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 int blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
, bool spin
);
42 static void blk_mq_poll_stats_start(struct request_queue
*q
);
43 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
);
45 static int blk_mq_poll_stats_bkt(const struct request
*rq
)
47 int ddir
, bytes
, bucket
;
49 ddir
= rq_data_dir(rq
);
50 bytes
= blk_rq_bytes(rq
);
52 bucket
= ddir
+ 2*(ilog2(bytes
) - 9);
56 else if (bucket
>= BLK_MQ_POLL_STATS_BKTS
)
57 return ddir
+ BLK_MQ_POLL_STATS_BKTS
- 2;
63 * Check if any of the ctx's have pending work in this hardware queue
65 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx
*hctx
)
67 return !list_empty_careful(&hctx
->dispatch
) ||
68 sbitmap_any_bit_set(&hctx
->ctx_map
) ||
69 blk_mq_sched_has_work(hctx
);
73 * Mark this ctx as having pending work in this hardware queue
75 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx
*hctx
,
76 struct blk_mq_ctx
*ctx
)
78 const int bit
= ctx
->index_hw
[hctx
->type
];
80 if (!sbitmap_test_bit(&hctx
->ctx_map
, bit
))
81 sbitmap_set_bit(&hctx
->ctx_map
, bit
);
84 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx
*hctx
,
85 struct blk_mq_ctx
*ctx
)
87 const int bit
= ctx
->index_hw
[hctx
->type
];
89 sbitmap_clear_bit(&hctx
->ctx_map
, bit
);
93 struct hd_struct
*part
;
94 unsigned int *inflight
;
97 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx
*hctx
,
98 struct request
*rq
, void *priv
,
101 struct mq_inflight
*mi
= priv
;
104 * index[0] counts the specific partition that was asked for. index[1]
105 * counts the ones that are active on the whole device, so increment
106 * that if mi->part is indeed a partition, and not a whole device.
108 if (rq
->part
== mi
->part
)
110 if (mi
->part
->partno
)
116 void blk_mq_in_flight(struct request_queue
*q
, struct hd_struct
*part
,
117 unsigned int inflight
[2])
119 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
121 inflight
[0] = inflight
[1] = 0;
122 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight
, &mi
);
125 static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx
*hctx
,
126 struct request
*rq
, void *priv
,
129 struct mq_inflight
*mi
= priv
;
131 if (rq
->part
== mi
->part
)
132 mi
->inflight
[rq_data_dir(rq
)]++;
137 void blk_mq_in_flight_rw(struct request_queue
*q
, struct hd_struct
*part
,
138 unsigned int inflight
[2])
140 struct mq_inflight mi
= { .part
= part
, .inflight
= inflight
, };
142 inflight
[0] = inflight
[1] = 0;
143 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_inflight_rw
, &mi
);
146 void blk_freeze_queue_start(struct request_queue
*q
)
150 freeze_depth
= atomic_inc_return(&q
->mq_freeze_depth
);
151 if (freeze_depth
== 1) {
152 percpu_ref_kill(&q
->q_usage_counter
);
154 blk_mq_run_hw_queues(q
, false);
157 EXPORT_SYMBOL_GPL(blk_freeze_queue_start
);
159 void blk_mq_freeze_queue_wait(struct request_queue
*q
)
161 wait_event(q
->mq_freeze_wq
, percpu_ref_is_zero(&q
->q_usage_counter
));
163 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait
);
165 int blk_mq_freeze_queue_wait_timeout(struct request_queue
*q
,
166 unsigned long timeout
)
168 return wait_event_timeout(q
->mq_freeze_wq
,
169 percpu_ref_is_zero(&q
->q_usage_counter
),
172 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout
);
175 * Guarantee no request is in use, so we can change any data structure of
176 * the queue afterward.
178 void blk_freeze_queue(struct request_queue
*q
)
181 * In the !blk_mq case we are only calling this to kill the
182 * q_usage_counter, otherwise this increases the freeze depth
183 * and waits for it to return to zero. For this reason there is
184 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
185 * exported to drivers as the only user for unfreeze is blk_mq.
187 blk_freeze_queue_start(q
);
188 blk_mq_freeze_queue_wait(q
);
191 void blk_mq_freeze_queue(struct request_queue
*q
)
194 * ...just an alias to keep freeze and unfreeze actions balanced
195 * in the blk_mq_* namespace
199 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue
);
201 void blk_mq_unfreeze_queue(struct request_queue
*q
)
205 freeze_depth
= atomic_dec_return(&q
->mq_freeze_depth
);
206 WARN_ON_ONCE(freeze_depth
< 0);
208 percpu_ref_resurrect(&q
->q_usage_counter
);
209 wake_up_all(&q
->mq_freeze_wq
);
212 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue
);
215 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
216 * mpt3sas driver such that this function can be removed.
218 void blk_mq_quiesce_queue_nowait(struct request_queue
*q
)
220 blk_queue_flag_set(QUEUE_FLAG_QUIESCED
, q
);
222 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait
);
225 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
228 * Note: this function does not prevent that the struct request end_io()
229 * callback function is invoked. Once this function is returned, we make
230 * sure no dispatch can happen until the queue is unquiesced via
231 * blk_mq_unquiesce_queue().
233 void blk_mq_quiesce_queue(struct request_queue
*q
)
235 struct blk_mq_hw_ctx
*hctx
;
239 blk_mq_quiesce_queue_nowait(q
);
241 queue_for_each_hw_ctx(q
, hctx
, i
) {
242 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
243 synchronize_srcu(hctx
->srcu
);
250 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue
);
253 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
256 * This function recovers queue into the state before quiescing
257 * which is done by blk_mq_quiesce_queue.
259 void blk_mq_unquiesce_queue(struct request_queue
*q
)
261 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED
, q
);
263 /* dispatch requests which are inserted during quiescing */
264 blk_mq_run_hw_queues(q
, true);
266 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue
);
268 void blk_mq_wake_waiters(struct request_queue
*q
)
270 struct blk_mq_hw_ctx
*hctx
;
273 queue_for_each_hw_ctx(q
, hctx
, i
)
274 if (blk_mq_hw_queue_mapped(hctx
))
275 blk_mq_tag_wakeup_all(hctx
->tags
, true);
278 bool blk_mq_can_queue(struct blk_mq_hw_ctx
*hctx
)
280 return blk_mq_has_free_tags(hctx
->tags
);
282 EXPORT_SYMBOL(blk_mq_can_queue
);
284 static struct request
*blk_mq_rq_ctx_init(struct blk_mq_alloc_data
*data
,
285 unsigned int tag
, unsigned int op
)
287 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
288 struct request
*rq
= tags
->static_rqs
[tag
];
289 req_flags_t rq_flags
= 0;
291 if (data
->flags
& BLK_MQ_REQ_INTERNAL
) {
293 rq
->internal_tag
= tag
;
295 if (data
->hctx
->flags
& BLK_MQ_F_TAG_SHARED
) {
296 rq_flags
= RQF_MQ_INFLIGHT
;
297 atomic_inc(&data
->hctx
->nr_active
);
300 rq
->internal_tag
= -1;
301 data
->hctx
->tags
->rqs
[rq
->tag
] = rq
;
304 /* csd/requeue_work/fifo_time is initialized before use */
306 rq
->mq_ctx
= data
->ctx
;
307 rq
->mq_hctx
= data
->hctx
;
308 rq
->rq_flags
= rq_flags
;
310 if (data
->flags
& BLK_MQ_REQ_PREEMPT
)
311 rq
->rq_flags
|= RQF_PREEMPT
;
312 if (blk_queue_io_stat(data
->q
))
313 rq
->rq_flags
|= RQF_IO_STAT
;
314 INIT_LIST_HEAD(&rq
->queuelist
);
315 INIT_HLIST_NODE(&rq
->hash
);
316 RB_CLEAR_NODE(&rq
->rb_node
);
319 rq
->start_time_ns
= ktime_get_ns();
320 rq
->io_start_time_ns
= 0;
321 rq
->nr_phys_segments
= 0;
322 #if defined(CONFIG_BLK_DEV_INTEGRITY)
323 rq
->nr_integrity_segments
= 0;
326 /* tag was already set */
328 WRITE_ONCE(rq
->deadline
, 0);
333 rq
->end_io_data
= NULL
;
336 data
->ctx
->rq_dispatched
[op_is_sync(op
)]++;
337 refcount_set(&rq
->ref
, 1);
341 static struct request
*blk_mq_get_request(struct request_queue
*q
,
343 struct blk_mq_alloc_data
*data
)
345 struct elevator_queue
*e
= q
->elevator
;
348 bool put_ctx_on_error
= false;
350 blk_queue_enter_live(q
);
352 if (likely(!data
->ctx
)) {
353 data
->ctx
= blk_mq_get_ctx(q
);
354 put_ctx_on_error
= true;
356 if (likely(!data
->hctx
))
357 data
->hctx
= blk_mq_map_queue(q
, data
->cmd_flags
,
359 if (data
->cmd_flags
& REQ_NOWAIT
)
360 data
->flags
|= BLK_MQ_REQ_NOWAIT
;
363 data
->flags
|= BLK_MQ_REQ_INTERNAL
;
366 * Flush requests are special and go directly to the
367 * dispatch list. Don't include reserved tags in the
368 * limiting, as it isn't useful.
370 if (!op_is_flush(data
->cmd_flags
) &&
371 e
->type
->ops
.limit_depth
&&
372 !(data
->flags
& BLK_MQ_REQ_RESERVED
))
373 e
->type
->ops
.limit_depth(data
->cmd_flags
, data
);
375 blk_mq_tag_busy(data
->hctx
);
378 tag
= blk_mq_get_tag(data
);
379 if (tag
== BLK_MQ_TAG_FAIL
) {
380 if (put_ctx_on_error
) {
381 blk_mq_put_ctx(data
->ctx
);
388 rq
= blk_mq_rq_ctx_init(data
, tag
, data
->cmd_flags
);
389 if (!op_is_flush(data
->cmd_flags
)) {
391 if (e
&& e
->type
->ops
.prepare_request
) {
392 if (e
->type
->icq_cache
)
393 blk_mq_sched_assign_ioc(rq
);
395 e
->type
->ops
.prepare_request(rq
, bio
);
396 rq
->rq_flags
|= RQF_ELVPRIV
;
399 data
->hctx
->queued
++;
403 struct request
*blk_mq_alloc_request(struct request_queue
*q
, unsigned int op
,
404 blk_mq_req_flags_t flags
)
406 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
, .cmd_flags
= op
};
410 ret
= blk_queue_enter(q
, flags
);
414 rq
= blk_mq_get_request(q
, NULL
, &alloc_data
);
418 return ERR_PTR(-EWOULDBLOCK
);
420 blk_mq_put_ctx(alloc_data
.ctx
);
423 rq
->__sector
= (sector_t
) -1;
424 rq
->bio
= rq
->biotail
= NULL
;
427 EXPORT_SYMBOL(blk_mq_alloc_request
);
429 struct request
*blk_mq_alloc_request_hctx(struct request_queue
*q
,
430 unsigned int op
, blk_mq_req_flags_t flags
, unsigned int hctx_idx
)
432 struct blk_mq_alloc_data alloc_data
= { .flags
= flags
, .cmd_flags
= op
};
438 * If the tag allocator sleeps we could get an allocation for a
439 * different hardware context. No need to complicate the low level
440 * allocator for this for the rare use case of a command tied to
443 if (WARN_ON_ONCE(!(flags
& BLK_MQ_REQ_NOWAIT
)))
444 return ERR_PTR(-EINVAL
);
446 if (hctx_idx
>= q
->nr_hw_queues
)
447 return ERR_PTR(-EIO
);
449 ret
= blk_queue_enter(q
, flags
);
454 * Check if the hardware context is actually mapped to anything.
455 * If not tell the caller that it should skip this queue.
