2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
29 #include <trace/events/block.h>
31 #include <linux/blk-mq.h>
34 #include "blk-mq-debugfs.h"
35 #include "blk-mq-tag.h"
38 #include "blk-mq-sched.h"
40 static DEFINE_MUTEX(all_q_mutex
);
41 static LIST_HEAD(all_q_list
);
43 static void blk_mq_poll_stats_start(struct request_queue
*q
);
44 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
45 static void __blk_mq_stop_hw_queues(struct request_queue
*q
, bool sync
);
47 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
49 int ddir
, bytes
, bucket
;
51 ddir
= rq_data_dir(rq
);
52 bytes
= blk_rq_bytes(rq
);
54 bucket
= ddir
+ 2*(ilog2(bytes
) - 9);
58 else if (bucket
>= BLK_MQ_POLL_STATS_BKTS
)
59 return ddir
+ BLK_MQ_POLL_STATS_BKTS
- 2;
65 * Check if any of the ctx's have pending work in this hardware queue
67 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
69 return sbitmap_any_bit_set(&hctx
->ctx_map
) ||
70 !list_empty_careful(&hctx
->dispatch
) ||
71 blk_mq_sched_has_work(hctx
);
75 * Mark this ctx as having pending work in this hardware queue
77 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
78 struct blk_mq_ctx
*ctx
)
80 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
81 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
84 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
85 struct blk_mq_ctx
*ctx
)
87 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
90 void blk_freeze_queue_start(struct request_queue
*q
)
94 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
95 if (freeze_depth
== 1) {
96 percpu_ref_kill(&q
->q_usage_counter
);
97 blk_mq_run_hw_queues(q
, false);
100 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
102 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
104 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
106 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
108 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
109 unsigned long timeout
)
111 return wait_event_timeout(q
->mq_freeze_wq
,
112 percpu_ref_is_zero(&q
->q_usage_counter
),
115 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
118 * Guarantee no request is in use, so we can change any data structure of
119 * the queue afterward.
121 void blk_freeze_queue(struct request_queue
*q
)
124 * In the !blk_mq case we are only calling this to kill the
125 * q_usage_counter, otherwise this increases the freeze depth
126 * and waits for it to return to zero. For this reason there is
127 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
128 * exported to drivers as the only user for unfreeze is blk_mq.
130 blk_freeze_queue_start(q
);
131 blk_mq_freeze_queue_wait(q
);
134 void blk_mq_freeze_queue(struct request_queue
*q
)
137 * ...just an alias to keep freeze and unfreeze actions balanced
138 * in the blk_mq_* namespace
142 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
144 void blk_mq_unfreeze_queue(struct request_queue
*q
)
148 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
149 WARN_ON_ONCE(freeze_depth
< 0);
151 percpu_ref_reinit(&q
->q_usage_counter
);
152 wake_up_all(&q
->mq_freeze_wq
);
155 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
158 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
161 * Note: this function does not prevent that the struct request end_io()
162 * callback function is invoked. Additionally, it is not prevented that
163 * new queue_rq() calls occur unless the queue has been stopped first.
165 void blk_mq_quiesce_queue(struct request_queue
*q
)
167 struct blk_mq_hw_ctx
*hctx
;
171 __blk_mq_stop_hw_queues(q
, true);
173 queue_for_each_hw_ctx(q
, hctx
, i
) {
174 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
175 synchronize_srcu(&hctx
->queue_rq_srcu
);
182 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
184 void blk_mq_wake_waiters(struct request_queue
*q
)
186 struct blk_mq_hw_ctx
*hctx
;
189 queue_for_each_hw_ctx(q
, hctx
, i
)
190 if (blk_mq_hw_queue_mapped(hctx
))
191 blk_mq_tag_wakeup_all(hctx
->tags
, true);
194 * If we are called because the queue has now been marked as
195 * dying, we need to ensure that processes currently waiting on
196 * the queue are notified as well.
198 wake_up_all(&q
->mq_freeze_wq
);
201 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
203 return blk_mq_has_free_tags(hctx
->tags
);
205 EXPORT_SYMBOL(blk_mq_can_queue
);
207 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
208 struct request
*rq
, unsigned int op
)
210 INIT_LIST_HEAD(&rq
->queuelist
);
211 /* csd/requeue_work/fifo_time is initialized before use */
215 if (blk_queue_io_stat(q
))
216 rq
->rq_flags
|= RQF_IO_STAT
;
217 /* do not touch atomic flags, it needs atomic ops against the timer */
219 INIT_HLIST_NODE(&rq
->hash
);
220 RB_CLEAR_NODE(&rq
->rb_node
);
223 rq
->start_time
= jiffies
;
224 #ifdef CONFIG_BLK_CGROUP
226 set_start_time_ns(rq
);
227 rq
->io_start_time_ns
= 0;
229 rq
->nr_phys_segments
= 0;
230 #if defined(CONFIG_BLK_DEV_INTEGRITY)
231 rq
->nr_integrity_segments
= 0;
234 /* tag was already set */
237 INIT_LIST_HEAD(&rq
->timeout_list
);
241 rq
->end_io_data
= NULL
;
244 ctx
->rq_dispatched
[op_is_sync(op
)]++;
247 struct request
*__blk_mq_alloc_request(struct blk_mq_alloc_data
*data
,
253 tag
= blk_mq_get_tag(data
);
254 if (tag
!= BLK_MQ_TAG_FAIL
) {
255 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
257 rq
= tags
->static_rqs
[tag
];
259 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
261 rq
->internal_tag
= tag
;
263 if (blk_mq_tag_busy(data
->hctx
)) {
264 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
265 atomic_inc(&data
->hctx
->nr_active
);
268 rq
->internal_tag
= -1;
269 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
272 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
);
278 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request
);
280 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
283 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
287 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
291 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
293 blk_mq_put_ctx(alloc_data
.ctx
);
297 return ERR_PTR(-EWOULDBLOCK
);
300 rq
->__sector
= (sector_t
) -1;
301 rq
->bio
= rq
->biotail
= NULL
;
304 EXPORT_SYMBOL(blk_mq_alloc_request
);
306 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
307 unsigned int flags
, unsigned int hctx_idx
)
309 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
315 * If the tag allocator sleeps we could get an allocation for a
316 * different hardware context. No need to complicate the low level
317 * allocator for this for the rare use case of a command tied to
320 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
321 return ERR_PTR(-EINVAL
);
323 if (hctx_idx
>= q
->nr_hw_queues
)
324 return ERR_PTR(-EIO
);
326 ret
= blk_queue_enter(q
, true);
331 * Check if the hardware context is actually mapped to anything.
332 * If not tell the caller that it should skip this queue.
