_request(set
->driver_data
,
1836 hctx
->fq
->flush_rq
, hctx_idx
,
1837 flush_start_tag
+ hctx_idx
);
1839 if (set
->ops
->exit_hctx
)
1840 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1842 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1843 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1845 blk_mq_remove_cpuhp(hctx
);
1846 blk_free_flush_queue(hctx
->fq
);
1847 sbitmap_free(&hctx
->ctx_map
);
1850 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1851 struct blk_mq_tag_set
*set
, int nr_queue
)
1853 struct blk_mq_hw_ctx
*hctx
;
1856 queue_for_each_hw_ctx(q
, hctx
, i
) {
1859 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1863 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1864 struct blk_mq_tag_set
*set
)
1866 struct blk_mq_hw_ctx
*hctx
;
1869 queue_for_each_hw_ctx(q
, hctx
, i
)
1870 free_cpumask_var(hctx
->cpumask
);
1873 static int blk_mq_init_hctx(struct request_queue
*q
,
1874 struct blk_mq_tag_set
*set
,
1875 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1878 unsigned flush_start_tag
= set
->queue_depth
;
1880 node
= hctx
->numa_node
;
1881 if (node
== NUMA_NO_NODE
)
1882 node
= hctx
->numa_node
= set
->numa_node
;
1884 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1885 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1886 spin_lock_init(&hctx
->lock
);
1887 INIT_LIST_HEAD(&hctx
->dispatch
);
1889 hctx
->queue_num
= hctx_idx
;
1890 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1892 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1894 hctx
->tags
= set
->tags
[hctx_idx
];
1897 * Allocate space for all possible cpus to avoid allocation at
1900 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1903 goto unregister_cpu_notifier
;
1905 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1911 if (set
->ops
->init_hctx
&&
1912 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1915 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1919 if (set
->ops
->init_request
&&
1920 set
->ops
->init_request(set
->driver_data
,
1921 hctx
->fq
->flush_rq
, hctx_idx
,
1922 flush_start_tag
+ hctx_idx
, node
))
1925 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1926 init_srcu_struct(&hctx
->queue_rq_srcu
);
1933 if (set
->ops
->exit_hctx
)
1934 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1936 sbitmap_free(&hctx
->ctx_map
);
1939 unregister_cpu_notifier
:
1940 blk_mq_remove_cpuhp(hctx
);
1944 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1945 unsigned int nr_hw_queues
)
1949 for_each_possible_cpu(i
) {
1950 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1951 struct blk_mq_hw_ctx
*hctx
;
1953 memset(__ctx
, 0, sizeof(*__ctx
));
1955 spin_lock_init(&__ctx
->lock
);
1956 INIT_LIST_HEAD(&__ctx
->rq_list
);
1958 blk_stat_init(&__ctx
->stat
[BLK_STAT_READ
]);
1959 blk_stat_init(&__ctx
->stat
[BLK_STAT_WRITE
]);
1961 /* If the cpu isn't online, the cpu is mapped to first hctx */
1965 hctx
= blk_mq_map_queue(q
, i
);
1968 * Set local node, IFF we have more than one hw queue. If
1969 * not, we remain on the home node of the device
1971 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1972 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1976 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
1980 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
1981 set
->queue_depth
, set
->reserved_tags
);
1982 if (!set
->tags
[hctx_idx
])
1985 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
1990 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
1991 set
->tags
[hctx_idx
] = NULL
;
1995 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
1996 unsigned int hctx_idx
)
1998 if (set
->tags
[hctx_idx
]) {
1999 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2000 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2001 set
->tags
[hctx_idx
] = NULL
;
2005 static void blk_mq_map_swqueue(struct request_queue
*q
,
2006 const struct cpumask
*online_mask
)
2008 unsigned int i
, hctx_idx
;
2009 struct blk_mq_hw_ctx
*hctx
;
2010 struct blk_mq_ctx
*ctx
;
2011 struct blk_mq_tag_set
*set
= q
->tag_set
;
2014 * Avoid others reading imcomplete hctx->cpumask through sysfs
2016 mutex_lock(&q
->sysfs_lock
);
2018 queue_for_each_hw_ctx(q
, hctx
, i
) {
2019 cpumask_clear(hctx
->cpumask
);
2024 * Map software to hardware queues
2026 for_each_possible_cpu(i
) {
2027 /* If the cpu isn't online, the cpu is mapped to first hctx */
2028 if (!cpumask_test_cpu(i
, online_mask
))
2031 hctx_idx
= q
->mq_map
[i
];
2032 /* unmapped hw queue can be remapped after CPU topo changed */
2033 if (!set
->tags
[hctx_idx
] &&
2034 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2036 * If tags initialization fail for some hctx,
2037 * that hctx won't be brought online. In this
2038 * case, remap the current ctx to hctx[0] which
2039 * is guaranteed to always have tags allocated
2044 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2045 hctx
= blk_mq_map_queue(q
, i
);
2047 cpumask_set_cpu(i
, hctx
->cpumask
);
2048 ctx
->index_hw
= hctx
->nr_ctx
;
2049 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2052 mutex_unlock(&q
->sysfs_lock
);
2054 queue_for_each_hw_ctx(q
, hctx
, i
) {
2056 * If no software queues are mapped to this hardware queue,
2057 * disable it and free the request entries.
