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 static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx
*hctx
,
123 struct request
*rq
, void *priv
,
126 struct mq_inflight
*mi
= priv
;
128 if (rq
->part
== mi
->part
)
129 mi
->inflight
[rq_data_dir(rq
)]++;
132 void blk_mq_in_flight_rw(struct request_queue
*q
, struct hd_struct
*part
,
133 unsigned int inflight
[2])
135 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
137 inflight
[0] = inflight
[1] = 0;
138 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight_rw
, &mi
);
141 void blk_freeze_queue_start(struct request_queue
*q
)
145 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
146 if (freeze_depth
== 1) {
147 percpu_ref_kill(&q
->q_usage_counter
);
149 blk_mq_run_hw_queues(q
, false);
152 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
154 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
156 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
158 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
160 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
161 unsigned long timeout
)
163 return wait_event_timeout(q
->mq_freeze_wq
,
164 percpu_ref_is_zero(&q
->q_usage_counter
),
167 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
170 * Guarantee no request is in use, so we can change any data structure of
171 * the queue afterward.
173 void blk_freeze_queue(struct request_queue
*q
)
176 * In the !blk_mq case we are only calling this to kill the
177 * q_usage_counter, otherwise this increases the freeze depth
178 * and waits for it to return to zero. For this reason there is
179 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
180 * exported to drivers as the only user for unfreeze is blk_mq.
182 blk_freeze_queue_start(q
);
185 blk_mq_freeze_queue_wait(q
);
188 void blk_mq_freeze_queue(struct request_queue
*q
)
191 * ...just an alias to keep freeze and unfreeze actions balanced
192 * in the blk_mq_* namespace
196 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
198 void blk_mq_unfreeze_queue(struct request_queue
*q
)
202 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
203 WARN_ON_ONCE(freeze_depth
< 0);
205 percpu_ref_reinit(&q
->q_usage_counter
);
206 wake_up_all(&q
->mq_freeze_wq
);
209 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
212 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
213 * mpt3sas driver such that this function can be removed.
215 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
219 spin_lock_irqsave(q
->queue_lock
, flags
);
220 queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
221 spin_unlock_irqrestore(q
->queue_lock
, flags
);
223 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
226 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
229 * Note: this function does not prevent that the struct request end_io()
230 * callback function is invoked. Once this function is returned, we make
231 * sure no dispatch can happen until the queue is unquiesced via
232 * blk_mq_unquiesce_queue().
234 void blk_mq_quiesce_queue(struct request_queue
*q
)
236 struct blk_mq_hw_ctx
*hctx
;
240 blk_mq_quiesce_queue_nowait(q
);
242 queue_for_each_hw_ctx(q
, hctx
, i
) {
243 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
244 synchronize_srcu(hctx
->queue_rq_srcu
);
251 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
254 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
257 * This function recovers queue into the state before quiescing
258 * which is done by blk_mq_quiesce_queue.
260 void blk_mq_unquiesce_queue(struct request_queue
*q
)
264 spin_lock_irqsave(q
->queue_lock
, flags
);
265 queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
266 spin_unlock_irqrestore(q
->queue_lock
, flags
);
268 /* dispatch requests which are inserted during quiescing */
269 blk_mq_run_hw_queues(q
, true);
271 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
273 void blk_mq_wake_waiters(struct request_queue
*q
)
275 struct blk_mq_hw_ctx
*hctx
;
278 queue_for_each_hw_ctx(q
, hctx
, i
)
279 if (blk_mq_hw_queue_mapped(hctx
))
280 blk_mq_tag_wakeup_all(hctx
->tags
, true);
283 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
285 return blk_mq_has_free_tags(hctx
->tags
);
287 EXPORT_SYMBOL(blk_mq_can_queue
);
289 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
290 unsigned int tag
, unsigned int op
)
292 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
293 struct request
*rq
= tags
->static_rqs
[tag
];
297 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
299 rq
->internal_tag
= tag
;
301 if (data
->hctx
->flags
& BLK_MQ_F_TAG_SHARED
) {
302 rq
->rq_flags
= RQF_MQ_INFLIGHT
;
303 atomic_inc(&data
->hctx
->nr_active
);
306 rq
->internal_tag
= -1;
307 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
310 INIT_LIST_HEAD(&rq
->queuelist
);
311 /* csd/requeue_work/fifo_time is initialized before use */
313 rq
->mq_ctx
= data
->ctx
;
315 if (data
->flags
& BLK_MQ_REQ_PREEMPT
)
316 rq
->rq_flags
|= RQF_PREEMPT
;
317 if (blk_queue_io_stat(data
->q
))
318 rq
->rq_flags
|= RQF_IO_STAT
;
319 /* do not touch atomic flags, it needs atomic ops against the timer */
321 INIT_HLIST_NODE(&rq
->hash
);
322 RB_CLEAR_NODE(&rq
->rb_node
);
325 rq
->start_time
= jiffies
;
326 #ifdef CONFIG_BLK_CGROUP
328 set_start_time_ns(rq
);
329 rq
->io_start_time_ns
= 0;
331 rq
->nr_phys_segments
= 0;
332 #if defined(CONFIG_BLK_DEV_INTEGRITY)
333 rq
->nr_integrity_segments
= 0;
336 /* tag was already set */
339 INIT_LIST_HEAD(&rq
->timeout_list
);
343 rq
->end_io_data
= NULL
;
346 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
350 static struct request
*blk_mq_get_request(struct request_queue
*q
,
351 struct bio
*bio
, unsigned int op
,
352 struct blk_mq_alloc_data
*data
)
354 struct elevator_queue
*e
= q
->elevator
;
357 bool put_ctx_on_error
= false;
359 blk_queue_enter_live(q
);
361 if (likely(!data
->ctx
)) {
362 data
->ctx
= blk_mq_get_ctx(q
);
363 put_ctx_on_error
= true;
365 if (likely(!data
->hctx
))
366 data
->hctx
= blk_mq_map_queue(q
, data
->ctx
->cpu
);
368 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
371 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
374 * Flush requests are special and go directly to the
377 if (!op_is_flush(op
) && e
->type
->ops
.mq
.limit_depth
)
378 e
->type
->ops
.mq
.limit_depth(op
, data
);
380 blk_mq_tag_busy(data
->hctx
);
383 tag
= blk_mq_get_tag(data
);
384 if (tag
== BLK_MQ_TAG_FAIL
) {
385 if (put_ctx_on_error
) {
386 blk_mq_put_ctx(data
->ctx
);
393 rq
= blk_mq_rq_ctx_init(data
, tag
, op
);
394 if (!op_is_flush(op
)) {
396 if (e
&& e
->type
->ops
.mq
.prepare_request
) {
397 if (e
->type
->icq_cache
&& rq_ioc(bio
))
398 blk_mq_sched_assign_ioc(rq
, bio
);
400 e
->type
->ops
.mq
.prepare_request(rq
, bio
);
401 rq
->rq_flags
|= RQF_ELVPRIV
;
404 data
->hctx
->queued
++;
408 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
409 blk_mq_req_flags_t flags
)
411 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
415 ret
= blk_queue_enter(q
, flags
);
419 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
423 return ERR_PTR(-EWOULDBLOCK
);
425 blk_mq_put_ctx(alloc_data
.ctx
);
428 rq
->__sector
= (sector_t
) -1;
429 rq
->bio
= rq
->biotail
= NULL
;
432 EXPORT_SYMBOL(blk_mq_alloc_request
);
434 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
435 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
437 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
443 * If the tag allocator sleeps we could get an allocation for a
444 * different hardware context. No need to complicate the low level
445 * allocator for this for the rare use case of a command tied to
448 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
449 return ERR_PTR(-EINVAL
);
451 if (hctx_idx
>= q
->nr_hw_queues
)
452 return ERR_PTR(-EIO
);
454 ret
= blk_queue_enter(q
, flags
);
459 * Check if the hardware context is actually mapped to anything.
460 * If not tell the caller that it should skip this queue.