334 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
335 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
337 return ERR_PTR(-EXDEV
);
339 cpu
= cpumask_first(alloc_data
.hctx
->cpumask
);
340 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
342 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
347 return ERR_PTR(-EWOULDBLOCK
);
351 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
353 void __blk_mq_finish_request(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
356 const int sched_tag
= rq
->internal_tag
;
357 struct request_queue
*q
= rq
->q
;
359 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
360 atomic_dec(&hctx
->nr_active
);
362 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
365 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
366 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
368 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
370 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
371 blk_mq_sched_restart(hctx
);
375 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx
*hctx
,
378 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
380 ctx
->rq_completed
[rq_is_sync(rq
)]++;
381 __blk_mq_finish_request(hctx
, ctx
, rq
);
384 void blk_mq_finish_request(struct request
*rq
)
386 blk_mq_finish_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
388 EXPORT_SYMBOL_GPL(blk_mq_finish_request
);
390 void blk_mq_free_request(struct request
*rq
)
392 blk_mq_sched_put_request(rq
);
394 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
396 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
398 blk_account_io_done(rq
);
401 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
402 rq
->end_io(rq
, error
);
404 if (unlikely(blk_bidi_rq(rq
)))
405 blk_mq_free_request(rq
->next_rq
);
406 blk_mq_free_request(rq
);
409 EXPORT_SYMBOL(__blk_mq_end_request
);
411 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
413 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
415 __blk_mq_end_request(rq
, error
);
417 EXPORT_SYMBOL(blk_mq_end_request
);
419 static void __blk_mq_complete_request_remote(void *data
)
421 struct request
*rq
= data
;
423 rq
->q
->softirq_done_fn(rq
);
426 static void __blk_mq_complete_request(struct request
*rq
)
428 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
432 if (rq
->internal_tag
!= -1)
433 blk_mq_sched_completed_request(rq
);
434 if (rq
->rq_flags
& RQF_STATS
) {
435 blk_mq_poll_stats_start(rq
->q
);
439 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
440 rq
->q
->softirq_done_fn(rq
);
445 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
446 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
448 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
449 rq
->csd
.func
= __blk_mq_complete_request_remote
;
452 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
454 rq
->q
->softirq_done_fn(rq
);
460 * blk_mq_complete_request - end I/O on a request
461 * @rq: the request being processed
464 * Ends all I/O on a request. It does not handle partial completions.
465 * The actual completion happens out-of-order, through a IPI handler.
467 void blk_mq_complete_request(struct request
*rq
)
469 struct request_queue
*q
= rq
->q
;
471 if (unlikely(blk_should_fake_timeout(q
)))
473 if (!blk_mark_rq_complete(rq
))
474 __blk_mq_complete_request(rq
);
476 EXPORT_SYMBOL(blk_mq_complete_request
);
478 int blk_mq_request_started(struct request
*rq
)
480 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
482 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
484 void blk_mq_start_request(struct request
*rq
)
486 struct request_queue
*q
= rq
->q
;
488 blk_mq_sched_started_request(rq
);
490 trace_block_rq_issue(q
, rq
);
492 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
493 blk_stat_set_issue(&rq
->issue_stat
, blk_rq_sectors(rq
));
494 rq
->rq_flags
|= RQF_STATS
;
495 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
501 * Ensure that ->deadline is visible before set the started
502 * flag and clear the completed flag.
504 smp_mb__before_atomic();
507 * Mark us as started and clear complete. Complete might have been
508 * set if requeue raced with timeout, which then marked it as
509 * complete. So be sure to clear complete again when we start
510 * the request, otherwise we'll ignore the completion event.
512 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
513 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
514 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
515 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
517 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
519 * Make sure space for the drain appears. We know we can do
520 * this because max_hw_segments has been adjusted to be one
521 * fewer than the device can handle.
523 rq
->nr_phys_segments
++;
526 EXPORT_SYMBOL(blk_mq_start_request
);
529 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
530 * flag isn't set yet, so there may be race with timeout handler,
531 * but given rq->deadline is just set in .queue_rq() under
532 * this situation, the race won't be possible in reality because
533 * rq->timeout should be set as big enough to cover the window
534 * between blk_mq_start_request() called from .queue_rq() and
535 * clearing REQ_ATOM_STARTED here.
537 static void __blk_mq_requeue_request(struct request
*rq
)
539 struct request_queue
*q
= rq
->q
;
541 trace_block_rq_requeue(q
, rq
);
542 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
543 blk_mq_sched_requeue_request(rq
);
545 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
546 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
547 rq
->nr_phys_segments
--;
551 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
553 __blk_mq_requeue_request(rq
);
555 BUG_ON(blk_queued_rq(rq
));
556 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
558 EXPORT_SYMBOL(blk_mq_requeue_request
);
560 static void blk_mq_requeue_work(struct work_struct
*work
)
562 struct request_queue
*q
=
563 container_of(work
, struct request_queue
, requeue_work
.work
);
565 struct request
*rq
, *next
;
568 spin_lock_irqsave(&q
->requeue_lock
, flags
);
569 list_splice_init(&q
->requeue_list
, &rq_list
);
570 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
572 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
573 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
576 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
577 list_del_init(&rq
->queuelist
);
578 blk_mq_sched_insert_request(rq
, true, false, false, true);
581 while (!list_empty(&rq_list
)) {
582 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
583 list_del_init(&rq
->queuelist
);
584 blk_mq_sched_insert_request(rq
, false, false, false, true);
587 blk_mq_run_hw_queues(q
, false);
590 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
591 bool kick_requeue_list
)
593 struct request_queue
*q
= rq
->q
;
597 * We abuse this flag that is otherwise used by the I/O scheduler to
598 * request head insertation from the workqueue.
600 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
602 spin_lock_irqsave(&q
->requeue_lock
, flags
);
604 rq
->rq_flags
|= RQF_SOFTBARRIER
;
605 list_add(&rq
->queuelist
, &q
->requeue_list
);
607 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
609 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
611 if (kick_requeue_list
)
612 blk_mq_kick_requeue_list(q
);
614 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
616 void blk_mq_kick_requeue_list(struct request_queue
*q
)
618 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
620 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
622 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
625 kblockd_schedule_delayed_work(&q
->requeue_work
,
626 msecs_to_jiffies(msecs
));
628 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
630 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
632 if (tag
< tags
->nr_tags
) {
633 prefetch(tags
->rqs
[tag
]);
634 return tags
->rqs
[tag
];
639 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
641 struct blk_mq_timeout_data
{
643 unsigned int next_set
;
646 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
648 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
649 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
652 * We know that complete is set at this point. If STARTED isn't set
653 * anymore, then the request isn't active and the "timeout" should
654 * just be ignored. This can happen due to the bitflag ordering.
655 * Timeout first checks if STARTED is set, and if it is, assumes
656 * the request is active. But if we race with completion, then
657 * both flags will get cleared. So check here again, and ignore
658 * a timeout event with a request that isn't active.
