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"
39 #include "blk-rq-qos.h"
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 const int bit
= ctx
->index_hw
[hctx
->type
];
79 if (!sbitmap_test_bit(&hctx
->ctx_map
, bit
))
80 sbitmap_set_bit(&hctx
->ctx_map
, bit
);
83 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
84 struct blk_mq_ctx
*ctx
)
86 const int bit
= ctx
->index_hw
[hctx
->type
];
88 sbitmap_clear_bit(&hctx
->ctx_map
, bit
);
92 struct hd_struct
*part
;
93 unsigned int *inflight
;
96 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx
*hctx
,
97 struct request
*rq
, void *priv
,
100 struct mq_inflight
*mi
= priv
;
103 * index[0] counts the specific partition that was asked for.
105 if (rq
->part
== mi
->part
)
111 unsigned int blk_mq_in_flight(struct request_queue
*q
, struct hd_struct
*part
)
113 unsigned inflight
[2];
114 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
116 inflight
[0] = inflight
[1] = 0;
117 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
122 static bool 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
)]++;
134 void blk_mq_in_flight_rw(struct request_queue
*q
, struct hd_struct
*part
,
135 unsigned int inflight
[2])
137 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
139 inflight
[0] = inflight
[1] = 0;
140 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight_rw
, &mi
);
143 void blk_freeze_queue_start(struct request_queue
*q
)
147 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
148 if (freeze_depth
== 1) {
149 percpu_ref_kill(&q
->q_usage_counter
);
151 blk_mq_run_hw_queues(q
, false);
154 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
156 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
158 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
160 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
162 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
163 unsigned long timeout
)
165 return wait_event_timeout(q
->mq_freeze_wq
,
166 percpu_ref_is_zero(&q
->q_usage_counter
),
169 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
172 * Guarantee no request is in use, so we can change any data structure of
173 * the queue afterward.
175 void blk_freeze_queue(struct request_queue
*q
)
178 * In the !blk_mq case we are only calling this to kill the
179 * q_usage_counter, otherwise this increases the freeze depth
180 * and waits for it to return to zero. For this reason there is
181 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
182 * exported to drivers as the only user for unfreeze is blk_mq.
184 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_resurrect(&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
)
217 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
219 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
222 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
225 * Note: this function does not prevent that the struct request end_io()
226 * callback function is invoked. Once this function is returned, we make
227 * sure no dispatch can happen until the queue is unquiesced via
228 * blk_mq_unquiesce_queue().
230 void blk_mq_quiesce_queue(struct request_queue
*q
)
232 struct blk_mq_hw_ctx
*hctx
;
236 blk_mq_quiesce_queue_nowait(q
);
238 queue_for_each_hw_ctx(q
, hctx
, i
) {
239 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
240 synchronize_srcu(hctx
->srcu
);
247 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
250 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
253 * This function recovers queue into the state before quiescing
254 * which is done by blk_mq_quiesce_queue.
256 void blk_mq_unquiesce_queue(struct request_queue
*q
)
258 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
260 /* dispatch requests which are inserted during quiescing */
261 blk_mq_run_hw_queues(q
, true);
263 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
265 void blk_mq_wake_waiters(struct request_queue
*q
)
267 struct blk_mq_hw_ctx
*hctx
;
270 queue_for_each_hw_ctx(q
, hctx
, i
)
271 if (blk_mq_hw_queue_mapped(hctx
))
272 blk_mq_tag_wakeup_all(hctx
->tags
, true);
275 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
277 return blk_mq_has_free_tags(hctx
->tags
);
279 EXPORT_SYMBOL(blk_mq_can_queue
);
282 * Only need start/end time stamping if we have stats enabled, or using
285 static inline bool blk_mq_need_time_stamp(struct request
*rq
)
287 return (rq
->rq_flags
& RQF_IO_STAT
) || rq
->q
->elevator
;
290 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
291 unsigned int tag
, unsigned int op
)
293 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
294 struct request
*rq
= tags
->static_rqs
[tag
];
295 req_flags_t rq_flags
= 0;
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_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 /* csd/requeue_work/fifo_time is initialized before use */
312 rq
->mq_ctx
= data
->ctx
;
313 rq
->mq_hctx
= data
->hctx
;
314 rq
->rq_flags
= rq_flags
;
316 if (data
->flags
& BLK_MQ_REQ_PREEMPT
)
317 rq
->rq_flags
|= RQF_PREEMPT
;
318 if (blk_queue_io_stat(data
->q
))
319 rq
->rq_flags
|= RQF_IO_STAT
;
320 INIT_LIST_HEAD(&rq
->queuelist
);
321 INIT_HLIST_NODE(&rq
->hash
);
322 RB_CLEAR_NODE(&rq
->rb_node
);
325 if (blk_mq_need_time_stamp(rq
))
326 rq
->start_time_ns
= ktime_get_ns();
328 rq
->start_time_ns
= 0;
329 rq
->io_start_time_ns
= 0;
330 rq
->nr_phys_segments
= 0;
331 #if defined(CONFIG_BLK_DEV_INTEGRITY)
332 rq
->nr_integrity_segments
= 0;
335 /* tag was already set */
337 WRITE_ONCE(rq
->deadline
, 0);
342 rq
->end_io_data
= NULL
;
345 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
346 refcount_set(&rq
->ref
, 1);
350 static struct request
*blk_mq_get_request(struct request_queue
*q
,
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
->cmd_flags
,
368 if (data
->cmd_flags
& REQ_NOWAIT
)
369 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
372 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
375 * Flush requests are special and go directly to the
376 * dispatch list. Don't include reserved tags in the
377 * limiting, as it isn't useful.
379 if (!op_is_flush(data
->cmd_flags
) &&
380 e
->type
->ops
.limit_depth
&&
381 !(data
->flags
& BLK_MQ_REQ_RESERVED
))
382 e
->type
->ops
.limit_depth(data
->cmd_flags
, data
);
384 blk_mq_tag_busy(data
->hctx
);
387 tag
= blk_mq_get_tag(data
);
388 if (tag
== BLK_MQ_TAG_FAIL
) {
389 if (put_ctx_on_error
) {
390 blk_mq_put_ctx(data
->ctx
);
397 rq
= blk_mq_rq_ctx_init(data
, tag
, data
->cmd_flags
);
398 if (!op_is_flush(data
->cmd_flags
)) {
400 if (e
&& e
->type
->ops
.prepare_request
) {
401 if (e
->type
->icq_cache
)
402 blk_mq_sched_assign_ioc(rq
);
404 e
->type
->ops
.prepare_request(rq
, bio
);
405 rq
->rq_flags
|= RQF_ELVPRIV
;
408 data
->hctx
->queued
++;
412 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
413 blk_mq_req_flags_t flags
)
415 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
, .cmd_flags
= op
};
419 ret
= blk_queue_enter(q
, flags
);
423 rq
= blk_mq_get_request(q
, NULL
, &alloc_data
);
427 return ERR_PTR(-EWOULDBLOCK
);
429 blk_mq_put_ctx(alloc_data
.ctx
);
432 rq
->__sector
= (sector_t
) -1;
433 rq
->bio
= rq
->biotail
= NULL
;
436 EXPORT_SYMBOL(blk_mq_alloc_request
);
438 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
439 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
441 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
, .cmd_flags
= op
};
447 * If the tag allocator sleeps we could get an allocation for a
448 * different hardware context. No need to complicate the low level
449 * allocator for this for the rare use case of a command tied to
452 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
453 return ERR_PTR(-EINVAL
);
455 if (hctx_idx
>= q
->nr_hw_queues
)
456 return ERR_PTR(-EIO
);
458 ret
= blk_queue_enter(q
, flags
);
463 * Check if the hardware context is actually mapped to anything.
464 * If not tell the caller that it should skip this queue.
466 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
467 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
469 return ERR_PTR(-EXDEV
);
471 cpu
= cpumask_first_and(alloc_data
.hctx
->cpumask
, cpu_online_mask
);
472 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
474 rq
= blk_mq_get_request(q
, NULL
, &alloc_data
);
478 return ERR_PTR(-EWOULDBLOCK
);
482 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
484 static void __blk_mq_free_request(struct request
*rq
)
486 struct request_queue
*q
= rq
->q
;
487 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
488 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
489 const int sched_tag
= rq
->internal_tag
;
491 blk_pm_mark_last_busy(rq
);
494 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
496 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
497 blk_mq_sched_restart(hctx
);
501 void blk_mq_free_request(struct request
*rq
)
503 struct request_queue
*q
= rq
->q
;
504 struct elevator_queue
*e
= q
->elevator
;
505 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
506 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
508 if (rq
->rq_flags
& RQF_ELVPRIV
) {
509 if (e
&& e
->type
->ops
.finish_request
)
510 e
->type
->ops
.finish_request(rq
);
512 put_io_context(rq
->elv
.icq
->ioc
);
517 ctx
->rq_completed
[rq_is_sync(rq
)]++;
518 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
519 atomic_dec(&hctx
->nr_active
);
521 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
522 laptop_io_completion(q
->backing_dev_info
);
526 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
527 if (refcount_dec_and_test(&rq
->ref
))
528 __blk_mq_free_request(rq
);
530 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
532 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
536 if (blk_mq_need_time_stamp(rq
))
537 now
= ktime_get_ns();
539 if (rq
->rq_flags
& RQF_STATS
) {
540 blk_mq_poll_stats_start(rq
->q
);
541 blk_stat_add(rq
, now
);
544 if (rq
->internal_tag
!= -1)
545 blk_mq_sched_completed_request(rq
, now
);
547 blk_account_io_done(rq
, now
);
550 rq_qos_done(rq
->q
, rq
);
551 rq
->end_io(rq
, error
);
553 if (unlikely(blk_bidi_rq(rq
)))
554 blk_mq_free_request(rq
->next_rq
);
555 blk_mq_free_request(rq
);
558 EXPORT_SYMBOL(__blk_mq_end_request
);
560 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
562 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
564 __blk_mq_end_request(rq
, error
);
566 EXPORT_SYMBOL(blk_mq_end_request
);
568 static void __blk_mq_complete_request_remote(void *data
)
570 struct request
*rq
= data
;
571 struct request_queue
*q
= rq
->q
;
573 q
->mq_ops
->complete(rq
);
576 static void __blk_mq_complete_request(struct request
*rq
)
578 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
579 struct request_queue
*q
= rq
->q
;
583 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
585 * Most of single queue controllers, there is only one irq vector
586 * for handling IO completion, and the only irq's affinity is set
587 * as all possible CPUs. On most of ARCHs, this affinity means the
588 * irq is handled on one specific CPU.
