2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
29 #include <trace/events/block.h>
31 #include <linux/blk-mq.h>
34 #include "blk-mq-tag.h"
37 #include "blk-mq-sched.h"
39 static DEFINE_MUTEX(all_q_mutex
);
40 static LIST_HEAD(all_q_list
);
43 * Check if any of the ctx's have pending work in this hardware queue
45 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
47 return sbitmap_any_bit_set(&hctx
->ctx_map
) ||
48 !list_empty_careful(&hctx
->dispatch
) ||
49 blk_mq_sched_has_work(hctx
);
53 * Mark this ctx as having pending work in this hardware queue
55 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
56 struct blk_mq_ctx
*ctx
)
58 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
59 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
62 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
63 struct blk_mq_ctx
*ctx
)
65 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
68 void blk_mq_freeze_queue_start(struct request_queue
*q
)
72 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
73 if (freeze_depth
== 1) {
74 percpu_ref_kill(&q
->q_usage_counter
);
75 blk_mq_run_hw_queues(q
, false);
78 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start
);
80 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
82 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
84 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
86 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
87 unsigned long timeout
)
89 return wait_event_timeout(q
->mq_freeze_wq
,
90 percpu_ref_is_zero(&q
->q_usage_counter
),
93 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
96 * Guarantee no request is in use, so we can change any data structure of
97 * the queue afterward.
99 void blk_freeze_queue(struct request_queue
*q
)
102 * In the !blk_mq case we are only calling this to kill the
103 * q_usage_counter, otherwise this increases the freeze depth
104 * and waits for it to return to zero. For this reason there is
105 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
106 * exported to drivers as the only user for unfreeze is blk_mq.
108 blk_mq_freeze_queue_start(q
);
109 blk_mq_freeze_queue_wait(q
);
112 void blk_mq_freeze_queue(struct request_queue
*q
)
115 * ...just an alias to keep freeze and unfreeze actions balanced
116 * in the blk_mq_* namespace
120 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
122 void blk_mq_unfreeze_queue(struct request_queue
*q
)
126 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
127 WARN_ON_ONCE(freeze_depth
< 0);
129 percpu_ref_reinit(&q
->q_usage_counter
);
130 wake_up_all(&q
->mq_freeze_wq
);
133 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
136 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
139 * Note: this function does not prevent that the struct request end_io()
140 * callback function is invoked. Additionally, it is not prevented that
141 * new queue_rq() calls occur unless the queue has been stopped first.
143 void blk_mq_quiesce_queue(struct request_queue
*q
)
145 struct blk_mq_hw_ctx
*hctx
;
149 blk_mq_stop_hw_queues(q
);
151 queue_for_each_hw_ctx(q
, hctx
, i
) {
152 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
153 synchronize_srcu(&hctx
->queue_rq_srcu
);
160 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
162 void blk_mq_wake_waiters(struct request_queue
*q
)
164 struct blk_mq_hw_ctx
*hctx
;
167 queue_for_each_hw_ctx(q
, hctx
, i
)
168 if (blk_mq_hw_queue_mapped(hctx
))
169 blk_mq_tag_wakeup_all(hctx
->tags
, true);
172 * If we are called because the queue has now been marked as
173 * dying, we need to ensure that processes currently waiting on
174 * the queue are notified as well.
176 wake_up_all(&q
->mq_freeze_wq
);
179 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
181 return blk_mq_has_free_tags(hctx
->tags
);
183 EXPORT_SYMBOL(blk_mq_can_queue
);
185 void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
186 struct request
*rq
, unsigned int op
)
188 INIT_LIST_HEAD(&rq
->queuelist
);
189 /* csd/requeue_work/fifo_time is initialized before use */
193 if (blk_queue_io_stat(q
))
194 rq
->rq_flags
|= RQF_IO_STAT
;
195 /* do not touch atomic flags, it needs atomic ops against the timer */
197 INIT_HLIST_NODE(&rq
->hash
);
198 RB_CLEAR_NODE(&rq
->rb_node
);
201 rq
->start_time
= jiffies
;
202 #ifdef CONFIG_BLK_CGROUP
204 set_start_time_ns(rq
);
205 rq
->io_start_time_ns
= 0;
207 rq
->nr_phys_segments
= 0;
208 #if defined(CONFIG_BLK_DEV_INTEGRITY)
209 rq
->nr_integrity_segments
= 0;
212 /* tag was already set */
216 INIT_LIST_HEAD(&rq
->timeout_list
);
220 rq
->end_io_data
= NULL
;
223 ctx
->rq_dispatched
[op_is_sync(op
)]++;
225 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init
);
227 struct request
*__blk_mq_alloc_request(struct blk_mq_alloc_data
*data
,
233 tag
= blk_mq_get_tag(data
);
234 if (tag
!= BLK_MQ_TAG_FAIL
) {
235 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
237 rq
= tags
->static_rqs
[tag
];
239 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
241 rq
->internal_tag
= tag
;
243 if (blk_mq_tag_busy(data
->hctx
)) {
244 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
245 atomic_inc(&data
->hctx
->nr_active
);
248 rq
->internal_tag
= -1;
249 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
252 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
);
258 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request
);
260 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
263 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
267 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
271 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
273 blk_mq_put_ctx(alloc_data
.ctx
);
277 return ERR_PTR(-EWOULDBLOCK
);
280 rq
->__sector
= (sector_t
) -1;
281 rq
->bio
= rq
->biotail
= NULL
;
284 EXPORT_SYMBOL(blk_mq_alloc_request
);
286 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
287 unsigned int flags
, unsigned int hctx_idx
)
289 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
295 * If the tag allocator sleeps we could get an allocation for a
296 * different hardware context. No need to complicate the low level
297 * allocator for this for the rare use case of a command tied to
300 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
301 return ERR_PTR(-EINVAL
);
303 if (hctx_idx
>= q
->nr_hw_queues
)
304 return ERR_PTR(-EIO
);
306 ret
= blk_queue_enter(q
, true);
311 * Check if the hardware context is actually mapped to anything.
312 * If not tell the caller that it should skip this queue.
314 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
315 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
317 return ERR_PTR(-EXDEV
);
319 cpu
= cpumask_first(alloc_data
.hctx
->cpumask
);
320 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
322 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
324 blk_mq_put_ctx(alloc_data
.ctx
);
328 return ERR_PTR(-EWOULDBLOCK
);
332 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
334 void __blk_mq_finish_request(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
337 const int sched_tag
= rq
->internal_tag
;
338 struct request_queue
*q
= rq
->q
;
340 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
341 atomic_dec(&hctx
->nr_active
);
343 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
346 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
347 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
349 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
351 blk_mq_sched_completed_request(hctx
, rq
);
352 blk_mq_sched_restart_queues(hctx
);
356 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx
*hctx
,
359 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
361 ctx
->rq_completed
[rq_is_sync(rq
)]++;
362 __blk_mq_finish_request(hctx
, ctx
, rq
);
365 void blk_mq_finish_request(struct request
*rq
)
367 blk_mq_finish_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
370 void blk_mq_free_request(struct request
*rq
)
372 blk_mq_sched_put_request(rq
);
374 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
376 inline void __blk_mq_end_request(struct request
*rq
, int error
)
378 blk_account_io_done(rq
);
381 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
382 rq
->end_io(rq
, error
);
384 if (unlikely(blk_bidi_rq(rq
)))
385 blk_mq_free_request(rq
->next_rq
);
386 blk_mq_free_request(rq
);
389 EXPORT_SYMBOL(__blk_mq_end_request
);
391 void blk_mq_end_request(struct request
*rq
, int error
)
393 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
395 __blk_mq_end_request(rq
, error
);
397 EXPORT_SYMBOL(blk_mq_end_request
);
399 static void __blk_mq_complete_request_remote(void *data
)
401 struct request
*rq
= data
;
403 rq
->q
->softirq_done_fn(rq
);
406 static void blk_mq_ipi_complete_request(struct request
*rq
)
408 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
412 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
413 rq
->q
->softirq_done_fn(rq
);
418 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
419 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
421 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
422 rq
->csd
.func
= __blk_mq_complete_request_remote
;
425 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
427 rq
->q
->softirq_done_fn(rq
);
432 static void blk_mq_stat_add(struct request
*rq
)
434 if (rq
->rq_flags
& RQF_STATS
) {
436 * We could rq->mq_ctx here, but there's less of a risk
437 * of races if we have the completion event add the stats
438 * to the local software queue.