660 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
664 ret
= ops
->timeout(req
, reserved
);
668 __blk_mq_complete_request(req
);
670 case BLK_EH_RESET_TIMER
:
672 blk_clear_rq_complete(req
);
674 case BLK_EH_NOT_HANDLED
:
677 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
682 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
683 struct request
*rq
, void *priv
, bool reserved
)
685 struct blk_mq_timeout_data
*data
= priv
;
687 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
691 * The rq being checked may have been freed and reallocated
692 * out already here, we avoid this race by checking rq->deadline
693 * and REQ_ATOM_COMPLETE flag together:
695 * - if rq->deadline is observed as new value because of
696 * reusing, the rq won't be timed out because of timing.
697 * - if rq->deadline is observed as previous value,
698 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
699 * because we put a barrier between setting rq->deadline
700 * and clearing the flag in blk_mq_start_request(), so
701 * this rq won't be timed out too.
703 if (time_after_eq(jiffies
, rq
->deadline
)) {
704 if (!blk_mark_rq_complete(rq
))
705 blk_mq_rq_timed_out(rq
, reserved
);
706 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
707 data
->next
= rq
->deadline
;
712 static void blk_mq_timeout_work(struct work_struct
*work
)
714 struct request_queue
*q
=
715 container_of(work
, struct request_queue
, timeout_work
);
716 struct blk_mq_timeout_data data
= {
722 /* A deadlock might occur if a request is stuck requiring a
723 * timeout at the same time a queue freeze is waiting
724 * completion, since the timeout code would not be able to
725 * acquire the queue reference here.
727 * That's why we don't use blk_queue_enter here; instead, we use
728 * percpu_ref_tryget directly, because we need to be able to
729 * obtain a reference even in the short window between the queue
730 * starting to freeze, by dropping the first reference in
731 * blk_freeze_queue_start, and the moment the last request is
732 * consumed, marked by the instant q_usage_counter reaches
735 if (!percpu_ref_tryget(&q
->q_usage_counter
))
738 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
741 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
742 mod_timer(&q
->timeout
, data
.next
);
744 struct blk_mq_hw_ctx
*hctx
;
746 queue_for_each_hw_ctx(q
, hctx
, i
) {
747 /* the hctx may be unmapped, so check it here */
748 if (blk_mq_hw_queue_mapped(hctx
))
749 blk_mq_tag_idle(hctx
);
755 struct flush_busy_ctx_data
{
756 struct blk_mq_hw_ctx
*hctx
;
757 struct list_head
*list
;
760 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
762 struct flush_busy_ctx_data
*flush_data
= data
;
763 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
764 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
766 sbitmap_clear_bit(sb
, bitnr
);
767 spin_lock(&ctx
->lock
);
768 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
769 spin_unlock(&ctx
->lock
);
774 * Process software queues that have been marked busy, splicing them
775 * to the for-dispatch
777 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
779 struct flush_busy_ctx_data data
= {
784 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
786 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
788 static inline unsigned int queued_to_index(unsigned int queued
)
793 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
796 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
799 struct blk_mq_alloc_data data
= {
801 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
802 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
805 might_sleep_if(wait
);
810 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
811 data
.flags
|= BLK_MQ_REQ_RESERVED
;
813 rq
->tag
= blk_mq_get_tag(&data
);
815 if (blk_mq_tag_busy(data
.hctx
)) {
816 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
817 atomic_inc(&data
.hctx
->nr_active
);
819 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
825 return rq
->tag
!= -1;
828 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx
*hctx
,
831 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
834 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
835 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
836 atomic_dec(&hctx
->nr_active
);
840 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx
*hctx
,
843 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
846 __blk_mq_put_driver_tag(hctx
, rq
);
849 static void blk_mq_put_driver_tag(struct request
*rq
)
851 struct blk_mq_hw_ctx
*hctx
;
853 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
856 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
857 __blk_mq_put_driver_tag(hctx
, rq
);
861 * If we fail getting a driver tag because all the driver tags are already
862 * assigned and on the dispatch list, BUT the first entry does not have a
863 * tag, then we could deadlock. For that case, move entries with assigned
864 * driver tags to the front, leaving the set of tagged requests in the
865 * same order, and the untagged set in the same order.
867 static bool reorder_tags_to_front(struct list_head
*list
)
869 struct request
*rq
, *tmp
, *first
= NULL
;
871 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
875 list_move(&rq
->queuelist
, list
);
881 return first
!= NULL
;
884 static int blk_mq_dispatch_wake(wait_queue_t
*wait
, unsigned mode
, int flags
,
887 struct blk_mq_hw_ctx
*hctx
;
889 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
891 list_del(&wait
->task_list
);
892 clear_bit_unlock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
);
893 blk_mq_run_hw_queue(hctx
, true);
897 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx
*hctx
)
899 struct sbq_wait_state
*ws
;
902 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
903 * The thread which wins the race to grab this bit adds the hardware
904 * queue to the wait queue.
906 if (test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
) ||
907 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
910 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
911 ws
= bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
);
914 * As soon as this returns, it's no longer safe to fiddle with
915 * hctx->dispatch_wait, since a completion can wake up the wait queue
916 * and unlock the bit.
918 add_wait_queue(&ws
->wait
, &hctx
->dispatch_wait
);
922 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
)
924 struct blk_mq_hw_ctx
*hctx
;
928 if (list_empty(list
))
932 * Now process all the entries, sending them to the driver.
936 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 * The initial allocation attempt failed, so we need to
946 * rerun the hardware queue when a tag is freed.
948 if (!blk_mq_dispatch_wait_add(hctx
))
952 * It's possible that a tag was freed in the window
953 * between the allocation failure and adding the
954 * hardware queue to the wait queue.
956 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
960 list_del_init(&rq
->queuelist
);
965 * Flag last if we have no more requests, or if we have more
966 * but can't assign a driver tag to it.
968 if (list_empty(list
))
973 nxt
= list_first_entry(list
, struct request
, queuelist
);
974 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
977 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
978 if (ret
== BLK_STS_RESOURCE
) {
979 blk_mq_put_driver_tag_hctx(hctx
, rq
);
980 list_add(&rq
->queuelist
, list
);
981 __blk_mq_requeue_request(rq
);
985 if (unlikely(ret
!= BLK_STS_OK
)) {
987 blk_mq_end_request(rq
, BLK_STS_IOERR
);
992 } while (!list_empty(list
));
994 hctx
->dispatched
[queued_to_index(queued
)]++;
997 * Any items that need requeuing? Stuff them into hctx->dispatch,
998 * that is where we will continue on next queue run.
1000 if (!list_empty(list
)) {
1002 * If an I/O scheduler has been configured and we got a driver
1003 * tag for the next request already, free it again.
1005 rq
= list_first_entry(list
, struct request
, queuelist
);
1006 blk_mq_put_driver_tag(rq
);
1008 spin_lock(&hctx
->lock
);
1009 list_splice_init(list
, &hctx
->dispatch
);
1010 spin_unlock(&hctx
->lock
);
1013 * If SCHED_RESTART was set by the caller of this function and
1014 * it is no longer set that means that it was cleared by another
1015 * thread and hence that a queue rerun is needed.
1017 * If TAG_WAITING is set that means that an I/O scheduler has
1018 * been configured and another thread is waiting for a driver
1019 * tag. To guarantee fairness, do not rerun this hardware queue
1020 * but let the other thread grab the driver tag.
