Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/sparc-2.6
[linux-2.6.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/delay.h>
35 #include <linux/blktrace_api.h>
36 #include <linux/hash.h>
37 #include <linux/uaccess.h>
38
39 #include <trace/events/block.h>
40
41 #include "blk.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->ops->elevator_allow_merge_fn)
66                 return e->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 int elv_rq_merge_ok(struct request *rq, struct bio *bio)
75 {
76         if (!rq_mergeable(rq))
77                 return 0;
78
79         /*
80          * Don't merge file system requests and discard requests
81          */
82         if (bio_rw_flagged(bio, BIO_RW_DISCARD) !=
83             bio_rw_flagged(rq->bio, BIO_RW_DISCARD))
84                 return 0;
85
86         /*
87          * different data direction or already started, don't merge
88          */
89         if (bio_data_dir(bio) != rq_data_dir(rq))
90                 return 0;
91
92         /*
93          * must be same device and not a special request
94          */
95         if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
96                 return 0;
97
98         /*
99          * only merge integrity protected bio into ditto rq
100          */
101         if (bio_integrity(bio) != blk_integrity_rq(rq))
102                 return 0;
103
104         if (!elv_iosched_allow_merge(rq, bio))
105                 return 0;
106
107         return 1;
108 }
109 EXPORT_SYMBOL(elv_rq_merge_ok);
110
111 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
112 {
113         int ret = ELEVATOR_NO_MERGE;
114
115         /*
116          * we can merge and sequence is ok, check if it's possible
117          */
118         if (elv_rq_merge_ok(__rq, bio)) {
119                 if (blk_rq_pos(__rq) + blk_rq_sectors(__rq) == bio->bi_sector)
120                         ret = ELEVATOR_BACK_MERGE;
121                 else if (blk_rq_pos(__rq) - bio_sectors(bio) == bio->bi_sector)
122                         ret = ELEVATOR_FRONT_MERGE;
123         }
124
125         return ret;
126 }
127
128 static struct elevator_type *elevator_find(const char *name)
129 {
130         struct elevator_type *e;
131
132         list_for_each_entry(e, &elv_list, list) {
133                 if (!strcmp(e->elevator_name, name))
134                         return e;
135         }
136
137         return NULL;
138 }
139
140 static void elevator_put(struct elevator_type *e)
141 {
142         module_put(e->elevator_owner);
143 }
144
145 static struct elevator_type *elevator_get(const char *name)
146 {
147         struct elevator_type *e;
148
149         spin_lock(&elv_list_lock);
150
151         e = elevator_find(name);
152         if (!e) {
153                 char elv[ELV_NAME_MAX + strlen("-iosched")];
154
155                 spin_unlock(&elv_list_lock);
156
157                 sprintf(elv, "%s-iosched", name);
158
159                 request_module("%s", elv);
160                 spin_lock(&elv_list_lock);
161                 e = elevator_find(name);
162         }
163
164         if (e && !try_module_get(e->elevator_owner))
165                 e = NULL;
166
167         spin_unlock(&elv_list_lock);
168
169         return e;
170 }
171
172 static void *elevator_init_queue(struct request_queue *q,
173                                  struct elevator_queue *eq)
174 {
175         return eq->ops->elevator_init_fn(q);
176 }
177
178 static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
179                            void *data)
180 {
181         q->elevator = eq;
182         eq->elevator_data = data;
183 }
184
185 static char chosen_elevator[16];
186
187 static int __init elevator_setup(char *str)
188 {
189         /*
190          * Be backwards-compatible with previous kernels, so users
191          * won't get the wrong elevator.
