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/delay.h>
35 #include <linux/blktrace_api.h>
36 #include <linux/hash.h>
38 #include <asm/uaccess.h>
40 static DEFINE_SPINLOCK(elv_list_lock
);
41 static LIST_HEAD(elv_list
);
46 static const int elv_hash_shift
= 6;
47 #define ELV_HASH_BLOCK(sec) ((sec) >> 3)
48 #define ELV_HASH_FN(sec) (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
49 #define ELV_HASH_ENTRIES (1 << elv_hash_shift)
50 #define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors)
51 #define ELV_ON_HASH(rq) (!hlist_unhashed(&(rq)->hash))
54 * Query io scheduler to see if the current process issuing bio may be
57 static int elv_iosched_allow_merge(struct request
*rq
, struct bio
*bio
)
59 request_queue_t
*q
= rq
->q
;
60 elevator_t
*e
= q
->elevator
;
62 if (e
->ops
->elevator_allow_merge_fn
)
63 return e
->ops
->elevator_allow_merge_fn(q
, rq
, bio
);
69 * can we safely merge with this request?
71 inline int elv_rq_merge_ok(struct request
*rq
, struct bio
*bio
)
73 if (!rq_mergeable(rq
))
77 * different data direction or already started, don't merge
79 if (bio_data_dir(bio
) != rq_data_dir(rq
))
83 * must be same device and not a special request
85 if (rq
->rq_disk
!= bio
->bi_bdev
->bd_disk
|| rq
->special
)
88 if (!elv_iosched_allow_merge(rq
, bio
))
93 EXPORT_SYMBOL(elv_rq_merge_ok
);
95 static inline int elv_try_merge(struct request
*__rq
, struct bio
*bio
)
97 int ret
= ELEVATOR_NO_MERGE
;
100 * we can merge and sequence is ok, check if it's possible
102 if (elv_rq_merge_ok(__rq
, bio
)) {
103 if (__rq
->sector
+ __rq
->nr_sectors
== bio
->bi_sector
)
104 ret
= ELEVATOR_BACK_MERGE
;
105 else if (__rq
->sector
- bio_sectors(bio
) == bio
->bi_sector
)
106 ret
= ELEVATOR_FRONT_MERGE
;
112 static struct elevator_type
*elevator_find(const char *name
)
114 struct elevator_type
*e
;
115 struct list_head
*entry
;
117 list_for_each(entry
, &elv_list
) {
119 e
= list_entry(entry
, struct elevator_type
, list
);
121 if (!strcmp(e
->elevator_name
, name
))
128 static void elevator_put(struct elevator_type
*e
)
130 module_put(e
->elevator_owner
);
133 static struct elevator_type
*elevator_get(const char *name
)
135 struct elevator_type
*e
;
137 spin_lock_irq(&elv_list_lock
);
139 e
= elevator_find(name
);
140 if (e
&& !try_module_get(e
->elevator_owner
))
143 spin_unlock_irq(&elv_list_lock
);
148 static void *elevator_init_queue(request_queue_t
*q
, struct elevator_queue
*eq
)
150 return eq
->ops
->elevator_init_fn(q
);
153 static void elevator_attach(request_queue_t
*q
, struct elevator_queue
*eq
,
157 eq
->elevator_data
= data
;
160 static char chosen_elevator
[16];
162 static int __init
elevator_setup(char *str
)
165 * Be backwards-compatible with previous kernels, so users
166 * won't get the wrong elevator.
