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)))
159 strcpy(chosen_elevator, "noop");
163 static int __init elevator_setup(char *str)
165 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
169 __setup("elevator=", elevator_setup);
171 int elevator_init(request_queue_t *q, char *name)
173 struct elevator_type *e = NULL;
174 struct elevator_queue *eq;
177 INIT_LIST_HEAD(&q->queue_head);
178 q->last_merge = NULL;
180 q->boundary_rq = NULL;
182 elevator_setup_default();
185 name = chosen_elevator;
187 e = elevator_get(name);
191 eq = kmalloc(sizeof(struct elevator_queue), GFP_KERNEL);
193 elevator_put(e->elevator_type);
197 ret = elevator_attach(q, e, eq);
200 elevator_put(e->elevator_type);
206 void elevator_exit(elevator_t *e)
208 if (e->ops->elevator_exit_fn)
209 e->ops->elevator_exit_fn(e);
211 elevator_put(e->elevator_type);
212 e->elevator_type = NULL;
217 * Insert rq into dispatch queue of q. Queue lock must be held on
218 * entry. If sort != 0, rq is sort-inserted; otherwise, rq will be
219 * appended to the dispatch queue. To be used by specific elevators.
221 void elv_dispatch_sort(request_queue_t *q, struct request *rq)
224 struct list_head *entry;
226 if (q->last_merge == rq)
227 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);
290 void elv_requeue_request(request_queue_t *q, struct request *rq)
292 elevator_t *e = q->elevator;
295 * it already went through dequeue, we need to decrement the
296 * in_flight count again
298 if (blk_account_rq(rq)) {
300 if (blk_sorted_rq(rq) && e->ops->elevator_deactivate_req_fn)
301 e->ops->elevator_deactivate_req_fn(q, rq);
304 rq->flags &= ~REQ_STARTED;
307 * if this is the flush, requeue the original instead and drop the flush
309 if (rq->flags & REQ_BAR_FLUSH) {
310 clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
311 rq = rq->end_io_data;
314 __elv_add_request(q, rq, ELEVATOR_INSERT_FRONT, 0);
317 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
320 if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
322 * barriers implicitly indicate back insertion
324 if (where == ELEVATOR_INSERT_SORT)
325 where = ELEVATOR_INSERT_BACK;
328 * this request is scheduling boundary, update end_sector
330 if (blk_fs_request(rq)) {
331 q->end_sector = rq_end_sector(rq);
334 } else if (!(rq->flags & REQ_ELVPRIV) && where == ELEVATOR_INSERT_SORT)
335 where = ELEVATOR_INSERT_BACK;
343 case ELEVATOR_INSERT_FRONT:
344 rq->flags |= REQ_SOFTBARRIER;
346 list_add(&rq->queuelist, &q->queue_head);
349 case ELEVATOR_INSERT_BACK:
350 rq->flags |= REQ_SOFTBARRIER;
352 while (q->elevator->ops->elevator_dispatch_fn(q, 1))
354 list_add_tail(&rq->queuelist, &q->queue_head);
356 * We kick the queue here for the following reasons.
357 * - The elevator might have returned NULL previously
358 * to delay requests and returned them now. As the
359 * queue wasn't empty before this request, ll_rw_blk
360 * won't run the queue on return, resulting in hang.
361 * - Usually, back inserted requests won't be merged
362 * with anything. There's no point in delaying queue
369 case ELEVATOR_INSERT_SORT:
370 BUG_ON(!blk_fs_request(rq));
371 rq->flags |= REQ_SORTED;
372 if (q->last_merge == NULL && rq_mergeable(rq))
375 * Some ioscheds (cfq) run q->request_fn directly, so
376 * rq cannot be accessed after calling
377 * elevator_add_req_fn.
379 q->elevator->ops->elevator_add_req_fn(q, rq);
383 printk(KERN_ERR "%s: bad insertion point %d\n",
384 __FUNCTION__, where);
388 if (blk_queue_plugged(q)) {
389 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
392 if (nrq >= q->unplug_thresh)
393 __generic_unplug_device(q);
397 void elv_add_request(request_queue_t *q, struct request *rq, int where,
402 spin_lock_irqsave(q->queue_lock, flags);
403 __elv_add_request(q, rq, where, plug);
404 spin_unlock_irqrestore(q->queue_lock, flags);
407 static inline struct request *__elv_next_request(request_queue_t *q)
411 if (unlikely(list_empty(&q->queue_head) &&
412 !q->elevator->ops->elevator_dispatch_fn(q, 0)))
415 rq = list_entry_rq(q->queue_head.next);
418 * if this is a barrier write and the device has to issue a
419 * flush sequence to support it, check how far we are
421 if (blk_fs_request(rq) && blk_barrier_rq(rq)) {
422 BUG_ON(q->ordered == QUEUE_ORDERED_NONE);
424 if (q->ordered == QUEUE_ORDERED_FLUSH &&
425 !blk_barrier_preflush(rq))
426 rq = blk_start_pre_flush(q, rq);
432 struct request *elv_next_request(request_queue_t *q)
437 while ((rq = __elv_next_request(q)) != NULL) {
438 if (!(rq->flags & REQ_STARTED)) {
439 elevator_t *e = q->elevator;
442 * This is the first time the device driver
443 * sees this request (possibly after
444 * requeueing). Notify IO scheduler.
