blkcg: shoot down blkio_groups on elevator switch
[linux-3.10.git] / block / elevator.c
1 /*
2  *  Block device elevator/IO-scheduler.
3  *
4  *  Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
5  *
6  * 30042000 Jens Axboe <axboe@kernel.dk> :
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/blktrace_api.h>
35 #include <linux/hash.h>
36 #include <linux/uaccess.h>
37
38 #include <trace/events/block.h>
39
40 #include "blk.h"
41 #include "blk-cgroup.h"
42
43 static DEFINE_SPINLOCK(elv_list_lock);
44 static LIST_HEAD(elv_list);
45
46 /*
47  * Merge hash stuff.
48  */
49 static const int elv_hash_shift = 6;
50 #define ELV_HASH_BLOCK(sec)     ((sec) >> 3)
51 #define ELV_HASH_FN(sec)        \
52                 (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
53 #define ELV_HASH_ENTRIES        (1 << elv_hash_shift)
54 #define rq_hash_key(rq)         (blk_rq_pos(rq) + blk_rq_sectors(rq))
55
56 /*
57  * Query io scheduler to see if the current process issuing bio may be
58  * merged with rq.
59  */
60 static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
61 {
62         struct request_queue *q = rq->q;
63         struct elevator_queue *e = q->elevator;
64
65         if (e->type->ops.elevator_allow_merge_fn)
66                 return e->type->ops.elevator_allow_merge_fn(q, rq, bio);
67
68         return 1;
69 }
70
71 /*
72  * can we safely merge with this request?
73  */
74 bool elv_rq_merge_ok(struct request *rq, struct bio *bio)
75 {
76         if (!blk_rq_merge_ok(rq, bio))
77                 return 0;
78
79         if (!elv_iosched_allow_merge(rq, bio))
80                 return 0;
81
82         return 1;
83 }
84 EXPORT_SYMBOL(elv_rq_merge_ok);
85
86 static struct elevator_type *elevator_find(const char *name)
87 {
88         struct elevator_type *e;
89
90         list_for_each_entry(e, &elv_list, list) {
91                 if (!strcmp(e->elevator_name, name))
92                         return e;
93         }
94
95         return NULL;
96 }
97
98 static void elevator_put(struct elevator_type *e)
99 {
100         module_put(e->elevator_owner);
101 }
102
103 static struct elevator_type *elevator_get(const char *name)
104 {
105         struct elevator_type *e;
106
107         spin_lock(&elv_list_lock);
108
109         e = elevator_find(name);
110         if (!e) {
111                 spin_unlock(&elv_list_lock);
112                 request_module("%s-iosched", name);
113                 spin_lock(&elv_list_lock);
114                 e = elevator_find(name);
115         }
116
117         if (e && !try_module_get(e->elevator_owner))
118                 e = NULL;
119
120         spin_unlock(&elv_list_lock);
121
122         return e;
123 }
124
125 static char chosen_elevator[ELV_NAME_MAX];
126
127 static int __init elevator_setup(char *str)
128 {
129         /*
130          * Be backwards-compatible with previous kernels, so users
131          * won't get the wrong elevator.
