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
;
99 * index[0] counts the specific partition that was asked for. index[1]
100 * counts the ones that are active on the whole device, so increment
101 * that if mi->part is indeed a partition, and not a whole device.
103 if (rq
->part
== mi
->part
)
105 if (mi
->part
->partno
)
109 void blk_mq_in_flight(struct request_queue
*q
, struct hd_struct
*part
,
110 unsigned int inflight
[2])
112 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
114 inflight
[0] = inflight
[1] = 0;
115 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
118 static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx
*hctx
,
119 struct request
*rq
, void *priv
,
122 struct mq_inflight
*mi
= priv
;
124 if (rq
->part
== mi
->part
)
125 mi
->inflight
[rq_data_dir(rq
)]++;
128 void blk_mq_in_flight_rw(struct request_queue
*q
, struct hd_struct
*part
,
129 unsigned int inflight
[2])
131 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
133 inflight
[0] = inflight
[1] = 0;
134 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight_rw
, &mi
);
137 void blk_freeze_queue_start(struct request_queue
*q
)
141 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
142 if (freeze_depth
== 1) {
143 percpu_ref_kill(&q
->q_usage_counter
);
145 blk_mq_run_hw_queues(q
, false);
148 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
150 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
152 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
154 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
156 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
157 unsigned long timeout
)
159 return wait_event_timeout(q
->mq_freeze_wq
,
160 percpu_ref_is_zero(&q
->q_usage_counter
),
163 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
166 * Guarantee no request is in use, so we can change any data structure of
167 * the queue afterward.
169 void blk_freeze_queue(struct request_queue
*q
)
172 * In the !blk_mq case we are only calling this to kill the
173 * q_usage_counter, otherwise this increases the freeze depth
174 * and waits for it to return to zero. For this reason there is
175 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
176 * exported to drivers as the only user for unfreeze is blk_mq.
178 blk_freeze_queue_start(q
);
181 blk_mq_freeze_queue_wait(q
);
184 void blk_mq_freeze_queue(struct request_queue
*q
)
187 * ...just an alias to keep freeze and unfreeze actions balanced
188 * in the blk_mq_* namespace
192 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
194 void blk_mq_unfreeze_queue(struct request_queue
*q
)
198 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
199 WARN_ON_ONCE(freeze_depth
< 0);
201 percpu_ref_reinit(&q
->q_usage_counter
);
202 wake_up_all(&q
->mq_freeze_wq
);
205 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
208 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
209 * mpt3sas driver such that this function can be removed.
211 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
213 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
215 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
218 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
221 * Note: this function does not prevent that the struct request end_io()
222 * callback function is invoked. Once this function is returned, we make
223 * sure no dispatch can happen until the queue is unquiesced via
224 * blk_mq_unquiesce_queue().
226 void blk_mq_quiesce_queue(struct request_queue
*q
)
228 struct blk_mq_hw_ctx
*hctx
;
232 blk_mq_quiesce_queue_nowait(q
);
234 queue_for_each_hw_ctx(q
, hctx
, i
) {
235 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
236 synchronize_srcu(hctx
->srcu
);
243 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
246 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
249 * This function recovers queue into the state before quiescing
250 * which is done by blk_mq_quiesce_queue.
252 void blk_mq_unquiesce_queue(struct request_queue
*q
)
254 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
256 /* dispatch requests which are inserted during quiescing */
257 blk_mq_run_hw_queues(q
, true);
259 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
261 void blk_mq_wake_waiters(struct request_queue
*q
)
263 struct blk_mq_hw_ctx
*hctx
;
266 queue_for_each_hw_ctx(q
, hctx
, i
)
267 if (blk_mq_hw_queue_mapped(hctx
))
268 blk_mq_tag_wakeup_all(hctx
->tags
, true);
271 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
273 return blk_mq_has_free_tags(hctx
->tags
);
275 EXPORT_SYMBOL(blk_mq_can_queue
);
277 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
278 unsigned int tag
, unsigned int op
)
280 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
281 struct request
*rq
= tags
->static_rqs
[tag
];
282 req_flags_t rq_flags
= 0;
284 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
286 rq
->internal_tag
= tag
;
288 if (blk_mq_tag_busy(data
->hctx
)) {
289 rq_flags
= RQF_MQ_INFLIGHT
;
290 atomic_inc(&data
->hctx
->nr_active
);
293 rq
->internal_tag
= -1;
294 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
297 /* csd/requeue_work/fifo_time is initialized before use */
299 rq
->mq_ctx
= data
->ctx
;
300 rq
->rq_flags
= rq_flags
;
303 if (data
->flags
& BLK_MQ_REQ_PREEMPT
)
304 rq
->rq_flags
|= RQF_PREEMPT
;
305 if (blk_queue_io_stat(data
->q
))
306 rq
->rq_flags
|= RQF_IO_STAT
;
307 INIT_LIST_HEAD(&rq
->queuelist
);
308 INIT_HLIST_NODE(&rq
->hash
);
309 RB_CLEAR_NODE(&rq
->rb_node
);
312 rq
->start_time_ns
= ktime_get_ns();
313 rq
->io_start_time_ns
= 0;
314 rq
->nr_phys_segments
= 0;
315 #if defined(CONFIG_BLK_DEV_INTEGRITY)
316 rq
->nr_integrity_segments
= 0;
319 /* tag was already set */
323 INIT_LIST_HEAD(&rq
->timeout_list
);
327 rq
->end_io_data
= NULL
;
330 #ifdef CONFIG_BLK_CGROUP
334 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
335 refcount_set(&rq
->ref
, 1);
339 static struct request
*blk_mq_get_request(struct request_queue
*q
,
340 struct bio
*bio
, unsigned int op
,
341 struct blk_mq_alloc_data
*data
)
343 struct elevator_queue
*e
= q
->elevator
;
346 bool put_ctx_on_error
= false;
348 blk_queue_enter_live(q
);
350 if (likely(!data
->ctx
)) {
351 data
->ctx
= blk_mq_get_ctx(q
);
352 put_ctx_on_error
= true;
354 if (likely(!data
->hctx
))
355 data
->hctx
= blk_mq_map_queue(q
, data
->ctx
->cpu
);
357 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
360 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
363 * Flush requests are special and go directly to the
364 * dispatch list. Don't include reserved tags in the
365 * limiting, as it isn't useful.
367 if (!op_is_flush(op
) && e
->type
->ops
.mq
.limit_depth
&&
368 !(data
->flags
& BLK_MQ_REQ_RESERVED
))
369 e
->type
->ops
.mq
.limit_depth(op
, data
);
372 tag
= blk_mq_get_tag(data
);
373 if (tag
== BLK_MQ_TAG_FAIL
) {
374 if (put_ctx_on_error
) {
375 blk_mq_put_ctx(data
->ctx
);
382 rq
= blk_mq_rq_ctx_init(data
, tag
, op
);
383 if (!op_is_flush(op
)) {
385 if (e
&& e
->type
->ops
.mq
.prepare_request
) {
386 if (e
->type
->icq_cache
&& rq_ioc(bio
))
387 blk_mq_sched_assign_ioc(rq
, bio
);
389 e
->type
->ops
.mq
.prepare_request(rq
, bio
);
390 rq
->rq_flags
|= RQF_ELVPRIV
;
393 data
->hctx
->queued
++;
397 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
398 blk_mq_req_flags_t flags
)
400 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
404 ret
= blk_queue_enter(q
, flags
);
408 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
412 return ERR_PTR(-EWOULDBLOCK
);
414 blk_mq_put_ctx(alloc_data
.ctx
);
417 rq
->__sector
= (sector_t
) -1;
418 rq
->bio
= rq
->biotail
= NULL
;
421 EXPORT_SYMBOL(blk_mq_alloc_request
);
423 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
424 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
426 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
432 * If the tag allocator sleeps we could get an allocation for a
433 * different hardware context. No need to complicate the low level
434 * allocator for this for the rare use case of a command tied to
437 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
438 return ERR_PTR(-EINVAL
);
440 if (hctx_idx
>= q
->nr_hw_queues
)
441 return ERR_PTR(-EIO
);
443 ret
= blk_queue_enter(q
, flags
);
448 * Check if the hardware context is actually mapped to anything.
