Merge branch 'topic/msnd' into for-linus
[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 <trace/block.h>
37 #include <linux/hash.h>
38 #include <linux/uaccess.h>
39
40 #include "blk.h"
41
42 static DEFINE_SPINLOCK(elv_list_lock);
43 static LIST_HEAD(elv_list);
44
45 DEFINE_TRACE(block_rq_abort);
46
47 /*
48  * Merge hash stuff.
49  */
50 static const int elv_hash_shift = 6;
51 #define ELV_HASH_BLOCK(sec)     ((sec) >> 3)
52 #define ELV_HASH_FN(sec)        \
53                 (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
54 #define ELV_HASH_ENTRIES        (1 << elv_hash_shift)
55 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
56 #define ELV_ON_HASH(rq)         (!hlist_unhashed(&(rq)->hash))
57
58 DEFINE_TRACE(block_rq_insert);
59 DEFINE_TRACE(block_rq_issue);
60
61 /*
62  * Query io scheduler to see if the current process issuing bio may be
63  * merged with rq.
64  */
65 static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
66 {
67         struct request_queue *q = rq->q;
68         struct elevator_queue *e = q->elevator;
69
70         if (e->ops->elevator_allow_merge_fn)
71                 return e->ops->elevator_allow_merge_fn(q, rq, bio);
72
73         return 1;
74 }
75
76 /*
77  * can we safely merge with this request?
78  */
79 int elv_rq_merge_ok(struct request *rq, struct bio *bio)
80 {
81         if (!rq_mergeable(rq))
82                 return 0;
83
84         /*
85          * Don't merge file system requests and discard requests
86          */
87         if (bio_discard(bio) != bio_discard(rq->bio))
88                 return 0;
89
90         /*
91          * different data direction or already started, don't merge
92          */
93         if (bio_data_dir(bio) != rq_data_dir(rq))
94                 return 0;
95
96         /*
97          * must be same device and not a special request
98          */
99         if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
100                 return 0;
101
102         /*
103          * only merge integrity protected bio into ditto rq
104          */
105         if (bio_integrity(bio) != blk_integrity_rq(rq))
106                 return 0;
107
108         if (!elv_iosched_allow_merge(rq, bio))
109                 return 0;
110
111         return 1;
112 }
113 EXPORT_SYMBOL(elv_rq_merge_ok);
114
115 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
116 {
117         int ret = ELEVATOR_NO_MERGE;
118
119         /*
120          * we can merge and sequence is ok, check if it's possible
121          */
122         if (elv_rq_merge_ok(__rq, bio)) {
123                 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
124                         ret = ELEVATOR_BACK_MERGE;
125                 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
126                         ret = ELEVATOR_FRONT_MERGE;
127         }
128
129         return ret;
130 }
131
132 static struct elevator_type *elevator_find(const char *name)
133 {
134         struct elevator_type *e;
135
136         list_for_each_entry(e, &elv_list, list) {
137                 if (!strcmp(e->elevator_name, name))
138                         return e;
139         }
140
141         return NULL;
142 }
143
144 static void elevator_put(struct elevator_type *e)
145 {
146         module_put(e->elevator_owner);
147 }
148
149 static struct elevator_type *elevator_get(const char *name)
150 {
151         struct elevator_type *e;
152
153         spin_lock(&elv_list_lock);
154
155         e = elevator_find(name);
156         if (!e) {
157                 char elv[ELV_NAME_MAX + strlen("-iosched")];
158
159                 spin_unlock(&elv_list_lock);
160
161                 if (!strcmp(name, "anticipatory"))
162                         sprintf(elv, "as-iosched");
163                 else
164                         sprintf(elv, "%s-iosched", name);
165
166                 request_module("%s", elv);
167                 spin_lock(&elv_list_lock);
168                 e = elevator_find(name);
169         }
170
171         if (e && !try_module_get(e->elevator_owner))
172                 e = NULL;
173
174         spin_unlock(&elv_list_lock);
175
176         return e;
177 }
178
179 static void *elevator_init_queue(struct request_queue *q,
180                                  struct elevator_queue *eq)
181 {
182         return eq->ops->elevator_init_fn(q);
183 }
184
185 static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
186                            void *data)
187 {
188         q->elevator = eq;
189         eq->elevator_data = data;
190 }
191
192 static char chosen_elevator[16];
193
194 static int __init elevator_setup(char *str)
195 {
196         /*
197          * Be backwards-compatible with previous kernels, so users
198          * won't get the wrong elevator.