132          */
133         strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
134         return 1;
135 }
136
137 __setup("elevator=", elevator_setup);
138
139 static struct kobj_type elv_ktype;
140
141 static struct elevator_queue *elevator_alloc(struct request_queue *q,
142                                   struct elevator_type *e)
143 {
144         struct elevator_queue *eq;
145         int i;
146
147         eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node);
148         if (unlikely(!eq))
149                 goto err;
150
151         eq->type = e;
152         kobject_init(&eq->kobj, &elv_ktype);
153         mutex_init(&eq->sysfs_lock);
154
155         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
156                                         GFP_KERNEL, q->node);
157         if (!eq->hash)
158                 goto err;
159
160         for (i = 0; i < ELV_HASH_ENTRIES; i++)
161                 INIT_HLIST_HEAD(&eq->hash[i]);
162
163         return eq;
164 err:
165         kfree(eq);
166         elevator_put(e);
167         return NULL;
168 }
169
170 static void elevator_release(struct kobject *kobj)
171 {
172         struct elevator_queue *e;
173
174         e = container_of(kobj, struct elevator_queue, kobj);
175         elevator_put(e->type);
176         kfree(e->hash);
177         kfree(e);
178 }
179
180 int elevator_init(struct request_queue *q, char *name)
181 {
182         struct elevator_type *e = NULL;
183         int err;
184
185         if (unlikely(q->elevator))
186                 return 0;
187
188         INIT_LIST_HEAD(&q->queue_head);
189         q->last_merge = NULL;
190         q->end_sector = 0;
191         q->boundary_rq = NULL;
192
193         if (name) {
194                 e = elevator_get(name);
195                 if (!e)
196                         return -EINVAL;
197         }
198
199         if (!e && *chosen_elevator) {
200                 e = elevator_get(chosen_elevator);
201                 if (!e)
202                         printk(KERN_ERR "I/O scheduler %s not found\n",
203                                                         chosen_elevator);
204         }
205
206         if (!e) {
207                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
208                 if (!e) {
209                         printk(KERN_ERR
210                                 "Default I/O scheduler not found. " \
211                                 "Using noop.\n");
212                         e = elevator_get("noop");
213                 }
214         }
215
216         q->elevator = elevator_alloc(q, e);
217         if (!q->elevator)
218                 return -ENOMEM;
219
220         err = e->ops.elevator_init_fn(q);
221         if (err) {
222                 kobject_put(&q->elevator->kobj);
223                 return err;
224         }
225
226         return 0;
227 }
228 EXPORT_SYMBOL(elevator_init);
229
230 void elevator_exit(struct elevator_queue *e)
231 {
232         mutex_lock(&e->sysfs_lock);
233         if (e->type->ops.elevator_exit_fn)
234                 e->type->ops.elevator_exit_fn(e);
235         mutex_unlock(&e->sysfs_lock);
236
237         kobject_put(&e->kobj);
238 }
239 EXPORT_SYMBOL(elevator_exit);
240
241 static inline void __elv_rqhash_del(struct request *rq)
242 {
243         hlist_del_init(&rq->hash);
244 }
245
246 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
247 {
248         if (ELV_ON_HASH(rq))
249                 __elv_rqhash_del(rq);
250 }
251
252 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
253 {
254         struct elevator_queue *e = q->elevator;
255
256         BUG_ON(ELV_ON_HASH(rq));
257         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
258 }
259
260 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
261 {
262         __elv_rqhash_del(rq);
263         elv_rqhash_add(q, rq);
264 }
265
266 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
267 {
268         struct elevator_queue *e = q->elevator;
269         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
270         struct hlist_node *entry, *next;
271         struct request *rq;
272
273         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
274                 BUG_ON(!ELV_ON_HASH(rq));
275
276                 if (unlikely(!rq_mergeable(rq))) {
277                         __elv_rqhash_del(rq);
278                         continue;
279                 }
280
281                 if (rq_hash_key(rq) == offset)
282                         return rq;
283         }
284
285         return NULL;
286 }
287
288 /*
289  * RB-tree support functions for inserting/lookup/removal of requests
290  * in a sorted RB tree.
291  */
292 void elv_rb_add(struct rb_root *root, struct request *rq)
293 {
294         struct rb_node **p = &root->rb_node;
295         struct rb_node *parent = NULL;
296         struct request *__rq;
297
298         while (*p) {
299                 parent = *p;
300                 __rq = rb_entry(parent, struct request, rb_node);
301
302                 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
303                         p = &(*p)->rb_left;
304                 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
305                         p = &(*p)->rb_right;
306         }
307
308         rb_link_node(&rq->rb_node, parent, p);
309         rb_insert_color(&rq->rb_node, root);
310 }
311 EXPORT_SYMBOL(elv_rb_add);
312
313 void elv_rb_del(struct rb_root *root, struct request *rq)
314 {
315         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
316         rb_erase(&rq->rb_node, root);
317         RB_CLEAR_NODE(&rq->rb_node);
318 }
319 EXPORT_SYMBOL(elv_rb_del);
320
321 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
322 {
323         struct rb_node *n = root->rb_node;
324         struct request *rq;
325
326         while (n) {
327                 rq = rb_entry(n, struct request, rb_node);
328
329                 if (sector < blk_rq_pos(rq))
330                         n = n->rb_left;
331                 else if (sector > blk_rq_pos(rq))
332                         n = n->rb_right;
333                 else
334                         return rq;
335         }
336
337         return NULL;
338 }
339 EXPORT_SYMBOL(elv_rb_find);
340
341 /*
342  * Insert rq into dispatch queue of q.  Queue lock must be held on
343  * entry.  rq is sort instead into the dispatch queue. To be used by
344  * specific elevators.
