2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
29 #include <trace/events/block.h>
31 #include <linux/blk-mq.h>
34 #include "blk-mq-tag.h"
37 #include "blk-mq-sched.h"
39 static DEFINE_MUTEX(all_q_mutex
);
40 static LIST_HEAD(all_q_list
);
42 static void blk_mq_poll_stats_start(struct request_queue
*q
);
43 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
45 /* Must be consisitent with function below */
46 #define BLK_MQ_POLL_STATS_BKTS 16
47 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
49 int ddir
, bytes
, bucket
;
51 ddir
= blk_stat_rq_ddir(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
);
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
,
872 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
873 data
.flags
|= BLK_MQ_REQ_RESERVED
;
875 rq
->tag
= blk_mq_get_tag(&data
);
877 if (blk_mq_tag_busy(data
.hctx
)) {
878 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
879 atomic_inc(&data
.hctx
->nr_active
);
881 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
887 return rq
->tag
!= -1;
890 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx
*hctx
,
893 blk_mq_put_tag(hctx
, hctx
->tags
, rq
->mq_ctx
, rq
->tag
);
896 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
) {
897 rq
->rq_flags
&= ~RQF_MQ_INFLIGHT
;
898 atomic_dec(&hctx
->nr_active
);
902 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx
*hctx
,
905 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
908 __blk_mq_put_driver_tag(hctx
, rq
);
911 static void blk_mq_put_driver_tag(struct request
*rq
)
913 struct blk_mq_hw_ctx
*hctx
;
915 if (rq
->tag
== -1 || rq
->internal_tag
== -1)
918 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
919 __blk_mq_put_driver_tag(hctx
, rq
);
923 * If we fail getting a driver tag because all the driver tags are already
924 * assigned and on the dispatch list, BUT the first entry does not have a
925 * tag, then we could deadlock. For that case, move entries with assigned
926 * driver tags to the front, leaving the set of tagged requests in the
927 * same order, and the untagged set in the same order.
929 static bool reorder_tags_to_front(struct list_head
*list
)
931 struct request
*rq
, *tmp
, *first
= NULL
;
933 list_for_each_entry_safe_reverse(rq
, tmp
, list
, queuelist
) {
937 list_move(&rq
->queuelist
, list
);
943 return first
!= NULL
;
946 static int blk_mq_dispatch_wake(wait_queue_t
*wait
, unsigned mode
, int flags
,
949 struct blk_mq_hw_ctx
*hctx
;
951 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
953 list_del(&wait
->task_list
);
954 clear_bit_unlock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
);
955 blk_mq_run_hw_queue(hctx
, true);
959 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx
*hctx
)
961 struct sbq_wait_state
*ws
;
964 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
965 * The thread which wins the race to grab this bit adds the hardware
966 * queue to the wait queue.
968 if (test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
) ||
969 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
972 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
973 ws
= bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
);
976 * As soon as this returns, it's no longer safe to fiddle with
977 * hctx->dispatch_wait, since a completion can wake up the wait queue
978 * and unlock the bit.
980 add_wait_queue(&ws
->wait
, &hctx
->dispatch_wait
);
984 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
)
986 struct blk_mq_hw_ctx
*hctx
;
988 int errors
, queued
, ret
= BLK_MQ_RQ_QUEUE_OK
;
990 if (list_empty(list
))
994 * Now process all the entries, sending them to the driver.
998 struct blk_mq_queue_data bd
;
1000 rq
= list_first_entry(list
, struct request
, queuelist
);
1001 if (!blk_mq_get_driver_tag(rq
, &hctx
, false)) {
1002 if (!queued
&& reorder_tags_to_front(list
))
1006 * The initial allocation attempt failed, so we need to
1007 * rerun the hardware queue when a tag is freed.
1009 if (!blk_mq_dispatch_wait_add(hctx
))
1013 * It's possible that a tag was freed in the window
1014 * between the allocation failure and adding the
1015 * hardware queue to the wait queue.
1017 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1021 list_del_init(&rq
->queuelist
);
1026 * Flag last if we have no more requests, or if we have more
1027 * but can't assign a driver tag to it.
1029 if (list_empty(list
))
1032 struct request
*nxt
;
1034 nxt
= list_first_entry(list
, struct request
, queuelist
);
1035 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1038 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1040 case BLK_MQ_RQ_QUEUE_OK
:
1043 case BLK_MQ_RQ_QUEUE_BUSY
:
1044 blk_mq_put_driver_tag_hctx(hctx
, rq
);
1045 list_add(&rq
->queuelist
, list
);
1046 __blk_mq_requeue_request(rq
);
1049 pr_err("blk-mq: bad return on queue: %d\n", ret
);
1050 case BLK_MQ_RQ_QUEUE_ERROR
:
1052 blk_mq_end_request(rq
, -EIO
);
1056 if (ret
== BLK_MQ_RQ_QUEUE_BUSY
)
1058 } while (!list_empty(list
));
1060 hctx
->dispatched
[queued_to_index(queued
)]++;
1063 * Any items that need requeuing? Stuff them into hctx->dispatch,
1064 * that is where we will continue on next queue run.
1066 if (!list_empty(list
)) {
1068 * If an I/O scheduler has been configured and we got a driver
1069 * tag for the next request already, free it again.
1071 rq
= list_first_entry(list
, struct request
, queuelist
);
1072 blk_mq_put_driver_tag(rq
);
1074 spin_lock(&hctx
->lock
);
1075 list_splice_init(list
, &hctx
->dispatch
);
1076 spin_unlock(&hctx
->lock
);
1079 * If SCHED_RESTART was set by the caller of this function and
1080 * it is no longer set that means that it was cleared by another
1081 * thread and hence that a queue rerun is needed.