1022 * If no I/O scheduler has been configured it is possible that
1023 * the hardware queue got stopped and restarted before requests
1024 * were pushed back onto the dispatch list. Rerun the queue to
1025 * avoid starvation. Notes:
1026 * - blk_mq_run_hw_queue() checks whether or not a queue has
1027 * been stopped before rerunning a queue.
1028 * - Some but not all block drivers stop a queue before
1029 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1032 if (!blk_mq_sched_needs_restart(hctx
) &&
1033 !test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1034 blk_mq_run_hw_queue(hctx
, true);
1037 return (queued
+ errors
) != 0;
1040 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1044 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1045 cpu_online(hctx
->next_cpu
));
1047 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1049 blk_mq_sched_dispatch_requests(hctx
);
1054 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1055 blk_mq_sched_dispatch_requests(hctx
);
1056 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1061 * It'd be great if the workqueue API had a way to pass
1062 * in a mask and had some smarts for more clever placement.
1063 * For now we just round-robin here, switching for every
1064 * BLK_MQ_CPU_WORK_BATCH queued items.
1066 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1068 if (hctx
->queue
->nr_hw_queues
== 1)
1069 return WORK_CPU_UNBOUND
;
1071 if (--hctx
->next_cpu_batch
<= 0) {
1074 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1075 if (next_cpu
>= nr_cpu_ids
)
1076 next_cpu
= cpumask_first(hctx
->cpumask
);
1078 hctx
->next_cpu
= next_cpu
;
1079 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1082 return hctx
->next_cpu
;
1085 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1086 unsigned long msecs
)
1088 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
1089 !blk_mq_hw_queue_mapped(hctx
)))
1092 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1093 int cpu
= get_cpu();
1094 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1095 __blk_mq_run_hw_queue(hctx
);
1103 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1105 msecs_to_jiffies(msecs
));
1108 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1110 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1112 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1114 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1116 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1118 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1120 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1122 struct blk_mq_hw_ctx
*hctx
;
1125 queue_for_each_hw_ctx(q
, hctx
, i
) {
1126 if (!blk_mq_hctx_has_pending(hctx
) ||
1127 blk_mq_hctx_stopped(hctx
))
1130 blk_mq_run_hw_queue(hctx
, async
);
1133 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1136 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1137 * @q: request queue.
1139 * The caller is responsible for serializing this function against
1140 * blk_mq_{start,stop}_hw_queue().
1142 bool blk_mq_queue_stopped(struct request_queue
*q
)
1144 struct blk_mq_hw_ctx
*hctx
;
1147 queue_for_each_hw_ctx(q
, hctx
, i
)
1148 if (blk_mq_hctx_stopped(hctx
))
1153 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1155 static void __blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool sync
)
1158 cancel_delayed_work_sync(&hctx
->run_work
);
1160 cancel_delayed_work(&hctx
->run_work
);
1162 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1165 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1167 __blk_mq_stop_hw_queue(hctx
, false);
1169 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1171 static void __blk_mq_stop_hw_queues(struct request_queue
*q
, bool sync
)
1173 struct blk_mq_hw_ctx
*hctx
;
1176 queue_for_each_hw_ctx(q
, hctx
, i
)
1177 __blk_mq_stop_hw_queue(hctx
, sync
);
1180 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1182 __blk_mq_stop_hw_queues(q
, false);
1184 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1186 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1188 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1190 blk_mq_run_hw_queue(hctx
, false);
1192 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1194 void blk_mq_start_hw_queues(struct request_queue
*q
)
1196 struct blk_mq_hw_ctx
*hctx
;
1199 queue_for_each_hw_ctx(q
, hctx
, i
)
1200 blk_mq_start_hw_queue(hctx
);
1202 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1204 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1206 if (!blk_mq_hctx_stopped(hctx
))
1209 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1210 blk_mq_run_hw_queue(hctx
, async
);
1212 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1214 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1216 struct blk_mq_hw_ctx
*hctx
;
1219 queue_for_each_hw_ctx(q
, hctx
, i
)
1220 blk_mq_start_stopped_hw_queue(hctx
, async
);
1222 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1224 static void blk_mq_run_work_fn(struct work_struct
*work
)
1226 struct blk_mq_hw_ctx
*hctx
;
1228 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1231 * If we are stopped, don't run the queue. The exception is if
1232 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1233 * the STOPPED bit and run it.
1235 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)) {
1236 if (!test_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
))
1239 clear_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1240 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1243 __blk_mq_run_hw_queue(hctx
);
1247 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1249 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1253 * Stop the hw queue, then modify currently delayed work.
1254 * This should prevent us from running the queue prematurely.
1255 * Mark the queue as auto-clearing STOPPED when it runs.
1257 blk_mq_stop_hw_queue(hctx
);
1258 set_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1259 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1261 msecs_to_jiffies(msecs
));
1263 EXPORT_SYMBOL(blk_mq_delay_queue
);
1265 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1269 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1271 trace_block_rq_insert(hctx
->queue
, rq
);
1274 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1276 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1279 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1282 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1284 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1285 blk_mq_hctx_mark_pending(hctx
, ctx
);
1288 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1289 struct list_head
*list
)
1293 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1296 spin_lock(&ctx
->lock
);
1297 while (!list_empty(list
)) {
1300 rq
= list_first_entry(list
, struct request
, queuelist
);
1301 BUG_ON(rq
->mq_ctx
!= ctx
);
1302 list_del_init(&rq
->queuelist
);
1303 __blk_mq_insert_req_list(hctx
, rq
, false);
1305 blk_mq_hctx_mark_pending(hctx
, ctx
);
1306 spin_unlock(&ctx
->lock
);
1309 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1311 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1312 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1314 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1315 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1316 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1319 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1321 struct blk_mq_ctx
*this_ctx
;
1322 struct request_queue
*this_q
;
1325 LIST_HEAD(ctx_list
);
1328 list_splice_init(&plug
->mq_list
, &list
);
1330 list_sort(NULL
, &list
, plug_ctx_cmp
);
1336 while (!list_empty(&list
)) {
1337 rq
= list_entry_rq(list
.next
);
1338 list_del_init(&rq
->queuelist
);
1340 if (rq
->mq_ctx
!= this_ctx
) {
1342 trace_block_unplug(this_q
, depth
, from_schedule
);
1343 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1348 this_ctx
= rq
->mq_ctx
;
1354 list_add_tail(&rq
->queuelist
, &ctx_list
);
1358 * If 'this_ctx' is set, we know we have entries to complete
1359 * on 'ctx_list'. Do those.
