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 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
633 if (tag
< tags
->nr_tags
) {
634 prefetch(tags
->rqs
[tag
]);
635 return tags
->rqs
[tag
];
640 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
642 struct blk_mq_timeout_data
{
644 unsigned int next_set
;
647 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
649 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
650 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
653 * We know that complete is set at this point. If STARTED isn't set
654 * anymore, then the request isn't active and the "timeout" should
655 * just be ignored. This can happen due to the bitflag ordering.
656 * Timeout first checks if STARTED is set, and if it is, assumes
657 * the request is active. But if we race with completion, then
658 * both flags will get cleared. So check here again, and ignore
659 * a timeout event with a request that isn't active.
661 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
665 ret
= ops
->timeout(req
, reserved
);
669 __blk_mq_complete_request(req
);
671 case BLK_EH_RESET_TIMER
:
673 blk_clear_rq_complete(req
);
675 case BLK_EH_NOT_HANDLED
:
678 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
683 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
684 struct request
*rq
, void *priv
, bool reserved
)
686 struct blk_mq_timeout_data
*data
= priv
;
688 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
692 * The rq being checked may have been freed and reallocated
693 * out already here, we avoid this race by checking rq->deadline
694 * and REQ_ATOM_COMPLETE flag together:
696 * - if rq->deadline is observed as new value because of
697 * reusing, the rq won't be timed out because of timing.
698 * - if rq->deadline is observed as previous value,
699 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
700 * because we put a barrier between setting rq->deadline
701 * and clearing the flag in blk_mq_start_request(), so
702 * this rq won't be timed out too.
704 if (time_after_eq(jiffies
, rq
->deadline
)) {
705 if (!blk_mark_rq_complete(rq
))
706 blk_mq_rq_timed_out(rq
, reserved
);
707 } else if (!data
->next_set
|| time_after(data
->next
, rq
->deadline
)) {
708 data
->next
= rq
->deadline
;
713 static void blk_mq_timeout_work(struct work_struct
*work
)
715 struct request_queue
*q
=
716 container_of(work
, struct request_queue
, timeout_work
);
717 struct blk_mq_timeout_data data
= {
723 /* A deadlock might occur if a request is stuck requiring a
724 * timeout at the same time a queue freeze is waiting
725 * completion, since the timeout code would not be able to
726 * acquire the queue reference here.
728 * That's why we don't use blk_queue_enter here; instead, we use
729 * percpu_ref_tryget directly, because we need to be able to
730 * obtain a reference even in the short window between the queue
731 * starting to freeze, by dropping the first reference in
732 * blk_freeze_queue_start, and the moment the last request is
733 * consumed, marked by the instant q_usage_counter reaches
736 if (!percpu_ref_tryget(&q
->q_usage_counter
))
739 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
742 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
743 mod_timer(&q
->timeout
, data
.next
);
745 struct blk_mq_hw_ctx
*hctx
;
747 queue_for_each_hw_ctx(q
, hctx
, i
) {
748 /* the hctx may be unmapped, so check it here */
749 if (blk_mq_hw_queue_mapped(hctx
))
750 blk_mq_tag_idle(hctx
);
757 * Reverse check our software queue for entries that we could potentially
758 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
759 * too much time checking for merges.
761 static bool blk_mq_attempt_merge(struct request_queue
*q
,
762 struct blk_mq_ctx
*ctx
, struct bio
*bio
)
767 list_for_each_entry_reverse(rq
, &ctx
->rq_list
, queuelist
) {
773 if (!blk_rq_merge_ok(rq
, bio
))
776 switch (blk_try_merge(rq
, bio
)) {
777 case ELEVATOR_BACK_MERGE
:
778 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
779 merged
= bio_attempt_back_merge(q
, rq
, bio
);
781 case ELEVATOR_FRONT_MERGE
:
782 if (blk_mq_sched_allow_merge(q
, rq
, bio
))
783 merged
= bio_attempt_front_merge(q
, rq
, bio
);
785 case ELEVATOR_DISCARD_MERGE
:
786 merged
= bio_attempt_discard_merge(q
, rq
, bio
);
800 struct flush_busy_ctx_data
{
801 struct blk_mq_hw_ctx
*hctx
;
802 struct list_head
*list
;
805 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
807 struct flush_busy_ctx_data
*flush_data
= data
;
808 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
809 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
811 sbitmap_clear_bit(sb
, bitnr
);
812 spin_lock(&ctx
->lock
);
813 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
814 spin_unlock(&ctx
->lock
);
819 * Process software queues that have been marked busy, splicing them
820 * to the for-dispatch
822 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
824 struct flush_busy_ctx_data data
= {
829 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
831 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
833 static inline unsigned int queued_to_index(unsigned int queued
)
838 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
841 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
844 struct blk_mq_alloc_data data
= {
846 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
847 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
850 might_sleep_if(wait
);
855 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
856 data
.flags
|= BLK_MQ_REQ_RESERVED
;
858 rq
->tag
= blk_mq_get_tag(&data
);
860 if (blk_mq_tag_busy(data
.hctx
)) {
861 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
862 atomic_inc(&data
.hctx
->nr_active
);
864 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
870 return rq
->tag
!= -1;
873 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx
*hctx
,
876 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
879 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
880 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
881 atomic_dec(&hctx
->nr_active
);
885 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx
*hctx
,
888 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
891 __blk_mq_put_driver_tag(hctx
, rq
);
894 static void blk_mq_put_driver_tag(struct request
*rq
)
896 struct blk_mq_hw_ctx
*hctx
;
898 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
901 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
902 __blk_mq_put_driver_tag(hctx
, rq
);
906 * If we fail getting a driver tag because all the driver tags are already
907 * assigned and on the dispatch list, BUT the first entry does not have a
908 * tag, then we could deadlock. For that case, move entries with assigned
909 * driver tags to the front, leaving the set of tagged requests in the
910 * same order, and the untagged set in the same order.