462 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
463 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
465 return ERR_PTR(-EXDEV
);
467 cpu
= cpumask_first_and(alloc_data
.hctx
->cpumask
, cpu_online_mask
);
468 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
470 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
474 return ERR_PTR(-EWOULDBLOCK
);
478 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
480 void blk_mq_free_request(struct request
*rq
)
482 struct request_queue
*q
= rq
->q
;
483 struct elevator_queue
*e
= q
->elevator
;
484 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
485 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
486 const int sched_tag
= rq
->internal_tag
;
488 if (rq
->rq_flags
& RQF_ELVPRIV
) {
489 if (e
&& e
->type
->ops
.mq
.finish_request
)
490 e
->type
->ops
.mq
.finish_request(rq
);
492 put_io_context(rq
->elv
.icq
->ioc
);
497 ctx
->rq_completed
[rq_is_sync(rq
)]++;
498 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
499 atomic_dec(&hctx
->nr_active
);
501 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
502 laptop_io_completion(q
->backing_dev_info
);
504 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
507 blk_put_rl(blk_rq_rl(rq
));
509 clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
510 clear_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
512 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
514 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
515 blk_mq_sched_restart(hctx
);
518 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
520 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
522 blk_account_io_done(rq
);
525 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
526 rq
->end_io(rq
, error
);
528 if (unlikely(blk_bidi_rq(rq
)))
529 blk_mq_free_request(rq
->next_rq
);
530 blk_mq_free_request(rq
);
533 EXPORT_SYMBOL(__blk_mq_end_request
);
535 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
537 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
539 __blk_mq_end_request(rq
, error
);
541 EXPORT_SYMBOL(blk_mq_end_request
);
543 static void __blk_mq_complete_request_remote(void *data
)
545 struct request
*rq
= data
;
547 rq
->q
->softirq_done_fn(rq
);
550 static void __blk_mq_complete_request(struct request
*rq
)
552 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
556 if (rq
->internal_tag
!= -1)
557 blk_mq_sched_completed_request(rq
);
558 if (rq
->rq_flags
& RQF_STATS
) {
559 blk_mq_poll_stats_start(rq
->q
);
563 if (!test_bit(QUEUE_FLAG_SAME_COMP
, &rq
->q
->queue_flags
)) {
564 rq
->q
->softirq_done_fn(rq
);
569 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &rq
->q
->queue_flags
))
570 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
572 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
573 rq
->csd
.func
= __blk_mq_complete_request_remote
;
576 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
578 rq
->q
->softirq_done_fn(rq
);
583 static void hctx_unlock(struct blk_mq_hw_ctx
*hctx
, int srcu_idx
)
585 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
588 srcu_read_unlock(hctx
->queue_rq_srcu
, srcu_idx
);
591 static void hctx_lock(struct blk_mq_hw_ctx
*hctx
, int *srcu_idx
)
593 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
596 *srcu_idx
= srcu_read_lock(hctx
->queue_rq_srcu
);
600 * blk_mq_complete_request - end I/O on a request
601 * @rq: the request being processed
604 * Ends all I/O on a request. It does not handle partial completions.
605 * The actual completion happens out-of-order, through a IPI handler.
607 void blk_mq_complete_request(struct request
*rq
)
609 struct request_queue
*q
= rq
->q
;
611 if (unlikely(blk_should_fake_timeout(q
)))
613 if (!blk_mark_rq_complete(rq
))
614 __blk_mq_complete_request(rq
);
616 EXPORT_SYMBOL(blk_mq_complete_request
);
618 int blk_mq_request_started(struct request
*rq
)
620 return test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
622 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
624 void blk_mq_start_request(struct request
*rq
)
626 struct request_queue
*q
= rq
->q
;
628 blk_mq_sched_started_request(rq
);
630 trace_block_rq_issue(q
, rq
);
632 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
633 blk_stat_set_issue(&rq
->issue_stat
, blk_rq_sectors(rq
));
634 rq
->rq_flags
|= RQF_STATS
;
635 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
640 WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
));
643 * Mark us as started and clear complete. Complete might have been
644 * set if requeue raced with timeout, which then marked it as
645 * complete. So be sure to clear complete again when we start
646 * the request, otherwise we'll ignore the completion event.
648 * Ensure that ->deadline is visible before we set STARTED, such that
649 * blk_mq_check_expired() is guaranteed to observe our ->deadline when
650 * it observes STARTED.
653 set_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
);
654 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
)) {
656 * Coherence order guarantees these consecutive stores to a
657 * single variable propagate in the specified order. Thus the
658 * clear_bit() is ordered _after_ the set bit. See
659 * blk_mq_check_expired().
661 * (the bits must be part of the same byte for this to be
664 clear_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
);
667 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
669 * Make sure space for the drain appears. We know we can do
670 * this because max_hw_segments has been adjusted to be one
671 * fewer than the device can handle.
673 rq
->nr_phys_segments
++;
676 EXPORT_SYMBOL(blk_mq_start_request
);
679 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
680 * flag isn't set yet, so there may be race with timeout handler,
681 * but given rq->deadline is just set in .queue_rq() under
682 * this situation, the race won't be possible in reality because
683 * rq->timeout should be set as big enough to cover the window
684 * between blk_mq_start_request() called from .queue_rq() and
685 * clearing REQ_ATOM_STARTED here.
687 static void __blk_mq_requeue_request(struct request
*rq
)
689 struct request_queue
*q
= rq
->q
;
691 blk_mq_put_driver_tag(rq
);
693 trace_block_rq_requeue(q
, rq
);
694 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
696 if (test_and_clear_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
)) {
697 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
698 rq
->nr_phys_segments
--;
702 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
704 __blk_mq_requeue_request(rq
);
706 /* this request will be re-inserted to io scheduler queue */
707 blk_mq_sched_requeue_request(rq
);
709 BUG_ON(blk_queued_rq(rq
));
710 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
712 EXPORT_SYMBOL(blk_mq_requeue_request
);
714 static void blk_mq_requeue_work(struct work_struct
*work
)
716 struct request_queue
*q
=
717 container_of(work
, struct request_queue
, requeue_work
.work
);
719 struct request
*rq
, *next
;
721 spin_lock_irq(&q
->requeue_lock
);
722 list_splice_init(&q
->requeue_list
, &rq_list
);
723 spin_unlock_irq(&q
->requeue_lock
);
725 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
726 if (!(rq
->rq_flags
& (RQF_SOFTBARRIER
| RQF_DONTPREP
)))
729 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
730 list_del_init(&rq
->queuelist
);
732 * If RQF_DONTPREP, rq has contained some driver specific
733 * data, so insert it to hctx dispatch list to avoid any
736 if (rq
->rq_flags
& RQF_DONTPREP
)
737 blk_mq_request_bypass_insert(rq
, false);
739 blk_mq_sched_insert_request(rq
, true, false, false);
742 while (!list_empty(&rq_list
)) {
743 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
744 list_del_init(&rq
->queuelist
);
745 blk_mq_sched_insert_request(rq
, false, false, false);
748 blk_mq_run_hw_queues(q
, false);
751 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
752 bool kick_requeue_list
)
754 struct request_queue
*q
= rq
->q
;
758 * We abuse this flag that is otherwise used by the I/O scheduler to
759 * request head insertion from the workqueue.
761 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
763 spin_lock_irqsave(&q
->requeue_lock
, flags
);
765 rq
->rq_flags
|= RQF_SOFTBARRIER
;
766 list_add(&rq
->queuelist
, &q
->requeue_list
);
768 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
770 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
772 if (kick_requeue_list
)
773 blk_mq_kick_requeue_list(q
);
775 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
777 void blk_mq_kick_requeue_list(struct request_queue
*q
)
779 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
781 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
783 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
786 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
787 msecs_to_jiffies(msecs
));
789 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
791 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
793 if (tag
< tags
->nr_tags
) {
794 prefetch(tags
->rqs
[tag
]);
795 return tags
->rqs
[tag
];
800 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
802 struct blk_mq_timeout_data
{
804 unsigned int next_set
;
807 void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
809 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
810 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
813 * We know that complete is set at this point. If STARTED isn't set
814 * anymore, then the request isn't active and the "timeout" should
815 * just be ignored. This can happen due to the bitflag ordering.
816 * Timeout first checks if STARTED is set, and if it is, assumes
817 * the request is active. But if we race with completion, then
818 * both flags will get cleared. So check here again, and ignore
819 * a timeout event with a request that isn't active.
821 if (!test_bit(REQ_ATOM_STARTED
, &req
->atomic_flags
))
825 ret
= ops
->timeout(req
, reserved
);
829 __blk_mq_complete_request(req
);
831 case BLK_EH_RESET_TIMER
:
833 blk_clear_rq_complete(req
);
835 case BLK_EH_NOT_HANDLED
:
838 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
843 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
844 struct request
*rq
, void *priv
, bool reserved
)
846 struct blk_mq_timeout_data
*data
= priv
;
847 unsigned long deadline
;
849 if (!test_bit(REQ_ATOM_STARTED
, &rq
->atomic_flags
))
853 * Ensures that if we see STARTED we must also see our
854 * up-to-date deadline, see blk_mq_start_request().
