2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
29 #include <trace/events/block.h>
31 #include <linux/blk-mq.h>
34 #include "blk-mq-debugfs.h"
35 #include "blk-mq-tag.h"
38 #include "blk-mq-sched.h"
40 static DEFINE_MUTEX(all_q_mutex
);
41 static LIST_HEAD(all_q_list
);
43 static void blk_mq_poll_stats_start(struct request_queue
*q
);
44 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
45 static void __blk_mq_stop_hw_queues(struct request_queue
*q
, bool sync
);
47 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
49 int ddir
, bytes
, bucket
;
51 ddir
= rq_data_dir(rq
);
52 bytes
= blk_rq_bytes(rq
);
54 bucket
= ddir
+ 2*(ilog2(bytes
) - 9);
58 else if (bucket
>= BLK_MQ_POLL_STATS_BKTS
)
59 return ddir
+ BLK_MQ_POLL_STATS_BKTS
- 2;
65 * Check if any of the ctx's have pending work in this hardware queue
67 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
69 return sbitmap_any_bit_set(&hctx
->ctx_map
) ||
70 !list_empty_careful(&hctx
->dispatch
) ||
71 blk_mq_sched_has_work(hctx
);
75 * Mark this ctx as having pending work in this hardware queue
77 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
78 struct blk_mq_ctx
*ctx
)
80 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
81 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
84 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
85 struct blk_mq_ctx
*ctx
)
87 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
90 void blk_freeze_queue_start(struct request_queue
*q
)
94 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
95 if (freeze_depth
== 1) {
96 percpu_ref_kill(&q
->q_usage_counter
);
97 blk_mq_run_hw_queues(q
, false);
100 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
102 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
104 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
106 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
108 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
109 unsigned long timeout
)
111 return wait_event_timeout(q
->mq_freeze_wq
,
112 percpu_ref_is_zero(&q
->q_usage_counter
),
115 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
118 * Guarantee no request is in use, so we can change any data structure of
119 * the queue afterward.
121 void blk_freeze_queue(struct request_queue
*q
)
124 * In the !blk_mq case we are only calling this to kill the
125 * q_usage_counter, otherwise this increases the freeze depth
126 * and waits for it to return to zero. For this reason there is
127 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
128 * exported to drivers as the only user for unfreeze is blk_mq.
130 blk_freeze_queue_start(q
);
131 blk_mq_freeze_queue_wait(q
);
134 void blk_mq_freeze_queue(struct request_queue
*q
)
137 * ...just an alias to keep freeze and unfreeze actions balanced
138 * in the blk_mq_* namespace
142 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
144 void blk_mq_unfreeze_queue(struct request_queue
*q
)
148 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
149 WARN_ON_ONCE(freeze_depth
< 0);
151 percpu_ref_reinit(&q
->q_usage_counter
);
152 wake_up_all(&q
->mq_freeze_wq
);
155 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
158 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
161 * Note: this function does not prevent that the struct request end_io()
162 * callback function is invoked. Additionally, it is not prevented that
163 * new queue_rq() calls occur unless the queue has been stopped first.
165 void blk_mq_quiesce_queue(struct request_queue
*q
)
167 struct blk_mq_hw_ctx
*hctx
;
171 __blk_mq_stop_hw_queues(q
, true);
173 queue_for_each_hw_ctx(q
, hctx
, i
) {
174 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
175 synchronize_srcu(&hctx
->queue_rq_srcu
);
182 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
184 void blk_mq_wake_waiters(struct request_queue
*q
)
186 struct blk_mq_hw_ctx
*hctx
;
189 queue_for_each_hw_ctx(q
, hctx
, i
)
190 if (blk_mq_hw_queue_mapped(hctx
))
191 blk_mq_tag_wakeup_all(hctx
->tags
, true);
194 * If we are called because the queue has now been marked as
195 * dying, we need to ensure that processes currently waiting on
196 * the queue are notified as well.
198 wake_up_all(&q
->mq_freeze_wq
);
201 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
203 return blk_mq_has_free_tags(hctx
->tags
);
205 EXPORT_SYMBOL(blk_mq_can_queue
);
207 void blk_mq_rq_ctx_init(struct request_queue
*q
, struct blk_mq_ctx
*ctx
,
208 struct request
*rq
, unsigned int op
)
210 INIT_LIST_HEAD(&rq
->queuelist
);
211 /* csd/requeue_work/fifo_time is initialized before use */
215 if (blk_queue_io_stat(q
))
216 rq
->rq_flags
|= RQF_IO_STAT
;
217 /* do not touch atomic flags, it needs atomic ops against the timer */
219 INIT_HLIST_NODE(&rq
->hash
);
220 RB_CLEAR_NODE(&rq
->rb_node
);
223 rq
->start_time
= jiffies
;
224 #ifdef CONFIG_BLK_CGROUP
226 set_start_time_ns(rq
);
227 rq
->io_start_time_ns
= 0;
229 rq
->nr_phys_segments
= 0;
230 #if defined(CONFIG_BLK_DEV_INTEGRITY)
231 rq
->nr_integrity_segments
= 0;
234 /* tag was already set */
237 INIT_LIST_HEAD(&rq
->timeout_list
);
241 rq
->end_io_data
= NULL
;
244 ctx
->rq_dispatched
[op_is_sync(op
)]++;
246 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init
);
248 struct request
*__blk_mq_alloc_request(struct blk_mq_alloc_data
*data
,
254 tag
= blk_mq_get_tag(data
);
255 if (tag
!= BLK_MQ_TAG_FAIL
) {
256 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
258 rq
= tags
->static_rqs
[tag
];
260 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
262 rq
->internal_tag
= tag
;
264 if (blk_mq_tag_busy(data
->hctx
)) {
265 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
266 atomic_inc(&data
->hctx
->nr_active
);
269 rq
->internal_tag
= -1;
270 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
273 blk_mq_rq_ctx_init(data
->q
, data
->ctx
, rq
, op
);
279 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request
);
281 struct request
*blk_mq_alloc_request(struct request_queue
*q
, int rw
,
284 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
288 ret
= blk_queue_enter(q
, flags
& BLK_MQ_REQ_NOWAIT
);
292 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
294 blk_mq_put_ctx(alloc_data
.ctx
);
298 return ERR_PTR(-EWOULDBLOCK
);
301 rq
->__sector
= (sector_t
) -1;
302 rq
->bio
= rq
->biotail
= NULL
;
305 EXPORT_SYMBOL(blk_mq_alloc_request
);
307 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
, int rw
,
308 unsigned int flags
, unsigned int hctx_idx
)
310 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
316 * If the tag allocator sleeps we could get an allocation for a
317 * different hardware context. No need to complicate the low level
318 * allocator for this for the rare use case of a command tied to
321 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
322 return ERR_PTR(-EINVAL
);
324 if (hctx_idx
>= q
->nr_hw_queues
)
325 return ERR_PTR(-EIO
);
327 ret
= blk_queue_enter(q
, true);
332 * Check if the hardware context is actually mapped to anything.
333 * If not tell the caller that it should skip this queue.
