2 * Block device elevator/IO-scheduler.
4 * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
6 * 30042000 Jens Axboe <axboe@kernel.dk> :
8 * Split the elevator a bit so that it is possible to choose a different
9 * one or even write a new "plug in". There are three pieces:
10 * - elevator_fn, inserts a new request in the queue list
11 * - elevator_merge_fn, decides whether a new buffer can be merged with
13 * - elevator_dequeue_fn, called when a request is taken off the active list
15 * 20082000 Dave Jones <davej@suse.de> :
16 * Removed tests for max-bomb-segments, which was breaking elvtune
17 * when run without -bN
20 * - Rework again to work with bio instead of buffer_heads
21 * - loose bi_dev comparisons, partition handling is right now
22 * - completely modularize elevator setup and teardown
25 #include <linux/kernel.h>
27 #include <linux/blkdev.h>
28 #include <linux/elevator.h>
29 #include <linux/bio.h>
30 #include <linux/module.h>
31 #include <linux/slab.h>
32 #include <linux/init.h>
33 #include <linux/compiler.h>
34 #include <linux/blktrace_api.h>
35 #include <linux/hash.h>
36 #include <linux/uaccess.h>
37 #include <linux/pm_runtime.h>
38 #include <linux/blk-cgroup.h>
40 #include <trace/events/block.h>
43 #include "blk-mq-sched.h"
45 static DEFINE_SPINLOCK(elv_list_lock
);
46 static LIST_HEAD(elv_list
);
51 #define rq_hash_key(rq) (blk_rq_pos(rq) + blk_rq_sectors(rq))
54 * Query io scheduler to see if the current process issuing bio may be
57 static int elv_iosched_allow_bio_merge(struct request
*rq
, struct bio
*bio
)
59 struct request_queue
*q
= rq
->q
;
60 struct elevator_queue
*e
= q
->elevator
;
62 if (e
->uses_mq
&& e
->type
->ops
.mq
.allow_merge
)
63 return e
->type
->ops
.mq
.allow_merge(q
, rq
, bio
);
64 else if (!e
->uses_mq
&& e
->type
->ops
.sq
.elevator_allow_bio_merge_fn
)
65 return e
->type
->ops
.sq
.elevator_allow_bio_merge_fn(q
, rq
, bio
);
71 * can we safely merge with this request?
73 bool elv_bio_merge_ok(struct request
*rq
, struct bio
*bio
)
75 if (!blk_rq_merge_ok(rq
, bio
))
78 if (!elv_iosched_allow_bio_merge(rq
, bio
))
83 EXPORT_SYMBOL(elv_bio_merge_ok
);
85 static struct elevator_type
*elevator_find(const char *name
)
87 struct elevator_type
*e
;
89 list_for_each_entry(e
, &elv_list
, list
) {
90 if (!strcmp(e
->elevator_name
, name
))
97 static void elevator_put(struct elevator_type
*e
)
99 module_put(e
->elevator_owner
);
102 static struct elevator_type
*elevator_get(const char *name
, bool try_loading
)
104 struct elevator_type
*e
;
106 spin_lock(&elv_list_lock
);
108 e
= elevator_find(name
);
109 if (!e
&& try_loading
) {
110 spin_unlock(&elv_list_lock
);
111 request_module("%s-iosched", name
);
112 spin_lock(&elv_list_lock
);
113 e
= elevator_find(name
);
116 if (e
&& !try_module_get(e
->elevator_owner
))
119 spin_unlock(&elv_list_lock
);
124 static char chosen_elevator
[ELV_NAME_MAX
];
126 static int __init
elevator_setup(char *str
)
129 * Be backwards-compatible with previous kernels, so users
130 * won't get the wrong elevator.
