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
);
162 blk_mq_freeze_queue_wait(q
);
165 void blk_mq_freeze_queue(struct request_queue
*q
)
168 * ...just an alias to keep freeze and unfreeze actions balanced
169 * in the blk_mq_* namespace
173 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
175 void blk_mq_unfreeze_queue(struct request_queue
*q
)
179 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
180 WARN_ON_ONCE(freeze_depth
< 0);
182 percpu_ref_reinit(&q
->q_usage_counter
);
183 wake_up_all(&q
->mq_freeze_wq
);
186 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
189 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
190 * mpt3sas driver such that this function can be removed.
192 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
196 spin_lock_irqsave(q
->queue_lock
, flags
);
197 queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
198 spin_unlock_irqrestore(q
->queue_lock
, flags
);
200 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
203 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
206 * Note: this function does not prevent that the struct request end_io()
207 * callback function is invoked. Once this function is returned, we make
208 * sure no dispatch can happen until the queue is unquiesced via
209 * blk_mq_unquiesce_queue().
211 void blk_mq_quiesce_queue(struct request_queue
*q
)
213 struct blk_mq_hw_ctx
*hctx
;
217 blk_mq_quiesce_queue_nowait(q
);
219 queue_for_each_hw_ctx(q
, hctx
, i
) {
220 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
221 synchronize_srcu(hctx
->queue_rq_srcu
);
228 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
231 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
234 * This function recovers queue into the state before quiescing
235 * which is done by blk_mq_quiesce_queue.
237 void blk_mq_unquiesce_queue(struct request_queue
*q
)
241 spin_lock_irqsave(q
->queue_lock
, flags
);
242 queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
243 spin_unlock_irqrestore(q
->queue_lock
, flags
);
245 /* dispatch requests which are inserted during quiescing */
246 blk_mq_run_hw_queues(q
, true);
248 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
250 void blk_mq_wake_waiters(struct request_queue
*q
)
252 struct blk_mq_hw_ctx
*hctx
;
255 queue_for_each_hw_ctx(q
, hctx
, i
)
256 if (blk_mq_hw_queue_mapped(hctx
))
257 blk_mq_tag_wakeup_all(hctx
->tags
, true);
260 * If we are called because the queue has now been marked as
261 * dying, we need to ensure that processes currently waiting on
262 * the queue are notified as well.
264 wake_up_all(&q
->mq_freeze_wq
);
267 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
269 return blk_mq_has_free_tags(hctx
->tags
);
271 EXPORT_SYMBOL(blk_mq_can_queue
);
273 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
274 unsigned int tag
, unsigned int op
)
276 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
277 struct request
*rq
= tags
->static_rqs
[tag
];
281 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
283 rq
->internal_tag
= tag
;
285 if (blk_mq_tag_busy(data
->hctx
)) {
286 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
287 atomic_inc(&data
->hctx
->nr_active
);
290 rq
->internal_tag
= -1;
291 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
294 INIT_LIST_HEAD(&rq
->queuelist
);
295 /* csd/requeue_work/fifo_time is initialized before use */
297 rq
->mq_ctx
= data
->ctx
;
299 if (blk_queue_io_stat(data
->q
))
300 rq
->rq_flags
|= RQF_IO_STAT
;
301 /* do not touch atomic flags, it needs atomic ops against the timer */
303 INIT_HLIST_NODE(&rq
->hash
);
304 RB_CLEAR_NODE(&rq
->rb_node
);
307 rq
->start_time
= jiffies
;
308 #ifdef CONFIG_BLK_CGROUP
310 set_start_time_ns(rq
);
311 rq
->io_start_time_ns
= 0;
313 rq
->nr_phys_segments
= 0;
314 #if defined(CONFIG_BLK_DEV_INTEGRITY)
315 rq
->nr_integrity_segments
= 0;
318 /* tag was already set */
321 INIT_LIST_HEAD(&rq
->timeout_list
);
325 rq
->end_io_data
= NULL
;
328 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
332 static struct request
*blk_mq_get_request(struct request_queue
*q
,
333 struct bio
*bio
, unsigned int op
,
334 struct blk_mq_alloc_data
*data
)
336 struct elevator_queue
*e
= q
->elevator
;
339 struct blk_mq_ctx
*local_ctx
= NULL
;
341 blk_queue_enter_live(q
);
343 if (likely(!data
->ctx
))
344 data
->ctx
= local_ctx
= blk_mq_get_ctx(q
);
345 if (likely(!data
->hctx
))
346 data
->hctx
= blk_mq_map_queue(q
, data
->ctx
->cpu
);
348 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
351 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
354 * Flush requests are special and go directly to the
357 if (!op_is_flush(op
) && e
->type
->ops
.mq
.limit_depth
)
358 e
->type
->ops
.mq
.limit_depth(op
, data
);
361 tag
= blk_mq_get_tag(data
);
362 if (tag
== BLK_MQ_TAG_FAIL
) {
364 blk_mq_put_ctx(local_ctx
);
371 rq
= blk_mq_rq_ctx_init(data
, tag
, op
);
372 if (!op_is_flush(op
)) {
374 if (e
&& e
->type
->ops
.mq
.prepare_request
) {
375 if (e
->type
->icq_cache
&& rq_ioc(bio
))
376 blk_mq_sched_assign_ioc(rq
, bio
);
378 e
->type
->ops
.mq
.prepare_request(rq
, bio
);
379 rq
->rq_flags
|= RQF_ELVPRIV
;
382 data
->hctx
->queued
++;
386 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
389 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
393 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
397 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
401 return ERR_PTR(-EWOULDBLOCK
);
403 blk_mq_put_ctx(alloc_data
.ctx
);
406 rq
->__sector
= (sector_t
) -1;
407 rq
->bio
= rq
->biotail
= NULL
;
410 EXPORT_SYMBOL(blk_mq_alloc_request
);
412 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
413 unsigned int op
, unsigned int flags
, unsigned int hctx_idx
)
415 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
421 * If the tag allocator sleeps we could get an allocation for a
422 * different hardware context. No need to complicate the low level
423 * allocator for this for the rare use case of a command tied to
426 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
427 return ERR_PTR(-EINVAL
);
429 if (hctx_idx
>= q
->nr_hw_queues
)
430 return ERR_PTR(-EIO
);
432 ret
= blk_queue_enter(q
, true);
437 * Check if the hardware context is actually mapped to anything.
438 * If not tell the caller that it should skip this queue.