335 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
336 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
338 return ERR_PTR(-EXDEV
);
340 cpu
= cpumask_first(alloc_data
.hctx
->cpumask
);
341 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
343 rq
= blk_mq_sched_get_request(q
, NULL
, rw
, &alloc_data
);
348 return ERR_PTR(-EWOULDBLOCK
);
352 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
354 void __blk_mq_finish_request(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
357 const int sched_tag
= rq
->internal_tag
;
358 struct request_queue
*q
= rq
->q
;
360 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
361 atomic_dec(&hctx
->nr_active
);
363 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
366 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
367 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
369 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
371 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
372 blk_mq_sched_restart(hctx
);
376 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx
*hctx
,
379 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
381 ctx
->rq_completed
[rq_is_sync(rq
)]++;
382 __blk_mq_finish_request(hctx
, ctx
, rq
);
385 void blk_mq_finish_request(struct request
*rq
)
387 blk_mq_finish_hctx_request(blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
), rq
);
389 EXPORT_SYMBOL_GPL(blk_mq_finish_request
);
391 void blk_mq_free_request(struct request
*rq
)
393 blk_mq_sched_put_request(rq
);
395 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
397 inline void __blk_mq_end_request(struct request
*rq
, int error
)
399 blk_account_io_done(rq
);
402 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
403 rq
->end_io(rq
, error
);
405 if (unlikely(blk_bidi_rq(rq
)))
406 blk_mq_free_request(rq
->next_rq
);
407 blk_mq_free_request(rq
);
410 EXPORT_SYMBOL(__blk_mq_end_request
);
412 void blk_mq_end_request(struct request
*rq
, int error
)
414 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
416 __blk_mq_end_request(rq
, error
);
418 EXPORT_SYMBOL(blk_mq_end_request
);
420 static void __blk_mq_complete_request_remote(void *data
)
422 struct request
*rq
= data
;
424 rq
->q
->softirq_done_fn(rq
);
427 static void __blk_mq_complete_request(struct request
*rq
)
429 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
433 if (rq
->internal_tag
!= -1)
434 blk_mq_sched_completed_request(rq
);
435 if (rq
->rq_flags
& RQF_STATS
) {
436 blk_mq_poll_stats_start(rq
->q
);
440 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
441 rq
->q
->softirq_done_fn(rq
);
446 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
447 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
449 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
450 rq
->csd
.func
= __blk_mq_complete_request_remote
;
453 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
455 rq
->q
->softirq_done_fn(rq
);
461 * blk_mq_complete_request - end I/O on a request
462 * @rq: the request being processed
465 * Ends all I/O on a request. It does not handle partial completions.
466 * The actual completion happens out-of-order, through a IPI handler.
468 void blk_mq_complete_request(struct request
*rq
)
470 struct request_queue
*q
= rq
->q
;
472 if (unlikely(blk_should_fake_timeout(q
)))
474 if (!blk_mark_rq_complete(rq
))
475 __blk_mq_complete_request(rq
);
477 EXPORT_SYMBOL(blk_mq_complete_request
);
479 int blk_mq_request_started(struct request
*rq
)
481 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
483 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
485 void blk_mq_start_request(struct request
*rq
)
487 struct request_queue
*q
= rq
->q
;
489 blk_mq_sched_started_request(rq
);
491 trace_block_rq_issue(q
, rq
);
493 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
494 blk_stat_set_issue(&rq
->issue_stat
, blk_rq_sectors(rq
));
495 rq
->rq_flags
|= RQF_STATS
;
496 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
502 * Ensure that ->deadline is visible before set the started
503 * flag and clear the completed flag.
505 smp_mb__before_atomic();
508 * Mark us as started and clear complete. Complete might have been
509 * set if requeue raced with timeout, which then marked it as
510 * complete. So be sure to clear complete again when we start
511 * the request, otherwise we'll ignore the completion event.
513 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
514 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
515 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
516 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
518 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
520 * Make sure space for the drain appears. We know we can do
521 * this because max_hw_segments has been adjusted to be one
522 * fewer than the device can handle.
524 rq
->nr_phys_segments
++;
527 EXPORT_SYMBOL(blk_mq_start_request
);
530 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
531 * flag isn't set yet, so there may be race with timeout handler,
532 * but given rq->deadline is just set in .queue_rq() under
533 * this situation, the race won't be possible in reality because
534 * rq->timeout should be set as big enough to cover the window
535 * between blk_mq_start_request() called from .queue_rq() and
536 * clearing REQ_ATOM_STARTED here.
538 static void __blk_mq_requeue_request(struct request
*rq
)
540 struct request_queue
*q
= rq
->q
;
542 trace_block_rq_requeue(q
, rq
);
543 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
544 blk_mq_sched_requeue_request(rq
);
546 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
547 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
548 rq
->nr_phys_segments
--;
552 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
554 __blk_mq_requeue_request(rq
);
556 BUG_ON(blk_queued_rq(rq
));
557 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
559 EXPORT_SYMBOL(blk_mq_requeue_request
);
561 static void blk_mq_requeue_work(struct work_struct
*work
)
563 struct request_queue
*q
=
564 container_of(work
, struct request_queue
, requeue_work
.work
);
566 struct request
*rq
, *next
;
569 spin_lock_irqsave(&q
->requeue_lock
, flags
);
570 list_splice_init(&q
->requeue_list
, &rq_list
);
571 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
573 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
574 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
577 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
578 list_del_init(&rq
->queuelist
);
579 blk_mq_sched_insert_request(rq
, true, false, false, true);
582 while (!list_empty(&rq_list
)) {
583 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
584 list_del_init(&rq
->queuelist
);
585 blk_mq_sched_insert_request(rq
, false, false, false, true);
588 blk_mq_run_hw_queues(q
, false);
591 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
592 bool kick_requeue_list
)
594 struct request_queue
*q
= rq
->q
;
598 * We abuse this flag that is otherwise used by the I/O scheduler to
599 * request head insertation from the workqueue.
601 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
603 spin_lock_irqsave(&q
->requeue_lock
, flags
);
605 rq
->rq_flags
|= RQF_SOFTBARRIER
;
606 list_add(&rq
->queuelist
, &q
->requeue_list
);
608 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
610 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
612 if (kick_requeue_list
)
613 blk_mq_kick_requeue_list(q
);
615 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
617 void blk_mq_kick_requeue_list(struct request_queue
*q
)
619 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
621 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
623 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
626 kblockd_schedule_delayed_work(&q
->requeue_work
,
627 msecs_to_jiffies(msecs
));
629 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
631 void blk_mq_abort_requeue_list(struct request_queue
*q
)
636 spin_lock_irqsave(&q
->requeue_lock
, flags
);
637 list_splice_init(&q
->requeue_list
, &rq_list
);
638 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
640 while (!list_empty(&rq_list
)) {
643 rq
= list_first_entry(&rq_list
, struct request
, queuelist
);
644 list_del_init(&rq
->queuelist
);
645 blk_mq_end_request(rq
, -EIO
);
648 EXPORT_SYMBOL(blk_mq_abort_requeue_list
);
650 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
652 if (tag
< tags
->nr_tags
) {
653 prefetch(tags
->rqs
[tag
]);
654 return tags
->rqs
[tag
];
659 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
661 struct blk_mq_timeout_data
{
663 unsigned int next_set
;
666 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
668 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
669 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
672 * We know that complete is set at this point. If STARTED isn't set
673 * anymore, then the request isn't active and the "timeout" should
674 * just be ignored. This can happen due to the bitflag ordering.
675 * Timeout first checks if STARTED is set, and if it is, assumes
676 * the request is active. But if we race with completion, then
677 * both flags will get cleared. So check here again, and ignore
678 * a timeout event with a request that isn't active.
