2 * linux/drivers/block/elevator.c
4 * Block device elevator/IO-scheduler.
6 * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
8 * 30042000 Jens Axboe <axboe@suse.de> :
10 * Split the elevator a bit so that it is possible to choose a different
11 * one or even write a new "plug in". There are three pieces:
12 * - elevator_fn, inserts a new request in the queue list
13 * - elevator_merge_fn, decides whether a new buffer can be merged with
15 * - elevator_dequeue_fn, called when a request is taken off the active list
17 * 20082000 Dave Jones <davej@suse.de> :
18 * Removed tests for max-bomb-segments, which was breaking elvtune
19 * when run without -bN
22 * - Rework again to work with bio instead of buffer_heads
23 * - loose bi_dev comparisons, partition handling is right now
24 * - completely modularize elevator setup and teardown
27 #include <linux/kernel.h>
29 #include <linux/blkdev.h>
30 #include <linux/elevator.h>
31 #include <linux/bio.h>
32 #include <linux/config.h>
33 #include <linux/module.h>
34 #include <linux/slab.h>
35 #include <linux/init.h>
36 #include <linux/compiler.h>
37 #include <linux/delay.h>
39 #include <asm/uaccess.h>
41 static DEFINE_SPINLOCK(elv_list_lock
);
42 static LIST_HEAD(elv_list
);
45 * can we safely merge with this request?
47 inline int elv_rq_merge_ok(struct request
*rq
, struct bio
*bio
)
49 if (!rq_mergeable(rq
))
53 * different data direction or already started, don't merge
55 if (bio_data_dir(bio
) != rq_data_dir(rq
))
59 * same device and no special stuff set, merge is ok
61 if (rq
->rq_disk
== bio
->bi_bdev
->bd_disk
&&
62 !rq
->waiting
&& !rq
->special
)
67 EXPORT_SYMBOL(elv_rq_merge_ok
);
69 inline int elv_try_merge(struct request
*__rq
, struct bio
*bio
)
71 int ret
= ELEVATOR_NO_MERGE
;
74 * we can merge and sequence is ok, check if it's possible
76 if (elv_rq_merge_ok(__rq
, bio
)) {
77 if (__rq
->sector
+ __rq
->nr_sectors
== bio
->bi_sector
)
78 ret
= ELEVATOR_BACK_MERGE
;
79 else if (__rq
->sector
- bio_sectors(bio
) == bio
->bi_sector
)
80 ret
= ELEVATOR_FRONT_MERGE
;
85 EXPORT_SYMBOL(elv_try_merge
);
87 static struct elevator_type
*elevator_find(const char *name
)
89 struct elevator_type
*e
= NULL
;
90 struct list_head
*entry
;
92 list_for_each(entry
, &elv_list
) {
93 struct elevator_type
*__e
;
95 __e
= list_entry(entry
, struct elevator_type
, list
);
97 if (!strcmp(__e
->elevator_name
, name
)) {
106 static void elevator_put(struct elevator_type
*e
)
108 module_put(e
->elevator_owner
);
111 static struct elevator_type
*elevator_get(const char *name
)
113 struct elevator_type
*e
;
115 spin_lock_irq(&elv_list_lock
);
117 e
= elevator_find(name
);
118 if (e
&& !try_module_get(e
->elevator_owner
))
121 spin_unlock_irq(&elv_list_lock
);
126 static int elevator_attach(request_queue_t
*q
, struct elevator_type
*e
,
127 struct elevator_queue
*eq
)
131 memset(eq
, 0, sizeof(*eq
));
133 eq
->elevator_type
= e
;
137 if (eq
->ops
->elevator_init_fn
)
138 ret
= eq
->ops
->elevator_init_fn(q
, eq
);
143 static char chosen_elevator
[16];
145 static void elevator_setup_default(void)
147 struct elevator_type
*e
;
150 * check if default is set and exists
152 if (chosen_elevator
[0] && (e
= elevator_get(chosen_elevator
))) {
157 #if defined(CONFIG_IOSCHED_AS)
158 strcpy(chosen_elevator
, "anticipatory");
159 #elif defined(CONFIG_IOSCHED_DEADLINE)
160 strcpy(chosen_elevator
, "deadline");
161 #elif defined(CONFIG_IOSCHED_CFQ)
162 strcpy(chosen_elevator
, "cfq");
163 #elif defined(CONFIG_IOSCHED_NOOP)
164 strcpy(chosen_elevator
, "noop");
166 #error "You must build at least 1 IO scheduler into the kernel"
170 static int __init
elevator_setup(char *str
)
172 strncpy(chosen_elevator
, str
, sizeof(chosen_elevator
) - 1);
