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 bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
);
41 static void blk_mq_poll_stats_start(struct request_queue
*q
);
42 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
44 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
46 int ddir
, bytes
, bucket
;
48 ddir
= rq_data_dir(rq
);
49 bytes
= blk_rq_bytes(rq
);
51 bucket
= ddir
+ 2*(ilog2(bytes
) - 9);
55 else if (bucket
>= BLK_MQ_POLL_STATS_BKTS
)
56 return ddir
+ BLK_MQ_POLL_STATS_BKTS
- 2;
62 * Check if any of the ctx's have pending work in this hardware queue
64 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
66 return !list_empty_careful(&hctx
->dispatch
) ||
67 sbitmap_any_bit_set(&hctx
->ctx_map
) ||
68 blk_mq_sched_has_work(hctx
);
72 * Mark this ctx as having pending work in this hardware queue
74 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
75 struct blk_mq_ctx
*ctx
)
77 if (!sbitmap_test_bit(&hctx
->ctx_map
, ctx
->index_hw
))
78 sbitmap_set_bit(&hctx
->ctx_map
, ctx
->index_hw
);
81 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
82 struct blk_mq_ctx
*ctx
)
84 sbitmap_clear_bit(&hctx
->ctx_map
, ctx
->index_hw
);
88 struct hd_struct
*part
;
89 unsigned int *inflight
;
92 static void blk_mq_check_inflight(struct blk_mq_hw_ctx
*hctx
,
93 struct request
*rq
, void *priv
,
96 struct mq_inflight
*mi
= priv
;
98 if (test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
) &&
99 !test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
)) {
101 * index[0] counts the specific partition that was asked
102 * for. index[1] counts the ones that are active on the
103 * whole device, so increment that if mi->part is indeed
104 * a partition, and not a whole device.
106 if (rq
->part
== mi
->part
)
108 if (mi
->part
->partno
)
113 void blk_mq_in_flight(struct request_queue
*q
, struct hd_struct
*part
,
114 unsigned int inflight
[2])
116 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
118 inflight
[0] = inflight
[1] = 0;
119 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
122 void blk_freeze_queue_start(struct request_queue
*q
)
126 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
127 if (freeze_depth
== 1) {
128 percpu_ref_kill(&q
->q_usage_counter
);
130 blk_mq_run_hw_queues(q
, false);
133 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
135 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
137 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
139 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
141 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
142 unsigned long timeout
)
144 return wait_event_timeout(q
->mq_freeze_wq
,
145 percpu_ref_is_zero(&q
->q_usage_counter
),
148 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
151 * Guarantee no request is in use, so we can change any data structure of
152 * the queue afterward.
154 void blk_freeze_queue(struct request_queue
*q
)
157 * In the !blk_mq case we are only calling this to kill the
158 * q_usage_counter, otherwise this increases the freeze depth
159 * and waits for it to return to zero. For this reason there is
160 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
161 * exported to drivers as the only user for unfreeze is blk_mq.
163 blk_freeze_queue_start(q
);
166 blk_mq_freeze_queue_wait(q
);
169 void blk_mq_freeze_queue(struct request_queue
*q
)
172 * ...just an alias to keep freeze and unfreeze actions balanced
173 * in the blk_mq_* namespace
177 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
179 void blk_mq_unfreeze_queue(struct request_queue
*q
)
183 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
184 WARN_ON_ONCE(freeze_depth
< 0);
186 percpu_ref_reinit(&q
->q_usage_counter
);
187 wake_up_all(&q
->mq_freeze_wq
);
190 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
193 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
194 * mpt3sas driver such that this function can be removed.
196 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
200 spin_lock_irqsave(q
->queue_lock
, flags
);
201 queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
202 spin_unlock_irqrestore(q
->queue_lock
, flags
);
204 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
207 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
210 * Note: this function does not prevent that the struct request end_io()
211 * callback function is invoked. Once this function is returned, we make
212 * sure no dispatch can happen until the queue is unquiesced via
213 * blk_mq_unquiesce_queue().
215 void blk_mq_quiesce_queue(struct request_queue
*q
)
217 struct blk_mq_hw_ctx
*hctx
;
221 blk_mq_quiesce_queue_nowait(q
);
223 queue_for_each_hw_ctx(q
, hctx
, i
) {
224 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
225 synchronize_srcu(hctx
->queue_rq_srcu
);
232 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
235 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
238 * This function recovers queue into the state before quiescing
239 * which is done by blk_mq_quiesce_queue.
241 void blk_mq_unquiesce_queue(struct request_queue
*q
)
245 spin_lock_irqsave(q
->queue_lock
, flags
);
246 queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
247 spin_unlock_irqrestore(q
->queue_lock
, flags
);
249 /* dispatch requests which are inserted during quiescing */
250 blk_mq_run_hw_queues(q
, true);
252 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
254 void blk_mq_wake_waiters(struct request_queue
*q
)
256 struct blk_mq_hw_ctx
*hctx
;
259 queue_for_each_hw_ctx(q
, hctx
, i
)
260 if (blk_mq_hw_queue_mapped(hctx
))
261 blk_mq_tag_wakeup_all(hctx
->tags
, true);
264 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
266 return blk_mq_has_free_tags(hctx
->tags
);
268 EXPORT_SYMBOL(blk_mq_can_queue
);
270 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
271 unsigned int tag
, unsigned int op
)
273 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
274 struct request
*rq
= tags
->static_rqs
[tag
];
278 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
280 rq
->internal_tag
= tag
;
282 if (blk_mq_tag_busy(data
->hctx
)) {
283 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
284 atomic_inc(&data
->hctx
->nr_active
);
287 rq
->internal_tag
= -1;
288 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
291 INIT_LIST_HEAD(&rq
->queuelist
);
292 /* csd/requeue_work/fifo_time is initialized before use */
294 rq
->mq_ctx
= data
->ctx
;
296 if (data
->flags
& BLK_MQ_REQ_PREEMPT
)
297 rq
->rq_flags
|= RQF_PREEMPT
;
298 if (blk_queue_io_stat(data
->q
))
299 rq
->rq_flags
|= RQF_IO_STAT
;
300 /* do not touch atomic flags, it needs atomic ops against the timer */
302 INIT_HLIST_NODE(&rq
->hash
);
303 RB_CLEAR_NODE(&rq
->rb_node
);
306 rq
->start_time
= jiffies
;
307 #ifdef CONFIG_BLK_CGROUP
309 set_start_time_ns(rq
);
310 rq
->io_start_time_ns
= 0;
312 rq
->nr_phys_segments
= 0;
313 #if defined(CONFIG_BLK_DEV_INTEGRITY)
314 rq
->nr_integrity_segments
= 0;
317 /* tag was already set */
320 INIT_LIST_HEAD(&rq
->timeout_list
);
324 rq
->end_io_data
= NULL
;
327 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
331 static struct request
*blk_mq_get_request(struct request_queue
*q
,
332 struct bio
*bio
, unsigned int op
,
333 struct blk_mq_alloc_data
*data
)
335 struct elevator_queue
*e
= q
->elevator
;
338 bool put_ctx_on_error
= false;
340 blk_queue_enter_live(q
);
342 if (likely(!data
->ctx
)) {
343 data
->ctx
= blk_mq_get_ctx(q
);
344 put_ctx_on_error
= true;
346 if (likely(!data
->hctx
))
347 data
->hctx
= blk_mq_map_queue(q
, data
->ctx
->cpu
);
349 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
352 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
355 * Flush requests are special and go directly to the
358 if (!op_is_flush(op
) && e
->type
->ops
.mq
.limit_depth
)
359 e
->type
->ops
.mq
.limit_depth(op
, data
);
362 tag
= blk_mq_get_tag(data
);
363 if (tag
== BLK_MQ_TAG_FAIL
) {
364 if (put_ctx_on_error
) {
365 blk_mq_put_ctx(data
->ctx
);
372 rq
= blk_mq_rq_ctx_init(data
, tag
, op
);
373 if (!op_is_flush(op
)) {
375 if (e
&& e
->type
->ops
.mq
.prepare_request
) {
376 if (e
->type
->icq_cache
&& rq_ioc(bio
))
377 blk_mq_sched_assign_ioc(rq
, bio
);
379 e
->type
->ops
.mq
.prepare_request(rq
, bio
);
380 rq
->rq_flags
|= RQF_ELVPRIV
;
383 data
->hctx
->queued
++;
387 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
388 blk_mq_req_flags_t flags
)
390 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
394 ret
= blk_queue_enter(q
, flags
);
398 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
402 return ERR_PTR(-EWOULDBLOCK
);
404 blk_mq_put_ctx(alloc_data
.ctx
);
407 rq
->__sector
= (sector_t
) -1;
408 rq
->bio
= rq
->biotail
= NULL
;
411 EXPORT_SYMBOL(blk_mq_alloc_request
);
413 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
414 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
416 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
422 * If the tag allocator sleeps we could get an allocation for a
423 * different hardware context. No need to complicate the low level
424 * allocator for this for the rare use case of a command tied to
427 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
428 return ERR_PTR(-EINVAL
);
430 if (hctx_idx
>= q
->nr_hw_queues
)
431 return ERR_PTR(-EIO
);
433 ret
= blk_queue_enter(q
, flags
);
438 * Check if the hardware context is actually mapped to anything.
