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 (blk_mq_rq_state(rq
) == MQ_RQ_IN_FLIGHT
) {
100 * index[0] counts the specific partition that was asked
101 * for. index[1] counts the ones that are active on the
102 * whole device, so increment that if mi->part is indeed
103 * a partition, and not a whole device.
105 if (rq
->part
== mi
->part
)
107 if (mi
->part
->partno
)
112 void blk_mq_in_flight(struct request_queue
*q
, struct hd_struct
*part
,
113 unsigned int inflight
[2])
115 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
117 inflight
[0] = inflight
[1] = 0;
118 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
121 void blk_freeze_queue_start(struct request_queue
*q
)
125 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
126 if (freeze_depth
== 1) {
127 percpu_ref_kill(&q
->q_usage_counter
);
129 blk_mq_run_hw_queues(q
, false);
132 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
134 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
136 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
138 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
140 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
141 unsigned long timeout
)
143 return wait_event_timeout(q
->mq_freeze_wq
,
144 percpu_ref_is_zero(&q
->q_usage_counter
),
147 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
150 * Guarantee no request is in use, so we can change any data structure of
151 * the queue afterward.
153 void blk_freeze_queue(struct request_queue
*q
)
156 * In the !blk_mq case we are only calling this to kill the
157 * q_usage_counter, otherwise this increases the freeze depth
158 * and waits for it to return to zero. For this reason there is
159 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
160 * exported to drivers as the only user for unfreeze is blk_mq.
162 blk_freeze_queue_start(q
);
163 blk_mq_freeze_queue_wait(q
);
166 void blk_mq_freeze_queue(struct request_queue
*q
)
169 * ...just an alias to keep freeze and unfreeze actions balanced
170 * in the blk_mq_* namespace
174 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
176 void blk_mq_unfreeze_queue(struct request_queue
*q
)
180 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
181 WARN_ON_ONCE(freeze_depth
< 0);
183 percpu_ref_reinit(&q
->q_usage_counter
);
184 wake_up_all(&q
->mq_freeze_wq
);
187 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
190 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
191 * mpt3sas driver such that this function can be removed.
193 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
197 spin_lock_irqsave(q
->queue_lock
, flags
);
198 queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
199 spin_unlock_irqrestore(q
->queue_lock
, flags
);
201 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
204 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
207 * Note: this function does not prevent that the struct request end_io()
208 * callback function is invoked. Once this function is returned, we make
209 * sure no dispatch can happen until the queue is unquiesced via
210 * blk_mq_unquiesce_queue().
212 void blk_mq_quiesce_queue(struct request_queue
*q
)
214 struct blk_mq_hw_ctx
*hctx
;
218 blk_mq_quiesce_queue_nowait(q
);
220 queue_for_each_hw_ctx(q
, hctx
, i
) {
221 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
222 synchronize_srcu(hctx
->srcu
);
229 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
232 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
235 * This function recovers queue into the state before quiescing
236 * which is done by blk_mq_quiesce_queue.
238 void blk_mq_unquiesce_queue(struct request_queue
*q
)
242 spin_lock_irqsave(q
->queue_lock
, flags
);
243 queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
244 spin_unlock_irqrestore(q
->queue_lock
, flags
);
246 /* dispatch requests which are inserted during quiescing */
247 blk_mq_run_hw_queues(q
, true);
249 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
251 void blk_mq_wake_waiters(struct request_queue
*q
)
253 struct blk_mq_hw_ctx
*hctx
;
256 queue_for_each_hw_ctx(q
, hctx
, i
)
257 if (blk_mq_hw_queue_mapped(hctx
))
258 blk_mq_tag_wakeup_all(hctx
->tags
, true);
261 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
263 return blk_mq_has_free_tags(hctx
->tags
);
265 EXPORT_SYMBOL(blk_mq_can_queue
);
267 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
268 unsigned int tag
, unsigned int op
)
270 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
271 struct request
*rq
= tags
->static_rqs
[tag
];
272 req_flags_t rq_flags
= 0;
274 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
276 rq
->internal_tag
= tag
;
278 if (blk_mq_tag_busy(data
->hctx
)) {
279 rq_flags
= RQF_MQ_INFLIGHT
;
280 atomic_inc(&data
->hctx
->nr_active
);
283 rq
->internal_tag
= -1;
284 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
287 /* csd/requeue_work/fifo_time is initialized before use */
289 rq
->mq_ctx
= data
->ctx
;
290 rq
->rq_flags
= rq_flags
;
293 if (data
->flags
& BLK_MQ_REQ_PREEMPT
)
294 rq
->rq_flags
|= RQF_PREEMPT
;
295 if (blk_queue_io_stat(data
->q
))
296 rq
->rq_flags
|= RQF_IO_STAT
;
297 INIT_LIST_HEAD(&rq
->queuelist
);
298 INIT_HLIST_NODE(&rq
->hash
);
299 RB_CLEAR_NODE(&rq
->rb_node
);
302 rq
->start_time
= jiffies
;
303 rq
->nr_phys_segments
= 0;
304 #if defined(CONFIG_BLK_DEV_INTEGRITY)
305 rq
->nr_integrity_segments
= 0;
308 /* tag was already set */
312 INIT_LIST_HEAD(&rq
->timeout_list
);
316 rq
->end_io_data
= NULL
;
319 #ifdef CONFIG_BLK_CGROUP
321 set_start_time_ns(rq
);
322 rq
->io_start_time_ns
= 0;
325 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
329 static struct request
*blk_mq_get_request(struct request_queue
*q
,
330 struct bio
*bio
, unsigned int op
,
331 struct blk_mq_alloc_data
*data
)
333 struct elevator_queue
*e
= q
->elevator
;
336 bool put_ctx_on_error
= false;
338 blk_queue_enter_live(q
);
340 if (likely(!data
->ctx
)) {
341 data
->ctx
= blk_mq_get_ctx(q
);
342 put_ctx_on_error
= true;
344 if (likely(!data
->hctx
))
345 data
->hctx
= blk_mq_map_queue(q
, data
->ctx
->cpu
);
347 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
350 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
353 * Flush requests are special and go directly to the
356 if (!op_is_flush(op
) && e
->type
->ops
.mq
.limit_depth
)
357 e
->type
->ops
.mq
.limit_depth(op
, data
);
360 tag
= blk_mq_get_tag(data
);
361 if (tag
== BLK_MQ_TAG_FAIL
) {
362 if (put_ctx_on_error
) {
363 blk_mq_put_ctx(data
->ctx
);
370 rq
= blk_mq_rq_ctx_init(data
, tag
, op
);
371 if (!op_is_flush(op
)) {
373 if (e
&& e
->type
->ops
.mq
.prepare_request
) {
374 if (e
->type
->icq_cache
&& rq_ioc(bio
))
375 blk_mq_sched_assign_ioc(rq
, bio
);
377 e
->type
->ops
.mq
.prepare_request(rq
, bio
);
378 rq
->rq_flags
|= RQF_ELVPRIV
;
381 data
->hctx
->queued
++;
385 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
386 blk_mq_req_flags_t flags
)
388 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
392 ret
= blk_queue_enter(q
, flags
);
396 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
400 return ERR_PTR(-EWOULDBLOCK
);
402 blk_mq_put_ctx(alloc_data
.ctx
);
405 rq
->__sector
= (sector_t
) -1;
406 rq
->bio
= rq
->biotail
= NULL
;
409 EXPORT_SYMBOL(blk_mq_alloc_request
);
411 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
412 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
414 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
};
420 * If the tag allocator sleeps we could get an allocation for a
421 * different hardware context. No need to complicate the low level
422 * allocator for this for the rare use case of a command tied to
425 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
426 return ERR_PTR(-EINVAL
);
428 if (hctx_idx
>= q
->nr_hw_queues
)
429 return ERR_PTR(-EIO
);
431 ret
= blk_queue_enter(q
, flags
);
436 * Check if the hardware context is actually mapped to anything.
437 * If not tell the caller that it should skip this queue.
