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 (blk_mq_rq_state(rq
) == MQ_RQ_IN_FLIGHT
) {
100 * index[0] counts the specific partition that was asked
101 * for. index[1] counts the ones that are active on the
102 * whole device, so increment that if mi->part is indeed
103 * a partition, and not a whole device.
105 if (rq
->part
== mi
->part
)
107 if (mi
->part
->partno
)
112 void blk_mq_in_flight(struct request_queue
*q
, struct hd_struct
*part
,
113 unsigned int inflight
[2])
115 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
117 inflight
[0] = inflight
[1] = 0;
118 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
121 void blk_freeze_queue_start(struct request_queue
*q
)
125 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
126 if (freeze_depth
== 1) {
127 percpu_ref_kill(&q
->q_usage_counter
);
129 blk_mq_run_hw_queues(q
, false);
132 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
134 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
136 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
138 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
140 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
141 unsigned long timeout
)
143 return wait_event_timeout(q
->mq_freeze_wq
,
144 percpu_ref_is_zero(&q
->q_usage_counter
),
147 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
150 * Guarantee no request is in use, so we can change any data structure of
151 * the queue afterward.
153 void blk_freeze_queue(struct request_queue
*q
)
156 * In the !blk_mq case we are only calling this to kill the
157 * q_usage_counter, otherwise this increases the freeze depth
158 * and waits for it to return to zero. For this reason there is
159 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
160 * exported to drivers as the only user for unfreeze is blk_mq.
162 blk_freeze_queue_start(q
);
163 blk_mq_freeze_queue_wait(q
);
166 void blk_mq_freeze_queue(struct request_queue
*q
)
169 * ...just an alias to keep freeze and unfreeze actions balanced
170 * in the blk_mq_* namespace
174 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
176 void blk_mq_unfreeze_queue(struct request_queue
*q
)
180 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
181 WARN_ON_ONCE(freeze_depth
< 0);
183 percpu_ref_reinit(&q
->q_usage_counter
);
184 wake_up_all(&q
->mq_freeze_wq
);
187 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
190 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
191 * mpt3sas driver such that this function can be removed.
193 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
197 spin_lock_irqsave(q
->queue_lock
, flags
);
198 queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
199 spin_unlock_irqrestore(q
->queue_lock
, flags
);
201 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
204 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
207 * Note: this function does not prevent that the struct request end_io()
208 * callback function is invoked. Once this function is returned, we make
209 * sure no dispatch can happen until the queue is unquiesced via
210 * blk_mq_unquiesce_queue().
212 void blk_mq_quiesce_queue(struct request_queue
*q
)
214 struct blk_mq_hw_ctx
*hctx
;
218 blk_mq_quiesce_queue_nowait(q
);
220 queue_for_each_hw_ctx(q
, hctx
, i
) {
221 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
222 synchronize_srcu(hctx
->queue_rq_srcu
);
229 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
232 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
235 * This function recovers queue into the state before quiescing
236 * which is done by blk_mq_quiesce_queue.
238 void blk_mq_unquiesce_queue(struct request_queue
*q
)
242 spin_lock_irqsave(q
->queue_lock
, flags
);
243 queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
244 spin_unlock_irqrestore(q
->queue_lock
, flags
);
246 /* dispatch requests which are inserted during quiescing */
247 blk_mq_run_hw_queues(q
, true);
249 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
251 void blk_mq_wake_waiters(struct request_queue
*q
)
253 struct blk_mq_hw_ctx
*hctx
;
256 queue_for_each_hw_ctx(q
, hctx
, i
)
257 if (blk_mq_hw_queue_mapped(hctx
))
258 blk_mq_tag_wakeup_all(hctx
->tags
, true);
261 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
263 return blk_mq_has_free_tags(hctx
->tags
);
265 EXPORT_SYMBOL(blk_mq_can_queue
);
267 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
268 unsigned int tag
, unsigned int op
)
270 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
271 struct request
*rq
= tags
->static_rqs
[tag
];
275 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
277 rq
->internal_tag
= tag
;
279 if (blk_mq_tag_busy(data
->hctx
)) {
280 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
281 atomic_inc(&data
->hctx
->nr_active
);
284 rq
->internal_tag
= -1;
285 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
288 INIT_LIST_HEAD(&rq
->queuelist
);
289 /* csd/requeue_work/fifo_time is initialized before use */
291 rq
->mq_ctx
= data
->ctx
;
293 if (data
->flags
& BLK_MQ_REQ_PREEMPT
)
294 rq
->rq_flags
|= RQF_PREEMPT
;
295 if (blk_queue_io_stat(data
->q
))
296 rq
->rq_flags
|= RQF_IO_STAT
;
297 /* do not touch atomic flags, it needs atomic ops against the timer */
299 INIT_HLIST_NODE(&rq
->hash
);
300 RB_CLEAR_NODE(&rq
->rb_node
);
303 rq
->start_time
= jiffies
;
304 #ifdef CONFIG_BLK_CGROUP
306 set_start_time_ns(rq
);
307 rq
->io_start_time_ns
= 0;
309 rq
->nr_phys_segments
= 0;
310 #if defined(CONFIG_BLK_DEV_INTEGRITY)
311 rq
->nr_integrity_segments
= 0;
314 /* tag was already set */
317 INIT_LIST_HEAD(&rq
->timeout_list
);
321 rq
->end_io_data
= NULL
;
324 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
328 static struct request
*blk_mq_get_request(struct request_queue
*q
,
329 struct bio
*bio
, unsigned int op
,
330 struct blk_mq_alloc_data
*data
)
332 struct elevator_queue
*e
= q
->elevator
;
335 bool put_ctx_on_error
= false;
337 blk_queue_enter_live(q
);
339 if (likely(!data
->ctx
)) {
340 data
->ctx
= blk_mq_get_ctx(q
);
341 put_ctx_on_error
= true;
343 if (likely(!data
->hctx
))
344 data
->hctx
= blk_mq_map_queue(q
, data
->ctx
->cpu
);
346 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
349 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
352 * Flush requests are special and go directly to the
355 if (!op_is_flush(op
) && e
->type
->ops
.mq
.limit_depth
)
356 e
->type
->ops
.mq
.limit_depth(op
, data
);
359 tag
= blk_mq_get_tag(data
);
360 if (tag
== BLK_MQ_TAG_FAIL
) {
361 if (put_ctx_on_error
) {
362 blk_mq_put_ctx(data
->ctx
);
369 rq
= blk_mq_rq_ctx_init(data
, tag
, op
);
370 if (!op_is_flush(op
)) {
372 if (e
&& e
->type
->ops
.mq
.prepare_request
) {
373 if (e
->type
->icq_cache
&& rq_ioc(bio
))
374 blk_mq_sched_assign_ioc(rq
, bio
);
376 e
->type
->ops
.mq
.prepare_request(rq
, bio
);
377 rq
->rq_flags
|= RQF_ELVPRIV
;
380 data
->hctx
->queued
++;
384 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
385 blk_mq_req_flags_t flags
)
387 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
391 ret
= blk_queue_enter(q
, flags
);
395 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
399 return ERR_PTR(-EWOULDBLOCK
);
401 blk_mq_put_ctx(alloc_data
.ctx
);
404 rq
->__sector
= (sector_t
) -1;
405 rq
->bio
= rq
->biotail
= NULL
;
408 EXPORT_SYMBOL(blk_mq_alloc_request
);
410 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
411 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
413 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
419 * If the tag allocator sleeps we could get an allocation for a
420 * different hardware context. No need to complicate the low level
421 * allocator for this for the rare use case of a command tied to
424 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
425 return ERR_PTR(-EINVAL
);
427 if (hctx_idx
>= q
->nr_hw_queues
)
428 return ERR_PTR(-EIO
);
430 ret
= blk_queue_enter(q
, flags
);
435 * Check if the hardware context is actually mapped to anything.
436 * If not tell the caller that it should skip this queue.
