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);
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);
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 static bool blk_mq_get_driver_tag(struct request
*rq
,
851 struct blk_mq_hw_ctx
**hctx
, bool wait
)
853 struct blk_mq_alloc_data data
= {
856 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
857 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
860 if (blk_mq_hctx_stopped(data
.hctx
))
870 rq
->tag
= blk_mq_get_tag(&data
);
872 if (blk_mq_tag_busy(data
.hctx
)) {
873 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
874 atomic_inc(&data
.hctx
->nr_active
);
876 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
883 static void blk_mq_put_driver_tag(struct blk_mq_hw_ctx
*hctx
,
886 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
889 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
892 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
893 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
894 atomic_dec(&hctx
->nr_active
);
899 * If we fail getting a driver tag because all the driver tags are already
900 * assigned and on the dispatch list, BUT the first entry does not have a
901 * tag, then we could deadlock. For that case, move entries with assigned
902 * driver tags to the front, leaving the set of tagged requests in the
903 * same order, and the untagged set in the same order.
905 static bool reorder_tags_to_front(struct list_head
*list
)
907 struct request
*rq
, *tmp
, *first
= NULL
;
909 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
913 list_move(&rq
->queuelist
, list
);
919 return first
!= NULL
;
922 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
924 struct request_queue
*q
= hctx
->queue
;
926 LIST_HEAD(driver_list
);
927 struct list_head
*dptr
;
928 int queued
, ret
= BLK_MQ_RQ_QUEUE_OK
;
931 * Start off with dptr being NULL, so we start the first request
932 * immediately, even if we have more pending.
937 * Now process all the entries, sending them to the driver.
940 while (!list_empty(list
)) {
941 struct blk_mq_queue_data bd
;
943 rq
= list_first_entry(list
, struct request
, queuelist
);
944 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
945 if (!queued
&& reorder_tags_to_front(list
))
949 * We failed getting a driver tag. Mark the queue(s)
950 * as needing a restart. Retry getting a tag again,
951 * in case the needed IO completed right before we
952 * marked the queue as needing a restart.
954 blk_mq_sched_mark_restart(hctx
);
955 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
958 list_del_init(&rq
->queuelist
);
962 bd
.last
= list_empty(list
);
964 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
966 case BLK_MQ_RQ_QUEUE_OK
:
969 case BLK_MQ_RQ_QUEUE_BUSY
:
970 blk_mq_put_driver_tag(hctx
, rq
);
971 list_add(&rq
->queuelist
, list
);
972 __blk_mq_requeue_request(rq
);
975 pr_err("blk-mq: bad return on queue: %d\n", ret
);
976 case BLK_MQ_RQ_QUEUE_ERROR
:
978 blk_mq_end_request(rq
, rq
->errors
);
982 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
986 * We've done the first request. If we have more than 1
987 * left in the list, set dptr to defer issue.
989 if (!dptr
&& list
->next
!= list
->prev
)
993 hctx
->dispatched
[queued_to_index(queued
)]++;
996 * Any items that need requeuing? Stuff them into hctx->dispatch,
997 * that is where we will continue on next queue run.
999 if (!list_empty(list
)) {
1000 spin_lock(&hctx
->lock
);
1001 list_splice_init(list
, &hctx
->dispatch
);
1002 spin_unlock(&hctx
->lock
);
1005 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
1006 * it's possible the queue is stopped and restarted again
1007 * before this. Queue restart will dispatch requests. And since
1008 * requests in rq_list aren't added into hctx->dispatch yet,
1009 * the requests in rq_list might get lost.
1011 * blk_mq_run_hw_queue() already checks the STOPPED bit
1013 * If RESTART is set, then let completion restart the queue
1014 * instead of potentially looping here.
1016 if (!blk_mq_sched_needs_restart(hctx
))
1017 blk_mq_run_hw_queue(hctx
, true);
1020 return ret
!= BLK_MQ_RQ_QUEUE_BUSY
;
1023 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1027 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1028 cpu_online(hctx
->next_cpu
));
1030 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1032 blk_mq_sched_dispatch_requests(hctx
);
1035 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1036 blk_mq_sched_dispatch_requests(hctx
);
1037 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1042 * It'd be great if the workqueue API had a way to pass
1043 * in a mask and had some smarts for more clever placement.
1044 * For now we just round-robin here, switching for every
1045 * BLK_MQ_CPU_WORK_BATCH queued items.
1047 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1049 if (hctx
->queue
->nr_hw_queues
== 1)
1050 return WORK_CPU_UNBOUND
;
1052 if (--hctx
->next_cpu_batch
<= 0) {
1055 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1056 if (next_cpu
>= nr_cpu_ids
)
1057 next_cpu
= cpumask_first(hctx
->cpumask
);
1059 hctx
->next_cpu
= next_cpu
;
1060 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1063 return hctx
->next_cpu
;
1066 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1068 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
1069 !blk_mq_hw_queue_mapped(hctx
)))
1072 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1073 int cpu
= get_cpu();
1074 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1075 __blk_mq_run_hw_queue(hctx
);
1083 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
);
1086 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1088 struct blk_mq_hw_ctx
*hctx
;
1091 queue_for_each_hw_ctx(q
, hctx
, i
) {
1092 if (!blk_mq_hctx_has_pending(hctx
) ||
1093 blk_mq_hctx_stopped(hctx
))
1096 blk_mq_run_hw_queue(hctx
, async
);
1099 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1102 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1103 * @q: request queue.