440 struct blk_mq_ctx
*ctx
;
442 ctx
= __blk_mq_get_ctx(rq
->q
, raw_smp_processor_id());
443 blk_stat_add(&ctx
->stat
[rq_data_dir(rq
)], rq
);
447 static void __blk_mq_complete_request(struct request
*rq
)
449 struct request_queue
*q
= rq
->q
;
453 if (!q
->softirq_done_fn
)
454 blk_mq_end_request(rq
, rq
->errors
);
456 blk_mq_ipi_complete_request(rq
);
460 * blk_mq_complete_request - end I/O on a request
461 * @rq: the request being processed
464 * Ends all I/O on a request. It does not handle partial completions.
465 * The actual completion happens out-of-order, through a IPI handler.
467 void blk_mq_complete_request(struct request
*rq
, int error
)
469 struct request_queue
*q
= rq
->q
;
471 if (unlikely(blk_should_fake_timeout(q
)))
473 if (!blk_mark_rq_complete(rq
)) {
475 __blk_mq_complete_request(rq
);
478 EXPORT_SYMBOL(blk_mq_complete_request
);
480 int blk_mq_request_started(struct request
*rq
)
482 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
484 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
486 void blk_mq_start_request(struct request
*rq
)
488 struct request_queue
*q
= rq
->q
;
490 blk_mq_sched_started_request(rq
);
492 trace_block_rq_issue(q
, rq
);
494 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
495 blk_stat_set_issue_time(&rq
->issue_stat
);
496 rq
->rq_flags
|= RQF_STATS
;
497 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
503 * Ensure that ->deadline is visible before set the started
504 * flag and clear the completed flag.
506 smp_mb__before_atomic();
509 * Mark us as started and clear complete. Complete might have been
510 * set if requeue raced with timeout, which then marked it as
511 * complete. So be sure to clear complete again when we start
512 * the request, otherwise we'll ignore the completion event.
514 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
515 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
516 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
517 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
519 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
521 * Make sure space for the drain appears. We know we can do
522 * this because max_hw_segments has been adjusted to be one
523 * fewer than the device can handle.
525 rq
->nr_phys_segments
++;
528 EXPORT_SYMBOL(blk_mq_start_request
);
530 static void __blk_mq_requeue_request(struct request
*rq
)
532 struct request_queue
*q
= rq
->q
;
534 trace_block_rq_requeue(q
, rq
);
535 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
536 blk_mq_sched_requeue_request(rq
);
538 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
539 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
540 rq
->nr_phys_segments
--;
544 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
546 __blk_mq_requeue_request(rq
);
548 BUG_ON(blk_queued_rq(rq
));
549 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
551 EXPORT_SYMBOL(blk_mq_requeue_request
);
553 static void blk_mq_requeue_work(struct work_struct
*work
)
555 struct request_queue
*q
=
556 container_of(work
, struct request_queue
, requeue_work
.work
);
558 struct request
*rq
, *next
;
561 spin_lock_irqsave(&q
->requeue_lock
, flags
);
562 list_splice_init(&q
->requeue_list
, &rq_list
);
563 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
565 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
566 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
569 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
570 list_del_init(&rq
->queuelist
);
571 blk_mq_sched_insert_request(rq
, true, false, false, true);
574 while (!list_empty(&rq_list
)) {
575 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
576 list_del_init(&rq
->queuelist
);
577 blk_mq_sched_insert_request(rq
, false, false, false, true);
580 blk_mq_run_hw_queues(q
, false);
583 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
584 bool kick_requeue_list
)
586 struct request_queue
*q
= rq
->q
;
590 * We abuse this flag that is otherwise used by the I/O scheduler to
591 * request head insertation from the workqueue.
593 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
595 spin_lock_irqsave(&q
->requeue_lock
, flags
);
597 rq
->rq_flags
|= RQF_SOFTBARRIER
;
598 list_add(&rq
->queuelist
, &q
->requeue_list
);
600 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
602 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
604 if (kick_requeue_list
)
605 blk_mq_kick_requeue_list(q
);
607 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
609 void blk_mq_kick_requeue_list(struct request_queue
*q
)
611 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
613 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
615 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
618 kblockd_schedule_delayed_work(&q
->requeue_work
,
619 msecs_to_jiffies(msecs
));
621 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
623 void blk_mq_abort_requeue_list(struct request_queue
*q
)
628 spin_lock_irqsave(&q
->requeue_lock
, flags
);
629 list_splice_init(&q
->requeue_list
, &rq_list
);
630 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
632 while (!list_empty(&rq_list
)) {
635 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
636 list_del_init(&rq
->queuelist
);
638 blk_mq_end_request(rq
, rq
->errors
);
641 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
643 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
645 if (tag
< tags
->nr_tags
) {
646 prefetch(tags
->rqs
[tag
]);
647 return tags
->rqs
[tag
];
652 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
654 struct blk_mq_timeout_data
{
656 unsigned int next_set
;
659 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
661 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
662 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
665 * We know that complete is set at this point. If STARTED isn't set
666 * anymore, then the request isn't active and the "timeout" should
667 * just be ignored. This can happen due to the bitflag ordering.
668 * Timeout first checks if STARTED is set, and if it is, assumes
669 * the request is active. But if we race with completion, then
670 * we both flags will get cleared. So check here again, and ignore
671 * a timeout event with a request that isn't active.
673 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
677 ret
= ops
->timeout(req
, reserved
);
681 __blk_mq_complete_request(req
);
683 case BLK_EH_RESET_TIMER
:
685 blk_clear_rq_complete(req
);
687 case BLK_EH_NOT_HANDLED
:
690 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
695 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
696 struct request
*rq
, void *priv
, bool reserved
)
698 struct blk_mq_timeout_data
*data
= priv
;
700 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
703 if (time_after_eq(jiffies
, rq
->deadline
)) {
704 if (!blk_mark_rq_complete(rq
))
705 blk_mq_rq_timed_out(rq
, reserved
);
706 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
707 data
->next
= rq
->deadline
;
712 static void blk_mq_timeout_work(struct work_struct
*work
)
714 struct request_queue
*q
=
715 container_of(work
, struct request_queue
, timeout_work
);
716 struct blk_mq_timeout_data data
= {
722 /* A deadlock might occur if a request is stuck requiring a
723 * timeout at the same time a queue freeze is waiting
724 * completion, since the timeout code would not be able to
725 * acquire the queue reference here.
