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. index[1]
104 * counts the ones that are active on the whole device, so increment
105 * that if mi->part is indeed a partition, and not a whole device.
107 if (rq
->part
== mi
->part
)
109 if (mi
->part
->partno
)
115 void blk_mq_in_flight(struct request_queue
*q
, struct hd_struct
*part
,
116 unsigned int inflight
[2])
118 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
120 inflight
[0] = inflight
[1] = 0;
121 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
124 static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx
*hctx
,
125 struct request
*rq
, void *priv
,
128 struct mq_inflight
*mi
= priv
;
130 if (rq
->part
== mi
->part
)
131 mi
->inflight
[rq_data_dir(rq
)]++;
136 void blk_mq_in_flight_rw(struct request_queue
*q
, struct hd_struct
*part
,
137 unsigned int inflight
[2])
139 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
141 inflight
[0] = inflight
[1] = 0;
142 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight_rw
, &mi
);
145 void blk_freeze_queue_start(struct request_queue
*q
)
149 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
150 if (freeze_depth
== 1) {
151 percpu_ref_kill(&q
->q_usage_counter
);
153 blk_mq_run_hw_queues(q
, false);
156 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
158 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
160 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
162 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
164 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
165 unsigned long timeout
)
167 return wait_event_timeout(q
->mq_freeze_wq
,
168 percpu_ref_is_zero(&q
->q_usage_counter
),
171 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
174 * Guarantee no request is in use, so we can change any data structure of
175 * the queue afterward.
177 void blk_freeze_queue(struct request_queue
*q
)
180 * In the !blk_mq case we are only calling this to kill the
181 * q_usage_counter, otherwise this increases the freeze depth
182 * and waits for it to return to zero. For this reason there is
183 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
184 * exported to drivers as the only user for unfreeze is blk_mq.
186 blk_freeze_queue_start(q
);
187 blk_mq_freeze_queue_wait(q
);
190 void blk_mq_freeze_queue(struct request_queue
*q
)
193 * ...just an alias to keep freeze and unfreeze actions balanced
194 * in the blk_mq_* namespace
198 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
200 void blk_mq_unfreeze_queue(struct request_queue
*q
)
204 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
205 WARN_ON_ONCE(freeze_depth
< 0);
207 percpu_ref_resurrect(&q
->q_usage_counter
);
208 wake_up_all(&q
->mq_freeze_wq
);
211 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
214 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
215 * mpt3sas driver such that this function can be removed.
217 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
219 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
221 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
224 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
227 * Note: this function does not prevent that the struct request end_io()
228 * callback function is invoked. Once this function is returned, we make
229 * sure no dispatch can happen until the queue is unquiesced via
230 * blk_mq_unquiesce_queue().
232 void blk_mq_quiesce_queue(struct request_queue
*q
)
234 struct blk_mq_hw_ctx
*hctx
;
238 blk_mq_quiesce_queue_nowait(q
);
240 queue_for_each_hw_ctx(q
, hctx
, i
) {
241 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
242 synchronize_srcu(hctx
->srcu
);
249 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
252 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
255 * This function recovers queue into the state before quiescing
256 * which is done by blk_mq_quiesce_queue.
258 void blk_mq_unquiesce_queue(struct request_queue
*q
)
260 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
262 /* dispatch requests which are inserted during quiescing */
263 blk_mq_run_hw_queues(q
, true);
265 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
267 void blk_mq_wake_waiters(struct request_queue
*q
)
269 struct blk_mq_hw_ctx
*hctx
;
272 queue_for_each_hw_ctx(q
, hctx
, i
)
273 if (blk_mq_hw_queue_mapped(hctx
))
274 blk_mq_tag_wakeup_all(hctx
->tags
, true);
277 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
279 return blk_mq_has_free_tags(hctx
->tags
);
281 EXPORT_SYMBOL(blk_mq_can_queue
);
284 * Only need start/end time stamping if we have stats enabled, or using
287 static inline bool blk_mq_need_time_stamp(struct request
*rq
)
289 return (rq
->rq_flags
& RQF_IO_STAT
) || rq
->q
->elevator
;
292 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
293 unsigned int tag
, unsigned int op
)
295 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
296 struct request
*rq
= tags
->static_rqs
[tag
];
297 req_flags_t rq_flags
= 0;
299 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
301 rq
->internal_tag
= tag
;
303 if (data
->hctx
->flags
& BLK_MQ_F_TAG_SHARED
) {
304 rq_flags
= RQF_MQ_INFLIGHT
;
305 atomic_inc(&data
->hctx
->nr_active
);
308 rq
->internal_tag
= -1;
309 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
312 /* csd/requeue_work/fifo_time is initialized before use */
314 rq
->mq_ctx
= data
->ctx
;
315 rq
->mq_hctx
= data
->hctx
;
316 rq
->rq_flags
= rq_flags
;
318 if (data
->flags
& BLK_MQ_REQ_PREEMPT
)
319 rq
->rq_flags
|= RQF_PREEMPT
;
320 if (blk_queue_io_stat(data
->q
))
321 rq
->rq_flags
|= RQF_IO_STAT
;
322 INIT_LIST_HEAD(&rq
->queuelist
);
323 INIT_HLIST_NODE(&rq
->hash
);
324 RB_CLEAR_NODE(&rq
->rb_node
);
327 if (blk_mq_need_time_stamp(rq
))
328 rq
->start_time_ns
= ktime_get_ns();
330 rq
->start_time_ns
= 0;
331 rq
->io_start_time_ns
= 0;
332 rq
->nr_phys_segments
= 0;
333 #if defined(CONFIG_BLK_DEV_INTEGRITY)
334 rq
->nr_integrity_segments
= 0;
337 /* tag was already set */
339 WRITE_ONCE(rq
->deadline
, 0);
344 rq
->end_io_data
= NULL
;
347 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
348 refcount_set(&rq
->ref
, 1);
352 static struct request
*blk_mq_get_request(struct request_queue
*q
,
354 struct blk_mq_alloc_data
*data
)
356 struct elevator_queue
*e
= q
->elevator
;
359 bool put_ctx_on_error
= false;
361 blk_queue_enter_live(q
);
363 if (likely(!data
->ctx
)) {
364 data
->ctx
= blk_mq_get_ctx(q
);
365 put_ctx_on_error
= true;
367 if (likely(!data
->hctx
))
368 data
->hctx
= blk_mq_map_queue(q
, data
->cmd_flags
,
370 if (data
->cmd_flags
& REQ_NOWAIT
)
371 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
374 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
377 * Flush requests are special and go directly to the
378 * dispatch list. Don't include reserved tags in the
379 * limiting, as it isn't useful.
381 if (!op_is_flush(data
->cmd_flags
) &&
382 e
->type
->ops
.limit_depth
&&
383 !(data
->flags
& BLK_MQ_REQ_RESERVED
))
384 e
->type
->ops
.limit_depth(data
->cmd_flags
, data
);
386 blk_mq_tag_busy(data
->hctx
);
389 tag
= blk_mq_get_tag(data
);
390 if (tag
== BLK_MQ_TAG_FAIL
) {
391 if (put_ctx_on_error
) {
392 blk_mq_put_ctx(data
->ctx
);
399 rq
= blk_mq_rq_ctx_init(data
, tag
, data
->cmd_flags
);
400 if (!op_is_flush(data
->cmd_flags
)) {
402 if (e
&& e
->type
->ops
.prepare_request
) {
403 if (e
->type
->icq_cache
)
404 blk_mq_sched_assign_ioc(rq
);
406 e
->type
->ops
.prepare_request(rq
, bio
);
407 rq
->rq_flags
|= RQF_ELVPRIV
;
410 data
->hctx
->queued
++;
414 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
415 blk_mq_req_flags_t flags
)
417 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
, .cmd_flags
= op
};
421 ret
= blk_queue_enter(q
, flags
);
425 rq
= blk_mq_get_request(q
, NULL
, &alloc_data
);
429 return ERR_PTR(-EWOULDBLOCK
);
431 blk_mq_put_ctx(alloc_data
.ctx
);
434 rq
->__sector
= (sector_t
) -1;
435 rq
->bio
= rq
->biotail
= NULL
;
438 EXPORT_SYMBOL(blk_mq_alloc_request
);
440 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
441 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
443 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
, .cmd_flags
= op
};
449 * If the tag allocator sleeps we could get an allocation for a
450 * different hardware context. No need to complicate the low level
451 * allocator for this for the rare use case of a command tied to
454 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
455 return ERR_PTR(-EINVAL
);
457 if (hctx_idx
>= q
->nr_hw_queues
)
458 return ERR_PTR(-EIO
);
460 ret
= blk_queue_enter(q
, flags
);
465 * Check if the hardware context is actually mapped to anything.
466 * If not tell the caller that it should skip this queue.