858 deadline
= READ_ONCE(rq
->deadline
);
861 * The rq being checked may have been freed and reallocated
862 * out already here, we avoid this race by checking rq->deadline
863 * and REQ_ATOM_COMPLETE flag together:
865 * - if rq->deadline is observed as new value because of
866 * reusing, the rq won't be timed out because of timing.
867 * - if rq->deadline is observed as previous value,
868 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
869 * because we put a barrier between setting rq->deadline
870 * and clearing the flag in blk_mq_start_request(), so
871 * this rq won't be timed out too.
873 if (time_after_eq(jiffies
, deadline
)) {
874 if (!blk_mark_rq_complete(rq
)) {
876 * Again coherence order ensures that consecutive reads
877 * from the same variable must be in that order. This
878 * ensures that if we see COMPLETE clear, we must then
879 * see STARTED set and we'll ignore this timeout.
881 * (There's also the MB implied by the test_and_clear())
883 blk_mq_rq_timed_out(rq
, reserved
);
885 } else if (!data
->next_set
|| time_after(data
->next
, deadline
)) {
886 data
->next
= deadline
;
891 static void blk_mq_timeout_work(struct work_struct
*work
)
893 struct request_queue
*q
=
894 container_of(work
, struct request_queue
, timeout_work
);
895 struct blk_mq_timeout_data data
= {
901 /* A deadlock might occur if a request is stuck requiring a
902 * timeout at the same time a queue freeze is waiting
903 * completion, since the timeout code would not be able to
904 * acquire the queue reference here.
906 * That's why we don't use blk_queue_enter here; instead, we use
907 * percpu_ref_tryget directly, because we need to be able to
908 * obtain a reference even in the short window between the queue
909 * starting to freeze, by dropping the first reference in
910 * blk_freeze_queue_start, and the moment the last request is
911 * consumed, marked by the instant q_usage_counter reaches
914 if (!percpu_ref_tryget(&q
->q_usage_counter
))
917 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
920 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
921 mod_timer(&q
->timeout
, data
.next
);
923 struct blk_mq_hw_ctx
*hctx
;
925 queue_for_each_hw_ctx(q
, hctx
, i
) {
926 /* the hctx may be unmapped, so check it here */
927 if (blk_mq_hw_queue_mapped(hctx
))
928 blk_mq_tag_idle(hctx
);
934 struct flush_busy_ctx_data
{
935 struct blk_mq_hw_ctx
*hctx
;
936 struct list_head
*list
;
939 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
941 struct flush_busy_ctx_data
*flush_data
= data
;
942 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
943 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
945 sbitmap_clear_bit(sb
, bitnr
);
946 spin_lock(&ctx
->lock
);
947 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
948 spin_unlock(&ctx
->lock
);
953 * Process software queues that have been marked busy, splicing them
954 * to the for-dispatch
956 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
958 struct flush_busy_ctx_data data
= {
963 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
965 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
967 struct dispatch_rq_data
{
968 struct blk_mq_hw_ctx
*hctx
;
972 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
975 struct dispatch_rq_data
*dispatch_data
= data
;
976 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
977 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
979 spin_lock(&ctx
->lock
);
980 if (unlikely(!list_empty(&ctx
->rq_list
))) {
981 dispatch_data
->rq
= list_entry_rq(ctx
->rq_list
.next
);
982 list_del_init(&dispatch_data
->rq
->queuelist
);
983 if (list_empty(&ctx
->rq_list
))
984 sbitmap_clear_bit(sb
, bitnr
);
986 spin_unlock(&ctx
->lock
);
988 return !dispatch_data
->rq
;
991 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
992 struct blk_mq_ctx
*start
)
994 unsigned off
= start
? start
->index_hw
: 0;
995 struct dispatch_rq_data data
= {
1000 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1001 dispatch_rq_from_ctx
, &data
);
1006 static inline unsigned int queued_to_index(unsigned int queued
)
1011 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
1014 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
1017 struct blk_mq_alloc_data data
= {
1019 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
1020 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
1024 might_sleep_if(wait
);
1029 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
1030 data
.flags
|= BLK_MQ_REQ_RESERVED
;
1032 shared
= blk_mq_tag_busy(data
.hctx
);
1033 rq
->tag
= blk_mq_get_tag(&data
);
1036 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1037 atomic_inc(&data
.hctx
->nr_active
);
1039 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
1045 return rq
->tag
!= -1;
1048 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1049 int flags
, void *key
)
1051 struct blk_mq_hw_ctx
*hctx
;
1053 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1055 list_del_init(&wait
->entry
);
1056 blk_mq_run_hw_queue(hctx
, true);
1061 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1062 * the tag wakeups. For non-shared tags, we can simply mark us nedeing a
1063 * restart. For both caes, take care to check the condition again after
1064 * marking us as waiting.
1066 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
**hctx
,
1069 struct blk_mq_hw_ctx
*this_hctx
= *hctx
;
1070 bool shared_tags
= (this_hctx
->flags
& BLK_MQ_F_TAG_SHARED
) != 0;
1071 struct sbq_wait_state
*ws
;
1072 wait_queue_entry_t
*wait
;
1076 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &this_hctx
->state
))
1077 set_bit(BLK_MQ_S_SCHED_RESTART
, &this_hctx
->state
);
1079 wait
= &this_hctx
->dispatch_wait
;
1080 if (!list_empty_careful(&wait
->entry
))
1083 spin_lock(&this_hctx
->lock
);
1084 if (!list_empty(&wait
->entry
)) {
1085 spin_unlock(&this_hctx
->lock
);
1089 ws
= bt_wait_ptr(&this_hctx
->tags
->bitmap_tags
, this_hctx
);
1090 add_wait_queue(&ws
->wait
, wait
);
1094 * It's possible that a tag was freed in the window between the
1095 * allocation failure and adding the hardware queue to the wait
1098 ret
= blk_mq_get_driver_tag(rq
, hctx
, false);
1102 * Don't clear RESTART here, someone else could have set it.
1103 * At most this will cost an extra queue run.
1108 spin_unlock(&this_hctx
->lock
);
1113 * We got a tag, remove ourselves from the wait queue to ensure
1114 * someone else gets the wakeup.
1116 spin_lock_irq(&ws
->wait
.lock
);
1117 list_del_init(&wait
->entry
);
1118 spin_unlock_irq(&ws
->wait
.lock
);
1119 spin_unlock(&this_hctx
->lock
);
1124 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1125 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1127 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1128 * - EWMA is one simple way to compute running average value
1129 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1130 * - take 4 as factor for avoiding to get too small(0) result, and this
1131 * factor doesn't matter because EWMA decreases exponentially
1133 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1137 if (hctx
->queue
->elevator
)
1140 ewma
= hctx
->dispatch_busy
;
1145 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1147 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1148 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1150 hctx
->dispatch_busy
= ewma
;
1153 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1156 * Returns true if we did some work AND can potentially do more.
1158 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
,
1161 struct blk_mq_hw_ctx
*hctx
;
1162 struct request
*rq
, *nxt
;
1163 bool no_tag
= false;
1165 blk_status_t ret
= BLK_STS_OK
;
1167 if (list_empty(list
))
1170 WARN_ON(!list_is_singular(list
) && got_budget
);
1173 * Now process all the entries, sending them to the driver.
1175 errors
= queued
= 0;
1177 struct blk_mq_queue_data bd
;
1179 rq
= list_first_entry(list
, struct request
, queuelist
);
1181 hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
);
1182 if (!got_budget
&& !blk_mq_get_dispatch_budget(hctx
))
1185 if (!blk_mq_get_driver_tag(rq
, NULL
, false)) {
1187 * The initial allocation attempt failed, so we need to
1188 * rerun the hardware queue when a tag is freed. The
1189 * waitqueue takes care of that. If the queue is run
1190 * before we add this entry back on the dispatch list,
1191 * we'll re-run it below.
1193 if (!blk_mq_mark_tag_wait(&hctx
, rq
)) {
1194 blk_mq_put_dispatch_budget(hctx
);
1196 * For non-shared tags, the RESTART check
1199 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
)
1205 list_del_init(&rq
->queuelist
);
1210 * Flag last if we have no more requests, or if we have more
1211 * but can't assign a driver tag to it.
1213 if (list_empty(list
))
1216 nxt
= list_first_entry(list
, struct request
, queuelist
);
1217 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1220 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1221 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
) {
1223 * If an I/O scheduler has been configured and we got a
1224 * driver tag for the next request already, free it
1227 if (!list_empty(list
)) {
1228 nxt
= list_first_entry(list
, struct request
, queuelist
);
1229 blk_mq_put_driver_tag(nxt
);
1231 list_add(&rq
->queuelist
, list
);
1232 __blk_mq_requeue_request(rq
);
1236 if (unlikely(ret
!= BLK_STS_OK
)) {
1238 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1243 } while (!list_empty(list
));
1245 hctx
->dispatched
[queued_to_index(queued
)]++;
1248 * Any items that need requeuing? Stuff them into hctx->dispatch,
1249 * that is where we will continue on next queue run.
