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 * If default has not been set, use the compiled-in selection.
152 if (!chosen_elevator[0])
153 strcpy(chosen_elevator, CONFIG_DEFAULT_IOSCHED);
156 * If the given scheduler is not available, fall back to no-op.
158 if ((e = elevator_find(chosen_elevator)))
161 strcpy(chosen_elevator, "noop");
164 static int __init elevator_setup(char *str)
166 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
170 __setup("elevator=", elevator_setup);
172 int elevator_init(request_queue_t *q, char *name)
174 struct elevator_type *e = NULL;
175 struct elevator_queue *eq;
178 INIT_LIST_HEAD(&q->queue_head);
179 q->last_merge = NULL;
181 q->boundary_rq = NULL;
183 elevator_setup_default();
186 name = chosen_elevator;
188 e = elevator_get(name);
192 eq = kmalloc(sizeof(struct elevator_queue), GFP_KERNEL);
198 ret = elevator_attach(q, e, eq);
207 void elevator_exit(elevator_t *e)
209 if (e->ops->elevator_exit_fn)
210 e->ops->elevator_exit_fn(e);
212 elevator_put(e->elevator_type);
213 e->elevator_type = NULL;
218 * Insert rq into dispatch queue of q. Queue lock must be held on
219 * entry. If sort != 0, rq is sort-inserted; otherwise, rq will be
220 * appended to the dispatch queue. To be used by specific elevators.
222 void elv_dispatch_sort(request_queue_t *q, struct request *rq)
225 struct list_head *entry;
227 if (q->last_merge == rq)
228 q->last_merge = NULL;
230 boundary = q->end_sector;
232 list_for_each_prev(entry, &q->queue_head) {
233 struct request *pos = list_entry_rq(entry);
235 if (pos->flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
237 if (rq->sector >= boundary) {
238 if (pos->sector < boundary)
241 if (pos->sector >= boundary)
244 if (rq->sector >= pos->sector)
248 list_add(&rq->queuelist, entry);
251 int elv_merge(request_queue_t *q, struct request **req, struct bio *bio)
253 elevator_t *e = q->elevator;
257 ret = elv_try_merge(q->last_merge, bio);
258 if (ret != ELEVATOR_NO_MERGE) {
259 *req = q->last_merge;
264 if (e->ops->elevator_merge_fn)
265 return e->ops->elevator_merge_fn(q, req, bio);
267 return ELEVATOR_NO_MERGE;
270 void elv_merged_request(request_queue_t *q, struct request *rq)
272 elevator_t *e = q->elevator;
274 if (e->ops->elevator_merged_fn)
275 e->ops->elevator_merged_fn(q, rq);
280 void elv_merge_requests(request_queue_t *q, struct request *rq,
281 struct request *next)
283 elevator_t *e = q->elevator;
285 if (e->ops->elevator_merge_req_fn)
286 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 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
321 if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
323 * barriers implicitly indicate back insertion
325 if (where == ELEVATOR_INSERT_SORT)
326 where = ELEVATOR_INSERT_BACK;
329 * this request is scheduling boundary, update end_sector
331 if (blk_fs_request(rq)) {
332 q->end_sector = rq_end_sector(rq);
335 } else if (!(rq->flags & REQ_ELVPRIV) && where == ELEVATOR_INSERT_SORT)
336 where = ELEVATOR_INSERT_BACK;
344 case ELEVATOR_INSERT_FRONT:
345 rq->flags |= REQ_SOFTBARRIER;
347 list_add(&rq->queuelist, &q->queue_head);
350 case ELEVATOR_INSERT_BACK:
351 rq->flags |= REQ_SOFTBARRIER;
353 while (q->elevator->ops->elevator_dispatch_fn(q, 1))
355 list_add_tail(&rq->queuelist, &q->queue_head);
357 * We kick the queue here for the following reasons.
358 * - The elevator might have returned NULL previously
359 * to delay requests and returned them now. As the
360 * queue wasn't empty before this request, ll_rw_blk
361 * won't run the queue on return, resulting in hang.
362 * - Usually, back inserted requests won't be merged
363 * with anything. There's no point in delaying queue
370 case ELEVATOR_INSERT_SORT:
371 BUG_ON(!blk_fs_request(rq));
372 rq->flags |= REQ_SORTED;
373 if (q->last_merge == NULL && rq_mergeable(rq))
376 * Some ioscheds (cfq) run q->request_fn directly, so
377 * rq cannot be accessed after calling
378 * elevator_add_req_fn.