439 * If not tell the caller that it should skip this queue.
441 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
442 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
444 return ERR_PTR(-EXDEV
);
446 cpu
= cpumask_first_and(alloc_data
.hctx
->cpumask
, cpu_online_mask
);
447 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
449 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
453 return ERR_PTR(-EWOULDBLOCK
);
457 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
459 void blk_mq_free_request(struct request
*rq
)
461 struct request_queue
*q
= rq
->q
;
462 struct elevator_queue
*e
= q
->elevator
;
463 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
464 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
465 const int sched_tag
= rq
->internal_tag
;
467 if (rq
->rq_flags
& RQF_ELVPRIV
) {
468 if (e
&& e
->type
->ops
.mq
.finish_request
)
469 e
->type
->ops
.mq
.finish_request(rq
);
471 put_io_context(rq
->elv
.icq
->ioc
);
476 ctx
->rq_completed
[rq_is_sync(rq
)]++;
477 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
478 atomic_dec(&hctx
->nr_active
);
480 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
481 laptop_io_completion(q
->backing_dev_info
);
483 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
486 blk_put_rl(blk_rq_rl(rq
));
488 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
489 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
491 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
493 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
494 blk_mq_sched_restart(hctx
);
497 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
499 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
501 blk_account_io_done(rq
);
504 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
505 rq
->end_io(rq
, error
);
507 if (unlikely(blk_bidi_rq(rq
)))
508 blk_mq_free_request(rq
->next_rq
);
509 blk_mq_free_request(rq
);
512 EXPORT_SYMBOL(__blk_mq_end_request
);
514 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
516 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
518 __blk_mq_end_request(rq
, error
);
520 EXPORT_SYMBOL(blk_mq_end_request
);
522 static void __blk_mq_complete_request_remote(void *data
)
524 struct request
*rq
= data
;
526 rq
->q
->softirq_done_fn(rq
);
529 static void __blk_mq_complete_request(struct request
*rq
)
531 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
535 if (rq
->internal_tag
!= -1)
536 blk_mq_sched_completed_request(rq
);
537 if (rq
->rq_flags
& RQF_STATS
) {
538 blk_mq_poll_stats_start(rq
->q
);
542 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
543 rq
->q
->softirq_done_fn(rq
);
548 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
549 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
551 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
552 rq
->csd
.func
= __blk_mq_complete_request_remote
;
555 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
557 rq
->q
->softirq_done_fn(rq
);
563 * blk_mq_complete_request - end I/O on a request
564 * @rq: the request being processed
567 * Ends all I/O on a request. It does not handle partial completions.
568 * The actual completion happens out-of-order, through a IPI handler.
570 void blk_mq_complete_request(struct request
*rq
)
572 struct request_queue
*q
= rq
->q
;
574 if (unlikely(blk_should_fake_timeout(q
)))
576 if (!blk_mark_rq_complete(rq
))
577 __blk_mq_complete_request(rq
);
579 EXPORT_SYMBOL(blk_mq_complete_request
);
581 int blk_mq_request_started(struct request
*rq
)
583 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
585 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
587 void blk_mq_start_request(struct request
*rq
)
589 struct request_queue
*q
= rq
->q
;
591 blk_mq_sched_started_request(rq
);
593 trace_block_rq_issue(q
, rq
);
595 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
596 blk_stat_set_issue(&rq
->issue_stat
, blk_rq_sectors(rq
));
597 rq
->rq_flags
|= RQF_STATS
;
598 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
603 WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
));
606 * Mark us as started and clear complete. Complete might have been
607 * set if requeue raced with timeout, which then marked it as
608 * complete. So be sure to clear complete again when we start
609 * the request, otherwise we'll ignore the completion event.
611 * Ensure that ->deadline is visible before we set STARTED, such that
612 * blk_mq_check_expired() is guaranteed to observe our ->deadline when
613 * it observes STARTED.
616 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
617 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
)) {
619 * Coherence order guarantees these consecutive stores to a
620 * single variable propagate in the specified order. Thus the
621 * clear_bit() is ordered _after_ the set bit. See
622 * blk_mq_check_expired().
624 * (the bits must be part of the same byte for this to be
627 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
630 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
632 * Make sure space for the drain appears. We know we can do
633 * this because max_hw_segments has been adjusted to be one
634 * fewer than the device can handle.
636 rq
->nr_phys_segments
++;
639 EXPORT_SYMBOL(blk_mq_start_request
);
642 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
643 * flag isn't set yet, so there may be race with timeout handler,
644 * but given rq->deadline is just set in .queue_rq() under
645 * this situation, the race won't be possible in reality because
646 * rq->timeout should be set as big enough to cover the window
647 * between blk_mq_start_request() called from .queue_rq() and
648 * clearing REQ_ATOM_STARTED here.
650 static void __blk_mq_requeue_request(struct request
*rq
)
652 struct request_queue
*q
= rq
->q
;
654 blk_mq_put_driver_tag(rq
);
656 trace_block_rq_requeue(q
, rq
);
657 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
659 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
660 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
661 rq
->nr_phys_segments
--;
665 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
667 __blk_mq_requeue_request(rq
);
669 /* this request will be re-inserted to io scheduler queue */
670 blk_mq_sched_requeue_request(rq
);
672 BUG_ON(blk_queued_rq(rq
));
673 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
675 EXPORT_SYMBOL(blk_mq_requeue_request
);
677 static void blk_mq_requeue_work(struct work_struct
*work
)
679 struct request_queue
*q
=
680 container_of(work
, struct request_queue
, requeue_work
.work
);
682 struct request
*rq
, *next
;
684 spin_lock_irq(&q
->requeue_lock
);
685 list_splice_init(&q
->requeue_list
, &rq_list
);
686 spin_unlock_irq(&q
->requeue_lock
);
688 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
689 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
692 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
693 list_del_init(&rq
->queuelist
);
694 blk_mq_sched_insert_request(rq
, true, false, false, true);
697 while (!list_empty(&rq_list
)) {
698 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
699 list_del_init(&rq
->queuelist
);
700 blk_mq_sched_insert_request(rq
, false, false, false, true);
703 blk_mq_run_hw_queues(q
, false);
706 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
707 bool kick_requeue_list
)
709 struct request_queue
*q
= rq
->q
;
713 * We abuse this flag that is otherwise used by the I/O scheduler to
714 * request head insertion from the workqueue.
716 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
718 spin_lock_irqsave(&q
->requeue_lock
, flags
);
720 rq
->rq_flags
|= RQF_SOFTBARRIER
;
721 list_add(&rq
->queuelist
, &q
->requeue_list
);
723 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
725 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
727 if (kick_requeue_list
)
728 blk_mq_kick_requeue_list(q
);
730 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
732 void blk_mq_kick_requeue_list(struct request_queue
*q
)
734 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
736 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
738 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
741 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
742 msecs_to_jiffies(msecs
));
744 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
746 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
748 if (tag
< tags
->nr_tags
) {
749 prefetch(tags
->rqs
[tag
]);
750 return tags
->rqs
[tag
];
755 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
757 struct blk_mq_timeout_data
{
759 unsigned int next_set
;
762 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
764 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
765 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
768 * We know that complete is set at this point. If STARTED isn't set
769 * anymore, then the request isn't active and the "timeout" should
770 * just be ignored. This can happen due to the bitflag ordering.