192          */
193         strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
194         return 1;
195 }
196
197 __setup("elevator=", elevator_setup);
198
199 static struct kobj_type elv_ktype;
200
201 static struct elevator_queue *elevator_alloc(struct request_queue *q,
202                                   struct elevator_type *e)
203 {
204         struct elevator_queue *eq;
205         int i;
206
207         eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node);
208         if (unlikely(!eq))
209                 goto err;
210
211         eq->ops = &e->ops;
212         eq->elevator_type = e;
213         kobject_init(&eq->kobj, &elv_ktype);
214         mutex_init(&eq->sysfs_lock);
215
216         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
217                                         GFP_KERNEL, q->node);
218         if (!eq->hash)
219                 goto err;
220
221         for (i = 0; i < ELV_HASH_ENTRIES; i++)
222                 INIT_HLIST_HEAD(&eq->hash[i]);
223
224         return eq;
225 err:
226         kfree(eq);
227         elevator_put(e);
228         return NULL;
229 }
230
231 static void elevator_release(struct kobject *kobj)
232 {
233         struct elevator_queue *e;
234
235         e = container_of(kobj, struct elevator_queue, kobj);
236         elevator_put(e->elevator_type);
237         kfree(e->hash);
238         kfree(e);
239 }
240
241 int elevator_init(struct request_queue *q, char *name)
242 {
243         struct elevator_type *e = NULL;
244         struct elevator_queue *eq;
245         int ret = 0;
246         void *data;
247
248         INIT_LIST_HEAD(&q->queue_head);
249         q->last_merge = NULL;
250         q->end_sector = 0;
251         q->boundary_rq = NULL;
252
253         if (name) {
254                 e = elevator_get(name);
255                 if (!e)
256                         return -EINVAL;
257         }
258
259         if (!e && *chosen_elevator) {
260                 e = elevator_get(chosen_elevator);
261                 if (!e)
262                         printk(KERN_ERR "I/O scheduler %s not found\n",
263                                                         chosen_elevator);
264         }
265
266         if (!e) {
267                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
268                 if (!e) {
269                         printk(KERN_ERR
270                                 "Default I/O scheduler not found. " \
271                                 "Using noop.\n");
272                         e = elevator_get("noop");
273                 }
274         }
275
276         eq = elevator_alloc(q, e);
277         if (!eq)
278                 return -ENOMEM;
279
280         data = elevator_init_queue(q, eq);
281         if (!data) {
282                 kobject_put(&eq->kobj);
283                 return -ENOMEM;
284         }
285
286         elevator_attach(q, eq, data);
287         return ret;
288 }
289 EXPORT_SYMBOL(elevator_init);
290
291 void elevator_exit(struct elevator_queue *e)
292 {
293         mutex_lock(&e->sysfs_lock);
294         if (e->ops->elevator_exit_fn)
295                 e->ops->elevator_exit_fn(e);
296         e->ops = NULL;
297         mutex_unlock(&e->sysfs_lock);
298
299         kobject_put(&e->kobj);
300 }
301 EXPORT_SYMBOL(elevator_exit);
302
303 static inline void __elv_rqhash_del(struct request *rq)
304 {
305         hlist_del_init(&rq->hash);
306 }
307
308 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
309 {
310         if (ELV_ON_HASH(rq))
311                 __elv_rqhash_del(rq);
312 }
313
314 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
315 {
316         struct elevator_queue *e = q->elevator;
317
318         BUG_ON(ELV_ON_HASH(rq));
319         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
320 }
321
322 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
323 {
324         __elv_rqhash_del(rq);
325         elv_rqhash_add(q, rq);
326 }
327
328 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
329 {
330         struct elevator_queue *e = q->elevator;
331         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
332         struct hlist_node *entry, *next;
333         struct request *rq;
334
335         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
336                 BUG_ON(!ELV_ON_HASH(rq));
337
338                 if (unlikely(!rq_mergeable(rq))) {
339                         __elv_rqhash_del(rq);
340                         continue;
341                 }
342
343                 if (rq_hash_key(rq) == offset)
344                         return rq;
345         }
346
347         return NULL;
348 }
349
350 /*
351  * RB-tree support functions for inserting/lookup/removal of requests
352  * in a sorted RB tree.