1083 * If TAG_WAITING is set that means that an I/O scheduler has
1084 * been configured and another thread is waiting for a driver
1085 * tag. To guarantee fairness, do not rerun this hardware queue
1086 * but let the other thread grab the driver tag.
1088 * If no I/O scheduler has been configured it is possible that
1089 * the hardware queue got stopped and restarted before requests
1090 * were pushed back onto the dispatch list. Rerun the queue to
1091 * avoid starvation. Notes:
1092 * - blk_mq_run_hw_queue() checks whether or not a queue has
1093 * been stopped before rerunning a queue.
1094 * - Some but not all block drivers stop a queue before
1095 * returning BLK_MQ_RQ_QUEUE_BUSY. Two exceptions are scsi-mq
1098 if (!blk_mq_sched_needs_restart(hctx
) &&
1099 !test_bit(BLK_MQ_S_TAG_WAITING
, &hctx
->state
))
1100 blk_mq_run_hw_queue(hctx
, true);
1103 return (queued
+ errors
) != 0;
1106 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1110 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1111 cpu_online(hctx
->next_cpu
));
1113 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1115 blk_mq_sched_dispatch_requests(hctx
);
1120 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1121 blk_mq_sched_dispatch_requests(hctx
);
1122 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1127 * It'd be great if the workqueue API had a way to pass
1128 * in a mask and had some smarts for more clever placement.
1129 * For now we just round-robin here, switching for every
1130 * BLK_MQ_CPU_WORK_BATCH queued items.
1132 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1134 if (hctx
->queue
->nr_hw_queues
== 1)
1135 return WORK_CPU_UNBOUND
;
1137 if (--hctx
->next_cpu_batch
<= 0) {
1140 next_cpu
= cpumask_next(hctx
->next_cpu
, hctx
->cpumask
);
1141 if (next_cpu
>= nr_cpu_ids
)
1142 next_cpu
= cpumask_first(hctx
->cpumask
);
1144 hctx
->next_cpu
= next_cpu
;
1145 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1148 return hctx
->next_cpu
;
1151 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1152 unsigned long msecs
)
1154 if (unlikely(blk_mq_hctx_stopped(hctx
) ||
1155 !blk_mq_hw_queue_mapped(hctx
)))
1158 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1159 int cpu
= get_cpu();
1160 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1161 __blk_mq_run_hw_queue(hctx
);
1170 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx
),
1173 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1174 &hctx
->delayed_run_work
,
1175 msecs_to_jiffies(msecs
));
1178 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1180 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1182 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1184 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1186 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1188 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1190 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1192 struct blk_mq_hw_ctx
*hctx
;
1195 queue_for_each_hw_ctx(q
, hctx
, i
) {
1196 if (!blk_mq_hctx_has_pending(hctx
) ||
1197 blk_mq_hctx_stopped(hctx
))
1200 blk_mq_run_hw_queue(hctx
, async
);
1203 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1206 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1207 * @q: request queue.
1209 * The caller is responsible for serializing this function against
1210 * blk_mq_{start,stop}_hw_queue().
1212 bool blk_mq_queue_stopped(struct request_queue
*q
)
1214 struct blk_mq_hw_ctx
*hctx
;
1217 queue_for_each_hw_ctx(q
, hctx
, i
)
1218 if (blk_mq_hctx_stopped(hctx
))
1223 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1225 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1227 cancel_work(&hctx
->run_work
);
1228 cancel_delayed_work(&hctx
->delay_work
);
1229 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1231 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1233 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1235 struct blk_mq_hw_ctx
*hctx
;
1238 queue_for_each_hw_ctx(q
, hctx
, i
)
1239 blk_mq_stop_hw_queue(hctx
);
1241 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1243 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1245 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1247 blk_mq_run_hw_queue(hctx
, false);
1249 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1251 void blk_mq_start_hw_queues(struct request_queue
*q
)
1253 struct blk_mq_hw_ctx
*hctx
;
1256 queue_for_each_hw_ctx(q
, hctx
, i
)
1257 blk_mq_start_hw_queue(hctx
);
1259 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1261 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1263 if (!blk_mq_hctx_stopped(hctx
))
1266 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1267 blk_mq_run_hw_queue(hctx
, async
);
1269 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1271 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1273 struct blk_mq_hw_ctx
*hctx
;
1276 queue_for_each_hw_ctx(q
, hctx
, i
)
1277 blk_mq_start_stopped_hw_queue(hctx
, async
);
1279 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1281 static void blk_mq_run_work_fn(struct work_struct
*work
)
1283 struct blk_mq_hw_ctx
*hctx
;
1285 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
);
1287 __blk_mq_run_hw_queue(hctx
);
1290 static void blk_mq_delayed_run_work_fn(struct work_struct
*work
)
1292 struct blk_mq_hw_ctx
*hctx
;
1294 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delayed_run_work
.work
);
1296 __blk_mq_run_hw_queue(hctx
);
1299 static void blk_mq_delay_work_fn(struct work_struct
*work
)
1301 struct blk_mq_hw_ctx
*hctx
;
1303 hctx
= container_of(work
, struct blk_mq_hw_ctx
, delay_work
.work
);
1305 if (test_and_clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1306 __blk_mq_run_hw_queue(hctx
);
1309 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1311 if (unlikely(!blk_mq_hw_queue_mapped(hctx
)))
1314 blk_mq_stop_hw_queue(hctx
);
1315 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1316 &hctx
->delay_work
, msecs_to_jiffies(msecs
));
1318 EXPORT_SYMBOL(blk_mq_delay_queue
);
1320 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1324 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1326 trace_block_rq_insert(hctx
->queue
, rq
);
1329 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1331 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1334 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1337 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1339 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1340 blk_mq_hctx_mark_pending(hctx
, ctx
);
1343 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1344 struct list_head
*list
)
1348 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1351 spin_lock(&ctx
->lock
);
1352 while (!list_empty(list
)) {
1355 rq
= list_first_entry(list
, struct request
, queuelist
);
1356 BUG_ON(rq
->mq_ctx
!= ctx
);
1357 list_del_init(&rq
->queuelist
);
1358 __blk_mq_insert_req_list(hctx
, rq
, false);
1360 blk_mq_hctx_mark_pending(hctx
, ctx
);
1361 spin_unlock(&ctx
->lock
);
1364 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1366 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1367 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1369 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1370 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1371 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1374 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1376 struct blk_mq_ctx
*this_ctx
;
1377 struct request_queue
*this_q
;
1380 LIST_HEAD(ctx_list
);
1383 list_splice_init(&plug
->mq_list
, &list
);
1385 list_sort(NULL
, &list
, plug_ctx_cmp
);
1391 while (!list_empty(&list
)) {
1392 rq
= list_entry_rq(list
.next
);
1393 list_del_init(&rq
->queuelist
);
1395 if (rq
->mq_ctx
!= this_ctx
) {
1397 trace_block_unplug(this_q
, depth
, from_schedule
);
1398 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1403 this_ctx
= rq
->mq_ctx
;
1409 list_add_tail(&rq
->queuelist
, &ctx_list
);
1413 * If 'this_ctx' is set, we know we have entries to complete
1414 * on 'ctx_list'. Do those.