132 strncpy(chosen_elevator
, str
, sizeof(chosen_elevator
) - 1);
136 __setup("elevator=", elevator_setup
);
138 /* called during boot to load the elevator chosen by the elevator param */
139 void __init
load_default_elevator_module(void)
141 struct elevator_type
*e
;
143 if (!chosen_elevator
[0])
146 spin_lock(&elv_list_lock
);
147 e
= elevator_find(chosen_elevator
);
148 spin_unlock(&elv_list_lock
);
151 request_module("%s-iosched", chosen_elevator
);
154 static struct kobj_type elv_ktype
;
156 struct elevator_queue
*elevator_alloc(struct request_queue
*q
,
157 struct elevator_type
*e
)
159 struct elevator_queue
*eq
;
161 eq
= kzalloc_node(sizeof(*eq
), GFP_KERNEL
, q
->node
);
166 kobject_init(&eq
->kobj
, &elv_ktype
);
167 mutex_init(&eq
->sysfs_lock
);
169 eq
->uses_mq
= e
->uses_mq
;
173 EXPORT_SYMBOL(elevator_alloc
);
175 static void elevator_release(struct kobject
*kobj
)
177 struct elevator_queue
*e
;
179 e
= container_of(kobj
, struct elevator_queue
, kobj
);
180 elevator_put(e
->type
);
184 int elevator_init(struct request_queue
*q
, char *name
)
186 struct elevator_type
*e
= NULL
;
190 * q->sysfs_lock must be held to provide mutual exclusion between
191 * elevator_switch() and here.
193 lockdep_assert_held(&q
->sysfs_lock
);
195 if (unlikely(q
->elevator
))
198 INIT_LIST_HEAD(&q
->queue_head
);
199 q
->last_merge
= NULL
;
201 q
->boundary_rq
= NULL
;
204 e
= elevator_get(name
, true);
210 * Use the default elevator specified by config boot param for
211 * non-mq devices, or by config option. Don't try to load modules
212 * as we could be running off async and request_module() isn't
213 * allowed from async.
215 if (!e
&& !q
->mq_ops
&& *chosen_elevator
) {
216 e
= elevator_get(chosen_elevator
, false);
218 printk(KERN_ERR
"I/O scheduler %s not found\n",
224 * For blk-mq devices, we default to using mq-deadline,
225 * if available, for single queue devices. If deadline
226 * isn't available OR we have multiple queues, default
230 if (q
->nr_hw_queues
== 1)
231 e
= elevator_get("mq-deadline", false);
235 e
= elevator_get(CONFIG_DEFAULT_IOSCHED
, false);
239 "Default I/O scheduler not found. " \
241 e
= elevator_get("noop", false);
246 err
= blk_mq_init_sched(q
, e
);
248 err
= e
->ops
.sq
.elevator_init_fn(q
, e
);
253 EXPORT_SYMBOL(elevator_init
);
255 void elevator_exit(struct request_queue
*q
, struct elevator_queue
*e
)
257 mutex_lock(&e
->sysfs_lock
);
258 if (e
->uses_mq
&& e
->type
->ops
.mq
.exit_sched
)
259 blk_mq_exit_sched(q
, e
);
260 else if (!e
->uses_mq
&& e
->type
->ops
.sq
.elevator_exit_fn
)
261 e
->type
->ops
.sq
.elevator_exit_fn(e
);
262 mutex_unlock(&e
->sysfs_lock
);
264 kobject_put(&e
->kobj
);
266 EXPORT_SYMBOL(elevator_exit
);
268 static inline void __elv_rqhash_del(struct request
*rq
)
271 rq
->rq_flags
&= ~RQF_HASHED
;
274 void elv_rqhash_del(struct request_queue
*q
, struct request
*rq
)
277 __elv_rqhash_del(rq
);
279 EXPORT_SYMBOL_GPL(elv_rqhash_del
);
281 void elv_rqhash_add(struct request_queue
*q
, struct request
*rq
)
283 struct elevator_queue
*e
= q
->elevator
;
285 BUG_ON(ELV_ON_HASH(rq
));
286 hash_add(e
->hash
, &rq
->hash
, rq_hash_key(rq
));
287 rq
->rq_flags
|= RQF_HASHED
;
289 EXPORT_SYMBOL_GPL(elv_rqhash_add
);
291 void elv_rqhash_reposition(struct request_queue
*q
, struct request
*rq
)
293 __elv_rqhash_del(rq
);
294 elv_rqhash_add(q
, rq
);
297 struct request
*elv_rqhash_find(struct request_queue
*q
, sector_t offset
)
299 struct elevator_queue
*e
= q
->elevator
;
300 struct hlist_node
*next
;
303 hash_for_each_possible_safe(e
->hash
, rq
, next
, hash
, offset
) {
304 BUG_ON(!ELV_ON_HASH(rq
));
306 if (unlikely(!rq_mergeable(rq
))) {
307 __elv_rqhash_del(rq
);
311 if (rq_hash_key(rq
) == offset
)
319 * RB-tree support functions for inserting/lookup/removal of requests
320 * in a sorted RB tree.