771 * Timeout first checks if STARTED is set, and if it is, assumes
772 * the request is active. But if we race with completion, then
773 * both flags will get cleared. So check here again, and ignore
774 * a timeout event with a request that isn't active.
776 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
780 ret
= ops
->timeout(req
, reserved
);
784 __blk_mq_complete_request(req
);
786 case BLK_EH_RESET_TIMER
:
788 blk_clear_rq_complete(req
);
790 case BLK_EH_NOT_HANDLED
:
793 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
798 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
799 struct request
*rq
, void *priv
, bool reserved
)
801 struct blk_mq_timeout_data
*data
= priv
;
802 unsigned long deadline
;
804 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
808 * Ensures that if we see STARTED we must also see our
809 * up-to-date deadline, see blk_mq_start_request().
813 deadline
= READ_ONCE(rq
->deadline
);
816 * The rq being checked may have been freed and reallocated
817 * out already here, we avoid this race by checking rq->deadline
818 * and REQ_ATOM_COMPLETE flag together:
820 * - if rq->deadline is observed as new value because of
821 * reusing, the rq won't be timed out because of timing.
822 * - if rq->deadline is observed as previous value,
823 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
824 * because we put a barrier between setting rq->deadline
825 * and clearing the flag in blk_mq_start_request(), so
826 * this rq won't be timed out too.
828 if (time_after_eq(jiffies
, deadline
)) {
829 if (!blk_mark_rq_complete(rq
)) {
831 * Again coherence order ensures that consecutive reads
832 * from the same variable must be in that order. This
833 * ensures that if we see COMPLETE clear, we must then
834 * see STARTED set and we'll ignore this timeout.
836 * (There's also the MB implied by the test_and_clear())
838 blk_mq_rq_timed_out(rq
, reserved
);
840 } else if (!data
->next_set
|| time_after(data
->next
, deadline
)) {
841 data
->next
= deadline
;
846 static void blk_mq_timeout_work(struct work_struct
*work
)
848 struct request_queue
*q
=
849 container_of(work
, struct request_queue
, timeout_work
);
850 struct blk_mq_timeout_data data
= {
856 /* A deadlock might occur if a request is stuck requiring a
857 * timeout at the same time a queue freeze is waiting
858 * completion, since the timeout code would not be able to
859 * acquire the queue reference here.
861 * That's why we don't use blk_queue_enter here; instead, we use
862 * percpu_ref_tryget directly, because we need to be able to
863 * obtain a reference even in the short window between the queue
864 * starting to freeze, by dropping the first reference in
865 * blk_freeze_queue_start, and the moment the last request is
866 * consumed, marked by the instant q_usage_counter reaches
869 if (!percpu_ref_tryget(&q
->q_usage_counter
))
872 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
875 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
876 mod_timer(&q
->timeout
, data
.next
);
878 struct blk_mq_hw_ctx
*hctx
;
880 queue_for_each_hw_ctx(q
, hctx
, i
) {
881 /* the hctx may be unmapped, so check it here */
882 if (blk_mq_hw_queue_mapped(hctx
))
883 blk_mq_tag_idle(hctx
);
889 struct flush_busy_ctx_data
{
890 struct blk_mq_hw_ctx
*hctx
;
891 struct list_head
*list
;
894 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
896 struct flush_busy_ctx_data
*flush_data
= data
;
897 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
898 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
900 sbitmap_clear_bit(sb
, bitnr
);
901 spin_lock(&ctx
->lock
);
902 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
903 spin_unlock(&ctx
->lock
);
908 * Process software queues that have been marked busy, splicing them
909 * to the for-dispatch
911 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
913 struct flush_busy_ctx_data data
= {
918 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
920 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
922 struct dispatch_rq_data
{
923 struct blk_mq_hw_ctx
*hctx
;
927 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
930 struct dispatch_rq_data
*dispatch_data
= data
;
931 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
932 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
934 spin_lock(&ctx
->lock
);
935 if (unlikely(!list_empty(&ctx
->rq_list
))) {
936 dispatch_data
->rq
= list_entry_rq(ctx
->rq_list
.next
);
937 list_del_init(&dispatch_data
->rq
->queuelist
);
938 if (list_empty(&ctx
->rq_list
))
939 sbitmap_clear_bit(sb
, bitnr
);
941 spin_unlock(&ctx
->lock
);
943 return !dispatch_data
->rq
;
946 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
947 struct blk_mq_ctx
*start
)
949 unsigned off
= start
? start
->index_hw
: 0;
950 struct dispatch_rq_data data
= {
955 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
956 dispatch_rq_from_ctx
, &data
);
961 static inline unsigned int queued_to_index(unsigned int queued
)
966 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
969 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
972 struct blk_mq_alloc_data data
= {
974 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
975 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
978 might_sleep_if(wait
);
983 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
984 data
.flags
|= BLK_MQ_REQ_RESERVED
;
986 rq
->tag
= blk_mq_get_tag(&data
);
988 if (blk_mq_tag_busy(data
.hctx
)) {
989 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
990 atomic_inc(&data
.hctx
->nr_active
);
992 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
998 return rq
->tag
!= -1;
1001 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1002 int flags
, void *key
)
1004 struct blk_mq_hw_ctx
*hctx
;
1006 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1008 list_del_init(&wait
->entry
);
1009 blk_mq_run_hw_queue(hctx
, true);
1014 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1015 * the tag wakeups. For non-shared tags, we can simply mark us nedeing a
1016 * restart. For both caes, take care to check the condition again after
1017 * marking us as waiting.
1019 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
**hctx
,
1022 struct blk_mq_hw_ctx
*this_hctx
= *hctx
;
1023 bool shared_tags
= (this_hctx
->flags
& BLK_MQ_F_TAG_SHARED
) != 0;
1024 struct sbq_wait_state
*ws
;
1025 wait_queue_entry_t
*wait
;
1029 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &this_hctx
->state
))
1030 set_bit(BLK_MQ_S_SCHED_RESTART
, &this_hctx
->state
);
1032 wait
= &this_hctx
->dispatch_wait
;
1033 if (!list_empty_careful(&wait
->entry
))
1036 spin_lock(&this_hctx
->lock
);
1037 if (!list_empty(&wait
->entry
)) {
1038 spin_unlock(&this_hctx
->lock
);
1042 ws
= bt_wait_ptr(&this_hctx
->tags
->bitmap_tags
, this_hctx
);
1043 add_wait_queue(&ws
->wait
, wait
);
1047 * It's possible that a tag was freed in the window between the
1048 * allocation failure and adding the hardware queue to the wait
1051 ret
= blk_mq_get_driver_tag(rq
, hctx
, false);
1055 * Don't clear RESTART here, someone else could have set it.
1056 * At most this will cost an extra queue run.
1061 spin_unlock(&this_hctx
->lock
);
1066 * We got a tag, remove ourselves from the wait queue to ensure
1067 * someone else gets the wakeup.
1069 spin_lock_irq(&ws
->wait
.lock
);
1070 list_del_init(&wait
->entry
);
1071 spin_unlock_irq(&ws
->wait
.lock
);
1072 spin_unlock(&this_hctx
->lock
);
1077 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
,
1080 struct blk_mq_hw_ctx
*hctx
;
1081 struct request
*rq
, *nxt
;
1082 bool no_tag
= false;
1085 if (list_empty(list
))
1088 WARN_ON(!list_is_singular(list
) && got_budget
);
1091 * Now process all the entries, sending them to the driver.
1093 errors
= queued
= 0;
1095 struct blk_mq_queue_data bd
;
1098 rq
= list_first_entry(list
, struct request
, queuelist
);
1100 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
1101 if (!got_budget
&& !blk_mq_get_dispatch_budget(hctx
))
1104 if (!blk_mq_get_driver_tag(rq
, NULL
, false)) {
1106 * The initial allocation attempt failed, so we need to
1107 * rerun the hardware queue when a tag is freed. The
1108 * waitqueue takes care of that. If the queue is run
1109 * before we add this entry back on the dispatch list,
1110 * we'll re-run it below.
