[PATCH] Drop useless bio passing in may_queue/set_request API
[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_queued = 8;                /* minimum rq allocate limit per-queue*/
21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
22 static const int cfq_back_max = 16 * 1024;      /* maximum backwards seek, in KiB */
23 static const int cfq_back_penalty = 2;          /* penalty of a backwards seek */
24
25 static const int cfq_slice_sync = HZ / 10;
26 static int cfq_slice_async = HZ / 25;
27 static const int cfq_slice_async_rq = 2;
28 static int cfq_slice_idle = HZ / 125;
29
30 #define CFQ_IDLE_GRACE          (HZ / 10)
31 #define CFQ_SLICE_SCALE         (5)
32
33 #define CFQ_KEY_ASYNC           (0)
34
35 static DEFINE_SPINLOCK(cfq_exit_lock);
36
37 /*
38  * for the hash of cfqq inside the cfqd
39  */
40 #define CFQ_QHASH_SHIFT         6
41 #define CFQ_QHASH_ENTRIES       (1 << CFQ_QHASH_SHIFT)
42 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
43
44 #define list_entry_cfqq(ptr)    list_entry((ptr), struct cfq_queue, cfq_list)
45
46 #define RQ_CIC(rq)              ((struct cfq_io_context*)(rq)->elevator_private)
47 #define RQ_CFQQ(rq)             ((rq)->elevator_private2)
48
49 static kmem_cache_t *cfq_pool;
50 static kmem_cache_t *cfq_ioc_pool;
51
52 static atomic_t ioc_count = ATOMIC_INIT(0);
53 static struct completion *ioc_gone;
54
55 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
56 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
57 #define cfq_class_be(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
58 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
59
60 #define ASYNC                   (0)
61 #define SYNC                    (1)
62
63 #define cfq_cfqq_dispatched(cfqq)       \
64         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
65
66 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
67
68 #define cfq_cfqq_sync(cfqq)             \
69         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
70
71 #define sample_valid(samples)   ((samples) > 80)
72
73 /*
74  * Per block device queue structure
75  */
76 struct cfq_data {
77         request_queue_t *queue;
78
79         /*
80          * rr list of queues with requests and the count of them
81          */
82         struct list_head rr_list[CFQ_PRIO_LISTS];
83         struct list_head busy_rr;
84         struct list_head cur_rr;
85         struct list_head idle_rr;
86         unsigned int busy_queues;
87
88         /*
89          * non-ordered list of empty cfqq's
90          */
91         struct list_head empty_list;
92
93         /*
94          * cfqq lookup hash
95          */
96         struct hlist_head *cfq_hash;
97
98         int rq_in_driver;
99         int hw_tag;
100
101         /*
102          * schedule slice state info
103          */
104         /*
105          * idle window management
106          */
107         struct timer_list idle_slice_timer;
108         struct work_struct unplug_work;
109
110         struct cfq_queue *active_queue;
111         struct cfq_io_context *active_cic;
112         int cur_prio, cur_end_prio;
113         unsigned int dispatch_slice;
114
115         struct timer_list idle_class_timer;
116
117         sector_t last_sector;
118         unsigned long last_end_request;
119
120         unsigned int rq_starved;
121
122         /*
123          * tunables, see top of file
124          */
125         unsigned int cfq_quantum;
126         unsigned int cfq_queued;
127         unsigned int cfq_fifo_expire[2];
128         unsigned int cfq_back_penalty;
129         unsigned int cfq_back_max;
130         unsigned int cfq_slice[2];
131         unsigned int cfq_slice_async_rq;
132         unsigned int cfq_slice_idle;
133
134         struct list_head cic_list;
135 };
136
137 /*
138  * Per process-grouping structure
139  */
140 struct cfq_queue {
141         /* reference count */
142         atomic_t ref;
143         /* parent cfq_data */
144         struct cfq_data *cfqd;
145         /* cfqq lookup hash */
146         struct hlist_node cfq_hash;
147         /* hash key */
148         unsigned int key;
149         /* on either rr or empty list of cfqd */
150         struct list_head cfq_list;
151         /* sorted list of pending requests */
152         struct rb_root sort_list;
153         /* if fifo isn't expired, next request to serve */
154         struct request *next_rq;
155         /* requests queued in sort_list */
156         int queued[2];
157         /* currently allocated requests */
158         int allocated[2];
159         /* fifo list of requests in sort_list */
160         struct list_head fifo;
161
162         unsigned long slice_start;
163         unsigned long slice_end;
164         unsigned long slice_left;
165         unsigned long service_last;