912 static bool reorder_tags_to_front(struct list_head
*list
)
914 struct request
*rq
, *tmp
, *first
= NULL
;
916 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
920 list_move(&rq
->queuelist
, list
);
926 return first
!= NULL
;
929 static int blk_mq_dispatch_wake(wait_queue_t
*wait
, unsigned mode
, int flags
,
932 struct blk_mq_hw_ctx
*hctx
;
934 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
936 list_del(&wait
->task_list
);
937 clear_bit_unlock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
);
938 blk_mq_run_hw_queue(hctx
, true);
942 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx
*hctx
)
944 struct sbq_wait_state
*ws
;
947 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
948 * The thread which wins the race to grab this bit adds the hardware
949 * queue to the wait queue.
951 if (test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
) ||
952 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
955 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
956 ws
= bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
);
959 * As soon as this returns, it's no longer safe to fiddle with
960 * hctx->dispatch_wait, since a completion can wake up the wait queue
961 * and unlock the bit.
963 add_wait_queue(&ws
->wait
, &hctx
->dispatch_wait
);
967 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
)
969 struct blk_mq_hw_ctx
*hctx
;
971 int errors
, queued
, ret
= BLK_MQ_RQ_QUEUE_OK
;
973 if (list_empty(list
))
977 * Now process all the entries, sending them to the driver.
981 struct blk_mq_queue_data bd
;
983 rq
= list_first_entry(list
, struct request
, queuelist
);
984 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
985 if (!queued
&& reorder_tags_to_front(list
))
989 * The initial allocation attempt failed, so we need to
990 * rerun the hardware queue when a tag is freed.
992 if (!blk_mq_dispatch_wait_add(hctx
))
996 * It's possible that a tag was freed in the window
997 * between the allocation failure and adding the
998 * hardware queue to the wait queue.
1000 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1004 list_del_init(&rq
->queuelist
);
1009 * Flag last if we have no more requests, or if we have more
1010 * but can't assign a driver tag to it.
1012 if (list_empty(list
))
1015 struct request
*nxt
;
1017 nxt
= list_first_entry(list
, struct request
, queuelist
);
1018 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1021 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1023 case BLK_MQ_RQ_QUEUE_OK
:
1026 case BLK_MQ_RQ_QUEUE_BUSY
:
1027 blk_mq_put_driver_tag_hctx(hctx
, rq
);
1028 list_add(&rq
->queuelist
, list
);
1029 __blk_mq_requeue_request(rq
);
1032 pr_err("blk-mq: bad return on queue: %d\n", ret
);
1033 case BLK_MQ_RQ_QUEUE_ERROR
:
1035 blk_mq_end_request(rq
, -EIO
);
1039 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
1041 } while (!list_empty(list
));
1043 hctx
->dispatched
[queued_to_index(queued
)]++;
1046 * Any items that need requeuing? Stuff them into hctx->dispatch,
1047 * that is where we will continue on next queue run.
1049 if (!list_empty(list
)) {
1051 * If an I/O scheduler has been configured and we got a driver
1052 * tag for the next request already, free it again.
1054 rq
= list_first_entry(list
, struct request
, queuelist
);
1055 blk_mq_put_driver_tag(rq
);
1057 spin_lock(&hctx
->lock
);
1058 list_splice_init(list
, &hctx
->dispatch
);
1059 spin_unlock(&hctx
->lock
);
1062 * If SCHED_RESTART was set by the caller of this function and
1063 * it is no longer set that means that it was cleared by another
1064 * thread and hence that a queue rerun is needed.
1066 * If TAG_WAITING is set that means that an I/O scheduler has
1067 * been configured and another thread is waiting for a driver
1068 * tag. To guarantee fairness, do not rerun this hardware queue
1069 * but let the other thread grab the driver tag.
1071 * If no I/O scheduler has been configured it is possible that
1072 * the hardware queue got stopped and restarted before requests
1073 * were pushed back onto the dispatch list. Rerun the queue to
1074 * avoid starvation. Notes:
1075 * - blk_mq_run_hw_queue() checks whether or not a queue has
1076 * been stopped before rerunning a queue.
1077 * - Some but not all block drivers stop a queue before
1078 * returning BLK_MQ_RQ_QUEUE_BUSY. Two exceptions are scsi-mq
1081 if (!blk_mq_sched_needs_restart(hctx
) &&
1082 !test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1083 blk_mq_run_hw_queue(hctx
, true);
1086 return (queued
+ errors
) != 0;
1089 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1093 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1094 cpu_online(hctx
->next_cpu
));
1096 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1098 blk_mq_sched_dispatch_requests(hctx
);
1103 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1104 blk_mq_sched_dispatch_requests(hctx
);
1105 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1110 * It'd be great if the workqueue API had a way to pass
1111 * in a mask and had some smarts for more clever placement.
1112 * For now we just round-robin here, switching for every
1113 * BLK_MQ_CPU_WORK_BATCH queued items.
1115 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1117 if (hctx
->queue
->nr_hw_queues
== 1)
1118 return WORK_CPU_UNBOUND
;
1120 if (--hctx
->next_cpu_batch
<= 0) {
1123 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1124 if (next_cpu
>= nr_cpu_ids
)
1125 next_cpu
= cpumask_first(hctx
->cpumask
);
1127 hctx
->next_cpu
= next_cpu
;
1128 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1131 return hctx
->next_cpu
;
1134 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1135 unsigned long msecs
)
1137 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
1138 !blk_mq_hw_queue_mapped(hctx
)))
1141 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1142 int cpu
= get_cpu();
1143 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1144 __blk_mq_run_hw_queue(hctx
);
1152 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1154 msecs_to_jiffies(msecs
));
1157 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1159 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1161 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1163 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1165 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1167 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1169 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1171 struct blk_mq_hw_ctx
*hctx
;
1174 queue_for_each_hw_ctx(q
, hctx
, i
) {
1175 if (!blk_mq_hctx_has_pending(hctx
) ||
1176 blk_mq_hctx_stopped(hctx
))
1179 blk_mq_run_hw_queue(hctx
, async
);
1182 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1185 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1186 * @q: request queue.