199          */
200         if (!strcmp(str, "as"))
201                 strcpy(chosen_elevator, "anticipatory");
202         else
203                 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
204         return 1;
205 }
206
207 __setup("elevator=", elevator_setup);
208
209 static struct kobj_type elv_ktype;
210
211 static struct elevator_queue *elevator_alloc(struct request_queue *q,
212                                   struct elevator_type *e)
213 {
214         struct elevator_queue *eq;
215         int i;
216
217         eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node);
218         if (unlikely(!eq))
219                 goto err;
220
221         eq->ops = &e->ops;
222         eq->elevator_type = e;
223         kobject_init(&eq->kobj, &elv_ktype);
224         mutex_init(&eq->sysfs_lock);
225
226         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
227                                         GFP_KERNEL, q->node);
228         if (!eq->hash)
229                 goto err;
230
231         for (i = 0; i < ELV_HASH_ENTRIES; i++)
232                 INIT_HLIST_HEAD(&eq->hash[i]);
233
234         return eq;
235 err:
236         kfree(eq);
237         elevator_put(e);
238         return NULL;
239 }
240
241 static void elevator_release(struct kobject *kobj)
242 {
243         struct elevator_queue *e;
244
245         e = container_of(kobj, struct elevator_queue, kobj);
246         elevator_put(e->elevator_type);
247         kfree(e->hash);
248         kfree(e);
249 }
250
251 int elevator_init(struct request_queue *q, char *name)
252 {
253         struct elevator_type *e = NULL;
254         struct elevator_queue *eq;
255         int ret = 0;
256         void *data;
257
258         INIT_LIST_HEAD(&q->queue_head);
259         q->last_merge = NULL;
260         q->end_sector = 0;
261         q->boundary_rq = NULL;
262
263         if (name) {
264                 e = elevator_get(name);
265                 if (!e)
266                         return -EINVAL;
267         }
268
269         if (!e && *chosen_elevator) {
270                 e = elevator_get(chosen_elevator);
271                 if (!e)
272                         printk(KERN_ERR "I/O scheduler %s not found\n",
273                                                         chosen_elevator);
274         }
275
276         if (!e) {
277                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
278                 if (!e) {
279                         printk(KERN_ERR
280                                 "Default I/O scheduler not found. " \
281                                 "Using noop.\n");
282                         e = elevator_get("noop");
283                 }
284         }
285
286         eq = elevator_alloc(q, e);
287         if (!eq)
288                 return -ENOMEM;
289
290         data = elevator_init_queue(q, eq);
291         if (!data) {
292                 kobject_put(&eq->kobj);
293                 return -ENOMEM;
294         }
295
296         elevator_attach(q, eq, data);
297         return ret;
298 }
299 EXPORT_SYMBOL(elevator_init);
300
301 void elevator_exit(struct elevator_queue *e)
302 {
303         mutex_lock(&e->sysfs_lock);
304         if (e->ops->elevator_exit_fn)
305                 e->ops->elevator_exit_fn(e);
306         e->ops = NULL;
307         mutex_unlock(&e->sysfs_lock);
308
309         kobject_put(&e->kobj);
310 }
311 EXPORT_SYMBOL(elevator_exit);
312
313 static void elv_activate_rq(struct request_queue *q, struct request *rq)
314 {
315         struct elevator_queue *e = q->elevator;
316
317         if (e->ops->elevator_activate_req_fn)
318                 e->ops->elevator_activate_req_fn(q, rq);
319 }
320
321 static void elv_deactivate_rq(struct request_queue *q, struct request *rq)
322 {
323         struct elevator_queue *e = q->elevator;
324
325         if (e->ops->elevator_deactivate_req_fn)
326                 e->ops->elevator_deactivate_req_fn(q, rq);
327 }
328
329 static inline void __elv_rqhash_del(struct request *rq)
330 {
331         hlist_del_init(&rq->hash);
332 }
333
334 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
335 {
336         if (ELV_ON_HASH(rq))
337                 __elv_rqhash_del(rq);
338 }
339
340 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
341 {
342         struct elevator_queue *e = q->elevator;
343
344         BUG_ON(ELV_ON_HASH(rq));
345         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
346 }
347
348 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
349 {
350         __elv_rqhash_del(rq);
351         elv_rqhash_add(q, rq);
352 }
353
354 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
355 {
356         struct elevator_queue *e = q->elevator;
357         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
358         struct hlist_node *entry, *next;
359         struct request *rq;
360
361         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
362                 BUG_ON(!ELV_ON_HASH(rq));
363
364                 if (unlikely(!rq_mergeable(rq))) {
365                         __elv_rqhash_del(rq);
366                         continue;
367                 }
368
369                 if (rq_hash_key(rq) == offset)
370                         return rq;
371         }
372
373         return NULL;
374 }
375
376 /*
377  * RB-tree support functions for inserting/lookup/removal of requests
378  * in a sorted RB tree.
