[PATCH] cfq-iosched: merging problem
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/hash.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15
16 /*
17  * tunables
18  */
19 static const int cfq_quantum = 4;               /* max queue in one round of service */
20 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
21 static const int cfq_back_max = 16 * 1024;      /* maximum backwards seek, in KiB */
22 static const int cfq_back_penalty = 2;          /* penalty of a backwards seek */
23
24 static const int cfq_slice_sync = HZ / 10;
25 static int cfq_slice_async = HZ / 25;
26 static const int cfq_slice_async_rq = 2;
27 static int cfq_slice_idle = HZ / 125;
28
29 #define CFQ_IDLE_GRACE          (HZ / 10)
30 #define CFQ_SLICE_SCALE         (5)
31
32 #define CFQ_KEY_ASYNC           (0)
33
34 /*
35  * for the hash of cfqq inside the cfqd
36  */
37 #define CFQ_QHASH_SHIFT         6
38 #define CFQ_QHASH_ENTRIES       (1 << CFQ_QHASH_SHIFT)
39 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
40
41 #define list_entry_cfqq(ptr)    list_entry((ptr), struct cfq_queue, cfq_list)
42
43 #define RQ_CIC(rq)              ((struct cfq_io_context*)(rq)->elevator_private)
44 #define RQ_CFQQ(rq)             ((rq)->elevator_private2)
45
46 static struct kmem_cache *cfq_pool;
47 static struct kmem_cache *cfq_ioc_pool;
48
49 static DEFINE_PER_CPU(unsigned long, ioc_count);
50 static struct completion *ioc_gone;
51
52 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
53 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
54 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
55
56 #define ASYNC                   (0)
57 #define SYNC                    (1)
58
59 #define cfq_cfqq_dispatched(cfqq)       \
60         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
61
62 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
63
64 #define cfq_cfqq_sync(cfqq)             \
65         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
66
67 #define sample_valid(samples)   ((samples) > 80)
68
69 /*
70  * Per block device queue structure
71  */
72 struct cfq_data {
73         request_queue_t *queue;
74
75         /*
76          * rr list of queues with requests and the count of them
77          */
78         struct list_head rr_list[CFQ_PRIO_LISTS];
79         struct list_head busy_rr;
80         struct list_head cur_rr;
81         struct list_head idle_rr;
82         unsigned int busy_queues;
83
84         /*
85          * cfqq lookup hash
86          */
87         struct hlist_head *cfq_hash;
88
89         int rq_in_driver;
90         int hw_tag;
91
92         /*
93          * idle window management
94          */
95         struct timer_list idle_slice_timer;
96         struct work_struct unplug_work;
97
98         struct cfq_queue *active_queue;
99         struct cfq_io_context *active_cic;
100         int cur_prio, cur_end_prio;
101         unsigned int dispatch_slice;
102
103         struct timer_list idle_class_timer;
104
105         sector_t last_sector;
106         unsigned long last_end_request;
107
108         /*
109          * tunables, see top of file
110          */
111         unsigned int cfq_quantum;
112         unsigned int cfq_fifo_expire[2];
113         unsigned int cfq_back_penalty;
114         unsigned int cfq_back_max;
115         unsigned int cfq_slice[2];
116         unsigned int cfq_slice_async_rq;
117         unsigned int cfq_slice_idle;
118
119         struct list_head cic_list;
120 };
121
122 /*
123  * Per process-grouping structure
124  */
125 struct cfq_queue {
126         /* reference count */
127         atomic_t ref;
128         /* parent cfq_data */
129         struct cfq_data *cfqd;
130         /* cfqq lookup hash */
131         struct hlist_node cfq_hash;
132         /* hash key */
133         unsigned int key;
134         /* member of the rr/busy/cur/idle cfqd list */
135         struct list_head cfq_list;
136         /* sorted list of pending requests */
137         struct rb_root sort_list;
138         /* if fifo isn't expired, next request to serve */
139         struct request *next_rq;
140         /* requests queued in sort_list */
141         int queued[2];
142         /* currently allocated requests */
143         int allocated[2];
144         /* pending metadata requests */
145         int meta_pending;
146         /* fifo list of requests in sort_list */
147         struct list_head fifo;
148
149         unsigned long slice_start;
150         unsigned long slice_end;
151         unsigned long slice_left;
152
153         /* number of requests that are on the dispatch list */
154         int on_dispatch[2];
155
156         /* io prio of this group */
157         unsigned short ioprio, org_ioprio;
158         unsigned short ioprio_class, org_ioprio_class;
159
160         /* various state flags, see below */
161         unsigned int flags;
162 };
163
164 enum cfqq_state_flags {
165         CFQ_CFQQ_FLAG_on_rr = 0,
166         CFQ_CFQQ_FLAG_wait_request,
167         CFQ_CFQQ_FLAG_must_alloc,
168         CFQ_CFQQ_FLAG_must_alloc_slice,
169         CFQ_CFQQ_FLAG_must_dispatch,
170         CFQ_CFQQ_FLAG_fifo_expire,
171         CFQ_CFQQ_FLAG_idle_window,
172         CFQ_CFQQ_FLAG_prio_changed,
173         CFQ_CFQQ_FLAG_queue_new,
174 };
175
176 #define CFQ_CFQQ_FNS(name)                                              \
177 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
178 {                                                                       \
179         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
180 }                                                                       \
181 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
182 {                                                                       \
183         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
184 }                                                                       \
185 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
186 {                                                                       \
187         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
188 }
189
190 CFQ_CFQQ_FNS(on_rr);
191 CFQ_CFQQ_FNS(wait_request);
192 CFQ_CFQQ_FNS(must_alloc);
193 CFQ_CFQQ_FNS(must_alloc_slice);
194 CFQ_CFQQ_FNS(must_dispatch);
195 CFQ_CFQQ_FNS(fifo_expire);
196 CFQ_CFQQ_FNS(idle_window);
197 CFQ_CFQQ_FNS(prio_changed);
198 CFQ_CFQQ_FNS(queue_new);
199 #undef CFQ_CFQQ_FNS
200
201 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
202 static void cfq_dispatch_insert(request_queue_t *, struct request *);
203 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
204
205 /*
206  * scheduler run of queue, if there are requests pending and no one in the
207  * driver that will restart queueing
208  */
209 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
210 {
211         if (cfqd->busy_queues)
212                 kblockd_schedule_work(&cfqd->unplug_work);
213 }
214
215 static int cfq_queue_empty(request_queue_t *q)
216 {
217         struct cfq_data *cfqd = q->elevator->elevator_data;
218
219         return !cfqd->busy_queues;
220 }
221
222 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw, int is_sync)
223 {
224         /*
225          * Use the per-process queue, for read requests and syncronous writes
226          */
227         if (!(rw & REQ_RW) || is_sync)
228                 return task->pid;
229
230         return CFQ_KEY_ASYNC;
231 }
232
233 /*
234  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
235  * We choose the request that is closest to the head right now. Distance
236  * behind the head is penalized and only allowed to a certain extent.
237  */
238 static struct request *
239 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
240 {
241         sector_t last, s1, s2, d1 = 0, d2 = 0;
242         unsigned long back_max;
243 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
244 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
245         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
246
247         if (rq1 == NULL || rq1 == rq2)
248                 return rq2;
249         if (rq2 == NULL)
250                 return rq1;
251
252         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
253                 return rq1;
254         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
255                 return rq2;
256         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
257                 return rq1;
258         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
259                 return rq2;
260
261         s1 = rq1->sector;
262         s2 = rq2->sector;
263
264         last = cfqd->last_sector;
265
266         /*
267          * by definition, 1KiB is 2 sectors
268          */
269         back_max = cfqd->cfq_back_max * 2;
270
271         /*
272          * Strict one way elevator _except_ in the case where we allow
273          * short backward seeks which are biased as twice the cost of a
274          * similar forward seek.