468 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
469 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
471 return ERR_PTR(-EXDEV
);
473 cpu
= cpumask_first_and(alloc_data
.hctx
->cpumask
, cpu_online_mask
);
474 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
476 rq
= blk_mq_get_request(q
, NULL
, &alloc_data
);
480 return ERR_PTR(-EWOULDBLOCK
);
484 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
486 static void __blk_mq_free_request(struct request
*rq
)
488 struct request_queue
*q
= rq
->q
;
489 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
490 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
491 const int sched_tag
= rq
->internal_tag
;
493 blk_pm_mark_last_busy(rq
);
496 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
498 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
499 blk_mq_sched_restart(hctx
);
503 void blk_mq_free_request(struct request
*rq
)
505 struct request_queue
*q
= rq
->q
;
506 struct elevator_queue
*e
= q
->elevator
;
507 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
508 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
510 if (rq
->rq_flags
& RQF_ELVPRIV
) {
511 if (e
&& e
->type
->ops
.finish_request
)
512 e
->type
->ops
.finish_request(rq
);
514 put_io_context(rq
->elv
.icq
->ioc
);
519 ctx
->rq_completed
[rq_is_sync(rq
)]++;
520 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
521 atomic_dec(&hctx
->nr_active
);
523 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
524 laptop_io_completion(q
->backing_dev_info
);
528 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
529 if (refcount_dec_and_test(&rq
->ref
))
530 __blk_mq_free_request(rq
);
532 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
534 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
538 if (blk_mq_need_time_stamp(rq
))
539 now
= ktime_get_ns();
541 if (rq
->rq_flags
& RQF_STATS
) {
542 blk_mq_poll_stats_start(rq
->q
);
543 blk_stat_add(rq
, now
);
546 if (rq
->internal_tag
!= -1)
547 blk_mq_sched_completed_request(rq
, now
);
549 blk_account_io_done(rq
, now
);
552 rq_qos_done(rq
->q
, rq
);
553 rq
->end_io(rq
, error
);
555 if (unlikely(blk_bidi_rq(rq
)))
556 blk_mq_free_request(rq
->next_rq
);
557 blk_mq_free_request(rq
);
560 EXPORT_SYMBOL(__blk_mq_end_request
);
562 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
564 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
566 __blk_mq_end_request(rq
, error
);
568 EXPORT_SYMBOL(blk_mq_end_request
);
570 static void __blk_mq_complete_request_remote(void *data
)
572 struct request
*rq
= data
;
573 struct request_queue
*q
= rq
->q
;
575 q
->mq_ops
->complete(rq
);
578 static void __blk_mq_complete_request(struct request
*rq
)
580 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
581 struct request_queue
*q
= rq
->q
;
585 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
587 * Most of single queue controllers, there is only one irq vector
588 * for handling IO completion, and the only irq's affinity is set
589 * as all possible CPUs. On most of ARCHs, this affinity means the
590 * irq is handled on one specific CPU.
592 * So complete IO reqeust in softirq context in case of single queue
593 * for not degrading IO performance by irqsoff latency.
595 if (q
->nr_hw_queues
== 1) {
596 __blk_complete_request(rq
);
601 * For a polled request, always complete locallly, it's pointless
602 * to redirect the completion.
604 if ((rq
->cmd_flags
& REQ_HIPRI
) ||
605 !test_bit(QUEUE_FLAG_SAME_COMP
, &q
->queue_flags
)) {
606 q
->mq_ops
->complete(rq
);
611 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &q
->queue_flags
))
612 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
614 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
615 rq
->csd
.func
= __blk_mq_complete_request_remote
;
618 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
620 q
->mq_ops
->complete(rq
);
625 static void hctx_unlock(struct blk_mq_hw_ctx
*hctx
, int srcu_idx
)
626 __releases(hctx
->srcu
)
628 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
631 srcu_read_unlock(hctx
->srcu
, srcu_idx
);
634 static void hctx_lock(struct blk_mq_hw_ctx
*hctx
, int *srcu_idx
)
635 __acquires(hctx
->srcu
)
637 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
638 /* shut up gcc false positive */
642 *srcu_idx
= srcu_read_lock(hctx
->srcu
);
646 * blk_mq_complete_request - end I/O on a request
647 * @rq: the request being processed
650 * Ends all I/O on a request. It does not handle partial completions.
651 * The actual completion happens out-of-order, through a IPI handler.
653 bool blk_mq_complete_request(struct request
*rq
)
655 if (unlikely(blk_should_fake_timeout(rq
->q
)))
657 __blk_mq_complete_request(rq
);
660 EXPORT_SYMBOL(blk_mq_complete_request
);
662 int blk_mq_request_started(struct request
*rq
)
664 return blk_mq_rq_state(rq
) != MQ_RQ_IDLE
;
666 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
668 void blk_mq_start_request(struct request
*rq
)
670 struct request_queue
*q
= rq
->q
;
672 blk_mq_sched_started_request(rq
);
674 trace_block_rq_issue(q
, rq
);
676 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
677 rq
->io_start_time_ns
= ktime_get_ns();
678 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
679 rq
->throtl_size
= blk_rq_sectors(rq
);
681 rq
->rq_flags
|= RQF_STATS
;
685 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
688 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
690 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
692 * Make sure space for the drain appears. We know we can do
693 * this because max_hw_segments has been adjusted to be one
694 * fewer than the device can handle.
696 rq
->nr_phys_segments
++;
699 EXPORT_SYMBOL(blk_mq_start_request
);
701 static void __blk_mq_requeue_request(struct request
*rq
)
703 struct request_queue
*q
= rq
->q
;
705 blk_mq_put_driver_tag(rq
);
707 trace_block_rq_requeue(q
, rq
);
708 rq_qos_requeue(q
, rq
);
710 if (blk_mq_request_started(rq
)) {
711 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
712 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
713 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
714 rq
->nr_phys_segments
--;
718 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
720 __blk_mq_requeue_request(rq
);
722 /* this request will be re-inserted to io scheduler queue */
723 blk_mq_sched_requeue_request(rq
);
725 BUG_ON(!list_empty(&rq
->queuelist
));
726 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
728 EXPORT_SYMBOL(blk_mq_requeue_request
);
730 static void blk_mq_requeue_work(struct work_struct
*work
)
732 struct request_queue
*q
=
733 container_of(work
, struct request_queue
, requeue_work
.work
);
735 struct request
*rq
, *next
;
737 spin_lock_irq(&q
->requeue_lock
);
738 list_splice_init(&q
->requeue_list
, &rq_list
);
739 spin_unlock_irq(&q
->requeue_lock
);
741 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
742 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
745 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
746 list_del_init(&rq
->queuelist
);
747 blk_mq_sched_insert_request(rq
, true, false, false);
750 while (!list_empty(&rq_list
)) {
751 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
752 list_del_init(&rq
->queuelist
);
753 blk_mq_sched_insert_request(rq
, false, false, false);
756 blk_mq_run_hw_queues(q
, false);
759 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
760 bool kick_requeue_list
)
762 struct request_queue
*q
= rq
->q
;
766 * We abuse this flag that is otherwise used by the I/O scheduler to
767 * request head insertion from the workqueue.
769 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
771 spin_lock_irqsave(&q
->requeue_lock
, flags
);
773 rq
->rq_flags
|= RQF_SOFTBARRIER
;
774 list_add(&rq
->queuelist
, &q
->requeue_list
);
776 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
778 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
780 if (kick_requeue_list
)
781 blk_mq_kick_requeue_list(q
);
783 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
785 void blk_mq_kick_requeue_list(struct request_queue
*q
)
787 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
789 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
791 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
794 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
795 msecs_to_jiffies(msecs
));
797 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
799 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
801 if (tag
< tags
->nr_tags
) {
802 prefetch(tags
->rqs
[tag
]);
803 return tags
->rqs
[tag
];
808 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
810 static bool blk_mq_check_busy(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
811 void *priv
, bool reserved
)
814 * If we find a request, we know the queue is busy. Return false
815 * to stop the iteration.
817 if (rq
->q
== hctx
->queue
) {
827 bool blk_mq_queue_busy(struct request_queue
*q
)
831 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_busy
, &busy
);
834 EXPORT_SYMBOL_GPL(blk_mq_queue_busy
);
836 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
838 req
->rq_flags
|= RQF_TIMED_OUT
;
839 if (req
->q
->mq_ops
->timeout
) {
840 enum blk_eh_timer_return ret
;
842 ret
= req
->q
->mq_ops
->timeout(req
, reserved
);
843 if (ret
== BLK_EH_DONE
)
845 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
851 static bool blk_mq_req_expired(struct request
*rq
, unsigned long *next
)
853 unsigned long deadline
;
855 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
857 if (rq
->rq_flags
& RQF_TIMED_OUT
)
860 deadline
= READ_ONCE(rq
->deadline
);
861 if (time_after_eq(jiffies
, deadline
))
866 else if (time_after(*next
, deadline
))
871 static bool blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
872 struct request
*rq
, void *priv
, bool reserved
)
874 unsigned long *next
= priv
;
877 * Just do a quick check if it is expired before locking the request in
878 * so we're not unnecessarilly synchronizing across CPUs.
880 if (!blk_mq_req_expired(rq
, next
))
884 * We have reason to believe the request may be expired. Take a
885 * reference on the request to lock this request lifetime into its
886 * currently allocated context to prevent it from being reallocated in
887 * the event the completion by-passes this timeout handler.
889 * If the reference was already released, then the driver beat the
890 * timeout handler to posting a natural completion.
892 if (!refcount_inc_not_zero(&rq
->ref
))
896 * The request is now locked and cannot be reallocated underneath the
897 * timeout handler's processing. Re-verify this exact request is truly
898 * expired; if it is not expired, then the request was completed and
899 * reallocated as a new request.
901 if (blk_mq_req_expired(rq
, next
))
902 blk_mq_rq_timed_out(rq
, reserved
);
903 if (refcount_dec_and_test(&rq
->ref
))
904 __blk_mq_free_request(rq
);
909 static void blk_mq_timeout_work(struct work_struct
*work
)
911 struct request_queue
*q
=
912 container_of(work
, struct request_queue
, timeout_work
);
913 unsigned long next
= 0;
914 struct blk_mq_hw_ctx
*hctx
;
917 /* A deadlock might occur if a request is stuck requiring a
918 * timeout at the same time a queue freeze is waiting
919 * completion, since the timeout code would not be able to
920 * acquire the queue reference here.