2059 if (!hctx
->nr_ctx
) {
2060 /* Never unmap queue 0. We need it as a
2061 * fallback in case of a new remap fails
2064 if (i
&& set
->tags
[i
])
2065 blk_mq_free_map_and_requests(set
, i
);
2071 hctx
->tags
= set
->tags
[i
];
2072 WARN_ON(!hctx
->tags
);
2075 * Set the map size to the number of mapped software queues.
2076 * This is more accurate and more efficient than looping
2077 * over all possibly mapped software queues.
2079 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2082 * Initialize batch roundrobin counts
2084 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2085 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2089 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2091 struct blk_mq_hw_ctx
*hctx
;
2094 queue_for_each_hw_ctx(q
, hctx
, i
) {
2096 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2098 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2102 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
2104 struct request_queue
*q
;
2106 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2107 blk_mq_freeze_queue(q
);
2108 queue_set_hctx_shared(q
, shared
);
2109 blk_mq_unfreeze_queue(q
);
2113 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2115 struct blk_mq_tag_set
*set
= q
->tag_set
;
2117 mutex_lock(&set
->tag_list_lock
);
2118 list_del_init(&q
->tag_set_list
);
2119 if (list_is_singular(&set
->tag_list
)) {
2120 /* just transitioned to unshared */
2121 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2122 /* update existing queue */
2123 blk_mq_update_tag_set_depth(set
, false);
2125 mutex_unlock(&set
->tag_list_lock
);
2128 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2129 struct request_queue
*q
)
2133 mutex_lock(&set
->tag_list_lock
);
2135 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2136 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2137 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2138 /* update existing queue */
2139 blk_mq_update_tag_set_depth(set
, true);
2141 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2142 queue_set_hctx_shared(q
, true);
2143 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
2145 mutex_unlock(&set
->tag_list_lock
);
2149 * It is the actual release handler for mq, but we do it from
2150 * request queue's release handler for avoiding use-after-free
2151 * and headache because q->mq_kobj shouldn't have been introduced,
2152 * but we can't group ctx/kctx kobj without it.
