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 blk_mq_put_tag(hctx
, ctx
, tag
);
339 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
341 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
343 ctx
->rq_completed
[rq_is_sync(rq
)]++;
344 __blk_mq_free_request(hctx
, ctx
, rq
);
347 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request
);
349 void blk_mq_free_request(struct request
*rq
)
351 blk_mq_free_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
353 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
355 inline void __blk_mq_end_request(struct request
*rq
, int error
)
357 blk_account_io_done(rq
);
360 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
361 rq
->end_io(rq
, error
);
363 if (unlikely(blk_bidi_rq(rq
)))
364 blk_mq_free_request(rq
->next_rq
);
365 blk_mq_free_request(rq
);
368 EXPORT_SYMBOL(__blk_mq_end_request
);
370 void blk_mq_end_request(struct request
*rq
, int error
)
372 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
374 __blk_mq_end_request(rq
, error
);
376 EXPORT_SYMBOL(blk_mq_end_request
);
378 static void __blk_mq_complete_request_remote(void *data
)
380 struct request
*rq
= data
;
382 rq
->q
->softirq_done_fn(rq
);
385 static void blk_mq_ipi_complete_request(struct request
*rq
)
387 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
391 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
392 rq
->q
->softirq_done_fn(rq
);
397 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
398 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
400 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
401 rq
->csd
.func
= __blk_mq_complete_request_remote
;
404 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
406 rq
->q
->softirq_done_fn(rq
);
411 static void blk_mq_stat_add(struct request
*rq
)
413 if (rq
->rq_flags
& RQF_STATS
) {
415 * We could rq->mq_ctx here, but there's less of a risk
416 * of races if we have the completion event add the stats
417 * to the local software queue.
419 struct blk_mq_ctx
*ctx
;
421 ctx
= __blk_mq_get_ctx(rq
->q
, raw_smp_processor_id());
422 blk_stat_add(&ctx
->stat
[rq_data_dir(rq
)], rq
);
426 static void __blk_mq_complete_request(struct request
*rq
)
428 struct request_queue
*q
= rq
->q
;
432 if (!q
->softirq_done_fn
)
433 blk_mq_end_request(rq
, rq
->errors
);
435 blk_mq_ipi_complete_request(rq
);
439 * blk_mq_complete_request - end I/O on a request
440 * @rq: the request being processed
443 * Ends all I/O on a request. It does not handle partial completions.
444 * The actual completion happens out-of-order, through a IPI handler.
446 void blk_mq_complete_request(struct request
*rq
, int error
)
448 struct request_queue
*q
= rq
->q
;
450 if (unlikely(blk_should_fake_timeout(q
)))
452 if (!blk_mark_rq_complete(rq
)) {
454 __blk_mq_complete_request(rq
);
457 EXPORT_SYMBOL(blk_mq_complete_request
);
459 int blk_mq_request_started(struct request
*rq
)
461 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
463 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
465 void blk_mq_start_request(struct request
*rq
)
467 struct request_queue
*q
= rq
->q
;
469 trace_block_rq_issue(q
, rq
);
471 rq
->resid_len
= blk_rq_bytes(rq
);
472 if (unlikely(blk_bidi_rq(rq
)))
473 rq
->next_rq
->resid_len
= blk_rq_bytes(rq
->next_rq
);
475 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
476 blk_stat_set_issue_time(&rq
->issue_stat
);
477 rq
->rq_flags
|= RQF_STATS
;
478 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
484 * Ensure that ->deadline is visible before set the started
485 * flag and clear the completed flag.
487 smp_mb__before_atomic();
490 * Mark us as started and clear complete. Complete might have been
491 * set if requeue raced with timeout, which then marked it as
492 * complete. So be sure to clear complete again when we start
493 * the request, otherwise we'll ignore the completion event.
495 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
496 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
497 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
498 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
500 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
502 * Make sure space for the drain appears. We know we can do
503 * this because max_hw_segments has been adjusted to be one
504 * fewer than the device can handle.
506 rq
->nr_phys_segments
++;
509 EXPORT_SYMBOL(blk_mq_start_request
);
511 static void __blk_mq_requeue_request(struct request
*rq
)
513 struct request_queue
*q
= rq
->q
;
515 trace_block_rq_requeue(q
, rq
);
516 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
518 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
519 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
520 rq
->nr_phys_segments
--;
524 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
526 __blk_mq_requeue_request(rq
);
528 BUG_ON(blk_queued_rq(rq
));
529 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
531 EXPORT_SYMBOL(blk_mq_requeue_request
);
533 static void blk_mq_requeue_work(struct work_struct
*work
)
535 struct request_queue
*q
=
536 container_of(work
, struct request_queue
, requeue_work
.work
);
538 struct request
*rq
, *next
;
541 spin_lock_irqsave(&q
->requeue_lock
, flags
);
542 list_splice_init(&q
->requeue_list
, &rq_list
);
543 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
545 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
546 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
549 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
550 list_del_init(&rq
->queuelist
);
551 blk_mq_insert_request(rq
, true, false, false);
554 while (!list_empty(&rq_list
)) {
555 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
556 list_del_init(&rq
->queuelist
);
557 blk_mq_insert_request(rq
, false, false, false);
560 blk_mq_run_hw_queues(q
, false);
563 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
564 bool kick_requeue_list
)
566 struct request_queue
*q
= rq
->q
;
570 * We abuse this flag that is otherwise used by the I/O scheduler to
571 * request head insertation from the workqueue.
