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/delay.h>
24 #include <linux/crash_dump.h>
25 #include <linux/prefetch.h>
27 #include <trace/events/block.h>
29 #include <linux/blk-mq.h>
32 #include "blk-mq-tag.h"
36 static DEFINE_MUTEX(all_q_mutex
);
37 static LIST_HEAD(all_q_list
);
40 * Check if any of the ctx's have pending work in this hardware queue
42 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
44 return sbitmap_any_bit_set(&hctx
->ctx_map
);
48 * Mark this ctx as having pending work in this hardware queue
50 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
51 struct blk_mq_ctx
*ctx
)
53 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
54 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
57 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
58 struct blk_mq_ctx
*ctx
)
60 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
63 void blk_mq_freeze_queue_start(struct request_queue
*q
)
67 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
68 if (freeze_depth
== 1) {
69 percpu_ref_kill(&q
->q_usage_counter
);
70 blk_mq_run_hw_queues(q
, false);
73 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start
);
75 static void blk_mq_freeze_queue_wait(struct request_queue
*q
)
77 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
81 * Guarantee no request is in use, so we can change any data structure of
82 * the queue afterward.
84 void blk_freeze_queue(struct request_queue
*q
)
87 * In the !blk_mq case we are only calling this to kill the
88 * q_usage_counter, otherwise this increases the freeze depth
89 * and waits for it to return to zero. For this reason there is
90 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
91 * exported to drivers as the only user for unfreeze is blk_mq.
93 blk_mq_freeze_queue_start(q
);
94 blk_mq_freeze_queue_wait(q
);
97 void blk_mq_freeze_queue(struct request_queue
*q
)
100 * ...just an alias to keep freeze and unfreeze actions balanced
101 * in the blk_mq_* namespace
105 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
107 void blk_mq_unfreeze_queue(struct request_queue
*q
)
111 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
112 WARN_ON_ONCE(freeze_depth
< 0);
114 percpu_ref_reinit(&q
->q_usage_counter
);
115 wake_up_all(&q
->mq_freeze_wq
);
118 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
121 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
124 * Note: this function does not prevent that the struct request end_io()
125 * callback function is invoked. Additionally, it is not prevented that
126 * new queue_rq() calls occur unless the queue has been stopped first.
128 void blk_mq_quiesce_queue(struct request_queue
*q
)
130 struct blk_mq_hw_ctx
*hctx
;
134 blk_mq_stop_hw_queues(q
);
136 queue_for_each_hw_ctx(q
, hctx
, i
) {
137 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
138 synchronize_srcu(&hctx
->queue_rq_srcu
);
145 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
147 void blk_mq_wake_waiters(struct request_queue
*q
)
149 struct blk_mq_hw_ctx
*hctx
;
152 queue_for_each_hw_ctx(q
, hctx
, i
)
153 if (blk_mq_hw_queue_mapped(hctx
))
154 blk_mq_tag_wakeup_all(hctx
->tags
, true);
157 * If we are called because the queue has now been marked as
158 * dying, we need to ensure that processes currently waiting on
159 * the queue are notified as well.
161 wake_up_all(&q
->mq_freeze_wq
);
164 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
166 return blk_mq_has_free_tags(hctx
->tags
);
168 EXPORT_SYMBOL(blk_mq_can_queue
);
170 static void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
171 struct request
*rq
, unsigned int op
)
173 INIT_LIST_HEAD(&rq
->queuelist
);
174 /* csd/requeue_work/fifo_time is initialized before use */
178 if (blk_queue_io_stat(q
))
179 rq
->rq_flags
|= RQF_IO_STAT
;
180 /* do not touch atomic flags, it needs atomic ops against the timer */
182 INIT_HLIST_NODE(&rq
->hash
);
183 RB_CLEAR_NODE(&rq
->rb_node
);
186 rq
->start_time
= jiffies
;
187 #ifdef CONFIG_BLK_CGROUP
189 set_start_time_ns(rq
);
190 rq
->io_start_time_ns
= 0;
192 rq
->nr_phys_segments
= 0;
193 #if defined(CONFIG_BLK_DEV_INTEGRITY)
194 rq
->nr_integrity_segments
= 0;
197 /* tag was already set */
207 INIT_LIST_HEAD(&rq
->timeout_list
);
211 rq
->end_io_data
= NULL
;
214 ctx
->rq_dispatched
[op_is_sync(op
)]++;
217 static struct request
*
218 __blk_mq_alloc_request(struct blk_mq_alloc_data
*data
, unsigned int op
)
223 tag
= blk_mq_get_tag(data
);
224 if (tag
!= BLK_MQ_TAG_FAIL
) {
225 rq
= data
->hctx
->tags
->rqs
[tag
];
227 if (blk_mq_tag_busy(data
->hctx
)) {
228 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
229 atomic_inc(&data
->hctx
->nr_active
);
233 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
);
240 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
243 struct blk_mq_ctx
*ctx
;
244 struct blk_mq_hw_ctx
*hctx
;
246 struct blk_mq_alloc_data alloc_data
;
249 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
253 ctx
= blk_mq_get_ctx(q
);
254 hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
255 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
256 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
261 return ERR_PTR(-EWOULDBLOCK
);
265 rq
->__sector
= (sector_t
) -1;
266 rq
->bio
= rq
->biotail
= NULL
;
269 EXPORT_SYMBOL(blk_mq_alloc_request
);
271 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
272 unsigned int flags
, unsigned int hctx_idx
)
274 struct blk_mq_hw_ctx
*hctx
;
275 struct blk_mq_ctx
*ctx
;
277 struct blk_mq_alloc_data alloc_data
;
281 * If the tag allocator sleeps we could get an allocation for a
282 * different hardware context. No need to complicate the low level
283 * allocator for this for the rare use case of a command tied to
286 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
287 return ERR_PTR(-EINVAL
);
289 if (hctx_idx
>= q
->nr_hw_queues
)
290 return ERR_PTR(-EIO
);
292 ret
= blk_queue_enter(q
, true);
297 * Check if the hardware context is actually mapped to anything.
298 * If not tell the caller that it should skip this queue.
