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