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 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
);
41 static void blk_mq_poll_stats_start(struct request_queue
*q
);
42 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
44 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
46 int ddir
, bytes
, bucket
;
48 ddir
= rq_data_dir(rq
);
49 bytes
= blk_rq_bytes(rq
);
51 bucket
= ddir
+ 2*(ilog2(bytes
) - 9);
55 else if (bucket
>= BLK_MQ_POLL_STATS_BKTS
)
56 return ddir
+ BLK_MQ_POLL_STATS_BKTS
- 2;
62 * Check if any of the ctx's have pending work in this hardware queue
64 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
66 return !list_empty_careful(&hctx
->dispatch
) ||
67 sbitmap_any_bit_set(&hctx
->ctx_map
) ||
68 blk_mq_sched_has_work(hctx
);
72 * Mark this ctx as having pending work in this hardware queue
74 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
75 struct blk_mq_ctx
*ctx
)
77 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
78 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
81 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
82 struct blk_mq_ctx
*ctx
)
84 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
88 struct hd_struct
*part
;
89 unsigned int *inflight
;
92 static void blk_mq_check_inflight(struct blk_mq_hw_ctx
*hctx
,
93 struct request
*rq
, void *priv
,
96 struct mq_inflight
*mi
= priv
;
98 if (test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
) &&
99 !test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
)) {
101 * index[0] counts the specific partition that was asked
102 * for. index[1] counts the ones that are active on the
103 * whole device, so increment that if mi->part is indeed
104 * a partition, and not a whole device.
106 if (rq
->part
== mi
->part
)
108 if (mi
->part
->partno
)
113 void blk_mq_in_flight(struct request_queue
*q
, struct hd_struct
*part
,
114 unsigned int inflight
[2])
116 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
118 inflight
[0] = inflight
[1] = 0;
119 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
122 static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx
*hctx
,
123 struct request
*rq
, void *priv
,
126 struct mq_inflight
*mi
= priv
;
128 if (rq
->part
== mi
->part
)
129 mi
->inflight
[rq_data_dir(rq
)]++;
132 void blk_mq_in_flight_rw(struct request_queue
*q
, struct hd_struct
*part
,
133 unsigned int inflight
[2])
135 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
137 inflight
[0] = inflight
[1] = 0;
138 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight_rw
, &mi
);
141 void blk_freeze_queue_start(struct request_queue
*q
)
145 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
146 if (freeze_depth
== 1) {
147 percpu_ref_kill(&q
->q_usage_counter
);
149 blk_mq_run_hw_queues(q
, false);
152 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
154 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
156 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
158 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
160 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
161 unsigned long timeout
)
163 return wait_event_timeout(q
->mq_freeze_wq
,
164 percpu_ref_is_zero(&q
->q_usage_counter
),
167 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
170 * Guarantee no request is in use, so we can change any data structure of
171 * the queue afterward.
173 void blk_freeze_queue(struct request_queue
*q
)
176 * In the !blk_mq case we are only calling this to kill the
177 * q_usage_counter, otherwise this increases the freeze depth
178 * and waits for it to return to zero. For this reason there is
179 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
180 * exported to drivers as the only user for unfreeze is blk_mq.
182 blk_freeze_queue_start(q
);
185 blk_mq_freeze_queue_wait(q
);
188 void blk_mq_freeze_queue(struct request_queue
*q
)
191 * ...just an alias to keep freeze and unfreeze actions balanced
192 * in the blk_mq_* namespace
196 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
198 void blk_mq_unfreeze_queue(struct request_queue
*q
)
202 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
203 WARN_ON_ONCE(freeze_depth
< 0);
205 percpu_ref_reinit(&q
->q_usage_counter
);
206 wake_up_all(&q
->mq_freeze_wq
);
209 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
212 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
213 * mpt3sas driver such that this function can be removed.
215 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
219 spin_lock_irqsave(q
->queue_lock
, flags
);
220 queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
221 spin_unlock_irqrestore(q
->queue_lock
, flags
);
223 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
226 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
229 * Note: this function does not prevent that the struct request end_io()
230 * callback function is invoked. Once this function is returned, we make
231 * sure no dispatch can happen until the queue is unquiesced via
232 * blk_mq_unquiesce_queue().
234 void blk_mq_quiesce_queue(struct request_queue
*q
)
236 struct blk_mq_hw_ctx
*hctx
;
240 blk_mq_quiesce_queue_nowait(q
);
242 queue_for_each_hw_ctx(q
, hctx
, i
) {
243 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
244 synchronize_srcu(hctx
->queue_rq_srcu
);
251 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
254 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
257 * This function recovers queue into the state before quiescing
258 * which is done by blk_mq_quiesce_queue.
260 void blk_mq_unquiesce_queue(struct request_queue
*q
)
264 spin_lock_irqsave(q
->queue_lock
, flags
);
265 queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
266 spin_unlock_irqrestore(q
->queue_lock
, flags
);
268 /* dispatch requests which are inserted during quiescing */
269 blk_mq_run_hw_queues(q
, true);
271 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
273 void blk_mq_wake_waiters(struct request_queue
*q
)
275 struct blk_mq_hw_ctx
*hctx
;
278 queue_for_each_hw_ctx(q
, hctx
, i
)
279 if (blk_mq_hw_queue_mapped(hctx
))
280 blk_mq_tag_wakeup_all(hctx
->tags
, true);
283 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
285 return blk_mq_has_free_tags(hctx
->tags
);
287 EXPORT_SYMBOL(blk_mq_can_queue
);
289 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
290 unsigned int tag
, unsigned int op
)
292 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
293 struct request
*rq
= tags
->static_rqs
[tag
];
297 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
299 rq
->internal_tag
= tag
;
301 if (blk_mq_tag_busy(data
->hctx
)) {
302 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
303 atomic_inc(&data
->hctx
->nr_active
);
306 rq
->internal_tag
= -1;
307 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
310 INIT_LIST_HEAD(&rq
->queuelist
);
311 /* csd/requeue_work/fifo_time is initialized before use */
313 rq
->mq_ctx
= data
->ctx
;
315 if (data
->flags
& BLK_MQ_REQ_PREEMPT
)
316 rq
->rq_flags
|= RQF_PREEMPT
;
317 if (blk_queue_io_stat(data
->q
))
318 rq
->rq_flags
|= RQF_IO_STAT
;
319 /* do not touch atomic flags, it needs atomic ops against the timer */
321 INIT_HLIST_NODE(&rq
->hash
);
322 RB_CLEAR_NODE(&rq
->rb_node
);
325 rq
->start_time
= jiffies
;
326 #ifdef CONFIG_BLK_CGROUP
328 set_start_time_ns(rq
);
329 rq
->io_start_time_ns
= 0;
331 rq
->nr_phys_segments
= 0;
332 #if defined(CONFIG_BLK_DEV_INTEGRITY)
333 rq
->nr_integrity_segments
= 0;
336 /* tag was already set */
339 INIT_LIST_HEAD(&rq
->timeout_list
);
343 rq
->end_io_data
= NULL
;
346 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
350 static struct request
*blk_mq_get_request(struct request_queue
*q
,
351 struct bio
*bio
, unsigned int op
,
352 struct blk_mq_alloc_data
*data
)
354 struct elevator_queue
*e
= q
->elevator
;
357 bool put_ctx_on_error
= false;
359 blk_queue_enter_live(q
);
361 if (likely(!data
->ctx
)) {
362 data
->ctx
= blk_mq_get_ctx(q
);
363 put_ctx_on_error
= true;
365 if (likely(!data
->hctx
))
366 data
->hctx
= blk_mq_map_queue(q
, data
->ctx
->cpu
);
368 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
371 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
374 * Flush requests are special and go directly to the
377 if (!op_is_flush(op
) && e
->type
->ops
.mq
.limit_depth
)
378 e
->type
->ops
.mq
.limit_depth(op
, data
);
381 tag
= blk_mq_get_tag(data
);
382 if (tag
== BLK_MQ_TAG_FAIL
) {
383 if (put_ctx_on_error
) {
384 blk_mq_put_ctx(data
->ctx
);
391 rq
= blk_mq_rq_ctx_init(data
, tag
, op
);
392 if (!op_is_flush(op
)) {
394 if (e
&& e
->type
->ops
.mq
.prepare_request
) {
395 if (e
->type
->icq_cache
&& rq_ioc(bio
))
396 blk_mq_sched_assign_ioc(rq
, bio
);
398 e
->type
->ops
.mq
.prepare_request(rq
, bio
);
399 rq
->rq_flags
|= RQF_ELVPRIV
;
402 data
->hctx
->queued
++;
406 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
407 blk_mq_req_flags_t flags
)
409 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
413 ret
= blk_queue_enter(q
, flags
);
417 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
421 return ERR_PTR(-EWOULDBLOCK
);
423 blk_mq_put_ctx(alloc_data
.ctx
);
426 rq
->__sector
= (sector_t
) -1;
427 rq
->bio
= rq
->biotail
= NULL
;
430 EXPORT_SYMBOL(blk_mq_alloc_request
);
432 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
433 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
435 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
441 * If the tag allocator sleeps we could get an allocation for a
442 * different hardware context. No need to complicate the low level
443 * allocator for this for the rare use case of a command tied to
446 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
447 return ERR_PTR(-EINVAL
);
449 if (hctx_idx
>= q
->nr_hw_queues
)
450 return ERR_PTR(-EIO
);
452 ret
= blk_queue_enter(q
, flags
);
457 * Check if the hardware context is actually mapped to anything.