300 hctx
= q
->queue_hw_ctx
[hctx_idx
];
301 if (!blk_mq_hw_queue_mapped(hctx
)) {
305 ctx
= __blk_mq_get_ctx(q
, cpumask_first(hctx
->cpumask
));
307 blk_mq_set_alloc_data(&alloc_data
, q
, flags
, ctx
, hctx
);
308 rq
= __blk_mq_alloc_request(&alloc_data
, rw
);
320 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
322 static void __blk_mq_free_request(struct blk_mq_hw_ctx
*hctx
,
323 struct blk_mq_ctx
*ctx
, struct request
*rq
)
325 const int tag
= rq
->tag
;
326 struct request_queue
*q
= rq
->q
;
328 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
329 atomic_dec(&hctx
->nr_active
);
331 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
334 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
335 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
336 blk_mq_put_tag(hctx
, ctx
, tag
);
340 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
342 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
344 ctx
->rq_completed
[rq_is_sync(rq
)]++;
345 __blk_mq_free_request(hctx
, ctx
, rq
);
348 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request
);
350 void blk_mq_free_request(struct request
*rq
)
352 blk_mq_free_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
354 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
356 inline void __blk_mq_end_request(struct request
*rq
, int error
)
358 blk_account_io_done(rq
);
361 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
362 rq
->end_io(rq
, error
);
364 if (unlikely(blk_bidi_rq(rq
)))
365 blk_mq_free_request(rq
->next_rq
);
366 blk_mq_free_request(rq
);
369 EXPORT_SYMBOL(__blk_mq_end_request
);
371 void blk_mq_end_request(struct request
*rq
, int error
)
373 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
375 __blk_mq_end_request(rq
, error
);
377 EXPORT_SYMBOL(blk_mq_end_request
);
379 static void __blk_mq_complete_request_remote(void *data
)
381 struct request
*rq
= data
;
383 rq
->q
->softirq_done_fn(rq
);
386 static void blk_mq_ipi_complete_request(struct request
*rq
)
388 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
392 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
393 rq
->q
->softirq_done_fn(rq
);
398 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
399 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
401 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
402 rq
->csd
.func
= __blk_mq_complete_request_remote
;
405 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
407 rq
->q
->softirq_done_fn(rq
);
412 static void blk_mq_stat_add(struct request
*rq
)
414 if (rq
->rq_flags
& RQF_STATS
) {
416 * We could rq->mq_ctx here, but there's less of a risk
417 * of races if we have the completion event add the stats
418 * to the local software queue.
420 struct blk_mq_ctx
*ctx
;
422 ctx
= __blk_mq_get_ctx(rq
->q
, raw_smp_processor_id());
423 blk_stat_add(&ctx
->stat
[rq_data_dir(rq
)], rq
);
427 static void __blk_mq_complete_request(struct request
*rq
)
429 struct request_queue
*q
= rq
->q
;
433 if (!q
->softirq_done_fn
)
434 blk_mq_end_request(rq
, rq
->errors
);
436 blk_mq_ipi_complete_request(rq
);
440 * blk_mq_complete_request - end I/O on a request
441 * @rq: the request being processed
444 * Ends all I/O on a request. It does not handle partial completions.
445 * The actual completion happens out-of-order, through a IPI handler.
447 void blk_mq_complete_request(struct request
*rq
, int error
)
449 struct request_queue
*q
= rq
->q
;
451 if (unlikely(blk_should_fake_timeout(q
)))
453 if (!blk_mark_rq_complete(rq
)) {
455 __blk_mq_complete_request(rq
);
458 EXPORT_SYMBOL(blk_mq_complete_request
);
460 int blk_mq_request_started(struct request
*rq
)
462 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
464 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
466 void blk_mq_start_request(struct request
*rq
)
468 struct request_queue
*q
= rq
->q
;
470 trace_block_rq_issue(q
, rq
);
472 rq
->resid_len
= blk_rq_bytes(rq
);
473 if (unlikely(blk_bidi_rq(rq
)))
474 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
476 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
477 blk_stat_set_issue_time(&rq
->issue_stat
);
478 rq
->rq_flags
|= RQF_STATS
;
479 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
485 * Ensure that ->deadline is visible before set the started
486 * flag and clear the completed flag.
488 smp_mb__before_atomic();
491 * Mark us as started and clear complete. Complete might have been
492 * set if requeue raced with timeout, which then marked it as
493 * complete. So be sure to clear complete again when we start
494 * the request, otherwise we'll ignore the completion event.
496 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
497 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
498 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
499 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
501 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
503 * Make sure space for the drain appears. We know we can do
504 * this because max_hw_segments has been adjusted to be one
505 * fewer than the device can handle.
507 rq
->nr_phys_segments
++;
510 EXPORT_SYMBOL(blk_mq_start_request
);
512 static void __blk_mq_requeue_request(struct request
*rq
)
514 struct request_queue
*q
= rq
->q
;
516 trace_block_rq_requeue(q
, rq
);
517 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
519 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
520 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
521 rq
->nr_phys_segments
--;
525 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
527 __blk_mq_requeue_request(rq
);
529 BUG_ON(blk_queued_rq(rq
));
530 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
532 EXPORT_SYMBOL(blk_mq_requeue_request
);
534 static void blk_mq_requeue_work(struct work_struct
*work
)
536 struct request_queue
*q
=
537 container_of(work
, struct request_queue
, requeue_work
.work
);
539 struct request
*rq
, *next
;
542 spin_lock_irqsave(&q
->requeue_lock
, flags
);
543 list_splice_init(&q
->requeue_list
, &rq_list
);
544 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
546 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
547 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
550 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
551 list_del_init(&rq
->queuelist
);
552 blk_mq_insert_request(rq
, true, false, false);
555 while (!list_empty(&rq_list
)) {
556 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
557 list_del_init(&rq
->queuelist
);
558 blk_mq_insert_request(rq
, false, false, false);
561 blk_mq_run_hw_queues(q
, false);
564 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
565 bool kick_requeue_list
)
567 struct request_queue
*q
= rq
->q
;
571 * We abuse this flag that is otherwise used by the I/O scheduler to
572 * request head insertation from the workqueue.
574 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
576 spin_lock_irqsave(&q
->requeue_lock
, flags
);
578 rq
->rq_flags
|= RQF_SOFTBARRIER
;
579 list_add(&rq
->queuelist
, &q
->requeue_list
);
581 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
583 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
585 if (kick_requeue_list
)
586 blk_mq_kick_requeue_list(q
);
588 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
590 void blk_mq_kick_requeue_list(struct request_queue
*q
)
592 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
594 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
596 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
599 kblockd_schedule_delayed_work(&q
->requeue_work
,
600 msecs_to_jiffies(msecs
));
602 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
604 void blk_mq_abort_requeue_list(struct request_queue
*q
)
609 spin_lock_irqsave(&q
->requeue_lock
, flags
);
610 list_splice_init(&q
->requeue_list
, &rq_list
);
611 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
613 while (!list_empty(&rq_list
)) {
616 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
617 list_del_init(&rq
->queuelist
);
619 blk_mq_end_request(rq
, rq
->errors
);
622 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
624 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
626 if (tag
< tags
->nr_tags
) {
627 prefetch(tags
->rqs
[tag
]);
628 return tags
->rqs
[tag
];
633 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
635 struct blk_mq_timeout_data
{
637 unsigned int next_set
;
640 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
642 struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
643 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
646 * We know that complete is set at this point. If STARTED isn't set
647 * anymore, then the request isn't active and the "timeout" should
648 * just be ignored. This can happen due to the bitflag ordering.
649 * Timeout first checks if STARTED is set, and if it is, assumes
650 * the request is active. But if we race with completion, then
651 * we both flags will get cleared. So check here again, and ignore
652 * a timeout event with a request that isn't active.
