ia64/linux-2.6.18-xen.hg

view block/elevator.c @ 417:7128fe32720e

block: backport Jens Axboe's commit from
Tue, 16 Oct 2007 09:03:56 +0000 (11:03 +0200)
bf2de6f5a4faf0197268f18d08969b003b87b6e8
Initial support for data-less (or empty) barrier support

blkback: permit and implement empty barrier.
Signed-off-by: Samuel Thibault <samuel.thibault@eu.citrix.com>
author Keir Fraser <keir.fraser@citrix.com>
date Mon Feb 11 10:08:57 2008 +0000 (2008-02-11)
parents 3e8752eb6d9c
children
line source
1 /*
2 * Block device elevator/IO-scheduler.
3 *
4 * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
5 *
6 * 30042000 Jens Axboe <axboe@suse.de> :
7 *
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
12 * an existing request
13 * - elevator_dequeue_fn, called when a request is taken off the active list
14 *
15 * 20082000 Dave Jones <davej@suse.de> :
16 * Removed tests for max-bomb-segments, which was breaking elvtune
17 * when run without -bN
18 *
19 * Jens:
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
23 *
24 */
25 #include <linux/kernel.h>
26 #include <linux/fs.h>
27 #include <linux/blkdev.h>
28 #include <linux/elevator.h>
29 #include <linux/bio.h>
30 #include <linux/module.h>
31 #include <linux/slab.h>
32 #include <linux/init.h>
33 #include <linux/compiler.h>
34 #include <linux/delay.h>
35 #include <linux/blktrace_api.h>
37 #include <asm/uaccess.h>
39 static DEFINE_SPINLOCK(elv_list_lock);
40 static LIST_HEAD(elv_list);
42 /*
43 * can we safely merge with this request?
44 */
45 inline int elv_rq_merge_ok(struct request *rq, struct bio *bio)
46 {
47 if (!rq_mergeable(rq))
48 return 0;
50 /*
51 * different data direction or already started, don't merge
52 */
53 if (bio_data_dir(bio) != rq_data_dir(rq))
54 return 0;
56 /*
57 * same device and no special stuff set, merge is ok
58 */
59 if (rq->rq_disk == bio->bi_bdev->bd_disk &&
60 !rq->waiting && !rq->special)
61 return 1;
63 return 0;
64 }
65 EXPORT_SYMBOL(elv_rq_merge_ok);
67 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
68 {
69 int ret = ELEVATOR_NO_MERGE;
71 /*
72 * we can merge and sequence is ok, check if it's possible
73 */
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;
79 }
81 return ret;
82 }
84 static struct elevator_type *elevator_find(const char *name)
85 {
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)) {
95 e = __e;
96 break;
97 }
98 }
100 return e;
101 }
103 static void elevator_put(struct elevator_type *e)
104 {
105 module_put(e->elevator_owner);
106 }
108 static struct elevator_type *elevator_get(const char *name)
109 {
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))
116 e = NULL;
118 spin_unlock_irq(&elv_list_lock);
120 return e;
121 }
123 static void *elevator_init_queue(request_queue_t *q, struct elevator_queue *eq)
124 {
125 return eq->ops->elevator_init_fn(q, eq);
126 }
128 static void elevator_attach(request_queue_t *q, struct elevator_queue *eq,
129 void *data)
130 {
131 q->elevator = eq;
132 eq->elevator_data = data;
133 }
135 static char chosen_elevator[16];
137 static int __init elevator_setup(char *str)
138 {
139 /*
140 * Be backwards-compatible with previous kernels, so users
141 * won't get the wrong elevator.