345  */
346 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
347 {
348         sector_t boundary;
349         struct list_head *entry;
350         int stop_flags;
351
352         if (q->last_merge == rq)
353                 q->last_merge = NULL;
354
355         elv_rqhash_del(q, rq);
356
357         q->nr_sorted--;
358
359         boundary = q->end_sector;
360         stop_flags = REQ_SOFTBARRIER | REQ_STARTED;
361         list_for_each_prev(entry, &q->queue_head) {
362                 struct request *pos = list_entry_rq(entry);
363
364                 if ((rq->cmd_flags & REQ_DISCARD) !=
365                     (pos->cmd_flags & REQ_DISCARD))
366                         break;
367                 if (rq_data_dir(rq) != rq_data_dir(pos))
368                         break;
369                 if (pos->cmd_flags & stop_flags)
370                         break;
371                 if (blk_rq_pos(rq) >= boundary) {
372                         if (blk_rq_pos(pos) < boundary)
373                                 continue;
374                 } else {
375                         if (blk_rq_pos(pos) >= boundary)
376                                 break;
377                 }
378                 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
379                         break;
380         }
381
382         list_add(&rq->queuelist, entry);
383 }
384 EXPORT_SYMBOL(elv_dispatch_sort);
385
386 /*
387  * Insert rq into dispatch queue of q.  Queue lock must be held on
388  * entry.  rq is added to the back of the dispatch queue. To be used by
389  * specific elevators.
390  */
391 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
392 {
393         if (q->last_merge == rq)
394                 q->last_merge = NULL;
395
396         elv_rqhash_del(q, rq);
397
398         q->nr_sorted--;
399
400         q->end_sector = rq_end_sector(rq);
401         q->boundary_rq = rq;
402         list_add_tail(&rq->queuelist, &q->queue_head);
403 }
404 EXPORT_SYMBOL(elv_dispatch_add_tail);
405
406 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
407 {
408         struct elevator_queue *e = q->elevator;
409         struct request *__rq;
410         int ret;
411
412         /*
413          * Levels of merges:
414          *      nomerges:  No merges at all attempted
415          *      noxmerges: Only simple one-hit cache try
416          *      merges:    All merge tries attempted
417          */
418         if (blk_queue_nomerges(q))
419                 return ELEVATOR_NO_MERGE;
420
421         /*
422          * First try one-hit cache.
423          */
424         if (q->last_merge && elv_rq_merge_ok(q->last_merge, bio)) {
425                 ret = blk_try_merge(q->last_merge, bio);
426                 if (ret != ELEVATOR_NO_MERGE) {
427                         *req = q->last_merge;
428                         return ret;
429                 }
430         }
431
432         if (blk_queue_noxmerges(q))
433                 return ELEVATOR_NO_MERGE;
434
435         /*
436          * See if our hash lookup can find a potential backmerge.
437          */
438         __rq = elv_rqhash_find(q, bio->bi_sector);
439         if (__rq && elv_rq_merge_ok(__rq, bio)) {
440                 *req = __rq;
441                 return ELEVATOR_BACK_MERGE;
442         }
443
444         if (e->type->ops.elevator_merge_fn)
445                 return e->type->ops.elevator_merge_fn(q, req, bio);
446
447         return ELEVATOR_NO_MERGE;
448 }
449
450 /*
451  * Attempt to do an insertion back merge. Only check for the case where
452  * we can append 'rq' to an existing request, so we can throw 'rq' away
453  * afterwards.
454  *
455  * Returns true if we merged, false otherwise
456  */
457 static bool elv_attempt_insert_merge(struct request_queue *q,
458                                      struct request *rq)
459 {
460         struct request *__rq;
461
462         if (blk_queue_nomerges(q))
463                 return false;
464
465         /*
466          * First try one-hit cache.
467          */
468         if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
469                 return true;
470
471         if (blk_queue_noxmerges(q))
472                 return false;
473
474         /*
475          * See if our hash lookup can find a potential backmerge.