1112 if (!blk_mq_mark_tag_wait(&hctx
, rq
)) {
1113 blk_mq_put_dispatch_budget(hctx
);
1115 * For non-shared tags, the RESTART check
1118 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
)
1124 list_del_init(&rq
->queuelist
);
1129 * Flag last if we have no more requests, or if we have more
1130 * but can't assign a driver tag to it.
1132 if (list_empty(list
))
1135 nxt
= list_first_entry(list
, struct request
, queuelist
);
1136 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1139 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1140 if (ret
== BLK_STS_RESOURCE
) {
1142 * If an I/O scheduler has been configured and we got a
1143 * driver tag for the next request already, free it
1146 if (!list_empty(list
)) {
1147 nxt
= list_first_entry(list
, struct request
, queuelist
);
1148 blk_mq_put_driver_tag(nxt
);
1150 list_add(&rq
->queuelist
, list
);
1151 __blk_mq_requeue_request(rq
);
1155 if (unlikely(ret
!= BLK_STS_OK
)) {
1157 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1162 } while (!list_empty(list
));
1164 hctx
->dispatched
[queued_to_index(queued
)]++;
1167 * Any items that need requeuing? Stuff them into hctx->dispatch,
1168 * that is where we will continue on next queue run.
1170 if (!list_empty(list
)) {
1171 spin_lock(&hctx
->lock
);
1172 list_splice_init(list
, &hctx
->dispatch
);
1173 spin_unlock(&hctx
->lock
);
1176 * If SCHED_RESTART was set by the caller of this function and
1177 * it is no longer set that means that it was cleared by another
1178 * thread and hence that a queue rerun is needed.
1180 * If 'no_tag' is set, that means that we failed getting
1181 * a driver tag with an I/O scheduler attached. If our dispatch
1182 * waitqueue is no longer active, ensure that we run the queue
1183 * AFTER adding our entries back to the list.
1185 * If no I/O scheduler has been configured it is possible that
1186 * the hardware queue got stopped and restarted before requests
1187 * were pushed back onto the dispatch list. Rerun the queue to
1188 * avoid starvation. Notes:
1189 * - blk_mq_run_hw_queue() checks whether or not a queue has
1190 * been stopped before rerunning a queue.
1191 * - Some but not all block drivers stop a queue before
1192 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1195 if (!blk_mq_sched_needs_restart(hctx
) ||
1196 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1197 blk_mq_run_hw_queue(hctx
, true);
1200 return (queued
+ errors
) != 0;
1203 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1208 * We should be running this queue from one of the CPUs that
1211 * There are at least two related races now between setting
1212 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1213 * __blk_mq_run_hw_queue():
1215 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1216 * but later it becomes online, then this warning is harmless
1219 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1220 * but later it becomes offline, then the warning can't be
1221 * triggered, and we depend on blk-mq timeout handler to
1222 * handle dispatched requests to this hctx
1224 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1225 cpu_online(hctx
->next_cpu
)) {
1226 printk(KERN_WARNING
"run queue from wrong CPU %d, hctx %s\n",
1227 raw_smp_processor_id(),
1228 cpumask_empty(hctx
->cpumask
) ? "inactive": "active");
1233 * We can't run the queue inline with ints disabled. Ensure that
1234 * we catch bad users of this early.
1236 WARN_ON_ONCE(in_interrupt());
1238 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1240 blk_mq_sched_dispatch_requests(hctx
);
1245 srcu_idx
= srcu_read_lock(hctx
->queue_rq_srcu
);
1246 blk_mq_sched_dispatch_requests(hctx
);
1247 srcu_read_unlock(hctx
->queue_rq_srcu
, srcu_idx
);
1251 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1253 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1255 if (cpu
>= nr_cpu_ids
)
1256 cpu
= cpumask_first(hctx
->cpumask
);
1261 * It'd be great if the workqueue API had a way to pass
1262 * in a mask and had some smarts for more clever placement.
1263 * For now we just round-robin here, switching for every
1264 * BLK_MQ_CPU_WORK_BATCH queued items.
1266 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1269 int next_cpu
= hctx
->next_cpu
;
1271 if (hctx
->queue
->nr_hw_queues
== 1)
1272 return WORK_CPU_UNBOUND
;
1274 if (--hctx
->next_cpu_batch
<= 0) {
1276 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1278 if (next_cpu
>= nr_cpu_ids
)
1279 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1280 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1284 * Do unbound schedule if we can't find a online CPU for this hctx,
1285 * and it should only happen in the path of handling CPU DEAD.
1287 if (!cpu_online(next_cpu
)) {
1294 * Make sure to re-select CPU next time once after CPUs
1295 * in hctx->cpumask become online again.
1297 hctx
->next_cpu
= next_cpu
;
1298 hctx
->next_cpu_batch
= 1;
1299 return WORK_CPU_UNBOUND
;
1302 hctx
->next_cpu
= next_cpu
;
1306 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1307 unsigned long msecs
)
1309 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1312 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1313 int cpu
= get_cpu();
1314 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1315 __blk_mq_run_hw_queue(hctx
);
1323 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1324 msecs_to_jiffies(msecs
));
1327 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1329 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1331 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1333 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1335 if (blk_mq_hctx_has_pending(hctx
)) {
1336 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1342 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1344 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1346 struct blk_mq_hw_ctx
*hctx
;
1349 queue_for_each_hw_ctx(q
, hctx
, i
) {
1350 if (blk_mq_hctx_stopped(hctx
))
1353 blk_mq_run_hw_queue(hctx
, async
);
1356 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1359 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1360 * @q: request queue.
1362 * The caller is responsible for serializing this function against
1363 * blk_mq_{start,stop}_hw_queue().
1365 bool blk_mq_queue_stopped(struct request_queue
*q
)
1367 struct blk_mq_hw_ctx
*hctx
;
1370 queue_for_each_hw_ctx(q
, hctx
, i
)
1371 if (blk_mq_hctx_stopped(hctx
))
1376 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1379 * This function is often used for pausing .queue_rq() by driver when
1380 * there isn't enough resource or some conditions aren't satisfied, and
1381 * BLK_STS_RESOURCE is usually returned.
1383 * We do not guarantee that dispatch can be drained or blocked
1384 * after blk_mq_stop_hw_queue() returns. Please use
1385 * blk_mq_quiesce_queue() for that requirement.
1387 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1389 cancel_delayed_work(&hctx
->run_work
);
1391 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1393 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1396 * This function is often used for pausing .queue_rq() by driver when
1397 * there isn't enough resource or some conditions aren't satisfied, and
1398 * BLK_STS_RESOURCE is usually returned.
1400 * We do not guarantee that dispatch can be drained or blocked
1401 * after blk_mq_stop_hw_queues() returns. Please use
1402 * blk_mq_quiesce_queue() for that requirement.
1404 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1406 struct blk_mq_hw_ctx
*hctx
;
1409 queue_for_each_hw_ctx(q
, hctx
, i
)
1410 blk_mq_stop_hw_queue(hctx
);
1412 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1414 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1416 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1418 blk_mq_run_hw_queue(hctx
, false);
1420 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1422 void blk_mq_start_hw_queues(struct request_queue
*q
)
1424 struct blk_mq_hw_ctx
*hctx
;
1427 queue_for_each_hw_ctx(q
, hctx
, i
)
1428 blk_mq_start_hw_queue(hctx
);
1430 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1432 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1434 if (!blk_mq_hctx_stopped(hctx
))
1437 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1438 blk_mq_run_hw_queue(hctx
, async
);
1440 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1442 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1444 struct blk_mq_hw_ctx
*hctx
;
1447 queue_for_each_hw_ctx(q
, hctx
, i
)
1448 blk_mq_start_stopped_hw_queue(hctx
, async
);
1450 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1452 static void blk_mq_run_work_fn(struct work_struct
*work
)
1454 struct blk_mq_hw_ctx
*hctx
;
1456 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1459 * If we are stopped, don't run the queue. The exception is if
1460 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1461 * the STOPPED bit and run it.
1463 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)) {
1464 if (!test_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
))
1467 clear_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1468 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1471 __blk_mq_run_hw_queue(hctx
);
1475 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1477 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx
)))
1481 * Stop the hw queue, then modify currently delayed work.
1482 * This should prevent us from running the queue prematurely.