1105 * The caller is responsible for serializing this function against
1106 * blk_mq_{start,stop}_hw_queue().
1108 bool blk_mq_queue_stopped(struct request_queue
*q
)
1110 struct blk_mq_hw_ctx
*hctx
;
1113 queue_for_each_hw_ctx(q
, hctx
, i
)
1114 if (blk_mq_hctx_stopped(hctx
))
1119 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1121 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1123 cancel_work(&hctx
->run_work
);
1124 cancel_delayed_work(&hctx
->delay_work
);
1125 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1127 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1129 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1131 struct blk_mq_hw_ctx
*hctx
;
1134 queue_for_each_hw_ctx(q
, hctx
, i
)
1135 blk_mq_stop_hw_queue(hctx
);
1137 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1139 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1141 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1143 blk_mq_run_hw_queue(hctx
, false);
1145 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1147 void blk_mq_start_hw_queues(struct request_queue
*q
)
1149 struct blk_mq_hw_ctx
*hctx
;
1152 queue_for_each_hw_ctx(q
, hctx
, i
)
1153 blk_mq_start_hw_queue(hctx
);
1155 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1157 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1159 if (!blk_mq_hctx_stopped(hctx
))
1162 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1163 blk_mq_run_hw_queue(hctx
, async
);
1165 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1167 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1169 struct blk_mq_hw_ctx
*hctx
;
1172 queue_for_each_hw_ctx(q
, hctx
, i
)
1173 blk_mq_start_stopped_hw_queue(hctx
, async
);
1175 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1177 static void blk_mq_run_work_fn(struct work_struct
*work
)
1179 struct blk_mq_hw_ctx
*hctx
;
1181 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1183 __blk_mq_run_hw_queue(hctx
);
1186 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1188 struct blk_mq_hw_ctx
*hctx
;
1190 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1192 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1193 __blk_mq_run_hw_queue(hctx
);
1196 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1198 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1201 blk_mq_stop_hw_queue(hctx
);
1202 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1203 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1205 EXPORT_SYMBOL(blk_mq_delay_queue
);
1207 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1211 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1213 trace_block_rq_insert(hctx
->queue
, rq
);
1216 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1218 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1221 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1224 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1226 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1227 blk_mq_hctx_mark_pending(hctx
, ctx
);
1230 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1231 struct list_head
*list
)
1235 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1238 spin_lock(&ctx
->lock
);
1239 while (!list_empty(list
)) {
1242 rq
= list_first_entry(list
, struct request
, queuelist
);
1243 BUG_ON(rq
->mq_ctx
!= ctx
);
1244 list_del_init(&rq
->queuelist
);
1245 __blk_mq_insert_req_list(hctx
, rq
, false);
1247 blk_mq_hctx_mark_pending(hctx
, ctx
);
1248 spin_unlock(&ctx
->lock
);
1251 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1253 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1254 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1256 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1257 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1258 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1261 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1263 struct blk_mq_ctx
*this_ctx
;
1264 struct request_queue
*this_q
;
1267 LIST_HEAD(ctx_list
);
1270 list_splice_init(&plug
->mq_list
, &list
);
1272 list_sort(NULL
, &list
, plug_ctx_cmp
);
1278 while (!list_empty(&list
)) {
1279 rq
= list_entry_rq(list
.next
);
1280 list_del_init(&rq
->queuelist
);
1282 if (rq
->mq_ctx
!= this_ctx
) {
1284 trace_block_unplug(this_q
, depth
, from_schedule
);
1285 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1290 this_ctx
= rq
->mq_ctx
;
1296 list_add_tail(&rq
->queuelist
, &ctx_list
);
1300 * If 'this_ctx' is set, we know we have entries to complete
1301 * on 'ctx_list'. Do those.
1304 trace_block_unplug(this_q
, depth
, from_schedule
);
1305 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1310 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1312 init_request_from_bio(rq
, bio
);
1314 blk_account_io_start(rq
, true);
1317 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1319 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1320 !blk_queue_nomerges(hctx
->queue
);
1323 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1324 struct blk_mq_ctx
*ctx
,
1325 struct request
*rq
, struct bio
*bio
)
1327 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1328 blk_mq_bio_to_request(rq
, bio
);
1329 spin_lock(&ctx
->lock
);
1331 __blk_mq_insert_request(hctx
, rq
, false);
1332 spin_unlock(&ctx
->lock
);
1335 struct request_queue
*q
= hctx
->queue
;
1337 spin_lock(&ctx
->lock
);
1338 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1339 blk_mq_bio_to_request(rq
, bio
);
1343 spin_unlock(&ctx
->lock
);
1344 __blk_mq_finish_request(hctx
, ctx
, rq
);
1349 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1352 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1354 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1357 static void blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
)
1359 struct request_queue
*q
= rq
->q
;
1360 struct blk_mq_queue_data bd
= {
1365 struct blk_mq_hw_ctx
*hctx
;
1366 blk_qc_t new_cookie
;
1372 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1375 new_cookie
= request_to_qc_t(hctx
, rq
);
1378 * For OK queue, we are done. For error, kill it. Any other
1379 * error (busy), just add it to our list as we previously
1382 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1383 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1384 *cookie
= new_cookie
;
1388 __blk_mq_requeue_request(rq
);
1390 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1391 *cookie
= BLK_QC_T_NONE
;
1393 blk_mq_end_request(rq
, rq
->errors
);
1398 blk_mq_sched_insert_request(rq
, false, true, true);
1402 * Multiple hardware queue variant. This will not use per-process plugs,
1403 * but will attempt to bypass the hctx queueing if we can go straight to
1404 * hardware for SYNC IO.