438 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
439 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
441 return ERR_PTR(-EXDEV
);
443 cpu
= cpumask_first(alloc_data
.hctx
->cpumask
);
444 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
446 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
450 return ERR_PTR(-EWOULDBLOCK
);
454 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
456 void blk_mq_free_request(struct request
*rq
)
458 struct request_queue
*q
= rq
->q
;
459 struct elevator_queue
*e
= q
->elevator
;
460 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
461 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
462 const int sched_tag
= rq
->internal_tag
;
464 if (rq
->rq_flags
& RQF_ELVPRIV
) {
465 if (e
&& e
->type
->ops
.mq
.finish_request
)
466 e
->type
->ops
.mq
.finish_request(rq
);
468 put_io_context(rq
->elv
.icq
->ioc
);
473 ctx
->rq_completed
[rq_is_sync(rq
)]++;
474 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
475 atomic_dec(&hctx
->nr_active
);
477 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
478 laptop_io_completion(q
->backing_dev_info
);
480 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
483 blk_put_rl(blk_rq_rl(rq
));
485 blk_mq_rq_update_state(rq
, MQ_RQ_IDLE
);
486 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
488 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
490 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
491 blk_mq_sched_restart(hctx
);
494 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
496 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
498 blk_account_io_done(rq
);
501 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
502 rq
->end_io(rq
, error
);
504 if (unlikely(blk_bidi_rq(rq
)))
505 blk_mq_free_request(rq
->next_rq
);
506 blk_mq_free_request(rq
);
509 EXPORT_SYMBOL(__blk_mq_end_request
);
511 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
513 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
515 __blk_mq_end_request(rq
, error
);
517 EXPORT_SYMBOL(blk_mq_end_request
);
519 static void __blk_mq_complete_request_remote(void *data
)
521 struct request
*rq
= data
;
523 rq
->q
->softirq_done_fn(rq
);
526 static void __blk_mq_complete_request(struct request
*rq
)
528 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
532 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
);
533 blk_mq_rq_update_state(rq
, MQ_RQ_COMPLETE
);
535 if (rq
->internal_tag
!= -1)
536 blk_mq_sched_completed_request(rq
);
537 if (rq
->rq_flags
& RQF_STATS
) {
538 blk_mq_poll_stats_start(rq
->q
);
542 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
543 rq
->q
->softirq_done_fn(rq
);
548 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
549 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
551 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
552 rq
->csd
.func
= __blk_mq_complete_request_remote
;
555 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
557 rq
->q
->softirq_done_fn(rq
);
562 static void hctx_unlock(struct blk_mq_hw_ctx
*hctx
, int srcu_idx
)
564 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
567 srcu_read_unlock(hctx
->queue_rq_srcu
, srcu_idx
);
570 static void hctx_lock(struct blk_mq_hw_ctx
*hctx
, int *srcu_idx
)
572 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
575 *srcu_idx
= srcu_read_lock(hctx
->queue_rq_srcu
);
578 static void blk_mq_rq_update_aborted_gstate(struct request
*rq
, u64 gstate
)
583 * blk_mq_rq_aborted_gstate() is used from the completion path and
584 * can thus be called from irq context. u64_stats_fetch in the
585 * middle of update on the same CPU leads to lockup. Disable irq
588 local_irq_save(flags
);
589 u64_stats_update_begin(&rq
->aborted_gstate_sync
);
590 rq
->aborted_gstate
= gstate
;
591 u64_stats_update_end(&rq
->aborted_gstate_sync
);
592 local_irq_restore(flags
);
595 static u64
blk_mq_rq_aborted_gstate(struct request
*rq
)
601 start
= u64_stats_fetch_begin(&rq
->aborted_gstate_sync
);
602 aborted_gstate
= rq
->aborted_gstate
;
603 } while (u64_stats_fetch_retry(&rq
->aborted_gstate_sync
, start
));
605 return aborted_gstate
;
609 * blk_mq_complete_request - end I/O on a request
610 * @rq: the request being processed
613 * Ends all I/O on a request. It does not handle partial completions.
614 * The actual completion happens out-of-order, through a IPI handler.
616 void blk_mq_complete_request(struct request
*rq
)
618 struct request_queue
*q
= rq
->q
;
619 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, rq
->mq_ctx
->cpu
);
622 if (unlikely(blk_should_fake_timeout(q
)))
626 * If @rq->aborted_gstate equals the current instance, timeout is
627 * claiming @rq and we lost. This is synchronized through
628 * hctx_lock(). See blk_mq_timeout_work() for details.
630 * Completion path never blocks and we can directly use RCU here
631 * instead of hctx_lock() which can be either RCU or SRCU.
632 * However, that would complicate paths which want to synchronize
633 * against us. Let stay in sync with the issue path so that
634 * hctx_lock() covers both issue and completion paths.
636 hctx_lock(hctx
, &srcu_idx
);
637 if (blk_mq_rq_aborted_gstate(rq
) != rq
->gstate
)
638 __blk_mq_complete_request(rq
);
639 hctx_unlock(hctx
, srcu_idx
);
641 EXPORT_SYMBOL(blk_mq_complete_request
);
643 int blk_mq_request_started(struct request
*rq
)
645 return blk_mq_rq_state(rq
) != MQ_RQ_IDLE
;
647 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
649 void blk_mq_start_request(struct request
*rq
)
651 struct request_queue
*q
= rq
->q
;
653 blk_mq_sched_started_request(rq
);
655 trace_block_rq_issue(q
, rq
);
657 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
658 blk_stat_set_issue(&rq
->issue_stat
, blk_rq_sectors(rq
));
659 rq
->rq_flags
|= RQF_STATS
;
660 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
663 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
666 * Mark @rq in-flight which also advances the generation number,
667 * and register for timeout. Protect with a seqcount to allow the
668 * timeout path to read both @rq->gstate and @rq->deadline
671 * This is the only place where a request is marked in-flight. If
672 * the timeout path reads an in-flight @rq->gstate, the
673 * @rq->deadline it reads together under @rq->gstate_seq is
674 * guaranteed to be the matching one.
677 write_seqcount_begin(&rq
->gstate_seq
);
679 blk_mq_rq_update_state(rq
, MQ_RQ_IN_FLIGHT
);
682 write_seqcount_end(&rq
->gstate_seq
);
685 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
687 * Make sure space for the drain appears. We know we can do
688 * this because max_hw_segments has been adjusted to be one
689 * fewer than the device can handle.
691 rq
->nr_phys_segments
++;
694 EXPORT_SYMBOL(blk_mq_start_request
);
697 * When we reach here because queue is busy, it's safe to change the state
698 * to IDLE without checking @rq->aborted_gstate because we should still be
699 * holding the RCU read lock and thus protected against timeout.
701 static void __blk_mq_requeue_request(struct request
*rq
)
703 struct request_queue
*q
= rq
->q
;
705 blk_mq_put_driver_tag(rq
);
707 trace_block_rq_requeue(q
, rq
);
708 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
709 blk_mq_sched_requeue_request(rq
);
711 if (blk_mq_rq_state(rq
) != MQ_RQ_IDLE
) {
712 blk_mq_rq_update_state(rq
, MQ_RQ_IDLE
);
713 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
714 rq
->nr_phys_segments
--;
718 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
720 __blk_mq_requeue_request(rq
);
722 BUG_ON(blk_queued_rq(rq
));
723 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
725 EXPORT_SYMBOL(blk_mq_requeue_request
);
727 static void blk_mq_requeue_work(struct work_struct
*work
)
729 struct request_queue
*q
=
730 container_of(work
, struct request_queue
, requeue_work
.work
);
732 struct request
*rq
, *next
;
734 spin_lock_irq(&q
->requeue_lock
);
735 list_splice_init(&q
->requeue_list
, &rq_list
);
736 spin_unlock_irq(&q
->requeue_lock
);
738 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
739 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
742 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
743 list_del_init(&rq
->queuelist
);
744 blk_mq_sched_insert_request(rq
, true, false, false, true);
747 while (!list_empty(&rq_list
)) {
748 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
749 list_del_init(&rq
->queuelist
);
750 blk_mq_sched_insert_request(rq
, false, false, false, true);
753 blk_mq_run_hw_queues(q
, false);
756 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
757 bool kick_requeue_list
)
759 struct request_queue
*q
= rq
->q
;
763 * We abuse this flag that is otherwise used by the I/O scheduler to
764 * request head insertion from the workqueue.
