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 void blk_mq_poll_stats_start(struct request_queue
*q
);
41 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
43 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
45 int ddir
, bytes
, bucket
;
47 ddir
= rq_data_dir(rq
);
48 bytes
= blk_rq_bytes(rq
);
50 bucket
= ddir
+ 2*(ilog2(bytes
) - 9);
54 else if (bucket
>= BLK_MQ_POLL_STATS_BKTS
)
55 return ddir
+ BLK_MQ_POLL_STATS_BKTS
- 2;
61 * Check if any of the ctx's have pending work in this hardware queue
63 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
65 return sbitmap_any_bit_set(&hctx
->ctx_map
) ||
66 !list_empty_careful(&hctx
->dispatch
) ||
67 blk_mq_sched_has_work(hctx
);
71 * Mark this ctx as having pending work in this hardware queue
73 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
74 struct blk_mq_ctx
*ctx
)
76 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
77 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
80 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
81 struct blk_mq_ctx
*ctx
)
83 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
87 struct hd_struct
*part
;
88 unsigned int *inflight
;
91 static void blk_mq_check_inflight(struct blk_mq_hw_ctx
*hctx
,
92 struct request
*rq
, void *priv
,
95 struct mq_inflight
*mi
= priv
;
97 if (test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
) &&
98 !test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
)) {
100 * index[0] counts the specific partition that was asked
101 * for. index[1] counts the ones that are active on the
102 * whole device, so increment that if mi->part is indeed
103 * a partition, and not a whole device.
105 if (rq
->part
== mi
->part
)
107 if (mi
->part
->partno
)
112 void blk_mq_in_flight(struct request_queue
*q
, struct hd_struct
*part
,
113 unsigned int inflight
[2])
115 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
117 inflight
[0] = inflight
[1] = 0;
118 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
121 void blk_freeze_queue_start(struct request_queue
*q
)
125 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
126 if (freeze_depth
== 1) {
127 percpu_ref_kill(&q
->q_usage_counter
);
128 blk_mq_run_hw_queues(q
, false);
131 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
133 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
135 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
137 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
139 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
140 unsigned long timeout
)
142 return wait_event_timeout(q
->mq_freeze_wq
,
143 percpu_ref_is_zero(&q
->q_usage_counter
),
146 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
149 * Guarantee no request is in use, so we can change any data structure of
150 * the queue afterward.
152 void blk_freeze_queue(struct request_queue
*q
)
155 * In the !blk_mq case we are only calling this to kill the
156 * q_usage_counter, otherwise this increases the freeze depth
157 * and waits for it to return to zero. For this reason there is
158 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
159 * exported to drivers as the only user for unfreeze is blk_mq.
161 blk_freeze_queue_start(q
);
164 blk_mq_freeze_queue_wait(q
);
167 void blk_mq_freeze_queue(struct request_queue
*q
)
170 * ...just an alias to keep freeze and unfreeze actions balanced
171 * in the blk_mq_* namespace
175 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
177 void blk_mq_unfreeze_queue(struct request_queue
*q
)
181 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
182 WARN_ON_ONCE(freeze_depth
< 0);
184 percpu_ref_reinit(&q
->q_usage_counter
);
185 wake_up_all(&q
->mq_freeze_wq
);
188 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
191 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
192 * mpt3sas driver such that this function can be removed.
194 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
198 spin_lock_irqsave(q
->queue_lock
, flags
);
199 queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
200 spin_unlock_irqrestore(q
->queue_lock
, flags
);
202 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
205 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
208 * Note: this function does not prevent that the struct request end_io()
209 * callback function is invoked. Once this function is returned, we make
210 * sure no dispatch can happen until the queue is unquiesced via
211 * blk_mq_unquiesce_queue().
213 void blk_mq_quiesce_queue(struct request_queue
*q
)
215 struct blk_mq_hw_ctx
*hctx
;
219 blk_mq_quiesce_queue_nowait(q
);
221 queue_for_each_hw_ctx(q
, hctx
, i
) {
222 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
223 synchronize_srcu(hctx
->queue_rq_srcu
);
230 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
233 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
236 * This function recovers queue into the state before quiescing
237 * which is done by blk_mq_quiesce_queue.
239 void blk_mq_unquiesce_queue(struct request_queue
*q
)
243 spin_lock_irqsave(q
->queue_lock
, flags
);
244 queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
245 spin_unlock_irqrestore(q
->queue_lock
, flags
);
247 /* dispatch requests which are inserted during quiescing */
248 blk_mq_run_hw_queues(q
, true);
250 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
252 void blk_mq_wake_waiters(struct request_queue
*q
)
254 struct blk_mq_hw_ctx
*hctx
;
257 queue_for_each_hw_ctx(q
, hctx
, i
)
258 if (blk_mq_hw_queue_mapped(hctx
))
259 blk_mq_tag_wakeup_all(hctx
->tags
, true);
262 * If we are called because the queue has now been marked as
263 * dying, we need to ensure that processes currently waiting on
264 * the queue are notified as well.
266 wake_up_all(&q
->mq_freeze_wq
);
269 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
271 return blk_mq_has_free_tags(hctx
->tags
);
273 EXPORT_SYMBOL(blk_mq_can_queue
);
275 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
276 unsigned int tag
, unsigned int op
)
278 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
279 struct request
*rq
= tags
->static_rqs
[tag
];
283 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
285 rq
->internal_tag
= tag
;
287 if (blk_mq_tag_busy(data
->hctx
)) {
288 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
289 atomic_inc(&data
->hctx
->nr_active
);
292 rq
->internal_tag
= -1;
293 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
296 INIT_LIST_HEAD(&rq
->queuelist
);
297 /* csd/requeue_work/fifo_time is initialized before use */
299 rq
->mq_ctx
= data
->ctx
;
301 if (blk_queue_io_stat(data
->q
))
302 rq
->rq_flags
|= RQF_IO_STAT
;
303 /* do not touch atomic flags, it needs atomic ops against the timer */
305 INIT_HLIST_NODE(&rq
->hash
);
306 RB_CLEAR_NODE(&rq
->rb_node
);
309 rq
->start_time
= jiffies
;
310 #ifdef CONFIG_BLK_CGROUP
312 set_start_time_ns(rq
);
313 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 */
323 INIT_LIST_HEAD(&rq
->timeout_list
);
327 rq
->end_io_data
= NULL
;
330 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
334 static struct request
*blk_mq_get_request(struct request_queue
*q
,
335 struct bio
*bio
, unsigned int op
,
336 struct blk_mq_alloc_data
*data
)
338 struct elevator_queue
*e
= q
->elevator
;
341 struct blk_mq_ctx
*local_ctx
= NULL
;
343 blk_queue_enter_live(q
);
345 if (likely(!data
->ctx
))
346 data
->ctx
= local_ctx
= blk_mq_get_ctx(q
);
347 if (likely(!data
->hctx
))
348 data
->hctx
= blk_mq_map_queue(q
, data
->ctx
->cpu
);
350 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
353 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
356 * Flush requests are special and go directly to the
359 if (!op_is_flush(op
) && e
->type
->ops
.mq
.limit_depth
)
360 e
->type
->ops
.mq
.limit_depth(op
, data
);
363 tag
= blk_mq_get_tag(data
);
364 if (tag
== BLK_MQ_TAG_FAIL
) {
366 blk_mq_put_ctx(local_ctx
);
373 rq
= blk_mq_rq_ctx_init(data
, tag
, op
);
374 if (!op_is_flush(op
)) {
376 if (e
&& e
->type
->ops
.mq
.prepare_request
) {
377 if (e
->type
->icq_cache
&& rq_ioc(bio
))
378 blk_mq_sched_assign_ioc(rq
, bio
);
380 e
->type
->ops
.mq
.prepare_request(rq
, bio
);
381 rq
->rq_flags
|= RQF_ELVPRIV
;
384 data
->hctx
->queued
++;
388 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
391 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
395 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
399 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
403 return ERR_PTR(-EWOULDBLOCK
);
405 blk_mq_put_ctx(alloc_data
.ctx
);
408 rq
->__sector
= (sector_t
) -1;
409 rq
->bio
= rq
->biotail
= NULL
;
412 EXPORT_SYMBOL(blk_mq_alloc_request
);
414 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
415 unsigned int op
, unsigned int flags
, unsigned int hctx_idx
)
417 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
423 * If the tag allocator sleeps we could get an allocation for a
424 * different hardware context. No need to complicate the low level
425 * allocator for this for the rare use case of a command tied to
428 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
429 return ERR_PTR(-EINVAL
);
431 if (hctx_idx
>= q
->nr_hw_queues
)
432 return ERR_PTR(-EIO
);
434 ret
= blk_queue_enter(q
, true);
439 * Check if the hardware context is actually mapped to anything.