379  */
380 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
381 {
382         struct rb_node **p = &root->rb_node;
383         struct rb_node *parent = NULL;
384         struct request *__rq;
385
386         while (*p) {
387                 parent = *p;
388                 __rq = rb_entry(parent, struct request, rb_node);
389
390                 if (rq->sector < __rq->sector)
391                         p = &(*p)->rb_left;
392                 else if (rq->sector > __rq->sector)
393                         p = &(*p)->rb_right;
394                 else
395                         return __rq;
396         }
397
398         rb_link_node(&rq->rb_node, parent, p);
399         rb_insert_color(&rq->rb_node, root);
400         return NULL;
401 }
402 EXPORT_SYMBOL(elv_rb_add);
403
404 void elv_rb_del(struct rb_root *root, struct request *rq)
405 {
406         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
407         rb_erase(&rq->rb_node, root);
408         RB_CLEAR_NODE(&rq->rb_node);
409 }
410 EXPORT_SYMBOL(elv_rb_del);
411
412 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
413 {
414         struct rb_node *n = root->rb_node;
415         struct request *rq;
416
417         while (n) {
418                 rq = rb_entry(n, struct request, rb_node);
419
420                 if (sector < rq->sector)
421                         n = n->rb_left;
422                 else if (sector > rq->sector)
423                         n = n->rb_right;
424                 else
425                         return rq;
426         }
427
428         return NULL;
429 }
430 EXPORT_SYMBOL(elv_rb_find);
431
432 /*
433  * Insert rq into dispatch queue of q.  Queue lock must be held on
434  * entry.  rq is sort instead into the dispatch queue. To be used by
435  * specific elevators.
436  */
437 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
438 {
439         sector_t boundary;
440         struct list_head *entry;
441         int stop_flags;
442
443         if (q->last_merge == rq)
444                 q->last_merge = NULL;
445
446         elv_rqhash_del(q, rq);
447
448         q->nr_sorted--;
449
450         boundary = q->end_sector;
451         stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
452         list_for_each_prev(entry, &q->queue_head) {
453                 struct request *pos = list_entry_rq(entry);
454
455                 if (blk_discard_rq(rq) != blk_discard_rq(pos))
456                         break;
457                 if (rq_data_dir(rq) != rq_data_dir(pos))
458                         break;
459                 if (pos->cmd_flags & stop_flags)
460                         break;
461                 if (rq->sector >= boundary) {
462                         if (pos->sector < boundary)
463                                 continue;
464                 } else {
465                         if (pos->sector >= boundary)
466                                 break;
467                 }
468                 if (rq->sector >= pos->sector)
469                         break;
470         }
471
472         list_add(&rq->queuelist, entry);
473 }
474 EXPORT_SYMBOL(elv_dispatch_sort);
475
476 /*
477  * Insert rq into dispatch queue of q.  Queue lock must be held on
478  * entry.  rq is added to the back of the dispatch queue. To be used by
479  * specific elevators.
480  */
481 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
482 {
483         if (q->last_merge == rq)
484                 q->last_merge = NULL;
485
486         elv_rqhash_del(q, rq);
487
488         q->nr_sorted--;
489
490         q->end_sector = rq_end_sector(rq);
491         q->boundary_rq = rq;
492         list_add_tail(&rq->queuelist, &q->queue_head);
493 }
494 EXPORT_SYMBOL(elv_dispatch_add_tail);
495
496 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
497 {
498         struct elevator_queue *e = q->elevator;
499         struct request *__rq;
500         int ret;
501
502         /*
503          * First try one-hit cache.
504          */
505         if (q->last_merge) {
506                 ret = elv_try_merge(q->last_merge, bio);
507                 if (ret != ELEVATOR_NO_MERGE) {
508                         *req = q->last_merge;
509                         return ret;
510                 }
511         }
512
513         if (blk_queue_nomerges(q))
514                 return ELEVATOR_NO_MERGE;
515
516         /*
517          * See if our hash lookup can find a potential backmerge.