166
167         /* number of requests that are on the dispatch list */
168         int on_dispatch[2];
169
170         /* io prio of this group */
171         unsigned short ioprio, org_ioprio;
172         unsigned short ioprio_class, org_ioprio_class;
173
174         /* various state flags, see below */
175         unsigned int flags;
176 };
177
178 enum cfqq_state_flags {
179         CFQ_CFQQ_FLAG_on_rr = 0,
180         CFQ_CFQQ_FLAG_wait_request,
181         CFQ_CFQQ_FLAG_must_alloc,
182         CFQ_CFQQ_FLAG_must_alloc_slice,
183         CFQ_CFQQ_FLAG_must_dispatch,
184         CFQ_CFQQ_FLAG_fifo_expire,
185         CFQ_CFQQ_FLAG_idle_window,
186         CFQ_CFQQ_FLAG_prio_changed,
187 };
188
189 #define CFQ_CFQQ_FNS(name)                                              \
190 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
191 {                                                                       \
192         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
193 }                                                                       \
194 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
195 {                                                                       \
196         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
197 }                                                                       \
198 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
199 {                                                                       \
200         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
201 }
202
203 CFQ_CFQQ_FNS(on_rr);
204 CFQ_CFQQ_FNS(wait_request);
205 CFQ_CFQQ_FNS(must_alloc);
206 CFQ_CFQQ_FNS(must_alloc_slice);
207 CFQ_CFQQ_FNS(must_dispatch);
208 CFQ_CFQQ_FNS(fifo_expire);
209 CFQ_CFQQ_FNS(idle_window);
210 CFQ_CFQQ_FNS(prio_changed);
211 #undef CFQ_CFQQ_FNS
212
213 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
214 static void cfq_dispatch_insert(request_queue_t *, struct request *);
215 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
216
217 /*
218  * scheduler run of queue, if there are requests pending and no one in the
219  * driver that will restart queueing
220  */
221 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
222 {
223         if (cfqd->busy_queues)
224                 kblockd_schedule_work(&cfqd->unplug_work);
225 }
226
227 static int cfq_queue_empty(request_queue_t *q)
228 {
229         struct cfq_data *cfqd = q->elevator->elevator_data;
230
231         return !cfqd->busy_queues;
232 }
233
234 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
235 {
236         if (rw == READ || rw == WRITE_SYNC)
237                 return task->pid;
238
239         return CFQ_KEY_ASYNC;
240 }
241
242 /*
243  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
244  * We choose the request that is closest to the head right now. Distance
245  * behind the head is penalized and only allowed to a certain extent.
246  */
247 static struct request *
248 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
249 {
250         sector_t last, s1, s2, d1 = 0, d2 = 0;
251         unsigned long back_max;
252 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
253 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
254         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
255
256         if (rq1 == NULL || rq1 == rq2)
257                 return rq2;
258         if (rq2 == NULL)
259                 return rq1;
260
261         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
262                 return rq1;
263         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
264                 return rq2;
265
266         s1 = rq1->sector;
267         s2 = rq2->sector;
268
269         last = cfqd->last_sector;
270
271         /*
272          * by definition, 1KiB is 2 sectors
273          */
274         back_max = cfqd->cfq_back_max * 2;
275
276         /*
277          * Strict one way elevator _except_ in the case where we allow
278          * short backward seeks which are biased as twice the cost of a
279          * similar forward seek.
280          */
281         if (s1 >= last)
282                 d1 = s1 - last;
283         else if (s1 + back_max >= last)
284                 d1 = (last - s1) * cfqd->cfq_back_penalty;
285         else
286                 wrap |= CFQ_RQ1_WRAP;
287
288         if (s2 >= last)
289                 d2 = s2 - last;
290         else if (s2 + back_max >= last)
291                 d2 = (last - s2) * cfqd->cfq_back_penalty;
292         else
293                 wrap |= CFQ_RQ2_WRAP;
294
295         /* Found required data */
296
297         /*
298          * By doing switch() on the bit mask "wrap" we avoid having to
299          * check two variables for all permutations: --> faster!
