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"
39 #include "blk-rq-qos.h"
41 static bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
);
42 static void blk_mq_poll_stats_start(struct request_queue
*q
);
43 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
45 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
47 int ddir
, bytes
, bucket
;
49 ddir
= rq_data_dir(rq
);
50 bytes
= blk_rq_bytes(rq
);
52 bucket
= ddir
+ 2*(ilog2(bytes
) - 9);
56 else if (bucket
>= BLK_MQ_POLL_STATS_BKTS
)
57 return ddir
+ BLK_MQ_POLL_STATS_BKTS
- 2;
63 * Check if any of the ctx's have pending work in this hardware queue
65 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
67 return !list_empty_careful(&hctx
->dispatch
) ||
68 sbitmap_any_bit_set(&hctx
->ctx_map
) ||
69 blk_mq_sched_has_work(hctx
);
73 * Mark this ctx as having pending work in this hardware queue
75 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
76 struct blk_mq_ctx
*ctx
)
78 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
79 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
82 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
83 struct blk_mq_ctx
*ctx
)
85 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
89 struct hd_struct
*part
;
90 unsigned int *inflight
;
93 static void blk_mq_check_inflight(struct blk_mq_hw_ctx
*hctx
,
94 struct request
*rq
, void *priv
,
97 struct mq_inflight
*mi
= priv
;
100 * index[0] counts the specific partition that was asked for. index[1]
101 * counts the ones that are active on the whole device, so increment
102 * that if mi->part is indeed a partition, and not a whole device.
104 if (rq
->part
== mi
->part
)
106 if (mi
->part
->partno
)
110 void blk_mq_in_flight(struct request_queue
*q
, struct hd_struct
*part
,
111 unsigned int inflight
[2])
113 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
115 inflight
[0] = inflight
[1] = 0;
116 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
119 static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx
*hctx
,
120 struct request
*rq
, void *priv
,
123 struct mq_inflight
*mi
= priv
;
125 if (rq
->part
== mi
->part
)
126 mi
->inflight
[rq_data_dir(rq
)]++;
129 void blk_mq_in_flight_rw(struct request_queue
*q
, struct hd_struct
*part
,
130 unsigned int inflight
[2])
132 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
134 inflight
[0] = inflight
[1] = 0;
135 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight_rw
, &mi
);
138 void blk_freeze_queue_start(struct request_queue
*q
)
142 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
143 if (freeze_depth
== 1) {
144 percpu_ref_kill(&q
->q_usage_counter
);
146 blk_mq_run_hw_queues(q
, false);
149 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
151 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
153 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
155 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
157 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
158 unsigned long timeout
)
160 return wait_event_timeout(q
->mq_freeze_wq
,
161 percpu_ref_is_zero(&q
->q_usage_counter
),
164 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
167 * Guarantee no request is in use, so we can change any data structure of
168 * the queue afterward.
170 void blk_freeze_queue(struct request_queue
*q
)
173 * In the !blk_mq case we are only calling this to kill the
174 * q_usage_counter, otherwise this increases the freeze depth
175 * and waits for it to return to zero. For this reason there is
176 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
177 * exported to drivers as the only user for unfreeze is blk_mq.
179 blk_freeze_queue_start(q
);
182 blk_mq_freeze_queue_wait(q
);
185 void blk_mq_freeze_queue(struct request_queue
*q
)
188 * ...just an alias to keep freeze and unfreeze actions balanced
189 * in the blk_mq_* namespace
193 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
195 void blk_mq_unfreeze_queue(struct request_queue
*q
)
199 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
200 WARN_ON_ONCE(freeze_depth
< 0);
202 percpu_ref_resurrect(&q
->q_usage_counter
);
203 wake_up_all(&q
->mq_freeze_wq
);
206 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
209 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
210 * mpt3sas driver such that this function can be removed.
212 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
214 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
216 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
219 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
222 * Note: this function does not prevent that the struct request end_io()
223 * callback function is invoked. Once this function is returned, we make
224 * sure no dispatch can happen until the queue is unquiesced via
225 * blk_mq_unquiesce_queue().
227 void blk_mq_quiesce_queue(struct request_queue
*q
)
229 struct blk_mq_hw_ctx
*hctx
;
233 blk_mq_quiesce_queue_nowait(q
);
235 queue_for_each_hw_ctx(q
, hctx
, i
) {
236 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
237 synchronize_srcu(hctx
->srcu
);
244 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
247 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
250 * This function recovers queue into the state before quiescing
251 * which is done by blk_mq_quiesce_queue.
253 void blk_mq_unquiesce_queue(struct request_queue
*q
)
255 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
257 /* dispatch requests which are inserted during quiescing */
258 blk_mq_run_hw_queues(q
, true);
260 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
262 void blk_mq_wake_waiters(struct request_queue
*q
)
264 struct blk_mq_hw_ctx
*hctx
;
267 queue_for_each_hw_ctx(q
, hctx
, i
)
268 if (blk_mq_hw_queue_mapped(hctx
))
269 blk_mq_tag_wakeup_all(hctx
->tags
, true);
272 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
274 return blk_mq_has_free_tags(hctx
->tags
);
276 EXPORT_SYMBOL(blk_mq_can_queue
);
278 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
279 unsigned int tag
, unsigned int op
)
281 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
282 struct request
*rq
= tags
->static_rqs
[tag
];
283 req_flags_t rq_flags
= 0;
285 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
287 rq
->internal_tag
= tag
;
289 if (data
->hctx
->flags
& BLK_MQ_F_TAG_SHARED
) {
290 rq_flags
= RQF_MQ_INFLIGHT
;
291 atomic_inc(&data
->hctx
->nr_active
);
294 rq
->internal_tag
= -1;
295 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
298 /* csd/requeue_work/fifo_time is initialized before use */
300 rq
->mq_ctx
= data
->ctx
;
301 rq
->rq_flags
= rq_flags
;
304 if (data
->flags
& BLK_MQ_REQ_PREEMPT
)
305 rq
->rq_flags
|= RQF_PREEMPT
;
306 if (blk_queue_io_stat(data
->q
))
307 rq
->rq_flags
|= RQF_IO_STAT
;
308 INIT_LIST_HEAD(&rq
->queuelist
);
309 INIT_HLIST_NODE(&rq
->hash
);
310 RB_CLEAR_NODE(&rq
->rb_node
);
313 rq
->start_time_ns
= ktime_get_ns();
314 rq
->io_start_time_ns
= 0;
315 rq
->nr_phys_segments
= 0;
316 #if defined(CONFIG_BLK_DEV_INTEGRITY)
317 rq
->nr_integrity_segments
= 0;
320 /* tag was already set */
324 INIT_LIST_HEAD(&rq
->timeout_list
);
328 rq
->end_io_data
= NULL
;
331 #ifdef CONFIG_BLK_CGROUP
335 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
336 refcount_set(&rq
->ref
, 1);
340 static struct request
*blk_mq_get_request(struct request_queue
*q
,
341 struct bio
*bio
, unsigned int op
,
342 struct blk_mq_alloc_data
*data
)
344 struct elevator_queue
*e
= q
->elevator
;
347 bool put_ctx_on_error
= false;
349 blk_queue_enter_live(q
);
351 if (likely(!data
->ctx
)) {
352 data
->ctx
= blk_mq_get_ctx(q
);
353 put_ctx_on_error
= true;
355 if (likely(!data
->hctx
))
356 data
->hctx
= blk_mq_map_queue(q
, data
->ctx
->cpu
);
358 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
361 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
364 * Flush requests are special and go directly to the
365 * dispatch list. Don't include reserved tags in the
366 * limiting, as it isn't useful.
368 if (!op_is_flush(op
) && e
->type
->ops
.mq
.limit_depth
&&
369 !(data
->flags
& BLK_MQ_REQ_RESERVED
))
370 e
->type
->ops
.mq
.limit_depth(op
, data
);
372 blk_mq_tag_busy(data
->hctx
);
375 tag
= blk_mq_get_tag(data
);
376 if (tag
== BLK_MQ_TAG_FAIL
) {
377 if (put_ctx_on_error
) {
378 blk_mq_put_ctx(data
->ctx
);
385 rq
= blk_mq_rq_ctx_init(data
, tag
, op
);
386 if (!op_is_flush(op
)) {
388 if (e
&& e
->type
->ops
.mq
.prepare_request
) {
389 if (e
->type
->icq_cache
&& rq_ioc(bio
))
390 blk_mq_sched_assign_ioc(rq
, bio
);
392 e
->type
->ops
.mq
.prepare_request(rq
, bio
);
393 rq
->rq_flags
|= RQF_ELVPRIV
;
396 data
->hctx
->queued
++;
400 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
401 blk_mq_req_flags_t flags
)
403 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
407 ret
= blk_queue_enter(q
, flags
);
411 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
415 return ERR_PTR(-EWOULDBLOCK
);
417 blk_mq_put_ctx(alloc_data
.ctx
);
420 rq
->__sector
= (sector_t
) -1;
421 rq
->bio
= rq
->biotail
= NULL
;
424 EXPORT_SYMBOL(blk_mq_alloc_request
);
426 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
427 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
429 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
435 * If the tag allocator sleeps we could get an allocation for a
436 * different hardware context. No need to complicate the low level
437 * allocator for this for the rare use case of a command tied to
440 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
441 return ERR_PTR(-EINVAL
);
443 if (hctx_idx
>= q
->nr_hw_queues
)
444 return ERR_PTR(-EIO
);
446 ret
= blk_queue_enter(q
, flags
);
451 * Check if the hardware context is actually mapped to anything.