1406 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1408 const int is_sync
= op_is_sync(bio
->bi_opf
);
1409 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1410 struct blk_mq_alloc_data data
= { .flags
= 0 };
1412 unsigned int request_count
= 0, srcu_idx
;
1413 struct blk_plug
*plug
;
1414 struct request
*same_queue_rq
= NULL
;
1416 unsigned int wb_acct
;
1418 blk_queue_bounce(q
, &bio
);
1420 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1422 return BLK_QC_T_NONE
;
1425 blk_queue_split(q
, &bio
, q
->bio_split
);
1427 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1428 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1429 return BLK_QC_T_NONE
;
1431 if (blk_mq_sched_bio_merge(q
, bio
))
1432 return BLK_QC_T_NONE
;
1434 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1436 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1438 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1439 if (unlikely(!rq
)) {
1440 __wbt_done(q
->rq_wb
, wb_acct
);
1441 return BLK_QC_T_NONE
;
1444 wbt_track(&rq
->issue_stat
, wb_acct
);
1446 cookie
= request_to_qc_t(data
.hctx
, rq
);
1448 if (unlikely(is_flush_fua
)) {
1449 blk_mq_bio_to_request(rq
, bio
);
1450 blk_mq_get_driver_tag(rq
, NULL
, true);
1451 blk_insert_flush(rq
);
1455 plug
= current
->plug
;
1457 * If the driver supports defer issued based on 'last', then
1458 * queue it up like normal since we can potentially save some
1461 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1462 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1463 struct request
*old_rq
= NULL
;
1465 blk_mq_bio_to_request(rq
, bio
);
1468 * We do limited plugging. If the bio can be merged, do that.
1469 * Otherwise the existing request in the plug list will be
1470 * issued. So the plug list will have one request at most
1474 * The plug list might get flushed before this. If that
1475 * happens, same_queue_rq is invalid and plug list is
1478 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1479 old_rq
= same_queue_rq
;
1480 list_del_init(&old_rq
->queuelist
);
1482 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1483 } else /* is_sync */
1485 blk_mq_put_ctx(data
.ctx
);
1489 if (!(data
.hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1491 blk_mq_try_issue_directly(old_rq
, &cookie
);
1494 srcu_idx
= srcu_read_lock(&data
.hctx
->queue_rq_srcu
);
1495 blk_mq_try_issue_directly(old_rq
, &cookie
);
1496 srcu_read_unlock(&data
.hctx
->queue_rq_srcu
, srcu_idx
);
1502 blk_mq_put_ctx(data
.ctx
);
1503 blk_mq_bio_to_request(rq
, bio
);
1504 blk_mq_sched_insert_request(rq
, false, true,
1505 !is_sync
|| is_flush_fua
);
1508 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1510 * For a SYNC request, send it to the hardware immediately. For
1511 * an ASYNC request, just ensure that we run it later on. The
1512 * latter allows for merging opportunities and more efficient
1516 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1518 blk_mq_put_ctx(data
.ctx
);
1524 * Single hardware queue variant. This will attempt to use any per-process
1525 * plug for merging and IO deferral.
1527 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1529 const int is_sync
= op_is_sync(bio
->bi_opf
);
1530 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1531 struct blk_plug
*plug
;
1532 unsigned int request_count
= 0;
1533 struct blk_mq_alloc_data data
= { .flags
= 0 };
1536 unsigned int wb_acct
;
1538 blk_queue_bounce(q
, &bio
);
1540 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1542 return BLK_QC_T_NONE
;
1545 blk_queue_split(q
, &bio
, q
->bio_split
);
1547 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1548 if (blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1549 return BLK_QC_T_NONE
;
1551 request_count
= blk_plug_queued_count(q
);
1553 if (blk_mq_sched_bio_merge(q
, bio
))
1554 return BLK_QC_T_NONE
;
1556 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1558 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1560 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1561 if (unlikely(!rq
)) {
1562 __wbt_done(q
->rq_wb
, wb_acct
);
1563 return BLK_QC_T_NONE
;
1566 wbt_track(&rq
->issue_stat
, wb_acct
);
1568 cookie
= request_to_qc_t(data
.hctx
, rq
);
1570 if (unlikely(is_flush_fua
)) {
1571 blk_mq_bio_to_request(rq
, bio
);
1572 blk_mq_get_driver_tag(rq
, NULL
, true);
1573 blk_insert_flush(rq
);
1578 * A task plug currently exists. Since this is completely lockless,
1579 * utilize that to temporarily store requests until the task is
1580 * either done or scheduled away.