440 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
441 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
443 return ERR_PTR(-EXDEV
);
445 cpu
= cpumask_first(alloc_data
.hctx
->cpumask
);
446 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
448 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
452 return ERR_PTR(-EWOULDBLOCK
);
456 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
458 void blk_mq_free_request(struct request
*rq
)
460 struct request_queue
*q
= rq
->q
;
461 struct elevator_queue
*e
= q
->elevator
;
462 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
463 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
464 const int sched_tag
= rq
->internal_tag
;
466 if (rq
->rq_flags
& RQF_ELVPRIV
) {
467 if (e
&& e
->type
->ops
.mq
.finish_request
)
468 e
->type
->ops
.mq
.finish_request(rq
);
470 put_io_context(rq
->elv
.icq
->ioc
);
475 ctx
->rq_completed
[rq_is_sync(rq
)]++;
476 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
477 atomic_dec(&hctx
->nr_active
);
479 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
480 laptop_io_completion(q
->backing_dev_info
);
482 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
484 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
485 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
487 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
489 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
490 blk_mq_sched_restart(hctx
);
493 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
495 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
497 blk_account_io_done(rq
);
500 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
501 rq
->end_io(rq
, error
);
503 if (unlikely(blk_bidi_rq(rq
)))
504 blk_mq_free_request(rq
->next_rq
);
505 blk_mq_free_request(rq
);
508 EXPORT_SYMBOL(__blk_mq_end_request
);
510 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
512 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
514 __blk_mq_end_request(rq
, error
);
516 EXPORT_SYMBOL(blk_mq_end_request
);
518 static void __blk_mq_complete_request_remote(void *data
)
520 struct request
*rq
= data
;
522 rq
->q
->softirq_done_fn(rq
);
525 static void __blk_mq_complete_request(struct request
*rq
)
527 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
531 if (rq
->internal_tag
!= -1)
532 blk_mq_sched_completed_request(rq
);
533 if (rq
->rq_flags
& RQF_STATS
) {
534 blk_mq_poll_stats_start(rq
->q
);
538 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
539 rq
->q
->softirq_done_fn(rq
);
544 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
545 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
547 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
548 rq
->csd
.func
= __blk_mq_complete_request_remote
;
551 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
553 rq
->q
->softirq_done_fn(rq
);
559 * blk_mq_complete_request - end I/O on a request
560 * @rq: the request being processed
563 * Ends all I/O on a request. It does not handle partial completions.
564 * The actual completion happens out-of-order, through a IPI handler.
566 void blk_mq_complete_request(struct request
*rq
)
568 struct request_queue
*q
= rq
->q
;
570 if (unlikely(blk_should_fake_timeout(q
)))
572 if (!blk_mark_rq_complete(rq
))
573 __blk_mq_complete_request(rq
);
575 EXPORT_SYMBOL(blk_mq_complete_request
);
577 int blk_mq_request_started(struct request
*rq
)
579 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
581 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
583 void blk_mq_start_request(struct request
*rq
)
585 struct request_queue
*q
= rq
->q
;
587 blk_mq_sched_started_request(rq
);
589 trace_block_rq_issue(q
, rq
);
591 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
592 blk_stat_set_issue(&rq
->issue_stat
, blk_rq_sectors(rq
));
593 rq
->rq_flags
|= RQF_STATS
;
594 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
600 * Ensure that ->deadline is visible before set the started
601 * flag and clear the completed flag.
603 smp_mb__before_atomic();
606 * Mark us as started and clear complete. Complete might have been
607 * set if requeue raced with timeout, which then marked it as
608 * complete. So be sure to clear complete again when we start
609 * the request, otherwise we'll ignore the completion event.
611 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
612 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
613 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
614 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
616 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
618 * Make sure space for the drain appears. We know we can do
619 * this because max_hw_segments has been adjusted to be one
620 * fewer than the device can handle.
622 rq
->nr_phys_segments
++;
625 EXPORT_SYMBOL(blk_mq_start_request
);
628 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
629 * flag isn't set yet, so there may be race with timeout handler,
630 * but given rq->deadline is just set in .queue_rq() under
631 * this situation, the race won't be possible in reality because
632 * rq->timeout should be set as big enough to cover the window
633 * between blk_mq_start_request() called from .queue_rq() and
634 * clearing REQ_ATOM_STARTED here.
636 static void __blk_mq_requeue_request(struct request
*rq
)
638 struct request_queue
*q
= rq
->q
;
640 trace_block_rq_requeue(q
, rq
);
641 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
642 blk_mq_sched_requeue_request(rq
);
644 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
645 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
646 rq
->nr_phys_segments
--;
650 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
652 __blk_mq_requeue_request(rq
);
654 BUG_ON(blk_queued_rq(rq
));
655 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
657 EXPORT_SYMBOL(blk_mq_requeue_request
);
659 static void blk_mq_requeue_work(struct work_struct
*work
)
661 struct request_queue
*q
=
662 container_of(work
, struct request_queue
, requeue_work
.work
);
664 struct request
*rq
, *next
;
666 spin_lock_irq(&q
->requeue_lock
);
667 list_splice_init(&q
->requeue_list
, &rq_list
);
668 spin_unlock_irq(&q
->requeue_lock
);
670 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
671 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
674 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
675 list_del_init(&rq
->queuelist
);
676 blk_mq_sched_insert_request(rq
, true, false, false, true);
679 while (!list_empty(&rq_list
)) {
680 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
681 list_del_init(&rq
->queuelist
);
682 blk_mq_sched_insert_request(rq
, false, false, false, true);
685 blk_mq_run_hw_queues(q
, false);
688 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
689 bool kick_requeue_list
)
691 struct request_queue
*q
= rq
->q
;
695 * We abuse this flag that is otherwise used by the I/O scheduler to
696 * request head insertation from the workqueue.
698 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
700 spin_lock_irqsave(&q
->requeue_lock
, flags
);
702 rq
->rq_flags
|= RQF_SOFTBARRIER
;
703 list_add(&rq
->queuelist
, &q
->requeue_list
);
705 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
707 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
709 if (kick_requeue_list
)
710 blk_mq_kick_requeue_list(q
);
712 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
714 void blk_mq_kick_requeue_list(struct request_queue
*q
)
716 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
718 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
720 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
723 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
724 msecs_to_jiffies(msecs
));
726 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
728 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
730 if (tag
< tags
->nr_tags
) {
731 prefetch(tags
->rqs
[tag
]);
732 return tags
->rqs
[tag
];
737 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
739 struct blk_mq_timeout_data
{
741 unsigned int next_set
;
744 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
746 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
747 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
750 * We know that complete is set at this point. If STARTED isn't set
751 * anymore, then the request isn't active and the "timeout" should
752 * just be ignored. This can happen due to the bitflag ordering.
753 * Timeout first checks if STARTED is set, and if it is, assumes
754 * the request is active. But if we race with completion, then
755 * both flags will get cleared. So check here again, and ignore
756 * a timeout event with a request that isn't active.
758 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
762 ret
= ops
->timeout(req
, reserved
);
766 __blk_mq_complete_request(req
);
768 case BLK_EH_RESET_TIMER
:
770 blk_clear_rq_complete(req
);
772 case BLK_EH_NOT_HANDLED
:
775 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
780 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
781 struct request
*rq
, void *priv
, bool reserved
)
783 struct blk_mq_timeout_data
*data
= priv
;
785 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
789 * The rq being checked may have been freed and reallocated
790 * out already here, we avoid this race by checking rq->deadline
791 * and REQ_ATOM_COMPLETE flag together:
793 * - if rq->deadline is observed as new value because of
794 * reusing, the rq won't be timed out because of timing.
795 * - if rq->deadline is observed as previous value,
796 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
797 * because we put a barrier between setting rq->deadline
798 * and clearing the flag in blk_mq_start_request(), so
799 * this rq won't be timed out too.
