2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
29 #include <trace/events/block.h>
31 #include <linux/blk-mq.h>
34 #include "blk-mq-tag.h"
37 #include "blk-mq-sched.h"
39 static DEFINE_MUTEX(all_q_mutex
);
40 static LIST_HEAD(all_q_list
);
42 static void blk_mq_poll_stats_start(struct request_queue
*q
);
43 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
46 * Check if any of the ctx's have pending work in this hardware queue
48 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
50 return sbitmap_any_bit_set(&hctx
->ctx_map
) ||
51 !list_empty_careful(&hctx
->dispatch
) ||
52 blk_mq_sched_has_work(hctx
);
56 * Mark this ctx as having pending work in this hardware queue
58 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
59 struct blk_mq_ctx
*ctx
)
61 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
62 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
65 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
66 struct blk_mq_ctx
*ctx
)
68 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
71 void blk_mq_freeze_queue_start(struct request_queue
*q
)
75 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
76 if (freeze_depth
== 1) {
77 percpu_ref_kill(&q
->q_usage_counter
);
78 blk_mq_run_hw_queues(q
, false);
81 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start
);
83 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
85 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
87 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
89 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
90 unsigned long timeout
)
92 return wait_event_timeout(q
->mq_freeze_wq
,
93 percpu_ref_is_zero(&q
->q_usage_counter
),
96 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
99 * Guarantee no request is in use, so we can change any data structure of
100 * the queue afterward.
102 void blk_freeze_queue(struct request_queue
*q
)
105 * In the !blk_mq case we are only calling this to kill the
106 * q_usage_counter, otherwise this increases the freeze depth
107 * and waits for it to return to zero. For this reason there is
108 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
109 * exported to drivers as the only user for unfreeze is blk_mq.
111 blk_mq_freeze_queue_start(q
);
112 blk_mq_freeze_queue_wait(q
);
115 void blk_mq_freeze_queue(struct request_queue
*q
)
118 * ...just an alias to keep freeze and unfreeze actions balanced
119 * in the blk_mq_* namespace
123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
125 void blk_mq_unfreeze_queue(struct request_queue
*q
)
129 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
130 WARN_ON_ONCE(freeze_depth
< 0);
132 percpu_ref_reinit(&q
->q_usage_counter
);
133 wake_up_all(&q
->mq_freeze_wq
);
136 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
139 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
142 * Note: this function does not prevent that the struct request end_io()
143 * callback function is invoked. Additionally, it is not prevented that
144 * new queue_rq() calls occur unless the queue has been stopped first.
146 void blk_mq_quiesce_queue(struct request_queue
*q
)
148 struct blk_mq_hw_ctx
*hctx
;
152 blk_mq_stop_hw_queues(q
);
154 queue_for_each_hw_ctx(q
, hctx
, i
) {
155 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
156 synchronize_srcu(&hctx
->queue_rq_srcu
);
163 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
165 void blk_mq_wake_waiters(struct request_queue
*q
)
167 struct blk_mq_hw_ctx
*hctx
;
170 queue_for_each_hw_ctx(q
, hctx
, i
)
171 if (blk_mq_hw_queue_mapped(hctx
))
172 blk_mq_tag_wakeup_all(hctx
->tags
, true);
175 * If we are called because the queue has now been marked as
176 * dying, we need to ensure that processes currently waiting on
177 * the queue are notified as well.
179 wake_up_all(&q
->mq_freeze_wq
);
182 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
184 return blk_mq_has_free_tags(hctx
->tags
);
186 EXPORT_SYMBOL(blk_mq_can_queue
);
188 void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
189 struct request
*rq
, unsigned int op
)
191 INIT_LIST_HEAD(&rq
->queuelist
);
192 /* csd/requeue_work/fifo_time is initialized before use */
196 if (blk_queue_io_stat(q
))
197 rq
->rq_flags
|= RQF_IO_STAT
;
198 /* do not touch atomic flags, it needs atomic ops against the timer */
200 INIT_HLIST_NODE(&rq
->hash
);
201 RB_CLEAR_NODE(&rq
->rb_node
);
204 rq
->start_time
= jiffies
;
205 #ifdef CONFIG_BLK_CGROUP
207 set_start_time_ns(rq
);
208 rq
->io_start_time_ns
= 0;
210 rq
->nr_phys_segments
= 0;
211 #if defined(CONFIG_BLK_DEV_INTEGRITY)
212 rq
->nr_integrity_segments
= 0;
215 /* tag was already set */
219 INIT_LIST_HEAD(&rq
->timeout_list
);
223 rq
->end_io_data
= NULL
;
226 ctx
->rq_dispatched
[op_is_sync(op
)]++;
228 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init
);
230 struct request
*__blk_mq_alloc_request(struct blk_mq_alloc_data
*data
,
236 tag
= blk_mq_get_tag(data
);
237 if (tag
!= BLK_MQ_TAG_FAIL
) {
238 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
240 rq
= tags
->static_rqs
[tag
];
242 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
244 rq
->internal_tag
= tag
;
246 if (blk_mq_tag_busy(data
->hctx
)) {
247 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
248 atomic_inc(&data
->hctx
->nr_active
);
251 rq
->internal_tag
= -1;
252 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
255 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
);
261 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request
);
263 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
266 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
270 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
274 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
276 blk_mq_put_ctx(alloc_data
.ctx
);
280 return ERR_PTR(-EWOULDBLOCK
);
283 rq
->__sector
= (sector_t
) -1;
284 rq
->bio
= rq
->biotail
= NULL
;
287 EXPORT_SYMBOL(blk_mq_alloc_request
);
289 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
290 unsigned int flags
, unsigned int hctx_idx
)
292 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
298 * If the tag allocator sleeps we could get an allocation for a
299 * different hardware context. No need to complicate the low level
300 * allocator for this for the rare use case of a command tied to
303 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
304 return ERR_PTR(-EINVAL
);
306 if (hctx_idx
>= q
->nr_hw_queues
)
307 return ERR_PTR(-EIO
);
309 ret
= blk_queue_enter(q
, true);
314 * Check if the hardware context is actually mapped to anything.
315 * If not tell the caller that it should skip this queue.
