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_type *e,
124 struct elevator_queue *eq)
128 memset(eq, 0, sizeof(*eq));
130 eq->elevator_type = e;
134 if (eq->ops->elevator_init_fn)
135 ret = eq->ops->elevator_init_fn(q, eq);
140 static char chosen_elevator[16];
142 static int __init elevator_setup(char *str)
145 * Be backwards-compatible with previous kernels, so users
146 * won't get the wrong elevator.
148 if (!strcmp(str, "as"))
149 strcpy(chosen_elevator, "anticipatory");
151 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
155 __setup("elevator=", elevator_setup);
157 int elevator_init(request_queue_t *q, char *name)
159 struct elevator_type *e = NULL;
160 struct elevator_queue *eq;
163 INIT_LIST_HEAD(&q->queue_head);
164 q->last_merge = NULL;
166 q->boundary_rq = NULL;
168 if (name && !(e = elevator_get(name)))
171 if (!e && *chosen_elevator && !(e = elevator_get(chosen_elevator)))
172 printk("I/O scheduler %s not found\n", chosen_elevator);
174 if (!e && !(e = elevator_get(CONFIG_DEFAULT_IOSCHED))) {
175 printk("Default I/O scheduler not found, using no-op\n");
176 e = elevator_get("noop");
179 eq = kmalloc(sizeof(struct elevator_queue), GFP_KERNEL);
185 ret = elevator_attach(q, e, eq);
194 void elevator_exit(elevator_t *e)
196 if (e->ops->elevator_exit_fn)
197 e->ops->elevator_exit_fn(e);
199 elevator_put(e->elevator_type);
200 e->elevator_type = NULL;
205 * Insert rq into dispatch queue of q. Queue lock must be held on
206 * entry. If sort != 0, rq is sort-inserted; otherwise, rq will be
207 * appended to the dispatch queue. To be used by specific elevators.
209 void elv_dispatch_sort(request_queue_t *q, struct request *rq)
212 struct list_head *entry;
214 if (q->last_merge == rq)
215 q->last_merge = NULL;
218 boundary = q->end_sector;
220 list_for_each_prev(entry, &q->queue_head) {
221 struct request *pos = list_entry_rq(entry);
223 if (pos->flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
225 if (rq->sector >= boundary) {
226 if (pos->sector < boundary)
229 if (pos->sector >= boundary)
232 if (rq->sector >= pos->sector)
236 list_add(&rq->queuelist, entry);
239 int elv_merge(request_queue_t *q, struct request **req, struct bio *bio)
241 elevator_t *e = q->elevator;
245 ret = elv_try_merge(q->last_merge, bio);
246 if (ret != ELEVATOR_NO_MERGE) {
247 *req = q->last_merge;
252 if (e->ops->elevator_merge_fn)
253 return e->ops->elevator_merge_fn(q, req, bio);
255 return ELEVATOR_NO_MERGE;
258 void elv_merged_request(request_queue_t *q, struct request *rq)
260 elevator_t *e = q->elevator;
262 if (e->ops->elevator_merged_fn)
263 e->ops->elevator_merged_fn(q, rq);
268 void elv_merge_requests(request_queue_t *q, struct request *rq,
269 struct request *next)
271 elevator_t *e = q->elevator;
273 if (e->ops->elevator_merge_req_fn)
274 e->ops->elevator_merge_req_fn(q, rq, next);
280 void elv_requeue_request(request_queue_t *q, struct request *rq)
282 elevator_t *e = q->elevator;
285 * it already went through dequeue, we need to decrement the
286 * in_flight count again
288 if (blk_account_rq(rq)) {
290 if (blk_sorted_rq(rq) && e->ops->elevator_deactivate_req_fn)
291 e->ops->elevator_deactivate_req_fn(q, rq);
294 rq->flags &= ~REQ_STARTED;
296 elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
299 static void elv_drain_elevator(request_queue_t *q)
302 while (q->elevator->ops->elevator_dispatch_fn(q, 1))
304 if (q->nr_sorted == 0)
306 if (printed++ < 10) {
307 printk(KERN_ERR "%s: forced dispatching is broken "
308 "(nr_sorted=%u), please report this\n",
309 q->elevator->elevator_type->elevator_name, q->nr_sorted);
313 void elv_insert(request_queue_t *q, struct request *rq, int where)
315 struct list_head *pos;
322 case ELEVATOR_INSERT_FRONT:
323 rq->flags |= REQ_SOFTBARRIER;
325 list_add(&rq->queuelist, &q->queue_head);
328 case ELEVATOR_INSERT_BACK:
329 rq->flags |= REQ_SOFTBARRIER;
330 elv_drain_elevator(q);
331 list_add_tail(&rq->queuelist, &q->queue_head);
333 * We kick the queue here for the following reasons.