322 void elv_rb_add(struct rb_root
*root
, struct request
*rq
)
324 struct rb_node
**p
= &root
->rb_node
;
325 struct rb_node
*parent
= NULL
;
326 struct request
*__rq
;
330 __rq
= rb_entry(parent
, struct request
, rb_node
);
332 if (blk_rq_pos(rq
) < blk_rq_pos(__rq
))
334 else if (blk_rq_pos(rq
) >= blk_rq_pos(__rq
))
338 rb_link_node(&rq
->rb_node
, parent
, p
);
339 rb_insert_color(&rq
->rb_node
, root
);
341 EXPORT_SYMBOL(elv_rb_add
);
343 void elv_rb_del(struct rb_root
*root
, struct request
*rq
)
345 BUG_ON(RB_EMPTY_NODE(&rq
->rb_node
));
346 rb_erase(&rq
->rb_node
, root
);
347 RB_CLEAR_NODE(&rq
->rb_node
);
349 EXPORT_SYMBOL(elv_rb_del
);
351 struct request
*elv_rb_find(struct rb_root
*root
, sector_t sector
)
353 struct rb_node
*n
= root
->rb_node
;
357 rq
= rb_entry(n
, struct request
, rb_node
);
359 if (sector
< blk_rq_pos(rq
))
361 else if (sector
> blk_rq_pos(rq
))
369 EXPORT_SYMBOL(elv_rb_find
);
372 * Insert rq into dispatch queue of q. Queue lock must be held on
373 * entry. rq is sort instead into the dispatch queue. To be used by
374 * specific elevators.
376 void elv_dispatch_sort(struct request_queue
*q
, struct request
*rq
)
379 struct list_head
*entry
;
381 if (q
->last_merge
== rq
)
382 q
->last_merge
= NULL
;
384 elv_rqhash_del(q
, rq
);
388 boundary
= q
->end_sector
;
389 list_for_each_prev(entry
, &q
->queue_head
) {
390 struct request
*pos
= list_entry_rq(entry
);
392 if (req_op(rq
) != req_op(pos
))
394 if (rq_data_dir(rq
) != rq_data_dir(pos
))
396 if (pos
->rq_flags
& (RQF_STARTED
| RQF_SOFTBARRIER
))
398 if (blk_rq_pos(rq
) >= boundary
) {
399 if (blk_rq_pos(pos
) < boundary
)
402 if (blk_rq_pos(pos
) >= boundary
)
405 if (blk_rq_pos(rq
) >= blk_rq_pos(pos
))
409 list_add(&rq
->queuelist
, entry
);
411 EXPORT_SYMBOL(elv_dispatch_sort
);
414 * Insert rq into dispatch queue of q. Queue lock must be held on
415 * entry. rq is added to the back of the dispatch queue. To be used by
416 * specific elevators.
418 void elv_dispatch_add_tail(struct request_queue
*q
, struct request
*rq
)
420 if (q
->last_merge
== rq
)
421 q
->last_merge
= NULL
;
423 elv_rqhash_del(q
, rq
);
427 q
->end_sector
= rq_end_sector(rq
);
429 list_add_tail(&rq
->queuelist
, &q
->queue_head
);
431 EXPORT_SYMBOL(elv_dispatch_add_tail
);
433 enum elv_merge
elv_merge(struct request_queue
*q
, struct request
**req
,
436 struct elevator_queue
*e
= q
->elevator
;
437 struct request
*__rq
;
441 * nomerges: No merges at all attempted
442 * noxmerges: Only simple one-hit cache try
443 * merges: All merge tries attempted
445 if (blk_queue_nomerges(q
) || !bio_mergeable(bio
))
446 return ELEVATOR_NO_MERGE
;
449 * First try one-hit cache.
451 if (q
->last_merge
&& elv_bio_merge_ok(q
->last_merge
, bio
)) {
452 enum elv_merge ret
= blk_try_merge(q
->last_merge
, bio
);
454 if (ret
!= ELEVATOR_NO_MERGE
) {
455 *req
= q
->last_merge
;
460 if (blk_queue_noxmerges(q
))
461 return ELEVATOR_NO_MERGE
;
464 * See if our hash lookup can find a potential backmerge.