518          */
519         __rq = elv_rqhash_find(q, bio->bi_sector);
520         if (__rq && elv_rq_merge_ok(__rq, bio)) {
521                 *req = __rq;
522                 return ELEVATOR_BACK_MERGE;
523         }
524
525         if (e->ops->elevator_merge_fn)
526                 return e->ops->elevator_merge_fn(q, req, bio);
527
528         return ELEVATOR_NO_MERGE;
529 }
530
531 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
532 {
533         struct elevator_queue *e = q->elevator;
534
535         if (e->ops->elevator_merged_fn)
536                 e->ops->elevator_merged_fn(q, rq, type);
537
538         if (type == ELEVATOR_BACK_MERGE)
539                 elv_rqhash_reposition(q, rq);
540
541         q->last_merge = rq;
542 }
543
544 void elv_merge_requests(struct request_queue *q, struct request *rq,
545                              struct request *next)
546 {
547         struct elevator_queue *e = q->elevator;
548
549         if (e->ops->elevator_merge_req_fn)
550                 e->ops->elevator_merge_req_fn(q, rq, next);
551
552         elv_rqhash_reposition(q, rq);
553         elv_rqhash_del(q, next);
554
555         q->nr_sorted--;
556         q->last_merge = rq;
557 }
558
559 void elv_requeue_request(struct request_queue *q, struct request *rq)
560 {
561         /*
562          * it already went through dequeue, we need to decrement the
563          * in_flight count again
564          */
565         if (blk_account_rq(rq)) {
566                 q->in_flight--;
567                 if (blk_sorted_rq(rq))
568                         elv_deactivate_rq(q, rq);
569         }
570
571         rq->cmd_flags &= ~REQ_STARTED;
572
573         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
574 }
575
576 static void elv_drain_elevator(struct request_queue *q)
577 {
578         static int printed;
579         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
580                 ;
581         if (q->nr_sorted == 0)
582                 return;
583         if (printed++ < 10) {
584                 printk(KERN_ERR "%s: forced dispatching is broken "
585                        "(nr_sorted=%u), please report this\n",
586                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
587         }
588 }
589
590 void elv_insert(struct request_queue *q, struct request *rq, int where)
591 {
592         struct list_head *pos;
593         unsigned ordseq;
594         int unplug_it = 1;
595
596         trace_block_rq_insert(q, rq);
597
598         rq->q = q;
599
600         switch (where) {
601         case ELEVATOR_INSERT_FRONT:
602                 rq->cmd_flags |= REQ_SOFTBARRIER;
603
604                 list_add(&rq->queuelist, &q->queue_head);
605                 break;
606
607         case ELEVATOR_INSERT_BACK:
608                 rq->cmd_flags |= REQ_SOFTBARRIER;
609                 elv_drain_elevator(q);
610                 list_add_tail(&rq->queuelist, &q->queue_head);
611                 /*
612                  * We kick the queue here for the following reasons.
613                  * - The elevator might have returned NULL previously
614                  *   to delay requests and returned them now.  As the
615                  *   queue wasn't empty before this request, ll_rw_blk
616                  *   won't run the queue on return, resulting in hang.
617                  * - Usually, back inserted requests won't be merged
618                  *   with anything.  There's no point in delaying queue
619                  *   processing.
620                  */
621                 blk_remove_plug(q);
622                 blk_start_queueing(q);
623                 break;
624
625         case ELEVATOR_INSERT_SORT:
626                 BUG_ON(!blk_fs_request(rq) && !blk_discard_rq(rq));
627                 rq->cmd_flags |= REQ_SORTED;
628                 q->nr_sorted++;
629                 if (rq_mergeable(rq)) {
630                         elv_rqhash_add(q, rq);
631                         if (!q->last_merge)
632                                 q->last_merge = rq;
633                 }
634
635                 /*
636                  * Some ioscheds (cfq) run q->request_fn directly, so
637                  * rq cannot be accessed after calling
638                  * elevator_add_req_fn.
639                  */
640                 q->elevator->ops->elevator_add_req_fn(q, rq);
641                 break;
642
643         case ELEVATOR_INSERT_REQUEUE:
644                 /*
645                  * If ordered flush isn't in progress, we do front
646                  * insertion; otherwise, requests should be requeued
647                  * in ordseq order.