317 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
318 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
320 return ERR_PTR(-EXDEV
);
322 cpu
= cpumask_first(alloc_data
.hctx
->cpumask
);
323 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
325 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
327 blk_mq_put_ctx(alloc_data
.ctx
);
331 return ERR_PTR(-EWOULDBLOCK
);
335 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
337 void __blk_mq_finish_request(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
340 const int sched_tag
= rq
->internal_tag
;
341 struct request_queue
*q
= rq
->q
;
343 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
344 atomic_dec(&hctx
->nr_active
);
346 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
349 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
350 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
352 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
354 blk_mq_sched_completed_request(hctx
, rq
);
355 blk_mq_sched_restart_queues(hctx
);
359 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx
*hctx
,
362 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
364 ctx
->rq_completed
[rq_is_sync(rq
)]++;
365 __blk_mq_finish_request(hctx
, ctx
, rq
);
368 void blk_mq_finish_request(struct request
*rq
)
370 blk_mq_finish_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
373 void blk_mq_free_request(struct request
*rq
)
375 blk_mq_sched_put_request(rq
);
377 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
379 inline void __blk_mq_end_request(struct request
*rq
, int error
)
381 blk_account_io_done(rq
);
384 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
385 rq
->end_io(rq
, error
);
387 if (unlikely(blk_bidi_rq(rq
)))
388 blk_mq_free_request(rq
->next_rq
);
389 blk_mq_free_request(rq
);
392 EXPORT_SYMBOL(__blk_mq_end_request
);
394 void blk_mq_end_request(struct request
*rq
, int error
)
396 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
398 __blk_mq_end_request(rq
, error
);
400 EXPORT_SYMBOL(blk_mq_end_request
);
402 static void __blk_mq_complete_request_remote(void *data
)
404 struct request
*rq
= data
;
406 rq
->q
->softirq_done_fn(rq
);
409 static void blk_mq_ipi_complete_request(struct request
*rq
)
411 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
415 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
416 rq
->q
->softirq_done_fn(rq
);
421 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
422 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
424 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
425 rq
->csd
.func
= __blk_mq_complete_request_remote
;
428 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
430 rq
->q
->softirq_done_fn(rq
);
435 static void blk_mq_stat_add(struct request
*rq
)
437 if (rq
->rq_flags
& RQF_STATS
) {
438 blk_mq_poll_stats_start(rq
->q
);
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 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
491 blk_stat_set_issue(&rq
->issue_stat
, blk_rq_sectors(rq
));
492 rq
->rq_flags
|= RQF_STATS
;
493 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
499 * Ensure that ->deadline is visible before set the started
500 * flag and clear the completed flag.
502 smp_mb__before_atomic();
505 * Mark us as started and clear complete. Complete might have been
506 * set if requeue raced with timeout, which then marked it as
507 * complete. So be sure to clear complete again when we start
508 * the request, otherwise we'll ignore the completion event.
510 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
511 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
512 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
513 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
515 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
517 * Make sure space for the drain appears. We know we can do
518 * this because max_hw_segments has been adjusted to be one
519 * fewer than the device can handle.
521 rq
->nr_phys_segments
++;
524 EXPORT_SYMBOL(blk_mq_start_request
);
527 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
528 * flag isn't set yet, so there may be race with timeout hanlder,
529 * but given rq->deadline is just set in .queue_rq() under
530 * this situation, the race won't be possible in reality because
531 * rq->timeout should be set as big enough to cover the window
532 * between blk_mq_start_request() called from .queue_rq() and
533 * clearing REQ_ATOM_STARTED here.
535 static void __blk_mq_requeue_request(struct request
*rq
)
537 struct request_queue
*q
= rq
->q
;
539 trace_block_rq_requeue(q
, rq
);
540 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
541 blk_mq_sched_requeue_request(rq
);
543 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
544 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
545 rq
->nr_phys_segments
--;
549 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
551 __blk_mq_requeue_request(rq
);
553 BUG_ON(blk_queued_rq(rq
));
554 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
556 EXPORT_SYMBOL(blk_mq_requeue_request
);
558 static void blk_mq_requeue_work(struct work_struct
*work
)
560 struct request_queue
*q
=
561 container_of(work
, struct request_queue
, requeue_work
.work
);
563 struct request
*rq
, *next
;
566 spin_lock_irqsave(&q
->requeue_lock
, flags
);
567 list_splice_init(&q
->requeue_list
, &rq_list
);
568 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
570 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
571 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
574 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
575 list_del_init(&rq
->queuelist
);
576 blk_mq_sched_insert_request(rq
, true, false, false, true);
579 while (!list_empty(&rq_list
)) {
580 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
581 list_del_init(&rq
->queuelist
);
582 blk_mq_sched_insert_request(rq
, false, false, false, true);
585 blk_mq_run_hw_queues(q
, false);
588 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
589 bool kick_requeue_list
)
591 struct request_queue
*q
= rq
->q
;
595 * We abuse this flag that is otherwise used by the I/O scheduler to
596 * request head insertation from the workqueue.
598 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
600 spin_lock_irqsave(&q
->requeue_lock
, flags
);
602 rq
->rq_flags
|= RQF_SOFTBARRIER
;
603 list_add(&rq
->queuelist
, &q
->requeue_list
);
605 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
607 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
609 if (kick_requeue_list
)
610 blk_mq_kick_requeue_list(q
);
612 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
614 void blk_mq_kick_requeue_list(struct request_queue
*q
)
616 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
618 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
620 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
623 kblockd_schedule_delayed_work(&q
->requeue_work
,
624 msecs_to_jiffies(msecs
));
626 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
628 void blk_mq_abort_requeue_list(struct request_queue
*q
)
633 spin_lock_irqsave(&q
->requeue_lock
, flags
);
634 list_splice_init(&q
->requeue_list
, &rq_list
);
635 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
637 while (!list_empty(&rq_list
)) {
640 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
641 list_del_init(&rq
->queuelist
);
643 blk_mq_end_request(rq
, rq
->errors
);
646 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
648 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
650 if (tag
< tags
->nr_tags
) {
651 prefetch(tags
->rqs
[tag
]);
652 return tags
->rqs
[tag
];
657 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
659 struct blk_mq_timeout_data
{
661 unsigned int next_set
;
664 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
666 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
667 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
670 * We know that complete is set at this point. If STARTED isn't set
671 * anymore, then the request isn't active and the "timeout" should
672 * just be ignored. This can happen due to the bitflag ordering.
673 * Timeout first checks if STARTED is set, and if it is, assumes
674 * the request is active. But if we race with completion, then
675 * we both flags will get cleared. So check here again, and ignore
676 * a timeout event with a request that isn't active.
678 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
682 ret
= ops
->timeout(req
, reserved
);
686 __blk_mq_complete_request(req
);
688 case BLK_EH_RESET_TIMER
:
690 blk_clear_rq_complete(req
);
692 case BLK_EH_NOT_HANDLED
:
695 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
700 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
701 struct request
*rq
, void *priv
, bool reserved
)
703 struct blk_mq_timeout_data
*data
= priv
;
705 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
709 * The rq being checked may have been freed and reallocated
710 * out already here, we avoid this race by checking rq->deadline
711 * and REQ_ATOM_COMPLETE flag together:
713 * - if rq->deadline is observed as new value because of
714 * reusing, the rq won't be timed out because of timing.
715 * - if rq->deadline is observed as previous value,
716 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
717 * because we put a barrier between setting rq->deadline
718 * and clearing the flag in blk_mq_start_request(), so
719 * this rq won't be timed out too.
721 if (time_after_eq(jiffies
, rq
->deadline
)) {
722 if (!blk_mark_rq_complete(rq
))
723 blk_mq_rq_timed_out(rq
, reserved
);
724 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
725 data
->next
= rq
->deadline
;
730 static void blk_mq_timeout_work(struct work_struct
*work
)
732 struct request_queue
*q
=
733 container_of(work
, struct request_queue
, timeout_work
);
734 struct blk_mq_timeout_data data
= {
740 /* A deadlock might occur if a request is stuck requiring a
741 * timeout at the same time a queue freeze is waiting
742 * completion, since the timeout code would not be able to
743 * acquire the queue reference here.