466 __rq
= elv_rqhash_find(q
, bio
->bi_iter
.bi_sector
);
467 if (__rq
&& elv_bio_merge_ok(__rq
, bio
)) {
469 return ELEVATOR_BACK_MERGE
;
472 if (e
->uses_mq
&& e
->type
->ops
.mq
.request_merge
)
473 return e
->type
->ops
.mq
.request_merge(q
, req
, bio
);
474 else if (!e
->uses_mq
&& e
->type
->ops
.sq
.elevator_merge_fn
)
475 return e
->type
->ops
.sq
.elevator_merge_fn(q
, req
, bio
);
477 return ELEVATOR_NO_MERGE
;
481 * Attempt to do an insertion back merge. Only check for the case where
482 * we can append 'rq' to an existing request, so we can throw 'rq' away
485 * Returns true if we merged, false otherwise
487 bool elv_attempt_insert_merge(struct request_queue
*q
, struct request
*rq
)
489 struct request
*__rq
;
492 if (blk_queue_nomerges(q
))
496 * First try one-hit cache.
498 if (q
->last_merge
&& blk_attempt_req_merge(q
, q
->last_merge
, rq
))
501 if (blk_queue_noxmerges(q
))
506 * See if our hash lookup can find a potential backmerge.
509 __rq
= elv_rqhash_find(q
, blk_rq_pos(rq
));
510 if (!__rq
|| !blk_attempt_req_merge(q
, __rq
, rq
))
513 /* The merged request could be merged with others, try again */
521 void elv_merged_request(struct request_queue
*q
, struct request
*rq
,
524 struct elevator_queue
*e
= q
->elevator
;
526 if (e
->uses_mq
&& e
->type
->ops
.mq
.request_merged
)
527 e
->type
->ops
.mq
.request_merged(q
, rq
, type
);
528 else if (!e
->uses_mq
&& e
->type
->ops
.sq
.elevator_merged_fn
)
529 e
->type
->ops
.sq
.elevator_merged_fn(q
, rq
, type
);
531 if (type
== ELEVATOR_BACK_MERGE
)
532 elv_rqhash_reposition(q
, rq
);
537 void elv_merge_requests(struct request_queue
*q
, struct request
*rq
,
538 struct request
*next
)
540 struct elevator_queue
*e
= q
->elevator
;
541 bool next_sorted
= false;
543 if (e
->uses_mq
&& e
->type
->ops
.mq
.requests_merged
)
544 e
->type
->ops
.mq
.requests_merged(q
, rq
, next
);
545 else if (e
->type
->ops
.sq
.elevator_merge_req_fn
) {
546 next_sorted
= (__force
bool)(next
->rq_flags
& RQF_SORTED
);
548 e
->type
->ops
.sq
.elevator_merge_req_fn(q
, rq
, next
);
551 elv_rqhash_reposition(q
, rq
);
554 elv_rqhash_del(q
, next
);
561 void elv_bio_merged(struct request_queue
*q
, struct request
*rq
,
564 struct elevator_queue
*e
= q
->elevator
;
566 if (WARN_ON_ONCE(e
->uses_mq
))
569 if (e
->type
->ops
.sq
.elevator_bio_merged_fn
)
570 e
->type
->ops
.sq
.elevator_bio_merged_fn(q
, rq
, bio
);
574 static void blk_pm_requeue_request(struct request
*rq
)
576 if (rq
->q
->dev
&& !(rq
->rq_flags
& RQF_PM
))
580 static void blk_pm_add_request(struct request_queue
*q
, struct request
*rq
)
582 if (q
->dev
&& !(rq
->rq_flags
& RQF_PM
) && q
->nr_pending
++ == 0 &&
583 (q
->rpm_status
== RPM_SUSPENDED
|| q
->rpm_status
== RPM_SUSPENDING
))
584 pm_request_resume(q