439 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
440 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
442 return ERR_PTR(-EXDEV
);
444 cpu
= cpumask_first_and(alloc_data
.hctx
->cpumask
, cpu_online_mask
);
445 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
447 rq
= blk_mq_get_request(q
, NULL
, op
, &alloc_data
);
451 return ERR_PTR(-EWOULDBLOCK
);
455 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
457 void blk_mq_free_request(struct request
*rq
)
459 struct request_queue
*q
= rq
->q
;
460 struct elevator_queue
*e
= q
->elevator
;
461 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
462 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, ctx
->cpu
);
463 const int sched_tag
= rq
->internal_tag
;
465 if (rq
->rq_flags
& RQF_ELVPRIV
) {
466 if (e
&& e
->type
->ops
.mq
.finish_request
)
467 e
->type
->ops
.mq
.finish_request(rq
);
469 put_io_context(rq
->elv
.icq
->ioc
);
474 ctx
->rq_completed
[rq_is_sync(rq
)]++;
475 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
476 atomic_dec(&hctx
->nr_active
);
478 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
479 laptop_io_completion(q
->backing_dev_info
);
481 wbt_done(q
->rq_wb
, &rq
->issue_stat
);
484 blk_put_rl(blk_rq_rl(rq
));
486 blk_mq_rq_update_state(rq
, MQ_RQ_IDLE
);
488 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
490 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
491 blk_mq_sched_restart(hctx
);
494 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
496 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
498 blk_account_io_done(rq
);
501 wbt_done(rq
->q
->rq_wb
, &rq
->issue_stat
);
502 rq
->end_io(rq
, error
);
504 if (unlikely(blk_bidi_rq(rq
)))
505 blk_mq_free_request(rq
->next_rq
);
506 blk_mq_free_request(rq
);
509 EXPORT_SYMBOL(__blk_mq_end_request
);
511 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
513 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
515 __blk_mq_end_request(rq
, error
);
517 EXPORT_SYMBOL(blk_mq_end_request
);
519 static void __blk_mq_complete_request_remote(void *data
)
521 struct request
*rq
= data
;
523 rq
->q
->softirq_done_fn(rq
);
526 static void __blk_mq_complete_request(struct request
*rq
)
528 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
532 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
);
533 blk_mq_rq_update_state(rq
, MQ_RQ_COMPLETE
);
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
);
562 static void hctx_unlock(struct blk_mq_hw_ctx
*hctx
, int srcu_idx
)
563 __releases(hctx
->srcu
)
565 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
568 srcu_read_unlock(hctx
->srcu
, srcu_idx
);
571 static void hctx_lock(struct blk_mq_hw_ctx
*hctx
, int *srcu_idx
)
572 __acquires(hctx
->srcu
)
574 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
575 /* shut up gcc false positive */
579 *srcu_idx
= srcu_read_lock(hctx
->srcu
);
582 static void blk_mq_rq_update_aborted_gstate(struct request
*rq
, u64 gstate
)
587 * blk_mq_rq_aborted_gstate() is used from the completion path and
588 * can thus be called from irq context. u64_stats_fetch in the
589 * middle of update on the same CPU leads to lockup. Disable irq
592 local_irq_save(flags
);
593 u64_stats_update_begin(&rq
->aborted_gstate_sync
);
594 rq
->aborted_gstate
= gstate
;
595 u64_stats_update_end(&rq
->aborted_gstate_sync
);
596 local_irq_restore(flags
);
599 static u64
blk_mq_rq_aborted_gstate(struct request
*rq
)
605 start
= u64_stats_fetch_begin(&rq
->aborted_gstate_sync
);
606 aborted_gstate
= rq
->aborted_gstate
;
607 } while (u64_stats_fetch_retry(&rq
->aborted_gstate_sync
, start
));
609 return aborted_gstate
;
613 * blk_mq_complete_request - end I/O on a request
614 * @rq: the request being processed
617 * Ends all I/O on a request. It does not handle partial completions.
618 * The actual completion happens out-of-order, through a IPI handler.
620 void blk_mq_complete_request(struct request
*rq
)
622 struct request_queue
*q
= rq
->q
;
623 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(q
, rq
->mq_ctx
->cpu
);
626 if (unlikely(blk_should_fake_timeout(q
)))
630 * If @rq->aborted_gstate equals the current instance, timeout is
631 * claiming @rq and we lost. This is synchronized through
632 * hctx_lock(). See blk_mq_timeout_work() for details.
634 * Completion path never blocks and we can directly use RCU here
635 * instead of hctx_lock() which can be either RCU or SRCU.
636 * However, that would complicate paths which want to synchronize
637 * against us. Let stay in sync with the issue path so that
638 * hctx_lock() covers both issue and completion paths.
640 hctx_lock(hctx
, &srcu_idx
);
641 if (blk_mq_rq_aborted_gstate(rq
) != rq
->gstate
)
642 __blk_mq_complete_request(rq
);
643 hctx_unlock(hctx
, srcu_idx
);
645 EXPORT_SYMBOL(blk_mq_complete_request
);
647 int blk_mq_request_started(struct request
*rq
)
649 return blk_mq_rq_state(rq
) != MQ_RQ_IDLE
;
651 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
653 void blk_mq_start_request(struct request
*rq
)
655 struct request_queue
*q
= rq
->q
;
657 blk_mq_sched_started_request(rq
);
659 trace_block_rq_issue(q
, rq
);
661 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
662 blk_stat_set_issue(&rq
->issue_stat
, blk_rq_sectors(rq
));
663 rq
->rq_flags
|= RQF_STATS
;
664 wbt_issue(q
->rq_wb
, &rq
->issue_stat
);
667 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
670 * Mark @rq in-flight which also advances the generation number,
671 * and register for timeout. Protect with a seqcount to allow the
672 * timeout path to read both @rq->gstate and @rq->deadline
675 * This is the only place where a request is marked in-flight. If
676 * the timeout path reads an in-flight @rq->gstate, the
677 * @rq->deadline it reads together under @rq->gstate_seq is
678 * guaranteed to be the matching one.
681 write_seqcount_begin(&rq
->gstate_seq
);
683 blk_mq_rq_update_state(rq
, MQ_RQ_IN_FLIGHT
);
686 write_seqcount_end(&rq
->gstate_seq
);
689 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
691 * Make sure space for the drain appears. We know we can do
692 * this because max_hw_segments has been adjusted to be one
693 * fewer than the device can handle.
695 rq
->nr_phys_segments
++;
698 EXPORT_SYMBOL(blk_mq_start_request
);
701 * When we reach here because queue is busy, it's safe to change the state
702 * to IDLE without checking @rq->aborted_gstate because we should still be
703 * holding the RCU read lock and thus protected against timeout.
705 static void __blk_mq_requeue_request(struct request
*rq
)
707 struct request_queue
*q
= rq
->q
;
709 blk_mq_put_driver_tag(rq
);
711 trace_block_rq_requeue(q
, rq
);
712 wbt_requeue(q
->rq_wb
, &rq
->issue_stat
);
713 blk_mq_sched_requeue_request(rq
);
715 if (blk_mq_rq_state(rq
) != MQ_RQ_IDLE
) {
716 blk_mq_rq_update_state(rq
, MQ_RQ_IDLE
);
717 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
718 rq
->nr_phys_segments
--;
722 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
724 __blk_mq_requeue_request(rq
);
726 BUG_ON(blk_queued_rq(rq
));
727 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
729 EXPORT_SYMBOL(blk_mq_requeue_request
);
731 static void blk_mq_requeue_work(struct work_struct
*work
)
733 struct request_queue
*q
=
734 container_of(work
, struct request_queue
, requeue_work
.work
);
736 struct request
*rq
, *next
;
738 spin_lock_irq(&q
->requeue_lock
);
739 list_splice_init(&q
->requeue_list
, &rq_list
);
740 spin_unlock_irq(&q
->requeue_lock
);
742 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
743 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
746 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
747 list_del_init(&rq
->queuelist
);
748 blk_mq_sched_insert_request(rq
, true, false, false, true);
751 while (!list_empty(&rq_list
)) {
752 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
753 list_del_init(&rq
->queuelist
);
754 blk_mq_sched_insert_request(rq
, false, false, false, true);
757 blk_mq_run_hw_queues(q
, false);
760 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
761 bool kick_requeue_list
)
763 struct request_queue
*q
= rq
->q
;
767 * We abuse this flag that is otherwise used by the I/O scheduler to
768 * request head insertion from the workqueue.
770 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
772 spin_lock_irqsave(&q
->requeue_lock
, flags
);
774 rq
->rq_flags
|= RQF_SOFTBARRIER
;
775 list_add(&rq
->queuelist
, &q
->requeue_list
);
777 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
779 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
781 if (kick_requeue_list
)
782 blk_mq_kick_requeue_list(q
);
784 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
786 void blk_mq_kick_requeue_list(struct request_queue
*q
)
788 kblockd_schedule_delayed_work(&q
->requeue_work
, 0);
790 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
792 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
795 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
796 msecs_to_jiffies(msecs
));
798 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
800 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
802 if (tag
< tags
->nr_tags
) {
803 prefetch(tags
->rqs
[tag
]);
804 return tags
->rqs
[tag
];
809 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
811 struct blk_mq_timeout_data
{
813 unsigned int next_set
;
814 unsigned int nr_expired
;
817 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
819 const struct blk_mq_ops
*ops
= req
->q
->mq_ops
;
820 enum blk_eh_timer_return ret
= BLK_EH_RESET_TIMER
;
822 req
->rq_flags
|= RQF_MQ_TIMEOUT_EXPIRED
;
825 ret
= ops
->timeout(req
, reserved
);
829 __blk_mq_complete_request(req
);
831 case BLK_EH_RESET_TIMER
:
833 * As nothing prevents from completion happening while
834 * ->aborted_gstate is set, this may lead to ignored
835 * completions and further spurious timeouts.