275          */
276         if (s1 >= last)
277                 d1 = s1 - last;
278         else if (s1 + back_max >= last)
279                 d1 = (last - s1) * cfqd->cfq_back_penalty;
280         else
281                 wrap |= CFQ_RQ1_WRAP;
282
283         if (s2 >= last)
284                 d2 = s2 - last;
285         else if (s2 + back_max >= last)
286                 d2 = (last - s2) * cfqd->cfq_back_penalty;
287         else
288                 wrap |= CFQ_RQ2_WRAP;
289
290         /* Found required data */
291
292         /*
293          * By doing switch() on the bit mask "wrap" we avoid having to
294          * check two variables for all permutations: --> faster!
295          */
296         switch (wrap) {
297         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
298                 if (d1 < d2)
299                         return rq1;
300                 else if (d2 < d1)
301                         return rq2;
302                 else {
303                         if (s1 >= s2)
304                                 return rq1;
305                         else
306                                 return rq2;
307                 }
308
309         case CFQ_RQ2_WRAP:
310                 return rq1;
311         case CFQ_RQ1_WRAP:
312                 return rq2;
313         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
314         default:
315                 /*
316                  * Since both rqs are wrapped,
317                  * start with the one that's further behind head
318                  * (--> only *one* back seek required),
319                  * since back seek takes more time than forward.
320                  */
321                 if (s1 <= s2)
322                         return rq1;
323                 else
324                         return rq2;
325         }
326 }
327
328 /*
329  * would be nice to take fifo expire time into account as well
330  */
331 static struct request *
332 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
333                   struct request *last)
334 {
335         struct rb_node *rbnext = rb_next(&last->rb_node);
336         struct rb_node *rbprev = rb_prev(&last->rb_node);
337         struct request *next = NULL, *prev = NULL;
338
339         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
340
341         if (rbprev)
342                 prev = rb_entry_rq(rbprev);
343
344         if (rbnext)
345                 next = rb_entry_rq(rbnext);
346         else {
347                 rbnext = rb_first(&cfqq->sort_list);
348                 if (rbnext && rbnext != &last->rb_node)
349                         next = rb_entry_rq(rbnext);
350         }
351
352         return cfq_choose_req(cfqd, next, prev);
353 }
354
355 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
356 {
357         struct cfq_data *cfqd = cfqq->cfqd;
358         struct list_head *list;
359
360         BUG_ON(!cfq_cfqq_on_rr(cfqq));
361
362         list_del(&cfqq->cfq_list);
363
364         if (cfq_class_rt(cfqq))
365                 list = &cfqd->cur_rr;
366         else if (cfq_class_idle(cfqq))
367                 list = &cfqd->idle_rr;
368         else {
369                 /*
370                  * if cfqq has requests in flight, don't allow it to be
371                  * found in cfq_set_active_queue before it has finished them.
372                  * this is done to increase fairness between a process that
373                  * has lots of io pending vs one that only generates one
374                  * sporadically or synchronously
375                  */
376                 if (cfq_cfqq_dispatched(cfqq))
377                         list = &cfqd->busy_rr;
378                 else
379                         list = &cfqd->rr_list[cfqq->ioprio];
380         }
381
382         /*
383          * If this queue was preempted or is new (never been serviced), let
384          * it be added first for fairness but beind other new queues.
385          * Otherwise, just add to the back  of the list.
386          */
387         if (preempted || cfq_cfqq_queue_new(cfqq)) {
388                 struct list_head *n = list;
389                 struct cfq_queue *__cfqq;
390
391                 while (n->next != list) {
392                         __cfqq = list_entry_cfqq(n->next);
393                         if (!cfq_cfqq_queue_new(__cfqq))
394                                 break;
395
396                         n = n->next;
397                 }
398
399                 list = n;
400         }
401
402         list_add_tail(&cfqq->cfq_list, list);
403 }
404
405 /*
406  * add to busy list of queues for service, trying to be fair in ordering
407  * the pending list according to last request service
408  */
409 static inline void
410 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
411 {
412         BUG_ON(cfq_cfqq_on_rr(cfqq));
413         cfq_mark_cfqq_on_rr(cfqq);
414         cfqd->busy_queues++;
415
416         cfq_resort_rr_list(cfqq, 0);
417 }
418
419 static inline void
420 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
421 {
422         BUG_ON(!cfq_cfqq_on_rr(cfqq));
423         cfq_clear_cfqq_on_rr(cfqq);
424         list_del_init(&cfqq->cfq_list);
425
426         BUG_ON(!cfqd->busy_queues);
427         cfqd->busy_queues--;
428 }
429
430 /*
431  * rb tree support functions
432  */
433 static inline void cfq_del_rq_rb(struct request *rq)
434 {
435         struct cfq_queue *cfqq = RQ_CFQQ(rq);
436         struct cfq_data *cfqd = cfqq->cfqd;
437         const int sync = rq_is_sync(rq);
438
439         BUG_ON(!cfqq->queued[sync]);
440         cfqq->queued[sync]--;
441
442         elv_rb_del(&cfqq->sort_list, rq);
443
444         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
445                 cfq_del_cfqq_rr(cfqd, cfqq);
446 }
447
448 static void cfq_add_rq_rb(struct request *rq)
449 {
450         struct cfq_queue *cfqq = RQ_CFQQ(rq);
451         struct cfq_data *cfqd = cfqq->cfqd;
452         struct request *__alias;
453
454         cfqq->queued[rq_is_sync(rq)]++;
455
456         /*
457          * looks a little odd, but the first insert might return an alias.
458          * if that happens, put the alias on the dispatch list
459          */
460         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
461                 cfq_dispatch_insert(cfqd->queue, __alias);
462
463         if (!cfq_cfqq_on_rr(cfqq))
464                 cfq_add_cfqq_rr(cfqd, cfqq);
465 }
466
467 static inline void
468 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
469 {
470         elv_rb_del(&cfqq->sort_list, rq);
471         cfqq->queued[rq_is_sync(rq)]--;
472         cfq_add_rq_rb(rq);
473 }
474
475 static struct request *
476 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
477 {
478         struct task_struct *tsk = current;
479         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio));
480         struct cfq_queue *cfqq;
481
482         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
483         if (cfqq) {
484                 sector_t sector = bio->bi_sector + bio_sectors(bio);
485
486                 return elv_rb_find(&cfqq->sort_list, sector);
487         }
488
489         return NULL;
490 }
491
492 static void cfq_activate_request(request_queue_t *q, struct request *rq)
493 {
494         struct cfq_data *cfqd = q->elevator->elevator_data;
495
496         cfqd->rq_in_driver++;
497
498         /*
499          * If the depth is larger 1, it really could be queueing. But lets
500          * make the mark a little higher - idling could still be good for
501          * low queueing, and a low queueing number could also just indicate
502          * a SCSI mid layer like behaviour where limit+1 is often seen.
503          */
504         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
505                 cfqd->hw_tag = 1;
506 }
507
508 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
509 {
510         struct cfq_data *cfqd = q->elevator->elevator_data;
511
512         WARN_ON(!cfqd->rq_in_driver);
513         cfqd->rq_in_driver--;
514 }
515
516 static void cfq_remove_request(struct request *rq)
517 {
518         struct cfq_queue *cfqq = RQ_CFQQ(rq);
519
520         if (cfqq->next_rq == rq)
521                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
522
523         list_del_init(&rq->queuelist);
524         cfq_del_rq_rb(rq);
525
526         if (rq_is_meta(rq)) {
527                 WARN_ON(!cfqq->meta_pending);
528                 cfqq->meta_pending--;
529         }
530 }
531
532 static int
533 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
534 {
535         struct cfq_data *cfqd = q->elevator->elevator_data;
536         struct request *__rq;
537
538         __rq = cfq_find_rq_fmerge(cfqd, bio);
539         if (__rq && elv_rq_merge_ok(__rq, bio)) {
540                 *req = __rq;
541                 return ELEVATOR_FRONT_MERGE;
542         }
543
544         return ELEVATOR_NO_MERGE;
545 }
546
547 static void cfq_merged_request(request_queue_t *q, struct request *req,
548                                int type)
549 {
550         if (type == ELEVATOR_FRONT_MERGE) {
551                 struct cfq_queue *cfqq = RQ_CFQQ(req);
552
553                 cfq_reposition_rq_rb(cfqq, req);
554         }
555 }
556
557 static void
558 cfq_merged_requests(request_queue_t *q, struct request *rq,
559                     struct request *next)
560 {
561         /*
562          * reposition in fifo if next is older than rq
563          */
564         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
565             time_before(next->start_time, rq->start_time))
566                 list_move(&rq->queuelist, &next->queuelist);
567
568         cfq_remove_request(next);
569 }
570
571 static int cfq_allow_merge(request_queue_t *q, struct request *rq,
572                            struct bio *bio)
573 {
574         struct cfq_data *cfqd = q->elevator->elevator_data;
575         const int rw = bio_data_dir(bio);
576         struct cfq_queue *cfqq;
577         pid_t key;
578
579         /*
580          * Disallow merge of a sync bio into an async request.
