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>
38 #include <asm/uaccess.h>
40 static DEFINE_SPINLOCK(elv_list_lock
);
41 static LIST_HEAD(elv_list
);
44 * can we safely merge with this request?
46 inline int elv_rq_merge_ok(struct request
*rq
, struct bio
*bio
)
48 if (!rq_mergeable(rq
))
52 * different data direction or already started, don't merge
54 if (bio_data_dir(bio
) != rq_data_dir(rq
))
58 * same device and no special stuff set, merge is ok
60 if (rq
->rq_disk
== bio
->bi_bdev
->bd_disk
&&
61 !rq
->waiting
&& !rq
->special
)
66 EXPORT_SYMBOL(elv_rq_merge_ok
);
68 inline int elv_try_merge(struct request
*__rq
, struct bio
*bio
)
70 int ret
= ELEVATOR_NO_MERGE
;
73 * we can merge and sequence is ok, check if it's possible
75 if (elv_rq_merge_ok(__rq
, bio
)) {
76 if (__rq
->sector
+ __rq
->nr_sectors
== bio
->bi_sector
)
77 ret
= ELEVATOR_BACK_MERGE
;
78 else if (__rq
->sector
- bio_sectors(bio
) == bio
->bi_sector
)
79 ret
= ELEVATOR_FRONT_MERGE
;
84 EXPORT_SYMBOL(elv_try_merge
);
86 static struct elevator_type
*elevator_find(const char *name
)
88 struct elevator_type
*e
= NULL
;
89 struct list_head
*entry
;
91 list_for_each(entry
, &elv_list
) {
92 struct elevator_type
*__e
;
94 __e
= list_entry(entry
, struct elevator_type
, list
);
96 if (!strcmp(__e
->elevator_name
, name
)) {
105 static void elevator_put(struct elevator_type
*e
)
107 module_put(e
->elevator_owner
);
110 static struct elevator_type
*elevator_get(const char *name
)
112 struct elevator_type
*e
;
114 spin_lock_irq(&elv_list_lock
);
116 e
= elevator_find(name
);
117 if (e
&& !try_module_get(e
->elevator_owner
))
120 spin_unlock_irq(&elv_list_lock
);
125 static int elevator_attach(request_queue_t
*q
, struct elevator_type
*e
,
126 struct elevator_queue
*eq
)
130 memset(eq
, 0, sizeof(*eq
));
132 eq
->elevator_type
= e
;
134 INIT_LIST_HEAD(&q
->queue_head
);
135 q
->last_merge
= NULL
;
138 q
->boundary_rq
= NULL
;
140 if (eq
->ops
->elevator_init_fn
)
141 ret
= eq
->ops
->elevator_init_fn(q
, eq
);
146 static char chosen_elevator
[16];
148 static void elevator_setup_default(void)
150 struct elevator_type
*e
;
153 * check if default is set and exists
155 if (chosen_elevator
[0] && (e
= elevator_get(chosen_elevator
))) {
160 #if defined(CONFIG_IOSCHED_AS)
161 strcpy(chosen_elevator
, "anticipatory");
162 #elif defined(CONFIG_IOSCHED_DEADLINE)
163 strcpy(chosen_elevator
, "deadline");
164 #elif defined(CONFIG_IOSCHED_CFQ)
165 strcpy(chosen_elevator
, "cfq");
166 #elif defined(CONFIG_IOSCHED_NOOP)
167 strcpy(chosen_elevator
, "noop");
169 #error "You must build at least 1 IO scheduler into the kernel"
173 static int __init
elevator_setup(char *str
)
175 strncpy(chosen_elevator
, str
, sizeof(chosen_elevator
) - 1);
179 __setup("elevator=", elevator_setup
);
181 int elevator_init(request_queue_t
*q
, char *name
)
183 struct elevator_type
*e
= NULL
;
184 struct elevator_queue
*eq
;
187 elevator_setup_default();
190 name
= chosen_elevator
;
192 e
= elevator_get(name
);
196 eq
= kmalloc(sizeof(struct elevator_queue
), GFP_KERNEL
);
198 elevator_put(e
->elevator_type
);
202 ret
= elevator_attach(q
, e
, eq
);
205 elevator_put(e
->elevator_type
);
211 void elevator_exit(elevator_t
*e
)
213 if (e
->ops
->elevator_exit_fn
)
214 e
->ops
->elevator_exit_fn(e
);
216 elevator_put(e
->elevator_type
);
217 e
->elevator_type
= NULL
;
222 * Insert rq into dispatch queue of q. Queue lock must be held on
223 * entry. If sort != 0, rq is sort-inserted; otherwise, rq will be
224 * appended to the dispatch queue. To be used by specific elevators.