353  */
354 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
355 {
356         struct rb_node **p = &root->rb_node;
357         struct rb_node *parent = NULL;
358         struct request *__rq;
359
360         while (*p) {
361                 parent = *p;
362                 __rq = rb_entry(parent, struct request, rb_node);
363
364                 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
365                         p = &(*p)->rb_left;
366                 else if (blk_rq_pos(rq) > blk_rq_pos(__rq))
367                         p = &(*p)->rb_right;
368                 else
369                         return __rq;
370         }
371
372         rb_link_node(&rq->rb_node, parent, p);
373         rb_insert_color(&rq->rb_node, root);
374         return NULL;
375 }
376 EXPORT_SYMBOL(elv_rb_add);
377
378 void elv_rb_del(struct rb_root *root, struct request *rq)
379 {
380         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
381         rb_erase(&rq->rb_node, root);
382         RB_CLEAR_NODE(&rq->rb_node);
383 }
384 EXPORT_SYMBOL(elv_rb_del);
385
386 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
387 {
388         struct rb_node *n = root->rb_node;
389         struct request *rq;
390
391         while (n) {
392                 rq = rb_entry(n, struct request, rb_node);
393
394                 if (sector < blk_rq_pos(rq))
395                         n = n->rb_left;
396                 else if (sector > blk_rq_pos(rq))
397                         n = n->rb_right;
398                 else
399                         return rq;
400         }
401
402         return NULL;
403 }
404 EXPORT_SYMBOL(elv_rb_find);
405
406 /*
407  * Insert rq into dispatch queue of q.  Queue lock must be held on
408  * entry.  rq is sort instead into the dispatch queue. To be used by
409  * specific elevators.
410  */
411 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
412 {
413         sector_t boundary;
414         struct list_head *entry;
415         int stop_flags;
416
417         if (q->last_merge == rq)
418                 q->last_merge = NULL;
419
420         elv_rqhash_del(q, rq);
421
422         q->nr_sorted--;
423
424         boundary = q->end_sector;
425         stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
426         list_for_each_prev(entry, &q->queue_head) {
427                 struct request *pos = list_entry_rq(entry);
428
429                 if (blk_discard_rq(rq) != blk_discard_rq(pos))
430                         break;
431                 if (rq_data_dir(rq) != rq_data_dir(pos))
432                         break;
433                 if (pos->cmd_flags & stop_flags)
434                         break;
435                 if (blk_rq_pos(rq) >= boundary) {
436                         if (blk_rq_pos(pos) < boundary)
437                                 continue;
438                 } else {
439                         if (blk_rq_pos(pos) >= boundary)
440                                 break;
441                 }
442                 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
443                         break;
444         }
445
446         list_add(&rq->queuelist, entry);
447 }
448 EXPORT_SYMBOL(elv_dispatch_sort);
449
450 /*
451  * Insert rq into dispatch queue of q.  Queue lock must be held on
452  * entry.  rq is added to the back of the dispatch queue. To be used by
453  * specific elevators.
454  */
455 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
456 {
457         if (q->last_merge == rq)
458                 q->last_merge = NULL;
459
460         elv_rqhash_del(q, rq);
461
462         q->nr_sorted--;
463
464         q->end_sector = rq_end_sector(rq);
465         q->boundary_rq = rq;
466         list_add_tail(&rq->queuelist, &q->queue_head);
467 }
468 EXPORT_SYMBOL(elv_dispatch_add_tail);
469
470 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
471 {
472         struct elevator_queue *e = q->elevator;
473         struct request *__rq;
474         int ret;
475
476         /*
477          * Levels of merges:
478          *      nomerges:  No merges at all attempted
479          *      noxmerges: Only simple one-hit cache try
480          *      merges:    All merge tries attempted
481          */
482         if (blk_queue_nomerges(q))
483                 return ELEVATOR_NO_MERGE;
484
485         /*
486          * First try one-hit cache.