334 * - The elevator might have returned NULL previously
335 * to delay requests and returned them now. As the
336 * queue wasn't empty before this request, ll_rw_blk
337 * won't run the queue on return, resulting in hang.
338 * - Usually, back inserted requests won't be merged
339 * with anything. There's no point in delaying queue
346 case ELEVATOR_INSERT_SORT:
347 BUG_ON(!blk_fs_request(rq));
348 rq->flags |= REQ_SORTED;
350 if (q->last_merge == NULL && rq_mergeable(rq))
353 * Some ioscheds (cfq) run q->request_fn directly, so
354 * rq cannot be accessed after calling
355 * elevator_add_req_fn.
357 q->elevator->ops->elevator_add_req_fn(q, rq);
360 case ELEVATOR_INSERT_REQUEUE:
362 * If ordered flush isn't in progress, we do front
363 * insertion; otherwise, requests should be requeued
366 rq->flags |= REQ_SOFTBARRIER;
368 if (q->ordseq == 0) {
369 list_add(&rq->queuelist, &q->queue_head);
373 ordseq = blk_ordered_req_seq(rq);
375 list_for_each(pos, &q->queue_head) {
376 struct request *pos_rq = list_entry_rq(pos);
377 if (ordseq <= blk_ordered_req_seq(pos_rq))
381 list_add_tail(&rq->queuelist, pos);
383 * most requeues happen because of a busy condition, don't
384 * force unplug of the queue for that case.
390 printk(KERN_ERR "%s: bad insertion point %d\n",
391 __FUNCTION__, where);
395 if (unplug_it && blk_queue_plugged(q)) {
396 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
399 if (nrq >= q->unplug_thresh)
400 __generic_unplug_device(q);
404 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
408 rq->flags |= REQ_ORDERED_COLOR;
410 if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
412 * toggle ordered color
414 if (blk_barrier_rq(rq))
418 * barriers implicitly indicate back insertion
420 if (where == ELEVATOR_INSERT_SORT)
421 where = ELEVATOR_INSERT_BACK;
424 * this request is scheduling boundary, update
427 if (blk_fs_request(rq)) {
428 q->end_sector = rq_end_sector(rq);
431 } else if (!(rq->flags & REQ_ELVPRIV) && where == ELEVATOR_INSERT_SORT)
432 where = ELEVATOR_INSERT_BACK;
437 elv_insert(q, rq, where);
440 void elv_add_request(request_queue_t *q, struct request *rq, int where,
445 spin_lock_irqsave(q->queue_lock, flags);
446 __elv_add_request(q, rq, where, plug);
447 spin_unlock_irqrestore(q->queue_lock, flags);
450 static inline struct request *__elv_next_request(request_queue_t *q)
455 while (!list_empty(&q->queue_head)) {
456 rq = list_entry_rq(q->queue_head.next);
457 if (blk_do_ordered(q, &rq))
461 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
466 struct request *elv_next_request(request_queue_t *q)
471 while ((rq = __elv_next_request(q)) != NULL) {
472 if (!(rq->flags & REQ_STARTED)) {
473 elevator_t *e = q->elevator;
476 * This is the first time the device driver
477 * sees this request (possibly after
478 * requeueing). Notify IO scheduler.
480 if (blk_sorted_rq(rq) &&
481 e->ops->elevator_activate_req_fn)
482 e->ops->elevator_activate_req_fn(q, rq);
485 * just mark as started even if we don't start
486 * it, a request that has been delayed should
487 * not be passed by new incoming requests
489 rq->flags |= REQ_STARTED;
492 if (!q->boundary_rq || q->boundary_rq == rq) {
493 q->end_sector = rq_end_sector(rq);
494 q->boundary_rq = NULL;
497 if ((rq->flags & REQ_DONTPREP) || !q->prep_rq_fn)
500 ret = q->prep_rq_fn(q, rq);
501 if (ret == BLKPREP_OK) {
503 } else if (ret == BLKPREP_DEFER) {
505 * the request may have been (partially) prepped.