458 * If not tell the caller that it should skip this queue.
460 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
461 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
463 return ERR_PTR(-EXDEV
);
465 cpu
= cpumask_first_and(alloc_data
.hctx
->cpumask
, cpu_online_mask
);
466 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
468 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
472 return ERR_PTR(-EWOULDBLOCK
);
476 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
478 void blk_mq_free_request(struct request
*rq
)
480 struct request_queue
*q
= rq
->q
;
481 struct elevator_queue
*e
= q
->elevator
;
482 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
483 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
484 const int sched_tag
= rq
->internal_tag
;
486 if (rq
->rq_flags
& RQF_ELVPRIV
) {
487 if (e
&& e
->type
->ops
.mq
.finish_request
)
488 e
->type
->ops
.mq
.finish_request(rq
);
490 put_io_context(rq
->elv
.icq
->ioc
);
495 ctx
->rq_completed
[rq_is_sync(rq
)]++;
496 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
497 atomic_dec(&hctx
->nr_active
);
499 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
500 laptop_io_completion(q
->backing_dev_info
);
502 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
505 blk_put_rl(blk_rq_rl(rq
));
507 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
508 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
510 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
512 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
513 blk_mq_sched_restart(hctx
);
516 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
518 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
520 blk_account_io_done(rq
);
523 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
524 rq
->end_io(rq
, error
);
526 if (unlikely(blk_bidi_rq(rq
)))
527 blk_mq_free_request(rq
->next_rq
);
528 blk_mq_free_request(rq
);
531 EXPORT_SYMBOL(__blk_mq_end_request
);
533 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
535 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
537 __blk_mq_end_request(rq
, error
);
539 EXPORT_SYMBOL(blk_mq_end_request
);
541 static void __blk_mq_complete_request_remote(void *data
)
543 struct request
*rq
= data
;
545 rq
->q
->softirq_done_fn(rq
);
548 static void __blk_mq_complete_request(struct request
*rq
)
550 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
554 if (rq
->internal_tag
!= -1)
555 blk_mq_sched_completed_request(rq
);
556 if (rq
->rq_flags
& RQF_STATS
) {
557 blk_mq_poll_stats_start(rq
->q
);
561 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
562 rq
->q
->softirq_done_fn(rq
);
567 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
568 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
570 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
571 rq
->csd
.func
= __blk_mq_complete_request_remote
;
574 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
576 rq
->q
->softirq_done_fn(rq
);
582 * blk_mq_complete_request - end I/O on a request
583 * @rq: the request being processed
586 * Ends all I/O on a request. It does not handle partial completions.
587 * The actual completion happens out-of-order, through a IPI handler.
589 void blk_mq_complete_request(struct request
*rq
)
591 struct request_queue
*q
= rq
->q
;
593 if (unlikely(blk_should_fake_timeout(q
)))
595 if (!blk_mark_rq_complete(rq
))
596 __blk_mq_complete_request(rq
);
598 EXPORT_SYMBOL(blk_mq_complete_request
);
600 int blk_mq_request_started(struct request
*rq
)
602 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
604 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
606 void blk_mq_start_request(struct request
*rq
)
608 struct request_queue
*q
= rq
->q
;
610 blk_mq_sched_started_request(rq
);
612 trace_block_rq_issue(q
, rq
);
614 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
615 blk_stat_set_issue(&rq
->issue_stat
, blk_rq_sectors(rq
));
616 rq
->rq_flags
|= RQF_STATS
;
617 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
622 WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
));
625 * Mark us as started and clear complete. Complete might have been
626 * set if requeue raced with timeout, which then marked it as
627 * complete. So be sure to clear complete again when we start
628 * the request, otherwise we'll ignore the completion event.
630 * Ensure that ->deadline is visible before we set STARTED, such that
631 * blk_mq_check_expired() is guaranteed to observe our ->deadline when
632 * it observes STARTED.
635 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
636 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
)) {
638 * Coherence order guarantees these consecutive stores to a
639 * single variable propagate in the specified order. Thus the
640 * clear_bit() is ordered _after_ the set bit. See
641 * blk_mq_check_expired().
643 * (the bits must be part of the same byte for this to be
646 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
649 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
651 * Make sure space for the drain appears. We know we can do
652 * this because max_hw_segments has been adjusted to be one
653 * fewer than the device can handle.
655 rq
->nr_phys_segments
++;
658 EXPORT_SYMBOL(blk_mq_start_request
);
661 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
662 * flag isn't set yet, so there may be race with timeout handler,
663 * but given rq->deadline is just set in .queue_rq() under
664 * this situation, the race won't be possible in reality because
665 * rq->timeout should be set as big enough to cover the window
666 * between blk_mq_start_request() called from .queue_rq() and
667 * clearing REQ_ATOM_STARTED here.
669 static void __blk_mq_requeue_request(struct request
*rq
)
671 struct request_queue
*q
= rq
->q
;
673 blk_mq_put_driver_tag(rq
);
675 trace_block_rq_requeue(q
, rq
);
676 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
678 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
679 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
680 rq
->nr_phys_segments
--;
684 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
686 __blk_mq_requeue_request(rq
);
688 /* this request will be re-inserted to io scheduler queue */
689 blk_mq_sched_requeue_request(rq
);
691 BUG_ON(blk_queued_rq(rq
));
692 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
694 EXPORT_SYMBOL(blk_mq_requeue_request
);
696 static void blk_mq_requeue_work(struct work_struct
*work
)
698 struct request_queue
*q
=
699 container_of(work
, struct request_queue
, requeue_work
.work
);
701 struct request
*rq
, *next
;
703 spin_lock_irq(&q
->requeue_lock
);
704 list_splice_init(&q
->requeue_list
, &rq_list
);
705 spin_unlock_irq(&q
->requeue_lock
);
707 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
708 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
711 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
712 list_del_init(&rq
->queuelist
);
713 blk_mq_sched_insert_request(rq
, true, false, false, true);
716 while (!list_empty(&rq_list
)) {
717 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
718 list_del_init(&rq
->queuelist
);
719 blk_mq_sched_insert_request(rq
, false, false, false, true);
722 blk_mq_run_hw_queues(q
, false);
725 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
726 bool kick_requeue_list
)
728 struct request_queue
*q
= rq
->q
;
732 * We abuse this flag that is otherwise used by the I/O scheduler to
733 * request head insertion from the workqueue.
735 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
737 spin_lock_irqsave(&q
->requeue_lock
, flags
);
739 rq
->rq_flags
|= RQF_SOFTBARRIER
;
740 list_add(&rq
->queuelist
, &q
->requeue_list
);
742 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
744 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
746 if (kick_requeue_list
)
747 blk_mq_kick_requeue_list(q
);
749 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
751 void blk_mq_kick_requeue_list(struct request_queue
*q
)
753 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
755 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
757 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
760 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
761 msecs_to_jiffies(msecs
));
763 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
765 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
767 if (tag
< tags
->nr_tags
) {
768 prefetch(tags
->rqs
[tag
]);
769 return tags
->rqs
[tag
];
774 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
776 struct blk_mq_timeout_data
{
778 unsigned int next_set
;
781 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
783 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
784 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
787 * We know that complete is set at this point. If STARTED isn't set
788 * anymore, then the request isn't active and the "timeout" should
789 * just be ignored. This can happen due to the bitflag ordering.