176 __setup("elevator=", elevator_setup
);
178 int elevator_init(request_queue_t
*q
, char *name
)
180 struct elevator_type
*e
= NULL
;
181 struct elevator_queue
*eq
;
184 INIT_LIST_HEAD(&q
->queue_head
);
185 q
->last_merge
= NULL
;
187 q
->boundary_rq
= NULL
;
189 elevator_setup_default();
192 name
= chosen_elevator
;
194 e
= elevator_get(name
);
198 eq
= kmalloc(sizeof(struct elevator_queue
), GFP_KERNEL
);
200 elevator_put(e
->elevator_type
);
204 ret
= elevator_attach(q
, e
, eq
);
207 elevator_put(e
->elevator_type
);
213 void elevator_exit(elevator_t
*e
)
215 if (e
->ops
->elevator_exit_fn
)
216 e
->ops
->elevator_exit_fn(e
);
218 elevator_put(e
->elevator_type
);
219 e
->elevator_type
= NULL
;
224 * Insert rq into dispatch queue of q. Queue lock must be held on
225 * entry. If sort != 0, rq is sort-inserted; otherwise, rq will be
226 * appended to the dispatch queue. To be used by specific elevators.
228 void elv_dispatch_sort(request_queue_t
*q
, struct request
*rq
)
231 struct list_head
*entry
;
233 if (q
->last_merge
== rq
)
234 q
->last_merge
= NULL
;
236 boundary
= q
->end_sector
;
238 list_for_each_prev(entry
, &q
->queue_head
) {
239 struct request
*pos
= list_entry_rq(entry
);
241 if (pos
->flags
& (REQ_SOFTBARRIER
|REQ_HARDBARRIER
|REQ_STARTED
))
243 if (rq
->sector
>= boundary
) {
244 if (pos
->sector
< boundary
)
247 if (pos
->sector
>= boundary
)
250 if (rq
->sector
>= pos
->sector
)
254 list_add(&rq
->queuelist
, entry
);
257 int elv_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
259 elevator_t
*e
= q
->elevator
;
263 ret
= elv_try_merge(q
->last_merge
, bio
);
264 if (ret
!= ELEVATOR_NO_MERGE
) {
265 *req
= q
->last_merge
;
270 if (e
->ops
->elevator_merge_fn
)
271 return e
->ops
->elevator_merge_fn(q
, req
, bio
);
273 return ELEVATOR_NO_MERGE
;
276 void elv_merged_request(request_queue_t
*q
, struct request
*rq
)
278 elevator_t
*e
= q
->elevator
;
280 if (e
->ops
->elevator_merged_fn
)
281 e
->ops
->elevator_merged_fn(q
, rq
);
286 void elv_merge_requests(request_queue_t
*q
, struct request
*rq
,
287 struct request
*next
)
289 elevator_t
*e
= q
->elevator
;
291 if (e
->ops
->elevator_merge_req_fn
)
292 e
->ops
->elevator_merge_req_fn(q
, rq
, next
);
297 void elv_requeue_request(request_queue_t
*q
, struct request
*rq
)
299 elevator_t
*e
= q
->elevator
;
302 * it already went through dequeue, we need to decrement the
303 * in_flight count again
305 if (blk_account_rq(rq
)) {
307 if (blk_sorted_rq(rq
) && e
->ops
->elevator_deactivate_req_fn
)
308 e
->ops
->elevator_deactivate_req_fn(q
, rq
);
311 rq
->flags
&= ~REQ_STARTED
;
314 * if this is the flush, requeue the original instead and drop the flush
316 if (rq
->flags
& REQ_BAR_FLUSH
) {
317 clear_bit(QUEUE_FLAG_FLUSH
, &q
->queue_flags
);
318 rq
= rq
->end_io_data
;
321 __elv_add_request(q
, rq
, ELEVATOR_INSERT_FRONT
, 0);
324 void __elv_add_request(request_queue_t
*q
, struct request
*rq
, int where
,
327 if (rq
->flags
& (REQ_SOFTBARRIER
| REQ_HARDBARRIER
)) {
329 * barriers implicitly indicate back insertion
331 if (where
== ELEVATOR_INSERT_SORT
)
332 where
= ELEVATOR_INSERT_BACK
;
335 * this request is scheduling boundary, update end_sector
337 if (blk_fs_request(rq
)) {
338 q
->end_sector
= rq_end_sector(rq
);
341 } else if (!(rq
->flags
& REQ_ELVPRIV
) && where
== ELEVATOR_INSERT_SORT
)
342 where
= ELEVATOR_INSERT_BACK
;
350 case ELEVATOR_INSERT_FRONT
:
351 rq
->flags
|= REQ_SOFTBARRIER
;
353 list_add(&rq
->queuelist
, &q
->queue_head
);
356 case ELEVATOR_INSERT_BACK
:
357 rq
->flags
|= REQ_SOFTBARRIER
;
359 while (q
->elevator
->ops
->elevator_dispatch_fn(q
, 1))
361 list_add_tail(&rq
->queuelist
, &q
->queue_head
);
363 * We kick the queue here for the following reasons.