168 if (!strcmp(str
, "as"))
169 strcpy(chosen_elevator
, "anticipatory");
171 strncpy(chosen_elevator
, str
, sizeof(chosen_elevator
) - 1);
175 __setup("elevator=", elevator_setup
);
177 static struct kobj_type elv_ktype
;
179 static elevator_t
*elevator_alloc(request_queue_t
*q
, struct elevator_type
*e
)
184 eq
= kmalloc_node(sizeof(elevator_t
), GFP_KERNEL
, q
->node
);
188 memset(eq
, 0, sizeof(*eq
));
190 eq
->elevator_type
= e
;
191 kobject_init(&eq
->kobj
);
192 snprintf(eq
->kobj
.name
, KOBJ_NAME_LEN
, "%s", "iosched");
193 eq
->kobj
.ktype
= &elv_ktype
;
194 mutex_init(&eq
->sysfs_lock
);
196 eq
->hash
= kmalloc_node(sizeof(struct hlist_head
) * ELV_HASH_ENTRIES
,
197 GFP_KERNEL
, q
->node
);
201 for (i
= 0; i
< ELV_HASH_ENTRIES
; i
++)
202 INIT_HLIST_HEAD(&eq
->hash
[i
]);
211 static void elevator_release(struct kobject
*kobj
)
213 elevator_t
*e
= container_of(kobj
, elevator_t
, kobj
);
215 elevator_put(e
->elevator_type
);
220 int elevator_init(request_queue_t
*q
, char *name
)
222 struct elevator_type
*e
= NULL
;
223 struct elevator_queue
*eq
;
227 INIT_LIST_HEAD(&q
->queue_head
);
228 q
->last_merge
= NULL
;
230 q
->boundary_rq
= NULL
;
232 if (name
&& !(e
= elevator_get(name
)))
235 if (!e
&& *chosen_elevator
&& !(e
= elevator_get(chosen_elevator
)))
236 printk("I/O scheduler %s not found\n", chosen_elevator
);
238 if (!e
&& !(e
= elevator_get(CONFIG_DEFAULT_IOSCHED
))) {
239 printk("Default I/O scheduler not found, using no-op\n");
240 e
= elevator_get("noop");
243 eq
= elevator_alloc(q
, e
);
247 data
= elevator_init_queue(q
, eq
);
249 kobject_put(&eq
->kobj
);
253 elevator_attach(q
, eq
, data
);
257 EXPORT_SYMBOL(elevator_init
);
259 void elevator_exit(elevator_t
*e
)
261 mutex_lock(&e
->sysfs_lock
);
262 if (e
->ops
->elevator_exit_fn
)
263 e
->ops
->elevator_exit_fn(e
);
265 mutex_unlock(&e
->sysfs_lock
);
267 kobject_put(&e
->kobj
);
270 EXPORT_SYMBOL(elevator_exit
);
272 static inline void __elv_rqhash_del(struct request
*rq
)
274 hlist_del_init(&rq
->hash
);
277 static void elv_rqhash_del(request_queue_t
*q
, struct request
*rq
)
280 __elv_rqhash_del(rq
);
283 static void elv_rqhash_add(request_queue_t
*q
, struct request
*rq
)
285 elevator_t
*e
= q
->elevator
;
287 BUG_ON(ELV_ON_HASH(rq
));
288 hlist_add_head(&rq
->hash
, &e
->hash
[ELV_HASH_FN(rq_hash_key(rq
))]);
291 static void elv_rqhash_reposition(request_queue_t
*q
, struct request
*rq
)
293 __elv_rqhash_del(rq
);
294 elv_rqhash_add(q
, rq
);
297 static struct request
*elv_rqhash_find(request_queue_t
*q
, sector_t offset
)
299 elevator_t
*e
= q
->elevator
;
300 struct hlist_head
*hash_list
= &e
->hash
[ELV_HASH_FN(offset
)];
301 struct hlist_node
*entry
, *next
;
304 hlist_for_each_entry_safe(rq
, entry
, next
, hash_list
, hash
) {
305 BUG_ON(!ELV_ON_HASH(rq
));
307 if (unlikely(!rq_mergeable(rq
))) {
308 __elv_rqhash_del(rq
);
312 if (rq_hash_key(rq
) == offset
)
320 * RB-tree support functions for inserting/lookup/removal of requests
321 * in a sorted RB tree.
323 struct request
*elv_rb_add(struct rb_root
*root
, struct request
*rq
)
325 struct rb_node
**p
= &root
->rb_node
;
326 struct rb_node
*parent
= NULL
;
327 struct request
*__rq
;
331 __rq
= rb_entry(parent
, struct request
, rb_node
);
333 if (rq
->sector
< __rq
->sector
)
335 else if (rq
->sector
> __rq
->sector
)
341 rb_link_node(&rq
->rb_node
, parent
, p
);
342 rb_insert_color(&rq
->rb_node
, root
);
346 EXPORT_SYMBOL(elv_rb_add
);
348 void elv_rb_del(struct rb_root
*root
, struct request
*rq
)
350 BUG_ON(RB_EMPTY_NODE(&rq
->rb_node
));
351 rb_erase(&rq
->rb_node
, root
);
352 RB_CLEAR_NODE(&rq
->rb_node
);
355 EXPORT_SYMBOL(elv_rb_del
);
357 struct request
*elv_rb_find(struct rb_root
*root
, sector_t sector
)
359 struct rb_node
*n
= root
->rb_node
;
363 rq
= rb_entry(n
, struct request
, rb_node
);
365 if (sector
< rq
->sector
)
367 else if (sector
> rq
->sector
)
376 EXPORT_SYMBOL(elv_rb_find
);
379 * Insert rq into dispatch queue of q. Queue lock must be held on
380 * entry. rq is sort insted into the dispatch queue. To be used by
381 * specific elevators.