2154 void blk_mq_release(struct request_queue
*q
)
2156 struct blk_mq_hw_ctx
*hctx
;
2159 blk_mq_sched_teardown(q
);
2161 /* hctx kobj stays in hctx */
2162 queue_for_each_hw_ctx(q
, hctx
, i
) {
2171 kfree(q
->queue_hw_ctx
);
2173 /* ctx kobj stays in queue_ctx */
2174 free_percpu(q
->queue_ctx
);
2177 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2179 struct request_queue
*uninit_q
, *q
;
2181 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2183 return ERR_PTR(-ENOMEM
);
2185 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2187 blk_cleanup_queue(uninit_q
);
2191 EXPORT_SYMBOL(blk_mq_init_queue
);
2193 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2194 struct request_queue
*q
)
2197 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2199 blk_mq_sysfs_unregister(q
);
2200 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2206 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2207 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2212 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2219 atomic_set(&hctxs
[i
]->nr_active
, 0);
2220 hctxs
[i
]->numa_node
= node
;
2221 hctxs
[i
]->queue_num
= i
;
2223 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2224 free_cpumask_var(hctxs
[i
]->cpumask
);
2229 blk_mq_hctx_kobj_init(hctxs
[i
]);
2231 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2232 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2236 blk_mq_free_map_and_requests(set
, j
);
2237 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2238 free_cpumask_var(hctx
->cpumask
);
2239 kobject_put(&hctx
->kobj
);
2246 q
->nr_hw_queues
= i
;
2247 blk_mq_sysfs_register(q
);
2250 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2251 struct request_queue
*q
)
2253 /* mark the queue as mq asap */
2254 q
->mq_ops
= set
->ops
;
2256 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2260 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2261 GFP_KERNEL
, set
->numa_node
);
2262 if (!q
->queue_hw_ctx
)
2265 q
->mq_map
= set
->mq_map
;
2267 blk_mq_realloc_hw_ctxs(set
, q
);
2268 if (!q
->nr_hw_queues
)
2271 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2272 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2274 q
->nr_queues
= nr_cpu_ids
;
2276 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2278 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2279 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2281 q
->sg_reserved_size
= INT_MAX
;
2283 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2284 INIT_LIST_HEAD(&q
->requeue_list
);
2285 spin_lock_init(&q
->requeue_lock
);
2287 if (q
->nr_hw_queues
> 1)
2288 blk_queue_make_request(q
, blk_mq_make_request
);
2290 blk_queue_make_request(q
, blk_sq_make_request
);
2293 * Do this after blk_queue_make_request() overrides it...
2295 q
->nr_requests
= set
->queue_depth
;
2298 * Default to classic polling
2302 if (set
->ops
->complete
)
2303 blk_queue_softirq_done(q
, set
->ops
->complete
);
2305 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2308 mutex_lock(&all_q_mutex
);
2310 list_add_tail(&q
->all_q_node
, &all_q_list
);
2311 blk_mq_add_queue_tag_set(set
, q
);
2312 blk_mq_map_swqueue(q
, cpu_online_mask
);
2314 mutex_unlock(&all_q_mutex
);
2317 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2320 ret
= blk_mq_sched_init(q
);
2322 return ERR_PTR(ret
);
2328 kfree(q
->queue_hw_ctx
);
2330 free_percpu(q
->queue_ctx
);
2333 return ERR_PTR(-ENOMEM
);
2335 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2337 void blk_mq_free_queue(struct request_queue
*q
)
2339 struct blk_mq_tag_set
*set
= q
->tag_set
;
2341 mutex_lock(&all_q_mutex
);
2342 list_del_init(&q
->all_q_node
);
2343 mutex_unlock(&all_q_mutex
);
2347 blk_mq_del_queue_tag_set(q
);
2349 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2350 blk_mq_free_hw_queues(q
, set
);
2353 /* Basically redo blk_mq_init_queue with queue frozen */
2354 static void blk_mq_queue_reinit(struct request_queue
*q
,
2355 const struct cpumask
*online_mask
)
2357 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2359 blk_mq_sysfs_unregister(q
);
2362 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2363 * we should change hctx numa_node according to new topology (this
2364 * involves free and re-allocate memory, worthy doing?)
2367 blk_mq_map_swqueue(q
, online_mask
);
2369 blk_mq_sysfs_register(q
);
2373 * New online cpumask which is going to be set in this hotplug event.
2374 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2375 * one-by-one and dynamically allocating this could result in a failure.
2377 static struct cpumask cpuhp_online_new
;
2379 static void blk_mq_queue_reinit_work(void)
2381 struct request_queue
*q
;
2383 mutex_lock(&all_q_mutex
);
2385 * We need to freeze and reinit all existing queues. Freezing
2386 * involves synchronous wait for an RCU grace period and doing it
2387 * one by one may take a long time. Start freezing all queues in
2388 * one swoop and then wait for the completions so that freezing can
2389 * take place in parallel.
2391 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2392 blk_mq_freeze_queue_start(q
);
2393 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2394 blk_mq_freeze_queue_wait(q
);
2396 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2397 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2399 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2400 blk_mq_unfreeze_queue(q
);
2402 mutex_unlock(&all_q_mutex
);
2405 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2407 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2408 blk_mq_queue_reinit_work();
2413 * Before hotadded cpu starts handling requests, new mappings must be
2414 * established. Otherwise, these requests in hw queue might never be
2417 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2418 * for CPU0, and ctx1 for CPU1).