440 * If not tell the caller that it should skip this queue.
442 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
443 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
445 return ERR_PTR(-EXDEV
);
447 cpu
= cpumask_first(alloc_data
.hctx
->cpumask
);
448 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
450 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
454 return ERR_PTR(-EWOULDBLOCK
);
458 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
460 void blk_mq_free_request(struct request
*rq
)
462 struct request_queue
*q
= rq
->q
;
463 struct elevator_queue
*e
= q
->elevator
;
464 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
465 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
466 const int sched_tag
= rq
->internal_tag
;
468 if (rq
->rq_flags
& RQF_ELVPRIV
) {
469 if (e
&& e
->type
->ops
.mq
.finish_request
)
470 e
->type
->ops
.mq
.finish_request(rq
);
472 put_io_context(rq
->elv
.icq
->ioc
);
477 ctx
->rq_completed
[rq_is_sync(rq
)]++;
478 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
479 atomic_dec(&hctx
->nr_active
);
481 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
483 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
484 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
486 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
488 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
489 blk_mq_sched_restart(hctx
);
492 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
494 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
496 blk_account_io_done(rq
);
499 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
500 rq
->end_io(rq
, error
);
502 if (unlikely(blk_bidi_rq(rq
)))
503 blk_mq_free_request(rq
->next_rq
);
504 blk_mq_free_request(rq
);
507 EXPORT_SYMBOL(__blk_mq_end_request
);
509 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
511 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
513 __blk_mq_end_request(rq
, error
);
515 EXPORT_SYMBOL(blk_mq_end_request
);
517 static void __blk_mq_complete_request_remote(void *data
)
519 struct request
*rq
= data
;
521 rq
->q
->softirq_done_fn(rq
);
524 static void __blk_mq_complete_request(struct request
*rq
)
526 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
530 if (rq
->internal_tag
!= -1)
531 blk_mq_sched_completed_request(rq
);
532 if (rq
->rq_flags
& RQF_STATS
) {
533 blk_mq_poll_stats_start(rq
->q
);
537 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
538 rq
->q
->softirq_done_fn(rq
);
543 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
544 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
546 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
547 rq
->csd
.func
= __blk_mq_complete_request_remote
;
550 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
552 rq
->q
->softirq_done_fn(rq
);
558 * blk_mq_complete_request - end I/O on a request
559 * @rq: the request being processed
562 * Ends all I/O on a request. It does not handle partial completions.
563 * The actual completion happens out-of-order, through a IPI handler.
565 void blk_mq_complete_request(struct request
*rq
)
567 struct request_queue
*q
= rq
->q
;
569 if (unlikely(blk_should_fake_timeout(q
)))
571 if (!blk_mark_rq_complete(rq
))
572 __blk_mq_complete_request(rq
);
574 EXPORT_SYMBOL(blk_mq_complete_request
);
576 int blk_mq_request_started(struct request
*rq
)
578 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
580 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
582 void blk_mq_start_request(struct request
*rq
)
584 struct request_queue
*q
= rq
->q
;
586 blk_mq_sched_started_request(rq
);
588 trace_block_rq_issue(q
, rq
);
590 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
591 blk_stat_set_issue(&rq
->issue_stat
, blk_rq_sectors(rq
));
592 rq
->rq_flags
|= RQF_STATS
;
593 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
599 * Ensure that ->deadline is visible before set the started
600 * flag and clear the completed flag.
602 smp_mb__before_atomic();
605 * Mark us as started and clear complete. Complete might have been
606 * set if requeue raced with timeout, which then marked it as
607 * complete. So be sure to clear complete again when we start
608 * the request, otherwise we'll ignore the completion event.
610 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
611 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
612 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
613 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
615 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
617 * Make sure space for the drain appears. We know we can do
618 * this because max_hw_segments has been adjusted to be one
619 * fewer than the device can handle.
621 rq
->nr_phys_segments
++;
624 EXPORT_SYMBOL(blk_mq_start_request
);
627 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
628 * flag isn't set yet, so there may be race with timeout handler,
629 * but given rq->deadline is just set in .queue_rq() under
630 * this situation, the race won't be possible in reality because
631 * rq->timeout should be set as big enough to cover the window
632 * between blk_mq_start_request() called from .queue_rq() and
633 * clearing REQ_ATOM_STARTED here.
635 static void __blk_mq_requeue_request(struct request
*rq
)
637 struct request_queue
*q
= rq
->q
;
639 trace_block_rq_requeue(q
, rq
);
640 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
641 blk_mq_sched_requeue_request(rq
);
643 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
644 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
645 rq
->nr_phys_segments
--;
649 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
651 __blk_mq_requeue_request(rq
);
653 BUG_ON(blk_queued_rq(rq
));
654 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
656 EXPORT_SYMBOL(blk_mq_requeue_request
);
658 static void blk_mq_requeue_work(struct work_struct
*work
)
660 struct request_queue
*q
=
661 container_of(work
, struct request_queue
, requeue_work
.work
);
663 struct request
*rq
, *next
;
665 spin_lock_irq(&q
->requeue_lock
);
666 list_splice_init(&q
->requeue_list
, &rq_list
);
667 spin_unlock_irq(&q
->requeue_lock
);
669 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
670 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
673 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
674 list_del_init(&rq
->queuelist
);
675 blk_mq_sched_insert_request(rq
, true, false, false, true);
678 while (!list_empty(&rq_list
)) {
679 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
680 list_del_init(&rq
->queuelist
);
681 blk_mq_sched_insert_request(rq
, false, false, false, true);
684 blk_mq_run_hw_queues(q
, false);
687 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
688 bool kick_requeue_list
)
690 struct request_queue
*q
= rq
->q
;
694 * We abuse this flag that is otherwise used by the I/O scheduler to
695 * request head insertation from the workqueue.
697 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
699 spin_lock_irqsave(&q
->requeue_lock
, flags
);
701 rq
->rq_flags
|= RQF_SOFTBARRIER
;
702 list_add(&rq
->queuelist
, &q
->requeue_list
);
704 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
706 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
708 if (kick_requeue_list
)
709 blk_mq_kick_requeue_list(q
);
711 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
713 void blk_mq_kick_requeue_list(struct request_queue
*q
)
715 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
717 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
719 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
722 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
723 msecs_to_jiffies(msecs
));
725 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
727 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
729 if (tag
< tags
->nr_tags
) {
730 prefetch(tags
->rqs
[tag
]);
731 return tags
->rqs
[tag
];
736 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
738 struct blk_mq_timeout_data
{
740 unsigned int next_set
;
743 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
745 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
746 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
749 * We know that complete is set at this point. If STARTED isn't set
750 * anymore, then the request isn't active and the "timeout" should
751 * just be ignored. This can happen due to the bitflag ordering.