680 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
684 ret
= ops
->timeout(req
, reserved
);
688 __blk_mq_complete_request(req
);
690 case BLK_EH_RESET_TIMER
:
692 blk_clear_rq_complete(req
);
694 case BLK_EH_NOT_HANDLED
:
697 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
702 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
703 struct request
*rq
, void *priv
, bool reserved
)
705 struct blk_mq_timeout_data
*data
= priv
;
707 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
711 * The rq being checked may have been freed and reallocated
712 * out already here, we avoid this race by checking rq->deadline
713 * and REQ_ATOM_COMPLETE flag together:
715 * - if rq->deadline is observed as new value because of
716 * reusing, the rq won't be timed out because of timing.
717 * - if rq->deadline is observed as previous value,
718 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
719 * because we put a barrier between setting rq->deadline
720 * and clearing the flag in blk_mq_start_request(), so
721 * this rq won't be timed out too.
723 if (time_after_eq(jiffies
, rq
->deadline
)) {
724 if (!blk_mark_rq_complete(rq
))
725 blk_mq_rq_timed_out(rq
, reserved
);
726 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
727 data
->next
= rq
->deadline
;
732 static void blk_mq_timeout_work(struct work_struct
*work
)
734 struct request_queue
*q
=
735 container_of(work
, struct request_queue
, timeout_work
);
736 struct blk_mq_timeout_data data
= {
742 /* A deadlock might occur if a request is stuck requiring a
743 * timeout at the same time a queue freeze is waiting
744 * completion, since the timeout code would not be able to
745 * acquire the queue reference here.
747 * That's why we don't use blk_queue_enter here; instead, we use
748 * percpu_ref_tryget directly, because we need to be able to
749 * obtain a reference even in the short window between the queue
750 * starting to freeze, by dropping the first reference in
751 * blk_freeze_queue_start, and the moment the last request is
752 * consumed, marked by the instant q_usage_counter reaches
755 if (!percpu_ref_tryget(&q
->q_usage_counter
))
758 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
761 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
762 mod_timer(&q
->timeout
, data
.next
);
764 struct blk_mq_hw_ctx
*hctx
;
766 queue_for_each_hw_ctx(q
, hctx
, i
) {
767 /* the hctx may be unmapped, so check it here */
768 if (blk_mq_hw_queue_mapped(hctx
))
769 blk_mq_tag_idle(hctx
);
776 * Reverse check our software queue for entries that we could potentially
777 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
778 * too much time checking for merges.
780 static bool blk_mq_attempt_merge(struct request_queue
*q
,
781 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
786 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
792 if (!blk_rq_merge_ok(rq
, bio
))
795 switch (blk_try_merge(rq
, bio
)) {
796 case ELEVATOR_BACK_MERGE
:
797 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
798 merged
= bio_attempt_back_merge(q
, rq
, bio
);
800 case ELEVATOR_FRONT_MERGE
:
801 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
802 merged
= bio_attempt_front_merge(q
, rq
, bio
);
804 case ELEVATOR_DISCARD_MERGE
:
805 merged
= bio_attempt_discard_merge(q
, rq
, bio
);
819 struct flush_busy_ctx_data
{
820 struct blk_mq_hw_ctx
*hctx
;
821 struct list_head
*list
;
824 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
826 struct flush_busy_ctx_data
*flush_data
= data
;
827 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
828 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
830 sbitmap_clear_bit(sb
, bitnr
);
831 spin_lock(&ctx
->lock
);
832 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
833 spin_unlock(&ctx
->lock
);
838 * Process software queues that have been marked busy, splicing them
839 * to the for-dispatch
841 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
843 struct flush_busy_ctx_data data
= {
848 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
850 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
852 static inline unsigned int queued_to_index(unsigned int queued
)
857 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
860 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
863 struct blk_mq_alloc_data data
= {
865 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
866 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
869 might_sleep_if(wait
);
874 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
875 data
.flags
|= BLK_MQ_REQ_RESERVED
;
877 rq
->tag
= blk_mq_get_tag(&data
);
879 if (blk_mq_tag_busy(data
.hctx
)) {
880 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
881 atomic_inc(&data
.hctx
->nr_active
);
883 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
889 return rq
->tag
!= -1;
892 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx
*hctx
,
895 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
898 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
899 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
900 atomic_dec(&hctx
->nr_active
);
904 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx
*hctx
,
907 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
910 __blk_mq_put_driver_tag(hctx
, rq
);
913 static void blk_mq_put_driver_tag(struct request
*rq
)
915 struct blk_mq_hw_ctx
*hctx
;
917 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
920 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
921 __blk_mq_put_driver_tag(hctx
, rq
);
925 * If we fail getting a driver tag because all the driver tags are already
926 * assigned and on the dispatch list, BUT the first entry does not have a
927 * tag, then we could deadlock. For that case, move entries with assigned
928 * driver tags to the front, leaving the set of tagged requests in the
929 * same order, and the untagged set in the same order.
931 static bool reorder_tags_to_front(struct list_head
*list
)
933 struct request
*rq
, *tmp
, *first
= NULL
;
935 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
939 list_move(&rq
->queuelist
, list
);
945 return first
!= NULL
;
948 static int blk_mq_dispatch_wake(wait_queue_t
*wait
, unsigned mode
, int flags
,
951 struct blk_mq_hw_ctx
*hctx
;
953 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
955 list_del(&wait
->task_list
);
956 clear_bit_unlock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
);
957 blk_mq_run_hw_queue(hctx
, true);
961 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx
*hctx
)
963 struct sbq_wait_state
*ws
;
966 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
967 * The thread which wins the race to grab this bit adds the hardware
968 * queue to the wait queue.
970 if (test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
) ||
971 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
974 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
975 ws
= bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
);
978 * As soon as this returns, it's no longer safe to fiddle with
979 * hctx->dispatch_wait, since a completion can wake up the wait queue
980 * and unlock the bit.
982 add_wait_queue(&ws
->wait
, &hctx
->dispatch_wait
);
986 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
)
988 struct blk_mq_hw_ctx
*hctx
;
990 int errors
, queued
, ret
= BLK_MQ_RQ_QUEUE_OK
;
992 if (list_empty(list
))
996 * Now process all the entries, sending them to the driver.
1000 struct blk_mq_queue_data bd
;
1002 rq
= list_first_entry(list
, struct request
, queuelist
);
1003 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
1004 if (!queued
&& reorder_tags_to_front(list
))
1008 * The initial allocation attempt failed, so we need to
1009 * rerun the hardware queue when a tag is freed.
1011 if (!blk_mq_dispatch_wait_add(hctx
))
1015 * It's possible that a tag was freed in the window
1016 * between the allocation failure and adding the
1017 * hardware queue to the wait queue.
1019 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1023 list_del_init(&rq
->queuelist
);
1028 * Flag last if we have no more requests, or if we have more
1029 * but can't assign a driver tag to it.
1031 if (list_empty(list
))
1034 struct request
*nxt
;
1036 nxt
= list_first_entry(list
, struct request
, queuelist
);
1037 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1040 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1042 case BLK_MQ_RQ_QUEUE_OK
:
1045 case BLK_MQ_RQ_QUEUE_BUSY
:
1046 blk_mq_put_driver_tag_hctx(hctx
, rq
);
1047 list_add(&rq
->queuelist
, list
);
1048 __blk_mq_requeue_request(rq
);
1051 pr_err("blk-mq: bad return on queue: %d\n", ret
);
1052 case BLK_MQ_RQ_QUEUE_ERROR
:
1054 blk_mq_end_request(rq
, -EIO
);
1058 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
1060 } while (!list_empty(list
));
1062 hctx
->dispatched
[queued_to_index(queued
)]++;
1065 * Any items that need requeuing? Stuff them into hctx->dispatch,
1066 * that is where we will continue on next queue run.