590 * So complete IO reqeust in softirq context in case of single queue
591 * for not degrading IO performance by irqsoff latency.
593 if (q
->nr_hw_queues
== 1) {
594 __blk_complete_request(rq
);
599 * For a polled request, always complete locallly, it's pointless
600 * to redirect the completion.
602 if ((rq
->cmd_flags
& REQ_HIPRI
) ||
603 !test_bit(QUEUE_FLAG_SAME_COMP
, &q
->queue_flags
)) {
604 q
->mq_ops
->complete(rq
);
609 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &q
->queue_flags
))
610 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
612 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
613 rq
->csd
.func
= __blk_mq_complete_request_remote
;
616 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
618 q
->mq_ops
->complete(rq
);
623 static void hctx_unlock(struct blk_mq_hw_ctx
*hctx
, int srcu_idx
)
624 __releases(hctx
->srcu
)
626 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
629 srcu_read_unlock(hctx
->srcu
, srcu_idx
);
632 static void hctx_lock(struct blk_mq_hw_ctx
*hctx
, int *srcu_idx
)
633 __acquires(hctx
->srcu
)
635 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
636 /* shut up gcc false positive */
640 *srcu_idx
= srcu_read_lock(hctx
->srcu
);
644 * blk_mq_complete_request - end I/O on a request
645 * @rq: the request being processed
648 * Ends all I/O on a request. It does not handle partial completions.
649 * The actual completion happens out-of-order, through a IPI handler.
651 bool blk_mq_complete_request(struct request
*rq
)
653 if (unlikely(blk_should_fake_timeout(rq
->q
)))
655 __blk_mq_complete_request(rq
);
658 EXPORT_SYMBOL(blk_mq_complete_request
);
660 int blk_mq_request_started(struct request
*rq
)
662 return blk_mq_rq_state(rq
) != MQ_RQ_IDLE
;
664 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
666 void blk_mq_start_request(struct request
*rq
)
668 struct request_queue
*q
= rq
->q
;
670 blk_mq_sched_started_request(rq
);
672 trace_block_rq_issue(q
, rq
);
674 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
675 rq
->io_start_time_ns
= ktime_get_ns();
676 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
677 rq
->throtl_size
= blk_rq_sectors(rq
);
679 rq
->rq_flags
|= RQF_STATS
;
683 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
686 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
688 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
690 * Make sure space for the drain appears. We know we can do
691 * this because max_hw_segments has been adjusted to be one
692 * fewer than the device can handle.
694 rq
->nr_phys_segments
++;
697 EXPORT_SYMBOL(blk_mq_start_request
);
699 static void __blk_mq_requeue_request(struct request
*rq
)
701 struct request_queue
*q
= rq
->q
;
703 blk_mq_put_driver_tag(rq
);
705 trace_block_rq_requeue(q
, rq
);
706 rq_qos_requeue(q
, rq
);
708 if (blk_mq_request_started(rq
)) {
709 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
710 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
711 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
712 rq
->nr_phys_segments
--;
716 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
718 __blk_mq_requeue_request(rq
);
720 /* this request will be re-inserted to io scheduler queue */
721 blk_mq_sched_requeue_request(rq
);
723 BUG_ON(!list_empty(&rq
->queuelist
));
724 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
726 EXPORT_SYMBOL(blk_mq_requeue_request
);
728 static void blk_mq_requeue_work(struct work_struct
*work
)
730 struct request_queue
*q
=
731 container_of(work
, struct request_queue
, requeue_work
.work
);
733 struct request
*rq
, *next
;
735 spin_lock_irq(&q
->requeue_lock
);
736 list_splice_init(&q
->requeue_list
, &rq_list
);
737 spin_unlock_irq(&q
->requeue_lock
);
739 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
740 if (!(rq
->rq_flags
& (RQF_SOFTBARRIER
| RQF_DONTPREP
)))
743 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
744 list_del_init(&rq
->queuelist
);
746 * If RQF_DONTPREP, rq has contained some driver specific
747 * data, so insert it to hctx dispatch list to avoid any
750 if (rq
->rq_flags
& RQF_DONTPREP
)
751 blk_mq_request_bypass_insert(rq
, false);
753 blk_mq_sched_insert_request(rq
, true, false, false);
756 while (!list_empty(&rq_list
)) {
757 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
758 list_del_init(&rq
->queuelist
);
759 blk_mq_sched_insert_request(rq
, false, false, false);
762 blk_mq_run_hw_queues(q
, false);
765 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
766 bool kick_requeue_list
)
768 struct request_queue
*q
= rq
->q
;
772 * We abuse this flag that is otherwise used by the I/O scheduler to
773 * request head insertion from the workqueue.
775 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
777 spin_lock_irqsave(&q
->requeue_lock
, flags
);
779 rq
->rq_flags
|= RQF_SOFTBARRIER
;
780 list_add(&rq
->queuelist
, &q
->requeue_list
);
782 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
784 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
786 if (kick_requeue_list
)
787 blk_mq_kick_requeue_list(q
);
789 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
791 void blk_mq_kick_requeue_list(struct request_queue
*q
)
793 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
795 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
797 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
800 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
801 msecs_to_jiffies(msecs
));
803 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
805 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
807 if (tag
< tags
->nr_tags
) {
808 prefetch(tags
->rqs
[tag
]);
809 return tags
->rqs
[tag
];
814 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
816 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
817 void *priv
, bool reserved
)
820 * If we find a request that is inflight and the queue matches,
821 * we know the queue is busy. Return false to stop the iteration.
823 if (rq
->state
== MQ_RQ_IN_FLIGHT
&& rq
->q
== hctx
->queue
) {
833 bool blk_mq_queue_inflight(struct request_queue
*q
)
837 blk_mq_queue_tag_busy_iter(q
, blk_mq_rq_inflight
, &busy
);
840 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight
);
842 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
844 req
->rq_flags
|= RQF_TIMED_OUT
;
845 if (req
->q
->mq_ops
->timeout
) {
846 enum blk_eh_timer_return ret
;
848 ret
= req
->q
->mq_ops
->timeout(req
, reserved
);
849 if (ret
== BLK_EH_DONE
)
851 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
857 static bool blk_mq_req_expired(struct request
*rq
, unsigned long *next
)
859 unsigned long deadline
;
861 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
863 if (rq
->rq_flags
& RQF_TIMED_OUT
)
866 deadline
= READ_ONCE(rq
->deadline
);
867 if (time_after_eq(jiffies
, deadline
))
872 else if (time_after(*next
, deadline
))
877 static bool blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
878 struct request
*rq
, void *priv
, bool reserved
)
880 unsigned long *next
= priv
;
883 * Just do a quick check if it is expired before locking the request in
884 * so we're not unnecessarilly synchronizing across CPUs.
886 if (!blk_mq_req_expired(rq
, next
))
890 * We have reason to believe the request may be expired. Take a
891 * reference on the request to lock this request lifetime into its
892 * currently allocated context to prevent it from being reallocated in
893 * the event the completion by-passes this timeout handler.
895 * If the reference was already released, then the driver beat the
896 * timeout handler to posting a natural completion.
898 if (!refcount_inc_not_zero(&rq
->ref
))
902 * The request is now locked and cannot be reallocated underneath the
903 * timeout handler's processing. Re-verify this exact request is truly
904 * expired; if it is not expired, then the request was completed and
905 * reallocated as a new request.
907 if (blk_mq_req_expired(rq
, next
))
908 blk_mq_rq_timed_out(rq
, reserved
);
909 if (refcount_dec_and_test(&rq
->ref
))
910 __blk_mq_free_request(rq
);
915 static void blk_mq_timeout_work(struct work_struct
*work
)
917 struct request_queue
*q
=
918 container_of(work
, struct request_queue
, timeout_work
);
919 unsigned long next
= 0;
920 struct blk_mq_hw_ctx
*hctx
;
923 /* A deadlock might occur if a request is stuck requiring a
924 * timeout at the same time a queue freeze is waiting
925 * completion, since the timeout code would not be able to
926 * acquire the queue reference here.
