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);
43 static inline sector_t rq_last_sector(struct request *rq)
45 return rq->sector + rq->nr_sectors;
49 * can we safely merge with this request?
51 inline int elv_rq_merge_ok(struct request *rq, struct bio *bio)
53 if (!rq_mergeable(rq))
57 * different data direction or already started, don't merge
59 if (bio_data_dir(bio) != rq_data_dir(rq))
63 * same device and no special stuff set, merge is ok
65 if (rq->rq_disk == bio->bi_bdev->bd_disk &&
66 !rq->waiting && !rq->special)
71 EXPORT_SYMBOL(elv_rq_merge_ok);
73 inline int elv_try_merge(struct request *__rq, struct bio *bio)
75 int ret = ELEVATOR_NO_MERGE;
78 * we can merge and sequence is ok, check if it's possible
80 if (elv_rq_merge_ok(__rq, bio)) {
81 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
82 ret = ELEVATOR_BACK_MERGE;
83 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
84 ret = ELEVATOR_FRONT_MERGE;
89 EXPORT_SYMBOL(elv_try_merge);
91 inline int elv_try_last_merge(request_queue_t *q, struct bio *bio)
94 return elv_try_merge(q->last_merge, bio);
96 return ELEVATOR_NO_MERGE;
98 EXPORT_SYMBOL(elv_try_last_merge);
100 static struct elevator_type *elevator_find(const char *name)
102 struct elevator_type *e = NULL;
103 struct list_head *entry;
105 list_for_each(entry, &elv_list) {
106 struct elevator_type *__e;
108 __e = list_entry(entry, struct elevator_type, list);
110 if (!strcmp(__e->elevator_name, name)) {
119 static void elevator_put(struct elevator_type *e)
121 module_put(e->elevator_owner);
124 static struct elevator_type *elevator_get(const char *name)
126 struct elevator_type *e;
128 spin_lock_irq(&elv_list_lock);
130 e = elevator_find(name);
131 if (e && !try_module_get(e->elevator_owner))
134 spin_unlock_irq(&elv_list_lock);
139 static int elevator_attach(request_queue_t *q, struct elevator_type *e,
140 struct elevator_queue *eq)
144 memset(eq, 0, sizeof(*eq));
146 eq->elevator_type = e;
148 INIT_LIST_HEAD(&q->queue_head);
149 q->last_merge = NULL;
152 q->boundary_rq = NULL;
155 if (eq->ops->elevator_init_fn)
156 ret = eq->ops->elevator_init_fn(q, eq);
161 static char chosen_elevator[16];
163 static void elevator_setup_default(void)
165 struct elevator_type *e;
168 * check if default is set and exists
170 if (chosen_elevator[0] && (e = elevator_get(chosen_elevator))) {
175 #if defined(CONFIG_IOSCHED_AS)
176 strcpy(chosen_elevator, "anticipatory");
177 #elif defined(CONFIG_IOSCHED_DEADLINE)
178 strcpy(chosen_elevator, "deadline");
179 #elif defined(CONFIG_IOSCHED_CFQ)
180 strcpy(chosen_elevator, "cfq");
181 #elif defined(CONFIG_IOSCHED_NOOP)
182 strcpy(chosen_elevator, "noop");
184 #error "You must build at least 1 IO scheduler into the kernel"
188 static int __init elevator_setup(char *str)
190 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
194 __setup("elevator=", elevator_setup);
196 int elevator_init(request_queue_t *q, char *name)
198 struct elevator_type *e = NULL;
199 struct elevator_queue *eq;
202 elevator_setup_default();
205 name = chosen_elevator;
207 e = elevator_get(name);
211 eq = kmalloc(sizeof(struct elevator_queue), GFP_KERNEL);
213 elevator_put(e->elevator_type);
217 ret = elevator_attach(q, e, eq);
220 elevator_put(e->elevator_type);
226 void elevator_exit(elevator_t *e)
228 if (e->ops->elevator_exit_fn)
229 e->ops->elevator_exit_fn(e);
231 elevator_put(e->elevator_type);
232 e->elevator_type = NULL;
237 * Insert rq into dispatch queue of q. Queue lock must be held on
238 * entry. If sort != 0, rq is sort-inserted; otherwise, rq will be
239 * appended to the dispatch queue. To be used by specific elevators.