1068 if (!list_empty(list
)) {
1070 * If an I/O scheduler has been configured and we got a driver
1071 * tag for the next request already, free it again.
1073 rq
= list_first_entry(list
, struct request
, queuelist
);
1074 blk_mq_put_driver_tag(rq
);
1076 spin_lock(&hctx
->lock
);
1077 list_splice_init(list
, &hctx
->dispatch
);
1078 spin_unlock(&hctx
->lock
);
1081 * If SCHED_RESTART was set by the caller of this function and
1082 * it is no longer set that means that it was cleared by another
1083 * thread and hence that a queue rerun is needed.
1085 * If TAG_WAITING is set that means that an I/O scheduler has
1086 * been configured and another thread is waiting for a driver
1087 * tag. To guarantee fairness, do not rerun this hardware queue
1088 * but let the other thread grab the driver tag.
1090 * If no I/O scheduler has been configured it is possible that
1091 * the hardware queue got stopped and restarted before requests
1092 * were pushed back onto the dispatch list. Rerun the queue to
1093 * avoid starvation. Notes:
1094 * - blk_mq_run_hw_queue() checks whether or not a queue has
1095 * been stopped before rerunning a queue.
1096 * - Some but not all block drivers stop a queue before
1097 * returning BLK_MQ_RQ_QUEUE_BUSY. Two exceptions are scsi-mq
1100 if (!blk_mq_sched_needs_restart(hctx
) &&
1101 !test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1102 blk_mq_run_hw_queue(hctx
, true);
1105 return (queued
+ errors
) != 0;
1108 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1112 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1113 cpu_online(hctx
->next_cpu
));
1115 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1117 blk_mq_sched_dispatch_requests(hctx
);
1122 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1123 blk_mq_sched_dispatch_requests(hctx
);
1124 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1129 * It'd be great if the workqueue API had a way to pass
1130 * in a mask and had some smarts for more clever placement.
1131 * For now we just round-robin here, switching for every
1132 * BLK_MQ_CPU_WORK_BATCH queued items.
1134 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1136 if (hctx
->queue
->nr_hw_queues
== 1)
1137 return WORK_CPU_UNBOUND
;
1139 if (--hctx
->next_cpu_batch
<= 0) {
1142 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1143 if (next_cpu
>= nr_cpu_ids
)
1144 next_cpu
= cpumask_first(hctx
->cpumask
);
1146 hctx
->next_cpu
= next_cpu
;
1147 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1150 return hctx
->next_cpu
;
1153 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1154 unsigned long msecs
)
1156 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
1157 !blk_mq_hw_queue_mapped(hctx
)))
1160 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1161 int cpu
= get_cpu();
1162 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1163 __blk_mq_run_hw_queue(hctx
);
1171 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
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 static void __blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool sync
)
1226 cancel_delayed_work_sync(&hctx
->run_work
);
1228 cancel_delayed_work(&hctx
->run_work
);
1230 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1233 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1235 __blk_mq_stop_hw_queue(hctx
, false);
1237 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1239 static void __blk_mq_stop_hw_queues(struct request_queue
*q
, bool sync
)
1241 struct blk_mq_hw_ctx
*hctx
;
1244 queue_for_each_hw_ctx(q
, hctx
, i
)
1245 __blk_mq_stop_hw_queue(hctx
, sync
);
1248 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1250 __blk_mq_stop_hw_queues(q
, false);
1252 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1254 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1256 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1258 blk_mq_run_hw_queue(hctx
, false);
1260 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1262 void blk_mq_start_hw_queues(struct request_queue
*q
)
1264 struct blk_mq_hw_ctx
*hctx
;
1267 queue_for_each_hw_ctx(q
, hctx
, i
)
1268 blk_mq_start_hw_queue(hctx
);
1270 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1272 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1274 if (!blk_mq_hctx_stopped(hctx
))
1277 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1278 blk_mq_run_hw_queue(hctx
, async
);
1280 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1282 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1284 struct blk_mq_hw_ctx
*hctx
;
1287 queue_for_each_hw_ctx(q
, hctx
, i
)
1288 blk_mq_start_stopped_hw_queue(hctx
, async
);
1290 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1292 static void blk_mq_run_work_fn(struct work_struct
*work
)
1294 struct blk_mq_hw_ctx
*hctx
;
1296 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1299 * If we are stopped, don't run the queue. The exception is if
1300 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1301 * the STOPPED bit and run it.
1303 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)) {
1304 if (!test_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
))
1307 clear_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1308 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1311 __blk_mq_run_hw_queue(hctx
);
1315 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1317 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1321 * Stop the hw queue, then modify currently delayed work.
1322 * This should prevent us from running the queue prematurely.
1323 * Mark the queue as auto-clearing STOPPED when it runs.
1325 blk_mq_stop_hw_queue(hctx
);
1326 set_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1327 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1329 msecs_to_jiffies(msecs
));
1331 EXPORT_SYMBOL(blk_mq_delay_queue
);
1333 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1337 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1339 trace_block_rq_insert(hctx
->queue
, rq
);
1342 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1344 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1347 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1350 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1352 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1353 blk_mq_hctx_mark_pending(hctx
, ctx
);
1356 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1357 struct list_head
*list
)
1361 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1364 spin_lock(&ctx
->lock
);
1365 while (!list_empty(list
)) {
1368 rq
= list_first_entry(list
, struct request
, queuelist
);
1369 BUG_ON(rq
->mq_ctx
!= ctx
);
1370 list_del_init(&rq
->queuelist
);
1371 __blk_mq_insert_req_list(hctx
, rq
, false);
1373 blk_mq_hctx_mark_pending(hctx
, ctx
);
1374 spin_unlock(&ctx
->lock
);
1377 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1379 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1380 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1382 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1383 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1384 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1387 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1389 struct blk_mq_ctx
*this_ctx
;
1390 struct request_queue
*this_q
;
1393 LIST_HEAD(ctx_list
);
1396 list_splice_init(&plug
->mq_list
, &list
);
1398 list_sort(NULL
, &list
, plug_ctx_cmp
);
1404 while (!list_empty(&list
)) {
1405 rq
= list_entry_rq(list
.next
);
1406 list_del_init(&rq
->queuelist
);
1408 if (rq
->mq_ctx
!= this_ctx
) {
1410 trace_block_unplug(this_q
, depth
, from_schedule
);
1411 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1416 this_ctx
= rq
->mq_ctx
;
1422 list_add_tail(&rq
->queuelist
, &ctx_list
);
1426 * If 'this_ctx' is set, we know we have entries to complete
1427 * on 'ctx_list'. Do those.