922 * That's why we don't use blk_queue_enter here; instead, we use
923 * percpu_ref_tryget directly, because we need to be able to
924 * obtain a reference even in the short window between the queue
925 * starting to freeze, by dropping the first reference in
926 * blk_freeze_queue_start, and the moment the last request is
927 * consumed, marked by the instant q_usage_counter reaches
930 if (!percpu_ref_tryget(&q
->q_usage_counter
))
933 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &next
);
936 mod_timer(&q
->timeout
, next
);
939 * Request timeouts are handled as a forward rolling timer. If
940 * we end up here it means that no requests are pending and
941 * also that no request has been pending for a while. Mark
944 queue_for_each_hw_ctx(q
, hctx
, i
) {
945 /* the hctx may be unmapped, so check it here */
946 if (blk_mq_hw_queue_mapped(hctx
))
947 blk_mq_tag_idle(hctx
);
953 struct flush_busy_ctx_data
{
954 struct blk_mq_hw_ctx
*hctx
;
955 struct list_head
*list
;
958 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
960 struct flush_busy_ctx_data
*flush_data
= data
;
961 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
962 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
964 spin_lock(&ctx
->lock
);
965 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
966 sbitmap_clear_bit(sb
, bitnr
);
967 spin_unlock(&ctx
->lock
);
972 * Process software queues that have been marked busy, splicing them
973 * to the for-dispatch
975 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
977 struct flush_busy_ctx_data data
= {
982 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
984 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
986 struct dispatch_rq_data
{
987 struct blk_mq_hw_ctx
*hctx
;
991 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
994 struct dispatch_rq_data
*dispatch_data
= data
;
995 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
996 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
998 spin_lock(&ctx
->lock
);
999 if (!list_empty(&ctx
->rq_list
)) {
1000 dispatch_data
->rq
= list_entry_rq(ctx
->rq_list
.next
);
1001 list_del_init(&dispatch_data
->rq
->queuelist
);
1002 if (list_empty(&ctx
->rq_list
))
1003 sbitmap_clear_bit(sb
, bitnr
);
1005 spin_unlock(&ctx
->lock
);
1007 return !dispatch_data
->rq
;
1010 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
1011 struct blk_mq_ctx
*start
)
1013 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1014 struct dispatch_rq_data data
= {
1019 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1020 dispatch_rq_from_ctx
, &data
);
1025 static inline unsigned int queued_to_index(unsigned int queued
)
1030 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
1033 bool blk_mq_get_driver_tag(struct request
*rq
)
1035 struct blk_mq_alloc_data data
= {
1037 .hctx
= rq
->mq_hctx
,
1038 .flags
= BLK_MQ_REQ_NOWAIT
,
1039 .cmd_flags
= rq
->cmd_flags
,
1046 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
1047 data
.flags
|= BLK_MQ_REQ_RESERVED
;
1049 shared
= blk_mq_tag_busy(data
.hctx
);
1050 rq
->tag
= blk_mq_get_tag(&data
);
1053 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1054 atomic_inc(&data
.hctx
->nr_active
);
1056 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
1060 return rq
->tag
!= -1;
1063 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1064 int flags
, void *key
)
1066 struct blk_mq_hw_ctx
*hctx
;
1068 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1070 spin_lock(&hctx
->dispatch_wait_lock
);
1071 list_del_init(&wait
->entry
);
1072 spin_unlock(&hctx
->dispatch_wait_lock
);
1074 blk_mq_run_hw_queue(hctx
, true);
1079 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1080 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1081 * restart. For both cases, take care to check the condition again after
1082 * marking us as waiting.
1084 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1087 struct wait_queue_head
*wq
;
1088 wait_queue_entry_t
*wait
;
1091 if (!(hctx
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1092 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
1093 set_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
1096 * It's possible that a tag was freed in the window between the
1097 * allocation failure and adding the hardware queue to the wait
1100 * Don't clear RESTART here, someone else could have set it.
1101 * At most this will cost an extra queue run.
1103 return blk_mq_get_driver_tag(rq
);
1106 wait
= &hctx
->dispatch_wait
;
1107 if (!list_empty_careful(&wait
->entry
))
1110 wq
= &bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
)->wait
;
1112 spin_lock_irq(&wq
->lock
);
1113 spin_lock(&hctx
->dispatch_wait_lock
);
1114 if (!list_empty(&wait
->entry
)) {
1115 spin_unlock(&hctx
->dispatch_wait_lock
);
1116 spin_unlock_irq(&wq
->lock
);
1120 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1121 __add_wait_queue(wq
, wait
);
1124 * It's possible that a tag was freed in the window between the
1125 * allocation failure and adding the hardware queue to the wait
1128 ret
= blk_mq_get_driver_tag(rq
);
1130 spin_unlock(&hctx
->dispatch_wait_lock
);
1131 spin_unlock_irq(&wq
->lock
);
1136 * We got a tag, remove ourselves from the wait queue to ensure
1137 * someone else gets the wakeup.
1139 list_del_init(&wait
->entry
);
1140 spin_unlock(&hctx
->dispatch_wait_lock
);
1141 spin_unlock_irq(&wq
->lock
);
1146 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1147 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1149 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1150 * - EWMA is one simple way to compute running average value
1151 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1152 * - take 4 as factor for avoiding to get too small(0) result, and this
1153 * factor doesn't matter because EWMA decreases exponentially
1155 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1159 if (hctx
->queue
->elevator
)
1162 ewma
= hctx
->dispatch_busy
;
1167 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1169 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1170 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1172 hctx
->dispatch_busy
= ewma
;
1175 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1178 * Returns true if we did some work AND can potentially do more.
1180 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
,
1183 struct blk_mq_hw_ctx
*hctx
;
1184 struct request
*rq
, *nxt
;
1185 bool no_tag
= false;
1187 blk_status_t ret
= BLK_STS_OK
;
1189 if (list_empty(list
))
1192 WARN_ON(!list_is_singular(list
) && got_budget
);
1195 * Now process all the entries, sending them to the driver.
1197 errors
= queued
= 0;
1199 struct blk_mq_queue_data bd
;
1201 rq
= list_first_entry(list
, struct request
, queuelist
);
1204 if (!got_budget
&& !blk_mq_get_dispatch_budget(hctx
))
1207 if (!blk_mq_get_driver_tag(rq
)) {
1209 * The initial allocation attempt failed, so we need to
1210 * rerun the hardware queue when a tag is freed. The
1211 * waitqueue takes care of that. If the queue is run
1212 * before we add this entry back on the dispatch list,
1213 * we'll re-run it below.
1215 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1216 blk_mq_put_dispatch_budget(hctx
);
1218 * For non-shared tags, the RESTART check
1221 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
)
1227 list_del_init(&rq
->queuelist
);
1232 * Flag last if we have no more requests, or if we have more
1233 * but can't assign a driver tag to it.
1235 if (list_empty(list
))
1238 nxt
= list_first_entry(list
, struct request
, queuelist
);
1239 bd
.last
= !blk_mq_get_driver_tag(nxt
);
1242 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1243 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
) {
1245 * If an I/O scheduler has been configured and we got a
1246 * driver tag for the next request already, free it
1249 if (!list_empty(list
)) {
1250 nxt
= list_first_entry(list
, struct request
, queuelist
);
1251 blk_mq_put_driver_tag(nxt
);
1253 list_add(&rq
->queuelist
, list
);
1254 __blk_mq_requeue_request(rq
);
1258 if (unlikely(ret
!= BLK_STS_OK
)) {
1260 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1265 } while (!list_empty(list
));
1267 hctx
->dispatched
[queued_to_index(queued
)]++;
1270 * Any items that need requeuing? Stuff them into hctx->dispatch,
1271 * that is where we will continue on next queue run.
1273 if (!list_empty(list
)) {
1277 * If we didn't flush the entire list, we could have told
1278 * the driver there was more coming, but that turned out to
1281 if (q
->mq_ops
->commit_rqs
)
1282 q
->mq_ops
->commit_rqs(hctx
);
1284 spin_lock(&hctx
->lock
);
1285 list_splice_init(list
, &hctx
->dispatch
);
1286 spin_unlock(&hctx
->lock
);
1289 * If SCHED_RESTART was set by the caller of this function and
1290 * it is no longer set that means that it was cleared by another
1291 * thread and hence that a queue rerun is needed.
1293 * If 'no_tag' is set, that means that we failed getting
1294 * a driver tag with an I/O scheduler attached. If our dispatch
1295 * waitqueue is no longer active, ensure that we run the queue
1296 * AFTER adding our entries back to the list.
1298 * If no I/O scheduler has been configured it is possible that
1299 * the hardware queue got stopped and restarted before requests
1300 * were pushed back onto the dispatch list. Rerun the queue to
1301 * avoid starvation. Notes:
1302 * - blk_mq_run_hw_queue() checks whether or not a queue has
1303 * been stopped before rerunning a queue.
1304 * - Some but not all block drivers stop a queue before
1305 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1308 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1309 * bit is set, run queue after a delay to avoid IO stalls
1310 * that could otherwise occur if the queue is idle.
1312 needs_restart
= blk_mq_sched_needs_restart(hctx
);
1313 if (!needs_restart
||
1314 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1315 blk_mq_run_hw_queue(hctx
, true);
1316 else if (needs_restart
&& (ret
== BLK_STS_RESOURCE
))
1317 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
1319 blk_mq_update_dispatch_busy(hctx
, true);
1322 blk_mq_update_dispatch_busy(hctx
, false);
1325 * If the host/device is unable to accept more work, inform the
1328 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1331 return (queued
+ errors
) != 0;
1334 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1339 * We should be running this queue from one of the CPUs that
1342 * There are at least two related races now between setting
1343 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1344 * __blk_mq_run_hw_queue():
1346 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1347 * but later it becomes online, then this warning is harmless
1350 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1351 * but later it becomes offline, then the warning can't be
1352 * triggered, and we depend on blk-mq timeout handler to
1353 * handle dispatched requests to this hctx
1355 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1356 cpu_online(hctx
->next_cpu
)) {
1357 printk(KERN_WARNING
"run queue from wrong CPU %d, hctx %s\n",
1358 raw_smp_processor_id(),
1359 cpumask_empty(hctx
->cpumask
) ? "inactive": "active");
1364 * We can't run the queue inline with ints disabled. Ensure that
1365 * we catch bad users of this early.