1582 plug
= current
->plug
;
1584 struct request
*last
= NULL
;
1586 blk_mq_bio_to_request(rq
, bio
);
1589 * @request_count may become stale because of schedule
1590 * out, so check the list again.
1592 if (list_empty(&plug
->mq_list
))
1595 trace_block_plug(q
);
1597 last
= list_entry_rq(plug
->mq_list
.prev
);
1599 blk_mq_put_ctx(data
.ctx
);
1601 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1602 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1603 blk_flush_plug_list(plug
, false);
1604 trace_block_plug(q
);
1607 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1612 blk_mq_put_ctx(data
.ctx
);
1613 blk_mq_bio_to_request(rq
, bio
);
1614 blk_mq_sched_insert_request(rq
, false, true,
1615 !is_sync
|| is_flush_fua
);
1618 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1620 * For a SYNC request, send it to the hardware immediately. For
1621 * an ASYNC request, just ensure that we run it later on. The
1622 * latter allows for merging opportunities and more efficient
1626 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1629 blk_mq_put_ctx(data
.ctx
);
1634 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1635 unsigned int hctx_idx
)
1639 if (tags
->rqs
&& set
->ops
->exit_request
) {
1642 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1643 struct request
*rq
= tags
->static_rqs
[i
];
1647 set
->ops
->exit_request(set
->driver_data
, rq
,
1649 tags
->static_rqs
[i
] = NULL
;
1653 while (!list_empty(&tags
->page_list
)) {
1654 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1655 list_del_init(&page
->lru
);
1657 * Remove kmemleak object previously allocated in
1658 * blk_mq_init_rq_map().
1660 kmemleak_free(page_address(page
));
1661 __free_pages(page
, page
->private);
1665 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1669 kfree(tags
->static_rqs
);
1670 tags
->static_rqs
= NULL
;
1672 blk_mq_free_tags(tags
);
1675 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1676 unsigned int hctx_idx
,
1677 unsigned int nr_tags
,
1678 unsigned int reserved_tags
)
1680 struct blk_mq_tags
*tags
;
1682 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
,
1684 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1688 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1689 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1692 blk_mq_free_tags(tags
);
1696 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1697 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1699 if (!tags
->static_rqs
) {
1701 blk_mq_free_tags(tags
);
1708 static size_t order_to_size(unsigned int order
)
1710 return (size_t)PAGE_SIZE
<< order
;
1713 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1714 unsigned int hctx_idx
, unsigned int depth
)
1716 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1717 size_t rq_size
, left
;
1719 INIT_LIST_HEAD(&tags
->page_list
);
1722 * rq_size is the size of the request plus driver payload, rounded
1723 * to the cacheline size
1725 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1727 left
= rq_size
* depth
;
1729 for (i
= 0; i
< depth
; ) {
1730 int this_order
= max_order
;
1735 while (this_order
&& left
< order_to_size(this_order
- 1))
1739 page
= alloc_pages_node(set
->numa_node
,
1740 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1746 if (order_to_size(this_order
) < rq_size
)
1753 page
->private = this_order
;
1754 list_add_tail(&page
->lru
, &tags
->page_list
);
1756 p
= page_address(page
);
1758 * Allow kmemleak to scan these pages as they contain pointers
1759 * to additional allocations like via ops->init_request().
1761 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1762 entries_per_page
= order_to_size(this_order
) / rq_size
;
1763 to_do
= min(entries_per_page
, depth
- i
);
1764 left
-= to_do
* rq_size
;
1765 for (j
= 0; j
< to_do
; j
++) {
1766 struct request
*rq
= p
;
1768 tags
->static_rqs
[i
] = rq
;
1769 if (set
->ops
->init_request
) {
1770 if (set
->ops
->init_request(set
->driver_data
,
1773 tags
->static_rqs
[i
] = NULL
;
1785 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1790 * 'cpu' is going away. splice any existing rq_list entries from this
1791 * software queue to the hw queue dispatch list, and ensure that it
1794 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1796 struct blk_mq_hw_ctx
*hctx
;
1797 struct blk_mq_ctx
*ctx
;
1800 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1801 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1803 spin_lock(&ctx
->lock
);
1804 if (!list_empty(&ctx
->rq_list
)) {
1805 list_splice_init(&ctx
->rq_list
, &tmp
);
1806 blk_mq_hctx_clear_pending(hctx
, ctx
);
1808 spin_unlock(&ctx
->lock
);
1810 if (list_empty(&tmp
))
1813 spin_lock(&hctx
->lock
);
1814 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1815 spin_unlock(&hctx
->lock
);
1817 blk_mq_run_hw_queue(hctx
, true);
1821 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1823 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1827 /* hctx->ctxs will be freed in queue's release handler */
1828 static void blk_mq_exit_hctx(struct request_queue
*q
,
1829 struct blk_mq_tag_set
*set
,
1830 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1832 unsigned flush_start_tag
= set
->queue_depth
;
1834 blk_mq_tag_idle(hctx
);
1836 if (set
->ops
->exit_request
)