487          */
488         if (q->last_merge) {
489                 ret = elv_try_merge(q->last_merge, bio);
490                 if (ret != ELEVATOR_NO_MERGE) {
491                         *req = q->last_merge;
492                         return ret;
493                 }
494         }
495
496         if (blk_queue_noxmerges(q))
497                 return ELEVATOR_NO_MERGE;
498
499         /*
500          * See if our hash lookup can find a potential backmerge.
501          */
502         __rq = elv_rqhash_find(q, bio->bi_sector);
503         if (__rq && elv_rq_merge_ok(__rq, bio)) {
504                 *req = __rq;
505                 return ELEVATOR_BACK_MERGE;
506         }
507
508         if (e->ops->elevator_merge_fn)
509                 return e->ops->elevator_merge_fn(q, req, bio);
510
511         return ELEVATOR_NO_MERGE;
512 }
513
514 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
515 {
516         struct elevator_queue *e = q->elevator;
517
518         if (e->ops->elevator_merged_fn)
519                 e->ops->elevator_merged_fn(q, rq, type);
520
521         if (type == ELEVATOR_BACK_MERGE)
522                 elv_rqhash_reposition(q, rq);
523
524         q->last_merge = rq;
525 }
526
527 void elv_merge_requests(struct request_queue *q, struct request *rq,
528                              struct request *next)
529 {
530         struct elevator_queue *e = q->elevator;
531
532         if (e->ops->elevator_merge_req_fn)
533                 e->ops->elevator_merge_req_fn(q, rq, next);
534
535         elv_rqhash_reposition(q, rq);
536         elv_rqhash_del(q, next);
537
538         q->nr_sorted--;
539         q->last_merge = rq;
540 }
541
542 void elv_requeue_request(struct request_queue *q, struct request *rq)
543 {
544         /*
545          * it already went through dequeue, we need to decrement the
546          * in_flight count again
547          */
548         if (blk_account_rq(rq)) {
549                 q->in_flight[rq_is_sync(rq)]--;
550                 if (blk_sorted_rq(rq))
551                         elv_deactivate_rq(q, rq);
552         }
553
554         rq->cmd_flags &= ~REQ_STARTED;
555
556         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
557 }
558
559 void elv_drain_elevator(struct request_queue *q)
560 {
561         static int printed;
562         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
563                 ;
564         if (q->nr_sorted == 0)
565                 return;
566         if (printed++ < 10) {
567                 printk(KERN_ERR "%s: forced dispatching is broken "
568                        "(nr_sorted=%u), please report this\n",
569                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
570         }
571 }
572
573 /*
574  * Call with queue lock held, interrupts disabled
575  */
576 void elv_quiesce_start(struct request_queue *q)
577 {
578         if (!q->elevator)
579                 return;
580
581         queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
582
583         /*
584          * make sure we don't have any requests in flight
585          */
586         elv_drain_elevator(q);
587         while (q->rq.elvpriv) {
588                 __blk_run_queue(q);
589                 spin_unlock_irq(q->queue_lock);
590                 msleep(10);
591                 spin_lock_irq(q->queue_lock);
592                 elv_drain_elevator(q);
593         }
594 }
595
596 void elv_quiesce_end(struct request_queue *q)
597 {
598         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
599 }
600
601 void elv_insert(struct request_queue *q, struct request *rq, int where)
602 {
603         struct list_head *pos;
604         unsigned ordseq;
605         int unplug_it = 1;
606
607         trace_block_rq_insert(q, rq);
608
609         rq->q = q;
610
611         switch (where) {
612         case ELEVATOR_INSERT_FRONT:
613                 rq->cmd_flags |= REQ_SOFTBARRIER;
614
615                 list_add(&rq->queuelist, &q->queue_head);
616                 break;
617
618         case ELEVATOR_INSERT_BACK:
619                 rq->cmd_flags |= REQ_SOFTBARRIER;
620                 elv_drain_elevator(q);
621                 list_add_tail(&rq->queuelist, &q->queue_head);
622                 /*
623                  * We kick the queue here for the following reasons.
