2 * Block device elevator/IO-scheduler.
4 * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
6 * 30042000 Jens Axboe <axboe@suse.de> :
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/config.h>
31 #include <linux/module.h>
32 #include <linux/slab.h>
33 #include <linux/init.h>
34 #include <linux/compiler.h>
35 #include <linux/delay.h>
37 #include <asm/uaccess.h>
39 static DEFINE_SPINLOCK(elv_list_lock
);
40 static LIST_HEAD(elv_list
);
43 * can we safely merge with this request?
45 inline int elv_rq_merge_ok(struct request
*rq
, struct bio
*bio
)
47 if (!rq_mergeable(rq
))
51 * different data direction or already started, don't merge
53 if (bio_data_dir(bio
) != rq_data_dir(rq
))
57 * same device and no special stuff set, merge is ok
59 if (rq
->rq_disk
== bio
->bi_bdev
->bd_disk
&&
60 !rq
->waiting
&& !rq
->special
)
65 EXPORT_SYMBOL(elv_rq_merge_ok
);
67 inline int elv_try_merge(struct request
*__rq
, struct bio
*bio
)
69 int ret
= ELEVATOR_NO_MERGE
;
72 * we can merge and sequence is ok, check if it's possible
74 if (elv_rq_merge_ok(__rq
, bio
)) {
75 if (__rq
->sector
+ __rq
->nr_sectors
== bio
->bi_sector
)
76 ret
= ELEVATOR_BACK_MERGE
;
77 else if (__rq
->sector
- bio_sectors(bio
) == bio
->bi_sector
)
78 ret
= ELEVATOR_FRONT_MERGE
;
83 EXPORT_SYMBOL(elv_try_merge
);
85 static struct elevator_type
*elevator_find(const char *name
)
87 struct elevator_type
*e
= NULL
;
88 struct list_head
*entry
;
90 list_for_each(entry
, &elv_list
) {
91 struct elevator_type
*__e
;
93 __e
= list_entry(entry
, struct elevator_type
, list
);
95 if (!strcmp(__e
->elevator_name
, name
)) {
104 static void elevator_put(struct elevator_type
*e
)
106 module_put(e
->elevator_owner
);
109 static struct elevator_type
*elevator_get(const char *name
)
111 struct elevator_type
*e
;
113 spin_lock_irq(&elv_list_lock
);
115 e
= elevator_find(name
);
116 if (e
&& !try_module_get(e
->elevator_owner
))
119 spin_unlock_irq(&elv_list_lock
);
124 static int elevator_attach(request_queue_t
*q
, struct elevator_type
*e
,
125 struct elevator_queue
*eq
)
129 memset(eq
, 0, sizeof(*eq
));
131 eq
->elevator_type
= e
;
135 if (eq
->ops
->elevator_init_fn
)
136 ret
= eq
->ops
->elevator_init_fn(q
, eq
);
141 static char chosen_elevator
[16];
143 static void elevator_setup_default(void)
145 struct elevator_type
*e
;
148 * If default has not been set, use the compiled-in selection.
150 if (!chosen_elevator
[0])
151 strcpy(chosen_elevator
, CONFIG_DEFAULT_IOSCHED
);
154 * If the given scheduler is not available, fall back to no-op.
156 if ((e
= elevator_find(chosen_elevator
)))
159 strcpy(chosen_elevator
, "noop");
162 static int __init
elevator_setup(char *str
)
164 strncpy(chosen_elevator
, str
, sizeof(chosen_elevator
) - 1);
168 __setup("elevator=", elevator_setup
);
170 int elevator_init(request_queue_t
*q
, char *name
)
172 struct elevator_type
*e
= NULL
;
173 struct elevator_queue
*eq
;
176 INIT_LIST_HEAD(&q
->queue_head
);
177 q
->last_merge
= NULL
;
179 q
->boundary_rq
= NULL
;
181 elevator_setup_default();
184 name
= chosen_elevator
;
186 e
= elevator_get(name
);
190 eq
= kmalloc(sizeof(struct elevator_queue
), GFP_KERNEL
);
196 ret
= elevator_attach(q
, e
, eq
);
205 void elevator_exit(elevator_t
*e
)
207 if (e
->ops
->elevator_exit_fn
)
208 e
->ops
->elevator_exit_fn(e
);
210 elevator_put(e
->elevator_type
);
211 e
->elevator_type
= NULL
;
216 * Insert rq into dispatch queue of q. Queue lock must be held on
217 * entry. If sort != 0, rq is sort-inserted; otherwise, rq will be
218 * appended to the dispatch queue. To be used by specific elevators.