573 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
575 spin_lock_irqsave(&q
->requeue_lock
, flags
);
577 rq
->rq_flags
|= RQF_SOFTBARRIER
;
578 list_add(&rq
->queuelist
, &q
->requeue_list
);
580 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
582 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
584 if (kick_requeue_list
)
585 blk_mq_kick_requeue_list(q
);
587 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
589 void blk_mq_kick_requeue_list(struct request_queue
*q
)
591 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
593 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
595 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
598 kblockd_schedule_delayed_work(&q
->requeue_work
,
599 msecs_to_jiffies(msecs
));
601 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
603 void blk_mq_abort_requeue_list(struct request_queue
*q
)
608 spin_lock_irqsave(&q
->requeue_lock
, flags
);
609 list_splice_init(&q
->requeue_list
, &rq_list
);
610 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
612 while (!list_empty(&rq_list
)) {
615 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
616 list_del_init(&rq
->queuelist
);
618 blk_mq_end_request(rq
, rq
->errors
);
621 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
623 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
625 if (tag
< tags
->nr_tags
) {
626 prefetch(tags
->rqs
[tag
]);
627 return tags
->rqs
[tag
];
632 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
634 struct blk_mq_timeout_data
{
636 unsigned int next_set
;
639 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
641 struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
642 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
645 * We know that complete is set at this point. If STARTED isn't set
646 * anymore, then the request isn't active and the "timeout" should
647 * just be ignored. This can happen due to the bitflag ordering.
648 * Timeout first checks if STARTED is set, and if it is, assumes
649 * the request is active. But if we race with completion, then
650 * we both flags will get cleared. So check here again, and ignore
651 * a timeout event with a request that isn't active.
653 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
657 ret
= ops
->timeout(req
, reserved
);
661 __blk_mq_complete_request(req
);
663 case BLK_EH_RESET_TIMER
:
665 blk_clear_rq_complete(req
);
667 case BLK_EH_NOT_HANDLED
:
670 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
675 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
676 struct request
*rq
, void *priv
, bool reserved
)
678 struct blk_mq_timeout_data
*data
= priv
;
680 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
682 * If a request wasn't started before the queue was
683 * marked dying, kill it here or it'll go unnoticed.
685 if (unlikely(blk_queue_dying(rq
->q
))) {
687 blk_mq_end_request(rq
, rq
->errors
);
692 if (time_after_eq(jiffies
, rq
->deadline
)) {
693 if (!blk_mark_rq_complete(rq
))
694 blk_mq_rq_timed_out(rq
, reserved
);
695 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
696 data
->next
= rq
->deadline
;
701 static void blk_mq_timeout_work(struct work_struct
*work
)
703 struct request_queue
*q
=
704 container_of(work
, struct request_queue
, timeout_work
);
705 struct blk_mq_timeout_data data
= {
711 /* A deadlock might occur if a request is stuck requiring a
712 * timeout at the same time a queue freeze is waiting
713 * completion, since the timeout code would not be able to
714 * acquire the queue reference here.
716 * That's why we don't use blk_queue_enter here; instead, we use
717 * percpu_ref_tryget directly, because we need to be able to
718 * obtain a reference even in the short window between the queue
719 * starting to freeze, by dropping the first reference in
720 * blk_mq_freeze_queue_start, and the moment the last request is
721 * consumed, marked by the instant q_usage_counter reaches
724 if (!percpu_ref_tryget(&q
->q_usage_counter
))
727 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
730 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
731 mod_timer(&q
->timeout
, data
.next
);
733 struct blk_mq_hw_ctx
*hctx
;
735 queue_for_each_hw_ctx(q
, hctx
, i
) {
736 /* the hctx may be unmapped, so check it here */
737 if (blk_mq_hw_queue_mapped(hctx
))
738 blk_mq_tag_idle(hctx
);
745 * Reverse check our software queue for entries that we could potentially
746 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
747 * too much time checking for merges.
749 static bool blk_mq_attempt_merge(struct request_queue
*q
,
750 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
755 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
761 if (!blk_rq_merge_ok(rq
, bio
))
764 el_ret
= blk_try_merge(rq
, bio
);
765 if (el_ret
== ELEVATOR_BACK_MERGE
) {
766 if (bio_attempt_back_merge(q
, rq
, bio
)) {
771 } else if (el_ret
== ELEVATOR_FRONT_MERGE
) {
772 if (bio_attempt_front_merge(q
, rq
, bio
)) {
783 struct flush_busy_ctx_data
{
784 struct blk_mq_hw_ctx
*hctx
;
785 struct list_head
*list
;
788 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
790 struct flush_busy_ctx_data
*flush_data
= data
;
791 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
792 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
794 sbitmap_clear_bit(sb
, bitnr
);
795 spin_lock(&ctx
->lock
);
796 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
797 spin_unlock(&ctx
->lock
);
802 * Process software queues that have been marked busy, splicing them
803 * to the for-dispatch
805 static void flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
807 struct flush_busy_ctx_data data
= {
812 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
815 static inline unsigned int queued_to_index(unsigned int queued
)
820 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
824 * Run this hardware queue, pulling any software queues mapped to it in.
825 * Note that this function currently has various problems around ordering
826 * of IO. In particular, we'd like FIFO behaviour on handling existing
827 * items on the hctx->dispatch list. Ignore that for now.
829 static void blk_mq_process_rq_list(struct blk_mq_hw_ctx
*hctx
)
831 struct request_queue
*q
= hctx
->queue
;
834 LIST_HEAD(driver_list
);
835 struct list_head
*dptr
;
838 if (unlikely(blk_mq_hctx_stopped(hctx
)))
844 * Touch any software queue that has pending entries.
846 flush_busy_ctxs(hctx
, &rq_list
);
849 * If we have previous entries on our dispatch list, grab them
850 * and stuff them at the front for more fair dispatch.