1430 trace_block_unplug(this_q
, depth
, from_schedule
);
1431 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1436 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1438 blk_init_request_from_bio(rq
, bio
);
1440 blk_account_io_start(rq
, true);
1443 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1445 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1446 !blk_queue_nomerges(hctx
->queue
);
1449 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1450 struct blk_mq_ctx
*ctx
,
1451 struct request
*rq
, struct bio
*bio
)
1453 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1454 blk_mq_bio_to_request(rq
, bio
);
1455 spin_lock(&ctx
->lock
);
1457 __blk_mq_insert_request(hctx
, rq
, false);
1458 spin_unlock(&ctx
->lock
);
1461 struct request_queue
*q
= hctx
->queue
;
1463 spin_lock(&ctx
->lock
);
1464 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1465 blk_mq_bio_to_request(rq
, bio
);
1469 spin_unlock(&ctx
->lock
);
1470 __blk_mq_finish_request(hctx
, ctx
, rq
);
1475 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1478 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1480 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1483 static void __blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
,
1486 struct request_queue
*q
= rq
->q
;
1487 struct blk_mq_queue_data bd
= {
1491 struct blk_mq_hw_ctx
*hctx
;
1492 blk_qc_t new_cookie
;
1498 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1501 new_cookie
= request_to_qc_t(hctx
, rq
);
1504 * For OK queue, we are done. For error, kill it. Any other
1505 * error (busy), just add it to our list as we previously
1508 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1509 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1510 *cookie
= new_cookie
;
1514 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1515 *cookie
= BLK_QC_T_NONE
;
1516 blk_mq_end_request(rq
, -EIO
);
1520 __blk_mq_requeue_request(rq
);
1522 blk_mq_sched_insert_request(rq
, false, true, false, may_sleep
);
1525 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1526 struct request
*rq
, blk_qc_t
*cookie
)
1528 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1530 __blk_mq_try_issue_directly(rq
, cookie
, false);
1533 unsigned int srcu_idx
;
1537 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1538 __blk_mq_try_issue_directly(rq
, cookie
, true);
1539 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1543 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1545 const int is_sync
= op_is_sync(bio
->bi_opf
);
1546 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1547 struct blk_mq_alloc_data data
= { .flags
= 0 };
1549 unsigned int request_count
= 0;
1550 struct blk_plug
*plug
;
1551 struct request
*same_queue_rq
= NULL
;
1553 unsigned int wb_acct
;
1555 blk_queue_bounce(q
, &bio
);
1557 blk_queue_split(q
, &bio
, q
->bio_split
);
1559 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1561 return BLK_QC_T_NONE
;
1564 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1565 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1566 return BLK_QC_T_NONE
;
1568 if (blk_mq_sched_bio_merge(q
, bio
))
1569 return BLK_QC_T_NONE
;
1571 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1573 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1575 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1576 if (unlikely(!rq
)) {
1577 __wbt_done(q
->rq_wb
, wb_acct
);
1578 return BLK_QC_T_NONE
;
1581 wbt_track(&rq
->issue_stat
, wb_acct
);
1583 cookie
= request_to_qc_t(data
.hctx
, rq
);
1585 plug
= current
->plug
;
1586 if (unlikely(is_flush_fua
)) {
1587 blk_mq_put_ctx(data
.ctx
);
1588 blk_mq_bio_to_request(rq
, bio
);
1590 blk_mq_sched_insert_request(rq
, false, true, true,
1593 blk_insert_flush(rq
);
1594 blk_mq_run_hw_queue(data
.hctx
, true);
1596 } else if (plug
&& q
->nr_hw_queues
== 1) {
1597 struct request
*last
= NULL
;
1599 blk_mq_put_ctx(data
.ctx
);
1600 blk_mq_bio_to_request(rq
, bio
);
1603 * @request_count may become stale because of schedule
1604 * out, so check the list again.
1606 if (list_empty(&plug
->mq_list
))
1608 else if (blk_queue_nomerges(q
))
1609 request_count
= blk_plug_queued_count(q
);
1612 trace_block_plug(q
);
1614 last
= list_entry_rq(plug
->mq_list
.prev
);
1616 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1617 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1618 blk_flush_plug_list(plug
, false);
1619 trace_block_plug(q
);
1622 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1623 } else if (plug
&& !blk_queue_nomerges(q
)) {
1624 blk_mq_bio_to_request(rq
, bio
);
1627 * We do limited plugging. If the bio can be merged, do that.
1628 * Otherwise the existing request in the plug list will be
1629 * issued. So the plug list will have one request at most
1630 * The plug list might get flushed before this. If that happens,
1631 * the plug list is empty, and same_queue_rq is invalid.
1633 if (list_empty(&plug
->mq_list
))
1634 same_queue_rq
= NULL
;
1636 list_del_init(&same_queue_rq
->queuelist
);
1637 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1639 blk_mq_put_ctx(data
.ctx
);
1642 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1644 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1645 blk_mq_put_ctx(data
.ctx
);
1646 blk_mq_bio_to_request(rq
, bio
);
1647 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1648 } else if (q
->elevator
) {
1649 blk_mq_put_ctx(data
.ctx
);
1650 blk_mq_bio_to_request(rq
, bio
);
1651 blk_mq_sched_insert_request(rq
, false, true, true, true);
1652 } else if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1653 blk_mq_put_ctx(data
.ctx
);
1654 blk_mq_run_hw_queue(data
.hctx
, true);
1656 blk_mq_put_ctx(data
.ctx
);
1661 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1662 unsigned int hctx_idx
)
1666 if (tags
->rqs
&& set
->ops
->exit_request
) {
1669 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1670 struct request
*rq
= tags
->static_rqs
[i
];
1674 set
->ops
->exit_request(set
, rq
, hctx_idx
);
1675 tags
->static_rqs
[i
] = NULL
;
1679 while (!list_empty(&tags
->page_list
)) {
1680 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1681 list_del_init(&page
->lru
);
1683 * Remove kmemleak object previously allocated in
1684 * blk_mq_init_rq_map().
1686 kmemleak_free(page_address(page
));
1687 __free_pages(page
, page
->private);
1691 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1695 kfree(tags
->static_rqs
);
1696 tags
->static_rqs
= NULL
;
1698 blk_mq_free_tags(tags
);
1701 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1702 unsigned int hctx_idx
,
1703 unsigned int nr_tags
,
1704 unsigned int reserved_tags
)
1706 struct blk_mq_tags
*tags
;
1709 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1710 if (node
== NUMA_NO_NODE
)
1711 node
= set
->numa_node
;
1713 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1714 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1718 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1719 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1722 blk_mq_free_tags(tags
);
1726 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1727 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1729 if (!tags
->static_rqs
) {
1731 blk_mq_free_tags(tags
);
1738 static size_t order_to_size(unsigned int order
)
1740 return (size_t)PAGE_SIZE
<< order
;
1743 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1744 unsigned int hctx_idx
, unsigned int depth
)
1746 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1747 size_t rq_size
, left
;
1750 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1751 if (node
== NUMA_NO_NODE
)
1752 node
= set
->numa_node
;
1754 INIT_LIST_HEAD(&tags
->page_list
);
1757 * rq_size is the size of the request plus driver payload, rounded
1758 * to the cacheline size
1760 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1762 left
= rq_size
* depth
;
1764 for (i
= 0; i
< depth
; ) {
1765 int this_order
= max_order
;
1770 while (this_order
&& left
< order_to_size(this_order
- 1))
1774 page
= alloc_pages_node(node
,
1775 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1781 if (order_to_size(this_order
) < rq_size
)
1788 page
->private = this_order
;
1789 list_add_tail(&page
->lru
, &tags
->page_list
);
1791 p
= page_address(page
);
1793 * Allow kmemleak to scan these pages as they contain pointers
1794 * to additional allocations like via ops->init_request().