1188 * The caller is responsible for serializing this function against
1189 * blk_mq_{start,stop}_hw_queue().
1191 bool blk_mq_queue_stopped(struct request_queue
*q
)
1193 struct blk_mq_hw_ctx
*hctx
;
1196 queue_for_each_hw_ctx(q
, hctx
, i
)
1197 if (blk_mq_hctx_stopped(hctx
))
1202 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1204 static void __blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool sync
)
1207 cancel_delayed_work_sync(&hctx
->run_work
);
1209 cancel_delayed_work(&hctx
->run_work
);
1211 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1214 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1216 __blk_mq_stop_hw_queue(hctx
, false);
1218 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1220 static void __blk_mq_stop_hw_queues(struct request_queue
*q
, bool sync
)
1222 struct blk_mq_hw_ctx
*hctx
;
1225 queue_for_each_hw_ctx(q
, hctx
, i
)
1226 __blk_mq_stop_hw_queue(hctx
, sync
);
1229 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1231 __blk_mq_stop_hw_queues(q
, false);
1233 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1235 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1237 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1239 blk_mq_run_hw_queue(hctx
, false);
1241 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1243 void blk_mq_start_hw_queues(struct request_queue
*q
)
1245 struct blk_mq_hw_ctx
*hctx
;
1248 queue_for_each_hw_ctx(q
, hctx
, i
)
1249 blk_mq_start_hw_queue(hctx
);
1251 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1253 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1255 if (!blk_mq_hctx_stopped(hctx
))
1258 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1259 blk_mq_run_hw_queue(hctx
, async
);
1261 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1263 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1265 struct blk_mq_hw_ctx
*hctx
;
1268 queue_for_each_hw_ctx(q
, hctx
, i
)
1269 blk_mq_start_stopped_hw_queue(hctx
, async
);
1271 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1273 static void blk_mq_run_work_fn(struct work_struct
*work
)
1275 struct blk_mq_hw_ctx
*hctx
;
1277 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1280 * If we are stopped, don't run the queue. The exception is if
1281 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1282 * the STOPPED bit and run it.
1284 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)) {
1285 if (!test_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
))
1288 clear_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1289 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1292 __blk_mq_run_hw_queue(hctx
);
1296 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1298 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1302 * Stop the hw queue, then modify currently delayed work.
1303 * This should prevent us from running the queue prematurely.
1304 * Mark the queue as auto-clearing STOPPED when it runs.
1306 blk_mq_stop_hw_queue(hctx
);
1307 set_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1308 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1310 msecs_to_jiffies(msecs
));
1312 EXPORT_SYMBOL(blk_mq_delay_queue
);
1314 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1318 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1320 trace_block_rq_insert(hctx
->queue
, rq
);
1323 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1325 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1328 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1331 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1333 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1334 blk_mq_hctx_mark_pending(hctx
, ctx
);
1337 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1338 struct list_head
*list
)
1342 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1345 spin_lock(&ctx
->lock
);
1346 while (!list_empty(list
)) {
1349 rq
= list_first_entry(list
, struct request
, queuelist
);
1350 BUG_ON(rq
->mq_ctx
!= ctx
);
1351 list_del_init(&rq
->queuelist
);
1352 __blk_mq_insert_req_list(hctx
, rq
, false);
1354 blk_mq_hctx_mark_pending(hctx
, ctx
);
1355 spin_unlock(&ctx
->lock
);
1358 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1360 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1361 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1363 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1364 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1365 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1368 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1370 struct blk_mq_ctx
*this_ctx
;
1371 struct request_queue
*this_q
;
1374 LIST_HEAD(ctx_list
);
1377 list_splice_init(&plug
->mq_list
, &list
);
1379 list_sort(NULL
, &list
, plug_ctx_cmp
);
1385 while (!list_empty(&list
)) {
1386 rq
= list_entry_rq(list
.next
);
1387 list_del_init(&rq
->queuelist
);
1389 if (rq
->mq_ctx
!= this_ctx
) {
1391 trace_block_unplug(this_q
, depth
, from_schedule
);
1392 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1397 this_ctx
= rq
->mq_ctx
;
1403 list_add_tail(&rq
->queuelist
, &ctx_list
);
1407 * If 'this_ctx' is set, we know we have entries to complete
1408 * on 'ctx_list'. Do those.