745 * That's why we don't use blk_queue_enter here; instead, we use
746 * percpu_ref_tryget directly, because we need to be able to
747 * obtain a reference even in the short window between the queue
748 * starting to freeze, by dropping the first reference in
749 * blk_mq_freeze_queue_start, and the moment the last request is
750 * consumed, marked by the instant q_usage_counter reaches
753 if (!percpu_ref_tryget(&q
->q_usage_counter
))
756 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
759 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
760 mod_timer(&q
->timeout
, data
.next
);
762 struct blk_mq_hw_ctx
*hctx
;
764 queue_for_each_hw_ctx(q
, hctx
, i
) {
765 /* the hctx may be unmapped, so check it here */
766 if (blk_mq_hw_queue_mapped(hctx
))
767 blk_mq_tag_idle(hctx
);
774 * Reverse check our software queue for entries that we could potentially
775 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
776 * too much time checking for merges.
778 static bool blk_mq_attempt_merge(struct request_queue
*q
,
779 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
784 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
790 if (!blk_rq_merge_ok(rq
, bio
))
793 switch (blk_try_merge(rq
, bio
)) {
794 case ELEVATOR_BACK_MERGE
:
795 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
796 merged
= bio_attempt_back_merge(q
, rq
, bio
);
798 case ELEVATOR_FRONT_MERGE
:
799 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
800 merged
= bio_attempt_front_merge(q
, rq
, bio
);
802 case ELEVATOR_DISCARD_MERGE
:
803 merged
= bio_attempt_discard_merge(q
, rq
, bio
);
817 struct flush_busy_ctx_data
{
818 struct blk_mq_hw_ctx
*hctx
;
819 struct list_head
*list
;
822 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
824 struct flush_busy_ctx_data
*flush_data
= data
;
825 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
826 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
828 sbitmap_clear_bit(sb
, bitnr
);
829 spin_lock(&ctx
->lock
);
830 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
831 spin_unlock(&ctx
->lock
);
836 * Process software queues that have been marked busy, splicing them
837 * to the for-dispatch
839 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
841 struct flush_busy_ctx_data data
= {
846 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
848 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
850 static inline unsigned int queued_to_index(unsigned int queued
)
855 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
858 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
861 struct blk_mq_alloc_data data
= {
863 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
864 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
874 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
875 data
.flags
|= BLK_MQ_REQ_RESERVED
;
877 rq
->tag
= blk_mq_get_tag(&data
);
879 if (blk_mq_tag_busy(data
.hctx
)) {
880 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
881 atomic_inc(&data
.hctx
->nr_active
);
883 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
890 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx
*hctx
,
893 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
896 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
897 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
898 atomic_dec(&hctx
->nr_active
);
902 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx
*hctx
,
905 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
908 __blk_mq_put_driver_tag(hctx
, rq
);
911 static void blk_mq_put_driver_tag(struct request
*rq
)
913 struct blk_mq_hw_ctx
*hctx
;
915 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
918 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
919 __blk_mq_put_driver_tag(hctx
, rq
);
923 * If we fail getting a driver tag because all the driver tags are already
924 * assigned and on the dispatch list, BUT the first entry does not have a
925 * tag, then we could deadlock. For that case, move entries with assigned
926 * driver tags to the front, leaving the set of tagged requests in the
927 * same order, and the untagged set in the same order.
929 static bool reorder_tags_to_front(struct list_head
*list
)
931 struct request
*rq
, *tmp
, *first
= NULL
;
933 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
937 list_move(&rq
->queuelist
, list
);
943 return first
!= NULL
;
946 static int blk_mq_dispatch_wake(wait_queue_t
*wait
, unsigned mode
, int flags
,
949 struct blk_mq_hw_ctx
*hctx
;
951 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
953 list_del(&wait
->task_list
);
954 clear_bit_unlock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
);
955 blk_mq_run_hw_queue(hctx
, true);
959 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx
*hctx
)
961 struct sbq_wait_state
*ws
;
964 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
965 * The thread which wins the race to grab this bit adds the hardware
966 * queue to the wait queue.
968 if (test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
) ||
969 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
972 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
973 ws
= bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
);
976 * As soon as this returns, it's no longer safe to fiddle with
977 * hctx->dispatch_wait, since a completion can wake up the wait queue
978 * and unlock the bit.
980 add_wait_queue(&ws
->wait
, &hctx
->dispatch_wait
);
984 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
986 struct request_queue
*q
= hctx
->queue
;
988 LIST_HEAD(driver_list
);
989 struct list_head
*dptr
;
990 int queued
, ret
= BLK_MQ_RQ_QUEUE_OK
;
993 * Start off with dptr being NULL, so we start the first request
994 * immediately, even if we have more pending.
999 * Now process all the entries, sending them to the driver.
1002 while (!list_empty(list
)) {
1003 struct blk_mq_queue_data bd
;
1005 rq
= list_first_entry(list
, struct request
, queuelist
);
1006 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
1007 if (!queued
&& reorder_tags_to_front(list
))
1011 * The initial allocation attempt failed, so we need to
1012 * rerun the hardware queue when a tag is freed.
1014 if (blk_mq_dispatch_wait_add(hctx
)) {
1016 * It's possible that a tag was freed in the
1017 * window between the allocation failure and
1018 * adding the hardware queue to the wait queue.
1020 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1027 list_del_init(&rq
->queuelist
);
1033 * Flag last if we have no more requests, or if we have more
1034 * but can't assign a driver tag to it.
1036 if (list_empty(list
))
1039 struct request
*nxt
;
1041 nxt
= list_first_entry(list
, struct request
, queuelist
);
1042 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1045 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1047 case BLK_MQ_RQ_QUEUE_OK
:
1050 case BLK_MQ_RQ_QUEUE_BUSY
:
1051 blk_mq_put_driver_tag_hctx(hctx
, rq
);
1052 list_add(&rq
->queuelist
, list
);
1053 __blk_mq_requeue_request(rq
);
1056 pr_err("blk-mq: bad return on queue: %d\n", ret
);
1057 case BLK_MQ_RQ_QUEUE_ERROR
:
1059 blk_mq_end_request(rq
, rq
->errors
);
1063 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
1067 * We've done the first request. If we have more than 1
1068 * left in the list, set dptr to defer issue.
1070 if (!dptr
&& list
->next
!= list
->prev
)
1071 dptr
= &driver_list
;
1074 hctx
->dispatched
[queued_to_index(queued
)]++;
1077 * Any items that need requeuing? Stuff them into hctx->dispatch,
1078 * that is where we will continue on next queue run.
1080 if (!list_empty(list
)) {
1082 * If we got a driver tag for the next request already,
1085 rq
= list_first_entry(list
, struct request
, queuelist
);
1086 blk_mq_put_driver_tag(rq
);
1088 spin_lock(&hctx
->lock
);
1089 list_splice_init(list
, &hctx
->dispatch
);
1090 spin_unlock(&hctx
->lock
);
1093 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
1094 * it's possible the queue is stopped and restarted again
1095 * before this. Queue restart will dispatch requests. And since
1096 * requests in rq_list aren't added into hctx->dispatch yet,
1097 * the requests in rq_list might get lost.