801 if (time_after_eq(jiffies
, rq
->deadline
)) {
802 if (!blk_mark_rq_complete(rq
))
803 blk_mq_rq_timed_out(rq
, reserved
);
804 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
805 data
->next
= rq
->deadline
;
810 static void blk_mq_timeout_work(struct work_struct
*work
)
812 struct request_queue
*q
=
813 container_of(work
, struct request_queue
, timeout_work
);
814 struct blk_mq_timeout_data data
= {
820 /* A deadlock might occur if a request is stuck requiring a
821 * timeout at the same time a queue freeze is waiting
822 * completion, since the timeout code would not be able to
823 * acquire the queue reference here.
825 * That's why we don't use blk_queue_enter here; instead, we use
826 * percpu_ref_tryget directly, because we need to be able to
827 * obtain a reference even in the short window between the queue
828 * starting to freeze, by dropping the first reference in
829 * blk_freeze_queue_start, and the moment the last request is
830 * consumed, marked by the instant q_usage_counter reaches
833 if (!percpu_ref_tryget(&q
->q_usage_counter
))
836 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
839 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
840 mod_timer(&q
->timeout
, data
.next
);
842 struct blk_mq_hw_ctx
*hctx
;
844 queue_for_each_hw_ctx(q
, hctx
, i
) {
845 /* the hctx may be unmapped, so check it here */
846 if (blk_mq_hw_queue_mapped(hctx
))
847 blk_mq_tag_idle(hctx
);
853 struct flush_busy_ctx_data
{
854 struct blk_mq_hw_ctx
*hctx
;
855 struct list_head
*list
;
858 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
860 struct flush_busy_ctx_data
*flush_data
= data
;
861 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
862 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
864 sbitmap_clear_bit(sb
, bitnr
);
865 spin_lock(&ctx
->lock
);
866 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
867 spin_unlock(&ctx
->lock
);
872 * Process software queues that have been marked busy, splicing them
873 * to the for-dispatch
875 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
877 struct flush_busy_ctx_data data
= {
882 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
884 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
886 static inline unsigned int queued_to_index(unsigned int queued
)
891 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
894 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
897 struct blk_mq_alloc_data data
= {
899 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
900 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
903 might_sleep_if(wait
);
908 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
909 data
.flags
|= BLK_MQ_REQ_RESERVED
;
911 rq
->tag
= blk_mq_get_tag(&data
);
913 if (blk_mq_tag_busy(data
.hctx
)) {
914 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
915 atomic_inc(&data
.hctx
->nr_active
);
917 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
923 return rq
->tag
!= -1;
926 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx
*hctx
,
929 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
932 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
933 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
934 atomic_dec(&hctx
->nr_active
);
938 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx
*hctx
,
941 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
944 __blk_mq_put_driver_tag(hctx
, rq
);
947 static void blk_mq_put_driver_tag(struct request
*rq
)
949 struct blk_mq_hw_ctx
*hctx
;
951 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
954 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
955 __blk_mq_put_driver_tag(hctx
, rq
);
959 * If we fail getting a driver tag because all the driver tags are already
960 * assigned and on the dispatch list, BUT the first entry does not have a
961 * tag, then we could deadlock. For that case, move entries with assigned
962 * driver tags to the front, leaving the set of tagged requests in the
963 * same order, and the untagged set in the same order.
965 static bool reorder_tags_to_front(struct list_head
*list
)
967 struct request
*rq
, *tmp
, *first
= NULL
;
969 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
973 list_move(&rq
->queuelist
, list
);
979 return first
!= NULL
;
982 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
, int flags
,
985 struct blk_mq_hw_ctx
*hctx
;
987 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
989 list_del(&wait
->entry
);
990 clear_bit_unlock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
);
991 blk_mq_run_hw_queue(hctx
, true);
995 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx
*hctx
)
997 struct sbq_wait_state
*ws
;
1000 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
1001 * The thread which wins the race to grab this bit adds the hardware
1002 * queue to the wait queue.
1004 if (test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
) ||
1005 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1008 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
1009 ws
= bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
);
1012 * As soon as this returns, it's no longer safe to fiddle with
1013 * hctx->dispatch_wait, since a completion can wake up the wait queue
1014 * and unlock the bit.
1016 add_wait_queue(&ws
->wait
, &hctx
->dispatch_wait
);
1020 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
)
1022 struct blk_mq_hw_ctx
*hctx
;
1026 if (list_empty(list
))
1030 * Now process all the entries, sending them to the driver.
1032 errors
= queued
= 0;
1034 struct blk_mq_queue_data bd
;
1037 rq
= list_first_entry(list
, struct request
, queuelist
);
1038 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
1039 if (!queued
&& reorder_tags_to_front(list
))
1043 * The initial allocation attempt failed, so we need to
1044 * rerun the hardware queue when a tag is freed.
1046 if (!blk_mq_dispatch_wait_add(hctx
))
1050 * It's possible that a tag was freed in the window
1051 * between the allocation failure and adding the
1052 * hardware queue to the wait queue.
1054 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1058 list_del_init(&rq
->queuelist
);
1063 * Flag last if we have no more requests, or if we have more
1064 * but can't assign a driver tag to it.
1066 if (list_empty(list
))
1069 struct request
*nxt
;
1071 nxt
= list_first_entry(list
, struct request
, queuelist
);
1072 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1075 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1076 if (ret
== BLK_STS_RESOURCE
) {
1077 blk_mq_put_driver_tag_hctx(hctx
, rq
);
1078 list_add(&rq
->queuelist
, list
);
1079 __blk_mq_requeue_request(rq
);
1083 if (unlikely(ret
!= BLK_STS_OK
)) {
1085 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1090 } while (!list_empty(list
));
1092 hctx
->dispatched
[queued_to_index(queued
)]++;
1095 * Any items that need requeuing? Stuff them into hctx->dispatch,
1096 * that is where we will continue on next queue run.
1098 if (!list_empty(list
)) {
1100 * If an I/O scheduler has been configured and we got a driver
1101 * tag for the next request already, free it again.
1103 rq
= list_first_entry(list
, struct request
, queuelist
);
1104 blk_mq_put_driver_tag(rq
);
1106 spin_lock(&hctx
->lock
);
1107 list_splice_init(list
, &hctx
->dispatch
);
1108 spin_unlock(&hctx
->lock
);
1111 * If SCHED_RESTART was set by the caller of this function and
1112 * it is no longer set that means that it was cleared by another
1113 * thread and hence that a queue rerun is needed.
1115 * If TAG_WAITING is set that means that an I/O scheduler has
1116 * been configured and another thread is waiting for a driver
1117 * tag. To guarantee fairness, do not rerun this hardware queue
1118 * but let the other thread grab the driver tag.
1120 * If no I/O scheduler has been configured it is possible that
1121 * the hardware queue got stopped and restarted before requests
1122 * were pushed back onto the dispatch list. Rerun the queue to
1123 * avoid starvation. Notes:
1124 * - blk_mq_run_hw_queue() checks whether or not a queue has
1125 * been stopped before rerunning a queue.
1126 * - Some but not all block drivers stop a queue before
1127 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1130 if (!blk_mq_sched_needs_restart(hctx
) &&
1131 !test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1132 blk_mq_run_hw_queue(hctx
, true);
1135 return (queued
+ errors
) != 0;
1138 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1143 * We should be running this queue from one of the CPUs that
1146 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1147 cpu_online(hctx
->next_cpu
));
1150 * We can't run the queue inline with ints disabled. Ensure that
1151 * we catch bad users of this early.