1367 WARN_ON_ONCE(in_interrupt());
1369 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1371 hctx_lock(hctx
, &srcu_idx
);
1372 blk_mq_sched_dispatch_requests(hctx
);
1373 hctx_unlock(hctx
, srcu_idx
);
1376 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1378 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1380 if (cpu
>= nr_cpu_ids
)
1381 cpu
= cpumask_first(hctx
->cpumask
);
1386 * It'd be great if the workqueue API had a way to pass
1387 * in a mask and had some smarts for more clever placement.
1388 * For now we just round-robin here, switching for every
1389 * BLK_MQ_CPU_WORK_BATCH queued items.
1391 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1394 int next_cpu
= hctx
->next_cpu
;
1396 if (hctx
->queue
->nr_hw_queues
== 1)
1397 return WORK_CPU_UNBOUND
;
1399 if (--hctx
->next_cpu_batch
<= 0) {
1401 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1403 if (next_cpu
>= nr_cpu_ids
)
1404 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1405 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1409 * Do unbound schedule if we can't find a online CPU for this hctx,
1410 * and it should only happen in the path of handling CPU DEAD.
1412 if (!cpu_online(next_cpu
)) {
1419 * Make sure to re-select CPU next time once after CPUs
1420 * in hctx->cpumask become online again.
1422 hctx
->next_cpu
= next_cpu
;
1423 hctx
->next_cpu_batch
= 1;
1424 return WORK_CPU_UNBOUND
;
1427 hctx
->next_cpu
= next_cpu
;
1431 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1432 unsigned long msecs
)
1434 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1437 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1438 int cpu
= get_cpu();
1439 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1440 __blk_mq_run_hw_queue(hctx
);
1448 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1449 msecs_to_jiffies(msecs
));
1452 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1454 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1456 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1458 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1464 * When queue is quiesced, we may be switching io scheduler, or
1465 * updating nr_hw_queues, or other things, and we can't run queue
1466 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1468 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1471 hctx_lock(hctx
, &srcu_idx
);
1472 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1473 blk_mq_hctx_has_pending(hctx
);
1474 hctx_unlock(hctx
, srcu_idx
);
1477 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1483 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1485 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1487 struct blk_mq_hw_ctx
*hctx
;
1490 queue_for_each_hw_ctx(q
, hctx
, i
) {
1491 if (blk_mq_hctx_stopped(hctx
))
1494 blk_mq_run_hw_queue(hctx
, async
);
1497 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1500 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1501 * @q: request queue.
1503 * The caller is responsible for serializing this function against
1504 * blk_mq_{start,stop}_hw_queue().
1506 bool blk_mq_queue_stopped(struct request_queue
*q
)
1508 struct blk_mq_hw_ctx
*hctx
;
1511 queue_for_each_hw_ctx(q
, hctx
, i
)
1512 if (blk_mq_hctx_stopped(hctx
))
1517 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1520 * This function is often used for pausing .queue_rq() by driver when
1521 * there isn't enough resource or some conditions aren't satisfied, and
1522 * BLK_STS_RESOURCE is usually returned.
1524 * We do not guarantee that dispatch can be drained or blocked
1525 * after blk_mq_stop_hw_queue() returns. Please use
1526 * blk_mq_quiesce_queue() for that requirement.
1528 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1530 cancel_delayed_work(&hctx
->run_work
);
1532 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1534 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1537 * This function is often used for pausing .queue_rq() by driver when
1538 * there isn't enough resource or some conditions aren't satisfied, and
1539 * BLK_STS_RESOURCE is usually returned.
1541 * We do not guarantee that dispatch can be drained or blocked
1542 * after blk_mq_stop_hw_queues() returns. Please use
1543 * blk_mq_quiesce_queue() for that requirement.
1545 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1547 struct blk_mq_hw_ctx
*hctx
;
1550 queue_for_each_hw_ctx(q
, hctx
, i
)
1551 blk_mq_stop_hw_queue(hctx
);
1553 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1555 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1557 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1559 blk_mq_run_hw_queue(hctx
, false);
1561 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1563 void blk_mq_start_hw_queues(struct request_queue
*q
)
1565 struct blk_mq_hw_ctx
*hctx
;
1568 queue_for_each_hw_ctx(q
, hctx
, i
)
1569 blk_mq_start_hw_queue(hctx
);
1571 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1573 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1575 if (!blk_mq_hctx_stopped(hctx
))
1578 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1579 blk_mq_run_hw_queue(hctx
, async
);
1581 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1583 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1585 struct blk_mq_hw_ctx
*hctx
;
1588 queue_for_each_hw_ctx(q
, hctx
, i
)
1589 blk_mq_start_stopped_hw_queue(hctx
, async
);
1591 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1593 static void blk_mq_run_work_fn(struct work_struct
*work
)
1595 struct blk_mq_hw_ctx
*hctx
;
1597 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1600 * If we are stopped, don't run the queue.
1602 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1605 __blk_mq_run_hw_queue(hctx
);
1608 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1612 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1614 lockdep_assert_held(&ctx
->lock
);
1616 trace_block_rq_insert(hctx
->queue
, rq
);
1619 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1621 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1624 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1627 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1629 lockdep_assert_held(&ctx
->lock
);
1631 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1632 blk_mq_hctx_mark_pending(hctx
, ctx
);
1636 * Should only be used carefully, when the caller knows we want to
1637 * bypass a potential IO scheduler on the target device.
1639 void blk_mq_request_bypass_insert(struct request
*rq
, bool run_queue
)
1641 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1643 spin_lock(&hctx
->lock
);
1644 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1645 spin_unlock(&hctx
->lock
);
1648 blk_mq_run_hw_queue(hctx
, false);
1651 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1652 struct list_head
*list
)
1658 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1661 list_for_each_entry(rq
, list
, queuelist
) {
1662 BUG_ON(rq
->mq_ctx
!= ctx
);
1663 trace_block_rq_insert(hctx
->queue
, rq
);
1666 spin_lock(&ctx
->lock
);
1667 list_splice_tail_init(list
, &ctx
->rq_list
);
1668 blk_mq_hctx_mark_pending(hctx
, ctx
);
1669 spin_unlock(&ctx
->lock
);
1672 static int plug_rq_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1674 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1675 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1677 if (rqa
->mq_ctx
< rqb
->mq_ctx
)
1679 else if (rqa
->mq_ctx
> rqb
->mq_ctx
)
1681 else if (rqa
->mq_hctx
< rqb
->mq_hctx
)
1683 else if (rqa
->mq_hctx
> rqb
->mq_hctx
)
1686 return blk_rq_pos(rqa
) > blk_rq_pos(rqb
);
1689 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1691 struct blk_mq_hw_ctx
*this_hctx
;
1692 struct blk_mq_ctx
*this_ctx
;
1693 struct request_queue
*this_q
;
1699 list_splice_init(&plug
->mq_list
, &list
);
1702 if (plug
->rq_count
> 2 && plug
->multiple_queues
)
1703 list_sort(NULL
, &list
, plug_rq_cmp
);
1710 while (!list_empty(&list
)) {
1711 rq
= list_entry_rq(list
.next
);
1712 list_del_init(&rq
->queuelist
);
1714 if (rq
->mq_hctx
!= this_hctx
|| rq
->mq_ctx
!= this_ctx
) {
1716 trace_block_unplug(this_q
, depth
, !from_schedule
);
1717 blk_mq_sched_insert_requests(this_hctx
, this_ctx
,
1723 this_ctx
= rq
->mq_ctx
;
1724 this_hctx
= rq
->mq_hctx
;
1729 list_add_tail(&rq
->queuelist
, &rq_list
);
1733 * If 'this_hctx' is set, we know we have entries to complete
1734 * on 'rq_list'. Do those.
1737 trace_block_unplug(this_q
, depth
, !from_schedule
);
1738 blk_mq_sched_insert_requests(this_hctx
, this_ctx
, &rq_list
,
1743 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1745 blk_init_request_from_bio(rq
, bio
);
1747 blk_account_io_start(rq
, true);
1750 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1753 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1755 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1758 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1760 blk_qc_t
*cookie
, bool last
)
1762 struct request_queue
*q
= rq
->q
;
1763 struct blk_mq_queue_data bd
= {
1767 blk_qc_t new_cookie
;
1770 new_cookie
= request_to_qc_t(hctx
, rq
);
1773 * For OK queue, we are done. For error, caller may kill it.
1774 * Any other error (busy), just add it to our list as we
1775 * previously would have done.
1777 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1780 blk_mq_update_dispatch_busy(hctx
, false);
1781 *cookie
= new_cookie
;
1783 case BLK_STS_RESOURCE
:
1784 case BLK_STS_DEV_RESOURCE
:
1785 blk_mq_update_dispatch_busy(hctx
, true);
1786 __blk_mq_requeue_request(rq
);
1789 blk_mq_update_dispatch_busy(hctx
, false);
1790 *cookie
= BLK_QC_T_NONE
;
1797 static blk_status_t
__blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1800 bool bypass_insert
, bool last
)
1802 struct request_queue
*q
= rq
->q
;
1803 bool run_queue
= true;
1806 * RCU or SRCU read lock is needed before checking quiesced flag.
1808 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1809 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1810 * and avoid driver to try to dispatch again.