383 void elv_dispatch_sort(request_queue_t
*q
, struct request
*rq
)
386 struct list_head
*entry
;
388 if (q
->last_merge
== rq
)
389 q
->last_merge
= NULL
;
391 elv_rqhash_del(q
, rq
);
395 boundary
= q
->end_sector
;
397 list_for_each_prev(entry
, &q
->queue_head
) {
398 struct request
*pos
= list_entry_rq(entry
);
400 if (pos
->cmd_flags
& (REQ_SOFTBARRIER
|REQ_HARDBARRIER
|REQ_STARTED
))
402 if (rq
->sector
>= boundary
) {
403 if (pos
->sector
< boundary
)
406 if (pos
->sector
>= boundary
)
409 if (rq
->sector
>= pos
->sector
)
413 list_add(&rq
->queuelist
, entry
);
416 EXPORT_SYMBOL(elv_dispatch_sort
);
419 * Insert rq into dispatch queue of q. Queue lock must be held on
420 * entry. rq is added to the back of the dispatch queue. To be used by
421 * specific elevators.
423 void elv_dispatch_add_tail(struct request_queue
*q
, struct request
*rq
)
425 if (q
->last_merge
== rq
)
426 q
->last_merge
= NULL
;
428 elv_rqhash_del(q
, rq
);
432 q
->end_sector
= rq_end_sector(rq
);
434 list_add_tail(&rq
->queuelist
, &q
->queue_head
);
437 EXPORT_SYMBOL(elv_dispatch_add_tail
);
439 int elv_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
441 elevator_t
*e
= q
->elevator
;
442 struct request
*__rq
;
446 * First try one-hit cache.
449 ret
= elv_try_merge(q
->last_merge
, bio
);
450 if (ret
!= ELEVATOR_NO_MERGE
) {
451 *req
= q
->last_merge
;
457 * See if our hash lookup can find a potential backmerge.
459 __rq
= elv_rqhash_find(q
, bio
->bi_sector
);
460 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
462 return ELEVATOR_BACK_MERGE
;
465 if (e
->ops
->elevator_merge_fn
)
466 return e
->ops
->elevator_merge_fn(q
, req
, bio
);
468 return ELEVATOR_NO_MERGE
;
471 void elv_merged_request(request_queue_t
*q
, struct request
*rq
, int type
)
473 elevator_t
*e
= q
->elevator
;
475 if (e
->ops
->elevator_merged_fn
)
476 e
->ops
->elevator_merged_fn(q
, rq
, type
);
478 if (type
== ELEVATOR_BACK_MERGE
)
479 elv_rqhash_reposition(q
, rq
);
484 void elv_merge_requests(request_queue_t
*q
, struct request
*rq
,
485 struct request
*next
)
487 elevator_t
*e
= q
->elevator
;
489 if (e
->ops
->elevator_merge_req_fn
)
490 e
->ops
->elevator_merge_req_fn(q
, rq
, next
);
492 elv_rqhash_reposition(q
, rq
);
493 elv_rqhash_del(q
, next
);
499 void elv_requeue_request(request_queue_t
*q
, struct request
*rq
)
501 elevator_t
*e
= q
->elevator
;
504 * it already went through dequeue, we need to decrement the
505 * in_flight count again
507 if (blk_account_rq(rq
)) {
509 if (blk_sorted_rq(rq
) && e
->ops
->elevator_deactivate_req_fn
)
510 e
->ops
->elevator_deactivate_req_fn(q
, rq
);
513 rq
->cmd_flags
&= ~REQ_STARTED
;
515 elv_insert(q
, rq
, ELEVATOR_INSERT_REQUEUE
);
518 static void elv_drain_elevator(request_queue_t
*q
)
521 while (q
->elevator
->ops
->elevator_dispatch_fn(q
, 1))
523 if (q
->nr_sorted
== 0)
525 if (printed
++ < 10) {
526 printk(KERN_ERR
"%s: forced dispatching is broken "
527 "(nr_sorted=%u), please report this\n",
528 q
->elevator
->elevator_type
->elevator_name
, q
->nr_sorted
);
532 void elv_insert(request_queue_t
*q
, struct request
*rq
, int where
)
534 struct list_head
*pos
;
538 blk_add_trace_rq(q
, rq
, BLK_TA_INSERT
);
543 case ELEVATOR_INSERT_FRONT
:
544 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
546 list_add(&rq
->queuelist
, &q
->queue_head
);
549 case ELEVATOR_INSERT_BACK
:
550 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
551 elv_drain_elevator(q
);
552 list_add_tail(&rq
->queuelist
, &q
->queue_head
);
554 * We kick the queue here for the following reasons.