142 */
143 if (!strcmp(str, "as"))
144 strcpy(chosen_elevator, "anticipatory");
145 else
146 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
147 return 1;
148 }
150 __setup("elevator=", elevator_setup);
152 static struct kobj_type elv_ktype;
154 static elevator_t *elevator_alloc(struct elevator_type *e)
155 {
156 elevator_t *eq = kmalloc(sizeof(elevator_t), GFP_KERNEL);
157 if (eq) {
158 memset(eq, 0, sizeof(*eq));
159 eq->ops = &e->ops;
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);
165 } else {
166 elevator_put(e);
167 }
168 return eq;
169 }
171 static void elevator_release(struct kobject *kobj)
172 {
173 elevator_t *e = container_of(kobj, elevator_t, kobj);
174 elevator_put(e->elevator_type);
175 kfree(e);
176 }
178 int elevator_init(request_queue_t *q, char *name)
179 {
180 struct elevator_type *e = NULL;
181 struct elevator_queue *eq;
182 int ret = 0;
183 void *data;
185 INIT_LIST_HEAD(&q->queue_head);
186 q->last_merge = NULL;
187 q->end_sector = 0;
188 q->boundary_rq = NULL;
190 if (name && !(e = elevator_get(name)))
191 return -EINVAL;
193 if (!e && *chosen_elevator && !(e = elevator_get(chosen_elevator)))
194 printk("I/O scheduler %s not found\n", chosen_elevator);
196 if (!e && !(e = elevator_get(CONFIG_DEFAULT_IOSCHED))) {
197 printk("Default I/O scheduler not found, using no-op\n");
198 e = elevator_get("noop");
199 }
201 eq = elevator_alloc(e);
202 if (!eq)
203 return -ENOMEM;
205 data = elevator_init_queue(q, eq);
206 if (!data) {
207 kobject_put(&eq->kobj);
208 return -ENOMEM;
209 }
211 elevator_attach(q, eq, data);
212 return ret;
213 }
215 void elevator_exit(elevator_t *e)
216 {
217 mutex_lock(&e->sysfs_lock);
218 if (e->ops->elevator_exit_fn)
219 e->ops->elevator_exit_fn(e);
220 e->ops = NULL;
221 mutex_unlock(&e->sysfs_lock);
223 kobject_put(&e->kobj);
224 }
226 /*
227 * Insert rq into dispatch queue of q. Queue lock must be held on
228 * entry. If sort != 0, rq is sort-inserted; otherwise, rq will be
229 * appended to the dispatch queue. To be used by specific elevators.
230 */
231 void elv_dispatch_sort(request_queue_t *q, struct request *rq)
232 {
233 sector_t boundary;
234 struct list_head *entry;
236 if (q->last_merge == rq)
237 q->last_merge = NULL;
238 q->nr_sorted--;
240 boundary = q->end_sector;
242 list_for_each_prev(entry, &q->queue_head) {
243 struct request *pos = list_entry_rq(entry);
245 if (pos->flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
246 break;
247 if (rq->sector >= boundary) {
248 if (pos->sector < boundary)
249 continue;
250 } else {
251 if (pos->sector >= boundary)
252 break;
253 }
254 if (rq->sector >= pos->sector)
255 break;
256 }
258 list_add(&rq->queuelist, entry);
259 }
261 int elv_merge(request_queue_t *q, struct request **req, struct bio *bio)
262 {
263 elevator_t *e = q->elevator;
264 int ret;
266 if (q->last_merge) {
267 ret = elv_try_merge(q->last_merge, bio);
268 if (ret != ELEVATOR_NO_MERGE) {
269 *req = q->last_merge;
270 return ret;
271 }
272 }
274 if (e->ops->elevator_merge_fn)
275 return e->ops->elevator_merge_fn(q, req, bio);
277 return ELEVATOR_NO_MERGE;
278 }
280 void elv_merged_request(request_queue_t *q, struct request *rq)
281 {
282 elevator_t *e = q->elevator;
284 if (e->ops->elevator_merged_fn)
285 e->ops->elevator_merged_fn(q, rq);
287 q->last_merge = rq;
288 }
290 void elv_merge_requests(request_queue_t *q, struct request *rq,
291 struct request *next)
292 {
293 elevator_t *e = q->elevator;
295 if (e->ops->elevator_merge_req_fn)
296 e->ops->elevator_merge_req_fn(q, rq, next);
297 q->nr_sorted--;
299 q->last_merge = rq;
300 }
302 void elv_requeue_request(request_queue_t *q, struct request *rq)
303 {
304 elevator_t *e = q->elevator;
306 /*
307 * it already went through dequeue, we need to decrement the
308 * in_flight count again
309 */
310 if (blk_account_rq(rq)) {
311 q->in_flight--;
312 if (blk_sorted_rq(rq) && e->ops->elevator_deactivate_req_fn)
313 e->ops->elevator_deactivate_req_fn(q, rq);
314 }
316 rq->flags &= ~REQ_STARTED;
318 elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
319 }
321 static void elv_drain_elevator(request_queue_t *q)
322 {
323 static int printed;
324 while (q->elevator->ops->elevator_dispatch_fn(q, 1))
325 ;
326 if (q->nr_sorted == 0)
327 return;
328 if (printed++ < 10) {
329 printk(KERN_ERR "%s: forced dispatching is broken "
330 "(nr_sorted=%u), please report this\n",
331 q->elevator->elevator_type->elevator_name, q->nr_sorted);
332 }
333 }
335 void elv_insert(request_queue_t *q, struct request *rq, int where)
336 {
337 struct list_head *pos;
338 unsigned ordseq;
339 int unplug_it = 1;
341 blk_add_trace_rq(q, rq, BLK_TA_INSERT);
343 rq->q = q;
345 switch (where) {
346 case ELEVATOR_INSERT_FRONT:
347 rq->flags |= REQ_SOFTBARRIER;
349 list_add(&rq->queuelist, &q->queue_head);
350 break;
352 case ELEVATOR_INSERT_BACK:
353 rq->flags |= REQ_SOFTBARRIER;
354 elv_drain_elevator(q);
355 list_add_tail(&rq->queuelist, &q->queue_head);
356 /*
357 * We kick the queue here for the following reasons.