1362 trace_block_unplug(this_q
, depth
, from_schedule
);
1363 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1368 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1370 blk_init_request_from_bio(rq
, bio
);
1372 blk_account_io_start(rq
, true);
1375 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1377 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1378 !blk_queue_nomerges(hctx
->queue
);
1381 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx
*hctx
,
1382 struct blk_mq_ctx
*ctx
,
1385 spin_lock(&ctx
->lock
);
1386 __blk_mq_insert_request(hctx
, rq
, false);
1387 spin_unlock(&ctx
->lock
);
1390 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1393 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1395 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1398 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1400 blk_qc_t
*cookie
, bool may_sleep
)
1402 struct request_queue
*q
= rq
->q
;
1403 struct blk_mq_queue_data bd
= {
1407 blk_qc_t new_cookie
;
1409 bool run_queue
= true;
1411 if (blk_mq_hctx_stopped(hctx
)) {
1419 if (!blk_mq_get_driver_tag(rq
, NULL
, false))
1422 new_cookie
= request_to_qc_t(hctx
, rq
);
1425 * For OK queue, we are done. For error, kill it. Any other
1426 * error (busy), just add it to our list as we previously
1429 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1432 *cookie
= new_cookie
;
1434 case BLK_STS_RESOURCE
:
1435 __blk_mq_requeue_request(rq
);
1438 *cookie
= BLK_QC_T_NONE
;
1439 blk_mq_end_request(rq
, ret
);
1444 blk_mq_sched_insert_request(rq
, false, run_queue
, false, may_sleep
);
1447 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1448 struct request
*rq
, blk_qc_t
*cookie
)
1450 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1452 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false);
1455 unsigned int srcu_idx
;
1459 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1460 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, true);
1461 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1465 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1467 const int is_sync
= op_is_sync(bio
->bi_opf
);
1468 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1469 struct blk_mq_alloc_data data
= { .flags
= 0 };
1471 unsigned int request_count
= 0;
1472 struct blk_plug
*plug
;
1473 struct request
*same_queue_rq
= NULL
;
1475 unsigned int wb_acct
;
1477 blk_queue_bounce(q
, &bio
);
1479 blk_queue_split(q
, &bio
, q
->bio_split
);
1481 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1483 return BLK_QC_T_NONE
;
1486 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1487 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1488 return BLK_QC_T_NONE
;
1490 if (blk_mq_sched_bio_merge(q
, bio
))
1491 return BLK_QC_T_NONE
;
1493 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1495 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1497 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1498 if (unlikely(!rq
)) {
1499 __wbt_done(q
->rq_wb
, wb_acct
);
1500 return BLK_QC_T_NONE
;
1503 wbt_track(&rq
->issue_stat
, wb_acct
);
1505 cookie
= request_to_qc_t(data
.hctx
, rq
);
1507 plug
= current
->plug
;
1508 if (unlikely(is_flush_fua
)) {
1509 blk_mq_put_ctx(data
.ctx
);
1510 blk_mq_bio_to_request(rq
, bio
);
1512 blk_mq_sched_insert_request(rq
, false, true, true,
1515 blk_insert_flush(rq
);
1516 blk_mq_run_hw_queue(data
.hctx
, true);
1518 } else if (plug
&& q
->nr_hw_queues
== 1) {
1519 struct request
*last
= NULL
;
1521 blk_mq_put_ctx(data
.ctx
);
1522 blk_mq_bio_to_request(rq
, bio
);
1525 * @request_count may become stale because of schedule
1526 * out, so check the list again.
1528 if (list_empty(&plug
->mq_list
))
1530 else if (blk_queue_nomerges(q
))
1531 request_count
= blk_plug_queued_count(q
);
1534 trace_block_plug(q
);
1536 last
= list_entry_rq(plug
->mq_list
.prev
);
1538 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1539 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1540 blk_flush_plug_list(plug
, false);
1541 trace_block_plug(q
);
1544 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1545 } else if (plug
&& !blk_queue_nomerges(q
)) {
1546 blk_mq_bio_to_request(rq
, bio
);
1549 * We do limited plugging. If the bio can be merged, do that.
1550 * Otherwise the existing request in the plug list will be
1551 * issued. So the plug list will have one request at most
1552 * The plug list might get flushed before this. If that happens,
1553 * the plug list is empty, and same_queue_rq is invalid.
1555 if (list_empty(&plug
->mq_list
))
1556 same_queue_rq
= NULL
;
1558 list_del_init(&same_queue_rq
->queuelist
);
1559 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1561 blk_mq_put_ctx(data
.ctx
);
1563 if (same_queue_rq
) {
1564 data
.hctx
= blk_mq_map_queue(q
,
1565 same_queue_rq
->mq_ctx
->cpu
);
1566 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1569 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1570 blk_mq_put_ctx(data
.ctx
);
1571 blk_mq_bio_to_request(rq
, bio
);
1572 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1573 } else if (q
->elevator
) {
1574 blk_mq_put_ctx(data
.ctx
);
1575 blk_mq_bio_to_request(rq
, bio
);
1576 blk_mq_sched_insert_request(rq
, false, true, true, true);
1578 blk_mq_put_ctx(data
.ctx
);
1579 blk_mq_bio_to_request(rq
, bio
);
1580 blk_mq_queue_io(data
.hctx
, data
.ctx
, rq
);
1581 blk_mq_run_hw_queue(data
.hctx
, true);
1587 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1588 unsigned int hctx_idx
)
1592 if (tags
->rqs
&& set
->ops
->exit_request
) {
1595 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1596 struct request
*rq
= tags
->static_rqs
[i
];
1600 set
->ops
->exit_request(set
, rq
, hctx_idx
);
1601 tags
->static_rqs
[i
] = NULL
;
1605 while (!list_empty(&tags
->page_list
)) {
1606 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1607 list_del_init(&page
->lru
);
1609 * Remove kmemleak object previously allocated in
1610 * blk_mq_init_rq_map().
1612 kmemleak_free(page_address(page
));
1613 __free_pages(page
, page
->private);
1617 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1621 kfree(tags
->static_rqs
);
1622 tags
->static_rqs
= NULL
;
1624 blk_mq_free_tags(tags
);
1627 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1628 unsigned int hctx_idx
,
1629 unsigned int nr_tags
,
1630 unsigned int reserved_tags
)
1632 struct blk_mq_tags
*tags
;
1635 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1636 if (node
== NUMA_NO_NODE
)
1637 node
= set
->numa_node
;
1639 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1640 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1644 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1645 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1648 blk_mq_free_tags(tags
);
1652 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1653 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1655 if (!tags
->static_rqs
) {
1657 blk_mq_free_tags(tags
);
1664 static size_t order_to_size(unsigned int order
)
1666 return (size_t)PAGE_SIZE
<< order
;
1669 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1670 unsigned int hctx_idx
, unsigned int depth
)
1672 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1673 size_t rq_size
, left
;
1676 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1677 if (node
== NUMA_NO_NODE
)
1678 node
= set
->numa_node
;
1680 INIT_LIST_HEAD(&tags
->page_list
);
1683 * rq_size is the size of the request plus driver payload, rounded
1684 * to the cacheline size
1686 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1688 left
= rq_size
* depth
;
1690 for (i
= 0; i
< depth
; ) {
1691 int this_order
= max_order
;
1696 while (this_order
&& left
< order_to_size(this_order
- 1))
1700 page
= alloc_pages_node(node
,
1701 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1707 if (order_to_size(this_order
) < rq_size
)
1714 page
->private = this_order
;
1715 list_add_tail(&page
->lru
, &tags
->page_list
);
1717 p
= page_address(page
);
1719 * Allow kmemleak to scan these pages as they contain pointers
1720 * to additional allocations like via ops->init_request().
