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 inline int elv_try_last_merge(request_queue_t
*q
, struct bio
*bio
)
89 return elv_try_merge(q
->last_merge
, bio
);
91 return ELEVATOR_NO_MERGE
;
93 EXPORT_SYMBOL(elv_try_last_merge
);
95 static struct elevator_type
*elevator_find(const char *name
)
97 struct elevator_type
*e
= NULL
;
98 struct list_head
*entry
;
100 list_for_each(entry
, &elv_list
) {
101 struct elevator_type
*__e
;
103 __e
= list_entry(entry
, struct elevator_type
, list
);
105 if (!strcmp(__e
->elevator_name
, name
)) {
114 static void elevator_put(struct elevator_type
*e
)
116 module_put(e
->elevator_owner
);
119 static struct elevator_type
*elevator_get(const char *name
)
121 struct elevator_type
*e
;
123 spin_lock_irq(&elv_list_lock
);
125 e
= elevator_find(name
);
126 if (e
&& !try_module_get(e
->elevator_owner
))
129 spin_unlock_irq(&elv_list_lock
);
134 static int elevator_attach(request_queue_t
*q
, struct elevator_type
*e
,
135 struct elevator_queue
*eq
)
139 memset(eq
, 0, sizeof(*eq
));
141 eq
->elevator_type
= e
;
143 INIT_LIST_HEAD(&q
->queue_head
);
144 q
->last_merge
= NULL
;
147 if (eq
->ops
->elevator_init_fn
)
148 ret
= eq
->ops
->elevator_init_fn(q
, eq
);
153 static char chosen_elevator
[16];
155 static void elevator_setup_default(void)
157 struct elevator_type
*e
;
160 * check if default is set and exists
162 if (chosen_elevator
[0] && (e
= elevator_get(chosen_elevator
))) {
167 #if defined(CONFIG_IOSCHED_AS)
168 strcpy(chosen_elevator
, "anticipatory");
169 #elif defined(CONFIG_IOSCHED_DEADLINE)
170 strcpy(chosen_elevator
, "deadline");
171 #elif defined(CONFIG_IOSCHED_CFQ)
172 strcpy(chosen_elevator
, "cfq");
173 #elif defined(CONFIG_IOSCHED_NOOP)
174 strcpy(chosen_elevator
, "noop");
176 #error "You must build at least 1 IO scheduler into the kernel"
180 static int __init
elevator_setup(char *str
)
182 strncpy(chosen_elevator
, str
, sizeof(chosen_elevator
) - 1);
186 __setup("elevator=", elevator_setup
);
188 int elevator_init(request_queue_t
*q
, char *name
)
190 struct elevator_type
*e
= NULL
;
191 struct elevator_queue
*eq
;
194 elevator_setup_default();
197 name
= chosen_elevator
;
199 e
= elevator_get(name
);
203 eq
= kmalloc(sizeof(struct elevator_queue
), GFP_KERNEL
);
205 elevator_put(e
->elevator_type
);
209 ret
= elevator_attach(q
, e
, eq
);
212 elevator_put(e
->elevator_type
);
218 void elevator_exit(elevator_t
*e
)
220 if (e
->ops
->elevator_exit_fn
)
221 e
->ops
->elevator_exit_fn(e
);
223 elevator_put(e
->elevator_type
);
224 e
->elevator_type
= NULL
;
228 int elv_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
230 elevator_t
*e
= q
->elevator
;
232 if (e
->ops
->elevator_merge_fn
)
233 return e
->ops
->elevator_merge_fn(q
, req
, bio
);
235 return ELEVATOR_NO_MERGE
;
238 void elv_merged_request(request_queue_t
*q
, struct request
*rq
)
240 elevator_t
*e
= q
->elevator
;
242 if (e
->ops
->elevator_merged_fn
)
243 e
->ops
->elevator_merged_fn(q
, rq
);
246 void elv_merge_requests(request_queue_t
*q
, struct request
*rq
,
247 struct request
*next
)
249 elevator_t
*e
= q
->elevator
;
251 if (q
->last_merge
== next
)
252 q
->last_merge
= NULL
;
254 if (e
->ops
->elevator_merge_req_fn
)
255 e
->ops
->elevator_merge_req_fn(q
, rq
, next
);
259 * For careful internal use by the block layer. Essentially the same as
260 * a requeue in that it tells the io scheduler that this request is not
261 * active in the driver or hardware anymore, but we don't want the request
262 * added back to the scheduler. Function is not exported.