457 alloc_data
.hctx
= q
->queue_hw_ctx
[hctx_idx
];
458 if (!blk_mq_hw_queue_mapped(alloc_data
.hctx
)) {
460 return ERR_PTR(-EXDEV
);
462 cpu
= cpumask_first_and(alloc_data
.hctx
->cpumask
, cpu_online_mask
);
463 alloc_data
.ctx
= __blk_mq_get_ctx(q
, cpu
);
465 rq
= blk_mq_get_request(q
, NULL
, &alloc_data
);
469 return ERR_PTR(-EWOULDBLOCK
);
473 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx
);
475 static void __blk_mq_free_request(struct request
*rq
)
477 struct request_queue
*q
= rq
->q
;
478 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
479 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
480 const int sched_tag
= rq
->internal_tag
;
482 blk_pm_mark_last_busy(rq
);
485 blk_mq_put_tag(hctx
, hctx
->tags
, ctx
, rq
->tag
);
487 blk_mq_put_tag(hctx
, hctx
->sched_tags
, ctx
, sched_tag
);
488 blk_mq_sched_restart(hctx
);
492 void blk_mq_free_request(struct request
*rq
)
494 struct request_queue
*q
= rq
->q
;
495 struct elevator_queue
*e
= q
->elevator
;
496 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
497 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
499 if (rq
->rq_flags
& RQF_ELVPRIV
) {
500 if (e
&& e
->type
->ops
.finish_request
)
501 e
->type
->ops
.finish_request(rq
);
503 put_io_context(rq
->elv
.icq
->ioc
);
508 ctx
->rq_completed
[rq_is_sync(rq
)]++;
509 if (rq
->rq_flags
& RQF_MQ_INFLIGHT
)
510 atomic_dec(&hctx
->nr_active
);
512 if (unlikely(laptop_mode
&& !blk_rq_is_passthrough(rq
)))
513 laptop_io_completion(q
->backing_dev_info
);
517 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
518 if (refcount_dec_and_test(&rq
->ref
))
519 __blk_mq_free_request(rq
);
521 EXPORT_SYMBOL_GPL(blk_mq_free_request
);
523 inline void __blk_mq_end_request(struct request
*rq
, blk_status_t error
)
525 u64 now
= ktime_get_ns();
527 if (rq
->rq_flags
& RQF_STATS
) {
528 blk_mq_poll_stats_start(rq
->q
);
529 blk_stat_add(rq
, now
);
532 if (rq
->internal_tag
!= -1)
533 blk_mq_sched_completed_request(rq
, now
);
535 blk_account_io_done(rq
, now
);
538 rq_qos_done(rq
->q
, rq
);
539 rq
->end_io(rq
, error
);
541 if (unlikely(blk_bidi_rq(rq
)))
542 blk_mq_free_request(rq
->next_rq
);
543 blk_mq_free_request(rq
);
546 EXPORT_SYMBOL(__blk_mq_end_request
);
548 void blk_mq_end_request(struct request
*rq
, blk_status_t error
)
550 if (blk_update_request(rq
, error
, blk_rq_bytes(rq
)))
552 __blk_mq_end_request(rq
, error
);
554 EXPORT_SYMBOL(blk_mq_end_request
);
556 static void __blk_mq_complete_request_remote(void *data
)
558 struct request
*rq
= data
;
559 struct request_queue
*q
= rq
->q
;
561 q
->mq_ops
->complete(rq
);
564 static void __blk_mq_complete_request(struct request
*rq
)
566 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
567 struct request_queue
*q
= rq
->q
;
571 WRITE_ONCE(rq
->state
, MQ_RQ_COMPLETE
);
573 * Most of single queue controllers, there is only one irq vector
574 * for handling IO completion, and the only irq's affinity is set
575 * as all possible CPUs. On most of ARCHs, this affinity means the
576 * irq is handled on one specific CPU.
578 * So complete IO reqeust in softirq context in case of single queue
579 * for not degrading IO performance by irqsoff latency.
581 if (q
->nr_hw_queues
== 1) {
582 __blk_complete_request(rq
);
587 * For a polled request, always complete locallly, it's pointless
588 * to redirect the completion.
590 if ((rq
->cmd_flags
& REQ_HIPRI
) ||
591 !test_bit(QUEUE_FLAG_SAME_COMP
, &q
->queue_flags
)) {
592 q
->mq_ops
->complete(rq
);
597 if (!test_bit(QUEUE_FLAG_SAME_FORCE
, &q
->queue_flags
))
598 shared
= cpus_share_cache(cpu
, ctx
->cpu
);
600 if (cpu
!= ctx
->cpu
&& !shared
&& cpu_online(ctx
->cpu
)) {
601 rq
->csd
.func
= __blk_mq_complete_request_remote
;
604 smp_call_function_single_async(ctx
->cpu
, &rq
->csd
);
606 q
->mq_ops
->complete(rq
);
611 static void hctx_unlock(struct blk_mq_hw_ctx
*hctx
, int srcu_idx
)
612 __releases(hctx
->srcu
)
614 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
))
617 srcu_read_unlock(hctx
->srcu
, srcu_idx
);
620 static void hctx_lock(struct blk_mq_hw_ctx
*hctx
, int *srcu_idx
)
621 __acquires(hctx
->srcu
)
623 if (!(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
624 /* shut up gcc false positive */
628 *srcu_idx
= srcu_read_lock(hctx
->srcu
);
632 * blk_mq_complete_request - end I/O on a request
633 * @rq: the request being processed
636 * Ends all I/O on a request. It does not handle partial completions.
637 * The actual completion happens out-of-order, through a IPI handler.
639 bool blk_mq_complete_request(struct request
*rq
)
641 if (unlikely(blk_should_fake_timeout(rq
->q
)))
643 __blk_mq_complete_request(rq
);
646 EXPORT_SYMBOL(blk_mq_complete_request
);
648 int blk_mq_request_started(struct request
*rq
)
650 return blk_mq_rq_state(rq
) != MQ_RQ_IDLE
;
652 EXPORT_SYMBOL_GPL(blk_mq_request_started
);
654 void blk_mq_start_request(struct request
*rq
)
656 struct request_queue
*q
= rq
->q
;
658 blk_mq_sched_started_request(rq
);
660 trace_block_rq_issue(q
, rq
);
662 if (test_bit(QUEUE_FLAG_STATS
, &q
->queue_flags
)) {
663 rq
->io_start_time_ns
= ktime_get_ns();
664 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
665 rq
->throtl_size
= blk_rq_sectors(rq
);
667 rq
->rq_flags
|= RQF_STATS
;
671 WARN_ON_ONCE(blk_mq_rq_state(rq
) != MQ_RQ_IDLE
);
674 WRITE_ONCE(rq
->state
, MQ_RQ_IN_FLIGHT
);
676 if (q
->dma_drain_size
&& blk_rq_bytes(rq
)) {
678 * Make sure space for the drain appears. We know we can do
679 * this because max_hw_segments has been adjusted to be one
680 * fewer than the device can handle.
682 rq
->nr_phys_segments
++;
685 EXPORT_SYMBOL(blk_mq_start_request
);
687 static void __blk_mq_requeue_request(struct request
*rq
)
689 struct request_queue
*q
= rq
->q
;
691 blk_mq_put_driver_tag(rq
);
693 trace_block_rq_requeue(q
, rq
);
694 rq_qos_requeue(q
, rq
);
696 if (blk_mq_request_started(rq
)) {
697 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
698 rq
->rq_flags
&= ~RQF_TIMED_OUT
;
699 if (q
->dma_drain_size
&& blk_rq_bytes(rq
))
700 rq
->nr_phys_segments
--;
704 void blk_mq_requeue_request(struct request
*rq
, bool kick_requeue_list
)
706 __blk_mq_requeue_request(rq
);
708 /* this request will be re-inserted to io scheduler queue */
709 blk_mq_sched_requeue_request(rq
);
711 BUG_ON(!list_empty(&rq
->queuelist
));
712 blk_mq_add_to_requeue_list(rq
, true, kick_requeue_list
);
714 EXPORT_SYMBOL(blk_mq_requeue_request
);
716 static void blk_mq_requeue_work(struct work_struct
*work
)
718 struct request_queue
*q
=
719 container_of(work
, struct request_queue
, requeue_work
.work
);
721 struct request
*rq
, *next
;
723 spin_lock_irq(&q
->requeue_lock
);
724 list_splice_init(&q
->requeue_list
, &rq_list
);
725 spin_unlock_irq(&q
->requeue_lock
);
727 list_for_each_entry_safe(rq
, next
, &rq_list
, queuelist
) {
728 if (!(rq
->rq_flags
& RQF_SOFTBARRIER
))
731 rq
->rq_flags
&= ~RQF_SOFTBARRIER
;
732 list_del_init(&rq
->queuelist
);
733 blk_mq_sched_insert_request(rq
, true, false, false);
736 while (!list_empty(&rq_list
)) {
737 rq
= list_entry(rq_list
.next
, struct request
, queuelist
);
738 list_del_init(&rq
->queuelist
);
739 blk_mq_sched_insert_request(rq
, false, false, false);
742 blk_mq_run_hw_queues(q
, false);
745 void blk_mq_add_to_requeue_list(struct request
*rq
, bool at_head
,
746 bool kick_requeue_list
)
748 struct request_queue
*q
= rq
->q
;
752 * We abuse this flag that is otherwise used by the I/O scheduler to
753 * request head insertion from the workqueue.
755 BUG_ON(rq
->rq_flags
& RQF_SOFTBARRIER
);
757 spin_lock_irqsave(&q
->requeue_lock
, flags
);
759 rq
->rq_flags
|= RQF_SOFTBARRIER
;
760 list_add(&rq
->queuelist
, &q
->requeue_list
);
762 list_add_tail(&rq
->queuelist
, &q
->requeue_list
);
764 spin_unlock_irqrestore(&q
->requeue_lock
, flags
);
766 if (kick_requeue_list
)
767 blk_mq_kick_requeue_list(q
);
769 EXPORT_SYMBOL(blk_mq_add_to_requeue_list
);
771 void blk_mq_kick_requeue_list(struct request_queue
*q
)
773 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
, 0);
775 EXPORT_SYMBOL(blk_mq_kick_requeue_list
);
777 void blk_mq_delay_kick_requeue_list(struct request_queue
*q
,
780 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND
, &q
->requeue_work
,
781 msecs_to_jiffies(msecs
));
783 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list
);
785 struct request
*blk_mq_tag_to_rq(struct blk_mq_tags
*tags
, unsigned int tag
)
787 if (tag
< tags
->nr_tags
) {
788 prefetch(tags
->rqs
[tag
]);
789 return tags
->rqs
[tag
];
794 EXPORT_SYMBOL(blk_mq_tag_to_rq
);
796 static bool blk_mq_check_busy(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
797 void *priv
, bool reserved
)
800 * If we find a request, we know the queue is busy. Return false
801 * to stop the iteration.
803 if (rq
->q
== hctx
->queue
) {
813 bool blk_mq_queue_busy(struct request_queue
*q
)
817 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_busy
, &busy
);
820 EXPORT_SYMBOL_GPL(blk_mq_queue_busy
);
822 static void blk_mq_rq_timed_out(struct request
*req
, bool reserved
)
824 req
->rq_flags
|= RQF_TIMED_OUT
;
825 if (req
->q
->mq_ops
->timeout
) {
826 enum blk_eh_timer_return ret
;
828 ret
= req
->q
->mq_ops
->timeout(req
, reserved
);
829 if (ret
== BLK_EH_DONE
)
831 WARN_ON_ONCE(ret
!= BLK_EH_RESET_TIMER
);
837 static bool blk_mq_req_expired(struct request
*rq
, unsigned long *next
)
839 unsigned long deadline
;
841 if (blk_mq_rq_state(rq
) != MQ_RQ_IN_FLIGHT
)
843 if (rq
->rq_flags
& RQF_TIMED_OUT
)
846 deadline
= READ_ONCE(rq
->deadline
);
847 if (time_after_eq(jiffies
, deadline
))
852 else if (time_after(*next
, deadline
))
857 static bool blk_mq_check_expired(struct blk_mq_hw_ctx
*hctx
,
858 struct request
*rq
, void *priv
, bool reserved
)
860 unsigned long *next
= priv
;
863 * Just do a quick check if it is expired before locking the request in
864 * so we're not unnecessarilly synchronizing across CPUs.
866 if (!blk_mq_req_expired(rq
, next
))
870 * We have reason to believe the request may be expired. Take a
871 * reference on the request to lock this request lifetime into its
872 * currently allocated context to prevent it from being reallocated in
873 * the event the completion by-passes this timeout handler.