555 * - The elevator might have returned NULL previously
556 * to delay requests and returned them now. As the
557 * queue wasn't empty before this request, ll_rw_blk
558 * won't run the queue on return, resulting in hang.
559 * - Usually, back inserted requests won't be merged
560 * with anything. There's no point in delaying queue
567 case ELEVATOR_INSERT_SORT
:
568 BUG_ON(!blk_fs_request(rq
));
569 rq
->cmd_flags
|= REQ_SORTED
;
571 if (rq_mergeable(rq
)) {
572 elv_rqhash_add(q
, rq
);
578 * Some ioscheds (cfq) run q->request_fn directly, so
579 * rq cannot be accessed after calling
580 * elevator_add_req_fn.
582 q
->elevator
->ops
->elevator_add_req_fn(q
, rq
);
585 case ELEVATOR_INSERT_REQUEUE
:
587 * If ordered flush isn't in progress, we do front
588 * insertion; otherwise, requests should be requeued
591 rq
->cmd_flags
|= REQ_SOFTBARRIER
;
593 if (q
->ordseq
== 0) {
594 list_add(&rq
->queuelist
, &q
->queue_head
);
598 ordseq
= blk_ordered_req_seq(rq
);
600 list_for_each(pos
, &q
->queue_head
) {
601 struct request
*pos_rq
= list_entry_rq(pos
);
602 if (ordseq
<= blk_ordered_req_seq(pos_rq
))
606 list_add_tail(&rq
->queuelist
, pos
);
608 * most requeues happen because of a busy condition, don't
609 * force unplug of the queue for that case.
615 printk(KERN_ERR
"%s: bad insertion point %d\n",
616 __FUNCTION__
, where
);
620 if (unplug_it
&& blk_queue_plugged(q
)) {
621 int nrq
= q
->rq
.count
[READ
] + q
->rq
.count
[WRITE
]
624 if (nrq
>= q
->unplug_thresh
)
625 __generic_unplug_device(q
);
629 void __elv_add_request(request_queue_t
*q
, struct request
*rq
, int where
,
633 rq
->cmd_flags
|= REQ_ORDERED_COLOR
;
635 if (rq
->cmd_flags
& (REQ_SOFTBARRIER
| REQ_HARDBARRIER
)) {
637 * toggle ordered color
639 if (blk_barrier_rq(rq
))
643 * barriers implicitly indicate back insertion
645 if (where
== ELEVATOR_INSERT_SORT
)
646 where
= ELEVATOR_INSERT_BACK
;
649 * this request is scheduling boundary, update
652 if (blk_fs_request(rq
)) {
653 q
->end_sector
= rq_end_sector(rq
);
656 } else if (!(rq
->cmd_flags
& REQ_ELVPRIV
) && where
== ELEVATOR_INSERT_SORT
)
657 where
= ELEVATOR_INSERT_BACK
;
662 elv_insert(q
, rq
, where
);
665 EXPORT_SYMBOL(__elv_add_request
);
667 void elv_add_request(request_queue_t
*q
, struct request
*rq
, int where
,
672 spin_lock_irqsave(q
->queue_lock
, flags
);
673 __elv_add_request(q
, rq
, where
, plug
);
674 spin_unlock_irqrestore(q
->queue_lock
, flags
);
677 EXPORT_SYMBOL(elv_add_request
);
679 static inline struct request
*__elv_next_request(request_queue_t
*q
)
684 while (!list_empty(&q
->queue_head
)) {
685 rq
= list_entry_rq(q
->queue_head
.next
);
686 if (blk_do_ordered(q
, &rq
))
690 if (!q
->elevator
->ops
->elevator_dispatch_fn(q
, 0))
695 struct request
*elv_next_request(request_queue_t
*q
)
700 while ((rq
= __elv_next_request(q
)) != NULL
) {
701 if (!(rq
->cmd_flags
& REQ_STARTED
)) {
702 elevator_t
*e
= q
->elevator
;
705 * This is the first time the device driver
706 * sees this request (possibly after
707 * requeueing). Notify IO scheduler.