928 * That's why we don't use blk_queue_enter here; instead, we use
929 * percpu_ref_tryget directly, because we need to be able to
930 * obtain a reference even in the short window between the queue
931 * starting to freeze, by dropping the first reference in
932 * blk_freeze_queue_start, and the moment the last request is
933 * consumed, marked by the instant q_usage_counter reaches
936 if (!percpu_ref_tryget(&q
->q_usage_counter
))
939 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &next
);
942 mod_timer(&q
->timeout
, next
);
945 * Request timeouts are handled as a forward rolling timer. If
946 * we end up here it means that no requests are pending and
947 * also that no request has been pending for a while. Mark
950 queue_for_each_hw_ctx(q
, hctx
, i
) {
951 /* the hctx may be unmapped, so check it here */
952 if (blk_mq_hw_queue_mapped(hctx
))
953 blk_mq_tag_idle(hctx
);
959 struct flush_busy_ctx_data
{
960 struct blk_mq_hw_ctx
*hctx
;
961 struct list_head
*list
;
964 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
966 struct flush_busy_ctx_data
*flush_data
= data
;
967 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
968 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
969 enum hctx_type type
= hctx
->type
;
971 spin_lock(&ctx
->lock
);
972 list_splice_tail_init(&ctx
->rq_lists
[type
], flush_data
->list
);
973 sbitmap_clear_bit(sb
, bitnr
);
974 spin_unlock(&ctx
->lock
);
979 * Process software queues that have been marked busy, splicing them
980 * to the for-dispatch
982 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
984 struct flush_busy_ctx_data data
= {
989 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
991 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
993 struct dispatch_rq_data
{
994 struct blk_mq_hw_ctx
*hctx
;
998 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
1001 struct dispatch_rq_data
*dispatch_data
= data
;
1002 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
1003 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
1004 enum hctx_type type
= hctx
->type
;
1006 spin_lock(&ctx
->lock
);
1007 if (!list_empty(&ctx
->rq_lists
[type
])) {
1008 dispatch_data
->rq
= list_entry_rq(ctx
->rq_lists
[type
].next
);
1009 list_del_init(&dispatch_data
->rq
->queuelist
);
1010 if (list_empty(&ctx
->rq_lists
[type
]))
1011 sbitmap_clear_bit(sb
, bitnr
);
1013 spin_unlock(&ctx
->lock
);
1015 return !dispatch_data
->rq
;
1018 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1019 struct blk_mq_ctx
*start
)
1021 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1022 struct dispatch_rq_data data
= {
1027 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1028 dispatch_rq_from_ctx
, &data
);
1033 static inline unsigned int queued_to_index(unsigned int queued
)
1038 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
1041 bool blk_mq_get_driver_tag(struct request
*rq
)
1043 struct blk_mq_alloc_data data
= {
1045 .hctx
= rq
->mq_hctx
,
1046 .flags
= BLK_MQ_REQ_NOWAIT
,
1047 .cmd_flags
= rq
->cmd_flags
,
1054 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
1055 data
.flags
|= BLK_MQ_REQ_RESERVED
;
1057 shared
= blk_mq_tag_busy(data
.hctx
);
1058 rq
->tag
= blk_mq_get_tag(&data
);
1061 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1062 atomic_inc(&data
.hctx
->nr_active
);
1064 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
1068 return rq
->tag
!= -1;
1071 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1072 int flags
, void *key
)
1074 struct blk_mq_hw_ctx
*hctx
;
1076 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1078 spin_lock(&hctx
->dispatch_wait_lock
);
1079 list_del_init(&wait
->entry
);
1080 spin_unlock(&hctx
->dispatch_wait_lock
);
1082 blk_mq_run_hw_queue(hctx
, true);
1087 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1088 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1089 * restart. For both cases, take care to check the condition again after
1090 * marking us as waiting.
1092 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1095 struct wait_queue_head
*wq
;
1096 wait_queue_entry_t
*wait
;
1099 if (!(hctx
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1100 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
1101 set_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
1104 * It's possible that a tag was freed in the window between the
1105 * allocation failure and adding the hardware queue to the wait
1108 * Don't clear RESTART here, someone else could have set it.
1109 * At most this will cost an extra queue run.
1111 return blk_mq_get_driver_tag(rq
);
1114 wait
= &hctx
->dispatch_wait
;
1115 if (!list_empty_careful(&wait
->entry
))
1118 wq
= &bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
)->wait
;
1120 spin_lock_irq(&wq
->lock
);
1121 spin_lock(&hctx
->dispatch_wait_lock
);
1122 if (!list_empty(&wait
->entry
)) {
1123 spin_unlock(&hctx
->dispatch_wait_lock
);
1124 spin_unlock_irq(&wq
->lock
);
1128 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1129 __add_wait_queue(wq
, wait
);
1132 * It's possible that a tag was freed in the window between the
1133 * allocation failure and adding the hardware queue to the wait
1136 ret
= blk_mq_get_driver_tag(rq
);
1138 spin_unlock(&hctx
->dispatch_wait_lock
);
1139 spin_unlock_irq(&wq
->lock
);
1144 * We got a tag, remove ourselves from the wait queue to ensure
1145 * someone else gets the wakeup.
1147 list_del_init(&wait
->entry
);
1148 spin_unlock(&hctx
->dispatch_wait_lock
);
1149 spin_unlock_irq(&wq
->lock
);
1154 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1155 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1157 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1158 * - EWMA is one simple way to compute running average value
1159 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1160 * - take 4 as factor for avoiding to get too small(0) result, and this
1161 * factor doesn't matter because EWMA decreases exponentially
1163 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1167 if (hctx
->queue
->elevator
)
1170 ewma
= hctx
->dispatch_busy
;
1175 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1177 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1178 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1180 hctx
->dispatch_busy
= ewma
;
1183 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1186 * Returns true if we did some work AND can potentially do more.
1188 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
,
1191 struct blk_mq_hw_ctx
*hctx
;
1192 struct request
*rq
, *nxt
;
1193 bool no_tag
= false;
1195 blk_status_t ret
= BLK_STS_OK
;
1197 if (list_empty(list
))
1200 WARN_ON(!list_is_singular(list
) && got_budget
);
1203 * Now process all the entries, sending them to the driver.
1205 errors
= queued
= 0;
1207 struct blk_mq_queue_data bd
;
1209 rq
= list_first_entry(list
, struct request
, queuelist
);
1212 if (!got_budget
&& !blk_mq_get_dispatch_budget(hctx
))
1215 if (!blk_mq_get_driver_tag(rq
)) {
1217 * The initial allocation attempt failed, so we need to
1218 * rerun the hardware queue when a tag is freed. The
1219 * waitqueue takes care of that. If the queue is run
1220 * before we add this entry back on the dispatch list,
1221 * we'll re-run it below.
1223 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1224 blk_mq_put_dispatch_budget(hctx
);
1226 * For non-shared tags, the RESTART check
1229 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
)
1235 list_del_init(&rq
->queuelist
);
1240 * Flag last if we have no more requests, or if we have more
1241 * but can't assign a driver tag to it.
1243 if (list_empty(list
))
1246 nxt
= list_first_entry(list
, struct request
, queuelist
);
1247 bd
.last
= !blk_mq_get_driver_tag(nxt
);
1250 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1251 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
) {
1253 * If an I/O scheduler has been configured and we got a
1254 * driver tag for the next request already, free it
1257 if (!list_empty(list
)) {
1258 nxt
= list_first_entry(list
, struct request
, queuelist
);
1259 blk_mq_put_driver_tag(nxt
);
1261 list_add(&rq
->queuelist
, list
);
1262 __blk_mq_requeue_request(rq
);
1266 if (unlikely(ret
!= BLK_STS_OK
)) {
1268 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1273 } while (!list_empty(list
));
1275 hctx
->dispatched
[queued_to_index(queued
)]++;
1278 * Any items that need requeuing? Stuff them into hctx->dispatch,
1279 * that is where we will continue on next queue run.
1281 if (!list_empty(list
)) {
1285 * If we didn't flush the entire list, we could have told
1286 * the driver there was more coming, but that turned out to
1289 if (q
->mq_ops
->commit_rqs
)
1290 q
->mq_ops
->commit_rqs(hctx
);
1292 spin_lock(&hctx
->lock
);
1293 list_splice_init(list
, &hctx
->dispatch
);
1294 spin_unlock(&hctx
->lock
);
1297 * If SCHED_RESTART was set by the caller of this function and
1298 * it is no longer set that means that it was cleared by another
1299 * thread and hence that a queue rerun is needed.
1301 * If 'no_tag' is set, that means that we failed getting
1302 * a driver tag with an I/O scheduler attached. If our dispatch
1303 * waitqueue is no longer active, ensure that we run the queue
1304 * AFTER adding our entries back to the list.
1306 * If no I/O scheduler has been configured it is possible that
1307 * the hardware queue got stopped and restarted before requests
1308 * were pushed back onto the dispatch list. Rerun the queue to
1309 * avoid starvation. Notes:
1310 * - blk_mq_run_hw_queue() checks whether or not a queue has
1311 * been stopped before rerunning a queue.
1312 * - Some but not all block drivers stop a queue before
1313 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1316 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1317 * bit is set, run queue after a delay to avoid IO stalls
1318 * that could otherwise occur if the queue is idle.
1320 needs_restart
= blk_mq_sched_needs_restart(hctx
);
1321 if (!needs_restart
||
1322 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1323 blk_mq_run_hw_queue(hctx
, true);
1324 else if (needs_restart
&& (ret
== BLK_STS_RESOURCE
))
1325 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
1327 blk_mq_update_dispatch_busy(hctx
, true);
1330 blk_mq_update_dispatch_busy(hctx
, false);
1333 * If the host/device is unable to accept more work, inform the
1336 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1339 return (queued
+ errors
) != 0;
1342 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1347 * We should be running this queue from one of the CPUs that
1350 * There are at least two related races now between setting
1351 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1352 * __blk_mq_run_hw_queue():
1354 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1355 * but later it becomes online, then this warning is harmless
1358 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1359 * but later it becomes offline, then the warning can't be
1360 * triggered, and we depend on blk-mq timeout handler to
1361 * handle dispatched requests to this hctx
1363 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1364 cpu_online(hctx
->next_cpu
)) {
1365 printk(KERN_WARNING
"run queue from wrong CPU %d, hctx %s\n",
1366 raw_smp_processor_id(),
1367 cpumask_empty(hctx
->cpumask
) ? "inactive": "active");
1372 * We can't run the queue inline with ints disabled. Ensure that
1373 * we catch bad users of this early.