1251 if (!list_empty(list
)) {
1254 spin_lock(&hctx
->lock
);
1255 list_splice_init(list
, &hctx
->dispatch
);
1256 spin_unlock(&hctx
->lock
);
1259 * If SCHED_RESTART was set by the caller of this function and
1260 * it is no longer set that means that it was cleared by another
1261 * thread and hence that a queue rerun is needed.
1263 * If 'no_tag' is set, that means that we failed getting
1264 * a driver tag with an I/O scheduler attached. If our dispatch
1265 * waitqueue is no longer active, ensure that we run the queue
1266 * AFTER adding our entries back to the list.
1268 * If no I/O scheduler has been configured it is possible that
1269 * the hardware queue got stopped and restarted before requests
1270 * were pushed back onto the dispatch list. Rerun the queue to
1271 * avoid starvation. Notes:
1272 * - blk_mq_run_hw_queue() checks whether or not a queue has
1273 * been stopped before rerunning a queue.
1274 * - Some but not all block drivers stop a queue before
1275 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1278 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1279 * bit is set, run queue after a delay to avoid IO stalls
1280 * that could otherwise occur if the queue is idle.
1282 needs_restart
= blk_mq_sched_needs_restart(hctx
);
1283 if (!needs_restart
||
1284 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1285 blk_mq_run_hw_queue(hctx
, true);
1286 else if (needs_restart
&& (ret
== BLK_STS_RESOURCE
))
1287 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
1289 blk_mq_update_dispatch_busy(hctx
, true);
1292 blk_mq_update_dispatch_busy(hctx
, false);
1295 * If the host/device is unable to accept more work, inform the
1298 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1301 return (queued
+ errors
) != 0;
1304 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1309 * We should be running this queue from one of the CPUs that
1312 * There are at least two related races now between setting
1313 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1314 * __blk_mq_run_hw_queue():
1316 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1317 * but later it becomes online, then this warning is harmless
1320 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1321 * but later it becomes offline, then the warning can't be
1322 * triggered, and we depend on blk-mq timeout handler to
1323 * handle dispatched requests to this hctx
1325 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1326 cpu_online(hctx
->next_cpu
)) {
1327 printk(KERN_WARNING
"run queue from wrong CPU %d, hctx %s\n",
1328 raw_smp_processor_id(),
1329 cpumask_empty(hctx
->cpumask
) ? "inactive": "active");
1334 * We can't run the queue inline with ints disabled. Ensure that
1335 * we catch bad users of this early.
1337 WARN_ON_ONCE(in_interrupt());
1339 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1341 hctx_lock(hctx
, &srcu_idx
);
1342 blk_mq_sched_dispatch_requests(hctx
);
1343 hctx_unlock(hctx
, srcu_idx
);
1346 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1348 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1350 if (cpu
>= nr_cpu_ids
)
1351 cpu
= cpumask_first(hctx
->cpumask
);
1356 * It'd be great if the workqueue API had a way to pass
1357 * in a mask and had some smarts for more clever placement.
1358 * For now we just round-robin here, switching for every
1359 * BLK_MQ_CPU_WORK_BATCH queued items.
1361 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1364 int next_cpu
= hctx
->next_cpu
;
1366 if (hctx
->queue
->nr_hw_queues
== 1)
1367 return WORK_CPU_UNBOUND
;
1369 if (--hctx
->next_cpu_batch
<= 0) {
1371 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1373 if (next_cpu
>= nr_cpu_ids
)
1374 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1375 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1379 * Do unbound schedule if we can't find a online CPU for this hctx,
1380 * and it should only happen in the path of handling CPU DEAD.
1382 if (!cpu_online(next_cpu
)) {
1389 * Make sure to re-select CPU next time once after CPUs
1390 * in hctx->cpumask become online again.
1392 hctx
->next_cpu
= next_cpu
;
1393 hctx
->next_cpu_batch
= 1;
1394 return WORK_CPU_UNBOUND
;
1397 hctx
->next_cpu
= next_cpu
;
1401 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1402 unsigned long msecs
)
1404 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1407 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1408 int cpu
= get_cpu();
1409 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1410 __blk_mq_run_hw_queue(hctx
);
1418 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1419 msecs_to_jiffies(msecs
));
1422 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1424 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1426 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1428 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1434 * When queue is quiesced, we may be switching io scheduler, or
1435 * updating nr_hw_queues, or other things, and we can't run queue
1436 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1438 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1441 hctx_lock(hctx
, &srcu_idx
);
1442 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1443 blk_mq_hctx_has_pending(hctx
);
1444 hctx_unlock(hctx
, srcu_idx
);
1447 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1453 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1455 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1457 struct blk_mq_hw_ctx
*hctx
;
1460 queue_for_each_hw_ctx(q
, hctx
, i
) {
1461 if (blk_mq_hctx_stopped(hctx
))
1464 blk_mq_run_hw_queue(hctx
, async
);
1467 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1470 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1471 * @q: request queue.
1473 * The caller is responsible for serializing this function against
1474 * blk_mq_{start,stop}_hw_queue().
1476 bool blk_mq_queue_stopped(struct request_queue
*q
)
1478 struct blk_mq_hw_ctx
*hctx
;
1481 queue_for_each_hw_ctx(q
, hctx
, i
)
1482 if (blk_mq_hctx_stopped(hctx
))
1487 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1490 * This function is often used for pausing .queue_rq() by driver when
1491 * there isn't enough resource or some conditions aren't satisfied, and
1492 * BLK_STS_RESOURCE is usually returned.
1494 * We do not guarantee that dispatch can be drained or blocked
1495 * after blk_mq_stop_hw_queue() returns. Please use
1496 * blk_mq_quiesce_queue() for that requirement.
1498 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1500 cancel_delayed_work(&hctx
->run_work
);
1502 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1504 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1507 * This function is often used for pausing .queue_rq() by driver when
1508 * there isn't enough resource or some conditions aren't satisfied, and
1509 * BLK_STS_RESOURCE is usually returned.
1511 * We do not guarantee that dispatch can be drained or blocked
1512 * after blk_mq_stop_hw_queues() returns. Please use
1513 * blk_mq_quiesce_queue() for that requirement.
1515 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1517 struct blk_mq_hw_ctx
*hctx
;
1520 queue_for_each_hw_ctx(q
, hctx
, i
)
1521 blk_mq_stop_hw_queue(hctx
);
1523 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1525 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1527 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1529 blk_mq_run_hw_queue(hctx
, false);
1531 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1533 void blk_mq_start_hw_queues(struct request_queue
*q
)
1535 struct blk_mq_hw_ctx
*hctx
;
1538 queue_for_each_hw_ctx(q
, hctx
, i
)
1539 blk_mq_start_hw_queue(hctx
);
1541 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1543 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1545 if (!blk_mq_hctx_stopped(hctx
))
1548 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1549 blk_mq_run_hw_queue(hctx
, async
);
1551 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1553 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1555 struct blk_mq_hw_ctx
*hctx
;
1558 queue_for_each_hw_ctx(q
, hctx
, i
)
1559 blk_mq_start_stopped_hw_queue(hctx
, async
);
1561 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1563 static void blk_mq_run_work_fn(struct work_struct
*work
)
1565 struct blk_mq_hw_ctx
*hctx
;
1567 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1570 * If we are stopped, don't run the queue. The exception is if
1571 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1572 * the STOPPED bit and run it.
1574 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)) {
1575 if (!test_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
))
1578 clear_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1579 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1582 __blk_mq_run_hw_queue(hctx
);
1586 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1588 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx
)))
1592 * Stop the hw queue, then modify currently delayed work.
1593 * This should prevent us from running the queue prematurely.
1594 * Mark the queue as auto-clearing STOPPED when it runs.
1596 blk_mq_stop_hw_queue(hctx
);
1597 set_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1598 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1600 msecs_to_jiffies(msecs
));
1602 EXPORT_SYMBOL(blk_mq_delay_queue
);
1604 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1608 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1610 lockdep_assert_held(&ctx
->lock
);
1612 trace_block_rq_insert(hctx
->queue
, rq
);
1615 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1617 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1620 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1623 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1625 lockdep_assert_held(&ctx
->lock
);
1627 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1628 blk_mq_hctx_mark_pending(hctx
, ctx
);
1632 * Should only be used carefully, when the caller knows we want to
1633 * bypass a potential IO scheduler on the target device.