709 if (blk_sorted_rq(rq
) &&
710 e
->ops
->elevator_activate_req_fn
)
711 e
->ops
->elevator_activate_req_fn(q
, rq
);
714 * just mark as started even if we don't start
715 * it, a request that has been delayed should
716 * not be passed by new incoming requests
718 rq
->cmd_flags
|= REQ_STARTED
;
719 blk_add_trace_rq(q
, rq
, BLK_TA_ISSUE
);
722 if (!q
->boundary_rq
|| q
->boundary_rq
== rq
) {
723 q
->end_sector
= rq_end_sector(rq
);
724 q
->boundary_rq
= NULL
;
727 if ((rq
->cmd_flags
& REQ_DONTPREP
) || !q
->prep_rq_fn
)
730 ret
= q
->prep_rq_fn(q
, rq
);
731 if (ret
== BLKPREP_OK
) {
733 } else if (ret
== BLKPREP_DEFER
) {
735 * the request may have been (partially) prepped.
736 * we need to keep this request in the front to
737 * avoid resource deadlock. REQ_STARTED will
738 * prevent other fs requests from passing this one.
742 } else if (ret
== BLKPREP_KILL
) {
743 int nr_bytes
= rq
->hard_nr_sectors
<< 9;
746 nr_bytes
= rq
->data_len
;
748 blkdev_dequeue_request(rq
);
749 rq
->cmd_flags
|= REQ_QUIET
;
750 end_that_request_chunk(rq
, 0, nr_bytes
);
751 end_that_request_last(rq
, 0);
753 printk(KERN_ERR
"%s: bad return=%d\n", __FUNCTION__
,
762 EXPORT_SYMBOL(elv_next_request
);
764 void elv_dequeue_request(request_queue_t
*q
, struct request
*rq
)
766 BUG_ON(list_empty(&rq
->queuelist
));
767 BUG_ON(ELV_ON_HASH(rq
));
769 list_del_init(&rq
->queuelist
);
772 * the time frame between a request being removed from the lists
773 * and to it is freed is accounted as io that is in progress at
776 if (blk_account_rq(rq
))
780 EXPORT_SYMBOL(elv_dequeue_request
);
782 int elv_queue_empty(request_queue_t
*q
)
784 elevator_t
*e
= q
->elevator
;
786 if (!list_empty(&q
->queue_head
))
789 if (e
->ops
->elevator_queue_empty_fn
)
790 return e
->ops
->elevator_queue_empty_fn(q
);
795 EXPORT_SYMBOL(elv_queue_empty
);
797 struct request
*elv_latter_request(request_queue_t
*q
, struct request
*rq
)
799 elevator_t
*e
= q
->elevator
;
801 if (e
->ops
->elevator_latter_req_fn
)
802 return e
->ops
->elevator_latter_req_fn(q
, rq
);
806 struct request
*elv_former_request(request_queue_t
*q
, struct request
*rq
)
808 elevator_t
*e
= q
->elevator
;
810 if (e
->ops
->elevator_former_req_fn
)
811 return e
->ops
->elevator_former_req_fn(q
, rq
);
815 int elv_set_request(request_queue_t
*q
, struct request
*rq
, gfp_t gfp_mask
)
817 elevator_t
*e
= q
->elevator
;
819 if (e
->ops
->elevator_set_req_fn
)
820 return e
->ops
->elevator_set_req_fn(q
, rq
, gfp_mask
);
822 rq
->elevator_private
= NULL
;
826 void elv_put_request(request_queue_t
*q
, struct request
*rq
)
828 elevator_t
*e
= q
->elevator
;
830 if (e
->ops
->elevator_put_req_fn
)
831 e
->ops
->elevator_put_req_fn(rq
);
834 int elv_may_queue(request_queue_t
*q
, int rw
)
836 elevator_t
*e
= q
->elevator
;
838 if (e
->ops
->elevator_may_queue_fn
)
839 return e
->ops
->elevator_may_queue_fn(q
, rw
);
841 return ELV_MQUEUE_MAY
;
844 void elv_completed_request(request_queue_t
*q
, struct request
*rq
)
846 elevator_t
*e
= q
->elevator
;
849 * request is released from the driver, io must be done
851 if (blk_account_rq(rq
)) {
853 if (blk_sorted_rq(rq
) && e
->ops
->elevator_completed_req_fn
)
854 e
->ops
->elevator_completed_req_fn(q
, rq
);
858 * Check if the queue is waiting for fs requests to be
859 * drained for flush sequence.