364 * - The elevator might have returned NULL previously
365 * to delay requests and returned them now. As the
366 * queue wasn't empty before this request, ll_rw_blk
367 * won't run the queue on return, resulting in hang.
368 * - Usually, back inserted requests won't be merged
369 * with anything. There's no point in delaying queue
376 case ELEVATOR_INSERT_SORT
:
377 BUG_ON(!blk_fs_request(rq
));
378 rq
->flags
|= REQ_SORTED
;
379 q
->elevator
->ops
->elevator_add_req_fn(q
, rq
);
380 if (q
->last_merge
== NULL
&& rq_mergeable(rq
))
385 printk(KERN_ERR
"%s: bad insertion point %d\n",
386 __FUNCTION__
, where
);
390 if (blk_queue_plugged(q
)) {
391 int nrq
= q
->rq
.count
[READ
] + q
->rq
.count
[WRITE
]
394 if (nrq
>= q
->unplug_thresh
)
395 __generic_unplug_device(q
);
399 void elv_add_request(request_queue_t
*q
, struct request
*rq
, int where
,
404 spin_lock_irqsave(q
->queue_lock
, flags
);
405 __elv_add_request(q
, rq
, where
, plug
);
406 spin_unlock_irqrestore(q
->queue_lock
, flags
);
409 static inline struct request
*__elv_next_request(request_queue_t
*q
)
413 if (unlikely(list_empty(&q
->queue_head
) &&
414 !q
->elevator
->ops
->elevator_dispatch_fn(q
, 0)))
417 rq
= list_entry_rq(q
->queue_head
.next
);
420 * if this is a barrier write and the device has to issue a
421 * flush sequence to support it, check how far we are
423 if (blk_fs_request(rq
) && blk_barrier_rq(rq
)) {
424 BUG_ON(q
->ordered
== QUEUE_ORDERED_NONE
);
426 if (q
->ordered
== QUEUE_ORDERED_FLUSH
&&
427 !blk_barrier_preflush(rq
))
428 rq
= blk_start_pre_flush(q
, rq
);
434 struct request
*elv_next_request(request_queue_t
*q
)
439 while ((rq
= __elv_next_request(q
)) != NULL
) {
440 if (!(rq
->flags
& REQ_STARTED
)) {
441 elevator_t
*e
= q
->elevator
;
444 * This is the first time the device driver
445 * sees this request (possibly after
446 * requeueing). Notify IO scheduler.
448 if (blk_sorted_rq(rq
) &&
449 e
->ops
->elevator_activate_req_fn
)
450 e
->ops
->elevator_activate_req_fn(q
, rq
);
453 * just mark as started even if we don't start
454 * it, a request that has been delayed should
455 * not be passed by new incoming requests
457 rq
->flags
|= REQ_STARTED
;
460 if (!q
->boundary_rq
|| q
->boundary_rq
== rq
) {
461 q
->end_sector
= rq_end_sector(rq
);
462 q
->boundary_rq
= NULL
;
465 if ((rq
->flags
& REQ_DONTPREP
) || !q
->prep_rq_fn
)
468 ret
= q
->prep_rq_fn(q
, rq
);
469 if (ret
== BLKPREP_OK
) {
471 } else if (ret
== BLKPREP_DEFER
) {
473 * the request may have been (partially) prepped.
474 * we need to keep this request in the front to
475 * avoid resource deadlock. REQ_STARTED will
476 * prevent other fs requests from passing this one.
