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_queue *eq)
129 if (eq->ops->elevator_init_fn)
130 ret = eq->ops->elevator_init_fn(q, eq);
135 static char chosen_elevator[16];
137 static int __init elevator_setup(char *str)
140 * Be backwards-compatible with previous kernels, so users
141 * won't get the wrong elevator.
143 if (!strcmp(str, "as"))
144 strcpy(chosen_elevator, "anticipatory");
146 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
150 __setup("elevator=", elevator_setup);
152 static struct kobj_type elv_ktype;
154 static elevator_t *elevator_alloc(struct elevator_type *e)
156 elevator_t *eq = kmalloc(sizeof(elevator_t), GFP_KERNEL);
158 memset(eq, 0, sizeof(*eq));
160 eq->elevator_type = e;
161 kobject_init(&eq->kobj);
162 snprintf(eq->kobj.name, KOBJ_NAME_LEN, "%s", "iosched");
163 eq->kobj.ktype = &elv_ktype;
164 mutex_init(&eq->sysfs_lock);
171 static void elevator_release(struct kobject *kobj)
173 elevator_t *e = container_of(kobj, elevator_t, kobj);
174 elevator_put(e->elevator_type);
178 int elevator_init(request_queue_t *q, char *name)
180 struct elevator_type *e = NULL;
181 struct elevator_queue *eq;
184 INIT_LIST_HEAD(&q->queue_head);
185 q->last_merge = NULL;
187 q->boundary_rq = NULL;
189 if (name && !(e = elevator_get(name)))
192 if (!e && *chosen_elevator && !(e = elevator_get(chosen_elevator)))
193 printk("I/O scheduler %s not found\n", chosen_elevator);
195 if (!e && !(e = elevator_get(CONFIG_DEFAULT_IOSCHED))) {
196 printk("Default I/O scheduler not found, using no-op\n");
197 e = elevator_get("noop");
200 eq = elevator_alloc(e);
204 ret = elevator_attach(q, eq);
206 kobject_put(&eq->kobj);
211 void elevator_exit(elevator_t *e)
213 mutex_lock(&e->sysfs_lock);
214 if (e->ops->elevator_exit_fn)
215 e->ops->elevator_exit_fn(e);
217 mutex_unlock(&e->sysfs_lock);
219 kobject_put(&e->kobj);
223 * Insert rq into dispatch queue of q. Queue lock must be held on
224 * entry. If sort != 0, rq is sort-inserted; otherwise, rq will be
225 * appended to the dispatch queue. To be used by specific elevators.
227 void elv_dispatch_sort(request_queue_t *q, struct request *rq)
230 struct list_head *entry;
232 if (q->last_merge == rq)
233 q->last_merge = NULL;
236 boundary = q->end_sector;
238 list_for_each_prev(entry, &q->queue_head) {
239 struct request *pos = list_entry_rq(entry);
241 if (pos->flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
243 if (rq->sector >= boundary) {
244 if (pos->sector < boundary)
247 if (pos->sector >= boundary)
250 if (rq->sector >= pos->sector)
254 list_add(&rq->queuelist, entry);
257 int elv_merge(request_queue_t *q, struct request **req, struct bio *bio)
259 elevator_t *e = q->elevator;
263 ret = elv_try_merge(q->last_merge, bio);
264 if (ret != ELEVATOR_NO_MERGE) {
265 *req = q->last_merge;
270 if (e->ops->elevator_merge_fn)
271 return e->ops->elevator_merge_fn(q, req, bio);
273 return ELEVATOR_NO_MERGE;
276 void elv_merged_request(request_queue_t *q, struct request *rq)
278 elevator_t *e = q->elevator;
280 if (e->ops->elevator_merged_fn)
281 e->ops->elevator_merged_fn(q, rq);
286 void elv_merge_requests(request_queue_t *q, struct request *rq,
287 struct request *next)
289 elevator_t *e = q->elevator;
291 if (e->ops->elevator_merge_req_fn)
292 e->ops->elevator_merge_req_fn(q, rq, next);
298 void elv_requeue_request(request_queue_t *q, struct request *rq)
300 elevator_t *e = q->elevator;
303 * it already went through dequeue, we need to decrement the
304 * in_flight count again
306 if (blk_account_rq(rq)) {
308 if (blk_sorted_rq(rq) && e->ops->elevator_deactivate_req_fn)
309 e->ops->elevator_deactivate_req_fn(q, rq);
312 rq->flags &= ~REQ_STARTED;
314 elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
317 static void elv_drain_elevator(request_queue_t *q)
320 while (q->elevator->ops->elevator_dispatch_fn(q, 1))
322 if (q->nr_sorted == 0)
324 if (printed++ < 10) {
325 printk(KERN_ERR "%s: forced dispatching is broken "
326 "(nr_sorted=%u), please report this\n",
327 q->elevator->elevator_type->elevator_name, q->nr_sorted);
331 void elv_insert(request_queue_t *q, struct request *rq, int where)
333 struct list_head *pos;
339 case ELEVATOR_INSERT_FRONT:
340 rq->flags |= REQ_SOFTBARRIER;
342 list_add(&rq->queuelist, &q->queue_head);
345 case ELEVATOR_INSERT_BACK:
346 rq->flags |= REQ_SOFTBARRIER;
347 elv_drain_elevator(q);
348 list_add_tail(&rq->queuelist, &q->queue_head);
350 * We kick the queue here for the following reasons.