449 * If not tell the caller that it should skip this queue.
451 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
452 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
454 return ERR_PTR(-EXDEV
);
456 cpu
= cpumask_first_and(alloc_data
.hctx
->cpumask
, cpu_online_mask
);
457 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
459 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
463 return ERR_PTR(-EWOULDBLOCK
);
467 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
469 static void __blk_mq_free_request(struct request
*rq
)
471 struct request_queue
*q
= rq
->q
;
472 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
473 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
474 const int sched_tag
= rq
->internal_tag
;
477 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
479 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
480 blk_mq_sched_restart(hctx
);
484 void blk_mq_free_request(struct request
*rq
)
486 struct request_queue
*q
= rq
->q
;
487 struct elevator_queue
*e
= q
->elevator
;
488 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
489 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
491 if (rq
->rq_flags
& RQF_ELVPRIV
) {
492 if (e
&& e
->type
->ops
.mq
.finish_request
)
493 e
->type
->ops
.mq
.finish_request(rq
);
495 put_io_context(rq
->elv
.icq
->ioc
);
500 ctx
->rq_completed
[rq_is_sync(rq
)]++;
501 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
502 atomic_dec(&hctx
->nr_active
);
504 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
505 laptop_io_completion(q
->backing_dev_info
);
507 wbt_done(q
->rq_wb
, rq
);
510 blk_put_rl(blk_rq_rl(rq
));
512 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
513 if (refcount_dec_and_test(&rq
->ref
))
514 __blk_mq_free_request(rq
);
516 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
518 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
520 u64 now
= ktime_get_ns();
522 if (rq
->rq_flags
& RQF_STATS
) {
523 blk_mq_poll_stats_start(rq
->q
);
524 blk_stat_add(rq
, now
);
527 blk_account_io_done(rq
, now
);
530 wbt_done(rq
->q
->rq_wb
, rq
);
531 rq
->end_io(rq
, error
);
533 if (unlikely(blk_bidi_rq(rq
)))
534 blk_mq_free_request(rq
->next_rq
);
535 blk_mq_free_request(rq
);
538 EXPORT_SYMBOL(__blk_mq_end_request
);
540 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
542 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
544 __blk_mq_end_request(rq
, error
);
546 EXPORT_SYMBOL(blk_mq_end_request
);
548 static void __blk_mq_complete_request_remote(void *data
)
550 struct request
*rq
= data
;
552 rq
->q
->softirq_done_fn(rq
);
555 static void __blk_mq_complete_request(struct request
*rq
)
557 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
561 if (cmpxchg(&rq
->state
, MQ_RQ_IN_FLIGHT
, MQ_RQ_COMPLETE
) !=
565 if (rq
->internal_tag
!= -1)
566 blk_mq_sched_completed_request(rq
);
568 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
569 rq
->q
->softirq_done_fn(rq
);
574 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
575 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
577 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
578 rq
->csd
.func
= __blk_mq_complete_request_remote
;
581 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
583 rq
->q
->softirq_done_fn(rq
);
588 static void hctx_unlock(struct blk_mq_hw_ctx
*hctx
, int srcu_idx
)
589 __releases(hctx
->srcu
)
591 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
594 srcu_read_unlock(hctx
->srcu
, srcu_idx
);
597 static void hctx_lock(struct blk_mq_hw_ctx
*hctx
, int *srcu_idx
)
598 __acquires(hctx
->srcu
)
600 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
601 /* shut up gcc false positive */
605 *srcu_idx
= srcu_read_lock(hctx
->srcu
);
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 if (unlikely(blk_should_fake_timeout(rq
->q
)))
620 __blk_mq_complete_request(rq
);
622 EXPORT_SYMBOL(blk_mq_complete_request
);
624 int blk_mq_request_started(struct request
*rq
)
626 return blk_mq_rq_state(rq
) != MQ_RQ_IDLE
;
628 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
630 void blk_mq_start_request(struct request
*rq
)
632 struct request_queue
*q
= rq
->q
;
634 blk_mq_sched_started_request(rq
);
636 trace_block_rq_issue(q
, rq
);
638 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
639 rq
->io_start_time_ns
= ktime_get_ns();
640 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
641 rq
->throtl_size
= blk_rq_sectors(rq
);
643 rq
->rq_flags
|= RQF_STATS
;
644 wbt_issue(q
->rq_wb
, rq
);
647 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
650 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
652 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
654 * Make sure space for the drain appears. We know we can do
655 * this because max_hw_segments has been adjusted to be one
656 * fewer than the device can handle.
658 rq
->nr_phys_segments
++;
661 EXPORT_SYMBOL(blk_mq_start_request
);
663 static void __blk_mq_requeue_request(struct request
*rq
)
665 struct request_queue
*q
= rq
->q
;
667 blk_mq_put_driver_tag(rq
);
669 trace_block_rq_requeue(q
, rq
);
670 wbt_requeue(q
->rq_wb
, rq
);
672 if (blk_mq_request_started(rq
)) {
673 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
674 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
675 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
676 rq
->nr_phys_segments
--;
680 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
682 __blk_mq_requeue_request(rq
);
684 /* this request will be re-inserted to io scheduler queue */
685 blk_mq_sched_requeue_request(rq
);
687 BUG_ON(blk_queued_rq(rq
));
688 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
690 EXPORT_SYMBOL(blk_mq_requeue_request
);
692 static void blk_mq_requeue_work(struct work_struct
*work
)
694 struct request_queue
*q
=
695 container_of(work
, struct request_queue
, requeue_work
.work
);
697 struct request
*rq
, *next
;
699 spin_lock_irq(&q
->requeue_lock
);
700 list_splice_init(&q
->requeue_list
, &rq_list
);
701 spin_unlock_irq(&q
->requeue_lock
);
703 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
704 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
707 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
708 list_del_init(&rq
->queuelist
);
709 blk_mq_sched_insert_request(rq
, true, false, false);
712 while (!list_empty(&rq_list
)) {
713 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
714 list_del_init(&rq
->queuelist
);
715 blk_mq_sched_insert_request(rq
, false, false, false);
718 blk_mq_run_hw_queues(q
, false);
721 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
722 bool kick_requeue_list
)
724 struct request_queue
*q
= rq
->q
;
728 * We abuse this flag that is otherwise used by the I/O scheduler to
729 * request head insertion from the workqueue.
731 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
733 spin_lock_irqsave(&q
->requeue_lock
, flags
);
735 rq
->rq_flags
|= RQF_SOFTBARRIER
;
736 list_add(&rq
->queuelist
, &q
->requeue_list
);
738 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
740 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
742 if (kick_requeue_list
)
743 blk_mq_kick_requeue_list(q
);
745 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
747 void blk_mq_kick_requeue_list(struct request_queue
*q
)
749 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
751 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
753 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
756 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
757 msecs_to_jiffies(msecs
));
759 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
761 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
763 if (tag
< tags
->nr_tags
) {
764 prefetch(tags
->rqs
[tag
]);
765 return tags
->rqs
[tag
];
770 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
772 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
774 req
->rq_flags
|= RQF_TIMED_OUT
;
775 if (req
->q
->mq_ops
->timeout
) {
776 enum blk_eh_timer_return ret
;
778 ret
= req
->q
->mq_ops
->timeout(req
, reserved
);
779 if (ret
== BLK_EH_DONE
)
781 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
784 req
->rq_flags
&= ~RQF_TIMED_OUT
;
788 static bool blk_mq_req_expired(struct request
*rq
, unsigned long *next
)
790 unsigned long deadline
;
792 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
794 if (rq
->rq_flags
& RQF_TIMED_OUT
)
797 deadline
= blk_rq_deadline(rq
);
798 if (time_after_eq(jiffies
, deadline
))
803 else if (time_after(*next
, deadline
))
808 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
809 struct request
*rq
, void *priv
, bool reserved
)
811 unsigned long *next
= priv
;
814 * Just do a quick check if it is expired before locking the request in
815 * so we're not unnecessarilly synchronizing across CPUs.