1411 trace_block_unplug(this_q
, depth
, from_schedule
);
1412 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1417 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1419 blk_init_request_from_bio(rq
, bio
);
1421 blk_account_io_start(rq
, true);
1424 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1426 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1427 !blk_queue_nomerges(hctx
->queue
);
1430 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1431 struct blk_mq_ctx
*ctx
,
1432 struct request
*rq
, struct bio
*bio
)
1434 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1435 blk_mq_bio_to_request(rq
, bio
);
1436 spin_lock(&ctx
->lock
);
1438 __blk_mq_insert_request(hctx
, rq
, false);
1439 spin_unlock(&ctx
->lock
);
1442 struct request_queue
*q
= hctx
->queue
;
1444 spin_lock(&ctx
->lock
);
1445 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1446 blk_mq_bio_to_request(rq
, bio
);
1450 spin_unlock(&ctx
->lock
);
1451 __blk_mq_finish_request(hctx
, ctx
, rq
);
1456 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1459 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1461 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1464 static void __blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
,
1467 struct request_queue
*q
= rq
->q
;
1468 struct blk_mq_queue_data bd
= {
1472 struct blk_mq_hw_ctx
*hctx
;
1473 blk_qc_t new_cookie
;
1479 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1482 new_cookie
= request_to_qc_t(hctx
, rq
);
1485 * For OK queue, we are done. For error, kill it. Any other
1486 * error (busy), just add it to our list as we previously
1489 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1490 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1491 *cookie
= new_cookie
;
1495 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1496 *cookie
= BLK_QC_T_NONE
;
1497 blk_mq_end_request(rq
, -EIO
);
1501 __blk_mq_requeue_request(rq
);
1503 blk_mq_sched_insert_request(rq
, false, true, false, may_sleep
);
1506 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1507 struct request
*rq
, blk_qc_t
*cookie
)
1509 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1511 __blk_mq_try_issue_directly(rq
, cookie
, false);
1514 unsigned int srcu_idx
;
1518 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1519 __blk_mq_try_issue_directly(rq
, cookie
, true);
1520 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1524 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1526 const int is_sync
= op_is_sync(bio
->bi_opf
);
1527 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1528 struct blk_mq_alloc_data data
= { .flags
= 0 };
1530 unsigned int request_count
= 0;
1531 struct blk_plug
*plug
;
1532 struct request
*same_queue_rq
= NULL
;
1534 unsigned int wb_acct
;
1536 blk_queue_bounce(q
, &bio
);
1538 blk_queue_split(q
, &bio
, q
->bio_split
);
1540 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1542 return BLK_QC_T_NONE
;
1545 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1546 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1547 return BLK_QC_T_NONE
;
1549 if (blk_mq_sched_bio_merge(q
, bio
))
1550 return BLK_QC_T_NONE
;
1552 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1554 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1556 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1557 if (unlikely(!rq
)) {
1558 __wbt_done(q
->rq_wb
, wb_acct
);
1559 return BLK_QC_T_NONE
;
1562 wbt_track(&rq
->issue_stat
, wb_acct
);
1564 cookie
= request_to_qc_t(data
.hctx
, rq
);
1566 plug
= current
->plug
;
1567 if (unlikely(is_flush_fua
)) {
1568 blk_mq_put_ctx(data
.ctx
);
1569 blk_mq_bio_to_request(rq
, bio
);
1571 blk_mq_sched_insert_request(rq
, false, true, true,
1574 blk_insert_flush(rq
);
1575 blk_mq_run_hw_queue(data
.hctx
, true);
1577 } else if (plug
&& q
->nr_hw_queues
== 1) {
1578 struct request
*last
= NULL
;
1580 blk_mq_put_ctx(data
.ctx
);
1581 blk_mq_bio_to_request(rq
, bio
);
1584 * @request_count may become stale because of schedule
1585 * out, so check the list again.
1587 if (list_empty(&plug
->mq_list
))
1589 else if (blk_queue_nomerges(q
))
1590 request_count
= blk_plug_queued_count(q
);
1593 trace_block_plug(q
);
1595 last
= list_entry_rq(plug
->mq_list
.prev
);
1597 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1598 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1599 blk_flush_plug_list(plug
, false);
1600 trace_block_plug(q
);
1603 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1604 } else if (plug
&& !blk_queue_nomerges(q
)) {
1605 blk_mq_bio_to_request(rq
, bio
);
1608 * We do limited plugging. If the bio can be merged, do that.
1609 * Otherwise the existing request in the plug list will be
1610 * issued. So the plug list will have one request at most
1611 * The plug list might get flushed before this. If that happens,
1612 * the plug list is empty, and same_queue_rq is invalid.
1614 if (list_empty(&plug
->mq_list
))
1615 same_queue_rq
= NULL
;
1617 list_del_init(&same_queue_rq
->queuelist
);
1618 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1620 blk_mq_put_ctx(data
.ctx
);
1623 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1625 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1626 blk_mq_put_ctx(data
.ctx
);
1627 blk_mq_bio_to_request(rq
, bio
);
1628 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1629 } else if (q
->elevator
) {
1630 blk_mq_put_ctx(data
.ctx
);
1631 blk_mq_bio_to_request(rq
, bio
);
1632 blk_mq_sched_insert_request(rq
, false, true, true, true);
1633 } else if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
)) {
1634 blk_mq_put_ctx(data
.ctx
);
1635 blk_mq_run_hw_queue(data
.hctx
, true);
1637 blk_mq_put_ctx(data
.ctx
);
1642 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1643 unsigned int hctx_idx
)
1647 if (tags
->rqs
&& set
->ops
->exit_request
) {
1650 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1651 struct request
*rq
= tags
->static_rqs
[i
];
1655 set
->ops
->exit_request(set
, rq
, hctx_idx
);
1656 tags
->static_rqs
[i
] = NULL
;
1660 while (!list_empty(&tags
->page_list
)) {
1661 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1662 list_del_init(&page
->lru
);
1664 * Remove kmemleak object previously allocated in
1665 * blk_mq_init_rq_map().