648                  */
649                 rq->cmd_flags |= REQ_SOFTBARRIER;
650
651                 /*
652                  * Most requeues happen because of a busy condition,
653                  * don't force unplug of the queue for that case.
654                  */
655                 unplug_it = 0;
656
657                 if (q->ordseq == 0) {
658                         list_add(&rq->queuelist, &q->queue_head);
659                         break;
660                 }
661
662                 ordseq = blk_ordered_req_seq(rq);
663
664                 list_for_each(pos, &q->queue_head) {
665                         struct request *pos_rq = list_entry_rq(pos);
666                         if (ordseq <= blk_ordered_req_seq(pos_rq))
667                                 break;
668                 }
669
670                 list_add_tail(&rq->queuelist, pos);
671                 break;
672
673         default:
674                 printk(KERN_ERR "%s: bad insertion point %d\n",
675                        __func__, where);
676                 BUG();
677         }
678
679         if (unplug_it && blk_queue_plugged(q)) {
680                 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
681                         - q->in_flight;
682
683                 if (nrq >= q->unplug_thresh)
684                         __generic_unplug_device(q);
685         }
686 }
687
688 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
689                        int plug)
690 {
691         if (q->ordcolor)
692                 rq->cmd_flags |= REQ_ORDERED_COLOR;
693
694         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
695                 /*
696                  * toggle ordered color
697                  */
698                 if (blk_barrier_rq(rq))
699                         q->ordcolor ^= 1;
700
701                 /*
702                  * barriers implicitly indicate back insertion
703                  */
704                 if (where == ELEVATOR_INSERT_SORT)
705                         where = ELEVATOR_INSERT_BACK;
706
707                 /*
708                  * this request is scheduling boundary, update
709                  * end_sector
710                  */
711                 if (blk_fs_request(rq) || blk_discard_rq(rq)) {
712                         q->end_sector = rq_end_sector(rq);
713                         q->boundary_rq = rq;
714                 }
715         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
716                     where == ELEVATOR_INSERT_SORT)
717                 where = ELEVATOR_INSERT_BACK;
718
719         if (plug)
720                 blk_plug_device(q);
721
722         elv_insert(q, rq, where);
723 }
724 EXPORT_SYMBOL(__elv_add_request);
725
726 void elv_add_request(struct request_queue *q, struct request *rq, int where,
727                      int plug)
728 {
729         unsigned long flags;
730
731         spin_lock_irqsave(q->queue_lock, flags);
732         __elv_add_request(q, rq, where, plug);
733         spin_unlock_irqrestore(q->queue_lock, flags);
734 }
735 EXPORT_SYMBOL(elv_add_request);
736
737 static inline struct request *__elv_next_request(struct request_queue *q)
738 {
739         struct request *rq;
740
741         while (1) {
742                 while (!list_empty(&q->queue_head)) {
743                         rq = list_entry_rq(q->queue_head.next);
744                         if (blk_do_ordered(q, &rq))
745                                 return rq;
746                 }
747
748                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
749                         return NULL;
750         }
751 }
752
753 struct request *elv_next_request(struct request_queue *q)
754 {
755         struct request *rq;
756         int ret;
757
758         while ((rq = __elv_next_request(q)) != NULL) {
759                 if (!(rq->cmd_flags & REQ_STARTED)) {
760                         /*
761                          * This is the first time the device driver
762                          * sees this request (possibly after
763                          * requeueing).  Notify IO scheduler.
764                          */
765                         if (blk_sorted_rq(rq))
766                                 elv_activate_rq(q, rq);
767
768                         /*
769                          * just mark as started even if we don't start
770                          * it, a request that has been delayed should
771                          * not be passed by new incoming requests
772                          */
773                         rq->cmd_flags |= REQ_STARTED;
774                         trace_block_rq_issue(q, rq);
775                 }
776
777                 if (!q->boundary_rq || q->boundary_rq == rq) {
778                         q->end_sector = rq_end_sector(rq);
779                         q->boundary_rq = NULL;
780                 }
781
782                 if (rq->cmd_flags & REQ_DONTPREP)
783                         break;
784
785                 if (q->dma_drain_size && rq->data_len) {
786                         /*
787                          * make sure space for the drain appears we
788                          * know we can do this because max_hw_segments
789                          * has been adjusted to be one fewer than the
790                          * device can handle
791                          */
792                         rq->nr_phys_segments++;
793                 }
794
795                 if (!q->prep_rq_fn)
796                         break;
797
798                 ret = q->prep_rq_fn(q, rq);
799                 if (ret == BLKPREP_OK) {
800                         break;
801                 } else if (ret == BLKPREP_DEFER) {
802                         /*
803                          * the request may have been (partially) prepped.