817 if (!blk_mq_req_expired(rq
, next
))
821 * We have reason to believe the request may be expired. Take a
822 * reference on the request to lock this request lifetime into its
823 * currently allocated context to prevent it from being reallocated in
824 * the event the completion by-passes this timeout handler.
826 * If the reference was already released, then the driver beat the
827 * timeout handler to posting a natural completion.
829 if (!refcount_inc_not_zero(&rq
->ref
))
833 * The request is now locked and cannot be reallocated underneath the
834 * timeout handler's processing. Re-verify this exact request is truly
835 * expired; if it is not expired, then the request was completed and
836 * reallocated as a new request.
838 if (blk_mq_req_expired(rq
, next
))
839 blk_mq_rq_timed_out(rq
, reserved
);
840 if (refcount_dec_and_test(&rq
->ref
))
841 __blk_mq_free_request(rq
);
844 static void blk_mq_timeout_work(struct work_struct
*work
)
846 struct request_queue
*q
=
847 container_of(work
, struct request_queue
, timeout_work
);
848 unsigned long next
= 0;
849 struct blk_mq_hw_ctx
*hctx
;
852 /* A deadlock might occur if a request is stuck requiring a
853 * timeout at the same time a queue freeze is waiting
854 * completion, since the timeout code would not be able to
855 * acquire the queue reference here.
857 * That's why we don't use blk_queue_enter here; instead, we use
858 * percpu_ref_tryget directly, because we need to be able to
859 * obtain a reference even in the short window between the queue
860 * starting to freeze, by dropping the first reference in
861 * blk_freeze_queue_start, and the moment the last request is
862 * consumed, marked by the instant q_usage_counter reaches
865 if (!percpu_ref_tryget(&q
->q_usage_counter
))
868 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &next
);
871 mod_timer(&q
->timeout
, next
);
874 * Request timeouts are handled as a forward rolling timer. If
875 * we end up here it means that no requests are pending and
876 * also that no request has been pending for a while. Mark
879 queue_for_each_hw_ctx(q
, hctx
, i
) {
880 /* the hctx may be unmapped, so check it here */
881 if (blk_mq_hw_queue_mapped(hctx
))
882 blk_mq_tag_idle(hctx
);
888 struct flush_busy_ctx_data
{
889 struct blk_mq_hw_ctx
*hctx
;
890 struct list_head
*list
;
893 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
895 struct flush_busy_ctx_data
*flush_data
= data
;
896 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
897 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
899 spin_lock(&ctx
->lock
);
900 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
901 sbitmap_clear_bit(sb
, bitnr
);
902 spin_unlock(&ctx
->lock
);
907 * Process software queues that have been marked busy, splicing them
908 * to the for-dispatch
910 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
912 struct flush_busy_ctx_data data
= {
917 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
919 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
921 struct dispatch_rq_data
{
922 struct blk_mq_hw_ctx
*hctx
;
926 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
929 struct dispatch_rq_data
*dispatch_data
= data
;
930 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
931 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
933 spin_lock(&ctx
->lock
);
934 if (!list_empty(&ctx
->rq_list
)) {
935 dispatch_data
->rq
= list_entry_rq(ctx
->rq_list
.next
);
936 list_del_init(&dispatch_data
->rq
->queuelist
);
937 if (list_empty(&ctx
->rq_list
))
938 sbitmap_clear_bit(sb
, bitnr
);
940 spin_unlock(&ctx
->lock
);
942 return !dispatch_data
->rq
;
945 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
946 struct blk_mq_ctx
*start
)
948 unsigned off
= start
? start
->index_hw
: 0;
949 struct dispatch_rq_data data
= {
954 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
955 dispatch_rq_from_ctx
, &data
);
960 static inline unsigned int queued_to_index(unsigned int queued
)
965 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
968 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
971 struct blk_mq_alloc_data data
= {
973 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
974 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
977 might_sleep_if(wait
);
982 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
983 data
.flags
|= BLK_MQ_REQ_RESERVED
;
985 rq
->tag
= blk_mq_get_tag(&data
);
987 if (blk_mq_tag_busy(data
.hctx
)) {
988 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
989 atomic_inc(&data
.hctx
->nr_active
);
991 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
997 return rq
->tag
!= -1;
1000 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1001 int flags
, void *key
)
1003 struct blk_mq_hw_ctx
*hctx
;
1005 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1007 list_del_init(&wait
->entry
);
1008 blk_mq_run_hw_queue(hctx
, true);
1013 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1014 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1015 * restart. For both cases, take care to check the condition again after
1016 * marking us as waiting.
1018 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
**hctx
,
1021 struct blk_mq_hw_ctx
*this_hctx
= *hctx
;
1022 struct sbq_wait_state
*ws
;
1023 wait_queue_entry_t
*wait
;
1026 if (!(this_hctx
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1027 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &this_hctx
->state
))
1028 set_bit(BLK_MQ_S_SCHED_RESTART
, &this_hctx
->state
);
1031 * It's possible that a tag was freed in the window between the
1032 * allocation failure and adding the hardware queue to the wait
1035 * Don't clear RESTART here, someone else could have set it.
1036 * At most this will cost an extra queue run.
1038 return blk_mq_get_driver_tag(rq
, hctx
, false);
1041 wait
= &this_hctx
->dispatch_wait
;
1042 if (!list_empty_careful(&wait
->entry
))
1045 spin_lock(&this_hctx
->lock
);
1046 if (!list_empty(&wait
->entry
)) {
1047 spin_unlock(&this_hctx
->lock
);
1051 ws
= bt_wait_ptr(&this_hctx
->tags
->bitmap_tags
, this_hctx
);
1052 add_wait_queue(&ws
->wait
, wait
);
1055 * It's possible that a tag was freed in the window between the
1056 * allocation failure and adding the hardware queue to the wait
1059 ret
= blk_mq_get_driver_tag(rq
, hctx
, false);
1061 spin_unlock(&this_hctx
->lock
);
1066 * We got a tag, remove ourselves from the wait queue to ensure
1067 * someone else gets the wakeup.
1069 spin_lock_irq(&ws
->wait
.lock
);
1070 list_del_init(&wait
->entry
);
1071 spin_unlock_irq(&ws
->wait
.lock
);
1072 spin_unlock(&this_hctx
->lock
);
1077 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1079 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
,
1082 struct blk_mq_hw_ctx
*hctx
;
1083 struct request
*rq
, *nxt
;
1084 bool no_tag
= false;
1086 blk_status_t ret
= BLK_STS_OK
;
1088 if (list_empty(list
))
1091 WARN_ON(!list_is_singular(list
) && got_budget
);
1094 * Now process all the entries, sending them to the driver.
1096 errors
= queued
= 0;
1098 struct blk_mq_queue_data bd
;
1100 rq
= list_first_entry(list
, struct request
, queuelist
);
1102 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
1103 if (!got_budget
&& !blk_mq_get_dispatch_budget(hctx
))
1106 if (!blk_mq_get_driver_tag(rq
, NULL
, false)) {
1108 * The initial allocation attempt failed, so we need to
1109 * rerun the hardware queue when a tag is freed. The
1110 * waitqueue takes care of that. If the queue is run
1111 * before we add this entry back on the dispatch list,
1112 * we'll re-run it below.
1114 if (!blk_mq_mark_tag_wait(&hctx
, rq
)) {
1115 blk_mq_put_dispatch_budget(hctx
);
1117 * For non-shared tags, the RESTART check
1120 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
)
1126 list_del_init(&rq
->queuelist
);
1131 * Flag last if we have no more requests, or if we have more
1132 * but can't assign a driver tag to it.
1134 if (list_empty(list
))
1137 nxt
= list_first_entry(list
, struct request
, queuelist
);
1138 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1141 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1142 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
) {
1144 * If an I/O scheduler has been configured and we got a
1145 * driver tag for the next request already, free it
1148 if (!list_empty(list
)) {
1149 nxt
= list_first_entry(list
, struct request
, queuelist
);
1150 blk_mq_put_driver_tag(nxt
);
1152 list_add(&rq
->queuelist
, list
);
1153 __blk_mq_requeue_request(rq
);
1157 if (unlikely(ret
!= BLK_STS_OK
)) {
1159 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1164 } while (!list_empty(list
));
1166 hctx
->dispatched
[queued_to_index(queued
)]++;
1169 * Any items that need requeuing? Stuff them into hctx->dispatch,
1170 * that is where we will continue on next queue run.