1796 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1797 entries_per_page
= order_to_size(this_order
) / rq_size
;
1798 to_do
= min(entries_per_page
, depth
- i
);
1799 left
-= to_do
* rq_size
;
1800 for (j
= 0; j
< to_do
; j
++) {
1801 struct request
*rq
= p
;
1803 tags
->static_rqs
[i
] = rq
;
1804 if (set
->ops
->init_request
) {
1805 if (set
->ops
->init_request(set
, rq
, hctx_idx
,
1807 tags
->static_rqs
[i
] = NULL
;
1819 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1824 * 'cpu' is going away. splice any existing rq_list entries from this
1825 * software queue to the hw queue dispatch list, and ensure that it
1828 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1830 struct blk_mq_hw_ctx
*hctx
;
1831 struct blk_mq_ctx
*ctx
;
1834 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1835 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1837 spin_lock(&ctx
->lock
);
1838 if (!list_empty(&ctx
->rq_list
)) {
1839 list_splice_init(&ctx
->rq_list
, &tmp
);
1840 blk_mq_hctx_clear_pending(hctx
, ctx
);
1842 spin_unlock(&ctx
->lock
);
1844 if (list_empty(&tmp
))
1847 spin_lock(&hctx
->lock
);
1848 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1849 spin_unlock(&hctx
->lock
);
1851 blk_mq_run_hw_queue(hctx
, true);
1855 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1857 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1861 /* hctx->ctxs will be freed in queue's release handler */
1862 static void blk_mq_exit_hctx(struct request_queue
*q
,
1863 struct blk_mq_tag_set
*set
,
1864 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1866 blk_mq_debugfs_unregister_hctx(hctx
);
1868 blk_mq_tag_idle(hctx
);
1870 if (set
->ops
->exit_request
)
1871 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
1873 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1875 if (set
->ops
->exit_hctx
)
1876 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1878 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1879 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1881 blk_mq_remove_cpuhp(hctx
);
1882 blk_free_flush_queue(hctx
->fq
);
1883 sbitmap_free(&hctx
->ctx_map
);
1886 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1887 struct blk_mq_tag_set
*set
, int nr_queue
)
1889 struct blk_mq_hw_ctx
*hctx
;
1892 queue_for_each_hw_ctx(q
, hctx
, i
) {
1895 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1899 static int blk_mq_init_hctx(struct request_queue
*q
,
1900 struct blk_mq_tag_set
*set
,
1901 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1905 node
= hctx
->numa_node
;
1906 if (node
== NUMA_NO_NODE
)
1907 node
= hctx
->numa_node
= set
->numa_node
;
1909 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1910 spin_lock_init(&hctx
->lock
);
1911 INIT_LIST_HEAD(&hctx
->dispatch
);
1913 hctx
->queue_num
= hctx_idx
;
1914 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1916 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1918 hctx
->tags
= set
->tags
[hctx_idx
];
1921 * Allocate space for all possible cpus to avoid allocation at
1924 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1927 goto unregister_cpu_notifier
;
1929 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1935 if (set
->ops
->init_hctx
&&
1936 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1939 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
1942 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1944 goto sched_exit_hctx
;
1946 if (set
->ops
->init_request
&&
1947 set
->ops
->init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
1951 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1952 init_srcu_struct(&hctx
->queue_rq_srcu
);
1954 blk_mq_debugfs_register_hctx(q
, hctx
);
1961 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1963 if (set
->ops
->exit_hctx
)
1964 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1966 sbitmap_free(&hctx
->ctx_map
);
1969 unregister_cpu_notifier
:
1970 blk_mq_remove_cpuhp(hctx
);
1974 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1975 unsigned int nr_hw_queues
)
1979 for_each_possible_cpu(i
) {
1980 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1981 struct blk_mq_hw_ctx
*hctx
;
1984 spin_lock_init(&__ctx
->lock
);
1985 INIT_LIST_HEAD(&__ctx
->rq_list
);
1988 /* If the cpu isn't online, the cpu is mapped to first hctx */
1992 hctx
= blk_mq_map_queue(q
, i
);
1995 * Set local node, IFF we have more than one hw queue. If
1996 * not, we remain on the home node of the device
1998 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1999 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2003 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2007 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2008 set
->queue_depth
, set
->reserved_tags
);
2009 if (!set
->tags
[hctx_idx
])
2012 ret
= blk_mq_alloc_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_free_map_and_requests(struct blk_mq_tag_set
*set
,
2023 unsigned int hctx_idx
)
2025 if (set
->tags
[hctx_idx
]) {
2026 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2027 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2028 set
->tags
[hctx_idx
] = NULL
;
2032 static void blk_mq_map_swqueue(struct request_queue
*q
,
2033 const struct cpumask
*online_mask
)
2035 unsigned int i
, hctx_idx
;
2036 struct blk_mq_hw_ctx
*hctx
;
2037 struct blk_mq_ctx
*ctx
;
2038 struct blk_mq_tag_set
*set
= q
->tag_set
;
2041 * Avoid others reading imcomplete hctx->cpumask through sysfs
2043 mutex_lock(&q
->sysfs_lock
);
2045 queue_for_each_hw_ctx(q
, hctx
, i
) {
2046 cpumask_clear(hctx
->cpumask
);
2051 * Map software to hardware queues
2053 for_each_possible_cpu(i
) {
2054 /* If the cpu isn't online, the cpu is mapped to first hctx */
2055 if (!cpumask_test_cpu(i
, online_mask
))
2058 hctx_idx
= q
->mq_map
[i
];
2059 /* unmapped hw queue can be remapped after CPU topo changed */
2060 if (!set
->tags
[hctx_idx
] &&
2061 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2063 * If tags initialization fail for some hctx,
2064 * that hctx won't be brought online. In this
2065 * case, remap the current ctx to hctx[0] which
2066 * is guaranteed to always have tags allocated
2071 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2072 hctx
= blk_mq_map_queue(q
, i
);
2074 cpumask_set_cpu(i
, hctx
->cpumask
);
2075 ctx
->index_hw
= hctx
->nr_ctx
;
2076 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2079 mutex_unlock(&q
->sysfs_lock
);
2081 queue_for_each_hw_ctx(q
, hctx
, i
) {
2083 * If no software queues are mapped to this hardware queue,
2084 * disable it and free the request entries.
2086 if (!hctx
->nr_ctx
) {
2087 /* Never unmap queue 0. We need it as a
2088 * fallback in case of a new remap fails
2091 if (i
&& set
->tags
[i
])
2092 blk_mq_free_map_and_requests(set
, i
);
2098 hctx
->tags
= set
->tags
[i
];
2099 WARN_ON(!hctx
->tags
);
2102 * Set the map size to the number of mapped software queues.
2103 * This is more accurate and more efficient than looping
2104 * over all possibly mapped software queues.
2106 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2109 * Initialize batch roundrobin counts
2111 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2112 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2116 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2118 struct blk_mq_hw_ctx
*hctx
;
2121 queue_for_each_hw_ctx(q
, hctx
, i
) {
2123 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2125 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2129 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
2131 struct request_queue
*q
;
2133 lockdep_assert_held(&set
->tag_list_lock
);
2135 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2136 blk_mq_freeze_queue(q
);
2137 queue_set_hctx_shared(q
, shared
);
2138 blk_mq_unfreeze_queue(q
);
2142 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2144 struct blk_mq_tag_set
*set
= q
->tag_set
;
2146 mutex_lock(&set
->tag_list_lock
);
2147 list_del_rcu(&q
->tag_set_list
);
2148 INIT_LIST_HEAD(&q
->tag_set_list
);
2149 if (list_is_singular(&set
->tag_list
)) {
2150 /* just transitioned to unshared */
2151 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2152 /* update existing queue */
2153 blk_mq_update_tag_set_depth(set
, false);
2155 mutex_unlock(&set
->tag_list_lock
);
2160 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2161 struct request_queue
*q
)
2165 mutex_lock(&set
->tag_list_lock
);
2167 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2168 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2169 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2170 /* update existing queue */
2171 blk_mq_update_tag_set_depth(set
, true);
2173 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2174 queue_set_hctx_shared(q
, true);
2175 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2177 mutex_unlock(&set
->tag_list_lock
);
2181 * It is the actual release handler for mq, but we do it from
2182 * request queue's release handler for avoiding use-after-free
2183 * and headache because q->mq_kobj shouldn't have been introduced,
2184 * but we can't group ctx/kctx kobj without it.