1099 * blk_mq_run_hw_queue() already checks the STOPPED bit
1101 * If RESTART or TAG_WAITING is set, then let completion restart
1102 * the queue instead of potentially looping here.
1104 if (!blk_mq_sched_needs_restart(hctx
) &&
1105 !test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1106 blk_mq_run_hw_queue(hctx
, true);
1112 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1116 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1117 cpu_online(hctx
->next_cpu
));
1119 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1121 blk_mq_sched_dispatch_requests(hctx
);
1124 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1125 blk_mq_sched_dispatch_requests(hctx
);
1126 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1131 * It'd be great if the workqueue API had a way to pass
1132 * in a mask and had some smarts for more clever placement.
1133 * For now we just round-robin here, switching for every
1134 * BLK_MQ_CPU_WORK_BATCH queued items.
1136 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1138 if (hctx
->queue
->nr_hw_queues
== 1)
1139 return WORK_CPU_UNBOUND
;
1141 if (--hctx
->next_cpu_batch
<= 0) {
1144 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1145 if (next_cpu
>= nr_cpu_ids
)
1146 next_cpu
= cpumask_first(hctx
->cpumask
);
1148 hctx
->next_cpu
= next_cpu
;
1149 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1152 return hctx
->next_cpu
;
1155 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1157 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
1158 !blk_mq_hw_queue_mapped(hctx
)))
1161 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1162 int cpu
= get_cpu();
1163 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1164 __blk_mq_run_hw_queue(hctx
);
1172 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
);
1175 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1177 struct blk_mq_hw_ctx
*hctx
;
1180 queue_for_each_hw_ctx(q
, hctx
, i
) {
1181 if (!blk_mq_hctx_has_pending(hctx
) ||
1182 blk_mq_hctx_stopped(hctx
))
1185 blk_mq_run_hw_queue(hctx
, async
);
1188 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1191 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1192 * @q: request queue.
1194 * The caller is responsible for serializing this function against
1195 * blk_mq_{start,stop}_hw_queue().
1197 bool blk_mq_queue_stopped(struct request_queue
*q
)
1199 struct blk_mq_hw_ctx
*hctx
;
1202 queue_for_each_hw_ctx(q
, hctx
, i
)
1203 if (blk_mq_hctx_stopped(hctx
))
1208 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1210 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1212 cancel_work(&hctx
->run_work
);
1213 cancel_delayed_work(&hctx
->delay_work
);
1214 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1216 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1218 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1220 struct blk_mq_hw_ctx
*hctx
;
1223 queue_for_each_hw_ctx(q
, hctx
, i
)
1224 blk_mq_stop_hw_queue(hctx
);
1226 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1228 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1230 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1232 blk_mq_run_hw_queue(hctx
, false);
1234 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1236 void blk_mq_start_hw_queues(struct request_queue
*q
)
1238 struct blk_mq_hw_ctx
*hctx
;
1241 queue_for_each_hw_ctx(q
, hctx
, i
)
1242 blk_mq_start_hw_queue(hctx
);
1244 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1246 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1248 if (!blk_mq_hctx_stopped(hctx
))
1251 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1252 blk_mq_run_hw_queue(hctx
, async
);
1254 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1256 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1258 struct blk_mq_hw_ctx
*hctx
;
1261 queue_for_each_hw_ctx(q
, hctx
, i
)
1262 blk_mq_start_stopped_hw_queue(hctx
, async
);
1264 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1266 static void blk_mq_run_work_fn(struct work_struct
*work
)
1268 struct blk_mq_hw_ctx
*hctx
;
1270 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1272 __blk_mq_run_hw_queue(hctx
);
1275 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1277 struct blk_mq_hw_ctx
*hctx
;
1279 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1281 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1282 __blk_mq_run_hw_queue(hctx
);
1285 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1287 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1290 blk_mq_stop_hw_queue(hctx
);
1291 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1292 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1294 EXPORT_SYMBOL(blk_mq_delay_queue
);
1296 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1300 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1302 trace_block_rq_insert(hctx
->queue
, rq
);
1305 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1307 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1310 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1313 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1315 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1316 blk_mq_hctx_mark_pending(hctx
, ctx
);
1319 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1320 struct list_head
*list
)
1324 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1327 spin_lock(&ctx
->lock
);
1328 while (!list_empty(list
)) {
1331 rq
= list_first_entry(list
, struct request
, queuelist
);
1332 BUG_ON(rq
->mq_ctx
!= ctx
);
1333 list_del_init(&rq
->queuelist
);
1334 __blk_mq_insert_req_list(hctx
, rq
, false);
1336 blk_mq_hctx_mark_pending(hctx
, ctx
);
1337 spin_unlock(&ctx
->lock
);
1340 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1342 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1343 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1345 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1346 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1347 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1350 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1352 struct blk_mq_ctx
*this_ctx
;
1353 struct request_queue
*this_q
;
1356 LIST_HEAD(ctx_list
);
1359 list_splice_init(&plug
->mq_list
, &list
);
1361 list_sort(NULL
, &list
, plug_ctx_cmp
);
1367 while (!list_empty(&list
)) {
1368 rq
= list_entry_rq(list
.next
);
1369 list_del_init(&rq
->queuelist
);
1371 if (rq
->mq_ctx
!= this_ctx
) {
1373 trace_block_unplug(this_q
, depth
, from_schedule
);
1374 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1379 this_ctx
= rq
->mq_ctx
;
1385 list_add_tail(&rq
->queuelist
, &ctx_list
);
1389 * If 'this_ctx' is set, we know we have entries to complete
1390 * on 'ctx_list'. Do those.