1722 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1723 entries_per_page
= order_to_size(this_order
) / rq_size
;
1724 to_do
= min(entries_per_page
, depth
- i
);
1725 left
-= to_do
* rq_size
;
1726 for (j
= 0; j
< to_do
; j
++) {
1727 struct request
*rq
= p
;
1729 tags
->static_rqs
[i
] = rq
;
1730 if (set
->ops
->init_request
) {
1731 if (set
->ops
->init_request(set
, rq
, hctx_idx
,
1733 tags
->static_rqs
[i
] = NULL
;
1745 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1750 * 'cpu' is going away. splice any existing rq_list entries from this
1751 * software queue to the hw queue dispatch list, and ensure that it
1754 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1756 struct blk_mq_hw_ctx
*hctx
;
1757 struct blk_mq_ctx
*ctx
;
1760 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1761 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1763 spin_lock(&ctx
->lock
);
1764 if (!list_empty(&ctx
->rq_list
)) {
1765 list_splice_init(&ctx
->rq_list
, &tmp
);
1766 blk_mq_hctx_clear_pending(hctx
, ctx
);
1768 spin_unlock(&ctx
->lock
);
1770 if (list_empty(&tmp
))
1773 spin_lock(&hctx
->lock
);
1774 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1775 spin_unlock(&hctx
->lock
);
1777 blk_mq_run_hw_queue(hctx
, true);
1781 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1783 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1787 /* hctx->ctxs will be freed in queue's release handler */
1788 static void blk_mq_exit_hctx(struct request_queue
*q
,
1789 struct blk_mq_tag_set
*set
,
1790 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1792 blk_mq_debugfs_unregister_hctx(hctx
);
1794 blk_mq_tag_idle(hctx
);
1796 if (set
->ops
->exit_request
)
1797 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
1799 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1801 if (set
->ops
->exit_hctx
)
1802 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1804 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1805 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1807 blk_mq_remove_cpuhp(hctx
);
1808 blk_free_flush_queue(hctx
->fq
);
1809 sbitmap_free(&hctx
->ctx_map
);
1812 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1813 struct blk_mq_tag_set
*set
, int nr_queue
)
1815 struct blk_mq_hw_ctx
*hctx
;
1818 queue_for_each_hw_ctx(q
, hctx
, i
) {
1821 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1825 static int blk_mq_init_hctx(struct request_queue
*q
,
1826 struct blk_mq_tag_set
*set
,
1827 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1831 node
= hctx
->numa_node
;
1832 if (node
== NUMA_NO_NODE
)
1833 node
= hctx
->numa_node
= set
->numa_node
;
1835 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1836 spin_lock_init(&hctx
->lock
);
1837 INIT_LIST_HEAD(&hctx
->dispatch
);
1839 hctx
->queue_num
= hctx_idx
;
1840 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1842 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1844 hctx
->tags
= set
->tags
[hctx_idx
];
1847 * Allocate space for all possible cpus to avoid allocation at
1850 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1853 goto unregister_cpu_notifier
;
1855 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1861 if (set
->ops
->init_hctx
&&
1862 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1865 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
1868 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1870 goto sched_exit_hctx
;
1872 if (set
->ops
->init_request
&&
1873 set
->ops
->init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
1877 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1878 init_srcu_struct(&hctx
->queue_rq_srcu
);
1880 blk_mq_debugfs_register_hctx(q
, hctx
);
1887 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1889 if (set
->ops
->exit_hctx
)
1890 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1892 sbitmap_free(&hctx
->ctx_map
);
1895 unregister_cpu_notifier
:
1896 blk_mq_remove_cpuhp(hctx
);
1900 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1901 unsigned int nr_hw_queues
)
1905 for_each_possible_cpu(i
) {
1906 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1907 struct blk_mq_hw_ctx
*hctx
;
1910 spin_lock_init(&__ctx
->lock
);
1911 INIT_LIST_HEAD(&__ctx
->rq_list
);
1914 /* If the cpu isn't online, the cpu is mapped to first hctx */
1918 hctx
= blk_mq_map_queue(q
, i
);
1921 * Set local node, IFF we have more than one hw queue. If
1922 * not, we remain on the home node of the device
1924 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1925 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1929 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
1933 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
1934 set
->queue_depth
, set
->reserved_tags
);
1935 if (!set
->tags
[hctx_idx
])
1938 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
1943 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
1944 set
->tags
[hctx_idx
] = NULL
;
1948 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
1949 unsigned int hctx_idx
)
1951 if (set
->tags
[hctx_idx
]) {
1952 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
1953 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
1954 set
->tags
[hctx_idx
] = NULL
;
1958 static void blk_mq_map_swqueue(struct request_queue
*q
,
1959 const struct cpumask
*online_mask
)
1961 unsigned int i
, hctx_idx
;
1962 struct blk_mq_hw_ctx
*hctx
;
1963 struct blk_mq_ctx
*ctx
;
1964 struct blk_mq_tag_set
*set
= q
->tag_set
;
1967 * Avoid others reading imcomplete hctx->cpumask through sysfs
1969 mutex_lock(&q
->sysfs_lock
);
1971 queue_for_each_hw_ctx(q
, hctx
, i
) {
1972 cpumask_clear(hctx
->cpumask
);
1977 * Map software to hardware queues
1979 for_each_possible_cpu(i
) {
1980 /* If the cpu isn't online, the cpu is mapped to first hctx */
1981 if (!cpumask_test_cpu(i
, online_mask
))
1984 hctx_idx
= q
->mq_map
[i
];
1985 /* unmapped hw queue can be remapped after CPU topo changed */
1986 if (!set
->tags
[hctx_idx
] &&
1987 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
1989 * If tags initialization fail for some hctx,
1990 * that hctx won't be brought online. In this
1991 * case, remap the current ctx to hctx[0] which
1992 * is guaranteed to always have tags allocated
1997 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1998 hctx
= blk_mq_map_queue(q
, i
);
2000 cpumask_set_cpu(i
, hctx
->cpumask
);
2001 ctx
->index_hw
= hctx
->nr_ctx
;
2002 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2005 mutex_unlock(&q
->sysfs_lock
);
2007 queue_for_each_hw_ctx(q
, hctx
, i
) {
2009 * If no software queues are mapped to this hardware queue,
2010 * disable it and free the request entries.
2012 if (!hctx
->nr_ctx
) {
2013 /* Never unmap queue 0. We need it as a
2014 * fallback in case of a new remap fails
2017 if (i
&& set
->tags
[i
])
2018 blk_mq_free_map_and_requests(set
, i
);
2024 hctx
->tags
= set
->tags
[i
];
2025 WARN_ON(!hctx
->tags
);
2028 * Set the map size to the number of mapped software queues.