790 * Timeout first checks if STARTED is set, and if it is, assumes
791 * the request is active. But if we race with completion, then
792 * both flags will get cleared. So check here again, and ignore
793 * a timeout event with a request that isn't active.
795 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
799 ret
= ops
->timeout(req
, reserved
);
803 __blk_mq_complete_request(req
);
805 case BLK_EH_RESET_TIMER
:
807 blk_clear_rq_complete(req
);
809 case BLK_EH_NOT_HANDLED
:
812 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
817 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
818 struct request
*rq
, void *priv
, bool reserved
)
820 struct blk_mq_timeout_data
*data
= priv
;
821 unsigned long deadline
;
823 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
827 * Ensures that if we see STARTED we must also see our
828 * up-to-date deadline, see blk_mq_start_request().
832 deadline
= READ_ONCE(rq
->deadline
);
835 * The rq being checked may have been freed and reallocated
836 * out already here, we avoid this race by checking rq->deadline
837 * and REQ_ATOM_COMPLETE flag together:
839 * - if rq->deadline is observed as new value because of
840 * reusing, the rq won't be timed out because of timing.
841 * - if rq->deadline is observed as previous value,
842 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
843 * because we put a barrier between setting rq->deadline
844 * and clearing the flag in blk_mq_start_request(), so
845 * this rq won't be timed out too.
847 if (time_after_eq(jiffies
, deadline
)) {
848 if (!blk_mark_rq_complete(rq
)) {
850 * Again coherence order ensures that consecutive reads
851 * from the same variable must be in that order. This
852 * ensures that if we see COMPLETE clear, we must then
853 * see STARTED set and we'll ignore this timeout.
855 * (There's also the MB implied by the test_and_clear())
857 blk_mq_rq_timed_out(rq
, reserved
);
859 } else if (!data
->next_set
|| time_after(data
->next
, deadline
)) {
860 data
->next
= deadline
;
865 static void blk_mq_timeout_work(struct work_struct
*work
)
867 struct request_queue
*q
=
868 container_of(work
, struct request_queue
, timeout_work
);
869 struct blk_mq_timeout_data data
= {
875 /* A deadlock might occur if a request is stuck requiring a
876 * timeout at the same time a queue freeze is waiting
877 * completion, since the timeout code would not be able to
878 * acquire the queue reference here.
880 * That's why we don't use blk_queue_enter here; instead, we use
881 * percpu_ref_tryget directly, because we need to be able to
882 * obtain a reference even in the short window between the queue
883 * starting to freeze, by dropping the first reference in
884 * blk_freeze_queue_start, and the moment the last request is
885 * consumed, marked by the instant q_usage_counter reaches
888 if (!percpu_ref_tryget(&q
->q_usage_counter
))
891 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
894 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
895 mod_timer(&q
->timeout
, data
.next
);
897 struct blk_mq_hw_ctx
*hctx
;
899 queue_for_each_hw_ctx(q
, hctx
, i
) {
900 /* the hctx may be unmapped, so check it here */
901 if (blk_mq_hw_queue_mapped(hctx
))
902 blk_mq_tag_idle(hctx
);
908 struct flush_busy_ctx_data
{
909 struct blk_mq_hw_ctx
*hctx
;
910 struct list_head
*list
;
913 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
915 struct flush_busy_ctx_data
*flush_data
= data
;
916 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
917 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
919 sbitmap_clear_bit(sb
, bitnr
);
920 spin_lock(&ctx
->lock
);
921 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
922 spin_unlock(&ctx
->lock
);
927 * Process software queues that have been marked busy, splicing them
928 * to the for-dispatch
930 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
932 struct flush_busy_ctx_data data
= {
937 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
939 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
941 struct dispatch_rq_data
{
942 struct blk_mq_hw_ctx
*hctx
;
946 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
949 struct dispatch_rq_data
*dispatch_data
= data
;
950 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
951 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
953 spin_lock(&ctx
->lock
);
954 if (unlikely(!list_empty(&ctx
->rq_list
))) {
955 dispatch_data
->rq
= list_entry_rq(ctx
->rq_list
.next
);
956 list_del_init(&dispatch_data
->rq
->queuelist
);
957 if (list_empty(&ctx
->rq_list
))
958 sbitmap_clear_bit(sb
, bitnr
);
960 spin_unlock(&ctx
->lock
);
962 return !dispatch_data
->rq
;
965 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
966 struct blk_mq_ctx
*start
)
968 unsigned off
= start
? start
->index_hw
: 0;
969 struct dispatch_rq_data data
= {
974 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
975 dispatch_rq_from_ctx
, &data
);
980 static inline unsigned int queued_to_index(unsigned int queued
)
985 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
988 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
991 struct blk_mq_alloc_data data
= {
993 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
994 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
997 might_sleep_if(wait
);
1002 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
1003 data
.flags
|= BLK_MQ_REQ_RESERVED
;
1005 rq
->tag
= blk_mq_get_tag(&data
);
1007 if (blk_mq_tag_busy(data
.hctx
)) {
1008 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1009 atomic_inc(&data
.hctx
->nr_active
);
1011 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
1017 return rq
->tag
!= -1;
1020 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1021 int flags
, void *key
)
1023 struct blk_mq_hw_ctx
*hctx
;
1025 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1027 list_del_init(&wait
->entry
);
1028 blk_mq_run_hw_queue(hctx
, true);
1033 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1034 * the tag wakeups. For non-shared tags, we can simply mark us nedeing a
1035 * restart. For both caes, take care to check the condition again after
1036 * marking us as waiting.
1038 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
**hctx
,
1041 struct blk_mq_hw_ctx
*this_hctx
= *hctx
;
1042 bool shared_tags
= (this_hctx
->flags
& BLK_MQ_F_TAG_SHARED
) != 0;
1043 struct sbq_wait_state
*ws
;
1044 wait_queue_entry_t
*wait
;
1048 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &this_hctx
->state
))
1049 set_bit(BLK_MQ_S_SCHED_RESTART
, &this_hctx
->state
);
1051 wait
= &this_hctx
->dispatch_wait
;
1052 if (!list_empty_careful(&wait
->entry
))
1055 spin_lock(&this_hctx
->lock
);
1056 if (!list_empty(&wait
->entry
)) {
1057 spin_unlock(&this_hctx
->lock
);
1061 ws
= bt_wait_ptr(&this_hctx
->tags
->bitmap_tags
, this_hctx
);
1062 add_wait_queue(&ws
->wait
, wait
);
1066 * It's possible that a tag was freed in the window between the
1067 * allocation failure and adding the hardware queue to the wait
1070 ret
= blk_mq_get_driver_tag(rq
, hctx
, false);
1074 * Don't clear RESTART here, someone else could have set it.
1075 * At most this will cost an extra queue run.
1080 spin_unlock(&this_hctx
->lock
);
1085 * We got a tag, remove ourselves from the wait queue to ensure
1086 * someone else gets the wakeup.
1088 spin_lock_irq(&ws
->wait
.lock
);
1089 list_del_init(&wait
->entry
);
1090 spin_unlock_irq(&ws
->wait
.lock
);
1091 spin_unlock(&this_hctx
->lock
);
1096 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
,
1099 struct blk_mq_hw_ctx
*hctx
;
1100 struct request
*rq
, *nxt
;
1101 bool no_tag
= false;
1104 if (list_empty(list
))
1107 WARN_ON(!list_is_singular(list
) && got_budget
);
1110 * Now process all the entries, sending them to the driver.
1112 errors
= queued
= 0;
1114 struct blk_mq_queue_data bd
;
1117 rq
= list_first_entry(list
, struct request
, queuelist
);
1119 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
1120 if (!got_budget
&& !blk_mq_get_dispatch_budget(hctx
))
1123 if (!blk_mq_get_driver_tag(rq
, NULL
, false)) {
1125 * The initial allocation attempt failed, so we need to
1126 * rerun the hardware queue when a tag is freed. The
1127 * waitqueue takes care of that. If the queue is run
1128 * before we add this entry back on the dispatch list,
1129 * we'll re-run it below.
