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
);
42 static void blk_mq_poll_stats_start(struct request_queue
*q
);
43 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
45 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
47 int ddir
, bytes
, bucket
;
49 ddir
= blk_stat_rq_ddir(rq
);
50 bytes
= blk_rq_bytes(rq
);
52 bucket
= ddir
+ 2*(ilog2(bytes
) - 9);
56 else if (bucket
>= BLK_MQ_POLL_STATS_BKTS
)
57 return ddir
+ BLK_MQ_POLL_STATS_BKTS
- 2;
63 * Check if any of the ctx's have pending work in this hardware queue
65 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
67 return sbitmap_any_bit_set(&hctx
->ctx_map
) ||
68 !list_empty_careful(&hctx
->dispatch
) ||
69 blk_mq_sched_has_work(hctx
);
73 * Mark this ctx as having pending work in this hardware queue
75 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
76 struct blk_mq_ctx
*ctx
)
78 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
79 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
82 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
83 struct blk_mq_ctx
*ctx
)
85 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
88 void blk_freeze_queue_start(struct request_queue
*q
)
92 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
93 if (freeze_depth
== 1) {
94 percpu_ref_kill(&q
->q_usage_counter
);
95 blk_mq_run_hw_queues(q
, false);
98 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
100 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
102 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
104 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
106 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
107 unsigned long timeout
)
109 return wait_event_timeout(q
->mq_freeze_wq
,
110 percpu_ref_is_zero(&q
->q_usage_counter
),
113 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
116 * Guarantee no request is in use, so we can change any data structure of
117 * the queue afterward.
119 void blk_freeze_queue(struct request_queue
*q
)
122 * In the !blk_mq case we are only calling this to kill the
123 * q_usage_counter, otherwise this increases the freeze depth
124 * and waits for it to return to zero. For this reason there is
125 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
126 * exported to drivers as the only user for unfreeze is blk_mq.
128 blk_freeze_queue_start(q
);
129 blk_mq_freeze_queue_wait(q
);
132 void blk_mq_freeze_queue(struct request_queue
*q
)
135 * ...just an alias to keep freeze and unfreeze actions balanced
136 * in the blk_mq_* namespace
140 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
142 void blk_mq_unfreeze_queue(struct request_queue
*q
)
146 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
147 WARN_ON_ONCE(freeze_depth
< 0);
149 percpu_ref_reinit(&q
->q_usage_counter
);
150 wake_up_all(&q
->mq_freeze_wq
);
153 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
156 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
159 * Note: this function does not prevent that the struct request end_io()
160 * callback function is invoked. Additionally, it is not prevented that
161 * new queue_rq() calls occur unless the queue has been stopped first.
163 void blk_mq_quiesce_queue(struct request_queue
*q
)
165 struct blk_mq_hw_ctx
*hctx
;
169 blk_mq_stop_hw_queues(q
);
171 queue_for_each_hw_ctx(q
, hctx
, i
) {
172 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
173 synchronize_srcu(&hctx
->queue_rq_srcu
);
180 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
182 void blk_mq_wake_waiters(struct request_queue
*q
)
184 struct blk_mq_hw_ctx
*hctx
;
187 queue_for_each_hw_ctx(q
, hctx
, i
)
188 if (blk_mq_hw_queue_mapped(hctx
))
189 blk_mq_tag_wakeup_all(hctx
->tags
, true);
192 * If we are called because the queue has now been marked as
193 * dying, we need to ensure that processes currently waiting on
194 * the queue are notified as well.
196 wake_up_all(&q
->mq_freeze_wq
);
199 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
201 return blk_mq_has_free_tags(hctx
->tags
);
203 EXPORT_SYMBOL(blk_mq_can_queue
);
205 void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
206 struct request
*rq
, unsigned int op
)
208 INIT_LIST_HEAD(&rq
->queuelist
);
209 /* csd/requeue_work/fifo_time is initialized before use */
213 if (blk_queue_io_stat(q
))
214 rq
->rq_flags
|= RQF_IO_STAT
;
215 /* do not touch atomic flags, it needs atomic ops against the timer */
217 INIT_HLIST_NODE(&rq
->hash
);
218 RB_CLEAR_NODE(&rq
->rb_node
);
221 rq
->start_time
= jiffies
;
222 #ifdef CONFIG_BLK_CGROUP
224 set_start_time_ns(rq
);
225 rq
->io_start_time_ns
= 0;
227 rq
->nr_phys_segments
= 0;
228 #if defined(CONFIG_BLK_DEV_INTEGRITY)
229 rq
->nr_integrity_segments
= 0;
232 /* tag was already set */
235 INIT_LIST_HEAD(&rq
->timeout_list
);
239 rq
->end_io_data
= NULL
;
242 ctx
->rq_dispatched
[op_is_sync(op
)]++;
244 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init
);
246 struct request
*__blk_mq_alloc_request(struct blk_mq_alloc_data
*data
,
252 tag
= blk_mq_get_tag(data
);
253 if (tag
!= BLK_MQ_TAG_FAIL
) {
254 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
256 rq
= tags
->static_rqs
[tag
];
258 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
260 rq
->internal_tag
= tag
;
262 if (blk_mq_tag_busy(data
->hctx
)) {
263 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
264 atomic_inc(&data
->hctx
->nr_active
);
267 rq
->internal_tag
= -1;
268 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
271 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
);
277 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request
);
279 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
282 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
286 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
290 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
292 blk_mq_put_ctx(alloc_data
.ctx
);
296 return ERR_PTR(-EWOULDBLOCK
);
299 rq
->__sector
= (sector_t
) -1;
300 rq
->bio
= rq
->biotail
= NULL
;
303 EXPORT_SYMBOL(blk_mq_alloc_request
);
305 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
306 unsigned int flags
, unsigned int hctx_idx
)
308 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
314 * If the tag allocator sleeps we could get an allocation for a
315 * different hardware context. No need to complicate the low level
316 * allocator for this for the rare use case of a command tied to
319 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
320 return ERR_PTR(-EINVAL
);
322 if (hctx_idx
>= q
->nr_hw_queues
)
323 return ERR_PTR(-EIO
);
325 ret
= blk_queue_enter(q
, true);
330 * Check if the hardware context is actually mapped to anything.
331 * If not tell the caller that it should skip this queue.