1483 * Mark the queue as auto-clearing STOPPED when it runs.
1485 blk_mq_stop_hw_queue(hctx
);
1486 set_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1487 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1489 msecs_to_jiffies(msecs
));
1491 EXPORT_SYMBOL(blk_mq_delay_queue
);
1493 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1497 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1499 lockdep_assert_held(&ctx
->lock
);
1501 trace_block_rq_insert(hctx
->queue
, rq
);
1504 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1506 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1509 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1512 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1514 lockdep_assert_held(&ctx
->lock
);
1516 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1517 blk_mq_hctx_mark_pending(hctx
, ctx
);
1521 * Should only be used carefully, when the caller knows we want to
1522 * bypass a potential IO scheduler on the target device.
1524 void blk_mq_request_bypass_insert(struct request
*rq
, bool run_queue
)
1526 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1527 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(rq
->q
, ctx
->cpu
);
1529 spin_lock(&hctx
->lock
);
1530 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1531 spin_unlock(&hctx
->lock
);
1534 blk_mq_run_hw_queue(hctx
, false);
1537 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1538 struct list_head
*list
)
1542 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1545 spin_lock(&ctx
->lock
);
1546 while (!list_empty(list
)) {
1549 rq
= list_first_entry(list
, struct request
, queuelist
);
1550 BUG_ON(rq
->mq_ctx
!= ctx
);
1551 list_del_init(&rq
->queuelist
);
1552 __blk_mq_insert_req_list(hctx
, rq
, false);
1554 blk_mq_hctx_mark_pending(hctx
, ctx
);
1555 spin_unlock(&ctx
->lock
);
1558 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1560 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1561 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1563 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1564 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1565 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1568 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1570 struct blk_mq_ctx
*this_ctx
;
1571 struct request_queue
*this_q
;
1574 LIST_HEAD(ctx_list
);
1577 list_splice_init(&plug
->mq_list
, &list
);
1579 list_sort(NULL
, &list
, plug_ctx_cmp
);
1585 while (!list_empty(&list
)) {
1586 rq
= list_entry_rq(list
.next
);
1587 list_del_init(&rq
->queuelist
);
1589 if (rq
->mq_ctx
!= this_ctx
) {
1591 trace_block_unplug(this_q
, depth
, from_schedule
);
1592 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1597 this_ctx
= rq
->mq_ctx
;
1603 list_add_tail(&rq
->queuelist
, &ctx_list
);
1607 * If 'this_ctx' is set, we know we have entries to complete
1608 * on 'ctx_list'. Do those.
1611 trace_block_unplug(this_q
, depth
, from_schedule
);
1612 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1617 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1619 blk_init_request_from_bio(rq
, bio
);
1621 blk_rq_set_rl(rq
, blk_get_rl(rq
->q
, bio
));
1623 blk_account_io_start(rq
, true);
1626 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx
*hctx
,
1627 struct blk_mq_ctx
*ctx
,
1630 spin_lock(&ctx
->lock
);
1631 __blk_mq_insert_request(hctx
, rq
, false);
1632 spin_unlock(&ctx
->lock
);
1635 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1638 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1640 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1643 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1645 blk_qc_t
*cookie
, bool may_sleep
)
1647 struct request_queue
*q
= rq
->q
;
1648 struct blk_mq_queue_data bd
= {
1652 blk_qc_t new_cookie
;
1654 bool run_queue
= true;
1656 /* RCU or SRCU read lock is needed before checking quiesced flag */
1657 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1665 if (!blk_mq_get_dispatch_budget(hctx
))
1668 if (!blk_mq_get_driver_tag(rq
, NULL
, false)) {
1669 blk_mq_put_dispatch_budget(hctx
);
1673 new_cookie
= request_to_qc_t(hctx
, rq
);
1676 * For OK queue, we are done. For error, kill it. Any other
1677 * error (busy), just add it to our list as we previously
1680 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1683 *cookie
= new_cookie
;
1685 case BLK_STS_RESOURCE
:
1686 __blk_mq_requeue_request(rq
);
1689 *cookie
= BLK_QC_T_NONE
;
1690 blk_mq_end_request(rq
, ret
);
1695 blk_mq_sched_insert_request(rq
, false, run_queue
, false, may_sleep
);
1698 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1699 struct request
*rq
, blk_qc_t
*cookie
)
1701 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1703 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false);
1706 unsigned int srcu_idx
;
1710 srcu_idx
= srcu_read_lock(hctx
->queue_rq_srcu
);
1711 __blk_mq_try_issue_directly(hctx
, rq
, cookie
, true);
1712 srcu_read_unlock(hctx
->queue_rq_srcu
, srcu_idx
);
1716 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1718 const int is_sync
= op_is_sync(bio
->bi_opf
);
1719 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1720 struct blk_mq_alloc_data data
= { .flags
= 0 };
1722 unsigned int request_count
= 0;
1723 struct blk_plug
*plug
;
1724 struct request
*same_queue_rq
= NULL
;
1726 unsigned int wb_acct
;
1728 blk_queue_bounce(q
, &bio
);
1730 blk_queue_split(q
, &bio
);
1732 if (!bio_integrity_prep(bio
))
1733 return BLK_QC_T_NONE
;
1735 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1736 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1737 return BLK_QC_T_NONE
;
1739 if (blk_mq_sched_bio_merge(q
, bio
))
1740 return BLK_QC_T_NONE
;
1742 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1744 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1746 rq
= blk_mq_get_request(q
, bio
, bio
->bi_opf
, &data
);
1747 if (unlikely(!rq
)) {
1748 __wbt_done(q
->rq_wb
, wb_acct
);
1749 if (bio
->bi_opf
& REQ_NOWAIT
)
1750 bio_wouldblock_error(bio
);
1751 return BLK_QC_T_NONE
;
1754 wbt_track(&rq
->issue_stat
, wb_acct
);
1756 cookie
= request_to_qc_t(data
.hctx
, rq
);
1758 plug
= current
->plug
;
1759 if (unlikely(is_flush_fua
)) {
1760 blk_mq_put_ctx(data
.ctx
);
1761 blk_mq_bio_to_request(rq
, bio
);
1763 /* bypass scheduler for flush rq */
1764 blk_insert_flush(rq
);
1765 blk_mq_run_hw_queue(data
.hctx
, true);
1766 } else if (plug
&& q
->nr_hw_queues
== 1) {
1767 struct request
*last
= NULL
;
1769 blk_mq_put_ctx(data
.ctx
);
1770 blk_mq_bio_to_request(rq
, bio
);
1773 * @request_count may become stale because of schedule
1774 * out, so check the list again.
1776 if (list_empty(&plug
->mq_list
))
1778 else if (blk_queue_nomerges(q
))
1779 request_count
= blk_plug_queued_count(q
);
1782 trace_block_plug(q
);
1784 last
= list_entry_rq(plug
->mq_list
.prev
);
1786 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1787 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1788 blk_flush_plug_list(plug
, false);
1789 trace_block_plug(q
);
1792 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1793 } else if (plug
&& !blk_queue_nomerges(q
)) {
1794 blk_mq_bio_to_request(rq
, bio
);
1797 * We do limited plugging. If the bio can be merged, do that.
1798 * Otherwise the existing request in the plug list will be
1799 * issued. So the plug list will have one request at most
1800 * The plug list might get flushed before this. If that happens,
1801 * the plug list is empty, and same_queue_rq is invalid.
1803 if (list_empty(&plug
->mq_list
))
1804 same_queue_rq
= NULL
;
1806 list_del_init(&same_queue_rq
->queuelist
);
1807 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1809 blk_mq_put_ctx(data
.ctx
);
1811 if (same_queue_rq
) {
1812 data
.hctx
= blk_mq_map_queue(q
,
1813 same_queue_rq
->mq_ctx
->cpu
);
1814 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1817 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1818 blk_mq_put_ctx(data
.ctx
);
1819 blk_mq_bio_to_request(rq
, bio
);
1820 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1821 } else if (q
->elevator
) {
1822 blk_mq_put_ctx(data
.ctx
);
1823 blk_mq_bio_to_request(rq
, bio
);
1824 blk_mq_sched_insert_request(rq
, false, true, true, true);
1826 blk_mq_put_ctx(data
.ctx
);
1827 blk_mq_bio_to_request(rq
, bio
);
1828 blk_mq_queue_io(data
.hctx
, data
.ctx
, rq
);
1829 blk_mq_run_hw_queue(data
.hctx
, true);
1835 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1836 unsigned int hctx_idx
)
1840 if (tags
->rqs
&& set
->ops
->exit_request
) {
1843 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1844 struct request
*rq
= tags
->static_rqs
[i
];
1848 set
->ops
->exit_request(set
, rq
, hctx_idx
);
1849 tags
->static_rqs
[i
] = NULL
;
1853 while (!list_empty(&tags
->page_list
)) {
1854 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1855 list_del_init(&page
->lru
);
1857 * Remove kmemleak object previously allocated in
1858 * blk_mq_init_rq_map().