727 * That's why we don't use blk_queue_enter here; instead, we use
728 * percpu_ref_tryget directly, because we need to be able to
729 * obtain a reference even in the short window between the queue
730 * starting to freeze, by dropping the first reference in
731 * blk_mq_freeze_queue_start, and the moment the last request is
732 * consumed, marked by the instant q_usage_counter reaches
735 if (!percpu_ref_tryget(&q
->q_usage_counter
))
738 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
741 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
742 mod_timer(&q
->timeout
, data
.next
);
744 struct blk_mq_hw_ctx
*hctx
;
746 queue_for_each_hw_ctx(q
, hctx
, i
) {
747 /* the hctx may be unmapped, so check it here */
748 if (blk_mq_hw_queue_mapped(hctx
))
749 blk_mq_tag_idle(hctx
);
756 * Reverse check our software queue for entries that we could potentially
757 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
758 * too much time checking for merges.
760 static bool blk_mq_attempt_merge(struct request_queue
*q
,
761 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
766 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
772 if (!blk_rq_merge_ok(rq
, bio
))
775 switch (blk_try_merge(rq
, bio
)) {
776 case ELEVATOR_BACK_MERGE
:
777 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
778 merged
= bio_attempt_back_merge(q
, rq
, bio
);
780 case ELEVATOR_FRONT_MERGE
:
781 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
782 merged
= bio_attempt_front_merge(q
, rq
, bio
);
784 case ELEVATOR_DISCARD_MERGE
:
785 merged
= bio_attempt_discard_merge(q
, rq
, bio
);
799 struct flush_busy_ctx_data
{
800 struct blk_mq_hw_ctx
*hctx
;
801 struct list_head
*list
;
804 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
806 struct flush_busy_ctx_data
*flush_data
= data
;
807 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
808 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
810 sbitmap_clear_bit(sb
, bitnr
);
811 spin_lock(&ctx
->lock
);
812 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
813 spin_unlock(&ctx
->lock
);
818 * Process software queues that have been marked busy, splicing them
819 * to the for-dispatch
821 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
823 struct flush_busy_ctx_data data
= {
828 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
830 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
832 static inline unsigned int queued_to_index(unsigned int queued
)
837 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
840 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
843 struct blk_mq_alloc_data data
= {
845 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
846 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
856 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
857 data
.flags
|= BLK_MQ_REQ_RESERVED
;
859 rq
->tag
= blk_mq_get_tag(&data
);
861 if (blk_mq_tag_busy(data
.hctx
)) {
862 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
863 atomic_inc(&data
.hctx
->nr_active
);
865 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
872 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx
*hctx
,
875 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
878 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
879 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
880 atomic_dec(&hctx
->nr_active
);
884 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx
*hctx
,
887 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
890 __blk_mq_put_driver_tag(hctx
, rq
);
893 static void blk_mq_put_driver_tag(struct request
*rq
)
895 struct blk_mq_hw_ctx
*hctx
;
897 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
900 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
901 __blk_mq_put_driver_tag(hctx
, rq
);
905 * If we fail getting a driver tag because all the driver tags are already
906 * assigned and on the dispatch list, BUT the first entry does not have a
907 * tag, then we could deadlock. For that case, move entries with assigned
908 * driver tags to the front, leaving the set of tagged requests in the
909 * same order, and the untagged set in the same order.
911 static bool reorder_tags_to_front(struct list_head
*list
)
913 struct request
*rq
, *tmp
, *first
= NULL
;
915 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
919 list_move(&rq
->queuelist
, list
);
925 return first
!= NULL
;
928 static int blk_mq_dispatch_wake(wait_queue_t
*wait
, unsigned mode
, int flags
,
931 struct blk_mq_hw_ctx
*hctx
;
933 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
935 list_del(&wait
->task_list
);
936 clear_bit_unlock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
);
937 blk_mq_run_hw_queue(hctx
, true);
941 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx
*hctx
)
943 struct sbq_wait_state
*ws
;
946 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
947 * The thread which wins the race to grab this bit adds the hardware
948 * queue to the wait queue.
950 if (test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
) ||
951 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
954 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
955 ws
= bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
);
958 * As soon as this returns, it's no longer safe to fiddle with
959 * hctx->dispatch_wait, since a completion can wake up the wait queue
960 * and unlock the bit.
962 add_wait_queue(&ws
->wait
, &hctx
->dispatch_wait
);
966 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
968 struct request_queue
*q
= hctx
->queue
;
970 LIST_HEAD(driver_list
);
971 struct list_head
*dptr
;
972 int queued
, ret
= BLK_MQ_RQ_QUEUE_OK
;
975 * Start off with dptr being NULL, so we start the first request
976 * immediately, even if we have more pending.
981 * Now process all the entries, sending them to the driver.
984 while (!list_empty(list
)) {
985 struct blk_mq_queue_data bd
;
987 rq
= list_first_entry(list
, struct request
, queuelist
);
988 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
989 if (!queued
&& reorder_tags_to_front(list
))
993 * The initial allocation attempt failed, so we need to
994 * rerun the hardware queue when a tag is freed.
996 if (blk_mq_dispatch_wait_add(hctx
)) {
998 * It's possible that a tag was freed in the
999 * window between the allocation failure and
1000 * adding the hardware queue to the wait queue.
1002 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1009 list_del_init(&rq
->queuelist
);
1015 * Flag last if we have no more requests, or if we have more
1016 * but can't assign a driver tag to it.
1018 if (list_empty(list
))
1021 struct request
*nxt
;
1023 nxt
= list_first_entry(list
, struct request
, queuelist
);
1024 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1027 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1029 case BLK_MQ_RQ_QUEUE_OK
:
1032 case BLK_MQ_RQ_QUEUE_BUSY
:
1033 blk_mq_put_driver_tag_hctx(hctx
, rq
);
1034 list_add(&rq
->queuelist
, list
);
1035 __blk_mq_requeue_request(rq
);
1038 pr_err("blk-mq: bad return on queue: %d\n", ret
);
1039 case BLK_MQ_RQ_QUEUE_ERROR
:
1041 blk_mq_end_request(rq
, rq
->errors
);
1045 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
1049 * We've done the first request. If we have more than 1
1050 * left in the list, set dptr to defer issue.
1052 if (!dptr
&& list
->next
!= list
->prev
)
1053 dptr
= &driver_list
;
1056 hctx
->dispatched
[queued_to_index(queued
)]++;
1059 * Any items that need requeuing? Stuff them into hctx->dispatch,
1060 * that is where we will continue on next queue run.
1062 if (!list_empty(list
)) {
1064 * If we got a driver tag for the next request already,
1067 rq
= list_first_entry(list
, struct request
, queuelist
);
1068 blk_mq_put_driver_tag(rq
);
1070 spin_lock(&hctx
->lock
);
1071 list_splice_init(list
, &hctx
->dispatch
);
1072 spin_unlock(&hctx
->lock
);
1075 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
1076 * it's possible the queue is stopped and restarted again
1077 * before this. Queue restart will dispatch requests. And since
1078 * requests in rq_list aren't added into hctx->dispatch yet,
1079 * the requests in rq_list might get lost.
1081 * blk_mq_run_hw_queue() already checks the STOPPED bit
1083 * If RESTART or TAG_WAITING is set, then let completion restart
1084 * the queue instead of potentially looping here.
1086 if (!blk_mq_sched_needs_restart(hctx
) &&
1087 !test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1088 blk_mq_run_hw_queue(hctx
, true);
1094 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1098 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1099 cpu_online(hctx
->next_cpu
));
1101 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1103 blk_mq_sched_dispatch_requests(hctx
);
1106 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1107 blk_mq_sched_dispatch_requests(hctx
);
1108 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1113 * It'd be great if the workqueue API had a way to pass
1114 * in a mask and had some smarts for more clever placement.