2186 void blk_mq_release(struct request_queue
*q
)
2188 struct blk_mq_hw_ctx
*hctx
;
2191 /* hctx kobj stays in hctx */
2192 queue_for_each_hw_ctx(q
, hctx
, i
) {
2195 kobject_put(&hctx
->kobj
);
2200 kfree(q
->queue_hw_ctx
);
2203 * release .mq_kobj and sw queue's kobject now because
2204 * both share lifetime with request queue.
2206 blk_mq_sysfs_deinit(q
);
2208 free_percpu(q
->queue_ctx
);
2211 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2213 struct request_queue
*uninit_q
, *q
;
2215 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2217 return ERR_PTR(-ENOMEM
);
2219 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2221 blk_cleanup_queue(uninit_q
);
2225 EXPORT_SYMBOL(blk_mq_init_queue
);
2227 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2228 struct request_queue
*q
)
2231 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2233 blk_mq_sysfs_unregister(q
);
2234 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2240 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2241 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2246 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2253 atomic_set(&hctxs
[i
]->nr_active
, 0);
2254 hctxs
[i
]->numa_node
= node
;
2255 hctxs
[i
]->queue_num
= i
;
2257 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2258 free_cpumask_var(hctxs
[i
]->cpumask
);
2263 blk_mq_hctx_kobj_init(hctxs
[i
]);
2265 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2266 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2270 blk_mq_free_map_and_requests(set
, j
);
2271 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2272 kobject_put(&hctx
->kobj
);
2277 q
->nr_hw_queues
= i
;
2278 blk_mq_sysfs_register(q
);
2281 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2282 struct request_queue
*q
)
2284 /* mark the queue as mq asap */
2285 q
->mq_ops
= set
->ops
;
2287 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2288 blk_mq_poll_stats_bkt
,
2289 BLK_MQ_POLL_STATS_BKTS
, q
);
2293 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2297 /* init q->mq_kobj and sw queues' kobjects */
2298 blk_mq_sysfs_init(q
);
2300 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2301 GFP_KERNEL
, set
->numa_node
);
2302 if (!q
->queue_hw_ctx
)
2305 q
->mq_map
= set
->mq_map
;
2307 blk_mq_realloc_hw_ctxs(set
, q
);
2308 if (!q
->nr_hw_queues
)
2311 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2312 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2314 q
->nr_queues
= nr_cpu_ids
;
2316 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2318 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2319 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2321 q
->sg_reserved_size
= INT_MAX
;
2323 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2324 INIT_LIST_HEAD(&q
->requeue_list
);
2325 spin_lock_init(&q
->requeue_lock
);
2327 blk_queue_make_request(q
, blk_mq_make_request
);
2330 * Do this after blk_queue_make_request() overrides it...
2332 q
->nr_requests
= set
->queue_depth
;
2335 * Default to classic polling
2339 if (set
->ops
->complete
)
2340 blk_queue_softirq_done(q
, set
->ops
->complete
);
2342 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2345 mutex_lock(&all_q_mutex
);
2347 list_add_tail(&q
->all_q_node
, &all_q_list
);
2348 blk_mq_add_queue_tag_set(set
, q
);
2349 blk_mq_map_swqueue(q
, cpu_online_mask
);
2351 mutex_unlock(&all_q_mutex
);
2354 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2357 ret
= blk_mq_sched_init(q
);
2359 return ERR_PTR(ret
);
2365 kfree(q
->queue_hw_ctx
);
2367 free_percpu(q
->queue_ctx
);
2370 return ERR_PTR(-ENOMEM
);
2372 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2374 void blk_mq_free_queue(struct request_queue
*q
)
2376 struct blk_mq_tag_set
*set
= q
->tag_set
;
2378 mutex_lock(&all_q_mutex
);
2379 list_del_init(&q
->all_q_node
);
2380 mutex_unlock(&all_q_mutex
);
2382 blk_mq_del_queue_tag_set(q
);
2384 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2387 /* Basically redo blk_mq_init_queue with queue frozen */
2388 static void blk_mq_queue_reinit(struct request_queue
*q
,
2389 const struct cpumask
*online_mask
)
2391 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2393 blk_mq_debugfs_unregister_hctxs(q
);
2394 blk_mq_sysfs_unregister(q
);
2397 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2398 * we should change hctx numa_node according to new topology (this
2399 * involves free and re-allocate memory, worthy doing?)
2402 blk_mq_map_swqueue(q
, online_mask
);
2404 blk_mq_sysfs_register(q
);
2405 blk_mq_debugfs_register_hctxs(q
);
2409 * New online cpumask which is going to be set in this hotplug event.
2410 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2411 * one-by-one and dynamically allocating this could result in a failure.
2413 static struct cpumask cpuhp_online_new
;
2415 static void blk_mq_queue_reinit_work(void)
2417 struct request_queue
*q
;
2419 mutex_lock(&all_q_mutex
);
2421 * We need to freeze and reinit all existing queues. Freezing
2422 * involves synchronous wait for an RCU grace period and doing it
2423 * one by one may take a long time. Start freezing all queues in
2424 * one swoop and then wait for the completions so that freezing can
2425 * take place in parallel.
2427 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2428 blk_freeze_queue_start(q
);
2429 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2430 blk_mq_freeze_queue_wait(q
);
2432 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2433 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2435 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2436 blk_mq_unfreeze_queue(q
);
2438 mutex_unlock(&all_q_mutex
);
2441 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2443 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2444 blk_mq_queue_reinit_work();
2449 * Before hotadded cpu starts handling requests, new mappings must be
2450 * established. Otherwise, these requests in hw queue might never be
2453 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2454 * for CPU0, and ctx1 for CPU1).
2456 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2457 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2459 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2460 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2461 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2464 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2466 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2467 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2468 blk_mq_queue_reinit_work();
2472 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2476 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2477 if (!__blk_mq_alloc_rq_map(set
, i
))
2484 blk_mq_free_rq_map(set
->tags
[i
]);
2490 * Allocate the request maps associated with this tag_set. Note that this
2491 * may reduce the depth asked for, if memory is tight. set->queue_depth
2492 * will be updated to reflect the allocated depth.
2494 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2499 depth
= set
->queue_depth
;
2501 err
= __blk_mq_alloc_rq_maps(set
);
2505 set
->queue_depth
>>= 1;
2506 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2510 } while (set
->queue_depth
);
2512 if (!set
->queue_depth
|| err
) {
2513 pr_err("blk-mq: failed to allocate request map\n");
2517 if (depth
!= set
->queue_depth
)
2518 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2519 depth
, set
->queue_depth
);
2524 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2526 if (set
->ops
->map_queues
)
2527 return set
->ops
->map_queues(set
);
2529 return blk_mq_map_queues(set
);
2533 * Alloc a tag set to be associated with one or more request queues.