452 * If not tell the caller that it should skip this queue.
454 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
455 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
457 return ERR_PTR(-EXDEV
);
459 cpu
= cpumask_first_and(alloc_data
.hctx
->cpumask
, cpu_online_mask
);
460 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
462 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
466 return ERR_PTR(-EWOULDBLOCK
);
470 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
472 static void __blk_mq_free_request(struct request
*rq
)
474 struct request_queue
*q
= rq
->q
;
475 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
476 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
477 const int sched_tag
= rq
->internal_tag
;
479 blk_pm_mark_last_busy(rq
);
481 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
483 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
484 blk_mq_sched_restart(hctx
);
488 void blk_mq_free_request(struct request
*rq
)
490 struct request_queue
*q
= rq
->q
;
491 struct elevator_queue
*e
= q
->elevator
;
492 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
493 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
495 if (rq
->rq_flags
& RQF_ELVPRIV
) {
496 if (e
&& e
->type
->ops
.mq
.finish_request
)
497 e
->type
->ops
.mq
.finish_request(rq
);
499 put_io_context(rq
->elv
.icq
->ioc
);
504 ctx
->rq_completed
[rq_is_sync(rq
)]++;
505 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
506 atomic_dec(&hctx
->nr_active
);
508 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
509 laptop_io_completion(q
->backing_dev_info
);
514 blk_put_rl(blk_rq_rl(rq
));
516 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
517 if (refcount_dec_and_test(&rq
->ref
))
518 __blk_mq_free_request(rq
);
520 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
522 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
524 u64 now
= ktime_get_ns();
526 if (rq
->rq_flags
& RQF_STATS
) {
527 blk_mq_poll_stats_start(rq
->q
);
528 blk_stat_add(rq
, now
);
531 if (rq
->internal_tag
!= -1)
532 blk_mq_sched_completed_request(rq
, now
);
534 blk_account_io_done(rq
, now
);
537 rq_qos_done(rq
->q
, rq
);
538 rq
->end_io(rq
, error
);
540 if (unlikely(blk_bidi_rq(rq
)))
541 blk_mq_free_request(rq
->next_rq
);
542 blk_mq_free_request(rq
);
545 EXPORT_SYMBOL(__blk_mq_end_request
);
547 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
549 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
551 __blk_mq_end_request(rq
, error
);
553 EXPORT_SYMBOL(blk_mq_end_request
);
555 static void __blk_mq_complete_request_remote(void *data
)
557 struct request
*rq
= data
;
559 rq
->q
->softirq_done_fn(rq
);
562 static void __blk_mq_complete_request(struct request
*rq
)
564 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
568 if (!blk_mq_mark_complete(rq
))
571 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
572 rq
->q
->softirq_done_fn(rq
);
577 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
578 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
580 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
581 rq
->csd
.func
= __blk_mq_complete_request_remote
;
584 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
586 rq
->q
->softirq_done_fn(rq
);
591 static void hctx_unlock(struct blk_mq_hw_ctx
*hctx
, int srcu_idx
)
592 __releases(hctx
->srcu
)
594 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
597 srcu_read_unlock(hctx
->srcu
, srcu_idx
);
600 static void hctx_lock(struct blk_mq_hw_ctx
*hctx
, int *srcu_idx
)
601 __acquires(hctx
->srcu
)
603 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
604 /* shut up gcc false positive */
608 *srcu_idx
= srcu_read_lock(hctx
->srcu
);
612 * blk_mq_complete_request - end I/O on a request
613 * @rq: the request being processed
616 * Ends all I/O on a request. It does not handle partial completions.
617 * The actual completion happens out-of-order, through a IPI handler.
619 void blk_mq_complete_request(struct request
*rq
)
621 if (unlikely(blk_should_fake_timeout(rq
->q
)))
623 __blk_mq_complete_request(rq
);
625 EXPORT_SYMBOL(blk_mq_complete_request
);
627 int blk_mq_request_started(struct request
*rq
)
629 return blk_mq_rq_state(rq
) != MQ_RQ_IDLE
;
631 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
633 void blk_mq_start_request(struct request
*rq
)
635 struct request_queue
*q
= rq
->q
;
637 blk_mq_sched_started_request(rq
);
639 trace_block_rq_issue(q
, rq
);
641 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
642 rq
->io_start_time_ns
= ktime_get_ns();
643 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
644 rq
->throtl_size
= blk_rq_sectors(rq
);
646 rq
->rq_flags
|= RQF_STATS
;
650 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
653 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
655 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
657 * Make sure space for the drain appears. We know we can do
658 * this because max_hw_segments has been adjusted to be one
659 * fewer than the device can handle.
661 rq
->nr_phys_segments
++;
664 EXPORT_SYMBOL(blk_mq_start_request
);
666 static void __blk_mq_requeue_request(struct request
*rq
)
668 struct request_queue
*q
= rq
->q
;
670 blk_mq_put_driver_tag(rq
);
672 trace_block_rq_requeue(q
, rq
);
673 rq_qos_requeue(q
, rq
);
675 if (blk_mq_request_started(rq
)) {
676 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
677 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
678 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
679 rq
->nr_phys_segments
--;
683 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
685 __blk_mq_requeue_request(rq
);
687 /* this request will be re-inserted to io scheduler queue */
688 blk_mq_sched_requeue_request(rq
);
690 BUG_ON(blk_queued_rq(rq
));
691 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
693 EXPORT_SYMBOL(blk_mq_requeue_request
);
695 static void blk_mq_requeue_work(struct work_struct
*work
)
697 struct request_queue
*q
=
698 container_of(work
, struct request_queue
, requeue_work
.work
);
700 struct request
*rq
, *next
;
702 spin_lock_irq(&q
->requeue_lock
);
703 list_splice_init(&q
->requeue_list
, &rq_list
);
704 spin_unlock_irq(&q
->requeue_lock
);
706 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
707 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
710 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
711 list_del_init(&rq
->queuelist
);
712 blk_mq_sched_insert_request(rq
, true, false, false);
715 while (!list_empty(&rq_list
)) {
716 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
717 list_del_init(&rq
->queuelist
);
718 blk_mq_sched_insert_request(rq
, false, false, false);
721 blk_mq_run_hw_queues(q
, false);
724 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
725 bool kick_requeue_list
)
727 struct request_queue
*q
= rq
->q
;
731 * We abuse this flag that is otherwise used by the I/O scheduler to
732 * request head insertion from the workqueue.
734 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
736 spin_lock_irqsave(&q
->requeue_lock
, flags
);
738 rq
->rq_flags
|= RQF_SOFTBARRIER
;
739 list_add(&rq
->queuelist
, &q
->requeue_list
);
741 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
743 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
745 if (kick_requeue_list
)
746 blk_mq_kick_requeue_list(q
);
748 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
750 void blk_mq_kick_requeue_list(struct request_queue
*q
)
752 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
754 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
756 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
759 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
760 msecs_to_jiffies(msecs
));
762 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
764 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
766 if (tag
< tags
->nr_tags
) {
767 prefetch(tags
->rqs
[tag
]);
768 return tags
->rqs
[tag
];
773 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
775 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
777 req
->rq_flags
|= RQF_TIMED_OUT
;
778 if (req
->q
->mq_ops
->timeout
) {
779 enum blk_eh_timer_return ret
;
781 ret
= req
->q
->mq_ops
->timeout(req
, reserved
);
782 if (ret
== BLK_EH_DONE
)
784 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
790 static bool blk_mq_req_expired(struct request
*rq
, unsigned long *next
)
792 unsigned long deadline
;
794 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
796 if (rq
->rq_flags
& RQF_TIMED_OUT
)
799 deadline
= blk_rq_deadline(rq
);
800 if (time_after_eq(jiffies
, deadline
))
805 else if (time_after(*next
, deadline
))
810 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
811 struct request
*rq
, void *priv
, bool reserved
)
813 unsigned long *next
= priv
;
816 * Just do a quick check if it is expired before locking the request in
817 * so we're not unnecessarilly synchronizing across CPUs.