333 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
334 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
336 return ERR_PTR(-EXDEV
);
338 cpu
= cpumask_first(alloc_data
.hctx
->cpumask
);
339 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
341 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
346 return ERR_PTR(-EWOULDBLOCK
);
350 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
352 void __blk_mq_finish_request(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
355 const int sched_tag
= rq
->internal_tag
;
356 struct request_queue
*q
= rq
->q
;
358 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
359 atomic_dec(&hctx
->nr_active
);
361 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
364 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
365 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
367 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
369 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
370 blk_mq_sched_restart(hctx
);
374 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx
*hctx
,
377 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
379 ctx
->rq_completed
[rq_is_sync(rq
)]++;
380 __blk_mq_finish_request(hctx
, ctx
, rq
);
383 void blk_mq_finish_request(struct request
*rq
)
385 blk_mq_finish_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
387 EXPORT_SYMBOL_GPL(blk_mq_finish_request
);
389 void blk_mq_free_request(struct request
*rq
)
391 blk_mq_sched_put_request(rq
);
393 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
395 inline void __blk_mq_end_request(struct request
*rq
, int error
)
397 blk_account_io_done(rq
);
400 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
401 rq
->end_io(rq
, error
);
403 if (unlikely(blk_bidi_rq(rq
)))
404 blk_mq_free_request(rq
->next_rq
);
405 blk_mq_free_request(rq
);
408 EXPORT_SYMBOL(__blk_mq_end_request
);
410 void blk_mq_end_request(struct request
*rq
, int error
)
412 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
414 __blk_mq_end_request(rq
, error
);
416 EXPORT_SYMBOL(blk_mq_end_request
);
418 static void __blk_mq_complete_request_remote(void *data
)
420 struct request
*rq
= data
;
422 rq
->q
->softirq_done_fn(rq
);
425 static void __blk_mq_complete_request(struct request
*rq
)
427 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
431 if (rq
->internal_tag
!= -1)
432 blk_mq_sched_completed_request(rq
);
433 if (rq
->rq_flags
& RQF_STATS
) {
434 blk_mq_poll_stats_start(rq
->q
);
438 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
439 rq
->q
->softirq_done_fn(rq
);
444 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
445 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
447 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
448 rq
->csd
.func
= __blk_mq_complete_request_remote
;
451 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
453 rq
->q
->softirq_done_fn(rq
);
459 * blk_mq_complete_request - end I/O on a request
460 * @rq: the request being processed
463 * Ends all I/O on a request. It does not handle partial completions.
464 * The actual completion happens out-of-order, through a IPI handler.
466 void blk_mq_complete_request(struct request
*rq
)
468 struct request_queue
*q
= rq
->q
;
470 if (unlikely(blk_should_fake_timeout(q
)))
472 if (!blk_mark_rq_complete(rq
))
473 __blk_mq_complete_request(rq
);
475 EXPORT_SYMBOL(blk_mq_complete_request
);
477 int blk_mq_request_started(struct request
*rq
)
479 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
481 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
483 void blk_mq_start_request(struct request
*rq
)
485 struct request_queue
*q
= rq
->q
;
487 blk_mq_sched_started_request(rq
);
489 trace_block_rq_issue(q
, rq
);
491 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
492 blk_stat_set_issue(&rq
->issue_stat
, blk_rq_sectors(rq
));
493 rq
->rq_flags
|= RQF_STATS
;
494 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
500 * Ensure that ->deadline is visible before set the started
501 * flag and clear the completed flag.
503 smp_mb__before_atomic();
506 * Mark us as started and clear complete. Complete might have been
507 * set if requeue raced with timeout, which then marked it as
508 * complete. So be sure to clear complete again when we start
509 * the request, otherwise we'll ignore the completion event.
511 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
512 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
513 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
514 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
516 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
518 * Make sure space for the drain appears. We know we can do
519 * this because max_hw_segments has been adjusted to be one
520 * fewer than the device can handle.
522 rq
->nr_phys_segments
++;
525 EXPORT_SYMBOL(blk_mq_start_request
);
528 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
529 * flag isn't set yet, so there may be race with timeout handler,
530 * but given rq->deadline is just set in .queue_rq() under
531 * this situation, the race won't be possible in reality because
532 * rq->timeout should be set as big enough to cover the window
533 * between blk_mq_start_request() called from .queue_rq() and
534 * clearing REQ_ATOM_STARTED here.
536 static void __blk_mq_requeue_request(struct request
*rq
)
538 struct request_queue
*q
= rq
->q
;
540 trace_block_rq_requeue(q
, rq
);
541 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
542 blk_mq_sched_requeue_request(rq
);
544 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
545 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
546 rq
->nr_phys_segments
--;
550 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
552 __blk_mq_requeue_request(rq
);
554 BUG_ON(blk_queued_rq(rq
));
555 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
557 EXPORT_SYMBOL(blk_mq_requeue_request
);
559 static void blk_mq_requeue_work(struct work_struct
*work
)
561 struct request_queue
*q
=
562 container_of(work
, struct request_queue
, requeue_work
.work
);
564 struct request
*rq
, *next
;
567 spin_lock_irqsave(&q
->requeue_lock
, flags
);
568 list_splice_init(&q
->requeue_list
, &rq_list
);
569 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
571 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
572 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
575 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
576 list_del_init(&rq
->queuelist
);
577 blk_mq_sched_insert_request(rq
, true, false, false, true);
580 while (!list_empty(&rq_list
)) {
581 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
582 list_del_init(&rq
->queuelist
);
583 blk_mq_sched_insert_request(rq
, false, false, false, true);
586 blk_mq_run_hw_queues(q
, false);
589 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
590 bool kick_requeue_list
)
592 struct request_queue
*q
= rq
->q
;
596 * We abuse this flag that is otherwise used by the I/O scheduler to
597 * request head insertation from the workqueue.
599 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
601 spin_lock_irqsave(&q
->requeue_lock
, flags
);
603 rq
->rq_flags
|= RQF_SOFTBARRIER
;
604 list_add(&rq
->queuelist
, &q
->requeue_list
);
606 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
608 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
610 if (kick_requeue_list
)
611 blk_mq_kick_requeue_list(q
);
613 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
615 void blk_mq_kick_requeue_list(struct request_queue
*q
)
617 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
619 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
621 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
624 kblockd_schedule_delayed_work(&q
->requeue_work
,
625 msecs_to_jiffies(msecs
));
627 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
629 void blk_mq_abort_requeue_list(struct request_queue
*q
)
634 spin_lock_irqsave(&q
->requeue_lock
, flags
);
635 list_splice_init(&q
->requeue_list
, &rq_list
);
636 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
638 while (!list_empty(&rq_list
)) {
641 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
642 list_del_init(&rq
->queuelist
);
643 blk_mq_end_request(rq
, -EIO
);
646 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
648 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
650 if (tag
< tags
->nr_tags
) {
651 prefetch(tags
->rqs
[tag
]);
652 return tags
->rqs
[tag
];
657 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
659 struct blk_mq_timeout_data
{
661 unsigned int next_set
;
664 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
666 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
667 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
670 * We know that complete is set at this point. If STARTED isn't set
671 * anymore, then the request isn't active and the "timeout" should
672 * just be ignored. This can happen due to the bitflag ordering.
673 * Timeout first checks if STARTED is set, and if it is, assumes
674 * the request is active. But if we race with completion, then
675 * both flags will get cleared. So check here again, and ignore
676 * a timeout event with a request that isn't active.
678 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
682 ret
= ops
->timeout(req
, reserved
);
686 __blk_mq_complete_request(req
);
688 case BLK_EH_RESET_TIMER
:
690 blk_clear_rq_complete(req
);
692 case BLK_EH_NOT_HANDLED
:
695 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
700 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
701 struct request
*rq
, void *priv
, bool reserved
)
703 struct blk_mq_timeout_data
*data
= priv
;
705 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
709 * The rq being checked may have been freed and reallocated
710 * out already here, we avoid this race by checking rq->deadline
711 * and REQ_ATOM_COMPLETE flag together:
713 * - if rq->deadline is observed as new value because of
714 * reusing, the rq won't be timed out because of timing.
715 * - if rq->deadline is observed as previous value,
716 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
717 * because we put a barrier between setting rq->deadline
718 * and clearing the flag in blk_mq_start_request(), so
719 * this rq won't be timed out too.