1417 trace_block_unplug(this_q
, depth
, from_schedule
);
1418 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1423 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1425 blk_init_request_from_bio(rq
, bio
);
1427 blk_account_io_start(rq
, true);
1430 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx
*hctx
)
1432 return (hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) &&
1433 !blk_queue_nomerges(hctx
->queue
);
1436 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx
*hctx
,
1437 struct blk_mq_ctx
*ctx
,
1438 struct request
*rq
, struct bio
*bio
)
1440 if (!hctx_allow_merges(hctx
) || !bio_mergeable(bio
)) {
1441 blk_mq_bio_to_request(rq
, bio
);
1442 spin_lock(&ctx
->lock
);
1444 __blk_mq_insert_request(hctx
, rq
, false);
1445 spin_unlock(&ctx
->lock
);
1448 struct request_queue
*q
= hctx
->queue
;
1450 spin_lock(&ctx
->lock
);
1451 if (!blk_mq_attempt_merge(q
, ctx
, bio
)) {
1452 blk_mq_bio_to_request(rq
, bio
);
1456 spin_unlock(&ctx
->lock
);
1457 __blk_mq_finish_request(hctx
, ctx
, rq
);
1462 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1465 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1467 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1470 static void __blk_mq_try_issue_directly(struct request
*rq
, blk_qc_t
*cookie
,
1473 struct request_queue
*q
= rq
->q
;
1474 struct blk_mq_queue_data bd
= {
1478 struct blk_mq_hw_ctx
*hctx
;
1479 blk_qc_t new_cookie
;
1485 if (!blk_mq_get_driver_tag(rq
, &hctx
, false))
1488 new_cookie
= request_to_qc_t(hctx
, rq
);
1491 * For OK queue, we are done. For error, kill it. Any other
1492 * error (busy), just add it to our list as we previously
1495 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1496 if (ret
== BLK_MQ_RQ_QUEUE_OK
) {
1497 *cookie
= new_cookie
;
1501 if (ret
== BLK_MQ_RQ_QUEUE_ERROR
) {
1502 *cookie
= BLK_QC_T_NONE
;
1503 blk_mq_end_request(rq
, -EIO
);
1507 __blk_mq_requeue_request(rq
);
1509 blk_mq_sched_insert_request(rq
, false, true, false, may_sleep
);
1512 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1513 struct request
*rq
, blk_qc_t
*cookie
)
1515 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1517 __blk_mq_try_issue_directly(rq
, cookie
, false);
1520 unsigned int srcu_idx
;
1524 srcu_idx
= srcu_read_lock(&hctx
->queue_rq_srcu
);
1525 __blk_mq_try_issue_directly(rq
, cookie
, true);
1526 srcu_read_unlock(&hctx
->queue_rq_srcu
, srcu_idx
);
1530 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1532 const int is_sync
= op_is_sync(bio
->bi_opf
);
1533 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1534 struct blk_mq_alloc_data data
= { .flags
= 0 };
1536 unsigned int request_count
= 0;
1537 struct blk_plug
*plug
;
1538 struct request
*same_queue_rq
= NULL
;
1540 unsigned int wb_acct
;
1542 blk_queue_bounce(q
, &bio
);
1544 if (bio_integrity_enabled(bio
) && bio_integrity_prep(bio
)) {
1546 return BLK_QC_T_NONE
;
1549 blk_queue_split(q
, &bio
, q
->bio_split
);
1551 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1552 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1553 return BLK_QC_T_NONE
;
1555 if (blk_mq_sched_bio_merge(q
, bio
))
1556 return BLK_QC_T_NONE
;
1558 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1560 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1562 rq
= blk_mq_sched_get_request(q
, bio
, bio
->bi_opf
, &data
);
1563 if (unlikely(!rq
)) {
1564 __wbt_done(q
->rq_wb
, wb_acct
);
1565 return BLK_QC_T_NONE
;
1568 wbt_track(&rq
->issue_stat
, wb_acct
);
1570 cookie
= request_to_qc_t(data
.hctx
, rq
);
1572 plug
= current
->plug
;
1573 if (unlikely(is_flush_fua
)) {
1574 blk_mq_bio_to_request(rq
, bio
);
1576 blk_mq_sched_insert_request(rq
, false, true, true,
1579 blk_insert_flush(rq
);
1580 blk_mq_run_hw_queue(data
.hctx
, true);
1582 } else if (plug
&& q
->nr_hw_queues
== 1) {
1583 struct request
*last
= NULL
;
1585 blk_mq_bio_to_request(rq
, bio
);
1588 * @request_count may become stale because of schedule
1589 * out, so check the list again.