358 * - The elevator might have returned NULL previously
359 * to delay requests and returned them now. As the
360 * queue wasn't empty before this request, ll_rw_blk
361 * won't run the queue on return, resulting in hang.
362 * - Usually, back inserted requests won't be merged
363 * with anything. There's no point in delaying queue
364 * processing.
365 */
366 blk_remove_plug(q);
367 q->request_fn(q);
368 break;
370 case ELEVATOR_INSERT_SORT:
371 BUG_ON(!blk_fs_request(rq));
372 rq->flags |= REQ_SORTED;
373 q->nr_sorted++;
374 if (q->last_merge == NULL && rq_mergeable(rq))
375 q->last_merge = rq;
376 /*
377 * Some ioscheds (cfq) run q->request_fn directly, so
378 * rq cannot be accessed after calling
379 * elevator_add_req_fn.
380 */
381 q->elevator->ops->elevator_add_req_fn(q, rq);
382 break;
384 case ELEVATOR_INSERT_REQUEUE:
385 /*
386 * If ordered flush isn't in progress, we do front
387 * insertion; otherwise, requests should be requeued
388 * in ordseq order.
389 */
390 rq->flags |= REQ_SOFTBARRIER;
392 if (q->ordseq == 0) {
393 list_add(&rq->queuelist, &q->queue_head);
394 break;
395 }
397 ordseq = blk_ordered_req_seq(rq);
399 list_for_each(pos, &q->queue_head) {
400 struct request *pos_rq = list_entry_rq(pos);
401 if (ordseq <= blk_ordered_req_seq(pos_rq))
402 break;
403 }
405 list_add_tail(&rq->queuelist, pos);
406 /*
407 * most requeues happen because of a busy condition, don't
408 * force unplug of the queue for that case.
409 */
410 unplug_it = 0;
411 break;
413 default:
414 printk(KERN_ERR "%s: bad insertion point %d\n",
415 __FUNCTION__, where);
416 BUG();
417 }
419 if (unplug_it && blk_queue_plugged(q)) {
420 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
421 - q->in_flight;
423 if (nrq >= q->unplug_thresh)
424 __generic_unplug_device(q);
425 }
426 }
428 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
429 int plug)
430 {
431 if (q->ordcolor)
432 rq->flags |= REQ_ORDERED_COLOR;
434 if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
435 /*
436 * toggle ordered color
437 */
438 if (blk_barrier_rq(rq))
439 q->ordcolor ^= 1;
441 /*
442 * barriers implicitly indicate back insertion
443 */
444 if (where == ELEVATOR_INSERT_SORT)
445 where = ELEVATOR_INSERT_BACK;
447 /*
448 * this request is scheduling boundary, update
449 * end_sector
450 */
451 if (blk_fs_request(rq)) {
452 q->end_sector = rq_end_sector(rq);
453 q->boundary_rq = rq;
454 }
455 } else if (!(rq->flags & REQ_ELVPRIV) && where == ELEVATOR_INSERT_SORT)
456 where = ELEVATOR_INSERT_BACK;
458 if (plug)
459 blk_plug_device(q);
461 elv_insert(q, rq, where);
462 }
464 void elv_add_request(request_queue_t *q, struct request *rq, int where,
465 int plug)
466 {
467 unsigned long flags;
469 spin_lock_irqsave(q->queue_lock, flags);
470 __elv_add_request(q, rq, where, plug);
471 spin_unlock_irqrestore(q->queue_lock, flags);
472 }
474 static inline struct request *__elv_next_request(request_queue_t *q)
475 {
476 struct request *rq;
478 while (1) {
479 while (!list_empty(&q->queue_head)) {
480 rq = list_entry_rq(q->queue_head.next);
481 if (blk_do_ordered(q, &rq))
482 return rq;
483 }
485 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
486 return NULL;
487 }
488 }
490 struct request *elv_next_request(request_queue_t *q)
491 {
492 struct request *rq;
493 int ret;
495 while ((rq = __elv_next_request(q)) != NULL) {
496 /*
497 * Kill the empty barrier place holder, the driver must
498 * not ever see it.