721 if (time_after_eq(jiffies
, rq
->deadline
)) {
722 if (!blk_mark_rq_complete(rq
))
723 blk_mq_rq_timed_out(rq
, reserved
);
724 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
725 data
->next
= rq
->deadline
;
730 static void blk_mq_timeout_work(struct work_struct
*work
)
732 struct request_queue
*q
=
733 container_of(work
, struct request_queue
, timeout_work
);
734 struct blk_mq_timeout_data data
= {
740 /* A deadlock might occur if a request is stuck requiring a
741 * timeout at the same time a queue freeze is waiting
742 * completion, since the timeout code would not be able to
743 * acquire the queue reference here.
745 * That's why we don't use blk_queue_enter here; instead, we use
746 * percpu_ref_tryget directly, because we need to be able to
747 * obtain a reference even in the short window between the queue
748 * starting to freeze, by dropping the first reference in
749 * blk_freeze_queue_start, and the moment the last request is
750 * consumed, marked by the instant q_usage_counter reaches
753 if (!percpu_ref_tryget(&q
->q_usage_counter
))
756 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
759 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
760 mod_timer(&q
->timeout
, data
.next
);
762 struct blk_mq_hw_ctx
*hctx
;
764 queue_for_each_hw_ctx(q
, hctx
, i
) {
765 /* the hctx may be unmapped, so check it here */
766 if (blk_mq_hw_queue_mapped(hctx
))
767 blk_mq_tag_idle(hctx
);
774 * Reverse check our software queue for entries that we could potentially
775 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
776 * too much time checking for merges.
778 static bool blk_mq_attempt_merge(struct request_queue
*q
,
779 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
784 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
790 if (!blk_rq_merge_ok(rq
, bio
))
793 switch (blk_try_merge(rq
, bio
)) {
794 case ELEVATOR_BACK_MERGE
:
795 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
796 merged
= bio_attempt_back_merge(q
, rq
, bio
);
798 case ELEVATOR_FRONT_MERGE
:
799 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
800 merged
= bio_attempt_front_merge(q
, rq
, bio
);
802 case ELEVATOR_DISCARD_MERGE
:
803 merged
= bio_attempt_discard_merge(q
, rq
, bio
);
817 struct flush_busy_ctx_data
{
818 struct blk_mq_hw_ctx
*hctx
;
819 struct list_head
*list
;
822 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
824 struct flush_busy_ctx_data
*flush_data
= data
;
825 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
826 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
828 sbitmap_clear_bit(sb
, bitnr
);
829 spin_lock(&ctx
->lock
);
830 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
831 spin_unlock(&ctx
->lock
);
836 * Process software queues that have been marked busy, splicing them
837 * to the for-dispatch
839 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
841 struct flush_busy_ctx_data data
= {
846 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
848 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
850 static inline unsigned int queued_to_index(unsigned int queued
)
855 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
858 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
861 struct blk_mq_alloc_data data
= {
863 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
864 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
870 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
871 data
.flags
|= BLK_MQ_REQ_RESERVED
;
873 rq
->tag
= blk_mq_get_tag(&data
);
875 if (blk_mq_tag_busy(data
.hctx
)) {
876 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
877 atomic_inc(&data
.hctx
->nr_active
);
879 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
885 return rq
->tag
!= -1;
888 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx
*hctx
,
891 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
894 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
895 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
896 atomic_dec(&hctx
->nr_active
);
900 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx
*hctx
,
903 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
906 __blk_mq_put_driver_tag(hctx
, rq
);
909 static void blk_mq_put_driver_tag(struct request
*rq
)
911 struct blk_mq_hw_ctx
*hctx
;
913 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
916 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
917 __blk_mq_put_driver_tag(hctx
, rq
);
921 * If we fail getting a driver tag because all the driver tags are already
922 * assigned and on the dispatch list, BUT the first entry does not have a
923 * tag, then we could deadlock. For that case, move entries with assigned
924 * driver tags to the front, leaving the set of tagged requests in the
925 * same order, and the untagged set in the same order.
927 static bool reorder_tags_to_front(struct list_head
*list
)
929 struct request
*rq
, *tmp
, *first
= NULL
;
931 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
935 list_move(&rq
->queuelist
, list
);
941 return first
!= NULL
;
944 static int blk_mq_dispatch_wake(wait_queue_t
*wait
, unsigned mode
, int flags
,
947 struct blk_mq_hw_ctx
*hctx
;
949 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
951 list_del(&wait
->task_list
);
952 clear_bit_unlock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
);
953 blk_mq_run_hw_queue(hctx
, true);
957 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx
*hctx
)
959 struct sbq_wait_state
*ws
;
962 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
963 * The thread which wins the race to grab this bit adds the hardware
964 * queue to the wait queue.
966 if (test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
) ||
967 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
970 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
971 ws
= bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
);
974 * As soon as this returns, it's no longer safe to fiddle with
975 * hctx->dispatch_wait, since a completion can wake up the wait queue
976 * and unlock the bit.
978 add_wait_queue(&ws
->wait
, &hctx
->dispatch_wait
);
982 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
)
984 struct blk_mq_hw_ctx
*hctx
;
986 int errors
, queued
, ret
= BLK_MQ_RQ_QUEUE_OK
;
988 if (list_empty(list
))
992 * Now process all the entries, sending them to the driver.
996 struct blk_mq_queue_data bd
;
998 rq
= list_first_entry(list
, struct request
, queuelist
);
999 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
1000 if (!queued
&& reorder_tags_to_front(list
))
1004 * The initial allocation attempt failed, so we need to
1005 * rerun the hardware queue when a tag is freed.
1007 if (!blk_mq_dispatch_wait_add(hctx
))
1011 * It's possible that a tag was freed in the window
1012 * between the allocation failure and adding the
1013 * hardware queue to the wait queue.
1015 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1019 list_del_init(&rq
->queuelist
);
1024 * Flag last if we have no more requests, or if we have more
1025 * but can't assign a driver tag to it.
1027 if (list_empty(list
))
1030 struct request
*nxt
;
1032 nxt
= list_first_entry(list
, struct request
, queuelist
);
1033 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1036 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1038 case BLK_MQ_RQ_QUEUE_OK
:
1041 case BLK_MQ_RQ_QUEUE_BUSY
:
1042 blk_mq_put_driver_tag_hctx(hctx
, rq
);
1043 list_add(&rq
->queuelist
, list
);
1044 __blk_mq_requeue_request(rq
);
1047 pr_err("blk-mq: bad return on queue: %d\n", ret
);
1048 case BLK_MQ_RQ_QUEUE_ERROR
:
1050 blk_mq_end_request(rq
, -EIO
);
1054 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
1056 } while (!list_empty(list
));
1058 hctx
->dispatched
[queued_to_index(queued
)]++;
1061 * Any items that need requeuing? Stuff them into hctx->dispatch,
1062 * that is where we will continue on next queue run.
1064 if (!list_empty(list
)) {
1066 * If an I/O scheduler has been configured and we got a driver
1067 * tag for the next request already, free it again.
1069 rq
= list_first_entry(list
, struct request
, queuelist
);
1070 blk_mq_put_driver_tag(rq
);
1072 spin_lock(&hctx
->lock
);
1073 list_splice_init(list
, &hctx
->dispatch
);
1074 spin_unlock(&hctx
->lock
);
1077 * If SCHED_RESTART was set by the caller of this function and
1078 * it is no longer set that means that it was cleared by another
1079 * thread and hence that a queue rerun is needed.
1081 * If TAG_WAITING is set that means that an I/O scheduler has
1082 * been configured and another thread is waiting for a driver
1083 * tag. To guarantee fairness, do not rerun this hardware queue
1084 * but let the other thread grab the driver tag.