1375 WARN_ON_ONCE(in_interrupt());
1377 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1379 hctx_lock(hctx
, &srcu_idx
);
1380 blk_mq_sched_dispatch_requests(hctx
);
1381 hctx_unlock(hctx
, srcu_idx
);
1384 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1386 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1388 if (cpu
>= nr_cpu_ids
)
1389 cpu
= cpumask_first(hctx
->cpumask
);
1394 * It'd be great if the workqueue API had a way to pass
1395 * in a mask and had some smarts for more clever placement.
1396 * For now we just round-robin here, switching for every
1397 * BLK_MQ_CPU_WORK_BATCH queued items.
1399 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1402 int next_cpu
= hctx
->next_cpu
;
1404 if (hctx
->queue
->nr_hw_queues
== 1)
1405 return WORK_CPU_UNBOUND
;
1407 if (--hctx
->next_cpu_batch
<= 0) {
1409 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1411 if (next_cpu
>= nr_cpu_ids
)
1412 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1413 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1417 * Do unbound schedule if we can't find a online CPU for this hctx,
1418 * and it should only happen in the path of handling CPU DEAD.
1420 if (!cpu_online(next_cpu
)) {
1427 * Make sure to re-select CPU next time once after CPUs
1428 * in hctx->cpumask become online again.
1430 hctx
->next_cpu
= next_cpu
;
1431 hctx
->next_cpu_batch
= 1;
1432 return WORK_CPU_UNBOUND
;
1435 hctx
->next_cpu
= next_cpu
;
1439 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1440 unsigned long msecs
)
1442 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1445 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1446 int cpu
= get_cpu();
1447 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1448 __blk_mq_run_hw_queue(hctx
);
1456 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1457 msecs_to_jiffies(msecs
));
1460 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1462 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1464 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1466 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1472 * When queue is quiesced, we may be switching io scheduler, or
1473 * updating nr_hw_queues, or other things, and we can't run queue
1474 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1476 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1479 hctx_lock(hctx
, &srcu_idx
);
1480 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1481 blk_mq_hctx_has_pending(hctx
);
1482 hctx_unlock(hctx
, srcu_idx
);
1485 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1491 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1493 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1495 struct blk_mq_hw_ctx
*hctx
;
1498 queue_for_each_hw_ctx(q
, hctx
, i
) {
1499 if (blk_mq_hctx_stopped(hctx
))
1502 blk_mq_run_hw_queue(hctx
, async
);
1505 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1508 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1509 * @q: request queue.
1511 * The caller is responsible for serializing this function against
1512 * blk_mq_{start,stop}_hw_queue().
1514 bool blk_mq_queue_stopped(struct request_queue
*q
)
1516 struct blk_mq_hw_ctx
*hctx
;
1519 queue_for_each_hw_ctx(q
, hctx
, i
)
1520 if (blk_mq_hctx_stopped(hctx
))
1525 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1528 * This function is often used for pausing .queue_rq() by driver when
1529 * there isn't enough resource or some conditions aren't satisfied, and
1530 * BLK_STS_RESOURCE is usually returned.
1532 * We do not guarantee that dispatch can be drained or blocked
1533 * after blk_mq_stop_hw_queue() returns. Please use
1534 * blk_mq_quiesce_queue() for that requirement.
1536 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1538 cancel_delayed_work(&hctx
->run_work
);
1540 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1542 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1545 * This function is often used for pausing .queue_rq() by driver when
1546 * there isn't enough resource or some conditions aren't satisfied, and
1547 * BLK_STS_RESOURCE is usually returned.
1549 * We do not guarantee that dispatch can be drained or blocked
1550 * after blk_mq_stop_hw_queues() returns. Please use
1551 * blk_mq_quiesce_queue() for that requirement.
1553 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1555 struct blk_mq_hw_ctx
*hctx
;
1558 queue_for_each_hw_ctx(q
, hctx
, i
)
1559 blk_mq_stop_hw_queue(hctx
);
1561 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1563 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1565 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1567 blk_mq_run_hw_queue(hctx
, false);
1569 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1571 void blk_mq_start_hw_queues(struct request_queue
*q
)
1573 struct blk_mq_hw_ctx
*hctx
;
1576 queue_for_each_hw_ctx(q
, hctx
, i
)
1577 blk_mq_start_hw_queue(hctx
);
1579 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1581 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1583 if (!blk_mq_hctx_stopped(hctx
))
1586 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1587 blk_mq_run_hw_queue(hctx
, async
);
1589 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1591 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1593 struct blk_mq_hw_ctx
*hctx
;
1596 queue_for_each_hw_ctx(q
, hctx
, i
)
1597 blk_mq_start_stopped_hw_queue(hctx
, async
);
1599 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1601 static void blk_mq_run_work_fn(struct work_struct
*work
)
1603 struct blk_mq_hw_ctx
*hctx
;
1605 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1608 * If we are stopped, don't run the queue.
1610 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1613 __blk_mq_run_hw_queue(hctx
);
1616 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1620 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1621 enum hctx_type type
= hctx
->type
;
1623 lockdep_assert_held(&ctx
->lock
);
1625 trace_block_rq_insert(hctx
->queue
, rq
);
1628 list_add(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1630 list_add_tail(&rq
->queuelist
, &ctx
->rq_lists
[type
]);
1633 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1636 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1638 lockdep_assert_held(&ctx
->lock
);
1640 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1641 blk_mq_hctx_mark_pending(hctx
, ctx
);
1645 * Should only be used carefully, when the caller knows we want to
1646 * bypass a potential IO scheduler on the target device.
1648 void blk_mq_request_bypass_insert(struct request
*rq
, bool run_queue
)
1650 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1652 spin_lock(&hctx
->lock
);
1653 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1654 spin_unlock(&hctx
->lock
);
1657 blk_mq_run_hw_queue(hctx
, false);
1660 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1661 struct list_head
*list
)
1665 enum hctx_type type
= hctx
->type
;
1668 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1671 list_for_each_entry(rq
, list
, queuelist
) {
1672 BUG_ON(rq
->mq_ctx
!= ctx
);
1673 trace_block_rq_insert(hctx
->queue
, rq
);
1676 spin_lock(&ctx
->lock
);
1677 list_splice_tail_init(list
, &ctx
->rq_lists
[type
]);
1678 blk_mq_hctx_mark_pending(hctx
, ctx
);
1679 spin_unlock(&ctx
->lock
);
1682 static int plug_rq_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1684 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1685 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1687 if (rqa
->mq_ctx
< rqb
->mq_ctx
)
1689 else if (rqa
->mq_ctx
> rqb
->mq_ctx
)
1691 else if (rqa
->mq_hctx
< rqb
->mq_hctx
)
1693 else if (rqa
->mq_hctx
> rqb
->mq_hctx
)
1696 return blk_rq_pos(rqa
) > blk_rq_pos(rqb
);
1699 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1701 struct blk_mq_hw_ctx
*this_hctx
;
1702 struct blk_mq_ctx
*this_ctx
;
1703 struct request_queue
*this_q
;
1709 list_splice_init(&plug
->mq_list
, &list
);
1712 if (plug
->rq_count
> 2 && plug
->multiple_queues
)
1713 list_sort(NULL
, &list
, plug_rq_cmp
);
1720 while (!list_empty(&list
)) {
1721 rq
= list_entry_rq(list
.next
);
1722 list_del_init(&rq
->queuelist
);
1724 if (rq
->mq_hctx
!= this_hctx
|| rq
->mq_ctx
!= this_ctx
) {
1726 trace_block_unplug(this_q
, depth
, !from_schedule
);
1727 blk_mq_sched_insert_requests(this_hctx
, this_ctx
,
1733 this_ctx
= rq
->mq_ctx
;
1734 this_hctx
= rq
->mq_hctx
;
1739 list_add_tail(&rq
->queuelist
, &rq_list
);
1743 * If 'this_hctx' is set, we know we have entries to complete
1744 * on 'rq_list'. Do those.
1747 trace_block_unplug(this_q
, depth
, !from_schedule
);
1748 blk_mq_sched_insert_requests(this_hctx
, this_ctx
, &rq_list
,
1753 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1755 blk_init_request_from_bio(rq
, bio
);
1757 blk_account_io_start(rq
, true);
1760 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1762 blk_qc_t
*cookie
, bool last
)
1764 struct request_queue
*q
= rq
->q
;
1765 struct blk_mq_queue_data bd
= {
1769 blk_qc_t new_cookie
;
1772 new_cookie
= request_to_qc_t(hctx
, rq
);
1775 * For OK queue, we are done. For error, caller may kill it.
1776 * Any other error (busy), just add it to our list as we
1777 * previously would have done.
1779 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1782 blk_mq_update_dispatch_busy(hctx
, false);
1783 *cookie
= new_cookie
;
1785 case BLK_STS_RESOURCE
:
1786 case BLK_STS_DEV_RESOURCE
:
1787 blk_mq_update_dispatch_busy(hctx
, true);
1788 __blk_mq_requeue_request(rq
);
1791 blk_mq_update_dispatch_busy(hctx
, false);
1792 *cookie
= BLK_QC_T_NONE
;
1799 blk_status_t
blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1802 bool bypass
, bool last
)
1804 struct request_queue
*q
= rq
->q
;
1805 bool run_queue
= true;
1806 blk_status_t ret
= BLK_STS_RESOURCE
;
1810 hctx_lock(hctx
, &srcu_idx
);
1812 * hctx_lock is needed before checking quiesced flag.