506 * we need to keep this request in the front to
507 * avoid resource deadlock. REQ_STARTED will
508 * prevent other fs requests from passing this one.
512 } else if (ret == BLKPREP_KILL) {
513 int nr_bytes = rq->hard_nr_sectors << 9;
516 nr_bytes = rq->data_len;
518 blkdev_dequeue_request(rq);
519 rq->flags |= REQ_QUIET;
520 end_that_request_chunk(rq, 0, nr_bytes);
521 end_that_request_last(rq, 0);
523 printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
532 void elv_dequeue_request(request_queue_t *q, struct request *rq)
534 BUG_ON(list_empty(&rq->queuelist));
536 list_del_init(&rq->queuelist);
539 * the time frame between a request being removed from the lists
540 * and to it is freed is accounted as io that is in progress at
543 if (blk_account_rq(rq))
547 int elv_queue_empty(request_queue_t *q)
549 elevator_t *e = q->elevator;
551 if (!list_empty(&q->queue_head))
554 if (e->ops->elevator_queue_empty_fn)
555 return e->ops->elevator_queue_empty_fn(q);
560 struct request *elv_latter_request(request_queue_t *q, struct request *rq)
562 elevator_t *e = q->elevator;
564 if (e->ops->elevator_latter_req_fn)
565 return e->ops->elevator_latter_req_fn(q, rq);
569 struct request *elv_former_request(request_queue_t *q, struct request *rq)
571 elevator_t *e = q->elevator;
573 if (e->ops->elevator_former_req_fn)
574 return e->ops->elevator_former_req_fn(q, rq);
578 int elv_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
581 elevator_t *e = q->elevator;
583 if (e->ops->elevator_set_req_fn)
584 return e->ops->elevator_set_req_fn(q, rq, bio, gfp_mask);
586 rq->elevator_private = NULL;
590 void elv_put_request(request_queue_t *q, struct request *rq)
592 elevator_t *e = q->elevator;
594 if (e->ops->elevator_put_req_fn)
595 e->ops->elevator_put_req_fn(q, rq);
598 int elv_may_queue(request_queue_t *q, int rw, struct bio *bio)
600 elevator_t *e = q->elevator;
602 if (e->ops->elevator_may_queue_fn)
603 return e->ops->elevator_may_queue_fn(q, rw, bio);
605 return ELV_MQUEUE_MAY;
608 void elv_completed_request(request_queue_t *q, struct request *rq)
610 elevator_t *e = q->elevator;
613 * request is released from the driver, io must be done
615 if (blk_account_rq(rq)) {
617 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
618 e->ops->elevator_completed_req_fn(q, rq);
622 * Check if the queue is waiting for fs requests to be
623 * drained for flush sequence.
625 if (unlikely(q->ordseq)) {
626 struct request *first_rq = list_entry_rq(q->queue_head.next);
627 if (q->in_flight == 0 &&
628 blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
629 blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
630 blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
636 int elv_register_queue(struct request_queue *q)
638 elevator_t *e = q->elevator;
640 e->kobj.parent = kobject_get(&q->kobj);
644 snprintf(e->kobj.name, KOBJ_NAME_LEN, "%s", "iosched");
645 e->kobj.ktype = e->elevator_type->elevator_ktype;
647 return kobject_register(&e->kobj);
650 void elv_unregister_queue(struct request_queue *q)
653 elevator_t *e = q->elevator;
654 kobject_unregister(&e->kobj);
655 kobject_put(&q->kobj);
659 int elv_register(struct elevator_type *e)
661 spin_lock_irq(&elv_list_lock);
662 if (elevator_find(e->elevator_name))
664 list_add_tail(&e->list, &elv_list);
665 spin_unlock_irq(&elv_list_lock);
667 printk(KERN_INFO "io scheduler %s registered", e->elevator_name);
668 if (!strcmp(e->elevator_name, chosen_elevator) ||
669 (!*chosen_elevator &&
670 !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
671 printk(" (default)");
675 EXPORT_SYMBOL_GPL(elv_register);
677 void elv_unregister(struct elevator_type *e)
679 struct task_struct *g, *p;
682 * Iterate every thread in the process to remove the io contexts.
