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 (data
->hctx
->flags
& BLK_MQ_F_TAG_SHARED
) {
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
);
380 blk_mq_tag_busy(data
->hctx
);
383 tag
= blk_mq_get_tag(data
);
384 if (tag
== BLK_MQ_TAG_FAIL
) {
385 if (put_ctx_on_error
) {
386 blk_mq_put_ctx(data
->ctx
);
393 rq
= blk_mq_rq_ctx_init(data
, tag
, op
);
394 if (!op_is_flush(op
)) {
396 if (e
&& e
->type
->ops
.mq
.prepare_request
) {
397 if (e
->type
->icq_cache
&& rq_ioc(bio
))
398 blk_mq_sched_assign_ioc(rq
, bio
);
400 e
->type
->ops
.mq
.prepare_request(rq
, bio
);
401 rq
->rq_flags
|= RQF_ELVPRIV
;
404 data
->hctx
->queued
++;
408 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
409 blk_mq_req_flags_t flags
)
411 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
415 ret
= blk_queue_enter(q
, flags
);
419 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
423 return ERR_PTR(-EWOULDBLOCK
);
425 blk_mq_put_ctx(alloc_data
.ctx
);
428 rq
->__sector
= (sector_t
) -1;
429 rq
->bio
= rq
->biotail
= NULL
;
432 EXPORT_SYMBOL(blk_mq_alloc_request
);
434 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
435 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
437 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
443 * If the tag allocator sleeps we could get an allocation for a
444 * different hardware context. No need to complicate the low level
445 * allocator for this for the rare use case of a command tied to
448 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
449 return ERR_PTR(-EINVAL
);
451 if (hctx_idx
>= q
->nr_hw_queues
)
452 return ERR_PTR(-EIO
);
454 ret
= blk_queue_enter(q
, flags
);
459 * Check if the hardware context is actually mapped to anything.
460 * If not tell the caller that it should skip this queue.
462 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
463 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
465 return ERR_PTR(-EXDEV
);
467 cpu
= cpumask_first_and(alloc_data
.hctx
->cpumask
, cpu_online_mask
);
468 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
470 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
474 return ERR_PTR(-EWOULDBLOCK
);
478 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
480 void blk_mq_free_request(struct request
*rq
)
482 struct request_queue
*q
= rq
->q
;
483 struct elevator_queue
*e
= q
->elevator
;
484 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
485 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
486 const int sched_tag
= rq
->internal_tag
;
488 if (rq
->rq_flags
& RQF_ELVPRIV
) {
489 if (e
&& e
->type
->ops
.mq
.finish_request
)
490 e
->type
->ops
.mq
.finish_request(rq
);
492 put_io_context(rq
->elv
.icq
->ioc
);
497 ctx
->rq_completed
[rq_is_sync(rq
)]++;
498 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
499 atomic_dec(&hctx
->nr_active
);
501 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
502 laptop_io_completion(q
->backing_dev_info
);
504 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
507 blk_put_rl(blk_rq_rl(rq
));
509 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
510 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
512 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
514 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
515 blk_mq_sched_restart(hctx
);
518 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
520 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
522 blk_account_io_done(rq
);
525 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
526 rq
->end_io(rq
, error
);
528 if (unlikely(blk_bidi_rq(rq
)))
529 blk_mq_free_request(rq
->next_rq
);
530 blk_mq_free_request(rq
);
533 EXPORT_SYMBOL(__blk_mq_end_request
);
535 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
537 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
539 __blk_mq_end_request(rq
, error
);
541 EXPORT_SYMBOL(blk_mq_end_request
);
543 static void __blk_mq_complete_request_remote(void *data
)
545 struct request
*rq
= data
;
547 rq
->q
->softirq_done_fn(rq
);
550 static void __blk_mq_complete_request(struct request
*rq
)
552 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
556 if (rq
->internal_tag
!= -1)
557 blk_mq_sched_completed_request(rq
);
558 if (rq
->rq_flags
& RQF_STATS
) {
559 blk_mq_poll_stats_start(rq
->q
);
563 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
564 rq
->q
->softirq_done_fn(rq
);
569 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
570 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
572 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
573 rq
->csd
.func
= __blk_mq_complete_request_remote
;
576 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
578 rq
->q
->softirq_done_fn(rq
);
584 * blk_mq_complete_request - end I/O on a request
585 * @rq: the request being processed
588 * Ends all I/O on a request. It does not handle partial completions.
589 * The actual completion happens out-of-order, through a IPI handler.
591 void blk_mq_complete_request(struct request
*rq
)
593 struct request_queue
*q
= rq
->q
;
595 if (unlikely(blk_should_fake_timeout(q
)))
597 if (!blk_mark_rq_complete(rq
))
598 __blk_mq_complete_request(rq
);
600 EXPORT_SYMBOL(blk_mq_complete_request
);
602 int blk_mq_request_started(struct request
*rq
)
604 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
606 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
608 void blk_mq_start_request(struct request
*rq
)
610 struct request_queue
*q
= rq
->q
;
612 blk_mq_sched_started_request(rq
);
614 trace_block_rq_issue(q
, rq
);
616 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
617 blk_stat_set_issue(&rq
->issue_stat
, blk_rq_sectors(rq
));
618 rq
->rq_flags
|= RQF_STATS
;
619 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
624 WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
));
627 * Mark us as started and clear complete. Complete might have been
628 * set if requeue raced with timeout, which then marked it as
629 * complete. So be sure to clear complete again when we start
630 * the request, otherwise we'll ignore the completion event.
632 * Ensure that ->deadline is visible before we set STARTED, such that
633 * blk_mq_check_expired() is guaranteed to observe our ->deadline when
634 * it observes STARTED.
637 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
638 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
)) {
640 * Coherence order guarantees these consecutive stores to a
641 * single variable propagate in the specified order. Thus the
642 * clear_bit() is ordered _after_ the set bit. See
643 * blk_mq_check_expired().
645 * (the bits must be part of the same byte for this to be
648 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
651 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
653 * Make sure space for the drain appears. We know we can do
654 * this because max_hw_segments has been adjusted to be one
655 * fewer than the device can handle.
657 rq
->nr_phys_segments
++;
660 EXPORT_SYMBOL(blk_mq_start_request
);
663 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
664 * flag isn't set yet, so there may be race with timeout handler,
665 * but given rq->deadline is just set in .queue_rq() under
666 * this situation, the race won't be possible in reality because
667 * rq->timeout should be set as big enough to cover the window
668 * between blk_mq_start_request() called from .queue_rq() and
669 * clearing REQ_ATOM_STARTED here.