752 * Timeout first checks if STARTED is set, and if it is, assumes
753 * the request is active. But if we race with completion, then
754 * both flags will get cleared. So check here again, and ignore
755 * a timeout event with a request that isn't active.
757 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
761 ret
= ops
->timeout(req
, reserved
);
765 __blk_mq_complete_request(req
);
767 case BLK_EH_RESET_TIMER
:
769 blk_clear_rq_complete(req
);
771 case BLK_EH_NOT_HANDLED
:
774 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
779 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
780 struct request
*rq
, void *priv
, bool reserved
)
782 struct blk_mq_timeout_data
*data
= priv
;
784 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
788 * The rq being checked may have been freed and reallocated
789 * out already here, we avoid this race by checking rq->deadline
790 * and REQ_ATOM_COMPLETE flag together:
792 * - if rq->deadline is observed as new value because of
793 * reusing, the rq won't be timed out because of timing.
794 * - if rq->deadline is observed as previous value,
795 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
796 * because we put a barrier between setting rq->deadline
797 * and clearing the flag in blk_mq_start_request(), so
798 * this rq won't be timed out too.
800 if (time_after_eq(jiffies
, rq
->deadline
)) {
801 if (!blk_mark_rq_complete(rq
))
802 blk_mq_rq_timed_out(rq
, reserved
);
803 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
804 data
->next
= rq
->deadline
;
809 static void blk_mq_timeout_work(struct work_struct
*work
)
811 struct request_queue
*q
=
812 container_of(work
, struct request_queue
, timeout_work
);
813 struct blk_mq_timeout_data data
= {
819 /* A deadlock might occur if a request is stuck requiring a
820 * timeout at the same time a queue freeze is waiting
821 * completion, since the timeout code would not be able to
822 * acquire the queue reference here.
824 * That's why we don't use blk_queue_enter here; instead, we use
825 * percpu_ref_tryget directly, because we need to be able to
826 * obtain a reference even in the short window between the queue
827 * starting to freeze, by dropping the first reference in
828 * blk_freeze_queue_start, and the moment the last request is
829 * consumed, marked by the instant q_usage_counter reaches
832 if (!percpu_ref_tryget(&q
->q_usage_counter
))
835 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
838 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
839 mod_timer(&q
->timeout
, data
.next
);
841 struct blk_mq_hw_ctx
*hctx
;
843 queue_for_each_hw_ctx(q
, hctx
, i
) {
844 /* the hctx may be unmapped, so check it here */
845 if (blk_mq_hw_queue_mapped(hctx
))
846 blk_mq_tag_idle(hctx
);
852 struct flush_busy_ctx_data
{
853 struct blk_mq_hw_ctx
*hctx
;
854 struct list_head
*list
;
857 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
859 struct flush_busy_ctx_data
*flush_data
= data
;
860 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
861 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
863 sbitmap_clear_bit(sb
, bitnr
);
864 spin_lock(&ctx
->lock
);
865 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
866 spin_unlock(&ctx
->lock
);
871 * Process software queues that have been marked busy, splicing them
872 * to the for-dispatch
874 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
876 struct flush_busy_ctx_data data
= {
881 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
883 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
885 static inline unsigned int queued_to_index(unsigned int queued
)
890 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
893 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
896 struct blk_mq_alloc_data data
= {
898 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
899 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
902 might_sleep_if(wait
);
907 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
908 data
.flags
|= BLK_MQ_REQ_RESERVED
;
910 rq
->tag
= blk_mq_get_tag(&data
);
912 if (blk_mq_tag_busy(data
.hctx
)) {
913 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
914 atomic_inc(&data
.hctx
->nr_active
);
916 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
922 return rq
->tag
!= -1;
925 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx
*hctx
,
928 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
931 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
932 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
933 atomic_dec(&hctx
->nr_active
);
937 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx
*hctx
,
940 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
943 __blk_mq_put_driver_tag(hctx
, rq
);
946 static void blk_mq_put_driver_tag(struct request
*rq
)
948 struct blk_mq_hw_ctx
*hctx
;
950 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
953 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
954 __blk_mq_put_driver_tag(hctx
, rq
);
958 * If we fail getting a driver tag because all the driver tags are already
959 * assigned and on the dispatch list, BUT the first entry does not have a
960 * tag, then we could deadlock. For that case, move entries with assigned
961 * driver tags to the front, leaving the set of tagged requests in the
962 * same order, and the untagged set in the same order.
964 static bool reorder_tags_to_front(struct list_head
*list
)
966 struct request
*rq
, *tmp
, *first
= NULL
;
968 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
972 list_move(&rq
->queuelist
, list
);
978 return first
!= NULL
;
981 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
, int flags
,
984 struct blk_mq_hw_ctx
*hctx
;
986 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
988 list_del(&wait
->entry
);
989 clear_bit_unlock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
);
990 blk_mq_run_hw_queue(hctx
, true);
994 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx
*hctx
)
996 struct sbq_wait_state
*ws
;
999 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
1000 * The thread which wins the race to grab this bit adds the hardware
1001 * queue to the wait queue.
1003 if (test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
) ||
1004 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1007 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
1008 ws
= bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
);
1011 * As soon as this returns, it's no longer safe to fiddle with
1012 * hctx->dispatch_wait, since a completion can wake up the wait queue
1013 * and unlock the bit.
1015 add_wait_queue(&ws
->wait
, &hctx
->dispatch_wait
);
1019 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
)
1021 struct blk_mq_hw_ctx
*hctx
;
1025 if (list_empty(list
))
1029 * Now process all the entries, sending them to the driver.
1031 errors
= queued
= 0;
1033 struct blk_mq_queue_data bd
;
1036 rq
= list_first_entry(list
, struct request
, queuelist
);
1037 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
1038 if (!queued
&& reorder_tags_to_front(list
))
1042 * The initial allocation attempt failed, so we need to
1043 * rerun the hardware queue when a tag is freed.
1045 if (!blk_mq_dispatch_wait_add(hctx
))
1049 * It's possible that a tag was freed in the window
1050 * between the allocation failure and adding the
1051 * hardware queue to the wait queue.
1053 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1057 list_del_init(&rq
->queuelist
);
1062 * Flag last if we have no more requests, or if we have more
1063 * but can't assign a driver tag to it.
1065 if (list_empty(list
))
1068 struct request
*nxt
;
1070 nxt
= list_first_entry(list
, struct request
, queuelist
);
1071 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1074 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1075 if (ret
== BLK_STS_RESOURCE
) {
1076 blk_mq_put_driver_tag_hctx(hctx
, rq
);
1077 list_add(&rq
->queuelist
, list
);
1078 __blk_mq_requeue_request(rq
);
1082 if (unlikely(ret
!= BLK_STS_OK
)) {
1084 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1089 } while (!list_empty(list
));
1091 hctx
->dispatched
[queued_to_index(queued
)]++;
1094 * Any items that need requeuing? Stuff them into hctx->dispatch,
1095 * that is where we will continue on next queue run.
1097 if (!list_empty(list
)) {
1099 * If an I/O scheduler has been configured and we got a driver
1100 * tag for the next request already, free it again.