1172 if (!list_empty(list
)) {
1175 spin_lock(&hctx
->lock
);
1176 list_splice_init(list
, &hctx
->dispatch
);
1177 spin_unlock(&hctx
->lock
);
1180 * If SCHED_RESTART was set by the caller of this function and
1181 * it is no longer set that means that it was cleared by another
1182 * thread and hence that a queue rerun is needed.
1184 * If 'no_tag' is set, that means that we failed getting
1185 * a driver tag with an I/O scheduler attached. If our dispatch
1186 * waitqueue is no longer active, ensure that we run the queue
1187 * AFTER adding our entries back to the list.
1189 * If no I/O scheduler has been configured it is possible that
1190 * the hardware queue got stopped and restarted before requests
1191 * were pushed back onto the dispatch list. Rerun the queue to
1192 * avoid starvation. Notes:
1193 * - blk_mq_run_hw_queue() checks whether or not a queue has
1194 * been stopped before rerunning a queue.
1195 * - Some but not all block drivers stop a queue before
1196 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1199 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1200 * bit is set, run queue after a delay to avoid IO stalls
1201 * that could otherwise occur if the queue is idle.
1203 needs_restart
= blk_mq_sched_needs_restart(hctx
);
1204 if (!needs_restart
||
1205 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1206 blk_mq_run_hw_queue(hctx
, true);
1207 else if (needs_restart
&& (ret
== BLK_STS_RESOURCE
))
1208 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
1211 return (queued
+ errors
) != 0;
1214 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1219 * We should be running this queue from one of the CPUs that
1222 * There are at least two related races now between setting
1223 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1224 * __blk_mq_run_hw_queue():
1226 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1227 * but later it becomes online, then this warning is harmless
1230 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1231 * but later it becomes offline, then the warning can't be
1232 * triggered, and we depend on blk-mq timeout handler to
1233 * handle dispatched requests to this hctx
1235 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1236 cpu_online(hctx
->next_cpu
)) {
1237 printk(KERN_WARNING
"run queue from wrong CPU %d, hctx %s\n",
1238 raw_smp_processor_id(),
1239 cpumask_empty(hctx
->cpumask
) ? "inactive": "active");
1244 * We can't run the queue inline with ints disabled. Ensure that
1245 * we catch bad users of this early.
1247 WARN_ON_ONCE(in_interrupt());
1249 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1251 hctx_lock(hctx
, &srcu_idx
);
1252 blk_mq_sched_dispatch_requests(hctx
);
1253 hctx_unlock(hctx
, srcu_idx
);
1256 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1258 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1260 if (cpu
>= nr_cpu_ids
)
1261 cpu
= cpumask_first(hctx
->cpumask
);
1266 * It'd be great if the workqueue API had a way to pass
1267 * in a mask and had some smarts for more clever placement.
1268 * For now we just round-robin here, switching for every
1269 * BLK_MQ_CPU_WORK_BATCH queued items.
1271 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1274 int next_cpu
= hctx
->next_cpu
;
1276 if (hctx
->queue
->nr_hw_queues
== 1)
1277 return WORK_CPU_UNBOUND
;
1279 if (--hctx
->next_cpu_batch
<= 0) {
1281 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1283 if (next_cpu
>= nr_cpu_ids
)
1284 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1285 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1289 * Do unbound schedule if we can't find a online CPU for this hctx,
1290 * and it should only happen in the path of handling CPU DEAD.
1292 if (!cpu_online(next_cpu
)) {
1299 * Make sure to re-select CPU next time once after CPUs
1300 * in hctx->cpumask become online again.
1302 hctx
->next_cpu
= next_cpu
;
1303 hctx
->next_cpu_batch
= 1;
1304 return WORK_CPU_UNBOUND
;
1307 hctx
->next_cpu
= next_cpu
;
1311 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1312 unsigned long msecs
)
1314 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1317 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1318 int cpu
= get_cpu();
1319 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1320 __blk_mq_run_hw_queue(hctx
);
1328 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1329 msecs_to_jiffies(msecs
));
1332 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1334 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1336 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1338 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1344 * When queue is quiesced, we may be switching io scheduler, or
1345 * updating nr_hw_queues, or other things, and we can't run queue
1346 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1348 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1351 hctx_lock(hctx
, &srcu_idx
);
1352 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1353 blk_mq_hctx_has_pending(hctx
);
1354 hctx_unlock(hctx
, srcu_idx
);
1357 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1363 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1365 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1367 struct blk_mq_hw_ctx
*hctx
;
1370 queue_for_each_hw_ctx(q
, hctx
, i
) {
1371 if (blk_mq_hctx_stopped(hctx
))
1374 blk_mq_run_hw_queue(hctx
, async
);
1377 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1380 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1381 * @q: request queue.
1383 * The caller is responsible for serializing this function against
1384 * blk_mq_{start,stop}_hw_queue().
1386 bool blk_mq_queue_stopped(struct request_queue
*q
)
1388 struct blk_mq_hw_ctx
*hctx
;
1391 queue_for_each_hw_ctx(q
, hctx
, i
)
1392 if (blk_mq_hctx_stopped(hctx
))
1397 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1400 * This function is often used for pausing .queue_rq() by driver when
1401 * there isn't enough resource or some conditions aren't satisfied, and
1402 * BLK_STS_RESOURCE is usually returned.
1404 * We do not guarantee that dispatch can be drained or blocked
1405 * after blk_mq_stop_hw_queue() returns. Please use
1406 * blk_mq_quiesce_queue() for that requirement.
1408 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1410 cancel_delayed_work(&hctx
->run_work
);
1412 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1414 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1417 * This function is often used for pausing .queue_rq() by driver when
1418 * there isn't enough resource or some conditions aren't satisfied, and
1419 * BLK_STS_RESOURCE is usually returned.
1421 * We do not guarantee that dispatch can be drained or blocked
1422 * after blk_mq_stop_hw_queues() returns. Please use
1423 * blk_mq_quiesce_queue() for that requirement.
1425 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1427 struct blk_mq_hw_ctx
*hctx
;
1430 queue_for_each_hw_ctx(q
, hctx
, i
)
1431 blk_mq_stop_hw_queue(hctx
);
1433 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1435 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1437 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1439 blk_mq_run_hw_queue(hctx
, false);
1441 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1443 void blk_mq_start_hw_queues(struct request_queue
*q
)
1445 struct blk_mq_hw_ctx
*hctx
;
1448 queue_for_each_hw_ctx(q
, hctx
, i
)
1449 blk_mq_start_hw_queue(hctx
);
1451 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1453 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1455 if (!blk_mq_hctx_stopped(hctx
))
1458 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1459 blk_mq_run_hw_queue(hctx
, async
);
1461 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1463 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1465 struct blk_mq_hw_ctx
*hctx
;
1468 queue_for_each_hw_ctx(q
, hctx
, i
)
1469 blk_mq_start_stopped_hw_queue(hctx
, async
);
1471 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1473 static void blk_mq_run_work_fn(struct work_struct
*work
)
1475 struct blk_mq_hw_ctx
*hctx
;
1477 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1480 * If we are stopped, don't run the queue.
1482 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1485 __blk_mq_run_hw_queue(hctx
);
1488 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1492 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1494 lockdep_assert_held(&ctx
->lock
);
1496 trace_block_rq_insert(hctx
->queue
, rq
);
1499 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1501 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1504 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1507 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1509 lockdep_assert_held(&ctx
->lock
);
1511 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1512 blk_mq_hctx_mark_pending(hctx
, ctx
);
1516 * Should only be used carefully, when the caller knows we want to
1517 * bypass a potential IO scheduler on the target device.