300          */
301         switch (wrap) {
302         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
303                 if (d1 < d2)
304                         return rq1;
305                 else if (d2 < d1)
306                         return rq2;
307                 else {
308                         if (s1 >= s2)
309                                 return rq1;
310                         else
311                                 return rq2;
312                 }
313
314         case CFQ_RQ2_WRAP:
315                 return rq1;
316         case CFQ_RQ1_WRAP:
317                 return rq2;
318         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
319         default:
320                 /*
321                  * Since both rqs are wrapped,
322                  * start with the one that's further behind head
323                  * (--> only *one* back seek required),
324                  * since back seek takes more time than forward.
325                  */
326                 if (s1 <= s2)
327                         return rq1;
328                 else
329                         return rq2;
330         }
331 }
332
333 /*
334  * would be nice to take fifo expire time into account as well
335  */
336 static struct request *
337 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
338                   struct request *last)
339 {
340         struct rb_node *rbnext = rb_next(&last->rb_node);
341         struct rb_node *rbprev = rb_prev(&last->rb_node);
342         struct request *next = NULL, *prev = NULL;
343
344         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
345
346         if (rbprev)
347                 prev = rb_entry_rq(rbprev);
348
349         if (rbnext)
350                 next = rb_entry_rq(rbnext);
351         else {
352                 rbnext = rb_first(&cfqq->sort_list);
353                 if (rbnext && rbnext != &last->rb_node)
354                         next = rb_entry_rq(rbnext);
355         }
356
357         return cfq_choose_req(cfqd, next, prev);
358 }
359
360 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
361 {
362         struct cfq_data *cfqd = cfqq->cfqd;
363         struct list_head *list, *entry;
364
365         BUG_ON(!cfq_cfqq_on_rr(cfqq));
366
367         list_del(&cfqq->cfq_list);
368
369         if (cfq_class_rt(cfqq))
370                 list = &cfqd->cur_rr;
371         else if (cfq_class_idle(cfqq))
372                 list = &cfqd->idle_rr;
373         else {
374                 /*
375                  * if cfqq has requests in flight, don't allow it to be
376                  * found in cfq_set_active_queue before it has finished them.
377                  * this is done to increase fairness between a process that
378                  * has lots of io pending vs one that only generates one
379                  * sporadically or synchronously
380                  */
381                 if (cfq_cfqq_dispatched(cfqq))
382                         list = &cfqd->busy_rr;
383                 else
384                         list = &cfqd->rr_list[cfqq->ioprio];
385         }
386
387         /*
388          * if queue was preempted, just add to front to be fair. busy_rr
389          * isn't sorted, but insert at the back for fairness.
390          */
391         if (preempted || list == &cfqd->busy_rr) {
392                 if (preempted)
393                         list = list->prev;
394
395                 list_add_tail(&cfqq->cfq_list, list);
396                 return;
397         }
398
399         /*
400          * sort by when queue was last serviced
401          */
402         entry = list;
403         while ((entry = entry->prev) != list) {
404                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
405
406                 if (!__cfqq->service_last)
407                         break;
408                 if (time_before(__cfqq->service_last, cfqq->service_last))
409                         break;
410         }
411
412         list_add(&cfqq->cfq_list, entry);
413 }
414
415 /*
416  * add to busy list of queues for service, trying to be fair in ordering
417  * the pending list according to last request service
418  */
419 static inline void
420 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
421 {
422         BUG_ON(cfq_cfqq_on_rr(cfqq));
423         cfq_mark_cfqq_on_rr(cfqq);
424         cfqd->busy_queues++;
425
426         cfq_resort_rr_list(cfqq, 0);
427 }
428
429 static inline void
430 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
431 {
432         BUG_ON(!cfq_cfqq_on_rr(cfqq));
433         cfq_clear_cfqq_on_rr(cfqq);
434         list_move(&cfqq->cfq_list, &cfqd->empty_list);
435
436         BUG_ON(!cfqd->busy_queues);
437         cfqd->busy_queues--;
438 }
439
440 /*
441  * rb tree support functions
442  */
443 static inline void cfq_del_rq_rb(struct request *rq)
444 {
445         struct cfq_queue *cfqq = RQ_CFQQ(rq);
446         struct cfq_data *cfqd = cfqq->cfqd;
447         const int sync = rq_is_sync(rq);
448
449         BUG_ON(!cfqq->queued[sync]);
450         cfqq->queued[sync]--;
451
452         elv_rb_del(&cfqq->sort_list, rq);
453
454         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
455                 cfq_del_cfqq_rr(cfqd, cfqq);
456 }
457
458 static void cfq_add_rq_rb(struct request *rq)
459 {
460         struct cfq_queue *cfqq = RQ_CFQQ(rq);
461         struct cfq_data *cfqd = cfqq->cfqd;
462         struct request *__alias;
463
464         cfqq->queued[rq_is_sync(rq)]++;
465
466         /*
467          * looks a little odd, but the first insert might return an alias.
