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 static 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
;
84 static struct elevator_type
*elevator_find(const char *name
)
86 struct elevator_type
*e
= NULL
;
87 struct list_head
*entry
;
89 list_for_each(entry
, &elv_list
) {
90 struct elevator_type
*__e
;
92 __e
= list_entry(entry
, struct elevator_type
, list
);
94 if (!strcmp(__e
->elevator_name
, name
)) {
103 static void elevator_put(struct elevator_type
*e
)
105 module_put(e
->elevator_owner
);
108 static struct elevator_type
*elevator_get(const char *name
)
110 struct elevator_type
*e
;
112 spin_lock_irq(&elv_list_lock
);
114 e
= elevator_find(name
);
115 if (e
&& !try_module_get(e
->elevator_owner
))
118 spin_unlock_irq(&elv_list_lock
);
123 static int elevator_attach(request_queue_t
*q
, struct elevator_type
*e
,
124 struct elevator_queue
*eq
)
128 memset(eq
, 0, sizeof(*eq
));
130 eq
->elevator_type
= e
;
134 if (eq
->ops
->elevator_init_fn
)
135 ret
= eq
->ops
->elevator_init_fn(q
, eq
);
140 static char chosen_elevator
[16];
142 static int __init
elevator_setup(char *str
)
145 * Be backwards-compatible with previous kernels, so users
146 * won't get the wrong elevator.
148 if (!strcmp(str
, "as"))
149 strcpy(chosen_elevator
, "anticipatory");
151 strncpy(chosen_elevator
, str
, sizeof(chosen_elevator
) - 1);
155 __setup("elevator=", elevator_setup
);
157 int elevator_init(request_queue_t
*q
, char *name
)
159 struct elevator_type
*e
= NULL
;
160 struct elevator_queue
*eq
;
163 INIT_LIST_HEAD(&q
->queue_head
);
164 q
->last_merge
= NULL
;
166 q
->boundary_rq
= NULL
;
168 if (name
&& !(e
= elevator_get(name
)))
171 if (!e
&& *chosen_elevator
&& !(e
= elevator_get(chosen_elevator
)))
172 printk("I/O scheduler %s not found\n", chosen_elevator
);
174 if (!e
&& !(e
= elevator_get(CONFIG_DEFAULT_IOSCHED
))) {
175 printk("Default I/O scheduler not found, using no-op\n");
176 e
= elevator_get("noop");
179 eq
= kmalloc(sizeof(struct elevator_queue
), GFP_KERNEL
);
185 ret
= elevator_attach(q
, e
, eq
);
194 void elevator_exit(elevator_t
*e
)
196 if (e
->ops
->elevator_exit_fn
)
197 e
->ops
->elevator_exit_fn(e
);
199 elevator_put(e
->elevator_type
);
200 e
->elevator_type
= NULL
;
205 * Insert rq into dispatch queue of q. Queue lock must be held on
206 * entry. If sort != 0, rq is sort-inserted; otherwise, rq will be
207 * appended to the dispatch queue. To be used by specific elevators.