852 if (!list_empty_careful(&hctx
->dispatch
)) {
853 spin_lock(&hctx
->lock
);
854 if (!list_empty(&hctx
->dispatch
))
855 list_splice_init(&hctx
->dispatch
, &rq_list
);
856 spin_unlock(&hctx
->lock
);
860 * Start off with dptr being NULL, so we start the first request
861 * immediately, even if we have more pending.
866 * Now process all the entries, sending them to the driver.
869 while (!list_empty(&rq_list
)) {
870 struct blk_mq_queue_data bd
;
873 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
874 list_del_init(&rq
->queuelist
);
878 bd
.last
= list_empty(&rq_list
);
880 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
882 case BLK_MQ_RQ_QUEUE_OK
:
885 case BLK_MQ_RQ_QUEUE_BUSY
:
886 list_add(&rq
->queuelist
, &rq_list
);
887 __blk_mq_requeue_request(rq
);
890 pr_err("blk-mq: bad return on queue: %d\n", ret
);
891 case BLK_MQ_RQ_QUEUE_ERROR
:
893 blk_mq_end_request(rq
, rq
->errors
);
897 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
901 * We've done the first request. If we have more than 1
902 * left in the list, set dptr to defer issue.
904 if (!dptr
&& rq_list
.next
!= rq_list
.prev
)
908 hctx
->dispatched
[queued_to_index(queued
)]++;
911 * Any items that need requeuing? Stuff them into hctx->dispatch,
912 * that is where we will continue on next queue run.
914 if (!list_empty(&rq_list
)) {
915 spin_lock(&hctx
->lock
);
916 list_splice(&rq_list
, &hctx
->dispatch
);
917 spin_unlock(&hctx
->lock
);
919 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
920 * it's possible the queue is stopped and restarted again
921 * before this. Queue restart will dispatch requests. And since
922 * requests in rq_list aren't added into hctx->dispatch yet,
923 * the requests in rq_list might get lost.
925 * blk_mq_run_hw_queue() already checks the STOPPED bit
927 blk_mq_run_hw_queue(hctx
, true);
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_queues(struct request_queue
*q
, bool async
)
1068 struct blk_mq_hw_ctx
*hctx
;
1071 queue_for_each_hw_ctx(q
, hctx
, i
) {
1072 if (!blk_mq_hctx_stopped(hctx
))
1075 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1076 blk_mq_run_hw_queue(hctx
, async
);
1079 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1081 static void blk_mq_run_work_fn(struct work_struct
*work
)
1083 struct blk_mq_hw_ctx
*hctx
;
1085 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1087 __blk_mq_run_hw_queue(hctx
);
1090 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1092 struct blk_mq_hw_ctx
*hctx
;
1094 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1096 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1097 __blk_mq_run_hw_queue(hctx
);
1100 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1102 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1105 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1106 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1108 EXPORT_SYMBOL(blk_mq_delay_queue
);
1110 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1114 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1116 trace_block_rq_insert(hctx
->queue
, rq
);
1119 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1121 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1124 static void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
,
1125 struct request
*rq
, bool at_head
)
1127 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1129 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1130 blk_mq_hctx_mark_pending(hctx
, ctx
);
1133 void blk_mq_insert_request(struct request
*rq
, bool at_head
, bool run_queue
,
1136 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1137 struct request_queue
*q
= rq
->q
;
1138 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1140 spin_lock(&ctx
->lock
);
1141 __blk_mq_insert_request(hctx
, rq
, at_head
);
1142 spin_unlock(&ctx
->lock
);
1145 blk_mq_run_hw_queue(hctx
, async
);
1148 static void blk_mq_insert_requests(struct request_queue
*q
,
1149 struct blk_mq_ctx
*ctx
,
1150 struct list_head
*list
,
1155 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1157 trace_block_unplug(q
, depth
, !from_schedule
);
1160 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1163 spin_lock(&ctx
->lock
);
1164 while (!list_empty(list
)) {
1167 rq
= list_first_entry(list
, struct request
, queuelist
);
1168 BUG_ON(rq
->mq_ctx
!= ctx
);
1169 list_del_init(&rq
->queuelist
);
1170 __blk_mq_insert_req_list(hctx
, rq
, false);
1172 blk_mq_hctx_mark_pending(hctx
, ctx
);
1173 spin_unlock(&ctx
->lock
);
1175 blk_mq_run_hw_queue(hctx
, from_schedule
);
1178 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1180 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1181 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1183 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1184 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1185 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1188 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1190 struct blk_mq_ctx
*this_ctx
;
1191 struct request_queue
*this_q
;
1194 LIST_HEAD(ctx_list
);
1197 list_splice_init(&plug
->mq_list
, &list
);
1199 list_sort(NULL
, &list
, plug_ctx_cmp
);
1205 while (!list_empty(&list
)) {
1206 rq
= list_entry_rq(list
.next
);
1207 list_del_init(&rq
->queuelist
);
1209 if (rq
->mq_ctx
!= this_ctx
) {
1211 blk_mq_insert_requests(this_q
, this_ctx
,
1216 this_ctx
= rq
->mq_ctx
;
1222 list_add_tail(&rq
->queuelist
, &ctx_list
);
1226 * If 'this_ctx' is set, we know we have entries to complete
1227 * on 'ctx_list'. Do those.