1102 rq
= list_first_entry(list
, struct request
, queuelist
);
1103 blk_mq_put_driver_tag(rq
);
1105 spin_lock(&hctx
->lock
);
1106 list_splice_init(list
, &hctx
->dispatch
);
1107 spin_unlock(&hctx
->lock
);
1110 * If SCHED_RESTART was set by the caller of this function and
1111 * it is no longer set that means that it was cleared by another
1112 * thread and hence that a queue rerun is needed.
1114 * If TAG_WAITING is set that means that an I/O scheduler has
1115 * been configured and another thread is waiting for a driver
1116 * tag. To guarantee fairness, do not rerun this hardware queue
1117 * but let the other thread grab the driver tag.
1119 * If no I/O scheduler has been configured it is possible that
1120 * the hardware queue got stopped and restarted before requests
1121 * were pushed back onto the dispatch list. Rerun the queue to
1122 * avoid starvation. Notes:
1123 * - blk_mq_run_hw_queue() checks whether or not a queue has
1124 * been stopped before rerunning a queue.
1125 * - Some but not all block drivers stop a queue before
1126 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1129 if (!blk_mq_sched_needs_restart(hctx
) &&
1130 !test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1131 blk_mq_run_hw_queue(hctx
, true);
1134 return (queued
+ errors
) != 0;
1137 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1142 * We should be running this queue from one of the CPUs that
1145 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1146 cpu_online(hctx
->next_cpu
));
1149 * We can't run the queue inline with ints disabled. Ensure that
1150 * we catch bad users of this early.
1152 WARN_ON_ONCE(in_interrupt());
1154 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1156 blk_mq_sched_dispatch_requests(hctx
);
1161 srcu_idx
= srcu_read_lock(hctx
->queue_rq_srcu
);
1162 blk_mq_sched_dispatch_requests(hctx
);
1163 srcu_read_unlock(hctx
->queue_rq_srcu
, srcu_idx
);
1168 * It'd be great if the workqueue API had a way to pass
1169 * in a mask and had some smarts for more clever placement.
1170 * For now we just round-robin here, switching for every
1171 * BLK_MQ_CPU_WORK_BATCH queued items.
1173 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1175 if (hctx
->queue
->nr_hw_queues
== 1)
1176 return WORK_CPU_UNBOUND
;
1178 if (--hctx
->next_cpu_batch
<= 0) {
1181 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1182 if (next_cpu
>= nr_cpu_ids
)
1183 next_cpu
= cpumask_first(hctx
->cpumask
);
1185 hctx
->next_cpu
= next_cpu
;
1186 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1189 return hctx
->next_cpu
;
1192 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1193 unsigned long msecs
)
1195 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx
)))
1198 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1201 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1202 int cpu
= get_cpu();
1203 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1204 __blk_mq_run_hw_queue(hctx
);
1212 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1214 msecs_to_jiffies(msecs
));
1217 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1219 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1221 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1223 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1225 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1227 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1229 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1231 struct blk_mq_hw_ctx
*hctx
;
1234 queue_for_each_hw_ctx(q
, hctx
, i
) {
1235 if (!blk_mq_hctx_has_pending(hctx
) ||
1236 blk_mq_hctx_stopped(hctx
))
1239 blk_mq_run_hw_queue(hctx
, async
);
1242 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1245 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1246 * @q: request queue.
1248 * The caller is responsible for serializing this function against
1249 * blk_mq_{start,stop}_hw_queue().
1251 bool blk_mq_queue_stopped(struct request_queue
*q
)
1253 struct blk_mq_hw_ctx
*hctx
;
1256 queue_for_each_hw_ctx(q
, hctx
, i
)
1257 if (blk_mq_hctx_stopped(hctx
))
1262 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1265 * This function is often used for pausing .queue_rq() by driver when
1266 * there isn't enough resource or some conditions aren't satisfied, and
1267 * BLK_STS_RESOURCE is usually returned.
1269 * We do not guarantee that dispatch can be drained or blocked
1270 * after blk_mq_stop_hw_queue() returns. Please use
1271 * blk_mq_quiesce_queue() for that requirement.
1273 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1275 cancel_delayed_work(&hctx
->run_work
);
1277 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1279 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1282 * This function is often used for pausing .queue_rq() by driver when
1283 * there isn't enough resource or some conditions aren't satisfied, and
1284 * BLK_STS_RESOURCE is usually returned.
1286 * We do not guarantee that dispatch can be drained or blocked
1287 * after blk_mq_stop_hw_queues() returns. Please use
1288 * blk_mq_quiesce_queue() for that requirement.
1290 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1292 struct blk_mq_hw_ctx
*hctx
;
1295 queue_for_each_hw_ctx(q
, hctx
, i
)
1296 blk_mq_stop_hw_queue(hctx
);
1298 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1300 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1302 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1304 blk_mq_run_hw_queue(hctx
, false);
1306 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1308 void blk_mq_start_hw_queues(struct request_queue
*q
)
1310 struct blk_mq_hw_ctx
*hctx
;
1313 queue_for_each_hw_ctx(q
, hctx
, i
)
1314 blk_mq_start_hw_queue(hctx
);
1316 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1318 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1320 if (!blk_mq_hctx_stopped(hctx
))
1323 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1324 blk_mq_run_hw_queue(hctx
, async
);
1326 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1328 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1330 struct blk_mq_hw_ctx
*hctx
;
1333 queue_for_each_hw_ctx(q
, hctx
, i
)
1334 blk_mq_start_stopped_hw_queue(hctx
, async
);
1336 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1338 static void blk_mq_run_work_fn(struct work_struct
*work
)
1340 struct blk_mq_hw_ctx
*hctx
;
1342 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1345 * If we are stopped, don't run the queue. The exception is if
1346 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1347 * the STOPPED bit and run it.
1349 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)) {
1350 if (!test_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
))
1353 clear_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1354 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1357 __blk_mq_run_hw_queue(hctx
);
1361 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1363 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx
)))
1367 * Stop the hw queue, then modify currently delayed work.
1368 * This should prevent us from running the queue prematurely.
1369 * Mark the queue as auto-clearing STOPPED when it runs.
1371 blk_mq_stop_hw_queue(hctx
);
1372 set_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1373 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1375 msecs_to_jiffies(msecs
));
1377 EXPORT_SYMBOL(blk_mq_delay_queue
);
1379 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1383 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1385 lockdep_assert_held(&ctx
->lock
);
1387 trace_block_rq_insert(hctx
->queue
, rq
);
1390 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1392 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1395 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1398 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1400 lockdep_assert_held(&ctx
->lock
);
1402 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1403 blk_mq_hctx_mark_pending(hctx
, ctx
);
1407 * Should only be used carefully, when the caller knows we want to
1408 * bypass a potential IO scheduler on the target device.