837 blk_mq_rq_update_aborted_gstate(req
, 0);
840 case BLK_EH_NOT_HANDLED
:
843 printk(KERN_ERR
"block: bad eh return: %d\n", ret
);
848 static void blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
849 struct request
*rq
, void *priv
, bool reserved
)
851 struct blk_mq_timeout_data
*data
= priv
;
852 unsigned long gstate
, deadline
;
857 if (rq
->rq_flags
& RQF_MQ_TIMEOUT_EXPIRED
)
860 /* read coherent snapshots of @rq->state_gen and @rq->deadline */
862 start
= read_seqcount_begin(&rq
->gstate_seq
);
863 gstate
= READ_ONCE(rq
->gstate
);
864 deadline
= blk_rq_deadline(rq
);
865 if (!read_seqcount_retry(&rq
->gstate_seq
, start
))
870 /* if in-flight && overdue, mark for abortion */
871 if ((gstate
& MQ_RQ_STATE_MASK
) == MQ_RQ_IN_FLIGHT
&&
872 time_after_eq(jiffies
, deadline
)) {
873 blk_mq_rq_update_aborted_gstate(rq
, gstate
);
876 } else if (!data
->next_set
|| time_after(data
->next
, deadline
)) {
877 data
->next
= deadline
;
882 static void blk_mq_terminate_expired(struct blk_mq_hw_ctx
*hctx
,
883 struct request
*rq
, void *priv
, bool reserved
)
886 * We marked @rq->aborted_gstate and waited for RCU. If there were
887 * completions that we lost to, they would have finished and
888 * updated @rq->gstate by now; otherwise, the completion path is
889 * now guaranteed to see @rq->aborted_gstate and yield. If
890 * @rq->aborted_gstate still matches @rq->gstate, @rq is ours.
892 if (!(rq
->rq_flags
& RQF_MQ_TIMEOUT_EXPIRED
) &&
893 READ_ONCE(rq
->gstate
) == rq
->aborted_gstate
)
894 blk_mq_rq_timed_out(rq
, reserved
);
897 static void blk_mq_timeout_work(struct work_struct
*work
)
899 struct request_queue
*q
=
900 container_of(work
, struct request_queue
, timeout_work
);
901 struct blk_mq_timeout_data data
= {
906 struct blk_mq_hw_ctx
*hctx
;
909 /* A deadlock might occur if a request is stuck requiring a
910 * timeout at the same time a queue freeze is waiting
911 * completion, since the timeout code would not be able to
912 * acquire the queue reference here.
914 * That's why we don't use blk_queue_enter here; instead, we use
915 * percpu_ref_tryget directly, because we need to be able to
916 * obtain a reference even in the short window between the queue
917 * starting to freeze, by dropping the first reference in
918 * blk_freeze_queue_start, and the moment the last request is
919 * consumed, marked by the instant q_usage_counter reaches
922 if (!percpu_ref_tryget(&q
->q_usage_counter
))
925 /* scan for the expired ones and set their ->aborted_gstate */
926 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &data
);
928 if (data
.nr_expired
) {
929 bool has_rcu
= false;
932 * Wait till everyone sees ->aborted_gstate. The
933 * sequential waits for SRCUs aren't ideal. If this ever
934 * becomes a problem, we can add per-hw_ctx rcu_head and
937 queue_for_each_hw_ctx(q
, hctx
, i
) {
938 if (!hctx
->nr_expired
)
941 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
944 synchronize_srcu(hctx
->srcu
);
946 hctx
->nr_expired
= 0;
951 /* terminate the ones we won */
952 blk_mq_queue_tag_busy_iter(q
, blk_mq_terminate_expired
, NULL
);
956 data
.next
= blk_rq_timeout(round_jiffies_up(data
.next
));
957 mod_timer(&q
->timeout
, data
.next
);
960 * Request timeouts are handled as a forward rolling timer. If
961 * we end up here it means that no requests are pending and
962 * also that no request has been pending for a while. Mark
965 queue_for_each_hw_ctx(q
, hctx
, i
) {
966 /* the hctx may be unmapped, so check it here */
967 if (blk_mq_hw_queue_mapped(hctx
))
968 blk_mq_tag_idle(hctx
);
974 struct flush_busy_ctx_data
{
975 struct blk_mq_hw_ctx
*hctx
;
976 struct list_head
*list
;
979 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
981 struct flush_busy_ctx_data
*flush_data
= data
;
982 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
983 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
985 sbitmap_clear_bit(sb
, bitnr
);
986 spin_lock(&ctx
->lock
);
987 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
988 spin_unlock(&ctx
->lock
);
993 * Process software queues that have been marked busy, splicing them
994 * to the for-dispatch
996 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
998 struct flush_busy_ctx_data data
= {
1003 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
1005 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
1007 struct dispatch_rq_data
{
1008 struct blk_mq_hw_ctx
*hctx
;
1012 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
1015 struct dispatch_rq_data
*dispatch_data
= data
;
1016 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
1017 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1019 spin_lock(&ctx
->lock
);
1020 if (unlikely(!list_empty(&ctx
->rq_list
))) {
1021 dispatch_data
->rq
= list_entry_rq(ctx
->rq_list
.next
);
1022 list_del_init(&dispatch_data
->rq
->queuelist
);
1023 if (list_empty(&ctx
->rq_list
))
1024 sbitmap_clear_bit(sb
, bitnr
);
1026 spin_unlock(&ctx
->lock
);
1028 return !dispatch_data
->rq
;
1031 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1032 struct blk_mq_ctx
*start
)
1034 unsigned off
= start
? start
->index_hw
: 0;
1035 struct dispatch_rq_data data
= {
1040 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1041 dispatch_rq_from_ctx
, &data
);
1046 static inline unsigned int queued_to_index(unsigned int queued
)
1051 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
1054 bool blk_mq_get_driver_tag(struct request
*rq
, struct blk_mq_hw_ctx
**hctx
,
1057 struct blk_mq_alloc_data data
= {
1059 .hctx
= blk_mq_map_queue(rq
->q
, rq
->mq_ctx
->cpu
),
1060 .flags
= wait
? 0 : BLK_MQ_REQ_NOWAIT
,
1063 might_sleep_if(wait
);
1068 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
1069 data
.flags
|= BLK_MQ_REQ_RESERVED
;
1071 rq
->tag
= blk_mq_get_tag(&data
);
1073 if (blk_mq_tag_busy(data
.hctx
)) {
1074 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1075 atomic_inc(&data
.hctx
->nr_active
);
1077 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
1083 return rq
->tag
!= -1;
1086 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1087 int flags
, void *key
)
1089 struct blk_mq_hw_ctx
*hctx
;
1091 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1093 list_del_init(&wait
->entry
);
1094 blk_mq_run_hw_queue(hctx
, true);
1099 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1100 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1101 * restart. For both cases, take care to check the condition again after
1102 * marking us as waiting.
1104 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
**hctx
,
1107 struct blk_mq_hw_ctx
*this_hctx
= *hctx
;
1108 struct sbq_wait_state
*ws
;
1109 wait_queue_entry_t
*wait
;
1112 if (!(this_hctx
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1113 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &this_hctx
->state
))
1114 set_bit(BLK_MQ_S_SCHED_RESTART
, &this_hctx
->state
);
1117 * It's possible that a tag was freed in the window between the
1118 * allocation failure and adding the hardware queue to the wait
1121 * Don't clear RESTART here, someone else could have set it.
1122 * At most this will cost an extra queue run.
1124 return blk_mq_get_driver_tag(rq
, hctx
, false);
1127 wait
= &this_hctx
->dispatch_wait
;
1128 if (!list_empty_careful(&wait
->entry
))
1131 spin_lock(&this_hctx
->lock
);
1132 if (!list_empty(&wait
->entry
)) {
1133 spin_unlock(&this_hctx
->lock
);
1137 ws
= bt_wait_ptr(&this_hctx
->tags
->bitmap_tags
, this_hctx
);
1138 add_wait_queue(&ws
->wait
, wait
);
1141 * It's possible that a tag was freed in the window between the
1142 * allocation failure and adding the hardware queue to the wait
1145 ret
= blk_mq_get_driver_tag(rq
, hctx
, false);
1147 spin_unlock(&this_hctx
->lock
);
1152 * We got a tag, remove ourselves from the wait queue to ensure
1153 * someone else gets the wakeup.