2029 * This is more accurate and more efficient than looping
2030 * over all possibly mapped software queues.
2032 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2035 * Initialize batch roundrobin counts
2037 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2038 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2042 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2044 struct blk_mq_hw_ctx
*hctx
;
2047 queue_for_each_hw_ctx(q
, hctx
, i
) {
2049 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2051 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2055 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
2057 struct request_queue
*q
;
2059 lockdep_assert_held(&set
->tag_list_lock
);
2061 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2062 blk_mq_freeze_queue(q
);
2063 queue_set_hctx_shared(q
, shared
);
2064 blk_mq_unfreeze_queue(q
);
2068 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2070 struct blk_mq_tag_set
*set
= q
->tag_set
;
2072 mutex_lock(&set
->tag_list_lock
);
2073 list_del_rcu(&q
->tag_set_list
);
2074 INIT_LIST_HEAD(&q
->tag_set_list
);
2075 if (list_is_singular(&set
->tag_list
)) {
2076 /* just transitioned to unshared */
2077 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2078 /* update existing queue */
2079 blk_mq_update_tag_set_depth(set
, false);
2081 mutex_unlock(&set
->tag_list_lock
);
2086 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2087 struct request_queue
*q
)
2091 mutex_lock(&set
->tag_list_lock
);
2093 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2094 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2095 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2096 /* update existing queue */
2097 blk_mq_update_tag_set_depth(set
, true);
2099 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2100 queue_set_hctx_shared(q
, true);
2101 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2103 mutex_unlock(&set
->tag_list_lock
);
2107 * It is the actual release handler for mq, but we do it from
2108 * request queue's release handler for avoiding use-after-free
2109 * and headache because q->mq_kobj shouldn't have been introduced,
2110 * but we can't group ctx/kctx kobj without it.
2112 void blk_mq_release(struct request_queue
*q
)
2114 struct blk_mq_hw_ctx
*hctx
;
2117 /* hctx kobj stays in hctx */
2118 queue_for_each_hw_ctx(q
, hctx
, i
) {
2121 kobject_put(&hctx
->kobj
);
2126 kfree(q
->queue_hw_ctx
);
2129 * release .mq_kobj and sw queue's kobject now because
2130 * both share lifetime with request queue.
2132 blk_mq_sysfs_deinit(q
);
2134 free_percpu(q
->queue_ctx
);
2137 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2139 struct request_queue
*uninit_q
, *q
;
2141 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2143 return ERR_PTR(-ENOMEM
);
2145 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2147 blk_cleanup_queue(uninit_q
);
2151 EXPORT_SYMBOL(blk_mq_init_queue
);
2153 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2154 struct request_queue
*q
)
2157 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2159 blk_mq_sysfs_unregister(q
);
2160 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2166 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2167 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2172 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2179 atomic_set(&hctxs
[i
]->nr_active
, 0);
2180 hctxs
[i
]->numa_node
= node
;
2181 hctxs
[i
]->queue_num
= i
;
2183 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2184 free_cpumask_var(hctxs
[i
]->cpumask
);
2189 blk_mq_hctx_kobj_init(hctxs
[i
]);
2191 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2192 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2196 blk_mq_free_map_and_requests(set
, j
);
2197 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2198 kobject_put(&hctx
->kobj
);
2203 q
->nr_hw_queues
= i
;
2204 blk_mq_sysfs_register(q
);
2207 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2208 struct request_queue
*q
)
2210 /* mark the queue as mq asap */
2211 q
->mq_ops
= set
->ops
;
2213 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2214 blk_mq_poll_stats_bkt
,
2215 BLK_MQ_POLL_STATS_BKTS
, q
);
2219 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2223 /* init q->mq_kobj and sw queues' kobjects */
2224 blk_mq_sysfs_init(q
);
2226 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2227 GFP_KERNEL
, set
->numa_node
);
2228 if (!q
->queue_hw_ctx
)
2231 q
->mq_map
= set
->mq_map
;
2233 blk_mq_realloc_hw_ctxs(set
, q
);
2234 if (!q
->nr_hw_queues
)
2237 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2238 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2240 q
->nr_queues
= nr_cpu_ids
;
2242 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2244 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2245 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2247 q
->sg_reserved_size
= INT_MAX
;
2249 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2250 INIT_LIST_HEAD(&q
->requeue_list
);
2251 spin_lock_init(&q
->requeue_lock
);
2253 blk_queue_make_request(q
, blk_mq_make_request
);
2256 * Do this after blk_queue_make_request() overrides it...
2258 q
->nr_requests
= set
->queue_depth
;
2261 * Default to classic polling
2265 if (set
->ops
->complete
)
2266 blk_queue_softirq_done(q
, set
->ops
->complete
);
2268 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2271 mutex_lock(&all_q_mutex
);
2273 list_add_tail(&q
->all_q_node
, &all_q_list
);
2274 blk_mq_add_queue_tag_set(set
, q
);
2275 blk_mq_map_swqueue(q
, cpu_online_mask
);
2277 mutex_unlock(&all_q_mutex
);
2280 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2283 ret
= blk_mq_sched_init(q
);
2285 return ERR_PTR(ret
);
2291 kfree(q
->queue_hw_ctx
);
2293 free_percpu(q
->queue_ctx
);
2296 return ERR_PTR(-ENOMEM
);
2298 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2300 void blk_mq_free_queue(struct request_queue
*q
)
2302 struct blk_mq_tag_set
*set
= q
->tag_set
;
2304 mutex_lock(&all_q_mutex
);
2305 list_del_init(&q
->all_q_node
);
2306 mutex_unlock(&all_q_mutex
);
2308 blk_mq_del_queue_tag_set(q
);
2310 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2313 /* Basically redo blk_mq_init_queue with queue frozen */
2314 static void blk_mq_queue_reinit(struct request_queue
*q
,
2315 const struct cpumask
*online_mask
)
2317 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2319 blk_mq_debugfs_unregister_hctxs(q
);
2320 blk_mq_sysfs_unregister(q
);
2323 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2324 * we should change hctx numa_node according to new topology (this
2325 * involves free and re-allocate memory, worthy doing?)
2328 blk_mq_map_swqueue(q
, online_mask
);
2330 blk_mq_sysfs_register(q
);
2331 blk_mq_debugfs_register_hctxs(q
);
2335 * New online cpumask which is going to be set in this hotplug event.
2336 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2337 * one-by-one and dynamically allocating this could result in a failure.
2339 static struct cpumask cpuhp_online_new
;
2341 static void blk_mq_queue_reinit_work(void)
2343 struct request_queue
*q
;
2345 mutex_lock(&all_q_mutex
);
2347 * We need to freeze and reinit all existing queues. Freezing
2348 * involves synchronous wait for an RCU grace period and doing it
2349 * one by one may take a long time. Start freezing all queues in
2350 * one swoop and then wait for the completions so that freezing can
2351 * take place in parallel.