1393 trace_block_unplug(this_q
, depth
, from_schedule
);
1394 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1399 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1401 init_request_from_bio(rq
, bio
);
1403 blk_account_io_start(rq
, true);
1406 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1408 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1409 !blk_queue_nomerges(hctx
->queue
);
1412 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1413 struct blk_mq_ctx
*ctx
,
1414 struct request
*rq
, struct bio
*bio
)
1416 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1417 blk_mq_bio_to_request(rq
, bio
);
1418 spin_lock(&ctx
->lock
);
1420 __blk_mq_insert_request(hctx
, rq
, false);
1421 spin_unlock(&ctx
->lock
);
1424 struct request_queue
*q
= hctx
->queue
;
1426 spin_lock(&ctx
->lock
);
1427 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1428 blk_mq_bio_to_request(rq
, bio
);
1432 spin_unlock(&ctx
->lock
);
1433 __blk_mq_finish_request(hctx
, ctx
, rq
);
1438 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1441 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1443 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1446 static void __blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
,
1449 struct request_queue
*q
= rq
->q
;
1450 struct blk_mq_queue_data bd
= {
1455 struct blk_mq_hw_ctx
*hctx
;
1456 blk_qc_t new_cookie
;
1462 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1465 new_cookie
= request_to_qc_t(hctx
, rq
);
1468 * For OK queue, we are done. For error, kill it. Any other
1469 * error (busy), just add it to our list as we previously
1472 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1473 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1474 *cookie
= new_cookie
;
1478 __blk_mq_requeue_request(rq
);
1480 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1481 *cookie
= BLK_QC_T_NONE
;
1483 blk_mq_end_request(rq
, rq
->errors
);
1488 blk_mq_sched_insert_request(rq
, false, true, false, may_sleep
);
1491 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1492 struct request
*rq
, blk_qc_t
*cookie
)
1494 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1496 __blk_mq_try_issue_directly(rq
, cookie
, false);
1499 unsigned int srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1500 __blk_mq_try_issue_directly(rq
, cookie
, true);
1501 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1505 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1507 const int is_sync
= op_is_sync(bio
->bi_opf
);
1508 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1509 struct blk_mq_alloc_data data
= { .flags
= 0 };
1511 unsigned int request_count
= 0;
1512 struct blk_plug
*plug
;
1513 struct request
*same_queue_rq
= NULL
;
1515 unsigned int wb_acct
;
1517 blk_queue_bounce(q
, &bio
);
1519 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1521 return BLK_QC_T_NONE
;
1524 blk_queue_split(q
, &bio
, q
->bio_split
);
1526 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1527 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1528 return BLK_QC_T_NONE
;
1530 if (blk_mq_sched_bio_merge(q
, bio
))
1531 return BLK_QC_T_NONE
;
1533 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1535 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1537 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1538 if (unlikely(!rq
)) {
1539 __wbt_done(q
->rq_wb
, wb_acct
);
1540 return BLK_QC_T_NONE
;
1543 wbt_track(&rq
->issue_stat
, wb_acct
);
1545 cookie
= request_to_qc_t(data
.hctx
, rq
);
1547 plug
= current
->plug
;
1548 if (unlikely(is_flush_fua
)) {
1549 blk_mq_bio_to_request(rq
, bio
);
1551 blk_mq_sched_insert_request(rq
, false, true, true,
1554 blk_insert_flush(rq
);
1555 blk_mq_run_hw_queue(data
.hctx
, true);
1557 } else if (plug
&& q
->nr_hw_queues
== 1) {
1558 struct request
*last
= NULL
;
1560 blk_mq_bio_to_request(rq
, bio
);
1563 * @request_count may become stale because of schedule
1564 * out, so check the list again.
1566 if (list_empty(&plug
->mq_list
))
1568 else if (blk_queue_nomerges(q
))
1569 request_count
= blk_plug_queued_count(q
);
1572 trace_block_plug(q
);
1574 last
= list_entry_rq(plug
->mq_list
.prev
);
1576 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1577 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1578 blk_flush_plug_list(plug
, false);
1579 trace_block_plug(q
);
1582 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1583 } else if (plug
&& !blk_queue_nomerges(q
)) {
1584 blk_mq_bio_to_request(rq
, bio
);
1587 * We do limited plugging. If the bio can be merged, do that.
1588 * Otherwise the existing request in the plug list will be
1589 * issued. So the plug list will have one request at most
1590 * The plug list might get flushed before this. If that happens,
1591 * the plug list is empty, and same_queue_rq is invalid.
1593 if (list_empty(&plug
->mq_list
))
1594 same_queue_rq
= NULL
;
1596 list_del_init(&same_queue_rq
->queuelist
);
1597 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1600 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1602 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1603 blk_mq_bio_to_request(rq
, bio
);
1604 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1605 } else if (q
->elevator
) {
1606 blk_mq_bio_to_request(rq
, bio
);
1607 blk_mq_sched_insert_request(rq
, false, true, true, true);
1608 } else if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1609 blk_mq_run_hw_queue(data
.hctx
, true);
1612 blk_mq_put_ctx(data
.ctx
);
1616 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1617 unsigned int hctx_idx
)
1621 if (tags
->rqs
&& set
->ops
->exit_request
) {
1624 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1625 struct request
*rq
= tags
->static_rqs
[i
];
1629 set
->ops
->exit_request(set
->driver_data
, rq
,
1631 tags
->static_rqs
[i
] = NULL
;
1635 while (!list_empty(&tags
->page_list
)) {
1636 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1637 list_del_init(&page
->lru
);
1639 * Remove kmemleak object previously allocated in
1640 * blk_mq_init_rq_map().
1642 kmemleak_free(page_address(page
));
1643 __free_pages(page
, page
->private);
1647 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1651 kfree(tags
->static_rqs
);
1652 tags
->static_rqs
= NULL
;
1654 blk_mq_free_tags(tags
);
1657 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1658 unsigned int hctx_idx
,
1659 unsigned int nr_tags
,
1660 unsigned int reserved_tags
)
1662 struct blk_mq_tags
*tags
;
1665 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1666 if (node
== NUMA_NO_NODE
)
1667 node
= set
->numa_node
;
1669 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1670 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1674 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1675 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1678 blk_mq_free_tags(tags
);
1682 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1683 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1685 if (!tags
->static_rqs
) {
1687 blk_mq_free_tags(tags
);
1694 static size_t order_to_size(unsigned int order
)
1696 return (size_t)PAGE_SIZE
<< order
;
1699 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1700 unsigned int hctx_idx
, unsigned int depth
)
1702 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1703 size_t rq_size
, left
;
1706 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1707 if (node
== NUMA_NO_NODE
)
1708 node
= set
->numa_node
;
1710 INIT_LIST_HEAD(&tags
->page_list
);
1713 * rq_size is the size of the request plus driver payload, rounded
1714 * to the cacheline size
1716 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1718 left
= rq_size
* depth
;
1720 for (i
= 0; i
< depth
; ) {
1721 int this_order
= max_order
;
1726 while (this_order
&& left
< order_to_size(this_order
- 1))
1730 page
= alloc_pages_node(node
,
1731 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1737 if (order_to_size(this_order
) < rq_size
)
1744 page
->private = this_order
;
1745 list_add_tail(&page
->lru
, &tags
->page_list
);
1747 p
= page_address(page
);
1749 * Allow kmemleak to scan these pages as they contain pointers
1750 * to additional allocations like via ops->init_request().