468          * if that happens, put the alias on the dispatch list
469          */
470         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
471                 cfq_dispatch_insert(cfqd->queue, __alias);
472 }
473
474 static inline void
475 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
476 {
477         elv_rb_del(&cfqq->sort_list, rq);
478         cfqq->queued[rq_is_sync(rq)]--;
479         cfq_add_rq_rb(rq);
480 }
481
482 static struct request *
483 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
484 {
485         struct task_struct *tsk = current;
486         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
487         sector_t sector = bio->bi_sector + bio_sectors(bio);
488         struct cfq_queue *cfqq;
489
490         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
491         if (cfqq)
492                 return elv_rb_find(&cfqq->sort_list, sector);
493
494         return NULL;
495 }
496
497 static void cfq_activate_request(request_queue_t *q, struct request *rq)
498 {
499         struct cfq_data *cfqd = q->elevator->elevator_data;
500
501         cfqd->rq_in_driver++;
502
503         /*
504          * If the depth is larger 1, it really could be queueing. But lets
505          * make the mark a little higher - idling could still be good for
506          * low queueing, and a low queueing number could also just indicate
507          * a SCSI mid layer like behaviour where limit+1 is often seen.
508          */
509         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
510                 cfqd->hw_tag = 1;
511 }
512
513 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
514 {
515         struct cfq_data *cfqd = q->elevator->elevator_data;
516
517         WARN_ON(!cfqd->rq_in_driver);
518         cfqd->rq_in_driver--;
519 }
520
521 static void cfq_remove_request(struct request *rq)
522 {
523         struct cfq_queue *cfqq = RQ_CFQQ(rq);
524
525         if (cfqq->next_rq == rq)
526                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
527
528         list_del_init(&rq->queuelist);
529         cfq_del_rq_rb(rq);
530 }
531
532 static int
533 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
534 {
535         struct cfq_data *cfqd = q->elevator->elevator_data;
536         struct request *__rq;
537
538         __rq = cfq_find_rq_fmerge(cfqd, bio);
539         if (__rq && elv_rq_merge_ok(__rq, bio)) {
540                 *req = __rq;
541                 return ELEVATOR_FRONT_MERGE;
542         }
543
544         return ELEVATOR_NO_MERGE;
545 }
546
547 static void cfq_merged_request(request_queue_t *q, struct request *req,
548                                int type)
549 {
550         if (type == ELEVATOR_FRONT_MERGE) {
551                 struct cfq_queue *cfqq = RQ_CFQQ(req);
552
553                 cfq_reposition_rq_rb(cfqq, req);
554         }
555 }
556
557 static void
558 cfq_merged_requests(request_queue_t *q, struct request *rq,
559                     struct request *next)
560 {
561         /*
562          * reposition in fifo if next is older than rq
563          */
564         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
565             time_before(next->start_time, rq->start_time))
566                 list_move(&rq->queuelist, &next->queuelist);
567
568         cfq_remove_request(next);
569 }
570
571 static inline void
572 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
573 {
574         if (cfqq) {
575                 /*
576                  * stop potential idle class queues waiting service
577                  */
578                 del_timer(&cfqd->idle_class_timer);
579
580                 cfqq->slice_start = jiffies;
581                 cfqq->slice_end = 0;
582                 cfqq->slice_left = 0;
583                 cfq_clear_cfqq_must_alloc_slice(cfqq);
584                 cfq_clear_cfqq_fifo_expire(cfqq);