875 * If the reference was already released, then the driver beat the
876 * timeout handler to posting a natural completion.
878 if (!refcount_inc_not_zero(&rq
->ref
))
882 * The request is now locked and cannot be reallocated underneath the
883 * timeout handler's processing. Re-verify this exact request is truly
884 * expired; if it is not expired, then the request was completed and
885 * reallocated as a new request.
887 if (blk_mq_req_expired(rq
, next
))
888 blk_mq_rq_timed_out(rq
, reserved
);
889 if (refcount_dec_and_test(&rq
->ref
))
890 __blk_mq_free_request(rq
);
895 static void blk_mq_timeout_work(struct work_struct
*work
)
897 struct request_queue
*q
=
898 container_of(work
, struct request_queue
, timeout_work
);
899 unsigned long next
= 0;
900 struct blk_mq_hw_ctx
*hctx
;
903 /* A deadlock might occur if a request is stuck requiring a
904 * timeout at the same time a queue freeze is waiting
905 * completion, since the timeout code would not be able to
906 * acquire the queue reference here.
908 * That's why we don't use blk_queue_enter here; instead, we use
909 * percpu_ref_tryget directly, because we need to be able to
910 * obtain a reference even in the short window between the queue
911 * starting to freeze, by dropping the first reference in
912 * blk_freeze_queue_start, and the moment the last request is
913 * consumed, marked by the instant q_usage_counter reaches
916 if (!percpu_ref_tryget(&q
->q_usage_counter
))
919 blk_mq_queue_tag_busy_iter(q
, blk_mq_check_expired
, &next
);
922 mod_timer(&q
->timeout
, next
);
925 * Request timeouts are handled as a forward rolling timer. If
926 * we end up here it means that no requests are pending and
927 * also that no request has been pending for a while. Mark
930 queue_for_each_hw_ctx(q
, hctx
, i
) {
931 /* the hctx may be unmapped, so check it here */
932 if (blk_mq_hw_queue_mapped(hctx
))
933 blk_mq_tag_idle(hctx
);
939 struct flush_busy_ctx_data
{
940 struct blk_mq_hw_ctx
*hctx
;
941 struct list_head
*list
;
944 static bool flush_busy_ctx(struct sbitmap
*sb
, unsigned int bitnr
, void *data
)
946 struct flush_busy_ctx_data
*flush_data
= data
;
947 struct blk_mq_hw_ctx
*hctx
= flush_data
->hctx
;
948 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
950 spin_lock(&ctx
->lock
);
951 list_splice_tail_init(&ctx
->rq_list
, flush_data
->list
);
952 sbitmap_clear_bit(sb
, bitnr
);
953 spin_unlock(&ctx
->lock
);
958 * Process software queues that have been marked busy, splicing them
959 * to the for-dispatch
961 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx
*hctx
, struct list_head
*list
)
963 struct flush_busy_ctx_data data
= {
968 sbitmap_for_each_set(&hctx
->ctx_map
, flush_busy_ctx
, &data
);
970 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs
);
972 struct dispatch_rq_data
{
973 struct blk_mq_hw_ctx
*hctx
;
977 static bool dispatch_rq_from_ctx(struct sbitmap
*sb
, unsigned int bitnr
,
980 struct dispatch_rq_data
*dispatch_data
= data
;
981 struct blk_mq_hw_ctx
*hctx
= dispatch_data
->hctx
;
982 struct blk_mq_ctx
*ctx
= hctx
->ctxs
[bitnr
];
984 spin_lock(&ctx
->lock
);
985 if (!list_empty(&ctx
->rq_list
)) {
986 dispatch_data
->rq
= list_entry_rq(ctx
->rq_list
.next
);
987 list_del_init(&dispatch_data
->rq
->queuelist
);
988 if (list_empty(&ctx
->rq_list
))
989 sbitmap_clear_bit(sb
, bitnr
);
991 spin_unlock(&ctx
->lock
);
993 return !dispatch_data
->rq
;
996 struct request
*blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx
*hctx
,
997 struct blk_mq_ctx
*start
)
999 unsigned off
= start
? start
->index_hw
[hctx
->type
] : 0;
1000 struct dispatch_rq_data data
= {
1005 __sbitmap_for_each_set(&hctx
->ctx_map
, off
,
1006 dispatch_rq_from_ctx
, &data
);
1011 static inline unsigned int queued_to_index(unsigned int queued
)
1016 return min(BLK_MQ_MAX_DISPATCH_ORDER
- 1, ilog2(queued
) + 1);
1019 bool blk_mq_get_driver_tag(struct request
*rq
)
1021 struct blk_mq_alloc_data data
= {
1023 .hctx
= rq
->mq_hctx
,
1024 .flags
= BLK_MQ_REQ_NOWAIT
,
1025 .cmd_flags
= rq
->cmd_flags
,
1032 if (blk_mq_tag_is_reserved(data
.hctx
->sched_tags
, rq
->internal_tag
))
1033 data
.flags
|= BLK_MQ_REQ_RESERVED
;
1035 shared
= blk_mq_tag_busy(data
.hctx
);
1036 rq
->tag
= blk_mq_get_tag(&data
);
1039 rq
->rq_flags
|= RQF_MQ_INFLIGHT
;
1040 atomic_inc(&data
.hctx
->nr_active
);
1042 data
.hctx
->tags
->rqs
[rq
->tag
] = rq
;
1046 return rq
->tag
!= -1;
1049 static int blk_mq_dispatch_wake(wait_queue_entry_t
*wait
, unsigned mode
,
1050 int flags
, void *key
)
1052 struct blk_mq_hw_ctx
*hctx
;
1054 hctx
= container_of(wait
, struct blk_mq_hw_ctx
, dispatch_wait
);
1056 spin_lock(&hctx
->dispatch_wait_lock
);
1057 list_del_init(&wait
->entry
);
1058 spin_unlock(&hctx
->dispatch_wait_lock
);
1060 blk_mq_run_hw_queue(hctx
, true);
1065 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1066 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1067 * restart. For both cases, take care to check the condition again after
1068 * marking us as waiting.
1070 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx
*hctx
,
1073 struct wait_queue_head
*wq
;
1074 wait_queue_entry_t
*wait
;
1077 if (!(hctx
->flags
& BLK_MQ_F_TAG_SHARED
)) {
1078 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
1079 set_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
1082 * It's possible that a tag was freed in the window between the
1083 * allocation failure and adding the hardware queue to the wait
1086 * Don't clear RESTART here, someone else could have set it.
1087 * At most this will cost an extra queue run.
1089 return blk_mq_get_driver_tag(rq
);
1092 wait
= &hctx
->dispatch_wait
;
1093 if (!list_empty_careful(&wait
->entry
))
1096 wq
= &bt_wait_ptr(&hctx
->tags
->bitmap_tags
, hctx
)->wait
;
1098 spin_lock_irq(&wq
->lock
);
1099 spin_lock(&hctx
->dispatch_wait_lock
);
1100 if (!list_empty(&wait
->entry
)) {
1101 spin_unlock(&hctx
->dispatch_wait_lock
);
1102 spin_unlock_irq(&wq
->lock
);
1106 wait
->flags
&= ~WQ_FLAG_EXCLUSIVE
;
1107 __add_wait_queue(wq
, wait
);
1110 * It's possible that a tag was freed in the window between the
1111 * allocation failure and adding the hardware queue to the wait
1114 ret
= blk_mq_get_driver_tag(rq
);
1116 spin_unlock(&hctx
->dispatch_wait_lock
);
1117 spin_unlock_irq(&wq
->lock
);
1122 * We got a tag, remove ourselves from the wait queue to ensure
1123 * someone else gets the wakeup.
1125 list_del_init(&wait
->entry
);
1126 spin_unlock(&hctx
->dispatch_wait_lock
);
1127 spin_unlock_irq(&wq
->lock
);
1132 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1133 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1135 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1136 * - EWMA is one simple way to compute running average value
1137 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1138 * - take 4 as factor for avoiding to get too small(0) result, and this
1139 * factor doesn't matter because EWMA decreases exponentially
1141 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx
*hctx
, bool busy
)
1145 if (hctx
->queue
->elevator
)
1148 ewma
= hctx
->dispatch_busy
;
1153 ewma
*= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
- 1;
1155 ewma
+= 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR
;
1156 ewma
/= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT
;
1158 hctx
->dispatch_busy
= ewma
;
1161 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1164 * Returns true if we did some work AND can potentially do more.
1166 bool blk_mq_dispatch_rq_list(struct request_queue
*q
, struct list_head
*list
,
1169 struct blk_mq_hw_ctx
*hctx
;
1170 struct request
*rq
, *nxt
;
1171 bool no_tag
= false;
1173 blk_status_t ret
= BLK_STS_OK
;
1175 if (list_empty(list
))
1178 WARN_ON(!list_is_singular(list
) && got_budget
);
1181 * Now process all the entries, sending them to the driver.
1183 errors
= queued
= 0;
1185 struct blk_mq_queue_data bd
;
1187 rq
= list_first_entry(list
, struct request
, queuelist
);
1190 if (!got_budget
&& !blk_mq_get_dispatch_budget(hctx
))
1193 if (!blk_mq_get_driver_tag(rq
)) {
1195 * The initial allocation attempt failed, so we need to
1196 * rerun the hardware queue when a tag is freed. The
1197 * waitqueue takes care of that. If the queue is run
1198 * before we add this entry back on the dispatch list,
1199 * we'll re-run it below.
1201 if (!blk_mq_mark_tag_wait(hctx
, rq
)) {
1202 blk_mq_put_dispatch_budget(hctx
);
1204 * For non-shared tags, the RESTART check
1207 if (hctx
->flags
& BLK_MQ_F_TAG_SHARED
)
1213 list_del_init(&rq
->queuelist
);
1218 * Flag last if we have no more requests, or if we have more
1219 * but can't assign a driver tag to it.
1221 if (list_empty(list
))
1224 nxt
= list_first_entry(list
, struct request
, queuelist
);
1225 bd
.last
= !blk_mq_get_driver_tag(nxt
);
1228 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1229 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
) {
1231 * If an I/O scheduler has been configured and we got a
1232 * driver tag for the next request already, free it
1235 if (!list_empty(list
)) {
1236 nxt
= list_first_entry(list
, struct request
, queuelist
);
1237 blk_mq_put_driver_tag(nxt
);
1239 list_add(&rq
->queuelist
, list
);
1240 __blk_mq_requeue_request(rq
);
1244 if (unlikely(ret
!= BLK_STS_OK
)) {
1246 blk_mq_end_request(rq
, BLK_STS_IOERR
);
1251 } while (!list_empty(list
));
1253 hctx
->dispatched
[queued_to_index(queued
)]++;
1256 * Any items that need requeuing? Stuff them into hctx->dispatch,
1257 * that is where we will continue on next queue run.
1259 if (!list_empty(list
)) {
1262 spin_lock(&hctx
->lock
);
1263 list_splice_init(list
, &hctx
->dispatch
);
1264 spin_unlock(&hctx
->lock
);
1267 * If SCHED_RESTART was set by the caller of this function and
1268 * it is no longer set that means that it was cleared by another
1269 * thread and hence that a queue rerun is needed.
1271 * If 'no_tag' is set, that means that we failed getting
1272 * a driver tag with an I/O scheduler attached. If our dispatch
1273 * waitqueue is no longer active, ensure that we run the queue
1274 * AFTER adding our entries back to the list.
1276 * If no I/O scheduler has been configured it is possible that
1277 * the hardware queue got stopped and restarted before requests
1278 * were pushed back onto the dispatch list. Rerun the queue to
1279 * avoid starvation. Notes:
1280 * - blk_mq_run_hw_queue() checks whether or not a queue has
1281 * been stopped before rerunning a queue.
1282 * - Some but not all block drivers stop a queue before
1283 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1286 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1287 * bit is set, run queue after a delay to avoid IO stalls
1288 * that could otherwise occur if the queue is idle.
