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