480 } else if (ret
== BLKPREP_KILL
) {
481 int nr_bytes
= rq
->hard_nr_sectors
<< 9;
484 nr_bytes
= rq
->data_len
;
486 blkdev_dequeue_request(rq
);
487 rq
->flags
|= REQ_QUIET
;
488 end_that_request_chunk(rq
, 0, nr_bytes
);
489 end_that_request_last(rq
);
491 printk(KERN_ERR
"%s: bad return=%d\n", __FUNCTION__
,
500 void elv_dequeue_request(request_queue_t
*q
, struct request
*rq
)
502 BUG_ON(list_empty(&rq
->queuelist
));
504 list_del_init(&rq
->queuelist
);
507 * the time frame between a request being removed from the lists
508 * and to it is freed is accounted as io that is in progress at
511 if (blk_account_rq(rq
))
515 int elv_queue_empty(request_queue_t
*q
)
517 elevator_t
*e
= q
->elevator
;
519 if (!list_empty(&q
->queue_head
))
522 if (e
->ops
->elevator_queue_empty_fn
)
523 return e
->ops
->elevator_queue_empty_fn(q
);
528 struct request
*elv_latter_request(request_queue_t
*q
, struct request
*rq
)
530 struct list_head
*next
;
532 elevator_t
*e
= q
->elevator
;
534 if (e
->ops
->elevator_latter_req_fn
)
535 return e
->ops
->elevator_latter_req_fn(q
, rq
);
537 next
= rq
->queuelist
.next
;
538 if (next
!= &q
->queue_head
&& next
!= &rq
->queuelist
)
539 return list_entry_rq(next
);
544 struct request
*elv_former_request(request_queue_t
*q
, struct request
*rq
)
546 struct list_head
*prev
;
548 elevator_t
*e
= q
->elevator
;
550 if (e
->ops
->elevator_former_req_fn
)
551 return e
->ops
->elevator_former_req_fn(q
, rq
);
553 prev
= rq
->queuelist
.prev
;
554 if (prev
!= &q
->queue_head
&& prev
!= &rq
->queuelist
)
555 return list_entry_rq(prev
);
560 int elv_set_request(request_queue_t
*q
, struct request
*rq
, struct bio
*bio
,
563 elevator_t
*e
= q
->elevator
;
565 if (e
->ops
->elevator_set_req_fn
)
566 return e
->ops
->elevator_set_req_fn(q
, rq
, bio
, gfp_mask
);
568 rq
->elevator_private
= NULL
;
572 void elv_put_request(request_queue_t
*q
, struct request
*rq
)
574 elevator_t
*e
= q
->elevator
;
576 if (e
->ops
->elevator_put_req_fn
)
577 e
->ops
->elevator_put_req_fn(q
, rq
);
580 int elv_may_queue(request_queue_t
*q
, int rw
, struct bio
*bio
)
582 elevator_t
*e
= q
->elevator
;
584 if (e
->ops
->elevator_may_queue_fn
)
585 return e
->ops
->elevator_may_queue_fn(q
, rw
, bio
);
587 return ELV_MQUEUE_MAY
;
590 void elv_completed_request(request_queue_t
*q
, struct request
*rq
)
592 elevator_t
*e
= q
->elevator
;
595 * request is released from the driver, io must be done
597 if (blk_account_rq(rq
)) {
599 if (blk_sorted_rq(rq
) && e
->ops
->elevator_completed_req_fn
)
600 e
->ops
->elevator_completed_req_fn(q
, rq
);
604 int elv_register_queue(struct request_queue
*q
)
606 elevator_t
*e
= q
->elevator
;
608 e
->kobj
.parent
= kobject_get(&q
->kobj
);
612 snprintf(e
->kobj
.name
, KOBJ_NAME_LEN
, "%s", "iosched");
613 e
->kobj
.ktype
= e
->elevator_type
->elevator_ktype
;
615 return kobject_register(&e
->kobj
);
618 void elv_unregister_queue(struct request_queue
*q
)
621 elevator_t
*e
= q
->elevator
;
622 kobject_unregister(&e
->kobj
);
623 kobject_put(&q
->kobj
);
627 int elv_register(struct elevator_type
*e
)
629 spin_lock_irq(&elv_list_lock
);
630 if (elevator_find(e
->elevator_name
))
632 list_add_tail(&e
->list
, &elv_list
);
633 spin_unlock_irq(&elv_list_lock
);
635 printk(KERN_INFO
"io scheduler %s registered", e
->elevator_name
);
636 if (!strcmp(e
->elevator_name
, chosen_elevator
))
637 printk(" (default)");
641 EXPORT_SYMBOL_GPL(elv_register
);
643 void elv_unregister(struct elevator_type
*e
)
645 spin_lock_irq(&elv_list_lock
);
646 list_del_init(&e
->list
);
647 spin_unlock_irq(&elv_list_lock
);
649 EXPORT_SYMBOL_GPL(elv_unregister
);
652 * switch to new_e io scheduler. be careful not to introduce deadlocks -
653 * we don't free the old io scheduler, before we have allocated what we
654 * need for the new one. this way we have a chance of going back to the old
655 * one, if the new one fails init for some reason.