585         }
586
587         cfqd->active_queue = cfqq;
588 }
589
590 /*
591  * current cfqq expired its slice (or was too idle), select new one
592  */
593 static void
594 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
595                     int preempted)
596 {
597         unsigned long now = jiffies;
598
599         if (cfq_cfqq_wait_request(cfqq))
600                 del_timer(&cfqd->idle_slice_timer);
601
602         if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
603                 cfqq->service_last = now;
604                 cfq_schedule_dispatch(cfqd);
605         }
606
607         cfq_clear_cfqq_must_dispatch(cfqq);
608         cfq_clear_cfqq_wait_request(cfqq);
609
610         /*
611          * store what was left of this slice, if the queue idled out
612          * or was preempted
613          */
614         if (time_after(cfqq->slice_end, now))
615                 cfqq->slice_left = cfqq->slice_end - now;
616         else
617                 cfqq->slice_left = 0;
618
619         if (cfq_cfqq_on_rr(cfqq))
620                 cfq_resort_rr_list(cfqq, preempted);
621
622         if (cfqq == cfqd->active_queue)
623                 cfqd->active_queue = NULL;
624
625         if (cfqd->active_cic) {
626                 put_io_context(cfqd->active_cic->ioc);
627                 cfqd->active_cic = NULL;
628         }
629
630         cfqd->dispatch_slice = 0;
631 }
632
633 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
634 {
635         struct cfq_queue *cfqq = cfqd->active_queue;
636
637         if (cfqq)
638                 __cfq_slice_expired(cfqd, cfqq, preempted);
639 }
640
641 /*
642  * 0
643  * 0,1
644  * 0,1,2
645  * 0,1,2,3
646  * 0,1,2,3,4
647  * 0,1,2,3,4,5
648  * 0,1,2,3,4,5,6
649  * 0,1,2,3,4,5,6,7
650  */
651 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
652 {
653         int prio, wrap;
654
655         prio = -1;
656         wrap = 0;
657         do {
658                 int p;
659
660                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
661                         if (!list_empty(&cfqd->rr_list[p])) {
662                                 prio = p;
663                                 break;
664                         }
665                 }
666
667                 if (prio != -1)
668                         break;
669                 cfqd->cur_prio = 0;
670                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
671                         cfqd->cur_end_prio = 0;
672                         if (wrap)
673                                 break;
674                         wrap = 1;
675                 }
676         } while (1);
677
678         if (unlikely(prio == -1))
679                 return -1;
680
681         BUG_ON(prio >= CFQ_PRIO_LISTS);
682
683         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
684
685         cfqd->cur_prio = prio + 1;
686         if (cfqd->cur_prio > cfqd->cur_end_prio) {
687                 cfqd->cur_end_prio = cfqd->cur_prio;
688                 cfqd->cur_prio = 0;
689         }
690         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
691                 cfqd->cur_prio = 0;
692                 cfqd->cur_end_prio = 0;
693         }
694
695         return prio;
696 }
697
698 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
699 {
700         struct cfq_queue *cfqq = NULL;
701
702         /*
703          * if current list is non-empty, grab first entry. if it is empty,
704          * get next prio level and grab first entry then if any are spliced
705          */
706         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
707                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
708
709         /*
710          * If no new queues are available, check if the busy list has some
711          * before falling back to idle io.