1115 * For now we just round-robin here, switching for every
1116 * BLK_MQ_CPU_WORK_BATCH queued items.
1118 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1120 if (hctx
->queue
->nr_hw_queues
== 1)
1121 return WORK_CPU_UNBOUND
;
1123 if (--hctx
->next_cpu_batch
<= 0) {
1126 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1127 if (next_cpu
>= nr_cpu_ids
)
1128 next_cpu
= cpumask_first(hctx
->cpumask
);
1130 hctx
->next_cpu
= next_cpu
;
1131 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1134 return hctx
->next_cpu
;
1137 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1139 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
1140 !blk_mq_hw_queue_mapped(hctx
)))
1143 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1144 int cpu
= get_cpu();
1145 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1146 __blk_mq_run_hw_queue(hctx
);
1154 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
);
1157 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1159 struct blk_mq_hw_ctx
*hctx
;
1162 queue_for_each_hw_ctx(q
, hctx
, i
) {
1163 if (!blk_mq_hctx_has_pending(hctx
) ||
1164 blk_mq_hctx_stopped(hctx
))
1167 blk_mq_run_hw_queue(hctx
, async
);
1170 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1173 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1174 * @q: request queue.
1176 * The caller is responsible for serializing this function against
1177 * blk_mq_{start,stop}_hw_queue().
1179 bool blk_mq_queue_stopped(struct request_queue
*q
)
1181 struct blk_mq_hw_ctx
*hctx
;
1184 queue_for_each_hw_ctx(q
, hctx
, i
)
1185 if (blk_mq_hctx_stopped(hctx
))
1190 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1192 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1194 cancel_work(&hctx
->run_work
);
1195 cancel_delayed_work(&hctx
->delay_work
);
1196 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1198 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1200 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1202 struct blk_mq_hw_ctx
*hctx
;
1205 queue_for_each_hw_ctx(q
, hctx
, i
)
1206 blk_mq_stop_hw_queue(hctx
);
1208 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1210 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1212 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1214 blk_mq_run_hw_queue(hctx
, false);
1216 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1218 void blk_mq_start_hw_queues(struct request_queue
*q
)
1220 struct blk_mq_hw_ctx
*hctx
;
1223 queue_for_each_hw_ctx(q
, hctx
, i
)
1224 blk_mq_start_hw_queue(hctx
);
1226 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1228 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1230 if (!blk_mq_hctx_stopped(hctx
))
1233 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1234 blk_mq_run_hw_queue(hctx
, async
);
1236 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1238 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1240 struct blk_mq_hw_ctx
*hctx
;
1243 queue_for_each_hw_ctx(q
, hctx
, i
)
1244 blk_mq_start_stopped_hw_queue(hctx
, async
);
1246 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1248 static void blk_mq_run_work_fn(struct work_struct
*work
)
1250 struct blk_mq_hw_ctx
*hctx
;
1252 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1254 __blk_mq_run_hw_queue(hctx
);
1257 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1259 struct blk_mq_hw_ctx
*hctx
;
1261 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1263 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1264 __blk_mq_run_hw_queue(hctx
);
1267 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1269 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1272 blk_mq_stop_hw_queue(hctx
);
1273 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1274 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1276 EXPORT_SYMBOL(blk_mq_delay_queue
);
1278 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1282 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1284 trace_block_rq_insert(hctx
->queue
, rq
);
1287 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1289 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1292 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1295 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1297 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1298 blk_mq_hctx_mark_pending(hctx
, ctx
);
1301 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1302 struct list_head
*list
)
1306 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1309 spin_lock(&ctx
->lock
);
1310 while (!list_empty(list
)) {
1313 rq
= list_first_entry(list
, struct request
, queuelist
);
1314 BUG_ON(rq
->mq_ctx
!= ctx
);
1315 list_del_init(&rq
->queuelist
);
1316 __blk_mq_insert_req_list(hctx
, rq
, false);
1318 blk_mq_hctx_mark_pending(hctx
, ctx
);
1319 spin_unlock(&ctx
->lock
);
1322 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1324 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1325 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1327 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1328 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1329 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1332 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1334 struct blk_mq_ctx
*this_ctx
;
1335 struct request_queue
*this_q
;
1338 LIST_HEAD(ctx_list
);
1341 list_splice_init(&plug
->mq_list
, &list
);
1343 list_sort(NULL
, &list
, plug_ctx_cmp
);
1349 while (!list_empty(&list
)) {
1350 rq
= list_entry_rq(list
.next
);
1351 list_del_init(&rq
->queuelist
);
1353 if (rq
->mq_ctx
!= this_ctx
) {
1355 trace_block_unplug(this_q
, depth
, from_schedule
);
1356 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1361 this_ctx
= rq
->mq_ctx
;
1367 list_add_tail(&rq
->queuelist
, &ctx_list
);
1371 * If 'this_ctx' is set, we know we have entries to complete
1372 * on 'ctx_list'. Do those.
1375 trace_block_unplug(this_q
, depth
, from_schedule
);
1376 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1381 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1383 init_request_from_bio(rq
, bio
);
1385 blk_account_io_start(rq
, true);
1388 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1390 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1391 !blk_queue_nomerges(hctx
->queue
);
1394 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1395 struct blk_mq_ctx
*ctx
,
1396 struct request
*rq
, struct bio
*bio
)
1398 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1399 blk_mq_bio_to_request(rq
, bio
);
1400 spin_lock(&ctx
->lock
);
1402 __blk_mq_insert_request(hctx
, rq
, false);
1403 spin_unlock(&ctx
->lock
);
1406 struct request_queue
*q
= hctx
->queue
;
1408 spin_lock(&ctx
->lock
);
1409 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1410 blk_mq_bio_to_request(rq
, bio
);
1414 spin_unlock(&ctx
->lock
);
1415 __blk_mq_finish_request(hctx
, ctx
, rq
);
1420 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1423 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1425 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1428 static void blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
,
1431 struct request_queue
*q
= rq
->q
;
1432 struct blk_mq_queue_data bd
= {
1437 struct blk_mq_hw_ctx
*hctx
;
1438 blk_qc_t new_cookie
;
1444 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1447 new_cookie
= request_to_qc_t(hctx
, rq
);
1450 * For OK queue, we are done. For error, kill it. Any other
1451 * error (busy), just add it to our list as we previously
1454 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1455 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1456 *cookie
= new_cookie
;
1460 __blk_mq_requeue_request(rq
);
1462 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1463 *cookie
= BLK_QC_T_NONE
;
1465 blk_mq_end_request(rq
, rq
->errors
);
1470 blk_mq_sched_insert_request(rq
, false, true, false, may_sleep
);
1474 * Multiple hardware queue variant. This will not use per-process plugs,
1475 * but will attempt to bypass the hctx queueing if we can go straight to
1476 * hardware for SYNC IO.