1752 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1753 entries_per_page
= order_to_size(this_order
) / rq_size
;
1754 to_do
= min(entries_per_page
, depth
- i
);
1755 left
-= to_do
* rq_size
;
1756 for (j
= 0; j
< to_do
; j
++) {
1757 struct request
*rq
= p
;
1759 tags
->static_rqs
[i
] = rq
;
1760 if (set
->ops
->init_request
) {
1761 if (set
->ops
->init_request(set
->driver_data
,
1764 tags
->static_rqs
[i
] = NULL
;
1776 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1781 * 'cpu' is going away. splice any existing rq_list entries from this
1782 * software queue to the hw queue dispatch list, and ensure that it
1785 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1787 struct blk_mq_hw_ctx
*hctx
;
1788 struct blk_mq_ctx
*ctx
;
1791 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1792 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1794 spin_lock(&ctx
->lock
);
1795 if (!list_empty(&ctx
->rq_list
)) {
1796 list_splice_init(&ctx
->rq_list
, &tmp
);
1797 blk_mq_hctx_clear_pending(hctx
, ctx
);
1799 spin_unlock(&ctx
->lock
);
1801 if (list_empty(&tmp
))
1804 spin_lock(&hctx
->lock
);
1805 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1806 spin_unlock(&hctx
->lock
);
1808 blk_mq_run_hw_queue(hctx
, true);
1812 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1814 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1818 /* hctx->ctxs will be freed in queue's release handler */
1819 static void blk_mq_exit_hctx(struct request_queue
*q
,
1820 struct blk_mq_tag_set
*set
,
1821 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1823 unsigned flush_start_tag
= set
->queue_depth
;
1825 blk_mq_tag_idle(hctx
);
1827 if (set
->ops
->exit_request
)
1828 set
->ops
->exit_request(set
->driver_data
,
1829 hctx
->fq
->flush_rq
, hctx_idx
,
1830 flush_start_tag
+ hctx_idx
);
1832 if (set
->ops
->exit_hctx
)
1833 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1835 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1836 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1838 blk_mq_remove_cpuhp(hctx
);
1839 blk_free_flush_queue(hctx
->fq
);
1840 sbitmap_free(&hctx
->ctx_map
);
1843 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1844 struct blk_mq_tag_set
*set
, int nr_queue
)
1846 struct blk_mq_hw_ctx
*hctx
;
1849 queue_for_each_hw_ctx(q
, hctx
, i
) {
1852 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1856 static int blk_mq_init_hctx(struct request_queue
*q
,
1857 struct blk_mq_tag_set
*set
,
1858 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1861 unsigned flush_start_tag
= set
->queue_depth
;
1863 node
= hctx
->numa_node
;
1864 if (node
== NUMA_NO_NODE
)
1865 node
= hctx
->numa_node
= set
->numa_node
;
1867 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1868 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1869 spin_lock_init(&hctx
->lock
);
1870 INIT_LIST_HEAD(&hctx
->dispatch
);
1872 hctx
->queue_num
= hctx_idx
;
1873 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1875 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1877 hctx
->tags
= set
->tags
[hctx_idx
];
1880 * Allocate space for all possible cpus to avoid allocation at
1883 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1886 goto unregister_cpu_notifier
;
1888 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1894 if (set
->ops
->init_hctx
&&
1895 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1898 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1902 if (set
->ops
->init_request
&&
1903 set
->ops
->init_request(set
->driver_data
,
1904 hctx
->fq
->flush_rq
, hctx_idx
,
1905 flush_start_tag
+ hctx_idx
, node
))
1908 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1909 init_srcu_struct(&hctx
->queue_rq_srcu
);
1916 if (set
->ops
->exit_hctx
)
1917 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1919 sbitmap_free(&hctx
->ctx_map
);
1922 unregister_cpu_notifier
:
1923 blk_mq_remove_cpuhp(hctx
);
1927 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1928 unsigned int nr_hw_queues
)
1932 for_each_possible_cpu(i
) {
1933 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1934 struct blk_mq_hw_ctx
*hctx
;
1937 spin_lock_init(&__ctx
->lock
);
1938 INIT_LIST_HEAD(&__ctx
->rq_list
);
1941 /* If the cpu isn't online, the cpu is mapped to first hctx */
1945 hctx
= blk_mq_map_queue(q
, i
);
1948 * Set local node, IFF we have more than one hw queue. If
1949 * not, we remain on the home node of the device
1951 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1952 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1956 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
1960 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
1961 set
->queue_depth
, set
->reserved_tags
);
1962 if (!set
->tags
[hctx_idx
])
1965 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
1970 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
1971 set
->tags
[hctx_idx
] = NULL
;
1975 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
1976 unsigned int hctx_idx
)
1978 if (set
->tags
[hctx_idx
]) {
1979 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
1980 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
1981 set
->tags
[hctx_idx
] = NULL
;
1985 static void blk_mq_map_swqueue(struct request_queue
*q
,
1986 const struct cpumask
*online_mask
)
1988 unsigned int i
, hctx_idx
;
1989 struct blk_mq_hw_ctx
*hctx
;
1990 struct blk_mq_ctx
*ctx
;
1991 struct blk_mq_tag_set
*set
= q
->tag_set
;
1994 * Avoid others reading imcomplete hctx->cpumask through sysfs
1996 mutex_lock(&q
->sysfs_lock
);
1998 queue_for_each_hw_ctx(q
, hctx
, i
) {
1999 cpumask_clear(hctx
->cpumask
);
2004 * Map software to hardware queues
2006 for_each_possible_cpu(i
) {
2007 /* If the cpu isn't online, the cpu is mapped to first hctx */
2008 if (!cpumask_test_cpu(i
, online_mask
))
2011 hctx_idx
= q
->mq_map
[i
];
2012 /* unmapped hw queue can be remapped after CPU topo changed */
2013 if (!set
->tags
[hctx_idx
] &&
2014 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2016 * If tags initialization fail for some hctx,
2017 * that hctx won't be brought online. In this
2018 * case, remap the current ctx to hctx[0] which
2019 * is guaranteed to always have tags allocated
2024 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2025 hctx
= blk_mq_map_queue(q
, i
);
2027 cpumask_set_cpu(i
, hctx
->cpumask
);
2028 ctx
->index_hw
= hctx
->nr_ctx
;
2029 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2032 mutex_unlock(&q
->sysfs_lock
);
2034 queue_for_each_hw_ctx(q
, hctx
, i
) {
2036 * If no software queues are mapped to this hardware queue,
2037 * disable it and free the request entries.
2039 if (!hctx
->nr_ctx
) {
2040 /* Never unmap queue 0. We need it as a
2041 * fallback in case of a new remap fails
2044 if (i
&& set
->tags
[i
])
2045 blk_mq_free_map_and_requests(set
, i
);
2051 hctx
->tags
= set
->tags
[i
];
2052 WARN_ON(!hctx
->tags
);
2055 * Set the map size to the number of mapped software queues.
2056 * This is more accurate and more efficient than looping
2057 * over all possibly mapped software queues.
2059 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2062 * Initialize batch roundrobin counts
2064 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2065 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2069 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2071 struct blk_mq_hw_ctx
*hctx
;
2074 queue_for_each_hw_ctx(q
, hctx
, i
) {
2076 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2078 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2082 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
2084 struct request_queue
*q
;
2086 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2087 blk_mq_freeze_queue(q
);
2088 queue_set_hctx_shared(q
, shared
);
2089 blk_mq_unfreeze_queue(q
);
2093 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2095 struct blk_mq_tag_set
*set
= q
->tag_set
;
2097 mutex_lock(&set
->tag_list_lock
);
2098 list_del_init(&q
->tag_set_list
);
2099 if (list_is_singular(&set
->tag_list
)) {
2100 /* just transitioned to unshared */
2101 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2102 /* update existing queue */
2103 blk_mq_update_tag_set_depth(set
, false);
2105 mutex_unlock(&set
->tag_list_lock
);
2108 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2109 struct request_queue
*q
)
2113 mutex_lock(&set
->tag_list_lock
);
2115 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2116 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2117 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2118 /* update existing queue */
2119 blk_mq_update_tag_set_depth(set
, true);
2121 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2122 queue_set_hctx_shared(q
, true);
2123 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
2125 mutex_unlock(&set
->tag_list_lock
);
2129 * It is the actual release handler for mq, but we do it from
2130 * request queue's release handler for avoiding use-after-free
2131 * and headache because q->mq_kobj shouldn't have been introduced,
2132 * but we can't group ctx/kctx kobj without it.