1860 kmemleak_free(page_address(page
));
1861 __free_pages(page
, page
->private);
1865 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1869 kfree(tags
->static_rqs
);
1870 tags
->static_rqs
= NULL
;
1872 blk_mq_free_tags(tags
);
1875 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1876 unsigned int hctx_idx
,
1877 unsigned int nr_tags
,
1878 unsigned int reserved_tags
)
1880 struct blk_mq_tags
*tags
;
1883 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1884 if (node
== NUMA_NO_NODE
)
1885 node
= set
->numa_node
;
1887 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1888 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1892 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1893 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1896 blk_mq_free_tags(tags
);
1900 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1901 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1903 if (!tags
->static_rqs
) {
1905 blk_mq_free_tags(tags
);
1912 static size_t order_to_size(unsigned int order
)
1914 return (size_t)PAGE_SIZE
<< order
;
1917 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1918 unsigned int hctx_idx
, unsigned int depth
)
1920 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1921 size_t rq_size
, left
;
1924 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1925 if (node
== NUMA_NO_NODE
)
1926 node
= set
->numa_node
;
1928 INIT_LIST_HEAD(&tags
->page_list
);
1931 * rq_size is the size of the request plus driver payload, rounded
1932 * to the cacheline size
1934 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
1936 left
= rq_size
* depth
;
1938 for (i
= 0; i
< depth
; ) {
1939 int this_order
= max_order
;
1944 while (this_order
&& left
< order_to_size(this_order
- 1))
1948 page
= alloc_pages_node(node
,
1949 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
1955 if (order_to_size(this_order
) < rq_size
)
1962 page
->private = this_order
;
1963 list_add_tail(&page
->lru
, &tags
->page_list
);
1965 p
= page_address(page
);
1967 * Allow kmemleak to scan these pages as they contain pointers
1968 * to additional allocations like via ops->init_request().
1970 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
1971 entries_per_page
= order_to_size(this_order
) / rq_size
;
1972 to_do
= min(entries_per_page
, depth
- i
);
1973 left
-= to_do
* rq_size
;
1974 for (j
= 0; j
< to_do
; j
++) {
1975 struct request
*rq
= p
;
1977 tags
->static_rqs
[i
] = rq
;
1978 if (set
->ops
->init_request
) {
1979 if (set
->ops
->init_request(set
, rq
, hctx_idx
,
1981 tags
->static_rqs
[i
] = NULL
;
1993 blk_mq_free_rqs(set
, tags
, hctx_idx
);
1998 * 'cpu' is going away. splice any existing rq_list entries from this
1999 * software queue to the hw queue dispatch list, and ensure that it
2002 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2004 struct blk_mq_hw_ctx
*hctx
;
2005 struct blk_mq_ctx
*ctx
;
2008 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2009 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2011 spin_lock(&ctx
->lock
);
2012 if (!list_empty(&ctx
->rq_list
)) {
2013 list_splice_init(&ctx
->rq_list
, &tmp
);
2014 blk_mq_hctx_clear_pending(hctx
, ctx
);
2016 spin_unlock(&ctx
->lock
);
2018 if (list_empty(&tmp
))
2021 spin_lock(&hctx
->lock
);
2022 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2023 spin_unlock(&hctx
->lock
);
2025 blk_mq_run_hw_queue(hctx
, true);
2029 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2031 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2035 /* hctx->ctxs will be freed in queue's release handler */
2036 static void blk_mq_exit_hctx(struct request_queue
*q
,
2037 struct blk_mq_tag_set
*set
,
2038 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2040 blk_mq_debugfs_unregister_hctx(hctx
);
2042 if (blk_mq_hw_queue_mapped(hctx
))
2043 blk_mq_tag_idle(hctx
);
2045 if (set
->ops
->exit_request
)
2046 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2048 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2050 if (set
->ops
->exit_hctx
)
2051 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2053 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2054 cleanup_srcu_struct(hctx
->queue_rq_srcu
);
2056 blk_mq_remove_cpuhp(hctx
);
2057 blk_free_flush_queue(hctx
->fq
);
2058 sbitmap_free(&hctx
->ctx_map
);
2061 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2062 struct blk_mq_tag_set
*set
, int nr_queue
)
2064 struct blk_mq_hw_ctx
*hctx
;
2067 queue_for_each_hw_ctx(q
, hctx
, i
) {
2070 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2074 static int blk_mq_init_hctx(struct request_queue
*q
,
2075 struct blk_mq_tag_set
*set
,
2076 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2080 node
= hctx
->numa_node
;
2081 if (node
== NUMA_NO_NODE
)
2082 node
= hctx
->numa_node
= set
->numa_node
;
2084 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2085 spin_lock_init(&hctx
->lock
);
2086 INIT_LIST_HEAD(&hctx
->dispatch
);
2088 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
2090 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2092 hctx
->tags
= set
->tags
[hctx_idx
];
2095 * Allocate space for all possible cpus to avoid allocation at
2098 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2101 goto unregister_cpu_notifier
;
2103 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
2109 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2110 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2112 if (set
->ops
->init_hctx
&&
2113 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2116 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
2119 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
2121 goto sched_exit_hctx
;
2123 if (set
->ops
->init_request
&&
2124 set
->ops
->init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
2128 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2129 init_srcu_struct(hctx
->queue_rq_srcu
);
2131 blk_mq_debugfs_register_hctx(q
, hctx
);
2138 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2140 if (set
->ops
->exit_hctx
)
2141 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2143 sbitmap_free(&hctx
->ctx_map
);
2146 unregister_cpu_notifier
:
2147 blk_mq_remove_cpuhp(hctx
);
2151 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2152 unsigned int nr_hw_queues
)
2156 for_each_possible_cpu(i
) {
2157 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2158 struct blk_mq_hw_ctx
*hctx
;
2161 spin_lock_init(&__ctx
->lock
);
2162 INIT_LIST_HEAD(&__ctx
->rq_list
);
2166 * Set local node, IFF we have more than one hw queue. If
2167 * not, we remain on the home node of the device
2169 hctx
= blk_mq_map_queue(q
, i
);
2170 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2171 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2175 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2179 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2180 set
->queue_depth
, set
->reserved_tags
);
2181 if (!set
->tags
[hctx_idx
])
2184 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2189 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2190 set
->tags
[hctx_idx
] = NULL
;
2194 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2195 unsigned int hctx_idx
)
2197 if (set
->tags
[hctx_idx
]) {
2198 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2199 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2200 set
->tags
[hctx_idx
] = NULL
;
2204 static void blk_mq_map_swqueue(struct request_queue
*q
)
2206 unsigned int i
, hctx_idx
;
2207 struct blk_mq_hw_ctx
*hctx
;
2208 struct blk_mq_ctx
*ctx
;
2209 struct blk_mq_tag_set
*set
= q
->tag_set
;
2212 * Avoid others reading imcomplete hctx->cpumask through sysfs
2214 mutex_lock(&q
->sysfs_lock
);
2216 queue_for_each_hw_ctx(q
, hctx
, i
) {
2217 cpumask_clear(hctx
->cpumask
);
2222 * Map software to hardware queues.
2224 * If the cpu isn't present, the cpu is mapped to first hctx.