220 void elv_dispatch_sort(request_queue_t
*q
, struct request
*rq
)
223 struct list_head
*entry
;
225 if (q
->last_merge
== rq
)
226 q
->last_merge
= NULL
;
229 boundary
= q
->end_sector
;
231 list_for_each_prev(entry
, &q
->queue_head
) {
232 struct request
*pos
= list_entry_rq(entry
);
234 if (pos
->flags
& (REQ_SOFTBARRIER
|REQ_HARDBARRIER
|REQ_STARTED
))
236 if (rq
->sector
>= boundary
) {
237 if (pos
->sector
< boundary
)
240 if (pos
->sector
>= boundary
)
243 if (rq
->sector
>= pos
->sector
)
247 list_add(&rq
->queuelist
, entry
);
250 int elv_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
252 elevator_t
*e
= q
->elevator
;
256 ret
= elv_try_merge(q
->last_merge
, bio
);
257 if (ret
!= ELEVATOR_NO_MERGE
) {
258 *req
= q
->last_merge
;
263 if (e
->ops
->elevator_merge_fn
)
264 return e
->ops
->elevator_merge_fn(q
, req
, bio
);
266 return ELEVATOR_NO_MERGE
;
269 void elv_merged_request(request_queue_t
*q
, struct request
*rq
)
271 elevator_t
*e
= q
->elevator
;
273 if (e
->ops
->elevator_merged_fn
)
274 e
->ops
->elevator_merged_fn(q
, rq
);
279 void elv_merge_requests(request_queue_t
*q
, struct request
*rq
,
280 struct request
*next
)
282 elevator_t
*e
= q
->elevator
;
284 if (e
->ops
->elevator_merge_req_fn
)
285 e
->ops
->elevator_merge_req_fn(q
, rq
, next
);
291 void elv_requeue_request(request_queue_t
*q
, struct request
*rq
)
293 elevator_t
*e
= q
->elevator
;
296 * it already went through dequeue, we need to decrement the
297 * in_flight count again
299 if (blk_account_rq(rq
)) {
301 if (blk_sorted_rq(rq
) && e
->ops
->elevator_deactivate_req_fn
)
302 e
->ops
->elevator_deactivate_req_fn(q
, rq
);
305 rq
->flags
&= ~REQ_STARTED
;
308 * if this is the flush, requeue the original instead and drop the flush
310 if (rq
->flags
& REQ_BAR_FLUSH
) {
311 clear_bit(QUEUE_FLAG_FLUSH
, &q
->queue_flags
);
312 rq
= rq
->end_io_data
;
315 __elv_add_request(q
, rq
, ELEVATOR_INSERT_FRONT
, 0);
318 static void elv_drain_elevator(request_queue_t
*q
)
321 while (q
->elevator
->ops
->elevator_dispatch_fn(q
, 1))
323 if (q
->nr_sorted
== 0)
325 if (printed
++ < 10) {
326 printk(KERN_ERR
"%s: forced dispatching is broken "
327 "(nr_sorted=%u), please report this\n",
328 q
->elevator
->elevator_type
->elevator_name
, q
->nr_sorted
);
332 void __elv_add_request(request_queue_t
*q
, struct request
*rq
, int where
,
335 if (rq
->flags
& (REQ_SOFTBARRIER
| REQ_HARDBARRIER
)) {
337 * barriers implicitly indicate back insertion
339 if (where
== ELEVATOR_INSERT_SORT
)
340 where
= ELEVATOR_INSERT_BACK
;
343 * this request is scheduling boundary, update end_sector
345 if (blk_fs_request(rq
)) {
346 q
->end_sector
= rq_end_sector(rq
);
349 } else if (!(rq
->flags
& REQ_ELVPRIV
) && where
== ELEVATOR_INSERT_SORT
)
350 where
= ELEVATOR_INSERT_BACK
;
358 case ELEVATOR_INSERT_FRONT
:
359 rq
->flags
|= REQ_SOFTBARRIER
;
361 list_add(&rq
->queuelist
, &q
->queue_head
);
364 case ELEVATOR_INSERT_BACK
:
365 rq
->flags
|= REQ_SOFTBARRIER
;
366 elv_drain_elevator(q
);
367 list_add_tail(&rq
->queuelist
, &q
->queue_head
);
369 * We kick the queue here for the following reasons.