712          */
713         if (!cfqq && !list_empty(&cfqd->busy_rr))
714                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
715
716         /*
717          * if we have idle queues and no rt or be queues had pending
718          * requests, either allow immediate service if the grace period
719          * has passed or arm the idle grace timer
720          */
721         if (!cfqq && !list_empty(&cfqd->idle_rr)) {
722                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
723
724                 if (time_after_eq(jiffies, end))
725                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
726                 else
727                         mod_timer(&cfqd->idle_class_timer, end);
728         }
729
730         __cfq_set_active_queue(cfqd, cfqq);
731         return cfqq;
732 }
733
734 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
735
736 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
737
738 {
739         struct cfq_io_context *cic;
740         unsigned long sl;
741
742         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
743         WARN_ON(cfqq != cfqd->active_queue);
744
745         /*
746          * idle is disabled, either manually or by past process history
747          */
748         if (!cfqd->cfq_slice_idle)
749                 return 0;
750         if (!cfq_cfqq_idle_window(cfqq))
751                 return 0;
752         /*
753          * task has exited, don't wait
754          */
755         cic = cfqd->active_cic;
756         if (!cic || !cic->ioc->task)
757                 return 0;
758
759         cfq_mark_cfqq_must_dispatch(cfqq);
760         cfq_mark_cfqq_wait_request(cfqq);
761
762         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
763
764         /*
765          * we don't want to idle for seeks, but we do want to allow
766          * fair distribution of slice time for a process doing back-to-back
767          * seeks. so allow a little bit of time for him to submit a new rq
768          */
769         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
770                 sl = min(sl, msecs_to_jiffies(2));
771
772         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
773         return 1;
774 }
775
776 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
777 {
778         struct cfq_data *cfqd = q->elevator->elevator_data;
779         struct cfq_queue *cfqq = RQ_CFQQ(rq);
780
781         cfq_remove_request(rq);
782         cfqq->on_dispatch[rq_is_sync(rq)]++;
783         elv_dispatch_sort(q, rq);
784
785         rq = list_entry(q->queue_head.prev, struct request, queuelist);
786         cfqd->last_sector = rq->sector + rq->nr_sectors;
787 }
788
789 /*
790  * return expired entry, or NULL to just start from scratch in rbtree
791  */
792 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
793 {
794         struct cfq_data *cfqd = cfqq->cfqd;
795         struct request *rq;
796
797         if (cfq_cfqq_fifo_expire(cfqq))
798                 return NULL;
799
800         if (!list_empty(&cfqq->fifo)) {
801                 int fifo = cfq_cfqq_class_sync(cfqq);
802
803                 rq = rq_entry_fifo(cfqq->fifo.next);
804                 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
805                         cfq_mark_cfqq_fifo_expire(cfqq);
806                         return rq;
807                 }
808         }
809
810         return NULL;
811 }
812
813 /*
814  * Scale schedule slice based on io priority. Use the sync time slice only
815  * if a queue is marked sync and has sync io queued. A sync queue with async
816  * io only, should not get full sync slice length.
817  */
818 static inline int
819 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
820 {
821         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
822
823         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
824
825         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
826 }
827
828 static inline void
829 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
830 {
831         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
832 }
833
834 static inline int
835 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
836 {
837         const int base_rq = cfqd->cfq_slice_async_rq;
838
839         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
840
841         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
842 }
843
844 /*
845  * get next queue for service
846  */
847 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
848 {
849         unsigned long now = jiffies;
850         struct cfq_queue *cfqq;
851
852         cfqq = cfqd->active_queue;
853         if (!cfqq)
854                 goto new_queue;
855
856         /*
857          * slice has expired
858          */
859         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
860                 goto expire;
861
862         /*
863          * if queue has requests, dispatch one. if not, check if
864          * enough slice is left to wait for one
865          */
866         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
867                 goto keep_queue;
868         else if (cfq_cfqq_dispatched(cfqq)) {
869                 cfqq = NULL;
870                 goto keep_queue;
871         } else if (cfq_cfqq_class_sync(cfqq)) {
872                 if (cfq_arm_slice_timer(cfqd, cfqq))
873                         return NULL;
874         }
875
876 expire:
877         cfq_slice_expired(cfqd, 0);
878 new_queue:
879         cfqq = cfq_set_active_queue(cfqd);
880 keep_queue:
881         return cfqq;
882 }
883
884 static int
885 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
886                         int max_dispatch)
887 {
888         int dispatched = 0;
889
890         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
891
892         do {
893                 struct request *rq;
894
895                 /*
896                  * follow expired path, else get first next available
897                  */
898                 if ((rq = cfq_check_fifo(cfqq)) == NULL)
899                         rq = cfqq->next_rq;
900
901                 /*
902                  * finally, insert request into driver dispatch list
903                  */
904                 cfq_dispatch_insert(cfqd->queue, rq);
905
906                 cfqd->dispatch_slice++;
907                 dispatched++;
908
909                 if (!cfqd->active_cic) {
910                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
911                         cfqd->active_cic = RQ_CIC(rq);
912                 }
913
914                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
915                         break;
916
917         } while (dispatched < max_dispatch);
918
919         /*
920          * if slice end isn't set yet, set it.