1635 void blk_mq_request_bypass_insert(struct request
*rq
, bool run_queue
)
1637 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1638 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(rq
->q
, ctx
->cpu
);
1640 spin_lock(&hctx
->lock
);
1641 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1642 spin_unlock(&hctx
->lock
);
1645 blk_mq_run_hw_queue(hctx
, false);
1648 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1649 struct list_head
*list
)
1653 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1656 spin_lock(&ctx
->lock
);
1657 while (!list_empty(list
)) {
1660 rq
= list_first_entry(list
, struct request
, queuelist
);
1661 BUG_ON(rq
->mq_ctx
!= ctx
);
1662 list_del_init(&rq
->queuelist
);
1663 __blk_mq_insert_req_list(hctx
, rq
, false);
1665 blk_mq_hctx_mark_pending(hctx
, ctx
);
1666 spin_unlock(&ctx
->lock
);
1669 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1671 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1672 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1674 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1675 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1676 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1679 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1681 struct blk_mq_ctx
*this_ctx
;
1682 struct request_queue
*this_q
;
1685 LIST_HEAD(ctx_list
);
1688 list_splice_init(&plug
->mq_list
, &list
);
1690 list_sort(NULL
, &list
, plug_ctx_cmp
);
1696 while (!list_empty(&list
)) {
1697 rq
= list_entry_rq(list
.next
);
1698 list_del_init(&rq
->queuelist
);
1700 if (rq
->mq_ctx
!= this_ctx
) {
1702 trace_block_unplug(this_q
, depth
, !from_schedule
);
1703 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1708 this_ctx
= rq
->mq_ctx
;
1714 list_add_tail(&rq
->queuelist
, &ctx_list
);
1718 * If 'this_ctx' is set, we know we have entries to complete
1719 * on 'ctx_list'. Do those.
1722 trace_block_unplug(this_q
, depth
, !from_schedule
);
1723 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1728 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1730 blk_init_request_from_bio(rq
, bio
);
1732 blk_rq_set_rl(rq
, blk_get_rl(rq
->q
, bio
));
1734 blk_account_io_start(rq
, true);
1737 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx
*hctx
,
1738 struct blk_mq_ctx
*ctx
,
1741 spin_lock(&ctx
->lock
);
1742 __blk_mq_insert_request(hctx
, rq
, false);
1743 spin_unlock(&ctx
->lock
);
1746 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1749 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1751 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1754 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1758 struct request_queue
*q
= rq
->q
;
1759 struct blk_mq_queue_data bd
= {
1763 blk_qc_t new_cookie
;
1766 new_cookie
= request_to_qc_t(hctx
, rq
);
1769 * For OK queue, we are done. For error, caller may kill it.
1770 * Any other error (busy), just add it to our list as we
1771 * previously would have done.
1773 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1776 blk_mq_update_dispatch_busy(hctx
, false);
1777 *cookie
= new_cookie
;
1779 case BLK_STS_RESOURCE
:
1780 case BLK_STS_DEV_RESOURCE
:
1781 blk_mq_update_dispatch_busy(hctx
, true);
1782 __blk_mq_requeue_request(rq
);
1785 blk_mq_update_dispatch_busy(hctx
, false);
1786 *cookie
= BLK_QC_T_NONE
;
1793 static blk_status_t
__blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1798 struct request_queue
*q
= rq
->q
;
1799 bool run_queue
= true;
1802 * RCU or SRCU read lock is needed before checking quiesced flag.
1804 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1805 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1806 * and avoid driver to try to dispatch again.
1808 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1810 bypass_insert
= false;
1814 if (q
->elevator
&& !bypass_insert
)
1817 if (!blk_mq_get_dispatch_budget(hctx
))
1820 if (!blk_mq_get_driver_tag(rq
, NULL
, false)) {
1821 blk_mq_put_dispatch_budget(hctx
);
1825 return __blk_mq_issue_directly(hctx
, rq
, cookie
);
1828 return BLK_STS_RESOURCE
;
1830 blk_mq_request_bypass_insert(rq
, run_queue
);
1834 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1835 struct request
*rq
, blk_qc_t
*cookie
)
1840 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1842 hctx_lock(hctx
, &srcu_idx
);
1844 ret
= __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false);
1845 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1846 blk_mq_request_bypass_insert(rq
, true);
1847 else if (ret
!= BLK_STS_OK
)
1848 blk_mq_end_request(rq
, ret
);
1850 hctx_unlock(hctx
, srcu_idx
);
1853 blk_status_t
blk_mq_request_issue_directly(struct request
*rq
)
1857 blk_qc_t unused_cookie
;
1858 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1859 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(rq
->q
, ctx
->cpu
);
1861 hctx_lock(hctx
, &srcu_idx
);
1862 ret
= __blk_mq_try_issue_directly(hctx
, rq
, &unused_cookie
, true);
1863 hctx_unlock(hctx
, srcu_idx
);
1868 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
1869 struct list_head
*list
)
1871 while (!list_empty(list
)) {
1873 struct request
*rq
= list_first_entry(list
, struct request
,
1876 list_del_init(&rq
->queuelist
);
1877 ret
= blk_mq_request_issue_directly(rq
);
1878 if (ret
!= BLK_STS_OK
) {
1879 if (ret
== BLK_STS_RESOURCE
||
1880 ret
== BLK_STS_DEV_RESOURCE
) {
1881 blk_mq_request_bypass_insert(rq
,
1885 blk_mq_end_request(rq
, ret
);
1890 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1892 const int is_sync
= op_is_sync(bio
->bi_opf
);
1893 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1894 struct blk_mq_alloc_data data
= { .flags
= 0 };
1896 unsigned int request_count
= 0;
1897 struct blk_plug
*plug
;
1898 struct request
*same_queue_rq
= NULL
;
1900 unsigned int wb_acct
;
1902 blk_queue_bounce(q
, &bio
);
1904 blk_queue_split(q
, &bio
);
1906 if (!bio_integrity_prep(bio
))
1907 return BLK_QC_T_NONE
;
1909 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1910 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1911 return BLK_QC_T_NONE
;
1913 if (blk_mq_sched_bio_merge(q
, bio
))
1914 return BLK_QC_T_NONE
;
1916 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1918 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1920 rq
= blk_mq_get_request(q
, bio
, bio
->bi_opf
, &data
);
1921 if (unlikely(!rq
)) {
1922 __wbt_done(q
->rq_wb
, wb_acct
);
1923 if (bio
->bi_opf
& REQ_NOWAIT
)
1924 bio_wouldblock_error(bio
);
1925 return BLK_QC_T_NONE
;
1928 wbt_track(&rq
->issue_stat
, wb_acct
);
1930 cookie
= request_to_qc_t(data
.hctx
, rq
);
1932 plug
= current
->plug
;
1933 if (unlikely(is_flush_fua
)) {
1934 blk_mq_put_ctx(data
.ctx
);
1935 blk_mq_bio_to_request(rq
, bio
);
1937 /* bypass scheduler for flush rq */
1938 blk_insert_flush(rq
);
1939 blk_mq_run_hw_queue(data
.hctx
, true);
1940 } else if (plug
&& q
->nr_hw_queues
== 1) {
1941 struct request
*last
= NULL
;
1943 blk_mq_put_ctx(data
.ctx
);
1944 blk_mq_bio_to_request(rq
, bio
);
1947 * @request_count may become stale because of schedule
1948 * out, so check the list again.
1950 if (list_empty(&plug
->mq_list
))
1952 else if (blk_queue_nomerges(q
))
1953 request_count
= blk_plug_queued_count(q
);
1956 trace_block_plug(q
);
1958 last
= list_entry_rq(plug
->mq_list
.prev
);
1960 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1961 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1962 blk_flush_plug_list(plug
, false);
1963 trace_block_plug(q
);
1966 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1967 } else if (plug
&& !blk_queue_nomerges(q
)) {
1968 blk_mq_bio_to_request(rq
, bio
);
1971 * We do limited plugging. If the bio can be merged, do that.
1972 * Otherwise the existing request in the plug list will be
1973 * issued. So the plug list will have one request at most
1974 * The plug list might get flushed before this. If that happens,
1975 * the plug list is empty, and same_queue_rq is invalid.