581          */
582         if ((bio_data_dir(bio) == READ || bio_sync(bio)) && !rq_is_sync(rq))
583                 return 0;
584
585         /*
586          * Lookup the cfqq that this bio will be queued with. Allow
587          * merge only if rq is queued there.
588          */
589         key = cfq_queue_pid(current, rw, bio_sync(bio));
590         cfqq = cfq_find_cfq_hash(cfqd, key, current->ioprio);
591
592         if (cfqq == RQ_CFQQ(rq))
593                 return 1;
594
595         return 0;
596 }
597
598 static inline void
599 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
600 {
601         if (cfqq) {
602                 /*
603                  * stop potential idle class queues waiting service
604                  */
605                 del_timer(&cfqd->idle_class_timer);
606
607                 cfqq->slice_start = jiffies;
608                 cfqq->slice_end = 0;
609                 cfqq->slice_left = 0;
610                 cfq_clear_cfqq_must_alloc_slice(cfqq);
611                 cfq_clear_cfqq_fifo_expire(cfqq);
612         }
613
614         cfqd->active_queue = cfqq;
615 }
616
617 /*
618  * current cfqq expired its slice (or was too idle), select new one
619  */
620 static void
621 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
622                     int preempted)
623 {
624         unsigned long now = jiffies;
625
626         if (cfq_cfqq_wait_request(cfqq))
627                 del_timer(&cfqd->idle_slice_timer);
628
629         if (!preempted && !cfq_cfqq_dispatched(cfqq))
630                 cfq_schedule_dispatch(cfqd);
631
632         cfq_clear_cfqq_must_dispatch(cfqq);
633         cfq_clear_cfqq_wait_request(cfqq);
634         cfq_clear_cfqq_queue_new(cfqq);
635
636         /*
637          * store what was left of this slice, if the queue idled out
638          * or was preempted
639          */
640         if (time_after(cfqq->slice_end, now))
641                 cfqq->slice_left = cfqq->slice_end - now;
642         else
643                 cfqq->slice_left = 0;
644
645         if (cfq_cfqq_on_rr(cfqq))
646                 cfq_resort_rr_list(cfqq, preempted);
647
648         if (cfqq == cfqd->active_queue)
649                 cfqd->active_queue = NULL;
650
651         if (cfqd->active_cic) {
652                 put_io_context(cfqd->active_cic->ioc);
653                 cfqd->active_cic = NULL;
654         }
655
656         cfqd->dispatch_slice = 0;
657 }
658
659 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
660 {
661         struct cfq_queue *cfqq = cfqd->active_queue;
662
663         if (cfqq)
664                 __cfq_slice_expired(cfqd, cfqq, preempted);
665 }
666
667 /*
668  * 0
669  * 0,1
670  * 0,1,2
671  * 0,1,2,3
672  * 0,1,2,3,4
673  * 0,1,2,3,4,5
674  * 0,1,2,3,4,5,6
675  * 0,1,2,3,4,5,6,7
676  */
677 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
678 {
679         int prio, wrap;
680
681         prio = -1;
682         wrap = 0;
683         do {
684                 int p;
685
686                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
687                         if (!list_empty(&cfqd->rr_list[p])) {
688                                 prio = p;
689                                 break;
690                         }
691                 }
692
693                 if (prio != -1)
694                         break;
695                 cfqd->cur_prio = 0;
696                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
697                         cfqd->cur_end_prio = 0;
698                         if (wrap)
699                                 break;
700                         wrap = 1;
701                 }
702         } while (1);
703
704         if (unlikely(prio == -1))
705                 return -1;
706
707         BUG_ON(prio >= CFQ_PRIO_LISTS);
708
709         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
710
711         cfqd->cur_prio = prio + 1;
712         if (cfqd->cur_prio > cfqd->cur_end_prio) {
713                 cfqd->cur_end_prio = cfqd->cur_prio;
714                 cfqd->cur_prio = 0;
715         }
716         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
717                 cfqd->cur_prio = 0;
718                 cfqd->cur_end_prio = 0;
719         }
720
721         return prio;
722 }
723
724 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
725 {
726         struct cfq_queue *cfqq = NULL;
727
728         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
729                 /*
730                  * if current list is non-empty, grab first entry. if it is
731                  * empty, get next prio level and grab first entry then if any
732                  * are spliced
733                  */
734                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
735         } else if (!list_empty(&cfqd->busy_rr)) {
736                 /*
737                  * If no new queues are available, check if the busy list has
738                  * some before falling back to idle io.
739                  */
740                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
741         } else if (!list_empty(&cfqd->idle_rr)) {
742                 /*
743                  * if we have idle queues and no rt or be queues had pending
744                  * requests, either allow immediate service if the grace period
745                  * has passed or arm the idle grace timer
746                  */
747                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
748
749                 if (time_after_eq(jiffies, end))
750                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
751                 else
752                         mod_timer(&cfqd->idle_class_timer, end);
753         }
754
755         __cfq_set_active_queue(cfqd, cfqq);
756         return cfqq;
757 }
758
759 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
760
761 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
762
763 {
764         struct cfq_io_context *cic;
765         unsigned long sl;
766
767         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
768         WARN_ON(cfqq != cfqd->active_queue);
769
770         /*
771          * idle is disabled, either manually or by past process history
772          */
773         if (!cfqd->cfq_slice_idle)
774                 return 0;
775         if (!cfq_cfqq_idle_window(cfqq))
776                 return 0;
777         /*
778          * task has exited, don't wait
779          */
780         cic = cfqd->active_cic;
781         if (!cic || !cic->ioc->task)
782                 return 0;
783
784         cfq_mark_cfqq_must_dispatch(cfqq);
785         cfq_mark_cfqq_wait_request(cfqq);
786
787         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
788
789         /*
790          * we don't want to idle for seeks, but we do want to allow
791          * fair distribution of slice time for a process doing back-to-back
792          * seeks. so allow a little bit of time for him to submit a new rq
793          */
794         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
795                 sl = min(sl, msecs_to_jiffies(2));
796
797         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
798         return 1;
799 }
800
801 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
802 {
803         struct cfq_data *cfqd = q->elevator->elevator_data;
804         struct cfq_queue *cfqq = RQ_CFQQ(rq);
805
806         cfq_remove_request(rq);
807         cfqq->on_dispatch[rq_is_sync(rq)]++;
808         elv_dispatch_sort(q, rq);
809
810         rq = list_entry(q->queue_head.prev, struct request, queuelist);
811         cfqd->last_sector = rq->sector + rq->nr_sectors;
812 }
813
814 /*
815  * return expired entry, or NULL to just start from scratch in rbtree
816  */
817 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
818 {
819         struct cfq_data *cfqd = cfqq->cfqd;
820         struct request *rq;
821         int fifo;
822
823         if (cfq_cfqq_fifo_expire(cfqq))
824                 return NULL;
825         if (list_empty(&cfqq->fifo))
826                 return NULL;
827
828         fifo = cfq_cfqq_class_sync(cfqq);
829         rq = rq_entry_fifo(cfqq->fifo.next);
830
831         if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
832                 cfq_mark_cfqq_fifo_expire(cfqq);
833                 return rq;
834         }
835
836         return NULL;
837 }
838
839 /*
840  * Scale schedule slice based on io priority. Use the sync time slice only
841  * if a queue is marked sync and has sync io queued. A sync queue with async
842  * io only, should not get full sync slice length.