1837 set
->ops
->exit_request(set
->driver_data
,
1838 hctx
->fq
->flush_rq
, hctx_idx
,
1839 flush_start_tag
+ hctx_idx
);
1841 if (set
->ops
->exit_hctx
)
1842 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1844 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1845 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1847 blk_mq_remove_cpuhp(hctx
);
1848 blk_free_flush_queue(hctx
->fq
);
1849 sbitmap_free(&hctx
->ctx_map
);
1852 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1853 struct blk_mq_tag_set
*set
, int nr_queue
)
1855 struct blk_mq_hw_ctx
*hctx
;
1858 queue_for_each_hw_ctx(q
, hctx
, i
) {
1861 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1865 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1866 struct blk_mq_tag_set
*set
)
1868 struct blk_mq_hw_ctx
*hctx
;
1871 queue_for_each_hw_ctx(q
, hctx
, i
)
1872 free_cpumask_var(hctx
->cpumask
);
1875 static int blk_mq_init_hctx(struct request_queue
*q
,
1876 struct blk_mq_tag_set
*set
,
1877 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1880 unsigned flush_start_tag
= set
->queue_depth
;
1882 node
= hctx
->numa_node
;
1883 if (node
== NUMA_NO_NODE
)
1884 node
= hctx
->numa_node
= set
->numa_node
;
1886 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1887 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1888 spin_lock_init(&hctx
->lock
);
1889 INIT_LIST_HEAD(&hctx
->dispatch
);
1891 hctx
->queue_num
= hctx_idx
;
1892 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1894 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1896 hctx
->tags
= set
->tags
[hctx_idx
];
1899 * Allocate space for all possible cpus to avoid allocation at
1902 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1905 goto unregister_cpu_notifier
;
1907 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1913 if (set
->ops
->init_hctx
&&
1914 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1917 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1921 if (set
->ops
->init_request
&&
1922 set
->ops
->init_request(set
->driver_data
,
1923 hctx
->fq
->flush_rq
, hctx_idx
,
1924 flush_start_tag
+ hctx_idx
, node
))
1927 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1928 init_srcu_struct(&hctx
->queue_rq_srcu
);
1935 if (set
->ops
->exit_hctx
)
1936 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1938 sbitmap_free(&hctx
->ctx_map
);
1941 unregister_cpu_notifier
:
1942 blk_mq_remove_cpuhp(hctx
);
1946 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1947 unsigned int nr_hw_queues
)
1951 for_each_possible_cpu(i
) {
1952 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1953 struct blk_mq_hw_ctx
*hctx
;
1955 memset(__ctx
, 0, sizeof(*__ctx
));
1957 spin_lock_init(&__ctx
->lock
);
1958 INIT_LIST_HEAD(&__ctx
->rq_list
);
1960 blk_stat_init(&__ctx
->stat
[BLK_STAT_READ
]);
1961 blk_stat_init(&__ctx
->stat
[BLK_STAT_WRITE
]);
1963 /* If the cpu isn't online, the cpu is mapped to first hctx */
1967 hctx
= blk_mq_map_queue(q
, i
);
1970 * Set local node, IFF we have more than one hw queue. If
1971 * not, we remain on the home node of the device
1973 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1974 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1978 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
1982 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
1983 set
->queue_depth
, set
->reserved_tags
);
1984 if (!set
->tags
[hctx_idx
])
1987 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
1992 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
1993 set
->tags
[hctx_idx
] = NULL
;
1997 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
1998 unsigned int hctx_idx
)
2000 if (set
->tags
[hctx_idx
]) {
2001 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2002 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2003 set
->tags
[hctx_idx
] = NULL
;
2007 static void blk_mq_map_swqueue(struct request_queue
*q
,
2008 const struct cpumask
*online_mask
)
2010 unsigned int i
, hctx_idx
;
2011 struct blk_mq_hw_ctx
*hctx
;
2012 struct blk_mq_ctx
*ctx
;
2013 struct blk_mq_tag_set
*set
= q
->tag_set
;
2016 * Avoid others reading imcomplete hctx->cpumask through sysfs
2018 mutex_lock(&q
->sysfs_lock
);
2020 queue_for_each_hw_ctx(q
, hctx
, i
) {
2021 cpumask_clear(hctx
->cpumask
);
2026 * Map software to hardware queues
2028 for_each_possible_cpu(i
) {
2029 /* If the cpu isn't online, the cpu is mapped to first hctx */
2030 if (!cpumask_test_cpu(i
, online_mask
))
2033 hctx_idx
= q
->mq_map
[i
];
2034 /* unmapped hw queue can be remapped after CPU topo changed */
2035 if (!set
->tags
[hctx_idx
] &&
2036 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2038 * If tags initialization fail for some hctx,
2039 * that hctx won't be brought online. In this
2040 * case, remap the current ctx to hctx[0] which
2041 * is guaranteed to always have tags allocated
2046 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2047 hctx
= blk_mq_map_queue(q
, i
);
2049 cpumask_set_cpu(i
, hctx
->cpumask
);
2050 ctx
->index_hw
= hctx
->nr_ctx
;
2051 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2054 mutex_unlock(&q
->sysfs_lock
);
2056 queue_for_each_hw_ctx(q
, hctx
, i
) {
2058 * If no software queues are mapped to this hardware queue,
2059 * disable it and free the request entries.