624                  * - The elevator might have returned NULL previously
625                  *   to delay requests and returned them now.  As the
626                  *   queue wasn't empty before this request, ll_rw_blk
627                  *   won't run the queue on return, resulting in hang.
628                  * - Usually, back inserted requests won't be merged
629                  *   with anything.  There's no point in delaying queue
630                  *   processing.
631                  */
632                 __blk_run_queue(q);
633                 break;
634
635         case ELEVATOR_INSERT_SORT:
636                 BUG_ON(!blk_fs_request(rq) && !blk_discard_rq(rq));
637                 rq->cmd_flags |= REQ_SORTED;
638                 q->nr_sorted++;
639                 if (rq_mergeable(rq)) {
640                         elv_rqhash_add(q, rq);
641                         if (!q->last_merge)
642                                 q->last_merge = rq;
643                 }
644
645                 /*
646                  * Some ioscheds (cfq) run q->request_fn directly, so
647                  * rq cannot be accessed after calling
648                  * elevator_add_req_fn.
649                  */
650                 q->elevator->ops->elevator_add_req_fn(q, rq);
651                 break;
652
653         case ELEVATOR_INSERT_REQUEUE:
654                 /*
655                  * If ordered flush isn't in progress, we do front
656                  * insertion; otherwise, requests should be requeued
657                  * in ordseq order.
658                  */
659                 rq->cmd_flags |= REQ_SOFTBARRIER;
660
661                 /*
662                  * Most requeues happen because of a busy condition,
663                  * don't force unplug of the queue for that case.
664                  */
665                 unplug_it = 0;
666
667                 if (q->ordseq == 0) {
668                         list_add(&rq->queuelist, &q->queue_head);
669                         break;
670                 }
671
672                 ordseq = blk_ordered_req_seq(rq);
673
674                 list_for_each(pos, &q->queue_head) {
675                         struct request *pos_rq = list_entry_rq(pos);
676                         if (ordseq <= blk_ordered_req_seq(pos_rq))
677                                 break;
678                 }
679
680                 list_add_tail(&rq->queuelist, pos);
681                 break;
682
683         default:
684                 printk(KERN_ERR "%s: bad insertion point %d\n",
685                        __func__, where);
686                 BUG();
687         }
688
689         if (unplug_it && blk_queue_plugged(q)) {
690                 int nrq = q->rq.count[BLK_RW_SYNC] + q->rq.count[BLK_RW_ASYNC]
691                                 - queue_in_flight(q);
692
693                 if (nrq >= q->unplug_thresh)
694                         __generic_unplug_device(q);
695         }
696 }
697
698 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
699                        int plug)
700 {
701         if (q->ordcolor)
702                 rq->cmd_flags |= REQ_ORDERED_COLOR;
703
704         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
705                 /*
706                  * toggle ordered color
707                  */
708                 if (blk_barrier_rq(rq))
709                         q->ordcolor ^= 1;
710
711                 /*
712                  * barriers implicitly indicate back insertion
713                  */
714                 if (where == ELEVATOR_INSERT_SORT)
715                         where = ELEVATOR_INSERT_BACK;
716
717                 /*
718                  * this request is scheduling boundary, update
719                  * end_sector
720                  */
721                 if (blk_fs_request(rq) || blk_discard_rq(rq)) {