1290 needs_restart
= blk_mq_sched_needs_restart(hctx
);
1291 if (!needs_restart
||
1292 (no_tag
&& list_empty_careful(&hctx
->dispatch_wait
.entry
)))
1293 blk_mq_run_hw_queue(hctx
, true);
1294 else if (needs_restart
&& (ret
== BLK_STS_RESOURCE
))
1295 blk_mq_delay_run_hw_queue(hctx
, BLK_MQ_RESOURCE_DELAY
);
1297 blk_mq_update_dispatch_busy(hctx
, true);
1300 blk_mq_update_dispatch_busy(hctx
, false);
1303 * If the host/device is unable to accept more work, inform the
1306 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1309 return (queued
+ errors
) != 0;
1312 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1317 * We should be running this queue from one of the CPUs that
1320 * There are at least two related races now between setting
1321 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1322 * __blk_mq_run_hw_queue():
1324 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1325 * but later it becomes online, then this warning is harmless
1328 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1329 * but later it becomes offline, then the warning can't be
1330 * triggered, and we depend on blk-mq timeout handler to
1331 * handle dispatched requests to this hctx
1333 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx
->cpumask
) &&
1334 cpu_online(hctx
->next_cpu
)) {
1335 printk(KERN_WARNING
"run queue from wrong CPU %d, hctx %s\n",
1336 raw_smp_processor_id(),
1337 cpumask_empty(hctx
->cpumask
) ? "inactive": "active");
1342 * We can't run the queue inline with ints disabled. Ensure that
1343 * we catch bad users of this early.
1345 WARN_ON_ONCE(in_interrupt());
1347 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1349 hctx_lock(hctx
, &srcu_idx
);
1350 blk_mq_sched_dispatch_requests(hctx
);
1351 hctx_unlock(hctx
, srcu_idx
);
1354 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx
*hctx
)
1356 int cpu
= cpumask_first_and(hctx
->cpumask
, cpu_online_mask
);
1358 if (cpu
>= nr_cpu_ids
)
1359 cpu
= cpumask_first(hctx
->cpumask
);
1364 * It'd be great if the workqueue API had a way to pass
1365 * in a mask and had some smarts for more clever placement.
1366 * For now we just round-robin here, switching for every
1367 * BLK_MQ_CPU_WORK_BATCH queued items.
1369 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx
*hctx
)
1372 int next_cpu
= hctx
->next_cpu
;
1374 if (hctx
->queue
->nr_hw_queues
== 1)
1375 return WORK_CPU_UNBOUND
;
1377 if (--hctx
->next_cpu_batch
<= 0) {
1379 next_cpu
= cpumask_next_and(next_cpu
, hctx
->cpumask
,
1381 if (next_cpu
>= nr_cpu_ids
)
1382 next_cpu
= blk_mq_first_mapped_cpu(hctx
);
1383 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
1387 * Do unbound schedule if we can't find a online CPU for this hctx,
1388 * and it should only happen in the path of handling CPU DEAD.
1390 if (!cpu_online(next_cpu
)) {
1397 * Make sure to re-select CPU next time once after CPUs
1398 * in hctx->cpumask become online again.
1400 hctx
->next_cpu
= next_cpu
;
1401 hctx
->next_cpu_batch
= 1;
1402 return WORK_CPU_UNBOUND
;
1405 hctx
->next_cpu
= next_cpu
;
1409 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
,
1410 unsigned long msecs
)
1412 if (unlikely(blk_mq_hctx_stopped(hctx
)))
1415 if (!async
&& !(hctx
->flags
& BLK_MQ_F_BLOCKING
)) {
1416 int cpu
= get_cpu();
1417 if (cpumask_test_cpu(cpu
, hctx
->cpumask
)) {
1418 __blk_mq_run_hw_queue(hctx
);
1426 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx
), &hctx
->run_work
,
1427 msecs_to_jiffies(msecs
));
1430 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, unsigned long msecs
)
1432 __blk_mq_delay_run_hw_queue(hctx
, true, msecs
);
1434 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue
);
1436 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1442 * When queue is quiesced, we may be switching io scheduler, or
1443 * updating nr_hw_queues, or other things, and we can't run queue
1444 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1446 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1449 hctx_lock(hctx
, &srcu_idx
);
1450 need_run
= !blk_queue_quiesced(hctx
->queue
) &&
1451 blk_mq_hctx_has_pending(hctx
);
1452 hctx_unlock(hctx
, srcu_idx
);
1455 __blk_mq_delay_run_hw_queue(hctx
, async
, 0);
1461 EXPORT_SYMBOL(blk_mq_run_hw_queue
);
1463 void blk_mq_run_hw_queues(struct request_queue
*q
, bool async
)
1465 struct blk_mq_hw_ctx
*hctx
;
1468 queue_for_each_hw_ctx(q
, hctx
, i
) {
1469 if (blk_mq_hctx_stopped(hctx
))
1472 blk_mq_run_hw_queue(hctx
, async
);
1475 EXPORT_SYMBOL(blk_mq_run_hw_queues
);
1478 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1479 * @q: request queue.
1481 * The caller is responsible for serializing this function against
1482 * blk_mq_{start,stop}_hw_queue().
1484 bool blk_mq_queue_stopped(struct request_queue
*q
)
1486 struct blk_mq_hw_ctx
*hctx
;
1489 queue_for_each_hw_ctx(q
, hctx
, i
)
1490 if (blk_mq_hctx_stopped(hctx
))
1495 EXPORT_SYMBOL(blk_mq_queue_stopped
);
1498 * This function is often used for pausing .queue_rq() by driver when
1499 * there isn't enough resource or some conditions aren't satisfied, and
1500 * BLK_STS_RESOURCE is usually returned.
1502 * We do not guarantee that dispatch can be drained or blocked
1503 * after blk_mq_stop_hw_queue() returns. Please use
1504 * blk_mq_quiesce_queue() for that requirement.
1506 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1508 cancel_delayed_work(&hctx
->run_work
);
1510 set_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1512 EXPORT_SYMBOL(blk_mq_stop_hw_queue
);
1515 * This function is often used for pausing .queue_rq() by driver when
1516 * there isn't enough resource or some conditions aren't satisfied, and
1517 * BLK_STS_RESOURCE is usually returned.
1519 * We do not guarantee that dispatch can be drained or blocked
1520 * after blk_mq_stop_hw_queues() returns. Please use
1521 * blk_mq_quiesce_queue() for that requirement.
1523 void blk_mq_stop_hw_queues(struct request_queue
*q
)
1525 struct blk_mq_hw_ctx
*hctx
;
1528 queue_for_each_hw_ctx(q
, hctx
, i
)
1529 blk_mq_stop_hw_queue(hctx
);
1531 EXPORT_SYMBOL(blk_mq_stop_hw_queues
);
1533 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx
*hctx
)
1535 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1537 blk_mq_run_hw_queue(hctx
, false);
1539 EXPORT_SYMBOL(blk_mq_start_hw_queue
);
1541 void blk_mq_start_hw_queues(struct request_queue
*q
)
1543 struct blk_mq_hw_ctx
*hctx
;
1546 queue_for_each_hw_ctx(q
, hctx
, i
)
1547 blk_mq_start_hw_queue(hctx
);
1549 EXPORT_SYMBOL(blk_mq_start_hw_queues
);
1551 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx
*hctx
, bool async
)
1553 if (!blk_mq_hctx_stopped(hctx
))
1556 clear_bit(BLK_MQ_S_STOPPED
, &hctx
->state
);
1557 blk_mq_run_hw_queue(hctx
, async
);
1559 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue
);
1561 void blk_mq_start_stopped_hw_queues(struct request_queue
*q
, bool async
)
1563 struct blk_mq_hw_ctx
*hctx
;
1566 queue_for_each_hw_ctx(q
, hctx
, i
)
1567 blk_mq_start_stopped_hw_queue(hctx
, async
);
1569 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues
);
1571 static void blk_mq_run_work_fn(struct work_struct
*work
)
1573 struct blk_mq_hw_ctx
*hctx
;
1575 hctx
= container_of(work
, struct blk_mq_hw_ctx
, run_work
.work
);
1578 * If we are stopped, don't run the queue.
1580 if (test_bit(BLK_MQ_S_STOPPED
, &hctx
->state
))
1583 __blk_mq_run_hw_queue(hctx
);
1586 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx
*hctx
,
1590 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1592 lockdep_assert_held(&ctx
->lock
);
1594 trace_block_rq_insert(hctx
->queue
, rq
);
1597 list_add(&rq
->queuelist
, &ctx
->rq_list
);
1599 list_add_tail(&rq
->queuelist
, &ctx
->rq_list
);
1602 void __blk_mq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
1605 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
1607 lockdep_assert_held(&ctx
->lock
);
1609 __blk_mq_insert_req_list(hctx
, rq
, at_head
);
1610 blk_mq_hctx_mark_pending(hctx
, ctx
);
1614 * Should only be used carefully, when the caller knows we want to
1615 * bypass a potential IO scheduler on the target device.
1617 void blk_mq_request_bypass_insert(struct request
*rq
, bool run_queue
)
1619 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1621 spin_lock(&hctx
->lock
);
1622 list_add_tail(&rq
->queuelist
, &hctx
->dispatch
);
1623 spin_unlock(&hctx
->lock
);
1626 blk_mq_run_hw_queue(hctx
, false);
1629 void blk_mq_insert_requests(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_ctx
*ctx
,
1630 struct list_head
*list
)
1636 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1639 list_for_each_entry(rq
, list
, queuelist
) {
1640 BUG_ON(rq
->mq_ctx
!= ctx
);
1641 trace_block_rq_insert(hctx
->queue
, rq
);
1644 spin_lock(&ctx
->lock
);
1645 list_splice_tail_init(list
, &ctx
->rq_list
);
1646 blk_mq_hctx_mark_pending(hctx
, ctx
);
1647 spin_unlock(&ctx
->lock
);
1650 static int plug_rq_cmp(void *priv
, struct list_head
*a
, struct list_head
*b
)
1652 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
1653 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
1655 if (rqa
->mq_ctx
< rqb
->mq_ctx
)
1657 else if (rqa
->mq_ctx
> rqb
->mq_ctx
)
1659 else if (rqa
->mq_hctx
< rqb
->mq_hctx
)
1661 else if (rqa
->mq_hctx
> rqb
->mq_hctx
)
1664 return blk_rq_pos(rqa
) > blk_rq_pos(rqb
);
1667 void blk_mq_flush_plug_list(struct blk_plug
*plug
, bool from_schedule
)
1669 struct blk_mq_hw_ctx
*this_hctx
;
1670 struct blk_mq_ctx
*this_ctx
;
1671 struct request_queue
*this_q
;
1677 list_splice_init(&plug
->mq_list
, &list
);
1680 list_sort(NULL
, &list
, plug_rq_cmp
);
1687 while (!list_empty(&list
)) {
1688 rq
= list_entry_rq(list
.next
);
1689 list_del_init(&rq
->queuelist
);
1691 if (rq
->mq_hctx
!= this_hctx
|| rq
->mq_ctx
!= this_ctx
) {
1693 trace_block_unplug(this_q
, depth
, !from_schedule
);
1694 blk_mq_sched_insert_requests(this_hctx
, this_ctx
,
1700 this_ctx
= rq
->mq_ctx
;
1701 this_hctx
= rq
->mq_hctx
;
1706 list_add_tail(&rq
->queuelist
, &rq_list
);
1710 * If 'this_hctx' is set, we know we have entries to complete
1711 * on 'rq_list'. Do those.
1714 trace_block_unplug(this_q
, depth
, !from_schedule
);
1715 blk_mq_sched_insert_requests(this_hctx
, this_ctx
, &rq_list
,
1720 static void blk_mq_bio_to_request(struct request
*rq
, struct bio
*bio
)
1722 blk_init_request_from_bio(rq
, bio
);
1724 blk_account_io_start(rq
, true);
1727 static blk_qc_t
request_to_qc_t(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
)
1730 return blk_tag_to_qc_t(rq
->tag
, hctx
->queue_num
, false);
1732 return blk_tag_to_qc_t(rq
->internal_tag
, hctx
->queue_num
, true);
1735 static blk_status_t
__blk_mq_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1739 struct request_queue
*q
= rq
->q
;
1740 struct blk_mq_queue_data bd
= {
1744 blk_qc_t new_cookie
;
1747 new_cookie
= request_to_qc_t(hctx
, rq
);
1750 * For OK queue, we are done. For error, caller may kill it.