766 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
768 spin_lock_irqsave(&q
->requeue_lock
, flags
);
770 rq
->rq_flags
|= RQF_SOFTBARRIER
;
771 list_add(&rq
->queuelist
, &q
->requeue_list
);
773 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
775 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
777 if (kick_requeue_list
)
778 blk_mq_kick_requeue_list(q
);
780 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
782 void blk_mq_kick_requeue_list(struct request_queue
*q
)
784 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
786 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
788 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
791 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
792 msecs_to_jiffies(msecs
));
794 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
796 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
798 if (tag
< tags
->nr_tags
) {
799 prefetch(tags
->rqs
[tag
]);
800 return tags
->rqs
[tag
];
805 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
807 struct blk_mq_timeout_data
{
809 unsigned int next_set
;
810 unsigned int nr_expired
;
813 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
815 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
816 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
818 req
->rq_flags
|= RQF_MQ_TIMEOUT_EXPIRED
;
821 ret
= ops
->timeout(req
, reserved
);
825 __blk_mq_complete_request(req
);
827 case BLK_EH_RESET_TIMER
:
829 * As nothing prevents from completion happening while
830 * ->aborted_gstate is set, this may lead to ignored
831 * completions and further spurious timeouts.
833 blk_mq_rq_update_aborted_gstate(req
, 0);
836 case BLK_EH_NOT_HANDLED
:
839 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
844 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
845 struct request
*rq
, void *priv
, bool reserved
)
847 struct blk_mq_timeout_data
*data
= priv
;
848 unsigned long gstate
, deadline
;
853 if (rq
->rq_flags
& RQF_MQ_TIMEOUT_EXPIRED
)
856 /* read coherent snapshots of @rq->state_gen and @rq->deadline */
858 start
= read_seqcount_begin(&rq
->gstate_seq
);
859 gstate
= READ_ONCE(rq
->gstate
);
860 deadline
= rq
->deadline
;
861 if (!read_seqcount_retry(&rq
->gstate_seq
, start
))
866 /* if in-flight && overdue, mark for abortion */
867 if ((gstate
& MQ_RQ_STATE_MASK
) == MQ_RQ_IN_FLIGHT
&&
868 time_after_eq(jiffies
, deadline
)) {
869 blk_mq_rq_update_aborted_gstate(rq
, gstate
);
872 } else if (!data
->next_set
|| time_after(data
->next
, deadline
)) {
873 data
->next
= deadline
;
878 static void blk_mq_terminate_expired(struct blk_mq_hw_ctx
*hctx
,
879 struct request
*rq
, void *priv
, bool reserved
)
882 * We marked @rq->aborted_gstate and waited for RCU. If there were
883 * completions that we lost to, they would have finished and
884 * updated @rq->gstate by now; otherwise, the completion path is
885 * now guaranteed to see @rq->aborted_gstate and yield. If
886 * @rq->aborted_gstate still matches @rq->gstate, @rq is ours.
888 if (!(rq
->rq_flags
& RQF_MQ_TIMEOUT_EXPIRED
) &&
889 READ_ONCE(rq
->gstate
) == rq
->aborted_gstate
)
890 blk_mq_rq_timed_out(rq
, reserved
);
893 static void blk_mq_timeout_work(struct work_struct
*work
)
895 struct request_queue
*q
=
896 container_of(work
, struct request_queue
, timeout_work
);
897 struct blk_mq_timeout_data data
= {
902 struct blk_mq_hw_ctx
*hctx
;
905 /* A deadlock might occur if a request is stuck requiring a
906 * timeout at the same time a queue freeze is waiting
907 * completion, since the timeout code would not be able to
908 * acquire the queue reference here.
910 * That's why we don't use blk_queue_enter here; instead, we use
911 * percpu_ref_tryget directly, because we need to be able to
912 * obtain a reference even in the short window between the queue
913 * starting to freeze, by dropping the first reference in
914 * blk_freeze_queue_start, and the moment the last request is
915 * consumed, marked by the instant q_usage_counter reaches
918 if (!percpu_ref_tryget(&q
->q_usage_counter
))
921 /* scan for the expired ones and set their ->aborted_gstate */
922 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
924 if (data
.nr_expired
) {
925 bool has_rcu
= false;
928 * Wait till everyone sees ->aborted_gstate. The
929 * sequential waits for SRCUs aren't ideal. If this ever
930 * becomes a problem, we can add per-hw_ctx rcu_head and
933 queue_for_each_hw_ctx(q
, hctx
, i
) {
934 if (!hctx
->nr_expired
)
937 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
940 synchronize_srcu(hctx
->queue_rq_srcu
);
942 hctx
->nr_expired
= 0;
947 /* terminate the ones we won */
948 blk_mq_queue_tag_busy_iter(q
, blk_mq_terminate_expired
, NULL
);
952 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
953 mod_timer(&q
->timeout
, data
.next
);
955 queue_for_each_hw_ctx(q
, hctx
, i
) {
956 /* the hctx may be unmapped, so check it here */
957 if (blk_mq_hw_queue_mapped(hctx
))
958 blk_mq_tag_idle(hctx
);
964 struct flush_busy_ctx_data
{
965 struct blk_mq_hw_ctx
*hctx
;
966 struct list_head
*list
;
969 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
971 struct flush_busy_ctx_data
*flush_data
= data
;
972 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
973 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
975 sbitmap_clear_bit(sb
, bitnr
);
976 spin_lock(&ctx
->lock
);
977 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
978 spin_unlock(&ctx
->lock
);
983 * Process software queues that have been marked busy, splicing them
984 * to the for-dispatch
986 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
988 struct flush_busy_ctx_data data
= {
993 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
995 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
997 struct dispatch_rq_data
{
998 struct blk_mq_hw_ctx
*hctx
;
1002 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
1005 struct dispatch_rq_data
*dispatch_data
= data
;
1006 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
1007 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1009 spin_lock(&ctx
->lock
);
1010 if (unlikely(!list_empty(&ctx
->rq_list
))) {
1011 dispatch_data
->rq
= list_entry_rq(ctx
->rq_list
.next
);
1012 list_del_init(&dispatch_data
->rq
->queuelist
);
1013 if (list_empty(&ctx
->rq_list
))
1014 sbitmap_clear_bit(sb
, bitnr
);
1016 spin_unlock(&ctx
->lock
);
1018 return !dispatch_data
->rq
;
1021 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1022 struct blk_mq_ctx
*start
)
1024 unsigned off
= start
? start
->index_hw
: 0;
1025 struct dispatch_rq_data data
= {
1030 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1031 dispatch_rq_from_ctx
, &data
);
1036 static inline unsigned int queued_to_index(unsigned int queued
)
1041 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
1044 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
1047 struct blk_mq_alloc_data data
= {
1049 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
1050 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
1053 might_sleep_if(wait
);
1058 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
1059 data
.flags
|= BLK_MQ_REQ_RESERVED
;
1061 rq
->tag
= blk_mq_get_tag(&data
);
1063 if (blk_mq_tag_busy(data
.hctx
)) {
1064 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1065 atomic_inc(&data
.hctx
->nr_active
);
1067 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
1073 return rq
->tag
!= -1;
1076 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1077 int flags
, void *key
)
1079 struct blk_mq_hw_ctx
*hctx
;
1081 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1083 list_del_init(&wait
->entry
);
1084 blk_mq_run_hw_queue(hctx
, true);
1089 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1090 * the tag wakeups. For non-shared tags, we can simply mark us nedeing a
1091 * restart. For both caes, take care to check the condition again after
1092 * marking us as waiting.
1094 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
**hctx
,
1097 struct blk_mq_hw_ctx
*this_hctx
= *hctx
;
1098 bool shared_tags
= (this_hctx
->flags
& BLK_MQ_F_TAG_SHARED
) != 0;
1099 struct sbq_wait_state
*ws
;
1100 wait_queue_entry_t
*wait
;
1104 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &this_hctx
->state
))
1105 set_bit(BLK_MQ_S_SCHED_RESTART
, &this_hctx
->state
);
1107 wait
= &this_hctx
->dispatch_wait
;
1108 if (!list_empty_careful(&wait
->entry
))
1111 spin_lock(&this_hctx
->lock
);
1112 if (!list_empty(&wait
->entry
)) {
1113 spin_unlock(&this_hctx
->lock
);
1117 ws
= bt_wait_ptr(&this_hctx
->tags
->bitmap_tags
, this_hctx
);
1118 add_wait_queue(&ws
->wait
, wait
);
1122 * It's possible that a tag was freed in the window between the
1123 * allocation failure and adding the hardware queue to the wait
1126 ret
= blk_mq_get_driver_tag(rq
, hctx
, false);
1130 * Don't clear RESTART here, someone else could have set it.