1086 * If no I/O scheduler has been configured it is possible that
1087 * the hardware queue got stopped and restarted before requests
1088 * were pushed back onto the dispatch list. Rerun the queue to
1089 * avoid starvation. Notes:
1090 * - blk_mq_run_hw_queue() checks whether or not a queue has
1091 * been stopped before rerunning a queue.
1092 * - Some but not all block drivers stop a queue before
1093 * returning BLK_MQ_RQ_QUEUE_BUSY. Two exceptions are scsi-mq
1096 if (!blk_mq_sched_needs_restart(hctx
) &&
1097 !test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1098 blk_mq_run_hw_queue(hctx
, true);
1101 return (queued
+ errors
) != 0;
1104 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1108 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1109 cpu_online(hctx
->next_cpu
));
1111 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1113 blk_mq_sched_dispatch_requests(hctx
);
1118 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1119 blk_mq_sched_dispatch_requests(hctx
);
1120 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1125 * It'd be great if the workqueue API had a way to pass
1126 * in a mask and had some smarts for more clever placement.
1127 * For now we just round-robin here, switching for every
1128 * BLK_MQ_CPU_WORK_BATCH queued items.
1130 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1132 if (hctx
->queue
->nr_hw_queues
== 1)
1133 return WORK_CPU_UNBOUND
;
1135 if (--hctx
->next_cpu_batch
<= 0) {
1138 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1139 if (next_cpu
>= nr_cpu_ids
)
1140 next_cpu
= cpumask_first(hctx
->cpumask
);
1142 hctx
->next_cpu
= next_cpu
;
1143 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1146 return hctx
->next_cpu
;
1149 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1150 unsigned long msecs
)
1152 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
1153 !blk_mq_hw_queue_mapped(hctx
)))
1156 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1157 int cpu
= get_cpu();
1158 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1159 __blk_mq_run_hw_queue(hctx
);
1168 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
),
1171 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1172 &hctx
->delayed_run_work
,
1173 msecs_to_jiffies(msecs
));
1176 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1178 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1180 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1182 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1184 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1186 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1188 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1190 struct blk_mq_hw_ctx
*hctx
;
1193 queue_for_each_hw_ctx(q
, hctx
, i
) {
1194 if (!blk_mq_hctx_has_pending(hctx
) ||
1195 blk_mq_hctx_stopped(hctx
))
1198 blk_mq_run_hw_queue(hctx
, async
);
1201 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1204 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1205 * @q: request queue.
1207 * The caller is responsible for serializing this function against
1208 * blk_mq_{start,stop}_hw_queue().
1210 bool blk_mq_queue_stopped(struct request_queue
*q
)
1212 struct blk_mq_hw_ctx
*hctx
;
1215 queue_for_each_hw_ctx(q
, hctx
, i
)
1216 if (blk_mq_hctx_stopped(hctx
))
1221 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1223 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1225 cancel_work(&hctx
->run_work
);
1226 cancel_delayed_work(&hctx
->delay_work
);
1227 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1229 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1231 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1233 struct blk_mq_hw_ctx
*hctx
;
1236 queue_for_each_hw_ctx(q
, hctx
, i
)
1237 blk_mq_stop_hw_queue(hctx
);
1239 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1241 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1243 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1245 blk_mq_run_hw_queue(hctx
, false);
1247 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1249 void blk_mq_start_hw_queues(struct request_queue
*q
)
1251 struct blk_mq_hw_ctx
*hctx
;
1254 queue_for_each_hw_ctx(q
, hctx
, i
)
1255 blk_mq_start_hw_queue(hctx
);
1257 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1259 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1261 if (!blk_mq_hctx_stopped(hctx
))
1264 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1265 blk_mq_run_hw_queue(hctx
, async
);
1267 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1269 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1271 struct blk_mq_hw_ctx
*hctx
;
1274 queue_for_each_hw_ctx(q
, hctx
, i
)
1275 blk_mq_start_stopped_hw_queue(hctx
, async
);
1277 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1279 static void blk_mq_run_work_fn(struct work_struct
*work
)
1281 struct blk_mq_hw_ctx
*hctx
;
1283 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1285 __blk_mq_run_hw_queue(hctx
);
1288 static void blk_mq_delayed_run_work_fn(struct work_struct
*work
)
1290 struct blk_mq_hw_ctx
*hctx
;
1292 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delayed_run_work
.work
);
1294 __blk_mq_run_hw_queue(hctx
);
1297 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1299 struct blk_mq_hw_ctx
*hctx
;
1301 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1303 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1304 __blk_mq_run_hw_queue(hctx
);
1307 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1309 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1312 blk_mq_stop_hw_queue(hctx
);
1313 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1314 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1316 EXPORT_SYMBOL(blk_mq_delay_queue
);
1318 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1322 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1324 trace_block_rq_insert(hctx
->queue
, rq
);
1327 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1329 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1332 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1335 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1337 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1338 blk_mq_hctx_mark_pending(hctx
, ctx
);
1341 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1342 struct list_head
*list
)
1346 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1349 spin_lock(&ctx
->lock
);
1350 while (!list_empty(list
)) {
1353 rq
= list_first_entry(list
, struct request
, queuelist
);
1354 BUG_ON(rq
->mq_ctx
!= ctx
);
1355 list_del_init(&rq
->queuelist
);
1356 __blk_mq_insert_req_list(hctx
, rq
, false);
1358 blk_mq_hctx_mark_pending(hctx
, ctx
);
1359 spin_unlock(&ctx
->lock
);
1362 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1364 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1365 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1367 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1368 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1369 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1372 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1374 struct blk_mq_ctx
*this_ctx
;
1375 struct request_queue
*this_q
;
1378 LIST_HEAD(ctx_list
);
1381 list_splice_init(&plug
->mq_list
, &list
);
1383 list_sort(NULL
, &list
, plug_ctx_cmp
);
1389 while (!list_empty(&list
)) {
1390 rq
= list_entry_rq(list
.next
);
1391 list_del_init(&rq
->queuelist
);
1393 if (rq
->mq_ctx
!= this_ctx
) {
1395 trace_block_unplug(this_q
, depth
, from_schedule
);
1396 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1401 this_ctx
= rq
->mq_ctx
;
1407 list_add_tail(&rq
->queuelist
, &ctx_list
);
1411 * If 'this_ctx' is set, we know we have entries to complete
1412 * on 'ctx_list'. Do those.