2420 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2421 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2423 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2424 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2425 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2428 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2430 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2431 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2432 blk_mq_queue_reinit_work();
2436 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2440 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2441 if (!__blk_mq_alloc_rq_map(set
, i
))
2448 blk_mq_free_rq_map(set
->tags
[i
]);
2454 * Allocate the request maps associated with this tag_set. Note that this
2455 * may reduce the depth asked for, if memory is tight. set->queue_depth
2456 * will be updated to reflect the allocated depth.
2458 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2463 depth
= set
->queue_depth
;
2465 err
= __blk_mq_alloc_rq_maps(set
);
2469 set
->queue_depth
>>= 1;
2470 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2474 } while (set
->queue_depth
);
2476 if (!set
->queue_depth
|| err
) {
2477 pr_err("blk-mq: failed to allocate request map\n");
2481 if (depth
!= set
->queue_depth
)
2482 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2483 depth
, set
->queue_depth
);
2489 * Alloc a tag set to be associated with one or more request queues.
2490 * May fail with EINVAL for various error conditions. May adjust the
2491 * requested depth down, if if it too large. In that case, the set
2492 * value will be stored in set->queue_depth.
2494 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2498 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2500 if (!set
->nr_hw_queues
)
2502 if (!set
->queue_depth
)
2504 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2507 if (!set
->ops
->queue_rq
)
2510 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2511 pr_info("blk-mq: reduced tag depth to %u\n",
2513 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2517 * If a crashdump is active, then we are potentially in a very
2518 * memory constrained environment. Limit us to 1 queue and
2519 * 64 tags to prevent using too much memory.
2521 if (is_kdump_kernel()) {
2522 set
->nr_hw_queues
= 1;
2523 set
->queue_depth
= min(64U, set
->queue_depth
);
2526 * There is no use for more h/w queues than cpus.
2528 if (set
->nr_hw_queues
> nr_cpu_ids
)
2529 set
->nr_hw_queues
= nr_cpu_ids
;
2531 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2532 GFP_KERNEL
, set
->numa_node
);
2537 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2538 GFP_KERNEL
, set
->numa_node
);
2542 if (set
->ops
->map_queues
)
2543 ret
= set
->ops
->map_queues(set
);
2545 ret
= blk_mq_map_queues(set
);
2547 goto out_free_mq_map
;
2549 ret
= blk_mq_alloc_rq_maps(set
);
2551 goto out_free_mq_map
;
2553 mutex_init(&set
->tag_list_lock
);
2554 INIT_LIST_HEAD(&set
->tag_list
);
2566 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2568 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2572 for (i
= 0; i
< nr_cpu_ids
; i
++)
2573 blk_mq_free_map_and_requests(set
, i
);
2581 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2583 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2585 struct blk_mq_tag_set
*set
= q
->tag_set
;
2586 struct blk_mq_hw_ctx
*hctx
;
2592 blk_mq_freeze_queue(q
);
2593 blk_mq_quiesce_queue(q
);
2596 queue_for_each_hw_ctx(q
, hctx
, i
) {
2600 * If we're using an MQ scheduler, just update the scheduler
2601 * queue depth. This is similar to what the old code would do.
2603 if (!hctx
->sched_tags
) {
2604 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2605 min(nr
, set
->queue_depth
),
2608 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2616 q
->nr_requests
= nr
;
2618 blk_mq_unfreeze_queue(q
);
2619 blk_mq_start_stopped_hw_queues(q
, true);
2624 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2626 struct request_queue
*q
;
2628 if (nr_hw_queues
> nr_cpu_ids
)
2629 nr_hw_queues
= nr_cpu_ids
;
2630 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2633 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2634 blk_mq_freeze_queue(q
);
2636 set
->nr_hw_queues
= nr_hw_queues
;
2637 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2638 blk_mq_realloc_hw_ctxs(set
, q
);
2640 if (q
->nr_hw_queues
> 1)
2641 blk_queue_make_request(q
, blk_mq_make_request
);
2643 blk_queue_make_request(q
, blk_sq_make_request
);
2645 blk_mq_queue_reinit(q
, cpu_online_mask
);
2648 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2649 blk_mq_unfreeze_queue(q
);
2651 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2653 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2654 struct blk_mq_hw_ctx
*hctx
,
2657 struct blk_rq_stat stat
[2];
2658 unsigned long ret
= 0;
2661 * If stats collection isn't on, don't sleep but turn it on for
2664 if (!blk_stat_enable(q
))
2668 * We don't have to do this once per IO, should optimize this
2669 * to just use the current window of stats until it changes
2671 memset(&stat
, 0, sizeof(stat
));
2672 blk_hctx_stat_get(hctx
, stat
);
2675 * As an optimistic guess, use half of the mean service time
2676 * for this type of request. We can (and should) make this smarter.