1977 if (list_empty(&plug
->mq_list
))
1978 same_queue_rq
= NULL
;
1980 list_del_init(&same_queue_rq
->queuelist
);
1981 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1983 blk_mq_put_ctx(data
.ctx
);
1985 if (same_queue_rq
) {
1986 data
.hctx
= blk_mq_map_queue(q
,
1987 same_queue_rq
->mq_ctx
->cpu
);
1988 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1991 } else if ((q
->nr_hw_queues
> 1 && is_sync
) || (!q
->elevator
&&
1992 !data
.hctx
->dispatch_busy
)) {
1993 blk_mq_put_ctx(data
.ctx
);
1994 blk_mq_bio_to_request(rq
, bio
);
1995 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1996 } else if (q
->elevator
) {
1997 blk_mq_put_ctx(data
.ctx
);
1998 blk_mq_bio_to_request(rq
, bio
);
1999 blk_mq_sched_insert_request(rq
, false, true, true);
2001 blk_mq_put_ctx(data
.ctx
);
2002 blk_mq_bio_to_request(rq
, bio
);
2003 blk_mq_queue_io(data
.hctx
, data
.ctx
, rq
);
2004 blk_mq_run_hw_queue(data
.hctx
, true);
2010 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2011 unsigned int hctx_idx
)
2015 if (tags
->rqs
&& set
->ops
->exit_request
) {
2018 for (i
= 0; i
< tags
->nr_tags
; i
++) {
2019 struct request
*rq
= tags
->static_rqs
[i
];
2023 set
->ops
->exit_request(set
, rq
, hctx_idx
);
2024 tags
->static_rqs
[i
] = NULL
;
2028 while (!list_empty(&tags
->page_list
)) {
2029 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
2030 list_del_init(&page
->lru
);
2032 * Remove kmemleak object previously allocated in
2033 * blk_mq_init_rq_map().
2035 kmemleak_free(page_address(page
));
2036 __free_pages(page
, page
->private);
2040 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
2044 kfree(tags
->static_rqs
);
2045 tags
->static_rqs
= NULL
;
2047 blk_mq_free_tags(tags
);
2050 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
2051 unsigned int hctx_idx
,
2052 unsigned int nr_tags
,
2053 unsigned int reserved_tags
)
2055 struct blk_mq_tags
*tags
;
2058 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
2059 if (node
== NUMA_NO_NODE
)
2060 node
= set
->numa_node
;
2062 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
2063 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
2067 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
2068 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2071 blk_mq_free_tags(tags
);
2075 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
2076 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2078 if (!tags
->static_rqs
) {
2080 blk_mq_free_tags(tags
);
2087 static size_t order_to_size(unsigned int order
)
2089 return (size_t)PAGE_SIZE
<< order
;
2092 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2093 unsigned int hctx_idx
, unsigned int depth
)
2095 unsigned int i
, j
, entries_per_page
, max_order
= 4;
2096 size_t rq_size
, left
;
2099 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
2100 if (node
== NUMA_NO_NODE
)
2101 node
= set
->numa_node
;
2103 INIT_LIST_HEAD(&tags
->page_list
);
2106 * rq_size is the size of the request plus driver payload, rounded
2107 * to the cacheline size
2109 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
2111 left
= rq_size
* depth
;
2113 for (i
= 0; i
< depth
; ) {
2114 int this_order
= max_order
;
2119 while (this_order
&& left
< order_to_size(this_order
- 1))
2123 page
= alloc_pages_node(node
,
2124 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
2130 if (order_to_size(this_order
) < rq_size
)
2137 page
->private = this_order
;
2138 list_add_tail(&page
->lru
, &tags
->page_list
);
2140 p
= page_address(page
);
2142 * Allow kmemleak to scan these pages as they contain pointers
2143 * to additional allocations like via ops->init_request().
2145 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2146 entries_per_page
= order_to_size(this_order
) / rq_size
;
2147 to_do
= min(entries_per_page
, depth
- i
);
2148 left
-= to_do
* rq_size
;
2149 for (j
= 0; j
< to_do
; j
++) {
2150 struct request
*rq
= p
;
2152 tags
->static_rqs
[i
] = rq
;
2153 if (set
->ops
->init_request
) {
2154 if (set
->ops
->init_request(set
, rq
, hctx_idx
,
2156 tags
->static_rqs
[i
] = NULL
;
2168 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2173 * 'cpu' is going away. splice any existing rq_list entries from this
2174 * software queue to the hw queue dispatch list, and ensure that it
2177 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2179 struct blk_mq_hw_ctx
*hctx
;
2180 struct blk_mq_ctx
*ctx
;
2183 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2184 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2186 spin_lock(&ctx
->lock
);
2187 if (!list_empty(&ctx
->rq_list
)) {
2188 list_splice_init(&ctx
->rq_list
, &tmp
);
2189 blk_mq_hctx_clear_pending(hctx
, ctx
);
2191 spin_unlock(&ctx
->lock
);
2193 if (list_empty(&tmp
))
2196 spin_lock(&hctx
->lock
);
2197 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2198 spin_unlock(&hctx
->lock
);
2200 blk_mq_run_hw_queue(hctx
, true);
2204 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2206 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2210 /* hctx->ctxs will be freed in queue's release handler */
2211 static void blk_mq_exit_hctx(struct request_queue
*q
,
2212 struct blk_mq_tag_set
*set
,
2213 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2215 blk_mq_debugfs_unregister_hctx(hctx
);
2217 if (blk_mq_hw_queue_mapped(hctx
))
2218 blk_mq_tag_idle(hctx
);
2220 if (set
->ops
->exit_request
)
2221 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2223 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2225 if (set
->ops
->exit_hctx
)
2226 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2228 blk_mq_remove_cpuhp(hctx
);
2231 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2232 struct blk_mq_tag_set
*set
, int nr_queue
)
2234 struct blk_mq_hw_ctx
*hctx
;
2237 queue_for_each_hw_ctx(q
, hctx
, i
) {
2240 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2244 static int blk_mq_init_hctx(struct request_queue
*q
,
2245 struct blk_mq_tag_set
*set
,
2246 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2250 node
= hctx
->numa_node
;
2251 if (node
== NUMA_NO_NODE
)
2252 node
= hctx
->numa_node
= set
->numa_node
;
2254 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2255 spin_lock_init(&hctx
->lock
);
2256 INIT_LIST_HEAD(&hctx
->dispatch
);
2258 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
2260 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2262 hctx
->tags
= set
->tags
[hctx_idx
];
2265 * Allocate space for all possible cpus to avoid allocation at
2268 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2269 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
, node
);
2271 goto unregister_cpu_notifier
;
2273 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
2274 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
, node
))
2279 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2280 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2282 if (set
->ops
->init_hctx
&&
2283 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2286 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
2289 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
,
2290 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
2292 goto sched_exit_hctx
;
2294 if (set
->ops
->init_request
&&
2295 set
->ops
->init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
,
2299 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2300 init_srcu_struct(hctx
->queue_rq_srcu
);
2302 blk_mq_debugfs_register_hctx(q
, hctx
);
2307 blk_free_flush_queue(hctx
->fq
);
2309 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2311 if (set
->ops
->exit_hctx
)
2312 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2314 sbitmap_free(&hctx
->ctx_map
);
2317 unregister_cpu_notifier
:
2318 blk_mq_remove_cpuhp(hctx
);
2322 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2323 unsigned int nr_hw_queues
)
2327 for_each_possible_cpu(i
) {
2328 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2329 struct blk_mq_hw_ctx
*hctx
;
2332 spin_lock_init(&__ctx
->lock
);
2333 INIT_LIST_HEAD(&__ctx
->rq_list
);
2337 * Set local node, IFF we have more than one hw queue. If
2338 * not, we remain on the home node of the device
2340 hctx
= blk_mq_map_queue(q
, i
);
2341 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2342 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2346 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2350 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2351 set
->queue_depth
, set
->reserved_tags
);
2352 if (!set
->tags
[hctx_idx
])
2355 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2360 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2361 set
->tags
[hctx_idx
] = NULL
;
2365 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2366 unsigned int hctx_idx
)
2368 if (set
->tags
[hctx_idx
]) {
2369 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2370 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2371 set
->tags
[hctx_idx
] = NULL
;
2375 static void blk_mq_map_swqueue(struct request_queue
*q
)
2377 unsigned int i
, hctx_idx
;
2378 struct blk_mq_hw_ctx
*hctx
;
2379 struct blk_mq_ctx
*ctx
;
2380 struct blk_mq_tag_set
*set
= q
->tag_set
;
2383 * Avoid others reading imcomplete hctx->cpumask through sysfs
2385 mutex_lock(&q
->sysfs_lock
);
2387 queue_for_each_hw_ctx(q
, hctx
, i
) {
2388 cpumask_clear(hctx
->cpumask
);
2393 * Map software to hardware queues.
2395 * If the cpu isn't present, the cpu is mapped to first hctx.