654 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
658 ret
= ops
->timeout(req
, reserved
);
662 __blk_mq_complete_request(req
);
664 case BLK_EH_RESET_TIMER
:
666 blk_clear_rq_complete(req
);
668 case BLK_EH_NOT_HANDLED
:
671 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
676 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
677 struct request
*rq
, void *priv
, bool reserved
)
679 struct blk_mq_timeout_data
*data
= priv
;
681 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
684 if (time_after_eq(jiffies
, rq
->deadline
)) {
685 if (!blk_mark_rq_complete(rq
))
686 blk_mq_rq_timed_out(rq
, reserved
);
687 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
688 data
->next
= rq
->deadline
;
693 static void blk_mq_timeout_work(struct work_struct
*work
)
695 struct request_queue
*q
=
696 container_of(work
, struct request_queue
, timeout_work
);
697 struct blk_mq_timeout_data data
= {
703 /* A deadlock might occur if a request is stuck requiring a
704 * timeout at the same time a queue freeze is waiting
705 * completion, since the timeout code would not be able to
706 * acquire the queue reference here.
708 * That's why we don't use blk_queue_enter here; instead, we use
709 * percpu_ref_tryget directly, because we need to be able to
710 * obtain a reference even in the short window between the queue
711 * starting to freeze, by dropping the first reference in
712 * blk_mq_freeze_queue_start, and the moment the last request is
713 * consumed, marked by the instant q_usage_counter reaches
716 if (!percpu_ref_tryget(&q
->q_usage_counter
))
719 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
722 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
723 mod_timer(&q
->timeout
, data
.next
);
725 struct blk_mq_hw_ctx
*hctx
;
727 queue_for_each_hw_ctx(q
, hctx
, i
) {
728 /* the hctx may be unmapped, so check it here */
729 if (blk_mq_hw_queue_mapped(hctx
))
730 blk_mq_tag_idle(hctx
);
737 * Reverse check our software queue for entries that we could potentially
738 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
739 * too much time checking for merges.
741 static bool blk_mq_attempt_merge(struct request_queue
*q
,
742 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
747 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
753 if (!blk_rq_merge_ok(rq
, bio
))
756 el_ret
= blk_try_merge(rq
, bio
);
757 if (el_ret
== ELEVATOR_BACK_MERGE
) {
758 if (bio_attempt_back_merge(q
, rq
, bio
)) {
763 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
764 if (bio_attempt_front_merge(q
, rq
, bio
)) {
775 struct flush_busy_ctx_data
{
776 struct blk_mq_hw_ctx
*hctx
;
777 struct list_head
*list
;
780 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
782 struct flush_busy_ctx_data
*flush_data
= data
;
783 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
784 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
786 sbitmap_clear_bit(sb
, bitnr
);
787 spin_lock(&ctx
->lock
);
788 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
789 spin_unlock(&ctx
->lock
);
794 * Process software queues that have been marked busy, splicing them
795 * to the for-dispatch
797 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
799 struct flush_busy_ctx_data data
= {
804 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
807 static inline unsigned int queued_to_index(unsigned int queued
)
812 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
815 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
817 struct request_queue
*q
= hctx
->queue
;
819 LIST_HEAD(driver_list
);
820 struct list_head
*dptr
;
821 int queued
, ret
= BLK_MQ_RQ_QUEUE_OK
;
824 * Start off with dptr being NULL, so we start the first request
825 * immediately, even if we have more pending.
830 * Now process all the entries, sending them to the driver.
833 while (!list_empty(list
)) {
834 struct blk_mq_queue_data bd
;
836 rq
= list_first_entry(list
, struct request
, queuelist
);
837 list_del_init(&rq
->queuelist
);
841 bd
.last
= list_empty(list
);
843 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
845 case BLK_MQ_RQ_QUEUE_OK
:
848 case BLK_MQ_RQ_QUEUE_BUSY
:
849 list_add(&rq
->queuelist
, list
);
850 __blk_mq_requeue_request(rq
);
853 pr_err("blk-mq: bad return on queue: %d\n", ret
);
854 case BLK_MQ_RQ_QUEUE_ERROR
:
856 blk_mq_end_request(rq
, rq
->errors
);
860 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
864 * We've done the first request. If we have more than 1
865 * left in the list, set dptr to defer issue.
867 if (!dptr
&& list
->next
!= list
->prev
)
871 hctx
->dispatched
[queued_to_index(queued
)]++;
874 * Any items that need requeuing? Stuff them into hctx->dispatch,
875 * that is where we will continue on next queue run.
877 if (!list_empty(list
)) {
878 spin_lock(&hctx
->lock
);
879 list_splice(list
, &hctx
->dispatch
);
880 spin_unlock(&hctx
->lock
);
883 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
884 * it's possible the queue is stopped and restarted again
885 * before this. Queue restart will dispatch requests. And since
886 * requests in rq_list aren't added into hctx->dispatch yet,
887 * the requests in rq_list might get lost.
889 * blk_mq_run_hw_queue() already checks the STOPPED bit
891 blk_mq_run_hw_queue(hctx
, true);
894 return ret
!= BLK_MQ_RQ_QUEUE_BUSY
;
898 * Run this hardware queue, pulling any software queues mapped to it in.
899 * Note that this function currently has various problems around ordering
900 * of IO. In particular, we'd like FIFO behaviour on handling existing
901 * items on the hctx->dispatch list. Ignore that for now.
903 static void blk_mq_process_rq_list(struct blk_mq_hw_ctx
*hctx
)
907 if (unlikely(blk_mq_hctx_stopped(hctx
)))
913 * Touch any software queue that has pending entries.
915 flush_busy_ctxs(hctx
, &rq_list
);
918 * If we have previous entries on our dispatch list, grab them
919 * and stuff them at the front for more fair dispatch.
921 if (!list_empty_careful(&hctx
->dispatch
)) {
922 spin_lock(&hctx
->lock
);
923 if (!list_empty(&hctx
->dispatch
))
924 list_splice_init(&hctx
->dispatch
, &rq_list
);
925 spin_unlock(&hctx
->lock
);
928 blk_mq_dispatch_rq_list(hctx
, &rq_list
);
931 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
935 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
936 cpu_online(hctx
->next_cpu
));
938 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
940 blk_mq_process_rq_list(hctx
);
943 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
944 blk_mq_process_rq_list(hctx
);
945 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
950 * It'd be great if the workqueue API had a way to pass
951 * in a mask and had some smarts for more clever placement.
952 * For now we just round-robin here, switching for every
953 * BLK_MQ_CPU_WORK_BATCH queued items.
955 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
957 if (hctx
->queue
->nr_hw_queues
== 1)
958 return WORK_CPU_UNBOUND
;
960 if (--hctx
->next_cpu_batch
<= 0) {
963 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
964 if (next_cpu
>= nr_cpu_ids
)
965 next_cpu
= cpumask_first(hctx
->cpumask
);
967 hctx
->next_cpu
= next_cpu
;
968 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
971 return hctx
->next_cpu
;
974 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
976 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
977 !blk_mq_hw_queue_mapped(hctx
)))
980 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
982 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
983 __blk_mq_run_hw_queue(hctx
);
991 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
);
994 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
996 struct blk_mq_hw_ctx
*hctx
;
999 queue_for_each_hw_ctx(q
, hctx
, i
) {
1000 if ((!blk_mq_hctx_has_pending(hctx
) &&
1001 list_empty_careful(&hctx
->dispatch
)) ||
1002 blk_mq_hctx_stopped(hctx
))
1005 blk_mq_run_hw_queue(hctx
, async
);
1008 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1011 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1012 * @q: request queue.
