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 void elevator_setup_default(void)
144 struct elevator_type *e;
147 * If default has not been set, use the compiled-in selection.
149 if (!chosen_elevator[0])
150 strcpy(chosen_elevator, CONFIG_DEFAULT_IOSCHED);
153 * Be backwards-compatible with previous kernels, so users
154 * won't get the wrong elevator.
156 if (!strcmp(chosen_elevator, "as"))
157 strcpy(chosen_elevator, "anticipatory");
160 * If the given scheduler is not available, fall back to the default
162 if ((e = elevator_find(chosen_elevator)))
165 strcpy(chosen_elevator, CONFIG_DEFAULT_IOSCHED);
168 static int __init elevator_setup(char *str)
170 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
174 __setup("elevator=", elevator_setup);
176 int elevator_init(request_queue_t *q, char *name)
178 struct elevator_type *e = NULL;
179 struct elevator_queue *eq;
182 INIT_LIST_HEAD(&q->queue_head);
183 q->last_merge = NULL;
185 q->boundary_rq = NULL;
187 elevator_setup_default();
190 name = chosen_elevator;
192 e = elevator_get(name);
196 eq = kmalloc(sizeof(struct elevator_queue), GFP_KERNEL);
202 ret = elevator_attach(q, e, eq);
211 void elevator_exit(elevator_t *e)
213 if (e->ops->elevator_exit_fn)
214 e->ops->elevator_exit_fn(e);
216 elevator_put(e->elevator_type);
217 e->elevator_type = NULL;
222 * Insert rq into dispatch queue of q. Queue lock must be held on
223 * entry. If sort != 0, rq is sort-inserted; otherwise, rq will be
224 * appended to the dispatch queue. To be used by specific elevators.
226 void elv_dispatch_sort(request_queue_t *q, struct request *rq)
229 struct list_head *entry;
231 if (q->last_merge == rq)
232 q->last_merge = NULL;
235 boundary = q->end_sector;
237 list_for_each_prev(entry, &q->queue_head) {
238 struct request *pos = list_entry_rq(entry);
240 if (pos->flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
242 if (rq->sector >= boundary) {
243 if (pos->sector < boundary)
246 if (pos->sector >= boundary)
249 if (rq->sector >= pos->sector)
253 list_add(&rq->queuelist, entry);
256 int elv_merge(request_queue_t *q, struct request **req, struct bio *bio)
258 elevator_t *e = q->elevator;
262 ret = elv_try_merge(q->last_merge, bio);
263 if (ret != ELEVATOR_NO_MERGE) {
264 *req = q->last_merge;
269 if (e->ops->elevator_merge_fn)
270 return e->ops->elevator_merge_fn(q, req, bio);
272 return ELEVATOR_NO_MERGE;
275 void elv_merged_request(request_queue_t *q, struct request *rq)
277 elevator_t *e = q->elevator;
279 if (e->ops->elevator_merged_fn)
280 e->ops->elevator_merged_fn(q, rq);
285 void elv_merge_requests(request_queue_t *q, struct request *rq,
286 struct request *next)
288 elevator_t *e = q->elevator;
290 if (e->ops->elevator_merge_req_fn)
291 e->ops->elevator_merge_req_fn(q, rq, next);
297 void elv_requeue_request(request_queue_t *q, struct request *rq)
299 elevator_t *e = q->elevator;
302 * it already went through dequeue, we need to decrement the
303 * in_flight count again
305 if (blk_account_rq(rq)) {
307 if (blk_sorted_rq(rq) && e->ops->elevator_deactivate_req_fn)
308 e->ops->elevator_deactivate_req_fn(q, rq);
311 rq->flags &= ~REQ_STARTED;
313 __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE, 0);
316 static void elv_drain_elevator(request_queue_t *q)
319 while (q->elevator->ops->elevator_dispatch_fn(q, 1))
321 if (q->nr_sorted == 0)
323 if (printed++ < 10) {
324 printk(KERN_ERR "%s: forced dispatching is broken "
325 "(nr_sorted=%u), please report this\n",
326 q->elevator->elevator_type->elevator_name, q->nr_sorted);
330 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
333 struct list_head *pos;
337 rq->flags |= REQ_ORDERED_COLOR;
339 if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
341 * toggle ordered color
346 * barriers implicitly indicate back insertion
348 if (where == ELEVATOR_INSERT_SORT)
349 where = ELEVATOR_INSERT_BACK;
352 * this request is scheduling boundary, update end_sector
354 if (blk_fs_request(rq)) {
355 q->end_sector = rq_end_sector(rq);
358 } else if (!(rq->flags & REQ_ELVPRIV) && where == ELEVATOR_INSERT_SORT)
359 where = ELEVATOR_INSERT_BACK;
367 case ELEVATOR_INSERT_FRONT:
368 rq->flags |= REQ_SOFTBARRIER;
370 list_add(&rq->queuelist, &q->queue_head);
373 case ELEVATOR_INSERT_BACK:
374 rq->flags |= REQ_SOFTBARRIER;
375 elv_drain_elevator(q);
376 list_add_tail(&rq->queuelist, &q->queue_head);
378 * We kick the queue here for the following reasons.