380 q->elevator->ops->elevator_add_req_fn(q, rq);
384 printk(KERN_ERR "%s: bad insertion point %d\n",
385 __FUNCTION__, where);
389 if (blk_queue_plugged(q)) {
390 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
393 if (nrq >= q->unplug_thresh)
394 __generic_unplug_device(q);
398 void elv_add_request(request_queue_t *q, struct request *rq, int where,
403 spin_lock_irqsave(q->queue_lock, flags);
404 __elv_add_request(q, rq, where, plug);
405 spin_unlock_irqrestore(q->queue_lock, flags);
408 static inline struct request *__elv_next_request(request_queue_t *q)
412 if (unlikely(list_empty(&q->queue_head) &&
413 !q->elevator->ops->elevator_dispatch_fn(q, 0)))
416 rq = list_entry_rq(q->queue_head.next);
419 * if this is a barrier write and the device has to issue a
420 * flush sequence to support it, check how far we are
422 if (blk_fs_request(rq) && blk_barrier_rq(rq)) {
423 BUG_ON(q->ordered == QUEUE_ORDERED_NONE);
425 if (q->ordered == QUEUE_ORDERED_FLUSH &&
426 !blk_barrier_preflush(rq))
427 rq = blk_start_pre_flush(q, rq);
433 struct request *elv_next_request(request_queue_t *q)
438 while ((rq = __elv_next_request(q)) != NULL) {
439 if (!(rq->flags & REQ_STARTED)) {
440 elevator_t *e = q->elevator;
443 * This is the first time the device driver
444 * sees this request (possibly after
445 * requeueing). Notify IO scheduler.
447 if (blk_sorted_rq(rq) &&
448 e->ops->elevator_activate_req_fn)
449 e->ops->elevator_activate_req_fn(q, rq);
452 * just mark as started even if we don't start
453 * it, a request that has been delayed should
454 * not be passed by new incoming requests
456 rq->flags |= REQ_STARTED;
459 if (!q->boundary_rq || q->boundary_rq == rq) {
460 q->end_sector = rq_end_sector(rq);
461 q->boundary_rq = NULL;
464 if ((rq->flags & REQ_DONTPREP) || !q->prep_rq_fn)
467 ret = q->prep_rq_fn(q, rq);
468 if (ret == BLKPREP_OK) {
470 } else if (ret == BLKPREP_DEFER) {
472 * the request may have been (partially) prepped.
473 * we need to keep this request in the front to
474 * avoid resource deadlock. REQ_STARTED will
475 * prevent other fs requests from passing this one.
479 } else if (ret == BLKPREP_KILL) {
480 int nr_bytes = rq->hard_nr_sectors << 9;
483 nr_bytes = rq->data_len;
485 blkdev_dequeue_request(rq);
486 rq->flags |= REQ_QUIET;
487 end_that_request_chunk(rq, 0, nr_bytes);
488 end_that_request_last(rq);
490 printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
499 void elv_dequeue_request(request_queue_t *q, struct request *rq)
501 BUG_ON(list_empty(&rq->queuelist));
503 list_del_init(&rq->queuelist);
506 * the time frame between a request being removed from the lists
507 * and to it is freed is accounted as io that is in progress at
510 if (blk_account_rq(rq))
514 int elv_queue_empty(request_queue_t *q)
516 elevator_t *e = q->elevator;
518 if (!list_empty(&q->queue_head))
521 if (e->ops->elevator_queue_empty_fn)
522 return e->ops->elevator_queue_empty_fn(q);
527 struct request *elv_latter_request(request_queue_t *q, struct request *rq)
529 struct list_head *next;
531 elevator_t *e = q->elevator;
533 if (e->ops->elevator_latter_req_fn)
534 return e->ops->elevator_latter_req_fn(q, rq);
536 next = rq->queuelist.next;
537 if (next != &q->queue_head && next != &rq->queuelist)
538 return list_entry_rq(next);
543 struct request *elv_former_request(request_queue_t *q, struct request *rq)
545 struct list_head *prev;
547 elevator_t *e = q->elevator;
549 if (e->ops->elevator_former_req_fn)
550 return e->ops->elevator_former_req_fn(q, rq);
552 prev = rq->queuelist.prev;
553 if (prev != &q->queue_head && prev != &rq->queuelist)
554 return list_entry_rq(prev);
559 int elv_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
562 elevator_t *e = q->elevator;
564 if (e->ops->elevator_set_req_fn)
565 return e->ops->elevator_set_req_fn(q, rq, bio, gfp_mask);
567 rq->elevator_private = NULL;
571 void elv_put_request(request_queue_t *q, struct request *rq)
573 elevator_t *e = q->elevator;
575 if (e->ops->elevator_put_req_fn)
576 e->ops->elevator_put_req_fn(q, rq);
579 int elv_may_queue(request_queue_t *q, int rw, struct bio *bio)
581 elevator_t *e = q->elevator;
583 if (e->ops->elevator_may_queue_fn)
584 return e->ops->elevator_may_queue_fn(q, rw, bio);
586 return ELV_MQUEUE_MAY;
589 void elv_completed_request(request_queue_t *q, struct request *rq)
591 elevator_t *e = q->elevator;
594 * request is released from the driver, io must be done
596 if (blk_account_rq(rq)) {
598 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
599 e->ops->elevator_completed_req_fn(q, rq);
603 int elv_register_queue(struct request_queue *q)
605 elevator_t *e = q->elevator;
607 e->kobj.parent = kobject_get(&q->kobj);
611 snprintf(e->kobj.name, KOBJ_NAME_LEN, "%s", "iosched");
612 e->kobj.ktype = e->elevator_type->elevator_ktype;
614 return kobject_register(&e->kobj);
617 void elv_unregister_queue(struct request_queue *q)
620 elevator_t *e = q->elevator;
621 kobject_unregister(&e->kobj);
622 kobject_put(&q->kobj);
626 int elv_register(struct elevator_type *e)
628 spin_lock_irq(&elv_list_lock);
629 if (elevator_find(e->elevator_name))
631 list_add_tail(&e->list, &elv_list);
632 spin_unlock_irq(&elv_list_lock);
634 printk(KERN_INFO "io scheduler %s registered", e->elevator_name);
635 if (!strcmp(e->elevator_name, chosen_elevator))
636 printk(" (default)");
640 EXPORT_SYMBOL_GPL(elv_register);
642 void elv_unregister(struct elevator_type *e)
644 struct task_struct *g, *p;
647 * Iterate every thread in the process to remove the io contexts.