861 if (unlikely(q
->ordseq
)) {
862 struct request
*first_rq
= list_entry_rq(q
->queue_head
.next
);
863 if (q
->in_flight
== 0 &&
864 blk_ordered_cur_seq(q
) == QUEUE_ORDSEQ_DRAIN
&&
865 blk_ordered_req_seq(first_rq
) > QUEUE_ORDSEQ_DRAIN
) {
866 blk_ordered_complete_seq(q
, QUEUE_ORDSEQ_DRAIN
, 0);
872 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
875 elv_attr_show(struct kobject
*kobj
, struct attribute
*attr
, char *page
)
877 elevator_t
*e
= container_of(kobj
, elevator_t
, kobj
);
878 struct elv_fs_entry
*entry
= to_elv(attr
);
884 mutex_lock(&e
->sysfs_lock
);
885 error
= e
->ops
? entry
->show(e
, page
) : -ENOENT
;
886 mutex_unlock(&e
->sysfs_lock
);
891 elv_attr_store(struct kobject
*kobj
, struct attribute
*attr
,
892 const char *page
, size_t length
)
894 elevator_t
*e
= container_of(kobj
, elevator_t
, kobj
);
895 struct elv_fs_entry
*entry
= to_elv(attr
);
901 mutex_lock(&e
->sysfs_lock
);
902 error
= e
->ops
? entry
->store(e
, page
, length
) : -ENOENT
;
903 mutex_unlock(&e
->sysfs_lock
);
907 static struct sysfs_ops elv_sysfs_ops
= {
908 .show
= elv_attr_show
,
909 .store
= elv_attr_store
,
912 static struct kobj_type elv_ktype
= {
913 .sysfs_ops
= &elv_sysfs_ops
,
914 .release
= elevator_release
,
917 int elv_register_queue(struct request_queue
*q
)
919 elevator_t
*e
= q
->elevator
;
922 e
->kobj
.parent
= &q
->kobj
;
924 error
= kobject_add(&e
->kobj
);
926 struct elv_fs_entry
*attr
= e
->elevator_type
->elevator_attrs
;
928 while (attr
->attr
.name
) {
929 if (sysfs_create_file(&e
->kobj
, &attr
->attr
))
934 kobject_uevent(&e
->kobj
, KOBJ_ADD
);
939 static void __elv_unregister_queue(elevator_t
*e
)
941 kobject_uevent(&e
->kobj
, KOBJ_REMOVE
);
942 kobject_del(&e
->kobj
);
945 void elv_unregister_queue(struct request_queue
*q
)
948 __elv_unregister_queue(q
->elevator
);
951 int elv_register(struct elevator_type
*e
)
953 spin_lock_irq(&elv_list_lock
);
954 BUG_ON(elevator_find(e
->elevator_name
));
955 list_add_tail(&e
->list
, &elv_list
);
956 spin_unlock_irq(&elv_list_lock
);
958 printk(KERN_INFO
"io scheduler %s registered", e
->elevator_name
);
959 if (!strcmp(e
->elevator_name
, chosen_elevator
) ||
960 (!*chosen_elevator
&&
961 !strcmp(e
->elevator_name
, CONFIG_DEFAULT_IOSCHED
)))
962 printk(" (default)");
966 EXPORT_SYMBOL_GPL(elv_register
);
968 void elv_unregister(struct elevator_type
*e
)
970 struct task_struct
*g
, *p
;
973 * Iterate every thread in the process to remove the io contexts.
976 read_lock(&tasklist_lock
);
977 do_each_thread(g
, p
) {
980 e
->ops
.trim(p
->io_context
);
982 } while_each_thread(g
, p
);
983 read_unlock(&tasklist_lock
);
986 spin_lock_irq(&elv_list_lock
);
987 list_del_init(&e
->list
);
988 spin_unlock_irq(&elv_list_lock
);
990 EXPORT_SYMBOL_GPL(elv_unregister
);
993 * switch to new_e io scheduler. be careful not to introduce deadlocks -
994 * we don't free the old io scheduler, before we have allocated what we
995 * need for the new one. this way we have a chance of going back to the old
996 * one, if the new one fails init for some reason.