1519 void blk_mq_request_bypass_insert(struct request
*rq
, bool run_queue
)
1521 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1522 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(rq
->q
, ctx
->cpu
);
1524 spin_lock(&hctx
->lock
);
1525 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1526 spin_unlock(&hctx
->lock
);
1529 blk_mq_run_hw_queue(hctx
, false);
1532 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1533 struct list_head
*list
)
1537 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1540 spin_lock(&ctx
->lock
);
1541 while (!list_empty(list
)) {
1544 rq
= list_first_entry(list
, struct request
, queuelist
);
1545 BUG_ON(rq
->mq_ctx
!= ctx
);
1546 list_del_init(&rq
->queuelist
);
1547 __blk_mq_insert_req_list(hctx
, rq
, false);
1549 blk_mq_hctx_mark_pending(hctx
, ctx
);
1550 spin_unlock(&ctx
->lock
);
1553 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1555 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1556 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1558 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1559 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1560 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1563 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1565 struct blk_mq_ctx
*this_ctx
;
1566 struct request_queue
*this_q
;
1569 LIST_HEAD(ctx_list
);
1572 list_splice_init(&plug
->mq_list
, &list
);
1574 list_sort(NULL
, &list
, plug_ctx_cmp
);
1580 while (!list_empty(&list
)) {
1581 rq
= list_entry_rq(list
.next
);
1582 list_del_init(&rq
->queuelist
);
1584 if (rq
->mq_ctx
!= this_ctx
) {
1586 trace_block_unplug(this_q
, depth
, from_schedule
);
1587 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1592 this_ctx
= rq
->mq_ctx
;
1598 list_add_tail(&rq
->queuelist
, &ctx_list
);
1602 * If 'this_ctx' is set, we know we have entries to complete
1603 * on 'ctx_list'. Do those.
1606 trace_block_unplug(this_q
, depth
, from_schedule
);
1607 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1612 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1614 blk_init_request_from_bio(rq
, bio
);
1616 blk_rq_set_rl(rq
, blk_get_rl(rq
->q
, bio
));
1618 blk_account_io_start(rq
, true);
1621 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1624 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1626 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1629 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1633 struct request_queue
*q
= rq
->q
;
1634 struct blk_mq_queue_data bd
= {
1638 blk_qc_t new_cookie
;
1641 new_cookie
= request_to_qc_t(hctx
, rq
);
1644 * For OK queue, we are done. For error, caller may kill it.
1645 * Any other error (busy), just add it to our list as we
1646 * previously would have done.
1648 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1651 *cookie
= new_cookie
;
1653 case BLK_STS_RESOURCE
:
1654 case BLK_STS_DEV_RESOURCE
:
1655 __blk_mq_requeue_request(rq
);
1658 *cookie
= BLK_QC_T_NONE
;
1665 static blk_status_t
__blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1670 struct request_queue
*q
= rq
->q
;
1671 bool run_queue
= true;
1674 * RCU or SRCU read lock is needed before checking quiesced flag.
1676 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1677 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1678 * and avoid driver to try to dispatch again.
1680 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1682 bypass_insert
= false;
1686 if (q
->elevator
&& !bypass_insert
)
1689 if (!blk_mq_get_dispatch_budget(hctx
))
1692 if (!blk_mq_get_driver_tag(rq
, NULL
, false)) {
1693 blk_mq_put_dispatch_budget(hctx
);
1697 return __blk_mq_issue_directly(hctx
, rq
, cookie
);
1700 return BLK_STS_RESOURCE
;
1702 blk_mq_sched_insert_request(rq
, false, run_queue
, false);
1706 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1707 struct request
*rq
, blk_qc_t
*cookie
)
1712 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1714 hctx_lock(hctx
, &srcu_idx
);
1716 ret
= __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false);
1717 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1718 blk_mq_sched_insert_request(rq
, false, true, false);
1719 else if (ret
!= BLK_STS_OK
)
1720 blk_mq_end_request(rq
, ret
);
1722 hctx_unlock(hctx
, srcu_idx
);
1725 blk_status_t
blk_mq_request_issue_directly(struct request
*rq
)
1729 blk_qc_t unused_cookie
;
1730 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1731 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(rq
->q
, ctx
->cpu
);
1733 hctx_lock(hctx
, &srcu_idx
);
1734 ret
= __blk_mq_try_issue_directly(hctx
, rq
, &unused_cookie
, true);
1735 hctx_unlock(hctx
, srcu_idx
);
1740 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1742 const int is_sync
= op_is_sync(bio
->bi_opf
);
1743 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1744 struct blk_mq_alloc_data data
= { .flags
= 0 };
1746 unsigned int request_count
= 0;
1747 struct blk_plug
*plug
;
1748 struct request
*same_queue_rq
= NULL
;
1750 unsigned int wb_acct
;
1752 blk_queue_bounce(q
, &bio
);
1754 blk_queue_split(q
, &bio
);
1756 if (!bio_integrity_prep(bio
))
1757 return BLK_QC_T_NONE
;
1759 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1760 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1761 return BLK_QC_T_NONE
;
1763 if (blk_mq_sched_bio_merge(q
, bio
))
1764 return BLK_QC_T_NONE
;
1766 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1768 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1770 rq
= blk_mq_get_request(q
, bio
, bio
->bi_opf
, &data
);
1771 if (unlikely(!rq
)) {
1772 __wbt_done(q
->rq_wb
, wb_acct
);
1773 if (bio
->bi_opf
& REQ_NOWAIT
)
1774 bio_wouldblock_error(bio
);
1775 return BLK_QC_T_NONE
;
1778 wbt_track(rq
, wb_acct
);
1780 cookie
= request_to_qc_t(data
.hctx
, rq
);
1782 plug
= current
->plug
;
1783 if (unlikely(is_flush_fua
)) {
1784 blk_mq_put_ctx(data
.ctx
);
1785 blk_mq_bio_to_request(rq
, bio
);
1787 /* bypass scheduler for flush rq */
1788 blk_insert_flush(rq
);
1789 blk_mq_run_hw_queue(data
.hctx
, true);
1790 } else if (plug
&& q
->nr_hw_queues
== 1) {
1791 struct request
*last
= NULL
;
1793 blk_mq_put_ctx(data
.ctx
);
1794 blk_mq_bio_to_request(rq
, bio
);
1797 * @request_count may become stale because of schedule
1798 * out, so check the list again.
1800 if (list_empty(&plug
->mq_list
))
1802 else if (blk_queue_nomerges(q
))
1803 request_count
= blk_plug_queued_count(q
);
1806 trace_block_plug(q
);
1808 last
= list_entry_rq(plug
->mq_list
.prev
);
1810 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1811 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1812 blk_flush_plug_list(plug
, false);
1813 trace_block_plug(q
);
1816 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1817 } else if (plug
&& !blk_queue_nomerges(q
)) {
1818 blk_mq_bio_to_request(rq
, bio
);
1821 * We do limited plugging. If the bio can be merged, do that.
1822 * Otherwise the existing request in the plug list will be
1823 * issued. So the plug list will have one request at most
1824 * The plug list might get flushed before this. If that happens,
1825 * the plug list is empty, and same_queue_rq is invalid.
1827 if (list_empty(&plug
->mq_list
))
1828 same_queue_rq
= NULL
;
1830 list_del_init(&same_queue_rq
->queuelist
);
1831 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1833 blk_mq_put_ctx(data
.ctx
);
1835 if (same_queue_rq
) {
1836 data
.hctx
= blk_mq_map_queue(q
,
1837 same_queue_rq
->mq_ctx
->cpu
);
1838 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1841 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1842 blk_mq_put_ctx(data
.ctx
);
1843 blk_mq_bio_to_request(rq
, bio
);
1844 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1846 blk_mq_put_ctx(data
.ctx
);
1847 blk_mq_bio_to_request(rq
, bio
);
1848 blk_mq_sched_insert_request(rq
, false, true, true);
1854 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1855 unsigned int hctx_idx
)
1859 if (tags
->rqs
&& set
->ops
->exit_request
) {
1862 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1863 struct request
*rq
= tags
->static_rqs
[i
];
1867 set
->ops
->exit_request(set
, rq
, hctx_idx
);
1868 tags
->static_rqs
[i
] = NULL
;
1872 while (!list_empty(&tags
->page_list
)) {
1873 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1874 list_del_init(&page
->lru
);
1876 * Remove kmemleak object previously allocated in
1877 * blk_mq_init_rq_map().