2534 * May fail with EINVAL for various error conditions. May adjust the
2535 * requested depth down, if if it too large. In that case, the set
2536 * value will be stored in set->queue_depth.
2538 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2542 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2544 if (!set
->nr_hw_queues
)
2546 if (!set
->queue_depth
)
2548 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2551 if (!set
->ops
->queue_rq
)
2554 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2555 pr_info("blk-mq: reduced tag depth to %u\n",
2557 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2561 * If a crashdump is active, then we are potentially in a very
2562 * memory constrained environment. Limit us to 1 queue and
2563 * 64 tags to prevent using too much memory.
2565 if (is_kdump_kernel()) {
2566 set
->nr_hw_queues
= 1;
2567 set
->queue_depth
= min(64U, set
->queue_depth
);
2570 * There is no use for more h/w queues than cpus.
2572 if (set
->nr_hw_queues
> nr_cpu_ids
)
2573 set
->nr_hw_queues
= nr_cpu_ids
;
2575 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2576 GFP_KERNEL
, set
->numa_node
);
2581 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2582 GFP_KERNEL
, set
->numa_node
);
2586 ret
= blk_mq_update_queue_map(set
);
2588 goto out_free_mq_map
;
2590 ret
= blk_mq_alloc_rq_maps(set
);
2592 goto out_free_mq_map
;
2594 mutex_init(&set
->tag_list_lock
);
2595 INIT_LIST_HEAD(&set
->tag_list
);
2607 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2609 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2613 for (i
= 0; i
< nr_cpu_ids
; i
++)
2614 blk_mq_free_map_and_requests(set
, i
);
2622 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2624 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2626 struct blk_mq_tag_set
*set
= q
->tag_set
;
2627 struct blk_mq_hw_ctx
*hctx
;
2633 blk_mq_freeze_queue(q
);
2636 queue_for_each_hw_ctx(q
, hctx
, i
) {
2640 * If we're using an MQ scheduler, just update the scheduler
2641 * queue depth. This is similar to what the old code would do.
2643 if (!hctx
->sched_tags
) {
2644 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2645 min(nr
, set
->queue_depth
),
2648 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2656 q
->nr_requests
= nr
;
2658 blk_mq_unfreeze_queue(q
);
2663 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2665 struct request_queue
*q
;
2667 lockdep_assert_held(&set
->tag_list_lock
);
2669 if (nr_hw_queues
> nr_cpu_ids
)
2670 nr_hw_queues
= nr_cpu_ids
;
2671 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2674 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2675 blk_mq_freeze_queue(q
);
2677 set
->nr_hw_queues
= nr_hw_queues
;
2678 blk_mq_update_queue_map(set
);
2679 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2680 blk_mq_realloc_hw_ctxs(set
, q
);
2681 blk_mq_queue_reinit(q
, cpu_online_mask
);
2684 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2685 blk_mq_unfreeze_queue(q
);
2687 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2689 /* Enable polling stats and return whether they were already enabled. */
2690 static bool blk_poll_stats_enable(struct request_queue
*q
)
2692 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2693 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
2695 blk_stat_add_callback(q
, q
->poll_cb
);
2699 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2702 * We don't arm the callback if polling stats are not enabled or the
2703 * callback is already active.
2705 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2706 blk_stat_is_active(q
->poll_cb
))
2709 blk_stat_activate_msecs(q
->poll_cb
, 100);
2712 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2714 struct request_queue
*q
= cb
->data
;
2717 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
2718 if (cb
->stat
[bucket
].nr_samples
)
2719 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
2723 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2724 struct blk_mq_hw_ctx
*hctx
,
2727 unsigned long ret
= 0;
2731 * If stats collection isn't on, don't sleep but turn it on for
2734 if (!blk_poll_stats_enable(q
))
2738 * As an optimistic guess, use half of the mean service time
2739 * for this type of request. We can (and should) make this smarter.
2740 * For instance, if the completion latencies are tight, we can
2741 * get closer than just half the mean. This is especially
2742 * important on devices where the completion latencies are longer
2743 * than ~10 usec. We do use the stats for the relevant IO size
2744 * if available which does lead to better estimates.
2746 bucket
= blk_mq_poll_stats_bkt(rq
);
2750 if (q
->poll_stat
[bucket
].nr_samples
)
2751 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
2756 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2757 struct blk_mq_hw_ctx
*hctx
,
2760 struct hrtimer_sleeper hs
;
2761 enum hrtimer_mode mode
;
2765 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2771 * -1: don't ever hybrid sleep
2772 * 0: use half of prev avg
2773 * >0: use this specific value
2775 if (q
->poll_nsec
== -1)
2777 else if (q
->poll_nsec
> 0)
2778 nsecs
= q
->poll_nsec
;
2780 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2785 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2788 * This will be replaced with the stats tracking code, using
2789 * 'avg_completion_time / 2' as the pre-sleep target.
2793 mode
= HRTIMER_MODE_REL
;
2794 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2795 hrtimer_set_expires(&hs
.timer
, kt
);
2797 hrtimer_init_sleeper(&hs
, current
);
2799 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2801 set_current_state(TASK_UNINTERRUPTIBLE
);
2802 hrtimer_start_expires(&hs
.timer
, mode
);
2805 hrtimer_cancel(&hs
.timer
);
2806 mode
= HRTIMER_MODE_ABS
;
2807 } while (hs
.task
&& !signal_pending(current
));
2809 __set_current_state(TASK_RUNNING
);
2810 destroy_hrtimer_on_stack(&hs
.timer
);
2814 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2816 struct request_queue
*q
= hctx
->queue
;
2820 * If we sleep, have the caller restart the poll loop to reset
2821 * the state. Like for the other success return cases, the
2822 * caller is responsible for checking if the IO completed. If
2823 * the IO isn't complete, we'll get called again and will go
2824 * straight to the busy poll loop.
2826 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2829 hctx
->poll_considered
++;
2831 state
= current
->state
;
2832 while (!need_resched()) {
2835 hctx
->poll_invoked
++;
2837 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2839 hctx
->poll_success
++;
2840 set_current_state(TASK_RUNNING
);
2844 if (signal_pending_state(state
, current
))
2845 set_current_state(TASK_RUNNING
);
2847 if (current
->state
== TASK_RUNNING
)
2857 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2859 struct blk_mq_hw_ctx
*hctx
;
2860 struct blk_plug
*plug
;
2863 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2864 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2867 plug
= current
->plug
;
2869 blk_flush_plug_list(plug
, false);
2871 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2872 if (!blk_qc_t_is_internal(cookie
))
2873 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2875 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2877 * With scheduling, if the request has completed, we'll
2878 * get a NULL return here, as we clear the sched tag when
2879 * that happens. The request still remains valid, like always,
2880 * so we should be safe with just the NULL check.
2886 return __blk_mq_poll(hctx
, rq
);
2888 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2890 void blk_mq_disable_hotplug(void)
2892 mutex_lock(&all_q_mutex
);
2895 void blk_mq_enable_hotplug(void)
2897 mutex_unlock(&all_q_mutex
);
2900 static int __init
blk_mq_init(void)
2902 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2903 blk_mq_hctx_notify_dead
);
2905 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2906 blk_mq_queue_reinit_prepare
,
2907 blk_mq_queue_reinit_dead
);
2910 subsys_initcall(blk_mq_init
);