1751 * Any other error (busy), just add it to our list as we
1752 * previously would have done.
1754 ret
= q
->mq_ops
->queue_rq(hctx
, &bd
);
1757 blk_mq_update_dispatch_busy(hctx
, false);
1758 *cookie
= new_cookie
;
1760 case BLK_STS_RESOURCE
:
1761 case BLK_STS_DEV_RESOURCE
:
1762 blk_mq_update_dispatch_busy(hctx
, true);
1763 __blk_mq_requeue_request(rq
);
1766 blk_mq_update_dispatch_busy(hctx
, false);
1767 *cookie
= BLK_QC_T_NONE
;
1774 static blk_status_t
__blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1779 struct request_queue
*q
= rq
->q
;
1780 bool run_queue
= true;
1783 * RCU or SRCU read lock is needed before checking quiesced flag.
1785 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1786 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1787 * and avoid driver to try to dispatch again.
1789 if (blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)) {
1791 bypass_insert
= false;
1795 if (q
->elevator
&& !bypass_insert
)
1798 if (!blk_mq_get_dispatch_budget(hctx
))
1801 if (!blk_mq_get_driver_tag(rq
)) {
1802 blk_mq_put_dispatch_budget(hctx
);
1806 return __blk_mq_issue_directly(hctx
, rq
, cookie
);
1809 return BLK_STS_RESOURCE
;
1811 blk_mq_sched_insert_request(rq
, false, run_queue
, false);
1815 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx
*hctx
,
1816 struct request
*rq
, blk_qc_t
*cookie
)
1821 might_sleep_if(hctx
->flags
& BLK_MQ_F_BLOCKING
);
1823 hctx_lock(hctx
, &srcu_idx
);
1825 ret
= __blk_mq_try_issue_directly(hctx
, rq
, cookie
, false);
1826 if (ret
== BLK_STS_RESOURCE
|| ret
== BLK_STS_DEV_RESOURCE
)
1827 blk_mq_sched_insert_request(rq
, false, true, false);
1828 else if (ret
!= BLK_STS_OK
)
1829 blk_mq_end_request(rq
, ret
);
1831 hctx_unlock(hctx
, srcu_idx
);
1834 blk_status_t
blk_mq_request_issue_directly(struct request
*rq
)
1838 blk_qc_t unused_cookie
;
1839 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
1841 hctx_lock(hctx
, &srcu_idx
);
1842 ret
= __blk_mq_try_issue_directly(hctx
, rq
, &unused_cookie
, true);
1843 hctx_unlock(hctx
, srcu_idx
);
1848 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx
*hctx
,
1849 struct list_head
*list
)
1851 while (!list_empty(list
)) {
1853 struct request
*rq
= list_first_entry(list
, struct request
,
1856 list_del_init(&rq
->queuelist
);
1857 ret
= blk_mq_request_issue_directly(rq
);
1858 if (ret
!= BLK_STS_OK
) {
1859 if (ret
== BLK_STS_RESOURCE
||
1860 ret
== BLK_STS_DEV_RESOURCE
) {
1861 list_add(&rq
->queuelist
, list
);
1864 blk_mq_end_request(rq
, ret
);
1869 static blk_qc_t
blk_mq_make_request(struct request_queue
*q
, struct bio
*bio
)
1871 const int is_sync
= op_is_sync(bio
->bi_opf
);
1872 const int is_flush_fua
= op_is_flush(bio
->bi_opf
);
1873 struct blk_mq_alloc_data data
= { .flags
= 0, .cmd_flags
= bio
->bi_opf
};
1875 struct blk_plug
*plug
;
1876 struct request
*same_queue_rq
= NULL
;
1879 blk_queue_bounce(q
, &bio
);
1881 blk_queue_split(q
, &bio
);
1883 if (!bio_integrity_prep(bio
))
1884 return BLK_QC_T_NONE
;
1886 if (!is_flush_fua
&& !blk_queue_nomerges(q
) &&
1887 blk_attempt_plug_merge(q
, bio
, &same_queue_rq
))
1888 return BLK_QC_T_NONE
;
1890 if (blk_mq_sched_bio_merge(q
, bio
))
1891 return BLK_QC_T_NONE
;
1893 rq_qos_throttle(q
, bio
);
1895 rq
= blk_mq_get_request(q
, bio
, &data
);
1896 if (unlikely(!rq
)) {
1897 rq_qos_cleanup(q
, bio
);
1898 if (bio
->bi_opf
& REQ_NOWAIT
)
1899 bio_wouldblock_error(bio
);
1900 return BLK_QC_T_NONE
;
1903 trace_block_getrq(q
, bio
, bio
->bi_opf
);
1905 rq_qos_track(q
, rq
, bio
);
1907 cookie
= request_to_qc_t(data
.hctx
, rq
);
1909 plug
= current
->plug
;
1910 if (unlikely(is_flush_fua
)) {
1911 blk_mq_put_ctx(data
.ctx
);
1912 blk_mq_bio_to_request(rq
, bio
);
1914 /* bypass scheduler for flush rq */
1915 blk_insert_flush(rq
);
1916 blk_mq_run_hw_queue(data
.hctx
, true);
1917 } else if (plug
&& q
->nr_hw_queues
== 1) {
1918 unsigned int request_count
= plug
->rq_count
;
1919 struct request
*last
= NULL
;
1921 blk_mq_put_ctx(data
.ctx
);
1922 blk_mq_bio_to_request(rq
, bio
);
1925 trace_block_plug(q
);
1927 last
= list_entry_rq(plug
->mq_list
.prev
);
1929 if (request_count
>= BLK_MAX_REQUEST_COUNT
|| (last
&&
1930 blk_rq_bytes(last
) >= BLK_PLUG_FLUSH_SIZE
)) {
1931 blk_flush_plug_list(plug
, false);
1932 trace_block_plug(q
);
1935 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1937 } else if (plug
&& !blk_queue_nomerges(q
)) {
1938 blk_mq_bio_to_request(rq
, bio
);
1941 * We do limited plugging. If the bio can be merged, do that.
1942 * Otherwise the existing request in the plug list will be
1943 * issued. So the plug list will have one request at most
1944 * The plug list might get flushed before this. If that happens,
1945 * the plug list is empty, and same_queue_rq is invalid.
1947 if (list_empty(&plug
->mq_list
))
1948 same_queue_rq
= NULL
;
1949 if (same_queue_rq
) {
1950 list_del_init(&same_queue_rq
->queuelist
);
1953 list_add_tail(&rq
->queuelist
, &plug
->mq_list
);
1956 blk_mq_put_ctx(data
.ctx
);
1958 if (same_queue_rq
) {
1959 data
.hctx
= same_queue_rq
->mq_hctx
;
1960 blk_mq_try_issue_directly(data
.hctx
, same_queue_rq
,
1963 } else if ((q
->nr_hw_queues
> 1 && is_sync
) || (!q
->elevator
&&
1964 !data
.hctx
->dispatch_busy
)) {
1965 blk_mq_put_ctx(data
.ctx
);
1966 blk_mq_bio_to_request(rq
, bio
);
1967 blk_mq_try_issue_directly(data
.hctx
, rq
, &cookie
);
1969 blk_mq_put_ctx(data
.ctx
);
1970 blk_mq_bio_to_request(rq
, bio
);
1971 blk_mq_sched_insert_request(rq
, false, true, true);
1977 void blk_mq_free_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
1978 unsigned int hctx_idx
)
1982 if (tags
->rqs
&& set
->ops
->exit_request
) {
1985 for (i
= 0; i
< tags
->nr_tags
; i
++) {
1986 struct request
*rq
= tags
->static_rqs
[i
];
1990 set
->ops
->exit_request(set
, rq
, hctx_idx
);
1991 tags
->static_rqs
[i
] = NULL
;
1995 while (!list_empty(&tags
->page_list
)) {
1996 page
= list_first_entry(&tags
->page_list
, struct page
, lru
);
1997 list_del_init(&page
->lru
);
1999 * Remove kmemleak object previously allocated in
2000 * blk_mq_init_rq_map().
2002 kmemleak_free(page_address(page
));
2003 __free_pages(page
, page
->private);
2007 void blk_mq_free_rq_map(struct blk_mq_tags
*tags
)
2011 kfree(tags
->static_rqs
);
2012 tags
->static_rqs
= NULL
;
2014 blk_mq_free_tags(tags
);
2017 struct blk_mq_tags
*blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
,
2018 unsigned int hctx_idx
,
2019 unsigned int nr_tags
,
2020 unsigned int reserved_tags
)
2022 struct blk_mq_tags
*tags
;
2025 node
= blk_mq_hw_queue_to_node(&set
->map
[0], hctx_idx
);
2026 if (node
== NUMA_NO_NODE
)
2027 node
= set
->numa_node
;
2029 tags
= blk_mq_init_tags(nr_tags
, reserved_tags
, node
,
2030 BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
));
2034 tags
->rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2035 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2038 blk_mq_free_tags(tags
);
2042 tags
->static_rqs
= kcalloc_node(nr_tags
, sizeof(struct request
*),
2043 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2045 if (!tags
->static_rqs
) {
2047 blk_mq_free_tags(tags
);
2054 static size_t order_to_size(unsigned int order
)
2056 return (size_t)PAGE_SIZE
<< order
;
2059 static int blk_mq_init_request(struct blk_mq_tag_set
*set
, struct request
*rq
,
2060 unsigned int hctx_idx
, int node
)
2064 if (set
->ops
->init_request
) {
2065 ret
= set
->ops
->init_request(set
, rq
, hctx_idx
, node
);
2070 WRITE_ONCE(rq
->state
, MQ_RQ_IDLE
);
2074 int blk_mq_alloc_rqs(struct blk_mq_tag_set
*set
, struct blk_mq_tags
*tags
,
2075 unsigned int hctx_idx
, unsigned int depth
)
2077 unsigned int i
, j
, entries_per_page
, max_order
= 4;
2078 size_t rq_size
, left
;
2081 node
= blk_mq_hw_queue_to_node(&set
->map
[0], hctx_idx
);
2082 if (node
== NUMA_NO_NODE
)
2083 node
= set
->numa_node
;
2085 INIT_LIST_HEAD(&tags
->page_list
);
2088 * rq_size is the size of the request plus driver payload, rounded
2089 * to the cacheline size
2091 rq_size
= round_up(sizeof(struct request
) + set
->cmd_size
,
2093 left
= rq_size
* depth
;
2095 for (i
= 0; i
< depth
; ) {
2096 int this_order
= max_order
;
2101 while (this_order
&& left
< order_to_size(this_order
- 1))
2105 page
= alloc_pages_node(node
,
2106 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
| __GFP_ZERO
,
2112 if (order_to_size(this_order
) < rq_size
)
2119 page
->private = this_order
;
2120 list_add_tail(&page
->lru
, &tags
->page_list
);
2122 p
= page_address(page
);
2124 * Allow kmemleak to scan these pages as they contain pointers
2125 * to additional allocations like via ops->init_request().