1415 trace_block_unplug(this_q
, depth
, from_schedule
);
1416 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1421 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1423 blk_init_request_from_bio(rq
, bio
);
1425 blk_account_io_start(rq
, true);
1428 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1430 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1431 !blk_queue_nomerges(hctx
->queue
);
1434 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1435 struct blk_mq_ctx
*ctx
,
1436 struct request
*rq
, struct bio
*bio
)
1438 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1439 blk_mq_bio_to_request(rq
, bio
);
1440 spin_lock(&ctx
->lock
);
1442 __blk_mq_insert_request(hctx
, rq
, false);
1443 spin_unlock(&ctx
->lock
);
1446 struct request_queue
*q
= hctx
->queue
;
1448 spin_lock(&ctx
->lock
);
1449 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1450 blk_mq_bio_to_request(rq
, bio
);
1454 spin_unlock(&ctx
->lock
);
1455 __blk_mq_finish_request(hctx
, ctx
, rq
);
1460 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1463 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1465 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1468 static void __blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
,
1471 struct request_queue
*q
= rq
->q
;
1472 struct blk_mq_queue_data bd
= {
1476 struct blk_mq_hw_ctx
*hctx
;
1477 blk_qc_t new_cookie
;
1483 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1486 new_cookie
= request_to_qc_t(hctx
, rq
);
1489 * For OK queue, we are done. For error, kill it. Any other
1490 * error (busy), just add it to our list as we previously
1493 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1494 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1495 *cookie
= new_cookie
;
1499 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1500 *cookie
= BLK_QC_T_NONE
;
1501 blk_mq_end_request(rq
, -EIO
);
1505 __blk_mq_requeue_request(rq
);
1507 blk_mq_sched_insert_request(rq
, false, true, false, may_sleep
);
1510 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1511 struct request
*rq
, blk_qc_t
*cookie
)
1513 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1515 __blk_mq_try_issue_directly(rq
, cookie
, false);
1518 unsigned int srcu_idx
;
1522 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1523 __blk_mq_try_issue_directly(rq
, cookie
, true);
1524 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1528 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1530 const int is_sync
= op_is_sync(bio
->bi_opf
);
1531 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1532 struct blk_mq_alloc_data data
= { .flags
= 0 };
1534 unsigned int request_count
= 0;
1535 struct blk_plug
*plug
;
1536 struct request
*same_queue_rq
= NULL
;
1538 unsigned int wb_acct
;
1540 blk_queue_bounce(q
, &bio
);
1542 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1544 return BLK_QC_T_NONE
;
1547 blk_queue_split(q
, &bio
, q
->bio_split
);
1549 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1550 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1551 return BLK_QC_T_NONE
;
1553 if (blk_mq_sched_bio_merge(q
, bio
))
1554 return BLK_QC_T_NONE
;
1556 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1558 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1560 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1561 if (unlikely(!rq
)) {
1562 __wbt_done(q
->rq_wb
, wb_acct
);
1563 return BLK_QC_T_NONE
;
1566 wbt_track(&rq
->issue_stat
, wb_acct
);
1568 cookie
= request_to_qc_t(data
.hctx
, rq
);
1570 plug
= current
->plug
;
1571 if (unlikely(is_flush_fua
)) {
1572 blk_mq_put_ctx(data
.ctx
);
1573 blk_mq_bio_to_request(rq
, bio
);
1575 blk_mq_sched_insert_request(rq
, false, true, true,
1578 blk_insert_flush(rq
);
1579 blk_mq_run_hw_queue(data
.hctx
, true);
1581 } else if (plug
&& q
->nr_hw_queues
== 1) {
1582 struct request
*last
= NULL
;
1584 blk_mq_put_ctx(data
.ctx
);
1585 blk_mq_bio_to_request(rq
, bio
);
1588 * @request_count may become stale because of schedule
1589 * out, so check the list again.
1591 if (list_empty(&plug
->mq_list
))
1593 else if (blk_queue_nomerges(q
))
1594 request_count
= blk_plug_queued_count(q
);
1597 trace_block_plug(q
);
1599 last
= list_entry_rq(plug
->mq_list
.prev
);
1601 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1602 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1603 blk_flush_plug_list(plug
, false);
1604 trace_block_plug(q
);
1607 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1608 } else if (plug
&& !blk_queue_nomerges(q
)) {
1609 blk_mq_bio_to_request(rq
, bio
);
1612 * We do limited plugging. If the bio can be merged, do that.
1613 * Otherwise the existing request in the plug list will be
1614 * issued. So the plug list will have one request at most
1615 * The plug list might get flushed before this. If that happens,
1616 * the plug list is empty, and same_queue_rq is invalid.
1618 if (list_empty(&plug
->mq_list
))
1619 same_queue_rq
= NULL
;
1621 list_del_init(&same_queue_rq
->queuelist
);
1622 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1624 blk_mq_put_ctx(data
.ctx
);
1627 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1629 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1630 blk_mq_put_ctx(data
.ctx
);
1631 blk_mq_bio_to_request(rq
, bio
);
1632 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1633 } else if (q
->elevator
) {
1634 blk_mq_put_ctx(data
.ctx
);
1635 blk_mq_bio_to_request(rq
, bio
);
1636 blk_mq_sched_insert_request(rq
, false, true, true, true);
1637 } else if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1638 blk_mq_put_ctx(data
.ctx
);
1639 blk_mq_run_hw_queue(data
.hctx
, true);
1645 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1646 unsigned int hctx_idx
)
1650 if (tags
->rqs
&& set
->ops
->exit_request
) {
1653 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1654 struct request
*rq
= tags
->static_rqs
[i
];
1658 set
->ops
->exit_request(set
->driver_data
, rq
,
1660 tags
->static_rqs
[i
] = NULL
;
1664 while (!list_empty(&tags
->page_list
)) {
1665 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1666 list_del_init(&page
->lru
);
1668 * Remove kmemleak object previously allocated in
1669 * blk_mq_init_rq_map().
1671 kmemleak_free(page_address(page
));
1672 __free_pages(page
, page
->private);
1676 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1680 kfree(tags
->static_rqs
);
1681 tags
->static_rqs
= NULL
;
1683 blk_mq_free_tags(tags
);
1686 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1687 unsigned int hctx_idx
,
1688 unsigned int nr_tags
,
1689 unsigned int reserved_tags
)
1691 struct blk_mq_tags
*tags
;
1694 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1695 if (node
== NUMA_NO_NODE
)
1696 node
= set
->numa_node
;
1698 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1699 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1703 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1704 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1707 blk_mq_free_tags(tags
);
1711 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1712 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1714 if (!tags
->static_rqs
) {
1716 blk_mq_free_tags(tags
);
1723 static size_t order_to_size(unsigned int order
)
1725 return (size_t)PAGE_SIZE
<< order
;
1728 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1729 unsigned int hctx_idx
, unsigned int depth
)
1731 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1732 size_t rq_size
, left
;
1735 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1736 if (node
== NUMA_NO_NODE
)
1737 node
= set
->numa_node
;
1739 INIT_LIST_HEAD(&tags
->page_list
);
1742 * rq_size is the size of the request plus driver payload, rounded
1743 * to the cacheline size
1745 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1747 left
= rq_size
* depth
;
1749 for (i
= 0; i
< depth
; ) {
1750 int this_order
= max_order
;
1755 while (this_order
&& left
< order_to_size(this_order
- 1))
1759 page
= alloc_pages_node(node
,
1760 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1766 if (order_to_size(this_order
) < rq_size
)
1773 page
->private = this_order
;
1774 list_add_tail(&page
->lru
, &tags
->page_list
);
1776 p
= page_address(page
);
1778 * Allow kmemleak to scan these pages as they contain pointers
1779 * to additional allocations like via ops->init_request().