370 * - The elevator might have returned NULL previously
371 * to delay requests and returned them now. As the
372 * queue wasn't empty before this request, ll_rw_blk
373 * won't run the queue on return, resulting in hang.
374 * - Usually, back inserted requests won't be merged
375 * with anything. There's no point in delaying queue
382 case ELEVATOR_INSERT_SORT
:
383 BUG_ON(!blk_fs_request(rq
));
384 rq
->flags
|= REQ_SORTED
;
386 if (q
->last_merge
== NULL
&& rq_mergeable(rq
))
389 * Some ioscheds (cfq) run q->request_fn directly, so
390 * rq cannot be accessed after calling
391 * elevator_add_req_fn.
393 q
->elevator
->ops
->elevator_add_req_fn(q
, rq
);
397 printk(KERN_ERR
"%s: bad insertion point %d\n",
398 __FUNCTION__
, where
);
402 if (blk_queue_plugged(q
)) {
403 int nrq
= q
->rq
.count
[READ
] + q
->rq
.count
[WRITE
]
406 if (nrq
>= q
->unplug_thresh
)
407 __generic_unplug_device(q
);
411 void elv_add_request(request_queue_t
*q
, struct request
*rq
, int where
,
416 spin_lock_irqsave(q
->queue_lock
, flags
);
417 __elv_add_request(q
, rq
, where
, plug
);
418 spin_unlock_irqrestore(q
->queue_lock
, flags
);
421 static inline struct request
*__elv_next_request(request_queue_t
*q
)
425 if (unlikely(list_empty(&q
->queue_head
) &&
426 !q
->elevator
->ops
->elevator_dispatch_fn(q
, 0)))
429 rq
= list_entry_rq(q
->queue_head
.next
);
432 * if this is a barrier write and the device has to issue a
433 * flush sequence to support it, check how far we are
435 if (blk_fs_request(rq
) && blk_barrier_rq(rq
)) {
436 BUG_ON(q
->ordered
== QUEUE_ORDERED_NONE
);
438 if (q
->ordered
== QUEUE_ORDERED_FLUSH
&&
439 !blk_barrier_preflush(rq
))
440 rq
= blk_start_pre_flush(q
, rq
);
446 struct request
*elv_next_request(request_queue_t
*q
)
451 while ((rq
= __elv_next_request(q
)) != NULL
) {
452 if (!(rq
->flags
& REQ_STARTED
)) {
453 elevator_t
*e
= q
->elevator
;
456 * This is the first time the device driver
457 * sees this request (possibly after
458 * requeueing). Notify IO scheduler.
460 if (blk_sorted_rq(rq
) &&
461 e
->ops
->elevator_activate_req_fn
)
462 e
->ops
->elevator_activate_req_fn(q
, rq
);
465 * just mark as started even if we don't start
466 * it, a request that has been delayed should
467 * not be passed by new incoming requests
469 rq
->flags
|= REQ_STARTED
;
472 if (!q
->boundary_rq
|| q
->boundary_rq
== rq
) {
473 q
->end_sector
= rq_end_sector(rq
);
474 q
->boundary_rq
= NULL
;
477 if ((rq
->flags
& REQ_DONTPREP
) || !q
->prep_rq_fn
)
480 ret
= q
->prep_rq_fn(q
, rq
);
481 if (ret
== BLKPREP_OK
) {
483 } else if (ret
== BLKPREP_DEFER
) {
485 * the request may have been (partially) prepped.
486 * we need to keep this request in the front to
487 * avoid resource deadlock. REQ_STARTED will
488 * prevent other fs requests from passing this one.