722                         q->end_sector = rq_end_sector(rq);
723                         q->boundary_rq = rq;
724                 }
725         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
726                     where == ELEVATOR_INSERT_SORT)
727                 where = ELEVATOR_INSERT_BACK;
728
729         if (plug)
730                 blk_plug_device(q);
731
732         elv_insert(q, rq, where);
733 }
734 EXPORT_SYMBOL(__elv_add_request);
735
736 void elv_add_request(struct request_queue *q, struct request *rq, int where,
737                      int plug)
738 {
739         unsigned long flags;
740
741         spin_lock_irqsave(q->queue_lock, flags);
742         __elv_add_request(q, rq, where, plug);
743         spin_unlock_irqrestore(q->queue_lock, flags);
744 }
745 EXPORT_SYMBOL(elv_add_request);
746
747 int elv_queue_empty(struct request_queue *q)
748 {
749         struct elevator_queue *e = q->elevator;
750
751         if (!list_empty(&q->queue_head))
752                 return 0;
753
754         if (e->ops->elevator_queue_empty_fn)
755                 return e->ops->elevator_queue_empty_fn(q);
756
757         return 1;
758 }
759 EXPORT_SYMBOL(elv_queue_empty);
760
761 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
762 {
763         struct elevator_queue *e = q->elevator;
764
765         if (e->ops->elevator_latter_req_fn)
766                 return e->ops->elevator_latter_req_fn(q, rq);
767         return NULL;
768 }
769
770 struct request *elv_former_request(struct request_queue *q, struct request *rq)
771 {
772         struct elevator_queue *e = q->elevator;
773
774         if (e->ops->elevator_former_req_fn)
775                 return e->ops->elevator_former_req_fn(q, rq);
776         return NULL;
777 }
778
779 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
780 {
781         struct elevator_queue *e = q->elevator;
782
783         if (e->ops->elevator_set_req_fn)
784                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
785
786         rq->elevator_private = NULL;
787         return 0;
788 }
789
790 void elv_put_request(struct request_queue *q, struct request *rq)
791 {
792         struct elevator_queue *e = q->elevator;
793
794         if (e->ops->elevator_put_req_fn)
795                 e->ops->elevator_put_req_fn(rq);
796 }
797
798 int elv_may_queue(struct request_queue *q, int rw)
799 {
800         struct elevator_queue *e = q->elevator;
801
802         if (e->ops->elevator_may_queue_fn)
803                 return e->ops->elevator_may_queue_fn(q, rw);
804
805         return ELV_MQUEUE_MAY;
806 }
807
808 void elv_abort_queue(struct request_queue *q)
809 {
810         struct request *rq;
811
812         while (!list_empty(&q->queue_head)) {
813                 rq = list_entry_rq(q->queue_head.next);
814                 rq->cmd_flags |= REQ_QUIET;
815                 trace_block_rq_abort(q, rq);
816                 /*
817                  * Mark this request as started so we don't trigger
818                  * any debug logic in the end I/O path.
819                  */
820                 blk_start_request(rq);
821                 __blk_end_request_all(rq, -EIO);
822         }
823 }
824 EXPORT_SYMBOL(elv_abort_queue);
825
826 void elv_completed_request(struct request_queue *q, struct request *rq)
827 {
828         struct elevator_queue *e = q->elevator;
829
830         /*
831          * request is released from the driver, io must be done
832          */
833         if (blk_account_rq(rq)) {
834                 q->in_flight[rq_is_sync(rq)]--;
835                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
836                         e->ops->elevator_completed_req_fn(q, rq);
837         }
838
839         /*
840          * Check if the queue is waiting for fs requests to be
841          * drained for flush sequence.