1014 * The caller is responsible for serializing this function against
1015 * blk_mq_{start,stop}_hw_queue().
1017 bool blk_mq_queue_stopped(struct request_queue
*q
)
1019 struct blk_mq_hw_ctx
*hctx
;
1022 queue_for_each_hw_ctx(q
, hctx
, i
)
1023 if (blk_mq_hctx_stopped(hctx
))
1028 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1030 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1032 cancel_work(&hctx
->run_work
);
1033 cancel_delayed_work(&hctx
->delay_work
);
1034 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1036 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1038 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1040 struct blk_mq_hw_ctx
*hctx
;
1043 queue_for_each_hw_ctx(q
, hctx
, i
)
1044 blk_mq_stop_hw_queue(hctx
);
1046 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1048 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1050 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1052 blk_mq_run_hw_queue(hctx
, false);
1054 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1056 void blk_mq_start_hw_queues(struct request_queue
*q
)
1058 struct blk_mq_hw_ctx
*hctx
;
1061 queue_for_each_hw_ctx(q
, hctx
, i
)
1062 blk_mq_start_hw_queue(hctx
);
1064 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1066 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1068 if (!blk_mq_hctx_stopped(hctx
))
1071 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1072 blk_mq_run_hw_queue(hctx
, async
);
1074 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1076 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1078 struct blk_mq_hw_ctx
*hctx
;
1081 queue_for_each_hw_ctx(q
, hctx
, i
)
1082 blk_mq_start_stopped_hw_queue(hctx
, async
);
1084 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1086 static void blk_mq_run_work_fn(struct work_struct
*work
)
1088 struct blk_mq_hw_ctx
*hctx
;
1090 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1092 __blk_mq_run_hw_queue(hctx
);
1095 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1097 struct blk_mq_hw_ctx
*hctx
;
1099 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1101 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1102 __blk_mq_run_hw_queue(hctx
);
1105 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1107 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1110 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1111 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1113 EXPORT_SYMBOL(blk_mq_delay_queue
);
1115 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1119 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1121 trace_block_rq_insert(hctx
->queue
, rq
);
1124 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1126 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1129 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
1130 struct request
*rq
, bool at_head
)
1132 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1134 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1135 blk_mq_hctx_mark_pending(hctx
, ctx
);
1138 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
1141 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1142 struct request_queue
*q
= rq
->q
;
1143 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1145 spin_lock(&ctx
->lock
);
1146 __blk_mq_insert_request(hctx
, rq
, at_head
);
1147 spin_unlock(&ctx
->lock
);
1150 blk_mq_run_hw_queue(hctx
, async
);
1153 static void blk_mq_insert_requests(struct request_queue
*q
,
1154 struct blk_mq_ctx
*ctx
,
1155 struct list_head
*list
,
1160 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1162 trace_block_unplug(q
, depth
, !from_schedule
);
1165 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1168 spin_lock(&ctx
->lock
);
1169 while (!list_empty(list
)) {
1172 rq
= list_first_entry(list
, struct request
, queuelist
);
1173 BUG_ON(rq
->mq_ctx
!= ctx
);
1174 list_del_init(&rq
->queuelist
);
1175 __blk_mq_insert_req_list(hctx
, rq
, false);
1177 blk_mq_hctx_mark_pending(hctx
, ctx
);
1178 spin_unlock(&ctx
->lock
);
1180 blk_mq_run_hw_queue(hctx
, from_schedule
);
1183 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1185 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1186 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1188 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1189 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1190 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1193 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1195 struct blk_mq_ctx
*this_ctx
;
1196 struct request_queue
*this_q
;
1199 LIST_HEAD(ctx_list
);
1202 list_splice_init(&plug
->mq_list
, &list
);
1204 list_sort(NULL
, &list
, plug_ctx_cmp
);
1210 while (!list_empty(&list
)) {
1211 rq
= list_entry_rq(list
.next
);
1212 list_del_init(&rq
->queuelist
);
1214 if (rq
->mq_ctx
!= this_ctx
) {
1216 blk_mq_insert_requests(this_q
, this_ctx
,
1221 this_ctx
= rq
->mq_ctx
;
1227 list_add_tail(&rq
->queuelist
, &ctx_list
);
1231 * If 'this_ctx' is set, we know we have entries to complete
1232 * on 'ctx_list'. Do those.
1235 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1240 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1242 init_request_from_bio(rq
, bio
);
1244 blk_account_io_start(rq
, true);
1247 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1249 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1250 !blk_queue_nomerges(hctx
->queue
);
1253 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1254 struct blk_mq_ctx
*ctx
,
1255 struct request
*rq
, struct bio
*bio
)
1257 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1258 blk_mq_bio_to_request(rq
, bio
);
1259 spin_lock(&ctx
->lock
);
1261 __blk_mq_insert_request(hctx
, rq
, false);
1262 spin_unlock(&ctx
->lock
);
1265 struct request_queue
*q
= hctx
->queue
;
1267 spin_lock(&ctx
->lock
);
1268 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1269 blk_mq_bio_to_request(rq
, bio
);
1273 spin_unlock(&ctx
->lock
);
1274 __blk_mq_free_request(hctx
, ctx
, rq
);
1279 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1281 struct blk_mq_alloc_data
*data
)
1283 struct blk_mq_hw_ctx
*hctx
;
1284 struct blk_mq_ctx
*ctx
;
1287 blk_queue_enter_live(q
);
1288 ctx
= blk_mq_get_ctx(q
);
1289 hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1291 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1292 blk_mq_set_alloc_data(data
, q
, 0, ctx
, hctx
);
1293 rq
= __blk_mq_alloc_request(data
, bio
->bi_opf
);
1295 data
->hctx
->queued
++;
1299 static void blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
)
1302 struct request_queue
*q
= rq
->q
;
1303 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, rq
->mq_ctx
->cpu
);
1304 struct blk_mq_queue_data bd
= {
1309 blk_qc_t new_cookie
= blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
);
1311 if (blk_mq_hctx_stopped(hctx
))
1315 * For OK queue, we are done. For error, kill it. Any other
1316 * error (busy), just add it to our list as we previously
1319 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1320 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1321 *cookie
= new_cookie
;
1325 __blk_mq_requeue_request(rq
);
1327 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1328 *cookie
= BLK_QC_T_NONE
;
1330 blk_mq_end_request(rq
, rq
->errors
);
1335 blk_mq_insert_request(rq
, false, true, true);
1339 * Multiple hardware queue variant. This will not use per-process plugs,
1340 * but will attempt to bypass the hctx queueing if we can go straight to
1341 * hardware for SYNC IO.