1667 kmemleak_free(page_address(page
));
1668 __free_pages(page
, page
->private);
1672 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1676 kfree(tags
->static_rqs
);
1677 tags
->static_rqs
= NULL
;
1679 blk_mq_free_tags(tags
);
1682 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1683 unsigned int hctx_idx
,
1684 unsigned int nr_tags
,
1685 unsigned int reserved_tags
)
1687 struct blk_mq_tags
*tags
;
1690 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1691 if (node
== NUMA_NO_NODE
)
1692 node
= set
->numa_node
;
1694 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1695 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1699 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1700 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1703 blk_mq_free_tags(tags
);
1707 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1708 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1710 if (!tags
->static_rqs
) {
1712 blk_mq_free_tags(tags
);
1719 static size_t order_to_size(unsigned int order
)
1721 return (size_t)PAGE_SIZE
<< order
;
1724 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1725 unsigned int hctx_idx
, unsigned int depth
)
1727 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1728 size_t rq_size
, left
;
1731 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1732 if (node
== NUMA_NO_NODE
)
1733 node
= set
->numa_node
;
1735 INIT_LIST_HEAD(&tags
->page_list
);
1738 * rq_size is the size of the request plus driver payload, rounded
1739 * to the cacheline size
1741 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1743 left
= rq_size
* depth
;
1745 for (i
= 0; i
< depth
; ) {
1746 int this_order
= max_order
;
1751 while (this_order
&& left
< order_to_size(this_order
- 1))
1755 page
= alloc_pages_node(node
,
1756 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1762 if (order_to_size(this_order
) < rq_size
)
1769 page
->private = this_order
;
1770 list_add_tail(&page
->lru
, &tags
->page_list
);
1772 p
= page_address(page
);
1774 * Allow kmemleak to scan these pages as they contain pointers
1775 * to additional allocations like via ops->init_request().
1777 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1778 entries_per_page
= order_to_size(this_order
) / rq_size
;
1779 to_do
= min(entries_per_page
, depth
- i
);
1780 left
-= to_do
* rq_size
;
1781 for (j
= 0; j
< to_do
; j
++) {
1782 struct request
*rq
= p
;
1784 tags
->static_rqs
[i
] = rq
;
1785 if (set
->ops
->init_request
) {
1786 if (set
->ops
->init_request(set
, rq
, hctx_idx
,
1788 tags
->static_rqs
[i
] = NULL
;
1800 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1805 * 'cpu' is going away. splice any existing rq_list entries from this
1806 * software queue to the hw queue dispatch list, and ensure that it
1809 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1811 struct blk_mq_hw_ctx
*hctx
;
1812 struct blk_mq_ctx
*ctx
;
1815 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1816 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1818 spin_lock(&ctx
->lock
);
1819 if (!list_empty(&ctx
->rq_list
)) {
1820 list_splice_init(&ctx
->rq_list
, &tmp
);
1821 blk_mq_hctx_clear_pending(hctx
, ctx
);
1823 spin_unlock(&ctx
->lock
);
1825 if (list_empty(&tmp
))
1828 spin_lock(&hctx
->lock
);
1829 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1830 spin_unlock(&hctx
->lock
);
1832 blk_mq_run_hw_queue(hctx
, true);
1836 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1838 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1842 /* hctx->ctxs will be freed in queue's release handler */
1843 static void blk_mq_exit_hctx(struct request_queue
*q
,
1844 struct blk_mq_tag_set
*set
,
1845 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1847 blk_mq_debugfs_unregister_hctx(hctx
);
1849 blk_mq_tag_idle(hctx
);
1851 if (set
->ops
->exit_request
)
1852 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
1854 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1856 if (set
->ops
->exit_hctx
)
1857 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1859 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1860 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1862 blk_mq_remove_cpuhp(hctx
);
1863 blk_free_flush_queue(hctx
->fq
);
1864 sbitmap_free(&hctx
->ctx_map
);
1867 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1868 struct blk_mq_tag_set
*set
, int nr_queue
)
1870 struct blk_mq_hw_ctx
*hctx
;
1873 queue_for_each_hw_ctx(q
, hctx
, i
) {
1876 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1880 static int blk_mq_init_hctx(struct request_queue
*q
,
1881 struct blk_mq_tag_set
*set
,
1882 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1886 node
= hctx
->numa_node
;
1887 if (node
== NUMA_NO_NODE
)
1888 node
= hctx
->numa_node
= set
->numa_node
;
1890 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1891 spin_lock_init(&hctx
->lock
);
1892 INIT_LIST_HEAD(&hctx
->dispatch
);
1894 hctx
->queue_num
= hctx_idx
;
1895 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1897 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1899 hctx
->tags
= set
->tags
[hctx_idx
];
1902 * Allocate space for all possible cpus to avoid allocation at
1905 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1908 goto unregister_cpu_notifier
;
1910 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1916 if (set
->ops
->init_hctx
&&
1917 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1920 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
1923 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1925 goto sched_exit_hctx
;
1927 if (set
->ops
->init_request
&&
1928 set
->ops
->init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
1932 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1933 init_srcu_struct(&hctx
->queue_rq_srcu
);
1935 blk_mq_debugfs_register_hctx(q
, hctx
);
1942 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1944 if (set
->ops
->exit_hctx
)
1945 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1947 sbitmap_free(&hctx
->ctx_map
);
1950 unregister_cpu_notifier
:
1951 blk_mq_remove_cpuhp(hctx
);
1955 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1956 unsigned int nr_hw_queues
)
1960 for_each_possible_cpu(i
) {
1961 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1962 struct blk_mq_hw_ctx
*hctx
;
1965 spin_lock_init(&__ctx
->lock
);
1966 INIT_LIST_HEAD(&__ctx
->rq_list
);
1969 /* If the cpu isn't online, the cpu is mapped to first hctx */
1973 hctx
= blk_mq_map_queue(q
, i
);
1976 * Set local node, IFF we have more than one hw queue. If
1977 * not, we remain on the home node of the device
1979 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1980 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1984 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
1988 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
1989 set
->queue_depth
, set
->reserved_tags
);
1990 if (!set
->tags
[hctx_idx
])
1993 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
1998 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
1999 set
->tags
[hctx_idx
] = NULL
;
2003 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2004 unsigned int hctx_idx
)
2006 if (set
->tags
[hctx_idx
]) {
2007 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2008 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2009 set
->tags
[hctx_idx
] = NULL
;
2013 static void blk_mq_map_swqueue(struct request_queue
*q
,
2014 const struct cpumask
*online_mask
)
2016 unsigned int i
, hctx_idx
;
2017 struct blk_mq_hw_ctx
*hctx
;
2018 struct blk_mq_ctx
*ctx
;
2019 struct blk_mq_tag_set
*set
= q
->tag_set
;
2022 * Avoid others reading imcomplete hctx->cpumask through sysfs
2024 mutex_lock(&q
->sysfs_lock
);
2026 queue_for_each_hw_ctx(q
, hctx
, i
) {
2027 cpumask_clear(hctx
->cpumask
);
2032 * Map software to hardware queues
2034 for_each_possible_cpu(i
) {
2035 /* If the cpu isn't online, the cpu is mapped to first hctx */
2036 if (!cpumask_test_cpu(i
, online_mask
))
2039 hctx_idx
= q
->mq_map
[i
];
2040 /* unmapped hw queue can be remapped after CPU topo changed */
2041 if (!set
->tags
[hctx_idx
] &&
2042 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2044 * If tags initialization fail for some hctx,
2045 * that hctx won't be brought online. In this
2046 * case, remap the current ctx to hctx[0] which
2047 * is guaranteed to always have tags allocated
2052 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2053 hctx
= blk_mq_map_queue(q
, i
);
2055 cpumask_set_cpu(i
, hctx
->cpumask
);
2056 ctx
->index_hw
= hctx
->nr_ctx
;
2057 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2060 mutex_unlock(&q
->sysfs_lock
);
2062 queue_for_each_hw_ctx(q
, hctx
, i
) {
2064 * If no software queues are mapped to this hardware queue,
2065 * disable it and free the request entries.