1153 WARN_ON_ONCE(in_interrupt());
1155 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1157 blk_mq_sched_dispatch_requests(hctx
);
1162 srcu_idx
= srcu_read_lock(hctx
->queue_rq_srcu
);
1163 blk_mq_sched_dispatch_requests(hctx
);
1164 srcu_read_unlock(hctx
->queue_rq_srcu
, srcu_idx
);
1169 * It'd be great if the workqueue API had a way to pass
1170 * in a mask and had some smarts for more clever placement.
1171 * For now we just round-robin here, switching for every
1172 * BLK_MQ_CPU_WORK_BATCH queued items.
1174 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1176 if (hctx
->queue
->nr_hw_queues
== 1)
1177 return WORK_CPU_UNBOUND
;
1179 if (--hctx
->next_cpu_batch
<= 0) {
1182 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1183 if (next_cpu
>= nr_cpu_ids
)
1184 next_cpu
= cpumask_first(hctx
->cpumask
);
1186 hctx
->next_cpu
= next_cpu
;
1187 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1190 return hctx
->next_cpu
;
1193 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1194 unsigned long msecs
)
1196 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx
)))
1199 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1202 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1203 int cpu
= get_cpu();
1204 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1205 __blk_mq_run_hw_queue(hctx
);
1213 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1215 msecs_to_jiffies(msecs
));
1218 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1220 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1222 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1224 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1226 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1228 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1230 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1232 struct blk_mq_hw_ctx
*hctx
;
1235 queue_for_each_hw_ctx(q
, hctx
, i
) {
1236 if (!blk_mq_hctx_has_pending(hctx
) ||
1237 blk_mq_hctx_stopped(hctx
))
1240 blk_mq_run_hw_queue(hctx
, async
);
1243 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1246 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1247 * @q: request queue.
1249 * The caller is responsible for serializing this function against
1250 * blk_mq_{start,stop}_hw_queue().
1252 bool blk_mq_queue_stopped(struct request_queue
*q
)
1254 struct blk_mq_hw_ctx
*hctx
;
1257 queue_for_each_hw_ctx(q
, hctx
, i
)
1258 if (blk_mq_hctx_stopped(hctx
))
1263 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1266 * This function is often used for pausing .queue_rq() by driver when
1267 * there isn't enough resource or some conditions aren't satisfied, and
1268 * BLK_STS_RESOURCE is usually returned.
1270 * We do not guarantee that dispatch can be drained or blocked
1271 * after blk_mq_stop_hw_queue() returns. Please use
1272 * blk_mq_quiesce_queue() for that requirement.
1274 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1276 cancel_delayed_work(&hctx
->run_work
);
1278 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1280 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1283 * This function is often used for pausing .queue_rq() by driver when
1284 * there isn't enough resource or some conditions aren't satisfied, and
1285 * BLK_STS_RESOURCE is usually returned.
1287 * We do not guarantee that dispatch can be drained or blocked
1288 * after blk_mq_stop_hw_queues() returns. Please use
1289 * blk_mq_quiesce_queue() for that requirement.
1291 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1293 struct blk_mq_hw_ctx
*hctx
;
1296 queue_for_each_hw_ctx(q
, hctx
, i
)
1297 blk_mq_stop_hw_queue(hctx
);
1299 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1301 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1303 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1305 blk_mq_run_hw_queue(hctx
, false);
1307 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1309 void blk_mq_start_hw_queues(struct request_queue
*q
)
1311 struct blk_mq_hw_ctx
*hctx
;
1314 queue_for_each_hw_ctx(q
, hctx
, i
)
1315 blk_mq_start_hw_queue(hctx
);
1317 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1319 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1321 if (!blk_mq_hctx_stopped(hctx
))
1324 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1325 blk_mq_run_hw_queue(hctx
, async
);
1327 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1329 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1331 struct blk_mq_hw_ctx
*hctx
;
1334 queue_for_each_hw_ctx(q
, hctx
, i
)
1335 blk_mq_start_stopped_hw_queue(hctx
, async
);
1337 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1339 static void blk_mq_run_work_fn(struct work_struct
*work
)
1341 struct blk_mq_hw_ctx
*hctx
;
1343 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1346 * If we are stopped, don't run the queue. The exception is if
1347 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1348 * the STOPPED bit and run it.
1350 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)) {
1351 if (!test_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
))
1354 clear_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1355 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1358 __blk_mq_run_hw_queue(hctx
);
1362 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1364 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx
)))
1368 * Stop the hw queue, then modify currently delayed work.
1369 * This should prevent us from running the queue prematurely.
1370 * Mark the queue as auto-clearing STOPPED when it runs.
1372 blk_mq_stop_hw_queue(hctx
);
1373 set_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1374 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1376 msecs_to_jiffies(msecs
));
1378 EXPORT_SYMBOL(blk_mq_delay_queue
);
1380 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1384 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1386 lockdep_assert_held(&ctx
->lock
);
1388 trace_block_rq_insert(hctx
->queue
, rq
);
1391 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1393 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1396 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1399 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1401 lockdep_assert_held(&ctx
->lock
);
1403 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1404 blk_mq_hctx_mark_pending(hctx
, ctx
);
1408 * Should only be used carefully, when the caller knows we want to
1409 * bypass a potential IO scheduler on the target device.
1411 void blk_mq_request_bypass_insert(struct request
*rq
)
1413 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1414 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(rq
->q
, ctx
->cpu
);
1416 spin_lock(&hctx
->lock
);
1417 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1418 spin_unlock(&hctx
->lock
);
1420 blk_mq_run_hw_queue(hctx
, false);
1423 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1424 struct list_head
*list
)
1428 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1431 spin_lock(&ctx
->lock
);
1432 while (!list_empty(list
)) {
1435 rq
= list_first_entry(list
, struct request
, queuelist
);
1436 BUG_ON(rq
->mq_ctx
!= ctx
);
1437 list_del_init(&rq
->queuelist
);
1438 __blk_mq_insert_req_list(hctx
, rq
, false);
1440 blk_mq_hctx_mark_pending(hctx
, ctx
);
1441 spin_unlock(&ctx
->lock
);
1444 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1446 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1447 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1449 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1450 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1451 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1454 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1456 struct blk_mq_ctx
*this_ctx
;
1457 struct request_queue
*this_q
;
1460 LIST_HEAD(ctx_list
);
1463 list_splice_init(&plug
->mq_list
, &list
);
1465 list_sort(NULL
, &list
, plug_ctx_cmp
);
1471 while (!list_empty(&list
)) {
1472 rq
= list_entry_rq(list
.next
);
1473 list_del_init(&rq
->queuelist
);
1475 if (rq
->mq_ctx
!= this_ctx
) {
1477 trace_block_unplug(this_q
, depth
, from_schedule
);
1478 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1483 this_ctx
= rq
->mq_ctx
;
1489 list_add_tail(&rq
->queuelist
, &ctx_list
);
1493 * If 'this_ctx' is set, we know we have entries to complete
1494 * on 'ctx_list'. Do those.