1131 if (!blk_mq_mark_tag_wait(&hctx
, rq
)) {
1132 blk_mq_put_dispatch_budget(hctx
);
1134 * For non-shared tags, the RESTART check
1137 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
)
1143 list_del_init(&rq
->queuelist
);
1148 * Flag last if we have no more requests, or if we have more
1149 * but can't assign a driver tag to it.
1151 if (list_empty(list
))
1154 nxt
= list_first_entry(list
, struct request
, queuelist
);
1155 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1158 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1159 if (ret
== BLK_STS_RESOURCE
) {
1161 * If an I/O scheduler has been configured and we got a
1162 * driver tag for the next request already, free it
1165 if (!list_empty(list
)) {
1166 nxt
= list_first_entry(list
, struct request
, queuelist
);
1167 blk_mq_put_driver_tag(nxt
);
1169 list_add(&rq
->queuelist
, list
);
1170 __blk_mq_requeue_request(rq
);
1174 if (unlikely(ret
!= BLK_STS_OK
)) {
1176 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1181 } while (!list_empty(list
));
1183 hctx
->dispatched
[queued_to_index(queued
)]++;
1186 * Any items that need requeuing? Stuff them into hctx->dispatch,
1187 * that is where we will continue on next queue run.
1189 if (!list_empty(list
)) {
1190 spin_lock(&hctx
->lock
);
1191 list_splice_init(list
, &hctx
->dispatch
);
1192 spin_unlock(&hctx
->lock
);
1195 * If SCHED_RESTART was set by the caller of this function and
1196 * it is no longer set that means that it was cleared by another
1197 * thread and hence that a queue rerun is needed.
1199 * If 'no_tag' is set, that means that we failed getting
1200 * a driver tag with an I/O scheduler attached. If our dispatch
1201 * waitqueue is no longer active, ensure that we run the queue
1202 * AFTER adding our entries back to the list.
1204 * If no I/O scheduler has been configured it is possible that
1205 * the hardware queue got stopped and restarted before requests
1206 * were pushed back onto the dispatch list. Rerun the queue to
1207 * avoid starvation. Notes:
1208 * - blk_mq_run_hw_queue() checks whether or not a queue has
1209 * been stopped before rerunning a queue.
1210 * - Some but not all block drivers stop a queue before
1211 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1214 if (!blk_mq_sched_needs_restart(hctx
) ||
1215 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1216 blk_mq_run_hw_queue(hctx
, true);
1219 return (queued
+ errors
) != 0;
1222 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1227 * We should be running this queue from one of the CPUs that
1230 * There are at least two related races now between setting
1231 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1232 * __blk_mq_run_hw_queue():
1234 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1235 * but later it becomes online, then this warning is harmless
1238 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1239 * but later it becomes offline, then the warning can't be
1240 * triggered, and we depend on blk-mq timeout handler to
1241 * handle dispatched requests to this hctx
1243 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1244 cpu_online(hctx
->next_cpu
)) {
1245 printk(KERN_WARNING
"run queue from wrong CPU %d, hctx %s\n",
1246 raw_smp_processor_id(),
1247 cpumask_empty(hctx
->cpumask
) ? "inactive": "active");
1252 * We can't run the queue inline with ints disabled. Ensure that
1253 * we catch bad users of this early.
1255 WARN_ON_ONCE(in_interrupt());
1257 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1259 blk_mq_sched_dispatch_requests(hctx
);
1264 srcu_idx
= srcu_read_lock(hctx
->queue_rq_srcu
);
1265 blk_mq_sched_dispatch_requests(hctx
);
1266 srcu_read_unlock(hctx
->queue_rq_srcu
, srcu_idx
);
1270 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1272 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1274 if (cpu
>= nr_cpu_ids
)
1275 cpu
= cpumask_first(hctx
->cpumask
);
1280 * It'd be great if the workqueue API had a way to pass
1281 * in a mask and had some smarts for more clever placement.
1282 * For now we just round-robin here, switching for every
1283 * BLK_MQ_CPU_WORK_BATCH queued items.
1285 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1288 int next_cpu
= hctx
->next_cpu
;
1290 if (hctx
->queue
->nr_hw_queues
== 1)
1291 return WORK_CPU_UNBOUND
;
1293 if (--hctx
->next_cpu_batch
<= 0) {
1295 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1297 if (next_cpu
>= nr_cpu_ids
)
1298 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1299 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1303 * Do unbound schedule if we can't find a online CPU for this hctx,
1304 * and it should only happen in the path of handling CPU DEAD.
1306 if (!cpu_online(next_cpu
)) {
1313 * Make sure to re-select CPU next time once after CPUs
1314 * in hctx->cpumask become online again.
1316 hctx
->next_cpu
= next_cpu
;
1317 hctx
->next_cpu_batch
= 1;
1318 return WORK_CPU_UNBOUND
;
1321 hctx
->next_cpu
= next_cpu
;
1325 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1326 unsigned long msecs
)
1328 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1331 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1332 int cpu
= get_cpu();
1333 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1334 __blk_mq_run_hw_queue(hctx
);
1342 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1343 msecs_to_jiffies(msecs
));
1346 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1348 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1350 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1352 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1354 if (blk_mq_hctx_has_pending(hctx
)) {
1355 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1361 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1363 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1365 struct blk_mq_hw_ctx
*hctx
;
1368 queue_for_each_hw_ctx(q
, hctx
, i
) {
1369 if (blk_mq_hctx_stopped(hctx
))
1372 blk_mq_run_hw_queue(hctx
, async
);
1375 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1378 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1379 * @q: request queue.
1381 * The caller is responsible for serializing this function against
1382 * blk_mq_{start,stop}_hw_queue().
1384 bool blk_mq_queue_stopped(struct request_queue
*q
)
1386 struct blk_mq_hw_ctx
*hctx
;
1389 queue_for_each_hw_ctx(q
, hctx
, i
)
1390 if (blk_mq_hctx_stopped(hctx
))
1395 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1398 * This function is often used for pausing .queue_rq() by driver when
1399 * there isn't enough resource or some conditions aren't satisfied, and
1400 * BLK_STS_RESOURCE is usually returned.
1402 * We do not guarantee that dispatch can be drained or blocked
1403 * after blk_mq_stop_hw_queue() returns. Please use
1404 * blk_mq_quiesce_queue() for that requirement.
1406 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1408 cancel_delayed_work(&hctx
->run_work
);
1410 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1412 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1415 * This function is often used for pausing .queue_rq() by driver when
1416 * there isn't enough resource or some conditions aren't satisfied, and
1417 * BLK_STS_RESOURCE is usually returned.
1419 * We do not guarantee that dispatch can be drained or blocked
1420 * after blk_mq_stop_hw_queues() returns. Please use
1421 * blk_mq_quiesce_queue() for that requirement.
1423 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1425 struct blk_mq_hw_ctx
*hctx
;
1428 queue_for_each_hw_ctx(q
, hctx
, i
)
1429 blk_mq_stop_hw_queue(hctx
);
1431 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1433 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1435 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1437 blk_mq_run_hw_queue(hctx
, false);
1439 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1441 void blk_mq_start_hw_queues(struct request_queue
*q
)
1443 struct blk_mq_hw_ctx
*hctx
;
1446 queue_for_each_hw_ctx(q
, hctx
, i
)
1447 blk_mq_start_hw_queue(hctx
);
1449 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1451 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1453 if (!blk_mq_hctx_stopped(hctx
))
1456 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1457 blk_mq_run_hw_queue(hctx
, async
);
1459 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1461 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1463 struct blk_mq_hw_ctx
*hctx
;
1466 queue_for_each_hw_ctx(q
, hctx
, i
)
1467 blk_mq_start_stopped_hw_queue(hctx
, async
);
1469 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1471 static void blk_mq_run_work_fn(struct work_struct
*work
)
1473 struct blk_mq_hw_ctx
*hctx
;
1475 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1478 * If we are stopped, don't run the queue. The exception is if
1479 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1480 * the STOPPED bit and run it.