1591 if (list_empty(&plug
->mq_list
))
1593 else if (blk_queue_nomerges(q
))
1594 request_count
= blk_plug_queued_count(q
);
1597 trace_block_plug(q
);
1599 last
= list_entry_rq(plug
->mq_list
.prev
);
1601 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1602 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1603 blk_flush_plug_list(plug
, false);
1604 trace_block_plug(q
);
1607 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1608 } else if (plug
&& !blk_queue_nomerges(q
)) {
1609 blk_mq_bio_to_request(rq
, bio
);
1612 * We do limited plugging. If the bio can be merged, do that.
1613 * Otherwise the existing request in the plug list will be
1614 * issued. So the plug list will have one request at most
1615 * The plug list might get flushed before this. If that happens,
1616 * the plug list is empty, and same_queue_rq is invalid.
1618 if (list_empty(&plug
->mq_list
))
1619 same_queue_rq
= NULL
;
1621 list_del_init(&same_queue_rq
->queuelist
);
1622 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1624 blk_mq_put_ctx(data
.ctx
);
1627 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1631 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1632 blk_mq_put_ctx(data
.ctx
);
1633 blk_mq_bio_to_request(rq
, bio
);
1634 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1636 } else if (q
->elevator
) {
1637 blk_mq_bio_to_request(rq
, bio
);
1638 blk_mq_sched_insert_request(rq
, false, true, true, true);
1639 } else if (!blk_mq_merge_queue_io(data
.hctx
, data
.ctx
, rq
, bio
))
1640 blk_mq_run_hw_queue(data
.hctx
, true);
1642 blk_mq_put_ctx(data
.ctx
);
1646 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1647 unsigned int hctx_idx
)
1651 if (tags
->rqs
&& set
->ops
->exit_request
) {
1654 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1655 struct request
*rq
= tags
->static_rqs
[i
];
1659 set
->ops
->exit_request(set
->driver_data
, rq
,
1661 tags
->static_rqs
[i
] = NULL
;
1665 while (!list_empty(&tags
->page_list
)) {
1666 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1667 list_del_init(&page
->lru
);
1669 * Remove kmemleak object previously allocated in
1670 * blk_mq_init_rq_map().
1672 kmemleak_free(page_address(page
));
1673 __free_pages(page
, page
->private);
1677 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1681 kfree(tags
->static_rqs
);
1682 tags
->static_rqs
= NULL
;
1684 blk_mq_free_tags(tags
);
1687 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1688 unsigned int hctx_idx
,
1689 unsigned int nr_tags
,
1690 unsigned int reserved_tags
)
1692 struct blk_mq_tags
*tags
;
1695 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1696 if (node
== NUMA_NO_NODE
)
1697 node
= set
->numa_node
;
1699 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1700 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1704 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1705 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1708 blk_mq_free_tags(tags
);
1712 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1713 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1715 if (!tags
->static_rqs
) {
1717 blk_mq_free_tags(tags
);
1724 static size_t order_to_size(unsigned int order
)
1726 return (size_t)PAGE_SIZE
<< order
;
1729 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1730 unsigned int hctx_idx
, unsigned int depth
)
1732 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1733 size_t rq_size
, left
;
1736 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1737 if (node
== NUMA_NO_NODE
)
1738 node
= set
->numa_node
;
1740 INIT_LIST_HEAD(&tags
->page_list
);
1743 * rq_size is the size of the request plus driver payload, rounded
1744 * to the cacheline size
1746 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1748 left
= rq_size
* depth
;
1750 for (i
= 0; i
< depth
; ) {
1751 int this_order
= max_order
;
1756 while (this_order
&& left
< order_to_size(this_order
- 1))
1760 page
= alloc_pages_node(node
,
1761 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1767 if (order_to_size(this_order
) < rq_size
)
1774 page
->private = this_order
;
1775 list_add_tail(&page
->lru
, &tags
->page_list
);
1777 p
= page_address(page
);
1779 * Allow kmemleak to scan these pages as they contain pointers
1780 * to additional allocations like via ops->init_request().