2226 for_each_possible_cpu(i
) {
2227 hctx_idx
= q
->mq_map
[i
];
2228 /* unmapped hw queue can be remapped after CPU topo changed */
2229 if (!set
->tags
[hctx_idx
] &&
2230 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2232 * If tags initialization fail for some hctx,
2233 * that hctx won't be brought online. In this
2234 * case, remap the current ctx to hctx[0] which
2235 * is guaranteed to always have tags allocated
2240 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2241 hctx
= blk_mq_map_queue(q
, i
);
2243 cpumask_set_cpu(i
, hctx
->cpumask
);
2244 ctx
->index_hw
= hctx
->nr_ctx
;
2245 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2248 mutex_unlock(&q
->sysfs_lock
);
2250 queue_for_each_hw_ctx(q
, hctx
, i
) {
2252 * If no software queues are mapped to this hardware queue,
2253 * disable it and free the request entries.
2255 if (!hctx
->nr_ctx
) {
2256 /* Never unmap queue 0. We need it as a
2257 * fallback in case of a new remap fails
2260 if (i
&& set
->tags
[i
])
2261 blk_mq_free_map_and_requests(set
, i
);
2267 hctx
->tags
= set
->tags
[i
];
2268 WARN_ON(!hctx
->tags
);
2271 * Set the map size to the number of mapped software queues.
2272 * This is more accurate and more efficient than looping
2273 * over all possibly mapped software queues.
2275 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2278 * Initialize batch roundrobin counts
2280 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2281 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2286 * Caller needs to ensure that we're either frozen/quiesced, or that
2287 * the queue isn't live yet.
2289 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2291 struct blk_mq_hw_ctx
*hctx
;
2294 queue_for_each_hw_ctx(q
, hctx
, i
) {
2296 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2297 atomic_inc(&q
->shared_hctx_restart
);
2298 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2300 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2301 atomic_dec(&q
->shared_hctx_restart
);
2302 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2307 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2310 struct request_queue
*q
;
2312 lockdep_assert_held(&set
->tag_list_lock
);
2314 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2315 blk_mq_freeze_queue(q
);
2316 queue_set_hctx_shared(q
, shared
);
2317 blk_mq_unfreeze_queue(q
);
2321 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2323 struct blk_mq_tag_set
*set
= q
->tag_set
;
2325 mutex_lock(&set
->tag_list_lock
);
2326 list_del_rcu(&q
->tag_set_list
);
2327 INIT_LIST_HEAD(&q
->tag_set_list
);
2328 if (list_is_singular(&set
->tag_list
)) {
2329 /* just transitioned to unshared */
2330 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2331 /* update existing queue */
2332 blk_mq_update_tag_set_depth(set
, false);
2334 mutex_unlock(&set
->tag_list_lock
);
2339 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2340 struct request_queue
*q
)
2344 mutex_lock(&set
->tag_list_lock
);
2347 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2349 if (!list_empty(&set
->tag_list
) &&
2350 !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2351 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2352 /* update existing queue */
2353 blk_mq_update_tag_set_depth(set
, true);
2355 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2356 queue_set_hctx_shared(q
, true);
2357 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2359 mutex_unlock(&set
->tag_list_lock
);
2363 * It is the actual release handler for mq, but we do it from
2364 * request queue's release handler for avoiding use-after-free
2365 * and headache because q->mq_kobj shouldn't have been introduced,
2366 * but we can't group ctx/kctx kobj without it.
2368 void blk_mq_release(struct request_queue
*q
)
2370 struct blk_mq_hw_ctx
*hctx
;
2373 /* hctx kobj stays in hctx */
2374 queue_for_each_hw_ctx(q
, hctx
, i
) {
2377 kobject_put(&hctx
->kobj
);
2382 kfree(q
->queue_hw_ctx
);
2385 * release .mq_kobj and sw queue's kobject now because
2386 * both share lifetime with request queue.
2388 blk_mq_sysfs_deinit(q
);
2390 free_percpu(q
->queue_ctx
);
2393 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2395 struct request_queue
*uninit_q
, *q
;
2397 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2399 return ERR_PTR(-ENOMEM
);
2401 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2403 blk_cleanup_queue(uninit_q
);
2407 EXPORT_SYMBOL(blk_mq_init_queue
);
2409 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2411 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2413 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, queue_rq_srcu
),
2414 __alignof__(struct blk_mq_hw_ctx
)) !=
2415 sizeof(struct blk_mq_hw_ctx
));
2417 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2418 hw_ctx_size
+= sizeof(struct srcu_struct
);
2423 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2424 struct request_queue
*q
)
2427 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2429 blk_mq_sysfs_unregister(q
);
2431 /* protect against switching io scheduler */
2432 mutex_lock(&q
->sysfs_lock
);
2433 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2439 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2440 hctxs
[i
] = kzalloc_node(blk_mq_hw_ctx_size(set
),
2445 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2452 atomic_set(&hctxs
[i
]->nr_active
, 0);
2453 hctxs
[i
]->numa_node
= node
;
2454 hctxs
[i
]->queue_num
= i
;
2456 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2457 free_cpumask_var(hctxs
[i
]->cpumask
);
2462 blk_mq_hctx_kobj_init(hctxs
[i
]);
2464 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2465 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2469 blk_mq_free_map_and_requests(set
, j
);
2470 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2471 kobject_put(&hctx
->kobj
);
2476 q
->nr_hw_queues
= i
;
2477 mutex_unlock(&q
->sysfs_lock
);
2478 blk_mq_sysfs_register(q
);
2481 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2482 struct request_queue
*q
)
2484 /* mark the queue as mq asap */
2485 q
->mq_ops
= set
->ops
;
2487 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2488 blk_mq_poll_stats_bkt
,
2489 BLK_MQ_POLL_STATS_BKTS
, q
);
2493 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2497 /* init q->mq_kobj and sw queues' kobjects */
2498 blk_mq_sysfs_init(q
);
2500 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2501 GFP_KERNEL
, set
->numa_node
);
2502 if (!q
->queue_hw_ctx
)
2505 q
->mq_map
= set
->mq_map
;
2507 blk_mq_realloc_hw_ctxs(set
, q
);
2508 if (!q
->nr_hw_queues
)
2511 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2512 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2514 q
->nr_queues
= nr_cpu_ids
;
2516 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2518 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2519 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2521 q
->sg_reserved_size
= INT_MAX
;
2523 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2524 INIT_LIST_HEAD(&q
->requeue_list
);
2525 spin_lock_init(&q
->requeue_lock
);
2527 blk_queue_make_request(q
, blk_mq_make_request
);
2528 if (q
->mq_ops
->poll
)
2529 q
->poll_fn
= blk_mq_poll
;
2532 * Do this after blk_queue_make_request() overrides it...
2534 q
->nr_requests
= set
->queue_depth
;
2537 * Default to classic polling
2541 if (set
->ops
->complete
)
2542 blk_queue_softirq_done(q
, set
->ops
->complete
);
2544 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2545 blk_mq_add_queue_tag_set(set
, q
);
2546 blk_mq_map_swqueue(q
);
2548 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2551 ret
= blk_mq_sched_init(q
);
2553 return ERR_PTR(ret
);
2559 kfree(q
->queue_hw_ctx
);
2561 free_percpu(q
->queue_ctx
);
2564 return ERR_PTR(-ENOMEM
);
2566 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2568 void blk_mq_free_queue(struct request_queue
*q
)
2570 struct blk_mq_tag_set
*set
= q
->tag_set
;
2572 blk_mq_del_queue_tag_set(q
);
2573 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2576 /* Basically redo blk_mq_init_queue with queue frozen */
2577 static void blk_mq_queue_reinit(struct request_queue
*q
)
2579 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2581 blk_mq_debugfs_unregister_hctxs(q
);
2582 blk_mq_sysfs_unregister(q
);
2585 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2586 * we should change hctx numa_node according to the new topology (this
2587 * involves freeing and re-allocating memory, worth doing?)
2589 blk_mq_map_swqueue(q
);
2591 blk_mq_sysfs_register(q
);
2592 blk_mq_debugfs_register_hctxs(q
);
2595 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2599 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2600 if (!__blk_mq_alloc_rq_map(set
, i
))
2607 blk_mq_free_rq_map(set
->tags
[i
]);
2613 * Allocate the request maps associated with this tag_set. Note that this
2614 * may reduce the depth asked for, if memory is tight. set->queue_depth
2615 * will be updated to reflect the allocated depth.