1482 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)) {
1483 if (!test_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
))
1486 clear_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1487 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1490 __blk_mq_run_hw_queue(hctx
);
1494 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1496 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx
)))
1500 * Stop the hw queue, then modify currently delayed work.
1501 * This should prevent us from running the queue prematurely.
1502 * Mark the queue as auto-clearing STOPPED when it runs.
1504 blk_mq_stop_hw_queue(hctx
);
1505 set_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1506 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1508 msecs_to_jiffies(msecs
));
1510 EXPORT_SYMBOL(blk_mq_delay_queue
);
1512 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1516 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1518 lockdep_assert_held(&ctx
->lock
);
1520 trace_block_rq_insert(hctx
->queue
, rq
);
1523 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1525 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1528 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1531 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1533 lockdep_assert_held(&ctx
->lock
);
1535 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1536 blk_mq_hctx_mark_pending(hctx
, ctx
);
1540 * Should only be used carefully, when the caller knows we want to
1541 * bypass a potential IO scheduler on the target device.
1543 void blk_mq_request_bypass_insert(struct request
*rq
, bool run_queue
)
1545 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1546 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(rq
->q
, ctx
->cpu
);
1548 spin_lock(&hctx
->lock
);
1549 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1550 spin_unlock(&hctx
->lock
);
1553 blk_mq_run_hw_queue(hctx
, false);
1556 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1557 struct list_head
*list
)
1561 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1564 spin_lock(&ctx
->lock
);
1565 while (!list_empty(list
)) {
1568 rq
= list_first_entry(list
, struct request
, queuelist
);
1569 BUG_ON(rq
->mq_ctx
!= ctx
);
1570 list_del_init(&rq
->queuelist
);
1571 __blk_mq_insert_req_list(hctx
, rq
, false);
1573 blk_mq_hctx_mark_pending(hctx
, ctx
);
1574 spin_unlock(&ctx
->lock
);
1577 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1579 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1580 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1582 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1583 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1584 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1587 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1589 struct blk_mq_ctx
*this_ctx
;
1590 struct request_queue
*this_q
;
1593 LIST_HEAD(ctx_list
);
1596 list_splice_init(&plug
->mq_list
, &list
);
1598 list_sort(NULL
, &list
, plug_ctx_cmp
);
1604 while (!list_empty(&list
)) {
1605 rq
= list_entry_rq(list
.next
);
1606 list_del_init(&rq
->queuelist
);
1608 if (rq
->mq_ctx
!= this_ctx
) {
1610 trace_block_unplug(this_q
, depth
, from_schedule
);
1611 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1616 this_ctx
= rq
->mq_ctx
;
1622 list_add_tail(&rq
->queuelist
, &ctx_list
);
1626 * If 'this_ctx' is set, we know we have entries to complete
1627 * on 'ctx_list'. Do those.
1630 trace_block_unplug(this_q
, depth
, from_schedule
);
1631 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1636 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1638 blk_init_request_from_bio(rq
, bio
);
1640 blk_rq_set_rl(rq
, blk_get_rl(rq
->q
, bio
));
1642 blk_account_io_start(rq
, true);
1645 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx
*hctx
,
1646 struct blk_mq_ctx
*ctx
,
1649 spin_lock(&ctx
->lock
);
1650 __blk_mq_insert_request(hctx
, rq
, false);
1651 spin_unlock(&ctx
->lock
);
1654 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1657 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1659 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1662 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1664 blk_qc_t
*cookie
, bool may_sleep
)
1666 struct request_queue
*q
= rq
->q
;
1667 struct blk_mq_queue_data bd
= {
1671 blk_qc_t new_cookie
;
1673 bool run_queue
= true;
1675 /* RCU or SRCU read lock is needed before checking quiesced flag */
1676 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1684 if (!blk_mq_get_dispatch_budget(hctx
))
1687 if (!blk_mq_get_driver_tag(rq
, NULL
, false)) {
1688 blk_mq_put_dispatch_budget(hctx
);
1692 new_cookie
= request_to_qc_t(hctx
, rq
);
1695 * For OK queue, we are done. For error, kill it. Any other
1696 * error (busy), just add it to our list as we previously
1699 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1702 *cookie
= new_cookie
;
1704 case BLK_STS_RESOURCE
:
1705 __blk_mq_requeue_request(rq
);
1708 *cookie
= BLK_QC_T_NONE
;
1709 blk_mq_end_request(rq
, ret
);
1714 blk_mq_sched_insert_request(rq
, false, run_queue
, false, may_sleep
);
1717 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1718 struct request
*rq
, blk_qc_t
*cookie
)
1720 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1722 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false);
1725 unsigned int srcu_idx
;
1729 srcu_idx
= srcu_read_lock(hctx
->queue_rq_srcu
);
1730 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, true);
1731 srcu_read_unlock(hctx
->queue_rq_srcu
, srcu_idx
);
1735 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1737 const int is_sync
= op_is_sync(bio
->bi_opf
);
1738 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1739 struct blk_mq_alloc_data data
= { .flags
= 0 };
1741 unsigned int request_count
= 0;
1742 struct blk_plug
*plug
;
1743 struct request
*same_queue_rq
= NULL
;
1745 unsigned int wb_acct
;
1747 blk_queue_bounce(q
, &bio
);
1749 blk_queue_split(q
, &bio
);
1751 if (!bio_integrity_prep(bio
))
1752 return BLK_QC_T_NONE
;
1754 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1755 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1756 return BLK_QC_T_NONE
;
1758 if (blk_mq_sched_bio_merge(q
, bio
))
1759 return BLK_QC_T_NONE
;
1761 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1763 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1765 rq
= blk_mq_get_request(q
, bio
, bio
->bi_opf
, &data
);
1766 if (unlikely(!rq
)) {
1767 __wbt_done(q
->rq_wb
, wb_acct
);
1768 if (bio
->bi_opf
& REQ_NOWAIT
)
1769 bio_wouldblock_error(bio
);
1770 return BLK_QC_T_NONE
;
1773 wbt_track(&rq
->issue_stat
, wb_acct
);
1775 cookie
= request_to_qc_t(data
.hctx
, rq
);
1777 plug
= current
->plug
;
1778 if (unlikely(is_flush_fua
)) {
1779 blk_mq_put_ctx(data
.ctx
);
1780 blk_mq_bio_to_request(rq
, bio
);
1782 /* bypass scheduler for flush rq */
1783 blk_insert_flush(rq
);
1784 blk_mq_run_hw_queue(data
.hctx
, true);
1785 } else if (plug
&& q
->nr_hw_queues
== 1) {
1786 struct request
*last
= NULL
;
1788 blk_mq_put_ctx(data
.ctx
);
1789 blk_mq_bio_to_request(rq
, bio
);
1792 * @request_count may become stale because of schedule
1793 * out, so check the list again.
1795 if (list_empty(&plug
->mq_list
))
1797 else if (blk_queue_nomerges(q
))
1798 request_count
= blk_plug_queued_count(q
);
1801 trace_block_plug(q
);
1803 last
= list_entry_rq(plug
->mq_list
.prev
);
1805 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1806 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1807 blk_flush_plug_list(plug
, false);
1808 trace_block_plug(q
);
1811 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1812 } else if (plug
&& !blk_queue_nomerges(q
)) {
1813 blk_mq_bio_to_request(rq
, bio
);
1816 * We do limited plugging. If the bio can be merged, do that.
1817 * Otherwise the existing request in the plug list will be
1818 * issued. So the plug list will have one request at most
1819 * The plug list might get flushed before this. If that happens,
1820 * the plug list is empty, and same_queue_rq is invalid.