1155 spin_lock_irq(&ws
->wait
.lock
);
1156 list_del_init(&wait
->entry
);
1157 spin_unlock_irq(&ws
->wait
.lock
);
1158 spin_unlock(&this_hctx
->lock
);
1163 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
,
1166 struct blk_mq_hw_ctx
*hctx
;
1167 struct request
*rq
, *nxt
;
1168 bool no_tag
= false;
1171 if (list_empty(list
))
1174 WARN_ON(!list_is_singular(list
) && got_budget
);
1177 * Now process all the entries, sending them to the driver.
1179 errors
= queued
= 0;
1181 struct blk_mq_queue_data bd
;
1184 rq
= list_first_entry(list
, struct request
, queuelist
);
1185 if (!blk_mq_get_driver_tag(rq
, &hctx
, 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
)) {
1195 blk_mq_put_dispatch_budget(hctx
);
1197 * For non-shared tags, the RESTART check
1200 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
)
1206 if (!got_budget
&& !blk_mq_get_dispatch_budget(hctx
)) {
1207 blk_mq_put_driver_tag(rq
);
1211 list_del_init(&rq
->queuelist
);
1216 * Flag last if we have no more requests, or if we have more
1217 * but can't assign a driver tag to it.
1219 if (list_empty(list
))
1222 nxt
= list_first_entry(list
, struct request
, queuelist
);
1223 bd
.last
= !blk_mq_get_driver_tag(nxt
, NULL
, false);
1226 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1227 if (ret
== BLK_STS_RESOURCE
) {
1229 * If an I/O scheduler has been configured and we got a
1230 * driver tag for the next request already, free it
1233 if (!list_empty(list
)) {
1234 nxt
= list_first_entry(list
, struct request
, queuelist
);
1235 blk_mq_put_driver_tag(nxt
);
1237 list_add(&rq
->queuelist
, list
);
1238 __blk_mq_requeue_request(rq
);
1242 if (unlikely(ret
!= BLK_STS_OK
)) {
1244 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1249 } while (!list_empty(list
));
1251 hctx
->dispatched
[queued_to_index(queued
)]++;
1254 * Any items that need requeuing? Stuff them into hctx->dispatch,
1255 * that is where we will continue on next queue run.
1257 if (!list_empty(list
)) {
1258 spin_lock(&hctx
->lock
);
1259 list_splice_init(list
, &hctx
->dispatch
);
1260 spin_unlock(&hctx
->lock
);
1263 * If SCHED_RESTART was set by the caller of this function and
1264 * it is no longer set that means that it was cleared by another
1265 * thread and hence that a queue rerun is needed.
1267 * If 'no_tag' is set, that means that we failed getting
1268 * a driver tag with an I/O scheduler attached. If our dispatch
1269 * waitqueue is no longer active, ensure that we run the queue
1270 * AFTER adding our entries back to the list.
1272 * If no I/O scheduler has been configured it is possible that
1273 * the hardware queue got stopped and restarted before requests
1274 * were pushed back onto the dispatch list. Rerun the queue to
1275 * avoid starvation. Notes:
1276 * - blk_mq_run_hw_queue() checks whether or not a queue has
1277 * been stopped before rerunning a queue.
1278 * - Some but not all block drivers stop a queue before
1279 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1282 if (!blk_mq_sched_needs_restart(hctx
) ||
1283 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1284 blk_mq_run_hw_queue(hctx
, true);
1287 return (queued
+ errors
) != 0;
1290 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1295 * We should be running this queue from one of the CPUs that
1298 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1299 cpu_online(hctx
->next_cpu
));
1302 * We can't run the queue inline with ints disabled. Ensure that
1303 * we catch bad users of this early.
1305 WARN_ON_ONCE(in_interrupt());
1307 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1309 hctx_lock(hctx
, &srcu_idx
);
1310 blk_mq_sched_dispatch_requests(hctx
);
1311 hctx_unlock(hctx
, srcu_idx
);
1315 * It'd be great if the workqueue API had a way to pass
1316 * in a mask and had some smarts for more clever placement.
1317 * For now we just round-robin here, switching for every
1318 * BLK_MQ_CPU_WORK_BATCH queued items.
1320 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1322 if (hctx
->queue
->nr_hw_queues
== 1)
1323 return WORK_CPU_UNBOUND
;
1325 if (--hctx
->next_cpu_batch
<= 0) {
1328 next_cpu
= cpumask_next_and(hctx
->next_cpu
, hctx
->cpumask
,
1330 if (next_cpu
>= nr_cpu_ids
)
1331 next_cpu
= cpumask_first_and(hctx
->cpumask
,cpu_online_mask
);
1333 hctx
->next_cpu
= next_cpu
;
1334 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1337 return hctx
->next_cpu
;
1340 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1341 unsigned long msecs
)
1343 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx
)))
1346 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1349 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1350 int cpu
= get_cpu();
1351 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1352 __blk_mq_run_hw_queue(hctx
);
1360 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1362 msecs_to_jiffies(msecs
));
1365 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1367 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1369 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1371 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1377 * When queue is quiesced, we may be switching io scheduler, or
1378 * updating nr_hw_queues, or other things, and we can't run queue
1379 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1381 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1384 hctx_lock(hctx
, &srcu_idx
);
1385 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1386 blk_mq_hctx_has_pending(hctx
);
1387 hctx_unlock(hctx
, srcu_idx
);
1390 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1396 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1398 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1400 struct blk_mq_hw_ctx
*hctx
;
1403 queue_for_each_hw_ctx(q
, hctx
, i
) {
1404 if (blk_mq_hctx_stopped(hctx
))
1407 blk_mq_run_hw_queue(hctx
, async
);
1410 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1413 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1414 * @q: request queue.
1416 * The caller is responsible for serializing this function against
1417 * blk_mq_{start,stop}_hw_queue().
1419 bool blk_mq_queue_stopped(struct request_queue
*q
)
1421 struct blk_mq_hw_ctx
*hctx
;
1424 queue_for_each_hw_ctx(q
, hctx
, i
)
1425 if (blk_mq_hctx_stopped(hctx
))
1430 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1433 * This function is often used for pausing .queue_rq() by driver when
1434 * there isn't enough resource or some conditions aren't satisfied, and
1435 * BLK_STS_RESOURCE is usually returned.
1437 * We do not guarantee that dispatch can be drained or blocked
1438 * after blk_mq_stop_hw_queue() returns. Please use
1439 * blk_mq_quiesce_queue() for that requirement.
1441 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1443 cancel_delayed_work(&hctx
->run_work
);
1445 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1447 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1450 * This function is often used for pausing .queue_rq() by driver when
1451 * there isn't enough resource or some conditions aren't satisfied, and
1452 * BLK_STS_RESOURCE is usually returned.
1454 * We do not guarantee that dispatch can be drained or blocked
1455 * after blk_mq_stop_hw_queues() returns. Please use
1456 * blk_mq_quiesce_queue() for that requirement.
1458 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1460 struct blk_mq_hw_ctx
*hctx
;
1463 queue_for_each_hw_ctx(q
, hctx
, i
)
1464 blk_mq_stop_hw_queue(hctx
);
1466 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1468 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1470 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1472 blk_mq_run_hw_queue(hctx
, false);
1474 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1476 void blk_mq_start_hw_queues(struct request_queue
*q
)
1478 struct blk_mq_hw_ctx
*hctx
;
1481 queue_for_each_hw_ctx(q
, hctx
, i
)
1482 blk_mq_start_hw_queue(hctx
);
1484 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1486 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1488 if (!blk_mq_hctx_stopped(hctx
))
1491 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1492 blk_mq_run_hw_queue(hctx
, async
);
1494 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1496 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1498 struct blk_mq_hw_ctx
*hctx
;
1501 queue_for_each_hw_ctx(q
, hctx
, i
)
1502 blk_mq_start_stopped_hw_queue(hctx
, async
);
1504 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1506 static void blk_mq_run_work_fn(struct work_struct
*work
)
1508 struct blk_mq_hw_ctx
*hctx
;
1510 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1513 * If we are stopped, don't run the queue. The exception is if
1514 * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1515 * the STOPPED bit and run it.
1517 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
)) {
1518 if (!test_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
))
1521 clear_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1522 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1525 __blk_mq_run_hw_queue(hctx
);
1529 void blk_mq_delay_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1531 if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx
)))
1535 * Stop the hw queue, then modify currently delayed work.
1536 * This should prevent us from running the queue prematurely.
1537 * Mark the queue as auto-clearing STOPPED when it runs.