351 * - The elevator might have returned NULL previously
352 * to delay requests and returned them now. As the
353 * queue wasn't empty before this request, ll_rw_blk
354 * won't run the queue on return, resulting in hang.
355 * - Usually, back inserted requests won't be merged
356 * with anything. There's no point in delaying queue
363 case ELEVATOR_INSERT_SORT:
364 BUG_ON(!blk_fs_request(rq));
365 rq->flags |= REQ_SORTED;
367 if (q->last_merge == NULL && rq_mergeable(rq))
370 * Some ioscheds (cfq) run q->request_fn directly, so
371 * rq cannot be accessed after calling
372 * elevator_add_req_fn.
374 q->elevator->ops->elevator_add_req_fn(q, rq);
377 case ELEVATOR_INSERT_REQUEUE:
379 * If ordered flush isn't in progress, we do front
380 * insertion; otherwise, requests should be requeued
383 rq->flags |= REQ_SOFTBARRIER;
385 if (q->ordseq == 0) {
386 list_add(&rq->queuelist, &q->queue_head);
390 ordseq = blk_ordered_req_seq(rq);
392 list_for_each(pos, &q->queue_head) {
393 struct request *pos_rq = list_entry_rq(pos);
394 if (ordseq <= blk_ordered_req_seq(pos_rq))
398 list_add_tail(&rq->queuelist, pos);
402 printk(KERN_ERR "%s: bad insertion point %d\n",
403 __FUNCTION__, where);
407 if (blk_queue_plugged(q)) {
408 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
411 if (nrq >= q->unplug_thresh)
412 __generic_unplug_device(q);
416 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
420 rq->flags |= REQ_ORDERED_COLOR;
422 if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
424 * toggle ordered color
426 if (blk_barrier_rq(rq))
430 * barriers implicitly indicate back insertion
432 if (where == ELEVATOR_INSERT_SORT)
433 where = ELEVATOR_INSERT_BACK;
436 * this request is scheduling boundary, update
439 if (blk_fs_request(rq)) {
440 q->end_sector = rq_end_sector(rq);
443 } else if (!(rq->flags & REQ_ELVPRIV) && where == ELEVATOR_INSERT_SORT)
444 where = ELEVATOR_INSERT_BACK;
449 elv_insert(q, rq, where);
452 void elv_add_request(request_queue_t *q, struct request *rq, int where,
457 spin_lock_irqsave(q->queue_lock, flags);
458 __elv_add_request(q, rq, where, plug);
459 spin_unlock_irqrestore(q->queue_lock, flags);
462 static inline struct request *__elv_next_request(request_queue_t *q)
467 while (!list_empty(&q->queue_head)) {
468 rq = list_entry_rq(q->queue_head.next);
469 if (blk_do_ordered(q, &rq))
473 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
478 struct request *elv_next_request(request_queue_t *q)
483 while ((rq = __elv_next_request(q)) != NULL) {
484 if (!(rq->flags & REQ_STARTED)) {
485 elevator_t *e = q->elevator;
488 * This is the first time the device driver
489 * sees this request (possibly after
490 * requeueing). Notify IO scheduler.
492 if (blk_sorted_rq(rq) &&
493 e->ops->elevator_activate_req_fn)
494 e->ops->elevator_activate_req_fn(q, rq);
497 * just mark as started even if we don't start
498 * it, a request that has been delayed should
499 * not be passed by new incoming requests
501 rq->flags |= REQ_STARTED;
504 if (!q->boundary_rq || q->boundary_rq == rq) {
505 q->end_sector = rq_end_sector(rq);
506 q->boundary_rq = NULL;
509 if ((rq->flags & REQ_DONTPREP) || !q->prep_rq_fn)
512 ret = q->prep_rq_fn(q, rq);
513 if (ret == BLKPREP_OK) {
515 } else if (ret == BLKPREP_DEFER) {
517 * the request may have been (partially) prepped.