241 void elv_dispatch_insert(request_queue_t *q, struct request *rq, int sort)
245 struct list_head *entry;
248 /* Specific elevator is performing sort. Step away. */
249 q->last_sector = rq_last_sector(rq);
251 list_add_tail(&rq->queuelist, &q->queue_head);
255 boundary = q->last_sector;
256 max_back = q->max_back_kb * 2;
257 boundary = boundary > max_back ? boundary - max_back : 0;
259 list_for_each_prev(entry, &q->queue_head) {
260 struct request *pos = list_entry_rq(entry);
262 if (pos->flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
264 if (rq->sector >= boundary) {
265 if (pos->sector < boundary)
268 if (pos->sector >= boundary)
271 if (rq->sector >= pos->sector)
275 list_add(&rq->queuelist, entry);
278 int elv_merge(request_queue_t *q, struct request **req, struct bio *bio)
280 elevator_t *e = q->elevator;
282 if (e->ops->elevator_merge_fn)
283 return e->ops->elevator_merge_fn(q, req, bio);
285 return ELEVATOR_NO_MERGE;
288 void elv_merged_request(request_queue_t *q, struct request *rq)
290 elevator_t *e = q->elevator;
292 if (e->ops->elevator_merged_fn)
293 e->ops->elevator_merged_fn(q, rq);
296 void elv_merge_requests(request_queue_t *q, struct request *rq,
297 struct request *next)
299 elevator_t *e = q->elevator;
301 if (q->last_merge == next)
302 q->last_merge = NULL;
304 if (e->ops->elevator_merge_req_fn)
305 e->ops->elevator_merge_req_fn(q, rq, next);
308 void elv_requeue_request(request_queue_t *q, struct request *rq)
310 elevator_t *e = q->elevator;
313 * it already went through dequeue, we need to decrement the
314 * in_flight count again
316 if (blk_account_rq(rq)) {
318 if (blk_sorted_rq(rq) && e->ops->elevator_deactivate_req_fn)
319 e->ops->elevator_deactivate_req_fn(q, rq);
322 rq->flags &= ~REQ_STARTED;
325 * if this is the flush, requeue the original instead and drop the flush
327 if (rq->flags & REQ_BAR_FLUSH) {
328 clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags);
329 rq = rq->end_io_data;
332 __elv_add_request(q, rq, ELEVATOR_INSERT_FRONT, 0);
335 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
338 if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
340 * barriers implicitly indicate back insertion
342 if (where == ELEVATOR_INSERT_SORT)
343 where = ELEVATOR_INSERT_BACK;
346 * this request is scheduling boundary, update last_sector
348 if (blk_fs_request(rq)) {
349 q->last_sector = rq_last_sector(rq);
359 if (unlikely(test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags))) {
361 * if drain is set, store the request "locally". when the drain
362 * is finished, the requests will be handed ordered to the io
365 list_add_tail(&rq->queuelist, &q->drain_list);
370 case ELEVATOR_INSERT_FRONT:
371 rq->flags |= REQ_SOFTBARRIER;
373 list_add(&rq->queuelist, &q->queue_head);
376 case ELEVATOR_INSERT_BACK:
377 rq->flags |= REQ_SOFTBARRIER;
379 while (q->elevator->ops->elevator_dispatch_fn(q, 1))
381 list_add_tail(&rq->queuelist, &q->queue_head);
383 * We kick the queue here for the following reasons.
384 * - The elevator might have returned NULL previously
385 * to delay requests and returned them now. As the
386 * queue wasn't empty before this request, ll_rw_blk
387 * won't run the queue on return, resulting in hang.
388 * - Usually, back inserted requests won't be merged
389 * with anything. There's no point in delaying queue
396 case ELEVATOR_INSERT_SORT:
397 BUG_ON(!blk_fs_request(rq));
398 rq->flags |= REQ_SORTED;
399 q->elevator->ops->elevator_add_req_fn(q, rq);
403 printk(KERN_ERR "%s: bad insertion point %d\n",
404 __FUNCTION__, where);
408 if (blk_queue_plugged(q)) {
409 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
412 if (nrq >= q->unplug_thresh)
413 __generic_unplug_device(q);
417 void elv_add_request(request_queue_t *q, struct request *rq, int where,
422 spin_lock_irqsave(q->queue_lock, flags);
423 __elv_add_request(q, rq, where, plug);
424 spin_unlock_irqrestore(q->queue_lock, flags);
427 static inline struct request *__elv_next_request(request_queue_t *q)
431 if (unlikely(list_empty(&q->queue_head) &&
432 !q->elevator->ops->elevator_dispatch_fn(q, 0)))
435 rq = list_entry_rq(q->queue_head.next);
438 * if this is a barrier write and the device has to issue a
439 * flush sequence to support it, check how far we are
441 if (blk_fs_request(rq) && blk_barrier_rq(rq)) {
442 BUG_ON(q->ordered == QUEUE_ORDERED_NONE);
444 if (q->ordered == QUEUE_ORDERED_FLUSH &&
445 !blk_barrier_preflush(rq))
446 rq = blk_start_pre_flush(q, rq);
452 struct request *elv_next_request(request_queue_t *q)
457 while ((rq = __elv_next_request(q)) != NULL) {
458 if (!(rq->flags & REQ_STARTED)) {
459 elevator_t *e = q->elevator;
462 * This is the first time the device driver
463 * sees this request (possibly after
464 * requeueing). Notify IO scheduler.