1782 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1783 entries_per_page
= order_to_size(this_order
) / rq_size
;
1784 to_do
= min(entries_per_page
, depth
- i
);
1785 left
-= to_do
* rq_size
;
1786 for (j
= 0; j
< to_do
; j
++) {
1787 struct request
*rq
= p
;
1789 tags
->static_rqs
[i
] = rq
;
1790 if (set
->ops
->init_request
) {
1791 if (set
->ops
->init_request(set
->driver_data
,
1794 tags
->static_rqs
[i
] = NULL
;
1806 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1811 * 'cpu' is going away. splice any existing rq_list entries from this
1812 * software queue to the hw queue dispatch list, and ensure that it
1815 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
1817 struct blk_mq_hw_ctx
*hctx
;
1818 struct blk_mq_ctx
*ctx
;
1821 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
1822 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
1824 spin_lock(&ctx
->lock
);
1825 if (!list_empty(&ctx
->rq_list
)) {
1826 list_splice_init(&ctx
->rq_list
, &tmp
);
1827 blk_mq_hctx_clear_pending(hctx
, ctx
);
1829 spin_unlock(&ctx
->lock
);
1831 if (list_empty(&tmp
))
1834 spin_lock(&hctx
->lock
);
1835 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
1836 spin_unlock(&hctx
->lock
);
1838 blk_mq_run_hw_queue(hctx
, true);
1842 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
1844 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
1848 /* hctx->ctxs will be freed in queue's release handler */
1849 static void blk_mq_exit_hctx(struct request_queue
*q
,
1850 struct blk_mq_tag_set
*set
,
1851 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
1853 unsigned flush_start_tag
= set
->queue_depth
;
1855 blk_mq_tag_idle(hctx
);
1857 if (set
->ops
->exit_request
)
1858 set
->ops
->exit_request(set
->driver_data
,
1859 hctx
->fq
->flush_rq
, hctx_idx
,
1860 flush_start_tag
+ hctx_idx
);
1862 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1864 if (set
->ops
->exit_hctx
)
1865 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1867 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1868 cleanup_srcu_struct(&hctx
->queue_rq_srcu
);
1870 blk_mq_remove_cpuhp(hctx
);
1871 blk_free_flush_queue(hctx
->fq
);
1872 sbitmap_free(&hctx
->ctx_map
);
1875 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
1876 struct blk_mq_tag_set
*set
, int nr_queue
)
1878 struct blk_mq_hw_ctx
*hctx
;
1881 queue_for_each_hw_ctx(q
, hctx
, i
) {
1884 blk_mq_exit_hctx(q
, set
, hctx
, i
);
1888 static int blk_mq_init_hctx(struct request_queue
*q
,
1889 struct blk_mq_tag_set
*set
,
1890 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
1893 unsigned flush_start_tag
= set
->queue_depth
;
1895 node
= hctx
->numa_node
;
1896 if (node
== NUMA_NO_NODE
)
1897 node
= hctx
->numa_node
= set
->numa_node
;
1899 INIT_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
1900 INIT_DELAYED_WORK(&hctx
->delayed_run_work
, blk_mq_delayed_run_work_fn
);
1901 INIT_DELAYED_WORK(&hctx
->delay_work
, blk_mq_delay_work_fn
);
1902 spin_lock_init(&hctx
->lock
);
1903 INIT_LIST_HEAD(&hctx
->dispatch
);
1905 hctx
->queue_num
= hctx_idx
;
1906 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
1908 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
1910 hctx
->tags
= set
->tags
[hctx_idx
];
1913 * Allocate space for all possible cpus to avoid allocation at
1916 hctx
->ctxs
= kmalloc_node(nr_cpu_ids
* sizeof(void *),
1919 goto unregister_cpu_notifier
;
1921 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
1927 if (set
->ops
->init_hctx
&&
1928 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
1931 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
1934 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
1936 goto sched_exit_hctx
;
1938 if (set
->ops
->init_request
&&
1939 set
->ops
->init_request(set
->driver_data
,
1940 hctx
->fq
->flush_rq
, hctx_idx
,
1941 flush_start_tag
+ hctx_idx
, node
))
1944 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
1945 init_srcu_struct(&hctx
->queue_rq_srcu
);
1952 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
1954 if (set
->ops
->exit_hctx
)
1955 set
->ops
->exit_hctx(hctx
, hctx_idx
);
1957 sbitmap_free(&hctx
->ctx_map
);
1960 unregister_cpu_notifier
:
1961 blk_mq_remove_cpuhp(hctx
);
1965 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
1966 unsigned int nr_hw_queues
)
1970 for_each_possible_cpu(i
) {
1971 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
1972 struct blk_mq_hw_ctx
*hctx
;
1975 spin_lock_init(&__ctx
->lock
);
1976 INIT_LIST_HEAD(&__ctx
->rq_list
);
1979 /* If the cpu isn't online, the cpu is mapped to first hctx */
1983 hctx
= blk_mq_map_queue(q
, i
);
1986 * Set local node, IFF we have more than one hw queue. If
1987 * not, we remain on the home node of the device
1989 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
1990 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
1994 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
1998 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
1999 set
->queue_depth
, set
->reserved_tags
);
2000 if (!set
->tags
[hctx_idx
])
2003 ret
= blk_mq_alloc_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_free_map_and_requests(struct blk_mq_tag_set
*set
,
2014 unsigned int hctx_idx
)
2016 if (set
->tags
[hctx_idx
]) {
2017 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2018 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2019 set
->tags
[hctx_idx
] = NULL
;
2023 static void blk_mq_map_swqueue(struct request_queue
*q
,
2024 const struct cpumask
*online_mask
)
2026 unsigned int i
, hctx_idx
;
2027 struct blk_mq_hw_ctx
*hctx
;
2028 struct blk_mq_ctx
*ctx
;
2029 struct blk_mq_tag_set
*set
= q
->tag_set
;
2032 * Avoid others reading imcomplete hctx->cpumask through sysfs
2034 mutex_lock(&q
->sysfs_lock
);
2036 queue_for_each_hw_ctx(q
, hctx
, i
) {
2037 cpumask_clear(hctx
->cpumask
);
2042 * Map software to hardware queues
2044 for_each_possible_cpu(i
) {
2045 /* If the cpu isn't online, the cpu is mapped to first hctx */
2046 if (!cpumask_test_cpu(i
, online_mask
))
2049 hctx_idx
= q
->mq_map
[i
];
2050 /* unmapped hw queue can be remapped after CPU topo changed */
2051 if (!set
->tags
[hctx_idx
] &&
2052 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2054 * If tags initialization fail for some hctx,
2055 * that hctx won't be brought online. In this
2056 * case, remap the current ctx to hctx[0] which
2057 * is guaranteed to always have tags allocated
2062 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2063 hctx
= blk_mq_map_queue(q
, i
);
2065 cpumask_set_cpu(i
, hctx
->cpumask
);
2066 ctx
->index_hw
= hctx
->nr_ctx
;
2067 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2070 mutex_unlock(&q
->sysfs_lock
);
2072 queue_for_each_hw_ctx(q
, hctx
, i
) {
2074 * If no software queues are mapped to this hardware queue,
2075 * disable it and free the request entries.