2397 for_each_possible_cpu(i
) {
2398 hctx_idx
= q
->mq_map
[i
];
2399 /* unmapped hw queue can be remapped after CPU topo changed */
2400 if (!set
->tags
[hctx_idx
] &&
2401 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2403 * If tags initialization fail for some hctx,
2404 * that hctx won't be brought online. In this
2405 * case, remap the current ctx to hctx[0] which
2406 * is guaranteed to always have tags allocated
2411 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2412 hctx
= blk_mq_map_queue(q
, i
);
2414 cpumask_set_cpu(i
, hctx
->cpumask
);
2415 ctx
->index_hw
= hctx
->nr_ctx
;
2416 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2419 mutex_unlock(&q
->sysfs_lock
);
2421 queue_for_each_hw_ctx(q
, hctx
, i
) {
2423 * If no software queues are mapped to this hardware queue,
2424 * disable it and free the request entries.
2426 if (!hctx
->nr_ctx
) {
2427 /* Never unmap queue 0. We need it as a
2428 * fallback in case of a new remap fails
2431 if (i
&& set
->tags
[i
])
2432 blk_mq_free_map_and_requests(set
, i
);
2438 hctx
->tags
= set
->tags
[i
];
2439 WARN_ON(!hctx
->tags
);
2442 * Set the map size to the number of mapped software queues.
2443 * This is more accurate and more efficient than looping
2444 * over all possibly mapped software queues.
2446 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2449 * Initialize batch roundrobin counts
2451 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2452 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2457 * Caller needs to ensure that we're either frozen/quiesced, or that
2458 * the queue isn't live yet.
2460 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2462 struct blk_mq_hw_ctx
*hctx
;
2465 queue_for_each_hw_ctx(q
, hctx
, i
) {
2467 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2468 atomic_inc(&q
->shared_hctx_restart
);
2469 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2471 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2472 atomic_dec(&q
->shared_hctx_restart
);
2473 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2478 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2481 struct request_queue
*q
;
2483 lockdep_assert_held(&set
->tag_list_lock
);
2485 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2486 blk_mq_freeze_queue(q
);
2487 queue_set_hctx_shared(q
, shared
);
2488 blk_mq_unfreeze_queue(q
);
2492 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2494 struct blk_mq_tag_set
*set
= q
->tag_set
;
2496 mutex_lock(&set
->tag_list_lock
);
2497 list_del_rcu(&q
->tag_set_list
);
2498 if (list_is_singular(&set
->tag_list
)) {
2499 /* just transitioned to unshared */
2500 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2501 /* update existing queue */
2502 blk_mq_update_tag_set_depth(set
, false);
2504 mutex_unlock(&set
->tag_list_lock
);
2506 INIT_LIST_HEAD(&q
->tag_set_list
);
2509 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2510 struct request_queue
*q
)
2514 mutex_lock(&set
->tag_list_lock
);
2517 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2519 if (!list_empty(&set
->tag_list
) &&
2520 !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2521 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2522 /* update existing queue */
2523 blk_mq_update_tag_set_depth(set
, true);
2525 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2526 queue_set_hctx_shared(q
, true);
2527 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2529 mutex_unlock(&set
->tag_list_lock
);
2533 * It is the actual release handler for mq, but we do it from
2534 * request queue's release handler for avoiding use-after-free
2535 * and headache because q->mq_kobj shouldn't have been introduced,
2536 * but we can't group ctx/kctx kobj without it.
2538 void blk_mq_release(struct request_queue
*q
)
2540 struct blk_mq_hw_ctx
*hctx
;
2543 /* hctx kobj stays in hctx */
2544 queue_for_each_hw_ctx(q
, hctx
, i
) {
2547 kobject_put(&hctx
->kobj
);
2552 kfree(q
->queue_hw_ctx
);
2555 * release .mq_kobj and sw queue's kobject now because
2556 * both share lifetime with request queue.
2558 blk_mq_sysfs_deinit(q
);
2560 free_percpu(q
->queue_ctx
);
2563 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2565 struct request_queue
*uninit_q
, *q
;
2567 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2569 return ERR_PTR(-ENOMEM
);
2571 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2573 blk_cleanup_queue(uninit_q
);
2577 EXPORT_SYMBOL(blk_mq_init_queue
);
2579 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2581 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2583 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, queue_rq_srcu
),
2584 __alignof__(struct blk_mq_hw_ctx
)) !=
2585 sizeof(struct blk_mq_hw_ctx
));
2587 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2588 hw_ctx_size
+= sizeof(struct srcu_struct
);
2593 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2594 struct request_queue
*q
)
2597 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2599 blk_mq_sysfs_unregister(q
);
2601 /* protect against switching io scheduler */
2602 mutex_lock(&q
->sysfs_lock
);
2603 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2609 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2610 hctxs
[i
] = kzalloc_node(blk_mq_hw_ctx_size(set
),
2611 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2616 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
,
2617 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2624 atomic_set(&hctxs
[i
]->nr_active
, 0);
2625 hctxs
[i
]->numa_node
= node
;
2626 hctxs
[i
]->queue_num
= i
;
2628 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2629 free_cpumask_var(hctxs
[i
]->cpumask
);
2634 blk_mq_hctx_kobj_init(hctxs
[i
]);
2636 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2637 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2641 blk_mq_free_map_and_requests(set
, j
);
2642 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2643 kobject_put(&hctx
->kobj
);
2648 q
->nr_hw_queues
= i
;
2649 mutex_unlock(&q
->sysfs_lock
);
2650 blk_mq_sysfs_register(q
);
2653 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2654 struct request_queue
*q
)
2656 /* mark the queue as mq asap */
2657 q
->mq_ops
= set
->ops
;
2659 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2660 blk_mq_poll_stats_bkt
,
2661 BLK_MQ_POLL_STATS_BKTS
, q
);
2665 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2669 /* init q->mq_kobj and sw queues' kobjects */
2670 blk_mq_sysfs_init(q
);
2672 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2673 GFP_KERNEL
, set
->numa_node
);
2674 if (!q
->queue_hw_ctx
)
2677 q
->mq_map
= set
->mq_map
;
2679 blk_mq_realloc_hw_ctxs(set
, q
);
2680 if (!q
->nr_hw_queues
)
2683 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2684 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2686 q
->nr_queues
= nr_cpu_ids
;
2688 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2690 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2691 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2693 q
->sg_reserved_size
= INT_MAX
;
2695 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2696 INIT_LIST_HEAD(&q
->requeue_list
);
2697 spin_lock_init(&q
->requeue_lock
);
2699 blk_queue_make_request(q
, blk_mq_make_request
);
2700 if (q
->mq_ops
->poll
)
2701 q
->poll_fn
= blk_mq_poll
;
2704 * Do this after blk_queue_make_request() overrides it...
2706 q
->nr_requests
= set
->queue_depth
;
2709 * Default to classic polling
2713 if (set
->ops
->complete
)
2714 blk_queue_softirq_done(q
, set
->ops
->complete
);
2716 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2717 blk_mq_add_queue_tag_set(set
, q
);
2718 blk_mq_map_swqueue(q
);
2720 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2723 ret
= blk_mq_sched_init(q
);
2725 return ERR_PTR(ret
);
2731 kfree(q
->queue_hw_ctx
);
2733 free_percpu(q
->queue_ctx
);
2736 return ERR_PTR(-ENOMEM
);
2738 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2740 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2741 void blk_mq_exit_queue(struct request_queue
*q
)
2743 struct blk_mq_tag_set
*set
= q
->tag_set
;
2745 blk_mq_del_queue_tag_set(q
);
2746 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2749 /* Basically redo blk_mq_init_queue with queue frozen */
2750 static void blk_mq_queue_reinit(struct request_queue
*q
)
2752 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2754 blk_mq_debugfs_unregister_hctxs(q
);
2755 blk_mq_sysfs_unregister(q
);
2758 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2759 * we should change hctx numa_node according to the new topology (this
2760 * involves freeing and re-allocating memory, worth doing?)
2762 blk_mq_map_swqueue(q
);
2764 blk_mq_sysfs_register(q
);
2765 blk_mq_debugfs_register_hctxs(q
);
2768 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2772 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2773 if (!__blk_mq_alloc_rq_map(set
, i
))
2780 blk_mq_free_rq_map(set
->tags
[i
]);
2786 * Allocate the request maps associated with this tag_set. Note that this
2787 * may reduce the depth asked for, if memory is tight. set->queue_depth
2788 * will be updated to reflect the allocated depth.