1478 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1480 const int is_sync
= op_is_sync(bio
->bi_opf
);
1481 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1482 struct blk_mq_alloc_data data
= { .flags
= 0 };
1484 unsigned int request_count
= 0, srcu_idx
;
1485 struct blk_plug
*plug
;
1486 struct request
*same_queue_rq
= NULL
;
1488 unsigned int wb_acct
;
1490 blk_queue_bounce(q
, &bio
);
1492 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1494 return BLK_QC_T_NONE
;
1497 blk_queue_split(q
, &bio
, q
->bio_split
);
1499 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1500 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1501 return BLK_QC_T_NONE
;
1503 if (blk_mq_sched_bio_merge(q
, bio
))
1504 return BLK_QC_T_NONE
;
1506 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1508 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1510 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1511 if (unlikely(!rq
)) {
1512 __wbt_done(q
->rq_wb
, wb_acct
);
1513 return BLK_QC_T_NONE
;
1516 wbt_track(&rq
->issue_stat
, wb_acct
);
1518 cookie
= request_to_qc_t(data
.hctx
, rq
);
1520 if (unlikely(is_flush_fua
)) {
1523 blk_mq_bio_to_request(rq
, bio
);
1524 blk_insert_flush(rq
);
1528 plug
= current
->plug
;
1530 * If the driver supports defer issued based on 'last', then
1531 * queue it up like normal since we can potentially save some
1534 if (((plug
&& !blk_queue_nomerges(q
)) || is_sync
) &&
1535 !(data
.hctx
->flags
& BLK_MQ_F_DEFER_ISSUE
)) {
1536 struct request
*old_rq
= NULL
;
1538 blk_mq_bio_to_request(rq
, bio
);
1541 * We do limited plugging. If the bio can be merged, do that.
1542 * Otherwise the existing request in the plug list will be
1543 * issued. So the plug list will have one request at most
1547 * The plug list might get flushed before this. If that
1548 * happens, same_queue_rq is invalid and plug list is
1551 if (same_queue_rq
&& !list_empty(&plug
->mq_list
)) {
1552 old_rq
= same_queue_rq
;
1553 list_del_init(&old_rq
->queuelist
);
1555 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1556 } else /* is_sync */
1558 blk_mq_put_ctx(data
.ctx
);
1562 if (!(data
.hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1564 blk_mq_try_issue_directly(old_rq
, &cookie
, false);
1567 srcu_idx
= srcu_read_lock(&data
.hctx
->queue_rq_srcu
);
1568 blk_mq_try_issue_directly(old_rq
, &cookie
, true);
1569 srcu_read_unlock(&data
.hctx
->queue_rq_srcu
, srcu_idx
);
1576 blk_mq_put_ctx(data
.ctx
);
1577 blk_mq_bio_to_request(rq
, bio
);
1578 blk_mq_sched_insert_request(rq
, false, true,
1579 !is_sync
|| is_flush_fua
, true);
1582 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1584 * For a SYNC request, send it to the hardware immediately. For
1585 * an ASYNC request, just ensure that we run it later on. The
1586 * latter allows for merging opportunities and more efficient
1590 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1592 blk_mq_put_ctx(data
.ctx
);
1598 * Single hardware queue variant. This will attempt to use any per-process
1599 * plug for merging and IO deferral.
1601 static blk_qc_t
blk_sq_make_request(struct request_queue
*q
, struct bio
*bio
)
1603 const int is_sync
= op_is_sync(bio
->bi_opf
);
1604 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1605 struct blk_plug
*plug
;
1606 unsigned int request_count
= 0;
1607 struct blk_mq_alloc_data data
= { .flags
= 0 };
1610 unsigned int wb_acct
;
1612 blk_queue_bounce(q
, &bio
);
1614 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1616 return BLK_QC_T_NONE
;
1619 blk_queue_split(q
, &bio
, q
->bio_split
);
1621 if (!is_flush_fua
&& !blk_queue_nomerges(q
)) {
1622 if (blk_attempt_plug_merge(q
, bio
, &request_count
, NULL
))
1623 return BLK_QC_T_NONE
;
1625 request_count
= blk_plug_queued_count(q
);
1627 if (blk_mq_sched_bio_merge(q
, bio
))
1628 return BLK_QC_T_NONE
;
1630 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1632 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1634 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1635 if (unlikely(!rq
)) {
1636 __wbt_done(q
->rq_wb
, wb_acct
);
1637 return BLK_QC_T_NONE
;
1640 wbt_track(&rq
->issue_stat
, wb_acct
);
1642 cookie
= request_to_qc_t(data
.hctx
, rq
);
1644 if (unlikely(is_flush_fua
)) {
1647 blk_mq_bio_to_request(rq
, bio
);
1648 blk_insert_flush(rq
);
1653 * A task plug currently exists. Since this is completely lockless,
1654 * utilize that to temporarily store requests until the task is
1655 * either done or scheduled away.
1657 plug
= current
->plug
;
1659 struct request
*last
= NULL
;
1661 blk_mq_bio_to_request(rq
, bio
);
1664 * @request_count may become stale because of schedule
1665 * out, so check the list again.
1667 if (list_empty(&plug
->mq_list
))
1670 trace_block_plug(q
);
1672 last
= list_entry_rq(plug
->mq_list
.prev
);
1674 blk_mq_put_ctx(data
.ctx
);
1676 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1677 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1678 blk_flush_plug_list(plug
, false);
1679 trace_block_plug(q
);
1682 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1688 blk_mq_put_ctx(data
.ctx
);
1689 blk_mq_bio_to_request(rq
, bio
);
1690 blk_mq_sched_insert_request(rq
, false, true,
1691 !is_sync
|| is_flush_fua
, true);
1694 if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1696 * For a SYNC request, send it to the hardware immediately. For
1697 * an ASYNC request, just ensure that we run it later on. The
1698 * latter allows for merging opportunities and more efficient
1702 blk_mq_run_hw_queue(data
.hctx
, !is_sync
|| is_flush_fua
);
1705 blk_mq_put_ctx(data
.ctx
);
1710 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1711 unsigned int hctx_idx
)
1715 if (tags
->rqs
&& set
->ops
->exit_request
) {
1718 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1719 struct request
*rq
= tags
->static_rqs
[i
];
1723 set
->ops
->exit_request(set
->driver_data
, rq
,
1725 tags
->static_rqs
[i
] = NULL
;
1729 while (!list_empty(&tags
->page_list
)) {
1730 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1731 list_del_init(&page
->lru
);
1733 * Remove kmemleak object previously allocated in
1734 * blk_mq_init_rq_map().
1736 kmemleak_free(page_address(page
));
1737 __free_pages(page
, page
->private);
1741 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1745 kfree(tags
->static_rqs
);
1746 tags
->static_rqs
= NULL
;
1748 blk_mq_free_tags(tags
);
1751 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1752 unsigned int hctx_idx
,
1753 unsigned int nr_tags
,
1754 unsigned int reserved_tags
)
1756 struct blk_mq_tags
*tags
;
1759 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1760 if (node
== NUMA_NO_NODE
)
1761 node
= set
->numa_node
;
1763 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1764 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1768 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1769 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1772 blk_mq_free_tags(tags
);
1776 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1777 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1779 if (!tags
->static_rqs
) {
1781 blk_mq_free_tags(tags
);
1788 static size_t order_to_size(unsigned int order
)
1790 return (size_t)PAGE_SIZE
<< order
;
1793 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1794 unsigned int hctx_idx
, unsigned int depth
)
1796 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1797 size_t rq_size
, left
;
1800 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1801 if (node
== NUMA_NO_NODE
)
1802 node
= set
->numa_node
;
1804 INIT_LIST_HEAD(&tags
->page_list
);
1807 * rq_size is the size of the request plus driver payload, rounded
1808 * to the cacheline size
1810 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1812 left
= rq_size
* depth
;
1814 for (i
= 0; i
< depth
; ) {
1815 int this_order
= max_order
;
1820 while (this_order
&& left
< order_to_size(this_order
- 1))
1824 page
= alloc_pages_node(node
,
1825 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1831 if (order_to_size(this_order
) < rq_size
)
1838 page
->private = this_order
;
1839 list_add_tail(&page
->lru
, &tags
->page_list
);
1841 p
= page_address(page
);
1843 * Allow kmemleak to scan these pages as they contain pointers
1844 * to additional allocations like via ops->init_request().