843  */
844 static inline int
845 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
846 {
847         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
848
849         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
850
851         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
852 }
853
854 static inline void
855 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
856 {
857         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
858 }
859
860 static inline int
861 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
862 {
863         const int base_rq = cfqd->cfq_slice_async_rq;
864
865         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
866
867         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
868 }
869
870 /*
871  * get next queue for service
872  */
873 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
874 {
875         unsigned long now = jiffies;
876         struct cfq_queue *cfqq;
877
878         cfqq = cfqd->active_queue;
879         if (!cfqq)
880                 goto new_queue;
881
882         /*
883          * slice has expired
884          */
885         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
886                 goto expire;
887
888         /*
889          * if queue has requests, dispatch one. if not, check if
890          * enough slice is left to wait for one
891          */
892         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
893                 goto keep_queue;
894         else if (cfq_cfqq_dispatched(cfqq)) {
895                 cfqq = NULL;
896                 goto keep_queue;
897         } else if (cfq_cfqq_class_sync(cfqq)) {
898                 if (cfq_arm_slice_timer(cfqd, cfqq))
899                         return NULL;
900         }
901
902 expire:
903         cfq_slice_expired(cfqd, 0);
904 new_queue:
905         cfqq = cfq_set_active_queue(cfqd);
906 keep_queue:
907         return cfqq;
908 }
909
910 static int
911 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
912                         int max_dispatch)
913 {
914         int dispatched = 0;
915
916         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
917
918         do {
919                 struct request *rq;
920
921                 /*
922                  * follow expired path, else get first next available
923                  */
924                 if ((rq = cfq_check_fifo(cfqq)) == NULL)
925                         rq = cfqq->next_rq;
926
927                 /*
928                  * finally, insert request into driver dispatch list
929                  */
930                 cfq_dispatch_insert(cfqd->queue, rq);
931
932                 cfqd->dispatch_slice++;
933                 dispatched++;
934
935                 if (!cfqd->active_cic) {
936                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
937                         cfqd->active_cic = RQ_CIC(rq);
938                 }
939
940                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
941                         break;
942
943         } while (dispatched < max_dispatch);
944
945         /*
946          * if slice end isn't set yet, set it.
947          */
948         if (!cfqq->slice_end)
949                 cfq_set_prio_slice(cfqd, cfqq);
950
951         /*
952          * expire an async queue immediately if it has used up its slice. idle
953          * queue always expire after 1 dispatch round.
954          */
955         if ((!cfq_cfqq_sync(cfqq) &&
956             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
957             cfq_class_idle(cfqq) ||
958             !cfq_cfqq_idle_window(cfqq))
959                 cfq_slice_expired(cfqd, 0);
960
961         return dispatched;
962 }
963
964 static int
965 cfq_forced_dispatch_cfqqs(struct list_head *list)
966 {
967         struct cfq_queue *cfqq, *next;
968         int dispatched;
969
970         dispatched = 0;
971         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
972                 while (cfqq->next_rq) {
973                         cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
974                         dispatched++;
975                 }
976                 BUG_ON(!list_empty(&cfqq->fifo));
977         }
978
979         return dispatched;
980 }
981
982 static int
983 cfq_forced_dispatch(struct cfq_data *cfqd)
984 {
985         int i, dispatched = 0;
986
987         for (i = 0; i < CFQ_PRIO_LISTS; i++)
988                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
989
990         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
991         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
992         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
993
994         cfq_slice_expired(cfqd, 0);
995
996         BUG_ON(cfqd->busy_queues);
997
998         return dispatched;
999 }
1000
1001 static int
1002 cfq_dispatch_requests(request_queue_t *q, int force)
1003 {
1004         struct cfq_data *cfqd = q->elevator->elevator_data;
1005         struct cfq_queue *cfqq, *prev_cfqq;
1006         int dispatched;
1007
1008         if (!cfqd->busy_queues)
1009                 return 0;
1010
1011         if (unlikely(force))
1012                 return cfq_forced_dispatch(cfqd);
1013
1014         dispatched = 0;
1015         prev_cfqq = NULL;
1016         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1017                 int max_dispatch;
1018
1019                 /*
1020                  * Don't repeat dispatch from the previous queue.
1021                  */
1022                 if (prev_cfqq == cfqq)
1023                         break;
1024
1025                 cfq_clear_cfqq_must_dispatch(cfqq);
1026                 cfq_clear_cfqq_wait_request(cfqq);
1027                 del_timer(&cfqd->idle_slice_timer);
1028
1029                 max_dispatch = cfqd->cfq_quantum;
1030                 if (cfq_class_idle(cfqq))
1031                         max_dispatch = 1;
1032
1033                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1034
1035                 /*
1036                  * If the dispatch cfqq has idling enabled and is still
1037                  * the active queue, break out.
1038                  */
1039                 if (cfq_cfqq_idle_window(cfqq) && cfqd->active_queue)
1040                         break;
1041
1042                 prev_cfqq = cfqq;
1043         }
1044
1045         return dispatched;
1046 }
1047
1048 /*
1049  * task holds one reference to the queue, dropped when task exits. each rq
1050  * in-flight on this queue also holds a reference, dropped when rq is freed.
1051  *
1052  * queue lock must be held here.
1053  */
1054 static void cfq_put_queue(struct cfq_queue *cfqq)
1055 {
1056         struct cfq_data *cfqd = cfqq->cfqd;
1057
1058         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1059
1060         if (!atomic_dec_and_test(&cfqq->ref))
1061                 return;
1062
1063         BUG_ON(rb_first(&cfqq->sort_list));
1064         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1065         BUG_ON(cfq_cfqq_on_rr(cfqq));
1066
1067         if (unlikely(cfqd->active_queue == cfqq))
1068                 __cfq_slice_expired(cfqd, cfqq, 0);
1069
1070         /*
1071          * it's on the empty list and still hashed
1072          */
1073         list_del(&cfqq->cfq_list);
1074         hlist_del(&cfqq->cfq_hash);
1075         kmem_cache_free(cfq_pool, cfqq);
1076 }
1077
1078 static struct cfq_queue *
1079 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1080                     const int hashval)
1081 {
1082         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1083         struct hlist_node *entry;
1084         struct cfq_queue *__cfqq;
1085
1086         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1087                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1088
1089                 if (__cfqq->key == key && (__p == prio || !prio))
1090                         return __cfqq;
1091         }
1092
1093         return NULL;
1094 }
1095
1096 static struct cfq_queue *
1097 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1098 {
1099         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1100 }
1101
1102 static void cfq_free_io_context(struct io_context *ioc)
1103 {
1104         struct cfq_io_context *__cic;
1105         struct rb_node *n;
1106         int freed = 0;
1107
1108         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1109                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1110                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1111                 kmem_cache_free(cfq_ioc_pool, __cic);
1112                 freed++;
1113         }
1114
1115         elv_ioc_count_mod(ioc_count, -freed);
1116
1117         if (ioc_gone && !elv_ioc_count_read(ioc_count))
1118                 complete(ioc_gone);
1119 }
1120
1121 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1122 {
1123         if (unlikely(cfqq == cfqd->active_queue))
1124                 __cfq_slice_expired(cfqd, cfqq, 0);
1125
1126         cfq_put_queue(cfqq);
1127 }
1128
1129 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1130                                          struct cfq_io_context *cic)
1131 {
1132         list_del_init(&cic->queue_list);
1133         smp_wmb();
1134         cic->key = NULL;
1135
1136         if (cic->cfqq[ASYNC]) {
1137                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1138                 cic->cfqq[ASYNC] = NULL;
1139         }
1140
1141         if (cic->cfqq[SYNC]) {
1142                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1143                 cic->cfqq[SYNC] = NULL;
1144         }
1145 }
1146
1147
1148 /*
1149  * Called with interrupts disabled
1150  */
1151 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1152 {
1153         struct cfq_data *cfqd = cic->key;
1154
1155         if (cfqd) {
1156                 request_queue_t *q = cfqd->queue;
1157
1158                 spin_lock_irq(q->queue_lock);
1159                 __cfq_exit_single_io_context(cfqd, cic);