1131 * At most this will cost an extra queue run.
1136 spin_unlock(&this_hctx
->lock
);
1141 * We got a tag, remove ourselves from the wait queue to ensure
1142 * someone else gets the wakeup.
1144 spin_lock_irq(&ws
->wait
.lock
);
1145 list_del_init(&wait
->entry
);
1146 spin_unlock_irq(&ws
->wait
.lock
);
1147 spin_unlock(&this_hctx
->lock
);
1152 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
,
1155 struct blk_mq_hw_ctx
*hctx
;
1156 struct request
*rq
, *nxt
;
1157 bool no_tag
= false;
1160 if (list_empty(list
))
1163 WARN_ON(!list_is_singular(list
) && got_budget
);
1166 * Now process all the entries, sending them to the driver.
1168 errors
= queued
= 0;
1170 struct blk_mq_queue_data bd
;
1173 rq
= list_first_entry(list
, struct request
, queuelist
);
1174 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
1176 * The initial allocation attempt failed, so we need to
1177 * rerun the hardware queue when a tag is freed. The
1178 * waitqueue takes care of that. If the queue is run
1179 * before we add this entry back on the dispatch list,
1180 * we'll re-run it below.
1182 if (!blk_mq_mark_tag_wait(&hctx
, rq
)) {
1184 blk_mq_put_dispatch_budget(hctx
);
1186 * For non-shared tags, the RESTART check
1189 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
)
1195 if (!got_budget
&& !blk_mq_get_dispatch_budget(hctx
)) {
1196 blk_mq_put_driver_tag(rq
);
1200 list_del_init(&rq
->queuelist
);
1205 * Flag last if we have no more requests, or if we have more
1206 * but can't assign a driver tag to it.
1208 if (list_empty(list
))
1211 nxt
= list_first_entry(list
, struct request
, queuelist
);
1212 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1215 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1216 if (ret
== BLK_STS_RESOURCE
) {
1218 * If an I/O scheduler has been configured and we got a
1219 * driver tag for the next request already, free it
1222 if (!list_empty(list
)) {
1223 nxt
= list_first_entry(list
, struct request
, queuelist
);
1224 blk_mq_put_driver_tag(nxt
);
1226 list_add(&rq
->queuelist
, list
);
1227 __blk_mq_requeue_request(rq
);
1231 if (unlikely(ret
!= BLK_STS_OK
)) {
1233 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1238 } while (!list_empty(list
));
1240 hctx
->dispatched
[queued_to_index(queued
)]++;
1243 * Any items that need requeuing? Stuff them into hctx->dispatch,
1244 * that is where we will continue on next queue run.
1246 if (!list_empty(list
)) {
1247 spin_lock(&hctx
->lock
);
1248 list_splice_init(list
, &hctx
->dispatch
);
1249 spin_unlock(&hctx
->lock
);
1252 * If SCHED_RESTART was set by the caller of this function and
1253 * it is no longer set that means that it was cleared by another
1254 * thread and hence that a queue rerun is needed.
1256 * If 'no_tag' is set, that means that we failed getting
1257 * a driver tag with an I/O scheduler attached. If our dispatch
1258 * waitqueue is no longer active, ensure that we run the queue
1259 * AFTER adding our entries back to the list.
1261 * If no I/O scheduler has been configured it is possible that
1262 * the hardware queue got stopped and restarted before requests
1263 * were pushed back onto the dispatch list. Rerun the queue to
1264 * avoid starvation. Notes:
1265 * - blk_mq_run_hw_queue() checks whether or not a queue has
1266 * been stopped before rerunning a queue.
1267 * - Some but not all block drivers stop a queue before
1268 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1271 if (!blk_mq_sched_needs_restart(hctx
) ||
1272 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1273 blk_mq_run_hw_queue(hctx
, true);
1276 return (queued
+ errors
) != 0;
1279 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1284 * We should be running this queue from one of the CPUs that
1287 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1288 cpu_online(hctx
->next_cpu
));
1291 * We can't run the queue inline with ints disabled. Ensure that
1292 * we catch bad users of this early.
1294 WARN_ON_ONCE(in_interrupt());
1296 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1298 hctx_lock(hctx
, &srcu_idx
);
1299 blk_mq_sched_dispatch_requests(hctx
);
1300 hctx_unlock(hctx
, srcu_idx
);
1304 * It'd be great if the workqueue API had a way to pass
1305 * in a mask and had some smarts for more clever placement.
1306 * For now we just round-robin here, switching for every
1307 * BLK_MQ_CPU_WORK_BATCH queued items.
1309 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1311 if (hctx
->queue
->nr_hw_queues
== 1)
1312 return WORK_CPU_UNBOUND
;
1314 if (--hctx
->next_cpu_batch
<= 0) {
1317 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1318 if (next_cpu
>= nr_cpu_ids
)
1319 next_cpu
= cpumask_first(hctx
->cpumask
);
1321 hctx
->next_cpu
= next_cpu
;
1322 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1325 return hctx
->next_cpu
;
1328 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1329 unsigned long msecs
)
1331 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx
)))
1334 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1337 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1338 int cpu
= get_cpu();
1339 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1340 __blk_mq_run_hw_queue(hctx
);
1348 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1350 msecs_to_jiffies(msecs
));
1353 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1355 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1357 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1359 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1365 * When queue is quiesced, we may be switching io scheduler, or
1366 * updating nr_hw_queues, or other things, and we can't run queue
1367 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1369 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1372 hctx_lock(hctx
, &srcu_idx
);
1373 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1374 blk_mq_hctx_has_pending(hctx
);
1375 hctx_unlock(hctx
, srcu_idx
);
1378 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1384 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1386 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1388 struct blk_mq_hw_ctx
*hctx
;
1391 queue_for_each_hw_ctx(q
, hctx
, i
) {
1392 if (blk_mq_hctx_stopped(hctx
))
1395 blk_mq_run_hw_queue(hctx
, async
);
1398 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1401 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1402 * @q: request queue.
1404 * The caller is responsible for serializing this function against
1405 * blk_mq_{start,stop}_hw_queue().
1407 bool blk_mq_queue_stopped(struct request_queue
*q
)
1409 struct blk_mq_hw_ctx
*hctx
;
1412 queue_for_each_hw_ctx(q
, hctx
, i
)
1413 if (blk_mq_hctx_stopped(hctx
))
1418 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1421 * This function is often used for pausing .queue_rq() by driver when
1422 * there isn't enough resource or some conditions aren't satisfied, and
1423 * BLK_STS_RESOURCE is usually returned.
1425 * We do not guarantee that dispatch can be drained or blocked
1426 * after blk_mq_stop_hw_queue() returns. Please use
1427 * blk_mq_quiesce_queue() for that requirement.
1429 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1431 cancel_delayed_work(&hctx
->run_work
);
1433 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1435 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1438 * This function is often used for pausing .queue_rq() by driver when
1439 * there isn't enough resource or some conditions aren't satisfied, and
1440 * BLK_STS_RESOURCE is usually returned.
1442 * We do not guarantee that dispatch can be drained or blocked
1443 * after blk_mq_stop_hw_queues() returns. Please use
1444 * blk_mq_quiesce_queue() for that requirement.
1446 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1448 struct blk_mq_hw_ctx
*hctx
;
1451 queue_for_each_hw_ctx(q
, hctx
, i
)
1452 blk_mq_stop_hw_queue(hctx
);
1454 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1456 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1458 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1460 blk_mq_run_hw_queue(hctx
, false);
1462 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1464 void blk_mq_start_hw_queues(struct request_queue
*q
)
1466 struct blk_mq_hw_ctx
*hctx
;
1469 queue_for_each_hw_ctx(q
, hctx
, i
)
1470 blk_mq_start_hw_queue(hctx
);
1472 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1474 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1476 if (!blk_mq_hctx_stopped(hctx
))
1479 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1480 blk_mq_run_hw_queue(hctx
, async
);
1482 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1484 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1486 struct blk_mq_hw_ctx
*hctx
;
1489 queue_for_each_hw_ctx(q
, hctx
, i
)
1490 blk_mq_start_stopped_hw_queue(hctx
, async
);
1492 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1494 static void blk_mq_run_work_fn(struct work_struct
*work
)
1496 struct blk_mq_hw_ctx
*hctx
;
1498 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1501 * If we are stopped, don't run the queue. The exception is if
1502 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1503 * the STOPPED bit and run it.