1822 if (list_empty(&plug
->mq_list
))
1823 same_queue_rq
= NULL
;
1825 list_del_init(&same_queue_rq
->queuelist
);
1826 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1828 blk_mq_put_ctx(data
.ctx
);
1830 if (same_queue_rq
) {
1831 data
.hctx
= blk_mq_map_queue(q
,
1832 same_queue_rq
->mq_ctx
->cpu
);
1833 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1836 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1837 blk_mq_put_ctx(data
.ctx
);
1838 blk_mq_bio_to_request(rq
, bio
);
1839 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1840 } else if (q
->elevator
) {
1841 blk_mq_put_ctx(data
.ctx
);
1842 blk_mq_bio_to_request(rq
, bio
);
1843 blk_mq_sched_insert_request(rq
, false, true, true, true);
1845 blk_mq_put_ctx(data
.ctx
);
1846 blk_mq_bio_to_request(rq
, bio
);
1847 blk_mq_queue_io(data
.hctx
, data
.ctx
, rq
);
1848 blk_mq_run_hw_queue(data
.hctx
, true);
1854 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1855 unsigned int hctx_idx
)
1859 if (tags
->rqs
&& set
->ops
->exit_request
) {
1862 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1863 struct request
*rq
= tags
->static_rqs
[i
];
1867 set
->ops
->exit_request(set
, rq
, hctx_idx
);
1868 tags
->static_rqs
[i
] = NULL
;
1872 while (!list_empty(&tags
->page_list
)) {
1873 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1874 list_del_init(&page
->lru
);
1876 * Remove kmemleak object previously allocated in
1877 * blk_mq_init_rq_map().
1879 kmemleak_free(page_address(page
));
1880 __free_pages(page
, page
->private);
1884 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1888 kfree(tags
->static_rqs
);
1889 tags
->static_rqs
= NULL
;
1891 blk_mq_free_tags(tags
);
1894 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1895 unsigned int hctx_idx
,
1896 unsigned int nr_tags
,
1897 unsigned int reserved_tags
)
1899 struct blk_mq_tags
*tags
;
1902 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1903 if (node
== NUMA_NO_NODE
)
1904 node
= set
->numa_node
;
1906 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1907 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1911 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1912 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1915 blk_mq_free_tags(tags
);
1919 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1920 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1922 if (!tags
->static_rqs
) {
1924 blk_mq_free_tags(tags
);
1931 static size_t order_to_size(unsigned int order
)
1933 return (size_t)PAGE_SIZE
<< order
;
1936 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1937 unsigned int hctx_idx
, unsigned int depth
)
1939 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1940 size_t rq_size
, left
;
1943 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1944 if (node
== NUMA_NO_NODE
)
1945 node
= set
->numa_node
;
1947 INIT_LIST_HEAD(&tags
->page_list
);
1950 * rq_size is the size of the request plus driver payload, rounded
1951 * to the cacheline size
1953 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1955 left
= rq_size
* depth
;
1957 for (i
= 0; i
< depth
; ) {
1958 int this_order
= max_order
;
1963 while (this_order
&& left
< order_to_size(this_order
- 1))
1967 page
= alloc_pages_node(node
,
1968 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1974 if (order_to_size(this_order
) < rq_size
)
1981 page
->private = this_order
;
1982 list_add_tail(&page
->lru
, &tags
->page_list
);
1984 p
= page_address(page
);
1986 * Allow kmemleak to scan these pages as they contain pointers
1987 * to additional allocations like via ops->init_request().
1989 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1990 entries_per_page
= order_to_size(this_order
) / rq_size
;
1991 to_do
= min(entries_per_page
, depth
- i
);
1992 left
-= to_do
* rq_size
;
1993 for (j
= 0; j
< to_do
; j
++) {
1994 struct request
*rq
= p
;
1996 tags
->static_rqs
[i
] = rq
;
1997 if (set
->ops
->init_request
) {
1998 if (set
->ops
->init_request(set
, rq
, hctx_idx
,
2000 tags
->static_rqs
[i
] = NULL
;
2012 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2017 * 'cpu' is going away. splice any existing rq_list entries from this
2018 * software queue to the hw queue dispatch list, and ensure that it
2021 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2023 struct blk_mq_hw_ctx
*hctx
;
2024 struct blk_mq_ctx
*ctx
;
2027 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2028 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2030 spin_lock(&ctx
->lock
);
2031 if (!list_empty(&ctx
->rq_list
)) {
2032 list_splice_init(&ctx
->rq_list
, &tmp
);
2033 blk_mq_hctx_clear_pending(hctx
, ctx
);
2035 spin_unlock(&ctx
->lock
);
2037 if (list_empty(&tmp
))
2040 spin_lock(&hctx
->lock
);
2041 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2042 spin_unlock(&hctx
->lock
);
2044 blk_mq_run_hw_queue(hctx
, true);
2048 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2050 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2054 /* hctx->ctxs will be freed in queue's release handler */
2055 static void blk_mq_exit_hctx(struct request_queue
*q
,
2056 struct blk_mq_tag_set
*set
,
2057 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2059 blk_mq_debugfs_unregister_hctx(hctx
);
2061 if (blk_mq_hw_queue_mapped(hctx
))
2062 blk_mq_tag_idle(hctx
);
2064 if (set
->ops
->exit_request
)
2065 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2067 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2069 if (set
->ops
->exit_hctx
)
2070 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2072 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2073 cleanup_srcu_struct(hctx
->queue_rq_srcu
);
2075 blk_mq_remove_cpuhp(hctx
);
2076 blk_free_flush_queue(hctx
->fq
);
2077 sbitmap_free(&hctx
->ctx_map
);
2080 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2081 struct blk_mq_tag_set
*set
, int nr_queue
)
2083 struct blk_mq_hw_ctx
*hctx
;
2086 queue_for_each_hw_ctx(q
, hctx
, i
) {
2089 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2093 static int blk_mq_init_hctx(struct request_queue
*q
,
2094 struct blk_mq_tag_set
*set
,
2095 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2099 node
= hctx
->numa_node
;
2100 if (node
== NUMA_NO_NODE
)
2101 node
= hctx
->numa_node
= set
->numa_node
;
2103 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2104 spin_lock_init(&hctx
->lock
);
2105 INIT_LIST_HEAD(&hctx
->dispatch
);
2107 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
2109 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2111 hctx
->tags
= set
->tags
[hctx_idx
];
2114 * Allocate space for all possible cpus to avoid allocation at
2117 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2120 goto unregister_cpu_notifier
;
2122 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
2128 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2129 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2131 if (set
->ops
->init_hctx
&&
2132 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2135 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
2138 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
2140 goto sched_exit_hctx
;
2142 if (set
->ops
->init_request
&&
2143 set
->ops
->init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
2147 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2148 init_srcu_struct(hctx
->queue_rq_srcu
);
2150 blk_mq_debugfs_register_hctx(q
, hctx
);
2157 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2159 if (set
->ops
->exit_hctx
)
2160 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2162 sbitmap_free(&hctx
->ctx_map
);
2165 unregister_cpu_notifier
:
2166 blk_mq_remove_cpuhp(hctx
);
2170 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2171 unsigned int nr_hw_queues
)
2175 for_each_possible_cpu(i
) {
2176 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2177 struct blk_mq_hw_ctx
*hctx
;
2180 spin_lock_init(&__ctx
->lock
);
2181 INIT_LIST_HEAD(&__ctx
->rq_list
);
2185 * Set local node, IFF we have more than one hw queue. If
2186 * not, we remain on the home node of the device
2188 hctx
= blk_mq_map_queue(q
, i
);
2189 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2190 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2194 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2198 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2199 set
->queue_depth
, set
->reserved_tags
);
2200 if (!set
->tags
[hctx_idx
])
2203 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2208 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2209 set
->tags
[hctx_idx
] = NULL
;
2213 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2214 unsigned int hctx_idx
)
2216 if (set
->tags
[hctx_idx
]) {
2217 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2218 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2219 set
->tags
[hctx_idx
] = NULL
;
2223 static void blk_mq_map_swqueue(struct request_queue
*q
)
2225 unsigned int i
, hctx_idx
;
2226 struct blk_mq_hw_ctx
*hctx
;
2227 struct blk_mq_ctx
*ctx
;
2228 struct blk_mq_tag_set
*set
= q
->tag_set
;
2231 * Avoid others reading imcomplete hctx->cpumask through sysfs
2233 mutex_lock(&q
->sysfs_lock
);
2235 queue_for_each_hw_ctx(q
, hctx
, i
) {
2236 cpumask_clear(hctx
->cpumask
);
2241 * Map software to hardware queues.
2243 * If the cpu isn't present, the cpu is mapped to first hctx.