264 void elv_deactivate_request(request_queue_t
*q
, struct request
*rq
)
266 elevator_t
*e
= q
->elevator
;
269 * it already went through dequeue, we need to decrement the
270 * in_flight count again
272 if (blk_account_rq(rq
))
275 rq
->flags
&= ~REQ_STARTED
;
277 if (e
->ops
->elevator_deactivate_req_fn
)
278 e
->ops
->elevator_deactivate_req_fn(q
, rq
);
281 void elv_requeue_request(request_queue_t
*q
, struct request
*rq
)
283 elv_deactivate_request(q
, rq
);
286 * if this is the flush, requeue the original instead and drop the flush
288 if (rq
->flags
& REQ_BAR_FLUSH
) {
289 clear_bit(QUEUE_FLAG_FLUSH
, &q
->queue_flags
);
290 rq
= rq
->end_io_data
;
294 * the request is prepped and may have some resources allocated.
295 * allowing unprepped requests to pass this one may cause resource
296 * deadlock. turn on softbarrier.
298 rq
->flags
|= REQ_SOFTBARRIER
;
301 * if iosched has an explicit requeue hook, then use that. otherwise
302 * just put the request at the front of the queue
304 if (q
->elevator
->ops
->elevator_requeue_req_fn
)
305 q
->elevator
->ops
->elevator_requeue_req_fn(q
, rq
);
307 __elv_add_request(q
, rq
, ELEVATOR_INSERT_FRONT
, 0);
310 void __elv_add_request(request_queue_t
*q
, struct request
*rq
, int where
,
314 * barriers implicitly indicate back insertion
316 if (rq
->flags
& (REQ_SOFTBARRIER
| REQ_HARDBARRIER
) &&
317 where
== ELEVATOR_INSERT_SORT
)
318 where
= ELEVATOR_INSERT_BACK
;
325 if (!test_bit(QUEUE_FLAG_DRAIN
, &q
->queue_flags
)) {
326 q
->elevator
->ops
->elevator_add_req_fn(q
, rq
, where
);
328 if (blk_queue_plugged(q
)) {
329 int nrq
= q
->rq
.count
[READ
] + q
->rq
.count
[WRITE
]
332 if (nrq
>= q
->unplug_thresh
)
333 __generic_unplug_device(q
);
337 * if drain is set, store the request "locally". when the drain
338 * is finished, the requests will be handed ordered to the io
341 list_add_tail(&rq
->queuelist
, &q
->drain_list
);
344 void elv_add_request(request_queue_t
*q
, struct request
*rq
, int where
,
349 spin_lock_irqsave(q
->queue_lock
, flags
);
350 __elv_add_request(q
, rq
, where
, plug
);
351 spin_unlock_irqrestore(q
->queue_lock
, flags
);
354 static inline struct request
*__elv_next_request(request_queue_t
*q
)
356 struct request
*rq
= q
->elevator
->ops
->elevator_next_req_fn(q
);
359 * if this is a barrier write and the device has to issue a
360 * flush sequence to support it, check how far we are
362 if (rq
&& blk_fs_request(rq
) && blk_barrier_rq(rq
)) {
363 BUG_ON(q
->ordered
== QUEUE_ORDERED_NONE
);
365 if (q
->ordered
== QUEUE_ORDERED_FLUSH
&&
366 !blk_barrier_preflush(rq
))
367 rq
= blk_start_pre_flush(q
, rq
);
373 struct request
*elv_next_request(request_queue_t
*q
)
378 while ((rq
= __elv_next_request(q
)) != NULL
) {
380 * just mark as started even if we don't start it, a request
381 * that has been delayed should not be passed by new incoming
384 rq
->flags
|= REQ_STARTED
;
386 if (rq
== q
->last_merge
)
387 q
->last_merge
= NULL
;
389 if ((rq
->flags
& REQ_DONTPREP
) || !q
->prep_rq_fn
)
392 ret
= q
->prep_rq_fn(q
, rq
);
393 if (ret
== BLKPREP_OK
) {
395 } else if (ret
== BLKPREP_DEFER
) {
397 * the request may have been (partially) prepped.
398 * we need to keep this request in the front to
399 * avoid resource deadlock. turn on softbarrier.