819 if (!blk_mq_req_expired(rq
, next
))
823 * We have reason to believe the request may be expired. Take a
824 * reference on the request to lock this request lifetime into its
825 * currently allocated context to prevent it from being reallocated in
826 * the event the completion by-passes this timeout handler.
828 * If the reference was already released, then the driver beat the
829 * timeout handler to posting a natural completion.
831 if (!refcount_inc_not_zero(&rq
->ref
))
835 * The request is now locked and cannot be reallocated underneath the
836 * timeout handler's processing. Re-verify this exact request is truly
837 * expired; if it is not expired, then the request was completed and
838 * reallocated as a new request.
840 if (blk_mq_req_expired(rq
, next
))
841 blk_mq_rq_timed_out(rq
, reserved
);
842 if (refcount_dec_and_test(&rq
->ref
))
843 __blk_mq_free_request(rq
);
846 static void blk_mq_timeout_work(struct work_struct
*work
)
848 struct request_queue
*q
=
849 container_of(work
, struct request_queue
, timeout_work
);
850 unsigned long next
= 0;
851 struct blk_mq_hw_ctx
*hctx
;
854 /* A deadlock might occur if a request is stuck requiring a
855 * timeout at the same time a queue freeze is waiting
856 * completion, since the timeout code would not be able to
857 * acquire the queue reference here.
859 * That's why we don't use blk_queue_enter here; instead, we use
860 * percpu_ref_tryget directly, because we need to be able to
861 * obtain a reference even in the short window between the queue
862 * starting to freeze, by dropping the first reference in
863 * blk_freeze_queue_start, and the moment the last request is
864 * consumed, marked by the instant q_usage_counter reaches
867 if (!percpu_ref_tryget(&q
->q_usage_counter
))
870 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &next
);
873 mod_timer(&q
->timeout
, next
);
876 * Request timeouts are handled as a forward rolling timer. If
877 * we end up here it means that no requests are pending and
878 * also that no request has been pending for a while. Mark
881 queue_for_each_hw_ctx(q
, hctx
, i
) {
882 /* the hctx may be unmapped, so check it here */
883 if (blk_mq_hw_queue_mapped(hctx
))
884 blk_mq_tag_idle(hctx
);
890 struct flush_busy_ctx_data
{
891 struct blk_mq_hw_ctx
*hctx
;
892 struct list_head
*list
;
895 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
897 struct flush_busy_ctx_data
*flush_data
= data
;
898 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
899 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
901 spin_lock(&ctx
->lock
);
902 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
903 sbitmap_clear_bit(sb
, bitnr
);
904 spin_unlock(&ctx
->lock
);
909 * Process software queues that have been marked busy, splicing them
910 * to the for-dispatch
912 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
914 struct flush_busy_ctx_data data
= {
919 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
921 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
923 struct dispatch_rq_data
{
924 struct blk_mq_hw_ctx
*hctx
;
928 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
931 struct dispatch_rq_data
*dispatch_data
= data
;
932 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
933 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
935 spin_lock(&ctx
->lock
);
936 if (!list_empty(&ctx
->rq_list
)) {
937 dispatch_data
->rq
= list_entry_rq(ctx
->rq_list
.next
);
938 list_del_init(&dispatch_data
->rq
->queuelist
);
939 if (list_empty(&ctx
->rq_list
))
940 sbitmap_clear_bit(sb
, bitnr
);
942 spin_unlock(&ctx
->lock
);
944 return !dispatch_data
->rq
;
947 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
948 struct blk_mq_ctx
*start
)
950 unsigned off
= start
? start
->index_hw
: 0;
951 struct dispatch_rq_data data
= {
956 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
957 dispatch_rq_from_ctx
, &data
);
962 static inline unsigned int queued_to_index(unsigned int queued
)
967 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
970 bool blk_mq_get_driver_tag(struct request
*rq
)
972 struct blk_mq_alloc_data data
= {
974 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
975 .flags
= BLK_MQ_REQ_NOWAIT
,
982 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
983 data
.flags
|= BLK_MQ_REQ_RESERVED
;
985 shared
= blk_mq_tag_busy(data
.hctx
);
986 rq
->tag
= blk_mq_get_tag(&data
);
989 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
990 atomic_inc(&data
.hctx
->nr_active
);
992 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
996 return rq
->tag
!= -1;
999 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1000 int flags
, void *key
)
1002 struct blk_mq_hw_ctx
*hctx
;
1004 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1006 spin_lock(&hctx
->dispatch_wait_lock
);
1007 list_del_init(&wait
->entry
);
1008 spin_unlock(&hctx
->dispatch_wait_lock
);
1010 blk_mq_run_hw_queue(hctx
, true);
1015 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1016 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1017 * restart. For both cases, take care to check the condition again after
1018 * marking us as waiting.
1020 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1023 struct wait_queue_head
*wq
;
1024 wait_queue_entry_t
*wait
;
1027 if (!(hctx
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1028 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
1029 set_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
1032 * It's possible that a tag was freed in the window between the
1033 * allocation failure and adding the hardware queue to the wait
1036 * Don't clear RESTART here, someone else could have set it.
1037 * At most this will cost an extra queue run.
1039 return blk_mq_get_driver_tag(rq
);
1042 wait
= &hctx
->dispatch_wait
;
1043 if (!list_empty_careful(&wait
->entry
))
1046 wq
= &bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
)->wait
;
1048 spin_lock_irq(&wq
->lock
);
1049 spin_lock(&hctx
->dispatch_wait_lock
);
1050 if (!list_empty(&wait
->entry
)) {
1051 spin_unlock(&hctx
->dispatch_wait_lock
);
1052 spin_unlock_irq(&wq
->lock
);
1056 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1057 __add_wait_queue(wq
, wait
);
1060 * It's possible that a tag was freed in the window between the
1061 * allocation failure and adding the hardware queue to the wait
1064 ret
= blk_mq_get_driver_tag(rq
);
1066 spin_unlock(&hctx
->dispatch_wait_lock
);
1067 spin_unlock_irq(&wq
->lock
);
1072 * We got a tag, remove ourselves from the wait queue to ensure
1073 * someone else gets the wakeup.
1075 list_del_init(&wait
->entry
);
1076 spin_unlock(&hctx
->dispatch_wait_lock
);
1077 spin_unlock_irq(&wq
->lock
);
1082 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1083 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1085 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1086 * - EWMA is one simple way to compute running average value
1087 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1088 * - take 4 as factor for avoiding to get too small(0) result, and this
1089 * factor doesn't matter because EWMA decreases exponentially
1091 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1095 if (hctx
->queue
->elevator
)
1098 ewma
= hctx
->dispatch_busy
;
1103 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1105 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1106 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1108 hctx
->dispatch_busy
= ewma
;
1111 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1114 * Returns true if we did some work AND can potentially do more.
1116 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
,
1119 struct blk_mq_hw_ctx
*hctx
;
1120 struct request
*rq
, *nxt
;
1121 bool no_tag
= false;
1123 blk_status_t ret
= BLK_STS_OK
;
1125 if (list_empty(list
))
1128 WARN_ON(!list_is_singular(list
) && got_budget
);
1131 * Now process all the entries, sending them to the driver.
1133 errors
= queued
= 0;
1135 struct blk_mq_queue_data bd
;
1137 rq
= list_first_entry(list
, struct request
, queuelist
);
1139 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
1140 if (!got_budget
&& !blk_mq_get_dispatch_budget(hctx
))
1143 if (!blk_mq_get_driver_tag(rq
)) {
1145 * The initial allocation attempt failed, so we need to
1146 * rerun the hardware queue when a tag is freed. The
1147 * waitqueue takes care of that. If the queue is run
1148 * before we add this entry back on the dispatch list,
1149 * we'll re-run it below.
1151 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1152 blk_mq_put_dispatch_budget(hctx
);
1154 * For non-shared tags, the RESTART check
1157 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
)
1163 list_del_init(&rq
->queuelist
);
1168 * Flag last if we have no more requests, or if we have more
1169 * but can't assign a driver tag to it.
1171 if (list_empty(list
))
1174 nxt
= list_first_entry(list
, struct request
, queuelist
);
1175 bd
.last
= !blk_mq_get_driver_tag(nxt
);
1178 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1179 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
) {
1181 * If an I/O scheduler has been configured and we got a
1182 * driver tag for the next request already, free it
1185 if (!list_empty(list
)) {
1186 nxt
= list_first_entry(list
, struct request
, queuelist
);
1187 blk_mq_put_driver_tag(nxt
);
1189 list_add(&rq
->queuelist
, list
);
1190 __blk_mq_requeue_request(rq
);
1194 if (unlikely(ret
!= BLK_STS_OK
)) {
1196 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1201 } while (!list_empty(list
));
1203 hctx
->dispatched
[queued_to_index(queued
)]++;
1206 * Any items that need requeuing? Stuff them into hctx->dispatch,
1207 * that is where we will continue on next queue run.