1410 void blk_mq_request_bypass_insert(struct request
*rq
)
1412 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1413 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(rq
->q
, ctx
->cpu
);
1415 spin_lock(&hctx
->lock
);
1416 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1417 spin_unlock(&hctx
->lock
);
1419 blk_mq_run_hw_queue(hctx
, false);
1422 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1423 struct list_head
*list
)
1427 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1430 spin_lock(&ctx
->lock
);
1431 while (!list_empty(list
)) {
1434 rq
= list_first_entry(list
, struct request
, queuelist
);
1435 BUG_ON(rq
->mq_ctx
!= ctx
);
1436 list_del_init(&rq
->queuelist
);
1437 __blk_mq_insert_req_list(hctx
, rq
, false);
1439 blk_mq_hctx_mark_pending(hctx
, ctx
);
1440 spin_unlock(&ctx
->lock
);
1443 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1445 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1446 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1448 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1449 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1450 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1453 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1455 struct blk_mq_ctx
*this_ctx
;
1456 struct request_queue
*this_q
;
1459 LIST_HEAD(ctx_list
);
1462 list_splice_init(&plug
->mq_list
, &list
);
1464 list_sort(NULL
, &list
, plug_ctx_cmp
);
1470 while (!list_empty(&list
)) {
1471 rq
= list_entry_rq(list
.next
);
1472 list_del_init(&rq
->queuelist
);
1474 if (rq
->mq_ctx
!= this_ctx
) {
1476 trace_block_unplug(this_q
, depth
, from_schedule
);
1477 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1482 this_ctx
= rq
->mq_ctx
;
1488 list_add_tail(&rq
->queuelist
, &ctx_list
);
1492 * If 'this_ctx' is set, we know we have entries to complete
1493 * on 'ctx_list'. Do those.
1496 trace_block_unplug(this_q
, depth
, from_schedule
);
1497 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1502 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1504 blk_init_request_from_bio(rq
, bio
);
1506 blk_account_io_start(rq
, true);
1509 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1511 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1512 !blk_queue_nomerges(hctx
->queue
);
1515 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx
*hctx
,
1516 struct blk_mq_ctx
*ctx
,
1519 spin_lock(&ctx
->lock
);
1520 __blk_mq_insert_request(hctx
, rq
, false);
1521 spin_unlock(&ctx
->lock
);
1524 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1527 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1529 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1532 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1534 blk_qc_t
*cookie
, bool may_sleep
)
1536 struct request_queue
*q
= rq
->q
;
1537 struct blk_mq_queue_data bd
= {
1541 blk_qc_t new_cookie
;
1543 bool run_queue
= true;
1545 /* RCU or SRCU read lock is needed before checking quiesced flag */
1546 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1554 if (!blk_mq_get_driver_tag(rq
, NULL
, false))
1557 new_cookie
= request_to_qc_t(hctx
, rq
);
1560 * For OK queue, we are done. For error, kill it. Any other
1561 * error (busy), just add it to our list as we previously
1564 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1567 *cookie
= new_cookie
;
1569 case BLK_STS_RESOURCE
:
1570 __blk_mq_requeue_request(rq
);
1573 *cookie
= BLK_QC_T_NONE
;
1574 blk_mq_end_request(rq
, ret
);
1579 blk_mq_sched_insert_request(rq
, false, run_queue
, false, may_sleep
);
1582 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1583 struct request
*rq
, blk_qc_t
*cookie
)
1585 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1587 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false);
1590 unsigned int srcu_idx
;
1594 srcu_idx
= srcu_read_lock(hctx
->queue_rq_srcu
);
1595 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, true);
1596 srcu_read_unlock(hctx
->queue_rq_srcu
, srcu_idx
);
1600 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1602 const int is_sync
= op_is_sync(bio
->bi_opf
);
1603 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1604 struct blk_mq_alloc_data data
= { .flags
= 0 };
1606 unsigned int request_count
= 0;
1607 struct blk_plug
*plug
;
1608 struct request
*same_queue_rq
= NULL
;
1610 unsigned int wb_acct
;
1612 blk_queue_bounce(q
, &bio
);
1614 blk_queue_split(q
, &bio
);
1616 if (!bio_integrity_prep(bio
))
1617 return BLK_QC_T_NONE
;
1619 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1620 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1621 return BLK_QC_T_NONE
;
1623 if (blk_mq_sched_bio_merge(q
, bio
))
1624 return BLK_QC_T_NONE
;
1626 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1628 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1630 rq
= blk_mq_get_request(q
, bio
, bio
->bi_opf
, &data
);
1631 if (unlikely(!rq
)) {
1632 __wbt_done(q
->rq_wb
, wb_acct
);
1633 if (bio
->bi_opf
& REQ_NOWAIT
)
1634 bio_wouldblock_error(bio
);
1635 return BLK_QC_T_NONE
;
1638 wbt_track(&rq
->issue_stat
, wb_acct
);
1640 cookie
= request_to_qc_t(data
.hctx
, rq
);
1642 plug
= current
->plug
;
1643 if (unlikely(is_flush_fua
)) {
1644 blk_mq_put_ctx(data
.ctx
);
1645 blk_mq_bio_to_request(rq
, bio
);
1647 blk_mq_sched_insert_request(rq
, false, true, true,
1650 blk_insert_flush(rq
);
1651 blk_mq_run_hw_queue(data
.hctx
, true);
1653 } else if (plug
&& q
->nr_hw_queues
== 1) {
1654 struct request
*last
= NULL
;
1656 blk_mq_put_ctx(data
.ctx
);
1657 blk_mq_bio_to_request(rq
, bio
);
1660 * @request_count may become stale because of schedule
1661 * out, so check the list again.
1663 if (list_empty(&plug
->mq_list
))
1665 else if (blk_queue_nomerges(q
))
1666 request_count
= blk_plug_queued_count(q
);
1669 trace_block_plug(q
);
1671 last
= list_entry_rq(plug
->mq_list
.prev
);
1673 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1674 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1675 blk_flush_plug_list(plug
, false);
1676 trace_block_plug(q
);
1679 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1680 } else if (plug
&& !blk_queue_nomerges(q
)) {
1681 blk_mq_bio_to_request(rq
, bio
);
1684 * We do limited plugging. If the bio can be merged, do that.
1685 * Otherwise the existing request in the plug list will be
1686 * issued. So the plug list will have one request at most
1687 * The plug list might get flushed before this. If that happens,
1688 * the plug list is empty, and same_queue_rq is invalid.
1690 if (list_empty(&plug
->mq_list
))
1691 same_queue_rq
= NULL
;
1693 list_del_init(&same_queue_rq
->queuelist
);
1694 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1696 blk_mq_put_ctx(data
.ctx
);
1698 if (same_queue_rq
) {
1699 data
.hctx
= blk_mq_map_queue(q
,
1700 same_queue_rq
->mq_ctx
->cpu
);
1701 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1704 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1705 blk_mq_put_ctx(data
.ctx
);
1706 blk_mq_bio_to_request(rq
, bio
);
1707 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1708 } else if (q
->elevator
) {
1709 blk_mq_put_ctx(data
.ctx
);
1710 blk_mq_bio_to_request(rq
, bio
);
1711 blk_mq_sched_insert_request(rq
, false, true, true, true);
1713 blk_mq_put_ctx(data
.ctx
);
1714 blk_mq_bio_to_request(rq
, bio
);
1715 blk_mq_queue_io(data
.hctx
, data
.ctx
, rq
);
1716 blk_mq_run_hw_queue(data
.hctx
, true);
1722 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1723 unsigned int hctx_idx
)
1727 if (tags
->rqs
&& set
->ops
->exit_request
) {
1730 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1731 struct request
*rq
= tags
->static_rqs
[i
];
1735 set
->ops
->exit_request(set
, rq
, hctx_idx
);
1736 tags
->static_rqs
[i
] = NULL
;
1740 while (!list_empty(&tags
->page_list
)) {
1741 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1742 list_del_init(&page
->lru
);
1744 * Remove kmemleak object previously allocated in
1745 * blk_mq_init_rq_map().