1497 trace_block_unplug(this_q
, depth
, from_schedule
);
1498 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1503 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1505 blk_init_request_from_bio(rq
, bio
);
1507 blk_account_io_start(rq
, true);
1510 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx
*hctx
,
1511 struct blk_mq_ctx
*ctx
,
1514 spin_lock(&ctx
->lock
);
1515 __blk_mq_insert_request(hctx
, rq
, false);
1516 spin_unlock(&ctx
->lock
);
1519 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1522 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1524 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1527 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1529 blk_qc_t
*cookie
, bool may_sleep
)
1531 struct request_queue
*q
= rq
->q
;
1532 struct blk_mq_queue_data bd
= {
1536 blk_qc_t new_cookie
;
1538 bool run_queue
= true;
1540 /* RCU or SRCU read lock is needed before checking quiesced flag */
1541 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1549 if (!blk_mq_get_driver_tag(rq
, NULL
, false))
1552 new_cookie
= request_to_qc_t(hctx
, rq
);
1555 * For OK queue, we are done. For error, kill it. Any other
1556 * error (busy), just add it to our list as we previously
1559 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1562 *cookie
= new_cookie
;
1564 case BLK_STS_RESOURCE
:
1565 __blk_mq_requeue_request(rq
);
1568 *cookie
= BLK_QC_T_NONE
;
1569 blk_mq_end_request(rq
, ret
);
1574 blk_mq_sched_insert_request(rq
, false, run_queue
, false, may_sleep
);
1577 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1578 struct request
*rq
, blk_qc_t
*cookie
)
1580 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1582 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false);
1585 unsigned int srcu_idx
;
1589 srcu_idx
= srcu_read_lock(hctx
->queue_rq_srcu
);
1590 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, true);
1591 srcu_read_unlock(hctx
->queue_rq_srcu
, srcu_idx
);
1595 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1597 const int is_sync
= op_is_sync(bio
->bi_opf
);
1598 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1599 struct blk_mq_alloc_data data
= { .flags
= 0 };
1601 unsigned int request_count
= 0;
1602 struct blk_plug
*plug
;
1603 struct request
*same_queue_rq
= NULL
;
1605 unsigned int wb_acct
;
1607 blk_queue_bounce(q
, &bio
);
1609 blk_queue_split(q
, &bio
);
1611 if (!bio_integrity_prep(bio
))
1612 return BLK_QC_T_NONE
;
1614 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1615 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1616 return BLK_QC_T_NONE
;
1618 if (blk_mq_sched_bio_merge(q
, bio
))
1619 return BLK_QC_T_NONE
;
1621 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1623 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1625 rq
= blk_mq_get_request(q
, bio
, bio
->bi_opf
, &data
);
1626 if (unlikely(!rq
)) {
1627 __wbt_done(q
->rq_wb
, wb_acct
);
1628 if (bio
->bi_opf
& REQ_NOWAIT
)
1629 bio_wouldblock_error(bio
);
1630 return BLK_QC_T_NONE
;
1633 wbt_track(&rq
->issue_stat
, wb_acct
);
1635 cookie
= request_to_qc_t(data
.hctx
, rq
);
1637 plug
= current
->plug
;
1638 if (unlikely(is_flush_fua
)) {
1639 blk_mq_put_ctx(data
.ctx
);
1640 blk_mq_bio_to_request(rq
, bio
);
1642 blk_mq_sched_insert_request(rq
, false, true, true,
1645 blk_insert_flush(rq
);
1646 blk_mq_run_hw_queue(data
.hctx
, true);
1648 } else if (plug
&& q
->nr_hw_queues
== 1) {
1649 struct request
*last
= NULL
;
1651 blk_mq_put_ctx(data
.ctx
);
1652 blk_mq_bio_to_request(rq
, bio
);
1655 * @request_count may become stale because of schedule
1656 * out, so check the list again.
1658 if (list_empty(&plug
->mq_list
))
1660 else if (blk_queue_nomerges(q
))
1661 request_count
= blk_plug_queued_count(q
);
1664 trace_block_plug(q
);
1666 last
= list_entry_rq(plug
->mq_list
.prev
);
1668 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1669 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1670 blk_flush_plug_list(plug
, false);
1671 trace_block_plug(q
);
1674 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1675 } else if (plug
&& !blk_queue_nomerges(q
)) {
1676 blk_mq_bio_to_request(rq
, bio
);
1679 * We do limited plugging. If the bio can be merged, do that.
1680 * Otherwise the existing request in the plug list will be
1681 * issued. So the plug list will have one request at most
1682 * The plug list might get flushed before this. If that happens,
1683 * the plug list is empty, and same_queue_rq is invalid.
1685 if (list_empty(&plug
->mq_list
))
1686 same_queue_rq
= NULL
;
1688 list_del_init(&same_queue_rq
->queuelist
);
1689 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1691 blk_mq_put_ctx(data
.ctx
);
1693 if (same_queue_rq
) {
1694 data
.hctx
= blk_mq_map_queue(q
,
1695 same_queue_rq
->mq_ctx
->cpu
);
1696 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1699 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1700 blk_mq_put_ctx(data
.ctx
);
1701 blk_mq_bio_to_request(rq
, bio
);
1702 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1703 } else if (q
->elevator
) {
1704 blk_mq_put_ctx(data
.ctx
);
1705 blk_mq_bio_to_request(rq
, bio
);
1706 blk_mq_sched_insert_request(rq
, false, true, true, true);
1708 blk_mq_put_ctx(data
.ctx
);
1709 blk_mq_bio_to_request(rq
, bio
);
1710 blk_mq_queue_io(data
.hctx
, data
.ctx
, rq
);
1711 blk_mq_run_hw_queue(data
.hctx
, true);
1717 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1718 unsigned int hctx_idx
)
1722 if (tags
->rqs
&& set
->ops
->exit_request
) {
1725 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1726 struct request
*rq
= tags
->static_rqs
[i
];
1730 set
->ops
->exit_request(set
, rq
, hctx_idx
);
1731 tags
->static_rqs
[i
] = NULL
;
1735 while (!list_empty(&tags
->page_list
)) {
1736 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1737 list_del_init(&page
->lru
);
1739 * Remove kmemleak object previously allocated in
1740 * blk_mq_init_rq_map().
1742 kmemleak_free(page_address(page
));
1743 __free_pages(page
, page
->private);
1747 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1751 kfree(tags
->static_rqs
);
1752 tags
->static_rqs
= NULL
;
1754 blk_mq_free_tags(tags
);
1757 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1758 unsigned int hctx_idx
,
1759 unsigned int nr_tags
,
1760 unsigned int reserved_tags
)
1762 struct blk_mq_tags
*tags
;
1765 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1766 if (node
== NUMA_NO_NODE
)
1767 node
= set
->numa_node
;
1769 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1770 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1774 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1775 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1778 blk_mq_free_tags(tags
);
1782 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1783 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1785 if (!tags
->static_rqs
) {
1787 blk_mq_free_tags(tags
);
1794 static size_t order_to_size(unsigned int order
)
1796 return (size_t)PAGE_SIZE
<< order
;
1799 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1800 unsigned int hctx_idx
, unsigned int depth
)
1802 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1803 size_t rq_size
, left
;
1806 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1807 if (node
== NUMA_NO_NODE
)
1808 node
= set
->numa_node
;
1810 INIT_LIST_HEAD(&tags
->page_list
);
1813 * rq_size is the size of the request plus driver payload, rounded
1814 * to the cacheline size
1816 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1818 left
= rq_size
* depth
;
1820 for (i
= 0; i
< depth
; ) {
1821 int this_order
= max_order
;
1826 while (this_order
&& left
< order_to_size(this_order
- 1))
1830 page
= alloc_pages_node(node
,
1831 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1837 if (order_to_size(this_order
) < rq_size
)
1844 page
->private = this_order
;
1845 list_add_tail(&page
->lru
, &tags
->page_list
);
1847 p
= page_address(page
);
1849 * Allow kmemleak to scan these pages as they contain pointers
1850 * to additional allocations like via ops->init_request().