476          */
477         __rq = elv_rqhash_find(q, blk_rq_pos(rq));
478         if (__rq && blk_attempt_req_merge(q, __rq, rq))
479                 return true;
480
481         return false;
482 }
483
484 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
485 {
486         struct elevator_queue *e = q->elevator;
487
488         if (e->type->ops.elevator_merged_fn)
489                 e->type->ops.elevator_merged_fn(q, rq, type);
490
491         if (type == ELEVATOR_BACK_MERGE)
492                 elv_rqhash_reposition(q, rq);
493
494         q->last_merge = rq;
495 }
496
497 void elv_merge_requests(struct request_queue *q, struct request *rq,
498                              struct request *next)
499 {
500         struct elevator_queue *e = q->elevator;
501         const int next_sorted = next->cmd_flags & REQ_SORTED;
502
503         if (next_sorted && e->type->ops.elevator_merge_req_fn)
504                 e->type->ops.elevator_merge_req_fn(q, rq, next);
505
506         elv_rqhash_reposition(q, rq);
507
508         if (next_sorted) {
509                 elv_rqhash_del(q, next);
510                 q->nr_sorted--;
511         }
512
513         q->last_merge = rq;
514 }
515
516 void elv_bio_merged(struct request_queue *q, struct request *rq,
517                         struct bio *bio)
518 {
519         struct elevator_queue *e = q->elevator;
520
521         if (e->type->ops.elevator_bio_merged_fn)
522                 e->type->ops.elevator_bio_merged_fn(q, rq, bio);
523 }
524
525 void elv_requeue_request(struct request_queue *q, struct request *rq)
526 {
527         /*
528          * it already went through dequeue, we need to decrement the
529          * in_flight count again
530          */
531         if (blk_account_rq(rq)) {
532                 q->in_flight[rq_is_sync(rq)]--;
533                 if (rq->cmd_flags & REQ_SORTED)
534                         elv_deactivate_rq(q, rq);
535         }
536
537         rq->cmd_flags &= ~REQ_STARTED;
538
539         __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
540 }
541
542 void elv_drain_elevator(struct request_queue *q)
543 {
544         static int printed;
545
546         lockdep_assert_held(q->queue_lock);
547
548         while (q->elevator->type->ops.elevator_dispatch_fn(q, 1))
549                 ;
550         if (q->nr_sorted && printed++ < 10) {
551                 printk(KERN_ERR "%s: forced dispatching is broken "
552                        "(nr_sorted=%u), please report this\n",
553                        q->elevator->type->elevator_name, q->nr_sorted);
554         }
555 }
556
557 void __elv_add_request(struct request_queue *q, struct request *rq, int where)
558 {
559         trace_block_rq_insert(q, rq);
560
561         rq->q = q;
562
563         if (rq->cmd_flags & REQ_SOFTBARRIER) {
564                 /* barriers are scheduling boundary, update end_sector */
565                 if (rq->cmd_type == REQ_TYPE_FS ||
566                     (rq->cmd_flags & REQ_DISCARD)) {
567                         q->end_sector = rq_end_sector(rq);
568                         q->boundary_rq = rq;
569                 }
570         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
571                     (where == ELEVATOR_INSERT_SORT ||
572                      where == ELEVATOR_INSERT_SORT_MERGE))
573                 where = ELEVATOR_INSERT_BACK;
574
575         switch (where) {
576         case ELEVATOR_INSERT_REQUEUE:
577         case ELEVATOR_INSERT_FRONT:
578                 rq->cmd_flags |= REQ_SOFTBARRIER;
579                 list_add(&rq->queuelist, &q->queue_head);
580                 break;
581
582         case ELEVATOR_INSERT_BACK:
583                 rq->cmd_flags |= REQ_SOFTBARRIER;
584                 elv_drain_elevator(q);
585                 list_add_tail(&rq->queuelist, &q->queue_head);
586                 /*
587                  * We kick the queue here for the following reasons.
588                  * - The elevator might have returned NULL previously
589                  *   to delay requests and returned them now.  As the
590                  *   queue wasn't empty before this request, ll_rw_blk
591                  *   won't run the queue on return, resulting in hang.
592                  * - Usually, back inserted requests won't be merged
593                  *   with anything.  There's no point in delaying queue
594                  *   processing.
595                  */
596                 __blk_run_queue(q);
597                 break;
598
599         case ELEVATOR_INSERT_SORT_MERGE:
600                 /*
601                  * If we succeed in merging this request with one in the
602                  * queue already, we are done - rq has now been freed,
603                  * so no need to do anything further.