379 * - The elevator might have returned NULL previously
380 * to delay requests and returned them now. As the
381 * queue wasn't empty before this request, ll_rw_blk
382 * won't run the queue on return, resulting in hang.
383 * - Usually, back inserted requests won't be merged
384 * with anything. There's no point in delaying queue
391 case ELEVATOR_INSERT_SORT:
392 BUG_ON(!blk_fs_request(rq));
393 rq->flags |= REQ_SORTED;
395 if (q->last_merge == NULL && rq_mergeable(rq))
398 * Some ioscheds (cfq) run q->request_fn directly, so
399 * rq cannot be accessed after calling
400 * elevator_add_req_fn.
402 q->elevator->ops->elevator_add_req_fn(q, rq);
405 case ELEVATOR_INSERT_REQUEUE:
407 * If ordered flush isn't in progress, we do front
408 * insertion; otherwise, requests should be requeued
411 rq->flags |= REQ_SOFTBARRIER;
413 if (q->ordseq == 0) {
414 list_add(&rq->queuelist, &q->queue_head);
418 ordseq = blk_ordered_req_seq(rq);
420 list_for_each(pos, &q->queue_head) {
421 struct request *pos_rq = list_entry_rq(pos);
422 if (ordseq <= blk_ordered_req_seq(pos_rq))
426 list_add_tail(&rq->queuelist, pos);
430 printk(KERN_ERR "%s: bad insertion point %d\n",
431 __FUNCTION__, where);
435 if (blk_queue_plugged(q)) {
436 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
439 if (nrq >= q->unplug_thresh)
440 __generic_unplug_device(q);
444 void elv_add_request(request_queue_t *q, struct request *rq, int where,
449 spin_lock_irqsave(q->queue_lock, flags);
450 __elv_add_request(q, rq, where, plug);
451 spin_unlock_irqrestore(q->queue_lock, flags);
454 static inline struct request *__elv_next_request(request_queue_t *q)
459 while (!list_empty(&q->queue_head)) {
460 rq = list_entry_rq(q->queue_head.next);
461 if (blk_do_ordered(q, &rq))
465 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
470 struct request *elv_next_request(request_queue_t *q)
475 while ((rq = __elv_next_request(q)) != NULL) {
476 if (!(rq->flags & REQ_STARTED)) {
477 elevator_t *e = q->elevator;
480 * This is the first time the device driver
481 * sees this request (possibly after
482 * requeueing). Notify IO scheduler.
484 if (blk_sorted_rq(rq) &&
485 e->ops->elevator_activate_req_fn)
486 e->ops->elevator_activate_req_fn(q, rq);
489 * just mark as started even if we don't start
490 * it, a request that has been delayed should
491 * not be passed by new incoming requests
493 rq->flags |= REQ_STARTED;
496 if (!q->boundary_rq || q->boundary_rq == rq) {
497 q->end_sector = rq_end_sector(rq);
498 q->boundary_rq = NULL;
501 if ((rq->flags & REQ_DONTPREP) || !q->prep_rq_fn)
504 ret = q->prep_rq_fn(q, rq);
505 if (ret == BLKPREP_OK) {
507 } else if (ret == BLKPREP_DEFER) {
509 * the request may have been (partially) prepped.