2245 for_each_possible_cpu(i
) {
2246 hctx_idx
= q
->mq_map
[i
];
2247 /* unmapped hw queue can be remapped after CPU topo changed */
2248 if (!set
->tags
[hctx_idx
] &&
2249 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2251 * If tags initialization fail for some hctx,
2252 * that hctx won't be brought online. In this
2253 * case, remap the current ctx to hctx[0] which
2254 * is guaranteed to always have tags allocated
2259 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2260 hctx
= blk_mq_map_queue(q
, i
);
2262 cpumask_set_cpu(i
, hctx
->cpumask
);
2263 ctx
->index_hw
= hctx
->nr_ctx
;
2264 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2267 mutex_unlock(&q
->sysfs_lock
);
2269 queue_for_each_hw_ctx(q
, hctx
, i
) {
2271 * If no software queues are mapped to this hardware queue,
2272 * disable it and free the request entries.
2274 if (!hctx
->nr_ctx
) {
2275 /* Never unmap queue 0. We need it as a
2276 * fallback in case of a new remap fails
2279 if (i
&& set
->tags
[i
])
2280 blk_mq_free_map_and_requests(set
, i
);
2286 hctx
->tags
= set
->tags
[i
];
2287 WARN_ON(!hctx
->tags
);
2290 * Set the map size to the number of mapped software queues.
2291 * This is more accurate and more efficient than looping
2292 * over all possibly mapped software queues.
2294 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2297 * Initialize batch roundrobin counts
2299 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2300 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2305 * Caller needs to ensure that we're either frozen/quiesced, or that
2306 * the queue isn't live yet.
2308 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2310 struct blk_mq_hw_ctx
*hctx
;
2313 queue_for_each_hw_ctx(q
, hctx
, i
) {
2315 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2316 atomic_inc(&q
->shared_hctx_restart
);
2317 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2319 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2320 atomic_dec(&q
->shared_hctx_restart
);
2321 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2326 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2329 struct request_queue
*q
;
2331 lockdep_assert_held(&set
->tag_list_lock
);
2333 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2334 blk_mq_freeze_queue(q
);
2335 queue_set_hctx_shared(q
, shared
);
2336 blk_mq_unfreeze_queue(q
);
2340 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2342 struct blk_mq_tag_set
*set
= q
->tag_set
;
2344 mutex_lock(&set
->tag_list_lock
);
2345 list_del_rcu(&q
->tag_set_list
);
2346 INIT_LIST_HEAD(&q
->tag_set_list
);
2347 if (list_is_singular(&set
->tag_list
)) {
2348 /* just transitioned to unshared */
2349 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2350 /* update existing queue */
2351 blk_mq_update_tag_set_depth(set
, false);
2353 mutex_unlock(&set
->tag_list_lock
);
2358 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2359 struct request_queue
*q
)
2363 mutex_lock(&set
->tag_list_lock
);
2366 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2368 if (!list_empty(&set
->tag_list
) &&
2369 !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2370 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2371 /* update existing queue */
2372 blk_mq_update_tag_set_depth(set
, true);
2374 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2375 queue_set_hctx_shared(q
, true);
2376 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2378 mutex_unlock(&set
->tag_list_lock
);
2382 * It is the actual release handler for mq, but we do it from
2383 * request queue's release handler for avoiding use-after-free
2384 * and headache because q->mq_kobj shouldn't have been introduced,
2385 * but we can't group ctx/kctx kobj without it.
2387 void blk_mq_release(struct request_queue
*q
)
2389 struct blk_mq_hw_ctx
*hctx
;
2392 /* hctx kobj stays in hctx */
2393 queue_for_each_hw_ctx(q
, hctx
, i
) {
2396 kobject_put(&hctx
->kobj
);
2401 kfree(q
->queue_hw_ctx
);
2404 * release .mq_kobj and sw queue's kobject now because
2405 * both share lifetime with request queue.
2407 blk_mq_sysfs_deinit(q
);
2409 free_percpu(q
->queue_ctx
);
2412 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2414 struct request_queue
*uninit_q
, *q
;
2416 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2418 return ERR_PTR(-ENOMEM
);
2420 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2422 blk_cleanup_queue(uninit_q
);
2426 EXPORT_SYMBOL(blk_mq_init_queue
);
2428 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2430 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2432 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, queue_rq_srcu
),
2433 __alignof__(struct blk_mq_hw_ctx
)) !=
2434 sizeof(struct blk_mq_hw_ctx
));
2436 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2437 hw_ctx_size
+= sizeof(struct srcu_struct
);
2442 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2443 struct request_queue
*q
)
2446 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2448 blk_mq_sysfs_unregister(q
);
2450 /* protect against switching io scheduler */
2451 mutex_lock(&q
->sysfs_lock
);
2452 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2458 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2459 hctxs
[i
] = kzalloc_node(blk_mq_hw_ctx_size(set
),
2464 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2471 atomic_set(&hctxs
[i
]->nr_active
, 0);
2472 hctxs
[i
]->numa_node
= node
;
2473 hctxs
[i
]->queue_num
= i
;
2475 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2476 free_cpumask_var(hctxs
[i
]->cpumask
);
2481 blk_mq_hctx_kobj_init(hctxs
[i
]);
2483 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2484 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2488 blk_mq_free_map_and_requests(set
, j
);
2489 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2490 kobject_put(&hctx
->kobj
);
2495 q
->nr_hw_queues
= i
;
2496 mutex_unlock(&q
->sysfs_lock
);
2497 blk_mq_sysfs_register(q
);
2500 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2501 struct request_queue
*q
)
2503 /* mark the queue as mq asap */
2504 q
->mq_ops
= set
->ops
;
2506 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2507 blk_mq_poll_stats_bkt
,
2508 BLK_MQ_POLL_STATS_BKTS
, q
);
2512 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2516 /* init q->mq_kobj and sw queues' kobjects */
2517 blk_mq_sysfs_init(q
);
2519 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2520 GFP_KERNEL
, set
->numa_node
);
2521 if (!q
->queue_hw_ctx
)
2524 q
->mq_map
= set
->mq_map
;
2526 blk_mq_realloc_hw_ctxs(set
, q
);
2527 if (!q
->nr_hw_queues
)
2530 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2531 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2533 q
->nr_queues
= nr_cpu_ids
;
2535 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2537 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2538 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2540 q
->sg_reserved_size
= INT_MAX
;
2542 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2543 INIT_LIST_HEAD(&q
->requeue_list
);
2544 spin_lock_init(&q
->requeue_lock
);
2546 blk_queue_make_request(q
, blk_mq_make_request
);
2547 if (q
->mq_ops
->poll
)
2548 q
->poll_fn
= blk_mq_poll
;
2551 * Do this after blk_queue_make_request() overrides it...
2553 q
->nr_requests
= set
->queue_depth
;
2556 * Default to classic polling
2560 if (set
->ops
->complete
)
2561 blk_queue_softirq_done(q
, set
->ops
->complete
);
2563 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2564 blk_mq_add_queue_tag_set(set
, q
);
2565 blk_mq_map_swqueue(q
);
2567 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2570 ret
= blk_mq_sched_init(q
);
2572 return ERR_PTR(ret
);
2578 kfree(q
->queue_hw_ctx
);
2580 free_percpu(q
->queue_ctx
);
2583 return ERR_PTR(-ENOMEM
);
2585 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2587 void blk_mq_free_queue(struct request_queue
*q
)
2589 struct blk_mq_tag_set
*set
= q
->tag_set
;
2591 blk_mq_del_queue_tag_set(q
);
2592 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2595 /* Basically redo blk_mq_init_queue with queue frozen */
2596 static void blk_mq_queue_reinit(struct request_queue
*q
)
2598 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2600 blk_mq_debugfs_unregister_hctxs(q
);
2601 blk_mq_sysfs_unregister(q
);
2604 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2605 * we should change hctx numa_node according to the new topology (this
2606 * involves freeing and re-allocating memory, worth doing?)
2608 blk_mq_map_swqueue(q
);
2610 blk_mq_sysfs_register(q
);
2611 blk_mq_debugfs_register_hctxs(q
);
2614 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2618 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2619 if (!__blk_mq_alloc_rq_map(set
, i
))
2626 blk_mq_free_rq_map(set
->tags
[i
]);
2632 * Allocate the request maps associated with this tag_set. Note that this
2633 * may reduce the depth asked for, if memory is tight. set->queue_depth
2634 * will be updated to reflect the allocated depth.