649 read_lock(&tasklist_lock);
650 do_each_thread(g, p) {
651 struct io_context *ioc = p->io_context;
652 if (ioc && ioc->cic) {
653 ioc->cic->exit(ioc->cic);
654 ioc->cic->dtor(ioc->cic);
657 if (ioc && ioc->aic) {
658 ioc->aic->exit(ioc->aic);
659 ioc->aic->dtor(ioc->aic);
662 } while_each_thread(g, p);
663 read_unlock(&tasklist_lock);
665 spin_lock_irq(&elv_list_lock);
666 list_del_init(&e->list);
667 spin_unlock_irq(&elv_list_lock);
669 EXPORT_SYMBOL_GPL(elv_unregister);
672 * switch to new_e io scheduler. be careful not to introduce deadlocks -
673 * we don't free the old io scheduler, before we have allocated what we
674 * need for the new one. this way we have a chance of going back to the old
675 * one, if the new one fails init for some reason.
677 static void elevator_switch(request_queue_t *q, struct elevator_type *new_e)
679 elevator_t *old_elevator, *e;
682 * Allocate new elevator
684 e = kmalloc(sizeof(elevator_t), GFP_KERNEL);
689 * Turn on BYPASS and drain all requests w/ elevator private data
691 spin_lock_irq(q->queue_lock);
693 set_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
695 while (q->elevator->ops->elevator_dispatch_fn(q, 1))
698 while (q->rq.elvpriv) {
701 spin_unlock_irq(q->queue_lock);
703 spin_lock_irq(q->queue_lock);
706 spin_unlock_irq(q->queue_lock);
709 * unregister old elevator data
711 elv_unregister_queue(q);
712 old_elevator = q->elevator;
715 * attach and start new elevator
717 if (elevator_attach(q, new_e, e))
720 if (elv_register_queue(q))
724 * finally exit old elevator and turn off BYPASS.
726 elevator_exit(old_elevator);
727 clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
732 * switch failed, exit the new io scheduler and reattach the old
733 * one again (along with re-adding the sysfs dir)
738 q->elevator = old_elevator;
739 elv_register_queue(q);
740 clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
744 printk(KERN_ERR "elevator: switch to %s failed\n",new_e->elevator_name);
747 ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count)
749 char elevator_name[ELV_NAME_MAX];
750 struct elevator_type *e;
752 memset(elevator_name, 0, sizeof(elevator_name));
753 strncpy(elevator_name, name, sizeof(elevator_name));
755 if (elevator_name[strlen(elevator_name) - 1] == '\n')
756 elevator_name[strlen(elevator_name) - 1] = '\0';
758 e = elevator_get(elevator_name);
760 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
764 if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
769 elevator_switch(q, e);
773 ssize_t elv_iosched_show(request_queue_t *q, char *name)
775 elevator_t *e = q->elevator;
776 struct elevator_type *elv = e->elevator_type;
777 struct list_head *entry;
780 spin_lock_irq(q->queue_lock);
781 list_for_each(entry, &elv_list) {
782 struct elevator_type *__e;
784 __e = list_entry(entry, struct elevator_type, list);
785 if (!strcmp(elv->elevator_name, __e->elevator_name))
786 len += sprintf(name+len, "[%s] ", elv->elevator_name);
788 len += sprintf(name+len, "%s ", __e->elevator_name);
790 spin_unlock_irq(q->queue_lock);
792 len += sprintf(len+name, "\n");
796 EXPORT_SYMBOL(elv_dispatch_sort);
797 EXPORT_SYMBOL(elv_add_request);
798 EXPORT_SYMBOL(__elv_add_request);
799 EXPORT_SYMBOL(elv_requeue_request);
800 EXPORT_SYMBOL(elv_next_request);
801 EXPORT_SYMBOL(elv_dequeue_request);
802 EXPORT_SYMBOL(elv_queue_empty);
803 EXPORT_SYMBOL(elv_completed_request);
804 EXPORT_SYMBOL(elevator_exit);
805 EXPORT_SYMBOL(elevator_init);