604                  */
605                 if (elv_attempt_insert_merge(q, rq))
606                         break;
607         case ELEVATOR_INSERT_SORT:
608                 BUG_ON(rq->cmd_type != REQ_TYPE_FS &&
609                        !(rq->cmd_flags & REQ_DISCARD));
610                 rq->cmd_flags |= REQ_SORTED;
611                 q->nr_sorted++;
612                 if (rq_mergeable(rq)) {
613                         elv_rqhash_add(q, rq);
614                         if (!q->last_merge)
615                                 q->last_merge = rq;
616                 }
617
618                 /*
619                  * Some ioscheds (cfq) run q->request_fn directly, so
620                  * rq cannot be accessed after calling
621                  * elevator_add_req_fn.
622                  */
623                 q->elevator->type->ops.elevator_add_req_fn(q, rq);
624                 break;
625
626         case ELEVATOR_INSERT_FLUSH:
627                 rq->cmd_flags |= REQ_SOFTBARRIER;
628                 blk_insert_flush(rq);
629                 break;
630         default:
631                 printk(KERN_ERR "%s: bad insertion point %d\n",
632                        __func__, where);
633                 BUG();
634         }
635 }
636 EXPORT_SYMBOL(__elv_add_request);
637
638 void elv_add_request(struct request_queue *q, struct request *rq, int where)
639 {
640         unsigned long flags;
641
642         spin_lock_irqsave(q->queue_lock, flags);
643         __elv_add_request(q, rq, where);
644         spin_unlock_irqrestore(q->queue_lock, flags);
645 }
646 EXPORT_SYMBOL(elv_add_request);
647
648 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
649 {
650         struct elevator_queue *e = q->elevator;
651
652         if (e->type->ops.elevator_latter_req_fn)
653                 return e->type->ops.elevator_latter_req_fn(q, rq);
654         return NULL;
655 }
656
657 struct request *elv_former_request(struct request_queue *q, struct request *rq)
658 {
659         struct elevator_queue *e = q->elevator;
660
661         if (e->type->ops.elevator_former_req_fn)
662                 return e->type->ops.elevator_former_req_fn(q, rq);
663         return NULL;
664 }
665
666 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
667 {
668         struct elevator_queue *e = q->elevator;
669
670         if (e->type->ops.elevator_set_req_fn)
671                 return e->type->ops.elevator_set_req_fn(q, rq, gfp_mask);
672         return 0;
673 }
674
675 void elv_put_request(struct request_queue *q, struct request *rq)
676 {
677         struct elevator_queue *e = q->elevator;
678
679         if (e->type->ops.elevator_put_req_fn)
680                 e->type->ops.elevator_put_req_fn(rq);
681 }
682
683 int elv_may_queue(struct request_queue *q, int rw)
684 {
685         struct elevator_queue *e = q->elevator;
686
687         if (e->type->ops.elevator_may_queue_fn)
688                 return e->type->ops.elevator_may_queue_fn(q, rw);
689
690         return ELV_MQUEUE_MAY;
691 }
692
693 void elv_abort_queue(struct request_queue *q)
694 {
695         struct request *rq;
696
697         blk_abort_flushes(q);
698
699         while (!list_empty(&q->queue_head)) {
700                 rq = list_entry_rq(q->queue_head.next);
701                 rq->cmd_flags |= REQ_QUIET;
702                 trace_block_rq_abort(q, rq);
703                 /*
704                  * Mark this request as started so we don't trigger
705                  * any debug logic in the end I/O path.