2134 void blk_mq_release(struct request_queue
*q
)
2136 struct blk_mq_hw_ctx
*hctx
;
2139 blk_mq_sched_teardown(q
);
2141 /* hctx kobj stays in hctx */
2142 queue_for_each_hw_ctx(q
, hctx
, i
) {
2145 kobject_put(&hctx
->kobj
);
2150 kfree(q
->queue_hw_ctx
);
2153 * release .mq_kobj and sw queue's kobject now because
2154 * both share lifetime with request queue.
2156 blk_mq_sysfs_deinit(q
);
2158 free_percpu(q
->queue_ctx
);
2161 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2163 struct request_queue
*uninit_q
, *q
;
2165 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2167 return ERR_PTR(-ENOMEM
);
2169 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2171 blk_cleanup_queue(uninit_q
);
2175 EXPORT_SYMBOL(blk_mq_init_queue
);
2177 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2178 struct request_queue
*q
)
2181 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2183 blk_mq_sysfs_unregister(q
);
2184 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2190 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2191 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2196 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2203 atomic_set(&hctxs
[i
]->nr_active
, 0);
2204 hctxs
[i
]->numa_node
= node
;
2205 hctxs
[i
]->queue_num
= i
;
2207 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2208 free_cpumask_var(hctxs
[i
]->cpumask
);
2213 blk_mq_hctx_kobj_init(hctxs
[i
]);
2215 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2216 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2220 blk_mq_free_map_and_requests(set
, j
);
2221 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2222 kobject_put(&hctx
->kobj
);
2227 q
->nr_hw_queues
= i
;
2228 blk_mq_sysfs_register(q
);
2231 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2232 struct request_queue
*q
)
2234 /* mark the queue as mq asap */
2235 q
->mq_ops
= set
->ops
;
2237 q
->stats
= blk_alloc_queue_stats();
2241 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2242 blk_stat_rq_ddir
, 2, q
);
2246 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2250 /* init q->mq_kobj and sw queues' kobjects */
2251 blk_mq_sysfs_init(q
);
2253 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2254 GFP_KERNEL
, set
->numa_node
);
2255 if (!q
->queue_hw_ctx
)
2258 q
->mq_map
= set
->mq_map
;
2260 blk_mq_realloc_hw_ctxs(set
, q
);
2261 if (!q
->nr_hw_queues
)
2264 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2265 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2267 q
->nr_queues
= nr_cpu_ids
;
2269 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2271 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2272 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2274 q
->sg_reserved_size
= INT_MAX
;
2276 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2277 INIT_LIST_HEAD(&q
->requeue_list
);
2278 spin_lock_init(&q
->requeue_lock
);
2280 blk_queue_make_request(q
, blk_mq_make_request
);
2283 * Do this after blk_queue_make_request() overrides it...
2285 q
->nr_requests
= set
->queue_depth
;
2288 * Default to classic polling
2292 if (set
->ops
->complete
)
2293 blk_queue_softirq_done(q
, set
->ops
->complete
);
2295 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2298 mutex_lock(&all_q_mutex
);
2300 list_add_tail(&q
->all_q_node
, &all_q_list
);
2301 blk_mq_add_queue_tag_set(set
, q
);
2302 blk_mq_map_swqueue(q
, cpu_online_mask
);
2304 mutex_unlock(&all_q_mutex
);
2307 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2310 ret
= blk_mq_sched_init(q
);
2312 return ERR_PTR(ret
);
2318 kfree(q
->queue_hw_ctx
);
2320 free_percpu(q
->queue_ctx
);
2323 return ERR_PTR(-ENOMEM
);
2325 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2327 void blk_mq_free_queue(struct request_queue
*q
)
2329 struct blk_mq_tag_set
*set
= q
->tag_set
;
2331 mutex_lock(&all_q_mutex
);
2332 list_del_init(&q
->all_q_node
);
2333 mutex_unlock(&all_q_mutex
);
2335 blk_mq_del_queue_tag_set(q
);
2337 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2340 /* Basically redo blk_mq_init_queue with queue frozen */
2341 static void blk_mq_queue_reinit(struct request_queue
*q
,
2342 const struct cpumask
*online_mask
)
2344 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2346 blk_mq_sysfs_unregister(q
);
2349 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2350 * we should change hctx numa_node according to new topology (this
2351 * involves free and re-allocate memory, worthy doing?)
2354 blk_mq_map_swqueue(q
, online_mask
);
2356 blk_mq_sysfs_register(q
);
2360 * New online cpumask which is going to be set in this hotplug event.
2361 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2362 * one-by-one and dynamically allocating this could result in a failure.
2364 static struct cpumask cpuhp_online_new
;
2366 static void blk_mq_queue_reinit_work(void)
2368 struct request_queue
*q
;
2370 mutex_lock(&all_q_mutex
);
2372 * We need to freeze and reinit all existing queues. Freezing
2373 * involves synchronous wait for an RCU grace period and doing it
2374 * one by one may take a long time. Start freezing all queues in
2375 * one swoop and then wait for the completions so that freezing can
2376 * take place in parallel.
2378 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2379 blk_mq_freeze_queue_start(q
);
2380 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2381 blk_mq_freeze_queue_wait(q
);
2383 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2384 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2386 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2387 blk_mq_unfreeze_queue(q
);
2389 mutex_unlock(&all_q_mutex
);
2392 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2394 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2395 blk_mq_queue_reinit_work();
2400 * Before hotadded cpu starts handling requests, new mappings must be
2401 * established. Otherwise, these requests in hw queue might never be
2404 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2405 * for CPU0, and ctx1 for CPU1).
2407 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2408 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2410 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2411 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2412 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2415 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2417 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2418 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2419 blk_mq_queue_reinit_work();
2423 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2427 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2428 if (!__blk_mq_alloc_rq_map(set
, i
))
2435 blk_mq_free_rq_map(set
->tags
[i
]);
2441 * Allocate the request maps associated with this tag_set. Note that this
2442 * may reduce the depth asked for, if memory is tight. set->queue_depth
2443 * will be updated to reflect the allocated depth.
2445 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2450 depth
= set
->queue_depth
;
2452 err
= __blk_mq_alloc_rq_maps(set
);
2456 set
->queue_depth
>>= 1;
2457 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2461 } while (set
->queue_depth
);
2463 if (!set
->queue_depth
|| err
) {
2464 pr_err("blk-mq: failed to allocate request map\n");
2468 if (depth
!= set
->queue_depth
)
2469 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2470 depth
, set
->queue_depth
);
2476 * Alloc a tag set to be associated with one or more request queues.
2477 * May fail with EINVAL for various error conditions. May adjust the
2478 * requested depth down, if if it too large. In that case, the set
2479 * value will be stored in set->queue_depth.
2481 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2485 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2487 if (!set
->nr_hw_queues
)
2489 if (!set
->queue_depth
)
2491 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2494 if (!set
->ops
->queue_rq
)
2497 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2498 pr_info("blk-mq: reduced tag depth to %u\n",
2500 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2504 * If a crashdump is active, then we are potentially in a very
2505 * memory constrained environment. Limit us to 1 queue and
2506 * 64 tags to prevent using too much memory.