684 read_lock(&tasklist_lock);
685 do_each_thread(g, p) {
686 struct io_context *ioc = p->io_context;
687 if (ioc && ioc->cic) {
688 ioc->cic->exit(ioc->cic);
689 ioc->cic->dtor(ioc->cic);
692 if (ioc && ioc->aic) {
693 ioc->aic->exit(ioc->aic);
694 ioc->aic->dtor(ioc->aic);
697 } while_each_thread(g, p);
698 read_unlock(&tasklist_lock);
700 spin_lock_irq(&elv_list_lock);
701 list_del_init(&e->list);
702 spin_unlock_irq(&elv_list_lock);
704 EXPORT_SYMBOL_GPL(elv_unregister);
707 * switch to new_e io scheduler. be careful not to introduce deadlocks -
708 * we don't free the old io scheduler, before we have allocated what we
709 * need for the new one. this way we have a chance of going back to the old
710 * one, if the new one fails init for some reason.
712 static void elevator_switch(request_queue_t *q, struct elevator_type *new_e)
714 elevator_t *old_elevator, *e;
717 * Allocate new elevator
719 e = kmalloc(sizeof(elevator_t), GFP_KERNEL);
724 * Turn on BYPASS and drain all requests w/ elevator private data
726 spin_lock_irq(q->queue_lock);
728 set_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
730 elv_drain_elevator(q);
732 while (q->rq.elvpriv) {
735 spin_unlock_irq(q->queue_lock);
737 spin_lock_irq(q->queue_lock);
738 elv_drain_elevator(q);
741 spin_unlock_irq(q->queue_lock);
744 * unregister old elevator data
746 elv_unregister_queue(q);
747 old_elevator = q->elevator;
750 * attach and start new elevator
752 if (elevator_attach(q, new_e, e))
755 if (elv_register_queue(q))
759 * finally exit old elevator and turn off BYPASS.
761 elevator_exit(old_elevator);
762 clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
767 * switch failed, exit the new io scheduler and reattach the old
768 * one again (along with re-adding the sysfs dir)
773 q->elevator = old_elevator;
774 elv_register_queue(q);
775 clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
779 printk(KERN_ERR "elevator: switch to %s failed\n",new_e->elevator_name);
782 ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count)
784 char elevator_name[ELV_NAME_MAX];
786 struct elevator_type *e;
788 elevator_name[sizeof(elevator_name) - 1] = '\0';
789 strncpy(elevator_name, name, sizeof(elevator_name) - 1);
790 len = strlen(elevator_name);
792 if (len && elevator_name[len - 1] == '\n')
793 elevator_name[len - 1] = '\0';
795 e = elevator_get(elevator_name);
797 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
801 if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
806 elevator_switch(q, e);
810 ssize_t elv_iosched_show(request_queue_t *q, char *name)
812 elevator_t *e = q->elevator;
813 struct elevator_type *elv = e->elevator_type;
814 struct list_head *entry;
817 spin_lock_irq(&elv_list_lock);
818 list_for_each(entry, &elv_list) {
819 struct elevator_type *__e;
821 __e = list_entry(entry, struct elevator_type, list);
822 if (!strcmp(elv->elevator_name, __e->elevator_name))
823 len += sprintf(name+len, "[%s] ", elv->elevator_name);
825 len += sprintf(name+len, "%s ", __e->elevator_name);
827 spin_unlock_irq(&elv_list_lock);
829 len += sprintf(len+name, "\n");
833 EXPORT_SYMBOL(elv_dispatch_sort);
834 EXPORT_SYMBOL(elv_add_request);
835 EXPORT_SYMBOL(__elv_add_request);
836 EXPORT_SYMBOL(elv_requeue_request);
837 EXPORT_SYMBOL(elv_next_request);
838 EXPORT_SYMBOL(elv_dequeue_request);
839 EXPORT_SYMBOL(elv_queue_empty);
840 EXPORT_SYMBOL(elv_completed_request);
841 EXPORT_SYMBOL(elevator_exit);
842 EXPORT_SYMBOL(elevator_init);