209 void elv_dispatch_sort(request_queue_t
*q
, struct request
*rq
)
212 struct list_head
*entry
;
214 if (q
->last_merge
== rq
)
215 q
->last_merge
= NULL
;
218 boundary
= q
->end_sector
;
220 list_for_each_prev(entry
, &q
->queue_head
) {
221 struct request
*pos
= list_entry_rq(entry
);
223 if (pos
->flags
& (REQ_SOFTBARRIER
|REQ_HARDBARRIER
|REQ_STARTED
))
225 if (rq
->sector
>= boundary
) {
226 if (pos
->sector
< boundary
)
229 if (pos
->sector
>= boundary
)
232 if (rq
->sector
>= pos
->sector
)
236 list_add(&rq
->queuelist
, entry
);
239 int elv_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
241 elevator_t
*e
= q
->elevator
;
245 ret
= elv_try_merge(q
->last_merge
, bio
);
246 if (ret
!= ELEVATOR_NO_MERGE
) {
247 *req
= q
->last_merge
;
252 if (e
->ops
->elevator_merge_fn
)
253 return e
->ops
->elevator_merge_fn(q
, req
, bio
);
255 return ELEVATOR_NO_MERGE
;
258 void elv_merged_request(request_queue_t
*q
, struct request
*rq
)
260 elevator_t
*e
= q
->elevator
;
262 if (e
->ops
->elevator_merged_fn
)
263 e
->ops
->elevator_merged_fn(q
, rq
);
268 void elv_merge_requests(request_queue_t
*q
, struct request
*rq
,
269 struct request
*next
)
271 elevator_t
*e
= q
->elevator
;
273 if (e
->ops
->elevator_merge_req_fn
)
274 e
->ops
->elevator_merge_req_fn(q
, rq
, next
);
280 void elv_requeue_request(request_queue_t
*q
, struct request
*rq
)
282 elevator_t
*e
= q
->elevator
;
285 * it already went through dequeue, we need to decrement the
286 * in_flight count again
288 if (blk_account_rq(rq
)) {
290 if (blk_sorted_rq(rq
) && e
->ops
->elevator_deactivate_req_fn
)
291 e
->ops
->elevator_deactivate_req_fn(q
, rq
);
294 rq
->flags
&= ~REQ_STARTED
;
296 __elv_add_request(q
, rq
, ELEVATOR_INSERT_REQUEUE
, 0);
299 static void elv_drain_elevator(request_queue_t
*q
)
302 while (q
->elevator
->ops
->elevator_dispatch_fn(q
, 1))
304 if (q
->nr_sorted
== 0)
306 if (printed
++ < 10) {
307 printk(KERN_ERR
"%s: forced dispatching is broken "
308 "(nr_sorted=%u), please report this\n",
309 q
->elevator
->elevator_type
->elevator_name
, q
->nr_sorted
);
313 void __elv_add_request(request_queue_t
*q
, struct request
*rq
, int where
,
316 struct list_head
*pos
;
320 rq
->flags
|= REQ_ORDERED_COLOR
;
322 if (rq
->flags
& (REQ_SOFTBARRIER
| REQ_HARDBARRIER
)) {
324 * toggle ordered color
326 if (blk_barrier_rq(rq
))
330 * barriers implicitly indicate back insertion
332 if (where
== ELEVATOR_INSERT_SORT
)
333 where
= ELEVATOR_INSERT_BACK
;
336 * this request is scheduling boundary, update end_sector
338 if (blk_fs_request(rq
)) {
339 q
->end_sector
= rq_end_sector(rq
);
342 } else if (!(rq
->flags
& REQ_ELVPRIV
) && where
== ELEVATOR_INSERT_SORT
)
343 where
= ELEVATOR_INSERT_BACK
;
351 case ELEVATOR_INSERT_FRONT
:
352 rq
->flags
|= REQ_SOFTBARRIER
;
354 list_add(&rq
->queuelist
, &q
->queue_head
);
357 case ELEVATOR_INSERT_BACK
:
358 rq
->flags
|= REQ_SOFTBARRIER
;
359 elv_drain_elevator(q
);
360 list_add_tail(&rq
->queuelist
, &q
->queue_head
);
362 * We kick the queue here for the following reasons.
363 * - The elevator might have returned NULL previously
364 * to delay requests and returned them now. As the
365 * queue wasn't empty before this request, ll_rw_blk
366 * won't run the queue on return, resulting in hang.
367 * - Usually, back inserted requests won't be merged
368 * with anything. There's no point in delaying queue
375 case ELEVATOR_INSERT_SORT
:
376 BUG_ON(!blk_fs_request(rq
));
377 rq
->flags
|= REQ_SORTED
;
379 if (q
->last_merge
== NULL
&& rq_mergeable(rq
))
382 * Some ioscheds (cfq) run q->request_fn directly, so
383 * rq cannot be accessed after calling
384 * elevator_add_req_fn.