1160                 spin_unlock_irq(q->queue_lock);
1161         }
1162 }
1163
1164 static void cfq_exit_io_context(struct io_context *ioc)
1165 {
1166         struct cfq_io_context *__cic;
1167         struct rb_node *n;
1168
1169         /*
1170          * put the reference this task is holding to the various queues
1171          */
1172
1173         n = rb_first(&ioc->cic_root);
1174         while (n != NULL) {
1175                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1176
1177                 cfq_exit_single_io_context(__cic);
1178                 n = rb_next(n);
1179         }
1180 }
1181
1182 static struct cfq_io_context *
1183 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1184 {
1185         struct cfq_io_context *cic;
1186
1187         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
1188         if (cic) {
1189                 memset(cic, 0, sizeof(*cic));
1190                 cic->last_end_request = jiffies;
1191                 INIT_LIST_HEAD(&cic->queue_list);
1192                 cic->dtor = cfq_free_io_context;
1193                 cic->exit = cfq_exit_io_context;
1194                 elv_ioc_count_inc(ioc_count);
1195         }
1196
1197         return cic;
1198 }
1199
1200 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1201 {
1202         struct task_struct *tsk = current;
1203         int ioprio_class;
1204
1205         if (!cfq_cfqq_prio_changed(cfqq))
1206                 return;
1207
1208         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1209         switch (ioprio_class) {
1210                 default:
1211                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1212                 case IOPRIO_CLASS_NONE:
1213                         /*
1214                          * no prio set, place us in the middle of the BE classes
1215                          */
1216                         cfqq->ioprio = task_nice_ioprio(tsk);
1217                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1218                         break;
1219                 case IOPRIO_CLASS_RT:
1220                         cfqq->ioprio = task_ioprio(tsk);
1221                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1222                         break;
1223                 case IOPRIO_CLASS_BE:
1224                         cfqq->ioprio = task_ioprio(tsk);
1225                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1226                         break;
1227                 case IOPRIO_CLASS_IDLE:
1228                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1229                         cfqq->ioprio = 7;
1230                         cfq_clear_cfqq_idle_window(cfqq);
1231                         break;
1232         }
1233
1234         /*
1235          * keep track of original prio settings in case we have to temporarily
1236          * elevate the priority of this queue
1237          */
1238         cfqq->org_ioprio = cfqq->ioprio;
1239         cfqq->org_ioprio_class = cfqq->ioprio_class;
1240
1241         if (cfq_cfqq_on_rr(cfqq))
1242                 cfq_resort_rr_list(cfqq, 0);
1243
1244         cfq_clear_cfqq_prio_changed(cfqq);
1245 }
1246
1247 static inline void changed_ioprio(struct cfq_io_context *cic)
1248 {
1249         struct cfq_data *cfqd = cic->key;
1250         struct cfq_queue *cfqq;
1251         unsigned long flags;
1252
1253         if (unlikely(!cfqd))
1254                 return;
1255
1256         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1257
1258         cfqq = cic->cfqq[ASYNC];
1259         if (cfqq) {
1260                 struct cfq_queue *new_cfqq;
1261                 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1262                                          GFP_ATOMIC);
1263                 if (new_cfqq) {
1264                         cic->cfqq[ASYNC] = new_cfqq;
1265                         cfq_put_queue(cfqq);
1266                 }
1267         }
1268
1269         cfqq = cic->cfqq[SYNC];
1270         if (cfqq)
1271                 cfq_mark_cfqq_prio_changed(cfqq);
1272
1273         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1274 }
1275
1276 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1277 {
1278         struct cfq_io_context *cic;
1279         struct rb_node *n;
1280
1281         ioc->ioprio_changed = 0;
1282
1283         n = rb_first(&ioc->cic_root);
1284         while (n != NULL) {
1285                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1286
1287                 changed_ioprio(cic);
1288                 n = rb_next(n);
1289         }
1290 }
1291
1292 static struct cfq_queue *
1293 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1294               gfp_t gfp_mask)
1295 {
1296         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1297         struct cfq_queue *cfqq, *new_cfqq = NULL;
1298         unsigned short ioprio;
1299
1300 retry:
1301         ioprio = tsk->ioprio;
1302         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1303
1304         if (!cfqq) {
1305                 if (new_cfqq) {
1306                         cfqq = new_cfqq;
1307                         new_cfqq = NULL;
1308                 } else if (gfp_mask & __GFP_WAIT) {
1309                         /*
1310                          * Inform the allocator of the fact that we will
1311                          * just repeat this allocation if it fails, to allow
1312                          * the allocator to do whatever it needs to attempt to
1313                          * free memory.
1314                          */
1315                         spin_unlock_irq(cfqd->queue->queue_lock);
1316                         new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
1317                         spin_lock_irq(cfqd->queue->queue_lock);
1318                         goto retry;
1319                 } else {
1320                         cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
1321                         if (!cfqq)
1322                                 goto out;
1323                 }
1324
1325                 memset(cfqq, 0, sizeof(*cfqq));
1326
1327                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1328                 INIT_LIST_HEAD(&cfqq->cfq_list);
1329                 INIT_LIST_HEAD(&cfqq->fifo);
1330
1331                 cfqq->key = key;
1332                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1333                 atomic_set(&cfqq->ref, 0);
1334                 cfqq->cfqd = cfqd;
1335                 /*
1336                  * set ->slice_left to allow preemption for a new process
1337                  */
1338                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1339                 cfq_mark_cfqq_idle_window(cfqq);
1340                 cfq_mark_cfqq_prio_changed(cfqq);
1341                 cfq_mark_cfqq_queue_new(cfqq);
1342                 cfq_init_prio_data(cfqq);
1343         }
1344
1345         if (new_cfqq)
1346                 kmem_cache_free(cfq_pool, new_cfqq);
1347
1348         atomic_inc(&cfqq->ref);
1349 out:
1350         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1351         return cfqq;
1352 }
1353
1354 static void
1355 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1356 {
1357         WARN_ON(!list_empty(&cic->queue_list));
1358         rb_erase(&cic->rb_node, &ioc->cic_root);
1359         kmem_cache_free(cfq_ioc_pool, cic);
1360         elv_ioc_count_dec(ioc_count);
1361 }
1362
1363 static struct cfq_io_context *
1364 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1365 {
1366         struct rb_node *n;
1367         struct cfq_io_context *cic;
1368         void *k, *key = cfqd;
1369
1370 restart:
1371         n = ioc->cic_root.rb_node;
1372         while (n) {
1373                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1374                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1375                 k = cic->key;
1376                 if (unlikely(!k)) {
1377                         cfq_drop_dead_cic(ioc, cic);
1378                         goto restart;
1379                 }
1380
1381                 if (key < k)
1382                         n = n->rb_left;
1383                 else if (key > k)
1384                         n = n->rb_right;
1385                 else
1386                         return cic;
1387         }
1388
1389         return NULL;
1390 }
1391
1392 static inline void
1393 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1394              struct cfq_io_context *cic)
1395 {
1396         struct rb_node **p;
1397         struct rb_node *parent;
1398         struct cfq_io_context *__cic;
1399         unsigned long flags;
1400         void *k;
1401
1402         cic->ioc = ioc;
1403         cic->key = cfqd;
1404
1405 restart:
1406         parent = NULL;
1407         p = &ioc->cic_root.rb_node;
1408         while (*p) {
1409                 parent = *p;
1410                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1411                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1412                 k = __cic->key;
1413                 if (unlikely(!k)) {
1414                         cfq_drop_dead_cic(ioc, __cic);
1415                         goto restart;
1416                 }
1417
1418                 if (cic->key < k)
1419                         p = &(*p)->rb_left;
1420                 else if (cic->key > k)
1421                         p = &(*p)->rb_right;
1422                 else
1423                         BUG();
1424         }
1425
1426         rb_link_node(&cic->rb_node, parent, p);
1427         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1428
1429         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1430         list_add(&cic->queue_list, &cfqd->cic_list);
1431         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1432 }
1433
1434 /*
1435  * Setup general io context and cfq io context. There can be several cfq
1436  * io contexts per general io context, if this process is doing io to more
1437  * than one device managed by cfq.