1846 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1847 entries_per_page
= order_to_size(this_order
) / rq_size
;
1848 to_do
= min(entries_per_page
, depth
- i
);
1849 left
-= to_do
* rq_size
;
1850 for (j
= 0; j
< to_do
; j
++) {
1851 struct request
*rq
= p
;
1853 tags
->static_rqs
[i
] = rq
;
1854 if (set
->ops
->init_request
) {
1855 if (set
->ops
->init_request(set
->driver_data
,
1858 tags
->static_rqs
[i
] = NULL
;
1870 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1875 * 'cpu' is going away. splice any existing rq_list entries from this
1876 * software queue to the hw queue dispatch list, and ensure that it
1879 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1881 struct blk_mq_hw_ctx
*hctx
;
1882 struct blk_mq_ctx
*ctx
;
1885 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1886 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1888 spin_lock(&ctx
->lock
);
1889 if (!list_empty(&ctx
->rq_list
)) {
1890 list_splice_init(&ctx
->rq_list
, &tmp
);
1891 blk_mq_hctx_clear_pending(hctx
, ctx
);
1893 spin_unlock(&ctx
->lock
);
1895 if (list_empty(&tmp
))
1898 spin_lock(&hctx
->lock
);
1899 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1900 spin_unlock(&hctx
->lock
);
1902 blk_mq_run_hw_queue(hctx
, true);
1906 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1908 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1912 /* hctx->ctxs will be freed in queue's release handler */
1913 static void blk_mq_exit_hctx(struct request_queue
*q
,
1914 struct blk_mq_tag_set
*set
,
1915 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1917 unsigned flush_start_tag
= set
->queue_depth
;
1919 blk_mq_tag_idle(hctx
);
1921 if (set
->ops
->exit_request
)
1922 set
->ops
->exit_request(set
->driver_data
,
1923 hctx
->fq
->flush_rq
, hctx_idx
,
1924 flush_start_tag
+ hctx_idx
);
1926 if (set
->ops
->exit_hctx
)
1927 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1929 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1930 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1932 blk_mq_remove_cpuhp(hctx
);
1933 blk_free_flush_queue(hctx
->fq
);
1934 sbitmap_free(&hctx
->ctx_map
);
1937 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1938 struct blk_mq_tag_set
*set
, int nr_queue
)
1940 struct blk_mq_hw_ctx
*hctx
;
1943 queue_for_each_hw_ctx(q
, hctx
, i
) {
1946 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1950 static int blk_mq_init_hctx(struct request_queue
*q
,
1951 struct blk_mq_tag_set
*set
,
1952 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1955 unsigned flush_start_tag
= set
->queue_depth
;
1957 node
= hctx
->numa_node
;
1958 if (node
== NUMA_NO_NODE
)
1959 node
= hctx
->numa_node
= set
->numa_node
;
1961 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1962 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1963 spin_lock_init(&hctx
->lock
);
1964 INIT_LIST_HEAD(&hctx
->dispatch
);
1966 hctx
->queue_num
= hctx_idx
;
1967 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1969 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1971 hctx
->tags
= set
->tags
[hctx_idx
];
1974 * Allocate space for all possible cpus to avoid allocation at
1977 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1980 goto unregister_cpu_notifier
;
1982 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1988 if (set
->ops
->init_hctx
&&
1989 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1992 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1996 if (set
->ops
->init_request
&&
1997 set
->ops
->init_request(set
->driver_data
,
1998 hctx
->fq
->flush_rq
, hctx_idx
,
1999 flush_start_tag
+ hctx_idx
, node
))
2002 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2003 init_srcu_struct(&hctx
->queue_rq_srcu
);
2010 if (set
->ops
->exit_hctx
)
2011 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2013 sbitmap_free(&hctx
->ctx_map
);
2016 unregister_cpu_notifier
:
2017 blk_mq_remove_cpuhp(hctx
);
2021 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2022 unsigned int nr_hw_queues
)
2026 for_each_possible_cpu(i
) {
2027 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2028 struct blk_mq_hw_ctx
*hctx
;
2031 spin_lock_init(&__ctx
->lock
);
2032 INIT_LIST_HEAD(&__ctx
->rq_list
);
2034 blk_stat_init(&__ctx
->stat
[BLK_STAT_READ
]);
2035 blk_stat_init(&__ctx
->stat
[BLK_STAT_WRITE
]);
2037 /* If the cpu isn't online, the cpu is mapped to first hctx */
2041 hctx
= blk_mq_map_queue(q
, i
);
2044 * Set local node, IFF we have more than one hw queue. If
2045 * not, we remain on the home node of the device
2047 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2048 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2052 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2056 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2057 set
->queue_depth
, set
->reserved_tags
);
2058 if (!set
->tags
[hctx_idx
])
2061 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2066 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2067 set
->tags
[hctx_idx
] = NULL
;
2071 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2072 unsigned int hctx_idx
)
2074 if (set
->tags
[hctx_idx
]) {
2075 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2076 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2077 set
->tags
[hctx_idx
] = NULL
;
2081 static void blk_mq_map_swqueue(struct request_queue
*q
,
2082 const struct cpumask
*online_mask
)
2084 unsigned int i
, hctx_idx
;
2085 struct blk_mq_hw_ctx
*hctx
;
2086 struct blk_mq_ctx
*ctx
;
2087 struct blk_mq_tag_set
*set
= q
->tag_set
;
2090 * Avoid others reading imcomplete hctx->cpumask through sysfs
2092 mutex_lock(&q
->sysfs_lock
);
2094 queue_for_each_hw_ctx(q
, hctx
, i
) {
2095 cpumask_clear(hctx
->cpumask
);
2100 * Map software to hardware queues
2102 for_each_possible_cpu(i
) {
2103 /* If the cpu isn't online, the cpu is mapped to first hctx */
2104 if (!cpumask_test_cpu(i
, online_mask
))
2107 hctx_idx
= q
->mq_map
[i
];
2108 /* unmapped hw queue can be remapped after CPU topo changed */
2109 if (!set
->tags
[hctx_idx
] &&
2110 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2112 * If tags initialization fail for some hctx,
2113 * that hctx won't be brought online. In this
2114 * case, remap the current ctx to hctx[0] which
2115 * is guaranteed to always have tags allocated
2120 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2121 hctx
= blk_mq_map_queue(q
, i
);
2123 cpumask_set_cpu(i
, hctx
->cpumask
);
2124 ctx
->index_hw
= hctx
->nr_ctx
;
2125 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2128 mutex_unlock(&q
->sysfs_lock
);
2130 queue_for_each_hw_ctx(q
, hctx
, i
) {
2132 * If no software queues are mapped to this hardware queue,
2133 * disable it and free the request entries.
2135 if (!hctx
->nr_ctx
) {
2136 /* Never unmap queue 0. We need it as a
2137 * fallback in case of a new remap fails
2140 if (i
&& set
->tags
[i
])
2141 blk_mq_free_map_and_requests(set
, i
);
2147 hctx
->tags
= set
->tags
[i
];
2148 WARN_ON(!hctx
->tags
);
2151 * Set the map size to the number of mapped software queues.
2152 * This is more accurate and more efficient than looping
2153 * over all possibly mapped software queues.