1539 blk_mq_stop_hw_queue(hctx
);
1540 set_bit(BLK_MQ_S_START_ON_RUN
, &hctx
->state
);
1541 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
),
1543 msecs_to_jiffies(msecs
));
1545 EXPORT_SYMBOL(blk_mq_delay_queue
);
1547 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1551 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1553 lockdep_assert_held(&ctx
->lock
);
1555 trace_block_rq_insert(hctx
->queue
, rq
);
1558 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1560 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1563 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1566 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1568 lockdep_assert_held(&ctx
->lock
);
1570 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1571 blk_mq_hctx_mark_pending(hctx
, ctx
);
1575 * Should only be used carefully, when the caller knows we want to
1576 * bypass a potential IO scheduler on the target device.
1578 void blk_mq_request_bypass_insert(struct request
*rq
, bool run_queue
)
1580 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1581 struct blk_mq_hw_ctx
*hctx
= blk_mq_map_queue(rq
->q
, ctx
->cpu
);
1583 spin_lock(&hctx
->lock
);
1584 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1585 spin_unlock(&hctx
->lock
);
1588 blk_mq_run_hw_queue(hctx
, false);
1591 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1592 struct list_head
*list
)
1596 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1599 spin_lock(&ctx
->lock
);
1600 while (!list_empty(list
)) {
1603 rq
= list_first_entry(list
, struct request
, queuelist
);
1604 BUG_ON(rq
->mq_ctx
!= ctx
);
1605 list_del_init(&rq
->queuelist
);
1606 __blk_mq_insert_req_list(hctx
, rq
, false);
1608 blk_mq_hctx_mark_pending(hctx
, ctx
);
1609 spin_unlock(&ctx
->lock
);
1612 static int plug_ctx_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1614 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1615 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1617 return !(rqa
->mq_ctx
< rqb
->mq_ctx
||
1618 (rqa
->mq_ctx
== rqb
->mq_ctx
&&
1619 blk_rq_pos(rqa
) < blk_rq_pos(rqb
)));
1622 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1624 struct blk_mq_ctx
*this_ctx
;
1625 struct request_queue
*this_q
;
1628 LIST_HEAD(ctx_list
);
1631 list_splice_init(&plug
->mq_list
, &list
);
1633 list_sort(NULL
, &list
, plug_ctx_cmp
);
1639 while (!list_empty(&list
)) {
1640 rq
= list_entry_rq(list
.next
);
1641 list_del_init(&rq
->queuelist
);
1643 if (rq
->mq_ctx
!= this_ctx
) {
1645 trace_block_unplug(this_q
, depth
, from_schedule
);
1646 blk_mq_sched_insert_requests(this_q
, this_ctx
,
1651 this_ctx
= rq
->mq_ctx
;
1657 list_add_tail(&rq
->queuelist
, &ctx_list
);
1661 * If 'this_ctx' is set, we know we have entries to complete
1662 * on 'ctx_list'. Do those.
1665 trace_block_unplug(this_q
, depth
, from_schedule
);
1666 blk_mq_sched_insert_requests(this_q
, this_ctx
, &ctx_list
,
1671 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1673 blk_init_request_from_bio(rq
, bio
);
1675 blk_rq_set_rl(rq
, blk_get_rl(rq
->q
, bio
));
1677 blk_account_io_start(rq
, true);
1680 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx
*hctx
,
1681 struct blk_mq_ctx
*ctx
,
1684 spin_lock(&ctx
->lock
);
1685 __blk_mq_insert_request(hctx
, rq
, false);
1686 spin_unlock(&ctx
->lock
);
1689 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1692 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1694 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1697 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1701 struct request_queue
*q
= rq
->q
;
1702 struct blk_mq_queue_data bd
= {
1706 blk_qc_t new_cookie
;
1708 bool run_queue
= true;
1710 /* RCU or SRCU read lock is needed before checking quiesced flag */
1711 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1719 if (!blk_mq_get_driver_tag(rq
, NULL
, false))
1722 if (!blk_mq_get_dispatch_budget(hctx
)) {
1723 blk_mq_put_driver_tag(rq
);
1727 new_cookie
= request_to_qc_t(hctx
, rq
);
1730 * For OK queue, we are done. For error, kill it. Any other
1731 * error (busy), just add it to our list as we previously
1734 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1737 *cookie
= new_cookie
;
1739 case BLK_STS_RESOURCE
:
1740 __blk_mq_requeue_request(rq
);
1743 *cookie
= BLK_QC_T_NONE
;
1744 blk_mq_end_request(rq
, ret
);
1749 blk_mq_sched_insert_request(rq
, false, run_queue
, false,
1750 hctx
->flags
& BLK_MQ_F_BLOCKING
);
1753 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1754 struct request
*rq
, blk_qc_t
*cookie
)
1758 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1760 hctx_lock(hctx
, &srcu_idx
);
1761 __blk_mq_try_issue_directly(hctx
, rq
, cookie
);
1762 hctx_unlock(hctx
, srcu_idx
);
1765 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1767 const int is_sync
= op_is_sync(bio
->bi_opf
);
1768 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1769 struct blk_mq_alloc_data data
= { .flags
= 0 };
1771 unsigned int request_count
= 0;
1772 struct blk_plug
*plug
;
1773 struct request
*same_queue_rq
= NULL
;
1775 unsigned int wb_acct
;
1777 blk_queue_bounce(q
, &bio
);
1779 blk_queue_split(q
, &bio
);
1781 if (!bio_integrity_prep(bio
))
1782 return BLK_QC_T_NONE
;
1784 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1785 blk_attempt_plug_merge(q
, bio
, &request_count
, &same_queue_rq
))
1786 return BLK_QC_T_NONE
;
1788 if (blk_mq_sched_bio_merge(q
, bio
))
1789 return BLK_QC_T_NONE
;
1791 wb_acct
= wbt_wait(q
->rq_wb
, bio
, NULL
);
1793 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1795 rq
= blk_mq_get_request(q
, bio
, bio
->bi_opf
, &data
);
1796 if (unlikely(!rq
)) {
1797 __wbt_done(q
->rq_wb
, wb_acct
);
1798 if (bio
->bi_opf
& REQ_NOWAIT
)
1799 bio_wouldblock_error(bio
);
1800 return BLK_QC_T_NONE
;
1803 wbt_track(&rq
->issue_stat
, wb_acct
);
1805 cookie
= request_to_qc_t(data
.hctx
, rq
);
1807 plug
= current
->plug
;
1808 if (unlikely(is_flush_fua
)) {
1809 blk_mq_put_ctx(data
.ctx
);
1810 blk_mq_bio_to_request(rq
, bio
);
1812 /* bypass scheduler for flush rq */
1813 blk_insert_flush(rq
);
1814 blk_mq_run_hw_queue(data
.hctx
, true);
1815 } else if (plug
&& q
->nr_hw_queues
== 1) {
1816 struct request
*last
= NULL
;
1818 blk_mq_put_ctx(data
.ctx
);
1819 blk_mq_bio_to_request(rq
, bio
);
1822 * @request_count may become stale because of schedule
1823 * out, so check the list again.
1825 if (list_empty(&plug
->mq_list
))
1827 else if (blk_queue_nomerges(q
))
1828 request_count
= blk_plug_queued_count(q
);
1831 trace_block_plug(q
);
1833 last
= list_entry_rq(plug
->mq_list
.prev
);
1835 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1836 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1837 blk_flush_plug_list(plug
, false);
1838 trace_block_plug(q
);
1841 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1842 } else if (plug
&& !blk_queue_nomerges(q
)) {
1843 blk_mq_bio_to_request(rq
, bio
);
1846 * We do limited plugging. If the bio can be merged, do that.
1847 * Otherwise the existing request in the plug list will be
1848 * issued. So the plug list will have one request at most
1849 * The plug list might get flushed before this. If that happens,
1850 * the plug list is empty, and same_queue_rq is invalid.
1852 if (list_empty(&plug
->mq_list
))
1853 same_queue_rq
= NULL
;
1855 list_del_init(&same_queue_rq
->queuelist
);
1856 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1858 blk_mq_put_ctx(data
.ctx
);
1860 if (same_queue_rq
) {
1861 data
.hctx
= blk_mq_map_queue(q
,
1862 same_queue_rq
->mq_ctx
->cpu
);
1863 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1866 } else if (q
->nr_hw_queues
> 1 && is_sync
) {
1867 blk_mq_put_ctx(data
.ctx
);
1868 blk_mq_bio_to_request(rq
, bio
);
1869 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1870 } else if (q
->elevator
) {
1871 blk_mq_put_ctx(data
.ctx
);
1872 blk_mq_bio_to_request(rq
, bio
);
1873 blk_mq_sched_insert_request(rq
, false, true, true, true);
1875 blk_mq_put_ctx(data
.ctx
);
1876 blk_mq_bio_to_request(rq
, bio
);
1877 blk_mq_queue_io(data
.hctx
, data
.ctx
, rq
);
1878 blk_mq_run_hw_queue(data
.hctx
, true);
1884 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1885 unsigned int hctx_idx
)
1889 if (tags
->rqs
&& set
->ops
->exit_request
) {
1892 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1893 struct request
*rq
= tags
->static_rqs
[i
];
1897 set
->ops
->exit_request(set
, rq
, hctx_idx
);
1898 tags
->static_rqs
[i
] = NULL
;
1902 while (!list_empty(&tags
->page_list
)) {
1903 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1904 list_del_init(&page
->lru
);
1906 * Remove kmemleak object previously allocated in
1907 * blk_mq_init_rq_map().