1852 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1853 entries_per_page
= order_to_size(this_order
) / rq_size
;
1854 to_do
= min(entries_per_page
, depth
- i
);
1855 left
-= to_do
* rq_size
;
1856 for (j
= 0; j
< to_do
; j
++) {
1857 struct request
*rq
= p
;
1859 tags
->static_rqs
[i
] = rq
;
1860 if (set
->ops
->init_request
) {
1861 if (set
->ops
->init_request(set
, rq
, hctx_idx
,
1863 tags
->static_rqs
[i
] = NULL
;
1875 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1880 * 'cpu' is going away. splice any existing rq_list entries from this
1881 * software queue to the hw queue dispatch list, and ensure that it
1884 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1886 struct blk_mq_hw_ctx
*hctx
;
1887 struct blk_mq_ctx
*ctx
;
1890 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1891 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1893 spin_lock(&ctx
->lock
);
1894 if (!list_empty(&ctx
->rq_list
)) {
1895 list_splice_init(&ctx
->rq_list
, &tmp
);
1896 blk_mq_hctx_clear_pending(hctx
, ctx
);
1898 spin_unlock(&ctx
->lock
);
1900 if (list_empty(&tmp
))
1903 spin_lock(&hctx
->lock
);
1904 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1905 spin_unlock(&hctx
->lock
);
1907 blk_mq_run_hw_queue(hctx
, true);
1911 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1913 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1917 /* hctx->ctxs will be freed in queue's release handler */
1918 static void blk_mq_exit_hctx(struct request_queue
*q
,
1919 struct blk_mq_tag_set
*set
,
1920 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1922 blk_mq_debugfs_unregister_hctx(hctx
);
1924 blk_mq_tag_idle(hctx
);
1926 if (set
->ops
->exit_request
)
1927 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
1929 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1931 if (set
->ops
->exit_hctx
)
1932 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1934 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1935 cleanup_srcu_struct(hctx
->queue_rq_srcu
);
1937 blk_mq_remove_cpuhp(hctx
);
1938 blk_free_flush_queue(hctx
->fq
);
1939 sbitmap_free(&hctx
->ctx_map
);
1942 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1943 struct blk_mq_tag_set
*set
, int nr_queue
)
1945 struct blk_mq_hw_ctx
*hctx
;
1948 queue_for_each_hw_ctx(q
, hctx
, i
) {
1951 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1955 static int blk_mq_init_hctx(struct request_queue
*q
,
1956 struct blk_mq_tag_set
*set
,
1957 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1961 node
= hctx
->numa_node
;
1962 if (node
== NUMA_NO_NODE
)
1963 node
= hctx
->numa_node
= set
->numa_node
;
1965 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1966 spin_lock_init(&hctx
->lock
);
1967 INIT_LIST_HEAD(&hctx
->dispatch
);
1969 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1971 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1973 hctx
->tags
= set
->tags
[hctx_idx
];
1976 * Allocate space for all possible cpus to avoid allocation at
1979 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1982 goto unregister_cpu_notifier
;
1984 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1990 if (set
->ops
->init_hctx
&&
1991 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1994 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
1997 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1999 goto sched_exit_hctx
;
2001 if (set
->ops
->init_request
&&
2002 set
->ops
->init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
2006 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2007 init_srcu_struct(hctx
->queue_rq_srcu
);
2009 blk_mq_debugfs_register_hctx(q
, hctx
);
2016 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2018 if (set
->ops
->exit_hctx
)
2019 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2021 sbitmap_free(&hctx
->ctx_map
);
2024 unregister_cpu_notifier
:
2025 blk_mq_remove_cpuhp(hctx
);
2029 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2030 unsigned int nr_hw_queues
)
2034 for_each_possible_cpu(i
) {
2035 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2036 struct blk_mq_hw_ctx
*hctx
;
2039 spin_lock_init(&__ctx
->lock
);
2040 INIT_LIST_HEAD(&__ctx
->rq_list
);
2043 /* If the cpu isn't present, the cpu is mapped to first hctx */
2044 if (!cpu_present(i
))
2047 hctx
= blk_mq_map_queue(q
, i
);
2050 * Set local node, IFF we have more than one hw queue. If
2051 * not, we remain on the home node of the device
2053 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2054 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2058 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2062 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2063 set
->queue_depth
, set
->reserved_tags
);
2064 if (!set
->tags
[hctx_idx
])
2067 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2072 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2073 set
->tags
[hctx_idx
] = NULL
;
2077 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2078 unsigned int hctx_idx
)
2080 if (set
->tags
[hctx_idx
]) {
2081 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2082 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2083 set
->tags
[hctx_idx
] = NULL
;
2087 static void blk_mq_map_swqueue(struct request_queue
*q
)
2089 unsigned int i
, hctx_idx
;
2090 struct blk_mq_hw_ctx
*hctx
;
2091 struct blk_mq_ctx
*ctx
;
2092 struct blk_mq_tag_set
*set
= q
->tag_set
;
2095 * Avoid others reading imcomplete hctx->cpumask through sysfs
2097 mutex_lock(&q
->sysfs_lock
);
2099 queue_for_each_hw_ctx(q
, hctx
, i
) {
2100 cpumask_clear(hctx
->cpumask
);
2105 * Map software to hardware queues.
2107 * If the cpu isn't present, the cpu is mapped to first hctx.
2109 for_each_present_cpu(i
) {
2110 hctx_idx
= q
->mq_map
[i
];
2111 /* unmapped hw queue can be remapped after CPU topo changed */
2112 if (!set
->tags
[hctx_idx
] &&
2113 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2115 * If tags initialization fail for some hctx,
2116 * that hctx won't be brought online. In this
2117 * case, remap the current ctx to hctx[0] which
2118 * is guaranteed to always have tags allocated
2123 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2124 hctx
= blk_mq_map_queue(q
, i
);
2126 cpumask_set_cpu(i
, hctx
->cpumask
);
2127 ctx
->index_hw
= hctx
->nr_ctx
;
2128 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2131 mutex_unlock(&q
->sysfs_lock
);
2133 queue_for_each_hw_ctx(q
, hctx
, i
) {
2135 * If no software queues are mapped to this hardware queue,
2136 * disable it and free the request entries.
2138 if (!hctx
->nr_ctx
) {
2139 /* Never unmap queue 0. We need it as a
2140 * fallback in case of a new remap fails
2143 if (i
&& set
->tags
[i
])
2144 blk_mq_free_map_and_requests(set
, i
);
2150 hctx
->tags
= set
->tags
[i
];
2151 WARN_ON(!hctx
->tags
);
2154 * Set the map size to the number of mapped software queues.
2155 * This is more accurate and more efficient than looping
2156 * over all possibly mapped software queues.