492 } else if (ret
== BLKPREP_KILL
) {
493 int nr_bytes
= rq
->hard_nr_sectors
<< 9;
496 nr_bytes
= rq
->data_len
;
498 blkdev_dequeue_request(rq
);
499 rq
->flags
|= REQ_QUIET
;
500 end_that_request_chunk(rq
, 0, nr_bytes
);
501 end_that_request_last(rq
);
503 printk(KERN_ERR
"%s: bad return=%d\n", __FUNCTION__
,
512 void elv_dequeue_request(request_queue_t
*q
, struct request
*rq
)
514 BUG_ON(list_empty(&rq
->queuelist
));
516 list_del_init(&rq
->queuelist
);
519 * the time frame between a request being removed from the lists
520 * and to it is freed is accounted as io that is in progress at
523 if (blk_account_rq(rq
))
527 int elv_queue_empty(request_queue_t
*q
)
529 elevator_t
*e
= q
->elevator
;
531 if (!list_empty(&q
->queue_head
))
534 if (e
->ops
->elevator_queue_empty_fn
)
535 return e
->ops
->elevator_queue_empty_fn(q
);
540 struct request
*elv_latter_request(request_queue_t
*q
, struct request
*rq
)
542 elevator_t
*e
= q
->elevator
;
544 if (e
->ops
->elevator_latter_req_fn
)
545 return e
->ops
->elevator_latter_req_fn(q
, rq
);
549 struct request
*elv_former_request(request_queue_t
*q
, struct request
*rq
)
551 elevator_t
*e
= q
->elevator
;
553 if (e
->ops
->elevator_former_req_fn
)
554 return e
->ops
->elevator_former_req_fn(q
, rq
);
558 int elv_set_request(request_queue_t
*q
, struct request
*rq
, struct bio
*bio
,
561 elevator_t
*e
= q
->elevator
;
563 if (e
->ops
->elevator_set_req_fn
)
564 return e
->ops
->elevator_set_req_fn(q
, rq
, bio
, gfp_mask
);
566 rq
->elevator_private
= NULL
;
570 void elv_put_request(request_queue_t
*q
, struct request
*rq
)
572 elevator_t
*e
= q
->elevator
;
574 if (e
->ops
->elevator_put_req_fn
)
575 e
->ops
->elevator_put_req_fn(q
, rq
);
578 int elv_may_queue(request_queue_t
*q
, int rw
, struct bio
*bio
)
580 elevator_t
*e
= q
->elevator
;
582 if (e
->ops
->elevator_may_queue_fn
)
583 return e
->ops
->elevator_may_queue_fn(q
, rw
, bio
);
585 return ELV_MQUEUE_MAY
;
588 void elv_completed_request(request_queue_t
*q
, struct request
*rq
)
590 elevator_t
*e
= q
->elevator
;
593 * request is released from the driver, io must be done
595 if (blk_account_rq(rq
)) {
597 if (blk_sorted_rq(rq
) && e
->ops
->elevator_completed_req_fn
)
598 e
->ops
->elevator_completed_req_fn(q
, rq
);
602 int elv_register_queue(struct request_queue
*q
)
604 elevator_t
*e
= q
->elevator
;
606 e
->kobj
.parent
= kobject_get(&q
->kobj
);
610 snprintf(e
->kobj
.name
, KOBJ_NAME_LEN
, "%s", "iosched");
611 e
->kobj
.ktype
= e
->elevator_type
->elevator_ktype
;
613 return kobject_register(&e
->kobj
);
616 void elv_unregister_queue(struct request_queue
*q
)
619 elevator_t
*e
= q
->elevator
;
620 kobject_unregister(&e
->kobj
);
621 kobject_put(&q
->kobj
);
625 int elv_register(struct elevator_type
*e
)
627 spin_lock_irq(&elv_list_lock
);
628 if (elevator_find(e
->elevator_name
))
630 list_add_tail(&e
->list
, &elv_list
);
631 spin_unlock_irq(&elv_list_lock
);
633 printk(KERN_INFO
"io scheduler %s registered", e
->elevator_name
);
634 if (!strcmp(e
->elevator_name
, chosen_elevator
))
635 printk(" (default)");
639 EXPORT_SYMBOL_GPL(elv_register
);
641 void elv_unregister(struct elevator_type
*e
)
643 struct task_struct
*g
, *p
;
646 * Iterate every thread in the process to remove the io contexts.
648 read_lock(&tasklist_lock
);
649 do_each_thread(g
, p
) {
650 struct io_context
*ioc
= p
->io_context
;
651 if (ioc
&& ioc
->cic
) {
652 ioc
->cic
->exit(ioc
->cic
);
653 ioc
->cic
->dtor(ioc
->cic
);
656 if (ioc
&& ioc
->aic
) {
657 ioc
->aic
->exit(ioc
->aic
);
658 ioc
->aic
->dtor(ioc
->aic
);
661 } while_each_thread(g
, p
);
662 read_unlock(&tasklist_lock
);
664 spin_lock_irq(&elv_list_lock
);
665 list_del_init(&e
->list
);
666 spin_unlock_irq(&elv_list_lock
);
668 EXPORT_SYMBOL_GPL(elv_unregister
);
671 * switch to new_e io scheduler. be careful not to introduce deadlocks -
672 * we don't free the old io scheduler, before we have allocated what we
673 * need for the new one. this way we have a chance of going back to the old
674 * one, if the new one fails init for some reason.