998 static int elevator_switch(request_queue_t
*q
, struct elevator_type
*new_e
)
1000 elevator_t
*old_elevator
, *e
;
1004 * Allocate new elevator
1006 e
= elevator_alloc(q
, new_e
);
1010 data
= elevator_init_queue(q
, e
);
1012 kobject_put(&e
->kobj
);
1017 * Turn on BYPASS and drain all requests w/ elevator private data
1019 spin_lock_irq(q
->queue_lock
);
1021 set_bit(QUEUE_FLAG_ELVSWITCH
, &q
->queue_flags
);
1023 elv_drain_elevator(q
);
1025 while (q
->rq
.elvpriv
) {
1028 spin_unlock_irq(q
->queue_lock
);
1030 spin_lock_irq(q
->queue_lock
);
1031 elv_drain_elevator(q
);
1035 * Remember old elevator.
1037 old_elevator
= q
->elevator
;
1040 * attach and start new elevator
1042 elevator_attach(q
, e
, data
);
1044 spin_unlock_irq(q
->queue_lock
);
1046 __elv_unregister_queue(old_elevator
);
1048 if (elv_register_queue(q
))
1052 * finally exit old elevator and turn off BYPASS.
1054 elevator_exit(old_elevator
);
1055 clear_bit(QUEUE_FLAG_ELVSWITCH
, &q
->queue_flags
);
1060 * switch failed, exit the new io scheduler and reattach the old
1061 * one again (along with re-adding the sysfs dir)
1064 q
->elevator
= old_elevator
;
1065 elv_register_queue(q
);
1066 clear_bit(QUEUE_FLAG_ELVSWITCH
, &q
->queue_flags
);
1070 ssize_t
elv_iosched_store(request_queue_t
*q
, const char *name
, size_t count
)
1072 char elevator_name
[ELV_NAME_MAX
];
1074 struct elevator_type
*e
;
1076 elevator_name
[sizeof(elevator_name
) - 1] = '\0';
1077 strncpy(elevator_name
, name
, sizeof(elevator_name
) - 1);
1078 len
= strlen(elevator_name
);
1080 if (len
&& elevator_name
[len
- 1] == '\n')
1081 elevator_name
[len
- 1] = '\0';
1083 e
= elevator_get(elevator_name
);
1085 printk(KERN_ERR
"elevator: type %s not found\n", elevator_name
);
1089 if (!strcmp(elevator_name
, q
->elevator
->elevator_type
->elevator_name
)) {
1094 if (!elevator_switch(q
, e
))
1095 printk(KERN_ERR
"elevator: switch to %s failed\n",elevator_name
);
1099 ssize_t
elv_iosched_show(request_queue_t
*q
, char *name
)
1101 elevator_t
*e
= q
->elevator
;
1102 struct elevator_type
*elv
= e
->elevator_type
;
1103 struct list_head
*entry
;
1106 spin_lock_irq(&elv_list_lock
);
1107 list_for_each(entry
, &elv_list
) {
1108 struct elevator_type
*__e
;
1110 __e
= list_entry(entry
, struct elevator_type
, list
);
1111 if (!strcmp(elv
->elevator_name
, __e
->elevator_name
))
1112 len
+= sprintf(name
+len
, "[%s] ", elv
->elevator_name
);
1114 len
+= sprintf(name
+len
, "%s ", __e
->elevator_name
);
1116 spin_unlock_irq(&elv_list_lock
);
1118 len
+= sprintf(len
+name
, "\n");
1122 struct request
*elv_rb_former_request(request_queue_t
*q
, struct request
*rq
)
1124 struct rb_node
*rbprev
= rb_prev(&rq
->rb_node
);
1127 return rb_entry_rq(rbprev
);
1132 EXPORT_SYMBOL(elv_rb_former_request
);
1134 struct request
*elv_rb_latter_request(request_queue_t
*q
, struct request
*rq
)
1136 struct rb_node
*rbnext
= rb_next(&rq
->rb_node
);
1139 return rb_entry_rq(rbnext
);
1144 EXPORT_SYMBOL(elv_rb_latter_request
);