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 struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
208 unsigned int tag
, unsigned int op
)
210 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
211 struct request
*rq
= tags
->static_rqs
[tag
];
213 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
215 rq
->internal_tag
= tag
;
217 if (blk_mq_tag_busy(data
->hctx
)) {
218 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
219 atomic_inc(&data
->hctx
->nr_active
);
222 rq
->internal_tag
= -1;
223 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
226 INIT_LIST_HEAD(&rq
->queuelist
);
227 /* csd/requeue_work/fifo_time is initialized before use */
229 rq
->mq_ctx
= data
->ctx
;
231 if (blk_queue_io_stat(data
->q
))
232 rq
->rq_flags
|= RQF_IO_STAT
;
233 /* do not touch atomic flags, it needs atomic ops against the timer */
235 INIT_HLIST_NODE(&rq
->hash
);
236 RB_CLEAR_NODE(&rq
->rb_node
);
239 rq
->start_time
= jiffies
;
240 #ifdef CONFIG_BLK_CGROUP
242 set_start_time_ns(rq
);
243 rq
->io_start_time_ns
= 0;
245 rq
->nr_phys_segments
= 0;
246 #if defined(CONFIG_BLK_DEV_INTEGRITY)
247 rq
->nr_integrity_segments
= 0;
250 /* tag was already set */
253 INIT_LIST_HEAD(&rq
->timeout_list
);
257 rq
->end_io_data
= NULL
;
260 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
264 static struct request
*blk_mq_get_request(struct request_queue
*q
,
265 struct bio
*bio
, unsigned int op
,
266 struct blk_mq_alloc_data
*data
)
268 struct elevator_queue
*e
= q
->elevator
;
272 blk_queue_enter_live(q
);
274 if (likely(!data
->ctx
))
275 data
->ctx
= blk_mq_get_ctx(q
);
276 if (likely(!data
->hctx
))
277 data
->hctx
= blk_mq_map_queue(q
, data
->ctx
->cpu
);
280 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
283 * Flush requests are special and go directly to the
286 if (!op_is_flush(op
) && e
->type
->ops
.mq
.limit_depth
)
287 e
->type
->ops
.mq
.limit_depth(op
, data
);
290 tag
= blk_mq_get_tag(data
);
291 if (tag
== BLK_MQ_TAG_FAIL
) {
296 rq
= blk_mq_rq_ctx_init(data
, tag
, op
);
297 if (!op_is_flush(op
)) {
299 if (e
&& e
->type
->ops
.mq
.prepare_request
) {
300 if (e
->type
->icq_cache
&& rq_ioc(bio
))
301 blk_mq_sched_assign_ioc(rq
, bio
);
303 e
->type
->ops
.mq
.prepare_request(rq
, bio
);
304 rq
->rq_flags
|= RQF_ELVPRIV
;
307 data
->hctx
->queued
++;
311 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
314 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
318 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
322 rq
= blk_mq_get_request(q
, NULL
, rw
, &alloc_data
);
324 blk_mq_put_ctx(alloc_data
.ctx
);
328 return ERR_PTR(-EWOULDBLOCK
);
331 rq
->__sector
= (sector_t
) -1;
332 rq
->bio
= rq
->biotail
= NULL
;
335 EXPORT_SYMBOL(blk_mq_alloc_request
);
337 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
338 unsigned int flags
, unsigned int hctx_idx
)
340 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
346 * If the tag allocator sleeps we could get an allocation for a
347 * different hardware context. No need to complicate the low level
348 * allocator for this for the rare use case of a command tied to
351 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
352 return ERR_PTR(-EINVAL
);
354 if (hctx_idx
>= q
->nr_hw_queues
)
355 return ERR_PTR(-EIO
);
357 ret
= blk_queue_enter(q
, true);
362 * Check if the hardware context is actually mapped to anything.
363 * If not tell the caller that it should skip this queue.
365 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
366 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
368 return ERR_PTR(-EXDEV
);
370 cpu
= cpumask_first(alloc_data
.hctx
->cpumask
);
371 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
373 rq
= blk_mq_get_request(q
, NULL
, rw
, &alloc_data
);
378 return ERR_PTR(-EWOULDBLOCK
);
382 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
384 void blk_mq_free_request(struct request
*rq
)
386 struct request_queue
*q
= rq
->q
;
387 struct elevator_queue
*e
= q
->elevator
;
388 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
389 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
390 const int sched_tag
= rq
->internal_tag
;
392 if (rq
->rq_flags
& RQF_ELVPRIV
) {
393 if (e
&& e
->type
->ops
.mq
.finish_request
)
394 e
->type
->ops
.mq
.finish_request(rq
);
396 put_io_context(rq
->elv
.icq
->ioc
);
401 ctx
->rq_completed
[rq_is_sync(rq
)]++;
402 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
403 atomic_dec(&hctx
->nr_active
);
405 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
408 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
409 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
411 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
413 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
414 blk_mq_sched_restart(hctx
);
417 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
419 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
421 blk_account_io_done(rq
);
424 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
425 rq
->end_io(rq
, error
);
427 if (unlikely(blk_bidi_rq(rq
)))
428 blk_mq_free_request(rq
->next_rq
);
429 blk_mq_free_request(rq
);
432 EXPORT_SYMBOL(__blk_mq_end_request
);
434 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
436 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
438 __blk_mq_end_request(rq
, error
);
440 EXPORT_SYMBOL(blk_mq_end_request
);
442 static void __blk_mq_complete_request_remote(void *data
)
444 struct request
*rq
= data
;
446 rq
->q
->softirq_done_fn(rq
);
449 static void __blk_mq_complete_request(struct request
*rq
)
451 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
455 if (rq
->internal_tag
!= -1)
456 blk_mq_sched_completed_request(rq
);
457 if (rq
->rq_flags
& RQF_STATS
) {
458 blk_mq_poll_stats_start(rq
->q
);
462 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
463 rq
->q
->softirq_done_fn(rq
);
468 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
469 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
471 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
472 rq
->csd
.func
= __blk_mq_complete_request_remote
;
475 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
477 rq
->q
->softirq_done_fn(rq
);
483 * blk_mq_complete_request - end I/O on a request
484 * @rq: the request being processed
487 * Ends all I/O on a request. It does not handle partial completions.
488 * The actual completion happens out-of-order, through a IPI handler.
490 void blk_mq_complete_request(struct request
*rq
)
492 struct request_queue
*q
= rq
->q
;
494 if (unlikely(blk_should_fake_timeout(q
)))
496 if (!blk_mark_rq_complete(rq
))
497 __blk_mq_complete_request(rq
);
499 EXPORT_SYMBOL(blk_mq_complete_request
);
501 int blk_mq_request_started(struct request
*rq
)
503 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
505 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
507 void blk_mq_start_request(struct request
*rq
)
509 struct request_queue
*q
= rq
->q
;
511 blk_mq_sched_started_request(rq
);
513 trace_block_rq_issue(q
, rq
);
515 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
516 blk_stat_set_issue(&rq
->issue_stat
, blk_rq_sectors(rq
));
517 rq
->rq_flags
|= RQF_STATS
;
518 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
524 * Ensure that ->deadline is visible before set the started
525 * flag and clear the completed flag.