842          */
843         if (unlikely(q->ordseq)) {
844                 struct request *next = NULL;
845
846                 if (!list_empty(&q->queue_head))
847                         next = list_entry_rq(q->queue_head.next);
848
849                 if (!queue_in_flight(q) &&
850                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
851                     (!next || blk_ordered_req_seq(next) > QUEUE_ORDSEQ_DRAIN)) {
852                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
853                         __blk_run_queue(q);
854                 }
855         }
856 }
857
858 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
859
860 static ssize_t
861 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
862 {
863         struct elv_fs_entry *entry = to_elv(attr);
864         struct elevator_queue *e;
865         ssize_t error;
866
867         if (!entry->show)
868                 return -EIO;
869
870         e = container_of(kobj, struct elevator_queue, kobj);
871         mutex_lock(&e->sysfs_lock);
872         error = e->ops ? entry->show(e, page) : -ENOENT;
873         mutex_unlock(&e->sysfs_lock);
874         return error;
875 }
876
877 static ssize_t
878 elv_attr_store(struct kobject *kobj, struct attribute *attr,
879                const char *page, size_t length)
880 {
881         struct elv_fs_entry *entry = to_elv(attr);
882         struct elevator_queue *e;
883         ssize_t error;
884
885         if (!entry->store)
886                 return -EIO;
887
888         e = container_of(kobj, struct elevator_queue, kobj);
889         mutex_lock(&e->sysfs_lock);
890         error = e->ops ? entry->store(e, page, length) : -ENOENT;
891         mutex_unlock(&e->sysfs_lock);
892         return error;
893 }
894
895 static const struct sysfs_ops elv_sysfs_ops = {
896         .show   = elv_attr_show,
897         .store  = elv_attr_store,
898 };
899
900 static struct kobj_type elv_ktype = {
901         .sysfs_ops      = &elv_sysfs_ops,
902         .release        = elevator_release,
903 };
904
905 int elv_register_queue(struct request_queue *q)
906 {
907         struct elevator_queue *e = q->elevator;
908         int error;
909
910         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
911         if (!error) {
912                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
913                 if (attr) {
914                         while (attr->attr.name) {
915                                 if (sysfs_create_file(&e->kobj, &attr->attr))
916                                         break;
917                                 attr++;
918                         }
919                 }
920                 kobject_uevent(&e->kobj, KOBJ_ADD);
921         }
922         return error;
923 }
924
925 static void __elv_unregister_queue(struct elevator_queue *e)
926 {
927         kobject_uevent(&e->kobj, KOBJ_REMOVE);
928         kobject_del(&e->kobj);
929 }
930
931 void elv_unregister_queue(struct request_queue *q)
932 {
933         if (q)
934                 __elv_unregister_queue(q->elevator);
935 }
936
937 void elv_register(struct elevator_type *e)
938 {
939         char *def = "";
940
941         spin_lock(&elv_list_lock);
942         BUG_ON(elevator_find(e->elevator_name));
943         list_add_tail(&e->list, &elv_list);
944         spin_unlock(&elv_list_lock);
945
946         if (!strcmp(e->elevator_name, chosen_elevator) ||
947                         (!*chosen_elevator &&
948                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
949                                 def = " (default)";
950
951         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
952                                                                 def);
953 }
954 EXPORT_SYMBOL_GPL(elv_register);
955
956 void elv_unregister(struct elevator_type *e)
957 {
958         struct task_struct *g, *p;
959
960         /*
961          * Iterate every thread in the process to remove the io contexts.
962          */
963         if (e->ops.trim) {
964                 read_lock(&tasklist_lock);
965                 do_each_thread(g, p) {
966                         task_lock(p);
967                         if (p->io_context)
968                                 e->ops.trim(p->io_context);
969                         task_unlock(p);
970                 } while_each_thread(g, p);
971                 read_unlock(&tasklist_lock);
972         }
973
974         spin_lock(&elv_list_lock);
975         list_del_init(&e->list);
976         spin_unlock(&elv_list_lock);
977 }
978 EXPORT_SYMBOL_GPL(elv_unregister);
979
980 /*
981  * switch to new_e io scheduler. be careful not to introduce deadlocks -
982  * we don't free the old io scheduler, before we have allocated what we
983  * need for the new one. this way we have a chance of going back to the old
984  * one, if the new one fails init for some reason.