1909 kmemleak_free(page_address(page
));
1910 __free_pages(page
, page
->private);
1914 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
1918 kfree(tags
->static_rqs
);
1919 tags
->static_rqs
= NULL
;
1921 blk_mq_free_tags(tags
);
1924 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
1925 unsigned int hctx_idx
,
1926 unsigned int nr_tags
,
1927 unsigned int reserved_tags
)
1929 struct blk_mq_tags
*tags
;
1932 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1933 if (node
== NUMA_NO_NODE
)
1934 node
= set
->numa_node
;
1936 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
1937 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
1941 tags
->rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1942 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1945 blk_mq_free_tags(tags
);
1949 tags
->static_rqs
= kzalloc_node(nr_tags
* sizeof(struct request
*),
1950 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
1952 if (!tags
->static_rqs
) {
1954 blk_mq_free_tags(tags
);
1961 static size_t order_to_size(unsigned int order
)
1963 return (size_t)PAGE_SIZE
<< order
;
1966 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
1967 unsigned int hctx_idx
, int node
)
1971 if (set
->ops
->init_request
) {
1972 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
1977 seqcount_init(&rq
->gstate_seq
);
1978 u64_stats_init(&rq
->aborted_gstate_sync
);
1982 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1983 unsigned int hctx_idx
, unsigned int depth
)
1985 unsigned int i
, j
, entries_per_page
, max_order
= 4;
1986 size_t rq_size
, left
;
1989 node
= blk_mq_hw_queue_to_node(set
->mq_map
, hctx_idx
);
1990 if (node
== NUMA_NO_NODE
)
1991 node
= set
->numa_node
;
1993 INIT_LIST_HEAD(&tags
->page_list
);
1996 * rq_size is the size of the request plus driver payload, rounded
1997 * to the cacheline size
1999 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
2001 left
= rq_size
* depth
;
2003 for (i
= 0; i
< depth
; ) {
2004 int this_order
= max_order
;
2009 while (this_order
&& left
< order_to_size(this_order
- 1))
2013 page
= alloc_pages_node(node
,
2014 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
2020 if (order_to_size(this_order
) < rq_size
)
2027 page
->private = this_order
;
2028 list_add_tail(&page
->lru
, &tags
->page_list
);
2030 p
= page_address(page
);
2032 * Allow kmemleak to scan these pages as they contain pointers
2033 * to additional allocations like via ops->init_request().
2035 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2036 entries_per_page
= order_to_size(this_order
) / rq_size
;
2037 to_do
= min(entries_per_page
, depth
- i
);
2038 left
-= to_do
* rq_size
;
2039 for (j
= 0; j
< to_do
; j
++) {
2040 struct request
*rq
= p
;
2042 tags
->static_rqs
[i
] = rq
;
2043 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
2044 tags
->static_rqs
[i
] = NULL
;
2055 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2060 * 'cpu' is going away. splice any existing rq_list entries from this
2061 * software queue to the hw queue dispatch list, and ensure that it
2064 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2066 struct blk_mq_hw_ctx
*hctx
;
2067 struct blk_mq_ctx
*ctx
;
2070 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2071 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2073 spin_lock(&ctx
->lock
);
2074 if (!list_empty(&ctx
->rq_list
)) {
2075 list_splice_init(&ctx
->rq_list
, &tmp
);
2076 blk_mq_hctx_clear_pending(hctx
, ctx
);
2078 spin_unlock(&ctx
->lock
);
2080 if (list_empty(&tmp
))
2083 spin_lock(&hctx
->lock
);
2084 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2085 spin_unlock(&hctx
->lock
);
2087 blk_mq_run_hw_queue(hctx
, true);
2091 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2093 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2097 /* hctx->ctxs will be freed in queue's release handler */
2098 static void blk_mq_exit_hctx(struct request_queue
*q
,
2099 struct blk_mq_tag_set
*set
,
2100 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2102 blk_mq_debugfs_unregister_hctx(hctx
);
2104 if (blk_mq_hw_queue_mapped(hctx
))
2105 blk_mq_tag_idle(hctx
);
2107 if (set
->ops
->exit_request
)
2108 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2110 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2112 if (set
->ops
->exit_hctx
)
2113 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2115 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2116 cleanup_srcu_struct(hctx
->srcu
);
2118 blk_mq_remove_cpuhp(hctx
);
2119 blk_free_flush_queue(hctx
->fq
);
2120 sbitmap_free(&hctx
->ctx_map
);
2123 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2124 struct blk_mq_tag_set
*set
, int nr_queue
)
2126 struct blk_mq_hw_ctx
*hctx
;
2129 queue_for_each_hw_ctx(q
, hctx
, i
) {
2132 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2136 static int blk_mq_init_hctx(struct request_queue
*q
,
2137 struct blk_mq_tag_set
*set
,
2138 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2142 node
= hctx
->numa_node
;
2143 if (node
== NUMA_NO_NODE
)
2144 node
= hctx
->numa_node
= set
->numa_node
;
2146 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2147 spin_lock_init(&hctx
->lock
);
2148 INIT_LIST_HEAD(&hctx
->dispatch
);
2150 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
2152 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2154 hctx
->tags
= set
->tags
[hctx_idx
];
2157 * Allocate space for all possible cpus to avoid allocation at
2160 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2163 goto unregister_cpu_notifier
;
2165 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8), GFP_KERNEL
,
2171 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2172 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2174 if (set
->ops
->init_hctx
&&
2175 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2178 if (blk_mq_sched_init_hctx(q
, hctx
, hctx_idx
))
2181 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
);
2183 goto sched_exit_hctx
;
2185 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
, node
))
2188 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2189 init_srcu_struct(hctx
->srcu
);
2191 blk_mq_debugfs_register_hctx(q
, hctx
);
2198 blk_mq_sched_exit_hctx(q
, hctx
, hctx_idx
);
2200 if (set
->ops
->exit_hctx
)
2201 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2203 sbitmap_free(&hctx
->ctx_map
);
2206 unregister_cpu_notifier
:
2207 blk_mq_remove_cpuhp(hctx
);
2211 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2212 unsigned int nr_hw_queues
)
2216 for_each_possible_cpu(i
) {
2217 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2218 struct blk_mq_hw_ctx
*hctx
;
2221 spin_lock_init(&__ctx
->lock
);
2222 INIT_LIST_HEAD(&__ctx
->rq_list
);
2226 * Set local node, IFF we have more than one hw queue. If
2227 * not, we remain on the home node of the device
2229 hctx
= blk_mq_map_queue(q
, i
);
2230 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2231 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2235 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2239 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2240 set
->queue_depth
, set
->reserved_tags
);
2241 if (!set
->tags
[hctx_idx
])
2244 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2249 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2250 set
->tags
[hctx_idx
] = NULL
;
2254 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2255 unsigned int hctx_idx
)
2257 if (set
->tags
[hctx_idx
]) {
2258 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2259 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2260 set
->tags
[hctx_idx
] = NULL
;
2264 static void blk_mq_map_swqueue(struct request_queue
*q
)
2266 unsigned int i
, hctx_idx
;
2267 struct blk_mq_hw_ctx
*hctx
;
2268 struct blk_mq_ctx
*ctx
;
2269 struct blk_mq_tag_set
*set
= q
->tag_set
;
2272 * Avoid others reading imcomplete hctx->cpumask through sysfs
2274 mutex_lock(&q
->sysfs_lock
);
2276 queue_for_each_hw_ctx(q
, hctx
, i
) {
2277 cpumask_clear(hctx
->cpumask
);
2282 * Map software to hardware queues.
2284 * If the cpu isn't present, the cpu is mapped to first hctx.
2286 for_each_possible_cpu(i
) {
2287 hctx_idx
= q
->mq_map
[i
];
2288 /* unmapped hw queue can be remapped after CPU topo changed */
2289 if (!set
->tags
[hctx_idx
] &&
2290 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2292 * If tags initialization fail for some hctx,
2293 * that hctx won't be brought online. In this
2294 * case, remap the current ctx to hctx[0] which
2295 * is guaranteed to always have tags allocated
2300 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2301 hctx
= blk_mq_map_queue(q
, i
);
2303 cpumask_set_cpu(i
, hctx
->cpumask
);
2304 ctx
->index_hw
= hctx
->nr_ctx
;
2305 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2308 mutex_unlock(&q
->sysfs_lock
);
2310 queue_for_each_hw_ctx(q
, hctx
, i
) {
2312 * If no software queues are mapped to this hardware queue,
2313 * disable it and free the request entries.