->dev
);
587 static inline void blk_pm_requeue_request(struct request
*rq
) {}
588 static inline void blk_pm_add_request(struct request_queue
*q
,
594 void elv_requeue_request(struct request_queue
*q
, struct request
*rq
)
597 * it already went through dequeue, we need to decrement the
598 * in_flight count again
600 if (blk_account_rq(rq
)) {
601 q
->in_flight
[rq_is_sync(rq
)]--;
602 if (rq
->rq_flags
& RQF_SORTED
)
603 elv_deactivate_rq(q
, rq
);
606 rq
->rq_flags
&= ~RQF_STARTED
;
608 blk_pm_requeue_request(rq
);
610 __elv_add_request(q
, rq
, ELEVATOR_INSERT_REQUEUE
);
613 void elv_drain_elevator(struct request_queue
*q
)
615 struct elevator_queue
*e
= q
->elevator
;
618 if (WARN_ON_ONCE(e
->uses_mq
))
621 lockdep_assert_held(q
->queue_lock
);
623 while (e
->type
->ops
.sq
.elevator_dispatch_fn(q
, 1))
625 if (q
->nr_sorted
&& printed
++ < 10) {
626 printk(KERN_ERR
"%s: forced dispatching is broken "
627 "(nr_sorted=%u), please report this\n",
628 q
->elevator
->type
->elevator_name
, q
->nr_sorted
);
632 void __elv_add_request(struct request_queue
*q
, struct request
*rq
, int where
)
634 trace_block_rq_insert(q
, rq
);
636 blk_pm_add_request(q
, rq
);
640 if (rq
->rq_flags
& RQF_SOFTBARRIER
) {
641 /* barriers are scheduling boundary, update end_sector */
642 if (!blk_rq_is_passthrough(rq
)) {
643 q
->end_sector
= rq_end_sector(rq
);
646 } else if (!(rq
->rq_flags
& RQF_ELVPRIV
) &&
647 (where
== ELEVATOR_INSERT_SORT
||
648 where
== ELEVATOR_INSERT_SORT_MERGE
))
649 where
= ELEVATOR_INSERT_BACK
;
652 case ELEVATOR_INSERT_REQUEUE
:
653 case ELEVATOR_INSERT_FRONT
:
654 rq
->rq_flags
|= RQF_SOFTBARRIER
;
655 list_add(&rq
->queuelist
, &q
->queue_head
);
658 case ELEVATOR_INSERT_BACK
:
659 rq
->rq_flags
|= RQF_SOFTBARRIER
;
660 elv_drain_elevator(q
);
661 list_add_tail(&rq
->queuelist
, &q
->queue_head
);
663 * We kick the queue here for the following reasons.
664 * - The elevator might have returned NULL previously
665 * to delay requests and returned them now. As the
666 * queue wasn't empty before this request, ll_rw_blk
667 * won't run the queue on return, resulting in hang.
668 * - Usually, back inserted requests won't be merged
669 * with anything. There's no point in delaying queue
675 case ELEVATOR_INSERT_SORT_MERGE
:
677 * If we succeed in merging this request with one in the
678 * queue already, we are done - rq has now been freed,
679 * so no need to do anything further.
681 if (elv_attempt_insert_merge(q
, rq
))
683 case ELEVATOR_INSERT_SORT
:
684 BUG_ON(blk_rq_is_passthrough(rq
));
685 rq
->rq_flags
|= RQF_SORTED
;
687 if (rq_mergeable(rq
)) {
688 elv_rqhash_add(q
, rq
);
694 * Some ioscheds (cfq) run q->request_fn directly, so
695 * rq cannot be accessed after calling
696 * elevator_add_req_fn.