1230 blk_mq_insert_requests(this_q
, this_ctx
, &ctx_list
, depth
,
1235 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1237 init_request_from_bio(rq
, bio
);
1239 blk_account_io_start(rq
, 1);
1242 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1244 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1245 !blk_queue_nomerges(hctx
->queue
);
1248 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1249 struct blk_mq_ctx
*ctx
,
1250 struct request
*rq
, struct bio
*bio
)
1252 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1253 blk_mq_bio_to_request(rq
, bio
);
1254 spin_lock(&ctx
->lock
);
1256 __blk_mq_insert_request(hctx
, rq
, false);
1257 spin_unlock(&ctx
->lock
);
1260 struct request_queue
*q
= hctx
->queue
;
1262 spin_lock(&ctx
->lock
);
1263 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1264 blk_mq_bio_to_request(rq
, bio
);
1268 spin_unlock(&ctx
->lock
);
1269 __blk_mq_free_request(hctx
, ctx
, rq
);
1274 static struct request
*blk_mq_map_request(struct request_queue
*q
,
1276 struct blk_mq_alloc_data
*data
)
1278 struct blk_mq_hw_ctx
*hctx
;
1279 struct blk_mq_ctx
*ctx
;
1282 blk_queue_enter_live(q
);
1283 ctx
= blk_mq_get_ctx(q
);
1284 hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
1286 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1287 blk_mq_set_alloc_data(data
, q
, 0, ctx
, hctx
);
1288 rq
= __blk_mq_alloc_request(data
, bio
->bi_opf
);
1290 data
->hctx
->queued
++;
1294 static void blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
)
1297 struct request_queue
*q
= rq
->q
;
1298 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, rq
->mq_ctx
->cpu
);
1299 struct blk_mq_queue_data bd
= {
1304 blk_qc_t new_cookie
= blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
);
1306 if (blk_mq_hctx_stopped(hctx
))
1310 * For OK queue, we are done. For error, kill it. Any other
1311 * error (busy), just add it to our list as we previously
1314 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1315 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1316 *cookie
= new_cookie
;
1320 __blk_mq_requeue_request(rq
);
1322 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1323 *cookie
= BLK_QC_T_NONE
;
1325 blk_mq_end_request(rq
, rq
->errors
);
1330 blk_mq_insert_request(rq
, false, true, true);
1334 * Multiple hardware queue variant. This will not use per-process plugs,
1335 * but will attempt to bypass the hctx queueing if we can go straight to
1336 * hardware for SYNC IO.
1338 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1340 const int is_sync
= op_is_sync(bio
->bi_opf
);
1341 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1342 struct blk_mq_alloc_data data
;
1344 unsigned int request_count
= 0, srcu_idx
;
1345 struct blk_plug
*plug
;
1346 struct request
*same_queue_rq
= NULL
;
1348 unsigned int wb_acct
;
1350 blk_queue_bounce(q
, &bio
);
1352 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1354 return BLK_QC_T_NONE
;
1357 blk_queue_split(q
, &bio
, q
->bio_split
);
1359 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1360 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1361 return BLK_QC_T_NONE
;
1363 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1365 rq
= blk_mq_map_request(q
, bio
, &data
);
1366 if (unlikely(!rq
)) {
1367 __wbt_done(q
->rq_wb
, wb_acct
);
1368 return BLK_QC_T_NONE
;
1371 wbt_track(&rq
->issue_stat
, wb_acct
);
1373 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1375 if (unlikely(is_flush_fua
)) {
1376 blk_mq_bio_to_request(rq
, bio
);
1377 blk_insert_flush(rq
);
1381 plug
= current
->plug
;
1383 * If the driver supports defer issued based on 'last', then
1384 * queue it up like normal since we can potentially save some
1387 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1388 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1389 struct request
*old_rq
= NULL
;
1391 blk_mq_bio_to_request(rq
, bio
);
1394 * We do limited plugging. If the bio can be merged, do that.
1395 * Otherwise the existing request in the plug list will be
1396 * issued. So the plug list will have one request at most
1400 * The plug list might get flushed before this. If that
1401 * happens, same_queue_rq is invalid and plug list is
1404 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1405 old_rq
= same_queue_rq
;
1406 list_del_init(&old_rq
->queuelist
);
1408 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1409 } else /* is_sync */
1411 blk_mq_put_ctx(data
.ctx
);
1415 if (!(data
.hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1417 blk_mq_try_issue_directly(old_rq
, &cookie
);
1420 srcu_idx
= srcu_read_lock(&data
.hctx
->queue_rq_srcu
);
1421 blk_mq_try_issue_directly(old_rq
, &cookie
);
1422 srcu_read_unlock(&data
.hctx
->queue_rq_srcu
, srcu_idx
);
1427 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1429 * For a SYNC request, send it to the hardware immediately. For
1430 * an ASYNC request, just ensure that we run it later on. The
1431 * latter allows for merging opportunities and more efficient
1435 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1437 blk_mq_put_ctx(data
.ctx
);
1443 * Single hardware queue variant. This will attempt to use any per-process
1444 * plug for merging and IO deferral.
1446 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1448 const int is_sync
= op_is_sync(bio
->bi_opf
);
1449 const int is_flush_fua
= bio
->bi_opf
& (REQ_PREFLUSH
| REQ_FUA
);
1450 struct blk_plug
*plug
;
1451 unsigned int request_count
= 0;
1452 struct blk_mq_alloc_data data
;
1455 unsigned int wb_acct
;
1457 blk_queue_bounce(q
, &bio
);
1459 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1461 return BLK_QC_T_NONE
;
1464 blk_queue_split(q
, &bio
, q
->bio_split
);
1466 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1467 if (blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1468 return BLK_QC_T_NONE
;
1470 request_count
= blk_plug_queued_count(q
);
1472 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1474 rq
= blk_mq_map_request(q
, bio
, &data
);
1475 if (unlikely(!rq
)) {
1476 __wbt_done(q
->rq_wb
, wb_acct
);
1477 return BLK_QC_T_NONE
;
1480 wbt_track(&rq
->issue_stat
, wb_acct
);
1482 cookie
= blk_tag_to_qc_t(rq
->tag
, data
.hctx
->queue_num
);
1484 if (unlikely(is_flush_fua
)) {
1485 blk_mq_bio_to_request(rq
, bio
);
1486 blk_insert_flush(rq
);
1491 * A task plug currently exists. Since this is completely lockless,
1492 * utilize that to temporarily store requests until the task is
1493 * either done or scheduled away.