499 */
500 if (blk_empty_barrier(rq)) {
501 blkdev_dequeue_request(rq);
502 end_that_request_chunk(rq, 1, 0);
503 end_that_request_last(rq, 1);
504 continue;
505 }
506 if (!(rq->flags & REQ_STARTED)) {
507 elevator_t *e = q->elevator;
509 /*
510 * This is the first time the device driver
511 * sees this request (possibly after
512 * requeueing). Notify IO scheduler.
513 */
514 if (blk_sorted_rq(rq) &&
515 e->ops->elevator_activate_req_fn)
516 e->ops->elevator_activate_req_fn(q, rq);
518 /*
519 * just mark as started even if we don't start
520 * it, a request that has been delayed should
521 * not be passed by new incoming requests
522 */
523 rq->flags |= REQ_STARTED;
524 blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
525 }
527 if (!q->boundary_rq || q->boundary_rq == rq) {
528 q->end_sector = rq_end_sector(rq);
529 q->boundary_rq = NULL;
530 }
532 if ((rq->flags & REQ_DONTPREP) || !q->prep_rq_fn)
533 break;
535 ret = q->prep_rq_fn(q, rq);
536 if (ret == BLKPREP_OK) {
537 break;
538 } else if (ret == BLKPREP_DEFER) {
539 /*
540 * the request may have been (partially) prepped.
541 * we need to keep this request in the front to
542 * avoid resource deadlock. REQ_STARTED will
543 * prevent other fs requests from passing this one.
544 */
545 rq = NULL;
546 break;
547 } else if (ret == BLKPREP_KILL) {
548 int nr_bytes = rq->hard_nr_sectors << 9;
550 if (!nr_bytes)
551 nr_bytes = rq->data_len;
553 blkdev_dequeue_request(rq);
554 rq->flags |= REQ_QUIET;
555 end_that_request_chunk(rq, 0, nr_bytes);
556 end_that_request_last(rq, 0);
557 } else {
558 printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
559 ret);
560 break;
561 }
562 }
564 return rq;
565 }
567 void elv_dequeue_request(request_queue_t *q, struct request *rq)
568 {
569 BUG_ON(list_empty(&rq->queuelist));
571 list_del_init(&rq->queuelist);
573 /*
574 * the time frame between a request being removed from the lists
575 * and to it is freed is accounted as io that is in progress at
576 * the driver side.
577 */
578 if (blk_account_rq(rq))
579 q->in_flight++;
580 }
582 int elv_queue_empty(request_queue_t *q)
583 {
584 elevator_t *e = q->elevator;
586 if (!list_empty(&q->queue_head))
587 return 0;
589 if (e->ops->elevator_queue_empty_fn)
590 return e->ops->elevator_queue_empty_fn(q);
592 return 1;
593 }
595 struct request *elv_latter_request(request_queue_t *q, struct request *rq)
596 {
597 elevator_t *e = q->elevator;
599 if (e->ops->elevator_latter_req_fn)
600 return e->ops->elevator_latter_req_fn(q, rq);
601 return NULL;
602 }
604 struct request *elv_former_request(request_queue_t *q, struct request *rq)
605 {
606 elevator_t *e = q->elevator;
608 if (e->ops->elevator_former_req_fn)
609 return e->ops->elevator_former_req_fn(q, rq);
610 return NULL;
611 }
613 int elv_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
614 gfp_t gfp_mask)
615 {
616 elevator_t *e = q->elevator;
618 if (e->ops->elevator_set_req_fn)
619 return e->ops->elevator_set_req_fn(q, rq, bio, gfp_mask);
621 rq->elevator_private = NULL;
622 return 0;
623 }
625 void elv_put_request(request_queue_t *q, struct request *rq)
626 {
627 elevator_t *e = q->elevator;
629 if (e->ops->elevator_put_req_fn)
630 e->ops->elevator_put_req_fn(q, rq);
631 }
633 int elv_may_queue(request_queue_t *q, int rw, struct bio *bio)
634 {
635 elevator_t *e = q->elevator;
637 if (e->ops->elevator_may_queue_fn)
638 return e->ops->elevator_may_queue_fn(q, rw, bio);
640 return ELV_MQUEUE_MAY;
641 }
643 void elv_completed_request(request_queue_t *q, struct request *rq)
644 {
645 elevator_t *e = q->elevator;
647 /*
648 * request is released from the driver, io must be done
649 */
650 if (blk_account_rq(rq)) {
651 q->in_flight--;
652 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
653 e->ops->elevator_completed_req_fn(q, rq);
654 }
656 /*
657 * Check if the queue is waiting for fs requests to be
658 * drained for flush sequence.