985  */
986 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
987 {
988         struct elevator_queue *old_elevator, *e;
989         void *data;
990
991         /*
992          * Allocate new elevator
993          */
994         e = elevator_alloc(q, new_e);
995         if (!e)
996                 return 0;
997
998         data = elevator_init_queue(q, e);
999         if (!data) {
1000                 kobject_put(&e->kobj);
1001                 return 0;
1002         }
1003
1004         /*
1005          * Turn on BYPASS and drain all requests w/ elevator private data
1006          */
1007         spin_lock_irq(q->queue_lock);
1008         elv_quiesce_start(q);
1009
1010         /*
1011          * Remember old elevator.
1012          */
1013         old_elevator = q->elevator;
1014
1015         /*
1016          * attach and start new elevator
1017          */
1018         elevator_attach(q, e, data);
1019
1020         spin_unlock_irq(q->queue_lock);
1021
1022         __elv_unregister_queue(old_elevator);
1023
1024         if (elv_register_queue(q))
1025                 goto fail_register;
1026
1027         /*
1028          * finally exit old elevator and turn off BYPASS.
1029          */
1030         elevator_exit(old_elevator);
1031         spin_lock_irq(q->queue_lock);
1032         elv_quiesce_end(q);
1033         spin_unlock_irq(q->queue_lock);
1034
1035         blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
1036
1037         return 1;
1038
1039 fail_register:
1040         /*
1041          * switch failed, exit the new io scheduler and reattach the old
1042          * one again (along with re-adding the sysfs dir)
1043          */
1044         elevator_exit(e);
1045         q->elevator = old_elevator;
1046         elv_register_queue(q);
1047
1048         spin_lock_irq(q->queue_lock);
1049         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1050         spin_unlock_irq(q->queue_lock);
1051
1052         return 0;
1053 }
1054
1055 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1056                           size_t count)
1057 {
1058         char elevator_name[ELV_NAME_MAX];
1059         struct elevator_type *e;
1060
1061         if (!q->elevator)
1062                 return count;
1063
1064         strlcpy(elevator_name, name, sizeof(elevator_name));
1065         e = elevator_get(strstrip(elevator_name));
1066         if (!e) {
1067                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1068                 return -EINVAL;
1069         }
1070
1071         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1072                 elevator_put(e);
1073                 return count;
1074         }
1075
1076         if (!elevator_switch(q, e))
1077                 printk(KERN_ERR "elevator: switch to %s failed\n",
1078                                                         elevator_name);
1079         return count;
1080 }
1081
1082 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1083 {
1084         struct elevator_queue *e = q->elevator;
1085         struct elevator_type *elv;
1086         struct elevator_type *__e;
1087         int len = 0;
1088
1089         if (!q->elevator)
1090                 return sprintf(name, "none\n");
1091
1092         elv = e->elevator_type;
1093
1094         spin_lock(&elv_list_lock);
1095         list_for_each_entry(__e, &elv_list, list) {
1096                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1097                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1098                 else
1099                         len += sprintf(name+len, "%s ", __e->elevator_name);
1100         }
1101         spin_unlock(&elv_list_lock);
1102
1103         len += sprintf(len+name, "\n");
1104         return len;
1105 }
1106
1107 struct request *elv_rb_former_request(struct request_queue *q,
1108                                       struct request *rq)
1109 {
1110         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1111
1112         if (rbprev)
1113                 return rb_entry_rq(rbprev);
1114
1115         return NULL;
1116 }
1117 EXPORT_SYMBOL(elv_rb_former_request);
1118
1119 struct request *elv_rb_latter_request(struct request_queue *q,
1120                                       struct request *rq)
1121 {
1122         struct rb_node *rbnext = rb_next(&rq->rb_node);
1123
1124         if (rbnext)
1125                 return rb_entry_rq(rbnext);
1126
1127         return NULL;
1128 }
1129 EXPORT_SYMBOL(elv_rb_latter_request);