466 if (blk_sorted_rq(rq) &&
467 e->ops->elevator_activate_req_fn)
468 e->ops->elevator_activate_req_fn(q, rq);
471 * just mark as started even if we don't start
472 * it, a request that has been delayed should
473 * not be passed by new incoming requests
475 rq->flags |= REQ_STARTED;
478 if (rq == q->last_merge)
479 q->last_merge = NULL;
481 if (!q->boundary_rq || q->boundary_rq == rq) {
482 q->last_sector = rq_last_sector(rq);
483 q->boundary_rq = NULL;
486 if ((rq->flags & REQ_DONTPREP) || !q->prep_rq_fn)
489 ret = q->prep_rq_fn(q, rq);
490 if (ret == BLKPREP_OK) {
492 } else if (ret == BLKPREP_DEFER) {
494 * the request may have been (partially) prepped.
495 * we need to keep this request in the front to
496 * avoid resource deadlock. REQ_STARTED will
497 * prevent other fs requests from passing this one.
501 } else if (ret == BLKPREP_KILL) {
502 int nr_bytes = rq->hard_nr_sectors << 9;
505 nr_bytes = rq->data_len;
507 blkdev_dequeue_request(rq);
508 rq->flags |= REQ_QUIET;
509 end_that_request_chunk(rq, 0, nr_bytes);
510 end_that_request_last(rq);
512 printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
521 void elv_dequeue_request(request_queue_t *q, struct request *rq)
523 BUG_ON(list_empty(&rq->queuelist));
525 list_del_init(&rq->queuelist);
528 * the time frame between a request being removed from the lists
529 * and to it is freed is accounted as io that is in progress at
532 if (blk_account_rq(rq))
536 * the main clearing point for q->last_merge is on retrieval of
537 * request by driver (it calls elv_next_request()), but it _can_
538 * also happen here if a request is added to the queue but later
539 * deleted without ever being given to driver (merged with another
542 if (rq == q->last_merge)
543 q->last_merge = NULL;
546 int elv_queue_empty(request_queue_t *q)
548 elevator_t *e = q->elevator;
550 if (!list_empty(&q->queue_head))
553 if (e->ops->elevator_queue_empty_fn)
554 return e->ops->elevator_queue_empty_fn(q);
559 struct request *elv_latter_request(request_queue_t *q, struct request *rq)
561 struct list_head *next;
563 elevator_t *e = q->elevator;
565 if (e->ops->elevator_latter_req_fn)
566 return e->ops->elevator_latter_req_fn(q, rq);
568 next = rq->queuelist.next;
569 if (next != &q->queue_head && next != &rq->queuelist)
570 return list_entry_rq(next);
575 struct request *elv_former_request(request_queue_t *q, struct request *rq)
577 struct list_head *prev;
579 elevator_t *e = q->elevator;
581 if (e->ops->elevator_former_req_fn)
582 return e->ops->elevator_former_req_fn(q, rq);
584 prev = rq->queuelist.prev;
585 if (prev != &q->queue_head && prev != &rq->queuelist)
586 return list_entry_rq(prev);
591 int elv_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
594 elevator_t *e = q->elevator;
596 if (e->ops->elevator_set_req_fn)
597 return e->ops->elevator_set_req_fn(q, rq, bio, gfp_mask);
599 rq->elevator_private = NULL;
603 void elv_put_request(request_queue_t *q, struct request *rq)
605 elevator_t *e = q->elevator;
607 if (e->ops->elevator_put_req_fn)
608 e->ops->elevator_put_req_fn(q, rq);
611 int elv_may_queue(request_queue_t *q, int rw, struct bio *bio)
613 elevator_t *e = q->elevator;
615 if (e->ops->elevator_may_queue_fn)
616 return e->ops->elevator_may_queue_fn(q, rw, bio);
618 return ELV_MQUEUE_MAY;
621 void elv_completed_request(request_queue_t *q, struct request *rq)
623 elevator_t *e = q->elevator;
626 * request is released from the driver, io must be done
628 if (blk_account_rq(rq)) {
630 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
631 e->ops->elevator_completed_req_fn(q, rq);
635 int elv_register_queue(struct request_queue *q)
637 elevator_t *e = q->elevator;
639 e->kobj.parent = kobject_get(&q->kobj);
643 snprintf(e->kobj.name, KOBJ_NAME_LEN, "%s", "iosched");
644 e->kobj.ktype = e->elevator_type->elevator_ktype;
646 return kobject_register(&e->kobj);
649 void elv_unregister_queue(struct request_queue *q)
652 elevator_t *e = q->elevator;
653 kobject_unregister(&e->kobj);
654 kobject_put(&q->kobj);