386 q
->elevator
->ops
->elevator_add_req_fn(q
, rq
);
389 case ELEVATOR_INSERT_REQUEUE
:
391 * If ordered flush isn't in progress, we do front
392 * insertion; otherwise, requests should be requeued
395 rq
->flags
|= REQ_SOFTBARRIER
;
397 if (q
->ordseq
== 0) {
398 list_add(&rq
->queuelist
, &q
->queue_head
);
402 ordseq
= blk_ordered_req_seq(rq
);
404 list_for_each(pos
, &q
->queue_head
) {
405 struct request
*pos_rq
= list_entry_rq(pos
);
406 if (ordseq
<= blk_ordered_req_seq(pos_rq
))
410 list_add_tail(&rq
->queuelist
, pos
);
414 printk(KERN_ERR
"%s: bad insertion point %d\n",
415 __FUNCTION__
, where
);
419 if (blk_queue_plugged(q
)) {
420 int nrq
= q
->rq
.count
[READ
] + q
->rq
.count
[WRITE
]
423 if (nrq
>= q
->unplug_thresh
)
424 __generic_unplug_device(q
);
428 void elv_add_request(request_queue_t
*q
, struct request
*rq
, int where
,
433 spin_lock_irqsave(q
->queue_lock
, flags
);
434 __elv_add_request(q
, rq
, where
, plug
);
435 spin_unlock_irqrestore(q
->queue_lock
, flags
);
438 static inline struct request
*__elv_next_request(request_queue_t
*q
)
443 while (!list_empty(&q
->queue_head
)) {
444 rq
= list_entry_rq(q
->queue_head
.next
);
445 if (blk_do_ordered(q
, &rq
))
449 if (!q
->elevator
->ops
->elevator_dispatch_fn(q
, 0))
454 struct request
*elv_next_request(request_queue_t
*q
)
459 while ((rq
= __elv_next_request(q
)) != NULL
) {
460 if (!(rq
->flags
& REQ_STARTED
)) {
461 elevator_t
*e
= q
->elevator
;
464 * This is the first time the device driver
465 * sees this request (possibly after
466 * requeueing). Notify IO scheduler.
468 if (blk_sorted_rq(rq
) &&
469 e
->ops
->elevator_activate_req_fn
)
470 e
->ops
->elevator_activate_req_fn(q
, rq
);
473 * just mark as started even if we don't start
474 * it, a request that has been delayed should
475 * not be passed by new incoming requests
477 rq
->flags
|= REQ_STARTED
;
480 if (!q
->boundary_rq
|| q
->boundary_rq
== rq
) {
481 q
->end_sector
= rq_end_sector(rq
);
482 q
->boundary_rq
= NULL
;
485 if ((rq
->flags
& REQ_DONTPREP
) || !q
->prep_rq_fn
)
488 ret
= q
->prep_rq_fn(q
, rq
);
489 if (ret
== BLKPREP_OK
) {
491 } else if (ret
== BLKPREP_DEFER
) {
493 * the request may have been (partially) prepped.
494 * we need to keep this request in the front to
495 * avoid resource deadlock. REQ_STARTED will
496 * prevent other fs requests from passing this one.
500 } else if (ret
== BLKPREP_KILL
) {
501 int nr_bytes
= rq
->hard_nr_sectors
<< 9;
504 nr_bytes
= rq
->data_len
;
506 blkdev_dequeue_request(rq
);
507 rq
->flags
|= REQ_QUIET
;
508 end_that_request_chunk(rq
, 0, nr_bytes
);
509 end_that_request_last(rq
, 0);
511 printk(KERN_ERR
"%s: bad return=%d\n", __FUNCTION__
,
520 void elv_dequeue_request(request_queue_t
*q
, struct request
*rq
)
522 BUG_ON(list_empty(&rq
->queuelist
));
524 list_del_init(&rq
->queuelist
);
527 * the time frame between a request being removed from the lists
528 * and to it is freed is accounted as io that is in progress at
531 if (blk_account_rq(rq
))
535 int elv_queue_empty(request_queue_t
*q
)
537 elevator_t
*e
= q
->elevator
;
539 if (!list_empty(&q
->queue_head
))
542 if (e
->ops
->elevator_queue_empty_fn
)
543 return e
->ops
->elevator_queue_empty_fn(q
);
548 struct request
*elv_latter_request(request_queue_t
*q
, struct request
*rq
)
550 elevator_t
*e
= q
->elevator
;
552 if (e
->ops
->elevator_latter_req_fn
)
553 return e
->ops
->elevator_latter_req_fn(q
, rq
);
557 struct request