510 * we need to keep this request in the front to
511 * avoid resource deadlock. REQ_STARTED will
512 * prevent other fs requests from passing this one.
516 } else if (ret == BLKPREP_KILL) {
517 int nr_bytes = rq->hard_nr_sectors << 9;
520 nr_bytes = rq->data_len;
522 blkdev_dequeue_request(rq);
523 rq->flags |= REQ_QUIET;
524 end_that_request_chunk(rq, 0, nr_bytes);
525 end_that_request_last(rq, 0);
527 printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
536 void elv_dequeue_request(request_queue_t *q, struct request *rq)
538 BUG_ON(list_empty(&rq->queuelist));
540 list_del_init(&rq->queuelist);
543 * the time frame between a request being removed from the lists
544 * and to it is freed is accounted as io that is in progress at
547 if (blk_account_rq(rq))
551 int elv_queue_empty(request_queue_t *q)
553 elevator_t *e = q->elevator;
555 if (!list_empty(&q->queue_head))
558 if (e->ops->elevator_queue_empty_fn)
559 return e->ops->elevator_queue_empty_fn(q);
564 struct request *elv_latter_request(request_queue_t *q, struct request *rq)
566 elevator_t *e = q->elevator;
568 if (e->ops->elevator_latter_req_fn)
569 return e->ops->elevator_latter_req_fn(q, rq);
573 struct request *elv_former_request(request_queue_t *q, struct request *rq)
575 elevator_t *e = q->elevator;
577 if (e->ops->elevator_former_req_fn)
578 return e->ops->elevator_former_req_fn(q, rq);
582 int elv_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
585 elevator_t *e = q->elevator;
587 if (e->ops->elevator_set_req_fn)
588 return e->ops->elevator_set_req_fn(q, rq, bio, gfp_mask);
590 rq->elevator_private = NULL;
594 void elv_put_request(request_queue_t *q, struct request *rq)
596 elevator_t *e = q->elevator;
598 if (e->ops->elevator_put_req_fn)
599 e->ops->elevator_put_req_fn(q, rq);
602 int elv_may_queue(request_queue_t *q, int rw, struct bio *bio)
604 elevator_t *e = q->elevator;
606 if (e->ops->elevator_may_queue_fn)
607 return e->ops->elevator_may_queue_fn(q, rw, bio);
609 return ELV_MQUEUE_MAY;
612 void elv_completed_request(request_queue_t *q, struct request *rq)
614 elevator_t *e = q->elevator;
617 * request is released from the driver, io must be done
619 if (blk_account_rq(rq)) {
621 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
622 e->ops->elevator_completed_req_fn(q, rq);
626 * Check if the queue is waiting for fs requests to be
627 * drained for flush sequence.
629 if (unlikely(q->ordseq)) {
630 struct request *first_rq = list_entry_rq(q->queue_head.next);
631 if (q->in_flight == 0 &&
632 blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
633 blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
634 blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
640 int elv_register_queue(struct request_queue *q)
642 elevator_t *e = q->elevator;
644 e->kobj.parent = kobject_get(&q->kobj);
648 snprintf(e->kobj.name, KOBJ_NAME_LEN, "%s", "iosched");
649 e->kobj.ktype = e->elevator_type->elevator_ktype;
651 return kobject_register(&e->kobj);
654 void elv_unregister_queue(struct request_queue *q)
657 elevator_t *e = q->elevator;
658 kobject_unregister(&e->kobj);
659 kobject_put(&q->kobj);
663 int elv_register(struct elevator_type *e)
665 spin_lock_irq(&elv_list_lock);
666 if (elevator_find(e->elevator_name))
668 list_add_tail(&e->list, &elv_list);
669 spin_unlock_irq(&elv_list_lock);
671 printk(KERN_INFO "io scheduler %s registered", e->elevator_name);
672 if (!strcmp(e->elevator_name, chosen_elevator))
673 printk(" (default)");
677 EXPORT_SYMBOL_GPL(elv_register);
679 void elv_unregister(struct elevator_type *e)
681 struct task_struct *g, *p;
684 * Iterate every thread in the process to remove the io contexts.