1879 kmemleak_free(page_address(page
));
1880 __free_pages(page
, page
->private);
1884 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1888 kfree(tags
->static_rqs
);
1889 tags
->static_rqs
= NULL
;
1891 blk_mq_free_tags(tags
);
1894 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1895 unsigned int hctx_idx
,
1896 unsigned int nr_tags
,
1897 unsigned int reserved_tags
)
1899 struct blk_mq_tags
*tags
;
1902 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1903 if (node
== NUMA_NO_NODE
)
1904 node
= set
->numa_node
;
1906 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1907 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1911 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1912 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1915 blk_mq_free_tags(tags
);
1919 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1920 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1922 if (!tags
->static_rqs
) {
1924 blk_mq_free_tags(tags
);
1931 static size_t order_to_size(unsigned int order
)
1933 return (size_t)PAGE_SIZE
<< order
;
1936 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
1937 unsigned int hctx_idx
, int node
)
1941 if (set
->ops
->init_request
) {
1942 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
1947 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
1951 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1952 unsigned int hctx_idx
, unsigned int depth
)
1954 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1955 size_t rq_size
, left
;
1958 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1959 if (node
== NUMA_NO_NODE
)
1960 node
= set
->numa_node
;
1962 INIT_LIST_HEAD(&tags
->page_list
);
1965 * rq_size is the size of the request plus driver payload, rounded
1966 * to the cacheline size
1968 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1970 left
= rq_size
* depth
;
1972 for (i
= 0; i
< depth
; ) {
1973 int this_order
= max_order
;
1978 while (this_order
&& left
< order_to_size(this_order
- 1))
1982 page
= alloc_pages_node(node
,
1983 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1989 if (order_to_size(this_order
) < rq_size
)
1996 page
->private = this_order
;
1997 list_add_tail(&page
->lru
, &tags
->page_list
);
1999 p
= page_address(page
);
2001 * Allow kmemleak to scan these pages as they contain pointers
2002 * to additional allocations like via ops->init_request().
2004 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2005 entries_per_page
= order_to_size(this_order
) / rq_size
;
2006 to_do
= min(entries_per_page
, depth
- i
);
2007 left
-= to_do
* rq_size
;
2008 for (j
= 0; j
< to_do
; j
++) {
2009 struct request
*rq
= p
;
2011 tags
->static_rqs
[i
] = rq
;
2012 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
2013 tags
->static_rqs
[i
] = NULL
;
2024 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2029 * 'cpu' is going away. splice any existing rq_list entries from this
2030 * software queue to the hw queue dispatch list, and ensure that it
2033 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2035 struct blk_mq_hw_ctx
*hctx
;
2036 struct blk_mq_ctx
*ctx
;
2039 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2040 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2042 spin_lock(&ctx
->lock
);
2043 if (!list_empty(&ctx
->rq_list
)) {
2044 list_splice_init(&ctx
->rq_list
, &tmp
);
2045 blk_mq_hctx_clear_pending(hctx
, ctx
);
2047 spin_unlock(&ctx
->lock
);
2049 if (list_empty(&tmp
))
2052 spin_lock(&hctx
->lock
);
2053 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2054 spin_unlock(&hctx
->lock
);
2056 blk_mq_run_hw_queue(hctx
, true);
2060 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2062 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2066 /* hctx->ctxs will be freed in queue's release handler */
2067 static void blk_mq_exit_hctx(struct request_queue
*q
,
2068 struct blk_mq_tag_set
*set
,
2069 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2071 blk_mq_debugfs_unregister_hctx(hctx
);
2073 if (blk_mq_hw_queue_mapped(hctx
))
2074 blk_mq_tag_idle(hctx
);
2076 if (set
->ops
->exit_request
)
2077 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2079 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2081 if (set
->ops
->exit_hctx
)
2082 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2084 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2085 cleanup_srcu_struct(hctx
->srcu
);
2087 blk_mq_remove_cpuhp(hctx
);
2088 blk_free_flush_queue(hctx
->fq
);
2089 sbitmap_free(&hctx
->ctx_map
);
2092 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2093 struct blk_mq_tag_set
*set
, int nr_queue
)
2095 struct blk_mq_hw_ctx
*hctx
;
2098 queue_for_each_hw_ctx(q
, hctx
, i
) {
2101 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2105 static int blk_mq_init_hctx(struct request_queue
*q
,
2106 struct blk_mq_tag_set
*set
,
2107 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2111 node
= hctx
->numa_node
;
2112 if (node
== NUMA_NO_NODE
)
2113 node
= hctx
->numa_node
= set
->numa_node
;
2115 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2116 spin_lock_init(&hctx
->lock
);
2117 INIT_LIST_HEAD(&hctx
->dispatch
);
2119 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
2121 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2123 hctx
->tags
= set
->tags
[hctx_idx
];
2126 * Allocate space for all possible cpus to avoid allocation at
2129 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2132 goto unregister_cpu_notifier
;
2134 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
2140 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2141 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2143 if (set
->ops
->init_hctx
&&
2144 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2147 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
2150 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
2152 goto sched_exit_hctx
;
2154 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
, node
))
2157 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2158 init_srcu_struct(hctx
->srcu
);
2160 blk_mq_debugfs_register_hctx(q
, hctx
);
2167 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2169 if (set
->ops
->exit_hctx
)
2170 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2172 sbitmap_free(&hctx
->ctx_map
);
2175 unregister_cpu_notifier
:
2176 blk_mq_remove_cpuhp(hctx
);
2180 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2181 unsigned int nr_hw_queues
)
2185 for_each_possible_cpu(i
) {
2186 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2187 struct blk_mq_hw_ctx
*hctx
;
2190 spin_lock_init(&__ctx
->lock
);
2191 INIT_LIST_HEAD(&__ctx
->rq_list
);
2195 * Set local node, IFF we have more than one hw queue. If
2196 * not, we remain on the home node of the device
2198 hctx
= blk_mq_map_queue(q
, i
);
2199 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2200 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2204 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2208 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2209 set
->queue_depth
, set
->reserved_tags
);
2210 if (!set
->tags
[hctx_idx
])
2213 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2218 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2219 set
->tags
[hctx_idx
] = NULL
;
2223 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2224 unsigned int hctx_idx
)
2226 if (set
->tags
[hctx_idx
]) {
2227 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2228 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2229 set
->tags
[hctx_idx
] = NULL
;
2233 static void blk_mq_map_swqueue(struct request_queue
*q
)
2235 unsigned int i
, hctx_idx
;
2236 struct blk_mq_hw_ctx
*hctx
;
2237 struct blk_mq_ctx
*ctx
;
2238 struct blk_mq_tag_set
*set
= q
->tag_set
;
2241 * Avoid others reading imcomplete hctx->cpumask through sysfs
2243 mutex_lock(&q
->sysfs_lock
);
2245 queue_for_each_hw_ctx(q
, hctx
, i
) {
2246 cpumask_clear(hctx
->cpumask
);
2248 hctx
->dispatch_from
= NULL
;
2252 * Map software to hardware queues.
2254 * If the cpu isn't present, the cpu is mapped to first hctx.
2256 for_each_possible_cpu(i
) {
2257 hctx_idx
= q
->mq_map
[i
];
2258 /* unmapped hw queue can be remapped after CPU topo changed */
2259 if (!set
->tags
[hctx_idx
] &&
2260 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2262 * If tags initialization fail for some hctx,
2263 * that hctx won't be brought online. In this
2264 * case, remap the current ctx to hctx[0] which
2265 * is guaranteed to always have tags allocated
2270 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2271 hctx
= blk_mq_map_queue(q
, i
);
2273 cpumask_set_cpu(i
, hctx
->cpumask
);
2274 ctx
->index_hw
= hctx
->nr_ctx
;
2275 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2278 mutex_unlock(&q
->sysfs_lock
);
2280 queue_for_each_hw_ctx(q
, hctx
, i
) {
2282 * If no software queues are mapped to this hardware queue,
2283 * disable it and free the request entries.