226 void elv_dispatch_sort(request_queue_t
*q
, struct request
*rq
)
229 struct list_head
*entry
;
231 if (q
->last_merge
== rq
)
232 q
->last_merge
= NULL
;
234 boundary
= q
->end_sector
;
236 list_for_each_prev(entry
, &q
->queue_head
) {
237 struct request
*pos
= list_entry_rq(entry
);
239 if (pos
->flags
& (REQ_SOFTBARRIER
|REQ_HARDBARRIER
|REQ_STARTED
))
241 if (rq
->sector
>= boundary
) {
242 if (pos
->sector
< boundary
)
245 if (pos
->sector
>= boundary
)
248 if (rq
->sector
>= pos
->sector
)
252 list_add(&rq
->queuelist
, entry
);
255 int elv_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
257 elevator_t
*e
= q
->elevator
;
261 ret
= elv_try_merge(q
->last_merge
, bio
);
262 if (ret
!= ELEVATOR_NO_MERGE
) {
263 *req
= q
->last_merge
;
268 if (e
->ops
->elevator_merge_fn
)
269 return e
->ops
->elevator_merge_fn(q
, req
, bio
);
271 return ELEVATOR_NO_MERGE
;
274 void elv_merged_request(request_queue_t
*q
, struct request
*rq
)
276 elevator_t
*e
= q
->elevator
;
278 if (e
->ops
->elevator_merged_fn
)
279 e
->ops
->elevator_merged_fn(q
, rq
);
284 void elv_merge_requests(request_queue_t
*q
, struct request
*rq
,
285 struct request
*next
)
287 elevator_t
*e
= q
->elevator
;
289 if (e
->ops
->elevator_merge_req_fn
)
290 e
->ops
->elevator_merge_req_fn(q
, rq
, next
);
295 void elv_requeue_request(request_queue_t
*q
, struct request
*rq
)
297 elevator_t
*e
= q
->elevator
;
300 * it already went through dequeue, we need to decrement the
301 * in_flight count again
303 if (blk_account_rq(rq
)) {
305 if (blk_sorted_rq(rq
) && e
->ops
->elevator_deactivate_req_fn
)
306 e
->ops
->elevator_deactivate_req_fn(q
, rq
);
309 rq
->flags
&= ~REQ_STARTED
;
312 * if this is the flush, requeue the original instead and drop the flush
314 if (rq
->flags
& REQ_BAR_FLUSH
) {
315 clear_bit(QUEUE_FLAG_FLUSH
, &q
->queue_flags
);
316 rq
= rq
->end_io_data
;
319 __elv_add_request(q
, rq
, ELEVATOR_INSERT_FRONT
, 0);
322 void __elv_add_request(request_queue_t
*q
, struct request
*rq
, int where
,
325 if (rq
->flags
& (REQ_SOFTBARRIER
| REQ_HARDBARRIER
)) {
327 * barriers implicitly indicate back insertion
329 if (where
== ELEVATOR_INSERT_SORT
)
330 where
= ELEVATOR_INSERT_BACK
;
333 * this request is scheduling boundary, update end_sector
335 if (blk_fs_request(rq
)) {
336 q
->end_sector
= rq_end_sector(rq
);
346 if (unlikely(test_bit(QUEUE_FLAG_DRAIN
, &q
->queue_flags
))) {
348 * if drain is set, store the request "locally". when the drain
349 * is finished, the requests will be handed ordered to the io
352 list_add_tail(&rq
->queuelist
, &q
->drain_list
);
357 case ELEVATOR_INSERT_FRONT
:
358 rq
->flags
|= REQ_SOFTBARRIER
;
360 list_add(&rq
->queuelist
, &q
->queue_head
);
363 case ELEVATOR_INSERT_BACK
:
364 rq
->flags
|= REQ_SOFTBARRIER
;
366 while (q
->elevator
->ops
->elevator_dispatch_fn(q
, 1))
368 list_add_tail(&rq
->queuelist
, &q
->queue_head
);
370 * We kick the queue here for the following reasons.