1505 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)) {
1506 if (!test_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
))
1509 clear_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1510 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1513 __blk_mq_run_hw_queue(hctx
);
1517 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1519 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx
)))
1523 * Stop the hw queue, then modify currently delayed work.
1524 * This should prevent us from running the queue prematurely.
1525 * Mark the queue as auto-clearing STOPPED when it runs.
1527 blk_mq_stop_hw_queue(hctx
);
1528 set_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1529 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1531 msecs_to_jiffies(msecs
));
1533 EXPORT_SYMBOL(blk_mq_delay_queue
);
1535 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1539 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1541 lockdep_assert_held(&ctx
->lock
);
1543 trace_block_rq_insert(hctx
->queue
, rq
);
1546 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1548 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1551 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1554 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1556 lockdep_assert_held(&ctx
->lock
);
1558 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1559 blk_mq_hctx_mark_pending(hctx
, ctx
);
1563 * Should only be used carefully, when the caller knows we want to
1564 * bypass a potential IO scheduler on the target device.
1566 void blk_mq_request_bypass_insert(struct request
*rq
, bool run_queue
)
1568 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1569 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(rq
->q
, ctx
->cpu
);
1571 spin_lock(&hctx
->lock
);
1572 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1573 spin_unlock(&hctx
->lock
);
1576 blk_mq_run_hw_queue(hctx
, false);
1579 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1580 struct list_head
*list
)
1584 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1587 spin_lock(&ctx
->lock
);
1588 while (!list_empty(list
)) {
1591 rq
= list_first_entry(list
, struct request
, queuelist
);
1592 BUG_ON(rq
->mq_ctx
!= ctx
);
1593 list_del_init(&rq
->queuelist
);
1594 __blk_mq_insert_req_list(hctx
, rq
, false);
1596 blk_mq_hctx_mark_pending(hctx
, ctx
);
1597 spin_unlock(&ctx
->lock
);
1600 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1602 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1603 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1605 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1606 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1607 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1610 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1612 struct blk_mq_ctx
*this_ctx
;
1613 struct request_queue
*this_q
;
1616 LIST_HEAD(ctx_list
);
1619 list_splice_init(&plug
->mq_list
, &list
);
1621 list_sort(NULL
, &list
, plug_ctx_cmp
);
1627 while (!list_empty(&list
)) {
1628 rq
= list_entry_rq(list
.next
);
1629 list_del_init(&rq
->queuelist
);
1631 if (rq
->mq_ctx
!= this_ctx
) {
1633 trace_block_unplug(this_q
, depth
, from_schedule
);
1634 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1639 this_ctx
= rq
->mq_ctx
;
1645 list_add_tail(&rq
->queuelist
, &ctx_list
);
1649 * If 'this_ctx' is set, we know we have entries to complete
1650 * on 'ctx_list'. Do those.
1653 trace_block_unplug(this_q
, depth
, from_schedule
);
1654 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1659 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1661 blk_init_request_from_bio(rq
, bio
);
1663 blk_rq_set_rl(rq
, blk_get_rl(rq
->q
, bio
));
1665 blk_account_io_start(rq
, true);
1668 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx
*hctx
,
1669 struct blk_mq_ctx
*ctx
,
1672 spin_lock(&ctx
->lock
);
1673 __blk_mq_insert_request(hctx
, rq
, false);
1674 spin_unlock(&ctx
->lock
);
1677 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1680 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1682 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1685 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1689 struct request_queue
*q
= rq
->q
;
1690 struct blk_mq_queue_data bd
= {
1694 blk_qc_t new_cookie
;
1696 bool run_queue
= true;
1698 /* RCU or SRCU read lock is needed before checking quiesced flag */
1699 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1707 if (!blk_mq_get_driver_tag(rq
, NULL
, false))
1710 if (!blk_mq_get_dispatch_budget(hctx
)) {
1711 blk_mq_put_driver_tag(rq
);
1715 new_cookie
= request_to_qc_t(hctx
, rq
);
1718 * For OK queue, we are done. For error, kill it. Any other
1719 * error (busy), just add it to our list as we previously
1722 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1725 *cookie
= new_cookie
;
1727 case BLK_STS_RESOURCE
:
1728 __blk_mq_requeue_request(rq
);
1731 *cookie
= BLK_QC_T_NONE
;
1732 blk_mq_end_request(rq
, ret
);
1737 blk_mq_sched_insert_request(rq
, false, run_queue
, false,
1738 hctx
->flags
& BLK_MQ_F_BLOCKING
);
1741 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1742 struct request
*rq
, blk_qc_t
*cookie
)
1746 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1748 hctx_lock(hctx
, &srcu_idx
);
1749 __blk_mq_try_issue_directly(hctx
, rq
, cookie
);
1750 hctx_unlock(hctx
, srcu_idx
);
1753 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1755 const int is_sync
= op_is_sync(bio
->bi_opf
);
1756 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1757 struct blk_mq_alloc_data data
= { .flags
= 0 };
1759 unsigned int request_count
= 0;
1760 struct blk_plug
*plug
;
1761 struct request
*same_queue_rq
= NULL
;
1763 unsigned int wb_acct
;
1765 blk_queue_bounce(q
, &bio
);
1767 blk_queue_split(q
, &bio
);
1769 if (!bio_integrity_prep(bio
))
1770 return BLK_QC_T_NONE
;
1772 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1773 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1774 return BLK_QC_T_NONE
;
1776 if (blk_mq_sched_bio_merge(q
, bio
))
1777 return BLK_QC_T_NONE
;
1779 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1781 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1783 rq
= blk_mq_get_request(q
, bio
, bio
->bi_opf
, &data
);
1784 if (unlikely(!rq
)) {
1785 __wbt_done(q
->rq_wb
, wb_acct
);
1786 if (bio
->bi_opf
& REQ_NOWAIT
)
1787 bio_wouldblock_error(bio
);
1788 return BLK_QC_T_NONE
;
1791 wbt_track(&rq
->issue_stat
, wb_acct
);
1793 cookie
= request_to_qc_t(data
.hctx
, rq
);
1795 plug
= current
->plug
;
1796 if (unlikely(is_flush_fua
)) {
1797 blk_mq_put_ctx(data
.ctx
);
1798 blk_mq_bio_to_request(rq
, bio
);
1800 /* bypass scheduler for flush rq */
1801 blk_insert_flush(rq
);
1802 blk_mq_run_hw_queue(data
.hctx
, true);
1803 } else if (plug
&& q
->nr_hw_queues
== 1) {
1804 struct request
*last
= NULL
;
1806 blk_mq_put_ctx(data
.ctx
);
1807 blk_mq_bio_to_request(rq
, bio
);
1810 * @request_count may become stale because of schedule
1811 * out, so check the list again.
1813 if (list_empty(&plug
->mq_list
))
1815 else if (blk_queue_nomerges(q
))
1816 request_count
= blk_plug_queued_count(q
);
1819 trace_block_plug(q
);
1821 last
= list_entry_rq(plug
->mq_list
.prev
);
1823 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1824 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1825 blk_flush_plug_list(plug
, false);
1826 trace_block_plug(q
);
1829 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1830 } else if (plug
&& !blk_queue_nomerges(q
)) {
1831 blk_mq_bio_to_request(rq
, bio
);
1834 * We do limited plugging. If the bio can be merged, do that.
1835 * Otherwise the existing request in the plug list will be
1836 * issued. So the plug list will have one request at most
1837 * The plug list might get flushed before this. If that happens,
1838 * the plug list is empty, and same_queue_rq is invalid.