1814 * When queue is stopped or quiesced, ignore 'bypass', insert
1815 * and return BLK_STS_OK to caller, and avoid driver to try to
1818 if (unlikely(blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
))) {
1824 if (unlikely(q
->elevator
&& !bypass
))
1827 if (!blk_mq_get_dispatch_budget(hctx
))
1830 if (!blk_mq_get_driver_tag(rq
)) {
1831 blk_mq_put_dispatch_budget(hctx
);
1836 * Always add a request that has been through
1837 *.queue_rq() to the hardware dispatch list.
1840 ret
= __blk_mq_issue_directly(hctx
, rq
, cookie
, last
);
1842 hctx_unlock(hctx
, srcu_idx
);
1846 case BLK_STS_DEV_RESOURCE
:
1847 case BLK_STS_RESOURCE
:
1849 blk_mq_request_bypass_insert(rq
, run_queue
);
1851 * We have to return BLK_STS_OK for the DM
1852 * to avoid livelock. Otherwise, we return
1853 * the real result to indicate whether the
1854 * request is direct-issued successfully.
1856 ret
= bypass
? BLK_STS_OK
: ret
;
1857 } else if (!bypass
) {
1858 blk_mq_sched_insert_request(rq
, false,
1864 blk_mq_end_request(rq
, ret
);
1871 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
1872 struct list_head
*list
)
1875 blk_status_t ret
= BLK_STS_OK
;
1877 while (!list_empty(list
)) {
1878 struct request
*rq
= list_first_entry(list
, struct request
,
1881 list_del_init(&rq
->queuelist
);
1882 if (ret
== BLK_STS_OK
)
1883 ret
= blk_mq_try_issue_directly(hctx
, rq
, &unused
,
1887 blk_mq_sched_insert_request(rq
, false, true, false);
1891 * If we didn't flush the entire list, we could have told
1892 * the driver there was more coming, but that turned out to
1895 if (ret
!= BLK_STS_OK
&& hctx
->queue
->mq_ops
->commit_rqs
)
1896 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
1899 static void blk_add_rq_to_plug(struct blk_plug
*plug
, struct request
*rq
)
1901 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1903 if (!plug
->multiple_queues
&& !list_is_singular(&plug
->mq_list
)) {
1904 struct request
*tmp
;
1906 tmp
= list_first_entry(&plug
->mq_list
, struct request
,
1908 if (tmp
->q
!= rq
->q
)
1909 plug
->multiple_queues
= true;
1913 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1915 const int is_sync
= op_is_sync(bio
->bi_opf
);
1916 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1917 struct blk_mq_alloc_data data
= { .flags
= 0};
1919 struct blk_plug
*plug
;
1920 struct request
*same_queue_rq
= NULL
;
1923 blk_queue_bounce(q
, &bio
);
1925 blk_queue_split(q
, &bio
);
1927 if (!bio_integrity_prep(bio
))
1928 return BLK_QC_T_NONE
;
1930 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1931 blk_attempt_plug_merge(q
, bio
, &same_queue_rq
))
1932 return BLK_QC_T_NONE
;
1934 if (blk_mq_sched_bio_merge(q
, bio
))
1935 return BLK_QC_T_NONE
;
1937 rq_qos_throttle(q
, bio
);
1939 data
.cmd_flags
= bio
->bi_opf
;
1940 rq
= blk_mq_get_request(q
, bio
, &data
);
1941 if (unlikely(!rq
)) {
1942 rq_qos_cleanup(q
, bio
);
1943 if (bio
->bi_opf
& REQ_NOWAIT
)
1944 bio_wouldblock_error(bio
);
1945 return BLK_QC_T_NONE
;
1948 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1950 rq_qos_track(q
, rq
, bio
);
1952 cookie
= request_to_qc_t(data
.hctx
, rq
);
1954 plug
= current
->plug
;
1955 if (unlikely(is_flush_fua
)) {
1956 blk_mq_put_ctx(data
.ctx
);
1957 blk_mq_bio_to_request(rq
, bio
);
1959 /* bypass scheduler for flush rq */
1960 blk_insert_flush(rq
);
1961 blk_mq_run_hw_queue(data
.hctx
, true);
1962 } else if (plug
&& (q
->nr_hw_queues
== 1 || q
->mq_ops
->commit_rqs
)) {
1964 * Use plugging if we have a ->commit_rqs() hook as well, as
1965 * we know the driver uses bd->last in a smart fashion.
1967 unsigned int request_count
= plug
->rq_count
;
1968 struct request
*last
= NULL
;
1970 blk_mq_put_ctx(data
.ctx
);
1971 blk_mq_bio_to_request(rq
, bio
);
1974 trace_block_plug(q
);
1976 last
= list_entry_rq(plug
->mq_list
.prev
);
1978 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1979 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1980 blk_flush_plug_list(plug
, false);
1981 trace_block_plug(q
);
1984 blk_add_rq_to_plug(plug
, rq
);
1985 } else if (plug
&& !blk_queue_nomerges(q
)) {
1986 blk_mq_bio_to_request(rq
, bio
);
1989 * We do limited plugging. If the bio can be merged, do that.
1990 * Otherwise the existing request in the plug list will be
1991 * issued. So the plug list will have one request at most
1992 * The plug list might get flushed before this. If that happens,
1993 * the plug list is empty, and same_queue_rq is invalid.
1995 if (list_empty(&plug
->mq_list
))
1996 same_queue_rq
= NULL
;
1997 if (same_queue_rq
) {
1998 list_del_init(&same_queue_rq
->queuelist
);
2001 blk_add_rq_to_plug(plug
, rq
);
2003 blk_mq_put_ctx(data
.ctx
);
2005 if (same_queue_rq
) {
2006 data
.hctx
= same_queue_rq
->mq_hctx
;
2007 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
2008 &cookie
, false, true);
2010 } else if ((q
->nr_hw_queues
> 1 && is_sync
) || (!q
->elevator
&&
2011 !data
.hctx
->dispatch_busy
)) {
2012 blk_mq_put_ctx(data
.ctx
);
2013 blk_mq_bio_to_request(rq
, bio
);
2014 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
, false, true);
2016 blk_mq_put_ctx(data
.ctx
);
2017 blk_mq_bio_to_request(rq
, bio
);
2018 blk_mq_sched_insert_request(rq
, false, true, true);
2024 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2025 unsigned int hctx_idx
)
2029 if (tags
->rqs
&& set
->ops
->exit_request
) {
2032 for (i
= 0; i
< tags
->nr_tags
; i
++) {
2033 struct request
*rq
= tags
->static_rqs
[i
];
2037 set
->ops
->exit_request(set
, rq
, hctx_idx
);
2038 tags
->static_rqs
[i
] = NULL
;
2042 while (!list_empty(&tags
->page_list
)) {
2043 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
2044 list_del_init(&page
->lru
);
2046 * Remove kmemleak object previously allocated in
2047 * blk_mq_init_rq_map().
2049 kmemleak_free(page_address(page
));
2050 __free_pages(page
, page
->private);
2054 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
2058 kfree(tags
->static_rqs
);
2059 tags
->static_rqs
= NULL
;
2061 blk_mq_free_tags(tags
);
2064 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
2065 unsigned int hctx_idx
,
2066 unsigned int nr_tags
,
2067 unsigned int reserved_tags
)
2069 struct blk_mq_tags
*tags
;
2072 node
= blk_mq_hw_queue_to_node(&set
->map
[0], hctx_idx
);
2073 if (node
== NUMA_NO_NODE
)
2074 node
= set
->numa_node
;
2076 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
2077 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
2081 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2082 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2085 blk_mq_free_tags(tags
);
2089 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2090 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2092 if (!tags
->static_rqs
) {
2094 blk_mq_free_tags(tags
);
2101 static size_t order_to_size(unsigned int order
)
2103 return (size_t)PAGE_SIZE
<< order
;
2106 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
2107 unsigned int hctx_idx
, int node
)
2111 if (set
->ops
->init_request
) {
2112 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
2117 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
2121 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2122 unsigned int hctx_idx
, unsigned int depth
)
2124 unsigned int i
, j
, entries_per_page
, max_order
= 4;
2125 size_t rq_size
, left
;
2128 node
= blk_mq_hw_queue_to_node(&set
->map
[0], hctx_idx
);
2129 if (node
== NUMA_NO_NODE
)
2130 node
= set
->numa_node
;
2132 INIT_LIST_HEAD(&tags
->page_list
);
2135 * rq_size is the size of the request plus driver payload, rounded
2136 * to the cacheline size
2138 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
2140 left
= rq_size
* depth
;
2142 for (i
= 0; i
< depth
; ) {
2143 int this_order
= max_order
;
2148 while (this_order
&& left
< order_to_size(this_order
- 1))
2152 page
= alloc_pages_node(node
,
2153 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
2159 if (order_to_size(this_order
) < rq_size
)
2166 page
->private = this_order
;
2167 list_add_tail(&page
->lru
, &tags
->page_list
);
2169 p
= page_address(page
);
2171 * Allow kmemleak to scan these pages as they contain pointers
2172 * to additional allocations like via ops->init_request().