401 rq
->flags
|= REQ_SOFTBARRIER
;
404 } else if (ret
== BLKPREP_KILL
) {
405 int nr_bytes
= rq
->hard_nr_sectors
<< 9;
408 nr_bytes
= rq
->data_len
;
410 blkdev_dequeue_request(rq
);
411 rq
->flags
|= REQ_QUIET
;
412 end_that_request_chunk(rq
, 0, nr_bytes
);
413 end_that_request_last(rq
);
415 printk(KERN_ERR
"%s: bad return=%d\n", __FUNCTION__
,
424 void elv_remove_request(request_queue_t
*q
, struct request
*rq
)
426 elevator_t
*e
= q
->elevator
;
429 * the time frame between a request being removed from the lists
430 * and to it is freed is accounted as io that is in progress at
431 * the driver side. note that we only account requests that the
432 * driver has seen (REQ_STARTED set), to avoid false accounting
433 * for request-request merges
435 if (blk_account_rq(rq
))
439 * the main clearing point for q->last_merge is on retrieval of
440 * request by driver (it calls elv_next_request()), but it _can_
441 * also happen here if a request is added to the queue but later
442 * deleted without ever being given to driver (merged with another
445 if (rq
== q
->last_merge
)
446 q
->last_merge
= NULL
;
448 if (e
->ops
->elevator_remove_req_fn
)
449 e
->ops
->elevator_remove_req_fn(q
, rq
);
452 int elv_queue_empty(request_queue_t
*q
)
454 elevator_t
*e
= q
->elevator
;
456 if (e
->ops
->elevator_queue_empty_fn
)
457 return e
->ops
->elevator_queue_empty_fn(q
);
459 return list_empty(&q
->queue_head
);
462 struct request
*elv_latter_request(request_queue_t
*q
, struct request
*rq
)
464 struct list_head
*next
;
466 elevator_t
*e
= q
->elevator
;
468 if (e
->ops
->elevator_latter_req_fn
)
469 return e
->ops
->elevator_latter_req_fn(q
, rq
);
471 next
= rq
->queuelist
.next
;
472 if (next
!= &q
->queue_head
&& next
!= &rq
->queuelist
)
473 return list_entry_rq(next
);
478 struct request
*elv_former_request(request_queue_t
*q
, struct request
*rq
)
480 struct list_head
*prev
;
482 elevator_t
*e
= q
->elevator
;
484 if (e
->ops
->elevator_former_req_fn
)
485 return e
->ops
->elevator_former_req_fn(q
, rq
);
487 prev
= rq
->queuelist
.prev
;
488 if (prev
!= &q
->queue_head
&& prev
!= &rq
->queuelist
)
489 return list_entry_rq(prev
);
494 int elv_set_request(request_queue_t
*q
, struct request
*rq
, struct bio
*bio
,
497 elevator_t
*e
= q
->elevator
;
499 if (e
->ops
->elevator_set_req_fn
)
500 return e
->ops
->elevator_set_req_fn(q
, rq
, bio
, gfp_mask
);
502 rq
->elevator_private
= NULL
;
506 void elv_put_request(request_queue_t
*q
, struct request
*rq
)
508 elevator_t
*e
= q
->elevator
;
510 if (e
->ops
->elevator_put_req_fn
)
511 e
->ops
->elevator_put_req_fn(q
, rq
);
514 int elv_may_queue(request_queue_t
*q
, int rw
, struct bio
*bio
)
516 elevator_t
*e
= q
->elevator
;
518 if (e
->ops
->elevator_may_queue_fn
)
519 return e
->ops
->elevator_may_queue_fn(q
, rw
, bio
);
521 return ELV_MQUEUE_MAY
;
524 void elv_completed_request(request_queue_t
*q
, struct request
*rq
)
526 elevator_t
*e
= q
->elevator
;
529 * request is released from the driver, io must be done
531 if (blk_account_rq(rq
))
534 if (e
->ops
->elevator_completed_req_fn
)
535 e
->ops
->elevator_completed_req_fn(q
, rq
);
538 int elv_register_queue(struct request_queue
*q
)
540 elevator_t
*e
= q
->elevator
;
542 e
->kobj
.parent
= kobject_get(&q
->kobj
);
546 snprintf(e
->kobj
.name
, KOBJ_NAME_LEN
, "%s", "iosched");
547 e
->kobj
.ktype
= e
->elevator_type
->elevator_ktype
;
549 return kobject_register(&e
->kobj
);
552 void elv_unregister_queue(struct request_queue
*q
)
555 elevator_t
*e
= q
->elevator
;
556 kobject_unregister(&e
->kobj
);
557 kobject_put(&q
->kobj
);
561 int elv_register(struct elevator_type
*e
)
563 spin_lock_irq(&elv_list_lock
);
564 if (elevator_find(e
->elevator_name
))
566 list_add_tail(&e
->list
, &elv_list
);
567 spin_unlock_irq(&elv_list_lock
);
569 printk(KERN_INFO
"io scheduler %s registered", e
->elevator_name
);
570 if (!strcmp(e
->elevator_name
, chosen_elevator
))
571 printk(" (default)");
575 EXPORT_SYMBOL_GPL(elv_register
);
577 void elv_unregister(struct elevator_type
*e
)
579 spin_lock_irq(&elv_list_lock
);
580 list_del_init(&e
->list
);
581 spin_unlock_irq(&elv_list_lock
);
583 EXPORT_SYMBOL_GPL(elv_unregister
);
586 * switch to new_e io scheduler. be careful not to introduce deadlocks -
587 * we don't free the old io scheduler, before we have allocated what we
588 * need for the new one. this way we have a chance of going back to the old
589 * one, if the new one fails init for some reason. we also do an intermediate
590 * switch to noop to ensure safety with stack-allocated requests, since they
591 * don't originate from the block layer allocator. noop is safe here, because
592 * it never needs to touch the elevator itself for completion events. DRAIN
593 * flags will make sure we don't touch it for additions either.