1343 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1345 const int is_sync
= op_is_sync(bio
->bi_opf
);
1346 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1347 struct blk_mq_alloc_data data
;
1349 unsigned int request_count
= 0, srcu_idx
;
1350 struct blk_plug
*plug
;
1351 struct request
*same_queue_rq
= NULL
;
1353 unsigned int wb_acct
;
1355 blk_queue_bounce(q
, &bio
);
1357 blk_queue_split(q
, &bio
, q
->bio_split
);
1359 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1361 return BLK_QC_T_NONE
;
1364 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1365 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1366 return BLK_QC_T_NONE
;
1368 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1370 rq
= blk_mq_map_request(q
, bio
, &data
);
1371 if (unlikely(!rq
)) {
1372 __wbt_done(q
->rq_wb
, wb_acct
);
1373 return BLK_QC_T_NONE
;
1376 wbt_track(&rq
->issue_stat
, wb_acct
);
1378 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1380 if (unlikely(is_flush_fua
)) {
1381 blk_mq_bio_to_request(rq
, bio
);
1382 blk_insert_flush(rq
);
1386 plug
= current
->plug
;
1388 * If the driver supports defer issued based on 'last', then
1389 * queue it up like normal since we can potentially save some
1392 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1393 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1394 struct request
*old_rq
= NULL
;
1396 blk_mq_bio_to_request(rq
, bio
);
1399 * We do limited plugging. If the bio can be merged, do that.
1400 * Otherwise the existing request in the plug list will be
1401 * issued. So the plug list will have one request at most
1405 * The plug list might get flushed before this. If that
1406 * happens, same_queue_rq is invalid and plug list is
1409 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1410 old_rq
= same_queue_rq
;
1411 list_del_init(&old_rq
->queuelist
);
1413 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1414 } else /* is_sync */
1416 blk_mq_put_ctx(data
.ctx
);
1420 if (!(data
.hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1422 blk_mq_try_issue_directly(old_rq
, &cookie
);
1425 srcu_idx
= srcu_read_lock(&data
.hctx
->queue_rq_srcu
);
1426 blk_mq_try_issue_directly(old_rq
, &cookie
);
1427 srcu_read_unlock(&data
.hctx
->queue_rq_srcu
, srcu_idx
);
1432 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1434 * For a SYNC request, send it to the hardware immediately. For
1435 * an ASYNC request, just ensure that we run it later on. The
1436 * latter allows for merging opportunities and more efficient
1440 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1442 blk_mq_put_ctx(data
.ctx
);
1448 * Single hardware queue variant. This will attempt to use any per-process
1449 * plug for merging and IO deferral.
1451 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1453 const int is_sync
= op_is_sync(bio
->bi_opf
);
1454 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1455 struct blk_plug
*plug
;
1456 unsigned int request_count
= 0;
1457 struct blk_mq_alloc_data data
;
1460 unsigned int wb_acct
;
1462 blk_queue_bounce(q
, &bio
);
1464 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1466 return BLK_QC_T_NONE
;
1469 blk_queue_split(q
, &bio
, q
->bio_split
);
1471 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1472 if (blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1473 return BLK_QC_T_NONE
;
1475 request_count
= blk_plug_queued_count(q
);
1477 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1479 rq
= blk_mq_map_request(q
, bio
, &data
);
1480 if (unlikely(!rq
)) {
1481 __wbt_done(q
->rq_wb
, wb_acct
);
1482 return BLK_QC_T_NONE
;
1485 wbt_track(&rq
->issue_stat
, wb_acct
);
1487 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1489 if (unlikely(is_flush_fua
)) {
1490 blk_mq_bio_to_request(rq
, bio
);
1491 blk_insert_flush(rq
);
1496 * A task plug currently exists. Since this is completely lockless,
1497 * utilize that to temporarily store requests until the task is
1498 * either done or scheduled away.
1500 plug
= current
->plug
;
1502 struct request
*last
= NULL
;
1504 blk_mq_bio_to_request(rq
, bio
);
1507 * @request_count may become stale because of schedule
1508 * out, so check the list again.
1510 if (list_empty(&plug
->mq_list
))
1513 trace_block_plug(q
);
1515 last
= list_entry_rq(plug
->mq_list
.prev
);
1517 blk_mq_put_ctx(data
.ctx
);
1519 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1520 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1521 blk_flush_plug_list(plug
, false);
1522 trace_block_plug(q
);
1525 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1529 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1531 * For a SYNC request, send it to the hardware immediately. For
1532 * an ASYNC request, just ensure that we run it later on. The
1533 * latter allows for merging opportunities and more efficient
1537 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1540 blk_mq_put_ctx(data
.ctx
);
1544 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1545 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1549 if (tags
->rqs
&& set
->ops
->exit_request
) {
1552 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1555 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1557 tags
->rqs
[i
] = NULL
;
1561 while (!list_empty(&tags
->page_list
)) {
1562 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1563 list_del_init(&page
->lru
);
1565 * Remove kmemleak object previously allocated in
1566 * blk_mq_init_rq_map().
1568 kmemleak_free(page_address(page
));
1569 __free_pages(page
, page
->private);
1574 blk_mq_free_tags(tags
);
1577 static size_t order_to_size(unsigned int order
)
1579 return (size_t)PAGE_SIZE
<< order
;
1582 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1583 unsigned int hctx_idx
)
1585 struct blk_mq_tags
*tags
;
1586 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1587 size_t rq_size
, left
;
1589 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1591 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1595 INIT_LIST_HEAD(&tags
->page_list
);
1597 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1598 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1601 blk_mq_free_tags(tags
);
1606 * rq_size is the size of the request plus driver payload, rounded
1607 * to the cacheline size
1609 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1611 left
= rq_size
* set
->queue_depth
;
1613 for (i
= 0; i
< set
->queue_depth
; ) {
1614 int this_order
= max_order
;
1619 while (this_order
&& left
< order_to_size(this_order
- 1))
1623 page
= alloc_pages_node(set
->numa_node
,
1624 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1630 if (order_to_size(this_order
) < rq_size
)
1637 page
->private = this_order
;
1638 list_add_tail(&page
->lru
, &tags
->page_list
);
1640 p
= page_address(page
);
1642 * Allow kmemleak to scan these pages as they contain pointers
1643 * to additional allocations like via ops->init_request().