1209 if (!list_empty(list
)) {
1212 spin_lock(&hctx
->lock
);
1213 list_splice_init(list
, &hctx
->dispatch
);
1214 spin_unlock(&hctx
->lock
);
1217 * If SCHED_RESTART was set by the caller of this function and
1218 * it is no longer set that means that it was cleared by another
1219 * thread and hence that a queue rerun is needed.
1221 * If 'no_tag' is set, that means that we failed getting
1222 * a driver tag with an I/O scheduler attached. If our dispatch
1223 * waitqueue is no longer active, ensure that we run the queue
1224 * AFTER adding our entries back to the list.
1226 * If no I/O scheduler has been configured it is possible that
1227 * the hardware queue got stopped and restarted before requests
1228 * were pushed back onto the dispatch list. Rerun the queue to
1229 * avoid starvation. Notes:
1230 * - blk_mq_run_hw_queue() checks whether or not a queue has
1231 * been stopped before rerunning a queue.
1232 * - Some but not all block drivers stop a queue before
1233 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1236 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1237 * bit is set, run queue after a delay to avoid IO stalls
1238 * that could otherwise occur if the queue is idle.
1240 needs_restart
= blk_mq_sched_needs_restart(hctx
);
1241 if (!needs_restart
||
1242 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1243 blk_mq_run_hw_queue(hctx
, true);
1244 else if (needs_restart
&& (ret
== BLK_STS_RESOURCE
))
1245 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
1247 blk_mq_update_dispatch_busy(hctx
, true);
1250 blk_mq_update_dispatch_busy(hctx
, false);
1253 * If the host/device is unable to accept more work, inform the
1256 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1259 return (queued
+ errors
) != 0;
1262 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1267 * We should be running this queue from one of the CPUs that
1270 * There are at least two related races now between setting
1271 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1272 * __blk_mq_run_hw_queue():
1274 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1275 * but later it becomes online, then this warning is harmless
1278 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1279 * but later it becomes offline, then the warning can't be
1280 * triggered, and we depend on blk-mq timeout handler to
1281 * handle dispatched requests to this hctx
1283 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1284 cpu_online(hctx
->next_cpu
)) {
1285 printk(KERN_WARNING
"run queue from wrong CPU %d, hctx %s\n",
1286 raw_smp_processor_id(),
1287 cpumask_empty(hctx
->cpumask
) ? "inactive": "active");
1292 * We can't run the queue inline with ints disabled. Ensure that
1293 * we catch bad users of this early.
1295 WARN_ON_ONCE(in_interrupt());
1297 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1299 hctx_lock(hctx
, &srcu_idx
);
1300 blk_mq_sched_dispatch_requests(hctx
);
1301 hctx_unlock(hctx
, srcu_idx
);
1304 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1306 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1308 if (cpu
>= nr_cpu_ids
)
1309 cpu
= cpumask_first(hctx
->cpumask
);
1314 * It'd be great if the workqueue API had a way to pass
1315 * in a mask and had some smarts for more clever placement.
1316 * For now we just round-robin here, switching for every
1317 * BLK_MQ_CPU_WORK_BATCH queued items.
1319 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1322 int next_cpu
= hctx
->next_cpu
;
1324 if (hctx
->queue
->nr_hw_queues
== 1)
1325 return WORK_CPU_UNBOUND
;
1327 if (--hctx
->next_cpu_batch
<= 0) {
1329 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1331 if (next_cpu
>= nr_cpu_ids
)
1332 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1333 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1337 * Do unbound schedule if we can't find a online CPU for this hctx,
1338 * and it should only happen in the path of handling CPU DEAD.
1340 if (!cpu_online(next_cpu
)) {
1347 * Make sure to re-select CPU next time once after CPUs
1348 * in hctx->cpumask become online again.
1350 hctx
->next_cpu
= next_cpu
;
1351 hctx
->next_cpu_batch
= 1;
1352 return WORK_CPU_UNBOUND
;
1355 hctx
->next_cpu
= next_cpu
;
1359 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1360 unsigned long msecs
)
1362 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1365 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1366 int cpu
= get_cpu();
1367 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1368 __blk_mq_run_hw_queue(hctx
);
1376 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1377 msecs_to_jiffies(msecs
));
1380 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1382 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1384 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1386 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1392 * When queue is quiesced, we may be switching io scheduler, or
1393 * updating nr_hw_queues, or other things, and we can't run queue
1394 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1396 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1399 hctx_lock(hctx
, &srcu_idx
);
1400 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1401 blk_mq_hctx_has_pending(hctx
);
1402 hctx_unlock(hctx
, srcu_idx
);
1405 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1411 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1413 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1415 struct blk_mq_hw_ctx
*hctx
;
1418 queue_for_each_hw_ctx(q
, hctx
, i
) {
1419 if (blk_mq_hctx_stopped(hctx
))
1422 blk_mq_run_hw_queue(hctx
, async
);
1425 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1428 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1429 * @q: request queue.
1431 * The caller is responsible for serializing this function against
1432 * blk_mq_{start,stop}_hw_queue().
1434 bool blk_mq_queue_stopped(struct request_queue
*q
)
1436 struct blk_mq_hw_ctx
*hctx
;
1439 queue_for_each_hw_ctx(q
, hctx
, i
)
1440 if (blk_mq_hctx_stopped(hctx
))
1445 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1448 * This function is often used for pausing .queue_rq() by driver when
1449 * there isn't enough resource or some conditions aren't satisfied, and
1450 * BLK_STS_RESOURCE is usually returned.
1452 * We do not guarantee that dispatch can be drained or blocked
1453 * after blk_mq_stop_hw_queue() returns. Please use
1454 * blk_mq_quiesce_queue() for that requirement.
1456 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1458 cancel_delayed_work(&hctx
->run_work
);
1460 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1462 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1465 * This function is often used for pausing .queue_rq() by driver when
1466 * there isn't enough resource or some conditions aren't satisfied, and
1467 * BLK_STS_RESOURCE is usually returned.
1469 * We do not guarantee that dispatch can be drained or blocked
1470 * after blk_mq_stop_hw_queues() returns. Please use
1471 * blk_mq_quiesce_queue() for that requirement.
1473 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1475 struct blk_mq_hw_ctx
*hctx
;
1478 queue_for_each_hw_ctx(q
, hctx
, i
)
1479 blk_mq_stop_hw_queue(hctx
);
1481 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1483 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1485 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1487 blk_mq_run_hw_queue(hctx
, false);
1489 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1491 void blk_mq_start_hw_queues(struct request_queue
*q
)
1493 struct blk_mq_hw_ctx
*hctx
;
1496 queue_for_each_hw_ctx(q
, hctx
, i
)
1497 blk_mq_start_hw_queue(hctx
);
1499 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1501 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1503 if (!blk_mq_hctx_stopped(hctx
))
1506 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1507 blk_mq_run_hw_queue(hctx
, async
);
1509 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1511 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1513 struct blk_mq_hw_ctx
*hctx
;
1516 queue_for_each_hw_ctx(q
, hctx
, i
)
1517 blk_mq_start_stopped_hw_queue(hctx
, async
);
1519 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1521 static void blk_mq_run_work_fn(struct work_struct
*work
)
1523 struct blk_mq_hw_ctx
*hctx
;
1525 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1528 * If we are stopped, don't run the queue.
1530 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1533 __blk_mq_run_hw_queue(hctx
);
1536 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1540 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1542 lockdep_assert_held(&ctx
->lock
);
1544 trace_block_rq_insert(hctx
->queue
, rq
);
1547 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1549 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1552 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1555 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1557 lockdep_assert_held(&ctx
->lock
);
1559 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1560 blk_mq_hctx_mark_pending(hctx
, ctx
);
1564 * Should only be used carefully, when the caller knows we want to
1565 * bypass a potential IO scheduler on the target device.