2127 kmemleak_alloc(p
, order_to_size(this_order
), 1, GFP_NOIO
);
2128 entries_per_page
= order_to_size(this_order
) / rq_size
;
2129 to_do
= min(entries_per_page
, depth
- i
);
2130 left
-= to_do
* rq_size
;
2131 for (j
= 0; j
< to_do
; j
++) {
2132 struct request
*rq
= p
;
2134 tags
->static_rqs
[i
] = rq
;
2135 if (blk_mq_init_request(set
, rq
, hctx_idx
, node
)) {
2136 tags
->static_rqs
[i
] = NULL
;
2147 blk_mq_free_rqs(set
, tags
, hctx_idx
);
2152 * 'cpu' is going away. splice any existing rq_list entries from this
2153 * software queue to the hw queue dispatch list, and ensure that it
2156 static int blk_mq_hctx_notify_dead(unsigned int cpu
, struct hlist_node
*node
)
2158 struct blk_mq_hw_ctx
*hctx
;
2159 struct blk_mq_ctx
*ctx
;
2162 hctx
= hlist_entry_safe(node
, struct blk_mq_hw_ctx
, cpuhp_dead
);
2163 ctx
= __blk_mq_get_ctx(hctx
->queue
, cpu
);
2165 spin_lock(&ctx
->lock
);
2166 if (!list_empty(&ctx
->rq_list
)) {
2167 list_splice_init(&ctx
->rq_list
, &tmp
);
2168 blk_mq_hctx_clear_pending(hctx
, ctx
);
2170 spin_unlock(&ctx
->lock
);
2172 if (list_empty(&tmp
))
2175 spin_lock(&hctx
->lock
);
2176 list_splice_tail_init(&tmp
, &hctx
->dispatch
);
2177 spin_unlock(&hctx
->lock
);
2179 blk_mq_run_hw_queue(hctx
, true);
2183 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx
*hctx
)
2185 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD
,
2189 /* hctx->ctxs will be freed in queue's release handler */
2190 static void blk_mq_exit_hctx(struct request_queue
*q
,
2191 struct blk_mq_tag_set
*set
,
2192 struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
2194 if (blk_mq_hw_queue_mapped(hctx
))
2195 blk_mq_tag_idle(hctx
);
2197 if (set
->ops
->exit_request
)
2198 set
->ops
->exit_request(set
, hctx
->fq
->flush_rq
, hctx_idx
);
2200 if (set
->ops
->exit_hctx
)
2201 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2203 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2204 cleanup_srcu_struct(hctx
->srcu
);
2206 blk_mq_remove_cpuhp(hctx
);
2207 blk_free_flush_queue(hctx
->fq
);
2208 sbitmap_free(&hctx
->ctx_map
);
2211 static void blk_mq_exit_hw_queues(struct request_queue
*q
,
2212 struct blk_mq_tag_set
*set
, int nr_queue
)
2214 struct blk_mq_hw_ctx
*hctx
;
2217 queue_for_each_hw_ctx(q
, hctx
, i
) {
2220 blk_mq_debugfs_unregister_hctx(hctx
);
2221 blk_mq_exit_hctx(q
, set
, hctx
, i
);
2225 static int blk_mq_init_hctx(struct request_queue
*q
,
2226 struct blk_mq_tag_set
*set
,
2227 struct blk_mq_hw_ctx
*hctx
, unsigned hctx_idx
)
2231 node
= hctx
->numa_node
;
2232 if (node
== NUMA_NO_NODE
)
2233 node
= hctx
->numa_node
= set
->numa_node
;
2235 INIT_DELAYED_WORK(&hctx
->run_work
, blk_mq_run_work_fn
);
2236 spin_lock_init(&hctx
->lock
);
2237 INIT_LIST_HEAD(&hctx
->dispatch
);
2239 hctx
->flags
= set
->flags
& ~BLK_MQ_F_TAG_SHARED
;
2241 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD
, &hctx
->cpuhp_dead
);
2243 hctx
->tags
= set
->tags
[hctx_idx
];
2246 * Allocate space for all possible cpus to avoid allocation at
2249 hctx
->ctxs
= kmalloc_array_node(nr_cpu_ids
, sizeof(void *),
2250 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
, node
);
2252 goto unregister_cpu_notifier
;
2254 if (sbitmap_init_node(&hctx
->ctx_map
, nr_cpu_ids
, ilog2(8),
2255 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
, node
))
2260 spin_lock_init(&hctx
->dispatch_wait_lock
);
2261 init_waitqueue_func_entry(&hctx
->dispatch_wait
, blk_mq_dispatch_wake
);
2262 INIT_LIST_HEAD(&hctx
->dispatch_wait
.entry
);
2264 if (set
->ops
->init_hctx
&&
2265 set
->ops
->init_hctx(hctx
, set
->driver_data
, hctx_idx
))
2268 hctx
->fq
= blk_alloc_flush_queue(q
, hctx
->numa_node
, set
->cmd_size
,
2269 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
2273 if (blk_mq_init_request(set
, hctx
->fq
->flush_rq
, hctx_idx
, node
))
2276 if (hctx
->flags
& BLK_MQ_F_BLOCKING
)
2277 init_srcu_struct(hctx
->srcu
);
2284 if (set
->ops
->exit_hctx
)
2285 set
->ops
->exit_hctx(hctx
, hctx_idx
);
2287 sbitmap_free(&hctx
->ctx_map
);
2290 unregister_cpu_notifier
:
2291 blk_mq_remove_cpuhp(hctx
);
2295 static void blk_mq_init_cpu_queues(struct request_queue
*q
,
2296 unsigned int nr_hw_queues
)
2298 struct blk_mq_tag_set
*set
= q
->tag_set
;
2301 for_each_possible_cpu(i
) {
2302 struct blk_mq_ctx
*__ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2303 struct blk_mq_hw_ctx
*hctx
;
2306 spin_lock_init(&__ctx
->lock
);
2307 INIT_LIST_HEAD(&__ctx
->rq_list
);
2311 * Set local node, IFF we have more than one hw queue. If
2312 * not, we remain on the home node of the device
2314 for (j
= 0; j
< set
->nr_maps
; j
++) {
2315 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2316 if (nr_hw_queues
> 1 && hctx
->numa_node
== NUMA_NO_NODE
)
2317 hctx
->numa_node
= local_memory_node(cpu_to_node(i
));
2322 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set
*set
, int hctx_idx
)
2326 set
->tags
[hctx_idx
] = blk_mq_alloc_rq_map(set
, hctx_idx
,
2327 set
->queue_depth
, set
->reserved_tags
);
2328 if (!set
->tags
[hctx_idx
])
2331 ret
= blk_mq_alloc_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
,
2336 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2337 set
->tags
[hctx_idx
] = NULL
;
2341 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set
*set
,
2342 unsigned int hctx_idx
)
2344 if (set
->tags
[hctx_idx
]) {
2345 blk_mq_free_rqs(set
, set
->tags
[hctx_idx
], hctx_idx
);
2346 blk_mq_free_rq_map(set
->tags
[hctx_idx
]);
2347 set
->tags
[hctx_idx
] = NULL
;
2351 static void blk_mq_map_swqueue(struct request_queue
*q
)
2353 unsigned int i
, j
, hctx_idx
;
2354 struct blk_mq_hw_ctx
*hctx
;
2355 struct blk_mq_ctx
*ctx
;
2356 struct blk_mq_tag_set
*set
= q
->tag_set
;
2359 * Avoid others reading imcomplete hctx->cpumask through sysfs
2361 mutex_lock(&q
->sysfs_lock
);
2363 queue_for_each_hw_ctx(q
, hctx
, i
) {
2364 cpumask_clear(hctx
->cpumask
);
2366 hctx
->dispatch_from
= NULL
;
2370 * Map software to hardware queues.
2372 * If the cpu isn't present, the cpu is mapped to first hctx.
2374 for_each_possible_cpu(i
) {
2375 hctx_idx
= set
->map
[0].mq_map
[i
];
2376 /* unmapped hw queue can be remapped after CPU topo changed */
2377 if (!set
->tags
[hctx_idx
] &&
2378 !__blk_mq_alloc_rq_map(set
, hctx_idx
)) {
2380 * If tags initialization fail for some hctx,
2381 * that hctx won't be brought online. In this
2382 * case, remap the current ctx to hctx[0] which
2383 * is guaranteed to always have tags allocated
2385 set
->map
[0].mq_map
[i
] = 0;
2388 ctx
= per_cpu_ptr(q
->queue_ctx
, i
);
2389 for (j
= 0; j
< set
->nr_maps
; j
++) {
2390 hctx
= blk_mq_map_queue_type(q
, j
, i
);
2393 * If the CPU is already set in the mask, then we've
2394 * mapped this one already. This can happen if
2395 * devices share queues across queue maps.
2397 if (cpumask_test_cpu(i
, hctx
->cpumask
))
2400 cpumask_set_cpu(i
, hctx
->cpumask
);
2402 ctx
->index_hw
[hctx
->type
] = hctx
->nr_ctx
;
2403 hctx
->ctxs
[hctx
->nr_ctx
++] = ctx
;
2406 * If the nr_ctx type overflows, we have exceeded the
2407 * amount of sw queues we can support.
2409 BUG_ON(!hctx
->nr_ctx
);
2413 mutex_unlock(&q
->sysfs_lock
);
2415 queue_for_each_hw_ctx(q
, hctx
, i
) {
2417 * If no software queues are mapped to this hardware queue,
2418 * disable it and free the request entries.
2420 if (!hctx
->nr_ctx
) {
2421 /* Never unmap queue 0. We need it as a
2422 * fallback in case of a new remap fails
2425 if (i
&& set
->tags
[i
])
2426 blk_mq_free_map_and_requests(set
, i
);
2432 hctx
->tags
= set
->tags
[i
];
2433 WARN_ON(!hctx
->tags
);
2436 * Set the map size to the number of mapped software queues.
2437 * This is more accurate and more efficient than looping
2438 * over all possibly mapped software queues.
2440 sbitmap_resize(&hctx
->ctx_map
, hctx
->nr_ctx
);
2443 * Initialize batch roundrobin counts
2445 hctx
->next_cpu
= blk_mq_first_mapped_cpu(hctx
);
2446 hctx
->next_cpu_batch
= BLK_MQ_CPU_WORK_BATCH
;
2451 * Caller needs to ensure that we're either frozen/quiesced, or that
2452 * the queue isn't live yet.
2454 static void queue_set_hctx_shared(struct request_queue
*q
, bool shared
)
2456 struct blk_mq_hw_ctx
*hctx
;
2459 queue_for_each_hw_ctx(q
, hctx
, i
) {
2461 hctx
->flags
|= BLK_MQ_F_TAG_SHARED
;
2463 hctx
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2467 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set
*set
,
2470 struct request_queue
*q
;
2472 lockdep_assert_held(&set
->tag_list_lock
);
2474 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
2475 blk_mq_freeze_queue(q
);
2476 queue_set_hctx_shared(q
, shared
);
2477 blk_mq_unfreeze_queue(q
);
2481 static void blk_mq_del_queue_tag_set(struct request_queue
*q
)
2483 struct blk_mq_tag_set
*set
= q
->tag_set
;
2485 mutex_lock(&set
->tag_list_lock
);
2486 list_del_rcu(&q
->tag_set_list
);
2487 if (list_is_singular(&set
->tag_list
)) {
2488 /* just transitioned to unshared */
2489 set
->flags
&= ~BLK_MQ_F_TAG_SHARED
;
2490 /* update existing queue */
2491 blk_mq_update_tag_set_depth(set
, false);
2493 mutex_unlock(&set
->tag_list_lock
);
2494 INIT_LIST_HEAD(&q
->tag_set_list
);
2497 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set
*set
,
2498 struct request_queue
*q
)
2500 mutex_lock(&set
->tag_list_lock
);
2503 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2505 if (!list_empty(&set
->tag_list
) &&
2506 !(set
->flags
& BLK_MQ_F_TAG_SHARED
)) {
2507 set
->flags
|= BLK_MQ_F_TAG_SHARED
;
2508 /* update existing queue */
2509 blk_mq_update_tag_set_depth(set
, true);
2511 if (set
->flags
& BLK_MQ_F_TAG_SHARED
)
2512 queue_set_hctx_shared(q
, true);
2513 list_add_tail_rcu(&q
->tag_set_list
, &set
->tag_list
);
2515 mutex_unlock(&set
->tag_list_lock
);
2518 /* All allocations will be freed in release handler of q->mq_kobj */
2519 static int blk_mq_alloc_ctxs(struct request_queue
*q
)
2521 struct blk_mq_ctxs
*ctxs
;
2524 ctxs
= kzalloc(sizeof(*ctxs
), GFP_KERNEL
);
2528 ctxs
->queue_ctx
= alloc_percpu(struct blk_mq_ctx
);
2529 if (!ctxs
->queue_ctx
)
2532 for_each_possible_cpu(cpu
) {
2533 struct blk_mq_ctx
*ctx
= per_cpu_ptr(ctxs
->queue_ctx
, cpu
);
2537 q
->mq_kobj
= &ctxs
->kobj
;
2538 q
->queue_ctx
= ctxs
->queue_ctx
;
2547 * It is the actual release handler for mq, but we do it from
2548 * request queue's release handler for avoiding use-after-free
2549 * and headache because q->mq_kobj shouldn't have been introduced,
2550 * but we can't group ctx/kctx kobj without it.