1812 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1814 bypass_insert
= false;
1818 if (q
->elevator
&& !bypass_insert
)
1821 if (!blk_mq_get_dispatch_budget(hctx
))
1824 if (!blk_mq_get_driver_tag(rq
)) {
1825 blk_mq_put_dispatch_budget(hctx
);
1829 return __blk_mq_issue_directly(hctx
, rq
, cookie
, last
);
1832 return BLK_STS_RESOURCE
;
1834 blk_mq_request_bypass_insert(rq
, run_queue
);
1838 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1839 struct request
*rq
, blk_qc_t
*cookie
)
1844 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1846 hctx_lock(hctx
, &srcu_idx
);
1848 ret
= __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false, true);
1849 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1850 blk_mq_request_bypass_insert(rq
, true);
1851 else if (ret
!= BLK_STS_OK
)
1852 blk_mq_end_request(rq
, ret
);
1854 hctx_unlock(hctx
, srcu_idx
);
1857 blk_status_t
blk_mq_request_issue_directly(struct request
*rq
, bool last
)
1861 blk_qc_t unused_cookie
;
1862 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1864 hctx_lock(hctx
, &srcu_idx
);
1865 ret
= __blk_mq_try_issue_directly(hctx
, rq
, &unused_cookie
, true, last
);
1866 hctx_unlock(hctx
, srcu_idx
);
1871 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
1872 struct list_head
*list
)
1874 while (!list_empty(list
)) {
1876 struct request
*rq
= list_first_entry(list
, struct request
,
1879 list_del_init(&rq
->queuelist
);
1880 ret
= blk_mq_request_issue_directly(rq
, list_empty(list
));
1881 if (ret
!= BLK_STS_OK
) {
1882 if (ret
== BLK_STS_RESOURCE
||
1883 ret
== BLK_STS_DEV_RESOURCE
) {
1884 blk_mq_request_bypass_insert(rq
,
1888 blk_mq_end_request(rq
, ret
);
1893 * If we didn't flush the entire list, we could have told
1894 * the driver there was more coming, but that turned out to
1897 if (!list_empty(list
) && hctx
->queue
->mq_ops
->commit_rqs
)
1898 hctx
->queue
->mq_ops
->commit_rqs(hctx
);
1901 static void blk_add_rq_to_plug(struct blk_plug
*plug
, struct request
*rq
)
1903 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1905 if (!plug
->multiple_queues
&& !list_is_singular(&plug
->mq_list
)) {
1906 struct request
*tmp
;
1908 tmp
= list_first_entry(&plug
->mq_list
, struct request
,
1910 if (tmp
->q
!= rq
->q
)
1911 plug
->multiple_queues
= true;
1915 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1917 const int is_sync
= op_is_sync(bio
->bi_opf
);
1918 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1919 struct blk_mq_alloc_data data
= { .flags
= 0, .cmd_flags
= bio
->bi_opf
};
1921 struct blk_plug
*plug
;
1922 struct request
*same_queue_rq
= NULL
;
1925 blk_queue_bounce(q
, &bio
);
1927 blk_queue_split(q
, &bio
);
1929 if (!bio_integrity_prep(bio
))
1930 return BLK_QC_T_NONE
;
1932 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1933 blk_attempt_plug_merge(q
, bio
, &same_queue_rq
))
1934 return BLK_QC_T_NONE
;
1936 if (blk_mq_sched_bio_merge(q
, bio
))
1937 return BLK_QC_T_NONE
;
1939 rq_qos_throttle(q
, bio
);
1941 rq
= blk_mq_get_request(q
, bio
, &data
);
1942 if (unlikely(!rq
)) {
1943 rq_qos_cleanup(q
, bio
);
1944 if (bio
->bi_opf
& REQ_NOWAIT
)
1945 bio_wouldblock_error(bio
);
1946 return BLK_QC_T_NONE
;
1949 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1951 rq_qos_track(q
, rq
, bio
);
1953 cookie
= request_to_qc_t(data
.hctx
, rq
);
1955 plug
= current
->plug
;
1956 if (unlikely(is_flush_fua
)) {
1957 blk_mq_put_ctx(data
.ctx
);
1958 blk_mq_bio_to_request(rq
, bio
);
1960 /* bypass scheduler for flush rq */
1961 blk_insert_flush(rq
);
1962 blk_mq_run_hw_queue(data
.hctx
, true);
1963 } else if (plug
&& (q
->nr_hw_queues
== 1 || q
->mq_ops
->commit_rqs
)) {
1965 * Use plugging if we have a ->commit_rqs() hook as well, as
1966 * we know the driver uses bd->last in a smart fashion.
1968 unsigned int request_count
= plug
->rq_count
;
1969 struct request
*last
= NULL
;
1971 blk_mq_put_ctx(data
.ctx
);
1972 blk_mq_bio_to_request(rq
, bio
);
1975 trace_block_plug(q
);
1977 last
= list_entry_rq(plug
->mq_list
.prev
);
1979 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1980 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1981 blk_flush_plug_list(plug
, false);
1982 trace_block_plug(q
);
1985 blk_add_rq_to_plug(plug
, rq
);
1986 } else if (plug
&& !blk_queue_nomerges(q
)) {
1987 blk_mq_bio_to_request(rq
, bio
);
1990 * We do limited plugging. If the bio can be merged, do that.
1991 * Otherwise the existing request in the plug list will be
1992 * issued. So the plug list will have one request at most
1993 * The plug list might get flushed before this. If that happens,
1994 * the plug list is empty, and same_queue_rq is invalid.
1996 if (list_empty(&plug
->mq_list
))
1997 same_queue_rq
= NULL
;
1998 if (same_queue_rq
) {
1999 list_del_init(&same_queue_rq
->queuelist
);
2002 blk_add_rq_to_plug(plug
, rq
);
2004 blk_mq_put_ctx(data
.ctx
);
2006 if (same_queue_rq
) {
2007 data
.hctx
= same_queue_rq
->mq_hctx
;
2008 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
2011 } else if ((q
->nr_hw_queues
> 1 && is_sync
) || (!q
->elevator
&&
2012 !data
.hctx
->dispatch_busy
)) {
2013 blk_mq_put_ctx(data
.ctx
);
2014 blk_mq_bio_to_request(rq
, bio
);
2015 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
2017 blk_mq_put_ctx(data
.ctx
);
2018 blk_mq_bio_to_request(rq
, bio
);
2019 blk_mq_sched_insert_request(rq
, false, true, true);
2025 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2026 unsigned int hctx_idx
)
2030 if (tags
->rqs
&& set
->ops
->exit_request
) {
2033 for (i
= 0; i
< tags
->nr_tags
; i
++) {
2034 struct request
*rq
= tags
->static_rqs
[i
];
2038 set
->ops
->exit_request(set
, rq
, hctx_idx
);
2039 tags
->static_rqs
[i
] = NULL
;
2043 while (!list_empty(&tags
->page_list
)) {
2044 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
2045 list_del_init(&page
->lru
);
2047 * Remove kmemleak object previously allocated in
2048 * blk_mq_init_rq_map().
2050 kmemleak_free(page_address(page
));
2051 __free_pages(page
, page
->private);
2055 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
2059 kfree(tags
->static_rqs
);
2060 tags
->static_rqs
= NULL
;
2062 blk_mq_free_tags(tags
);
2065 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
2066 unsigned int hctx_idx
,
2067 unsigned int nr_tags
,
2068 unsigned int reserved_tags
)
2070 struct blk_mq_tags
*tags
;
2073 node
= blk_mq_hw_queue_to_node(&set
->map
[0], hctx_idx
);
2074 if (node
== NUMA_NO_NODE
)
2075 node
= set
->numa_node
;
2077 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
2078 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
2082 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2083 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2086 blk_mq_free_tags(tags
);
2090 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2091 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2093 if (!tags
->static_rqs
) {
2095 blk_mq_free_tags(tags
);
2102 static size_t order_to_size(unsigned int order
)
2104 return (size_t)PAGE_SIZE
<< order
;
2107 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
2108 unsigned int hctx_idx
, int node
)
2112 if (set
->ops
->init_request
) {
2113 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
2118 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
2122 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2123 unsigned int hctx_idx
, unsigned int depth
)
2125 unsigned int i
, j
, entries_per_page
, max_order
= 4;
2126 size_t rq_size
, left
;
2129 node
= blk_mq_hw_queue_to_node(&set
->map
[0], hctx_idx
);
2130 if (node
== NUMA_NO_NODE
)
2131 node
= set
->numa_node
;
2133 INIT_LIST_HEAD(&tags
->page_list
);
2136 * rq_size is the size of the request plus driver payload, rounded
2137 * to the cacheline size
2139 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
2141 left
= rq_size
* depth
;
2143 for (i
= 0; i
< depth
; ) {
2144 int this_order
= max_order
;
2149 while (this_order
&& left
< order_to_size(this_order
- 1))
2153 page
= alloc_pages_node(node
,
2154 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
2160 if (order_to_size(this_order
) < rq_size
)
2167 page
->private = this_order
;
2168 list_add_tail(&page
->lru
, &tags
->page_list
);
2170 p
= page_address(page
);
2172 * Allow kmemleak to scan these pages as they contain pointers
2173 * to additional allocations like via ops->init_request().