518 * we need to keep this request in the front to
519 * avoid resource deadlock. REQ_STARTED will
520 * prevent other fs requests from passing this one.
524 } else if (ret == BLKPREP_KILL) {
525 int nr_bytes = rq->hard_nr_sectors << 9;
528 nr_bytes = rq->data_len;
530 blkdev_dequeue_request(rq);
531 rq->flags |= REQ_QUIET;
532 end_that_request_chunk(rq, 0, nr_bytes);
533 end_that_request_last(rq, 0);
535 printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
544 void elv_dequeue_request(request_queue_t *q, struct request *rq)
546 BUG_ON(list_empty(&rq->queuelist));
548 list_del_init(&rq->queuelist);
551 * the time frame between a request being removed from the lists
552 * and to it is freed is accounted as io that is in progress at
555 if (blk_account_rq(rq))
559 int elv_queue_empty(request_queue_t *q)
561 elevator_t *e = q->elevator;
563 if (!list_empty(&q->queue_head))
566 if (e->ops->elevator_queue_empty_fn)
567 return e->ops->elevator_queue_empty_fn(q);
572 struct request *elv_latter_request(request_queue_t *q, struct request *rq)
574 elevator_t *e = q->elevator;
576 if (e->ops->elevator_latter_req_fn)
577 return e->ops->elevator_latter_req_fn(q, rq);
581 struct request *elv_former_request(request_queue_t *q, struct request *rq)
583 elevator_t *e = q->elevator;
585 if (e->ops->elevator_former_req_fn)
586 return e->ops->elevator_former_req_fn(q, rq);
590 int elv_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
593 elevator_t *e = q->elevator;
595 if (e->ops->elevator_set_req_fn)
596 return e->ops->elevator_set_req_fn(q, rq, bio, gfp_mask);
598 rq->elevator_private = NULL;
602 void elv_put_request(request_queue_t *q, struct request *rq)
604 elevator_t *e = q->elevator;
606 if (e->ops->elevator_put_req_fn)
607 e->ops->elevator_put_req_fn(q, rq);
610 int elv_may_queue(request_queue_t *q, int rw, struct bio *bio)
612 elevator_t *e = q->elevator;
614 if (e->ops->elevator_may_queue_fn)
615 return e->ops->elevator_may_queue_fn(q, rw, bio);
617 return ELV_MQUEUE_MAY;
620 void elv_completed_request(request_queue_t *q, struct request *rq)
622 elevator_t *e = q->elevator;
625 * request is released from the driver, io must be done
627 if (blk_account_rq(rq)) {
629 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
630 e->ops->elevator_completed_req_fn(q, rq);
634 * Check if the queue is waiting for fs requests to be
635 * drained for flush sequence.
637 if (unlikely(q->ordseq)) {
638 struct request *first_rq = list_entry_rq(q->queue_head.next);
639 if (q->in_flight == 0 &&
640 blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
641 blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
642 blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
648 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
651 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
653 elevator_t *e = container_of(kobj, elevator_t, kobj);
654 struct elv_fs_entry *entry = to_elv(attr);
660 mutex_lock(&e->sysfs_lock);
661 error = e->ops ? entry->show(e, page) : -ENOENT;
662 mutex_unlock(&e->sysfs_lock);
667 elv_attr_store(struct kobject *kobj, struct attribute *attr,
668 const char *page, size_t length)
670 elevator_t *e = container_of(kobj, elevator_t, kobj);
671 struct elv_fs_entry *entry = to_elv(attr);
677 mutex_lock(&e->sysfs_lock);
678 error = e->ops ? entry->store(e, page, length) : -ENOENT;
679 mutex_unlock(&e->sysfs_lock);
683 static struct sysfs_ops elv_sysfs_ops = {
684 .show = elv_attr_show,
685 .store = elv_attr_store,
688 static struct kobj_type elv_ktype = {
689 .sysfs_ops = &elv_sysfs_ops,
690 .release = elevator_release,
693 int elv_register_queue(struct request_queue *q)
695 elevator_t *e = q->elevator;
698 e->kobj.parent = &q->kobj;
700 error = kobject_add(&e->kobj);
702 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
704 while (attr->attr.name) {
705 if (sysfs_create_file(&e->kobj, &attr->attr))
710 kobject_uevent(&e->kobj, KOBJ_ADD);
715 void elv_unregister_queue(struct request_queue *q)
718 elevator_t *e = q->elevator;
719 kobject_uevent(&e->kobj, KOBJ_REMOVE);
720 kobject_del(&e->kobj);
724 int elv_register(struct elevator_type *e)
726 spin_lock_irq(&elv_list_lock);
727 if (elevator_find(e->elevator_name))
729 list_add_tail(&e->list, &elv_list);
730 spin_unlock_irq(&elv_list_lock);
732 printk(KERN_INFO "io scheduler %s registered", e->elevator_name);
733 if (!strcmp(e->elevator_name, chosen_elevator) ||
734 (!*chosen_elevator &&
735 !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
736 printk(" (default)");
740 EXPORT_SYMBOL_GPL(elv_register);
742 void elv_unregister(struct elevator_type *e)
744 struct task_struct *g, *p;
747 * Iterate every thread in the process to remove the io contexts.