1747 kmemleak_free(page_address(page
));
1748 __free_pages(page
, page
->private);
1752 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1756 kfree(tags
->static_rqs
);
1757 tags
->static_rqs
= NULL
;
1759 blk_mq_free_tags(tags
);
1762 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1763 unsigned int hctx_idx
,
1764 unsigned int nr_tags
,
1765 unsigned int reserved_tags
)
1767 struct blk_mq_tags
*tags
;
1770 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1771 if (node
== NUMA_NO_NODE
)
1772 node
= set
->numa_node
;
1774 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1775 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1779 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1780 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1783 blk_mq_free_tags(tags
);
1787 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1788 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1790 if (!tags
->static_rqs
) {
1792 blk_mq_free_tags(tags
);
1799 static size_t order_to_size(unsigned int order
)
1801 return (size_t)PAGE_SIZE
<< order
;
1804 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1805 unsigned int hctx_idx
, unsigned int depth
)
1807 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1808 size_t rq_size
, left
;
1811 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1812 if (node
== NUMA_NO_NODE
)
1813 node
= set
->numa_node
;
1815 INIT_LIST_HEAD(&tags
->page_list
);
1818 * rq_size is the size of the request plus driver payload, rounded
1819 * to the cacheline size
1821 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1823 left
= rq_size
* depth
;
1825 for (i
= 0; i
< depth
; ) {
1826 int this_order
= max_order
;
1831 while (this_order
&& left
< order_to_size(this_order
- 1))
1835 page
= alloc_pages_node(node
,
1836 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1842 if (order_to_size(this_order
) < rq_size
)
1849 page
->private = this_order
;
1850 list_add_tail(&page
->lru
, &tags
->page_list
);
1852 p
= page_address(page
);
1854 * Allow kmemleak to scan these pages as they contain pointers
1855 * to additional allocations like via ops->init_request().
1857 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1858 entries_per_page
= order_to_size(this_order
) / rq_size
;
1859 to_do
= min(entries_per_page
, depth
- i
);
1860 left
-= to_do
* rq_size
;
1861 for (j
= 0; j
< to_do
; j
++) {
1862 struct request
*rq
= p
;
1864 tags
->static_rqs
[i
] = rq
;
1865 if (set
->ops
->init_request
) {
1866 if (set
->ops
->init_request(set
, rq
, hctx_idx
,
1868 tags
->static_rqs
[i
] = NULL
;
1880 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1885 * 'cpu' is going away. splice any existing rq_list entries from this
1886 * software queue to the hw queue dispatch list, and ensure that it
1889 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1891 struct blk_mq_hw_ctx
*hctx
;
1892 struct blk_mq_ctx
*ctx
;
1895 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1896 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1898 spin_lock(&ctx
->lock
);
1899 if (!list_empty(&ctx
->rq_list
)) {
1900 list_splice_init(&ctx
->rq_list
, &tmp
);
1901 blk_mq_hctx_clear_pending(hctx
, ctx
);
1903 spin_unlock(&ctx
->lock
);
1905 if (list_empty(&tmp
))
1908 spin_lock(&hctx
->lock
);
1909 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1910 spin_unlock(&hctx
->lock
);
1912 blk_mq_run_hw_queue(hctx
, true);
1916 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1918 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1922 /* hctx->ctxs will be freed in queue's release handler */
1923 static void blk_mq_exit_hctx(struct request_queue
*q
,
1924 struct blk_mq_tag_set
*set
,
1925 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1927 blk_mq_debugfs_unregister_hctx(hctx
);
1929 blk_mq_tag_idle(hctx
);
1931 if (set
->ops
->exit_request
)
1932 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
1934 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1936 if (set
->ops
->exit_hctx
)
1937 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1939 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1940 cleanup_srcu_struct(hctx
->queue_rq_srcu
);
1942 blk_mq_remove_cpuhp(hctx
);
1943 blk_free_flush_queue(hctx
->fq
);
1944 sbitmap_free(&hctx
->ctx_map
);
1947 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1948 struct blk_mq_tag_set
*set
, int nr_queue
)
1950 struct blk_mq_hw_ctx
*hctx
;
1953 queue_for_each_hw_ctx(q
, hctx
, i
) {
1956 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1960 static int blk_mq_init_hctx(struct request_queue
*q
,
1961 struct blk_mq_tag_set
*set
,
1962 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1966 node
= hctx
->numa_node
;
1967 if (node
== NUMA_NO_NODE
)
1968 node
= hctx
->numa_node
= set
->numa_node
;
1970 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1971 spin_lock_init(&hctx
->lock
);
1972 INIT_LIST_HEAD(&hctx
->dispatch
);
1974 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1976 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1978 hctx
->tags
= set
->tags
[hctx_idx
];
1981 * Allocate space for all possible cpus to avoid allocation at
1984 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1987 goto unregister_cpu_notifier
;
1989 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1995 if (set
->ops
->init_hctx
&&
1996 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1999 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
2002 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
2004 goto sched_exit_hctx
;
2006 if (set
->ops
->init_request
&&
2007 set
->ops
->init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
2011 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2012 init_srcu_struct(hctx
->queue_rq_srcu
);
2014 blk_mq_debugfs_register_hctx(q
, hctx
);
2021 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2023 if (set
->ops
->exit_hctx
)
2024 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2026 sbitmap_free(&hctx
->ctx_map
);
2029 unregister_cpu_notifier
:
2030 blk_mq_remove_cpuhp(hctx
);
2034 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2035 unsigned int nr_hw_queues
)
2039 for_each_possible_cpu(i
) {
2040 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2041 struct blk_mq_hw_ctx
*hctx
;
2044 spin_lock_init(&__ctx
->lock
);
2045 INIT_LIST_HEAD(&__ctx
->rq_list
);
2048 /* If the cpu isn't present, the cpu is mapped to first hctx */
2049 if (!cpu_present(i
))
2052 hctx
= blk_mq_map_queue(q
, i
);
2055 * Set local node, IFF we have more than one hw queue. If
2056 * not, we remain on the home node of the device
2058 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2059 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2063 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2067 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2068 set
->queue_depth
, set
->reserved_tags
);
2069 if (!set
->tags
[hctx_idx
])
2072 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2077 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2078 set
->tags
[hctx_idx
] = NULL
;
2082 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2083 unsigned int hctx_idx
)
2085 if (set
->tags
[hctx_idx
]) {
2086 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2087 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2088 set
->tags
[hctx_idx
] = NULL
;
2092 static void blk_mq_map_swqueue(struct request_queue
*q
)
2094 unsigned int i
, hctx_idx
;
2095 struct blk_mq_hw_ctx
*hctx
;
2096 struct blk_mq_ctx
*ctx
;
2097 struct blk_mq_tag_set
*set
= q
->tag_set
;
2100 * Avoid others reading imcomplete hctx->cpumask through sysfs
2102 mutex_lock(&q
->sysfs_lock
);
2104 queue_for_each_hw_ctx(q
, hctx
, i
) {
2105 cpumask_clear(hctx
->cpumask
);
2110 * Map software to hardware queues.
2112 * If the cpu isn't present, the cpu is mapped to first hctx.
2114 for_each_present_cpu(i
) {
2115 hctx_idx
= q
->mq_map
[i
];
2116 /* unmapped hw queue can be remapped after CPU topo changed */
2117 if (!set
->tags
[hctx_idx
] &&
2118 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2120 * If tags initialization fail for some hctx,
2121 * that hctx won't be brought online. In this
2122 * case, remap the current ctx to hctx[0] which
2123 * is guaranteed to always have tags allocated
2128 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2129 hctx
= blk_mq_map_queue(q
, i
);
2131 cpumask_set_cpu(i
, hctx
->cpumask
);
2132 ctx
->index_hw
= hctx
->nr_ctx
;
2133 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2136 mutex_unlock(&q
->sysfs_lock
);
2138 queue_for_each_hw_ctx(q
, hctx
, i
) {
2140 * If no software queues are mapped to this hardware queue,
2141 * disable it and free the request entries.