1645 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1646 entries_per_page
= order_to_size(this_order
) / rq_size
;
1647 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1648 left
-= to_do
* rq_size
;
1649 for (j
= 0; j
< to_do
; j
++) {
1651 if (set
->ops
->init_request
) {
1652 if (set
->ops
->init_request(set
->driver_data
,
1653 tags
->rqs
[i
], hctx_idx
, i
,
1655 tags
->rqs
[i
] = NULL
;
1667 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1672 * 'cpu' is going away. splice any existing rq_list entries from this
1673 * software queue to the hw queue dispatch list, and ensure that it
1676 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1678 struct blk_mq_hw_ctx
*hctx
;
1679 struct blk_mq_ctx
*ctx
;
1682 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1683 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1685 spin_lock(&ctx
->lock
);
1686 if (!list_empty(&ctx
->rq_list
)) {
1687 list_splice_init(&ctx
->rq_list
, &tmp
);
1688 blk_mq_hctx_clear_pending(hctx
, ctx
);
1690 spin_unlock(&ctx
->lock
);
1692 if (list_empty(&tmp
))
1695 spin_lock(&hctx
->lock
);
1696 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1697 spin_unlock(&hctx
->lock
);
1699 blk_mq_run_hw_queue(hctx
, true);
1703 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1705 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1709 /* hctx->ctxs will be freed in queue's release handler */
1710 static void blk_mq_exit_hctx(struct request_queue
*q
,
1711 struct blk_mq_tag_set
*set
,
1712 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1714 unsigned flush_start_tag
= set
->queue_depth
;
1716 blk_mq_tag_idle(hctx
);
1718 if (set
->ops
->exit_request
)
1719 set
->ops
->exit_request(set
->driver_data
,
1720 hctx
->fq
->flush_rq
, hctx_idx
,
1721 flush_start_tag
+ hctx_idx
);
1723 if (set
->ops
->exit_hctx
)
1724 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1726 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1727 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1729 blk_mq_remove_cpuhp(hctx
);
1730 blk_free_flush_queue(hctx
->fq
);
1731 sbitmap_free(&hctx
->ctx_map
);
1734 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1735 struct blk_mq_tag_set
*set
, int nr_queue
)
1737 struct blk_mq_hw_ctx
*hctx
;
1740 queue_for_each_hw_ctx(q
, hctx
, i
) {
1743 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1747 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1748 struct blk_mq_tag_set
*set
)
1750 struct blk_mq_hw_ctx
*hctx
;
1753 queue_for_each_hw_ctx(q
, hctx
, i
)
1754 free_cpumask_var(hctx
->cpumask
);
1757 static int blk_mq_init_hctx(struct request_queue
*q
,
1758 struct blk_mq_tag_set
*set
,
1759 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1762 unsigned flush_start_tag
= set
->queue_depth
;
1764 node
= hctx
->numa_node
;
1765 if (node
== NUMA_NO_NODE
)
1766 node
= hctx
->numa_node
= set
->numa_node
;
1768 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1769 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1770 spin_lock_init(&hctx
->lock
);
1771 INIT_LIST_HEAD(&hctx
->dispatch
);
1773 hctx
->queue_num
= hctx_idx
;
1774 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1776 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1778 hctx
->tags
= set
->tags
[hctx_idx
];
1781 * Allocate space for all possible cpus to avoid allocation at
1784 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1787 goto unregister_cpu_notifier
;
1789 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1795 if (set
->ops
->init_hctx
&&
1796 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1799 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1803 if (set
->ops
->init_request
&&
1804 set
->ops
->init_request(set
->driver_data
,
1805 hctx
->fq
->flush_rq
, hctx_idx
,
1806 flush_start_tag
+ hctx_idx
, node
))
1809 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1810 init_srcu_struct(&hctx
->queue_rq_srcu
);
1817 if (set
->ops
->exit_hctx
)
1818 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1820 sbitmap_free(&hctx
->ctx_map
);
1823 unregister_cpu_notifier
:
1824 blk_mq_remove_cpuhp(hctx
);
1828 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1829 unsigned int nr_hw_queues
)
1833 for_each_possible_cpu(i
) {
1834 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1835 struct blk_mq_hw_ctx
*hctx
;
1837 memset(__ctx
, 0, sizeof(*__ctx
));
1839 spin_lock_init(&__ctx
->lock
);
1840 INIT_LIST_HEAD(&__ctx
->rq_list
);
1842 blk_stat_init(&__ctx
->stat
[BLK_STAT_READ
]);
1843 blk_stat_init(&__ctx
->stat
[BLK_STAT_WRITE
]);
1845 /* If the cpu isn't online, the cpu is mapped to first hctx */
1849 hctx
= blk_mq_map_queue(q
, i
);
1852 * Set local node, IFF we have more than one hw queue. If
1853 * not, we remain on the home node of the device
1855 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1856 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1860 static void blk_mq_map_swqueue(struct request_queue
*q
,
1861 const struct cpumask
*online_mask
)
1863 unsigned int i
, hctx_idx
;
1864 struct blk_mq_hw_ctx
*hctx
;
1865 struct blk_mq_ctx
*ctx
;
1866 struct blk_mq_tag_set
*set
= q
->tag_set
;
1869 * Avoid others reading imcomplete hctx->cpumask through sysfs
1871 mutex_lock(&q
->sysfs_lock
);
1873 queue_for_each_hw_ctx(q
, hctx
, i
) {
1874 cpumask_clear(hctx
->cpumask
);
1879 * Map software to hardware queues
1881 for_each_possible_cpu(i
) {
1882 /* If the cpu isn't online, the cpu is mapped to first hctx */
1883 if (!cpumask_test_cpu(i
, online_mask
))
1886 hctx_idx
= q
->mq_map
[i
];
1887 /* unmapped hw queue can be remapped after CPU topo changed */
1888 if (!set
->tags
[hctx_idx
]) {
1889 set
->tags
[hctx_idx
] = blk_mq_init_rq_map(set
, hctx_idx
);
1892 * If tags initialization fail for some hctx,
1893 * that hctx won't be brought online. In this
1894 * case, remap the current ctx to hctx[0] which
1895 * is guaranteed to always have tags allocated
1897 if (!set
->tags
[hctx_idx
])
1901 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1902 hctx
= blk_mq_map_queue(q
, i
);
1904 cpumask_set_cpu(i
, hctx
->cpumask
);
1905 ctx
->index_hw
= hctx
->nr_ctx
;
1906 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1909 mutex_unlock(&q
->sysfs_lock
);
1911 queue_for_each_hw_ctx(q
, hctx
, i
) {
1913 * If no software queues are mapped to this hardware queue,
1914 * disable it and free the request entries.
1916 if (!hctx
->nr_ctx
) {
1917 /* Never unmap queue 0. We need it as a
1918 * fallback in case of a new remap fails
1921 if (i
&& set
->tags
[i
]) {
1922 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1923 set
->tags
[i
] = NULL
;
1929 hctx
->tags
= set
->tags
[i
];
1930 WARN_ON(!hctx
->tags
);
1933 * Set the map size to the number of mapped software queues.
1934 * This is more accurate and more efficient than looping
1935 * over all possibly mapped software queues.