527 smp_mb__before_atomic();
530 * Mark us as started and clear complete. Complete might have been
531 * set if requeue raced with timeout, which then marked it as
532 * complete. So be sure to clear complete again when we start
533 * the request, otherwise we'll ignore the completion event.
535 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
536 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
537 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
538 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
540 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
542 * Make sure space for the drain appears. We know we can do
543 * this because max_hw_segments has been adjusted to be one
544 * fewer than the device can handle.
546 rq
->nr_phys_segments
++;
549 EXPORT_SYMBOL(blk_mq_start_request
);
552 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
553 * flag isn't set yet, so there may be race with timeout handler,
554 * but given rq->deadline is just set in .queue_rq() under
555 * this situation, the race won't be possible in reality because
556 * rq->timeout should be set as big enough to cover the window
557 * between blk_mq_start_request() called from .queue_rq() and
558 * clearing REQ_ATOM_STARTED here.
560 static void __blk_mq_requeue_request(struct request
*rq
)
562 struct request_queue
*q
= rq
->q
;
564 trace_block_rq_requeue(q
, rq
);
565 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
566 blk_mq_sched_requeue_request(rq
);
568 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
569 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
570 rq
->nr_phys_segments
--;
574 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
576 __blk_mq_requeue_request(rq
);
578 BUG_ON(blk_queued_rq(rq
));
579 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
581 EXPORT_SYMBOL(blk_mq_requeue_request
);
583 static void blk_mq_requeue_work(struct work_struct
*work
)
585 struct request_queue
*q
=
586 container_of(work
, struct request_queue
, requeue_work
.work
);
588 struct request
*rq
, *next
;
591 spin_lock_irqsave(&q
->requeue_lock
, flags
);
592 list_splice_init(&q
->requeue_list
, &rq_list
);
593 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
595 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
596 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
599 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
600 list_del_init(&rq
->queuelist
);
601 blk_mq_sched_insert_request(rq
, true, false, false, true);
604 while (!list_empty(&rq_list
)) {
605 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
606 list_del_init(&rq
->queuelist
);
607 blk_mq_sched_insert_request(rq
, false, false, false, true);
610 blk_mq_run_hw_queues(q
, false);
613 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
614 bool kick_requeue_list
)
616 struct request_queue
*q
= rq
->q
;
620 * We abuse this flag that is otherwise used by the I/O scheduler to
621 * request head insertation from the workqueue.
623 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
625 spin_lock_irqsave(&q
->requeue_lock
, flags
);
627 rq
->rq_flags
|= RQF_SOFTBARRIER
;
628 list_add(&rq
->queuelist
, &q
->requeue_list
);
630 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
632 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
634 if (kick_requeue_list
)
635 blk_mq_kick_requeue_list(q
);
637 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
639 void blk_mq_kick_requeue_list(struct request_queue
*q
)
641 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
643 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
645 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
648 kblockd_schedule_delayed_work(&q
->requeue_work
,
649 msecs_to_jiffies(msecs
));
651 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
653 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
655 if (tag
< tags
->nr_tags
) {
656 prefetch(tags
->rqs
[tag
]);
657 return tags
->rqs
[tag
];
662 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
664 struct blk_mq_timeout_data
{
666 unsigned int next_set
;
669 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
671 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
672 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
675 * We know that complete is set at this point. If STARTED isn't set
676 * anymore, then the request isn't active and the "timeout" should
677 * just be ignored. This can happen due to the bitflag ordering.
678 * Timeout first checks if STARTED is set, and if it is, assumes
679 * the request is active. But if we race with completion, then
680 * both flags will get cleared. So check here again, and ignore
681 * a timeout event with a request that isn't active.
683 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
687 ret
= ops
->timeout(req
, reserved
);
691 __blk_mq_complete_request(req
);
693 case BLK_EH_RESET_TIMER
:
695 blk_clear_rq_complete(req
);
697 case BLK_EH_NOT_HANDLED
:
700 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
705 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
706 struct request
*rq
, void *priv
, bool reserved
)
708 struct blk_mq_timeout_data
*data
= priv
;
710 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
714 * The rq being checked may have been freed and reallocated
715 * out already here, we avoid this race by checking rq->deadline
716 * and REQ_ATOM_COMPLETE flag together:
718 * - if rq->deadline is observed as new value because of
719 * reusing, the rq won't be timed out because of timing.
720 * - if rq->deadline is observed as previous value,
721 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
722 * because we put a barrier between setting rq->deadline
723 * and clearing the flag in blk_mq_start_request(), so
724 * this rq won't be timed out too.
726 if (time_after_eq(jiffies
, rq
->deadline
)) {
727 if (!blk_mark_rq_complete(rq
))
728 blk_mq_rq_timed_out(rq
, reserved
);
729 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
730 data
->next
= rq
->deadline
;
735 static void blk_mq_timeout_work(struct work_struct
*work
)
737 struct request_queue
*q
=
738 container_of(work
, struct request_queue
, timeout_work
);
739 struct blk_mq_timeout_data data
= {
745 /* A deadlock might occur if a request is stuck requiring a
746 * timeout at the same time a queue freeze is waiting
747 * completion, since the timeout code would not be able to
748 * acquire the queue reference here.
750 * That's why we don't use blk_queue_enter here; instead, we use
751 * percpu_ref_tryget directly, because we need to be able to
752 * obtain a reference even in the short window between the queue
753 * starting to freeze, by dropping the first reference in
754 * blk_freeze_queue_start, and the moment the last request is
755 * consumed, marked by the instant q_usage_counter reaches
758 if (!percpu_ref_tryget(&q
->q_usage_counter
))
761 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
764 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
765 mod_timer(&q
->timeout
, data
.next
);
767 struct blk_mq_hw_ctx
*hctx
;
769 queue_for_each_hw_ctx(q
, hctx
, i
) {
770 /* the hctx may be unmapped, so check it here */
771 if (blk_mq_hw_queue_mapped(hctx
))
772 blk_mq_tag_idle(hctx
);
778 struct flush_busy_ctx_data
{
779 struct blk_mq_hw_ctx
*hctx
;
780 struct list_head
*list
;
783 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
785 struct flush_busy_ctx_data
*flush_data
= data
;
786 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
787 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
789 sbitmap_clear_bit(sb
, bitnr
);
790 spin_lock(&ctx
->lock
);
791 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
792 spin_unlock(&ctx
->lock
);
797 * Process software queues that have been marked busy, splicing them
798 * to the for-dispatch
800 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
802 struct flush_busy_ctx_data data
= {
807 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
809 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
811 static inline unsigned int queued_to_index(unsigned int queued
)
816 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
819 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
822 struct blk_mq_alloc_data data
= {
824 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
825 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
828 might_sleep_if(wait
);
833 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
834 data
.flags
|= BLK_MQ_REQ_RESERVED
;
836 rq
->tag
= blk_mq_get_tag(&data
);
838 if (blk_mq_tag_busy(data
.hctx
)) {
839 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
840 atomic_inc(&data
.hctx
->nr_active
);
842 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
848 return rq
->tag
!= -1;
851 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx
*hctx
,
854 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
857 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
858 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
859 atomic_dec(&hctx
->nr_active
);
863 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx
*hctx
,
866 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
869 __blk_mq_put_driver_tag(hctx
, rq
);
872 static void blk_mq_put_driver_tag(struct request
*rq
)
874 struct blk_mq_hw_ctx
*hctx
;
876 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
879 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
880 __blk_mq_put_driver_tag(hctx
, rq
);
884 * If we fail getting a driver tag because all the driver tags are already
885 * assigned and on the dispatch list, BUT the first entry does not have a
886 * tag, then we could deadlock. For that case, move entries with assigned
887 * driver tags to the front, leaving the set of tagged requests in the
888 * same order, and the untagged set in the same order.