804                          * we need to keep this request in the front to
805                          * avoid resource deadlock.  REQ_STARTED will
806                          * prevent other fs requests from passing this one.
807                          */
808                         if (q->dma_drain_size && rq->data_len &&
809                             !(rq->cmd_flags & REQ_DONTPREP)) {
810                                 /*
811                                  * remove the space for the drain we added
812                                  * so that we don't add it again
813                                  */
814                                 --rq->nr_phys_segments;
815                         }
816
817                         rq = NULL;
818                         break;
819                 } else if (ret == BLKPREP_KILL) {
820                         rq->cmd_flags |= REQ_QUIET;
821                         __blk_end_request(rq, -EIO, blk_rq_bytes(rq));
822                 } else {
823                         printk(KERN_ERR "%s: bad return=%d\n", __func__, ret);
824                         break;
825                 }
826         }
827
828         return rq;
829 }
830 EXPORT_SYMBOL(elv_next_request);
831
832 void elv_dequeue_request(struct request_queue *q, struct request *rq)
833 {
834         BUG_ON(list_empty(&rq->queuelist));
835         BUG_ON(ELV_ON_HASH(rq));
836
837         list_del_init(&rq->queuelist);
838
839         /*
840          * the time frame between a request being removed from the lists
841          * and to it is freed is accounted as io that is in progress at
842          * the driver side.
843          */
844         if (blk_account_rq(rq))
845                 q->in_flight++;
846 }
847
848 int elv_queue_empty(struct request_queue *q)
849 {
850         struct elevator_queue *e = q->elevator;
851
852         if (!list_empty(&q->queue_head))
853                 return 0;
854
855         if (e->ops->elevator_queue_empty_fn)
856                 return e->ops->elevator_queue_empty_fn(q);
857
858         return 1;
859 }
860 EXPORT_SYMBOL(elv_queue_empty);
861
862 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
863 {
864         struct elevator_queue *e = q->elevator;
865
866         if (e->ops->elevator_latter_req_fn)
867                 return e->ops->elevator_latter_req_fn(q, rq);
868         return NULL;
869 }
870
871 struct request *elv_former_request(struct request_queue *q, struct request *rq)
872 {
873         struct elevator_queue *e = q->elevator;
874
875         if (e->ops->elevator_former_req_fn)
876                 return e->ops->elevator_former_req_fn(q, rq);
877         return NULL;
878 }
879
880 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
881 {
882         struct elevator_queue *e = q->elevator;
883
884         if (e->ops->elevator_set_req_fn)
885                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
886
887         rq->elevator_private = NULL;
888         return 0;
889 }
890
891 void elv_put_request(struct request_queue *q, struct request *rq)
892 {
893         struct elevator_queue *e = q->elevator;
894
895         if (e->ops->elevator_put_req_fn)
896                 e->ops->elevator_put_req_fn(rq);
897 }
898
899 int elv_may_queue(struct request_queue *q, int rw)
900 {
901         struct elevator_queue *e = q->elevator;
902
903         if (e->ops->elevator_may_queue_fn)
904                 return e->ops->elevator_may_queue_fn(q, rw);
905
906         return ELV_MQUEUE_MAY;
907 }
908
909 void elv_abort_queue(struct request_queue *q)
910 {
911         struct request *rq;
912
913         while (!list_empty(&q->queue_head)) {
914                 rq = list_entry_rq(q->queue_head.next);
915                 rq->cmd_flags |= REQ_QUIET;
916                 trace_block_rq_abort(q, rq);
917                 __blk_end_request(rq, -EIO, blk_rq_bytes(rq));
918         }
919 }
920 EXPORT_SYMBOL(elv_abort_queue);
921
922 void elv_completed_request(struct request_queue *q, struct request *rq)
923 {
924         struct elevator_queue *e = q->elevator;
925
926         /*
927          * request is released from the driver, io must be done
928          */
929         if (blk_account_rq(rq)) {
930                 q->in_flight--;
931                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
932                         e->ops->elevator_completed_req_fn(q, rq);
933         }
934
935         /*
936          * Check if the queue is waiting for fs requests to be
937          * drained for flush sequence.