371 * - The elevator might have returned NULL previously
372 * to delay requests and returned them now. As the
373 * queue wasn't empty before this request, ll_rw_blk
374 * won't run the queue on return, resulting in hang.
375 * - Usually, back inserted requests won't be merged
376 * with anything. There's no point in delaying queue
383 case ELEVATOR_INSERT_SORT
:
384 BUG_ON(!blk_fs_request(rq
));
385 rq
->flags
|= REQ_SORTED
;
386 q
->elevator
->ops
->elevator_add_req_fn(q
, rq
);
387 if (q
->last_merge
== NULL
&& rq_mergeable(rq
))
392 printk(KERN_ERR
"%s: bad insertion point %d\n",
393 __FUNCTION__
, where
);
397 if (blk_queue_plugged(q
)) {
398 int nrq
= q
->rq
.count
[READ
] + q
->rq
.count
[WRITE
]
401 if (nrq
>= q
->unplug_thresh
)
402 __generic_unplug_device(q
);
406 void elv_add_request(request_queue_t
*q
, struct request
*rq
, int where
,
411 spin_lock_irqsave(q
->queue_lock
, flags
);
412 __elv_add_request(q
, rq
, where
, plug
);
413 spin_unlock_irqrestore(q
->queue_lock
, flags
);
416 static inline struct request
*__elv_next_request(request_queue_t
*q
)
420 if (unlikely(list_empty(&q
->queue_head
) &&
421 !q
->elevator
->ops
->elevator_dispatch_fn(q
, 0)))
424 rq
= list_entry_rq(q
->queue_head
.next
);
427 * if this is a barrier write and the device has to issue a
428 * flush sequence to support it, check how far we are
430 if (blk_fs_request(rq
) && blk_barrier_rq(rq
)) {
431 BUG_ON(q
->ordered
== QUEUE_ORDERED_NONE
);
433 if (q
->ordered
== QUEUE_ORDERED_FLUSH
&&
434 !blk_barrier_preflush(rq
))
435 rq
= blk_start_pre_flush(q
, rq
);
441 struct request
*elv_next_request(request_queue_t
*q
)
446 while ((rq
= __elv_next_request(q
)) != NULL
) {
447 if (!(rq
->flags
& REQ_STARTED
)) {
448 elevator_t
*e
= q
->elevator
;
451 * This is the first time the device driver
452 * sees this request (possibly after
453 * requeueing). Notify IO scheduler.
455 if (blk_sorted_rq(rq
) &&
456 e
->ops
->elevator_activate_req_fn
)
457 e
->ops
->elevator_activate_req_fn(q
, rq
);
460 * just mark as started even if we don't start
461 * it, a request that has been delayed should
462 * not be passed by new incoming requests
464 rq
->flags
|= REQ_STARTED
;
467 if (!q
->boundary_rq
|| q
->boundary_rq
== rq
) {
468 q
->end_sector
= rq_end_sector(rq
);
469 q
->boundary_rq
= NULL
;
472 if ((rq
->flags
& REQ_DONTPREP
) || !q
->prep_rq_fn
)
475 ret
= q
->prep_rq_fn(q
, rq
);
476 if (ret
== BLKPREP_OK
) {
478 } else if (ret
== BLKPREP_DEFER
) {
480 * the request may have been (partially) prepped.
481 * we need to keep this request in the front to
482 * avoid resource deadlock. REQ_STARTED will
483 * prevent other fs requests from passing this one.