2174 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2175 entries_per_page
= order_to_size(this_order
) / rq_size
;
2176 to_do
= min(entries_per_page
, depth
- i
);
2177 left
-= to_do
* rq_size
;
2178 for (j
= 0; j
< to_do
; j
++) {
2179 struct request
*rq
= p
;
2181 tags
->static_rqs
[i
] = rq
;
2182 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
2183 tags
->static_rqs
[i
] = NULL
;
2194 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2199 * 'cpu' is going away. splice any existing rq_list entries from this
2200 * software queue to the hw queue dispatch list, and ensure that it
2203 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2205 struct blk_mq_hw_ctx
*hctx
;
2206 struct blk_mq_ctx
*ctx
;
2208 enum hctx_type type
;
2210 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2211 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2214 spin_lock(&ctx
->lock
);
2215 if (!list_empty(&ctx
->rq_lists
[type
])) {
2216 list_splice_init(&ctx
->rq_lists
[type
], &tmp
);
2217 blk_mq_hctx_clear_pending(hctx
, ctx
);
2219 spin_unlock(&ctx
->lock
);
2221 if (list_empty(&tmp
))
2224 spin_lock(&hctx
->lock
);
2225 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2226 spin_unlock(&hctx
->lock
);
2228 blk_mq_run_hw_queue(hctx
, true);
2232 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2234 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2238 /* hctx->ctxs will be freed in queue's release handler */
2239 static void blk_mq_exit_hctx(struct request_queue
*q
,
2240 struct blk_mq_tag_set
*set
,
2241 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2243 if (blk_mq_hw_queue_mapped(hctx
))
2244 blk_mq_tag_idle(hctx
);
2246 if (set
->ops
->exit_request
)
2247 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2249 if (set
->ops
->exit_hctx
)
2250 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2252 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2253 cleanup_srcu_struct(hctx
->srcu
);
2255 blk_mq_remove_cpuhp(hctx
);
2256 blk_free_flush_queue(hctx
->fq
);
2257 sbitmap_free(&hctx
->ctx_map
);
2260 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2261 struct blk_mq_tag_set
*set
, int nr_queue
)
2263 struct blk_mq_hw_ctx
*hctx
;
2266 queue_for_each_hw_ctx(q
, hctx
, i
) {
2269 blk_mq_debugfs_unregister_hctx(hctx
);
2270 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2274 static int blk_mq_init_hctx(struct request_queue
*q
,
2275 struct blk_mq_tag_set
*set
,
2276 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2280 node
= hctx
->numa_node
;
2281 if (node
== NUMA_NO_NODE
)
2282 node
= hctx
->numa_node
= set
->numa_node
;
2284 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2285 spin_lock_init(&hctx
->lock
);
2286 INIT_LIST_HEAD(&hctx
->dispatch
);
2288 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
2290 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2292 hctx
->tags
= set
->tags
[hctx_idx
];
2295 * Allocate space for all possible cpus to avoid allocation at
2298 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2299 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
, node
);
2301 goto unregister_cpu_notifier
;
2303 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
2304 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
, node
))
2309 spin_lock_init(&hctx
->dispatch_wait_lock
);
2310 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2311 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2313 if (set
->ops
->init_hctx
&&
2314 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2317 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
,
2318 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
2322 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
, node
))
2325 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2326 init_srcu_struct(hctx
->srcu
);
2333 if (set
->ops
->exit_hctx
)
2334 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2336 sbitmap_free(&hctx
->ctx_map
);
2339 unregister_cpu_notifier
:
2340 blk_mq_remove_cpuhp(hctx
);
2344 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2345 unsigned int nr_hw_queues
)
2347 struct blk_mq_tag_set
*set
= q
->tag_set
;
2350 for_each_possible_cpu(i
) {
2351 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2352 struct blk_mq_hw_ctx
*hctx
;
2356 spin_lock_init(&__ctx
->lock
);
2357 for (k
= HCTX_TYPE_DEFAULT
; k
< HCTX_MAX_TYPES
; k
++)
2358 INIT_LIST_HEAD(&__ctx
->rq_lists
[k
]);
2363 * Set local node, IFF we have more than one hw queue. If
2364 * not, we remain on the home node of the device
2366 for (j
= 0; j
< set
->nr_maps
; j
++) {
2367 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2368 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2369 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2374 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2378 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2379 set
->queue_depth
, set
->reserved_tags
);
2380 if (!set
->tags
[hctx_idx
])
2383 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2388 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2389 set
->tags
[hctx_idx
] = NULL
;
2393 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2394 unsigned int hctx_idx
)
2396 if (set
->tags
&& set
->tags
[hctx_idx
]) {
2397 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2398 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2399 set
->tags
[hctx_idx
] = NULL
;
2403 static void blk_mq_map_swqueue(struct request_queue
*q
)
2405 unsigned int i
, j
, hctx_idx
;
2406 struct blk_mq_hw_ctx
*hctx
;
2407 struct blk_mq_ctx
*ctx
;
2408 struct blk_mq_tag_set
*set
= q
->tag_set
;
2411 * Avoid others reading imcomplete hctx->cpumask through sysfs
2413 mutex_lock(&q
->sysfs_lock
);
2415 queue_for_each_hw_ctx(q
, hctx
, i
) {
2416 cpumask_clear(hctx
->cpumask
);
2418 hctx
->dispatch_from
= NULL
;
2422 * Map software to hardware queues.
2424 * If the cpu isn't present, the cpu is mapped to first hctx.
2426 for_each_possible_cpu(i
) {
2427 hctx_idx
= set
->map
[0].mq_map
[i
];
2428 /* unmapped hw queue can be remapped after CPU topo changed */
2429 if (!set
->tags
[hctx_idx
] &&
2430 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2432 * If tags initialization fail for some hctx,
2433 * that hctx won't be brought online. In this
2434 * case, remap the current ctx to hctx[0] which
2435 * is guaranteed to always have tags allocated
2437 set
->map
[0].mq_map
[i
] = 0;
2440 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2441 for (j
= 0; j
< set
->nr_maps
; j
++) {
2442 if (!set
->map
[j
].nr_queues
)
2445 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2448 * If the CPU is already set in the mask, then we've
2449 * mapped this one already. This can happen if
2450 * devices share queues across queue maps.
2452 if (cpumask_test_cpu(i
, hctx
->cpumask
))
2455 cpumask_set_cpu(i
, hctx
->cpumask
);
2457 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
2458 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2461 * If the nr_ctx type overflows, we have exceeded the
2462 * amount of sw queues we can support.
2464 BUG_ON(!hctx
->nr_ctx
);
2468 mutex_unlock(&q
->sysfs_lock
);
2470 queue_for_each_hw_ctx(q
, hctx
, i
) {
2472 * If no software queues are mapped to this hardware queue,
2473 * disable it and free the request entries.
2475 if (!hctx
->nr_ctx
) {
2476 /* Never unmap queue 0. We need it as a
2477 * fallback in case of a new remap fails
2480 if (i
&& set
->tags
[i
])
2481 blk_mq_free_map_and_requests(set
, i
);
2487 hctx
->tags
= set
->tags
[i
];
2488 WARN_ON(!hctx
->tags
);
2491 * Set the map size to the number of mapped software queues.
2492 * This is more accurate and more efficient than looping
2493 * over all possibly mapped software queues.
2495 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2498 * Initialize batch roundrobin counts
2500 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2501 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2506 * Caller needs to ensure that we're either frozen/quiesced, or that
2507 * the queue isn't live yet.
2509 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2511 struct blk_mq_hw_ctx
*hctx
;
2514 queue_for_each_hw_ctx(q
, hctx
, i
) {
2516 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2518 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2522 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2525 struct request_queue
*q
;
2527 lockdep_assert_held(&set
->tag_list_lock
);
2529 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2530 blk_mq_freeze_queue(q
);
2531 queue_set_hctx_shared(q
, shared
);
2532 blk_mq_unfreeze_queue(q
);
2536 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2538 struct blk_mq_tag_set
*set
= q
->tag_set
;
2540 mutex_lock(&set
->tag_list_lock
);
2541 list_del_rcu(&q
->tag_set_list
);
2542 if (list_is_singular(&set
->tag_list
)) {
2543 /* just transitioned to unshared */
2544 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2545 /* update existing queue */
2546 blk_mq_update_tag_set_depth(set
, false);
2548 mutex_unlock(&set
->tag_list_lock
);
2549 INIT_LIST_HEAD(&q
->tag_set_list
);
2552 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2553 struct request_queue
*q
)
2555 mutex_lock(&set
->tag_list_lock
);
2558 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2560 if (!list_empty(&set
->tag_list
) &&
2561 !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2562 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2563 /* update existing queue */
2564 blk_mq_update_tag_set_depth(set
, true);
2566 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2567 queue_set_hctx_shared(q
, true);
2568 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2570 mutex_unlock(&set
->tag_list_lock
);
2573 /* All allocations will be freed in release handler of q->mq_kobj */
2574 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
2576 struct blk_mq_ctxs
*ctxs
;
2579 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
2583 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2584 if (!ctxs
->queue_ctx
)
2587 for_each_possible_cpu(cpu
) {
2588 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
2592 q
->mq_kobj
= &ctxs
->kobj
;
2593 q
->queue_ctx
= ctxs
->queue_ctx
;
2602 * It is the actual release handler for mq, but we do it from
2603 * request queue's release handler for avoiding use-after-free
2604 * and headache because q->mq_kobj shouldn't have been introduced,
2605 * but we can't group ctx/kctx kobj without it.
2607 void blk_mq_release(struct request_queue
*q
)
2609 struct blk_mq_hw_ctx
*hctx
;
2612 /* hctx kobj stays in hctx */
2613 queue_for_each_hw_ctx(q
, hctx
, i
) {
2616 kobject_put(&hctx
->kobj
);
2619 kfree(q
->queue_hw_ctx
);
2622 * release .mq_kobj and sw queue's kobject now because
2623 * both share lifetime with request queue.
2625 blk_mq_sysfs_deinit(q
);
2628 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2630 struct request_queue
*uninit_q
, *q
;
2632 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2634 return ERR_PTR(-ENOMEM
);
2636 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2638 blk_cleanup_queue(uninit_q
);
2642 EXPORT_SYMBOL(blk_mq_init_queue
);
2645 * Helper for setting up a queue with mq ops, given queue depth, and
2646 * the passed in mq ops flags.