890 static bool reorder_tags_to_front(struct list_head
*list
)
892 struct request
*rq
, *tmp
, *first
= NULL
;
894 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
898 list_move(&rq
->queuelist
, list
);
904 return first
!= NULL
;
907 static int blk_mq_dispatch_wake(wait_queue_t
*wait
, unsigned mode
, int flags
,
910 struct blk_mq_hw_ctx
*hctx
;
912 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
914 list_del(&wait
->task_list
);
915 clear_bit_unlock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
);
916 blk_mq_run_hw_queue(hctx
, true);
920 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx
*hctx
)
922 struct sbq_wait_state
*ws
;
925 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
926 * The thread which wins the race to grab this bit adds the hardware
927 * queue to the wait queue.
929 if (test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
) ||
930 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
933 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
934 ws
= bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
);
937 * As soon as this returns, it's no longer safe to fiddle with
938 * hctx->dispatch_wait, since a completion can wake up the wait queue
939 * and unlock the bit.
941 add_wait_queue(&ws
->wait
, &hctx
->dispatch_wait
);
945 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
)
947 struct blk_mq_hw_ctx
*hctx
;
951 if (list_empty(list
))
955 * Now process all the entries, sending them to the driver.
959 struct blk_mq_queue_data bd
;
962 rq
= list_first_entry(list
, struct request
, queuelist
);
963 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
964 if (!queued
&& reorder_tags_to_front(list
))
968 * The initial allocation attempt failed, so we need to
969 * rerun the hardware queue when a tag is freed.
971 if (!blk_mq_dispatch_wait_add(hctx
))
975 * It's possible that a tag was freed in the window
976 * between the allocation failure and adding the
977 * hardware queue to the wait queue.
979 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
983 list_del_init(&rq
->queuelist
);
988 * Flag last if we have no more requests, or if we have more
989 * but can't assign a driver tag to it.
991 if (list_empty(list
))
996 nxt
= list_first_entry(list
, struct request
, queuelist
);
997 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1000 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1001 if (ret
== BLK_STS_RESOURCE
) {
1002 blk_mq_put_driver_tag_hctx(hctx
, rq
);
1003 list_add(&rq
->queuelist
, list
);
1004 __blk_mq_requeue_request(rq
);
1008 if (unlikely(ret
!= BLK_STS_OK
)) {
1010 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1015 } while (!list_empty(list
));
1017 hctx
->dispatched
[queued_to_index(queued
)]++;
1020 * Any items that need requeuing? Stuff them into hctx->dispatch,
1021 * that is where we will continue on next queue run.
1023 if (!list_empty(list
)) {
1025 * If an I/O scheduler has been configured and we got a driver
1026 * tag for the next request already, free it again.
1028 rq
= list_first_entry(list
, struct request
, queuelist
);
1029 blk_mq_put_driver_tag(rq
);
1031 spin_lock(&hctx
->lock
);
1032 list_splice_init(list
, &hctx
->dispatch
);
1033 spin_unlock(&hctx
->lock
);
1036 * If SCHED_RESTART was set by the caller of this function and
1037 * it is no longer set that means that it was cleared by another
1038 * thread and hence that a queue rerun is needed.
1040 * If TAG_WAITING is set that means that an I/O scheduler has
1041 * been configured and another thread is waiting for a driver
1042 * tag. To guarantee fairness, do not rerun this hardware queue
1043 * but let the other thread grab the driver tag.
1045 * If no I/O scheduler has been configured it is possible that
1046 * the hardware queue got stopped and restarted before requests
1047 * were pushed back onto the dispatch list. Rerun the queue to
1048 * avoid starvation. Notes:
1049 * - blk_mq_run_hw_queue() checks whether or not a queue has
1050 * been stopped before rerunning a queue.
1051 * - Some but not all block drivers stop a queue before
1052 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1055 if (!blk_mq_sched_needs_restart(hctx
) &&
1056 !test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1057 blk_mq_run_hw_queue(hctx
, true);
1060 return (queued
+ errors
) != 0;
1063 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1067 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1068 cpu_online(hctx
->next_cpu
));
1070 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1072 blk_mq_sched_dispatch_requests(hctx
);
1077 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1078 blk_mq_sched_dispatch_requests(hctx
);
1079 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1084 * It'd be great if the workqueue API had a way to pass
1085 * in a mask and had some smarts for more clever placement.
1086 * For now we just round-robin here, switching for every
1087 * BLK_MQ_CPU_WORK_BATCH queued items.
1089 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1091 if (hctx
->queue
->nr_hw_queues
== 1)
1092 return WORK_CPU_UNBOUND
;
1094 if (--hctx
->next_cpu_batch
<= 0) {
1097 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1098 if (next_cpu
>= nr_cpu_ids
)
1099 next_cpu
= cpumask_first(hctx
->cpumask
);
1101 hctx
->next_cpu
= next_cpu
;
1102 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1105 return hctx
->next_cpu
;
1108 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1109 unsigned long msecs
)
1111 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
1112 !blk_mq_hw_queue_mapped(hctx
)))
1115 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1116 int cpu
= get_cpu();
1117 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1118 __blk_mq_run_hw_queue(hctx
);
1126 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1128 msecs_to_jiffies(msecs
));
1131 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1133 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1135 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1137 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1139 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1141 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1143 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1145 struct blk_mq_hw_ctx
*hctx
;
1148 queue_for_each_hw_ctx(q
, hctx
, i
) {
1149 if (!blk_mq_hctx_has_pending(hctx
) ||
1150 blk_mq_hctx_stopped(hctx
))
1153 blk_mq_run_hw_queue(hctx
, async
);
1156 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1159 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1160 * @q: request queue.
1162 * The caller is responsible for serializing this function against
1163 * blk_mq_{start,stop}_hw_queue().