2175 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2176 entries_per_page
= order_to_size(this_order
) / rq_size
;
2177 to_do
= min(entries_per_page
, depth
- i
);
2178 left
-= to_do
* rq_size
;
2179 for (j
= 0; j
< to_do
; j
++) {
2180 struct request
*rq
= p
;
2182 tags
->static_rqs
[i
] = rq
;
2183 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
2184 tags
->static_rqs
[i
] = NULL
;
2195 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2200 * 'cpu' is going away. splice any existing rq_list entries from this
2201 * software queue to the hw queue dispatch list, and ensure that it
2204 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2206 struct blk_mq_hw_ctx
*hctx
;
2207 struct blk_mq_ctx
*ctx
;
2210 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2211 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2213 spin_lock(&ctx
->lock
);
2214 if (!list_empty(&ctx
->rq_list
)) {
2215 list_splice_init(&ctx
->rq_list
, &tmp
);
2216 blk_mq_hctx_clear_pending(hctx
, ctx
);
2218 spin_unlock(&ctx
->lock
);
2220 if (list_empty(&tmp
))
2223 spin_lock(&hctx
->lock
);
2224 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2225 spin_unlock(&hctx
->lock
);
2227 blk_mq_run_hw_queue(hctx
, true);
2231 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2233 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2237 /* hctx->ctxs will be freed in queue's release handler */
2238 static void blk_mq_exit_hctx(struct request_queue
*q
,
2239 struct blk_mq_tag_set
*set
,
2240 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2242 if (blk_mq_hw_queue_mapped(hctx
))
2243 blk_mq_tag_idle(hctx
);
2245 if (set
->ops
->exit_request
)
2246 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2248 if (set
->ops
->exit_hctx
)
2249 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2251 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2252 cleanup_srcu_struct(hctx
->srcu
);
2254 blk_mq_remove_cpuhp(hctx
);
2255 blk_free_flush_queue(hctx
->fq
);
2256 sbitmap_free(&hctx
->ctx_map
);
2259 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2260 struct blk_mq_tag_set
*set
, int nr_queue
)
2262 struct blk_mq_hw_ctx
*hctx
;
2265 queue_for_each_hw_ctx(q
, hctx
, i
) {
2268 blk_mq_debugfs_unregister_hctx(hctx
);
2269 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2273 static int blk_mq_init_hctx(struct request_queue
*q
,
2274 struct blk_mq_tag_set
*set
,
2275 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2279 node
= hctx
->numa_node
;
2280 if (node
== NUMA_NO_NODE
)
2281 node
= hctx
->numa_node
= set
->numa_node
;
2283 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2284 spin_lock_init(&hctx
->lock
);
2285 INIT_LIST_HEAD(&hctx
->dispatch
);
2287 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
2289 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2291 hctx
->tags
= set
->tags
[hctx_idx
];
2294 * Allocate space for all possible cpus to avoid allocation at
2297 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2298 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
, node
);
2300 goto unregister_cpu_notifier
;
2302 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
2303 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
, node
))
2308 spin_lock_init(&hctx
->dispatch_wait_lock
);
2309 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2310 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2312 if (set
->ops
->init_hctx
&&
2313 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2316 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
,
2317 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
2321 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
, node
))
2324 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2325 init_srcu_struct(hctx
->srcu
);
2332 if (set
->ops
->exit_hctx
)
2333 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2335 sbitmap_free(&hctx
->ctx_map
);
2338 unregister_cpu_notifier
:
2339 blk_mq_remove_cpuhp(hctx
);
2343 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2344 unsigned int nr_hw_queues
)
2346 struct blk_mq_tag_set
*set
= q
->tag_set
;
2349 for_each_possible_cpu(i
) {
2350 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2351 struct blk_mq_hw_ctx
*hctx
;
2354 spin_lock_init(&__ctx
->lock
);
2355 INIT_LIST_HEAD(&__ctx
->rq_list
);
2359 * Set local node, IFF we have more than one hw queue. If
2360 * not, we remain on the home node of the device
2362 for (j
= 0; j
< set
->nr_maps
; j
++) {
2363 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2364 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2365 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2370 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2374 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2375 set
->queue_depth
, set
->reserved_tags
);
2376 if (!set
->tags
[hctx_idx
])
2379 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2384 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2385 set
->tags
[hctx_idx
] = NULL
;
2389 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2390 unsigned int hctx_idx
)
2392 if (set
->tags
&& set
->tags
[hctx_idx
]) {
2393 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2394 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2395 set
->tags
[hctx_idx
] = NULL
;
2399 static void blk_mq_map_swqueue(struct request_queue
*q
)
2401 unsigned int i
, j
, hctx_idx
;
2402 struct blk_mq_hw_ctx
*hctx
;
2403 struct blk_mq_ctx
*ctx
;
2404 struct blk_mq_tag_set
*set
= q
->tag_set
;
2407 * Avoid others reading imcomplete hctx->cpumask through sysfs
2409 mutex_lock(&q
->sysfs_lock
);
2411 queue_for_each_hw_ctx(q
, hctx
, i
) {
2412 cpumask_clear(hctx
->cpumask
);
2414 hctx
->dispatch_from
= NULL
;
2418 * Map software to hardware queues.
2420 * If the cpu isn't present, the cpu is mapped to first hctx.
2422 for_each_possible_cpu(i
) {
2423 hctx_idx
= set
->map
[0].mq_map
[i
];
2424 /* unmapped hw queue can be remapped after CPU topo changed */
2425 if (!set
->tags
[hctx_idx
] &&
2426 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2428 * If tags initialization fail for some hctx,
2429 * that hctx won't be brought online. In this
2430 * case, remap the current ctx to hctx[0] which
2431 * is guaranteed to always have tags allocated
2433 set
->map
[0].mq_map
[i
] = 0;
2436 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2437 for (j
= 0; j
< set
->nr_maps
; j
++) {
2438 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2441 * If the CPU is already set in the mask, then we've
2442 * mapped this one already. This can happen if
2443 * devices share queues across queue maps.
2445 if (cpumask_test_cpu(i
, hctx
->cpumask
))
2448 cpumask_set_cpu(i
, hctx
->cpumask
);
2450 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
2451 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2454 * If the nr_ctx type overflows, we have exceeded the
2455 * amount of sw queues we can support.
2457 BUG_ON(!hctx
->nr_ctx
);
2461 mutex_unlock(&q
->sysfs_lock
);
2463 queue_for_each_hw_ctx(q
, hctx
, i
) {
2465 * If no software queues are mapped to this hardware queue,
2466 * disable it and free the request entries.
2468 if (!hctx
->nr_ctx
) {
2469 /* Never unmap queue 0. We need it as a
2470 * fallback in case of a new remap fails
2473 if (i
&& set
->tags
[i
])
2474 blk_mq_free_map_and_requests(set
, i
);
2480 hctx
->tags
= set
->tags
[i
];
2481 WARN_ON(!hctx
->tags
);
2484 * Set the map size to the number of mapped software queues.
2485 * This is more accurate and more efficient than looping
2486 * over all possibly mapped software queues.
2488 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2491 * Initialize batch roundrobin counts
2493 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2494 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2499 * Caller needs to ensure that we're either frozen/quiesced, or that
2500 * the queue isn't live yet.
2502 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2504 struct blk_mq_hw_ctx
*hctx
;
2507 queue_for_each_hw_ctx(q
, hctx
, i
) {
2509 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2511 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2515 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2518 struct request_queue
*q
;
2520 lockdep_assert_held(&set
->tag_list_lock
);
2522 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2523 blk_mq_freeze_queue(q
);
2524 queue_set_hctx_shared(q
, shared
);
2525 blk_mq_unfreeze_queue(q
);
2529 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2531 struct blk_mq_tag_set
*set
= q
->tag_set
;
2533 mutex_lock(&set
->tag_list_lock
);
2534 list_del_rcu(&q
->tag_set_list
);
2535 if (list_is_singular(&set
->tag_list
)) {
2536 /* just transitioned to unshared */
2537 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2538 /* update existing queue */
2539 blk_mq_update_tag_set_depth(set
, false);
2541 mutex_unlock(&set
->tag_list_lock
);
2542 INIT_LIST_HEAD(&q
->tag_set_list
);
2545 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2546 struct request_queue
*q
)
2548 mutex_lock(&set
->tag_list_lock
);
2551 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2553 if (!list_empty(&set
->tag_list
) &&
2554 !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2555 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2556 /* update existing queue */
2557 blk_mq_update_tag_set_depth(set
, true);
2559 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2560 queue_set_hctx_shared(q
, true);
2561 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2563 mutex_unlock(&set
->tag_list_lock
);
2566 /* All allocations will be freed in release handler of q->mq_kobj */
2567 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
2569 struct blk_mq_ctxs
*ctxs
;
2572 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
2576 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2577 if (!ctxs
->queue_ctx
)
2580 for_each_possible_cpu(cpu
) {
2581 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
2585 q
->mq_kobj
= &ctxs
->kobj
;
2586 q
->queue_ctx
= ctxs
->queue_ctx
;
2595 * It is the actual release handler for mq, but we do it from
2596 * request queue's release handler for avoiding use-after-free
2597 * and headache because q->mq_kobj shouldn't have been introduced,
2598 * but we can't group ctx/kctx kobj without it.
2600 void blk_mq_release(struct request_queue
*q
)
2602 struct blk_mq_hw_ctx
*hctx
;
2605 /* hctx kobj stays in hctx */
2606 queue_for_each_hw_ctx(q
, hctx
, i
) {
2609 kobject_put(&hctx
->kobj
);
2612 kfree(q
->queue_hw_ctx
);
2615 * release .mq_kobj and sw queue's kobject now because
2616 * both share lifetime with request queue.
2618 blk_mq_sysfs_deinit(q
);
2621 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2623 struct request_queue
*uninit_q
, *q
;
2625 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2627 return ERR_PTR(-ENOMEM
);
2629 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2631 blk_cleanup_queue(uninit_q
);
2635 EXPORT_SYMBOL(blk_mq_init_queue
);
2638 * Helper for setting up a queue with mq ops, given queue depth, and
2639 * the passed in mq ops flags.