595 static void elevator_switch(request_queue_t
*q
, struct elevator_type
*new_e
)
597 elevator_t
*e
= kmalloc(sizeof(elevator_t
), GFP_KERNEL
);
598 struct elevator_type
*noop_elevator
= NULL
;
599 elevator_t
*old_elevator
;
605 * first step, drain requests from the block freelist
607 blk_wait_queue_drained(q
, 0);
610 * unregister old elevator data
612 elv_unregister_queue(q
);
613 old_elevator
= q
->elevator
;
616 * next step, switch to noop since it uses no private rq structures
617 * and doesn't allocate any memory for anything. then wait for any
618 * non-fs requests in-flight
620 noop_elevator
= elevator_get("noop");
621 spin_lock_irq(q
->queue_lock
);
622 elevator_attach(q
, noop_elevator
, e
);
623 spin_unlock_irq(q
->queue_lock
);
625 blk_wait_queue_drained(q
, 1);
628 * attach and start new elevator
630 if (elevator_attach(q
, new_e
, e
))
633 if (elv_register_queue(q
))
637 * finally exit old elevator and start queue again
639 elevator_exit(old_elevator
);
640 blk_finish_queue_drain(q
);
641 elevator_put(noop_elevator
);
646 * switch failed, exit the new io scheduler and reattach the old
647 * one again (along with re-adding the sysfs dir)
651 q
->elevator
= old_elevator
;
652 elv_register_queue(q
);
653 blk_finish_queue_drain(q
);
656 elevator_put(noop_elevator
);
658 printk(KERN_ERR
"elevator: switch to %s failed\n",new_e
->elevator_name
);
661 ssize_t
elv_iosched_store(request_queue_t
*q
, const char *name
, size_t count
)
663 char elevator_name
[ELV_NAME_MAX
];
664 struct elevator_type
*e
;
666 memset(elevator_name
, 0, sizeof(elevator_name
));
667 strncpy(elevator_name
, name
, sizeof(elevator_name
));
669 if (elevator_name
[strlen(elevator_name
) - 1] == '\n')
670 elevator_name
[strlen(elevator_name
) - 1] = '\0';
672 e
= elevator_get(elevator_name
);
674 printk(KERN_ERR
"elevator: type %s not found\n", elevator_name
);
678 if (!strcmp(elevator_name
, q
->elevator
->elevator_type
->elevator_name
))
681 elevator_switch(q
, e
);
685 ssize_t
elv_iosched_show(request_queue_t
*q
, char *name
)
687 elevator_t
*e
= q
->elevator
;
688 struct elevator_type
*elv
= e
->elevator_type
;
689 struct list_head
*entry
;
692 spin_lock_irq(q
->queue_lock
);
693 list_for_each(entry
, &elv_list
) {
694 struct elevator_type
*__e
;
696 __e
= list_entry(entry
, struct elevator_type
, list
);
697 if (!strcmp(elv
->elevator_name
, __e
->elevator_name
))
698 len
+= sprintf(name
+len
, "[%s] ", elv
->elevator_name
);
700 len
+= sprintf(name
+len
, "%s ", __e
->elevator_name
);
702 spin_unlock_irq(q
->queue_lock
);
704 len
+= sprintf(len
+name
, "\n");
708 EXPORT_SYMBOL(elv_add_request
);
709 EXPORT_SYMBOL(__elv_add_request
);
710 EXPORT_SYMBOL(elv_requeue_request
);
711 EXPORT_SYMBOL(elv_next_request
);
712 EXPORT_SYMBOL(elv_remove_request
);
713 EXPORT_SYMBOL(elv_queue_empty
);
714 EXPORT_SYMBOL(elv_completed_request
);
715 EXPORT_SYMBOL(elevator_exit
);
716 EXPORT_SYMBOL(elevator_init
);