2790 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2795 depth
= set
->queue_depth
;
2797 err
= __blk_mq_alloc_rq_maps(set
);
2801 set
->queue_depth
>>= 1;
2802 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2806 } while (set
->queue_depth
);
2808 if (!set
->queue_depth
|| err
) {
2809 pr_err("blk-mq: failed to allocate request map\n");
2813 if (depth
!= set
->queue_depth
)
2814 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2815 depth
, set
->queue_depth
);
2820 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2822 if (set
->ops
->map_queues
) {
2825 * transport .map_queues is usually done in the following
2828 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2829 * mask = get_cpu_mask(queue)
2830 * for_each_cpu(cpu, mask)
2831 * set->mq_map[cpu] = queue;
2834 * When we need to remap, the table has to be cleared for
2835 * killing stale mapping since one CPU may not be mapped
2838 for_each_possible_cpu(cpu
)
2839 set
->mq_map
[cpu
] = 0;
2841 return set
->ops
->map_queues(set
);
2843 return blk_mq_map_queues(set
);
2847 * Alloc a tag set to be associated with one or more request queues.
2848 * May fail with EINVAL for various error conditions. May adjust the
2849 * requested depth down, if if it too large. In that case, the set
2850 * value will be stored in set->queue_depth.
2852 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2856 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2858 if (!set
->nr_hw_queues
)
2860 if (!set
->queue_depth
)
2862 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2865 if (!set
->ops
->queue_rq
)
2868 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
2871 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2872 pr_info("blk-mq: reduced tag depth to %u\n",
2874 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2878 * If a crashdump is active, then we are potentially in a very
2879 * memory constrained environment. Limit us to 1 queue and
2880 * 64 tags to prevent using too much memory.
2882 if (is_kdump_kernel()) {
2883 set
->nr_hw_queues
= 1;
2884 set
->queue_depth
= min(64U, set
->queue_depth
);
2887 * There is no use for more h/w queues than cpus.
2889 if (set
->nr_hw_queues
> nr_cpu_ids
)
2890 set
->nr_hw_queues
= nr_cpu_ids
;
2892 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2893 GFP_KERNEL
, set
->numa_node
);
2898 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2899 GFP_KERNEL
, set
->numa_node
);
2903 ret
= blk_mq_update_queue_map(set
);
2905 goto out_free_mq_map
;
2907 ret
= blk_mq_alloc_rq_maps(set
);
2909 goto out_free_mq_map
;
2911 mutex_init(&set
->tag_list_lock
);
2912 INIT_LIST_HEAD(&set
->tag_list
);
2924 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2926 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2930 for (i
= 0; i
< nr_cpu_ids
; i
++)
2931 blk_mq_free_map_and_requests(set
, i
);
2939 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2941 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2943 struct blk_mq_tag_set
*set
= q
->tag_set
;
2944 struct blk_mq_hw_ctx
*hctx
;
2950 if (q
->nr_requests
== nr
)
2953 blk_mq_freeze_queue(q
);
2954 blk_mq_quiesce_queue(q
);
2957 queue_for_each_hw_ctx(q
, hctx
, i
) {
2961 * If we're using an MQ scheduler, just update the scheduler
2962 * queue depth. This is similar to what the old code would do.
2964 if (!hctx
->sched_tags
) {
2965 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
2968 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2976 q
->nr_requests
= nr
;
2978 blk_mq_unquiesce_queue(q
);
2979 blk_mq_unfreeze_queue(q
);
2984 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
2987 struct request_queue
*q
;
2989 lockdep_assert_held(&set
->tag_list_lock
);
2991 if (nr_hw_queues
> nr_cpu_ids
)
2992 nr_hw_queues
= nr_cpu_ids
;
2993 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2996 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2997 blk_mq_freeze_queue(q
);
2999 set
->nr_hw_queues
= nr_hw_queues
;
3000 blk_mq_update_queue_map(set
);
3001 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3002 blk_mq_realloc_hw_ctxs(set
, q
);
3003 blk_mq_queue_reinit(q
);
3006 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3007 blk_mq_unfreeze_queue(q
);
3010 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
3012 mutex_lock(&set
->tag_list_lock
);
3013 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
3014 mutex_unlock(&set
->tag_list_lock
);
3016 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
3018 /* Enable polling stats and return whether they were already enabled. */
3019 static bool blk_poll_stats_enable(struct request_queue
*q
)
3021 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3022 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
3024 blk_stat_add_callback(q
, q
->poll_cb
);
3028 static void blk_mq_poll_stats_start(struct request_queue
*q
)
3031 * We don't arm the callback if polling stats are not enabled or the
3032 * callback is already active.
3034 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3035 blk_stat_is_active(q
->poll_cb
))
3038 blk_stat_activate_msecs(q
->poll_cb
, 100);
3041 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
3043 struct request_queue
*q
= cb
->data
;
3046 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
3047 if (cb
->stat
[bucket
].nr_samples
)
3048 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
3052 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
3053 struct blk_mq_hw_ctx
*hctx
,
3056 unsigned long ret
= 0;
3060 * If stats collection isn't on, don't sleep but turn it on for
3063 if (!blk_poll_stats_enable(q
))
3067 * As an optimistic guess, use half of the mean service time
3068 * for this type of request. We can (and should) make this smarter.
3069 * For instance, if the completion latencies are tight, we can
3070 * get closer than just half the mean. This is especially
3071 * important on devices where the completion latencies are longer
3072 * than ~10 usec. We do use the stats for the relevant IO size
3073 * if available which does lead to better estimates.
3075 bucket
= blk_mq_poll_stats_bkt(rq
);
3079 if (q
->poll_stat
[bucket
].nr_samples
)
3080 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
3085 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
3086 struct blk_mq_hw_ctx
*hctx
,
3089 struct hrtimer_sleeper hs
;
3090 enum hrtimer_mode mode
;
3094 if (test_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
))
3100 * -1: don't ever hybrid sleep
3101 * 0: use half of prev avg
3102 * >0: use this specific value
3104 if (q
->poll_nsec
== -1)
3106 else if (q
->poll_nsec
> 0)
3107 nsecs
= q
->poll_nsec
;
3109 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
3114 set_bit(REQ_ATOM_POLL_SLEPT
, &rq
->atomic_flags
);
3117 * This will be replaced with the stats tracking code, using
3118 * 'avg_completion_time / 2' as the pre-sleep target.
3122 mode
= HRTIMER_MODE_REL
;
3123 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
3124 hrtimer_set_expires(&hs
.timer
, kt
);
3126 hrtimer_init_sleeper(&hs
, current
);
3128 if (test_bit(REQ_ATOM_COMPLETE
, &rq
->atomic_flags
))
3130 set_current_state(TASK_UNINTERRUPTIBLE
);
3131 hrtimer_start_expires(&hs
.timer
, mode
);
3134 hrtimer_cancel(&hs
.timer
);
3135 mode
= HRTIMER_MODE_ABS
;
3136 } while (hs
.task
&& !signal_pending(current
));
3138 __set_current_state(TASK_RUNNING
);
3139 destroy_hrtimer_on_stack(&hs
.timer
);
3143 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
3145 struct request_queue
*q
= hctx
->queue
;
3149 * If we sleep, have the caller restart the poll loop to reset
3150 * the state. Like for the other success return cases, the
3151 * caller is responsible for checking if the IO completed. If
3152 * the IO isn't complete, we'll get called again and will go
3153 * straight to the busy poll loop.
3155 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
3158 hctx
->poll_considered
++;
3160 state
= current
->state
;
3161 while (!need_resched()) {
3164 hctx
->poll_invoked
++;
3166 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
3168 hctx
->poll_success
++;
3169 set_current_state(TASK_RUNNING
);
3173 if (signal_pending_state(state
, current
))
3174 set_current_state(TASK_RUNNING
);
3176 if (current
->state
== TASK_RUNNING
)
3186 static bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
3188 struct blk_mq_hw_ctx
*hctx
;
3191 if (!test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3194 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3195 if (!blk_qc_t_is_internal(cookie
))
3196 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3198 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3200 * With scheduling, if the request has completed, we'll
3201 * get a NULL return here, as we clear the sched tag when
3202 * that happens. The request still remains valid, like always,
3203 * so we should be safe with just the NULL check.
3209 return __blk_mq_poll(hctx
, rq
);
3212 static int __init
blk_mq_init(void)
3215 * See comment in block/blk.h rq_atomic_flags enum
3217 BUILD_BUG_ON((REQ_ATOM_STARTED
/ BITS_PER_BYTE
) !=
3218 (REQ_ATOM_COMPLETE
/ BITS_PER_BYTE
));
3220 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3221 blk_mq_hctx_notify_dead
);
3224 subsys_initcall(blk_mq_init
);