1438  */
1439 static struct cfq_io_context *
1440 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1441 {
1442         struct io_context *ioc = NULL;
1443         struct cfq_io_context *cic;
1444
1445         might_sleep_if(gfp_mask & __GFP_WAIT);
1446
1447         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1448         if (!ioc)
1449                 return NULL;
1450
1451         cic = cfq_cic_rb_lookup(cfqd, ioc);
1452         if (cic)
1453                 goto out;
1454
1455         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1456         if (cic == NULL)
1457                 goto err;
1458
1459         cfq_cic_link(cfqd, ioc, cic);
1460 out:
1461         smp_read_barrier_depends();
1462         if (unlikely(ioc->ioprio_changed))
1463                 cfq_ioc_set_ioprio(ioc);
1464
1465         return cic;
1466 err:
1467         put_io_context(ioc);
1468         return NULL;
1469 }
1470
1471 static void
1472 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1473 {
1474         unsigned long elapsed, ttime;
1475
1476         /*
1477          * if this context already has stuff queued, thinktime is from
1478          * last queue not last end
1479          */
1480 #if 0
1481         if (time_after(cic->last_end_request, cic->last_queue))
1482                 elapsed = jiffies - cic->last_end_request;
1483         else
1484                 elapsed = jiffies - cic->last_queue;
1485 #else
1486                 elapsed = jiffies - cic->last_end_request;
1487 #endif
1488
1489         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1490
1491         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1492         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1493         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1494 }
1495
1496 static void
1497 cfq_update_io_seektime(struct cfq_io_context *cic, struct request *rq)
1498 {
1499         sector_t sdist;
1500         u64 total;
1501
1502         if (cic->last_request_pos < rq->sector)
1503                 sdist = rq->sector - cic->last_request_pos;
1504         else
1505                 sdist = cic->last_request_pos - rq->sector;
1506
1507         /*
1508          * Don't allow the seek distance to get too large from the
1509          * odd fragment, pagein, etc
1510          */
1511         if (cic->seek_samples <= 60) /* second&third seek */
1512                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1513         else
1514                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1515
1516         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1517         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1518         total = cic->seek_total + (cic->seek_samples/2);
1519         do_div(total, cic->seek_samples);
1520         cic->seek_mean = (sector_t)total;
1521 }
1522
1523 /*
1524  * Disable idle window if the process thinks too long or seeks so much that
1525  * it doesn't matter
1526  */
1527 static void
1528 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1529                        struct cfq_io_context *cic)
1530 {
1531         int enable_idle = cfq_cfqq_idle_window(cfqq);
1532
1533         if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1534             (cfqd->hw_tag && CIC_SEEKY(cic)))
1535                 enable_idle = 0;
1536         else if (sample_valid(cic->ttime_samples)) {
1537                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1538                         enable_idle = 0;
1539                 else
1540                         enable_idle = 1;
1541         }
1542
1543         if (enable_idle)
1544                 cfq_mark_cfqq_idle_window(cfqq);
1545         else
1546                 cfq_clear_cfqq_idle_window(cfqq);
1547 }
1548
1549
1550 /*
1551  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1552  * no or if we aren't sure, a 1 will cause a preempt.
1553  */
1554 static int
1555 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1556                    struct request *rq)
1557 {
1558         struct cfq_queue *cfqq = cfqd->active_queue;
1559
1560         if (cfq_class_idle(new_cfqq))
1561                 return 0;
1562
1563         if (!cfqq)
1564                 return 0;
1565
1566         if (cfq_class_idle(cfqq))
1567                 return 1;
1568         if (!cfq_cfqq_wait_request(new_cfqq))
1569                 return 0;
1570         /*
1571          * if it doesn't have slice left, forget it
1572          */
1573         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1574                 return 0;
1575         /*
1576          * if the new request is sync, but the currently running queue is
1577          * not, let the sync request have priority.
1578          */
1579         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1580                 return 1;
1581         /*
1582          * So both queues are sync. Let the new request get disk time if
1583          * it's a metadata request and the current queue is doing regular IO.
1584          */
1585         if (rq_is_meta(rq) && !cfqq->meta_pending)
1586                 return 1;
1587
1588         return 0;
1589 }
1590
1591 /*
1592  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1593  * let it have half of its nominal slice.
1594  */
1595 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1596 {
1597         cfq_slice_expired(cfqd, 1);
1598
1599         if (!cfqq->slice_left)
1600                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1601
1602         /*
1603          * Put the new queue at the front of the of the current list,
1604          * so we know that it will be selected next.
1605          */
1606         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1607         list_move(&cfqq->cfq_list, &cfqd->cur_rr);
1608
1609         cfqq->slice_end = cfqq->slice_left + jiffies;
1610 }
1611
1612 /*
1613  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1614  * something we should do about it
1615  */
1616 static void
1617 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1618                 struct request *rq)
1619 {
1620         struct cfq_io_context *cic = RQ_CIC(rq);
1621
1622         if (rq_is_meta(rq))
1623                 cfqq->meta_pending++;
1624
1625         /*
1626          * check if this request is a better next-serve candidate)) {
1627          */
1628         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
1629         BUG_ON(!cfqq->next_rq);
1630
1631         /*
1632          * we never wait for an async request and we don't allow preemption
1633          * of an async request. so just return early
1634          */
1635         if (!rq_is_sync(rq)) {
1636                 /*
1637                  * sync process issued an async request, if it's waiting
1638                  * then expire it and kick rq handling.
1639                  */
1640                 if (cic == cfqd->active_cic &&
1641                     del_timer(&cfqd->idle_slice_timer)) {
1642                         cfq_slice_expired(cfqd, 0);
1643                         blk_start_queueing(cfqd->queue);
1644                 }
1645                 return;
1646         }
1647
1648         cfq_update_io_thinktime(cfqd, cic);
1649         cfq_update_io_seektime(cic, rq);
1650         cfq_update_idle_window(cfqd, cfqq, cic);
1651
1652         cic->last_queue = jiffies;
1653         cic->last_request_pos = rq->sector + rq->nr_sectors;
1654
1655         if (cfqq == cfqd->active_queue) {
1656                 /*
1657                  * if we are waiting for a request for this queue, let it rip
1658                  * immediately and flag that we must not expire this queue
1659                  * just now
1660                  */
1661                 if (cfq_cfqq_wait_request(cfqq)) {
1662                         cfq_mark_cfqq_must_dispatch(cfqq);
1663                         del_timer(&cfqd->idle_slice_timer);
1664                         blk_start_queueing(cfqd->queue);
1665                 }
1666         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1667                 /*
1668                  * not the active queue - expire current slice if it is
1669                  * idle and has expired it's mean thinktime or this new queue
1670                  * has some old slice time left and is of higher priority
1671                  */
1672                 cfq_preempt_queue(cfqd, cfqq);
1673                 cfq_mark_cfqq_must_dispatch(cfqq);
1674                 blk_start_queueing(cfqd->queue);
1675         }
1676 }
1677
1678 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1679 {
1680         struct cfq_data *cfqd = q->elevator->elevator_data;
1681         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1682
1683         cfq_init_prio_data(cfqq);
1684
1685         cfq_add_rq_rb(rq);
1686
1687         list_add_tail(&rq->queuelist, &cfqq->fifo);
1688
1689         cfq_rq_enqueued(cfqd, cfqq, rq);
1690 }
1691
1692 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1693 {
1694         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1695         struct cfq_data *cfqd = cfqq->cfqd;
1696         const int sync = rq_is_sync(rq);
1697         unsigned long now;
1698
1699         now = jiffies;
1700
1701         WARN_ON(!cfqd->rq_in_driver);
1702         WARN_ON(!cfqq->on_dispatch[sync]);
1703         cfqd->rq_in_driver--;
1704         cfqq->on_dispatch[sync]--;
1705
1706         if (!cfq_class_idle(cfqq))
1707                 cfqd->last_end_request = now;
1708
1709         if (!cfq_cfqq_dispatched(cfqq) && cfq_cfqq_on_rr(cfqq))
1710                 cfq_resort_rr_list(cfqq, 0);
1711
1712         if (sync)
1713                 RQ_CIC(rq)->last_end_request = now;
1714
1715         /*
1716          * If this is the active queue, check if it needs to be expired,
1717          * or if we want to idle in case it has no pending requests.