2067 if (!hctx
->nr_ctx
) {
2068 /* Never unmap queue 0. We need it as a
2069 * fallback in case of a new remap fails
2072 if (i
&& set
->tags
[i
])
2073 blk_mq_free_map_and_requests(set
, i
);
2079 hctx
->tags
= set
->tags
[i
];
2080 WARN_ON(!hctx
->tags
);
2083 * Set the map size to the number of mapped software queues.
2084 * This is more accurate and more efficient than looping
2085 * over all possibly mapped software queues.
2087 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2090 * Initialize batch roundrobin counts
2092 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2093 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2097 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2099 struct blk_mq_hw_ctx
*hctx
;
2102 queue_for_each_hw_ctx(q
, hctx
, i
) {
2104 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2106 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2110 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
2112 struct request_queue
*q
;
2114 lockdep_assert_held(&set
->tag_list_lock
);
2116 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2117 blk_mq_freeze_queue(q
);
2118 queue_set_hctx_shared(q
, shared
);
2119 blk_mq_unfreeze_queue(q
);
2123 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2125 struct blk_mq_tag_set
*set
= q
->tag_set
;
2127 mutex_lock(&set
->tag_list_lock
);
2128 list_del_rcu(&q
->tag_set_list
);
2129 INIT_LIST_HEAD(&q
->tag_set_list
);
2130 if (list_is_singular(&set
->tag_list
)) {
2131 /* just transitioned to unshared */
2132 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2133 /* update existing queue */
2134 blk_mq_update_tag_set_depth(set
, false);
2136 mutex_unlock(&set
->tag_list_lock
);
2141 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2142 struct request_queue
*q
)
2146 mutex_lock(&set
->tag_list_lock
);
2148 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2149 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2150 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2151 /* update existing queue */
2152 blk_mq_update_tag_set_depth(set
, true);
2154 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2155 queue_set_hctx_shared(q
, true);
2156 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2158 mutex_unlock(&set
->tag_list_lock
);
2162 * It is the actual release handler for mq, but we do it from
2163 * request queue's release handler for avoiding use-after-free
2164 * and headache because q->mq_kobj shouldn't have been introduced,
2165 * but we can't group ctx/kctx kobj without it.
2167 void blk_mq_release(struct request_queue
*q
)
2169 struct blk_mq_hw_ctx
*hctx
;
2172 /* hctx kobj stays in hctx */
2173 queue_for_each_hw_ctx(q
, hctx
, i
) {
2176 kobject_put(&hctx
->kobj
);
2181 kfree(q
->queue_hw_ctx
);
2184 * release .mq_kobj and sw queue's kobject now because
2185 * both share lifetime with request queue.
2187 blk_mq_sysfs_deinit(q
);
2189 free_percpu(q
->queue_ctx
);
2192 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2194 struct request_queue
*uninit_q
, *q
;
2196 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2198 return ERR_PTR(-ENOMEM
);
2200 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2202 blk_cleanup_queue(uninit_q
);
2206 EXPORT_SYMBOL(blk_mq_init_queue
);
2208 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2209 struct request_queue
*q
)
2212 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2214 blk_mq_sysfs_unregister(q
);
2215 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2221 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2222 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2227 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2234 atomic_set(&hctxs
[i
]->nr_active
, 0);
2235 hctxs
[i
]->numa_node
= node
;
2236 hctxs
[i
]->queue_num
= i
;
2238 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2239 free_cpumask_var(hctxs
[i
]->cpumask
);
2244 blk_mq_hctx_kobj_init(hctxs
[i
]);
2246 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2247 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2251 blk_mq_free_map_and_requests(set
, j
);
2252 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2253 kobject_put(&hctx
->kobj
);
2258 q
->nr_hw_queues
= i
;
2259 blk_mq_sysfs_register(q
);
2262 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2263 struct request_queue
*q
)
2265 /* mark the queue as mq asap */
2266 q
->mq_ops
= set
->ops
;
2268 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2269 blk_mq_poll_stats_bkt
,
2270 BLK_MQ_POLL_STATS_BKTS
, q
);
2274 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2278 /* init q->mq_kobj and sw queues' kobjects */
2279 blk_mq_sysfs_init(q
);
2281 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2282 GFP_KERNEL
, set
->numa_node
);
2283 if (!q
->queue_hw_ctx
)
2286 q
->mq_map
= set
->mq_map
;
2288 blk_mq_realloc_hw_ctxs(set
, q
);
2289 if (!q
->nr_hw_queues
)
2292 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2293 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2295 q
->nr_queues
= nr_cpu_ids
;
2297 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2299 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2300 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2302 q
->sg_reserved_size
= INT_MAX
;
2304 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2305 INIT_LIST_HEAD(&q
->requeue_list
);
2306 spin_lock_init(&q
->requeue_lock
);
2308 blk_queue_make_request(q
, blk_mq_make_request
);
2311 * Do this after blk_queue_make_request() overrides it...