1495 plug
= current
->plug
;
1497 struct request
*last
= NULL
;
1499 blk_mq_bio_to_request(rq
, bio
);
1502 * @request_count may become stale because of schedule
1503 * out, so check the list again.
1505 if (list_empty(&plug
->mq_list
))
1508 trace_block_plug(q
);
1510 last
= list_entry_rq(plug
->mq_list
.prev
);
1512 blk_mq_put_ctx(data
.ctx
);
1514 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1515 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1516 blk_flush_plug_list(plug
, false);
1517 trace_block_plug(q
);
1520 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1524 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1526 * For a SYNC request, send it to the hardware immediately. For
1527 * an ASYNC request, just ensure that we run it later on. The
1528 * latter allows for merging opportunities and more efficient
1532 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1535 blk_mq_put_ctx(data
.ctx
);
1539 static void blk_mq_free_rq_map(struct blk_mq_tag_set
*set
,
1540 struct blk_mq_tags
*tags
, unsigned int hctx_idx
)
1544 if (tags
->rqs
&& set
->ops
->exit_request
) {
1547 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1550 set
->ops
->exit_request(set
->driver_data
, tags
->rqs
[i
],
1552 tags
->rqs
[i
] = NULL
;
1556 while (!list_empty(&tags
->page_list
)) {
1557 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1558 list_del_init(&page
->lru
);
1560 * Remove kmemleak object previously allocated in
1561 * blk_mq_init_rq_map().
1563 kmemleak_free(page_address(page
));
1564 __free_pages(page
, page
->private);
1569 blk_mq_free_tags(tags
);
1572 static size_t order_to_size(unsigned int order
)
1574 return (size_t)PAGE_SIZE
<< order
;
1577 static struct blk_mq_tags
*blk_mq_init_rq_map(struct blk_mq_tag_set
*set
,
1578 unsigned int hctx_idx
)
1580 struct blk_mq_tags
*tags
;
1581 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1582 size_t rq_size
, left
;
1584 tags
= blk_mq_init_tags(set
->queue_depth
, set
->reserved_tags
,
1586 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1590 INIT_LIST_HEAD(&tags
->page_list
);
1592 tags
->rqs
= kzalloc_node(set
->queue_depth
* sizeof(struct request
*),
1593 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
,
1596 blk_mq_free_tags(tags
);
1601 * rq_size is the size of the request plus driver payload, rounded
1602 * to the cacheline size
1604 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1606 left
= rq_size
* set
->queue_depth
;
1608 for (i
= 0; i
< set
->queue_depth
; ) {
1609 int this_order
= max_order
;
1614 while (this_order
&& left
< order_to_size(this_order
- 1))
1618 page
= alloc_pages_node(set
->numa_node
,
1619 GFP_KERNEL
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1625 if (order_to_size(this_order
) < rq_size
)
1632 page
->private = this_order
;
1633 list_add_tail(&page
->lru
, &tags
->page_list
);
1635 p
= page_address(page
);
1637 * Allow kmemleak to scan these pages as they contain pointers
1638 * to additional allocations like via ops->init_request().
1640 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_KERNEL
);
1641 entries_per_page
= order_to_size(this_order
) / rq_size
;
1642 to_do
= min(entries_per_page
, set
->queue_depth
- i
);
1643 left
-= to_do
* rq_size
;
1644 for (j
= 0; j
< to_do
; j
++) {
1646 if (set
->ops
->init_request
) {
1647 if (set
->ops
->init_request(set
->driver_data
,
1648 tags
->rqs
[i
], hctx_idx
, i
,
1650 tags
->rqs
[i
] = NULL
;
1662 blk_mq_free_rq_map(set
, tags
, hctx_idx
);
1667 * 'cpu' is going away. splice any existing rq_list entries from this
1668 * software queue to the hw queue dispatch list, and ensure that it
1671 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1673 struct blk_mq_hw_ctx
*hctx
;
1674 struct blk_mq_ctx
*ctx
;
1677 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1678 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1680 spin_lock(&ctx
->lock
);
1681 if (!list_empty(&ctx
->rq_list
)) {
1682 list_splice_init(&ctx
->rq_list
, &tmp
);
1683 blk_mq_hctx_clear_pending(hctx
, ctx
);
1685 spin_unlock(&ctx
->lock
);
1687 if (list_empty(&tmp
))
1690 spin_lock(&hctx
->lock
);
1691 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1692 spin_unlock(&hctx
->lock
);
1694 blk_mq_run_hw_queue(hctx
, true);
1698 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1700 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1704 /* hctx->ctxs will be freed in queue's release handler */
1705 static void blk_mq_exit_hctx(struct request_queue
*q
,
1706 struct blk_mq_tag_set
*set
,
1707 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1709 unsigned flush_start_tag
= set
->queue_depth
;
1711 blk_mq_tag_idle(hctx
);
1713 if (set
->ops
->exit_request
)
1714 set
->ops
->exit_request(set
->driver_data
,
1715 hctx
->fq
->flush_rq
, hctx_idx
,
1716 flush_start_tag
+ hctx_idx
);
1718 if (set
->ops
->exit_hctx
)
1719 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1721 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1722 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1724 blk_mq_remove_cpuhp(hctx
);
1725 blk_free_flush_queue(hctx
->fq
);
1726 sbitmap_free(&hctx
->ctx_map
);
1729 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1730 struct blk_mq_tag_set
*set
, int nr_queue
)
1732 struct blk_mq_hw_ctx
*hctx
;
1735 queue_for_each_hw_ctx(q
, hctx
, i
) {
1738 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1742 static void blk_mq_free_hw_queues(struct request_queue
*q
,
1743 struct blk_mq_tag_set
*set
)
1745 struct blk_mq_hw_ctx
*hctx
;
1748 queue_for_each_hw_ctx(q
, hctx
, i
)
1749 free_cpumask_var(hctx
->cpumask
);
1752 static int blk_mq_init_hctx(struct request_queue
*q
,
1753 struct blk_mq_tag_set
*set
,