2143 if (!hctx
->nr_ctx
) {
2144 /* Never unmap queue 0. We need it as a
2145 * fallback in case of a new remap fails
2148 if (i
&& set
->tags
[i
])
2149 blk_mq_free_map_and_requests(set
, i
);
2155 hctx
->tags
= set
->tags
[i
];
2156 WARN_ON(!hctx
->tags
);
2159 * Set the map size to the number of mapped software queues.
2160 * This is more accurate and more efficient than looping
2161 * over all possibly mapped software queues.
2163 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2166 * Initialize batch roundrobin counts
2168 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2169 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2174 * Caller needs to ensure that we're either frozen/quiesced, or that
2175 * the queue isn't live yet.
2177 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2179 struct blk_mq_hw_ctx
*hctx
;
2182 queue_for_each_hw_ctx(q
, hctx
, i
) {
2184 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2185 atomic_inc(&q
->shared_hctx_restart
);
2186 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2188 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2189 atomic_dec(&q
->shared_hctx_restart
);
2190 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2195 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2198 struct request_queue
*q
;
2200 lockdep_assert_held(&set
->tag_list_lock
);
2202 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2203 blk_mq_freeze_queue(q
);
2204 queue_set_hctx_shared(q
, shared
);
2205 blk_mq_unfreeze_queue(q
);
2209 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2211 struct blk_mq_tag_set
*set
= q
->tag_set
;
2213 mutex_lock(&set
->tag_list_lock
);
2214 list_del_rcu(&q
->tag_set_list
);
2215 INIT_LIST_HEAD(&q
->tag_set_list
);
2216 if (list_is_singular(&set
->tag_list
)) {
2217 /* just transitioned to unshared */
2218 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2219 /* update existing queue */
2220 blk_mq_update_tag_set_depth(set
, false);
2222 mutex_unlock(&set
->tag_list_lock
);
2227 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2228 struct request_queue
*q
)
2232 mutex_lock(&set
->tag_list_lock
);
2234 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2235 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2236 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2237 /* update existing queue */
2238 blk_mq_update_tag_set_depth(set
, true);
2240 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2241 queue_set_hctx_shared(q
, true);
2242 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2244 mutex_unlock(&set
->tag_list_lock
);
2248 * It is the actual release handler for mq, but we do it from
2249 * request queue's release handler for avoiding use-after-free
2250 * and headache because q->mq_kobj shouldn't have been introduced,
2251 * but we can't group ctx/kctx kobj without it.
2253 void blk_mq_release(struct request_queue
*q
)
2255 struct blk_mq_hw_ctx
*hctx
;
2258 /* hctx kobj stays in hctx */
2259 queue_for_each_hw_ctx(q
, hctx
, i
) {
2262 kobject_put(&hctx
->kobj
);
2267 kfree(q
->queue_hw_ctx
);
2270 * release .mq_kobj and sw queue's kobject now because
2271 * both share lifetime with request queue.
2273 blk_mq_sysfs_deinit(q
);
2275 free_percpu(q
->queue_ctx
);
2278 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2280 struct request_queue
*uninit_q
, *q
;
2282 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2284 return ERR_PTR(-ENOMEM
);
2286 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2288 blk_cleanup_queue(uninit_q
);
2292 EXPORT_SYMBOL(blk_mq_init_queue
);
2294 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2296 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2298 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, queue_rq_srcu
),
2299 __alignof__(struct blk_mq_hw_ctx
)) !=
2300 sizeof(struct blk_mq_hw_ctx
));
2302 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2303 hw_ctx_size
+= sizeof(struct srcu_struct
);
2308 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2309 struct request_queue
*q
)
2312 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2314 blk_mq_sysfs_unregister(q
);
2315 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2321 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2322 hctxs
[i
] = kzalloc_node(blk_mq_hw_ctx_size(set
),
2327 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2334 atomic_set(&hctxs
[i
]->nr_active
, 0);
2335 hctxs
[i
]->numa_node
= node
;
2336 hctxs
[i
]->queue_num
= i
;
2338 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2339 free_cpumask_var(hctxs
[i
]->cpumask
);
2344 blk_mq_hctx_kobj_init(hctxs
[i
]);
2346 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2347 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2351 blk_mq_free_map_and_requests(set
, j
);
2352 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2353 kobject_put(&hctx
->kobj
);
2358 q
->nr_hw_queues
= i
;
2359 blk_mq_sysfs_register(q
);
2362 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2363 struct request_queue
*q
)
2365 /* mark the queue as mq asap */
2366 q
->mq_ops
= set
->ops
;
2368 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2369 blk_mq_poll_stats_bkt
,
2370 BLK_MQ_POLL_STATS_BKTS
, q
);
2374 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2378 /* init q->mq_kobj and sw queues' kobjects */
2379 blk_mq_sysfs_init(q
);
2381 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2382 GFP_KERNEL
, set
->numa_node
);
2383 if (!q
->queue_hw_ctx
)
2386 q
->mq_map
= set
->mq_map
;
2388 blk_mq_realloc_hw_ctxs(set
, q
);
2389 if (!q
->nr_hw_queues
)
2392 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2393 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2395 q
->nr_queues
= nr_cpu_ids
;
2397 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2399 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2400 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2402 q
->sg_reserved_size
= INT_MAX
;
2404 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2405 INIT_LIST_HEAD(&q
->requeue_list
);
2406 spin_lock_init(&q
->requeue_lock
);
2408 blk_queue_make_request(q
, blk_mq_make_request
);
2411 * Do this after blk_queue_make_request() overrides it...
2413 q
->nr_requests
= set
->queue_depth
;
2416 * Default to classic polling
2420 if (set
->ops
->complete
)
2421 blk_queue_softirq_done(q
, set
->ops
->complete
);
2423 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2424 blk_mq_add_queue_tag_set(set
, q
);
2425 blk_mq_map_swqueue(q
);
2427 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2430 ret
= blk_mq_sched_init(q
);
2432 return ERR_PTR(ret
);
2438 kfree(q
->queue_hw_ctx
);
2440 free_percpu(q
->queue_ctx
);
2443 return ERR_PTR(-ENOMEM
);
2445 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2447 void blk_mq_free_queue(struct request_queue
*q
)
2449 struct blk_mq_tag_set
*set
= q
->tag_set
;
2451 blk_mq_del_queue_tag_set(q
);
2452 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2455 /* Basically redo blk_mq_init_queue with queue frozen */
2456 static void blk_mq_queue_reinit(struct request_queue
*q
)
2458 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2460 blk_mq_debugfs_unregister_hctxs(q
);
2461 blk_mq_sysfs_unregister(q
);
2464 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2465 * we should change hctx numa_node according to new topology (this
2466 * involves free and re-allocate memory, worthy doing?)
2469 blk_mq_map_swqueue(q
);
2471 blk_mq_sysfs_register(q
);
2472 blk_mq_debugfs_register_hctxs(q
);
2475 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2479 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2480 if (!__blk_mq_alloc_rq_map(set
, i
))
2487 blk_mq_free_rq_map(set
->tags
[i
]);
2493 * Allocate the request maps associated with this tag_set. Note that this
2494 * may reduce the depth asked for, if memory is tight. set->queue_depth
2495 * will be updated to reflect the allocated depth.