1165 bool blk_mq_queue_stopped(struct request_queue
*q
)
1167 struct blk_mq_hw_ctx
*hctx
;
1170 queue_for_each_hw_ctx(q
, hctx
, i
)
1171 if (blk_mq_hctx_stopped(hctx
))
1176 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1178 static void __blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool sync
)
1181 cancel_delayed_work_sync(&hctx
->run_work
);
1183 cancel_delayed_work(&hctx
->run_work
);
1185 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1188 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1190 __blk_mq_stop_hw_queue(hctx
, false);
1192 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1194 static void __blk_mq_stop_hw_queues(struct request_queue
*q
, bool sync
)
1196 struct blk_mq_hw_ctx
*hctx
;
1199 queue_for_each_hw_ctx(q
, hctx
, i
)
1200 __blk_mq_stop_hw_queue(hctx
, sync
);
1203 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1205 __blk_mq_stop_hw_queues(q
, false);
1207 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1209 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1211 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1213 blk_mq_run_hw_queue(hctx
, false);
1215 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1217 void blk_mq_start_hw_queues(struct request_queue
*q
)
1219 struct blk_mq_hw_ctx
*hctx
;
1222 queue_for_each_hw_ctx(q
, hctx
, i
)
1223 blk_mq_start_hw_queue(hctx
);
1225 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1227 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1229 if (!blk_mq_hctx_stopped(hctx
))
1232 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1233 blk_mq_run_hw_queue(hctx
, async
);
1235 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1237 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1239 struct blk_mq_hw_ctx
*hctx
;
1242 queue_for_each_hw_ctx(q
, hctx
, i
)
1243 blk_mq_start_stopped_hw_queue(hctx
, async
);
1245 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1247 static void blk_mq_run_work_fn(struct work_struct
*work
)
1249 struct blk_mq_hw_ctx
*hctx
;
1251 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1254 * If we are stopped, don't run the queue. The exception is if
1255 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1256 * the STOPPED bit and run it.
1258 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)) {
1259 if (!test_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
))
1262 clear_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1263 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1266 __blk_mq_run_hw_queue(hctx
);
1270 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1272 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1276 * Stop the hw queue, then modify currently delayed work.
1277 * This should prevent us from running the queue prematurely.
1278 * Mark the queue as auto-clearing STOPPED when it runs.
1280 blk_mq_stop_hw_queue(hctx
);
1281 set_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1282 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1284 msecs_to_jiffies(msecs
));
1286 EXPORT_SYMBOL(blk_mq_delay_queue
);
1288 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1292 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1294 trace_block_rq_insert(hctx
->queue
, rq
);
1297 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1299 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1302 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1305 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1307 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1308 blk_mq_hctx_mark_pending(hctx
, ctx
);
1311 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1312 struct list_head
*list
)
1316 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1319 spin_lock(&ctx
->lock
);
1320 while (!list_empty(list
)) {
1323 rq
= list_first_entry(list
, struct request
, queuelist
);
1324 BUG_ON(rq
->mq_ctx
!= ctx
);
1325 list_del_init(&rq
->queuelist
);
1326 __blk_mq_insert_req_list(hctx
, rq
, false);
1328 blk_mq_hctx_mark_pending(hctx
, ctx
);
1329 spin_unlock(&ctx
->lock
);
1332 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1334 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1335 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1337 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1338 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1339 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1342 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1344 struct blk_mq_ctx
*this_ctx
;
1345 struct request_queue
*this_q
;
1348 LIST_HEAD(ctx_list
);
1351 list_splice_init(&plug
->mq_list
, &list
);
1353 list_sort(NULL
, &list
, plug_ctx_cmp
);
1359 while (!list_empty(&list
)) {
1360 rq
= list_entry_rq(list
.next
);
1361 list_del_init(&rq
->queuelist
);
1363 if (rq
->mq_ctx
!= this_ctx
) {
1365 trace_block_unplug(this_q
, depth
, from_schedule
);
1366 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1371 this_ctx
= rq
->mq_ctx
;
1377 list_add_tail(&rq
->queuelist
, &ctx_list
);
1381 * If 'this_ctx' is set, we know we have entries to complete
1382 * on 'ctx_list'. Do those.
1385 trace_block_unplug(this_q
, depth
, from_schedule
);
1386 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1391 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1393 blk_init_request_from_bio(rq
, bio
);
1395 blk_account_io_start(rq
, true);
1398 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1400 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1401 !blk_queue_nomerges(hctx
->queue
);
1404 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx
*hctx
,
1405 struct blk_mq_ctx
*ctx
,
1408 spin_lock(&ctx
->lock
);
1409 __blk_mq_insert_request(hctx
, rq
, false);
1410 spin_unlock(&ctx
->lock
);
1413 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1416 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1418 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1421 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1423 blk_qc_t
*cookie
, bool may_sleep
)
1425 struct request_queue
*q
= rq
->q
;
1426 struct blk_mq_queue_data bd
= {
1430 blk_qc_t new_cookie
;
1432 bool run_queue
= true;
1434 if (blk_mq_hctx_stopped(hctx
)) {
1442 if (!blk_mq_get_driver_tag(rq
, NULL
, false))
1445 new_cookie
= request_to_qc_t(hctx
, rq
);
1448 * For OK queue, we are done. For error, kill it. Any other
1449 * error (busy), just add it to our list as we previously
1452 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1455 *cookie
= new_cookie
;
1457 case BLK_STS_RESOURCE
:
1458 __blk_mq_requeue_request(rq
);
1461 *cookie
= BLK_QC_T_NONE
;
1462 blk_mq_end_request(rq
, ret
);
1467 blk_mq_sched_insert_request(rq
, false, run_queue
, false, may_sleep
);
1470 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1471 struct request
*rq
, blk_qc_t
*cookie
)
1473 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1475 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false);
1478 unsigned int srcu_idx
;
1482 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1483 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, true);
1484 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1488 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1490 const int is_sync
= op_is_sync(bio
->bi_opf
);
1491 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1492 struct blk_mq_alloc_data data
= { .flags
= 0 };
1494 unsigned int request_count
= 0;
1495 struct blk_plug
*plug
;
1496 struct request
*same_queue_rq
= NULL
;
1498 unsigned int wb_acct
;
1500 blk_queue_bounce(q
, &bio
);
1502 blk_queue_split(q
, &bio
);
1504 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1506 return BLK_QC_T_NONE
;
1509 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1510 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1511 return BLK_QC_T_NONE
;
1513 if (blk_mq_sched_bio_merge(q
, bio
))
1514 return BLK_QC_T_NONE
;
1516 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1518 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1520 rq
= blk_mq_get_request(q
, bio
, bio
->bi_opf
, &data
);
1521 if (unlikely(!rq
)) {
1522 __wbt_done(q
->rq_wb
, wb_acct
);
1523 return BLK_QC_T_NONE
;
1526 wbt_track(&rq
->issue_stat
, wb_acct
);
1528 cookie
= request_to_qc_t(data
.hctx
, rq
);
1530 plug
= current
->plug
;
1531 if (unlikely(is_flush_fua
)) {
1532 blk_mq_put_ctx(data
.ctx
);
1533 blk_mq_bio_to_request(rq
, bio
);
1535 blk_mq_sched_insert_request(rq
, false, true, true,
1538 blk_insert_flush(rq
);
1539 blk_mq_run_hw_queue(data
.hctx
, true);
1541 } else if (plug
&& q
->nr_hw_queues
== 1) {
1542 struct request
*last
= NULL
;
1544 blk_mq_put_ctx(data
.ctx
);
1545 blk_mq_bio_to_request(rq
, bio
);
1548 * @request_count may become stale because of schedule
1549 * out, so check the list again.