676 static void elevator_switch(request_queue_t
*q
, struct elevator_type
*new_e
)
678 elevator_t
*old_elevator
, *e
;
681 * Allocate new elevator
683 e
= kmalloc(sizeof(elevator_t
), GFP_KERNEL
);
688 * Turn on BYPASS and drain all requests w/ elevator private data
690 spin_lock_irq(q
->queue_lock
);
692 set_bit(QUEUE_FLAG_ELVSWITCH
, &q
->queue_flags
);
694 elv_drain_elevator(q
);
696 while (q
->rq
.elvpriv
) {
699 spin_unlock_irq(q
->queue_lock
);
701 spin_lock_irq(q
->queue_lock
);
702 elv_drain_elevator(q
);
705 spin_unlock_irq(q
->queue_lock
);
708 * unregister old elevator data
710 elv_unregister_queue(q
);
711 old_elevator
= q
->elevator
;
714 * attach and start new elevator
716 if (elevator_attach(q
, new_e
, e
))
719 if (elv_register_queue(q
))
723 * finally exit old elevator and turn off BYPASS.
725 elevator_exit(old_elevator
);
726 clear_bit(QUEUE_FLAG_ELVSWITCH
, &q
->queue_flags
);
731 * switch failed, exit the new io scheduler and reattach the old
732 * one again (along with re-adding the sysfs dir)
737 q
->elevator
= old_elevator
;
738 elv_register_queue(q
);
739 clear_bit(QUEUE_FLAG_ELVSWITCH
, &q
->queue_flags
);
743 printk(KERN_ERR
"elevator: switch to %s failed\n",new_e
->elevator_name
);
746 ssize_t
elv_iosched_store(request_queue_t
*q
, const char *name
, size_t count
)
748 char elevator_name
[ELV_NAME_MAX
];
750 struct elevator_type
*e
;
752 elevator_name
[sizeof(elevator_name
) - 1] = '\0';
753 strncpy(elevator_name
, name
, sizeof(elevator_name
) - 1);
754 len
= strlen(elevator_name
);
756 if (len
&& elevator_name
[len
- 1] == '\n')
757 elevator_name
[len
- 1] = '\0';
759 e
= elevator_get(elevator_name
);
761 printk(KERN_ERR
"elevator: type %s not found\n", elevator_name
);
765 if (!strcmp(elevator_name
, q
->elevator
->elevator_type
->elevator_name
)) {
770 elevator_switch(q
, e
);
774 ssize_t
elv_iosched_show(request_queue_t
*q
, char *name
)
776 elevator_t
*e
= q
->elevator
;
777 struct elevator_type
*elv
= e
->elevator_type
;
778 struct list_head
*entry
;
781 spin_lock_irq(q
->queue_lock
);
782 list_for_each(entry
, &elv_list
) {
783 struct elevator_type
*__e
;
785 __e
= list_entry(entry
, struct elevator_type
, list
);
786 if (!strcmp(elv
->elevator_name
, __e
->elevator_name
))
787 len
+= sprintf(name
+len
, "[%s] ", elv
->elevator_name
);
789 len
+= sprintf(name
+len
, "%s ", __e
->elevator_name
);
791 spin_unlock_irq(q
->queue_lock
);
793 len
+= sprintf(len
+name
, "\n");
797 EXPORT_SYMBOL(elv_dispatch_sort
);
798 EXPORT_SYMBOL(elv_add_request
);
799 EXPORT_SYMBOL(__elv_add_request
);
800 EXPORT_SYMBOL(elv_requeue_request
);
801 EXPORT_SYMBOL(elv_next_request
);
802 EXPORT_SYMBOL(elv_dequeue_request
);
803 EXPORT_SYMBOL(elv_queue_empty
);
804 EXPORT_SYMBOL(elv_completed_request
);
805 EXPORT_SYMBOL(elevator_exit
);
806 EXPORT_SYMBOL(elevator_init
);