2636 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2641 depth
= set
->queue_depth
;
2643 err
= __blk_mq_alloc_rq_maps(set
);
2647 set
->queue_depth
>>= 1;
2648 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2652 } while (set
->queue_depth
);
2654 if (!set
->queue_depth
|| err
) {
2655 pr_err("blk-mq: failed to allocate request map\n");
2659 if (depth
!= set
->queue_depth
)
2660 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2661 depth
, set
->queue_depth
);
2666 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2668 if (set
->ops
->map_queues
) {
2671 * transport .map_queues is usually done in the following
2674 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2675 * mask = get_cpu_mask(queue)
2676 * for_each_cpu(cpu, mask)
2677 * set->mq_map[cpu] = queue;
2680 * When we need to remap, the table has to be cleared for
2681 * killing stale mapping since one CPU may not be mapped
2684 for_each_possible_cpu(cpu
)
2685 set
->mq_map
[cpu
] = 0;
2687 return set
->ops
->map_queues(set
);
2689 return blk_mq_map_queues(set
);
2693 * Alloc a tag set to be associated with one or more request queues.
2694 * May fail with EINVAL for various error conditions. May adjust the
2695 * requested depth down, if if it too large. In that case, the set
2696 * value will be stored in set->queue_depth.
2698 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2702 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2704 if (!set
->nr_hw_queues
)
2706 if (!set
->queue_depth
)
2708 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2711 if (!set
->ops
->queue_rq
)
2714 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
2717 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2718 pr_info("blk-mq: reduced tag depth to %u\n",
2720 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2724 * If a crashdump is active, then we are potentially in a very
2725 * memory constrained environment. Limit us to 1 queue and
2726 * 64 tags to prevent using too much memory.
2728 if (is_kdump_kernel()) {
2729 set
->nr_hw_queues
= 1;
2730 set
->queue_depth
= min(64U, set
->queue_depth
);
2733 * There is no use for more h/w queues than cpus.
2735 if (set
->nr_hw_queues
> nr_cpu_ids
)
2736 set
->nr_hw_queues
= nr_cpu_ids
;
2738 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2739 GFP_KERNEL
, set
->numa_node
);
2744 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2745 GFP_KERNEL
, set
->numa_node
);
2749 ret
= blk_mq_update_queue_map(set
);
2751 goto out_free_mq_map
;
2753 ret
= blk_mq_alloc_rq_maps(set
);
2755 goto out_free_mq_map
;
2757 mutex_init(&set
->tag_list_lock
);
2758 INIT_LIST_HEAD(&set
->tag_list
);
2770 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2772 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2776 for (i
= 0; i
< nr_cpu_ids
; i
++)
2777 blk_mq_free_map_and_requests(set
, i
);
2785 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2787 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2789 struct blk_mq_tag_set
*set
= q
->tag_set
;
2790 struct blk_mq_hw_ctx
*hctx
;
2796 blk_mq_freeze_queue(q
);
2799 queue_for_each_hw_ctx(q
, hctx
, i
) {
2803 * If we're using an MQ scheduler, just update the scheduler
2804 * queue depth. This is similar to what the old code would do.
2806 if (!hctx
->sched_tags
) {
2807 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
2810 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2818 q
->nr_requests
= nr
;
2820 blk_mq_unfreeze_queue(q
);
2825 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
2828 struct request_queue
*q
;
2830 lockdep_assert_held(&set
->tag_list_lock
);
2832 if (nr_hw_queues
> nr_cpu_ids
)
2833 nr_hw_queues
= nr_cpu_ids
;
2834 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2837 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2838 blk_mq_freeze_queue(q
);
2840 set
->nr_hw_queues
= nr_hw_queues
;
2841 blk_mq_update_queue_map(set
);
2842 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2843 blk_mq_realloc_hw_ctxs(set
, q
);
2844 blk_mq_queue_reinit(q
);
2847 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2848 blk_mq_unfreeze_queue(q
);
2851 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2853 mutex_lock(&set
->tag_list_lock
);
2854 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
2855 mutex_unlock(&set
->tag_list_lock
);
2857 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2859 /* Enable polling stats and return whether they were already enabled. */
2860 static bool blk_poll_stats_enable(struct request_queue
*q
)
2862 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2863 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
2865 blk_stat_add_callback(q
, q
->poll_cb
);
2869 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2872 * We don't arm the callback if polling stats are not enabled or the
2873 * callback is already active.
2875 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2876 blk_stat_is_active(q
->poll_cb
))
2879 blk_stat_activate_msecs(q
->poll_cb
, 100);
2882 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2884 struct request_queue
*q
= cb
->data
;
2887 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
2888 if (cb
->stat
[bucket
].nr_samples
)
2889 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
2893 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2894 struct blk_mq_hw_ctx
*hctx
,
2897 unsigned long ret
= 0;
2901 * If stats collection isn't on, don't sleep but turn it on for
2904 if (!blk_poll_stats_enable(q
))
2908 * As an optimistic guess, use half of the mean service time
2909 * for this type of request. We can (and should) make this smarter.
2910 * For instance, if the completion latencies are tight, we can
2911 * get closer than just half the mean. This is especially
2912 * important on devices where the completion latencies are longer
2913 * than ~10 usec. We do use the stats for the relevant IO size
2914 * if available which does lead to better estimates.
2916 bucket
= blk_mq_poll_stats_bkt(rq
);
2920 if (q
->poll_stat
[bucket
].nr_samples
)
2921 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
2926 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2927 struct blk_mq_hw_ctx
*hctx
,
2930 struct hrtimer_sleeper hs
;
2931 enum hrtimer_mode mode
;
2935 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2941 * -1: don't ever hybrid sleep
2942 * 0: use half of prev avg
2943 * >0: use this specific value
2945 if (q
->poll_nsec
== -1)
2947 else if (q
->poll_nsec
> 0)
2948 nsecs
= q
->poll_nsec
;
2950 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2955 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2958 * This will be replaced with the stats tracking code, using
2959 * 'avg_completion_time / 2' as the pre-sleep target.
2963 mode
= HRTIMER_MODE_REL
;
2964 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2965 hrtimer_set_expires(&hs
.timer
, kt
);
2967 hrtimer_init_sleeper(&hs
, current
);
2969 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2971 set_current_state(TASK_UNINTERRUPTIBLE
);
2972 hrtimer_start_expires(&hs
.timer
, mode
);
2975 hrtimer_cancel(&hs
.timer
);
2976 mode
= HRTIMER_MODE_ABS
;
2977 } while (hs
.task
&& !signal_pending(current
));
2979 __set_current_state(TASK_RUNNING
);
2980 destroy_hrtimer_on_stack(&hs
.timer
);
2984 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2986 struct request_queue
*q
= hctx
->queue
;
2990 * If we sleep, have the caller restart the poll loop to reset
2991 * the state. Like for the other success return cases, the
2992 * caller is responsible for checking if the IO completed. If
2993 * the IO isn't complete, we'll get called again and will go
2994 * straight to the busy poll loop.
2996 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2999 hctx
->poll_considered
++;
3001 state
= current
->state
;
3002 while (!need_resched()) {
3005 hctx
->poll_invoked
++;
3007 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
3009 hctx
->poll_success
++;
3010 set_current_state(TASK_RUNNING
);
3014 if (signal_pending_state(state
, current
))
3015 set_current_state(TASK_RUNNING
);
3017 if (current
->state
== TASK_RUNNING
)
3027 static bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
3029 struct blk_mq_hw_ctx
*hctx
;
3032 if (!test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3035 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3036 if (!blk_qc_t_is_internal(cookie
))
3037 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3039 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3041 * With scheduling, if the request has completed, we'll
3042 * get a NULL return here, as we clear the sched tag when
3043 * that happens. The request still remains valid, like always,
3044 * so we should be safe with just the NULL check.
3050 return __blk_mq_poll(hctx
, rq
);
3053 static int __init
blk_mq_init(void)
3056 * See comment in block/blk.h rq_atomic_flags enum
3058 BUILD_BUG_ON((REQ_ATOM_STARTED
/ BITS_PER_BYTE
) !=
3059 (REQ_ATOM_COMPLETE
/ BITS_PER_BYTE
));
3061 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3062 blk_mq_hctx_notify_dead
);
3065 subsys_initcall(blk_mq_init
);