2641 struct request_queue
*blk_mq_init_sq_queue(struct blk_mq_tag_set
*set
,
2642 const struct blk_mq_ops
*ops
,
2643 unsigned int queue_depth
,
2644 unsigned int set_flags
)
2646 struct request_queue
*q
;
2649 memset(set
, 0, sizeof(*set
));
2651 set
->nr_hw_queues
= 1;
2653 set
->queue_depth
= queue_depth
;
2654 set
->numa_node
= NUMA_NO_NODE
;
2655 set
->flags
= set_flags
;
2657 ret
= blk_mq_alloc_tag_set(set
);
2659 return ERR_PTR(ret
);
2661 q
= blk_mq_init_queue(set
);
2663 blk_mq_free_tag_set(set
);
2669 EXPORT_SYMBOL(blk_mq_init_sq_queue
);
2671 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2673 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2675 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, srcu
),
2676 __alignof__(struct blk_mq_hw_ctx
)) !=
2677 sizeof(struct blk_mq_hw_ctx
));
2679 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2680 hw_ctx_size
+= sizeof(struct srcu_struct
);
2685 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
2686 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
2687 int hctx_idx
, int node
)
2689 struct blk_mq_hw_ctx
*hctx
;
2691 hctx
= kzalloc_node(blk_mq_hw_ctx_size(set
),
2692 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2697 if (!zalloc_cpumask_var_node(&hctx
->cpumask
,
2698 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2704 atomic_set(&hctx
->nr_active
, 0);
2705 hctx
->numa_node
= node
;
2706 hctx
->queue_num
= hctx_idx
;
2708 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
)) {
2709 free_cpumask_var(hctx
->cpumask
);
2713 blk_mq_hctx_kobj_init(hctx
);
2718 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2719 struct request_queue
*q
)
2722 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2724 /* protect against switching io scheduler */
2725 mutex_lock(&q
->sysfs_lock
);
2726 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2728 struct blk_mq_hw_ctx
*hctx
;
2730 node
= blk_mq_hw_queue_to_node(&set
->map
[0], i
);
2732 * If the hw queue has been mapped to another numa node,
2733 * we need to realloc the hctx. If allocation fails, fallback
2734 * to use the previous one.
2736 if (hctxs
[i
] && (hctxs
[i
]->numa_node
== node
))
2739 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, node
);
2742 blk_mq_exit_hctx(q
, set
, hctxs
[i
], i
);
2743 kobject_put(&hctxs
[i
]->kobj
);
2748 pr_warn("Allocate new hctx on node %d fails,\
2749 fallback to previous one on node %d\n",
2750 node
, hctxs
[i
]->numa_node
);
2756 * Increasing nr_hw_queues fails. Free the newly allocated
2757 * hctxs and keep the previous q->nr_hw_queues.
2759 if (i
!= set
->nr_hw_queues
) {
2760 j
= q
->nr_hw_queues
;
2764 end
= q
->nr_hw_queues
;
2765 q
->nr_hw_queues
= set
->nr_hw_queues
;
2768 for (; j
< end
; j
++) {
2769 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2773 blk_mq_free_map_and_requests(set
, j
);
2774 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2775 kobject_put(&hctx
->kobj
);
2780 mutex_unlock(&q
->sysfs_lock
);
2784 * Maximum number of hardware queues we support. For single sets, we'll never
2785 * have more than the CPUs (software queues). For multiple sets, the tag_set
2786 * user may have set ->nr_hw_queues larger.
2788 static unsigned int nr_hw_queues(struct blk_mq_tag_set
*set
)
2790 if (set
->nr_maps
== 1)
2793 return max(set
->nr_hw_queues
, nr_cpu_ids
);
2796 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2797 struct request_queue
*q
)
2799 /* mark the queue as mq asap */
2800 q
->mq_ops
= set
->ops
;
2802 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2803 blk_mq_poll_stats_bkt
,
2804 BLK_MQ_POLL_STATS_BKTS
, q
);
2808 if (blk_mq_alloc_ctxs(q
))
2811 /* init q->mq_kobj and sw queues' kobjects */
2812 blk_mq_sysfs_init(q
);
2814 q
->nr_queues
= nr_hw_queues(set
);
2815 q
->queue_hw_ctx
= kcalloc_node(q
->nr_queues
, sizeof(*(q
->queue_hw_ctx
)),
2816 GFP_KERNEL
, set
->numa_node
);
2817 if (!q
->queue_hw_ctx
)
2820 blk_mq_realloc_hw_ctxs(set
, q
);
2821 if (!q
->nr_hw_queues
)
2824 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2825 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2829 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2830 if (set
->nr_maps
> HCTX_TYPE_POLL
)
2831 blk_queue_flag_set(QUEUE_FLAG_POLL
, q
);
2833 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2834 blk_queue_flag_set(QUEUE_FLAG_NO_SG_MERGE
, q
);
2836 q
->sg_reserved_size
= INT_MAX
;
2838 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2839 INIT_LIST_HEAD(&q
->requeue_list
);
2840 spin_lock_init(&q
->requeue_lock
);
2842 blk_queue_make_request(q
, blk_mq_make_request
);
2845 * Do this after blk_queue_make_request() overrides it...
2847 q
->nr_requests
= set
->queue_depth
;
2850 * Default to classic polling
2854 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2855 blk_mq_add_queue_tag_set(set
, q
);
2856 blk_mq_map_swqueue(q
);
2858 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2861 ret
= elevator_init_mq(q
);
2863 return ERR_PTR(ret
);
2869 kfree(q
->queue_hw_ctx
);
2871 blk_mq_sysfs_deinit(q
);
2874 return ERR_PTR(-ENOMEM
);
2876 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2878 void blk_mq_free_queue(struct request_queue
*q
)
2880 struct blk_mq_tag_set
*set
= q
->tag_set
;
2882 blk_mq_del_queue_tag_set(q
);
2883 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2886 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2890 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2891 if (!__blk_mq_alloc_rq_map(set
, i
))
2898 blk_mq_free_rq_map(set
->tags
[i
]);
2904 * Allocate the request maps associated with this tag_set. Note that this
2905 * may reduce the depth asked for, if memory is tight. set->queue_depth
2906 * will be updated to reflect the allocated depth.
2908 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2913 depth
= set
->queue_depth
;
2915 err
= __blk_mq_alloc_rq_maps(set
);
2919 set
->queue_depth
>>= 1;
2920 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2924 } while (set
->queue_depth
);
2926 if (!set
->queue_depth
|| err
) {
2927 pr_err("blk-mq: failed to allocate request map\n");
2931 if (depth
!= set
->queue_depth
)
2932 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2933 depth
, set
->queue_depth
);
2938 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2940 if (set
->ops
->map_queues
&& !is_kdump_kernel()) {
2944 * transport .map_queues is usually done in the following
2947 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2948 * mask = get_cpu_mask(queue)
2949 * for_each_cpu(cpu, mask)
2950 * set->map[x].mq_map[cpu] = queue;
2953 * When we need to remap, the table has to be cleared for
2954 * killing stale mapping since one CPU may not be mapped
2957 for (i
= 0; i
< set
->nr_maps
; i
++)
2958 blk_mq_clear_mq_map(&set
->map
[i
]);
2960 return set
->ops
->map_queues(set
);
2962 BUG_ON(set
->nr_maps
> 1);
2963 return blk_mq_map_queues(&set
->map
[0]);
2968 * Alloc a tag set to be associated with one or more request queues.
2969 * May fail with EINVAL for various error conditions. May adjust the
2970 * requested depth down, if it's too large. In that case, the set
2971 * value will be stored in set->queue_depth.
2973 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2977 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2979 if (!set
->nr_hw_queues
)
2981 if (!set
->queue_depth
)
2983 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2986 if (!set
->ops
->queue_rq
)
2989 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
2992 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2993 pr_info("blk-mq: reduced tag depth to %u\n",
2995 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
3000 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
3004 * If a crashdump is active, then we are potentially in a very
3005 * memory constrained environment. Limit us to 1 queue and
3006 * 64 tags to prevent using too much memory.
3008 if (is_kdump_kernel()) {
3009 set
->nr_hw_queues
= 1;
3011 set
->queue_depth
= min(64U, set
->queue_depth
);
3014 * There is no use for more h/w queues than cpus if we just have
3017 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
3018 set
->nr_hw_queues
= nr_cpu_ids
;
3020 set
->tags
= kcalloc_node(nr_hw_queues(set
), sizeof(struct blk_mq_tags
*),
3021 GFP_KERNEL
, set
->numa_node
);
3026 for (i
= 0; i
< set
->nr_maps
; i
++) {
3027 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
3028 sizeof(struct blk_mq_queue_map
),
3029 GFP_KERNEL
, set
->numa_node
);
3030 if (!set
->map
[i
].mq_map
)
3031 goto out_free_mq_map
;
3032 set
->map
[i
].nr_queues
= is_kdump_kernel() ? 1 : set
->nr_hw_queues
;
3035 ret
= blk_mq_update_queue_map(set
);
3037 goto out_free_mq_map
;
3039 ret
= blk_mq_alloc_rq_maps(set
);
3041 goto out_free_mq_map
;
3043 mutex_init(&set
->tag_list_lock
);
3044 INIT_LIST_HEAD(&set
->tag_list
);
3049 for (i
= 0; i
< set
->nr_maps
; i
++) {
3050 kfree(set
->map
[i
].mq_map
);
3051 set
->map
[i
].mq_map
= NULL
;
3057 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
3059 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
3063 for (i
= 0; i
< nr_hw_queues(set
); i
++)
3064 blk_mq_free_map_and_requests(set
, i
);
3066 for (j
= 0; j
< set
->nr_maps
; j
++) {
3067 kfree(set
->map
[j
].mq_map
);
3068 set
->map
[j
].mq_map
= NULL
;
3074 EXPORT_SYMBOL(blk_mq_free_tag_set
);
3076 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
3078 struct blk_mq_tag_set
*set
= q
->tag_set
;
3079 struct blk_mq_hw_ctx
*hctx
;
3085 blk_mq_freeze_queue(q
);
3086 blk_mq_quiesce_queue(q
);
3089 queue_for_each_hw_ctx(q
, hctx
, i
) {
3093 * If we're using an MQ scheduler, just update the scheduler
3094 * queue depth. This is similar to what the old code would do.