1567 void blk_mq_request_bypass_insert(struct request
*rq
, bool run_queue
)
1569 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1570 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(rq
->q
, ctx
->cpu
);
1572 spin_lock(&hctx
->lock
);
1573 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1574 spin_unlock(&hctx
->lock
);
1577 blk_mq_run_hw_queue(hctx
, false);
1580 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1581 struct list_head
*list
)
1587 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1590 list_for_each_entry(rq
, list
, queuelist
) {
1591 BUG_ON(rq
->mq_ctx
!= ctx
);
1592 trace_block_rq_insert(hctx
->queue
, rq
);
1595 spin_lock(&ctx
->lock
);
1596 list_splice_tail_init(list
, &ctx
->rq_list
);
1597 blk_mq_hctx_mark_pending(hctx
, ctx
);
1598 spin_unlock(&ctx
->lock
);
1601 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1603 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1604 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1606 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1607 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1608 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1611 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1613 struct blk_mq_ctx
*this_ctx
;
1614 struct request_queue
*this_q
;
1617 LIST_HEAD(ctx_list
);
1620 list_splice_init(&plug
->mq_list
, &list
);
1622 list_sort(NULL
, &list
, plug_ctx_cmp
);
1628 while (!list_empty(&list
)) {
1629 rq
= list_entry_rq(list
.next
);
1630 list_del_init(&rq
->queuelist
);
1632 if (rq
->mq_ctx
!= this_ctx
) {
1634 trace_block_unplug(this_q
, depth
, !from_schedule
);
1635 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1640 this_ctx
= rq
->mq_ctx
;
1646 list_add_tail(&rq
->queuelist
, &ctx_list
);
1650 * If 'this_ctx' is set, we know we have entries to complete
1651 * on 'ctx_list'. Do those.
1654 trace_block_unplug(this_q
, depth
, !from_schedule
);
1655 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1660 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1662 blk_init_request_from_bio(rq
, bio
);
1664 blk_rq_set_rl(rq
, blk_get_rl(rq
->q
, bio
));
1666 blk_account_io_start(rq
, true);
1669 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1672 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1674 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1677 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1681 struct request_queue
*q
= rq
->q
;
1682 struct blk_mq_queue_data bd
= {
1686 blk_qc_t new_cookie
;
1689 new_cookie
= request_to_qc_t(hctx
, rq
);
1692 * For OK queue, we are done. For error, caller may kill it.
1693 * Any other error (busy), just add it to our list as we
1694 * previously would have done.
1696 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1699 blk_mq_update_dispatch_busy(hctx
, false);
1700 *cookie
= new_cookie
;
1702 case BLK_STS_RESOURCE
:
1703 case BLK_STS_DEV_RESOURCE
:
1704 blk_mq_update_dispatch_busy(hctx
, true);
1705 __blk_mq_requeue_request(rq
);
1708 blk_mq_update_dispatch_busy(hctx
, false);
1709 *cookie
= BLK_QC_T_NONE
;
1716 static blk_status_t
__blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1721 struct request_queue
*q
= rq
->q
;
1722 bool run_queue
= true;
1725 * RCU or SRCU read lock is needed before checking quiesced flag.
1727 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1728 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1729 * and avoid driver to try to dispatch again.
1731 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1733 bypass_insert
= false;
1737 if (q
->elevator
&& !bypass_insert
)
1740 if (!blk_mq_get_dispatch_budget(hctx
))
1743 if (!blk_mq_get_driver_tag(rq
)) {
1744 blk_mq_put_dispatch_budget(hctx
);
1748 return __blk_mq_issue_directly(hctx
, rq
, cookie
);
1751 return BLK_STS_RESOURCE
;
1753 blk_mq_sched_insert_request(rq
, false, run_queue
, false);
1757 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1758 struct request
*rq
, blk_qc_t
*cookie
)
1763 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1765 hctx_lock(hctx
, &srcu_idx
);
1767 ret
= __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false);
1768 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1769 blk_mq_sched_insert_request(rq
, false, true, false);
1770 else if (ret
!= BLK_STS_OK
)
1771 blk_mq_end_request(rq
, ret
);
1773 hctx_unlock(hctx
, srcu_idx
);
1776 blk_status_t
blk_mq_request_issue_directly(struct request
*rq
)
1780 blk_qc_t unused_cookie
;
1781 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1782 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(rq
->q
, ctx
->cpu
);
1784 hctx_lock(hctx
, &srcu_idx
);
1785 ret
= __blk_mq_try_issue_directly(hctx
, rq
, &unused_cookie
, true);
1786 hctx_unlock(hctx
, srcu_idx
);
1791 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
1792 struct list_head
*list
)
1794 while (!list_empty(list
)) {
1796 struct request
*rq
= list_first_entry(list
, struct request
,
1799 list_del_init(&rq
->queuelist
);
1800 ret
= blk_mq_request_issue_directly(rq
);
1801 if (ret
!= BLK_STS_OK
) {
1802 if (ret
== BLK_STS_RESOURCE
||
1803 ret
== BLK_STS_DEV_RESOURCE
) {
1804 list_add(&rq
->queuelist
, list
);
1807 blk_mq_end_request(rq
, ret
);
1812 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1814 const int is_sync
= op_is_sync(bio
->bi_opf
);
1815 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1816 struct blk_mq_alloc_data data
= { .flags
= 0 };
1818 unsigned int request_count
= 0;
1819 struct blk_plug
*plug
;
1820 struct request
*same_queue_rq
= NULL
;
1823 blk_queue_bounce(q
, &bio
);
1825 blk_queue_split(q
, &bio
);
1827 if (!bio_integrity_prep(bio
))
1828 return BLK_QC_T_NONE
;
1830 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1831 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1832 return BLK_QC_T_NONE
;
1834 if (blk_mq_sched_bio_merge(q
, bio
))
1835 return BLK_QC_T_NONE
;
1837 rq_qos_throttle(q
, bio
, NULL
);
1839 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1841 rq
= blk_mq_get_request(q
, bio
, bio
->bi_opf
, &data
);
1842 if (unlikely(!rq
)) {
1843 rq_qos_cleanup(q
, bio
);
1844 if (bio
->bi_opf
& REQ_NOWAIT
)
1845 bio_wouldblock_error(bio
);
1846 return BLK_QC_T_NONE
;
1849 rq_qos_track(q
, rq
, bio
);
1851 cookie
= request_to_qc_t(data
.hctx
, rq
);
1853 plug
= current
->plug
;
1854 if (unlikely(is_flush_fua
)) {
1855 blk_mq_put_ctx(data
.ctx
);
1856 blk_mq_bio_to_request(rq
, bio
);
1858 /* bypass scheduler for flush rq */
1859 blk_insert_flush(rq
);
1860 blk_mq_run_hw_queue(data
.hctx
, true);
1861 } else if (plug
&& q
->nr_hw_queues
== 1) {
1862 struct request
*last
= NULL
;
1864 blk_mq_put_ctx(data
.ctx
);
1865 blk_mq_bio_to_request(rq
, bio
);
1868 * @request_count may become stale because of schedule
1869 * out, so check the list again.
1871 if (list_empty(&plug
->mq_list
))
1873 else if (blk_queue_nomerges(q
))
1874 request_count
= blk_plug_queued_count(q
);
1877 trace_block_plug(q
);
1879 last
= list_entry_rq(plug
->mq_list
.prev
);
1881 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1882 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1883 blk_flush_plug_list(plug
, false);
1884 trace_block_plug(q
);
1887 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1888 } else if (plug
&& !blk_queue_nomerges(q
)) {
1889 blk_mq_bio_to_request(rq
, bio
);
1892 * We do limited plugging. If the bio can be merged, do that.
1893 * Otherwise the existing request in the plug list will be
1894 * issued. So the plug list will have one request at most
1895 * The plug list might get flushed before this. If that happens,
1896 * the plug list is empty, and same_queue_rq is invalid.
1898 if (list_empty(&plug
->mq_list
))
1899 same_queue_rq
= NULL
;
1901 list_del_init(&same_queue_rq
->queuelist
);
1902 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1904 blk_mq_put_ctx(data
.ctx
);
1906 if (same_queue_rq
) {
1907 data
.hctx
= blk_mq_map_queue(q
,
1908 same_queue_rq
->mq_ctx
->cpu
);
1909 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1912 } else if ((q
->nr_hw_queues
> 1 && is_sync
) || (!q
->elevator
&&
1913 !data
.hctx
->dispatch_busy
)) {
1914 blk_mq_put_ctx(data
.ctx
);
1915 blk_mq_bio_to_request(rq
, bio
);
1916 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1918 blk_mq_put_ctx(data
.ctx
);
1919 blk_mq_bio_to_request(rq
, bio
);
1920 blk_mq_sched_insert_request(rq
, false, true, true);
1926 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1927 unsigned int hctx_idx
)
1931 if (tags
->rqs
&& set
->ops
->exit_request
) {
1934 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1935 struct request
*rq
= tags
->static_rqs
[i
];
1939 set
->ops
->exit_request(set
, rq
, hctx_idx
);
1940 tags
->static_rqs
[i
] = NULL
;
1944 while (!list_empty(&tags
->page_list
)) {
1945 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1946 list_del_init(&page
->lru
);
1948 * Remove kmemleak object previously allocated in
1949 * blk_mq_init_rq_map().