*elv_former_request(request_queue_t
*q
, struct request
*rq
)
559 elevator_t
*e
= q
->elevator
;
561 if (e
->ops
->elevator_former_req_fn
)
562 return e
->ops
->elevator_former_req_fn(q
, rq
);
566 int elv_set_request(request_queue_t
*q
, struct request
*rq
, struct bio
*bio
,
569 elevator_t
*e
= q
->elevator
;
571 if (e
->ops
->elevator_set_req_fn
)
572 return e
->ops
->elevator_set_req_fn(q
, rq
, bio
, gfp_mask
);
574 rq
->elevator_private
= NULL
;
578 void elv_put_request(request_queue_t
*q
, struct request
*rq
)
580 elevator_t
*e
= q
->elevator
;
582 if (e
->ops
->elevator_put_req_fn
)
583 e
->ops
->elevator_put_req_fn(q
, rq
);
586 int elv_may_queue(request_queue_t
*q
, int rw
, struct bio
*bio
)
588 elevator_t
*e
= q
->elevator
;
590 if (e
->ops
->elevator_may_queue_fn
)
591 return e
->ops
->elevator_may_queue_fn(q
, rw
, bio
);
593 return ELV_MQUEUE_MAY
;
596 void elv_completed_request(request_queue_t
*q
, struct request
*rq
)
598 elevator_t
*e
= q
->elevator
;
601 * request is released from the driver, io must be done
603 if (blk_account_rq(rq
)) {
605 if (blk_sorted_rq(rq
) && e
->ops
->elevator_completed_req_fn
)
606 e
->ops
->elevator_completed_req_fn(q
, rq
);
610 * Check if the queue is waiting for fs requests to be
611 * drained for flush sequence.
613 if (unlikely(q
->ordseq
)) {
614 struct request
*first_rq
= list_entry_rq(q
->queue_head
.next
);
615 if (q
->in_flight
== 0 &&
616 blk_ordered_cur_seq(q
) == QUEUE_ORDSEQ_DRAIN
&&
617 blk_ordered_req_seq(first_rq
) > QUEUE_ORDSEQ_DRAIN
) {
618 blk_ordered_complete_seq(q
, QUEUE_ORDSEQ_DRAIN
, 0);
624 int elv_register_queue(struct request_queue
*q
)
626 elevator_t
*e
= q
->elevator
;
628 e
->kobj
.parent
= kobject_get(&q
->kobj
);
632 snprintf(e
->kobj
.name
, KOBJ_NAME_LEN
, "%s", "iosched");
633 e
->kobj
.ktype
= e
->elevator_type
->elevator_ktype
;
635 return kobject_register(&e
->kobj
);
638 void elv_unregister_queue(struct request_queue
*q
)
641 elevator_t
*e
= q
->elevator
;
642 kobject_unregister(&e
->kobj
);
643 kobject_put(&q
->kobj
);
647 int elv_register(struct elevator_type
*e
)
649 spin_lock_irq(&elv_list_lock
);
650 if (elevator_find(e
->elevator_name
))
652 list_add_tail(&e
->list
, &elv_list
);
653 spin_unlock_irq(&elv_list_lock
);
655 printk(KERN_INFO
"io scheduler %s registered", e
->elevator_name
);
656 if (!strcmp(e
->elevator_name
, chosen_elevator
) ||
657 (!*chosen_elevator
&&
658 !strcmp(e
->elevator_name
, CONFIG_DEFAULT_IOSCHED
)))
659 printk(" (default)");
663 EXPORT_SYMBOL_GPL(elv_register
);
665 void elv_unregister(struct elevator_type
*e
)
667 struct task_struct
*g
, *p
;
670 * Iterate every thread in the process to remove the io contexts.
672 read_lock(&tasklist_lock
);
673 do_each_thread(g
, p
) {
674 struct io_context
*ioc
= p
->io_context
;
675 if (ioc
&& ioc
->cic
) {
676 ioc
->cic
->exit(ioc
->cic
);
677 ioc
->cic
->dtor(ioc
->cic
);
680 if (ioc
&& ioc
->aic
) {
681 ioc
->aic
->exit(ioc
->aic
);
682 ioc
->aic
->dtor(ioc
->aic
);
685 } while_each_thread(g
, p
);
686 read_unlock(&tasklist_lock
);
688 spin_lock_irq(&elv_list_lock
);
689 list_del_init(&e
->list
);
690 spin_unlock_irq(&elv_list_lock
);
692 EXPORT_SYMBOL_GPL(elv_unregister
);
695 * switch to new_e io scheduler. be careful not to introduce deadlocks -
696 * we don't free the old io scheduler, before we have allocated what we
697 * need for the new one. this way we have a chance of going back to the old
698 * one, if the new one fails init for some reason.
