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