446 if (blk_sorted_rq(rq) &&
447 e->ops->elevator_activate_req_fn)
448 e->ops->elevator_activate_req_fn(q, rq);
451 * just mark as started even if we don't start
452 * it, a request that has been delayed should
453 * not be passed by new incoming requests
455 rq->flags |= REQ_STARTED;
458 if (!q->boundary_rq || q->boundary_rq == rq) {
459 q->end_sector = rq_end_sector(rq);
460 q->boundary_rq = NULL;
463 if ((rq->flags & REQ_DONTPREP) || !q->prep_rq_fn)
466 ret = q->prep_rq_fn(q, rq);
467 if (ret == BLKPREP_OK) {
469 } else if (ret == BLKPREP_DEFER) {
471 * the request may have been (partially) prepped.
472 * we need to keep this request in the front to
473 * avoid resource deadlock. REQ_STARTED will
474 * prevent other fs requests from passing this one.
478 } else if (ret == BLKPREP_KILL) {
479 int nr_bytes = rq->hard_nr_sectors << 9;
482 nr_bytes = rq->data_len;
484 blkdev_dequeue_request(rq);
485 rq->flags |= REQ_QUIET;
486 end_that_request_chunk(rq, 0, nr_bytes);
487 end_that_request_last(rq);
489 printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
498 void elv_dequeue_request(request_queue_t *q, struct request *rq)
500 BUG_ON(list_empty(&rq->queuelist));
502 list_del_init(&rq->queuelist);
505 * the time frame between a request being removed from the lists
506 * and to it is freed is accounted as io that is in progress at
509 if (blk_account_rq(rq))
513 int elv_queue_empty(request_queue_t *q)
515 elevator_t *e = q->elevator;
517 if (!list_empty(&q->queue_head))
520 if (e->ops->elevator_queue_empty_fn)
521 return e->ops->elevator_queue_empty_fn(q);
526 struct request *elv_latter_request(request_queue_t *q, struct request *rq)
528 struct list_head *next;
530 elevator_t *e = q->elevator;
532 if (e->ops->elevator_latter_req_fn)
533 return e->ops->elevator_latter_req_fn(q, rq);
535 next = rq->queuelist.next;
536 if (next != &q->queue_head && next != &rq->queuelist)
537 return list_entry_rq(next);
542 struct request *elv_former_request(request_queue_t *q, struct request *rq)
544 struct list_head *prev;
546 elevator_t *e = q->elevator;
548 if (e->ops->elevator_former_req_fn)
549 return e->ops->elevator_former_req_fn(q, rq);
551 prev = rq->queuelist.prev;
552 if (prev != &q->queue_head && prev != &rq->queuelist)
553 return list_entry_rq(prev);
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 while (q->elevator->ops->elevator_dispatch_fn(q, 1))
697 while (q->rq.elvpriv) {
698 spin_unlock_irq(q->queue_lock);
700 spin_lock_irq(q->queue_lock);
703 spin_unlock_irq(q->queue_lock);
706 * unregister old elevator data
708 elv_unregister_queue(q);
709 old_elevator = q->elevator;
712 * attach and start new elevator
714 if (elevator_attach(q, new_e, e))
717 if (elv_register_queue(q))
721 * finally exit old elevator and turn off BYPASS.
723 elevator_exit(old_elevator);
724 clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
729 * switch failed, exit the new io scheduler and reattach the old
730 * one again (along with re-adding the sysfs dir)
735 q->elevator = old_elevator;
736 elv_register_queue(q);
737 clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
741 printk(KERN_ERR "elevator: switch to %s failed\n",new_e->elevator_name);
744 ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count)
746 char elevator_name[ELV_NAME_MAX];
747 struct elevator_type *e;
749 memset(elevator_name, 0, sizeof(elevator_name));
750 strncpy(elevator_name, name, sizeof(elevator_name));
752 if (elevator_name[strlen(elevator_name) - 1] == '\n')
753 elevator_name[strlen(elevator_name) - 1] = '\0';
755 e = elevator_get(elevator_name);
757 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
761 if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
766 elevator_switch(q, e);
770 ssize_t elv_iosched_show(request_queue_t *q, char *name)
772 elevator_t *e = q->elevator;
773 struct elevator_type *elv = e->elevator_type;
774 struct list_head *entry;
777 spin_lock_irq(q->queue_lock);
778 list_for_each(entry, &elv_list) {
779 struct elevator_type *__e;
781 __e = list_entry(entry, struct elevator_type, list);
782 if (!strcmp(elv->elevator_name, __e->elevator_name))
783 len += sprintf(name+len, "[%s] ", elv->elevator_name);
785 len += sprintf(name+len, "%s ", __e->elevator_name);
787 spin_unlock_irq(q->queue_lock);
789 len += sprintf(len+name, "\n");
793 EXPORT_SYMBOL(elv_dispatch_sort);
794 EXPORT_SYMBOL(elv_add_request);
795 EXPORT_SYMBOL(__elv_add_request);
796 EXPORT_SYMBOL(elv_requeue_request);
797 EXPORT_SYMBOL(elv_next_request);
798 EXPORT_SYMBOL(elv_dequeue_request);
799 EXPORT_SYMBOL(elv_queue_empty);
800 EXPORT_SYMBOL(elv_completed_request);
801 EXPORT_SYMBOL(elevator_exit);
802 EXPORT_SYMBOL(elevator_init);