1718          */
1719         if (cfqd->active_queue == cfqq) {
1720                 if (time_after(now, cfqq->slice_end))
1721                         cfq_slice_expired(cfqd, 0);
1722                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1723                         if (!cfq_arm_slice_timer(cfqd, cfqq))
1724                                 cfq_schedule_dispatch(cfqd);
1725                 }
1726         }
1727 }
1728
1729 /*
1730  * we temporarily boost lower priority queues if they are holding fs exclusive
1731  * resources. they are boosted to normal prio (CLASS_BE/4)
1732  */
1733 static void cfq_prio_boost(struct cfq_queue *cfqq)
1734 {
1735         const int ioprio_class = cfqq->ioprio_class;
1736         const int ioprio = cfqq->ioprio;
1737
1738         if (has_fs_excl()) {
1739                 /*
1740                  * boost idle prio on transactions that would lock out other
1741                  * users of the filesystem
1742                  */
1743                 if (cfq_class_idle(cfqq))
1744                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1745                 if (cfqq->ioprio > IOPRIO_NORM)
1746                         cfqq->ioprio = IOPRIO_NORM;
1747         } else {
1748                 /*
1749                  * check if we need to unboost the queue
1750                  */
1751                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1752                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1753                 if (cfqq->ioprio != cfqq->org_ioprio)
1754                         cfqq->ioprio = cfqq->org_ioprio;
1755         }
1756
1757         /*
1758          * refile between round-robin lists if we moved the priority class
1759          */
1760         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1761             cfq_cfqq_on_rr(cfqq))
1762                 cfq_resort_rr_list(cfqq, 0);
1763 }
1764
1765 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1766 {
1767         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1768             !cfq_cfqq_must_alloc_slice(cfqq)) {
1769                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1770                 return ELV_MQUEUE_MUST;
1771         }
1772
1773         return ELV_MQUEUE_MAY;
1774 }
1775
1776 static int cfq_may_queue(request_queue_t *q, int rw)
1777 {
1778         struct cfq_data *cfqd = q->elevator->elevator_data;
1779         struct task_struct *tsk = current;
1780         struct cfq_queue *cfqq;
1781         unsigned int key;
1782
1783         key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC);
1784
1785         /*
1786          * don't force setup of a queue from here, as a call to may_queue
1787          * does not necessarily imply that a request actually will be queued.
1788          * so just lookup a possibly existing queue, or return 'may queue'
1789          * if that fails
1790          */
1791         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
1792         if (cfqq) {
1793                 cfq_init_prio_data(cfqq);
1794                 cfq_prio_boost(cfqq);
1795
1796                 return __cfq_may_queue(cfqq);
1797         }
1798
1799         return ELV_MQUEUE_MAY;
1800 }
1801
1802 /*
1803  * queue lock held here
1804  */
1805 static void cfq_put_request(struct request *rq)
1806 {
1807         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1808
1809         if (cfqq) {
1810                 const int rw = rq_data_dir(rq);
1811
1812                 BUG_ON(!cfqq->allocated[rw]);
1813                 cfqq->allocated[rw]--;
1814
1815                 put_io_context(RQ_CIC(rq)->ioc);
1816
1817                 rq->elevator_private = NULL;
1818                 rq->elevator_private2 = NULL;
1819
1820                 cfq_put_queue(cfqq);
1821         }
1822 }
1823
1824 /*
1825  * Allocate cfq data structures associated with this request.
1826  */
1827 static int
1828 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1829 {
1830         struct cfq_data *cfqd = q->elevator->elevator_data;
1831         struct task_struct *tsk = current;
1832         struct cfq_io_context *cic;
1833         const int rw = rq_data_dir(rq);
1834         const int is_sync = rq_is_sync(rq);
1835         pid_t key = cfq_queue_pid(tsk, rw, is_sync);
1836         struct cfq_queue *cfqq;
1837         unsigned long flags;
1838
1839         might_sleep_if(gfp_mask & __GFP_WAIT);
1840
1841         cic = cfq_get_io_context(cfqd, gfp_mask);
1842
1843         spin_lock_irqsave(q->queue_lock, flags);
1844
1845         if (!cic)
1846                 goto queue_fail;
1847
1848         if (!cic->cfqq[is_sync]) {
1849                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1850                 if (!cfqq)
1851                         goto queue_fail;
1852
1853                 cic->cfqq[is_sync] = cfqq;
1854         } else
1855                 cfqq = cic->cfqq[is_sync];
1856
1857         cfqq->allocated[rw]++;
1858         cfq_clear_cfqq_must_alloc(cfqq);
1859         atomic_inc(&cfqq->ref);
1860
1861         spin_unlock_irqrestore(q->queue_lock, flags);
1862
1863         rq->elevator_private = cic;
1864         rq->elevator_private2 = cfqq;
1865         return 0;
1866
1867 queue_fail:
1868         if (cic)
1869                 put_io_context(cic->ioc);
1870
1871         cfq_schedule_dispatch(cfqd);
1872         spin_unlock_irqrestore(q->queue_lock, flags);
1873         return 1;
1874 }
1875
1876 static void cfq_kick_queue(struct work_struct *work)
1877 {
1878         struct cfq_data *cfqd =
1879                 container_of(work, struct cfq_data, unplug_work);
1880         request_queue_t *q = cfqd->queue;
1881         unsigned long flags;
1882
1883         spin_lock_irqsave(q->queue_lock, flags);
1884         blk_start_queueing(q);
1885         spin_unlock_irqrestore(q->queue_lock, flags);
1886 }
1887
1888 /*
1889  * Timer running if the active_queue is currently idling inside its time slice
1890  */
1891 static void cfq_idle_slice_timer(unsigned long data)
1892 {
1893         struct cfq_data *cfqd = (struct cfq_data *) data;
1894         struct cfq_queue *cfqq;
1895         unsigned long flags;
1896
1897         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1898
1899         if ((cfqq = cfqd->active_queue) != NULL) {
1900                 unsigned long now = jiffies;
1901
1902                 /*
1903                  * expired
1904                  */
1905                 if (time_after(now, cfqq->slice_end))
1906                         goto expire;
1907
1908                 /*
1909                  * only expire and reinvoke request handler, if there are
1910                  * other queues with pending requests
1911                  */
1912                 if (!cfqd->busy_queues)
1913                         goto out_cont;
1914
1915                 /*
1916                  * not expired and it has a request pending, let it dispatch
1917                  */
1918                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1919                         cfq_mark_cfqq_must_dispatch(cfqq);
1920                         goto out_kick;
1921                 }
1922         }
1923 expire:
1924         cfq_slice_expired(cfqd, 0);
1925 out_kick:
1926         cfq_schedule_dispatch(cfqd);
1927 out_cont:
1928         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1929 }
1930
1931 /*
1932  * Timer running if an idle class queue is waiting for service
1933  */
1934 static void cfq_idle_class_timer(unsigned long data)
1935 {
1936         struct cfq_data *cfqd = (struct cfq_data *) data;
1937         unsigned long flags, end;
1938
1939         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1940
1941         /*
1942          * race with a non-idle queue, reset timer
1943          */
1944         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
1945         if (!time_after_eq(jiffies, end))
1946                 mod_timer(&cfqd->idle_class_timer, end);
1947         else
1948                 cfq_schedule_dispatch(cfqd);
1949
1950         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1951 }
1952
1953 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1954 {
1955         del_timer_sync(&cfqd->idle_slice_timer);
1956         del_timer_sync(&cfqd->idle_class_timer);
1957         blk_sync_queue(cfqd->queue);
1958 }
1959
1960 static void cfq_exit_queue(elevator_t *e)
1961 {
1962         struct cfq_data *cfqd = e->elevator_data;
1963         request_queue_t *q = cfqd->queue;
1964
1965         cfq_shutdown_timer_wq(cfqd);
1966
1967         spin_lock_irq(q->queue_lock);
1968
1969         if (cfqd->active_queue)
1970                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
1971
1972         while (!list_empty(&cfqd->cic_list)) {
1973                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1974                                                         struct cfq_io_context,
1975                                                         queue_list);
1976
1977                 __cfq_exit_single_io_context(cfqd, cic);
1978         }
1979
1980         spin_unlock_irq(q->queue_lock);
1981
1982         cfq_shutdown_timer_wq(cfqd);
1983
1984         kfree(cfqd->cfq_hash);
1985         kfree(cfqd);
1986 }
1987