1754 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1757 unsigned flush_start_tag
= set
->queue_depth
;
1759 node
= hctx
->numa_node
;
1760 if (node
== NUMA_NO_NODE
)
1761 node
= hctx
->numa_node
= set
->numa_node
;
1763 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1764 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1765 spin_lock_init(&hctx
->lock
);
1766 INIT_LIST_HEAD(&hctx
->dispatch
);
1768 hctx
->queue_num
= hctx_idx
;
1769 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1771 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1773 hctx
->tags
= set
->tags
[hctx_idx
];
1776 * Allocate space for all possible cpus to avoid allocation at
1779 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1782 goto unregister_cpu_notifier
;
1784 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1790 if (set
->ops
->init_hctx
&&
1791 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1794 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1798 if (set
->ops
->init_request
&&
1799 set
->ops
->init_request(set
->driver_data
,
1800 hctx
->fq
->flush_rq
, hctx_idx
,
1801 flush_start_tag
+ hctx_idx
, node
))
1804 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1805 init_srcu_struct(&hctx
->queue_rq_srcu
);
1812 if (set
->ops
->exit_hctx
)
1813 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1815 sbitmap_free(&hctx
->ctx_map
);
1818 unregister_cpu_notifier
:
1819 blk_mq_remove_cpuhp(hctx
);
1823 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1824 unsigned int nr_hw_queues
)
1828 for_each_possible_cpu(i
) {
1829 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1830 struct blk_mq_hw_ctx
*hctx
;
1832 memset(__ctx
, 0, sizeof(*__ctx
));
1834 spin_lock_init(&__ctx
->lock
);
1835 INIT_LIST_HEAD(&__ctx
->rq_list
);
1837 blk_stat_init(&__ctx
->stat
[BLK_STAT_READ
]);
1838 blk_stat_init(&__ctx
->stat
[BLK_STAT_WRITE
]);
1840 /* If the cpu isn't online, the cpu is mapped to first hctx */
1844 hctx
= blk_mq_map_queue(q
, i
);
1847 * Set local node, IFF we have more than one hw queue. If
1848 * not, we remain on the home node of the device
1850 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1851 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1855 static void blk_mq_map_swqueue(struct request_queue
*q
,
1856 const struct cpumask
*online_mask
)
1859 struct blk_mq_hw_ctx
*hctx
;
1860 struct blk_mq_ctx
*ctx
;
1861 struct blk_mq_tag_set
*set
= q
->tag_set
;
1864 * Avoid others reading imcomplete hctx->cpumask through sysfs
1866 mutex_lock(&q
->sysfs_lock
);
1868 queue_for_each_hw_ctx(q
, hctx
, i
) {
1869 cpumask_clear(hctx
->cpumask
);
1874 * Map software to hardware queues
1876 for_each_possible_cpu(i
) {
1877 /* If the cpu isn't online, the cpu is mapped to first hctx */
1878 if (!cpumask_test_cpu(i
, online_mask
))
1881 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1882 hctx
= blk_mq_map_queue(q
, i
);
1884 cpumask_set_cpu(i
, hctx
->cpumask
);
1885 ctx
->index_hw
= hctx
->nr_ctx
;
1886 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
1889 mutex_unlock(&q
->sysfs_lock
);
1891 queue_for_each_hw_ctx(q
, hctx
, i
) {
1893 * If no software queues are mapped to this hardware queue,
1894 * disable it and free the request entries.
1896 if (!hctx
->nr_ctx
) {
1898 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
1899 set
->tags
[i
] = NULL
;
1905 /* unmapped hw queue can be remapped after CPU topo changed */
1907 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
1908 hctx
->tags
= set
->tags
[i
];
1909 WARN_ON(!hctx
->tags
);
1912 * Set the map size to the number of mapped software queues.
1913 * This is more accurate and more efficient than looping
1914 * over all possibly mapped software queues.
1916 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
1919 * Initialize batch roundrobin counts
1921 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
1922 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1926 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
1928 struct blk_mq_hw_ctx
*hctx
;
1931 queue_for_each_hw_ctx(q
, hctx
, i
) {
1933 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
1935 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1939 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
1941 struct request_queue
*q
;
1943 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
1944 blk_mq_freeze_queue(q
);
1945 queue_set_hctx_shared(q
, shared
);
1946 blk_mq_unfreeze_queue(q
);
1950 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
1952 struct blk_mq_tag_set
*set
= q
->tag_set
;
1954 mutex_lock(&set
->tag_list_lock
);
1955 list_del_init(&q
->tag_set_list
);
1956 if (list_is_singular(&set
->tag_list
)) {
1957 /* just transitioned to unshared */
1958 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
1959 /* update existing queue */
1960 blk_mq_update_tag_set_depth(set
, false);
1962 mutex_unlock(&set
->tag_list_lock
);
1965 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
1966 struct request_queue
*q
)
1970 mutex_lock(&set
->tag_list_lock
);
1972 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1973 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1974 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
1975 /* update existing queue */
1976 blk_mq_update_tag_set_depth(set
, true);
1978 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
1979 queue_set_hctx_shared(q
, true);
1980 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
1982 mutex_unlock(&set
->tag_list_lock
);
1986 * It is the actual release handler for mq, but we do it from
1987 * request queue's release handler for avoiding use-after-free
1988 * and headache because q->mq_kobj shouldn't have been introduced,
1989 * but we can't group ctx/kctx kobj without it.