2158 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2161 * Initialize batch roundrobin counts
2163 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2164 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2169 * Caller needs to ensure that we're either frozen/quiesced, or that
2170 * the queue isn't live yet.
2172 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2174 struct blk_mq_hw_ctx
*hctx
;
2177 queue_for_each_hw_ctx(q
, hctx
, i
) {
2179 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2180 atomic_inc(&q
->shared_hctx_restart
);
2181 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2183 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2184 atomic_dec(&q
->shared_hctx_restart
);
2185 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2190 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2193 struct request_queue
*q
;
2195 lockdep_assert_held(&set
->tag_list_lock
);
2197 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2198 blk_mq_freeze_queue(q
);
2199 queue_set_hctx_shared(q
, shared
);
2200 blk_mq_unfreeze_queue(q
);
2204 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2206 struct blk_mq_tag_set
*set
= q
->tag_set
;
2208 mutex_lock(&set
->tag_list_lock
);
2209 list_del_rcu(&q
->tag_set_list
);
2210 INIT_LIST_HEAD(&q
->tag_set_list
);
2211 if (list_is_singular(&set
->tag_list
)) {
2212 /* just transitioned to unshared */
2213 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2214 /* update existing queue */
2215 blk_mq_update_tag_set_depth(set
, false);
2217 mutex_unlock(&set
->tag_list_lock
);
2222 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2223 struct request_queue
*q
)
2227 mutex_lock(&set
->tag_list_lock
);
2229 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2230 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2231 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2232 /* update existing queue */
2233 blk_mq_update_tag_set_depth(set
, true);
2235 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2236 queue_set_hctx_shared(q
, true);
2237 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2239 mutex_unlock(&set
->tag_list_lock
);
2243 * It is the actual release handler for mq, but we do it from
2244 * request queue's release handler for avoiding use-after-free
2245 * and headache because q->mq_kobj shouldn't have been introduced,
2246 * but we can't group ctx/kctx kobj without it.
2248 void blk_mq_release(struct request_queue
*q
)
2250 struct blk_mq_hw_ctx
*hctx
;
2253 /* hctx kobj stays in hctx */
2254 queue_for_each_hw_ctx(q
, hctx
, i
) {
2257 kobject_put(&hctx
->kobj
);
2262 kfree(q
->queue_hw_ctx
);
2265 * release .mq_kobj and sw queue's kobject now because
2266 * both share lifetime with request queue.
2268 blk_mq_sysfs_deinit(q
);
2270 free_percpu(q
->queue_ctx
);
2273 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2275 struct request_queue
*uninit_q
, *q
;
2277 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2279 return ERR_PTR(-ENOMEM
);
2281 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2283 blk_cleanup_queue(uninit_q
);
2287 EXPORT_SYMBOL(blk_mq_init_queue
);
2289 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2291 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2293 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, queue_rq_srcu
),
2294 __alignof__(struct blk_mq_hw_ctx
)) !=
2295 sizeof(struct blk_mq_hw_ctx
));
2297 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2298 hw_ctx_size
+= sizeof(struct srcu_struct
);
2303 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2304 struct request_queue
*q
)
2307 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2309 blk_mq_sysfs_unregister(q
);
2310 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2316 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2317 hctxs
[i
] = kzalloc_node(blk_mq_hw_ctx_size(set
),
2322 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2329 atomic_set(&hctxs
[i
]->nr_active
, 0);
2330 hctxs
[i
]->numa_node
= node
;
2331 hctxs
[i
]->queue_num
= i
;
2333 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2334 free_cpumask_var(hctxs
[i
]->cpumask
);
2339 blk_mq_hctx_kobj_init(hctxs
[i
]);
2341 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2342 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2346 blk_mq_free_map_and_requests(set
, j
);
2347 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2348 kobject_put(&hctx
->kobj
);
2353 q
->nr_hw_queues
= i
;
2354 blk_mq_sysfs_register(q
);
2357 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2358 struct request_queue
*q
)
2360 /* mark the queue as mq asap */
2361 q
->mq_ops
= set
->ops
;
2363 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2364 blk_mq_poll_stats_bkt
,
2365 BLK_MQ_POLL_STATS_BKTS
, q
);
2369 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2373 /* init q->mq_kobj and sw queues' kobjects */
2374 blk_mq_sysfs_init(q
);
2376 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2377 GFP_KERNEL
, set
->numa_node
);
2378 if (!q
->queue_hw_ctx
)
2381 q
->mq_map
= set
->mq_map
;
2383 blk_mq_realloc_hw_ctxs(set
, q
);
2384 if (!q
->nr_hw_queues
)
2387 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2388 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2390 q
->nr_queues
= nr_cpu_ids
;
2392 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2394 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2395 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2397 q
->sg_reserved_size
= INT_MAX
;
2399 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2400 INIT_LIST_HEAD(&q
->requeue_list
);
2401 spin_lock_init(&q
->requeue_lock
);
2403 blk_queue_make_request(q
, blk_mq_make_request
);
2406 * Do this after blk_queue_make_request() overrides it...
2408 q
->nr_requests
= set
->queue_depth
;
2411 * Default to classic polling
2415 if (set
->ops
->complete
)
2416 blk_queue_softirq_done(q
, set
->ops
->complete
);
2418 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2419 blk_mq_add_queue_tag_set(set
, q
);
2420 blk_mq_map_swqueue(q
);
2422 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2425 ret
= blk_mq_sched_init(q
);
2427 return ERR_PTR(ret
);
2433 kfree(q
->queue_hw_ctx
);
2435 free_percpu(q
->queue_ctx
);
2438 return ERR_PTR(-ENOMEM
);
2440 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2442 void blk_mq_free_queue(struct request_queue
*q
)
2444 struct blk_mq_tag_set
*set
= q
->tag_set
;
2446 blk_mq_del_queue_tag_set(q
);
2447 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2450 /* Basically redo blk_mq_init_queue with queue frozen */
2451 static void blk_mq_queue_reinit(struct request_queue
*q
)
2453 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2455 blk_mq_debugfs_unregister_hctxs(q
);
2456 blk_mq_sysfs_unregister(q
);
2459 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2460 * we should change hctx numa_node according to new topology (this
2461 * involves free and re-allocate memory, worthy doing?)
2464 blk_mq_map_swqueue(q
);
2466 blk_mq_sysfs_register(q
);
2467 blk_mq_debugfs_register_hctxs(q
);
2470 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2474 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2475 if (!__blk_mq_alloc_rq_map(set
, i
))
2482 blk_mq_free_rq_map(set
->tags
[i
]);
2488 * Allocate the request maps associated with this tag_set. Note that this
2489 * may reduce the depth asked for, if memory is tight. set->queue_depth
2490 * will be updated to reflect the allocated depth.
2492 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2497 depth
= set
->queue_depth
;
2499 err
= __blk_mq_alloc_rq_maps(set
);
2503 set
->queue_depth
>>= 1;
2504 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2508 } while (set
->queue_depth
);
2510 if (!set
->queue_depth
|| err
) {
2511 pr_err("blk-mq: failed to allocate request map\n");
2515 if (depth
!= set
->queue_depth
)
2516 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2517 depth
, set
->queue_depth
);
2522 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2524 if (set
->ops
->map_queues
)
2525 return set
->ops
->map_queues(set
);
2527 return blk_mq_map_queues(set
);
2531 * Alloc a tag set to be associated with one or more request queues.