2677 * For instance, if the completion latencies are tight, we can
2678 * get closer than just half the mean. This is especially
2679 * important on devices where the completion latencies are longer
2682 if (req_op(rq
) == REQ_OP_READ
&& stat
[BLK_STAT_READ
].nr_samples
)
2683 ret
= (stat
[BLK_STAT_READ
].mean
+ 1) / 2;
2684 else if (req_op(rq
) == REQ_OP_WRITE
&& stat
[BLK_STAT_WRITE
].nr_samples
)
2685 ret
= (stat
[BLK_STAT_WRITE
].mean
+ 1) / 2;
2690 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2691 struct blk_mq_hw_ctx
*hctx
,
2694 struct hrtimer_sleeper hs
;
2695 enum hrtimer_mode mode
;
2699 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2705 * -1: don't ever hybrid sleep
2706 * 0: use half of prev avg
2707 * >0: use this specific value
2709 if (q
->poll_nsec
== -1)
2711 else if (q
->poll_nsec
> 0)
2712 nsecs
= q
->poll_nsec
;
2714 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2719 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2722 * This will be replaced with the stats tracking code, using
2723 * 'avg_completion_time / 2' as the pre-sleep target.
2727 mode
= HRTIMER_MODE_REL
;
2728 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2729 hrtimer_set_expires(&hs
.timer
, kt
);
2731 hrtimer_init_sleeper(&hs
, current
);
2733 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2735 set_current_state(TASK_UNINTERRUPTIBLE
);
2736 hrtimer_start_expires(&hs
.timer
, mode
);
2739 hrtimer_cancel(&hs
.timer
);
2740 mode
= HRTIMER_MODE_ABS
;
2741 } while (hs
.task
&& !signal_pending(current
));
2743 __set_current_state(TASK_RUNNING
);
2744 destroy_hrtimer_on_stack(&hs
.timer
);
2748 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2750 struct request_queue
*q
= hctx
->queue
;
2754 * If we sleep, have the caller restart the poll loop to reset
2755 * the state. Like for the other success return cases, the
2756 * caller is responsible for checking if the IO completed. If
2757 * the IO isn't complete, we'll get called again and will go
2758 * straight to the busy poll loop.
2760 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2763 hctx
->poll_considered
++;
2765 state
= current
->state
;
2766 while (!need_resched()) {
2769 hctx
->poll_invoked
++;
2771 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2773 hctx
->poll_success
++;
2774 set_current_state(TASK_RUNNING
);
2778 if (signal_pending_state(state
, current
))
2779 set_current_state(TASK_RUNNING
);
2781 if (current
->state
== TASK_RUNNING
)
2791 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2793 struct blk_mq_hw_ctx
*hctx
;
2794 struct blk_plug
*plug
;
2797 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2798 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2801 plug
= current
->plug
;
2803 blk_flush_plug_list(plug
, false);
2805 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2806 if (!blk_qc_t_is_internal(cookie
))
2807 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2809 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2811 return __blk_mq_poll(hctx
, rq
);
2813 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2815 void blk_mq_disable_hotplug(void)
2817 mutex_lock(&all_q_mutex
);
2820 void blk_mq_enable_hotplug(void)
2822 mutex_unlock(&all_q_mutex
);
2825 static int __init
blk_mq_init(void)
2827 blk_mq_debugfs_init();
2829 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2830 blk_mq_hctx_notify_dead
);
2832 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2833 blk_mq_queue_reinit_prepare
,
2834 blk_mq_queue_reinit_dead
);
2837 subsys_initcall(blk_mq_init
);