2077 if (!hctx
->nr_ctx
) {
2078 /* Never unmap queue 0. We need it as a
2079 * fallback in case of a new remap fails
2082 if (i
&& set
->tags
[i
])
2083 blk_mq_free_map_and_requests(set
, i
);
2089 hctx
->tags
= set
->tags
[i
];
2090 WARN_ON(!hctx
->tags
);
2093 * Set the map size to the number of mapped software queues.
2094 * This is more accurate and more efficient than looping
2095 * over all possibly mapped software queues.
2097 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2100 * Initialize batch roundrobin counts
2102 hctx
->next_cpu
= cpumask_first(hctx
->cpumask
);
2103 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2107 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2109 struct blk_mq_hw_ctx
*hctx
;
2112 queue_for_each_hw_ctx(q
, hctx
, i
) {
2114 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2116 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2120 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
, bool shared
)
2122 struct request_queue
*q
;
2124 lockdep_assert_held(&set
->tag_list_lock
);
2126 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2127 blk_mq_freeze_queue(q
);
2128 queue_set_hctx_shared(q
, shared
);
2129 blk_mq_unfreeze_queue(q
);
2133 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2135 struct blk_mq_tag_set
*set
= q
->tag_set
;
2137 mutex_lock(&set
->tag_list_lock
);
2138 list_del_rcu(&q
->tag_set_list
);
2139 INIT_LIST_HEAD(&q
->tag_set_list
);
2140 if (list_is_singular(&set
->tag_list
)) {
2141 /* just transitioned to unshared */
2142 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2143 /* update existing queue */
2144 blk_mq_update_tag_set_depth(set
, false);
2146 mutex_unlock(&set
->tag_list_lock
);
2151 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2152 struct request_queue
*q
)
2156 mutex_lock(&set
->tag_list_lock
);
2158 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2159 if (!list_empty(&set
->tag_list
) && !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2160 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2161 /* update existing queue */
2162 blk_mq_update_tag_set_depth(set
, true);
2164 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2165 queue_set_hctx_shared(q
, true);
2166 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2168 mutex_unlock(&set
->tag_list_lock
);
2172 * It is the actual release handler for mq, but we do it from
2173 * request queue's release handler for avoiding use-after-free
2174 * and headache because q->mq_kobj shouldn't have been introduced,
2175 * but we can't group ctx/kctx kobj without it.
2177 void blk_mq_release(struct request_queue
*q
)
2179 struct blk_mq_hw_ctx
*hctx
;
2182 /* hctx kobj stays in hctx */
2183 queue_for_each_hw_ctx(q
, hctx
, i
) {
2186 kobject_put(&hctx
->kobj
);
2191 kfree(q
->queue_hw_ctx
);
2194 * release .mq_kobj and sw queue's kobject now because
2195 * both share lifetime with request queue.
2197 blk_mq_sysfs_deinit(q
);
2199 free_percpu(q
->queue_ctx
);
2202 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2204 struct request_queue
*uninit_q
, *q
;
2206 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2208 return ERR_PTR(-ENOMEM
);
2210 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2212 blk_cleanup_queue(uninit_q
);
2216 EXPORT_SYMBOL(blk_mq_init_queue
);
2218 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2219 struct request_queue
*q
)
2222 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2224 blk_mq_sysfs_unregister(q
);
2225 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2231 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2232 hctxs
[i
] = kzalloc_node(sizeof(struct blk_mq_hw_ctx
),
2237 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2244 atomic_set(&hctxs
[i
]->nr_active
, 0);
2245 hctxs
[i
]->numa_node
= node
;
2246 hctxs
[i
]->queue_num
= i
;
2248 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2249 free_cpumask_var(hctxs
[i
]->cpumask
);
2254 blk_mq_hctx_kobj_init(hctxs
[i
]);
2256 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2257 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2261 blk_mq_free_map_and_requests(set
, j
);
2262 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2263 kobject_put(&hctx
->kobj
);
2268 q
->nr_hw_queues
= i
;
2269 blk_mq_sysfs_register(q
);
2272 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2273 struct request_queue
*q
)
2275 /* mark the queue as mq asap */
2276 q
->mq_ops
= set
->ops
;
2278 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2279 blk_mq_poll_stats_bkt
,
2280 BLK_MQ_POLL_STATS_BKTS
, q
);
2284 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2288 /* init q->mq_kobj and sw queues' kobjects */
2289 blk_mq_sysfs_init(q
);
2291 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2292 GFP_KERNEL
, set
->numa_node
);
2293 if (!q
->queue_hw_ctx
)
2296 q
->mq_map
= set
->mq_map
;
2298 blk_mq_realloc_hw_ctxs(set
, q
);
2299 if (!q
->nr_hw_queues
)
2302 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2303 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2305 q
->nr_queues
= nr_cpu_ids
;
2307 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2309 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2310 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2312 q
->sg_reserved_size
= INT_MAX
;
2314 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2315 INIT_LIST_HEAD(&q
->requeue_list
);
2316 spin_lock_init(&q
->requeue_lock
);
2318 blk_queue_make_request(q
, blk_mq_make_request
);
2321 * Do this after blk_queue_make_request() overrides it...