706                  */
707                 blk_start_request(rq);
708                 __blk_end_request_all(rq, -EIO);
709         }
710 }
711 EXPORT_SYMBOL(elv_abort_queue);
712
713 void elv_completed_request(struct request_queue *q, struct request *rq)
714 {
715         struct elevator_queue *e = q->elevator;
716
717         /*
718          * request is released from the driver, io must be done
719          */
720         if (blk_account_rq(rq)) {
721                 q->in_flight[rq_is_sync(rq)]--;
722                 if ((rq->cmd_flags & REQ_SORTED) &&
723                     e->type->ops.elevator_completed_req_fn)
724                         e->type->ops.elevator_completed_req_fn(q, rq);
725         }
726 }
727
728 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
729
730 static ssize_t
731 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
732 {
733         struct elv_fs_entry *entry = to_elv(attr);
734         struct elevator_queue *e;
735         ssize_t error;
736
737         if (!entry->show)
738                 return -EIO;
739
740         e = container_of(kobj, struct elevator_queue, kobj);
741         mutex_lock(&e->sysfs_lock);
742         error = e->type ? entry->show(e, page) : -ENOENT;
743         mutex_unlock(&e->sysfs_lock);
744         return error;
745 }
746
747 static ssize_t
748 elv_attr_store(struct kobject *kobj, struct attribute *attr,
749                const char *page, size_t length)
750 {
751         struct elv_fs_entry *entry = to_elv(attr);
752         struct elevator_queue *e;
753         ssize_t error;
754
755         if (!entry->store)
756                 return -EIO;
757
758         e = container_of(kobj, struct elevator_queue, kobj);
759         mutex_lock(&e->sysfs_lock);
760         error = e->type ? entry->store(e, page, length) : -ENOENT;
761         mutex_unlock(&e->sysfs_lock);
762         return error;
763 }
764
765 static const struct sysfs_ops elv_sysfs_ops = {
766         .show   = elv_attr_show,
767         .store  = elv_attr_store,
768 };
769
770 static struct kobj_type elv_ktype = {
771         .sysfs_ops      = &elv_sysfs_ops,
772         .release        = elevator_release,
773 };
774
775 int elv_register_queue(struct request_queue *q)
776 {
777         struct elevator_queue *e = q->elevator;
778         int error;
779
780         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
781         if (!error) {
782                 struct elv_fs_entry *attr = e->type->elevator_attrs;
783                 if (attr) {
784                         while (attr->attr.name) {
785                                 if (sysfs_create_file(&e->kobj, &attr->attr))
786                                         break;
787                                 attr++;
788                         }
789                 }
790                 kobject_uevent(&e->kobj, KOBJ_ADD);
791                 e->registered = 1;
792         }
793         return error;
794 }
795 EXPORT_SYMBOL(elv_register_queue);
796
797 void elv_unregister_queue(struct request_queue *q)
798 {
799         if (q) {
800                 struct elevator_queue *e = q->elevator;
801
802                 kobject_uevent(&e->kobj, KOBJ_REMOVE);
803                 kobject_del(&e->kobj);
804                 e->registered = 0;
805         }
806 }
807 EXPORT_SYMBOL(elv_unregister_queue);
808
809 int elv_register(struct elevator_type *e)
810 {
811         char *def = "";
812
813         /* create icq_cache if requested */
814         if (e->icq_size) {
815                 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
816                     WARN_ON(e->icq_align < __alignof__(struct io_cq)))
817                         return -EINVAL;
818
819                 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
820                          "%s_io_cq", e->elevator_name);
821                 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
822                                                  e->icq_align, 0, NULL);
823                 if (!e->icq_cache)
824                         return -ENOMEM;
825         }
826
827         /* register, don't allow duplicate names */
828         spin_lock(&elv_list_lock);
829         if (elevator_find(e->elevator_name)) {
830                 spin_unlock(&elv_list_lock);
831                 if (e->icq_cache)
832                         kmem_cache_destroy(e->icq_cache);
833                 return -EBUSY;
834         }
835         list_add_tail(&e->list, &elv_list);
836         spin_unlock(&elv_list_lock);
837
838         /* print pretty message */
839         if (!strcmp(e->elevator_name, chosen_elevator) ||
840                         (!*chosen_elevator &&
841                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
842                                 def = " (default)";
843
844         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
845                                                                 def);
846         return 0;
847 }
848 EXPORT_SYMBOL_GPL(elv_register);
849
850 void elv_unregister(struct elevator_type *e)
851 {
852         /* unregister */
853         spin_lock(&elv_list_lock);
854         list_del_init(&e->list);
855         spin_unlock(&elv_list_lock);
856
857         /*
858          * Destroy icq_cache if it exists.  icq's are RCU managed.  Make
859          * sure all RCU operations are complete before proceeding.
860          */
861         if (e->icq_cache) {
862                 rcu_barrier();
863                 kmem_cache_destroy(e->icq_cache);
864                 e->icq_cache = NULL;
865         }
866 }
867 EXPORT_SYMBOL_GPL(elv_unregister);
868
869 /*
870  * switch to new_e io scheduler. be careful not to introduce deadlocks -
871  * we don't free the old io scheduler, before we have allocated what we
872  * need for the new one. this way we have a chance of going back to the old
873  * one, if the new one fails init for some reason.