2353 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2354 blk_freeze_queue_start(q
);
2355 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2356 blk_mq_freeze_queue_wait(q
);
2358 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2359 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2361 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2362 blk_mq_unfreeze_queue(q
);
2364 mutex_unlock(&all_q_mutex
);
2367 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2369 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2370 blk_mq_queue_reinit_work();
2375 * Before hotadded cpu starts handling requests, new mappings must be
2376 * established. Otherwise, these requests in hw queue might never be
2379 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2380 * for CPU0, and ctx1 for CPU1).
2382 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2383 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2385 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2386 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2387 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2390 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2392 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2393 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2394 blk_mq_queue_reinit_work();
2398 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2402 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2403 if (!__blk_mq_alloc_rq_map(set
, i
))
2410 blk_mq_free_rq_map(set
->tags
[i
]);
2416 * Allocate the request maps associated with this tag_set. Note that this
2417 * may reduce the depth asked for, if memory is tight. set->queue_depth
2418 * will be updated to reflect the allocated depth.
2420 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2425 depth
= set
->queue_depth
;
2427 err
= __blk_mq_alloc_rq_maps(set
);
2431 set
->queue_depth
>>= 1;
2432 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2436 } while (set
->queue_depth
);
2438 if (!set
->queue_depth
|| err
) {
2439 pr_err("blk-mq: failed to allocate request map\n");
2443 if (depth
!= set
->queue_depth
)
2444 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2445 depth
, set
->queue_depth
);
2450 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2452 if (set
->ops
->map_queues
)
2453 return set
->ops
->map_queues(set
);
2455 return blk_mq_map_queues(set
);
2459 * Alloc a tag set to be associated with one or more request queues.
2460 * May fail with EINVAL for various error conditions. May adjust the
2461 * requested depth down, if if it too large. In that case, the set
2462 * value will be stored in set->queue_depth.
2464 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2468 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2470 if (!set
->nr_hw_queues
)
2472 if (!set
->queue_depth
)
2474 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2477 if (!set
->ops
->queue_rq
)
2480 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2481 pr_info("blk-mq: reduced tag depth to %u\n",
2483 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2487 * If a crashdump is active, then we are potentially in a very
2488 * memory constrained environment. Limit us to 1 queue and
2489 * 64 tags to prevent using too much memory.
2491 if (is_kdump_kernel()) {
2492 set
->nr_hw_queues
= 1;
2493 set
->queue_depth
= min(64U, set
->queue_depth
);
2496 * There is no use for more h/w queues than cpus.
2498 if (set
->nr_hw_queues
> nr_cpu_ids
)
2499 set
->nr_hw_queues
= nr_cpu_ids
;
2501 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2502 GFP_KERNEL
, set
->numa_node
);
2507 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2508 GFP_KERNEL
, set
->numa_node
);
2512 ret
= blk_mq_update_queue_map(set
);
2514 goto out_free_mq_map
;
2516 ret
= blk_mq_alloc_rq_maps(set
);
2518 goto out_free_mq_map
;
2520 mutex_init(&set
->tag_list_lock
);
2521 INIT_LIST_HEAD(&set
->tag_list
);
2533 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2535 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2539 for (i
= 0; i
< nr_cpu_ids
; i
++)
2540 blk_mq_free_map_and_requests(set
, i
);
2548 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2550 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2552 struct blk_mq_tag_set
*set
= q
->tag_set
;
2553 struct blk_mq_hw_ctx
*hctx
;
2559 blk_mq_freeze_queue(q
);
2562 queue_for_each_hw_ctx(q
, hctx
, i
) {
2566 * If we're using an MQ scheduler, just update the scheduler
2567 * queue depth. This is similar to what the old code would do.
2569 if (!hctx
->sched_tags
) {
2570 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2571 min(nr
, set
->queue_depth
),
2574 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2582 q
->nr_requests
= nr
;
2584 blk_mq_unfreeze_queue(q
);
2589 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
2592 struct request_queue
*q
;
2594 lockdep_assert_held(&set
->tag_list_lock
);
2596 if (nr_hw_queues
> nr_cpu_ids
)
2597 nr_hw_queues
= nr_cpu_ids
;
2598 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2601 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2602 blk_mq_freeze_queue(q
);
2604 set
->nr_hw_queues
= nr_hw_queues
;
2605 blk_mq_update_queue_map(set
);
2606 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2607 blk_mq_realloc_hw_ctxs(set
, q
);
2608 blk_mq_queue_reinit(q
, cpu_online_mask
);
2611 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2612 blk_mq_unfreeze_queue(q
);
2615 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2617 mutex_lock(&set
->tag_list_lock
);
2618 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
2619 mutex_unlock(&set
->tag_list_lock
);
2621 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2623 /* Enable polling stats and return whether they were already enabled. */
2624 static bool blk_poll_stats_enable(struct request_queue
*q
)
2626 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2627 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
2629 blk_stat_add_callback(q
, q
->poll_cb
);
2633 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2636 * We don't arm the callback if polling stats are not enabled or the
2637 * callback is already active.
2639 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2640 blk_stat_is_active(q
->poll_cb
))
2643 blk_stat_activate_msecs(q
->poll_cb
, 100);
2646 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2648 struct request_queue
*q
= cb
->data
;
2651 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
2652 if (cb
->stat
[bucket
].nr_samples
)
2653 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
2657 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2658 struct blk_mq_hw_ctx
*hctx
,
2661 unsigned long ret
= 0;
2665 * If stats collection isn't on, don't sleep but turn it on for
2668 if (!blk_poll_stats_enable(q
))
2672 * As an optimistic guess, use half of the mean service time
2673 * for this type of request. We can (and should) make this smarter.
2674 * For instance, if the completion latencies are tight, we can
2675 * get closer than just half the mean. This is especially
2676 * important on devices where the completion latencies are longer
2677 * than ~10 usec. We do use the stats for the relevant IO size
2678 * if available which does lead to better estimates.
2680 bucket
= blk_mq_poll_stats_bkt(rq
);
2684 if (q
->poll_stat
[bucket
].nr_samples
)
2685 ret
= (q
->poll_stat
[bucket
].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 * With scheduling, if the request has completed, we'll
2812 * get a NULL return here, as we clear the sched tag when
2813 * that happens. The request still remains valid, like always,
2814 * so we should be safe with just the NULL check.
2820 return __blk_mq_poll(hctx
, rq
);
2822 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2824 void blk_mq_disable_hotplug(void)
2826 mutex_lock(&all_q_mutex
);
2829 void blk_mq_enable_hotplug(void)
2831 mutex_unlock(&all_q_mutex
);
2834 static int __init
blk_mq_init(void)
2836 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2837 blk_mq_hctx_notify_dead
);
2839 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2840 blk_mq_queue_reinit_prepare
,
2841 blk_mq_queue_reinit_dead
);
2844 subsys_initcall(blk_mq_init
);