2285 if (!hctx
->nr_ctx
) {
2286 /* Never unmap queue 0. We need it as a
2287 * fallback in case of a new remap fails
2290 if (i
&& set
->tags
[i
])
2291 blk_mq_free_map_and_requests(set
, i
);
2297 hctx
->tags
= set
->tags
[i
];
2298 WARN_ON(!hctx
->tags
);
2301 * Set the map size to the number of mapped software queues.
2302 * This is more accurate and more efficient than looping
2303 * over all possibly mapped software queues.
2305 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2308 * Initialize batch roundrobin counts
2310 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2311 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2316 * Caller needs to ensure that we're either frozen/quiesced, or that
2317 * the queue isn't live yet.
2319 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2321 struct blk_mq_hw_ctx
*hctx
;
2324 queue_for_each_hw_ctx(q
, hctx
, i
) {
2326 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2327 atomic_inc(&q
->shared_hctx_restart
);
2328 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2330 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2331 atomic_dec(&q
->shared_hctx_restart
);
2332 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2337 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2340 struct request_queue
*q
;
2342 lockdep_assert_held(&set
->tag_list_lock
);
2344 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2345 blk_mq_freeze_queue(q
);
2346 queue_set_hctx_shared(q
, shared
);
2347 blk_mq_unfreeze_queue(q
);
2351 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2353 struct blk_mq_tag_set
*set
= q
->tag_set
;
2355 mutex_lock(&set
->tag_list_lock
);
2356 list_del_rcu(&q
->tag_set_list
);
2357 if (list_is_singular(&set
->tag_list
)) {
2358 /* just transitioned to unshared */
2359 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2360 /* update existing queue */
2361 blk_mq_update_tag_set_depth(set
, false);
2363 mutex_unlock(&set
->tag_list_lock
);
2365 INIT_LIST_HEAD(&q
->tag_set_list
);
2368 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2369 struct request_queue
*q
)
2373 mutex_lock(&set
->tag_list_lock
);
2376 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2378 if (!list_empty(&set
->tag_list
) &&
2379 !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2380 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2381 /* update existing queue */
2382 blk_mq_update_tag_set_depth(set
, true);
2384 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2385 queue_set_hctx_shared(q
, true);
2386 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2388 mutex_unlock(&set
->tag_list_lock
);
2392 * It is the actual release handler for mq, but we do it from
2393 * request queue's release handler for avoiding use-after-free
2394 * and headache because q->mq_kobj shouldn't have been introduced,
2395 * but we can't group ctx/kctx kobj without it.
2397 void blk_mq_release(struct request_queue
*q
)
2399 struct blk_mq_hw_ctx
*hctx
;
2402 /* hctx kobj stays in hctx */
2403 queue_for_each_hw_ctx(q
, hctx
, i
) {
2406 kobject_put(&hctx
->kobj
);
2411 kfree(q
->queue_hw_ctx
);
2414 * release .mq_kobj and sw queue's kobject now because
2415 * both share lifetime with request queue.
2417 blk_mq_sysfs_deinit(q
);
2419 free_percpu(q
->queue_ctx
);
2422 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2424 struct request_queue
*uninit_q
, *q
;
2426 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
, NULL
);
2428 return ERR_PTR(-ENOMEM
);
2430 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2432 blk_cleanup_queue(uninit_q
);
2436 EXPORT_SYMBOL(blk_mq_init_queue
);
2438 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2440 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2442 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, srcu
),
2443 __alignof__(struct blk_mq_hw_ctx
)) !=
2444 sizeof(struct blk_mq_hw_ctx
));
2446 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2447 hw_ctx_size
+= sizeof(struct srcu_struct
);
2452 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2453 struct request_queue
*q
)
2456 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2458 blk_mq_sysfs_unregister(q
);
2460 /* protect against switching io scheduler */
2461 mutex_lock(&q
->sysfs_lock
);
2462 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2468 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2469 hctxs
[i
] = kzalloc_node(blk_mq_hw_ctx_size(set
),
2474 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2481 atomic_set(&hctxs
[i
]->nr_active
, 0);
2482 hctxs
[i
]->numa_node
= node
;
2483 hctxs
[i
]->queue_num
= i
;
2485 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2486 free_cpumask_var(hctxs
[i
]->cpumask
);
2491 blk_mq_hctx_kobj_init(hctxs
[i
]);
2493 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2494 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2498 blk_mq_free_map_and_requests(set
, j
);
2499 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2500 kobject_put(&hctx
->kobj
);
2505 q
->nr_hw_queues
= i
;
2506 mutex_unlock(&q
->sysfs_lock
);
2507 blk_mq_sysfs_register(q
);
2510 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2511 struct request_queue
*q
)
2513 /* mark the queue as mq asap */
2514 q
->mq_ops
= set
->ops
;
2516 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2517 blk_mq_poll_stats_bkt
,
2518 BLK_MQ_POLL_STATS_BKTS
, q
);
2522 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2526 /* init q->mq_kobj and sw queues' kobjects */
2527 blk_mq_sysfs_init(q
);
2529 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2530 GFP_KERNEL
, set
->numa_node
);
2531 if (!q
->queue_hw_ctx
)
2534 q
->mq_map
= set
->mq_map
;
2536 blk_mq_realloc_hw_ctxs(set
, q
);
2537 if (!q
->nr_hw_queues
)
2540 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2541 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2543 q
->nr_queues
= nr_cpu_ids
;
2545 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2547 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2548 queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE
, q
);
2550 q
->sg_reserved_size
= INT_MAX
;
2552 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2553 INIT_LIST_HEAD(&q
->requeue_list
);
2554 spin_lock_init(&q
->requeue_lock
);
2556 blk_queue_make_request(q
, blk_mq_make_request
);
2557 if (q
->mq_ops
->poll
)
2558 q
->poll_fn
= blk_mq_poll
;
2561 * Do this after blk_queue_make_request() overrides it...
2563 q
->nr_requests
= set
->queue_depth
;
2566 * Default to classic polling
2570 if (set
->ops
->complete
)
2571 blk_queue_softirq_done(q
, set
->ops
->complete
);
2573 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2574 blk_mq_add_queue_tag_set(set
, q
);
2575 blk_mq_map_swqueue(q
);
2577 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2580 ret
= elevator_init_mq(q
);
2582 return ERR_PTR(ret
);
2588 kfree(q
->queue_hw_ctx
);
2590 free_percpu(q
->queue_ctx
);
2593 return ERR_PTR(-ENOMEM
);
2595 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2597 void blk_mq_free_queue(struct request_queue
*q
)
2599 struct blk_mq_tag_set
*set
= q
->tag_set
;
2601 blk_mq_del_queue_tag_set(q
);
2602 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2605 /* Basically redo blk_mq_init_queue with queue frozen */
2606 static void blk_mq_queue_reinit(struct request_queue
*q
)
2608 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2610 blk_mq_debugfs_unregister_hctxs(q
);
2611 blk_mq_sysfs_unregister(q
);
2614 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2615 * we should change hctx numa_node according to the new topology (this
2616 * involves freeing and re-allocating memory, worth doing?)
2618 blk_mq_map_swqueue(q
);
2620 blk_mq_sysfs_register(q
);
2621 blk_mq_debugfs_register_hctxs(q
);
2624 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2628 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2629 if (!__blk_mq_alloc_rq_map(set
, i
))
2636 blk_mq_free_rq_map(set
->tags
[i
]);
2642 * Allocate the request maps associated with this tag_set. Note that this
2643 * may reduce the depth asked for, if memory is tight. set->queue_depth
2644 * will be updated to reflect the allocated depth.