2061 if (!hctx
->nr_ctx
) {
2062 /* Never unmap queue 0. We need it as a
2063 * fallback in case of a new remap fails
2066 if (i
&& set
->tags
[i
])
2067 blk_mq_free_map_and_requests(set
, i
);
2073 hctx
->tags
= set
->tags
[i
];
2074 WARN_ON(!hctx
->tags
);
2077 * Set the map size to the number of mapped software queues.
2078 * This is more accurate and more efficient than looping
2079 * over all possibly mapped software queues.
2081 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2084 * Initialize batch roundrobin counts
2086 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2087 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2091 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2093 struct blk_mq_hw_ctx
*hctx
;
2096 queue_for_each_hw_ctx(q
, hctx
, i
) {
2098 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2100 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2104 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
2106 struct request_queue
*q
;
2108 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2109 blk_mq_freeze_queue(q
);
2110 queue_set_hctx_shared(q
, shared
);
2111 blk_mq_unfreeze_queue(q
);
2115 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2117 struct blk_mq_tag_set
*set
= q
->tag_set
;
2119 mutex_lock(&set
->tag_list_lock
);
2120 list_del_init(&q
->tag_set_list
);
2121 if (list_is_singular(&set
->tag_list
)) {
2122 /* just transitioned to unshared */
2123 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2124 /* update existing queue */
2125 blk_mq_update_tag_set_depth(set
, false);
2127 mutex_unlock(&set
->tag_list_lock
);
2130 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2131 struct request_queue
*q
)
2135 mutex_lock(&set
->tag_list_lock
);
2137 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2138 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2139 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2140 /* update existing queue */
2141 blk_mq_update_tag_set_depth(set
, true);
2143 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2144 queue_set_hctx_shared(q
, true);
2145 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
2147 mutex_unlock(&set
->tag_list_lock
);
2151 * It is the actual release handler for mq, but we do it from
2152 * request queue's release handler for avoiding use-after-free
2153 * and headache because q->mq_kobj shouldn't have been introduced,
2154 * but we can't group ctx/kctx kobj without it.
2156 void blk_mq_release(struct request_queue
*q
)
2158 struct blk_mq_hw_ctx
*hctx
;
2161 blk_mq_sched_teardown(q
);
2163 /* hctx kobj stays in hctx */
2164 queue_for_each_hw_ctx(q
, hctx
, i
) {
2173 kfree(q
->queue_hw_ctx
);
2175 /* ctx kobj stays in queue_ctx */
2176 free_percpu(q
->queue_ctx
);
2179 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2181 struct request_queue
*uninit_q
, *q
;
2183 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2185 return ERR_PTR(-ENOMEM
);
2187 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2189 blk_cleanup_queue(uninit_q
);
2193 EXPORT_SYMBOL(blk_mq_init_queue
);
2195 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2196 struct request_queue
*q
)
2199 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2201 blk_mq_sysfs_unregister(q
);
2202 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2208 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2209 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2214 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2221 atomic_set(&hctxs
[i
]->nr_active
, 0);
2222 hctxs
[i
]->numa_node
= node
;
2223 hctxs
[i
]->queue_num
= i
;
2225 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2226 free_cpumask_var(hctxs
[i
]->cpumask
);
2231 blk_mq_hctx_kobj_init(hctxs
[i
]);
2233 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2234 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2238 blk_mq_free_map_and_requests(set
, j
);
2239 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2240 free_cpumask_var(hctx
->cpumask
);
2241 kobject_put(&hctx
->kobj
);
2248 q
->nr_hw_queues
= i
;
2249 blk_mq_sysfs_register(q
);
2252 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2253 struct request_queue
*q
)
2255 /* mark the queue as mq asap */
2256 q
->mq_ops
= set
->ops
;
2258 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2262 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2263 GFP_KERNEL
, set
->numa_node
);
2264 if (!q
->queue_hw_ctx
)
2267 q
->mq_map
= set
->mq_map
;
2269 blk_mq_realloc_hw_ctxs(set
, q
);
2270 if (!q
->nr_hw_queues
)
2273 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2274 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2276 q
->nr_queues
= nr_cpu_ids
;
2278 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2280 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2281 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2283 q
->sg_reserved_size
= INT_MAX
;
2285 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2286 INIT_LIST_HEAD(&q
->requeue_list
);
2287 spin_lock_init(&q
->requeue_lock
);
2289 if (q
->nr_hw_queues
> 1)
2290 blk_queue_make_request(q
, blk_mq_make_request
);
2292 blk_queue_make_request(q
, blk_sq_make_request
);
2295 * Do this after blk_queue_make_request() overrides it...