2617 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2622 depth
= set
->queue_depth
;
2624 err
= __blk_mq_alloc_rq_maps(set
);
2628 set
->queue_depth
>>= 1;
2629 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2633 } while (set
->queue_depth
);
2635 if (!set
->queue_depth
|| err
) {
2636 pr_err("blk-mq: failed to allocate request map\n");
2640 if (depth
!= set
->queue_depth
)
2641 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2642 depth
, set
->queue_depth
);
2647 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2649 if (set
->ops
->map_queues
) {
2652 * transport .map_queues is usually done in the following
2655 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2656 * mask = get_cpu_mask(queue)
2657 * for_each_cpu(cpu, mask)
2658 * set->mq_map[cpu] = queue;
2661 * When we need to remap, the table has to be cleared for
2662 * killing stale mapping since one CPU may not be mapped
2665 for_each_possible_cpu(cpu
)
2666 set
->mq_map
[cpu
] = 0;
2668 return set
->ops
->map_queues(set
);
2670 return blk_mq_map_queues(set
);
2674 * Alloc a tag set to be associated with one or more request queues.
2675 * May fail with EINVAL for various error conditions. May adjust the
2676 * requested depth down, if if it too large. In that case, the set
2677 * value will be stored in set->queue_depth.
2679 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2683 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2685 if (!set
->nr_hw_queues
)
2687 if (!set
->queue_depth
)
2689 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2692 if (!set
->ops
->queue_rq
)
2695 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
2698 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2699 pr_info("blk-mq: reduced tag depth to %u\n",
2701 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2705 * If a crashdump is active, then we are potentially in a very
2706 * memory constrained environment. Limit us to 1 queue and
2707 * 64 tags to prevent using too much memory.
2709 if (is_kdump_kernel()) {
2710 set
->nr_hw_queues
= 1;
2711 set
->queue_depth
= min(64U, set
->queue_depth
);
2714 * There is no use for more h/w queues than cpus.
2716 if (set
->nr_hw_queues
> nr_cpu_ids
)
2717 set
->nr_hw_queues
= nr_cpu_ids
;
2719 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2720 GFP_KERNEL
, set
->numa_node
);
2725 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2726 GFP_KERNEL
, set
->numa_node
);
2730 ret
= blk_mq_update_queue_map(set
);
2732 goto out_free_mq_map
;
2734 ret
= blk_mq_alloc_rq_maps(set
);
2736 goto out_free_mq_map
;
2738 mutex_init(&set
->tag_list_lock
);
2739 INIT_LIST_HEAD(&set
->tag_list
);
2751 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2753 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2757 for (i
= 0; i
< nr_cpu_ids
; i
++)
2758 blk_mq_free_map_and_requests(set
, i
);
2766 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2768 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2770 struct blk_mq_tag_set
*set
= q
->tag_set
;
2771 struct blk_mq_hw_ctx
*hctx
;
2777 blk_mq_freeze_queue(q
);
2780 queue_for_each_hw_ctx(q
, hctx
, i
) {
2784 * If we're using an MQ scheduler, just update the scheduler
2785 * queue depth. This is similar to what the old code would do.
2787 if (!hctx
->sched_tags
) {
2788 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
2791 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2799 q
->nr_requests
= nr
;
2801 blk_mq_unfreeze_queue(q
);
2806 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
2809 struct request_queue
*q
;
2811 lockdep_assert_held(&set
->tag_list_lock
);
2813 if (nr_hw_queues
> nr_cpu_ids
)
2814 nr_hw_queues
= nr_cpu_ids
;
2815 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2818 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2819 blk_mq_freeze_queue(q
);
2821 set
->nr_hw_queues
= nr_hw_queues
;
2822 blk_mq_update_queue_map(set
);
2823 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2824 blk_mq_realloc_hw_ctxs(set
, q
);
2825 blk_mq_queue_reinit(q
);
2828 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2829 blk_mq_unfreeze_queue(q
);
2832 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2834 mutex_lock(&set
->tag_list_lock
);
2835 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
2836 mutex_unlock(&set
->tag_list_lock
);
2838 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2840 /* Enable polling stats and return whether they were already enabled. */
2841 static bool blk_poll_stats_enable(struct request_queue
*q
)
2843 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2844 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
2846 blk_stat_add_callback(q
, q
->poll_cb
);
2850 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2853 * We don't arm the callback if polling stats are not enabled or the
2854 * callback is already active.
2856 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2857 blk_stat_is_active(q
->poll_cb
))
2860 blk_stat_activate_msecs(q
->poll_cb
, 100);
2863 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2865 struct request_queue
*q
= cb
->data
;
2868 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
2869 if (cb
->stat
[bucket
].nr_samples
)
2870 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
2874 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2875 struct blk_mq_hw_ctx
*hctx
,
2878 unsigned long ret
= 0;
2882 * If stats collection isn't on, don't sleep but turn it on for
2885 if (!blk_poll_stats_enable(q
))
2889 * As an optimistic guess, use half of the mean service time
2890 * for this type of request. We can (and should) make this smarter.
2891 * For instance, if the completion latencies are tight, we can
2892 * get closer than just half the mean. This is especially
2893 * important on devices where the completion latencies are longer
2894 * than ~10 usec. We do use the stats for the relevant IO size
2895 * if available which does lead to better estimates.
2897 bucket
= blk_mq_poll_stats_bkt(rq
);
2901 if (q
->poll_stat
[bucket
].nr_samples
)
2902 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
2907 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2908 struct blk_mq_hw_ctx
*hctx
,
2911 struct hrtimer_sleeper hs
;
2912 enum hrtimer_mode mode
;
2916 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
2922 * -1: don't ever hybrid sleep
2923 * 0: use half of prev avg
2924 * >0: use this specific value
2926 if (q
->poll_nsec
== -1)
2928 else if (q
->poll_nsec
> 0)
2929 nsecs
= q
->poll_nsec
;
2931 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2936 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
2939 * This will be replaced with the stats tracking code, using
2940 * 'avg_completion_time / 2' as the pre-sleep target.
2944 mode
= HRTIMER_MODE_REL
;
2945 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
2946 hrtimer_set_expires(&hs
.timer
, kt
);
2948 hrtimer_init_sleeper(&hs
, current
);
2950 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
2952 set_current_state(TASK_UNINTERRUPTIBLE
);
2953 hrtimer_start_expires(&hs
.timer
, mode
);
2956 hrtimer_cancel(&hs
.timer
);
2957 mode
= HRTIMER_MODE_ABS
;
2958 } while (hs
.task
&& !signal_pending(current
));
2960 __set_current_state(TASK_RUNNING
);
2961 destroy_hrtimer_on_stack(&hs
.timer
);
2965 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
2967 struct request_queue
*q
= hctx
->queue
;
2971 * If we sleep, have the caller restart the poll loop to reset
2972 * the state. Like for the other success return cases, the
2973 * caller is responsible for checking if the IO completed. If
2974 * the IO isn't complete, we'll get called again and will go
2975 * straight to the busy poll loop.
2977 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
2980 hctx
->poll_considered
++;
2982 state
= current
->state
;
2983 while (!need_resched()) {
2986 hctx
->poll_invoked
++;
2988 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
2990 hctx
->poll_success
++;
2991 set_current_state(TASK_RUNNING
);
2995 if (signal_pending_state(state
, current
))
2996 set_current_state(TASK_RUNNING
);
2998 if (current
->state
== TASK_RUNNING
)
3008 static bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
3010 struct blk_mq_hw_ctx
*hctx
;
3013 if (!test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3016 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3017 if (!blk_qc_t_is_internal(cookie
))
3018 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3020 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3022 * With scheduling, if the request has completed, we'll
3023 * get a NULL return here, as we clear the sched tag when
3024 * that happens. The request still remains valid, like always,
3025 * so we should be safe with just the NULL check.
3031 return __blk_mq_poll(hctx
, rq
);
3034 static int __init
blk_mq_init(void)
3037 * See comment in block/blk.h rq_atomic_flags enum
3039 BUILD_BUG_ON((REQ_ATOM_STARTED
/ BITS_PER_BYTE
) !=
3040 (REQ_ATOM_COMPLETE
/ BITS_PER_BYTE
));
3042 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3043 blk_mq_hctx_notify_dead
);
3046 subsys_initcall(blk_mq_init
);