2313 q
->nr_requests
= set
->queue_depth
;
2316 * Default to classic polling
2320 if (set
->ops
->complete
)
2321 blk_queue_softirq_done(q
, set
->ops
->complete
);
2323 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2326 mutex_lock(&all_q_mutex
);
2328 list_add_tail(&q
->all_q_node
, &all_q_list
);
2329 blk_mq_add_queue_tag_set(set
, q
);
2330 blk_mq_map_swqueue(q
, cpu_online_mask
);
2332 mutex_unlock(&all_q_mutex
);
2335 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2338 ret
= blk_mq_sched_init(q
);
2340 return ERR_PTR(ret
);
2346 kfree(q
->queue_hw_ctx
);
2348 free_percpu(q
->queue_ctx
);
2351 return ERR_PTR(-ENOMEM
);
2353 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2355 void blk_mq_free_queue(struct request_queue
*q
)
2357 struct blk_mq_tag_set
*set
= q
->tag_set
;
2359 mutex_lock(&all_q_mutex
);
2360 list_del_init(&q
->all_q_node
);
2361 mutex_unlock(&all_q_mutex
);
2363 blk_mq_del_queue_tag_set(q
);
2365 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2368 /* Basically redo blk_mq_init_queue with queue frozen */
2369 static void blk_mq_queue_reinit(struct request_queue
*q
,
2370 const struct cpumask
*online_mask
)
2372 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2374 blk_mq_debugfs_unregister_hctxs(q
);
2375 blk_mq_sysfs_unregister(q
);
2378 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2379 * we should change hctx numa_node according to new topology (this
2380 * involves free and re-allocate memory, worthy doing?)
2383 blk_mq_map_swqueue(q
, online_mask
);
2385 blk_mq_sysfs_register(q
);
2386 blk_mq_debugfs_register_hctxs(q
);
2390 * New online cpumask which is going to be set in this hotplug event.
2391 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2392 * one-by-one and dynamically allocating this could result in a failure.
2394 static struct cpumask cpuhp_online_new
;
2396 static void blk_mq_queue_reinit_work(void)
2398 struct request_queue
*q
;
2400 mutex_lock(&all_q_mutex
);
2402 * We need to freeze and reinit all existing queues. Freezing
2403 * involves synchronous wait for an RCU grace period and doing it
2404 * one by one may take a long time. Start freezing all queues in
2405 * one swoop and then wait for the completions so that freezing can
2406 * take place in parallel.
2408 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2409 blk_freeze_queue_start(q
);
2410 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2411 blk_mq_freeze_queue_wait(q
);
2413 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2414 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2416 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2417 blk_mq_unfreeze_queue(q
);
2419 mutex_unlock(&all_q_mutex
);
2422 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2424 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2425 blk_mq_queue_reinit_work();
2430 * Before hotadded cpu starts handling requests, new mappings must be
2431 * established. Otherwise, these requests in hw queue might never be
2434 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2435 * for CPU0, and ctx1 for CPU1).
2437 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2438 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2440 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2441 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2442 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2445 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2447 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2448 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2449 blk_mq_queue_reinit_work();
2453 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2457 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2458 if (!__blk_mq_alloc_rq_map(set
, i
))
2465 blk_mq_free_rq_map(set
->tags
[i
]);
2471 * Allocate the request maps associated with this tag_set. Note that this
2472 * may reduce the depth asked for, if memory is tight. set->queue_depth
2473 * will be updated to reflect the allocated depth.
2475 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2480 depth
= set
->queue_depth
;
2482 err
= __blk_mq_alloc_rq_maps(set
);
2486 set
->queue_depth
>>= 1;
2487 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2491 } while (set
->queue_depth
);
2493 if (!set
->queue_depth
|| err
) {
2494 pr_err("blk-mq: failed to allocate request map\n");
2498 if (depth
!= set
->queue_depth
)
2499 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2500 depth
, set
->queue_depth
);
2505 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2507 if (set
->ops
->map_queues
)
2508 return set
->ops
->map_queues(set
);
2510 return blk_mq_map_queues(set
);
2514 * Alloc a tag set to be associated with one or more request queues.
2515 * May fail with EINVAL for various error conditions. May adjust the
2516 * requested depth down, if if it too large. In that case, the set
2517 * value will be stored in set->queue_depth.
2519 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2523 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2525 if (!set
->nr_hw_queues
)
2527 if (!set
->queue_depth
)
2529 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2532 if (!set
->ops
->queue_rq
)
2535 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2536 pr_info("blk-mq: reduced tag depth to %u\n",
2538 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2542 * If a crashdump is active, then we are potentially in a very
2543 * memory constrained environment. Limit us to 1 queue and
2544 * 64 tags to prevent using too much memory.
2546 if (is_kdump_kernel()) {
2547 set
->nr_hw_queues
= 1;
2548 set
->queue_depth
= min(64U, set
->queue_depth
);
2551 * There is no use for more h/w queues than cpus.