2532 * May fail with EINVAL for various error conditions. May adjust the
2533 * requested depth down, if if it too large. In that case, the set
2534 * value will be stored in set->queue_depth.
2536 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2540 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2542 if (!set
->nr_hw_queues
)
2544 if (!set
->queue_depth
)
2546 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2549 if (!set
->ops
->queue_rq
)
2552 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2553 pr_info("blk-mq: reduced tag depth to %u\n",
2555 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2559 * If a crashdump is active, then we are potentially in a very
2560 * memory constrained environment. Limit us to 1 queue and
2561 * 64 tags to prevent using too much memory.
2563 if (is_kdump_kernel()) {
2564 set
->nr_hw_queues
= 1;
2565 set
->queue_depth
= min(64U, set
->queue_depth
);
2568 * There is no use for more h/w queues than cpus.
2570 if (set
->nr_hw_queues
> nr_cpu_ids
)
2571 set
->nr_hw_queues
= nr_cpu_ids
;
2573 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2574 GFP_KERNEL
, set
->numa_node
);
2579 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2580 GFP_KERNEL
, set
->numa_node
);
2584 ret
= blk_mq_update_queue_map(set
);
2586 goto out_free_mq_map
;
2588 ret
= blk_mq_alloc_rq_maps(set
);
2590 goto out_free_mq_map
;
2592 mutex_init(&set
->tag_list_lock
);
2593 INIT_LIST_HEAD(&set
->tag_list
);
2605 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2607 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2611 for (i
= 0; i
< nr_cpu_ids
; i
++)
2612 blk_mq_free_map_and_requests(set
, i
);
2620 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2622 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2624 struct blk_mq_tag_set
*set
= q
->tag_set
;
2625 struct blk_mq_hw_ctx
*hctx
;
2631 blk_mq_freeze_queue(q
);
2634 queue_for_each_hw_ctx(q
, hctx
, i
) {
2638 * If we're using an MQ scheduler, just update the scheduler
2639 * queue depth. This is similar to what the old code would do.
2641 if (!hctx
->sched_tags
) {
2642 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2643 min(nr
, set
->queue_depth
),
2646 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2654 q
->nr_requests
= nr
;
2656 blk_mq_unfreeze_queue(q
);
2661 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
2664 struct request_queue
*q
;
2666 lockdep_assert_held(&set
->tag_list_lock
);
2668 if (nr_hw_queues
> nr_cpu_ids
)
2669 nr_hw_queues
= nr_cpu_ids
;
2670 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2673 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2674 blk_mq_freeze_queue(q
);
2676 set
->nr_hw_queues
= nr_hw_queues
;
2677 blk_mq_update_queue_map(set
);
2678 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2679 blk_mq_realloc_hw_ctxs(set
, q
);
2680 blk_mq_queue_reinit(q
);
2683 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2684 blk_mq_unfreeze_queue(q
);
2687 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2689 mutex_lock(&set
->tag_list_lock
);
2690 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
2691 mutex_unlock(&set
->tag_list_lock
);
2693 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2695 /* Enable polling stats and return whether they were already enabled. */
2696 static bool blk_poll_stats_enable(struct request_queue
*q
)
2698 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2699 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
2701 blk_stat_add_callback(q
, q
->poll_cb
);
2705 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2708 * We don't arm the callback if polling stats are not enabled or the
2709 * callback is already active.
2711 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2712 blk_stat_is_active(q
->poll_cb
))
2715 blk_stat_activate_msecs(q
->poll_cb
, 100);
2718 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2720 struct request_queue
*q
= cb
->data
;
2723 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
2724 if (cb
->stat
[bucket
].nr_samples
)
2725 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
2729 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2730 struct blk_mq_hw_ctx
*hctx
,
2733 unsigned long ret
= 0;
2737 * If stats collection isn't on, don't sleep but turn it on for
2740 if (!blk_poll_stats_enable(q
))
2744 * As an optimistic guess, use half of the mean service time
2745 * for this type of request. We can (and should) make this smarter.
2746 * For instance, if the completion latencies are tight, we can
2747 * get closer than just half the mean. This is especially
2748 * important on devices where the completion latencies are longer
2749 * than ~10 usec. We do use the stats for the relevant IO size
2750 * if available which does lead to better estimates.
2752 bucket
= blk_mq_poll_stats_bkt(rq
);
2756 if (q
->poll_stat
[bucket
].nr_samples
)
2757 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
2762 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2763 struct blk_mq_hw_ctx
*hctx
,
2766 struct hrtimer_sleeper hs
;
2767 enum hrtimer_mode mode
;
2771 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2777 * -1: don't ever hybrid sleep
2778 * 0: use half of prev avg
2779 * >0: use this specific value
2781 if (q
->poll_nsec
== -1)
2783 else if (q
->poll_nsec
> 0)
2784 nsecs
= q
->poll_nsec
;
2786 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2791 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2794 * This will be replaced with the stats tracking code, using
2795 * 'avg_completion_time / 2' as the pre-sleep target.
2799 mode
= HRTIMER_MODE_REL
;
2800 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2801 hrtimer_set_expires(&hs
.timer
, kt
);
2803 hrtimer_init_sleeper(&hs
, current
);
2805 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2807 set_current_state(TASK_UNINTERRUPTIBLE
);
2808 hrtimer_start_expires(&hs
.timer
, mode
);
2811 hrtimer_cancel(&hs
.timer
);
2812 mode
= HRTIMER_MODE_ABS
;
2813 } while (hs
.task
&& !signal_pending(current
));
2815 __set_current_state(TASK_RUNNING
);
2816 destroy_hrtimer_on_stack(&hs
.timer
);
2820 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2822 struct request_queue
*q
= hctx
->queue
;
2826 * If we sleep, have the caller restart the poll loop to reset
2827 * the state. Like for the other success return cases, the
2828 * caller is responsible for checking if the IO completed. If
2829 * the IO isn't complete, we'll get called again and will go
2830 * straight to the busy poll loop.
2832 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2835 hctx
->poll_considered
++;
2837 state
= current
->state
;
2838 while (!need_resched()) {
2841 hctx
->poll_invoked
++;
2843 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2845 hctx
->poll_success
++;
2846 set_current_state(TASK_RUNNING
);
2850 if (signal_pending_state(state
, current
))
2851 set_current_state(TASK_RUNNING
);
2853 if (current
->state
== TASK_RUNNING
)
2863 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2865 struct blk_mq_hw_ctx
*hctx
;
2866 struct blk_plug
*plug
;
2869 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2870 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2873 plug
= current
->plug
;
2875 blk_flush_plug_list(plug
, false);
2877 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2878 if (!blk_qc_t_is_internal(cookie
))
2879 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2881 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2883 * With scheduling, if the request has completed, we'll
2884 * get a NULL return here, as we clear the sched tag when
2885 * that happens. The request still remains valid, like always,
2886 * so we should be safe with just the NULL check.
2892 return __blk_mq_poll(hctx
, rq
);
2894 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2896 static int __init
blk_mq_init(void)
2898 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2899 blk_mq_hctx_notify_dead
);
2902 subsys_initcall(blk_mq_init
);