938          */
939         if (unlikely(q->ordseq)) {
940                 struct request *next = NULL;
941
942                 if (!list_empty(&q->queue_head))
943                         next = list_entry_rq(q->queue_head.next);
944
945                 if (!q->in_flight &&
946                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
947                     (!next || blk_ordered_req_seq(next) > QUEUE_ORDSEQ_DRAIN)) {
948                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
949                         blk_start_queueing(q);
950                 }
951         }
952 }
953
954 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
955
956 static ssize_t
957 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
958 {
959         struct elv_fs_entry *entry = to_elv(attr);
960         struct elevator_queue *e;
961         ssize_t error;
962
963         if (!entry->show)
964                 return -EIO;
965
966         e = container_of(kobj, struct elevator_queue, kobj);
967         mutex_lock(&e->sysfs_lock);
968         error = e->ops ? entry->show(e, page) : -ENOENT;
969         mutex_unlock(&e->sysfs_lock);
970         return error;
971 }
972
973 static ssize_t
974 elv_attr_store(struct kobject *kobj, struct attribute *attr,
975                const char *page, size_t length)
976 {
977         struct elv_fs_entry *entry = to_elv(attr);
978         struct elevator_queue *e;
979         ssize_t error;
980
981         if (!entry->store)
982                 return -EIO;
983
984         e = container_of(kobj, struct elevator_queue, kobj);
985         mutex_lock(&e->sysfs_lock);
986         error = e->ops ? entry->store(e, page, length) : -ENOENT;
987         mutex_unlock(&e->sysfs_lock);
988         return error;
989 }
990
991 static struct sysfs_ops elv_sysfs_ops = {
992         .show   = elv_attr_show,
993         .store  = elv_attr_store,
994 };
995
996 static struct kobj_type elv_ktype = {
997         .sysfs_ops      = &elv_sysfs_ops,
998         .release        = elevator_release,
999 };
1000
1001 int elv_register_queue(struct request_queue *q)
1002 {
1003         struct elevator_queue *e = q->elevator;
1004         int error;
1005
1006         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
1007         if (!error) {
1008                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
1009                 if (attr) {
1010                         while (attr->attr.name) {
1011                                 if (sysfs_create_file(&e->kobj, &attr->attr))
1012                                         break;
1013                                 attr++;
1014                         }
1015                 }
1016                 kobject_uevent(&e->kobj, KOBJ_ADD);
1017         }
1018         return error;
1019 }
1020
1021 static void __elv_unregister_queue(struct elevator_queue *e)
1022 {
1023         kobject_uevent(&e->kobj, KOBJ_REMOVE);
1024         kobject_del(&e->kobj);
1025 }
1026
1027 void elv_unregister_queue(struct request_queue *q)
1028 {
1029         if (q)
1030                 __elv_unregister_queue(q->elevator);
1031 }
1032
1033 void elv_register(struct elevator_type *e)
1034 {
1035         char *def = "";
1036
1037         spin_lock(&elv_list_lock);
1038         BUG_ON(elevator_find(e->elevator_name));
1039         list_add_tail(&e->list, &elv_list);
1040         spin_unlock(&elv_list_lock);
1041
1042         if (!strcmp(e->elevator_name, chosen_elevator) ||
1043                         (!*chosen_elevator &&
1044                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
1045                                 def = " (default)";
1046
1047         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
1048                                                                 def);
1049 }
1050 EXPORT_SYMBOL_GPL(elv_register);
1051
1052 void elv_unregister(struct elevator_type *e)
1053 {
1054         struct task_struct *g, *p;
1055
1056         /*
1057          * Iterate every thread in the process to remove the io contexts.
1058          */
1059         if (e->ops.trim) {
1060                 read_lock(&tasklist_lock);
1061                 do_each_thread(g, p) {
1062                         task_lock(p);
1063                         if (p->io_context)
1064                                 e->ops.trim(p->io_context);
1065                         task_unlock(p);
1066                 } while_each_thread(g, p);
1067                 read_unlock(&tasklist_lock);
1068         }
1069
1070         spin_lock(&elv_list_lock);
1071         list_del_init(&e->list);
1072         spin_unlock(&elv_list_lock);
1073 }
1074 EXPORT_SYMBOL_GPL(elv_unregister);
1075
1076 /*
1077  * switch to new_e io scheduler. be careful not to introduce deadlocks -
1078  * we don't free the old io scheduler, before we have allocated what we
1079  * need for the new one. this way we have a chance of going back to the old
1080  * one, if the new one fails init for some reason.