657 static void elevator_switch(request_queue_t
*q
, struct elevator_type
*new_e
)
659 elevator_t
*old_elevator
, *e
;
662 * Allocate new elevator
664 e
= kmalloc(sizeof(elevator_t
), GFP_KERNEL
);
669 * Turn on BYPASS and drain all requests w/ elevator private data
671 spin_lock_irq(q
->queue_lock
);
673 set_bit(QUEUE_FLAG_ELVSWITCH
, &q
->queue_flags
);
675 while (q
->elevator
->ops
->elevator_dispatch_fn(q
, 1))
678 while (q
->rq
.elvpriv
) {
679 spin_unlock_irq(q
->queue_lock
);
681 spin_lock_irq(q
->queue_lock
);
684 spin_unlock_irq(q
->queue_lock
);
687 * unregister old elevator data
689 elv_unregister_queue(q
);
690 old_elevator
= q
->elevator
;
693 * attach and start new elevator
695 if (elevator_attach(q
, new_e
, e
))
698 if (elv_register_queue(q
))
702 * finally exit old elevator and turn off BYPASS.
704 elevator_exit(old_elevator
);
705 clear_bit(QUEUE_FLAG_ELVSWITCH
, &q
->queue_flags
);
710 * switch failed, exit the new io scheduler and reattach the old
711 * one again (along with re-adding the sysfs dir)
716 q
->elevator
= old_elevator
;
717 elv_register_queue(q
);
718 clear_bit(QUEUE_FLAG_ELVSWITCH
, &q
->queue_flags
);
722 printk(KERN_ERR
"elevator: switch to %s failed\n",new_e
->elevator_name
);
725 ssize_t
elv_iosched_store(request_queue_t
*q
, const char *name
, size_t count
)
727 char elevator_name
[ELV_NAME_MAX
];
728 struct elevator_type
*e
;
730 memset(elevator_name
, 0, sizeof(elevator_name
));
731 strncpy(elevator_name
, name
, sizeof(elevator_name
));
733 if (elevator_name
[strlen(elevator_name
) - 1] == '\n')
734 elevator_name
[strlen(elevator_name
) - 1] = '\0';
736 e
= elevator_get(elevator_name
);
738 printk(KERN_ERR
"elevator: type %s not found\n", elevator_name
);
742 if (!strcmp(elevator_name
, q
->elevator
->elevator_type
->elevator_name
))
745 elevator_switch(q
, e
);
749 ssize_t
elv_iosched_show(request_queue_t
*q
, char *name
)
751 elevator_t
*e
= q
->elevator
;
752 struct elevator_type
*elv
= e
->elevator_type
;
753 struct list_head
*entry
;
756 spin_lock_irq(q
->queue_lock
);
757 list_for_each(entry
, &elv_list
) {
758 struct elevator_type
*__e
;
760 __e
= list_entry(entry
, struct elevator_type
, list
);
761 if (!strcmp(elv
->elevator_name
, __e
->elevator_name
))
762 len
+= sprintf(name
+len
, "[%s] ", elv
->elevator_name
);
764 len
+= sprintf(name
+len
, "%s ", __e
->elevator_name
);
766 spin_unlock_irq(q
->queue_lock
);
768 len
+= sprintf(len
+name
, "\n");
772 EXPORT_SYMBOL(elv_dispatch_sort
);
773 EXPORT_SYMBOL(elv_add_request
);
774 EXPORT_SYMBOL(__elv_add_request
);
775 EXPORT_SYMBOL(elv_requeue_request
);
776 EXPORT_SYMBOL(elv_next_request
);
777 EXPORT_SYMBOL(elv_dequeue_request
);
778 EXPORT_SYMBOL(elv_queue_empty
);
779 EXPORT_SYMBOL(elv_completed_request
);
780 EXPORT_SYMBOL(elevator_exit
);
781 EXPORT_SYMBOL(elevator_init
);