2155 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2158 * Initialize batch roundrobin counts
2160 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2161 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2165 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2167 struct blk_mq_hw_ctx
*hctx
;
2170 queue_for_each_hw_ctx(q
, hctx
, i
) {
2172 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2174 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2178 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
2180 struct request_queue
*q
;
2182 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2183 blk_mq_freeze_queue(q
);
2184 queue_set_hctx_shared(q
, shared
);
2185 blk_mq_unfreeze_queue(q
);
2189 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2191 struct blk_mq_tag_set
*set
= q
->tag_set
;
2193 mutex_lock(&set
->tag_list_lock
);
2194 list_del_init(&q
->tag_set_list
);
2195 if (list_is_singular(&set
->tag_list
)) {
2196 /* just transitioned to unshared */
2197 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2198 /* update existing queue */
2199 blk_mq_update_tag_set_depth(set
, false);
2201 mutex_unlock(&set
->tag_list_lock
);
2204 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2205 struct request_queue
*q
)
2209 mutex_lock(&set
->tag_list_lock
);
2211 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2212 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2213 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2214 /* update existing queue */
2215 blk_mq_update_tag_set_depth(set
, true);
2217 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2218 queue_set_hctx_shared(q
, true);
2219 list_add_tail(&q
->tag_set_list
, &set
->tag_list
);
2221 mutex_unlock(&set
->tag_list_lock
);
2225 * It is the actual release handler for mq, but we do it from
2226 * request queue's release handler for avoiding use-after-free
2227 * and headache because q->mq_kobj shouldn't have been introduced,
2228 * but we can't group ctx/kctx kobj without it.
2230 void blk_mq_release(struct request_queue
*q
)
2232 struct blk_mq_hw_ctx
*hctx
;
2235 blk_mq_sched_teardown(q
);
2237 /* hctx kobj stays in hctx */
2238 queue_for_each_hw_ctx(q
, hctx
, i
) {
2241 kobject_put(&hctx
->kobj
);
2246 kfree(q
->queue_hw_ctx
);
2249 * release .mq_kobj and sw queue's kobject now because
2250 * both share lifetime with request queue.
2252 blk_mq_sysfs_deinit(q
);
2254 free_percpu(q
->queue_ctx
);
2257 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2259 struct request_queue
*uninit_q
, *q
;
2261 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2263 return ERR_PTR(-ENOMEM
);
2265 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2267 blk_cleanup_queue(uninit_q
);
2271 EXPORT_SYMBOL(blk_mq_init_queue
);
2273 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2274 struct request_queue
*q
)
2277 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2279 blk_mq_sysfs_unregister(q
);
2280 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2286 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2287 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2292 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2299 atomic_set(&hctxs
[i
]->nr_active
, 0);
2300 hctxs
[i
]->numa_node
= node
;
2301 hctxs
[i
]->queue_num
= i
;
2303 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2304 free_cpumask_var(hctxs
[i
]->cpumask
);
2309 blk_mq_hctx_kobj_init(hctxs
[i
]);
2311 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2312 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2316 blk_mq_free_map_and_requests(set
, j
);
2317 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2318 kobject_put(&hctx
->kobj
);
2323 q
->nr_hw_queues
= i
;
2324 blk_mq_sysfs_register(q
);
2327 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2328 struct request_queue
*q
)
2330 /* mark the queue as mq asap */
2331 q
->mq_ops
= set
->ops
;
2333 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2337 /* init q->mq_kobj and sw queues' kobjects */
2338 blk_mq_sysfs_init(q
);
2340 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2341 GFP_KERNEL
, set
->numa_node
);
2342 if (!q
->queue_hw_ctx
)
2345 q
->mq_map
= set
->mq_map
;
2347 blk_mq_realloc_hw_ctxs(set
, q
);
2348 if (!q
->nr_hw_queues
)
2351 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2352 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2354 q
->nr_queues
= nr_cpu_ids
;
2356 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2358 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2359 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2361 q
->sg_reserved_size
= INT_MAX
;
2363 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2364 INIT_LIST_HEAD(&q
->requeue_list
);
2365 spin_lock_init(&q
->requeue_lock
);
2367 if (q
->nr_hw_queues
> 1)
2368 blk_queue_make_request(q
, blk_mq_make_request
);
2370 blk_queue_make_request(q
, blk_sq_make_request
);
2373 * Do this after blk_queue_make_request() overrides it...
2375 q
->nr_requests
= set
->queue_depth
;
2378 * Default to classic polling
2382 if (set
->ops
->complete
)
2383 blk_queue_softirq_done(q
, set
->ops
->complete
);
2385 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2388 mutex_lock(&all_q_mutex
);
2390 list_add_tail(&q
->all_q_node
, &all_q_list
);
2391 blk_mq_add_queue_tag_set(set
, q
);
2392 blk_mq_map_swqueue(q
, cpu_online_mask
);
2394 mutex_unlock(&all_q_mutex
);
2397 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2400 ret
= blk_mq_sched_init(q
);
2402 return ERR_PTR(ret
);
2408 kfree(q
->queue_hw_ctx
);
2410 free_percpu(q
->queue_ctx
);
2413 return ERR_PTR(-ENOMEM
);
2415 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2417 void blk_mq_free_queue(struct request_queue
*q
)
2419 struct blk_mq_tag_set
*set
= q
->tag_set
;
2421 mutex_lock(&all_q_mutex
);
2422 list_del_init(&q
->all_q_node
);
2423 mutex_unlock(&all_q_mutex
);
2427 blk_mq_del_queue_tag_set(q
);
2429 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2432 /* Basically redo blk_mq_init_queue with queue frozen */
2433 static void blk_mq_queue_reinit(struct request_queue
*q
,
2434 const struct cpumask
*online_mask
)
2436 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2438 blk_mq_sysfs_unregister(q
);
2441 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2442 * we should change hctx numa_node according to new topology (this
2443 * involves free and re-allocate memory, worthy doing?)
2446 blk_mq_map_swqueue(q
, online_mask
);
2448 blk_mq_sysfs_register(q
);
2452 * New online cpumask which is going to be set in this hotplug event.
2453 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2454 * one-by-one and dynamically allocating this could result in a failure.
2456 static struct cpumask cpuhp_online_new
;
2458 static void blk_mq_queue_reinit_work(void)
2460 struct request_queue
*q
;
2462 mutex_lock(&all_q_mutex
);
2464 * We need to freeze and reinit all existing queues. Freezing
2465 * involves synchronous wait for an RCU grace period and doing it
2466 * one by one may take a long time. Start freezing all queues in
2467 * one swoop and then wait for the completions so that freezing can
2468 * take place in parallel.
2470 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2471 blk_mq_freeze_queue_start(q
);
2472 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2473 blk_mq_freeze_queue_wait(q
);
2475 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2476 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2478 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2479 blk_mq_unfreeze_queue(q
);
2481 mutex_unlock(&all_q_mutex
);
2484 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2486 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2487 blk_mq_queue_reinit_work();
2492 * Before hotadded cpu starts handling requests, new mappings must be
2493 * established. Otherwise, these requests in hw queue might never be
2496 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2497 * for CPU0, and ctx1 for CPU1).
2499 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2500 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2502 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2503 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2504 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2507 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2509 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2510 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2511 blk_mq_queue_reinit_work();
2515 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2519 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2520 if (!__blk_mq_alloc_rq_map(set
, i
))
2527 blk_mq_free_rq_map(set
->tags
[i
]);
2533 * Allocate the request maps associated with this tag_set. Note that this
2534 * may reduce the depth asked for, if memory is tight. set->queue_depth
2535 * will be updated to reflect the allocated depth.
2537 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2542 depth
= set
->queue_depth
;
2544 err
= __blk_mq_alloc_rq_maps(set
);
2548 set
->queue_depth
>>= 1;
2549 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2553 } while (set
->queue_depth
);
2555 if (!set
->queue_depth
|| err
) {
2556 pr_err("blk-mq: failed to allocate request map\n");
2560 if (depth
!= set
->queue_depth
)
2561 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2562 depth
, set
->queue_depth
);
2568 * Alloc a tag set to be associated with one or more request queues.