750 read_lock(&tasklist_lock);
751 do_each_thread(g, p) {
753 e->ops.trim(p->io_context);
755 } while_each_thread(g, p);
756 read_unlock(&tasklist_lock);
759 spin_lock_irq(&elv_list_lock);
760 list_del_init(&e->list);
761 spin_unlock_irq(&elv_list_lock);
763 EXPORT_SYMBOL_GPL(elv_unregister);
766 * switch to new_e io scheduler. be careful not to introduce deadlocks -
767 * we don't free the old io scheduler, before we have allocated what we
768 * need for the new one. this way we have a chance of going back to the old
769 * one, if the new one fails init for some reason.
771 static int elevator_switch(request_queue_t *q, struct elevator_type *new_e)
773 elevator_t *old_elevator, *e;
776 * Allocate new elevator
778 e = elevator_alloc(new_e);
783 * Turn on BYPASS and drain all requests w/ elevator private data
785 spin_lock_irq(q->queue_lock);
787 set_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
789 elv_drain_elevator(q);
791 while (q->rq.elvpriv) {
794 spin_unlock_irq(q->queue_lock);
796 spin_lock_irq(q->queue_lock);
797 elv_drain_elevator(q);
800 spin_unlock_irq(q->queue_lock);
803 * unregister old elevator data
805 elv_unregister_queue(q);
806 old_elevator = q->elevator;
809 * attach and start new elevator
811 if (elevator_attach(q, e))
814 if (elv_register_queue(q))
818 * finally exit old elevator and turn off BYPASS.
820 elevator_exit(old_elevator);
821 clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
826 * switch failed, exit the new io scheduler and reattach the old
827 * one again (along with re-adding the sysfs dir)
832 q->elevator = old_elevator;
833 elv_register_queue(q);
834 clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
836 kobject_put(&e->kobj);
840 ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count)
842 char elevator_name[ELV_NAME_MAX];
844 struct elevator_type *e;
846 elevator_name[sizeof(elevator_name) - 1] = '\0';
847 strncpy(elevator_name, name, sizeof(elevator_name) - 1);
848 len = strlen(elevator_name);
850 if (len && elevator_name[len - 1] == '\n')
851 elevator_name[len - 1] = '\0';
853 e = elevator_get(elevator_name);
855 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
859 if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
864 if (!elevator_switch(q, e))
865 printk(KERN_ERR "elevator: switch to %s failed\n",elevator_name);
869 ssize_t elv_iosched_show(request_queue_t *q, char *name)
871 elevator_t *e = q->elevator;
872 struct elevator_type *elv = e->elevator_type;
873 struct list_head *entry;
876 spin_lock_irq(q->queue_lock);
877 list_for_each(entry, &elv_list) {
878 struct elevator_type *__e;
880 __e = list_entry(entry, struct elevator_type, list);
881 if (!strcmp(elv->elevator_name, __e->elevator_name))
882 len += sprintf(name+len, "[%s] ", elv->elevator_name);
884 len += sprintf(name+len, "%s ", __e->elevator_name);
886 spin_unlock_irq(q->queue_lock);
888 len += sprintf(len+name, "\n");
892 EXPORT_SYMBOL(elv_dispatch_sort);
893 EXPORT_SYMBOL(elv_add_request);
894 EXPORT_SYMBOL(__elv_add_request);
895 EXPORT_SYMBOL(elv_requeue_request);
896 EXPORT_SYMBOL(elv_next_request);
897 EXPORT_SYMBOL(elv_dequeue_request);
898 EXPORT_SYMBOL(elv_queue_empty);
899 EXPORT_SYMBOL(elv_completed_request);
900 EXPORT_SYMBOL(elevator_exit);
901 EXPORT_SYMBOL(elevator_init);