686 read_lock(&tasklist_lock);
687 do_each_thread(g, p) {
688 struct io_context *ioc = p->io_context;
689 if (ioc && ioc->cic) {
690 ioc->cic->exit(ioc->cic);
691 ioc->cic->dtor(ioc->cic);
694 if (ioc && ioc->aic) {
695 ioc->aic->exit(ioc->aic);
696 ioc->aic->dtor(ioc->aic);
699 } while_each_thread(g, p);
700 read_unlock(&tasklist_lock);
702 spin_lock_irq(&elv_list_lock);
703 list_del_init(&e->list);
704 spin_unlock_irq(&elv_list_lock);
706 EXPORT_SYMBOL_GPL(elv_unregister);
709 * switch to new_e io scheduler. be careful not to introduce deadlocks -
710 * we don't free the old io scheduler, before we have allocated what we
711 * need for the new one. this way we have a chance of going back to the old
712 * one, if the new one fails init for some reason.
714 static void elevator_switch(request_queue_t *q, struct elevator_type *new_e)
716 elevator_t *old_elevator, *e;
719 * Allocate new elevator
721 e = kmalloc(sizeof(elevator_t), GFP_KERNEL);
726 * Turn on BYPASS and drain all requests w/ elevator private data
728 spin_lock_irq(q->queue_lock);
730 set_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
732 elv_drain_elevator(q);
734 while (q->rq.elvpriv) {
737 spin_unlock_irq(q->queue_lock);
739 spin_lock_irq(q->queue_lock);
740 elv_drain_elevator(q);
743 spin_unlock_irq(q->queue_lock);
746 * unregister old elevator data
748 elv_unregister_queue(q);
749 old_elevator = q->elevator;
752 * attach and start new elevator
754 if (elevator_attach(q, new_e, e))
757 if (elv_register_queue(q))
761 * finally exit old elevator and turn off BYPASS.
763 elevator_exit(old_elevator);
764 clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
769 * switch failed, exit the new io scheduler and reattach the old
770 * one again (along with re-adding the sysfs dir)
775 q->elevator = old_elevator;
776 elv_register_queue(q);
777 clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
781 printk(KERN_ERR "elevator: switch to %s failed\n",new_e->elevator_name);
784 ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count)
786 char elevator_name[ELV_NAME_MAX];
788 struct elevator_type *e;
790 elevator_name[sizeof(elevator_name) - 1] = '\0';
791 strncpy(elevator_name, name, sizeof(elevator_name) - 1);
792 len = strlen(elevator_name);
794 if (len && elevator_name[len - 1] == '\n')
795 elevator_name[len - 1] = '\0';
797 e = elevator_get(elevator_name);
799 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
803 if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
808 elevator_switch(q, e);
812 ssize_t elv_iosched_show(request_queue_t *q, char *name)
814 elevator_t *e = q->elevator;
815 struct elevator_type *elv = e->elevator_type;
816 struct list_head *entry;
819 spin_lock_irq(q->queue_lock);
820 list_for_each(entry, &elv_list) {
821 struct elevator_type *__e;
823 __e = list_entry(entry, struct elevator_type, list);
824 if (!strcmp(elv->elevator_name, __e->elevator_name))
825 len += sprintf(name+len, "[%s] ", elv->elevator_name);
827 len += sprintf(name+len, "%s ", __e->elevator_name);
829 spin_unlock_irq(q->queue_lock);
831 len += sprintf(len+name, "\n");
835 EXPORT_SYMBOL(elv_dispatch_sort);
836 EXPORT_SYMBOL(elv_add_request);
837 EXPORT_SYMBOL(__elv_add_request);
838 EXPORT_SYMBOL(elv_requeue_request);
839 EXPORT_SYMBOL(elv_next_request);
840 EXPORT_SYMBOL(elv_dequeue_request);
841 EXPORT_SYMBOL(elv_queue_empty);
842 EXPORT_SYMBOL(elv_completed_request);
843 EXPORT_SYMBOL(elevator_exit);
844 EXPORT_SYMBOL(elevator_init);