700 static void elevator_switch(request_queue_t
*q
, struct elevator_type
*new_e
)
702 elevator_t
*old_elevator
, *e
;
705 * Allocate new elevator
707 e
= kmalloc(sizeof(elevator_t
), GFP_KERNEL
);
712 * Turn on BYPASS and drain all requests w/ elevator private data
714 spin_lock_irq(q
->queue_lock
);
716 set_bit(QUEUE_FLAG_ELVSWITCH
, &q
->queue_flags
);
718 elv_drain_elevator(q
);
720 while (q
->rq
.elvpriv
) {
723 spin_unlock_irq(q
->queue_lock
);
725 spin_lock_irq(q
->queue_lock
);
726 elv_drain_elevator(q
);
729 spin_unlock_irq(q
->queue_lock
);
732 * unregister old elevator data
734 elv_unregister_queue(q
);
735 old_elevator
= q
->elevator
;
738 * attach and start new elevator
740 if (elevator_attach(q
, new_e
, e
))
743 if (elv_register_queue(q
))
747 * finally exit old elevator and turn off BYPASS.
749 elevator_exit(old_elevator
);
750 clear_bit(QUEUE_FLAG_ELVSWITCH
, &q
->queue_flags
);
755 * switch failed, exit the new io scheduler and reattach the old
756 * one again (along with re-adding the sysfs dir)
761 q
->elevator
= old_elevator
;
762 elv_register_queue(q
);
763 clear_bit(QUEUE_FLAG_ELVSWITCH
, &q
->queue_flags
);
767 printk(KERN_ERR
"elevator: switch to %s failed\n",new_e
->elevator_name
);
770 ssize_t
elv_iosched_store(request_queue_t
*q
, const char *name
, size_t count
)
772 char elevator_name
[ELV_NAME_MAX
];
774 struct elevator_type
*e
;
776 elevator_name
[sizeof(elevator_name
) - 1] = '\0';
777 strncpy(elevator_name
, name
, sizeof(elevator_name
) - 1);
778 len
= strlen(elevator_name
);
780 if (len
&& elevator_name
[len
- 1] == '\n')
781 elevator_name
[len
- 1] = '\0';
783 e
= elevator_get(elevator_name
);
785 printk(KERN_ERR
"elevator: type %s not found\n", elevator_name
);
789 if (!strcmp(elevator_name
, q
->elevator
->elevator_type
->elevator_name
)) {
794 elevator_switch(q
, e
);
798 ssize_t
elv_iosched_show(request_queue_t
*q
, char *name
)
800 elevator_t
*e
= q
->elevator
;
801 struct elevator_type
*elv
= e
->elevator_type
;
802 struct list_head
*entry
;
805 spin_lock_irq(q
->queue_lock
);
806 list_for_each(entry
, &elv_list
) {
807 struct elevator_type
*__e
;
809 __e
= list_entry(entry
, struct elevator_type
, list
);
810 if (!strcmp(elv
->elevator_name
, __e
->elevator_name
))
811 len
+= sprintf(name
+len
, "[%s] ", elv
->elevator_name
);
813 len
+= sprintf(name
+len
, "%s ", __e
->elevator_name
);
815 spin_unlock_irq(q
->queue_lock
);
817 len
+= sprintf(len
+name
, "\n");
821 EXPORT_SYMBOL(elv_dispatch_sort
);
822 EXPORT_SYMBOL(elv_add_request
);
823 EXPORT_SYMBOL(__elv_add_request
);
824 EXPORT_SYMBOL(elv_requeue_request
);
825 EXPORT_SYMBOL(elv_next_request
);
826 EXPORT_SYMBOL(elv_dequeue_request
);
827 EXPORT_SYMBOL(elv_queue_empty
);
828 EXPORT_SYMBOL(elv_completed_request
);
829 EXPORT_SYMBOL(elevator_exit
);
830 EXPORT_SYMBOL(elevator_init
);