698 q
->elevator
->type
->ops
.sq
.elevator_add_req_fn(q
, rq
);
701 case ELEVATOR_INSERT_FLUSH
:
702 rq
->rq_flags
|= RQF_SOFTBARRIER
;
703 blk_insert_flush(rq
);
706 printk(KERN_ERR
"%s: bad insertion point %d\n",
711 EXPORT_SYMBOL(__elv_add_request
);
713 void elv_add_request(struct request_queue
*q
, struct request
*rq
, int where
)
717 spin_lock_irqsave(q
->queue_lock
, flags
);
718 __elv_add_request(q
, rq
, where
);
719 spin_unlock_irqrestore(q
->queue_lock
, flags
);
721 EXPORT_SYMBOL(elv_add_request
);
723 struct request
*elv_latter_request(struct request_queue
*q
, struct request
*rq
)
725 struct elevator_queue
*e
= q
->elevator
;
727 if (e
->uses_mq
&& e
->type
->ops
.mq
.next_request
)
728 return e
->type
->ops
.mq
.next_request(q
, rq
);
729 else if (!e
->uses_mq
&& e
->type
->ops
.sq
.elevator_latter_req_fn
)
730 return e
->type
->ops
.sq
.elevator_latter_req_fn(q
, rq
);
735 struct request
*elv_former_request(struct request_queue
*q
, struct request
*rq
)
737 struct elevator_queue
*e
= q
->elevator
;
739 if (e
->uses_mq
&& e
->type
->ops
.mq
.former_request
)
740 return e
->type
->ops
.mq
.former_request(q
, rq
);
741 if (!e
->uses_mq
&& e
->type
->ops
.sq
.elevator_former_req_fn
)
742 return e
->type
->ops
.sq
.elevator_former_req_fn(q
, rq
);
746 int elv_set_request(struct request_queue
*q
, struct request
*rq
,
747 struct bio
*bio
, gfp_t gfp_mask
)
749 struct elevator_queue
*e
= q
->elevator
;
751 if (WARN_ON_ONCE(e
->uses_mq
))
754 if (e
->type
->ops
.sq
.elevator_set_req_fn
)
755 return e
->type
->ops
.sq
.elevator_set_req_fn(q
, rq
, bio
, gfp_mask
);
759 void elv_put_request(struct request_queue
*q
, struct request
*rq
)
761 struct elevator_queue
*e
= q
->elevator
;
763 if (WARN_ON_ONCE(e
->uses_mq
))
766 if (e
->type
->ops
.sq
.elevator_put_req_fn
)
767 e
->type
->ops
.sq
.elevator_put_req_fn(rq
);
770 int elv_may_queue(struct request_queue
*q
, unsigned int op
)
772 struct elevator_queue
*e
= q
->elevator
;
774 if (WARN_ON_ONCE(e
->uses_mq
))
777 if (e
->type
->ops
.sq
.elevator_may_queue_fn
)
778 return e
->type
->ops
.sq
.elevator_may_queue_fn(q
, op
);
780 return ELV_MQUEUE_MAY
;
783 void elv_completed_request(struct request_queue
*q
, struct request
*rq
)
785 struct elevator_queue
*e
= q
->elevator
;
787 if (WARN_ON_ONCE(e
->uses_mq
))
791 * request is released from the driver, io must be done
793 if (blk_account_rq(rq
)) {
794 q
->in_flight
[rq_is_sync(rq
)]--;
795 if ((rq
->rq_flags
& RQF_SORTED
) &&
796 e
->type
->ops
.sq
.elevator_completed_req_fn
)
797 e
->type
->ops
.sq
.elevator_completed_req_fn(q
, rq
);
801 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
804 elv_attr_show(struct kobject
*kobj
, struct attribute
*attr
, char *page
)
806 struct elv_fs_entry
*entry
= to_elv(attr
);
807 struct elevator_queue
*e
;
813 e
= container_of(kobj
, struct elevator_queue
, kobj
);
814 mutex_lock(&e
->sysfs_lock
);
815 error
= e
->type
? entry
->show(e
, page
) : -ENOENT
;
816 mutex_unlock(&e
->sysfs_lock
);
821 elv_attr_store(struct kobject
*kobj
, struct attribute
*attr
,
822 const char *page
, size_t length
)
824 struct elv_fs_entry
*entry
= to_elv(attr
);
825 struct elevator_queue
*e
;
831 e
= container_of(kobj
, struct elevator_queue
, kobj
);
832 mutex_lock(&e
->sysfs_lock
);
833 error
= e
->type
? entry
->store(e
, page
, length
) : -ENOENT
;
834 mutex_unlock(&e
->sysfs_lock
);
838 static const struct sysfs_ops elv_sysfs_ops
= {
839 .show
= elv_attr_show
,
840 .store
= elv_attr_store
,
843 static struct kobj_type elv_ktype