2297 q
->nr_requests
= set
->queue_depth
;
2300 * Default to classic polling
2304 if (set
->ops
->complete
)
2305 blk_queue_softirq_done(q
, set
->ops
->complete
);
2307 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2310 mutex_lock(&all_q_mutex
);
2312 list_add_tail(&q
->all_q_node
, &all_q_list
);
2313 blk_mq_add_queue_tag_set(set
, q
);
2314 blk_mq_map_swqueue(q
, cpu_online_mask
);
2316 mutex_unlock(&all_q_mutex
);
2319 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2322 ret
= blk_mq_sched_init(q
);
2324 return ERR_PTR(ret
);
2330 kfree(q
->queue_hw_ctx
);
2332 free_percpu(q
->queue_ctx
);
2335 return ERR_PTR(-ENOMEM
);
2337 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2339 void blk_mq_free_queue(struct request_queue
*q
)
2341 struct blk_mq_tag_set
*set
= q
->tag_set
;
2343 mutex_lock(&all_q_mutex
);
2344 list_del_init(&q
->all_q_node
);
2345 mutex_unlock(&all_q_mutex
);
2349 blk_mq_del_queue_tag_set(q
);
2351 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2352 blk_mq_free_hw_queues(q
, set
);
2355 /* Basically redo blk_mq_init_queue with queue frozen */
2356 static void blk_mq_queue_reinit(struct request_queue
*q
,
2357 const struct cpumask
*online_mask
)
2359 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2361 blk_mq_sysfs_unregister(q
);
2364 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2365 * we should change hctx numa_node according to new topology (this
2366 * involves free and re-allocate memory, worthy doing?)
2369 blk_mq_map_swqueue(q
, online_mask
);
2371 blk_mq_sysfs_register(q
);
2375 * New online cpumask which is going to be set in this hotplug event.
2376 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2377 * one-by-one and dynamically allocating this could result in a failure.
2379 static struct cpumask cpuhp_online_new
;
2381 static void blk_mq_queue_reinit_work(void)
2383 struct request_queue
*q
;
2385 mutex_lock(&all_q_mutex
);
2387 * We need to freeze and reinit all existing queues. Freezing
2388 * involves synchronous wait for an RCU grace period and doing it
2389 * one by one may take a long time. Start freezing all queues in
2390 * one swoop and then wait for the completions so that freezing can
2391 * take place in parallel.
2393 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2394 blk_mq_freeze_queue_start(q
);
2395 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2396 blk_mq_freeze_queue_wait(q
);
2398 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2399 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2401 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2402 blk_mq_unfreeze_queue(q
);
2404 mutex_unlock(&all_q_mutex
);
2407 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2409 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2410 blk_mq_queue_reinit_work();
2415 * Before hotadded cpu starts handling requests, new mappings must be
2416 * established. Otherwise, these requests in hw queue might never be
2419 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2420 * for CPU0, and ctx1 for CPU1).
2422 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2423 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2425 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2426 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2427 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2430 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2432 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2433 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2434 blk_mq_queue_reinit_work();
2438 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2442 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2443 if (!__blk_mq_alloc_rq_map(set
, i
))
2450 blk_mq_free_rq_map(set
->tags
[i
]);
2456 * Allocate the request maps associated with this tag_set. Note that this
2457 * may reduce the depth asked for, if memory is tight. set->queue_depth
2458 * will be updated to reflect the allocated depth.
2460 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2465 depth
= set
->queue_depth
;
2467 err
= __blk_mq_alloc_rq_maps(set
);
2471 set
->queue_depth
>>= 1;
2472 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2476 } while (set
->queue_depth
);
2478 if (!set
->queue_depth
|| err
) {
2479 pr_err("blk-mq: failed to allocate request map\n");
2483 if (depth
!= set
->queue_depth
)
2484 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2485 depth
, set
->queue_depth
);
2491 * Alloc a tag set to be associated with one or more request queues.
2492 * May fail with EINVAL for various error conditions. May adjust the
2493 * requested depth down, if if it too large. In that case, the set
2494 * value will be stored in set->queue_depth.
2496 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2500 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2502 if (!set
->nr_hw_queues
)
2504 if (!set
->queue_depth
)
2506 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2509 if (!set
->ops
->queue_rq
)
2512 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2513 pr_info("blk-mq: reduced tag depth to %u\n",
2515 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2519 * If a crashdump is active, then we are potentially in a very
2520 * memory constrained environment. Limit us to 1 queue and
2521 * 64 tags to prevent using too much memory.
2523 if (is_kdump_kernel()) {
2524 set
->nr_hw_queues
= 1;
2525 set
->queue_depth
= min(64U, set
->queue_depth
);
2528 * There is no use for more h/w queues than cpus.
2530 if (set
->nr_hw_queues
> nr_cpu_ids
)
2531 set
->nr_hw_queues
= nr_cpu_ids
;
2533 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2534 GFP_KERNEL
, set
->numa_node
);
2539 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2540 GFP_KERNEL
, set
->numa_node
);
2544 if (set
->ops
->map_queues
)
2545 ret
= set
->ops
->map_queues(set
);
2547 ret
= blk_mq_map_queues(set
);
2549 goto out_free_mq_map
;
2551 ret
= blk_mq_alloc_rq_maps(set
);
2553 goto out_free_mq_map
;
2555 mutex_init(&set
->tag_list_lock
);
2556 INIT_LIST_HEAD(&set
->tag_list
);
2568 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2570 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2574 for (i
= 0; i
< nr_cpu_ids
; i
++)
2575 blk_mq_free_map_and_requests(set
, i
);
2583 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2585 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2587 struct blk_mq_tag_set
*set
= q
->tag_set
;
2588 struct blk_mq_hw_ctx
*hctx
;
2594 blk_mq_freeze_queue(q
);
2595 blk_mq_quiesce_queue(q
);
2598 queue_for_each_hw_ctx(q
, hctx
, i
) {
2602 * If we're using an MQ scheduler, just update the scheduler
2603 * queue depth. This is similar to what the old code would do.