874  */
875 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
876 {
877         struct elevator_queue *old = q->elevator;
878         bool registered = old->registered;
879         int err;
880
881         /*
882          * Turn on BYPASS and drain all requests w/ elevator private data.
883          * Block layer doesn't call into a quiesced elevator - all requests
884          * are directly put on the dispatch list without elevator data
885          * using INSERT_BACK.  All requests have SOFTBARRIER set and no
886          * merge happens either.
887          */
888         blk_queue_bypass_start(q);
889
890         /* unregister and clear all auxiliary data of the old elevator */
891         if (registered)
892                 elv_unregister_queue(q);
893
894         spin_lock_irq(q->queue_lock);
895         ioc_clear_queue(q);
896         spin_unlock_irq(q->queue_lock);
897
898         blkg_destroy_all(q);
899
900         /* allocate, init and register new elevator */
901         err = -ENOMEM;
902         q->elevator = elevator_alloc(q, new_e);
903         if (!q->elevator)
904                 goto fail_init;
905
906         err = new_e->ops.elevator_init_fn(q);
907         if (err) {
908                 kobject_put(&q->elevator->kobj);
909                 goto fail_init;
910         }
911
912         if (registered) {
913                 err = elv_register_queue(q);
914                 if (err)
915                         goto fail_register;
916         }
917
918         /* done, kill the old one and finish */
919         elevator_exit(old);
920         blk_queue_bypass_end(q);
921
922         blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
923
924         return 0;
925
926 fail_register:
927         elevator_exit(q->elevator);
928 fail_init:
929         /* switch failed, restore and re-register old elevator */
930         q->elevator = old;
931         elv_register_queue(q);
932         blk_queue_bypass_end(q);
933
934         return err;
935 }
936
937 /*
938  * Switch this queue to the given IO scheduler.
939  */
940 int elevator_change(struct request_queue *q, const char *name)
941 {
942         char elevator_name[ELV_NAME_MAX];
943         struct elevator_type *e;
944
945         if (!q->elevator)
946                 return -ENXIO;
947
948         strlcpy(elevator_name, name, sizeof(elevator_name));
949         e = elevator_get(strstrip(elevator_name));
950         if (!e) {
951                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
952                 return -EINVAL;
953         }
954
955         if (!strcmp(elevator_name, q->elevator->type->elevator_name)) {
956                 elevator_put(e);
957                 return 0;
958         }
959
960         return elevator_switch(q, e);
961 }
962 EXPORT_SYMBOL(elevator_change);
963
964 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
965                           size_t count)
966 {
967         int ret;
968
969         if (!q->elevator)
970                 return count;
971
972         ret = elevator_change(q, name);
973         if (!ret)
974                 return count;
975
976         printk(KERN_ERR "elevator: switch to %s failed\n", name);
977         return ret;
978 }
979
980 ssize_t elv_iosched_show(struct request_queue *q, char *name)
981 {
982         struct elevator_queue *e = q->elevator;
983         struct elevator_type *elv;
984         struct elevator_type *__e;
985         int len = 0;
986
987         if (!q->elevator || !blk_queue_stackable(q))
988                 return sprintf(name, "none\n");
989
990         elv = e->type;
991
992         spin_lock(&elv_list_lock);
993         list_for_each_entry(__e, &elv_list, list) {
994                 if (!strcmp(elv->elevator_name, __e->elevator_name))
995                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
996                 else
997                         len += sprintf(name+len, "%s ", __e->elevator_name);
998         }
999         spin_unlock(&elv_list_lock);
1000
1001         len += sprintf(len+name, "\n");
1002         return len;
1003 }
1004
1005 struct request *elv_rb_former_request(struct request_queue *q,
1006                                       struct request *rq)
1007 {
1008         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1009
1010         if (rbprev)
1011                 return rb_entry_rq(rbprev);
1012
1013         return NULL;
1014 }
1015 EXPORT_SYMBOL(elv_rb_former_request);
1016
1017 struct request *elv_rb_latter_request(struct request_queue *q,
1018                                       struct request *rq)
1019 {
1020         struct rb_node *rbnext = rb_next(&rq->rb_node);
1021
1022         if (rbnext)
1023                 return rb_entry_rq(rbnext);
1024
1025         return NULL;
1026 }
1027 EXPORT_SYMBOL(elv_rb_latter_request);