1840 if (list_empty(&plug
->mq_list
))
1841 same_queue_rq
= NULL
;
1843 list_del_init(&same_queue_rq
->queuelist
);
1844 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1846 blk_mq_put_ctx(data
.ctx
);
1848 if (same_queue_rq
) {
1849 data
.hctx
= blk_mq_map_queue(q
,
1850 same_queue_rq
->mq_ctx
->cpu
);
1851 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1854 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1855 blk_mq_put_ctx(data
.ctx
);
1856 blk_mq_bio_to_request(rq
, bio
);
1857 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1858 } else if (q
->elevator
) {
1859 blk_mq_put_ctx(data
.ctx
);
1860 blk_mq_bio_to_request(rq
, bio
);
1861 blk_mq_sched_insert_request(rq
, false, true, true, true);
1863 blk_mq_put_ctx(data
.ctx
);
1864 blk_mq_bio_to_request(rq
, bio
);
1865 blk_mq_queue_io(data
.hctx
, data
.ctx
, rq
);
1866 blk_mq_run_hw_queue(data
.hctx
, true);
1872 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1873 unsigned int hctx_idx
)
1877 if (tags
->rqs
&& set
->ops
->exit_request
) {
1880 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1881 struct request
*rq
= tags
->static_rqs
[i
];
1885 set
->ops
->exit_request(set
, rq
, hctx_idx
);
1886 tags
->static_rqs
[i
] = NULL
;
1890 while (!list_empty(&tags
->page_list
)) {
1891 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1892 list_del_init(&page
->lru
);
1894 * Remove kmemleak object previously allocated in
1895 * blk_mq_init_rq_map().
1897 kmemleak_free(page_address(page
));
1898 __free_pages(page
, page
->private);
1902 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1906 kfree(tags
->static_rqs
);
1907 tags
->static_rqs
= NULL
;
1909 blk_mq_free_tags(tags
);
1912 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1913 unsigned int hctx_idx
,
1914 unsigned int nr_tags
,
1915 unsigned int reserved_tags
)
1917 struct blk_mq_tags
*tags
;
1920 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1921 if (node
== NUMA_NO_NODE
)
1922 node
= set
->numa_node
;
1924 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1925 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1929 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1930 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1933 blk_mq_free_tags(tags
);
1937 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1938 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1940 if (!tags
->static_rqs
) {
1942 blk_mq_free_tags(tags
);
1949 static size_t order_to_size(unsigned int order
)
1951 return (size_t)PAGE_SIZE
<< order
;
1954 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
1955 unsigned int hctx_idx
, int node
)
1959 if (set
->ops
->init_request
) {
1960 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
1965 seqcount_init(&rq
->gstate_seq
);
1966 u64_stats_init(&rq
->aborted_gstate_sync
);
1970 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1971 unsigned int hctx_idx
, unsigned int depth
)
1973 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1974 size_t rq_size
, left
;
1977 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1978 if (node
== NUMA_NO_NODE
)
1979 node
= set
->numa_node
;
1981 INIT_LIST_HEAD(&tags
->page_list
);
1984 * rq_size is the size of the request plus driver payload, rounded
1985 * to the cacheline size
1987 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1989 left
= rq_size
* depth
;
1991 for (i
= 0; i
< depth
; ) {
1992 int this_order
= max_order
;
1997 while (this_order
&& left
< order_to_size(this_order
- 1))
2001 page
= alloc_pages_node(node
,
2002 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
2008 if (order_to_size(this_order
) < rq_size
)
2015 page
->private = this_order
;
2016 list_add_tail(&page
->lru
, &tags
->page_list
);
2018 p
= page_address(page
);
2020 * Allow kmemleak to scan these pages as they contain pointers
2021 * to additional allocations like via ops->init_request().
2023 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2024 entries_per_page
= order_to_size(this_order
) / rq_size
;
2025 to_do
= min(entries_per_page
, depth
- i
);
2026 left
-= to_do
* rq_size
;
2027 for (j
= 0; j
< to_do
; j
++) {
2028 struct request
*rq
= p
;
2030 tags
->static_rqs
[i
] = rq
;
2031 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
2032 tags
->static_rqs
[i
] = NULL
;
2043 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2048 * 'cpu' is going away. splice any existing rq_list entries from this
2049 * software queue to the hw queue dispatch list, and ensure that it
2052 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2054 struct blk_mq_hw_ctx
*hctx
;
2055 struct blk_mq_ctx
*ctx
;
2058 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2059 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2061 spin_lock(&ctx
->lock
);
2062 if (!list_empty(&ctx
->rq_list
)) {
2063 list_splice_init(&ctx
->rq_list
, &tmp
);
2064 blk_mq_hctx_clear_pending(hctx
, ctx
);
2066 spin_unlock(&ctx
->lock
);
2068 if (list_empty(&tmp
))
2071 spin_lock(&hctx
->lock
);
2072 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2073 spin_unlock(&hctx
->lock
);
2075 blk_mq_run_hw_queue(hctx
, true);
2079 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2081 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2085 /* hctx->ctxs will be freed in queue's release handler */
2086 static void blk_mq_exit_hctx(struct request_queue
*q
,
2087 struct blk_mq_tag_set
*set
,
2088 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2090 blk_mq_debugfs_unregister_hctx(hctx
);
2092 if (blk_mq_hw_queue_mapped(hctx
))
2093 blk_mq_tag_idle(hctx
);
2095 if (set
->ops
->exit_request
)
2096 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2098 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2100 if (set
->ops
->exit_hctx
)
2101 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2103 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2104 cleanup_srcu_struct(hctx
->queue_rq_srcu
);
2106 blk_mq_remove_cpuhp(hctx
);
2107 blk_free_flush_queue(hctx
->fq
);
2108 sbitmap_free(&hctx
->ctx_map
);
2111 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2112 struct blk_mq_tag_set
*set
, int nr_queue
)
2114 struct blk_mq_hw_ctx
*hctx
;
2117 queue_for_each_hw_ctx(q
, hctx
, i
) {
2120 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2124 static int blk_mq_init_hctx(struct request_queue
*q
,
2125 struct blk_mq_tag_set
*set
,
2126 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2130 node
= hctx
->numa_node
;
2131 if (node
== NUMA_NO_NODE
)
2132 node
= hctx
->numa_node
= set
->numa_node
;
2134 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2135 spin_lock_init(&hctx
->lock
);
2136 INIT_LIST_HEAD(&hctx
->dispatch
);
2138 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
2140 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2142 hctx
->tags
= set
->tags
[hctx_idx
];
2145 * Allocate space for all possible cpus to avoid allocation at
2148 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2151 goto unregister_cpu_notifier
;
2153 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
2159 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2160 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2162 if (set
->ops
->init_hctx
&&
2163 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2166 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
2169 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
2171 goto sched_exit_hctx
;
2173 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
, node
))
2176 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2177 init_srcu_struct(hctx
->queue_rq_srcu
);
2179 blk_mq_debugfs_register_hctx(q
, hctx
);
2186 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2188 if (set
->ops
->exit_hctx
)
2189 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2191 sbitmap_free(&hctx
->ctx_map
);
2194 unregister_cpu_notifier
:
2195 blk_mq_remove_cpuhp(hctx
);
2199 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2200 unsigned int nr_hw_queues
)
2204 for_each_possible_cpu(i
) {
2205 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2206 struct blk_mq_hw_ctx
*hctx
;
2209 spin_lock_init(&__ctx
->lock
);
2210 INIT_LIST_HEAD(&__ctx
->rq_list
);
2213 /* If the cpu isn't present, the cpu is mapped to first hctx */
2214 if (!cpu_present(i
))
2217 hctx
= blk_mq_map_queue(q
, i
);
2220 * Set local node, IFF we have more than one hw queue. If
2221 * not, we remain on the home node of the device
2223 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2224 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2228 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2232 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2233 set
->queue_depth
, set
->reserved_tags
);
2234 if (!set
->tags
[hctx_idx
])
2237 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2242 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2243 set
->tags
[hctx_idx
] = NULL
;
2247 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2248 unsigned int hctx_idx
)
2250 if (set
->tags
[hctx_idx
]) {
2251 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2252 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2253 set
->tags
[hctx_idx
] = NULL
;
2257 static void blk_mq_map_swqueue(struct request_queue
*q
)
2259 unsigned int i
, hctx_idx
;
2260 struct blk_mq_hw_ctx
*hctx
;
2261 struct blk_mq_ctx
*ctx
;
2262 struct blk_mq_tag_set
*set
= q
->tag_set
;
2265 * Avoid others reading imcomplete hctx->cpumask through sysfs
2267 mutex_lock(&q
->sysfs_lock
);
2269 queue_for_each_hw_ctx(q
, hctx
, i
) {
2270 cpumask_clear(hctx
->cpumask
);
2275 * Map software to hardware queues.
2277 * If the cpu isn't present, the cpu is mapped to first hctx.