1551 if (list_empty(&plug
->mq_list
))
1553 else if (blk_queue_nomerges(q
))
1554 request_count
= blk_plug_queued_count(q
);
1557 trace_block_plug(q
);
1559 last
= list_entry_rq(plug
->mq_list
.prev
);
1561 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1562 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1563 blk_flush_plug_list(plug
, false);
1564 trace_block_plug(q
);
1567 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1568 } else if (plug
&& !blk_queue_nomerges(q
)) {
1569 blk_mq_bio_to_request(rq
, bio
);
1572 * We do limited plugging. If the bio can be merged, do that.
1573 * Otherwise the existing request in the plug list will be
1574 * issued. So the plug list will have one request at most
1575 * The plug list might get flushed before this. If that happens,
1576 * the plug list is empty, and same_queue_rq is invalid.
1578 if (list_empty(&plug
->mq_list
))
1579 same_queue_rq
= NULL
;
1581 list_del_init(&same_queue_rq
->queuelist
);
1582 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1584 blk_mq_put_ctx(data
.ctx
);
1586 if (same_queue_rq
) {
1587 data
.hctx
= blk_mq_map_queue(q
,
1588 same_queue_rq
->mq_ctx
->cpu
);
1589 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1592 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1593 blk_mq_put_ctx(data
.ctx
);
1594 blk_mq_bio_to_request(rq
, bio
);
1595 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1596 } else if (q
->elevator
) {
1597 blk_mq_put_ctx(data
.ctx
);
1598 blk_mq_bio_to_request(rq
, bio
);
1599 blk_mq_sched_insert_request(rq
, false, true, true, true);
1601 blk_mq_put_ctx(data
.ctx
);
1602 blk_mq_bio_to_request(rq
, bio
);
1603 blk_mq_queue_io(data
.hctx
, data
.ctx
, rq
);
1604 blk_mq_run_hw_queue(data
.hctx
, true);
1610 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1611 unsigned int hctx_idx
)
1615 if (tags
->rqs
&& set
->ops
->exit_request
) {
1618 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1619 struct request
*rq
= tags
->static_rqs
[i
];
1623 set
->ops
->exit_request(set
, rq
, hctx_idx
);
1624 tags
->static_rqs
[i
] = NULL
;
1628 while (!list_empty(&tags
->page_list
)) {
1629 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1630 list_del_init(&page
->lru
);
1632 * Remove kmemleak object previously allocated in
1633 * blk_mq_init_rq_map().
1635 kmemleak_free(page_address(page
));
1636 __free_pages(page
, page
->private);
1640 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1644 kfree(tags
->static_rqs
);
1645 tags
->static_rqs
= NULL
;
1647 blk_mq_free_tags(tags
);
1650 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1651 unsigned int hctx_idx
,
1652 unsigned int nr_tags
,
1653 unsigned int reserved_tags
)
1655 struct blk_mq_tags
*tags
;
1658 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1659 if (node
== NUMA_NO_NODE
)
1660 node
= set
->numa_node
;
1662 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1663 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1667 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1668 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1671 blk_mq_free_tags(tags
);
1675 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1676 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1678 if (!tags
->static_rqs
) {
1680 blk_mq_free_tags(tags
);
1687 static size_t order_to_size(unsigned int order
)
1689 return (size_t)PAGE_SIZE
<< order
;
1692 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1693 unsigned int hctx_idx
, unsigned int depth
)
1695 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1696 size_t rq_size
, left
;
1699 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1700 if (node
== NUMA_NO_NODE
)
1701 node
= set
->numa_node
;
1703 INIT_LIST_HEAD(&tags
->page_list
);
1706 * rq_size is the size of the request plus driver payload, rounded
1707 * to the cacheline size
1709 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1711 left
= rq_size
* depth
;
1713 for (i
= 0; i
< depth
; ) {
1714 int this_order
= max_order
;
1719 while (this_order
&& left
< order_to_size(this_order
- 1))
1723 page
= alloc_pages_node(node
,
1724 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1730 if (order_to_size(this_order
) < rq_size
)
1737 page
->private = this_order
;
1738 list_add_tail(&page
->lru
, &tags
->page_list
);
1740 p
= page_address(page
);
1742 * Allow kmemleak to scan these pages as they contain pointers
1743 * to additional allocations like via ops->init_request().