2648 struct request_queue
*blk_mq_init_sq_queue(struct blk_mq_tag_set
*set
,
2649 const struct blk_mq_ops
*ops
,
2650 unsigned int queue_depth
,
2651 unsigned int set_flags
)
2653 struct request_queue
*q
;
2656 memset(set
, 0, sizeof(*set
));
2658 set
->nr_hw_queues
= 1;
2660 set
->queue_depth
= queue_depth
;
2661 set
->numa_node
= NUMA_NO_NODE
;
2662 set
->flags
= set_flags
;
2664 ret
= blk_mq_alloc_tag_set(set
);
2666 return ERR_PTR(ret
);
2668 q
= blk_mq_init_queue(set
);
2670 blk_mq_free_tag_set(set
);
2676 EXPORT_SYMBOL(blk_mq_init_sq_queue
);
2678 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2680 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2682 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, srcu
),
2683 __alignof__(struct blk_mq_hw_ctx
)) !=
2684 sizeof(struct blk_mq_hw_ctx
));
2686 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2687 hw_ctx_size
+= sizeof(struct srcu_struct
);
2692 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
2693 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
2694 int hctx_idx
, int node
)
2696 struct blk_mq_hw_ctx
*hctx
;
2698 hctx
= kzalloc_node(blk_mq_hw_ctx_size(set
),
2699 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2704 if (!zalloc_cpumask_var_node(&hctx
->cpumask
,
2705 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2711 atomic_set(&hctx
->nr_active
, 0);
2712 hctx
->numa_node
= node
;
2713 hctx
->queue_num
= hctx_idx
;
2715 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
)) {
2716 free_cpumask_var(hctx
->cpumask
);
2720 blk_mq_hctx_kobj_init(hctx
);
2725 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2726 struct request_queue
*q
)
2729 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2731 /* protect against switching io scheduler */
2732 mutex_lock(&q
->sysfs_lock
);
2733 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2735 struct blk_mq_hw_ctx
*hctx
;
2737 node
= blk_mq_hw_queue_to_node(&set
->map
[0], i
);
2739 * If the hw queue has been mapped to another numa node,
2740 * we need to realloc the hctx. If allocation fails, fallback
2741 * to use the previous one.
2743 if (hctxs
[i
] && (hctxs
[i
]->numa_node
== node
))
2746 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, node
);
2749 blk_mq_exit_hctx(q
, set
, hctxs
[i
], i
);
2750 kobject_put(&hctxs
[i
]->kobj
);
2755 pr_warn("Allocate new hctx on node %d fails,\
2756 fallback to previous one on node %d\n",
2757 node
, hctxs
[i
]->numa_node
);
2763 * Increasing nr_hw_queues fails. Free the newly allocated
2764 * hctxs and keep the previous q->nr_hw_queues.
2766 if (i
!= set
->nr_hw_queues
) {
2767 j
= q
->nr_hw_queues
;
2771 end
= q
->nr_hw_queues
;
2772 q
->nr_hw_queues
= set
->nr_hw_queues
;
2775 for (; j
< end
; j
++) {
2776 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2780 blk_mq_free_map_and_requests(set
, j
);
2781 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2782 kobject_put(&hctx
->kobj
);
2787 mutex_unlock(&q
->sysfs_lock
);
2791 * Maximum number of hardware queues we support. For single sets, we'll never
2792 * have more than the CPUs (software queues). For multiple sets, the tag_set
2793 * user may have set ->nr_hw_queues larger.
2795 static unsigned int nr_hw_queues(struct blk_mq_tag_set
*set
)
2797 if (set
->nr_maps
== 1)
2800 return max(set
->nr_hw_queues
, nr_cpu_ids
);
2803 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2804 struct request_queue
*q
)
2806 /* mark the queue as mq asap */
2807 q
->mq_ops
= set
->ops
;
2809 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2810 blk_mq_poll_stats_bkt
,
2811 BLK_MQ_POLL_STATS_BKTS
, q
);
2815 if (blk_mq_alloc_ctxs(q
))
2818 /* init q->mq_kobj and sw queues' kobjects */
2819 blk_mq_sysfs_init(q
);
2821 q
->nr_queues
= nr_hw_queues(set
);
2822 q
->queue_hw_ctx
= kcalloc_node(q
->nr_queues
, sizeof(*(q
->queue_hw_ctx
)),
2823 GFP_KERNEL
, set
->numa_node
);
2824 if (!q
->queue_hw_ctx
)
2827 blk_mq_realloc_hw_ctxs(set
, q
);
2828 if (!q
->nr_hw_queues
)
2831 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2832 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2836 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2837 if (set
->nr_maps
> HCTX_TYPE_POLL
&&
2838 set
->map
[HCTX_TYPE_POLL
].nr_queues
)
2839 blk_queue_flag_set(QUEUE_FLAG_POLL
, q
);
2841 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2842 blk_queue_flag_set(QUEUE_FLAG_NO_SG_MERGE
, q
);
2844 q
->sg_reserved_size
= INT_MAX
;
2846 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2847 INIT_LIST_HEAD(&q
->requeue_list
);
2848 spin_lock_init(&q
->requeue_lock
);
2850 blk_queue_make_request(q
, blk_mq_make_request
);
2853 * Do this after blk_queue_make_request() overrides it...
2855 q
->nr_requests
= set
->queue_depth
;
2858 * Default to classic polling
2862 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2863 blk_mq_add_queue_tag_set(set
, q
);
2864 blk_mq_map_swqueue(q
);
2866 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2869 ret
= elevator_init_mq(q
);
2871 return ERR_PTR(ret
);
2877 kfree(q
->queue_hw_ctx
);
2879 blk_mq_sysfs_deinit(q
);
2882 return ERR_PTR(-ENOMEM
);
2884 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2886 void blk_mq_free_queue(struct request_queue
*q
)
2888 struct blk_mq_tag_set
*set
= q
->tag_set
;
2890 blk_mq_del_queue_tag_set(q
);
2891 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2894 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2898 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2899 if (!__blk_mq_alloc_rq_map(set
, i
))
2906 blk_mq_free_rq_map(set
->tags
[i
]);
2912 * Allocate the request maps associated with this tag_set. Note that this
2913 * may reduce the depth asked for, if memory is tight. set->queue_depth
2914 * will be updated to reflect the allocated depth.
2916 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2921 depth
= set
->queue_depth
;
2923 err
= __blk_mq_alloc_rq_maps(set
);
2927 set
->queue_depth
>>= 1;
2928 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2932 } while (set
->queue_depth
);
2934 if (!set
->queue_depth
|| err
) {
2935 pr_err("blk-mq: failed to allocate request map\n");
2939 if (depth
!= set
->queue_depth
)
2940 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2941 depth
, set
->queue_depth
);
2946 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2948 if (set
->ops
->map_queues
&& !is_kdump_kernel()) {
2952 * transport .map_queues is usually done in the following
2955 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2956 * mask = get_cpu_mask(queue)
2957 * for_each_cpu(cpu, mask)
2958 * set->map[x].mq_map[cpu] = queue;
2961 * When we need to remap, the table has to be cleared for
2962 * killing stale mapping since one CPU may not be mapped
2965 for (i
= 0; i
< set
->nr_maps
; i
++)
2966 blk_mq_clear_mq_map(&set
->map
[i
]);
2968 return set
->ops
->map_queues(set
);
2970 BUG_ON(set
->nr_maps
> 1);
2971 return blk_mq_map_queues(&set
->map
[0]);
2976 * Alloc a tag set to be associated with one or more request queues.
2977 * May fail with EINVAL for various error conditions. May adjust the
2978 * requested depth down, if it's too large. In that case, the set
2979 * value will be stored in set->queue_depth.
2981 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2985 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2987 if (!set
->nr_hw_queues
)
2989 if (!set
->queue_depth
)
2991 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2994 if (!set
->ops
->queue_rq
)
2997 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
3000 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
3001 pr_info("blk-mq: reduced tag depth to %u\n",
3003 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
3008 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
3012 * If a crashdump is active, then we are potentially in a very
3013 * memory constrained environment. Limit us to 1 queue and
3014 * 64 tags to prevent using too much memory.
3016 if (is_kdump_kernel()) {
3017 set
->nr_hw_queues
= 1;
3019 set
->queue_depth
= min(64U, set
->queue_depth
);
3022 * There is no use for more h/w queues than cpus if we just have
3025 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
3026 set
->nr_hw_queues
= nr_cpu_ids
;
3028 set
->tags
= kcalloc_node(nr_hw_queues(set
), sizeof(struct blk_mq_tags
*),
3029 GFP_KERNEL
, set
->numa_node
);
3034 for (i
= 0; i
< set
->nr_maps
; i
++) {
3035 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
3036 sizeof(set
->map
[i
].mq_map
[0]),
3037 GFP_KERNEL
, set
->numa_node
);
3038 if (!set
->map
[i
].mq_map
)
3039 goto out_free_mq_map
;
3040 set
->map
[i
].nr_queues
= is_kdump_kernel() ? 1 : set
->nr_hw_queues
;
3043 ret
= blk_mq_update_queue_map(set
);
3045 goto out_free_mq_map
;
3047 ret
= blk_mq_alloc_rq_maps(set
);
3049 goto out_free_mq_map
;
3051 mutex_init(&set
->tag_list_lock
);
3052 INIT_LIST_HEAD(&set
->tag_list
);
3057 for (i
= 0; i
< set
->nr_maps
; i
++) {
3058 kfree(set
->map
[i
].mq_map
);
3059 set
->map
[i
].mq_map
= NULL
;
3065 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
3067 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
3071 for (i
= 0; i
< nr_hw_queues(set
); i
++)
3072 blk_mq_free_map_and_requests(set
, i
);
3074 for (j
= 0; j
< set
->nr_maps
; j
++) {
3075 kfree(set
->map
[j
].mq_map
);
3076 set
->map
[j
].mq_map
= NULL
;
3082 EXPORT_SYMBOL(blk_mq_free_tag_set
);
3084 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
3086 struct blk_mq_tag_set
*set
= q
->tag_set
;
3087 struct blk_mq_hw_ctx
*hctx
;
3093 blk_mq_freeze_queue(q
);
3094 blk_mq_quiesce_queue(q
);
3097 queue_for_each_hw_ctx(q
, hctx
, i
) {
3101 * If we're using an MQ scheduler, just update the scheduler
3102 * queue depth. This is similar to what the old code would do.