671 static void __blk_mq_requeue_request(struct request
*rq
)
673 struct request_queue
*q
= rq
->q
;
675 blk_mq_put_driver_tag(rq
);
677 trace_block_rq_requeue(q
, rq
);
678 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
680 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
681 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
682 rq
->nr_phys_segments
--;
686 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
688 __blk_mq_requeue_request(rq
);
690 /* this request will be re-inserted to io scheduler queue */
691 blk_mq_sched_requeue_request(rq
);
693 BUG_ON(blk_queued_rq(rq
));
694 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
696 EXPORT_SYMBOL(blk_mq_requeue_request
);
698 static void blk_mq_requeue_work(struct work_struct
*work
)
700 struct request_queue
*q
=
701 container_of(work
, struct request_queue
, requeue_work
.work
);
703 struct request
*rq
, *next
;
705 spin_lock_irq(&q
->requeue_lock
);
706 list_splice_init(&q
->requeue_list
, &rq_list
);
707 spin_unlock_irq(&q
->requeue_lock
);
709 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
710 if (!(rq
->rq_flags
& (RQF_SOFTBARRIER
| RQF_DONTPREP
)))
713 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
714 list_del_init(&rq
->queuelist
);
716 * If RQF_DONTPREP, rq has contained some driver specific
717 * data, so insert it to hctx dispatch list to avoid any
720 if (rq
->rq_flags
& RQF_DONTPREP
)
721 blk_mq_request_bypass_insert(rq
, false);
723 blk_mq_sched_insert_request(rq
, true, false, false, true);
726 while (!list_empty(&rq_list
)) {
727 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
728 list_del_init(&rq
->queuelist
);
729 blk_mq_sched_insert_request(rq
, false, false, false, true);
732 blk_mq_run_hw_queues(q
, false);
735 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
736 bool kick_requeue_list
)
738 struct request_queue
*q
= rq
->q
;
742 * We abuse this flag that is otherwise used by the I/O scheduler to
743 * request head insertion from the workqueue.
745 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
747 spin_lock_irqsave(&q
->requeue_lock
, flags
);
749 rq
->rq_flags
|= RQF_SOFTBARRIER
;
750 list_add(&rq
->queuelist
, &q
->requeue_list
);
752 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
754 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
756 if (kick_requeue_list
)
757 blk_mq_kick_requeue_list(q
);
759 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
761 void blk_mq_kick_requeue_list(struct request_queue
*q
)
763 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
765 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
767 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
770 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
771 msecs_to_jiffies(msecs
));
773 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
775 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
777 if (tag
< tags
->nr_tags
) {
778 prefetch(tags
->rqs
[tag
]);
779 return tags
->rqs
[tag
];
784 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
786 struct blk_mq_timeout_data
{
788 unsigned int next_set
;
791 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
793 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
794 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
797 * We know that complete is set at this point. If STARTED isn't set
798 * anymore, then the request isn't active and the "timeout" should
799 * just be ignored. This can happen due to the bitflag ordering.
800 * Timeout first checks if STARTED is set, and if it is, assumes
801 * the request is active. But if we race with completion, then
802 * both flags will get cleared. So check here again, and ignore
803 * a timeout event with a request that isn't active.
805 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
809 ret
= ops
->timeout(req
, reserved
);
813 __blk_mq_complete_request(req
);
815 case BLK_EH_RESET_TIMER
:
817 blk_clear_rq_complete(req
);
819 case BLK_EH_NOT_HANDLED
:
822 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
827 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
828 struct request
*rq
, void *priv
, bool reserved
)
830 struct blk_mq_timeout_data
*data
= priv
;
831 unsigned long deadline
;
833 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
837 * Ensures that if we see STARTED we must also see our
838 * up-to-date deadline, see blk_mq_start_request().
842 deadline
= READ_ONCE(rq
->deadline
);
845 * The rq being checked may have been freed and reallocated
846 * out already here, we avoid this race by checking rq->deadline
847 * and REQ_ATOM_COMPLETE flag together:
849 * - if rq->deadline is observed as new value because of
850 * reusing, the rq won't be timed out because of timing.
851 * - if rq->deadline is observed as previous value,
852 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
853 * because we put a barrier between setting rq->deadline
854 * and clearing the flag in blk_mq_start_request(), so
855 * this rq won't be timed out too.
857 if (time_after_eq(jiffies
, deadline
)) {
858 if (!blk_mark_rq_complete(rq
)) {
860 * Again coherence order ensures that consecutive reads
861 * from the same variable must be in that order. This
862 * ensures that if we see COMPLETE clear, we must then
863 * see STARTED set and we'll ignore this timeout.
865 * (There's also the MB implied by the test_and_clear())
867 blk_mq_rq_timed_out(rq
, reserved
);
869 } else if (!data
->next_set
|| time_after(data
->next
, deadline
)) {
870 data
->next
= deadline
;
875 static void blk_mq_timeout_work(struct work_struct
*work
)
877 struct request_queue
*q
=
878 container_of(work
, struct request_queue
, timeout_work
);
879 struct blk_mq_timeout_data data
= {
885 /* A deadlock might occur if a request is stuck requiring a
886 * timeout at the same time a queue freeze is waiting
887 * completion, since the timeout code would not be able to
888 * acquire the queue reference here.
890 * That's why we don't use blk_queue_enter here; instead, we use
891 * percpu_ref_tryget directly, because we need to be able to
892 * obtain a reference even in the short window between the queue
893 * starting to freeze, by dropping the first reference in
894 * blk_freeze_queue_start, and the moment the last request is
895 * consumed, marked by the instant q_usage_counter reaches
898 if (!percpu_ref_tryget(&q
->q_usage_counter
))
901 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
904 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
905 mod_timer(&q
->timeout
, data
.next
);
907 struct blk_mq_hw_ctx
*hctx
;
909 queue_for_each_hw_ctx(q
, hctx
, i
) {
910 /* the hctx may be unmapped, so check it here */
911 if (blk_mq_hw_queue_mapped(hctx
))
912 blk_mq_tag_idle(hctx
);
918 struct flush_busy_ctx_data
{
919 struct blk_mq_hw_ctx
*hctx
;
920 struct list_head
*list
;
923 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
925 struct flush_busy_ctx_data
*flush_data
= data
;
926 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
927 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
929 sbitmap_clear_bit(sb
, bitnr
);
930 spin_lock(&ctx
->lock
);
931 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
932 spin_unlock(&ctx
->lock
);
937 * Process software queues that have been marked busy, splicing them
938 * to the for-dispatch
940 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
942 struct flush_busy_ctx_data data
= {
947 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
949 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
951 struct dispatch_rq_data
{
952 struct blk_mq_hw_ctx
*hctx
;
956 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
959 struct dispatch_rq_data
*dispatch_data
= data
;
960 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
961 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
963 spin_lock(&ctx
->lock
);
964 if (unlikely(!list_empty(&ctx
->rq_list
))) {
965 dispatch_data
->rq
= list_entry_rq(ctx
->rq_list
.next
);
966 list_del_init(&dispatch_data
->rq
->queuelist
);
967 if (list_empty(&ctx
->rq_list
))
968 sbitmap_clear_bit(sb
, bitnr
);
970 spin_unlock(&ctx
->lock
);
972 return !dispatch_data
->rq
;
975 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
976 struct blk_mq_ctx
*start
)
978 unsigned off
= start
? start
->index_hw
: 0;
979 struct dispatch_rq_data data
= {
984 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
985 dispatch_rq_from_ctx
, &data
);
990 static inline unsigned int queued_to_index(unsigned int queued
)
995 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
998 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
1001 struct blk_mq_alloc_data data
= {
1003 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
1004 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
1008 might_sleep_if(wait
);
1013 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
1014 data
.flags
|= BLK_MQ_REQ_RESERVED
;
1016 shared
= blk_mq_tag_busy(data
.hctx
);
1017 rq
->tag
= blk_mq_get_tag(&data
);
1020 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1021 atomic_inc(&data
.hctx
->nr_active
);
1023 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
1029 return rq
->tag
!= -1;
1032 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1033 int flags
, void *key
)
1035 struct blk_mq_hw_ctx
*hctx
;
1037 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1039 list_del_init(&wait
->entry
);
1040 blk_mq_run_hw_queue(hctx
, true);
1045 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1046 * the tag wakeups. For non-shared tags, we can simply mark us nedeing a
1047 * restart. For both caes, take care to check the condition again after
1048 * marking us as waiting.