= {
844 .sysfs_ops
= &elv_sysfs_ops
,
845 .release
= elevator_release
,
848 int elv_register_queue(struct request_queue
*q
)
850 struct elevator_queue
*e
= q
->elevator
;
853 error
= kobject_add(&e
->kobj
, &q
->kobj
, "%s", "iosched");
855 struct elv_fs_entry
*attr
= e
->type
->elevator_attrs
;
857 while (attr
->attr
.name
) {
858 if (sysfs_create_file(&e
->kobj
, &attr
->attr
))
863 kobject_uevent(&e
->kobj
, KOBJ_ADD
);
865 if (!e
->uses_mq
&& e
->type
->ops
.sq
.elevator_registered_fn
)
866 e
->type
->ops
.sq
.elevator_registered_fn(q
);
870 EXPORT_SYMBOL(elv_register_queue
);
872 void elv_unregister_queue(struct request_queue
*q
)
875 struct elevator_queue
*e
= q
->elevator
;
877 kobject_uevent(&e
->kobj
, KOBJ_REMOVE
);
878 kobject_del(&e
->kobj
);
882 EXPORT_SYMBOL(elv_unregister_queue
);
884 int elv_register(struct elevator_type
*e
)
888 /* create icq_cache if requested */
890 if (WARN_ON(e
->icq_size
< sizeof(struct io_cq
)) ||
891 WARN_ON(e
->icq_align
< __alignof__(struct io_cq
)))
894 snprintf(e
->icq_cache_name
, sizeof(e
->icq_cache_name
),
895 "%s_io_cq", e
->elevator_name
);
896 e
->icq_cache
= kmem_cache_create(e
->icq_cache_name
, e
->icq_size
,
897 e
->icq_align
, 0, NULL
);
902 /* register, don't allow duplicate names */
903 spin_lock(&elv_list_lock
);
904 if (elevator_find(e
->elevator_name
)) {
905 spin_unlock(&elv_list_lock
);
907 kmem_cache_destroy(e
->icq_cache
);
910 list_add_tail(&e
->list
, &elv_list
);
911 spin_unlock(&elv_list_lock
);
913 /* print pretty message */
914 if (!strcmp(e
->elevator_name
, chosen_elevator
) ||
915 (!*chosen_elevator
&&
916 !strcmp(e
->elevator_name
, CONFIG_DEFAULT_IOSCHED
)))
919 printk(KERN_INFO
"io scheduler %s registered%s\n", e
->elevator_name
,
923 EXPORT_SYMBOL_GPL(elv_register
);
925 void elv_unregister(struct elevator_type
*e
)
928 spin_lock(&elv_list_lock
);
929 list_del_init(&e
->list
);
930 spin_unlock(&elv_list_lock
);
933 * Destroy icq_cache if it exists. icq's are RCU managed. Make
934 * sure all RCU operations are complete before proceeding.
938 kmem_cache_destroy(e
->icq_cache
);
942 EXPORT_SYMBOL_GPL(elv_unregister
);
944 static int elevator_switch_mq(struct request_queue
*q
,
945 struct elevator_type
*new_e
)
949 blk_mq_freeze_queue(q
);
950 blk_mq_quiesce_queue(q
);
953 if (q
->elevator
->registered
)
954 elv_unregister_queue(q
);
956 elevator_exit(q
, q
->elevator
);
959 ret
= blk_mq_init_sched(q
, new_e
);
964 ret
= elv_register_queue(q
);
966 elevator_exit(q
, q
->elevator
);
972 blk_add_trace_msg(q
, "elv switch: %s", new_e
->elevator_name
);
974 blk_add_trace_msg(q
, "elv switch: none");
977 blk_mq_unfreeze_queue(q
);
978 blk_mq_start_stopped_hw_queues(q
, true);
984 * switch to new_e io scheduler. be careful not to introduce deadlocks -
985 * we don't free the old io scheduler, before we have allocated what we
986 * need for the new one. this way we have a chance of going back to the old
987 * one, if the new one fails init for some reason.
989 static int elevator_switch(struct request_queue
*q
, struct elevator_type
*new_e
)
991 struct elevator_queue
*old
= q
->elevator
;
992 bool old_registered
= false;
996 return elevator_switch_mq(q
, new_e
);
999 * Turn on BYPASS and drain all requests w/ elevator private data.
1000 * Block layer doesn't call into a quiesced elevator - all requests
1001 * are directly put on the dispatch list without elevator data
1002 * using INSERT_BACK. All requests have SOFTBARRIER set and no
1003 * merge happens either.