487 } else if (ret
== BLKPREP_KILL
) {
488 int nr_bytes
= rq
->hard_nr_sectors
<< 9;
491 nr_bytes
= rq
->data_len
;
493 blkdev_dequeue_request(rq
);
494 rq
->flags
|= REQ_QUIET
;
495 end_that_request_chunk(rq
, 0, nr_bytes
);
496 end_that_request_last(rq
);
498 printk(KERN_ERR
"%s: bad return=%d\n", __FUNCTION__
,
507 void elv_dequeue_request(request_queue_t
*q
, struct request
*rq
)
509 BUG_ON(list_empty(&rq
->queuelist
));
511 list_del_init(&rq
->queuelist
);
514 * the time frame between a request being removed from the lists
515 * and to it is freed is accounted as io that is in progress at
518 if (blk_account_rq(rq
))
522 int elv_queue_empty(request_queue_t
*q
)
524 elevator_t
*e
= q
->elevator
;
526 if (!list_empty(&q
->queue_head
))
529 if (e
->ops
->elevator_queue_empty_fn
)
530 return e
->ops
->elevator_queue_empty_fn(q
);
535 struct request
*elv_latter_request(request_queue_t
*q
, struct request
*rq
)
537 struct list_head
*next
;
539 elevator_t
*e
= q
->elevator
;
541 if (e
->ops
->elevator_latter_req_fn
)
542 return e
->ops
->elevator_latter_req_fn(q
, rq
);
544 next
= rq
->queuelist
.next
;
545 if (next
!= &q
->queue_head
&& next
!= &rq
->queuelist
)
546 return list_entry_rq(next
);
551 struct request
*elv_former_request(request_queue_t
*q
, struct request
*rq
)
553 struct list_head
*prev
;
555 elevator_t
*e
= q
->elevator
;
557 if (e
->ops
->elevator_former_req_fn
)
558 return e
->ops
->elevator_former_req_fn(q
, rq
);
560 prev
= rq
->queuelist
.prev
;
561 if (prev
!= &q
->queue_head
&& prev
!= &rq
->queuelist
)
562 return list_entry_rq(prev
);
567 int elv_set_request(request_queue_t
*q
, struct request
*rq
, struct bio
*bio
,
570 elevator_t
*e
= q
->elevator
;
572 if (e
->ops
->elevator_set_req_fn
)
573 return e
->ops
->elevator_set_req_fn(q
, rq
, bio
, gfp_mask
);
575 rq
->elevator_private
= NULL
;
579 void elv_put_request(request_queue_t
*q
, struct request
*rq
)
581 elevator_t
*e
= q
->elevator
;
583 if (e
->ops
->elevator_put_req_fn
)
584 e
->ops
->elevator_put_req_fn(q
, rq
);
587 int elv_may_queue(request_queue_t
*q
, int rw
, struct bio
*bio
)
589 elevator_t
*e
= q
->elevator
;
591 if (e
->ops
->elevator_may_queue_fn
)
592 return e
->ops
->elevator_may_queue_fn(q
, rw
, bio
);
594 return ELV_MQUEUE_MAY
;
597 void elv_completed_request(request_queue_t
*q
, struct request
*rq
)
599 elevator_t
*e
= q
->elevator
;
602 * request is released from the driver, io must be done
604 if (blk_account_rq(rq
)) {
606 if (blk_sorted_rq(rq
) && e
->ops
->elevator_completed_req_fn
)
607 e
->ops
->elevator_completed_req_fn(q
, rq
);
611 int elv_register_queue(struct request_queue
*q
)
613 elevator_t
*e
= q
->elevator
;
615 e
->kobj
.parent
= kobject_get(&q
->kobj
);
619 snprintf(e
->kobj
.name
, KOBJ_NAME_LEN
, "%s", "iosched");
620 e
->kobj
.ktype
= e
->elevator_type
->elevator_ktype
;
622 return kobject_register(&e
->kobj
);
625 void elv_unregister_queue(struct request_queue
*q
)
628 elevator_t
*e
= q
->elevator
;
629 kobject_unregister(&e
->kobj
);
630 kobject_put(&q
->kobj
);
634 int elv_register(struct elevator_type
*e
)
636 spin_lock_irq(&elv_list_lock
);
637 if (elevator_find(e
->elevator_name
))
639 list_add_tail(&e
->list
, &elv_list
);
640 spin_unlock_irq(&elv_list_lock
);
642 printk(KERN_INFO
"io scheduler %s registered", e
->elevator_name
);
643 if (!strcmp(e
->elevator_name
, chosen_elevator
))
644 printk(" (default)");
648 EXPORT_SYMBOL_GPL(elv_register
);
650 void elv_unregister(struct elevator_type
*e
)
652 spin_lock_irq(&elv_list_lock
);
653 list_del_init(&e
->list
);
654 spin_unlock_irq(&elv_list_lock
);
656 EXPORT_SYMBOL_GPL(elv_unregister
);
659 * switch to new_e io scheduler. be careful not to introduce deadlocks -
660 * we don't free the old io scheduler, before we have allocated what we
661 * need for the new one. this way we have a chance of going back to the old
662 * one, if the new one fails init for some reason. we also do an intermediate
663 * switch to noop to ensure safety with stack-allocated requests, since they
664 * don't originate from the block layer allocator. noop is safe here, because
665 * it never needs to touch the elevator itself for completion events. DRAIN
666 * flags will make sure we don't touch it for additions either.