1937 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
1940 * Initialize batch roundrobin counts
1942 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1943 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1947 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
1949 struct blk_mq_hw_ctx
*hctx
;
1952 queue_for_each_hw_ctx(q
, hctx
, i
) {
1954 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1956 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1960 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
1962 struct request_queue
*q
;
1964 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1965 blk_mq_freeze_queue(q
);
1966 queue_set_hctx_shared(q
, shared
);
1967 blk_mq_unfreeze_queue(q
);
1971 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1973 struct blk_mq_tag_set
*set
= q
->tag_set
;
1975 mutex_lock(&set
->tag_list_lock
);
1976 list_del_init(&q
->tag_set_list
);
1977 if (list_is_singular(&set
->tag_list
)) {
1978 /* just transitioned to unshared */
1979 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1980 /* update existing queue */
1981 blk_mq_update_tag_set_depth(set
, false);
1983 mutex_unlock(&set
->tag_list_lock
);
1986 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1987 struct request_queue
*q
)
1991 mutex_lock(&set
->tag_list_lock
);
1993 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1994 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1995 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
1996 /* update existing queue */
1997 blk_mq_update_tag_set_depth(set
, true);
1999 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2000 queue_set_hctx_shared(q
, true);
2001 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
2003 mutex_unlock(&set
->tag_list_lock
);
2007 * It is the actual release handler for mq, but we do it from
2008 * request queue's release handler for avoiding use-after-free
2009 * and headache because q->mq_kobj shouldn't have been introduced,
2010 * but we can't group ctx/kctx kobj without it.
2012 void blk_mq_release(struct request_queue
*q
)
2014 struct blk_mq_hw_ctx
*hctx
;
2017 /* hctx kobj stays in hctx */
2018 queue_for_each_hw_ctx(q
, hctx
, i
) {
2027 kfree(q
->queue_hw_ctx
);
2029 /* ctx kobj stays in queue_ctx */
2030 free_percpu(q
->queue_ctx
);
2033 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2035 struct request_queue
*uninit_q
, *q
;
2037 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2039 return ERR_PTR(-ENOMEM
);
2041 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2043 blk_cleanup_queue(uninit_q
);
2047 EXPORT_SYMBOL(blk_mq_init_queue
);
2049 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2050 struct request_queue
*q
)
2053 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2055 blk_mq_sysfs_unregister(q
);
2056 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2062 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2063 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2068 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2075 atomic_set(&hctxs
[i
]->nr_active
, 0);
2076 hctxs
[i
]->numa_node
= node
;
2077 hctxs
[i
]->queue_num
= i
;
2079 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2080 free_cpumask_var(hctxs
[i
]->cpumask
);
2085 blk_mq_hctx_kobj_init(hctxs
[i
]);
2087 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2088 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2092 blk_mq_free_rq_map(set
, hctx
->tags
, j
);
2093 set
->tags
[j
] = NULL
;
2095 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2096 free_cpumask_var(hctx
->cpumask
);
2097 kobject_put(&hctx
->kobj
);
2104 q
->nr_hw_queues
= i
;
2105 blk_mq_sysfs_register(q
);
2108 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2109 struct request_queue
*q
)
2111 /* mark the queue as mq asap */
2112 q
->mq_ops
= set
->ops
;
2114 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2118 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2119 GFP_KERNEL
, set
->numa_node
);
2120 if (!q
->queue_hw_ctx
)
2123 q
->mq_map
= set
->mq_map
;
2125 blk_mq_realloc_hw_ctxs(set
, q
);
2126 if (!q
->nr_hw_queues
)
2129 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2130 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2132 q
->nr_queues
= nr_cpu_ids
;
2134 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2136 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2137 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2139 q
->sg_reserved_size
= INT_MAX
;
2141 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2142 INIT_LIST_HEAD(&q
->requeue_list
);
2143 spin_lock_init(&q
->requeue_lock
);
2145 if (q
->nr_hw_queues
> 1)
2146 blk_queue_make_request(q
, blk_mq_make_request
);
2148 blk_queue_make_request(q
, blk_sq_make_request
);
2151 * Do this after blk_queue_make_request() overrides it...
2153 q
->nr_requests
= set
->queue_depth
;
2156 * Default to classic polling
2160 if (set
->ops
->complete
)
2161 blk_queue_softirq_done(q
, set
->ops
->complete
);
2163 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2166 mutex_lock(&all_q_mutex
);
2168 list_add_tail(&q
->all_q_node
, &all_q_list
);
2169 blk_mq_add_queue_tag_set(set
, q
);
2170 blk_mq_map_swqueue(q
, cpu_online_mask
);
2172 mutex_unlock(&all_q_mutex
);
2178 kfree(q
->queue_hw_ctx
);
2180 free_percpu(q
->queue_ctx
);
2183 return ERR_PTR(-ENOMEM
);
2185 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2187 void blk_mq_free_queue(struct request_queue
*q
)
2189 struct blk_mq_tag_set
*set
= q
->tag_set
;
2191 mutex_lock(&all_q_mutex
);
2192 list_del_init(&q
->all_q_node
);
2193 mutex_unlock(&all_q_mutex
);
2197 blk_mq_del_queue_tag_set(q
);
2199 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2200 blk_mq_free_hw_queues(q
, set
);
2203 /* Basically redo blk_mq_init_queue with queue frozen */
2204 static void blk_mq_queue_reinit(struct request_queue
*q
,
2205 const struct cpumask
*online_mask
)
2207 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2209 blk_mq_sysfs_unregister(q
);
2212 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2213 * we should change hctx numa_node according to new topology (this
2214 * involves free and re-allocate memory, worthy doing?)
2217 blk_mq_map_swqueue(q
, online_mask
);
2219 blk_mq_sysfs_register(q
);
2223 * New online cpumask which is going to be set in this hotplug event.
2224 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2225 * one-by-one and dynamically allocating this could result in a failure.
2227 static struct cpumask cpuhp_online_new
;
2229 static void blk_mq_queue_reinit_work(void)
2231 struct request_queue
*q
;
2233 mutex_lock(&all_q_mutex
);
2235 * We need to freeze and reinit all existing queues. Freezing
2236 * involves synchronous wait for an RCU grace period and doing it
2237 * one by one may take a long time. Start freezing all queues in
2238 * one swoop and then wait for the completions so that freezing can
2239 * take place in parallel.
2241 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2242 blk_mq_freeze_queue_start(q
);
2243 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2244 blk_mq_freeze_queue_wait(q
);
2246 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2247 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2249 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2250 blk_mq_unfreeze_queue(q
);
2252 mutex_unlock(&all_q_mutex
);
2255 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2257 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2258 blk_mq_queue_reinit_work();
2263 * Before hotadded cpu starts handling requests, new mappings must be
2264 * established. Otherwise, these requests in hw queue might never be
2267 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2268 * for CPU0, and ctx1 for CPU1).
2270 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2271 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2273 * And then while running hw queue, flush_busy_ctxs() finds bit0 is set in
2274 * pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2275 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list
2278 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2280 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2281 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2282 blk_mq_queue_reinit_work();
2286 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2290 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2291 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2300 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2306 * Allocate the request maps associated with this tag_set. Note that this
2307 * may reduce the depth asked for, if memory is tight. set->queue_depth
2308 * will be updated to reflect the allocated depth.