1081  */
1082 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1083 {
1084         struct elevator_queue *old_elevator, *e;
1085         void *data;
1086
1087         /*
1088          * Allocate new elevator
1089          */
1090         e = elevator_alloc(q, new_e);
1091         if (!e)
1092                 return 0;
1093
1094         data = elevator_init_queue(q, e);
1095         if (!data) {
1096                 kobject_put(&e->kobj);
1097                 return 0;
1098         }
1099
1100         /*
1101          * Turn on BYPASS and drain all requests w/ elevator private data
1102          */
1103         spin_lock_irq(q->queue_lock);
1104
1105         queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
1106
1107         elv_drain_elevator(q);
1108
1109         while (q->rq.elvpriv) {
1110                 blk_start_queueing(q);
1111                 spin_unlock_irq(q->queue_lock);
1112                 msleep(10);
1113                 spin_lock_irq(q->queue_lock);
1114                 elv_drain_elevator(q);
1115         }
1116
1117         /*
1118          * Remember old elevator.
1119          */
1120         old_elevator = q->elevator;
1121
1122         /*
1123          * attach and start new elevator
1124          */
1125         elevator_attach(q, e, data);
1126
1127         spin_unlock_irq(q->queue_lock);
1128
1129         __elv_unregister_queue(old_elevator);
1130
1131         if (elv_register_queue(q))
1132                 goto fail_register;
1133
1134         /*
1135          * finally exit old elevator and turn off BYPASS.
1136          */
1137         elevator_exit(old_elevator);
1138         spin_lock_irq(q->queue_lock);
1139         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1140         spin_unlock_irq(q->queue_lock);
1141
1142         blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
1143
1144         return 1;
1145
1146 fail_register:
1147         /*
1148          * switch failed, exit the new io scheduler and reattach the old
1149          * one again (along with re-adding the sysfs dir)
1150          */
1151         elevator_exit(e);
1152         q->elevator = old_elevator;
1153         elv_register_queue(q);
1154
1155         spin_lock_irq(q->queue_lock);
1156         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1157         spin_unlock_irq(q->queue_lock);
1158
1159         return 0;
1160 }
1161
1162 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1163                           size_t count)
1164 {
1165         char elevator_name[ELV_NAME_MAX];
1166         struct elevator_type *e;
1167
1168         strlcpy(elevator_name, name, sizeof(elevator_name));
1169         strstrip(elevator_name);
1170
1171         e = elevator_get(elevator_name);
1172         if (!e) {
1173                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1174                 return -EINVAL;
1175         }
1176
1177         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1178                 elevator_put(e);
1179                 return count;
1180         }
1181
1182         if (!elevator_switch(q, e))
1183                 printk(KERN_ERR "elevator: switch to %s failed\n",
1184                                                         elevator_name);
1185         return count;
1186 }
1187
1188 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1189 {
1190         struct elevator_queue *e = q->elevator;
1191         struct elevator_type *elv = e->elevator_type;
1192         struct elevator_type *__e;
1193         int len = 0;
1194
1195         spin_lock(&elv_list_lock);
1196         list_for_each_entry(__e, &elv_list, list) {
1197                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1198                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1199                 else
1200                         len += sprintf(name+len, "%s ", __e->elevator_name);
1201         }
1202         spin_unlock(&elv_list_lock);
1203
1204         len += sprintf(len+name, "\n");
1205         return len;
1206 }
1207
1208 struct request *elv_rb_former_request(struct request_queue *q,
1209                                       struct request *rq)
1210 {
1211         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1212
1213         if (rbprev)
1214                 return rb_entry_rq(rbprev);
1215
1216         return NULL;
1217 }
1218 EXPORT_SYMBOL(elv_rb_former_request);
1219
1220 struct request *elv_rb_latter_request(struct request_queue *q,
1221                                       struct request *rq)
1222 {
1223         struct rb_node *rbnext = rb_next(&rq->rb_node);
1224
1225         if (rbnext)
1226                 return rb_entry_rq(rbnext);
1227
1228         return NULL;
1229 }
1230 EXPORT_SYMBOL(elv_rb_latter_request);