2553 if (set
->nr_hw_queues
> nr_cpu_ids
)
2554 set
->nr_hw_queues
= nr_cpu_ids
;
2556 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2557 GFP_KERNEL
, set
->numa_node
);
2562 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2563 GFP_KERNEL
, set
->numa_node
);
2567 ret
= blk_mq_update_queue_map(set
);
2569 goto out_free_mq_map
;
2571 ret
= blk_mq_alloc_rq_maps(set
);
2573 goto out_free_mq_map
;
2575 mutex_init(&set
->tag_list_lock
);
2576 INIT_LIST_HEAD(&set
->tag_list
);
2588 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2590 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2594 for (i
= 0; i
< nr_cpu_ids
; i
++)
2595 blk_mq_free_map_and_requests(set
, i
);
2603 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2605 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2607 struct blk_mq_tag_set
*set
= q
->tag_set
;
2608 struct blk_mq_hw_ctx
*hctx
;
2614 blk_mq_freeze_queue(q
);
2617 queue_for_each_hw_ctx(q
, hctx
, i
) {
2621 * If we're using an MQ scheduler, just update the scheduler
2622 * queue depth. This is similar to what the old code would do.
2624 if (!hctx
->sched_tags
) {
2625 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2626 min(nr
, set
->queue_depth
),
2629 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2637 q
->nr_requests
= nr
;
2639 blk_mq_unfreeze_queue(q
);
2644 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2646 struct request_queue
*q
;
2648 lockdep_assert_held(&set
->tag_list_lock
);
2650 if (nr_hw_queues
> nr_cpu_ids
)
2651 nr_hw_queues
= nr_cpu_ids
;
2652 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2655 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2656 blk_mq_freeze_queue(q
);
2658 set
->nr_hw_queues
= nr_hw_queues
;
2659 blk_mq_update_queue_map(set
);
2660 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2661 blk_mq_realloc_hw_ctxs(set
, q
);
2662 blk_mq_queue_reinit(q
, cpu_online_mask
);
2665 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2666 blk_mq_unfreeze_queue(q
);
2668 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2670 /* Enable polling stats and return whether they were already enabled. */
2671 static bool blk_poll_stats_enable(struct request_queue
*q
)
2673 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2674 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
2676 blk_stat_add_callback(q
, q
->poll_cb
);
2680 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2683 * We don't arm the callback if polling stats are not enabled or the
2684 * callback is already active.
2686 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2687 blk_stat_is_active(q
->poll_cb
))
2690 blk_stat_activate_msecs(q
->poll_cb
, 100);
2693 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2695 struct request_queue
*q
= cb
->data
;
2698 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
2699 if (cb
->stat
[bucket
].nr_samples
)
2700 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
2704 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2705 struct blk_mq_hw_ctx
*hctx
,
2708 unsigned long ret
= 0;
2712 * If stats collection isn't on, don't sleep but turn it on for
2715 if (!blk_poll_stats_enable(q
))
2719 * As an optimistic guess, use half of the mean service time
2720 * for this type of request. We can (and should) make this smarter.
2721 * For instance, if the completion latencies are tight, we can
2722 * get closer than just half the mean. This is especially
2723 * important on devices where the completion latencies are longer
2724 * than ~10 usec. We do use the stats for the relevant IO size
2725 * if available which does lead to better estimates.
2727 bucket
= blk_mq_poll_stats_bkt(rq
);
2731 if (q
->poll_stat
[bucket
].nr_samples
)
2732 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
2737 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2738 struct blk_mq_hw_ctx
*hctx
,
2741 struct hrtimer_sleeper hs
;
2742 enum hrtimer_mode mode
;
2746 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2752 * -1: don't ever hybrid sleep
2753 * 0: use half of prev avg
2754 * >0: use this specific value
2756 if (q
->poll_nsec
== -1)
2758 else if (q
->poll_nsec
> 0)
2759 nsecs
= q
->poll_nsec
;
2761 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2766 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2769 * This will be replaced with the stats tracking code, using
2770 * 'avg_completion_time / 2' as the pre-sleep target.
2774 mode
= HRTIMER_MODE_REL
;
2775 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2776 hrtimer_set_expires(&hs
.timer
, kt
);
2778 hrtimer_init_sleeper(&hs
, current
);
2780 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2782 set_current_state(TASK_UNINTERRUPTIBLE
);
2783 hrtimer_start_expires(&hs
.timer
, mode
);
2786 hrtimer_cancel(&hs
.timer
);
2787 mode
= HRTIMER_MODE_ABS
;
2788 } while (hs
.task
&& !signal_pending(current
));
2790 __set_current_state(TASK_RUNNING
);
2791 destroy_hrtimer_on_stack(&hs
.timer
);
2795 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2797 struct request_queue
*q
= hctx
->queue
;
2801 * If we sleep, have the caller restart the poll loop to reset
2802 * the state. Like for the other success return cases, the
2803 * caller is responsible for checking if the IO completed. If
2804 * the IO isn't complete, we'll get called again and will go
2805 * straight to the busy poll loop.
2807 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2810 hctx
->poll_considered
++;
2812 state
= current
->state
;
2813 while (!need_resched()) {
2816 hctx
->poll_invoked
++;
2818 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2820 hctx
->poll_success
++;
2821 set_current_state(TASK_RUNNING
);
2825 if (signal_pending_state(state
, current
))
2826 set_current_state(TASK_RUNNING
);
2828 if (current
->state
== TASK_RUNNING
)
2838 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2840 struct blk_mq_hw_ctx
*hctx
;
2841 struct blk_plug
*plug
;
2844 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2845 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2848 plug
= current
->plug
;
2850 blk_flush_plug_list(plug
, false);
2852 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2853 if (!blk_qc_t_is_internal(cookie
))
2854 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2856 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2858 * With scheduling, if the request has completed, we'll
2859 * get a NULL return here, as we clear the sched tag when
2860 * that happens. The request still remains valid, like always,
2861 * so we should be safe with just the NULL check.
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
);