3104 if (!hctx
->sched_tags
) {
3105 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
3108 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
3116 q
->nr_requests
= nr
;
3118 blk_mq_unquiesce_queue(q
);
3119 blk_mq_unfreeze_queue(q
);
3125 * request_queue and elevator_type pair.
3126 * It is just used by __blk_mq_update_nr_hw_queues to cache
3127 * the elevator_type associated with a request_queue.
3129 struct blk_mq_qe_pair
{
3130 struct list_head node
;
3131 struct request_queue
*q
;
3132 struct elevator_type
*type
;
3136 * Cache the elevator_type in qe pair list and switch the
3137 * io scheduler to 'none'
3139 static bool blk_mq_elv_switch_none(struct list_head
*head
,
3140 struct request_queue
*q
)
3142 struct blk_mq_qe_pair
*qe
;
3147 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
3151 INIT_LIST_HEAD(&qe
->node
);
3153 qe
->type
= q
->elevator
->type
;
3154 list_add(&qe
->node
, head
);
3156 mutex_lock(&q
->sysfs_lock
);
3158 * After elevator_switch_mq, the previous elevator_queue will be
3159 * released by elevator_release. The reference of the io scheduler
3160 * module get by elevator_get will also be put. So we need to get
3161 * a reference of the io scheduler module here to prevent it to be
3164 __module_get(qe
->type
->elevator_owner
);
3165 elevator_switch_mq(q
, NULL
);
3166 mutex_unlock(&q
->sysfs_lock
);
3171 static void blk_mq_elv_switch_back(struct list_head
*head
,
3172 struct request_queue
*q
)
3174 struct blk_mq_qe_pair
*qe
;
3175 struct elevator_type
*t
= NULL
;
3177 list_for_each_entry(qe
, head
, node
)
3186 list_del(&qe
->node
);
3189 mutex_lock(&q
->sysfs_lock
);
3190 elevator_switch_mq(q
, t
);
3191 mutex_unlock(&q
->sysfs_lock
);
3194 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
3197 struct request_queue
*q
;
3199 int prev_nr_hw_queues
;
3201 lockdep_assert_held(&set
->tag_list_lock
);
3203 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
3204 nr_hw_queues
= nr_cpu_ids
;
3205 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
3208 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3209 blk_mq_freeze_queue(q
);
3211 * Sync with blk_mq_queue_tag_busy_iter.
3215 * Switch IO scheduler to 'none', cleaning up the data associated
3216 * with the previous scheduler. We will switch back once we are done
3217 * updating the new sw to hw queue mappings.
3219 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3220 if (!blk_mq_elv_switch_none(&head
, q
))
3223 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3224 blk_mq_debugfs_unregister_hctxs(q
);
3225 blk_mq_sysfs_unregister(q
);
3228 prev_nr_hw_queues
= set
->nr_hw_queues
;
3229 set
->nr_hw_queues
= nr_hw_queues
;
3230 blk_mq_update_queue_map(set
);
3232 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3233 blk_mq_realloc_hw_ctxs(set
, q
);
3234 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
3235 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3236 nr_hw_queues
, prev_nr_hw_queues
);
3237 set
->nr_hw_queues
= prev_nr_hw_queues
;
3238 blk_mq_map_queues(&set
->map
[0]);
3241 blk_mq_map_swqueue(q
);
3244 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3245 blk_mq_sysfs_register(q
);
3246 blk_mq_debugfs_register_hctxs(q
);
3250 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3251 blk_mq_elv_switch_back(&head
, q
);
3253 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3254 blk_mq_unfreeze_queue(q
);
3257 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
3259 mutex_lock(&set
->tag_list_lock
);
3260 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
3261 mutex_unlock(&set
->tag_list_lock
);
3263 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
3265 /* Enable polling stats and return whether they were already enabled. */
3266 static bool blk_poll_stats_enable(struct request_queue
*q
)
3268 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3269 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS
, q
))
3271 blk_stat_add_callback(q
, q
->poll_cb
);
3275 static void blk_mq_poll_stats_start(struct request_queue
*q
)
3278 * We don't arm the callback if polling stats are not enabled or the
3279 * callback is already active.
3281 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3282 blk_stat_is_active(q
->poll_cb
))
3285 blk_stat_activate_msecs(q
->poll_cb
, 100);
3288 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
3290 struct request_queue
*q
= cb
->data
;
3293 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
3294 if (cb
->stat
[bucket
].nr_samples
)
3295 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
3299 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
3300 struct blk_mq_hw_ctx
*hctx
,
3303 unsigned long ret
= 0;
3307 * If stats collection isn't on, don't sleep but turn it on for
3310 if (!blk_poll_stats_enable(q
))
3314 * As an optimistic guess, use half of the mean service time
3315 * for this type of request. We can (and should) make this smarter.
3316 * For instance, if the completion latencies are tight, we can
3317 * get closer than just half the mean. This is especially
3318 * important on devices where the completion latencies are longer
3319 * than ~10 usec. We do use the stats for the relevant IO size
3320 * if available which does lead to better estimates.
3322 bucket
= blk_mq_poll_stats_bkt(rq
);
3326 if (q
->poll_stat
[bucket
].nr_samples
)
3327 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
3332 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
3333 struct blk_mq_hw_ctx
*hctx
,
3336 struct hrtimer_sleeper hs
;
3337 enum hrtimer_mode mode
;
3341 if (rq
->rq_flags
& RQF_MQ_POLL_SLEPT
)
3345 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3347 * 0: use half of prev avg
3348 * >0: use this specific value
3350 if (q
->poll_nsec
> 0)
3351 nsecs
= q
->poll_nsec
;
3353 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
3358 rq
->rq_flags
|= RQF_MQ_POLL_SLEPT
;
3361 * This will be replaced with the stats tracking code, using
3362 * 'avg_completion_time / 2' as the pre-sleep target.
3366 mode
= HRTIMER_MODE_REL
;
3367 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
3368 hrtimer_set_expires(&hs
.timer
, kt
);
3370 hrtimer_init_sleeper(&hs
, current
);
3372 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
3374 set_current_state(TASK_UNINTERRUPTIBLE
);
3375 hrtimer_start_expires(&hs
.timer
, mode
);
3378 hrtimer_cancel(&hs
.timer
);
3379 mode
= HRTIMER_MODE_ABS
;
3380 } while (hs
.task
&& !signal_pending(current
));
3382 __set_current_state(TASK_RUNNING
);
3383 destroy_hrtimer_on_stack(&hs
.timer
);
3387 static bool blk_mq_poll_hybrid(struct request_queue
*q
,
3388 struct blk_mq_hw_ctx
*hctx
, blk_qc_t cookie
)
3392 if (q
->poll_nsec
== -1)
3395 if (!blk_qc_t_is_internal(cookie
))
3396 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3398 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3400 * With scheduling, if the request has completed, we'll
3401 * get a NULL return here, as we clear the sched tag when
3402 * that happens. The request still remains valid, like always,
3403 * so we should be safe with just the NULL check.
3409 return blk_mq_poll_hybrid_sleep(q
, hctx
, rq
);
3413 * blk_poll - poll for IO completions
3415 * @cookie: cookie passed back at IO submission time
3416 * @spin: whether to spin for completions
3419 * Poll for completions on the passed in queue. Returns number of
3420 * completed entries found. If @spin is true, then blk_poll will continue
3421 * looping until at least one completion is found, unless the task is
3422 * otherwise marked running (or we need to reschedule).
3424 int blk_poll(struct request_queue
*q
, blk_qc_t cookie
, bool spin
)
3426 struct blk_mq_hw_ctx
*hctx
;
3429 if (!blk_qc_t_valid(cookie
) ||
3430 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3434 blk_flush_plug_list(current
->plug
, false);
3436 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3439 * If we sleep, have the caller restart the poll loop to reset
3440 * the state. Like for the other success return cases, the
3441 * caller is responsible for checking if the IO completed. If
3442 * the IO isn't complete, we'll get called again and will go
3443 * straight to the busy poll loop.
3445 if (blk_mq_poll_hybrid(q
, hctx
, cookie
))
3448 hctx
->poll_considered
++;
3450 state
= current
->state
;
3454 hctx
->poll_invoked
++;
3456 ret
= q
->mq_ops
->poll(hctx
);
3458 hctx
->poll_success
++;
3459 __set_current_state(TASK_RUNNING
);
3463 if (signal_pending_state(state
, current
))
3464 __set_current_state(TASK_RUNNING
);
3466 if (current
->state
== TASK_RUNNING
)
3468 if (ret
< 0 || !spin
)
3471 } while (!need_resched());
3473 __set_current_state(TASK_RUNNING
);
3476 EXPORT_SYMBOL_GPL(blk_poll
);
3478 unsigned int blk_mq_rq_cpu(struct request
*rq
)
3480 return rq
->mq_ctx
->cpu
;
3482 EXPORT_SYMBOL(blk_mq_rq_cpu
);
3484 static int __init
blk_mq_init(void)
3486 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3487 blk_mq_hctx_notify_dead
);
3490 subsys_initcall(blk_mq_init
);