1951 kmemleak_free(page_address(page
));
1952 __free_pages(page
, page
->private);
1956 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1960 kfree(tags
->static_rqs
);
1961 tags
->static_rqs
= NULL
;
1963 blk_mq_free_tags(tags
);
1966 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1967 unsigned int hctx_idx
,
1968 unsigned int nr_tags
,
1969 unsigned int reserved_tags
)
1971 struct blk_mq_tags
*tags
;
1974 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1975 if (node
== NUMA_NO_NODE
)
1976 node
= set
->numa_node
;
1978 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1979 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1983 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
1984 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1987 blk_mq_free_tags(tags
);
1991 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
1992 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1994 if (!tags
->static_rqs
) {
1996 blk_mq_free_tags(tags
);
2003 static size_t order_to_size(unsigned int order
)
2005 return (size_t)PAGE_SIZE
<< order
;
2008 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
2009 unsigned int hctx_idx
, int node
)
2013 if (set
->ops
->init_request
) {
2014 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
2019 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
2023 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2024 unsigned int hctx_idx
, unsigned int depth
)
2026 unsigned int i
, j
, entries_per_page
, max_order
= 4;
2027 size_t rq_size
, left
;
2030 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
2031 if (node
== NUMA_NO_NODE
)
2032 node
= set
->numa_node
;
2034 INIT_LIST_HEAD(&tags
->page_list
);
2037 * rq_size is the size of the request plus driver payload, rounded
2038 * to the cacheline size
2040 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
2042 left
= rq_size
* depth
;
2044 for (i
= 0; i
< depth
; ) {
2045 int this_order
= max_order
;
2050 while (this_order
&& left
< order_to_size(this_order
- 1))
2054 page
= alloc_pages_node(node
,
2055 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
2061 if (order_to_size(this_order
) < rq_size
)
2068 page
->private = this_order
;
2069 list_add_tail(&page
->lru
, &tags
->page_list
);
2071 p
= page_address(page
);
2073 * Allow kmemleak to scan these pages as they contain pointers
2074 * to additional allocations like via ops->init_request().
2076 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2077 entries_per_page
= order_to_size(this_order
) / rq_size
;
2078 to_do
= min(entries_per_page
, depth
- i
);
2079 left
-= to_do
* rq_size
;
2080 for (j
= 0; j
< to_do
; j
++) {
2081 struct request
*rq
= p
;
2083 tags
->static_rqs
[i
] = rq
;
2084 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
2085 tags
->static_rqs
[i
] = NULL
;
2096 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2101 * 'cpu' is going away. splice any existing rq_list entries from this
2102 * software queue to the hw queue dispatch list, and ensure that it
2105 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2107 struct blk_mq_hw_ctx
*hctx
;
2108 struct blk_mq_ctx
*ctx
;
2111 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2112 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2114 spin_lock(&ctx
->lock
);
2115 if (!list_empty(&ctx
->rq_list
)) {
2116 list_splice_init(&ctx
->rq_list
, &tmp
);
2117 blk_mq_hctx_clear_pending(hctx
, ctx
);
2119 spin_unlock(&ctx
->lock
);
2121 if (list_empty(&tmp
))
2124 spin_lock(&hctx
->lock
);
2125 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2126 spin_unlock(&hctx
->lock
);
2128 blk_mq_run_hw_queue(hctx
, true);
2132 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2134 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2138 /* hctx->ctxs will be freed in queue's release handler */
2139 static void blk_mq_exit_hctx(struct request_queue
*q
,
2140 struct blk_mq_tag_set
*set
,
2141 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2143 blk_mq_debugfs_unregister_hctx(hctx
);
2145 if (blk_mq_hw_queue_mapped(hctx
))
2146 blk_mq_tag_idle(hctx
);
2148 if (set
->ops
->exit_request
)
2149 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2151 if (set
->ops
->exit_hctx
)
2152 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2154 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2155 cleanup_srcu_struct(hctx
->srcu
);
2157 blk_mq_remove_cpuhp(hctx
);
2158 blk_free_flush_queue(hctx
->fq
);
2159 sbitmap_free(&hctx
->ctx_map
);
2162 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2163 struct blk_mq_tag_set
*set
, int nr_queue
)
2165 struct blk_mq_hw_ctx
*hctx
;
2168 queue_for_each_hw_ctx(q
, hctx
, i
) {
2171 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2175 static int blk_mq_init_hctx(struct request_queue
*q
,
2176 struct blk_mq_tag_set
*set
,
2177 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2181 node
= hctx
->numa_node
;
2182 if (node
== NUMA_NO_NODE
)
2183 node
= hctx
->numa_node
= set
->numa_node
;
2185 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2186 spin_lock_init(&hctx
->lock
);
2187 INIT_LIST_HEAD(&hctx
->dispatch
);
2189 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
2191 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2193 hctx
->tags
= set
->tags
[hctx_idx
];
2196 * Allocate space for all possible cpus to avoid allocation at
2199 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2202 goto unregister_cpu_notifier
;
2204 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
2210 spin_lock_init(&hctx
->dispatch_wait_lock
);
2211 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2212 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2214 if (set
->ops
->init_hctx
&&
2215 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2218 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
2222 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
, node
))
2225 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2226 init_srcu_struct(hctx
->srcu
);
2228 blk_mq_debugfs_register_hctx(q
, hctx
);
2235 if (set
->ops
->exit_hctx
)
2236 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2238 sbitmap_free(&hctx
->ctx_map
);
2241 unregister_cpu_notifier
:
2242 blk_mq_remove_cpuhp(hctx
);
2246 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2247 unsigned int nr_hw_queues
)
2251 for_each_possible_cpu(i
) {
2252 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2253 struct blk_mq_hw_ctx
*hctx
;
2256 spin_lock_init(&__ctx
->lock
);
2257 INIT_LIST_HEAD(&__ctx
->rq_list
);
2261 * Set local node, IFF we have more than one hw queue. If
2262 * not, we remain on the home node of the device
2264 hctx
= blk_mq_map_queue(q
, i
);
2265 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2266 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2270 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2274 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2275 set
->queue_depth
, set
->reserved_tags
);
2276 if (!set
->tags
[hctx_idx
])
2279 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2284 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2285 set
->tags
[hctx_idx
] = NULL
;
2289 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2290 unsigned int hctx_idx
)
2292 if (set
->tags
[hctx_idx
]) {
2293 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2294 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2295 set
->tags
[hctx_idx
] = NULL
;
2299 static void blk_mq_map_swqueue(struct request_queue
*q
)
2301 unsigned int i
, hctx_idx
;
2302 struct blk_mq_hw_ctx
*hctx
;
2303 struct blk_mq_ctx
*ctx
;
2304 struct blk_mq_tag_set
*set
= q
->tag_set
;
2307 * Avoid others reading imcomplete hctx->cpumask through sysfs
2309 mutex_lock(&q
->sysfs_lock
);
2311 queue_for_each_hw_ctx(q
, hctx
, i
) {
2312 cpumask_clear(hctx
->cpumask
);
2314 hctx
->dispatch_from
= NULL
;
2318 * Map software to hardware queues.
2320 * If the cpu isn't present, the cpu is mapped to first hctx.
2322 for_each_possible_cpu(i
) {
2323 hctx_idx
= q
->mq_map
[i
];
2324 /* unmapped hw queue can be remapped after CPU topo changed */
2325 if (!set
->tags
[hctx_idx
] &&
2326 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2328 * If tags initialization fail for some hctx,
2329 * that hctx won't be brought online. In this
2330 * case, remap the current ctx to hctx[0] which
2331 * is guaranteed to always have tags allocated
2336 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2337 hctx
= blk_mq_map_queue(q
, i
);
2339 cpumask_set_cpu(i
, hctx
->cpumask
);
2340 ctx
->index_hw
= hctx
->nr_ctx
;
2341 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2344 mutex_unlock(&q
->sysfs_lock
);
2346 queue_for_each_hw_ctx(q
, hctx
, i
) {
2348 * If no software queues are mapped to this hardware queue,
2349 * disable it and free the request entries.
2351 if (!hctx
->nr_ctx
) {
2352 /* Never unmap queue 0. We need it as a
2353 * fallback in case of a new remap fails
2356 if (i
&& set
->tags
[i
])
2357 blk_mq_free_map_and_requests(set
, i
);
2363 hctx
->tags
= set
->tags
[i
];
2364 WARN_ON(!hctx
->tags
);
2367 * Set the map size to the number of mapped software queues.
2368 * This is more accurate and more efficient than looping
2369 * over all possibly mapped software queues.
2371 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2374 * Initialize batch roundrobin counts
2376 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2377 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2382 * Caller needs to ensure that we're either frozen/quiesced, or that
2383 * the queue isn't live yet.