2569 * May fail with EINVAL for various error conditions. May adjust the
2570 * requested depth down, if if it too large. In that case, the set
2571 * value will be stored in set->queue_depth.
2573 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2577 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2579 if (!set
->nr_hw_queues
)
2581 if (!set
->queue_depth
)
2583 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2586 if (!set
->ops
->queue_rq
)
2589 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2590 pr_info("blk-mq: reduced tag depth to %u\n",
2592 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2596 * If a crashdump is active, then we are potentially in a very
2597 * memory constrained environment. Limit us to 1 queue and
2598 * 64 tags to prevent using too much memory.
2600 if (is_kdump_kernel()) {
2601 set
->nr_hw_queues
= 1;
2602 set
->queue_depth
= min(64U, set
->queue_depth
);
2605 * There is no use for more h/w queues than cpus.
2607 if (set
->nr_hw_queues
> nr_cpu_ids
)
2608 set
->nr_hw_queues
= nr_cpu_ids
;
2610 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2611 GFP_KERNEL
, set
->numa_node
);
2616 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2617 GFP_KERNEL
, set
->numa_node
);
2621 if (set
->ops
->map_queues
)
2622 ret
= set
->ops
->map_queues(set
);
2624 ret
= blk_mq_map_queues(set
);
2626 goto out_free_mq_map
;
2628 ret
= blk_mq_alloc_rq_maps(set
);
2630 goto out_free_mq_map
;
2632 mutex_init(&set
->tag_list_lock
);
2633 INIT_LIST_HEAD(&set
->tag_list
);
2645 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2647 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2651 for (i
= 0; i
< nr_cpu_ids
; i
++)
2652 blk_mq_free_map_and_requests(set
, i
);
2660 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2662 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2664 struct blk_mq_tag_set
*set
= q
->tag_set
;
2665 struct blk_mq_hw_ctx
*hctx
;
2671 blk_mq_freeze_queue(q
);
2672 blk_mq_quiesce_queue(q
);
2675 queue_for_each_hw_ctx(q
, hctx
, i
) {
2679 * If we're using an MQ scheduler, just update the scheduler
2680 * queue depth. This is similar to what the old code would do.
2682 if (!hctx
->sched_tags
) {
2683 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2684 min(nr
, set
->queue_depth
),
2687 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2695 q
->nr_requests
= nr
;
2697 blk_mq_unfreeze_queue(q
);
2698 blk_mq_start_stopped_hw_queues(q
, true);
2703 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2705 struct request_queue
*q
;
2707 if (nr_hw_queues
> nr_cpu_ids
)
2708 nr_hw_queues
= nr_cpu_ids
;
2709 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2712 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2713 blk_mq_freeze_queue(q
);
2715 set
->nr_hw_queues
= nr_hw_queues
;
2716 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2717 blk_mq_realloc_hw_ctxs(set
, q
);
2720 * Manually set the make_request_fn as blk_queue_make_request
2721 * resets a lot of the queue settings.
2723 if (q
->nr_hw_queues
> 1)
2724 q
->make_request_fn
= blk_mq_make_request
;
2726 q
->make_request_fn
= blk_sq_make_request
;
2728 blk_mq_queue_reinit(q
, cpu_online_mask
);
2731 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2732 blk_mq_unfreeze_queue(q
);
2734 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2736 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2737 struct blk_mq_hw_ctx
*hctx
,
2740 struct blk_rq_stat stat
[2];
2741 unsigned long ret
= 0;
2744 * If stats collection isn't on, don't sleep but turn it on for
2747 if (!blk_stat_enable(q
))
2751 * We don't have to do this once per IO, should optimize this
2752 * to just use the current window of stats until it changes
2754 memset(&stat
, 0, sizeof(stat
));
2755 blk_hctx_stat_get(hctx
, stat
);
2758 * As an optimistic guess, use half of the mean service time
2759 * for this type of request. We can (and should) make this smarter.
2760 * For instance, if the completion latencies are tight, we can
2761 * get closer than just half the mean. This is especially
2762 * important on devices where the completion latencies are longer
2765 if (req_op(rq
) == REQ_OP_READ
&& stat
[BLK_STAT_READ
].nr_samples
)
2766 ret
= (stat
[BLK_STAT_READ
].mean
+ 1) / 2;
2767 else if (req_op(rq
) == REQ_OP_WRITE
&& stat
[BLK_STAT_WRITE
].nr_samples
)
2768 ret
= (stat
[BLK_STAT_WRITE
].mean
+ 1) / 2;
2773 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2774 struct blk_mq_hw_ctx
*hctx
,
2777 struct hrtimer_sleeper hs
;
2778 enum hrtimer_mode mode
;
2782 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2788 * -1: don't ever hybrid sleep
2789 * 0: use half of prev avg
2790 * >0: use this specific value
2792 if (q
->poll_nsec
== -1)
2794 else if (q
->poll_nsec
> 0)
2795 nsecs
= q
->poll_nsec
;
2797 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2802 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2805 * This will be replaced with the stats tracking code, using
2806 * 'avg_completion_time / 2' as the pre-sleep target.
2810 mode
= HRTIMER_MODE_REL
;
2811 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2812 hrtimer_set_expires(&hs
.timer
, kt
);
2814 hrtimer_init_sleeper(&hs
, current
);
2816 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2818 set_current_state(TASK_UNINTERRUPTIBLE
);
2819 hrtimer_start_expires(&hs
.timer
, mode
);
2822 hrtimer_cancel(&hs
.timer
);
2823 mode
= HRTIMER_MODE_ABS
;
2824 } while (hs
.task
&& !signal_pending(current
));
2826 __set_current_state(TASK_RUNNING
);
2827 destroy_hrtimer_on_stack(&hs
.timer
);
2831 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2833 struct request_queue
*q
= hctx
->queue
;
2837 * If we sleep, have the caller restart the poll loop to reset
2838 * the state. Like for the other success return cases, the
2839 * caller is responsible for checking if the IO completed. If
2840 * the IO isn't complete, we'll get called again and will go
2841 * straight to the busy poll loop.
2843 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2846 hctx
->poll_considered
++;
2848 state
= current
->state
;
2849 while (!need_resched()) {
2852 hctx
->poll_invoked
++;
2854 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2856 hctx
->poll_success
++;
2857 set_current_state(TASK_RUNNING
);
2861 if (signal_pending_state(state
, current
))
2862 set_current_state(TASK_RUNNING
);
2864 if (current
->state
== TASK_RUNNING
)
2874 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2876 struct blk_mq_hw_ctx
*hctx
;
2877 struct blk_plug
*plug
;
2880 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2881 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2884 plug
= current
->plug
;
2886 blk_flush_plug_list(plug
, false);
2888 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2889 if (!blk_qc_t_is_internal(cookie
))
2890 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2892 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2894 return __blk_mq_poll(hctx
, rq
);
2896 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2898 void blk_mq_disable_hotplug(void)
2900 mutex_lock(&all_q_mutex
);
2903 void blk_mq_enable_hotplug(void)
2905 mutex_unlock(&all_q_mutex
);
2908 static int __init
blk_mq_init(void)
2910 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2911 blk_mq_hctx_notify_dead
);
2913 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2914 blk_mq_queue_reinit_prepare
,
2915 blk_mq_queue_reinit_dead
);
2918 subsys_initcall(blk_mq_init
);