1050 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
**hctx
,
1053 struct blk_mq_hw_ctx
*this_hctx
= *hctx
;
1054 bool shared_tags
= (this_hctx
->flags
& BLK_MQ_F_TAG_SHARED
) != 0;
1055 struct sbq_wait_state
*ws
;
1056 wait_queue_entry_t
*wait
;
1060 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &this_hctx
->state
))
1061 set_bit(BLK_MQ_S_SCHED_RESTART
, &this_hctx
->state
);
1063 wait
= &this_hctx
->dispatch_wait
;
1064 if (!list_empty_careful(&wait
->entry
))
1067 spin_lock(&this_hctx
->lock
);
1068 if (!list_empty(&wait
->entry
)) {
1069 spin_unlock(&this_hctx
->lock
);
1073 ws
= bt_wait_ptr(&this_hctx
->tags
->bitmap_tags
, this_hctx
);
1074 add_wait_queue(&ws
->wait
, wait
);
1078 * It's possible that a tag was freed in the window between the
1079 * allocation failure and adding the hardware queue to the wait
1082 ret
= blk_mq_get_driver_tag(rq
, hctx
, false);
1086 * Don't clear RESTART here, someone else could have set it.
1087 * At most this will cost an extra queue run.
1092 spin_unlock(&this_hctx
->lock
);
1097 * We got a tag, remove ourselves from the wait queue to ensure
1098 * someone else gets the wakeup.
1100 spin_lock_irq(&ws
->wait
.lock
);
1101 list_del_init(&wait
->entry
);
1102 spin_unlock_irq(&ws
->wait
.lock
);
1103 spin_unlock(&this_hctx
->lock
);
1108 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
,
1111 struct blk_mq_hw_ctx
*hctx
;
1112 struct request
*rq
, *nxt
;
1113 bool no_tag
= false;
1116 if (list_empty(list
))
1119 WARN_ON(!list_is_singular(list
) && got_budget
);
1122 * Now process all the entries, sending them to the driver.
1124 errors
= queued
= 0;
1126 struct blk_mq_queue_data bd
;
1129 rq
= list_first_entry(list
, struct request
, queuelist
);
1131 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
1132 if (!got_budget
&& !blk_mq_get_dispatch_budget(hctx
))
1135 if (!blk_mq_get_driver_tag(rq
, NULL
, false)) {
1137 * The initial allocation attempt failed, so we need to
1138 * rerun the hardware queue when a tag is freed. The
1139 * waitqueue takes care of that. If the queue is run
1140 * before we add this entry back on the dispatch list,
1141 * we'll re-run it below.
1143 if (!blk_mq_mark_tag_wait(&hctx
, rq
)) {
1144 blk_mq_put_dispatch_budget(hctx
);
1146 * For non-shared tags, the RESTART check
1149 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
)
1155 list_del_init(&rq
->queuelist
);
1160 * Flag last if we have no more requests, or if we have more
1161 * but can't assign a driver tag to it.
1163 if (list_empty(list
))
1166 nxt
= list_first_entry(list
, struct request
, queuelist
);
1167 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1170 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1171 if (ret
== BLK_STS_RESOURCE
) {
1173 * If an I/O scheduler has been configured and we got a
1174 * driver tag for the next request already, free it
1177 if (!list_empty(list
)) {
1178 nxt
= list_first_entry(list
, struct request
, queuelist
);
1179 blk_mq_put_driver_tag(nxt
);
1181 list_add(&rq
->queuelist
, list
);
1182 __blk_mq_requeue_request(rq
);
1186 if (unlikely(ret
!= BLK_STS_OK
)) {
1188 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1193 } while (!list_empty(list
));
1195 hctx
->dispatched
[queued_to_index(queued
)]++;
1198 * Any items that need requeuing? Stuff them into hctx->dispatch,
1199 * that is where we will continue on next queue run.
1201 if (!list_empty(list
)) {
1202 spin_lock(&hctx
->lock
);
1203 list_splice_init(list
, &hctx
->dispatch
);
1204 spin_unlock(&hctx
->lock
);
1207 * If SCHED_RESTART was set by the caller of this function and
1208 * it is no longer set that means that it was cleared by another
1209 * thread and hence that a queue rerun is needed.
1211 * If 'no_tag' is set, that means that we failed getting
1212 * a driver tag with an I/O scheduler attached. If our dispatch
1213 * waitqueue is no longer active, ensure that we run the queue
1214 * AFTER adding our entries back to the list.
1216 * If no I/O scheduler has been configured it is possible that
1217 * the hardware queue got stopped and restarted before requests
1218 * were pushed back onto the dispatch list. Rerun the queue to
1219 * avoid starvation. Notes:
1220 * - blk_mq_run_hw_queue() checks whether or not a queue has
1221 * been stopped before rerunning a queue.
1222 * - Some but not all block drivers stop a queue before
1223 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1226 if (!blk_mq_sched_needs_restart(hctx
) ||
1227 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1228 blk_mq_run_hw_queue(hctx
, true);
1231 return (queued
+ errors
) != 0;
1234 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1239 * We should be running this queue from one of the CPUs that
1242 * There are at least two related races now between setting
1243 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1244 * __blk_mq_run_hw_queue():
1246 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1247 * but later it becomes online, then this warning is harmless
1250 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1251 * but later it becomes offline, then the warning can't be
1252 * triggered, and we depend on blk-mq timeout handler to
1253 * handle dispatched requests to this hctx
1255 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1256 cpu_online(hctx
->next_cpu
)) {
1257 printk(KERN_WARNING
"run queue from wrong CPU %d, hctx %s\n",
1258 raw_smp_processor_id(),
1259 cpumask_empty(hctx
->cpumask
) ? "inactive": "active");
1264 * We can't run the queue inline with ints disabled. Ensure that
1265 * we catch bad users of this early.
1267 WARN_ON_ONCE(in_interrupt());
1269 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1271 blk_mq_sched_dispatch_requests(hctx
);
1276 srcu_idx
= srcu_read_lock(hctx
->queue_rq_srcu
);
1277 blk_mq_sched_dispatch_requests(hctx
);
1278 srcu_read_unlock(hctx
->queue_rq_srcu
, srcu_idx
);
1282 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1284 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1286 if (cpu
>= nr_cpu_ids
)
1287 cpu
= cpumask_first(hctx
->cpumask
);
1292 * It'd be great if the workqueue API had a way to pass
1293 * in a mask and had some smarts for more clever placement.
1294 * For now we just round-robin here, switching for every
1295 * BLK_MQ_CPU_WORK_BATCH queued items.
1297 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1300 int next_cpu
= hctx
->next_cpu
;
1302 if (hctx
->queue
->nr_hw_queues
== 1)
1303 return WORK_CPU_UNBOUND
;
1305 if (--hctx
->next_cpu_batch
<= 0) {
1307 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1309 if (next_cpu
>= nr_cpu_ids
)
1310 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1311 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1315 * Do unbound schedule if we can't find a online CPU for this hctx,
1316 * and it should only happen in the path of handling CPU DEAD.