2497 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2502 depth
= set
->queue_depth
;
2504 err
= __blk_mq_alloc_rq_maps(set
);
2508 set
->queue_depth
>>= 1;
2509 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2513 } while (set
->queue_depth
);
2515 if (!set
->queue_depth
|| err
) {
2516 pr_err("blk-mq: failed to allocate request map\n");
2520 if (depth
!= set
->queue_depth
)
2521 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2522 depth
, set
->queue_depth
);
2527 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2529 if (set
->ops
->map_queues
)
2530 return set
->ops
->map_queues(set
);
2532 return blk_mq_map_queues(set
);
2536 * Alloc a tag set to be associated with one or more request queues.
2537 * May fail with EINVAL for various error conditions. May adjust the
2538 * requested depth down, if if it too large. In that case, the set
2539 * value will be stored in set->queue_depth.
2541 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2545 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2547 if (!set
->nr_hw_queues
)
2549 if (!set
->queue_depth
)
2551 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2554 if (!set
->ops
->queue_rq
)
2557 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2558 pr_info("blk-mq: reduced tag depth to %u\n",
2560 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2564 * If a crashdump is active, then we are potentially in a very
2565 * memory constrained environment. Limit us to 1 queue and
2566 * 64 tags to prevent using too much memory.
2568 if (is_kdump_kernel()) {
2569 set
->nr_hw_queues
= 1;
2570 set
->queue_depth
= min(64U, set
->queue_depth
);
2573 * There is no use for more h/w queues than cpus.
2575 if (set
->nr_hw_queues
> nr_cpu_ids
)
2576 set
->nr_hw_queues
= nr_cpu_ids
;
2578 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2579 GFP_KERNEL
, set
->numa_node
);
2584 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2585 GFP_KERNEL
, set
->numa_node
);
2589 ret
= blk_mq_update_queue_map(set
);
2591 goto out_free_mq_map
;
2593 ret
= blk_mq_alloc_rq_maps(set
);
2595 goto out_free_mq_map
;
2597 mutex_init(&set
->tag_list_lock
);
2598 INIT_LIST_HEAD(&set
->tag_list
);
2610 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2612 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2616 for (i
= 0; i
< nr_cpu_ids
; i
++)
2617 blk_mq_free_map_and_requests(set
, i
);
2625 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2627 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2629 struct blk_mq_tag_set
*set
= q
->tag_set
;
2630 struct blk_mq_hw_ctx
*hctx
;
2636 blk_mq_freeze_queue(q
);
2639 queue_for_each_hw_ctx(q
, hctx
, i
) {
2643 * If we're using an MQ scheduler, just update the scheduler
2644 * queue depth. This is similar to what the old code would do.
2646 if (!hctx
->sched_tags
) {
2647 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2648 min(nr
, set
->queue_depth
),
2651 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2659 q
->nr_requests
= nr
;
2661 blk_mq_unfreeze_queue(q
);
2666 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
2669 struct request_queue
*q
;
2671 lockdep_assert_held(&set
->tag_list_lock
);
2673 if (nr_hw_queues
> nr_cpu_ids
)
2674 nr_hw_queues
= nr_cpu_ids
;
2675 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2678 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2679 blk_mq_freeze_queue(q
);
2681 set
->nr_hw_queues
= nr_hw_queues
;
2682 blk_mq_update_queue_map(set
);
2683 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2684 blk_mq_realloc_hw_ctxs(set
, q
);
2685 blk_mq_queue_reinit(q
);
2688 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2689 blk_mq_unfreeze_queue(q
);
2692 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2694 mutex_lock(&set
->tag_list_lock
);
2695 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
2696 mutex_unlock(&set
->tag_list_lock
);
2698 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2700 /* Enable polling stats and return whether they were already enabled. */
2701 static bool blk_poll_stats_enable(struct request_queue
*q
)
2703 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2704 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
2706 blk_stat_add_callback(q
, q
->poll_cb
);
2710 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2713 * We don't arm the callback if polling stats are not enabled or the
2714 * callback is already active.
2716 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2717 blk_stat_is_active(q
->poll_cb
))
2720 blk_stat_activate_msecs(q
->poll_cb
, 100);
2723 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2725 struct request_queue
*q
= cb
->data
;
2728 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
2729 if (cb
->stat
[bucket
].nr_samples
)
2730 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
2734 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2735 struct blk_mq_hw_ctx
*hctx
,
2738 unsigned long ret
= 0;
2742 * If stats collection isn't on, don't sleep but turn it on for
2745 if (!blk_poll_stats_enable(q
))
2749 * As an optimistic guess, use half of the mean service time
2750 * for this type of request. We can (and should) make this smarter.
2751 * For instance, if the completion latencies are tight, we can
2752 * get closer than just half the mean. This is especially
2753 * important on devices where the completion latencies are longer
2754 * than ~10 usec. We do use the stats for the relevant IO size
2755 * if available which does lead to better estimates.
2757 bucket
= blk_mq_poll_stats_bkt(rq
);
2761 if (q
->poll_stat
[bucket
].nr_samples
)
2762 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
2767 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2768 struct blk_mq_hw_ctx
*hctx
,
2771 struct hrtimer_sleeper hs
;
2772 enum hrtimer_mode mode
;
2776 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2782 * -1: don't ever hybrid sleep
2783 * 0: use half of prev avg
2784 * >0: use this specific value
2786 if (q
->poll_nsec
== -1)
2788 else if (q
->poll_nsec
> 0)
2789 nsecs
= q
->poll_nsec
;
2791 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2796 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2799 * This will be replaced with the stats tracking code, using
2800 * 'avg_completion_time / 2' as the pre-sleep target.
2804 mode
= HRTIMER_MODE_REL
;
2805 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2806 hrtimer_set_expires(&hs
.timer
, kt
);
2808 hrtimer_init_sleeper(&hs
, current
);
2810 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2812 set_current_state(TASK_UNINTERRUPTIBLE
);
2813 hrtimer_start_expires(&hs
.timer
, mode
);
2816 hrtimer_cancel(&hs
.timer
);
2817 mode
= HRTIMER_MODE_ABS
;
2818 } while (hs
.task
&& !signal_pending(current
));
2820 __set_current_state(TASK_RUNNING
);
2821 destroy_hrtimer_on_stack(&hs
.timer
);
2825 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2827 struct request_queue
*q
= hctx
->queue
;
2831 * If we sleep, have the caller restart the poll loop to reset
2832 * the state. Like for the other success return cases, the
2833 * caller is responsible for checking if the IO completed. If
2834 * the IO isn't complete, we'll get called again and will go
2835 * straight to the busy poll loop.
2837 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2840 hctx
->poll_considered
++;
2842 state
= current
->state
;
2843 while (!need_resched()) {
2846 hctx
->poll_invoked
++;
2848 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2850 hctx
->poll_success
++;
2851 set_current_state(TASK_RUNNING
);
2855 if (signal_pending_state(state
, current
))
2856 set_current_state(TASK_RUNNING
);
2858 if (current
->state
== TASK_RUNNING
)
2868 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2870 struct blk_mq_hw_ctx
*hctx
;
2871 struct blk_plug
*plug
;
2874 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2875 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2878 plug
= current
->plug
;
2880 blk_flush_plug_list(plug
, false);
2882 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2883 if (!blk_qc_t_is_internal(cookie
))
2884 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2886 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2888 * With scheduling, if the request has completed, we'll
2889 * get a NULL return here, as we clear the sched tag when
2890 * that happens. The request still remains valid, like always,
2891 * so we should be safe with just the NULL check.
2897 return __blk_mq_poll(hctx
, rq
);
2899 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2901 static int __init
blk_mq_init(void)
2903 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2904 blk_mq_hctx_notify_dead
);
2907 subsys_initcall(blk_mq_init
);