2385 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2387 struct blk_mq_hw_ctx
*hctx
;
2390 queue_for_each_hw_ctx(q
, hctx
, i
) {
2392 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2394 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2398 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2401 struct request_queue
*q
;
2403 lockdep_assert_held(&set
->tag_list_lock
);
2405 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2406 blk_mq_freeze_queue(q
);
2407 queue_set_hctx_shared(q
, shared
);
2408 blk_mq_unfreeze_queue(q
);
2412 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2414 struct blk_mq_tag_set
*set
= q
->tag_set
;
2416 mutex_lock(&set
->tag_list_lock
);
2417 list_del_rcu(&q
->tag_set_list
);
2418 if (list_is_singular(&set
->tag_list
)) {
2419 /* just transitioned to unshared */
2420 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2421 /* update existing queue */
2422 blk_mq_update_tag_set_depth(set
, false);
2424 mutex_unlock(&set
->tag_list_lock
);
2425 INIT_LIST_HEAD(&q
->tag_set_list
);
2428 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2429 struct request_queue
*q
)
2433 mutex_lock(&set
->tag_list_lock
);
2436 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2438 if (!list_empty(&set
->tag_list
) &&
2439 !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2440 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2441 /* update existing queue */
2442 blk_mq_update_tag_set_depth(set
, true);
2444 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2445 queue_set_hctx_shared(q
, true);
2446 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2448 mutex_unlock(&set
->tag_list_lock
);
2452 * It is the actual release handler for mq, but we do it from
2453 * request queue's release handler for avoiding use-after-free
2454 * and headache because q->mq_kobj shouldn't have been introduced,
2455 * but we can't group ctx/kctx kobj without it.
2457 void blk_mq_release(struct request_queue
*q
)
2459 struct blk_mq_hw_ctx
*hctx
;
2462 /* hctx kobj stays in hctx */
2463 queue_for_each_hw_ctx(q
, hctx
, i
) {
2466 kobject_put(&hctx
->kobj
);
2471 kfree(q
->queue_hw_ctx
);
2474 * release .mq_kobj and sw queue's kobject now because
2475 * both share lifetime with request queue.
2477 blk_mq_sysfs_deinit(q
);
2479 free_percpu(q
->queue_ctx
);
2482 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2484 struct request_queue
*uninit_q
, *q
;
2486 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
, NULL
);
2488 return ERR_PTR(-ENOMEM
);
2490 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2492 blk_cleanup_queue(uninit_q
);
2496 EXPORT_SYMBOL(blk_mq_init_queue
);
2498 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2500 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2502 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, srcu
),
2503 __alignof__(struct blk_mq_hw_ctx
)) !=
2504 sizeof(struct blk_mq_hw_ctx
));
2506 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2507 hw_ctx_size
+= sizeof(struct srcu_struct
);
2512 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2513 struct request_queue
*q
)
2516 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2518 blk_mq_sysfs_unregister(q
);
2520 /* protect against switching io scheduler */
2521 mutex_lock(&q
->sysfs_lock
);
2522 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2528 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2529 hctxs
[i
] = kzalloc_node(blk_mq_hw_ctx_size(set
),
2534 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2541 atomic_set(&hctxs
[i
]->nr_active
, 0);
2542 hctxs
[i
]->numa_node
= node
;
2543 hctxs
[i
]->queue_num
= i
;
2545 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2546 free_cpumask_var(hctxs
[i
]->cpumask
);
2551 blk_mq_hctx_kobj_init(hctxs
[i
]);
2553 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2554 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2558 blk_mq_free_map_and_requests(set
, j
);
2559 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2560 kobject_put(&hctx
->kobj
);
2565 q
->nr_hw_queues
= i
;
2566 mutex_unlock(&q
->sysfs_lock
);
2567 blk_mq_sysfs_register(q
);
2570 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2571 struct request_queue
*q
)
2573 /* mark the queue as mq asap */
2574 q
->mq_ops
= set
->ops
;
2576 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2577 blk_mq_poll_stats_bkt
,
2578 BLK_MQ_POLL_STATS_BKTS
, q
);
2582 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2586 /* init q->mq_kobj and sw queues' kobjects */
2587 blk_mq_sysfs_init(q
);
2589 q
->queue_hw_ctx
= kcalloc_node(nr_cpu_ids
, sizeof(*(q
->queue_hw_ctx
)),
2590 GFP_KERNEL
, set
->numa_node
);
2591 if (!q
->queue_hw_ctx
)
2594 q
->mq_map
= set
->mq_map
;
2596 blk_mq_realloc_hw_ctxs(set
, q
);
2597 if (!q
->nr_hw_queues
)
2600 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2601 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2603 q
->nr_queues
= nr_cpu_ids
;
2605 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2607 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2608 queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE
, q
);
2610 q
->sg_reserved_size
= INT_MAX
;
2612 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2613 INIT_LIST_HEAD(&q
->requeue_list
);
2614 spin_lock_init(&q
->requeue_lock
);
2616 blk_queue_make_request(q
, blk_mq_make_request
);
2617 if (q
->mq_ops
->poll
)
2618 q
->poll_fn
= blk_mq_poll
;
2621 * Do this after blk_queue_make_request() overrides it...
2623 q
->nr_requests
= set
->queue_depth
;
2626 * Default to classic polling
2630 if (set
->ops
->complete
)
2631 blk_queue_softirq_done(q
, set
->ops
->complete
);
2633 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2634 blk_mq_add_queue_tag_set(set
, q
);
2635 blk_mq_map_swqueue(q
);
2637 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2640 ret
= elevator_init_mq(q
);
2642 return ERR_PTR(ret
);
2648 kfree(q
->queue_hw_ctx
);
2650 free_percpu(q
->queue_ctx
);
2653 return ERR_PTR(-ENOMEM
);
2655 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2657 void blk_mq_free_queue(struct request_queue
*q
)
2659 struct blk_mq_tag_set
*set
= q
->tag_set
;
2661 blk_mq_del_queue_tag_set(q
);
2662 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2665 /* Basically redo blk_mq_init_queue with queue frozen */
2666 static void blk_mq_queue_reinit(struct request_queue
*q
)
2668 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2670 blk_mq_debugfs_unregister_hctxs(q
);
2671 blk_mq_sysfs_unregister(q
);
2674 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2675 * we should change hctx numa_node according to the new topology (this
2676 * involves freeing and re-allocating memory, worth doing?)
2678 blk_mq_map_swqueue(q
);
2680 blk_mq_sysfs_register(q
);
2681 blk_mq_debugfs_register_hctxs(q
);
2684 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2688 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2689 if (!__blk_mq_alloc_rq_map(set
, i
))
2696 blk_mq_free_rq_map(set
->tags
[i
]);
2702 * Allocate the request maps associated with this tag_set. Note that this
2703 * may reduce the depth asked for, if memory is tight. set->queue_depth
2704 * will be updated to reflect the allocated depth.
2706 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2711 depth
= set
->queue_depth
;
2713 err
= __blk_mq_alloc_rq_maps(set
);
2717 set
->queue_depth
>>= 1;
2718 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2722 } while (set
->queue_depth
);
2724 if (!set
->queue_depth
|| err
) {
2725 pr_err("blk-mq: failed to allocate request map\n");
2729 if (depth
!= set
->queue_depth
)
2730 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2731 depth
, set
->queue_depth
);
2736 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2738 if (set
->ops
->map_queues
) {
2740 * transport .map_queues is usually done in the following
2743 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2744 * mask = get_cpu_mask(queue)
2745 * for_each_cpu(cpu, mask)
2746 * set->mq_map[cpu] = queue;
2749 * When we need to remap, the table has to be cleared for
2750 * killing stale mapping since one CPU may not be mapped
2753 blk_mq_clear_mq_map(set
);
2755 return set
->ops
->map_queues(set
);
2757 return blk_mq_map_queues(set
);
2761 * Alloc a tag set to be associated with one or more request queues.
2762 * May fail with EINVAL for various error conditions. May adjust the
2763 * requested depth down, if it's too large. In that case, the set
2764 * value will be stored in set->queue_depth.
2766 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2770 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2772 if (!set
->nr_hw_queues
)
2774 if (!set
->queue_depth
)
2776 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2779 if (!set
->ops
->queue_rq
)
2782 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
2785 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2786 pr_info("blk-mq: reduced tag depth to %u\n",
2788 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2792 * If a crashdump is active, then we are potentially in a very
2793 * memory constrained environment. Limit us to 1 queue and
2794 * 64 tags to prevent using too much memory.
2796 if (is_kdump_kernel()) {
2797 set
->nr_hw_queues
= 1;
2798 set
->queue_depth
= min(64U, set
->queue_depth
);
2801 * There is no use for more h/w queues than cpus.