1318 if (!cpu_online(next_cpu
)) {
1325 * Make sure to re-select CPU next time once after CPUs
1326 * in hctx->cpumask become online again.
1328 hctx
->next_cpu
= next_cpu
;
1329 hctx
->next_cpu_batch
= 1;
1330 return WORK_CPU_UNBOUND
;
1333 hctx
->next_cpu
= next_cpu
;
1337 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1338 unsigned long msecs
)
1340 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1343 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1344 int cpu
= get_cpu();
1345 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1346 __blk_mq_run_hw_queue(hctx
);
1354 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1355 msecs_to_jiffies(msecs
));
1358 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1360 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1362 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1364 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1366 if (blk_mq_hctx_has_pending(hctx
)) {
1367 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1373 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1375 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1377 struct blk_mq_hw_ctx
*hctx
;
1380 queue_for_each_hw_ctx(q
, hctx
, i
) {
1381 if (blk_mq_hctx_stopped(hctx
))
1384 blk_mq_run_hw_queue(hctx
, async
);
1387 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1390 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1391 * @q: request queue.
1393 * The caller is responsible for serializing this function against
1394 * blk_mq_{start,stop}_hw_queue().
1396 bool blk_mq_queue_stopped(struct request_queue
*q
)
1398 struct blk_mq_hw_ctx
*hctx
;
1401 queue_for_each_hw_ctx(q
, hctx
, i
)
1402 if (blk_mq_hctx_stopped(hctx
))
1407 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1410 * This function is often used for pausing .queue_rq() by driver when
1411 * there isn't enough resource or some conditions aren't satisfied, and
1412 * BLK_STS_RESOURCE is usually returned.
1414 * We do not guarantee that dispatch can be drained or blocked
1415 * after blk_mq_stop_hw_queue() returns. Please use
1416 * blk_mq_quiesce_queue() for that requirement.
1418 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1420 cancel_delayed_work(&hctx
->run_work
);
1422 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1424 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1427 * This function is often used for pausing .queue_rq() by driver when
1428 * there isn't enough resource or some conditions aren't satisfied, and
1429 * BLK_STS_RESOURCE is usually returned.
1431 * We do not guarantee that dispatch can be drained or blocked
1432 * after blk_mq_stop_hw_queues() returns. Please use
1433 * blk_mq_quiesce_queue() for that requirement.
1435 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1437 struct blk_mq_hw_ctx
*hctx
;
1440 queue_for_each_hw_ctx(q
, hctx
, i
)
1441 blk_mq_stop_hw_queue(hctx
);
1443 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1445 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1447 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1449 blk_mq_run_hw_queue(hctx
, false);
1451 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1453 void blk_mq_start_hw_queues(struct request_queue
*q
)
1455 struct blk_mq_hw_ctx
*hctx
;
1458 queue_for_each_hw_ctx(q
, hctx
, i
)
1459 blk_mq_start_hw_queue(hctx
);
1461 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1463 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1465 if (!blk_mq_hctx_stopped(hctx
))
1468 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1469 blk_mq_run_hw_queue(hctx
, async
);
1471 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1473 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1475 struct blk_mq_hw_ctx
*hctx
;
1478 queue_for_each_hw_ctx(q
, hctx
, i
)
1479 blk_mq_start_stopped_hw_queue(hctx
, async
);
1481 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1483 static void blk_mq_run_work_fn(struct work_struct
*work
)
1485 struct blk_mq_hw_ctx
*hctx
;
1487 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1490 * If we are stopped, don't run the queue. The exception is if
1491 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1492 * the STOPPED bit and run it.
1494 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)) {
1495 if (!test_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
))
1498 clear_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1499 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1502 __blk_mq_run_hw_queue(hctx
);
1506 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1508 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx
)))
1512 * Stop the hw queue, then modify currently delayed work.
1513 * This should prevent us from running the queue prematurely.
1514 * Mark the queue as auto-clearing STOPPED when it runs.
1516 blk_mq_stop_hw_queue(hctx
);
1517 set_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1518 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1520 msecs_to_jiffies(msecs
));
1522 EXPORT_SYMBOL(blk_mq_delay_queue
);
1524 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1528 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1530 lockdep_assert_held(&ctx
->lock
);
1532 trace_block_rq_insert(hctx
->queue
, rq
);
1535 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1537 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1540 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1543 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1545 lockdep_assert_held(&ctx
->lock
);
1547 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1548 blk_mq_hctx_mark_pending(hctx
, ctx
);
1552 * Should only be used carefully, when the caller knows we want to
1553 * bypass a potential IO scheduler on the target device.
1555 void blk_mq_request_bypass_insert(struct request
*rq
, bool run_queue
)
1557 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1558 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(rq
->q
, ctx
->cpu
);
1560 spin_lock(&hctx
->lock
);
1561 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1562 spin_unlock(&hctx
->lock
);
1565 blk_mq_run_hw_queue(hctx
, false);
1568 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1569 struct list_head
*list
)
1573 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1576 spin_lock(&ctx
->lock
);
1577 while (!list_empty(list
)) {
1580 rq
= list_first_entry(list
, struct request
, queuelist
);
1581 BUG_ON(rq
->mq_ctx
!= ctx
);
1582 list_del_init(&rq
->queuelist
);
1583 __blk_mq_insert_req_list(hctx
, rq
, false);
1585 blk_mq_hctx_mark_pending(hctx
, ctx
);
1586 spin_unlock(&ctx
->lock
);
1589 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1591 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1592 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1594 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1595 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1596 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1599 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1601 struct blk_mq_ctx
*this_ctx
;
1602 struct request_queue
*this_q
;
1605 LIST_HEAD(ctx_list
);
1608 list_splice_init(&plug
->mq_list
, &list
);
1610 list_sort(NULL
, &list
, plug_ctx_cmp
);
1616 while (!list_empty(&list
)) {
1617 rq
= list_entry_rq(list
.next
);
1618 list_del_init(&rq
->queuelist
);
1620 if (rq
->mq_ctx
!= this_ctx
) {
1622 trace_block_unplug(this_q
, depth
, !from_schedule
);
1623 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1628 this_ctx
= rq
->mq_ctx
;
1634 list_add_tail(&rq
->queuelist
, &ctx_list
);
1638 * If 'this_ctx' is set, we know we have entries to complete
1639 * on 'ctx_list'. Do those.