2323 q
->nr_requests
= set
->queue_depth
;
2326 * Default to classic polling
2330 if (set
->ops
->complete
)
2331 blk_queue_softirq_done(q
, set
->ops
->complete
);
2333 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2336 mutex_lock(&all_q_mutex
);
2338 list_add_tail(&q
->all_q_node
, &all_q_list
);
2339 blk_mq_add_queue_tag_set(set
, q
);
2340 blk_mq_map_swqueue(q
, cpu_online_mask
);
2342 mutex_unlock(&all_q_mutex
);
2345 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2348 ret
= blk_mq_sched_init(q
);
2350 return ERR_PTR(ret
);
2356 kfree(q
->queue_hw_ctx
);
2358 free_percpu(q
->queue_ctx
);
2361 return ERR_PTR(-ENOMEM
);
2363 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2365 void blk_mq_free_queue(struct request_queue
*q
)
2367 struct blk_mq_tag_set
*set
= q
->tag_set
;
2369 mutex_lock(&all_q_mutex
);
2370 list_del_init(&q
->all_q_node
);
2371 mutex_unlock(&all_q_mutex
);
2373 blk_mq_del_queue_tag_set(q
);
2375 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2378 /* Basically redo blk_mq_init_queue with queue frozen */
2379 static void blk_mq_queue_reinit(struct request_queue
*q
,
2380 const struct cpumask
*online_mask
)
2382 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2384 blk_mq_sysfs_unregister(q
);
2387 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2388 * we should change hctx numa_node according to new topology (this
2389 * involves free and re-allocate memory, worthy doing?)
2392 blk_mq_map_swqueue(q
, online_mask
);
2394 blk_mq_sysfs_register(q
);
2398 * New online cpumask which is going to be set in this hotplug event.
2399 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2400 * one-by-one and dynamically allocating this could result in a failure.
2402 static struct cpumask cpuhp_online_new
;
2404 static void blk_mq_queue_reinit_work(void)
2406 struct request_queue
*q
;
2408 mutex_lock(&all_q_mutex
);
2410 * We need to freeze and reinit all existing queues. Freezing
2411 * involves synchronous wait for an RCU grace period and doing it
2412 * one by one may take a long time. Start freezing all queues in
2413 * one swoop and then wait for the completions so that freezing can
2414 * take place in parallel.
2416 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2417 blk_freeze_queue_start(q
);
2418 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2419 blk_mq_freeze_queue_wait(q
);
2421 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2422 blk_mq_queue_reinit(q
, &cpuhp_online_new
);
2424 list_for_each_entry(q
, &all_q_list
, all_q_node
)
2425 blk_mq_unfreeze_queue(q
);
2427 mutex_unlock(&all_q_mutex
);
2430 static int blk_mq_queue_reinit_dead(unsigned int cpu
)
2432 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2433 blk_mq_queue_reinit_work();
2438 * Before hotadded cpu starts handling requests, new mappings must be
2439 * established. Otherwise, these requests in hw queue might never be
2442 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2443 * for CPU0, and ctx1 for CPU1).
2445 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2446 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2448 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2449 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2450 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2453 static int blk_mq_queue_reinit_prepare(unsigned int cpu
)
2455 cpumask_copy(&cpuhp_online_new
, cpu_online_mask
);
2456 cpumask_set_cpu(cpu
, &cpuhp_online_new
);
2457 blk_mq_queue_reinit_work();
2461 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2465 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2466 if (!__blk_mq_alloc_rq_map(set
, i
))
2473 blk_mq_free_rq_map(set
->tags
[i
]);
2479 * Allocate the request maps associated with this tag_set. Note that this
2480 * may reduce the depth asked for, if memory is tight. set->queue_depth
2481 * will be updated to reflect the allocated depth.
2483 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2488 depth
= set
->queue_depth
;
2490 err
= __blk_mq_alloc_rq_maps(set
);
2494 set
->queue_depth
>>= 1;
2495 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2499 } while (set
->queue_depth
);
2501 if (!set
->queue_depth
|| err
) {
2502 pr_err("blk-mq: failed to allocate request map\n");
2506 if (depth
!= set
->queue_depth
)
2507 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2508 depth
, set
->queue_depth
);
2513 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2515 if (set
->ops
->map_queues
)
2516 return set
->ops
->map_queues(set
);
2518 return blk_mq_map_queues(set
);
2522 * Alloc a tag set to be associated with one or more request queues.
2523 * May fail with EINVAL for various error conditions. May adjust the
2524 * requested depth down, if if it too large. In that case, the set
2525 * value will be stored in set->queue_depth.
2527 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2531 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2533 if (!set
->nr_hw_queues
)
2535 if (!set
->queue_depth
)
2537 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2540 if (!set
->ops
->queue_rq
)
2543 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2544 pr_info("blk-mq: reduced tag depth to %u\n",
2546 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2550 * If a crashdump is active, then we are potentially in a very
2551 * memory constrained environment. Limit us to 1 queue and
2552 * 64 tags to prevent using too much memory.
2554 if (is_kdump_kernel()) {
2555 set
->nr_hw_queues
= 1;
2556 set
->queue_depth
= min(64U, set
->queue_depth
);
2559 * There is no use for more h/w queues than cpus.
2561 if (set
->nr_hw_queues
> nr_cpu_ids
)
2562 set
->nr_hw_queues
= nr_cpu_ids
;
2564 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2565 GFP_KERNEL
, set
->numa_node
);
2570 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2571 GFP_KERNEL
, set
->numa_node
);
2575 ret
= blk_mq_update_queue_map(set
);
2577 goto out_free_mq_map
;
2579 ret
= blk_mq_alloc_rq_maps(set
);
2581 goto out_free_mq_map
;
2583 mutex_init(&set
->tag_list_lock
);
2584 INIT_LIST_HEAD(&set
->tag_list
);
2596 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2598 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2602 for (i
= 0; i
< nr_cpu_ids
; i
++)
2603 blk_mq_free_map_and_requests(set
, i
);
2611 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2613 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2615 struct blk_mq_tag_set
*set
= q
->tag_set
;
2616 struct blk_mq_hw_ctx
*hctx
;
2622 blk_mq_freeze_queue(q
);
2623 blk_mq_quiesce_queue(q
);
2626 queue_for_each_hw_ctx(q
, hctx
, i
) {
2630 * If we're using an MQ scheduler, just update the scheduler
2631 * queue depth. This is similar to what the old code would do.