1781 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1782 entries_per_page
= order_to_size(this_order
) / rq_size
;
1783 to_do
= min(entries_per_page
, depth
- i
);
1784 left
-= to_do
* rq_size
;
1785 for (j
= 0; j
< to_do
; j
++) {
1786 struct request
*rq
= p
;
1788 tags
->static_rqs
[i
] = rq
;
1789 if (set
->ops
->init_request
) {
1790 if (set
->ops
->init_request(set
->driver_data
,
1793 tags
->static_rqs
[i
] = NULL
;
1805 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1810 * 'cpu' is going away. splice any existing rq_list entries from this
1811 * software queue to the hw queue dispatch list, and ensure that it
1814 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1816 struct blk_mq_hw_ctx
*hctx
;
1817 struct blk_mq_ctx
*ctx
;
1820 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1821 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1823 spin_lock(&ctx
->lock
);
1824 if (!list_empty(&ctx
->rq_list
)) {
1825 list_splice_init(&ctx
->rq_list
, &tmp
);
1826 blk_mq_hctx_clear_pending(hctx
, ctx
);
1828 spin_unlock(&ctx
->lock
);
1830 if (list_empty(&tmp
))
1833 spin_lock(&hctx
->lock
);
1834 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1835 spin_unlock(&hctx
->lock
);
1837 blk_mq_run_hw_queue(hctx
, true);
1841 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1843 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1847 /* hctx->ctxs will be freed in queue's release handler */
1848 static void blk_mq_exit_hctx(struct request_queue
*q
,
1849 struct blk_mq_tag_set
*set
,
1850 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1852 unsigned flush_start_tag
= set
->queue_depth
;
1854 blk_mq_tag_idle(hctx
);
1856 if (set
->ops
->exit_request
)
1857 set
->ops
->exit_request(set
->driver_data
,
1858 hctx
->fq
->flush_rq
, hctx_idx
,
1859 flush_start_tag
+ hctx_idx
);
1861 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1863 if (set
->ops
->exit_hctx
)
1864 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1866 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1867 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1869 blk_mq_remove_cpuhp(hctx
);
1870 blk_free_flush_queue(hctx
->fq
);
1871 sbitmap_free(&hctx
->ctx_map
);
1874 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1875 struct blk_mq_tag_set
*set
, int nr_queue
)
1877 struct blk_mq_hw_ctx
*hctx
;
1880 queue_for_each_hw_ctx(q
, hctx
, i
) {
1883 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1887 static int blk_mq_init_hctx(struct request_queue
*q
,
1888 struct blk_mq_tag_set
*set
,
1889 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1892 unsigned flush_start_tag
= set
->queue_depth
;
1894 node
= hctx
->numa_node
;
1895 if (node
== NUMA_NO_NODE
)
1896 node
= hctx
->numa_node
= set
->numa_node
;
1898 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1899 INIT_DELAYED_WORK(&hctx
->delayed_run_work
, blk_mq_delayed_run_work_fn
);
1900 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1901 spin_lock_init(&hctx
->lock
);
1902 INIT_LIST_HEAD(&hctx
->dispatch
);
1904 hctx
->queue_num
= hctx_idx
;
1905 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1907 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1909 hctx
->tags
= set
->tags
[hctx_idx
];
1912 * Allocate space for all possible cpus to avoid allocation at
1915 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1918 goto unregister_cpu_notifier
;
1920 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1926 if (set
->ops
->init_hctx
&&
1927 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1930 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
1933 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1935 goto sched_exit_hctx
;
1937 if (set
->ops
->init_request
&&
1938 set
->ops
->init_request(set
->driver_data
,
1939 hctx
->fq
->flush_rq
, hctx_idx
,
1940 flush_start_tag
+ hctx_idx
, node
))
1943 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1944 init_srcu_struct(&hctx
->queue_rq_srcu
);
1951 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1953 if (set
->ops
->exit_hctx
)
1954 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1956 sbitmap_free(&hctx
->ctx_map
);
1959 unregister_cpu_notifier
:
1960 blk_mq_remove_cpuhp(hctx
);
1964 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1965 unsigned int nr_hw_queues
)
1969 for_each_possible_cpu(i
) {
1970 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1971 struct blk_mq_hw_ctx
*hctx
;
1974 spin_lock_init(&__ctx
->lock
);
1975 INIT_LIST_HEAD(&__ctx
->rq_list
);
1978 /* If the cpu isn't online, the cpu is mapped to first hctx */
1982 hctx
= blk_mq_map_queue(q
, i
);
1985 * Set local node, IFF we have more than one hw queue. If
1986 * not, we remain on the home node of the device
1988 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1989 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1993 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
1997 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
1998 set
->queue_depth
, set
->reserved_tags
);
1999 if (!set
->tags
[hctx_idx
])
2002 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2007 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2008 set
->tags
[hctx_idx
] = NULL
;
2012 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2013 unsigned int hctx_idx
)
2015 if (set
->tags
[hctx_idx
]) {
2016 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2017 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2018 set
->tags
[hctx_idx
] = NULL
;
2022 static void blk_mq_map_swqueue(struct request_queue
*q
,
2023 const struct cpumask
*online_mask
)
2025 unsigned int i
, hctx_idx
;
2026 struct blk_mq_hw_ctx
*hctx
;
2027 struct blk_mq_ctx
*ctx
;
2028 struct blk_mq_tag_set
*set
= q
->tag_set
;
2031 * Avoid others reading imcomplete hctx->cpumask through sysfs
2033 mutex_lock(&q
->sysfs_lock
);
2035 queue_for_each_hw_ctx(q
, hctx
, i
) {
2036 cpumask_clear(hctx
->cpumask
);
2041 * Map software to hardware queues
2043 for_each_possible_cpu(i
) {
2044 /* If the cpu isn't online, the cpu is mapped to first hctx */
2045 if (!cpumask_test_cpu(i
, online_mask
))
2048 hctx_idx
= q
->mq_map
[i
];
2049 /* unmapped hw queue can be remapped after CPU topo changed */
2050 if (!set
->tags
[hctx_idx
] &&
2051 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2053 * If tags initialization fail for some hctx,
2054 * that hctx won't be brought online. In this
2055 * case, remap the current ctx to hctx[0] which
2056 * is guaranteed to always have tags allocated
2061 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2062 hctx
= blk_mq_map_queue(q
, i
);
2064 cpumask_set_cpu(i
, hctx
->cpumask
);
2065 ctx
->index_hw
= hctx
->nr_ctx
;
2066 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2069 mutex_unlock(&q
->sysfs_lock
);
2071 queue_for_each_hw_ctx(q
, hctx
, i
) {
2073 * If no software queues are mapped to this hardware queue,
2074 * disable it and free the request entries.
2076 if (!hctx
->nr_ctx
) {
2077 /* Never unmap queue 0. We need it as a
2078 * fallback in case of a new remap fails
2081 if (i
&& set
->tags
[i
])
2082 blk_mq_free_map_and_requests(set
, i
);
2088 hctx
->tags
= set
->tags
[i
];
2089 WARN_ON(!hctx
->tags
);
2092 * Set the map size to the number of mapped software queues.
2093 * This is more accurate and more efficient than looping
2094 * over all possibly mapped software queues.