1745 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1746 entries_per_page
= order_to_size(this_order
) / rq_size
;
1747 to_do
= min(entries_per_page
, depth
- i
);
1748 left
-= to_do
* rq_size
;
1749 for (j
= 0; j
< to_do
; j
++) {
1750 struct request
*rq
= p
;
1752 tags
->static_rqs
[i
] = rq
;
1753 if (set
->ops
->init_request
) {
1754 if (set
->ops
->init_request(set
, rq
, hctx_idx
,
1756 tags
->static_rqs
[i
] = NULL
;
1768 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1773 * 'cpu' is going away. splice any existing rq_list entries from this
1774 * software queue to the hw queue dispatch list, and ensure that it
1777 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1779 struct blk_mq_hw_ctx
*hctx
;
1780 struct blk_mq_ctx
*ctx
;
1783 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1784 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1786 spin_lock(&ctx
->lock
);
1787 if (!list_empty(&ctx
->rq_list
)) {
1788 list_splice_init(&ctx
->rq_list
, &tmp
);
1789 blk_mq_hctx_clear_pending(hctx
, ctx
);
1791 spin_unlock(&ctx
->lock
);
1793 if (list_empty(&tmp
))
1796 spin_lock(&hctx
->lock
);
1797 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1798 spin_unlock(&hctx
->lock
);
1800 blk_mq_run_hw_queue(hctx
, true);
1804 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1806 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1810 /* hctx->ctxs will be freed in queue's release handler */
1811 static void blk_mq_exit_hctx(struct request_queue
*q
,
1812 struct blk_mq_tag_set
*set
,
1813 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1815 blk_mq_debugfs_unregister_hctx(hctx
);
1817 blk_mq_tag_idle(hctx
);
1819 if (set
->ops
->exit_request
)
1820 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
1822 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1824 if (set
->ops
->exit_hctx
)
1825 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1827 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1828 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1830 blk_mq_remove_cpuhp(hctx
);
1831 blk_free_flush_queue(hctx
->fq
);
1832 sbitmap_free(&hctx
->ctx_map
);
1835 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1836 struct blk_mq_tag_set
*set
, int nr_queue
)
1838 struct blk_mq_hw_ctx
*hctx
;
1841 queue_for_each_hw_ctx(q
, hctx
, i
) {
1844 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1848 static int blk_mq_init_hctx(struct request_queue
*q
,
1849 struct blk_mq_tag_set
*set
,
1850 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1854 node
= hctx
->numa_node
;
1855 if (node
== NUMA_NO_NODE
)
1856 node
= hctx
->numa_node
= set
->numa_node
;
1858 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1859 spin_lock_init(&hctx
->lock
);
1860 INIT_LIST_HEAD(&hctx
->dispatch
);
1862 hctx
->queue_num
= hctx_idx
;
1863 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1865 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1867 hctx
->tags
= set
->tags
[hctx_idx
];
1870 * Allocate space for all possible cpus to avoid allocation at
1873 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1876 goto unregister_cpu_notifier
;
1878 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1884 if (set
->ops
->init_hctx
&&
1885 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1888 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
1891 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1893 goto sched_exit_hctx
;
1895 if (set
->ops
->init_request
&&
1896 set
->ops
->init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
1900 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1901 init_srcu_struct(&hctx
->queue_rq_srcu
);
1903 blk_mq_debugfs_register_hctx(q
, hctx
);
1910 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1912 if (set
->ops
->exit_hctx
)
1913 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1915 sbitmap_free(&hctx
->ctx_map
);
1918 unregister_cpu_notifier
:
1919 blk_mq_remove_cpuhp(hctx
);
1923 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1924 unsigned int nr_hw_queues
)
1928 for_each_possible_cpu(i
) {
1929 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1930 struct blk_mq_hw_ctx
*hctx
;
1933 spin_lock_init(&__ctx
->lock
);
1934 INIT_LIST_HEAD(&__ctx
->rq_list
);
1937 /* If the cpu isn't online, the cpu is mapped to first hctx */
1941 hctx
= blk_mq_map_queue(q
, i
);
1944 * Set local node, IFF we have more than one hw queue. If
1945 * not, we remain on the home node of the device
1947 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1948 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1952 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
1956 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
1957 set
->queue_depth
, set
->reserved_tags
);
1958 if (!set
->tags
[hctx_idx
])
1961 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
1966 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
1967 set
->tags
[hctx_idx
] = NULL
;
1971 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
1972 unsigned int hctx_idx
)
1974 if (set
->tags
[hctx_idx
]) {
1975 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
1976 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
1977 set
->tags
[hctx_idx
] = NULL
;
1981 static void blk_mq_map_swqueue(struct request_queue
*q
,
1982 const struct cpumask
*online_mask
)
1984 unsigned int i
, hctx_idx
;
1985 struct blk_mq_hw_ctx
*hctx
;
1986 struct blk_mq_ctx
*ctx
;
1987 struct blk_mq_tag_set
*set
= q
->tag_set
;
1990 * Avoid others reading imcomplete hctx->cpumask through sysfs
1992 mutex_lock(&q
->sysfs_lock
);
1994 queue_for_each_hw_ctx(q
, hctx
, i
) {
1995 cpumask_clear(hctx
->cpumask
);
2000 * Map software to hardware queues
2002 for_each_possible_cpu(i
) {
2003 /* If the cpu isn't online, the cpu is mapped to first hctx */
2004 if (!cpumask_test_cpu(i
, online_mask
))
2007 hctx_idx
= q
->mq_map
[i
];
2008 /* unmapped hw queue can be remapped after CPU topo changed */
2009 if (!set
->tags
[hctx_idx
] &&
2010 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2012 * If tags initialization fail for some hctx,
2013 * that hctx won't be brought online. In this
2014 * case, remap the current ctx to hctx[0] which
2015 * is guaranteed to always have tags allocated
2020 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2021 hctx
= blk_mq_map_queue(q
, i
);
2023 cpumask_set_cpu(i
, hctx
->cpumask
);
2024 ctx
->index_hw
= hctx
->nr_ctx
;
2025 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2028 mutex_unlock(&q
->sysfs_lock
);
2030 queue_for_each_hw_ctx(q
, hctx
, i
) {
2032 * If no software queues are mapped to this hardware queue,
2033 * disable it and free the request entries.
2035 if (!hctx
->nr_ctx
) {
2036 /* Never unmap queue 0. We need it as a
2037 * fallback in case of a new remap fails
2040 if (i
&& set
->tags
[i
])
2041 blk_mq_free_map_and_requests(set
, i
);
2047 hctx
->tags
= set
->tags
[i
];
2048 WARN_ON(!hctx
->tags
);
2051 * Set the map size to the number of mapped software queues.
2052 * This is more accurate and more efficient than looping
2053 * over all possibly mapped software queues.
2055 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2058 * Initialize batch roundrobin counts
2060 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2061 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2065 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2067 struct blk_mq_hw_ctx
*hctx
;
2070 queue_for_each_hw_ctx(q
, hctx
, i
) {
2072 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2074 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2078 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
2080 struct request_queue
*q
;
2082 lockdep_assert_held(&set
->tag_list_lock
);
2084 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2085 blk_mq_freeze_queue(q
);
2086 queue_set_hctx_shared(q
, shared
);
2087 blk_mq_unfreeze_queue(q
);
2091 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2093 struct blk_mq_tag_set
*set
= q
->tag_set
;
2095 mutex_lock(&set
->tag_list_lock
);
2096 list_del_rcu(&q
->tag_set_list
);
2097 INIT_LIST_HEAD(&q
->tag_set_list
);
2098 if (list_is_singular(&set
->tag_list
)) {
2099 /* just transitioned to unshared */
2100 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2101 /* update existing queue */
2102 blk_mq_update_tag_set_depth(set
, false);
2104 mutex_unlock(&set
->tag_list_lock
);
2109 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2110 struct request_queue
*q
)
2114 mutex_lock(&set
->tag_list_lock
);
2116 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2117 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2118 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2119 /* update existing queue */
2120 blk_mq_update_tag_set_depth(set
, true);
2122 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2123 queue_set_hctx_shared(q
, true);
2124 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2126 mutex_unlock(&set
->tag_list_lock
);
2130 * It is the actual release handler for mq, but we do it from
2131 * request queue's release handler for avoiding use-after-free
2132 * and headache because q->mq_kobj shouldn't have been introduced,
2133 * but we can't group ctx/kctx kobj without it.
2135 void blk_mq_release(struct request_queue
*q
)
2137 struct blk_mq_hw_ctx
*hctx
;
2140 /* hctx kobj stays in hctx */
2141 queue_for_each_hw_ctx(q
, hctx
, i
) {
2144 kobject_put(&hctx
->kobj
);
2149 kfree(q
->queue_hw_ctx
);
2152 * release .mq_kobj and sw queue's kobject now because
2153 * both share lifetime with request queue.