1642 trace_block_unplug(this_q
, depth
, !from_schedule
);
1643 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1648 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1650 blk_init_request_from_bio(rq
, bio
);
1652 blk_rq_set_rl(rq
, blk_get_rl(rq
->q
, bio
));
1654 blk_account_io_start(rq
, true);
1657 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx
*hctx
,
1658 struct blk_mq_ctx
*ctx
,
1661 spin_lock(&ctx
->lock
);
1662 __blk_mq_insert_request(hctx
, rq
, false);
1663 spin_unlock(&ctx
->lock
);
1666 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1669 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1671 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1674 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1676 blk_qc_t
*cookie
, bool may_sleep
)
1678 struct request_queue
*q
= rq
->q
;
1679 struct blk_mq_queue_data bd
= {
1683 blk_qc_t new_cookie
;
1685 bool run_queue
= true;
1687 /* RCU or SRCU read lock is needed before checking quiesced flag */
1688 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1696 if (!blk_mq_get_dispatch_budget(hctx
))
1699 if (!blk_mq_get_driver_tag(rq
, NULL
, false)) {
1700 blk_mq_put_dispatch_budget(hctx
);
1704 new_cookie
= request_to_qc_t(hctx
, rq
);
1707 * For OK queue, we are done. For error, kill it. Any other
1708 * error (busy), just add it to our list as we previously
1711 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1714 *cookie
= new_cookie
;
1716 case BLK_STS_RESOURCE
:
1717 __blk_mq_requeue_request(rq
);
1720 *cookie
= BLK_QC_T_NONE
;
1721 blk_mq_end_request(rq
, ret
);
1726 blk_mq_sched_insert_request(rq
, false, run_queue
, false, may_sleep
);
1729 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1730 struct request
*rq
, blk_qc_t
*cookie
)
1732 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1734 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false);
1737 unsigned int srcu_idx
;
1741 srcu_idx
= srcu_read_lock(hctx
->queue_rq_srcu
);
1742 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, true);
1743 srcu_read_unlock(hctx
->queue_rq_srcu
, srcu_idx
);
1747 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1749 const int is_sync
= op_is_sync(bio
->bi_opf
);
1750 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1751 struct blk_mq_alloc_data data
= { .flags
= 0 };
1753 unsigned int request_count
= 0;
1754 struct blk_plug
*plug
;
1755 struct request
*same_queue_rq
= NULL
;
1757 unsigned int wb_acct
;
1759 blk_queue_bounce(q
, &bio
);
1761 blk_queue_split(q
, &bio
);
1763 if (!bio_integrity_prep(bio
))
1764 return BLK_QC_T_NONE
;
1766 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1767 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1768 return BLK_QC_T_NONE
;
1770 if (blk_mq_sched_bio_merge(q
, bio
))
1771 return BLK_QC_T_NONE
;
1773 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1775 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1777 rq
= blk_mq_get_request(q
, bio
, bio
->bi_opf
, &data
);
1778 if (unlikely(!rq
)) {
1779 __wbt_done(q
->rq_wb
, wb_acct
);
1780 if (bio
->bi_opf
& REQ_NOWAIT
)
1781 bio_wouldblock_error(bio
);
1782 return BLK_QC_T_NONE
;
1785 wbt_track(&rq
->issue_stat
, wb_acct
);
1787 cookie
= request_to_qc_t(data
.hctx
, rq
);
1789 plug
= current
->plug
;
1790 if (unlikely(is_flush_fua
)) {
1791 blk_mq_put_ctx(data
.ctx
);
1792 blk_mq_bio_to_request(rq
, bio
);
1794 /* bypass scheduler for flush rq */
1795 blk_insert_flush(rq
);
1796 blk_mq_run_hw_queue(data
.hctx
, true);
1797 } else if (plug
&& q
->nr_hw_queues
== 1) {
1798 struct request
*last
= NULL
;
1800 blk_mq_put_ctx(data
.ctx
);
1801 blk_mq_bio_to_request(rq
, bio
);
1804 * @request_count may become stale because of schedule
1805 * out, so check the list again.
1807 if (list_empty(&plug
->mq_list
))
1809 else if (blk_queue_nomerges(q
))
1810 request_count
= blk_plug_queued_count(q
);
1813 trace_block_plug(q
);
1815 last
= list_entry_rq(plug
->mq_list
.prev
);
1817 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1818 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1819 blk_flush_plug_list(plug
, false);
1820 trace_block_plug(q
);
1823 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1824 } else if (plug
&& !blk_queue_nomerges(q
)) {
1825 blk_mq_bio_to_request(rq
, bio
);
1828 * We do limited plugging. If the bio can be merged, do that.
1829 * Otherwise the existing request in the plug list will be
1830 * issued. So the plug list will have one request at most
1831 * The plug list might get flushed before this. If that happens,
1832 * the plug list is empty, and same_queue_rq is invalid.
1834 if (list_empty(&plug
->mq_list
))
1835 same_queue_rq
= NULL
;
1837 list_del_init(&same_queue_rq
->queuelist
);
1838 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1840 blk_mq_put_ctx(data
.ctx
);
1842 if (same_queue_rq
) {
1843 data
.hctx
= blk_mq_map_queue(q
,
1844 same_queue_rq
->mq_ctx
->cpu
);
1845 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1848 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1849 blk_mq_put_ctx(data
.ctx
);
1850 blk_mq_bio_to_request(rq
, bio
);
1851 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1852 } else if (q
->elevator
) {
1853 blk_mq_put_ctx(data
.ctx
);
1854 blk_mq_bio_to_request(rq
, bio
);
1855 blk_mq_sched_insert_request(rq
, false, true, true, true);
1857 blk_mq_put_ctx(data
.ctx
);
1858 blk_mq_bio_to_request(rq
, bio
);
1859 blk_mq_queue_io(data
.hctx
, data
.ctx
, rq
);
1860 blk_mq_run_hw_queue(data
.hctx
, true);
1866 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1867 unsigned int hctx_idx
)
1871 if (tags
->rqs
&& set
->ops
->exit_request
) {
1874 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1875 struct request
*rq
= tags
->static_rqs
[i
];
1879 set
->ops
->exit_request(set
, rq
, hctx_idx
);
1880 tags
->static_rqs
[i
] = NULL
;
1884 while (!list_empty(&tags
->page_list
)) {
1885 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1886 list_del_init(&page
->lru
);
1888 * Remove kmemleak object previously allocated in
1889 * blk_mq_init_rq_map().
1891 kmemleak_free(page_address(page
));
1892 __free_pages(page
, page
->private);
1896 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1900 kfree(tags
->static_rqs
);
1901 tags
->static_rqs
= NULL
;
1903 blk_mq_free_tags(tags
);
1906 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1907 unsigned int hctx_idx
,
1908 unsigned int nr_tags
,
1909 unsigned int reserved_tags
)
1911 struct blk_mq_tags
*tags
;
1914 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1915 if (node
== NUMA_NO_NODE
)
1916 node
= set
->numa_node
;
1918 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1919 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1923 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1924 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1927 blk_mq_free_tags(tags
);
1931 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1932 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1934 if (!tags
->static_rqs
) {
1936 blk_mq_free_tags(tags
);
1943 static size_t order_to_size(unsigned int order
)
1945 return (size_t)PAGE_SIZE
<< order
;
1948 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1949 unsigned int hctx_idx
, unsigned int depth
)
1951 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1952 size_t rq_size
, left
;
1955 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1956 if (node
== NUMA_NO_NODE
)
1957 node
= set
->numa_node
;
1959 INIT_LIST_HEAD(&tags
->page_list
);
1962 * rq_size is the size of the request plus driver payload, rounded
1963 * to the cacheline size
1965 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1967 left
= rq_size
* depth
;
1969 for (i
= 0; i
< depth
; ) {
1970 int this_order
= max_order
;
1975 while (this_order
&& left
< order_to_size(this_order
- 1))
1979 page
= alloc_pages_node(node
,
1980 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1986 if (order_to_size(this_order
) < rq_size
)
1993 page
->private = this_order
;
1994 list_add_tail(&page
->lru
, &tags
->page_list
);
1996 p
= page_address(page
);
1998 * Allow kmemleak to scan these pages as they contain pointers
1999 * to additional allocations like via ops->init_request().