2096 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2099 * Initialize batch roundrobin counts
2101 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2102 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2106 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2108 struct blk_mq_hw_ctx
*hctx
;
2111 queue_for_each_hw_ctx(q
, hctx
, i
) {
2113 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2115 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2119 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
2121 struct request_queue
*q
;
2123 lockdep_assert_held(&set
->tag_list_lock
);
2125 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2126 blk_mq_freeze_queue(q
);
2127 queue_set_hctx_shared(q
, shared
);
2128 blk_mq_unfreeze_queue(q
);
2132 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2134 struct blk_mq_tag_set
*set
= q
->tag_set
;
2136 mutex_lock(&set
->tag_list_lock
);
2137 list_del_rcu(&q
->tag_set_list
);
2138 INIT_LIST_HEAD(&q
->tag_set_list
);
2139 if (list_is_singular(&set
->tag_list
)) {
2140 /* just transitioned to unshared */
2141 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2142 /* update existing queue */
2143 blk_mq_update_tag_set_depth(set
, false);
2145 mutex_unlock(&set
->tag_list_lock
);
2150 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2151 struct request_queue
*q
)
2155 mutex_lock(&set
->tag_list_lock
);
2157 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2158 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2159 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2160 /* update existing queue */
2161 blk_mq_update_tag_set_depth(set
, true);
2163 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2164 queue_set_hctx_shared(q
, true);
2165 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2167 mutex_unlock(&set
->tag_list_lock
);
2171 * It is the actual release handler for mq, but we do it from
2172 * request queue's release handler for avoiding use-after-free
2173 * and headache because q->mq_kobj shouldn't have been introduced,
2174 * but we can't group ctx/kctx kobj without it.
2176 void blk_mq_release(struct request_queue
*q
)
2178 struct blk_mq_hw_ctx
*hctx
;
2181 /* hctx kobj stays in hctx */
2182 queue_for_each_hw_ctx(q
, hctx
, i
) {
2185 kobject_put(&hctx
->kobj
);
2190 kfree(q
->queue_hw_ctx
);
2193 * release .mq_kobj and sw queue's kobject now because
2194 * both share lifetime with request queue.
2196 blk_mq_sysfs_deinit(q
);
2198 free_percpu(q
->queue_ctx
);
2201 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2203 struct request_queue
*uninit_q
, *q
;
2205 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2207 return ERR_PTR(-ENOMEM
);
2209 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2211 blk_cleanup_queue(uninit_q
);
2215 EXPORT_SYMBOL(blk_mq_init_queue
);
2217 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2218 struct request_queue
*q
)
2221 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2223 blk_mq_sysfs_unregister(q
);
2224 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2230 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2231 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2236 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2243 atomic_set(&hctxs
[i
]->nr_active
, 0);
2244 hctxs
[i
]->numa_node
= node
;
2245 hctxs
[i
]->queue_num
= i
;
2247 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2248 free_cpumask_var(hctxs
[i
]->cpumask
);
2253 blk_mq_hctx_kobj_init(hctxs
[i
]);
2255 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2256 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2260 blk_mq_free_map_and_requests(set
, j
);
2261 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2262 kobject_put(&hctx
->kobj
);
2267 q
->nr_hw_queues
= i
;
2268 blk_mq_sysfs_register(q
);
2271 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2272 struct request_queue
*q
)
2274 /* mark the queue as mq asap */
2275 q
->mq_ops
= set
->ops
;
2277 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2278 blk_mq_poll_stats_bkt
,
2279 BLK_MQ_POLL_STATS_BKTS
, q
);
2283 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2287 /* init q->mq_kobj and sw queues' kobjects */
2288 blk_mq_sysfs_init(q
);
2290 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2291 GFP_KERNEL
, set
->numa_node
);
2292 if (!q
->queue_hw_ctx
)
2295 q
->mq_map
= set
->mq_map
;
2297 blk_mq_realloc_hw_ctxs(set
, q
);
2298 if (!q
->nr_hw_queues
)
2301 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2302 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2304 q
->nr_queues
= nr_cpu_ids
;
2306 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2308 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2309 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2311 q
->sg_reserved_size
= INT_MAX
;
2313 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2314 INIT_LIST_HEAD(&q
->requeue_list
);
2315 spin_lock_init(&q
->requeue_lock
);
2317 blk_queue_make_request(q
, blk_mq_make_request
);
2320 * Do this after blk_queue_make_request() overrides it...
2322 q
->nr_requests
= set
->queue_depth
;
2325 * Default to classic polling
2329 if (set
->ops
->complete
)
2330 blk_queue_softirq_done(q
, set
->ops
->complete
);
2332 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2335 mutex_lock(&all_q_mutex
);
2337 list_add_tail(&q
->all_q_node
, &all_q_list
);
2338 blk_mq_add_queue_tag_set(set
, q
);
2339 blk_mq_map_swqueue(q
, cpu_online_mask
);
2341 mutex_unlock(&all_q_mutex
);
2344 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2347 ret
= blk_mq_sched_init(q
);
2349 return ERR_PTR(ret
);
2355 kfree(q
->queue_hw_ctx
);
2357 free_percpu(q
->queue_ctx
);
2360 return ERR_PTR(-ENOMEM
);
2362 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2364 void blk_mq_free_queue(struct request_queue
*q
)
2366 struct blk_mq_tag_set
*set
= q
->tag_set
;
2368 mutex_lock(&all_q_mutex
);
2369 list_del_init(&q
->all_q_node
);
2370 mutex_unlock(&all_q_mutex
);
2372 blk_mq_del_queue_tag_set(q
);
2374 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2377 /* Basically redo blk_mq_init_queue with queue frozen */
2378 static void blk_mq_queue_reinit(struct request_queue
*q
,
2379 const struct cpumask
*online_mask
)
2381 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2383 blk_mq_sysfs_unregister(q
);
2386 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2387 * we should change hctx numa_node according to new topology (this
2388 * involves free and re-allocate memory, worthy doing?)
2391 blk_mq_map_swqueue(q
, online_mask
);
2393 blk_mq_sysfs_register(q
);
2397 * New online cpumask which is going to be set in this hotplug event.
2398 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2399 * one-by-one and dynamically allocating this could result in a failure.
2401 static struct cpumask cpuhp_online_new
;
2403 static void blk_mq_queue_reinit_work(void)
2405 struct request_queue
*q
;
2407 mutex_lock(&all_q_mutex
);
2409 * We need to freeze and reinit all existing queues. Freezing
2410 * involves synchronous wait for an RCU grace period and doing it
2411 * one by one may take a long time. Start freezing all queues in
2412 * one swoop and then wait for the completions so that freezing can
2413 * take place in parallel.
2415 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2416 blk_freeze_queue_start(q
);
2417 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2418 blk_mq_freeze_queue_wait(q
);
2420 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2421 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2423 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2424 blk_mq_unfreeze_queue(q
);
2426 mutex_unlock(&all_q_mutex
);
2429 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2431 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2432 blk_mq_queue_reinit_work();
2437 * Before hotadded cpu starts handling requests, new mappings must be
2438 * established. Otherwise, these requests in hw queue might never be
2441 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2442 * for CPU0, and ctx1 for CPU1).
2444 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2445 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2447 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2448 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2449 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2452 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2454 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2455 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2456 blk_mq_queue_reinit_work();
2460 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2464 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2465 if (!__blk_mq_alloc_rq_map(set
, i
))
2472 blk_mq_free_rq_map(set
->tags
[i
]);
2478 * Allocate the request maps associated with this tag_set. Note that this
2479 * may reduce the depth asked for, if memory is tight. set->queue_depth
2480 * will be updated to reflect the allocated depth.
2482 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2487 depth
= set
->queue_depth
;
2489 err
= __blk_mq_alloc_rq_maps(set
);
2493 set
->queue_depth
>>= 1;
2494 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2498 } while (set
->queue_depth
);
2500 if (!set
->queue_depth
|| err
) {
2501 pr_err("blk-mq: failed to allocate request map\n");
2505 if (depth
!= set
->queue_depth
)
2506 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2507 depth
, set
->queue_depth
);
2512 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2514 if (set
->ops
->map_queues
)
2515 return set
->ops
->map_queues(set
);
2517 return blk_mq_map_queues(set
);
2521 * Alloc a tag set to be associated with one or more request queues.