2155 blk_mq_sysfs_deinit(q
);
2157 free_percpu(q
->queue_ctx
);
2160 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2162 struct request_queue
*uninit_q
, *q
;
2164 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2166 return ERR_PTR(-ENOMEM
);
2168 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2170 blk_cleanup_queue(uninit_q
);
2174 EXPORT_SYMBOL(blk_mq_init_queue
);
2176 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2177 struct request_queue
*q
)
2180 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2182 blk_mq_sysfs_unregister(q
);
2183 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2189 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2190 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2195 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2202 atomic_set(&hctxs
[i
]->nr_active
, 0);
2203 hctxs
[i
]->numa_node
= node
;
2204 hctxs
[i
]->queue_num
= i
;
2206 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2207 free_cpumask_var(hctxs
[i
]->cpumask
);
2212 blk_mq_hctx_kobj_init(hctxs
[i
]);
2214 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2215 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2219 blk_mq_free_map_and_requests(set
, j
);
2220 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2221 kobject_put(&hctx
->kobj
);
2226 q
->nr_hw_queues
= i
;
2227 blk_mq_sysfs_register(q
);
2230 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2231 struct request_queue
*q
)
2233 /* mark the queue as mq asap */
2234 q
->mq_ops
= set
->ops
;
2236 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2237 blk_mq_poll_stats_bkt
,
2238 BLK_MQ_POLL_STATS_BKTS
, q
);
2242 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2246 /* init q->mq_kobj and sw queues' kobjects */
2247 blk_mq_sysfs_init(q
);
2249 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2250 GFP_KERNEL
, set
->numa_node
);
2251 if (!q
->queue_hw_ctx
)
2254 q
->mq_map
= set
->mq_map
;
2256 blk_mq_realloc_hw_ctxs(set
, q
);
2257 if (!q
->nr_hw_queues
)
2260 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2261 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2263 q
->nr_queues
= nr_cpu_ids
;
2265 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2267 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2268 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2270 q
->sg_reserved_size
= INT_MAX
;
2272 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2273 INIT_LIST_HEAD(&q
->requeue_list
);
2274 spin_lock_init(&q
->requeue_lock
);
2276 blk_queue_make_request(q
, blk_mq_make_request
);
2279 * Do this after blk_queue_make_request() overrides it...
2281 q
->nr_requests
= set
->queue_depth
;
2284 * Default to classic polling
2288 if (set
->ops
->complete
)
2289 blk_queue_softirq_done(q
, set
->ops
->complete
);
2291 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2294 mutex_lock(&all_q_mutex
);
2296 list_add_tail(&q
->all_q_node
, &all_q_list
);
2297 blk_mq_add_queue_tag_set(set
, q
);
2298 blk_mq_map_swqueue(q
, cpu_online_mask
);
2300 mutex_unlock(&all_q_mutex
);
2303 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2306 ret
= blk_mq_sched_init(q
);
2308 return ERR_PTR(ret
);
2314 kfree(q
->queue_hw_ctx
);
2316 free_percpu(q
->queue_ctx
);
2319 return ERR_PTR(-ENOMEM
);
2321 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2323 void blk_mq_free_queue(struct request_queue
*q
)
2325 struct blk_mq_tag_set
*set
= q
->tag_set
;
2327 mutex_lock(&all_q_mutex
);
2328 list_del_init(&q
->all_q_node
);
2329 mutex_unlock(&all_q_mutex
);
2331 blk_mq_del_queue_tag_set(q
);
2333 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2336 /* Basically redo blk_mq_init_queue with queue frozen */
2337 static void blk_mq_queue_reinit(struct request_queue
*q
,
2338 const struct cpumask
*online_mask
)
2340 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2342 blk_mq_debugfs_unregister_hctxs(q
);
2343 blk_mq_sysfs_unregister(q
);
2346 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2347 * we should change hctx numa_node according to new topology (this
2348 * involves free and re-allocate memory, worthy doing?)
2351 blk_mq_map_swqueue(q
, online_mask
);
2353 blk_mq_sysfs_register(q
);
2354 blk_mq_debugfs_register_hctxs(q
);
2358 * New online cpumask which is going to be set in this hotplug event.
2359 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2360 * one-by-one and dynamically allocating this could result in a failure.
2362 static struct cpumask cpuhp_online_new
;
2364 static void blk_mq_queue_reinit_work(void)
2366 struct request_queue
*q
;
2368 mutex_lock(&all_q_mutex
);
2370 * We need to freeze and reinit all existing queues. Freezing
2371 * involves synchronous wait for an RCU grace period and doing it
2372 * one by one may take a long time. Start freezing all queues in
2373 * one swoop and then wait for the completions so that freezing can
2374 * take place in parallel.
2376 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2377 blk_freeze_queue_start(q
);
2378 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2379 blk_mq_freeze_queue_wait(q
);
2381 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2382 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2384 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2385 blk_mq_unfreeze_queue(q
);
2387 mutex_unlock(&all_q_mutex
);
2390 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2392 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2393 blk_mq_queue_reinit_work();
2398 * Before hotadded cpu starts handling requests, new mappings must be
2399 * established. Otherwise, these requests in hw queue might never be
2402 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2403 * for CPU0, and ctx1 for CPU1).
2405 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2406 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2408 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2409 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2410 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2413 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2415 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2416 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2417 blk_mq_queue_reinit_work();
2421 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2425 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2426 if (!__blk_mq_alloc_rq_map(set
, i
))
2433 blk_mq_free_rq_map(set
->tags
[i
]);
2439 * Allocate the request maps associated with this tag_set. Note that this
2440 * may reduce the depth asked for, if memory is tight. set->queue_depth
2441 * will be updated to reflect the allocated depth.
2443 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2448 depth
= set
->queue_depth
;
2450 err
= __blk_mq_alloc_rq_maps(set
);
2454 set
->queue_depth
>>= 1;
2455 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2459 } while (set
->queue_depth
);
2461 if (!set
->queue_depth
|| err
) {
2462 pr_err("blk-mq: failed to allocate request map\n");
2466 if (depth
!= set
->queue_depth
)
2467 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2468 depth
, set
->queue_depth
);
2473 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2475 if (set
->ops
->map_queues
)
2476 return set
->ops
->map_queues(set
);
2478 return blk_mq_map_queues(set
);
2482 * Alloc a tag set to be associated with one or more request queues.
2483 * May fail with EINVAL for various error conditions. May adjust the
2484 * requested depth down, if if it too large. In that case, the set
2485 * value will be stored in set->queue_depth.
2487 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2491 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2493 if (!set
->nr_hw_queues
)
2495 if (!set
->queue_depth
)
2497 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2500 if (!set
->ops
->queue_rq
)
2503 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2504 pr_info("blk-mq: reduced tag depth to %u\n",
2506 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2510 * If a crashdump is active, then we are potentially in a very
2511 * memory constrained environment. Limit us to 1 queue and
2512 * 64 tags to prevent using too much memory.