2803 if (set
->nr_hw_queues
> nr_cpu_ids
)
2804 set
->nr_hw_queues
= nr_cpu_ids
;
2806 set
->tags
= kcalloc_node(nr_cpu_ids
, sizeof(struct blk_mq_tags
*),
2807 GFP_KERNEL
, set
->numa_node
);
2812 set
->mq_map
= kcalloc_node(nr_cpu_ids
, sizeof(*set
->mq_map
),
2813 GFP_KERNEL
, set
->numa_node
);
2817 ret
= blk_mq_update_queue_map(set
);
2819 goto out_free_mq_map
;
2821 ret
= blk_mq_alloc_rq_maps(set
);
2823 goto out_free_mq_map
;
2825 mutex_init(&set
->tag_list_lock
);
2826 INIT_LIST_HEAD(&set
->tag_list
);
2838 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2840 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2844 for (i
= 0; i
< nr_cpu_ids
; i
++)
2845 blk_mq_free_map_and_requests(set
, i
);
2853 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2855 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2857 struct blk_mq_tag_set
*set
= q
->tag_set
;
2858 struct blk_mq_hw_ctx
*hctx
;
2864 blk_mq_freeze_queue(q
);
2865 blk_mq_quiesce_queue(q
);
2868 queue_for_each_hw_ctx(q
, hctx
, i
) {
2872 * If we're using an MQ scheduler, just update the scheduler
2873 * queue depth. This is similar to what the old code would do.
2875 if (!hctx
->sched_tags
) {
2876 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
2879 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2887 q
->nr_requests
= nr
;
2889 blk_mq_unquiesce_queue(q
);
2890 blk_mq_unfreeze_queue(q
);
2896 * request_queue and elevator_type pair.
2897 * It is just used by __blk_mq_update_nr_hw_queues to cache
2898 * the elevator_type associated with a request_queue.
2900 struct blk_mq_qe_pair
{
2901 struct list_head node
;
2902 struct request_queue
*q
;
2903 struct elevator_type
*type
;
2907 * Cache the elevator_type in qe pair list and switch the
2908 * io scheduler to 'none'
2910 static bool blk_mq_elv_switch_none(struct list_head
*head
,
2911 struct request_queue
*q
)
2913 struct blk_mq_qe_pair
*qe
;
2918 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
2922 INIT_LIST_HEAD(&qe
->node
);
2924 qe
->type
= q
->elevator
->type
;
2925 list_add(&qe
->node
, head
);
2927 mutex_lock(&q
->sysfs_lock
);
2929 * After elevator_switch_mq, the previous elevator_queue will be
2930 * released by elevator_release. The reference of the io scheduler
2931 * module get by elevator_get will also be put. So we need to get
2932 * a reference of the io scheduler module here to prevent it to be
2935 __module_get(qe
->type
->elevator_owner
);
2936 elevator_switch_mq(q
, NULL
);
2937 mutex_unlock(&q
->sysfs_lock
);
2942 static void blk_mq_elv_switch_back(struct list_head
*head
,
2943 struct request_queue
*q
)
2945 struct blk_mq_qe_pair
*qe
;
2946 struct elevator_type
*t
= NULL
;
2948 list_for_each_entry(qe
, head
, node
)
2957 list_del(&qe
->node
);
2960 mutex_lock(&q
->sysfs_lock
);
2961 elevator_switch_mq(q
, t
);
2962 mutex_unlock(&q
->sysfs_lock
);
2965 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
2968 struct request_queue
*q
;
2971 lockdep_assert_held(&set
->tag_list_lock
);
2973 if (nr_hw_queues
> nr_cpu_ids
)
2974 nr_hw_queues
= nr_cpu_ids
;
2975 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2978 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2979 blk_mq_freeze_queue(q
);
2981 * Sync with blk_mq_queue_tag_busy_iter.
2985 * Switch IO scheduler to 'none', cleaning up the data associated
2986 * with the previous scheduler. We will switch back once we are done
2987 * updating the new sw to hw queue mappings.
2989 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2990 if (!blk_mq_elv_switch_none(&head
, q
))
2993 set
->nr_hw_queues
= nr_hw_queues
;
2994 blk_mq_update_queue_map(set
);
2995 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2996 blk_mq_realloc_hw_ctxs(set
, q
);
2997 blk_mq_queue_reinit(q
);
3001 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3002 blk_mq_elv_switch_back(&head
, q
);
3004 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3005 blk_mq_unfreeze_queue(q
);
3008 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
3010 mutex_lock(&set
->tag_list_lock
);
3011 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
3012 mutex_unlock(&set
->tag_list_lock
);
3014 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
3016 /* Enable polling stats and return whether they were already enabled. */
3017 static bool blk_poll_stats_enable(struct request_queue
*q
)
3019 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3020 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS
, q
))
3022 blk_stat_add_callback(q
, q
->poll_cb
);
3026 static void blk_mq_poll_stats_start(struct request_queue
*q
)
3029 * We don't arm the callback if polling stats are not enabled or the
3030 * callback is already active.
3032 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3033 blk_stat_is_active(q
->poll_cb
))
3036 blk_stat_activate_msecs(q
->poll_cb
, 100);
3039 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
3041 struct request_queue
*q
= cb
->data
;
3044 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
3045 if (cb
->stat
[bucket
].nr_samples
)
3046 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
3050 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
3051 struct blk_mq_hw_ctx
*hctx
,
3054 unsigned long ret
= 0;
3058 * If stats collection isn't on, don't sleep but turn it on for
3061 if (!blk_poll_stats_enable(q
))
3065 * As an optimistic guess, use half of the mean service time
3066 * for this type of request. We can (and should) make this smarter.
3067 * For instance, if the completion latencies are tight, we can
3068 * get closer than just half the mean. This is especially
3069 * important on devices where the completion latencies are longer
3070 * than ~10 usec. We do use the stats for the relevant IO size
3071 * if available which does lead to better estimates.
3073 bucket
= blk_mq_poll_stats_bkt(rq
);
3077 if (q
->poll_stat
[bucket
].nr_samples
)
3078 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
3083 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
3084 struct blk_mq_hw_ctx
*hctx
,
3087 struct hrtimer_sleeper hs
;
3088 enum hrtimer_mode mode
;
3092 if (rq
->rq_flags
& RQF_MQ_POLL_SLEPT
)
3098 * -1: don't ever hybrid sleep
3099 * 0: use half of prev avg
3100 * >0: use this specific value
3102 if (q
->poll_nsec
== -1)
3104 else if (q
->poll_nsec
> 0)
3105 nsecs
= q
->poll_nsec
;
3107 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
3112 rq
->rq_flags
|= RQF_MQ_POLL_SLEPT
;
3115 * This will be replaced with the stats tracking code, using
3116 * 'avg_completion_time / 2' as the pre-sleep target.
3120 mode
= HRTIMER_MODE_REL
;
3121 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
3122 hrtimer_set_expires(&hs
.timer
, kt
);
3124 hrtimer_init_sleeper(&hs
, current
);
3126 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
3128 set_current_state(TASK_UNINTERRUPTIBLE
);
3129 hrtimer_start_expires(&hs
.timer
, mode
);
3132 hrtimer_cancel(&hs
.timer
);
3133 mode
= HRTIMER_MODE_ABS
;
3134 } while (hs
.task
&& !signal_pending(current
));
3136 __set_current_state(TASK_RUNNING
);
3137 destroy_hrtimer_on_stack(&hs
.timer
);
3141 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
3143 struct request_queue
*q
= hctx
->queue
;
3147 * If we sleep, have the caller restart the poll loop to reset
3148 * the state. Like for the other success return cases, the
3149 * caller is responsible for checking if the IO completed. If
3150 * the IO isn't complete, we'll get called again and will go
3151 * straight to the busy poll loop.
3153 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
3156 hctx
->poll_considered
++;
3158 state
= current
->state
;
3159 while (!need_resched()) {
3162 hctx
->poll_invoked
++;
3164 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
3166 hctx
->poll_success
++;
3167 set_current_state(TASK_RUNNING
);
3171 if (signal_pending_state(state
, current
))
3172 set_current_state(TASK_RUNNING
);
3174 if (current
->state
== TASK_RUNNING
)
3181 __set_current_state(TASK_RUNNING
);
3185 static bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
3187 struct blk_mq_hw_ctx
*hctx
;
3190 if (!test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3193 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3194 if (!blk_qc_t_is_internal(cookie
))
3195 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3197 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3199 * With scheduling, if the request has completed, we'll
3200 * get a NULL return here, as we clear the sched tag when
3201 * that happens. The request still remains valid, like always,
3202 * so we should be safe with just the NULL check.
3208 return __blk_mq_poll(hctx
, rq
);
3211 static int __init
blk_mq_init(void)
3213 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3214 blk_mq_hctx_notify_dead
);
3217 subsys_initcall(blk_mq_init
);