658 int elv_register(struct elevator_type *e)
660 spin_lock_irq(&elv_list_lock);
661 if (elevator_find(e->elevator_name))
663 list_add_tail(&e->list, &elv_list);
664 spin_unlock_irq(&elv_list_lock);
666 printk(KERN_INFO "io scheduler %s registered", e->elevator_name);
667 if (!strcmp(e->elevator_name, chosen_elevator))
668 printk(" (default)");
672 EXPORT_SYMBOL_GPL(elv_register);
674 void elv_unregister(struct elevator_type *e)
676 spin_lock_irq(&elv_list_lock);
677 list_del_init(&e->list);
678 spin_unlock_irq(&elv_list_lock);
680 EXPORT_SYMBOL_GPL(elv_unregister);
683 * switch to new_e io scheduler. be careful not to introduce deadlocks -
684 * we don't free the old io scheduler, before we have allocated what we
685 * need for the new one. this way we have a chance of going back to the old
686 * one, if the new one fails init for some reason. we also do an intermediate
687 * switch to noop to ensure safety with stack-allocated requests, since they
688 * don't originate from the block layer allocator. noop is safe here, because
689 * it never needs to touch the elevator itself for completion events. DRAIN
690 * flags will make sure we don't touch it for additions either.
692 static void elevator_switch(request_queue_t *q, struct elevator_type *new_e)
694 elevator_t *e = kmalloc(sizeof(elevator_t), GFP_KERNEL);
695 struct elevator_type *noop_elevator = NULL;
696 elevator_t *old_elevator;
702 * first step, drain requests from the block freelist
704 blk_wait_queue_drained(q, 0);
707 * unregister old elevator data
709 elv_unregister_queue(q);
710 old_elevator = q->elevator;
713 * next step, switch to noop since it uses no private rq structures
714 * and doesn't allocate any memory for anything. then wait for any
715 * non-fs requests in-flight
717 noop_elevator = elevator_get("noop");
718 spin_lock_irq(q->queue_lock);
719 elevator_attach(q, noop_elevator, e);
720 spin_unlock_irq(q->queue_lock);
722 blk_wait_queue_drained(q, 1);
725 * attach and start new elevator
727 if (elevator_attach(q, new_e, e))
730 if (elv_register_queue(q))
734 * finally exit old elevator and start queue again
736 elevator_exit(old_elevator);
737 blk_finish_queue_drain(q);
738 elevator_put(noop_elevator);
743 * switch failed, exit the new io scheduler and reattach the old
744 * one again (along with re-adding the sysfs dir)
748 q->elevator = old_elevator;
749 elv_register_queue(q);
750 blk_finish_queue_drain(q);
753 elevator_put(noop_elevator);
755 printk(KERN_ERR "elevator: switch to %s failed\n",new_e->elevator_name);
758 ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count)
760 char elevator_name[ELV_NAME_MAX];
761 struct elevator_type *e;
763 memset(elevator_name, 0, sizeof(elevator_name));
764 strncpy(elevator_name, name, sizeof(elevator_name));
766 if (elevator_name[strlen(elevator_name) - 1] == '\n')
767 elevator_name[strlen(elevator_name) - 1] = '\0';
769 e = elevator_get(elevator_name);
771 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
775 if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name))
778 elevator_switch(q, e);
782 ssize_t elv_iosched_show(request_queue_t *q, char *name)
784 elevator_t *e = q->elevator;
785 struct elevator_type *elv = e->elevator_type;
786 struct list_head *entry;
789 spin_lock_irq(q->queue_lock);
790 list_for_each(entry, &elv_list) {
791 struct elevator_type *__e;
793 __e = list_entry(entry, struct elevator_type, list);
794 if (!strcmp(elv->elevator_name, __e->elevator_name))
795 len += sprintf(name+len, "[%s] ", elv->elevator_name);
797 len += sprintf(name+len, "%s ", __e->elevator_name);
799 spin_unlock_irq(q->queue_lock);
801 len += sprintf(len+name, "\n");
805 EXPORT_SYMBOL(elv_dispatch_insert);
806 EXPORT_SYMBOL(elv_add_request);
807 EXPORT_SYMBOL(__elv_add_request);
808 EXPORT_SYMBOL(elv_requeue_request);
809 EXPORT_SYMBOL(elv_next_request);
810 EXPORT_SYMBOL(elv_dequeue_request);
811 EXPORT_SYMBOL(elv_queue_empty);
812 EXPORT_SYMBOL(elv_completed_request);
813 EXPORT_SYMBOL(elevator_exit);
814 EXPORT_SYMBOL(elevator_init);