1988 static void *cfq_init_queue(request_queue_t *q)
1989 {
1990         struct cfq_data *cfqd;
1991         int i;
1992
1993         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
1994         if (!cfqd)
1995                 return NULL;
1996
1997         memset(cfqd, 0, sizeof(*cfqd));
1998
1999         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2000                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2001
2002         INIT_LIST_HEAD(&cfqd->busy_rr);
2003         INIT_LIST_HEAD(&cfqd->cur_rr);
2004         INIT_LIST_HEAD(&cfqd->idle_rr);
2005         INIT_LIST_HEAD(&cfqd->cic_list);
2006
2007         cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
2008         if (!cfqd->cfq_hash)
2009                 goto out_free;
2010
2011         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2012                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2013
2014         cfqd->queue = q;
2015
2016         init_timer(&cfqd->idle_slice_timer);
2017         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2018         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2019
2020         init_timer(&cfqd->idle_class_timer);
2021         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2022         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2023
2024         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2025
2026         cfqd->cfq_quantum = cfq_quantum;
2027         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2028         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2029         cfqd->cfq_back_max = cfq_back_max;
2030         cfqd->cfq_back_penalty = cfq_back_penalty;
2031         cfqd->cfq_slice[0] = cfq_slice_async;
2032         cfqd->cfq_slice[1] = cfq_slice_sync;
2033         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2034         cfqd->cfq_slice_idle = cfq_slice_idle;
2035
2036         return cfqd;
2037 out_free:
2038         kfree(cfqd);
2039         return NULL;
2040 }
2041
2042 static void cfq_slab_kill(void)
2043 {
2044         if (cfq_pool)
2045                 kmem_cache_destroy(cfq_pool);
2046         if (cfq_ioc_pool)
2047                 kmem_cache_destroy(cfq_ioc_pool);
2048 }
2049
2050 static int __init cfq_slab_setup(void)
2051 {
2052         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2053                                         NULL, NULL);
2054         if (!cfq_pool)
2055                 goto fail;
2056
2057         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2058                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2059         if (!cfq_ioc_pool)
2060                 goto fail;
2061
2062         return 0;
2063 fail:
2064         cfq_slab_kill();
2065         return -ENOMEM;
2066 }
2067
2068 /*
2069  * sysfs parts below -->
2070  */
2071
2072 static ssize_t
2073 cfq_var_show(unsigned int var, char *page)
2074 {
2075         return sprintf(page, "%d\n", var);
2076 }
2077
2078 static ssize_t
2079 cfq_var_store(unsigned int *var, const char *page, size_t count)
2080 {
2081         char *p = (char *) page;
2082
2083         *var = simple_strtoul(p, &p, 10);
2084         return count;
2085 }
2086
2087 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2088 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2089 {                                                                       \
2090         struct cfq_data *cfqd = e->elevator_data;                       \
2091         unsigned int __data = __VAR;                                    \
2092         if (__CONV)                                                     \
2093                 __data = jiffies_to_msecs(__data);                      \
2094         return cfq_var_show(__data, (page));                            \
2095 }
2096 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2097 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2098 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2099 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2100 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2101 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2102 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2103 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2104 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2105 #undef SHOW_FUNCTION
2106
2107 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2108 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2109 {                                                                       \
2110         struct cfq_data *cfqd = e->elevator_data;                       \
2111         unsigned int __data;                                            \
2112         int ret = cfq_var_store(&__data, (page), count);                \
2113         if (__data < (MIN))                                             \
2114                 __data = (MIN);                                         \
2115         else if (__data > (MAX))                                        \
2116                 __data = (MAX);                                         \
2117         if (__CONV)                                                     \
2118                 *(__PTR) = msecs_to_jiffies(__data);                    \
2119         else                                                            \
2120                 *(__PTR) = __data;                                      \
2121         return ret;                                                     \
2122 }
2123 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2124 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2125 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2126 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2127 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2128 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2129 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2130 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2131 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2132 #undef STORE_FUNCTION
2133
2134 #define CFQ_ATTR(name) \
2135         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2136
2137 static struct elv_fs_entry cfq_attrs[] = {
2138         CFQ_ATTR(quantum),
2139         CFQ_ATTR(fifo_expire_sync),
2140         CFQ_ATTR(fifo_expire_async),
2141         CFQ_ATTR(back_seek_max),
2142         CFQ_ATTR(back_seek_penalty),
2143         CFQ_ATTR(slice_sync),
2144         CFQ_ATTR(slice_async),
2145         CFQ_ATTR(slice_async_rq),
2146         CFQ_ATTR(slice_idle),
2147         __ATTR_NULL
2148 };
2149
2150 static struct elevator_type iosched_cfq = {
2151         .ops = {
2152                 .elevator_merge_fn =            cfq_merge,
2153                 .elevator_merged_fn =           cfq_merged_request,
2154                 .elevator_merge_req_fn =        cfq_merged_requests,
2155                 .elevator_allow_merge_fn =      cfq_allow_merge,
2156                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2157                 .elevator_add_req_fn =          cfq_insert_request,
2158                 .elevator_activate_req_fn =     cfq_activate_request,
2159                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2160                 .elevator_queue_empty_fn =      cfq_queue_empty,
2161                 .elevator_completed_req_fn =    cfq_completed_request,
2162                 .elevator_former_req_fn =       elv_rb_former_request,
2163                 .elevator_latter_req_fn =       elv_rb_latter_request,
2164                 .elevator_set_req_fn =          cfq_set_request,
2165                 .elevator_put_req_fn =          cfq_put_request,
2166                 .elevator_may_queue_fn =        cfq_may_queue,
2167                 .elevator_init_fn =             cfq_init_queue,
2168                 .elevator_exit_fn =             cfq_exit_queue,
2169                 .trim =                         cfq_free_io_context,
2170         },
2171         .elevator_attrs =       cfq_attrs,
2172         .elevator_name =        "cfq",
2173         .elevator_owner =       THIS_MODULE,
2174 };
2175
2176 static int __init cfq_init(void)
2177 {
2178         int ret;
2179
2180         /*
2181          * could be 0 on HZ < 1000 setups
2182          */
2183         if (!cfq_slice_async)
2184                 cfq_slice_async = 1;
2185         if (!cfq_slice_idle)
2186                 cfq_slice_idle = 1;
2187
2188         if (cfq_slab_setup())
2189                 return -ENOMEM;
2190
2191         ret = elv_register(&iosched_cfq);
2192         if (ret)
2193                 cfq_slab_kill();
2194
2195         return ret;
2196 }
2197
2198 static void __exit cfq_exit(void)
2199 {
2200         DECLARE_COMPLETION_ONSTACK(all_gone);
2201         elv_unregister(&iosched_cfq);
2202         ioc_gone = &all_gone;
2203         /* ioc_gone's update must be visible before reading ioc_count */
2204         smp_wmb();
2205         if (elv_ioc_count_read(ioc_count))
2206                 wait_for_completion(ioc_gone);
2207         synchronize_rcu();
2208         cfq_slab_kill();
2209 }
2210
2211 module_init(cfq_init);
2212 module_exit(cfq_exit);
2213
2214 MODULE_AUTHOR("Jens Axboe");
2215 MODULE_LICENSE("GPL");
2216 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");