2514 if (is_kdump_kernel()) {
2515 set
->nr_hw_queues
= 1;
2516 set
->queue_depth
= min(64U, set
->queue_depth
);
2519 * There is no use for more h/w queues than cpus.
2521 if (set
->nr_hw_queues
> nr_cpu_ids
)
2522 set
->nr_hw_queues
= nr_cpu_ids
;
2524 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2525 GFP_KERNEL
, set
->numa_node
);
2530 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2531 GFP_KERNEL
, set
->numa_node
);
2535 ret
= blk_mq_update_queue_map(set
);
2537 goto out_free_mq_map
;
2539 ret
= blk_mq_alloc_rq_maps(set
);
2541 goto out_free_mq_map
;
2543 mutex_init(&set
->tag_list_lock
);
2544 INIT_LIST_HEAD(&set
->tag_list
);
2556 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2558 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2562 for (i
= 0; i
< nr_cpu_ids
; i
++)
2563 blk_mq_free_map_and_requests(set
, i
);
2571 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2573 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2575 struct blk_mq_tag_set
*set
= q
->tag_set
;
2576 struct blk_mq_hw_ctx
*hctx
;
2582 blk_mq_freeze_queue(q
);
2585 queue_for_each_hw_ctx(q
, hctx
, i
) {
2589 * If we're using an MQ scheduler, just update the scheduler
2590 * queue depth. This is similar to what the old code would do.
2592 if (!hctx
->sched_tags
) {
2593 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2594 min(nr
, set
->queue_depth
),
2597 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2605 q
->nr_requests
= nr
;
2607 blk_mq_unfreeze_queue(q
);
2612 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
2615 struct request_queue
*q
;
2617 lockdep_assert_held(&set
->tag_list_lock
);
2619 if (nr_hw_queues
> nr_cpu_ids
)
2620 nr_hw_queues
= nr_cpu_ids
;
2621 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2624 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2625 blk_mq_freeze_queue(q
);
2627 set
->nr_hw_queues
= nr_hw_queues
;
2628 blk_mq_update_queue_map(set
);
2629 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2630 blk_mq_realloc_hw_ctxs(set
, q
);
2631 blk_mq_queue_reinit(q
, cpu_online_mask
);
2634 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2635 blk_mq_unfreeze_queue(q
);
2638 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2640 mutex_lock(&set
->tag_list_lock
);
2641 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
2642 mutex_unlock(&set
->tag_list_lock
);
2644 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2646 /* Enable polling stats and return whether they were already enabled. */
2647 static bool blk_poll_stats_enable(struct request_queue
*q
)
2649 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2650 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
2652 blk_stat_add_callback(q
, q
->poll_cb
);
2656 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2659 * We don't arm the callback if polling stats are not enabled or the
2660 * callback is already active.
2662 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2663 blk_stat_is_active(q
->poll_cb
))
2666 blk_stat_activate_msecs(q
->poll_cb
, 100);
2669 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2671 struct request_queue
*q
= cb
->data
;
2674 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
2675 if (cb
->stat
[bucket
].nr_samples
)
2676 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
2680 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2681 struct blk_mq_hw_ctx
*hctx
,
2684 unsigned long ret
= 0;
2688 * If stats collection isn't on, don't sleep but turn it on for
2691 if (!blk_poll_stats_enable(q
))
2695 * As an optimistic guess, use half of the mean service time
2696 * for this type of request. We can (and should) make this smarter.
2697 * For instance, if the completion latencies are tight, we can
2698 * get closer than just half the mean. This is especially
2699 * important on devices where the completion latencies are longer
2700 * than ~10 usec. We do use the stats for the relevant IO size
2701 * if available which does lead to better estimates.
2703 bucket
= blk_mq_poll_stats_bkt(rq
);
2707 if (q
->poll_stat
[bucket
].nr_samples
)
2708 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
2713 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2714 struct blk_mq_hw_ctx
*hctx
,
2717 struct hrtimer_sleeper hs
;
2718 enum hrtimer_mode mode
;
2722 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2728 * -1: don't ever hybrid sleep
2729 * 0: use half of prev avg
2730 * >0: use this specific value
2732 if (q
->poll_nsec
== -1)
2734 else if (q
->poll_nsec
> 0)
2735 nsecs
= q
->poll_nsec
;
2737 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2742 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2745 * This will be replaced with the stats tracking code, using
2746 * 'avg_completion_time / 2' as the pre-sleep target.
2750 mode
= HRTIMER_MODE_REL
;
2751 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2752 hrtimer_set_expires(&hs
.timer
, kt
);
2754 hrtimer_init_sleeper(&hs
, current
);
2756 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2758 set_current_state(TASK_UNINTERRUPTIBLE
);
2759 hrtimer_start_expires(&hs
.timer
, mode
);
2762 hrtimer_cancel(&hs
.timer
);
2763 mode
= HRTIMER_MODE_ABS
;
2764 } while (hs
.task
&& !signal_pending(current
));
2766 __set_current_state(TASK_RUNNING
);
2767 destroy_hrtimer_on_stack(&hs
.timer
);
2771 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2773 struct request_queue
*q
= hctx
->queue
;
2777 * If we sleep, have the caller restart the poll loop to reset
2778 * the state. Like for the other success return cases, the
2779 * caller is responsible for checking if the IO completed. If
2780 * the IO isn't complete, we'll get called again and will go
2781 * straight to the busy poll loop.
2783 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2786 hctx
->poll_considered
++;
2788 state
= current
->state
;
2789 while (!need_resched()) {
2792 hctx
->poll_invoked
++;
2794 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2796 hctx
->poll_success
++;
2797 set_current_state(TASK_RUNNING
);
2801 if (signal_pending_state(state
, current
))
2802 set_current_state(TASK_RUNNING
);
2804 if (current
->state
== TASK_RUNNING
)
2814 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2816 struct blk_mq_hw_ctx
*hctx
;
2817 struct blk_plug
*plug
;
2820 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2821 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2824 plug
= current
->plug
;
2826 blk_flush_plug_list(plug
, false);
2828 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2829 if (!blk_qc_t_is_internal(cookie
))
2830 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2832 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2834 * With scheduling, if the request has completed, we'll
2835 * get a NULL return here, as we clear the sched tag when
2836 * that happens. The request still remains valid, like always,
2837 * so we should be safe with just the NULL check.
2843 return __blk_mq_poll(hctx
, rq
);
2845 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2847 void blk_mq_disable_hotplug(void)
2849 mutex_lock(&all_q_mutex
);
2852 void blk_mq_enable_hotplug(void)
2854 mutex_unlock(&all_q_mutex
);
2857 static int __init
blk_mq_init(void)
2859 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2860 blk_mq_hctx_notify_dead
);
2862 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2863 blk_mq_queue_reinit_prepare
,
2864 blk_mq_queue_reinit_dead
);
2867 subsys_initcall(blk_mq_init
);