1006 old_registered
= old
->registered
;
1008 blk_queue_bypass_start(q
);
1010 /* unregister and clear all auxiliary data of the old elevator */
1012 elv_unregister_queue(q
);
1017 /* allocate, init and register new elevator */
1018 err
= new_e
->ops
.sq
.elevator_init_fn(q
, new_e
);
1022 err
= elv_register_queue(q
);
1026 /* done, kill the old one and finish */
1028 elevator_exit(q
, old
);
1029 blk_queue_bypass_end(q
);
1032 blk_add_trace_msg(q
, "elv switch: %s", new_e
->elevator_name
);
1037 elevator_exit(q
, q
->elevator
);
1039 /* switch failed, restore and re-register old elevator */
1042 elv_register_queue(q
);
1043 blk_queue_bypass_end(q
);
1050 * Switch this queue to the given IO scheduler.
1052 static int __elevator_change(struct request_queue
*q
, const char *name
)
1054 char elevator_name
[ELV_NAME_MAX
];
1055 struct elevator_type
*e
;
1058 * Special case for mq, turn off scheduling
1060 if (q
->mq_ops
&& !strncmp(name
, "none", 4))
1061 return elevator_switch(q
, NULL
);
1063 strlcpy(elevator_name
, name
, sizeof(elevator_name
));
1064 e
= elevator_get(strstrip(elevator_name
), true);
1066 printk(KERN_ERR
"elevator: type %s not found\n", elevator_name
);
1071 !strcmp(elevator_name
, q
->elevator
->type
->elevator_name
)) {
1076 if (!e
->uses_mq
&& q
->mq_ops
) {
1080 if (e
->uses_mq
&& !q
->mq_ops
) {
1085 return elevator_switch(q
, e
);
1088 int elevator_change(struct request_queue
*q
, const char *name
)
1092 /* Protect q->elevator from elevator_init() */
1093 mutex_lock(&q
->sysfs_lock
);
1094 ret
= __elevator_change(q
, name
);
1095 mutex_unlock(&q
->sysfs_lock
);
1099 EXPORT_SYMBOL(elevator_change
);
1101 ssize_t
elv_iosched_store(struct request_queue
*q
, const char *name
,
1106 if (!(q
->mq_ops
|| q
->request_fn
))
1109 ret
= __elevator_change(q
, name
);
1113 printk(KERN_ERR
"elevator: switch to %s failed\n", name
);
1117 ssize_t
elv_iosched_show(struct request_queue
*q
, char *name
)
1119 struct elevator_queue
*e
= q
->elevator
;
1120 struct elevator_type
*elv
= NULL
;
1121 struct elevator_type
*__e
;
1124 if (!blk_queue_stackable(q
))
1125 return sprintf(name
, "none\n");
1128 len
+= sprintf(name
+len
, "[none] ");
1132 spin_lock(&elv_list_lock
);
1133 list_for_each_entry(__e
, &elv_list
, list
) {
1134 if (elv
&& !strcmp(elv
->elevator_name
, __e
->elevator_name
)) {
1135 len
+= sprintf(name
+len
, "[%s] ", elv
->elevator_name
);
1138 if (__e
->uses_mq
&& q
->mq_ops
)
1139 len
+= sprintf(name
+len
, "%s ", __e
->elevator_name
);
1140 else if (!__e
->uses_mq
&& !q
->mq_ops
)
1141 len
+= sprintf(name
+len
, "%s ", __e
->elevator_name
);
1143 spin_unlock(&elv_list_lock
);
1145 if (q
->mq_ops
&& q
->elevator
)
1146 len
+= sprintf(name
+len
, "none");
1148 len
+= sprintf(len
+name
, "\n");
1152 struct request
*elv_rb_former_request(struct request_queue
*q
,
1155 struct rb_node
*rbprev
= rb_prev(&rq
->rb_node
);
1158 return rb_entry_rq(rbprev
);
1162 EXPORT_SYMBOL(elv_rb_former_request
);
1164 struct request
*elv_rb_latter_request(struct request_queue
*q
,
1167 struct rb_node
*rbnext
= rb_next(&rq
->rb_node
);
1170 return rb_entry_rq(rbnext
);
1174 EXPORT_SYMBOL(elv_rb_latter_request
);