2552 void blk_mq_release(struct request_queue
*q
)
2554 struct blk_mq_hw_ctx
*hctx
;
2557 /* hctx kobj stays in hctx */
2558 queue_for_each_hw_ctx(q
, hctx
, i
) {
2561 kobject_put(&hctx
->kobj
);
2564 kfree(q
->queue_hw_ctx
);
2567 * release .mq_kobj and sw queue's kobject now because
2568 * both share lifetime with request queue.
2570 blk_mq_sysfs_deinit(q
);
2573 struct request_queue
*blk_mq_init_queue(struct blk_mq_tag_set
*set
)
2575 struct request_queue
*uninit_q
, *q
;
2577 uninit_q
= blk_alloc_queue_node(GFP_KERNEL
, set
->numa_node
);
2579 return ERR_PTR(-ENOMEM
);
2581 q
= blk_mq_init_allocated_queue(set
, uninit_q
);
2583 blk_cleanup_queue(uninit_q
);
2587 EXPORT_SYMBOL(blk_mq_init_queue
);
2590 * Helper for setting up a queue with mq ops, given queue depth, and
2591 * the passed in mq ops flags.
2593 struct request_queue
*blk_mq_init_sq_queue(struct blk_mq_tag_set
*set
,
2594 const struct blk_mq_ops
*ops
,
2595 unsigned int queue_depth
,
2596 unsigned int set_flags
)
2598 struct request_queue
*q
;
2601 memset(set
, 0, sizeof(*set
));
2603 set
->nr_hw_queues
= 1;
2605 set
->queue_depth
= queue_depth
;
2606 set
->numa_node
= NUMA_NO_NODE
;
2607 set
->flags
= set_flags
;
2609 ret
= blk_mq_alloc_tag_set(set
);
2611 return ERR_PTR(ret
);
2613 q
= blk_mq_init_queue(set
);
2615 blk_mq_free_tag_set(set
);
2621 EXPORT_SYMBOL(blk_mq_init_sq_queue
);
2623 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set
*tag_set
)
2625 int hw_ctx_size
= sizeof(struct blk_mq_hw_ctx
);
2627 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx
, srcu
),
2628 __alignof__(struct blk_mq_hw_ctx
)) !=
2629 sizeof(struct blk_mq_hw_ctx
));
2631 if (tag_set
->flags
& BLK_MQ_F_BLOCKING
)
2632 hw_ctx_size
+= sizeof(struct srcu_struct
);
2637 static struct blk_mq_hw_ctx
*blk_mq_alloc_and_init_hctx(
2638 struct blk_mq_tag_set
*set
, struct request_queue
*q
,
2639 int hctx_idx
, int node
)
2641 struct blk_mq_hw_ctx
*hctx
;
2643 hctx
= kzalloc_node(blk_mq_hw_ctx_size(set
),
2644 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2649 if (!zalloc_cpumask_var_node(&hctx
->cpumask
,
2650 GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
,
2656 atomic_set(&hctx
->nr_active
, 0);
2657 hctx
->numa_node
= node
;
2658 hctx
->queue_num
= hctx_idx
;
2660 if (blk_mq_init_hctx(q
, set
, hctx
, hctx_idx
)) {
2661 free_cpumask_var(hctx
->cpumask
);
2665 blk_mq_hctx_kobj_init(hctx
);
2670 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set
*set
,
2671 struct request_queue
*q
)
2674 struct blk_mq_hw_ctx
**hctxs
= q
->queue_hw_ctx
;
2676 /* protect against switching io scheduler */
2677 mutex_lock(&q
->sysfs_lock
);
2678 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
2680 struct blk_mq_hw_ctx
*hctx
;
2682 node
= blk_mq_hw_queue_to_node(&set
->map
[0], i
);
2684 * If the hw queue has been mapped to another numa node,
2685 * we need to realloc the hctx. If allocation fails, fallback
2686 * to use the previous one.
2688 if (hctxs
[i
] && (hctxs
[i
]->numa_node
== node
))
2691 hctx
= blk_mq_alloc_and_init_hctx(set
, q
, i
, node
);
2694 blk_mq_exit_hctx(q
, set
, hctxs
[i
], i
);
2695 kobject_put(&hctxs
[i
]->kobj
);
2700 pr_warn("Allocate new hctx on node %d fails,\
2701 fallback to previous one on node %d\n",
2702 node
, hctxs
[i
]->numa_node
);
2708 * Increasing nr_hw_queues fails. Free the newly allocated
2709 * hctxs and keep the previous q->nr_hw_queues.
2711 if (i
!= set
->nr_hw_queues
) {
2712 j
= q
->nr_hw_queues
;
2716 end
= q
->nr_hw_queues
;
2717 q
->nr_hw_queues
= set
->nr_hw_queues
;
2720 for (; j
< end
; j
++) {
2721 struct blk_mq_hw_ctx
*hctx
= hctxs
[j
];
2725 blk_mq_free_map_and_requests(set
, j
);
2726 blk_mq_exit_hctx(q
, set
, hctx
, j
);
2727 kobject_put(&hctx
->kobj
);
2732 mutex_unlock(&q
->sysfs_lock
);
2736 * Maximum number of hardware queues we support. For single sets, we'll never
2737 * have more than the CPUs (software queues). For multiple sets, the tag_set
2738 * user may have set ->nr_hw_queues larger.
2740 static unsigned int nr_hw_queues(struct blk_mq_tag_set
*set
)
2742 if (set
->nr_maps
== 1)
2745 return max(set
->nr_hw_queues
, nr_cpu_ids
);
2748 struct request_queue
*blk_mq_init_allocated_queue(struct blk_mq_tag_set
*set
,
2749 struct request_queue
*q
)
2751 /* mark the queue as mq asap */
2752 q
->mq_ops
= set
->ops
;
2754 q
->poll_cb
= blk_stat_alloc_callback(blk_mq_poll_stats_fn
,
2755 blk_mq_poll_stats_bkt
,
2756 BLK_MQ_POLL_STATS_BKTS
, q
);
2760 if (blk_mq_alloc_ctxs(q
))
2763 /* init q->mq_kobj and sw queues' kobjects */
2764 blk_mq_sysfs_init(q
);
2766 q
->nr_queues
= nr_hw_queues(set
);
2767 q
->queue_hw_ctx
= kcalloc_node(q
->nr_queues
, sizeof(*(q
->queue_hw_ctx
)),
2768 GFP_KERNEL
, set
->numa_node
);
2769 if (!q
->queue_hw_ctx
)
2772 blk_mq_realloc_hw_ctxs(set
, q
);
2773 if (!q
->nr_hw_queues
)
2776 INIT_WORK(&q
->timeout_work
, blk_mq_timeout_work
);
2777 blk_queue_rq_timeout(q
, set
->timeout
? set
->timeout
: 30 * HZ
);
2781 q
->queue_flags
|= QUEUE_FLAG_MQ_DEFAULT
;
2783 if (!(set
->flags
& BLK_MQ_F_SG_MERGE
))
2784 blk_queue_flag_set(QUEUE_FLAG_NO_SG_MERGE
, q
);
2786 q
->sg_reserved_size
= INT_MAX
;
2788 INIT_DELAYED_WORK(&q
->requeue_work
, blk_mq_requeue_work
);
2789 INIT_LIST_HEAD(&q
->requeue_list
);
2790 spin_lock_init(&q
->requeue_lock
);
2792 blk_queue_make_request(q
, blk_mq_make_request
);
2793 if (q
->mq_ops
->poll
)
2794 q
->poll_fn
= blk_mq_poll
;
2797 * Do this after blk_queue_make_request() overrides it...
2799 q
->nr_requests
= set
->queue_depth
;
2802 * Default to classic polling
2806 blk_mq_init_cpu_queues(q
, set
->nr_hw_queues
);
2807 blk_mq_add_queue_tag_set(set
, q
);
2808 blk_mq_map_swqueue(q
);
2810 if (!(set
->flags
& BLK_MQ_F_NO_SCHED
)) {
2813 ret
= elevator_init_mq(q
);
2815 return ERR_PTR(ret
);
2821 kfree(q
->queue_hw_ctx
);
2823 blk_mq_sysfs_deinit(q
);
2826 return ERR_PTR(-ENOMEM
);
2828 EXPORT_SYMBOL(blk_mq_init_allocated_queue
);
2830 void blk_mq_free_queue(struct request_queue
*q
)
2832 struct blk_mq_tag_set
*set
= q
->tag_set
;
2834 blk_mq_del_queue_tag_set(q
);
2835 blk_mq_exit_hw_queues(q
, set
, set
->nr_hw_queues
);
2838 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2842 for (i
= 0; i
< set
->nr_hw_queues
; i
++)
2843 if (!__blk_mq_alloc_rq_map(set
, i
))
2850 blk_mq_free_rq_map(set
->tags
[i
]);
2856 * Allocate the request maps associated with this tag_set. Note that this
2857 * may reduce the depth asked for, if memory is tight. set->queue_depth
2858 * will be updated to reflect the allocated depth.
2860 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set
*set
)
2865 depth
= set
->queue_depth
;
2867 err
= __blk_mq_alloc_rq_maps(set
);
2871 set
->queue_depth
>>= 1;
2872 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
) {
2876 } while (set
->queue_depth
);
2878 if (!set
->queue_depth
|| err
) {
2879 pr_err("blk-mq: failed to allocate request map\n");
2883 if (depth
!= set
->queue_depth
)
2884 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2885 depth
, set
->queue_depth
);
2890 static int blk_mq_update_queue_map(struct blk_mq_tag_set
*set
)
2892 if (set
->ops
->map_queues
) {
2896 * transport .map_queues is usually done in the following
2899 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2900 * mask = get_cpu_mask(queue)
2901 * for_each_cpu(cpu, mask)
2902 * set->map[x].mq_map[cpu] = queue;
2905 * When we need to remap, the table has to be cleared for
2906 * killing stale mapping since one CPU may not be mapped
2909 for (i
= 0; i
< set
->nr_maps
; i
++)
2910 blk_mq_clear_mq_map(&set
->map
[i
]);
2912 return set
->ops
->map_queues(set
);
2914 BUG_ON(set
->nr_maps
> 1);
2915 return blk_mq_map_queues(&set
->map
[0]);
2920 * Alloc a tag set to be associated with one or more request queues.
2921 * May fail with EINVAL for various error conditions. May adjust the
2922 * requested depth down, if it's too large. In that case, the set
2923 * value will be stored in set->queue_depth.
2925 int blk_mq_alloc_tag_set(struct blk_mq_tag_set
*set
)
2929 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH
> 1 << BLK_MQ_UNIQUE_TAG_BITS
);
2931 if (!set
->nr_hw_queues
)
2933 if (!set
->queue_depth
)
2935 if (set
->queue_depth
< set
->reserved_tags
+ BLK_MQ_TAG_MIN
)
2938 if (!set
->ops
->queue_rq
)
2941 if (!set
->ops
->get_budget
^ !set
->ops
->put_budget
)
2944 if (set
->queue_depth
> BLK_MQ_MAX_DEPTH
) {
2945 pr_info("blk-mq: reduced tag depth to %u\n",
2947 set
->queue_depth
= BLK_MQ_MAX_DEPTH
;
2952 else if (set
->nr_maps
> HCTX_MAX_TYPES
)
2956 * If a crashdump is active, then we are potentially in a very
2957 * memory constrained environment. Limit us to 1 queue and
2958 * 64 tags to prevent using too much memory.