1991 void blk_mq_release(struct request_queue
*q
)
1993 struct blk_mq_hw_ctx
*hctx
;
1996 /* hctx kobj stays in hctx */
1997 queue_for_each_hw_ctx(q
, hctx
, i
) {
2006 kfree(q
->queue_hw_ctx
);
2008 /* ctx kobj stays in queue_ctx */
2009 free_percpu(q
->queue_ctx
);
2012 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2014 struct request_queue
*uninit_q
, *q
;
2016 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2018 return ERR_PTR(-ENOMEM
);
2020 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2022 blk_cleanup_queue(uninit_q
);
2026 EXPORT_SYMBOL(blk_mq_init_queue
);
2028 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2029 struct request_queue
*q
)
2032 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2034 blk_mq_sysfs_unregister(q
);
2035 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2041 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2042 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2047 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2054 atomic_set(&hctxs
[i
]->nr_active
, 0);
2055 hctxs
[i
]->numa_node
= node
;
2056 hctxs
[i
]->queue_num
= i
;
2058 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2059 free_cpumask_var(hctxs
[i
]->cpumask
);
2064 blk_mq_hctx_kobj_init(hctxs
[i
]);
2066 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2067 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2071 blk_mq_free_rq_map(set
, hctx
->tags
, j
);
2072 set
->tags
[j
] = NULL
;
2074 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2075 free_cpumask_var(hctx
->cpumask
);
2076 kobject_put(&hctx
->kobj
);
2083 q
->nr_hw_queues
= i
;
2084 blk_mq_sysfs_register(q
);
2087 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2088 struct request_queue
*q
)
2090 /* mark the queue as mq asap */
2091 q
->mq_ops
= set
->ops
;
2093 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2097 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2098 GFP_KERNEL
, set
->numa_node
);
2099 if (!q
->queue_hw_ctx
)
2102 q
->mq_map
= set
->mq_map
;
2104 blk_mq_realloc_hw_ctxs(set
, q
);
2105 if (!q
->nr_hw_queues
)
2108 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2109 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2111 q
->nr_queues
= nr_cpu_ids
;
2113 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2115 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2116 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2118 q
->sg_reserved_size
= INT_MAX
;
2120 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2121 INIT_LIST_HEAD(&q
->requeue_list
);
2122 spin_lock_init(&q
->requeue_lock
);
2124 if (q
->nr_hw_queues
> 1)
2125 blk_queue_make_request(q
, blk_mq_make_request
);
2127 blk_queue_make_request(q
, blk_sq_make_request
);
2130 * Do this after blk_queue_make_request() overrides it...
2132 q
->nr_requests
= set
->queue_depth
;
2134 if (set
->ops
->complete
)
2135 blk_queue_softirq_done(q
, set
->ops
->complete
);
2137 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2140 mutex_lock(&all_q_mutex
);
2142 list_add_tail(&q
->all_q_node
, &all_q_list
);
2143 blk_mq_add_queue_tag_set(set
, q
);
2144 blk_mq_map_swqueue(q
, cpu_online_mask
);
2146 mutex_unlock(&all_q_mutex
);
2152 kfree(q
->queue_hw_ctx
);
2154 free_percpu(q
->queue_ctx
);
2157 return ERR_PTR(-ENOMEM
);
2159 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2161 void blk_mq_free_queue(struct request_queue
*q
)
2163 struct blk_mq_tag_set
*set
= q
->tag_set
;
2165 mutex_lock(&all_q_mutex
);
2166 list_del_init(&q
->all_q_node
);
2167 mutex_unlock(&all_q_mutex
);
2171 blk_mq_del_queue_tag_set(q
);
2173 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2174 blk_mq_free_hw_queues(q
, set
);
2177 /* Basically redo blk_mq_init_queue with queue frozen */
2178 static void blk_mq_queue_reinit(struct request_queue
*q
,
2179 const struct cpumask
*online_mask
)
2181 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2183 blk_mq_sysfs_unregister(q
);
2186 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2187 * we should change hctx numa_node according to new topology (this
2188 * involves free and re-allocate memory, worthy doing?)
2191 blk_mq_map_swqueue(q
, online_mask
);
2193 blk_mq_sysfs_register(q
);
2197 * New online cpumask which is going to be set in this hotplug event.
2198 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2199 * one-by-one and dynamically allocating this could result in a failure.
2201 static struct cpumask cpuhp_online_new
;
2203 static void blk_mq_queue_reinit_work(void)
2205 struct request_queue
*q
;
2207 mutex_lock(&all_q_mutex
);
2209 * We need to freeze and reinit all existing queues. Freezing
2210 * involves synchronous wait for an RCU grace period and doing it
2211 * one by one may take a long time. Start freezing all queues in
2212 * one swoop and then wait for the completions so that freezing can
2213 * take place in parallel.