2001 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2002 entries_per_page
= order_to_size(this_order
) / rq_size
;
2003 to_do
= min(entries_per_page
, depth
- i
);
2004 left
-= to_do
* rq_size
;
2005 for (j
= 0; j
< to_do
; j
++) {
2006 struct request
*rq
= p
;
2008 tags
->static_rqs
[i
] = rq
;
2009 if (set
->ops
->init_request
) {
2010 if (set
->ops
->init_request(set
, rq
, hctx_idx
,
2012 tags
->static_rqs
[i
] = NULL
;
2024 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2029 * 'cpu' is going away. splice any existing rq_list entries from this
2030 * software queue to the hw queue dispatch list, and ensure that it
2033 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2035 struct blk_mq_hw_ctx
*hctx
;
2036 struct blk_mq_ctx
*ctx
;
2039 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2040 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2042 spin_lock(&ctx
->lock
);
2043 if (!list_empty(&ctx
->rq_list
)) {
2044 list_splice_init(&ctx
->rq_list
, &tmp
);
2045 blk_mq_hctx_clear_pending(hctx
, ctx
);
2047 spin_unlock(&ctx
->lock
);
2049 if (list_empty(&tmp
))
2052 spin_lock(&hctx
->lock
);
2053 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2054 spin_unlock(&hctx
->lock
);
2056 blk_mq_run_hw_queue(hctx
, true);
2060 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2062 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2066 /* hctx->ctxs will be freed in queue's release handler */
2067 static void blk_mq_exit_hctx(struct request_queue
*q
,
2068 struct blk_mq_tag_set
*set
,
2069 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2071 blk_mq_debugfs_unregister_hctx(hctx
);
2073 if (blk_mq_hw_queue_mapped(hctx
))
2074 blk_mq_tag_idle(hctx
);
2076 if (set
->ops
->exit_request
)
2077 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2079 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2081 if (set
->ops
->exit_hctx
)
2082 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2084 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2085 cleanup_srcu_struct(hctx
->queue_rq_srcu
);
2087 blk_mq_remove_cpuhp(hctx
);
2088 blk_free_flush_queue(hctx
->fq
);
2089 sbitmap_free(&hctx
->ctx_map
);
2092 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2093 struct blk_mq_tag_set
*set
, int nr_queue
)
2095 struct blk_mq_hw_ctx
*hctx
;
2098 queue_for_each_hw_ctx(q
, hctx
, i
) {
2101 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2105 static int blk_mq_init_hctx(struct request_queue
*q
,
2106 struct blk_mq_tag_set
*set
,
2107 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2111 node
= hctx
->numa_node
;
2112 if (node
== NUMA_NO_NODE
)
2113 node
= hctx
->numa_node
= set
->numa_node
;
2115 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2116 spin_lock_init(&hctx
->lock
);
2117 INIT_LIST_HEAD(&hctx
->dispatch
);
2119 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
2121 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2123 hctx
->tags
= set
->tags
[hctx_idx
];
2126 * Allocate space for all possible cpus to avoid allocation at
2129 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2132 goto unregister_cpu_notifier
;
2134 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
2140 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2141 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2143 if (set
->ops
->init_hctx
&&
2144 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2147 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
2150 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
2152 goto sched_exit_hctx
;
2154 if (set
->ops
->init_request
&&
2155 set
->ops
->init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
2159 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2160 init_srcu_struct(hctx
->queue_rq_srcu
);
2162 blk_mq_debugfs_register_hctx(q
, hctx
);
2167 blk_free_flush_queue(hctx
->fq
);
2169 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2171 if (set
->ops
->exit_hctx
)
2172 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2174 sbitmap_free(&hctx
->ctx_map
);
2177 unregister_cpu_notifier
:
2178 blk_mq_remove_cpuhp(hctx
);
2182 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2183 unsigned int nr_hw_queues
)
2187 for_each_possible_cpu(i
) {
2188 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2189 struct blk_mq_hw_ctx
*hctx
;
2192 spin_lock_init(&__ctx
->lock
);
2193 INIT_LIST_HEAD(&__ctx
->rq_list
);
2197 * Set local node, IFF we have more than one hw queue. If
2198 * not, we remain on the home node of the device
2200 hctx
= blk_mq_map_queue(q
, i
);
2201 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2202 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2206 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2210 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2211 set
->queue_depth
, set
->reserved_tags
);
2212 if (!set
->tags
[hctx_idx
])
2215 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2220 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2221 set
->tags
[hctx_idx
] = NULL
;
2225 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2226 unsigned int hctx_idx
)
2228 if (set
->tags
[hctx_idx
]) {
2229 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2230 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2231 set
->tags
[hctx_idx
] = NULL
;
2235 static void blk_mq_map_swqueue(struct request_queue
*q
)
2237 unsigned int i
, hctx_idx
;
2238 struct blk_mq_hw_ctx
*hctx
;
2239 struct blk_mq_ctx
*ctx
;
2240 struct blk_mq_tag_set
*set
= q
->tag_set
;
2243 * Avoid others reading imcomplete hctx->cpumask through sysfs
2245 mutex_lock(&q
->sysfs_lock
);
2247 queue_for_each_hw_ctx(q
, hctx
, i
) {
2248 cpumask_clear(hctx
->cpumask
);
2253 * Map software to hardware queues.
2255 * If the cpu isn't present, the cpu is mapped to first hctx.
2257 for_each_possible_cpu(i
) {
2258 hctx_idx
= q
->mq_map
[i
];
2259 /* unmapped hw queue can be remapped after CPU topo changed */
2260 if (!set
->tags
[hctx_idx
] &&
2261 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2263 * If tags initialization fail for some hctx,
2264 * that hctx won't be brought online. In this
2265 * case, remap the current ctx to hctx[0] which
2266 * is guaranteed to always have tags allocated
2271 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2272 hctx
= blk_mq_map_queue(q
, i
);
2274 cpumask_set_cpu(i
, hctx
->cpumask
);
2275 ctx
->index_hw
= hctx
->nr_ctx
;
2276 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2279 mutex_unlock(&q
->sysfs_lock
);
2281 queue_for_each_hw_ctx(q
, hctx
, i
) {
2283 * If no software queues are mapped to this hardware queue,
2284 * disable it and free the request entries.
2286 if (!hctx
->nr_ctx
) {
2287 /* Never unmap queue 0. We need it as a
2288 * fallback in case of a new remap fails
2291 if (i
&& set
->tags
[i
])
2292 blk_mq_free_map_and_requests(set
, i
);
2298 hctx
->tags
= set
->tags
[i
];
2299 WARN_ON(!hctx
->tags
);
2302 * Set the map size to the number of mapped software queues.
2303 * This is more accurate and more efficient than looping
2304 * over all possibly mapped software queues.
2306 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2309 * Initialize batch roundrobin counts
2311 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2312 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2317 * Caller needs to ensure that we're either frozen/quiesced, or that
2318 * the queue isn't live yet.