2279 for_each_present_cpu(i
) {
2280 hctx_idx
= q
->mq_map
[i
];
2281 /* unmapped hw queue can be remapped after CPU topo changed */
2282 if (!set
->tags
[hctx_idx
] &&
2283 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2285 * If tags initialization fail for some hctx,
2286 * that hctx won't be brought online. In this
2287 * case, remap the current ctx to hctx[0] which
2288 * is guaranteed to always have tags allocated
2293 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2294 hctx
= blk_mq_map_queue(q
, i
);
2296 cpumask_set_cpu(i
, hctx
->cpumask
);
2297 ctx
->index_hw
= hctx
->nr_ctx
;
2298 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2301 mutex_unlock(&q
->sysfs_lock
);
2303 queue_for_each_hw_ctx(q
, hctx
, i
) {
2305 * If no software queues are mapped to this hardware queue,
2306 * disable it and free the request entries.
2308 if (!hctx
->nr_ctx
) {
2309 /* Never unmap queue 0. We need it as a
2310 * fallback in case of a new remap fails
2313 if (i
&& set
->tags
[i
])
2314 blk_mq_free_map_and_requests(set
, i
);
2320 hctx
->tags
= set
->tags
[i
];
2321 WARN_ON(!hctx
->tags
);
2324 * Set the map size to the number of mapped software queues.
2325 * This is more accurate and more efficient than looping
2326 * over all possibly mapped software queues.
2328 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2331 * Initialize batch roundrobin counts
2333 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2334 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2339 * Caller needs to ensure that we're either frozen/quiesced, or that
2340 * the queue isn't live yet.
2342 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2344 struct blk_mq_hw_ctx
*hctx
;
2347 queue_for_each_hw_ctx(q
, hctx
, i
) {
2349 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2350 atomic_inc(&q
->shared_hctx_restart
);
2351 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2353 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2354 atomic_dec(&q
->shared_hctx_restart
);
2355 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2360 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2363 struct request_queue
*q
;
2365 lockdep_assert_held(&set
->tag_list_lock
);
2367 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2368 blk_mq_freeze_queue(q
);
2369 queue_set_hctx_shared(q
, shared
);
2370 blk_mq_unfreeze_queue(q
);
2374 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2376 struct blk_mq_tag_set
*set
= q
->tag_set
;
2378 mutex_lock(&set
->tag_list_lock
);
2379 list_del_rcu(&q
->tag_set_list
);
2380 INIT_LIST_HEAD(&q
->tag_set_list
);
2381 if (list_is_singular(&set
->tag_list
)) {
2382 /* just transitioned to unshared */
2383 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2384 /* update existing queue */
2385 blk_mq_update_tag_set_depth(set
, false);
2387 mutex_unlock(&set
->tag_list_lock
);
2392 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2393 struct request_queue
*q
)
2397 mutex_lock(&set
->tag_list_lock
);
2400 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2402 if (!list_empty(&set
->tag_list
) &&
2403 !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2404 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2405 /* update existing queue */
2406 blk_mq_update_tag_set_depth(set
, true);
2408 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2409 queue_set_hctx_shared(q
, true);
2410 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2412 mutex_unlock(&set
->tag_list_lock
);
2416 * It is the actual release handler for mq, but we do it from
2417 * request queue's release handler for avoiding use-after-free
2418 * and headache because q->mq_kobj shouldn't have been introduced,
2419 * but we can't group ctx/kctx kobj without it.
2421 void blk_mq_release(struct request_queue
*q
)
2423 struct blk_mq_hw_ctx
*hctx
;
2426 /* hctx kobj stays in hctx */
2427 queue_for_each_hw_ctx(q
, hctx
, i
) {
2430 kobject_put(&hctx
->kobj
);
2435 kfree(q
->queue_hw_ctx
);
2438 * release .mq_kobj and sw queue's kobject now because
2439 * both share lifetime with request queue.
2441 blk_mq_sysfs_deinit(q
);
2443 free_percpu(q
->queue_ctx
);
2446 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2448 struct request_queue
*uninit_q
, *q
;
2450 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2452 return ERR_PTR(-ENOMEM
);
2454 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2456 blk_cleanup_queue(uninit_q
);
2460 EXPORT_SYMBOL(blk_mq_init_queue
);
2462 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2464 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2466 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, queue_rq_srcu
),
2467 __alignof__(struct blk_mq_hw_ctx
)) !=
2468 sizeof(struct blk_mq_hw_ctx
));
2470 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2471 hw_ctx_size
+= sizeof(struct srcu_struct
);
2476 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2477 struct request_queue
*q
)
2480 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2482 blk_mq_sysfs_unregister(q
);
2484 /* protect against switching io scheduler */
2485 mutex_lock(&q
->sysfs_lock
);
2486 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2492 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2493 hctxs
[i
] = kzalloc_node(blk_mq_hw_ctx_size(set
),
2498 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2505 atomic_set(&hctxs
[i
]->nr_active
, 0);
2506 hctxs
[i
]->numa_node
= node
;
2507 hctxs
[i
]->queue_num
= i
;
2509 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2510 free_cpumask_var(hctxs
[i
]->cpumask
);
2515 blk_mq_hctx_kobj_init(hctxs
[i
]);
2517 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2518 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2522 blk_mq_free_map_and_requests(set
, j
);
2523 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2524 kobject_put(&hctx
->kobj
);
2529 q
->nr_hw_queues
= i
;
2530 mutex_unlock(&q
->sysfs_lock
);
2531 blk_mq_sysfs_register(q
);
2534 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2535 struct request_queue
*q
)
2537 /* mark the queue as mq asap */
2538 q
->mq_ops
= set
->ops
;
2540 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2541 blk_mq_poll_stats_bkt
,
2542 BLK_MQ_POLL_STATS_BKTS
, q
);
2546 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2550 /* init q->mq_kobj and sw queues' kobjects */
2551 blk_mq_sysfs_init(q
);
2553 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2554 GFP_KERNEL
, set
->numa_node
);
2555 if (!q
->queue_hw_ctx
)
2558 q
->mq_map
= set
->mq_map
;
2560 blk_mq_realloc_hw_ctxs(set
, q
);
2561 if (!q
->nr_hw_queues
)
2564 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2565 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2567 q
->nr_queues
= nr_cpu_ids
;
2569 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2571 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2572 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2574 q
->sg_reserved_size
= INT_MAX
;
2576 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2577 INIT_LIST_HEAD(&q
->requeue_list
);
2578 spin_lock_init(&q
->requeue_lock
);
2580 blk_queue_make_request(q
, blk_mq_make_request
);
2581 if (q
->mq_ops
->poll
)
2582 q
->poll_fn
= blk_mq_poll
;
2585 * Do this after blk_queue_make_request() overrides it...
2587 q
->nr_requests
= set
->queue_depth
;
2590 * Default to classic polling
2594 if (set
->ops
->complete
)
2595 blk_queue_softirq_done(q
, set
->ops
->complete
);
2597 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2598 blk_mq_add_queue_tag_set(set
, q
);
2599 blk_mq_map_swqueue(q
);
2601 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2604 ret
= blk_mq_sched_init(q
);
2606 return ERR_PTR(ret
);
2612 kfree(q
->queue_hw_ctx
);
2614 free_percpu(q
->queue_ctx
);
2617 return ERR_PTR(-ENOMEM
);
2619 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2621 void blk_mq_free_queue(struct request_queue
*q
)
2623 struct blk_mq_tag_set
*set
= q
->tag_set
;
2625 blk_mq_del_queue_tag_set(q
);
2626 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2629 /* Basically redo blk_mq_init_queue with queue frozen */
2630 static void blk_mq_queue_reinit(struct request_queue
*q
)
2632 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2634 blk_mq_debugfs_unregister_hctxs(q
);
2635 blk_mq_sysfs_unregister(q
);
2638 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2639 * we should change hctx numa_node according to the new topology (this
2640 * involves freeing and re-allocating memory, worth doing?)
2642 blk_mq_map_swqueue(q
);
2644 blk_mq_sysfs_register(q
);
2645 blk_mq_debugfs_register_hctxs(q
);
2648 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2652 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2653 if (!__blk_mq_alloc_rq_map(set
, i
))
2660 blk_mq_free_rq_map(set
->tags
[i
]);
2666 * Allocate the request maps associated with this tag_set. Note that this
2667 * may reduce the depth asked for, if memory is tight. set->queue_depth
2668 * will be updated to reflect the allocated depth.