659 */
660 if (unlikely(q->ordseq)) {
661 struct request *first_rq = list_entry_rq(q->queue_head.next);
662 if (q->in_flight == 0 &&
663 blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
664 blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
665 blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
666 q->request_fn(q);
667 }
668 }
669 }
671 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
673 static ssize_t
674 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
675 {
676 elevator_t *e = container_of(kobj, elevator_t, kobj);
677 struct elv_fs_entry *entry = to_elv(attr);
678 ssize_t error;
680 if (!entry->show)
681 return -EIO;
683 mutex_lock(&e->sysfs_lock);
684 error = e->ops ? entry->show(e, page) : -ENOENT;
685 mutex_unlock(&e->sysfs_lock);
686 return error;
687 }
689 static ssize_t
690 elv_attr_store(struct kobject *kobj, struct attribute *attr,
691 const char *page, size_t length)
692 {
693 elevator_t *e = container_of(kobj, elevator_t, kobj);
694 struct elv_fs_entry *entry = to_elv(attr);
695 ssize_t error;
697 if (!entry->store)
698 return -EIO;
700 mutex_lock(&e->sysfs_lock);
701 error = e->ops ? entry->store(e, page, length) : -ENOENT;
702 mutex_unlock(&e->sysfs_lock);
703 return error;
704 }
706 static struct sysfs_ops elv_sysfs_ops = {
707 .show = elv_attr_show,
708 .store = elv_attr_store,
709 };
711 static struct kobj_type elv_ktype = {
712 .sysfs_ops = &elv_sysfs_ops,
713 .release = elevator_release,
714 };
716 int elv_register_queue(struct request_queue *q)
717 {
718 elevator_t *e = q->elevator;
719 int error;
721 e->kobj.parent = &q->kobj;
723 error = kobject_add(&e->kobj);
724 if (!error) {
725 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
726 if (attr) {
727 while (attr->attr.name) {
728 if (sysfs_create_file(&e->kobj, &attr->attr))
729 break;
730 attr++;
731 }
732 }
733 kobject_uevent(&e->kobj, KOBJ_ADD);
734 }
735 return error;
736 }
738 static void __elv_unregister_queue(elevator_t *e)
739 {
740 kobject_uevent(&e->kobj, KOBJ_REMOVE);
741 kobject_del(&e->kobj);
742 }
744 void elv_unregister_queue(struct request_queue *q)
745 {
746 if (q)
747 __elv_unregister_queue(q->elevator);
748 }
750 int elv_register(struct elevator_type *e)
751 {
752 spin_lock_irq(&elv_list_lock);
753 BUG_ON(elevator_find(e->elevator_name));
754 list_add_tail(&e->list, &elv_list);
755 spin_unlock_irq(&elv_list_lock);
757 printk(KERN_INFO "io scheduler %s registered", e->elevator_name);
758 if (!strcmp(e->elevator_name, chosen_elevator) ||
759 (!*chosen_elevator &&
760 !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
761 printk(" (default)");
762 printk("\n");
763 return 0;
764 }
765 EXPORT_SYMBOL_GPL(elv_register);
767 void elv_unregister(struct elevator_type *e)
768 {
769 struct task_struct *g, *p;
771 /*
772 * Iterate every thread in the process to remove the io contexts.