2310 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2315 depth
= set
->queue_depth
;
2317 err
= __blk_mq_alloc_rq_maps(set
);
2321 set
->queue_depth
>>= 1;
2322 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2326 } while (set
->queue_depth
);
2328 if (!set
->queue_depth
|| err
) {
2329 pr_err("blk-mq: failed to allocate request map\n");
2333 if (depth
!= set
->queue_depth
)
2334 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2335 depth
, set
->queue_depth
);
2341 * Alloc a tag set to be associated with one or more request queues.
2342 * May fail with EINVAL for various error conditions. May adjust the
2343 * requested depth down, if if it too large. In that case, the set
2344 * value will be stored in set->queue_depth.
2346 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2350 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2352 if (!set
->nr_hw_queues
)
2354 if (!set
->queue_depth
)
2356 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2359 if (!set
->ops
->queue_rq
)
2362 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2363 pr_info("blk-mq: reduced tag depth to %u\n",
2365 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2369 * If a crashdump is active, then we are potentially in a very
2370 * memory constrained environment. Limit us to 1 queue and
2371 * 64 tags to prevent using too much memory.
2373 if (is_kdump_kernel()) {
2374 set
->nr_hw_queues
= 1;
2375 set
->queue_depth
= min(64U, set
->queue_depth
);
2378 * There is no use for more h/w queues than cpus.
2380 if (set
->nr_hw_queues
> nr_cpu_ids
)
2381 set
->nr_hw_queues
= nr_cpu_ids
;
2383 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2384 GFP_KERNEL
, set
->numa_node
);
2389 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2390 GFP_KERNEL
, set
->numa_node
);
2394 if (set
->ops
->map_queues
)
2395 ret
= set
->ops
->map_queues(set
);
2397 ret
= blk_mq_map_queues(set
);
2399 goto out_free_mq_map
;
2401 ret
= blk_mq_alloc_rq_maps(set
);
2403 goto out_free_mq_map
;
2405 mutex_init(&set
->tag_list_lock
);
2406 INIT_LIST_HEAD(&set
->tag_list
);
2418 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2420 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2424 for (i
= 0; i
< nr_cpu_ids
; i
++) {
2426 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2435 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2437 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2439 struct blk_mq_tag_set
*set
= q
->tag_set
;
2440 struct blk_mq_hw_ctx
*hctx
;
2443 if (!set
|| nr
> set
->queue_depth
)
2447 queue_for_each_hw_ctx(q
, hctx
, i
) {
2450 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2456 q
->nr_requests
= nr
;
2461 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2463 struct request_queue
*q
;
2465 if (nr_hw_queues
> nr_cpu_ids
)
2466 nr_hw_queues
= nr_cpu_ids
;
2467 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2470 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2471 blk_mq_freeze_queue(q
);
2473 set
->nr_hw_queues
= nr_hw_queues
;
2474 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2475 blk_mq_realloc_hw_ctxs(set
, q
);
2477 if (q
->nr_hw_queues
> 1)
2478 blk_queue_make_request(q
, blk_mq_make_request
);
2480 blk_queue_make_request(q
, blk_sq_make_request
);
2482 blk_mq_queue_reinit(q
, cpu_online_mask
);
2485 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2486 blk_mq_unfreeze_queue(q
);
2488 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2490 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2491 struct blk_mq_hw_ctx
*hctx
,
2494 struct blk_rq_stat stat
[2];
2495 unsigned long ret
= 0;
2498 * If stats collection isn't on, don't sleep but turn it on for
2501 if (!blk_stat_enable(q
))
2505 * We don't have to do this once per IO, should optimize this
2506 * to just use the current window of stats until it changes
2508 memset(&stat
, 0, sizeof(stat
));
2509 blk_hctx_stat_get(hctx
, stat
);
2512 * As an optimistic guess, use half of the mean service time
2513 * for this type of request. We can (and should) make this smarter.
2514 * For instance, if the completion latencies are tight, we can
2515 * get closer than just half the mean. This is especially
2516 * important on devices where the completion latencies are longer
2519 if (req_op(rq
) == REQ_OP_READ
&& stat
[BLK_STAT_READ
].nr_samples
)
2520 ret
= (stat
[BLK_STAT_READ
].mean
+ 1) / 2;
2521 else if (req_op(rq
) == REQ_OP_WRITE
&& stat
[BLK_STAT_WRITE
].nr_samples
)
2522 ret
= (stat
[BLK_STAT_WRITE
].mean
+ 1) / 2;
2527 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2528 struct blk_mq_hw_ctx
*hctx
,
2531 struct hrtimer_sleeper hs
;
2532 enum hrtimer_mode mode
;
2536 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2542 * -1: don't ever hybrid sleep
2543 * 0: use half of prev avg
2544 * >0: use this specific value
2546 if (q
->poll_nsec
== -1)
2548 else if (q
->poll_nsec
> 0)
2549 nsecs
= q
->poll_nsec
;
2551 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2556 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2559 * This will be replaced with the stats tracking code, using
2560 * 'avg_completion_time / 2' as the pre-sleep target.
2564 mode
= HRTIMER_MODE_REL
;
2565 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2566 hrtimer_set_expires(&hs
.timer
, kt
);
2568 hrtimer_init_sleeper(&hs
, current
);
2570 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2572 set_current_state(TASK_UNINTERRUPTIBLE
);
2573 hrtimer_start_expires(&hs
.timer
, mode
);
2576 hrtimer_cancel(&hs
.timer
);
2577 mode
= HRTIMER_MODE_ABS
;
2578 } while (hs
.task
&& !signal_pending(current
));
2580 __set_current_state(TASK_RUNNING
);
2581 destroy_hrtimer_on_stack(&hs
.timer
);
2585 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2587 struct request_queue
*q
= hctx
->queue
;
2591 * If we sleep, have the caller restart the poll loop to reset
2592 * the state. Like for the other success return cases, the
2593 * caller is responsible for checking if the IO completed. If
2594 * the IO isn't complete, we'll get called again and will go
2595 * straight to the busy poll loop.
2597 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2600 hctx
->poll_considered
++;
2602 state
= current
->state
;
2603 while (!need_resched()) {
2606 hctx
->poll_invoked
++;
2608 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2610 hctx
->poll_success
++;
2611 set_current_state(TASK_RUNNING
);
2615 if (signal_pending_state(state
, current
))
2616 set_current_state(TASK_RUNNING
);
2618 if (current
->state
== TASK_RUNNING
)
2628 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2630 struct blk_mq_hw_ctx
*hctx
;
2631 struct blk_plug
*plug
;
2634 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2635 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2638 plug
= current
->plug
;
2640 blk_flush_plug_list(plug
, false);
2642 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2643 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2645 return __blk_mq_poll(hctx
, rq
);
2647 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2649 void blk_mq_disable_hotplug(void)
2651 mutex_lock(&all_q_mutex
);
2654 void blk_mq_enable_hotplug(void)
2656 mutex_unlock(&all_q_mutex
);
2659 static int __init
blk_mq_init(void)
2661 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2662 blk_mq_hctx_notify_dead
);
2664 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2665 blk_mq_queue_reinit_prepare
,
2666 blk_mq_queue_reinit_dead
);
2669 subsys_initcall(blk_mq_init
);