2320 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2322 struct blk_mq_hw_ctx
*hctx
;
2325 queue_for_each_hw_ctx(q
, hctx
, i
) {
2327 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2328 atomic_inc(&q
->shared_hctx_restart
);
2329 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2331 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2332 atomic_dec(&q
->shared_hctx_restart
);
2333 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2338 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2341 struct request_queue
*q
;
2343 lockdep_assert_held(&set
->tag_list_lock
);
2345 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2346 blk_mq_freeze_queue(q
);
2347 queue_set_hctx_shared(q
, shared
);
2348 blk_mq_unfreeze_queue(q
);
2352 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2354 struct blk_mq_tag_set
*set
= q
->tag_set
;
2356 mutex_lock(&set
->tag_list_lock
);
2357 list_del_rcu(&q
->tag_set_list
);
2358 if (list_is_singular(&set
->tag_list
)) {
2359 /* just transitioned to unshared */
2360 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2361 /* update existing queue */
2362 blk_mq_update_tag_set_depth(set
, false);
2364 mutex_unlock(&set
->tag_list_lock
);
2366 INIT_LIST_HEAD(&q
->tag_set_list
);
2369 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2370 struct request_queue
*q
)
2374 mutex_lock(&set
->tag_list_lock
);
2377 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2379 if (!list_empty(&set
->tag_list
) &&
2380 !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2381 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2382 /* update existing queue */
2383 blk_mq_update_tag_set_depth(set
, true);
2385 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2386 queue_set_hctx_shared(q
, true);
2387 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2389 mutex_unlock(&set
->tag_list_lock
);
2393 * It is the actual release handler for mq, but we do it from
2394 * request queue's release handler for avoiding use-after-free
2395 * and headache because q->mq_kobj shouldn't have been introduced,
2396 * but we can't group ctx/kctx kobj without it.
2398 void blk_mq_release(struct request_queue
*q
)
2400 struct blk_mq_hw_ctx
*hctx
;
2403 cancel_delayed_work_sync(&q
->requeue_work
);
2405 /* hctx kobj stays in hctx */
2406 queue_for_each_hw_ctx(q
, hctx
, i
) {
2409 kobject_put(&hctx
->kobj
);
2414 kfree(q
->queue_hw_ctx
);
2417 * release .mq_kobj and sw queue's kobject now because
2418 * both share lifetime with request queue.
2420 blk_mq_sysfs_deinit(q
);
2422 free_percpu(q
->queue_ctx
);
2425 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2427 struct request_queue
*uninit_q
, *q
;
2429 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2431 return ERR_PTR(-ENOMEM
);
2433 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2435 blk_cleanup_queue(uninit_q
);
2439 EXPORT_SYMBOL(blk_mq_init_queue
);
2441 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2443 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2445 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, queue_rq_srcu
),
2446 __alignof__(struct blk_mq_hw_ctx
)) !=
2447 sizeof(struct blk_mq_hw_ctx
));
2449 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2450 hw_ctx_size
+= sizeof(struct srcu_struct
);
2455 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2456 struct request_queue
*q
)
2459 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2461 blk_mq_sysfs_unregister(q
);
2463 /* protect against switching io scheduler */
2464 mutex_lock(&q
->sysfs_lock
);
2465 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2471 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2472 hctxs
[i
] = kzalloc_node(blk_mq_hw_ctx_size(set
),
2477 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2484 atomic_set(&hctxs
[i
]->nr_active
, 0);
2485 hctxs
[i
]->numa_node
= node
;
2486 hctxs
[i
]->queue_num
= i
;
2488 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2489 free_cpumask_var(hctxs
[i
]->cpumask
);
2494 blk_mq_hctx_kobj_init(hctxs
[i
]);
2496 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2497 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2501 blk_mq_free_map_and_requests(set
, j
);
2502 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2503 kobject_put(&hctx
->kobj
);
2508 q
->nr_hw_queues
= i
;
2509 mutex_unlock(&q
->sysfs_lock
);
2510 blk_mq_sysfs_register(q
);
2513 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2514 struct request_queue
*q
)
2516 /* mark the queue as mq asap */
2517 q
->mq_ops
= set
->ops
;
2519 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2520 blk_mq_poll_stats_bkt
,
2521 BLK_MQ_POLL_STATS_BKTS
, q
);
2525 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2529 /* init q->mq_kobj and sw queues' kobjects */
2530 blk_mq_sysfs_init(q
);
2532 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2533 GFP_KERNEL
, set
->numa_node
);
2534 if (!q
->queue_hw_ctx
)
2537 q
->mq_map
= set
->mq_map
;
2539 blk_mq_realloc_hw_ctxs(set
, q
);
2540 if (!q
->nr_hw_queues
)
2543 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2544 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2546 q
->nr_queues
= nr_cpu_ids
;
2548 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2550 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2551 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2553 q
->sg_reserved_size
= INT_MAX
;
2555 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2556 INIT_LIST_HEAD(&q
->requeue_list
);
2557 spin_lock_init(&q
->requeue_lock
);
2559 blk_queue_make_request(q
, blk_mq_make_request
);
2560 if (q
->mq_ops
->poll
)
2561 q
->poll_fn
= blk_mq_poll
;
2564 * Do this after blk_queue_make_request() overrides it...
2566 q
->nr_requests
= set
->queue_depth
;
2569 * Default to classic polling
2573 if (set
->ops
->complete
)
2574 blk_queue_softirq_done(q
, set
->ops
->complete
);
2576 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2577 blk_mq_add_queue_tag_set(set
, q
);
2578 blk_mq_map_swqueue(q
);
2580 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2583 ret
= blk_mq_sched_init(q
);
2585 return ERR_PTR(ret
);
2591 kfree(q
->queue_hw_ctx
);
2593 free_percpu(q
->queue_ctx
);
2596 return ERR_PTR(-ENOMEM
);
2598 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2600 void blk_mq_free_queue(struct request_queue
*q
)
2602 struct blk_mq_tag_set
*set
= q
->tag_set
;
2604 blk_mq_del_queue_tag_set(q
);
2605 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2608 /* Basically redo blk_mq_init_queue with queue frozen */
2609 static void blk_mq_queue_reinit(struct request_queue
*q
)
2611 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2613 blk_mq_debugfs_unregister_hctxs(q
);
2614 blk_mq_sysfs_unregister(q
);
2617 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2618 * we should change hctx numa_node according to the new topology (this
2619 * involves freeing and re-allocating memory, worth doing?)
2621 blk_mq_map_swqueue(q
);
2623 blk_mq_sysfs_register(q
);
2624 blk_mq_debugfs_register_hctxs(q
);
2627 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2631 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2632 if (!__blk_mq_alloc_rq_map(set
, i
))
2639 blk_mq_free_rq_map(set
->tags
[i
]);
2645 * Allocate the request maps associated with this tag_set. Note that this
2646 * may reduce the depth asked for, if memory is tight. set->queue_depth
2647 * will be updated to reflect the allocated depth.
2649 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2654 depth
= set
->queue_depth
;
2656 err
= __blk_mq_alloc_rq_maps(set
);
2660 set
->queue_depth
>>= 1;
2661 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2665 } while (set
->queue_depth
);
2667 if (!set
->queue_depth
|| err
) {
2668 pr_err("blk-mq: failed to allocate request map\n");
2672 if (depth
!= set
->queue_depth
)
2673 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2674 depth
, set
->queue_depth
);
2679 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2681 if (set
->ops
->map_queues
) {
2684 * transport .map_queues is usually done in the following
2687 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2688 * mask = get_cpu_mask(queue)
2689 * for_each_cpu(cpu, mask)
2690 * set->mq_map[cpu] = queue;
2693 * When we need to remap, the table has to be cleared for
2694 * killing stale mapping since one CPU may not be mapped
2697 for_each_possible_cpu(cpu
)
2698 set
->mq_map
[cpu
] = 0;
2700 return set
->ops
->map_queues(set
);
2702 return blk_mq_map_queues(set
);
2706 * Alloc a tag set to be associated with one or more request queues.