2960 if (is_kdump_kernel()) {
2961 set
->nr_hw_queues
= 1;
2962 set
->queue_depth
= min(64U, set
->queue_depth
);
2965 * There is no use for more h/w queues than cpus if we just have
2968 if (set
->nr_maps
== 1 && set
->nr_hw_queues
> nr_cpu_ids
)
2969 set
->nr_hw_queues
= nr_cpu_ids
;
2971 set
->tags
= kcalloc_node(nr_hw_queues(set
), sizeof(struct blk_mq_tags
*),
2972 GFP_KERNEL
, set
->numa_node
);
2977 for (i
= 0; i
< set
->nr_maps
; i
++) {
2978 set
->map
[i
].mq_map
= kcalloc_node(nr_cpu_ids
,
2979 sizeof(struct blk_mq_queue_map
),
2980 GFP_KERNEL
, set
->numa_node
);
2981 if (!set
->map
[i
].mq_map
)
2982 goto out_free_mq_map
;
2983 set
->map
[i
].nr_queues
= set
->nr_hw_queues
;
2986 ret
= blk_mq_update_queue_map(set
);
2988 goto out_free_mq_map
;
2990 ret
= blk_mq_alloc_rq_maps(set
);
2992 goto out_free_mq_map
;
2994 mutex_init(&set
->tag_list_lock
);
2995 INIT_LIST_HEAD(&set
->tag_list
);
3000 for (i
= 0; i
< set
->nr_maps
; i
++) {
3001 kfree(set
->map
[i
].mq_map
);
3002 set
->map
[i
].mq_map
= NULL
;
3008 EXPORT_SYMBOL(blk_mq_alloc_tag_set
);
3010 void blk_mq_free_tag_set(struct blk_mq_tag_set
*set
)
3014 for (i
= 0; i
< nr_hw_queues(set
); i
++)
3015 blk_mq_free_map_and_requests(set
, i
);
3017 for (j
= 0; j
< set
->nr_maps
; j
++) {
3018 kfree(set
->map
[j
].mq_map
);
3019 set
->map
[j
].mq_map
= NULL
;
3025 EXPORT_SYMBOL(blk_mq_free_tag_set
);
3027 int blk_mq_update_nr_requests(struct request_queue
*q
, unsigned int nr
)
3029 struct blk_mq_tag_set
*set
= q
->tag_set
;
3030 struct blk_mq_hw_ctx
*hctx
;
3036 blk_mq_freeze_queue(q
);
3037 blk_mq_quiesce_queue(q
);
3040 queue_for_each_hw_ctx(q
, hctx
, i
) {
3044 * If we're using an MQ scheduler, just update the scheduler
3045 * queue depth. This is similar to what the old code would do.
3047 if (!hctx
->sched_tags
) {
3048 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->tags
, nr
,
3051 ret
= blk_mq_tag_update_depth(hctx
, &hctx
->sched_tags
,
3059 q
->nr_requests
= nr
;
3061 blk_mq_unquiesce_queue(q
);
3062 blk_mq_unfreeze_queue(q
);
3068 * request_queue and elevator_type pair.
3069 * It is just used by __blk_mq_update_nr_hw_queues to cache
3070 * the elevator_type associated with a request_queue.
3072 struct blk_mq_qe_pair
{
3073 struct list_head node
;
3074 struct request_queue
*q
;
3075 struct elevator_type
*type
;
3079 * Cache the elevator_type in qe pair list and switch the
3080 * io scheduler to 'none'
3082 static bool blk_mq_elv_switch_none(struct list_head
*head
,
3083 struct request_queue
*q
)
3085 struct blk_mq_qe_pair
*qe
;
3090 qe
= kmalloc(sizeof(*qe
), GFP_NOIO
| __GFP_NOWARN
| __GFP_NORETRY
);
3094 INIT_LIST_HEAD(&qe
->node
);
3096 qe
->type
= q
->elevator
->type
;
3097 list_add(&qe
->node
, head
);
3099 mutex_lock(&q
->sysfs_lock
);
3101 * After elevator_switch_mq, the previous elevator_queue will be
3102 * released by elevator_release. The reference of the io scheduler
3103 * module get by elevator_get will also be put. So we need to get
3104 * a reference of the io scheduler module here to prevent it to be
3107 __module_get(qe
->type
->elevator_owner
);
3108 elevator_switch_mq(q
, NULL
);
3109 mutex_unlock(&q
->sysfs_lock
);
3114 static void blk_mq_elv_switch_back(struct list_head
*head
,
3115 struct request_queue
*q
)
3117 struct blk_mq_qe_pair
*qe
;
3118 struct elevator_type
*t
= NULL
;
3120 list_for_each_entry(qe
, head
, node
)
3129 list_del(&qe
->node
);
3132 mutex_lock(&q
->sysfs_lock
);
3133 elevator_switch_mq(q
, t
);
3134 mutex_unlock(&q
->sysfs_lock
);
3137 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
,
3140 struct request_queue
*q
;
3142 int prev_nr_hw_queues
;
3144 lockdep_assert_held(&set
->tag_list_lock
);
3146 if (set
->nr_maps
== 1 && nr_hw_queues
> nr_cpu_ids
)
3147 nr_hw_queues
= nr_cpu_ids
;
3148 if (nr_hw_queues
< 1 || nr_hw_queues
== set
->nr_hw_queues
)
3151 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3152 blk_mq_freeze_queue(q
);
3154 * Sync with blk_mq_queue_tag_busy_iter.
3158 * Switch IO scheduler to 'none', cleaning up the data associated
3159 * with the previous scheduler. We will switch back once we are done
3160 * updating the new sw to hw queue mappings.
3162 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3163 if (!blk_mq_elv_switch_none(&head
, q
))
3166 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3167 blk_mq_debugfs_unregister_hctxs(q
);
3168 blk_mq_sysfs_unregister(q
);
3171 prev_nr_hw_queues
= set
->nr_hw_queues
;
3172 set
->nr_hw_queues
= nr_hw_queues
;
3173 blk_mq_update_queue_map(set
);
3175 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3176 blk_mq_realloc_hw_ctxs(set
, q
);
3177 if (q
->nr_hw_queues
!= set
->nr_hw_queues
) {
3178 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3179 nr_hw_queues
, prev_nr_hw_queues
);
3180 set
->nr_hw_queues
= prev_nr_hw_queues
;
3181 blk_mq_map_queues(&set
->map
[0]);
3184 blk_mq_map_swqueue(q
);
3187 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
) {
3188 blk_mq_sysfs_register(q
);
3189 blk_mq_debugfs_register_hctxs(q
);
3193 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3194 blk_mq_elv_switch_back(&head
, q
);
3196 list_for_each_entry(q
, &set
->tag_list
, tag_set_list
)
3197 blk_mq_unfreeze_queue(q
);
3200 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set
*set
, int nr_hw_queues
)
3202 mutex_lock(&set
->tag_list_lock
);
3203 __blk_mq_update_nr_hw_queues(set
, nr_hw_queues
);
3204 mutex_unlock(&set
->tag_list_lock
);
3206 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues
);
3208 /* Enable polling stats and return whether they were already enabled. */
3209 static bool blk_poll_stats_enable(struct request_queue
*q
)
3211 if (test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3212 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS
, q
))
3214 blk_stat_add_callback(q
, q
->poll_cb
);
3218 static void blk_mq_poll_stats_start(struct request_queue
*q
)
3221 * We don't arm the callback if polling stats are not enabled or the
3222 * callback is already active.
3224 if (!test_bit(QUEUE_FLAG_POLL_STATS
, &q
->queue_flags
) ||
3225 blk_stat_is_active(q
->poll_cb
))
3228 blk_stat_activate_msecs(q
->poll_cb
, 100);
3231 static void blk_mq_poll_stats_fn(struct blk_stat_callback
*cb
)
3233 struct request_queue
*q
= cb
->data
;
3236 for (bucket
= 0; bucket
< BLK_MQ_POLL_STATS_BKTS
; bucket
++) {
3237 if (cb
->stat
[bucket
].nr_samples
)
3238 q
->poll_stat
[bucket
] = cb
->stat
[bucket
];
3242 static unsigned long blk_mq_poll_nsecs(struct request_queue
*q
,
3243 struct blk_mq_hw_ctx
*hctx
,
3246 unsigned long ret
= 0;
3250 * If stats collection isn't on, don't sleep but turn it on for
3253 if (!blk_poll_stats_enable(q
))
3257 * As an optimistic guess, use half of the mean service time
3258 * for this type of request. We can (and should) make this smarter.
3259 * For instance, if the completion latencies are tight, we can
3260 * get closer than just half the mean. This is especially
3261 * important on devices where the completion latencies are longer
3262 * than ~10 usec. We do use the stats for the relevant IO size
3263 * if available which does lead to better estimates.
3265 bucket
= blk_mq_poll_stats_bkt(rq
);
3269 if (q
->poll_stat
[bucket
].nr_samples
)
3270 ret
= (q
->poll_stat
[bucket
].mean
+ 1) / 2;
3275 static bool blk_mq_poll_hybrid_sleep(struct request_queue
*q
,
3276 struct blk_mq_hw_ctx
*hctx
,
3279 struct hrtimer_sleeper hs
;
3280 enum hrtimer_mode mode
;
3284 if (rq
->rq_flags
& RQF_MQ_POLL_SLEPT
)
3288 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3290 * 0: use half of prev avg
3291 * >0: use this specific value
3293 if (q
->poll_nsec
> 0)
3294 nsecs
= q
->poll_nsec
;
3296 nsecs
= blk_mq_poll_nsecs(q
, hctx
, rq
);
3301 rq
->rq_flags
|= RQF_MQ_POLL_SLEPT
;
3304 * This will be replaced with the stats tracking code, using
3305 * 'avg_completion_time / 2' as the pre-sleep target.
3309 mode
= HRTIMER_MODE_REL
;
3310 hrtimer_init_on_stack(&hs
.timer
, CLOCK_MONOTONIC
, mode
);
3311 hrtimer_set_expires(&hs
.timer
, kt
);
3313 hrtimer_init_sleeper(&hs
, current
);
3315 if (blk_mq_rq_state(rq
) == MQ_RQ_COMPLETE
)
3317 set_current_state(TASK_UNINTERRUPTIBLE
);
3318 hrtimer_start_expires(&hs
.timer
, mode
);
3321 hrtimer_cancel(&hs
.timer
);
3322 mode
= HRTIMER_MODE_ABS
;
3323 } while (hs
.task
&& !signal_pending(current
));
3325 __set_current_state(TASK_RUNNING
);
3326 destroy_hrtimer_on_stack(&hs
.timer
);
3330 static bool blk_mq_poll_hybrid(struct request_queue
*q
,
3331 struct blk_mq_hw_ctx
*hctx
, blk_qc_t cookie
)
3335 if (q
->poll_nsec
== -1)
3338 if (!blk_qc_t_is_internal(cookie
))
3339 rq
= blk_mq_tag_to_rq(hctx
->tags
, blk_qc_t_to_tag(cookie
));
3341 rq
= blk_mq_tag_to_rq(hctx
->sched_tags
, blk_qc_t_to_tag(cookie
));
3343 * With scheduling, if the request has completed, we'll
3344 * get a NULL return here, as we clear the sched tag when
3345 * that happens. The request still remains valid, like always,
3346 * so we should be safe with just the NULL check.
3352 return blk_mq_poll_hybrid_sleep(q
, hctx
, rq
);
3355 static int blk_mq_poll(struct request_queue
*q
, blk_qc_t cookie
, bool spin
)
3357 struct blk_mq_hw_ctx
*hctx
;
3360 if (!test_bit(QUEUE_FLAG_POLL
, &q
->queue_flags
))
3363 hctx
= q
->queue_hw_ctx
[blk_qc_t_to_queue_num(cookie
)];
3366 * If we sleep, have the caller restart the poll loop to reset
3367 * the state. Like for the other success return cases, the
3368 * caller is responsible for checking if the IO completed. If
3369 * the IO isn't complete, we'll get called again and will go
3370 * straight to the busy poll loop.
3372 if (blk_mq_poll_hybrid(q
, hctx
, cookie
))
3375 hctx
->poll_considered
++;
3377 state
= current
->state
;
3381 hctx
->poll_invoked
++;
3383 ret
= q
->mq_ops
->poll(hctx
);
3385 hctx
->poll_success
++;
3386 __set_current_state(TASK_RUNNING
);
3390 if (signal_pending_state(state
, current
))
3391 __set_current_state(TASK_RUNNING
);
3393 if (current
->state
== TASK_RUNNING
)
3395 if (ret
< 0 || !spin
)
3398 } while (!need_resched());
3400 __set_current_state(TASK_RUNNING
);
3404 unsigned int blk_mq_rq_cpu(struct request
*rq
)
3406 return rq
->mq_ctx
->cpu
;
3408 EXPORT_SYMBOL(blk_mq_rq_cpu
);
3410 static int __init
blk_mq_init(void)
3412 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD
, "block/mq:dead", NULL
,
3413 blk_mq_hctx_notify_dead
);
3416 subsys_initcall(blk_mq_init
);