2670 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2675 depth
= set
->queue_depth
;
2677 err
= __blk_mq_alloc_rq_maps(set
);
2681 set
->queue_depth
>>= 1;
2682 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2686 } while (set
->queue_depth
);
2688 if (!set
->queue_depth
|| err
) {
2689 pr_err("blk-mq: failed to allocate request map\n");
2693 if (depth
!= set
->queue_depth
)
2694 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2695 depth
, set
->queue_depth
);
2700 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2702 if (set
->ops
->map_queues
) {
2705 * transport .map_queues is usually done in the following
2708 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2709 * mask = get_cpu_mask(queue)
2710 * for_each_cpu(cpu, mask)
2711 * set->mq_map[cpu] = queue;
2714 * When we need to remap, the table has to be cleared for
2715 * killing stale mapping since one CPU may not be mapped
2718 for_each_possible_cpu(cpu
)
2719 set
->mq_map
[cpu
] = 0;
2721 return set
->ops
->map_queues(set
);
2723 return blk_mq_map_queues(set
);
2727 * Alloc a tag set to be associated with one or more request queues.
2728 * May fail with EINVAL for various error conditions. May adjust the
2729 * requested depth down, if if it too large. In that case, the set
2730 * value will be stored in set->queue_depth.
2732 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2736 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2738 if (!set
->nr_hw_queues
)
2740 if (!set
->queue_depth
)
2742 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2745 if (!set
->ops
->queue_rq
)
2748 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
2751 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2752 pr_info("blk-mq: reduced tag depth to %u\n",
2754 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2758 * If a crashdump is active, then we are potentially in a very
2759 * memory constrained environment. Limit us to 1 queue and
2760 * 64 tags to prevent using too much memory.
2762 if (is_kdump_kernel()) {
2763 set
->nr_hw_queues
= 1;
2764 set
->queue_depth
= min(64U, set
->queue_depth
);
2767 * There is no use for more h/w queues than cpus.
2769 if (set
->nr_hw_queues
> nr_cpu_ids
)
2770 set
->nr_hw_queues
= nr_cpu_ids
;
2772 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2773 GFP_KERNEL
, set
->numa_node
);
2778 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2779 GFP_KERNEL
, set
->numa_node
);
2783 ret
= blk_mq_update_queue_map(set
);
2785 goto out_free_mq_map
;
2787 ret
= blk_mq_alloc_rq_maps(set
);
2789 goto out_free_mq_map
;
2791 mutex_init(&set
->tag_list_lock
);
2792 INIT_LIST_HEAD(&set
->tag_list
);
2804 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2806 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2810 for (i
= 0; i
< nr_cpu_ids
; i
++)
2811 blk_mq_free_map_and_requests(set
, i
);
2819 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2821 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2823 struct blk_mq_tag_set
*set
= q
->tag_set
;
2824 struct blk_mq_hw_ctx
*hctx
;
2830 blk_mq_freeze_queue(q
);
2831 blk_mq_quiesce_queue(q
);
2834 queue_for_each_hw_ctx(q
, hctx
, i
) {
2838 * If we're using an MQ scheduler, just update the scheduler
2839 * queue depth. This is similar to what the old code would do.
2841 if (!hctx
->sched_tags
) {
2842 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
2845 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2853 q
->nr_requests
= nr
;
2855 blk_mq_unquiesce_queue(q
);
2856 blk_mq_unfreeze_queue(q
);
2861 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
2864 struct request_queue
*q
;
2866 lockdep_assert_held(&set
->tag_list_lock
);
2868 if (nr_hw_queues
> nr_cpu_ids
)
2869 nr_hw_queues
= nr_cpu_ids
;
2870 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2873 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2874 blk_mq_freeze_queue(q
);
2876 set
->nr_hw_queues
= nr_hw_queues
;
2877 blk_mq_update_queue_map(set
);
2878 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2879 blk_mq_realloc_hw_ctxs(set
, q
);
2880 blk_mq_queue_reinit(q
);
2883 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2884 blk_mq_unfreeze_queue(q
);
2887 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2889 mutex_lock(&set
->tag_list_lock
);
2890 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
2891 mutex_unlock(&set
->tag_list_lock
);
2893 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2895 /* Enable polling stats and return whether they were already enabled. */
2896 static bool blk_poll_stats_enable(struct request_queue
*q
)
2898 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2899 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
2901 blk_stat_add_callback(q
, q
->poll_cb
);
2905 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2908 * We don't arm the callback if polling stats are not enabled or the
2909 * callback is already active.
2911 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2912 blk_stat_is_active(q
->poll_cb
))
2915 blk_stat_activate_msecs(q
->poll_cb
, 100);
2918 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2920 struct request_queue
*q
= cb
->data
;
2923 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
2924 if (cb
->stat
[bucket
].nr_samples
)
2925 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
2929 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2930 struct blk_mq_hw_ctx
*hctx
,
2933 unsigned long ret
= 0;
2937 * If stats collection isn't on, don't sleep but turn it on for
2940 if (!blk_poll_stats_enable(q
))
2944 * As an optimistic guess, use half of the mean service time
2945 * for this type of request. We can (and should) make this smarter.
2946 * For instance, if the completion latencies are tight, we can
2947 * get closer than just half the mean. This is especially
2948 * important on devices where the completion latencies are longer
2949 * than ~10 usec. We do use the stats for the relevant IO size
2950 * if available which does lead to better estimates.
2952 bucket
= blk_mq_poll_stats_bkt(rq
);
2956 if (q
->poll_stat
[bucket
].nr_samples
)
2957 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
2962 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2963 struct blk_mq_hw_ctx
*hctx
,
2966 struct hrtimer_sleeper hs
;
2967 enum hrtimer_mode mode
;
2971 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2977 * -1: don't ever hybrid sleep
2978 * 0: use half of prev avg
2979 * >0: use this specific value
2981 if (q
->poll_nsec
== -1)
2983 else if (q
->poll_nsec
> 0)
2984 nsecs
= q
->poll_nsec
;
2986 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2991 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2994 * This will be replaced with the stats tracking code, using
2995 * 'avg_completion_time / 2' as the pre-sleep target.
2999 mode
= HRTIMER_MODE_REL
;
3000 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
3001 hrtimer_set_expires(&hs
.timer
, kt
);
3003 hrtimer_init_sleeper(&hs
, current
);
3005 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
3007 set_current_state(TASK_UNINTERRUPTIBLE
);
3008 hrtimer_start_expires(&hs
.timer
, mode
);
3011 hrtimer_cancel(&hs
.timer
);
3012 mode
= HRTIMER_MODE_ABS
;
3013 } while (hs
.task
&& !signal_pending(current
));
3015 __set_current_state(TASK_RUNNING
);
3016 destroy_hrtimer_on_stack(&hs
.timer
);
3020 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
3022 struct request_queue
*q
= hctx
->queue
;
3026 * If we sleep, have the caller restart the poll loop to reset
3027 * the state. Like for the other success return cases, the
3028 * caller is responsible for checking if the IO completed. If
3029 * the IO isn't complete, we'll get called again and will go
3030 * straight to the busy poll loop.
3032 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
3035 hctx
->poll_considered
++;
3037 state
= current
->state
;
3038 while (!need_resched()) {
3041 hctx
->poll_invoked
++;
3043 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
3045 hctx
->poll_success
++;
3046 set_current_state(TASK_RUNNING
);
3050 if (signal_pending_state(state
, current
))
3051 set_current_state(TASK_RUNNING
);
3053 if (current
->state
== TASK_RUNNING
)
3063 static bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
3065 struct blk_mq_hw_ctx
*hctx
;
3068 if (!test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3071 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3072 if (!blk_qc_t_is_internal(cookie
))
3073 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3075 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3077 * With scheduling, if the request has completed, we'll
3078 * get a NULL return here, as we clear the sched tag when
3079 * that happens. The request still remains valid, like always,
3080 * so we should be safe with just the NULL check.
3086 return __blk_mq_poll(hctx
, rq
);
3089 static int __init
blk_mq_init(void)
3091 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3092 blk_mq_hctx_notify_dead
);
3095 subsys_initcall(blk_mq_init
);