2707 * May fail with EINVAL for various error conditions. May adjust the
2708 * requested depth down, if if it too large. In that case, the set
2709 * value will be stored in set->queue_depth.
2711 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2715 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2717 if (!set
->nr_hw_queues
)
2719 if (!set
->queue_depth
)
2721 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2724 if (!set
->ops
->queue_rq
)
2727 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
2730 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2731 pr_info("blk-mq: reduced tag depth to %u\n",
2733 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2737 * If a crashdump is active, then we are potentially in a very
2738 * memory constrained environment. Limit us to 1 queue and
2739 * 64 tags to prevent using too much memory.
2741 if (is_kdump_kernel()) {
2742 set
->nr_hw_queues
= 1;
2743 set
->queue_depth
= min(64U, set
->queue_depth
);
2746 * There is no use for more h/w queues than cpus.
2748 if (set
->nr_hw_queues
> nr_cpu_ids
)
2749 set
->nr_hw_queues
= nr_cpu_ids
;
2751 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2752 GFP_KERNEL
, set
->numa_node
);
2757 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2758 GFP_KERNEL
, set
->numa_node
);
2762 ret
= blk_mq_update_queue_map(set
);
2764 goto out_free_mq_map
;
2766 ret
= blk_mq_alloc_rq_maps(set
);
2768 goto out_free_mq_map
;
2770 mutex_init(&set
->tag_list_lock
);
2771 INIT_LIST_HEAD(&set
->tag_list
);
2783 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2785 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2789 for (i
= 0; i
< nr_cpu_ids
; i
++)
2790 blk_mq_free_map_and_requests(set
, i
);
2798 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2800 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2802 struct blk_mq_tag_set
*set
= q
->tag_set
;
2803 struct blk_mq_hw_ctx
*hctx
;
2809 if (q
->nr_requests
== nr
)
2812 blk_mq_freeze_queue(q
);
2815 queue_for_each_hw_ctx(q
, hctx
, i
) {
2819 * If we're using an MQ scheduler, just update the scheduler
2820 * queue depth. This is similar to what the old code would do.
2822 if (!hctx
->sched_tags
) {
2823 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
2826 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2834 q
->nr_requests
= nr
;
2836 blk_mq_unfreeze_queue(q
);
2841 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
2844 struct request_queue
*q
;
2846 lockdep_assert_held(&set
->tag_list_lock
);
2848 if (nr_hw_queues
> nr_cpu_ids
)
2849 nr_hw_queues
= nr_cpu_ids
;
2850 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2853 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2854 blk_mq_freeze_queue(q
);
2856 set
->nr_hw_queues
= nr_hw_queues
;
2857 blk_mq_update_queue_map(set
);
2858 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2859 blk_mq_realloc_hw_ctxs(set
, q
);
2860 blk_mq_queue_reinit(q
);
2863 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2864 blk_mq_unfreeze_queue(q
);
2867 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2869 mutex_lock(&set
->tag_list_lock
);
2870 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
2871 mutex_unlock(&set
->tag_list_lock
);
2873 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2875 /* Enable polling stats and return whether they were already enabled. */
2876 static bool blk_poll_stats_enable(struct request_queue
*q
)
2878 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2879 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
2881 blk_stat_add_callback(q
, q
->poll_cb
);
2885 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2888 * We don't arm the callback if polling stats are not enabled or the
2889 * callback is already active.
2891 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2892 blk_stat_is_active(q
->poll_cb
))
2895 blk_stat_activate_msecs(q
->poll_cb
, 100);
2898 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2900 struct request_queue
*q
= cb
->data
;
2903 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
2904 if (cb
->stat
[bucket
].nr_samples
)
2905 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
2909 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2910 struct blk_mq_hw_ctx
*hctx
,
2913 unsigned long ret
= 0;
2917 * If stats collection isn't on, don't sleep but turn it on for
2920 if (!blk_poll_stats_enable(q
))
2924 * As an optimistic guess, use half of the mean service time
2925 * for this type of request. We can (and should) make this smarter.
2926 * For instance, if the completion latencies are tight, we can
2927 * get closer than just half the mean. This is especially
2928 * important on devices where the completion latencies are longer
2929 * than ~10 usec. We do use the stats for the relevant IO size
2930 * if available which does lead to better estimates.
2932 bucket
= blk_mq_poll_stats_bkt(rq
);
2936 if (q
->poll_stat
[bucket
].nr_samples
)
2937 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
2942 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2943 struct blk_mq_hw_ctx
*hctx
,
2946 struct hrtimer_sleeper hs
;
2947 enum hrtimer_mode mode
;
2951 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2957 * -1: don't ever hybrid sleep
2958 * 0: use half of prev avg
2959 * >0: use this specific value
2961 if (q
->poll_nsec
== -1)
2963 else if (q
->poll_nsec
> 0)
2964 nsecs
= q
->poll_nsec
;
2966 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2971 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2974 * This will be replaced with the stats tracking code, using
2975 * 'avg_completion_time / 2' as the pre-sleep target.
2979 mode
= HRTIMER_MODE_REL
;
2980 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2981 hrtimer_set_expires(&hs
.timer
, kt
);
2983 hrtimer_init_sleeper(&hs
, current
);
2985 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2987 set_current_state(TASK_UNINTERRUPTIBLE
);
2988 hrtimer_start_expires(&hs
.timer
, mode
);
2991 hrtimer_cancel(&hs
.timer
);
2992 mode
= HRTIMER_MODE_ABS
;
2993 } while (hs
.task
&& !signal_pending(current
));
2995 __set_current_state(TASK_RUNNING
);
2996 destroy_hrtimer_on_stack(&hs
.timer
);
3000 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
3002 struct request_queue
*q
= hctx
->queue
;
3006 * If we sleep, have the caller restart the poll loop to reset
3007 * the state. Like for the other success return cases, the
3008 * caller is responsible for checking if the IO completed. If
3009 * the IO isn't complete, we'll get called again and will go
3010 * straight to the busy poll loop.
3012 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
3015 hctx
->poll_considered
++;
3017 state
= current
->state
;
3018 while (!need_resched()) {
3021 hctx
->poll_invoked
++;
3023 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
3025 hctx
->poll_success
++;
3026 set_current_state(TASK_RUNNING
);
3030 if (signal_pending_state(state
, current
))
3031 set_current_state(TASK_RUNNING
);
3033 if (current
->state
== TASK_RUNNING
)
3043 static bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
3045 struct blk_mq_hw_ctx
*hctx
;
3048 if (!test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3051 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3052 if (!blk_qc_t_is_internal(cookie
))
3053 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3055 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3057 * With scheduling, if the request has completed, we'll
3058 * get a NULL return here, as we clear the sched tag when
3059 * that happens. The request still remains valid, like always,
3060 * so we should be safe with just the NULL check.
3066 return __blk_mq_poll(hctx
, rq
);
3069 static int __init
blk_mq_init(void)
3072 * See comment in block/blk.h rq_atomic_flags enum
3074 BUILD_BUG_ON((REQ_ATOM_STARTED
/ BITS_PER_BYTE
) !=
3075 (REQ_ATOM_COMPLETE
/ BITS_PER_BYTE
));
3077 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3078 blk_mq_hctx_notify_dead
);
3081 subsys_initcall(blk_mq_init
);