668 static void elevator_switch(request_queue_t
*q
, struct elevator_type
*new_e
)
670 elevator_t
*e
= kmalloc(sizeof(elevator_t
), GFP_KERNEL
);
671 struct elevator_type
*noop_elevator
= NULL
;
672 elevator_t
*old_elevator
;
678 * first step, drain requests from the block freelist
680 blk_wait_queue_drained(q
, 0);
683 * unregister old elevator data
685 elv_unregister_queue(q
);
686 old_elevator
= q
->elevator
;
689 * next step, switch to noop since it uses no private rq structures
690 * and doesn't allocate any memory for anything. then wait for any
691 * non-fs requests in-flight
693 noop_elevator
= elevator_get("noop");
694 spin_lock_irq(q
->queue_lock
);
695 elevator_attach(q
, noop_elevator
, e
);
696 spin_unlock_irq(q
->queue_lock
);
698 blk_wait_queue_drained(q
, 1);
701 * attach and start new elevator
703 if (elevator_attach(q
, new_e
, e
))
706 if (elv_register_queue(q
))
710 * finally exit old elevator and start queue again
712 elevator_exit(old_elevator
);
713 blk_finish_queue_drain(q
);
714 elevator_put(noop_elevator
);
719 * switch failed, exit the new io scheduler and reattach the old
720 * one again (along with re-adding the sysfs dir)
724 q
->elevator
= old_elevator
;
725 elv_register_queue(q
);
726 blk_finish_queue_drain(q
);
729 elevator_put(noop_elevator
);
731 printk(KERN_ERR
"elevator: switch to %s failed\n",new_e
->elevator_name
);
734 ssize_t
elv_iosched_store(request_queue_t
*q
, const char *name
, size_t count
)
736 char elevator_name
[ELV_NAME_MAX
];
737 struct elevator_type
*e
;
739 memset(elevator_name
, 0, sizeof(elevator_name
));
740 strncpy(elevator_name
, name
, sizeof(elevator_name
));
742 if (elevator_name
[strlen(elevator_name
) - 1] == '\n')
743 elevator_name
[strlen(elevator_name
) - 1] = '\0';
745 e
= elevator_get(elevator_name
);
747 printk(KERN_ERR
"elevator: type %s not found\n", elevator_name
);
751 if (!strcmp(elevator_name
, q
->elevator
->elevator_type
->elevator_name
))
754 elevator_switch(q
, e
);
758 ssize_t
elv_iosched_show(request_queue_t
*q
, char *name
)
760 elevator_t
*e
= q
->elevator
;
761 struct elevator_type
*elv
= e
->elevator_type
;
762 struct list_head
*entry
;
765 spin_lock_irq(q
->queue_lock
);
766 list_for_each(entry
, &elv_list
) {
767 struct elevator_type
*__e
;
769 __e
= list_entry(entry
, struct elevator_type
, list
);
770 if (!strcmp(elv
->elevator_name
, __e
->elevator_name
))
771 len
+= sprintf(name
+len
, "[%s] ", elv
->elevator_name
);
773 len
+= sprintf(name
+len
, "%s ", __e
->elevator_name
);
775 spin_unlock_irq(q
->queue_lock
);
777 len
+= sprintf(len
+name
, "\n");
781 EXPORT_SYMBOL(elv_dispatch_sort
);
782 EXPORT_SYMBOL(elv_add_request
);
783 EXPORT_SYMBOL(__elv_add_request
);
784 EXPORT_SYMBOL(elv_requeue_request
);
785 EXPORT_SYMBOL(elv_next_request
);
786 EXPORT_SYMBOL(elv_dequeue_request
);
787 EXPORT_SYMBOL(elv_queue_empty
);
788 EXPORT_SYMBOL(elv_completed_request
);
789 EXPORT_SYMBOL(elevator_exit
);
790 EXPORT_SYMBOL(elevator_init
);