2605 if (!hctx
->sched_tags
) {
2606 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2607 min(nr
, set
->queue_depth
),
2610 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2618 q
->nr_requests
= nr
;
2620 blk_mq_unfreeze_queue(q
);
2621 blk_mq_start_stopped_hw_queues(q
, true);
2626 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2628 struct request_queue
*q
;
2630 if (nr_hw_queues
> nr_cpu_ids
)
2631 nr_hw_queues
= nr_cpu_ids
;
2632 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2635 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2636 blk_mq_freeze_queue(q
);
2638 set
->nr_hw_queues
= nr_hw_queues
;
2639 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2640 blk_mq_realloc_hw_ctxs(set
, q
);
2642 if (q
->nr_hw_queues
> 1)
2643 blk_queue_make_request(q
, blk_mq_make_request
);
2645 blk_queue_make_request(q
, blk_sq_make_request
);
2647 blk_mq_queue_reinit(q
, cpu_online_mask
);
2650 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2651 blk_mq_unfreeze_queue(q
);
2653 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2655 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2656 struct blk_mq_hw_ctx
*hctx
,
2659 struct blk_rq_stat stat
[2];
2660 unsigned long ret
= 0;
2663 * If stats collection isn't on, don't sleep but turn it on for
2666 if (!blk_stat_enable(q
))
2670 * We don't have to do this once per IO, should optimize this
2671 * to just use the current window of stats until it changes
2673 memset(&stat
, 0, sizeof(stat
));
2674 blk_hctx_stat_get(hctx
, stat
);
2677 * As an optimistic guess, use half of the mean service time
2678 * for this type of request. We can (and should) make this smarter.
2679 * For instance, if the completion latencies are tight, we can
2680 * get closer than just half the mean. This is especially
2681 * important on devices where the completion latencies are longer
2684 if (req_op(rq
) == REQ_OP_READ
&& stat
[BLK_STAT_READ
].nr_samples
)
2685 ret
= (stat
[BLK_STAT_READ
].mean
+ 1) / 2;
2686 else if (req_op(rq
) == REQ_OP_WRITE
&& stat
[BLK_STAT_WRITE
].nr_samples
)
2687 ret
= (stat
[BLK_STAT_WRITE
].mean
+ 1) / 2;
2692 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2693 struct blk_mq_hw_ctx
*hctx
,
2696 struct hrtimer_sleeper hs
;
2697 enum hrtimer_mode mode
;
2701 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2707 * -1: don't ever hybrid sleep
2708 * 0: use half of prev avg
2709 * >0: use this specific value
2711 if (q
->poll_nsec
== -1)
2713 else if (q
->poll_nsec
> 0)
2714 nsecs
= q
->poll_nsec
;
2716 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2721 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2724 * This will be replaced with the stats tracking code, using
2725 * 'avg_completion_time / 2' as the pre-sleep target.
2729 mode
= HRTIMER_MODE_REL
;
2730 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2731 hrtimer_set_expires(&hs
.timer
, kt
);
2733 hrtimer_init_sleeper(&hs
, current
);
2735 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2737 set_current_state(TASK_UNINTERRUPTIBLE
);
2738 hrtimer_start_expires(&hs
.timer
, mode
);
2741 hrtimer_cancel(&hs
.timer
);
2742 mode
= HRTIMER_MODE_ABS
;
2743 } while (hs
.task
&& !signal_pending(current
));
2745 __set_current_state(TASK_RUNNING
);
2746 destroy_hrtimer_on_stack(&hs
.timer
);
2750 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2752 struct request_queue
*q
= hctx
->queue
;
2756 * If we sleep, have the caller restart the poll loop to reset
2757 * the state. Like for the other success return cases, the
2758 * caller is responsible for checking if the IO completed. If
2759 * the IO isn't complete, we'll get called again and will go
2760 * straight to the busy poll loop.
2762 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2765 hctx
->poll_considered
++;
2767 state
= current
->state
;
2768 while (!need_resched()) {
2771 hctx
->poll_invoked
++;
2773 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2775 hctx
->poll_success
++;
2776 set_current_state(TASK_RUNNING
);
2780 if (signal_pending_state(state
, current
))
2781 set_current_state(TASK_RUNNING
);
2783 if (current
->state
== TASK_RUNNING
)
2793 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2795 struct blk_mq_hw_ctx
*hctx
;
2796 struct blk_plug
*plug
;
2799 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2800 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2803 plug
= current
->plug
;
2805 blk_flush_plug_list(plug
, false);
2807 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2808 if (!blk_qc_t_is_internal(cookie
))
2809 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2811 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2813 return __blk_mq_poll(hctx
, rq
);
2815 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2817 void blk_mq_disable_hotplug(void)
2819 mutex_lock(&all_q_mutex
);
2822 void blk_mq_enable_hotplug(void)
2824 mutex_unlock(&all_q_mutex
);
2827 static int __init
blk_mq_init(void)
2829 blk_mq_debugfs_init();
2831 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2832 blk_mq_hctx_notify_dead
);
2834 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2835 blk_mq_queue_reinit_prepare
,
2836 blk_mq_queue_reinit_dead
);
2839 subsys_initcall(blk_mq_init
);