2215 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2216 blk_mq_freeze_queue_start(q
);
2217 list_for_each_entry(q
, &all_q_list
, all_q_node
) {
2218 blk_mq_freeze_queue_wait(q
);
2221 * timeout handler can't touch hw queue during the
2224 del_timer_sync(&q
->timeout
);
2227 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2228 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2230 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2231 blk_mq_unfreeze_queue(q
);
2233 mutex_unlock(&all_q_mutex
);
2236 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2238 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2239 blk_mq_queue_reinit_work();
2244 * Before hotadded cpu starts handling requests, new mappings must be
2245 * established. Otherwise, these requests in hw queue might never be
2248 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2249 * for CPU0, and ctx1 for CPU1).
2251 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2252 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2254 * And then while running hw queue, flush_busy_ctxs() finds bit0 is set in
2255 * pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2256 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list
2259 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2261 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2262 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2263 blk_mq_queue_reinit_work();
2267 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2271 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2272 set
->tags
[i
] = blk_mq_init_rq_map(set
, i
);
2281 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2287 * Allocate the request maps associated with this tag_set. Note that this
2288 * may reduce the depth asked for, if memory is tight. set->queue_depth
2289 * will be updated to reflect the allocated depth.
2291 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2296 depth
= set
->queue_depth
;
2298 err
= __blk_mq_alloc_rq_maps(set
);
2302 set
->queue_depth
>>= 1;
2303 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2307 } while (set
->queue_depth
);
2309 if (!set
->queue_depth
|| err
) {
2310 pr_err("blk-mq: failed to allocate request map\n");
2314 if (depth
!= set
->queue_depth
)
2315 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2316 depth
, set
->queue_depth
);
2322 * Alloc a tag set to be associated with one or more request queues.
2323 * May fail with EINVAL for various error conditions. May adjust the
2324 * requested depth down, if if it too large. In that case, the set
2325 * value will be stored in set->queue_depth.
2327 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2331 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2333 if (!set
->nr_hw_queues
)
2335 if (!set
->queue_depth
)
2337 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2340 if (!set
->ops
->queue_rq
)
2343 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2344 pr_info("blk-mq: reduced tag depth to %u\n",
2346 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2350 * If a crashdump is active, then we are potentially in a very
2351 * memory constrained environment. Limit us to 1 queue and
2352 * 64 tags to prevent using too much memory.
2354 if (is_kdump_kernel()) {
2355 set
->nr_hw_queues
= 1;
2356 set
->queue_depth
= min(64U, set
->queue_depth
);
2359 * There is no use for more h/w queues than cpus.
2361 if (set
->nr_hw_queues
> nr_cpu_ids
)
2362 set
->nr_hw_queues
= nr_cpu_ids
;
2364 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2365 GFP_KERNEL
, set
->numa_node
);
2370 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2371 GFP_KERNEL
, set
->numa_node
);
2375 if (set
->ops
->map_queues
)
2376 ret
= set
->ops
->map_queues(set
);
2378 ret
= blk_mq_map_queues(set
);
2380 goto out_free_mq_map
;
2382 ret
= blk_mq_alloc_rq_maps(set
);
2384 goto out_free_mq_map
;
2386 mutex_init(&set
->tag_list_lock
);
2387 INIT_LIST_HEAD(&set
->tag_list
);
2399 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2401 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2405 for (i
= 0; i
< nr_cpu_ids
; i
++) {
2407 blk_mq_free_rq_map(set
, set
->tags
[i
], i
);
2416 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2418 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2420 struct blk_mq_tag_set
*set
= q
->tag_set
;
2421 struct blk_mq_hw_ctx
*hctx
;
2424 if (!set
|| nr
> set
->queue_depth
)
2428 queue_for_each_hw_ctx(q
, hctx
, i
) {
2431 ret
= blk_mq_tag_update_depth(hctx
->tags
, nr
);
2437 q
->nr_requests
= nr
;
2442 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2444 struct request_queue
*q
;
2446 if (nr_hw_queues
> nr_cpu_ids
)
2447 nr_hw_queues
= nr_cpu_ids
;
2448 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2451 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2452 blk_mq_freeze_queue(q
);
2454 set
->nr_hw_queues
= nr_hw_queues
;
2455 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2456 blk_mq_realloc_hw_ctxs(set
, q
);
2458 if (q
->nr_hw_queues
> 1)
2459 blk_queue_make_request(q
, blk_mq_make_request
);
2461 blk_queue_make_request(q
, blk_sq_make_request
);
2463 blk_mq_queue_reinit(q
, cpu_online_mask
);
2466 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2467 blk_mq_unfreeze_queue(q
);
2469 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2471 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2473 struct request_queue
*q
= hctx
->queue
;
2476 hctx
->poll_considered
++;
2478 state
= current
->state
;
2479 while (!need_resched()) {
2482 hctx
->poll_invoked
++;
2484 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2486 hctx
->poll_success
++;
2487 set_current_state(TASK_RUNNING
);
2491 if (signal_pending_state(state
, current
))
2492 set_current_state(TASK_RUNNING
);
2494 if (current
->state
== TASK_RUNNING
)
2504 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2506 struct blk_mq_hw_ctx
*hctx
;
2507 struct blk_plug
*plug
;
2510 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2511 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2514 plug
= current
->plug
;
2516 blk_flush_plug_list(plug
, false);
2518 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2519 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2521 return __blk_mq_poll(hctx
, rq
);
2523 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2525 void blk_mq_disable_hotplug(void)
2527 mutex_lock(&all_q_mutex
);
2530 void blk_mq_enable_hotplug(void)
2532 mutex_unlock(&all_q_mutex
);
2535 static int __init
blk_mq_init(void)
2537 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2538 blk_mq_hctx_notify_dead
);
2540 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2541 blk_mq_queue_reinit_prepare
,
2542 blk_mq_queue_reinit_dead
);
2545 subsys_initcall(blk_mq_init
);