2646 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2651 depth
= set
->queue_depth
;
2653 err
= __blk_mq_alloc_rq_maps(set
);
2657 set
->queue_depth
>>= 1;
2658 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2662 } while (set
->queue_depth
);
2664 if (!set
->queue_depth
|| err
) {
2665 pr_err("blk-mq: failed to allocate request map\n");
2669 if (depth
!= set
->queue_depth
)
2670 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2671 depth
, set
->queue_depth
);
2676 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2678 if (set
->ops
->map_queues
) {
2681 * transport .map_queues is usually done in the following
2684 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2685 * mask = get_cpu_mask(queue)
2686 * for_each_cpu(cpu, mask)
2687 * set->mq_map[cpu] = queue;
2690 * When we need to remap, the table has to be cleared for
2691 * killing stale mapping since one CPU may not be mapped
2694 for_each_possible_cpu(cpu
)
2695 set
->mq_map
[cpu
] = 0;
2697 return set
->ops
->map_queues(set
);
2699 return blk_mq_map_queues(set
);
2703 * Alloc a tag set to be associated with one or more request queues.
2704 * May fail with EINVAL for various error conditions. May adjust the
2705 * requested depth down, if if it too large. In that case, the set
2706 * value will be stored in set->queue_depth.
2708 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2712 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2714 if (!set
->nr_hw_queues
)
2716 if (!set
->queue_depth
)
2718 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2721 if (!set
->ops
->queue_rq
)
2724 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
2727 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2728 pr_info("blk-mq: reduced tag depth to %u\n",
2730 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2734 * If a crashdump is active, then we are potentially in a very
2735 * memory constrained environment. Limit us to 1 queue and
2736 * 64 tags to prevent using too much memory.
2738 if (is_kdump_kernel()) {
2739 set
->nr_hw_queues
= 1;
2740 set
->queue_depth
= min(64U, set
->queue_depth
);
2743 * There is no use for more h/w queues than cpus.
2745 if (set
->nr_hw_queues
> nr_cpu_ids
)
2746 set
->nr_hw_queues
= nr_cpu_ids
;
2748 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2749 GFP_KERNEL
, set
->numa_node
);
2754 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2755 GFP_KERNEL
, set
->numa_node
);
2759 ret
= blk_mq_update_queue_map(set
);
2761 goto out_free_mq_map
;
2763 ret
= blk_mq_alloc_rq_maps(set
);
2765 goto out_free_mq_map
;
2767 mutex_init(&set
->tag_list_lock
);
2768 INIT_LIST_HEAD(&set
->tag_list
);
2780 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2782 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2786 for (i
= 0; i
< nr_cpu_ids
; i
++)
2787 blk_mq_free_map_and_requests(set
, i
);
2795 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2797 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2799 struct blk_mq_tag_set
*set
= q
->tag_set
;
2800 struct blk_mq_hw_ctx
*hctx
;
2806 blk_mq_freeze_queue(q
);
2807 blk_mq_quiesce_queue(q
);
2810 queue_for_each_hw_ctx(q
, hctx
, i
) {
2814 * If we're using an MQ scheduler, just update the scheduler
2815 * queue depth. This is similar to what the old code would do.
2817 if (!hctx
->sched_tags
) {
2818 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
2821 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2829 q
->nr_requests
= nr
;
2831 blk_mq_unquiesce_queue(q
);
2832 blk_mq_unfreeze_queue(q
);
2837 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
2840 struct request_queue
*q
;
2842 lockdep_assert_held(&set
->tag_list_lock
);
2844 if (nr_hw_queues
> nr_cpu_ids
)
2845 nr_hw_queues
= nr_cpu_ids
;
2846 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2849 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2850 blk_mq_freeze_queue(q
);
2852 set
->nr_hw_queues
= nr_hw_queues
;
2853 blk_mq_update_queue_map(set
);
2854 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2855 blk_mq_realloc_hw_ctxs(set
, q
);
2856 blk_mq_queue_reinit(q
);
2859 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2860 blk_mq_unfreeze_queue(q
);
2863 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2865 mutex_lock(&set
->tag_list_lock
);
2866 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
2867 mutex_unlock(&set
->tag_list_lock
);
2869 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2871 /* Enable polling stats and return whether they were already enabled. */
2872 static bool blk_poll_stats_enable(struct request_queue
*q
)
2874 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2875 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS
, q
))
2877 blk_stat_add_callback(q
, q
->poll_cb
);
2881 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2884 * We don't arm the callback if polling stats are not enabled or the
2885 * callback is already active.
2887 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2888 blk_stat_is_active(q
->poll_cb
))
2891 blk_stat_activate_msecs(q
->poll_cb
, 100);
2894 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2896 struct request_queue
*q
= cb
->data
;
2899 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
2900 if (cb
->stat
[bucket
].nr_samples
)
2901 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
2905 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2906 struct blk_mq_hw_ctx
*hctx
,
2909 unsigned long ret
= 0;
2913 * If stats collection isn't on, don't sleep but turn it on for
2916 if (!blk_poll_stats_enable(q
))
2920 * As an optimistic guess, use half of the mean service time
2921 * for this type of request. We can (and should) make this smarter.
2922 * For instance, if the completion latencies are tight, we can
2923 * get closer than just half the mean. This is especially
2924 * important on devices where the completion latencies are longer
2925 * than ~10 usec. We do use the stats for the relevant IO size
2926 * if available which does lead to better estimates.
2928 bucket
= blk_mq_poll_stats_bkt(rq
);
2932 if (q
->poll_stat
[bucket
].nr_samples
)
2933 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
2938 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2939 struct blk_mq_hw_ctx
*hctx
,
2942 struct hrtimer_sleeper hs
;
2943 enum hrtimer_mode mode
;
2947 if (rq
->rq_flags
& RQF_MQ_POLL_SLEPT
)
2953 * -1: don't ever hybrid sleep
2954 * 0: use half of prev avg
2955 * >0: use this specific value
2957 if (q
->poll_nsec
== -1)
2959 else if (q
->poll_nsec
> 0)
2960 nsecs
= q
->poll_nsec
;
2962 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2967 rq
->rq_flags
|= RQF_MQ_POLL_SLEPT
;
2970 * This will be replaced with the stats tracking code, using
2971 * 'avg_completion_time / 2' as the pre-sleep target.
2975 mode
= HRTIMER_MODE_REL
;
2976 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2977 hrtimer_set_expires(&hs
.timer
, kt
);
2979 hrtimer_init_sleeper(&hs
, current
);
2981 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
2983 set_current_state(TASK_UNINTERRUPTIBLE
);
2984 hrtimer_start_expires(&hs
.timer
, mode
);
2987 hrtimer_cancel(&hs
.timer
);
2988 mode
= HRTIMER_MODE_ABS
;
2989 } while (hs
.task
&& !signal_pending(current
));
2991 __set_current_state(TASK_RUNNING
);
2992 destroy_hrtimer_on_stack(&hs
.timer
);
2996 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2998 struct request_queue
*q
= hctx
->queue
;
3002 * If we sleep, have the caller restart the poll loop to reset
3003 * the state. Like for the other success return cases, the
3004 * caller is responsible for checking if the IO completed. If
3005 * the IO isn't complete, we'll get called again and will go
3006 * straight to the busy poll loop.
3008 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
3011 hctx
->poll_considered
++;
3013 state
= current
->state
;
3014 while (!need_resched()) {
3017 hctx
->poll_invoked
++;
3019 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
3021 hctx
->poll_success
++;
3022 set_current_state(TASK_RUNNING
);
3026 if (signal_pending_state(state
, current
))
3027 set_current_state(TASK_RUNNING
);
3029 if (current
->state
== TASK_RUNNING
)
3036 __set_current_state(TASK_RUNNING
);
3040 static bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
3042 struct blk_mq_hw_ctx
*hctx
;
3045 if (!test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3048 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3049 if (!blk_qc_t_is_internal(cookie
))
3050 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3052 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3054 * With scheduling, if the request has completed, we'll
3055 * get a NULL return here, as we clear the sched tag when
3056 * that happens. The request still remains valid, like always,
3057 * so we should be safe with just the NULL check.
3063 return __blk_mq_poll(hctx
, rq
);
3066 static int __init
blk_mq_init(void)
3068 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3069 blk_mq_hctx_notify_dead
);
3072 subsys_initcall(blk_mq_init
);