2522 * May fail with EINVAL for various error conditions. May adjust the
2523 * requested depth down, if if it too large. In that case, the set
2524 * value will be stored in set->queue_depth.
2526 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2530 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2532 if (!set
->nr_hw_queues
)
2534 if (!set
->queue_depth
)
2536 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2539 if (!set
->ops
->queue_rq
)
2542 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2543 pr_info("blk-mq: reduced tag depth to %u\n",
2545 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2549 * If a crashdump is active, then we are potentially in a very
2550 * memory constrained environment. Limit us to 1 queue and
2551 * 64 tags to prevent using too much memory.
2553 if (is_kdump_kernel()) {
2554 set
->nr_hw_queues
= 1;
2555 set
->queue_depth
= min(64U, set
->queue_depth
);
2558 * There is no use for more h/w queues than cpus.
2560 if (set
->nr_hw_queues
> nr_cpu_ids
)
2561 set
->nr_hw_queues
= nr_cpu_ids
;
2563 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2564 GFP_KERNEL
, set
->numa_node
);
2569 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2570 GFP_KERNEL
, set
->numa_node
);
2574 ret
= blk_mq_update_queue_map(set
);
2576 goto out_free_mq_map
;
2578 ret
= blk_mq_alloc_rq_maps(set
);
2580 goto out_free_mq_map
;
2582 mutex_init(&set
->tag_list_lock
);
2583 INIT_LIST_HEAD(&set
->tag_list
);
2595 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2597 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2601 for (i
= 0; i
< nr_cpu_ids
; i
++)
2602 blk_mq_free_map_and_requests(set
, i
);
2610 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2612 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2614 struct blk_mq_tag_set
*set
= q
->tag_set
;
2615 struct blk_mq_hw_ctx
*hctx
;
2621 blk_mq_freeze_queue(q
);
2622 blk_mq_quiesce_queue(q
);
2625 queue_for_each_hw_ctx(q
, hctx
, i
) {
2629 * If we're using an MQ scheduler, just update the scheduler
2630 * queue depth. This is similar to what the old code would do.
2632 if (!hctx
->sched_tags
) {
2633 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2634 min(nr
, set
->queue_depth
),
2637 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2645 q
->nr_requests
= nr
;
2647 blk_mq_unfreeze_queue(q
);
2648 blk_mq_start_stopped_hw_queues(q
, true);
2653 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2655 struct request_queue
*q
;
2657 lockdep_assert_held(&set
->tag_list_lock
);
2659 if (nr_hw_queues
> nr_cpu_ids
)
2660 nr_hw_queues
= nr_cpu_ids
;
2661 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2664 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2665 blk_mq_freeze_queue(q
);
2667 set
->nr_hw_queues
= nr_hw_queues
;
2668 blk_mq_update_queue_map(set
);
2669 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2670 blk_mq_realloc_hw_ctxs(set
, q
);
2671 blk_mq_queue_reinit(q
, cpu_online_mask
);
2674 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2675 blk_mq_unfreeze_queue(q
);
2677 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2679 /* Enable polling stats and return whether they were already enabled. */
2680 static bool blk_poll_stats_enable(struct request_queue
*q
)
2682 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2683 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
2685 blk_stat_add_callback(q
, q
->poll_cb
);
2689 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2692 * We don't arm the callback if polling stats are not enabled or the
2693 * callback is already active.
2695 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2696 blk_stat_is_active(q
->poll_cb
))
2699 blk_stat_activate_msecs(q
->poll_cb
, 100);
2702 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2704 struct request_queue
*q
= cb
->data
;
2707 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
2708 if (cb
->stat
[bucket
].nr_samples
)
2709 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
2713 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2714 struct blk_mq_hw_ctx
*hctx
,
2717 unsigned long ret
= 0;
2721 * If stats collection isn't on, don't sleep but turn it on for
2724 if (!blk_poll_stats_enable(q
))
2728 * As an optimistic guess, use half of the mean service time
2729 * for this type of request. We can (and should) make this smarter.
2730 * For instance, if the completion latencies are tight, we can
2731 * get closer than just half the mean. This is especially
2732 * important on devices where the completion latencies are longer
2733 * than ~10 usec. We do use the stats for the relevant IO size
2734 * if available which does lead to better estimates.
2736 bucket
= blk_mq_poll_stats_bkt(rq
);
2740 if (q
->poll_stat
[bucket
].nr_samples
)
2741 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
2746 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2747 struct blk_mq_hw_ctx
*hctx
,
2750 struct hrtimer_sleeper hs
;
2751 enum hrtimer_mode mode
;
2755 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2761 * -1: don't ever hybrid sleep
2762 * 0: use half of prev avg
2763 * >0: use this specific value
2765 if (q
->poll_nsec
== -1)
2767 else if (q
->poll_nsec
> 0)
2768 nsecs
= q
->poll_nsec
;
2770 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2775 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2778 * This will be replaced with the stats tracking code, using
2779 * 'avg_completion_time / 2' as the pre-sleep target.
2783 mode
= HRTIMER_MODE_REL
;
2784 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2785 hrtimer_set_expires(&hs
.timer
, kt
);
2787 hrtimer_init_sleeper(&hs
, current
);
2789 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2791 set_current_state(TASK_UNINTERRUPTIBLE
);
2792 hrtimer_start_expires(&hs
.timer
, mode
);
2795 hrtimer_cancel(&hs
.timer
);
2796 mode
= HRTIMER_MODE_ABS
;
2797 } while (hs
.task
&& !signal_pending(current
));
2799 __set_current_state(TASK_RUNNING
);
2800 destroy_hrtimer_on_stack(&hs
.timer
);
2804 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2806 struct request_queue
*q
= hctx
->queue
;
2810 * If we sleep, have the caller restart the poll loop to reset
2811 * the state. Like for the other success return cases, the
2812 * caller is responsible for checking if the IO completed. If
2813 * the IO isn't complete, we'll get called again and will go
2814 * straight to the busy poll loop.
2816 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2819 hctx
->poll_considered
++;
2821 state
= current
->state
;
2822 while (!need_resched()) {
2825 hctx
->poll_invoked
++;
2827 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2829 hctx
->poll_success
++;
2830 set_current_state(TASK_RUNNING
);
2834 if (signal_pending_state(state
, current
))
2835 set_current_state(TASK_RUNNING
);
2837 if (current
->state
== TASK_RUNNING
)
2847 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2849 struct blk_mq_hw_ctx
*hctx
;
2850 struct blk_plug
*plug
;
2853 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2854 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2857 plug
= current
->plug
;
2859 blk_flush_plug_list(plug
, false);
2861 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2862 if (!blk_qc_t_is_internal(cookie
))
2863 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2865 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2867 return __blk_mq_poll(hctx
, rq
);
2869 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2871 void blk_mq_disable_hotplug(void)
2873 mutex_lock(&all_q_mutex
);
2876 void blk_mq_enable_hotplug(void)
2878 mutex_unlock(&all_q_mutex
);
2881 static int __init
blk_mq_init(void)
2883 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2884 blk_mq_hctx_notify_dead
);
2886 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2887 blk_mq_queue_reinit_prepare
,
2888 blk_mq_queue_reinit_dead
);
2891 subsys_initcall(blk_mq_init
);