2315 if (!hctx
->nr_ctx
) {
2316 /* Never unmap queue 0. We need it as a
2317 * fallback in case of a new remap fails
2320 if (i
&& set
->tags
[i
])
2321 blk_mq_free_map_and_requests(set
, i
);
2327 hctx
->tags
= set
->tags
[i
];
2328 WARN_ON(!hctx
->tags
);
2331 * Set the map size to the number of mapped software queues.
2332 * This is more accurate and more efficient than looping
2333 * over all possibly mapped software queues.
2335 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2338 * Initialize batch roundrobin counts
2340 hctx
->next_cpu
= cpumask_first_and(hctx
->cpumask
,
2342 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2347 * Caller needs to ensure that we're either frozen/quiesced, or that
2348 * the queue isn't live yet.
2350 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2352 struct blk_mq_hw_ctx
*hctx
;
2355 queue_for_each_hw_ctx(q
, hctx
, i
) {
2357 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2358 atomic_inc(&q
->shared_hctx_restart
);
2359 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2361 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
2362 atomic_dec(&q
->shared_hctx_restart
);
2363 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2368 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2371 struct request_queue
*q
;
2373 lockdep_assert_held(&set
->tag_list_lock
);
2375 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2376 blk_mq_freeze_queue(q
);
2377 queue_set_hctx_shared(q
, shared
);
2378 blk_mq_unfreeze_queue(q
);
2382 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2384 struct blk_mq_tag_set
*set
= q
->tag_set
;
2386 mutex_lock(&set
->tag_list_lock
);
2387 list_del_rcu(&q
->tag_set_list
);
2388 INIT_LIST_HEAD(&q
->tag_set_list
);
2389 if (list_is_singular(&set
->tag_list
)) {
2390 /* just transitioned to unshared */
2391 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2392 /* update existing queue */
2393 blk_mq_update_tag_set_depth(set
, false);
2395 mutex_unlock(&set
->tag_list_lock
);
2400 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2401 struct request_queue
*q
)
2405 mutex_lock(&set
->tag_list_lock
);
2408 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2410 if (!list_empty(&set
->tag_list
) &&
2411 !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2412 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2413 /* update existing queue */
2414 blk_mq_update_tag_set_depth(set
, true);
2416 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2417 queue_set_hctx_shared(q
, true);
2418 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2420 mutex_unlock(&set
->tag_list_lock
);
2424 * It is the actual release handler for mq, but we do it from
2425 * request queue's release handler for avoiding use-after-free
2426 * and headache because q->mq_kobj shouldn't have been introduced,
2427 * but we can't group ctx/kctx kobj without it.
2429 void blk_mq_release(struct request_queue
*q
)
2431 struct blk_mq_hw_ctx
*hctx
;
2434 /* hctx kobj stays in hctx */
2435 queue_for_each_hw_ctx(q
, hctx
, i
) {
2438 kobject_put(&hctx
->kobj
);
2443 kfree(q
->queue_hw_ctx
);
2446 * release .mq_kobj and sw queue's kobject now because
2447 * both share lifetime with request queue.
2449 blk_mq_sysfs_deinit(q
);
2451 free_percpu(q
->queue_ctx
);
2454 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2456 struct request_queue
*uninit_q
, *q
;
2458 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2460 return ERR_PTR(-ENOMEM
);
2462 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2464 blk_cleanup_queue(uninit_q
);
2468 EXPORT_SYMBOL(blk_mq_init_queue
);
2470 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2472 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2474 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, srcu
),
2475 __alignof__(struct blk_mq_hw_ctx
)) !=
2476 sizeof(struct blk_mq_hw_ctx
));
2478 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2479 hw_ctx_size
+= sizeof(struct srcu_struct
);
2484 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2485 struct request_queue
*q
)
2488 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2490 blk_mq_sysfs_unregister(q
);
2492 /* protect against switching io scheduler */
2493 mutex_lock(&q
->sysfs_lock
);
2494 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2500 node
= blk_mq_hw_queue_to_node(q
->mq_map
, i
);
2501 hctxs
[i
] = kzalloc_node(blk_mq_hw_ctx_size(set
),
2506 if (!zalloc_cpumask_var_node(&hctxs
[i
]->cpumask
, GFP_KERNEL
,
2513 atomic_set(&hctxs
[i
]->nr_active
, 0);
2514 hctxs
[i
]->numa_node
= node
;
2515 hctxs
[i
]->queue_num
= i
;
2517 if (blk_mq_init_hctx(q
, set
, hctxs
[i
], i
)) {
2518 free_cpumask_var(hctxs
[i
]->cpumask
);
2523 blk_mq_hctx_kobj_init(hctxs
[i
]);
2525 for (j
= i
; j
< q
->nr_hw_queues
; j
++) {
2526 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2530 blk_mq_free_map_and_requests(set
, j
);
2531 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2532 kobject_put(&hctx
->kobj
);
2537 q
->nr_hw_queues
= i
;
2538 mutex_unlock(&q
->sysfs_lock
);
2539 blk_mq_sysfs_register(q
);
2542 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2543 struct request_queue
*q
)
2545 /* mark the queue as mq asap */
2546 q
->mq_ops
= set
->ops
;
2548 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2549 blk_mq_poll_stats_bkt
,
2550 BLK_MQ_POLL_STATS_BKTS
, q
);
2554 q
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2558 /* init q->mq_kobj and sw queues' kobjects */
2559 blk_mq_sysfs_init(q
);
2561 q
->queue_hw_ctx
= kzalloc_node(nr_cpu_ids
* sizeof(*(q
->queue_hw_ctx
)),
2562 GFP_KERNEL
, set
->numa_node
);
2563 if (!q
->queue_hw_ctx
)
2566 q
->mq_map
= set
->mq_map
;
2568 blk_mq_realloc_hw_ctxs(set
, q
);
2569 if (!q
->nr_hw_queues
)
2572 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2573 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2575 q
->nr_queues
= nr_cpu_ids
;
2577 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2579 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2580 q
->queue_flags
|= 1 << QUEUE_FLAG_NO_SG_MERGE
;
2582 q
->sg_reserved_size
= INT_MAX
;
2584 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2585 INIT_LIST_HEAD(&q
->requeue_list
);
2586 spin_lock_init(&q
->requeue_lock
);
2588 blk_queue_make_request(q
, blk_mq_make_request
);
2589 if (q
->mq_ops
->poll
)
2590 q
->poll_fn
= blk_mq_poll
;
2593 * Do this after blk_queue_make_request() overrides it...
2595 q
->nr_requests
= set
->queue_depth
;
2598 * Default to classic polling
2602 if (set
->ops
->complete
)
2603 blk_queue_softirq_done(q
, set
->ops
->complete
);
2605 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2606 blk_mq_add_queue_tag_set(set
, q
);
2607 blk_mq_map_swqueue(q
);
2609 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2612 ret
= blk_mq_sched_init(q
);
2614 return ERR_PTR(ret
);
2620 kfree(q
->queue_hw_ctx
);
2622 free_percpu(q
->queue_ctx
);
2625 return ERR_PTR(-ENOMEM
);
2627 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2629 void blk_mq_free_queue(struct request_queue
*q
)
2631 struct blk_mq_tag_set
*set
= q
->tag_set
;
2633 blk_mq_del_queue_tag_set(q
);
2634 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2637 /* Basically redo blk_mq_init_queue with queue frozen */
2638 static void blk_mq_queue_reinit(struct request_queue
*q
)
2640 WARN_ON_ONCE(!atomic_read(&q
->mq_freeze_depth
));
2642 blk_mq_debugfs_unregister_hctxs(q
);
2643 blk_mq_sysfs_unregister(q
);
2646 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2647 * we should change hctx numa_node according to the new topology (this
2648 * involves freeing and re-allocating memory, worth doing?)
2650 blk_mq_map_swqueue(q
);
2652 blk_mq_sysfs_register(q
);
2653 blk_mq_debugfs_register_hctxs(q
);
2656 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2660 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2661 if (!__blk_mq_alloc_rq_map(set
, i
))
2668 blk_mq_free_rq_map(set
->tags
[i
]);
2674 * Allocate the request maps associated with this tag_set. Note that this
2675 * may reduce the depth asked for, if memory is tight. set->queue_depth
2676 * will be updated to reflect the allocated depth.