3096 if (!hctx
->sched_tags
) {
3097 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
3100 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
3108 q
->nr_requests
= nr
;
3110 blk_mq_unquiesce_queue(q
);
3111 blk_mq_unfreeze_queue(q
);
3117 * request_queue and elevator_type pair.
3118 * It is just used by __blk_mq_update_nr_hw_queues to cache
3119 * the elevator_type associated with a request_queue.
3121 struct blk_mq_qe_pair
{
3122 struct list_head node
;
3123 struct request_queue
*q
;
3124 struct elevator_type
*type
;
3128 * Cache the elevator_type in qe pair list and switch the
3129 * io scheduler to 'none'
3131 static bool blk_mq_elv_switch_none(struct list_head
*head
,
3132 struct request_queue
*q
)
3134 struct blk_mq_qe_pair
*qe
;
3139 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
3143 INIT_LIST_HEAD(&qe
->node
);
3145 qe
->type
= q
->elevator
->type
;
3146 list_add(&qe
->node
, head
);
3148 mutex_lock(&q
->sysfs_lock
);
3150 * After elevator_switch_mq, the previous elevator_queue will be
3151 * released by elevator_release. The reference of the io scheduler
3152 * module get by elevator_get will also be put. So we need to get
3153 * a reference of the io scheduler module here to prevent it to be
3156 __module_get(qe
->type
->elevator_owner
);
3157 elevator_switch_mq(q
, NULL
);
3158 mutex_unlock(&q
->sysfs_lock
);
3163 static void blk_mq_elv_switch_back(struct list_head
*head
,
3164 struct request_queue
*q
)
3166 struct blk_mq_qe_pair
*qe
;
3167 struct elevator_type
*t
= NULL
;
3169 list_for_each_entry(qe
, head
, node
)
3178 list_del(&qe
->node
);
3181 mutex_lock(&q
->sysfs_lock
);
3182 elevator_switch_mq(q
, t
);
3183 mutex_unlock(&q
->sysfs_lock
);
3186 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
3189 struct request_queue
*q
;
3191 int prev_nr_hw_queues
;
3193 lockdep_assert_held(&set
->tag_list_lock
);
3195 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
3196 nr_hw_queues
= nr_cpu_ids
;
3197 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
3200 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3201 blk_mq_freeze_queue(q
);
3203 * Sync with blk_mq_queue_tag_busy_iter.
3207 * Switch IO scheduler to 'none', cleaning up the data associated
3208 * with the previous scheduler. We will switch back once we are done
3209 * updating the new sw to hw queue mappings.
3211 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3212 if (!blk_mq_elv_switch_none(&head
, q
))
3215 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3216 blk_mq_debugfs_unregister_hctxs(q
);
3217 blk_mq_sysfs_unregister(q
);
3220 prev_nr_hw_queues
= set
->nr_hw_queues
;
3221 set
->nr_hw_queues
= nr_hw_queues
;
3222 blk_mq_update_queue_map(set
);
3224 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3225 blk_mq_realloc_hw_ctxs(set
, q
);
3226 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
3227 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3228 nr_hw_queues
, prev_nr_hw_queues
);
3229 set
->nr_hw_queues
= prev_nr_hw_queues
;
3230 blk_mq_map_queues(&set
->map
[0]);
3233 blk_mq_map_swqueue(q
);
3236 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3237 blk_mq_sysfs_register(q
);
3238 blk_mq_debugfs_register_hctxs(q
);
3242 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3243 blk_mq_elv_switch_back(&head
, q
);
3245 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3246 blk_mq_unfreeze_queue(q
);
3249 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
3251 mutex_lock(&set
->tag_list_lock
);
3252 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
3253 mutex_unlock(&set
->tag_list_lock
);
3255 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
3257 /* Enable polling stats and return whether they were already enabled. */
3258 static bool blk_poll_stats_enable(struct request_queue
*q
)
3260 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3261 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS
, q
))
3263 blk_stat_add_callback(q
, q
->poll_cb
);
3267 static void blk_mq_poll_stats_start(struct request_queue
*q
)
3270 * We don't arm the callback if polling stats are not enabled or the
3271 * callback is already active.
3273 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3274 blk_stat_is_active(q
->poll_cb
))
3277 blk_stat_activate_msecs(q
->poll_cb
, 100);
3280 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
3282 struct request_queue
*q
= cb
->data
;
3285 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
3286 if (cb
->stat
[bucket
].nr_samples
)
3287 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
3291 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
3292 struct blk_mq_hw_ctx
*hctx
,
3295 unsigned long ret
= 0;
3299 * If stats collection isn't on, don't sleep but turn it on for
3302 if (!blk_poll_stats_enable(q
))
3306 * As an optimistic guess, use half of the mean service time
3307 * for this type of request. We can (and should) make this smarter.
3308 * For instance, if the completion latencies are tight, we can
3309 * get closer than just half the mean. This is especially
3310 * important on devices where the completion latencies are longer
3311 * than ~10 usec. We do use the stats for the relevant IO size
3312 * if available which does lead to better estimates.
3314 bucket
= blk_mq_poll_stats_bkt(rq
);
3318 if (q
->poll_stat
[bucket
].nr_samples
)
3319 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
3324 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
3325 struct blk_mq_hw_ctx
*hctx
,
3328 struct hrtimer_sleeper hs
;
3329 enum hrtimer_mode mode
;
3333 if (rq
->rq_flags
& RQF_MQ_POLL_SLEPT
)
3337 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3339 * 0: use half of prev avg
3340 * >0: use this specific value
3342 if (q
->poll_nsec
> 0)
3343 nsecs
= q
->poll_nsec
;
3345 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
3350 rq
->rq_flags
|= RQF_MQ_POLL_SLEPT
;
3353 * This will be replaced with the stats tracking code, using
3354 * 'avg_completion_time / 2' as the pre-sleep target.
3358 mode
= HRTIMER_MODE_REL
;
3359 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
3360 hrtimer_set_expires(&hs
.timer
, kt
);
3362 hrtimer_init_sleeper(&hs
, current
);
3364 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
3366 set_current_state(TASK_UNINTERRUPTIBLE
);
3367 hrtimer_start_expires(&hs
.timer
, mode
);
3370 hrtimer_cancel(&hs
.timer
);
3371 mode
= HRTIMER_MODE_ABS
;
3372 } while (hs
.task
&& !signal_pending(current
));
3374 __set_current_state(TASK_RUNNING
);
3375 destroy_hrtimer_on_stack(&hs
.timer
);
3379 static bool blk_mq_poll_hybrid(struct request_queue
*q
,
3380 struct blk_mq_hw_ctx
*hctx
, blk_qc_t cookie
)
3384 if (q
->poll_nsec
== -1)
3387 if (!blk_qc_t_is_internal(cookie
))
3388 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3390 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3392 * With scheduling, if the request has completed, we'll
3393 * get a NULL return here, as we clear the sched tag when
3394 * that happens. The request still remains valid, like always,
3395 * so we should be safe with just the NULL check.
3401 return blk_mq_poll_hybrid_sleep(q
, hctx
, rq
);
3405 * blk_poll - poll for IO completions
3407 * @cookie: cookie passed back at IO submission time
3408 * @spin: whether to spin for completions
3411 * Poll for completions on the passed in queue. Returns number of
3412 * completed entries found. If @spin is true, then blk_poll will continue
3413 * looping until at least one completion is found, unless the task is
3414 * otherwise marked running (or we need to reschedule).
3416 int blk_poll(struct request_queue
*q
, blk_qc_t cookie
, bool spin
)
3418 struct blk_mq_hw_ctx
*hctx
;
3421 if (!blk_qc_t_valid(cookie
) ||
3422 !test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3426 blk_flush_plug_list(current
->plug
, false);
3428 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3431 * If we sleep, have the caller restart the poll loop to reset
3432 * the state. Like for the other success return cases, the
3433 * caller is responsible for checking if the IO completed. If
3434 * the IO isn't complete, we'll get called again and will go
3435 * straight to the busy poll loop.
3437 if (blk_mq_poll_hybrid(q
, hctx
, cookie
))
3440 hctx
->poll_considered
++;
3442 state
= current
->state
;
3446 hctx
->poll_invoked
++;
3448 ret
= q
->mq_ops
->poll(hctx
);
3450 hctx
->poll_success
++;
3451 __set_current_state(TASK_RUNNING
);
3455 if (signal_pending_state(state
, current
))
3456 __set_current_state(TASK_RUNNING
);
3458 if (current
->state
== TASK_RUNNING
)
3460 if (ret
< 0 || !spin
)
3463 } while (!need_resched());
3465 __set_current_state(TASK_RUNNING
);
3468 EXPORT_SYMBOL_GPL(blk_poll
);
3470 unsigned int blk_mq_rq_cpu(struct request
*rq
)
3472 return rq
->mq_ctx
->cpu
;
3474 EXPORT_SYMBOL(blk_mq_rq_cpu
);
3476 static int __init
blk_mq_init(void)
3478 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3479 blk_mq_hctx_notify_dead
);
3482 subsys_initcall(blk_mq_init
);