2508 if (is_kdump_kernel()) {
2509 set
->nr_hw_queues
= 1;
2510 set
->queue_depth
= min(64U, set
->queue_depth
);
2513 * There is no use for more h/w queues than cpus.
2515 if (set
->nr_hw_queues
> nr_cpu_ids
)
2516 set
->nr_hw_queues
= nr_cpu_ids
;
2518 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2519 GFP_KERNEL
, set
->numa_node
);
2524 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2525 GFP_KERNEL
, set
->numa_node
);
2529 if (set
->ops
->map_queues
)
2530 ret
= set
->ops
->map_queues(set
);
2532 ret
= blk_mq_map_queues(set
);
2534 goto out_free_mq_map
;
2536 ret
= blk_mq_alloc_rq_maps(set
);
2538 goto out_free_mq_map
;
2540 mutex_init(&set
->tag_list_lock
);
2541 INIT_LIST_HEAD(&set
->tag_list
);
2553 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2555 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2559 for (i
= 0; i
< nr_cpu_ids
; i
++)
2560 blk_mq_free_map_and_requests(set
, i
);
2568 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2570 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2572 struct blk_mq_tag_set
*set
= q
->tag_set
;
2573 struct blk_mq_hw_ctx
*hctx
;
2579 blk_mq_freeze_queue(q
);
2580 blk_mq_quiesce_queue(q
);
2583 queue_for_each_hw_ctx(q
, hctx
, i
) {
2587 * If we're using an MQ scheduler, just update the scheduler
2588 * queue depth. This is similar to what the old code would do.
2590 if (!hctx
->sched_tags
) {
2591 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2592 min(nr
, set
->queue_depth
),
2595 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2603 q
->nr_requests
= nr
;
2605 blk_mq_unfreeze_queue(q
);
2606 blk_mq_start_stopped_hw_queues(q
, true);
2611 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2613 struct request_queue
*q
;
2615 if (nr_hw_queues
> nr_cpu_ids
)
2616 nr_hw_queues
= nr_cpu_ids
;
2617 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2620 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2621 blk_mq_freeze_queue(q
);
2623 set
->nr_hw_queues
= nr_hw_queues
;
2624 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2625 blk_mq_realloc_hw_ctxs(set
, q
);
2626 blk_mq_queue_reinit(q
, cpu_online_mask
);
2629 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2630 blk_mq_unfreeze_queue(q
);
2632 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2634 /* Enable polling stats and return whether they were already enabled. */
2635 static bool blk_poll_stats_enable(struct request_queue
*q
)
2637 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2638 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
2640 blk_stat_add_callback(q
, q
->poll_cb
);
2644 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2647 * We don't arm the callback if polling stats are not enabled or the
2648 * callback is already active.
2650 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2651 blk_stat_is_active(q
->poll_cb
))
2654 blk_stat_activate_msecs(q
->poll_cb
, 100);
2657 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2659 struct request_queue
*q
= cb
->data
;
2661 if (cb
->stat
[READ
].nr_samples
)
2662 q
->poll_stat
[READ
] = cb
->stat
[READ
];
2663 if (cb
->stat
[WRITE
].nr_samples
)
2664 q
->poll_stat
[WRITE
] = cb
->stat
[WRITE
];
2667 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2668 struct blk_mq_hw_ctx
*hctx
,
2671 unsigned long ret
= 0;
2674 * If stats collection isn't on, don't sleep but turn it on for
2677 if (!blk_poll_stats_enable(q
))
2681 * As an optimistic guess, use half of the mean service time
2682 * for this type of request. We can (and should) make this smarter.
2683 * For instance, if the completion latencies are tight, we can
2684 * get closer than just half the mean. This is especially
2685 * important on devices where the completion latencies are longer
2688 if (req_op(rq
) == REQ_OP_READ
&& q
->poll_stat
[READ
].nr_samples
)
2689 ret
= (q
->poll_stat
[READ
].mean
+ 1) / 2;
2690 else if (req_op(rq
) == REQ_OP_WRITE
&& q
->poll_stat
[WRITE
].nr_samples
)
2691 ret
= (q
->poll_stat
[WRITE
].mean
+ 1) / 2;
2696 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2697 struct blk_mq_hw_ctx
*hctx
,
2700 struct hrtimer_sleeper hs
;
2701 enum hrtimer_mode mode
;
2705 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2711 * -1: don't ever hybrid sleep
2712 * 0: use half of prev avg
2713 * >0: use this specific value
2715 if (q
->poll_nsec
== -1)
2717 else if (q
->poll_nsec
> 0)
2718 nsecs
= q
->poll_nsec
;
2720 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2725 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2728 * This will be replaced with the stats tracking code, using
2729 * 'avg_completion_time / 2' as the pre-sleep target.
2733 mode
= HRTIMER_MODE_REL
;
2734 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2735 hrtimer_set_expires(&hs
.timer
, kt
);
2737 hrtimer_init_sleeper(&hs
, current
);
2739 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2741 set_current_state(TASK_UNINTERRUPTIBLE
);
2742 hrtimer_start_expires(&hs
.timer
, mode
);
2745 hrtimer_cancel(&hs
.timer
);
2746 mode
= HRTIMER_MODE_ABS
;
2747 } while (hs
.task
&& !signal_pending(current
));
2749 __set_current_state(TASK_RUNNING
);
2750 destroy_hrtimer_on_stack(&hs
.timer
);
2754 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2756 struct request_queue
*q
= hctx
->queue
;
2760 * If we sleep, have the caller restart the poll loop to reset
2761 * the state. Like for the other success return cases, the
2762 * caller is responsible for checking if the IO completed. If
2763 * the IO isn't complete, we'll get called again and will go
2764 * straight to the busy poll loop.
2766 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2769 hctx
->poll_considered
++;
2771 state
= current
->state
;
2772 while (!need_resched()) {
2775 hctx
->poll_invoked
++;
2777 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2779 hctx
->poll_success
++;
2780 set_current_state(TASK_RUNNING
);
2784 if (signal_pending_state(state
, current
))
2785 set_current_state(TASK_RUNNING
);
2787 if (current
->state
== TASK_RUNNING
)
2797 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2799 struct blk_mq_hw_ctx
*hctx
;
2800 struct blk_plug
*plug
;
2803 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2804 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2807 plug
= current
->plug
;
2809 blk_flush_plug_list(plug
, false);
2811 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2812 if (!blk_qc_t_is_internal(cookie
))
2813 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2815 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2817 return __blk_mq_poll(hctx
, rq
);
2819 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2821 void blk_mq_disable_hotplug(void)
2823 mutex_lock(&all_q_mutex
);
2826 void blk_mq_enable_hotplug(void)
2828 mutex_unlock(&all_q_mutex
);
2831 static int __init
blk_mq_init(void)
2833 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2834 blk_mq_hctx_notify_dead
);
2836 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2837 blk_mq_queue_reinit_prepare
,
2838 blk_mq_queue_reinit_dead
);
2841 subsys_initcall(blk_mq_init
);