773 */
774 if (e->ops.trim) {
775 read_lock(&tasklist_lock);
776 do_each_thread(g, p) {
777 task_lock(p);
778 if (p->io_context)
779 e->ops.trim(p->io_context);
780 task_unlock(p);
781 } while_each_thread(g, p);
782 read_unlock(&tasklist_lock);
783 }
785 spin_lock_irq(&elv_list_lock);
786 list_del_init(&e->list);
787 spin_unlock_irq(&elv_list_lock);
788 }
789 EXPORT_SYMBOL_GPL(elv_unregister);
791 /*
792 * switch to new_e io scheduler. be careful not to introduce deadlocks -
793 * we don't free the old io scheduler, before we have allocated what we
794 * need for the new one. this way we have a chance of going back to the old
795 * one, if the new one fails init for some reason.
796 */
797 static int elevator_switch(request_queue_t *q, struct elevator_type *new_e)
798 {
799 elevator_t *old_elevator, *e;
800 void *data;
802 /*
803 * Allocate new elevator
804 */
805 e = elevator_alloc(new_e);
806 if (!e)
807 return 0;
809 data = elevator_init_queue(q, e);
810 if (!data) {
811 kobject_put(&e->kobj);
812 return 0;
813 }
815 /*
816 * Turn on BYPASS and drain all requests w/ elevator private data
817 */
818 spin_lock_irq(q->queue_lock);
820 set_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
822 elv_drain_elevator(q);
824 while (q->rq.elvpriv) {
825 blk_remove_plug(q);
826 q->request_fn(q);
827 spin_unlock_irq(q->queue_lock);
828 msleep(10);
829 spin_lock_irq(q->queue_lock);
830 elv_drain_elevator(q);
831 }
833 /*
834 * Remember old elevator.
835 */
836 old_elevator = q->elevator;
838 /*
839 * attach and start new elevator
840 */
841 elevator_attach(q, e, data);
843 spin_unlock_irq(q->queue_lock);
845 __elv_unregister_queue(old_elevator);
847 if (elv_register_queue(q))
848 goto fail_register;
850 /*
851 * finally exit old elevator and turn off BYPASS.
852 */
853 elevator_exit(old_elevator);
854 clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
855 return 1;
857 fail_register:
858 /*
859 * switch failed, exit the new io scheduler and reattach the old
860 * one again (along with re-adding the sysfs dir)
861 */
862 elevator_exit(e);
863 q->elevator = old_elevator;
864 elv_register_queue(q);
865 clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
866 return 0;
867 }
869 ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count)
870 {
871 char elevator_name[ELV_NAME_MAX];
872 size_t len;
873 struct elevator_type *e;
875 elevator_name[sizeof(elevator_name) - 1] = '\0';
876 strncpy(elevator_name, name, sizeof(elevator_name) - 1);
877 len = strlen(elevator_name);
879 if (len && elevator_name[len - 1] == '\n')
880 elevator_name[len - 1] = '\0';
882 e = elevator_get(elevator_name);
883 if (!e) {
884 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
885 return -EINVAL;
886 }
888 if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
889 elevator_put(e);
890 return count;
891 }
893 if (!elevator_switch(q, e))
894 printk(KERN_ERR "elevator: switch to %s failed\n",elevator_name);
895 return count;
896 }
898 ssize_t elv_iosched_show(request_queue_t *q, char *name)
899 {
900 elevator_t *e = q->elevator;
901 struct elevator_type *elv = e->elevator_type;
902 struct list_head *entry;
903 int len = 0;
905 spin_lock_irq(&elv_list_lock);
906 list_for_each(entry, &elv_list) {
907 struct elevator_type *__e;
909 __e = list_entry(entry, struct elevator_type, list);
910 if (!strcmp(elv->elevator_name, __e->elevator_name))
911 len += sprintf(name+len, "[%s] ", elv->elevator_name);
912 else
913 len += sprintf(name+len, "%s ", __e->elevator_name);
914 }
915 spin_unlock_irq(&elv_list_lock);
917 len += sprintf(len+name, "\n");
918 return len;
919 }
921 EXPORT_SYMBOL(elv_dispatch_sort);
922 EXPORT_SYMBOL(elv_add_request);
923 EXPORT_SYMBOL(__elv_add_request);
924 EXPORT_SYMBOL(elv_next_request);
925 EXPORT_SYMBOL(elv_dequeue_request);
926 EXPORT_SYMBOL(elv_queue_empty);
927 EXPORT_SYMBOL(elevator_exit);
928 EXPORT_SYMBOL(elevator_init);