2678 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2683 depth
= set
->queue_depth
;
2685 err
= __blk_mq_alloc_rq_maps(set
);
2689 set
->queue_depth
>>= 1;
2690 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2694 } while (set
->queue_depth
);
2696 if (!set
->queue_depth
|| err
) {
2697 pr_err("blk-mq: failed to allocate request map\n");
2701 if (depth
!= set
->queue_depth
)
2702 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2703 depth
, set
->queue_depth
);
2708 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2710 if (set
->ops
->map_queues
) {
2713 * transport .map_queues is usually done in the following
2716 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2717 * mask = get_cpu_mask(queue)
2718 * for_each_cpu(cpu, mask)
2719 * set->mq_map[cpu] = queue;
2722 * When we need to remap, the table has to be cleared for
2723 * killing stale mapping since one CPU may not be mapped
2726 for_each_possible_cpu(cpu
)
2727 set
->mq_map
[cpu
] = 0;
2729 return set
->ops
->map_queues(set
);
2731 return blk_mq_map_queues(set
);
2735 * Alloc a tag set to be associated with one or more request queues.
2736 * May fail with EINVAL for various error conditions. May adjust the
2737 * requested depth down, if if it too large. In that case, the set
2738 * value will be stored in set->queue_depth.
2740 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2744 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2746 if (!set
->nr_hw_queues
)
2748 if (!set
->queue_depth
)
2750 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2753 if (!set
->ops
->queue_rq
)
2756 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
2759 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2760 pr_info("blk-mq: reduced tag depth to %u\n",
2762 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2766 * If a crashdump is active, then we are potentially in a very
2767 * memory constrained environment. Limit us to 1 queue and
2768 * 64 tags to prevent using too much memory.
2770 if (is_kdump_kernel()) {
2771 set
->nr_hw_queues
= 1;
2772 set
->queue_depth
= min(64U, set
->queue_depth
);
2775 * There is no use for more h/w queues than cpus.
2777 if (set
->nr_hw_queues
> nr_cpu_ids
)
2778 set
->nr_hw_queues
= nr_cpu_ids
;
2780 set
->tags
= kzalloc_node(nr_cpu_ids
* sizeof(struct blk_mq_tags
*),
2781 GFP_KERNEL
, set
->numa_node
);
2786 set
->mq_map
= kzalloc_node(sizeof(*set
->mq_map
) * nr_cpu_ids
,
2787 GFP_KERNEL
, set
->numa_node
);
2791 ret
= blk_mq_update_queue_map(set
);
2793 goto out_free_mq_map
;
2795 ret
= blk_mq_alloc_rq_maps(set
);
2797 goto out_free_mq_map
;
2799 mutex_init(&set
->tag_list_lock
);
2800 INIT_LIST_HEAD(&set
->tag_list
);
2812 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
2814 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
2818 for (i
= 0; i
< nr_cpu_ids
; i
++)
2819 blk_mq_free_map_and_requests(set
, i
);
2827 EXPORT_SYMBOL(blk_mq_free_tag_set
);
2829 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
2831 struct blk_mq_tag_set
*set
= q
->tag_set
;
2832 struct blk_mq_hw_ctx
*hctx
;
2838 blk_mq_freeze_queue(q
);
2839 blk_mq_quiesce_queue(q
);
2842 queue_for_each_hw_ctx(q
, hctx
, i
) {
2846 * If we're using an MQ scheduler, just update the scheduler
2847 * queue depth. This is similar to what the old code would do.
2849 if (!hctx
->sched_tags
) {
2850 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
2853 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
2861 q
->nr_requests
= nr
;
2863 blk_mq_unquiesce_queue(q
);
2864 blk_mq_unfreeze_queue(q
);
2869 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
2872 struct request_queue
*q
;
2874 lockdep_assert_held(&set
->tag_list_lock
);
2876 if (nr_hw_queues
> nr_cpu_ids
)
2877 nr_hw_queues
= nr_cpu_ids
;
2878 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
2881 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2882 blk_mq_freeze_queue(q
);
2884 set
->nr_hw_queues
= nr_hw_queues
;
2885 blk_mq_update_queue_map(set
);
2886 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2887 blk_mq_realloc_hw_ctxs(set
, q
);
2888 blk_mq_queue_reinit(q
);
2891 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
2892 blk_mq_unfreeze_queue(q
);
2895 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
2897 mutex_lock(&set
->tag_list_lock
);
2898 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
2899 mutex_unlock(&set
->tag_list_lock
);
2901 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
2903 /* Enable polling stats and return whether they were already enabled. */
2904 static bool blk_poll_stats_enable(struct request_queue
*q
)
2906 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2907 test_and_set_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
))
2909 blk_stat_add_callback(q
, q
->poll_cb
);
2913 static void blk_mq_poll_stats_start(struct request_queue
*q
)
2916 * We don't arm the callback if polling stats are not enabled or the
2917 * callback is already active.
2919 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
2920 blk_stat_is_active(q
->poll_cb
))
2923 blk_stat_activate_msecs(q
->poll_cb
, 100);
2926 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
2928 struct request_queue
*q
= cb
->data
;
2931 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
2932 if (cb
->stat
[bucket
].nr_samples
)
2933 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
2937 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
2938 struct blk_mq_hw_ctx
*hctx
,
2941 unsigned long ret
= 0;
2945 * If stats collection isn't on, don't sleep but turn it on for
2948 if (!blk_poll_stats_enable(q
))
2952 * As an optimistic guess, use half of the mean service time
2953 * for this type of request. We can (and should) make this smarter.
2954 * For instance, if the completion latencies are tight, we can
2955 * get closer than just half the mean. This is especially
2956 * important on devices where the completion latencies are longer
2957 * than ~10 usec. We do use the stats for the relevant IO size
2958 * if available which does lead to better estimates.
2960 bucket
= blk_mq_poll_stats_bkt(rq
);
2964 if (q
->poll_stat
[bucket
].nr_samples
)
2965 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
2970 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
2971 struct blk_mq_hw_ctx
*hctx
,
2974 struct hrtimer_sleeper hs
;
2975 enum hrtimer_mode mode
;
2979 if (rq
->rq_flags
& RQF_MQ_POLL_SLEPT
)
2985 * -1: don't ever hybrid sleep
2986 * 0: use half of prev avg
2987 * >0: use this specific value
2989 if (q
->poll_nsec
== -1)
2991 else if (q
->poll_nsec
> 0)
2992 nsecs
= q
->poll_nsec
;
2994 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
2999 rq
->rq_flags
|= RQF_MQ_POLL_SLEPT
;
3002 * This will be replaced with the stats tracking code, using
3003 * 'avg_completion_time / 2' as the pre-sleep target.
3007 mode
= HRTIMER_MODE_REL
;
3008 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
3009 hrtimer_set_expires(&hs
.timer
, kt
);
3011 hrtimer_init_sleeper(&hs
, current
);
3013 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
3015 set_current_state(TASK_UNINTERRUPTIBLE
);
3016 hrtimer_start_expires(&hs
.timer
, mode
);
3019 hrtimer_cancel(&hs
.timer
);
3020 mode
= HRTIMER_MODE_ABS
;
3021 } while (hs
.task
&& !signal_pending(current
));
3023 __set_current_state(TASK_RUNNING
);
3024 destroy_hrtimer_on_stack(&hs
.timer
);
3028 static bool __blk_mq_poll(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
3030 struct request_queue
*q
= hctx
->queue
;
3034 * If we sleep, have the caller restart the poll loop to reset
3035 * the state. Like for the other success return cases, the
3036 * caller is responsible for checking if the IO completed. If
3037 * the IO isn't complete, we'll get called again and will go
3038 * straight to the busy poll loop.
3040 if (blk_mq_poll_hybrid_sleep(q
, hctx
, rq
))
3043 hctx
->poll_considered
++;
3045 state
= current
->state
;
3046 while (!need_resched()) {
3049 hctx
->poll_invoked
++;
3051 ret
= q
->mq_ops
->poll(hctx
, rq
->tag
);
3053 hctx
->poll_success
++;
3054 set_current_state(TASK_RUNNING
);
3058 if (signal_pending_state(state
, current
))
3059 set_current_state(TASK_RUNNING
);
3061 if (current
->state
== TASK_RUNNING
)
3071 static bool blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
)
3073 struct blk_mq_hw_ctx
*hctx
;
3076 if (!test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3079 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3080 if (!blk_qc_t_is_internal(cookie
))
3081 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3083 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3085 * With scheduling, if the request has completed, we'll
3086 * get a NULL return here, as we clear the sched tag when
3087 * that happens. The request still remains valid, like always,
3088 * so we should be safe with just the NULL check.
3094 return __blk_mq_poll(hctx
, rq
);
3097 static int __init
blk_mq_init(void)
3099 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3100 blk_mq_hctx_notify_dead
);
3103 subsys_initcall(blk_mq_init
);