2633 if (!hctx
->sched_tags
) {
2634 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
,
2635 min(nr
, set
->queue_depth
),
2638 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2646 q
->nr_requests
= nr
;
2648 blk_mq_unfreeze_queue(q
);
2649 blk_mq_start_stopped_hw_queues(q
, true);
2654 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2656 struct request_queue
*q
;
2658 lockdep_assert_held(&set
->tag_list_lock
);
2660 if (nr_hw_queues
> nr_cpu_ids
)
2661 nr_hw_queues
= nr_cpu_ids
;
2662 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2665 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2666 blk_mq_freeze_queue(q
);
2668 set
->nr_hw_queues
= nr_hw_queues
;
2669 blk_mq_update_queue_map(set
);
2670 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2671 blk_mq_realloc_hw_ctxs(set
, q
);
2672 blk_mq_queue_reinit(q
, cpu_online_mask
);
2675 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2676 blk_mq_unfreeze_queue(q
);
2678 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2680 /* Enable polling stats and return whether they were already enabled. */
2681 static bool blk_poll_stats_enable(struct request_queue
*q
)
2683 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2684 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
2686 blk_stat_add_callback(q
, q
->poll_cb
);
2690 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2693 * We don't arm the callback if polling stats are not enabled or the
2694 * callback is already active.
2696 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2697 blk_stat_is_active(q
->poll_cb
))
2700 blk_stat_activate_msecs(q
->poll_cb
, 100);
2703 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2705 struct request_queue
*q
= cb
->data
;
2708 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
2709 if (cb
->stat
[bucket
].nr_samples
)
2710 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
2714 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2715 struct blk_mq_hw_ctx
*hctx
,
2718 unsigned long ret
= 0;
2722 * If stats collection isn't on, don't sleep but turn it on for
2725 if (!blk_poll_stats_enable(q
))
2729 * As an optimistic guess, use half of the mean service time
2730 * for this type of request. We can (and should) make this smarter.
2731 * For instance, if the completion latencies are tight, we can
2732 * get closer than just half the mean. This is especially
2733 * important on devices where the completion latencies are longer
2734 * than ~10 usec. We do use the stats for the relevant IO size
2735 * if available which does lead to better estimates.
2737 bucket
= blk_mq_poll_stats_bkt(rq
);
2741 if (q
->poll_stat
[bucket
].nr_samples
)
2742 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
2747 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2748 struct blk_mq_hw_ctx
*hctx
,
2751 struct hrtimer_sleeper hs
;
2752 enum hrtimer_mode mode
;
2756 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2762 * -1: don't ever hybrid sleep
2763 * 0: use half of prev avg
2764 * >0: use this specific value
2766 if (q
->poll_nsec
== -1)
2768 else if (q
->poll_nsec
> 0)
2769 nsecs
= q
->poll_nsec
;
2771 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2776 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2779 * This will be replaced with the stats tracking code, using
2780 * 'avg_completion_time / 2' as the pre-sleep target.
2784 mode
= HRTIMER_MODE_REL
;
2785 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2786 hrtimer_set_expires(&hs
.timer
, kt
);
2788 hrtimer_init_sleeper(&hs
, current
);
2790 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2792 set_current_state(TASK_UNINTERRUPTIBLE
);
2793 hrtimer_start_expires(&hs
.timer
, mode
);
2796 hrtimer_cancel(&hs
.timer
);
2797 mode
= HRTIMER_MODE_ABS
;
2798 } while (hs
.task
&& !signal_pending(current
));
2800 __set_current_state(TASK_RUNNING
);
2801 destroy_hrtimer_on_stack(&hs
.timer
);
2805 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2807 struct request_queue
*q
= hctx
->queue
;
2811 * If we sleep, have the caller restart the poll loop to reset
2812 * the state. Like for the other success return cases, the
2813 * caller is responsible for checking if the IO completed. If
2814 * the IO isn't complete, we'll get called again and will go
2815 * straight to the busy poll loop.
2817 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2820 hctx
->poll_considered
++;
2822 state
= current
->state
;
2823 while (!need_resched()) {
2826 hctx
->poll_invoked
++;
2828 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2830 hctx
->poll_success
++;
2831 set_current_state(TASK_RUNNING
);
2835 if (signal_pending_state(state
, current
))
2836 set_current_state(TASK_RUNNING
);
2838 if (current
->state
== TASK_RUNNING
)
2848 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
2850 struct blk_mq_hw_ctx
*hctx
;
2851 struct blk_plug
*plug
;
2854 if (!q
->mq_ops
|| !q
->mq_ops
->poll
|| !blk_qc_t_valid(cookie
) ||
2855 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
2858 plug
= current
->plug
;
2860 blk_flush_plug_list(plug
, false);
2862 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
2863 if (!blk_qc_t_is_internal(cookie
))
2864 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
2866 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
2868 return __blk_mq_poll(hctx
, rq
);
2870 EXPORT_SYMBOL_GPL(blk_mq_poll
);
2872 void blk_mq_disable_hotplug(void)
2874 mutex_lock(&all_q_mutex
);
2877 void blk_mq_enable_hotplug(void)
2879 mutex_unlock(&all_q_mutex
);
2882 static int __init
blk_mq_init(void)
2884 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
2885 blk_mq_hctx_notify_dead
);
2887 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE
, "block/mq:prepare",
2888 blk_mq_queue_reinit_prepare
,
2889 blk_mq_queue_reinit_dead
);
2892 subsys_initcall(blk_mq_init
);