921          */
922         if (!cfqq->slice_end)
923                 cfq_set_prio_slice(cfqd, cfqq);
924
925         /*
926          * expire an async queue immediately if it has used up its slice. idle
927          * queue always expire after 1 dispatch round.
928          */
929         if ((!cfq_cfqq_sync(cfqq) &&
930             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
931             cfq_class_idle(cfqq) ||
932             !cfq_cfqq_idle_window(cfqq))
933                 cfq_slice_expired(cfqd, 0);
934
935         return dispatched;
936 }
937
938 static int
939 cfq_forced_dispatch_cfqqs(struct list_head *list)
940 {
941         struct cfq_queue *cfqq, *next;
942         int dispatched;
943
944         dispatched = 0;
945         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
946                 while (cfqq->next_rq) {
947                         cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
948                         dispatched++;
949                 }
950                 BUG_ON(!list_empty(&cfqq->fifo));
951         }
952
953         return dispatched;
954 }
955
956 static int
957 cfq_forced_dispatch(struct cfq_data *cfqd)
958 {
959         int i, dispatched = 0;
960
961         for (i = 0; i < CFQ_PRIO_LISTS; i++)
962                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
963
964         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
965         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
966         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
967
968         cfq_slice_expired(cfqd, 0);
969
970         BUG_ON(cfqd->busy_queues);
971
972         return dispatched;
973 }
974
975 static int
976 cfq_dispatch_requests(request_queue_t *q, int force)
977 {
978         struct cfq_data *cfqd = q->elevator->elevator_data;
979         struct cfq_queue *cfqq, *prev_cfqq;
980         int dispatched;
981
982         if (!cfqd->busy_queues)
983                 return 0;
984
985         if (unlikely(force))
986                 return cfq_forced_dispatch(cfqd);
987
988         dispatched = 0;
989         prev_cfqq = NULL;
990         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
991                 int max_dispatch;
992
993                 /*
994                  * Don't repeat dispatch from the previous queue.
995                  */
996                 if (prev_cfqq == cfqq)
997                         break;
998
999                 cfq_clear_cfqq_must_dispatch(cfqq);
1000                 cfq_clear_cfqq_wait_request(cfqq);
1001                 del_timer(&cfqd->idle_slice_timer);
1002
1003                 max_dispatch = cfqd->cfq_quantum;
1004                 if (cfq_class_idle(cfqq))
1005                         max_dispatch = 1;
1006
1007                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1008
1009                 /*
1010                  * If the dispatch cfqq has idling enabled and is still
1011                  * the active queue, break out.
1012                  */
1013                 if (cfq_cfqq_idle_window(cfqq) && cfqd->active_queue)
1014                         break;
1015
1016                 prev_cfqq = cfqq;
1017         }
1018
1019         return dispatched;
1020 }
1021
1022 /*
1023  * task holds one reference to the queue, dropped when task exits. each rq
1024  * in-flight on this queue also holds a reference, dropped when rq is freed.
1025  *
1026  * queue lock must be held here.
1027  */
1028 static void cfq_put_queue(struct cfq_queue *cfqq)
1029 {
1030         struct cfq_data *cfqd = cfqq->cfqd;
1031
1032         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1033
1034         if (!atomic_dec_and_test(&cfqq->ref))
1035                 return;
1036
1037         BUG_ON(rb_first(&cfqq->sort_list));
1038         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1039         BUG_ON(cfq_cfqq_on_rr(cfqq));
1040
1041         if (unlikely(cfqd->active_queue == cfqq))
1042                 __cfq_slice_expired(cfqd, cfqq, 0);
1043
1044         /*
1045          * it's on the empty list and still hashed
1046          */
1047         list_del(&cfqq->cfq_list);
1048         hlist_del(&cfqq->cfq_hash);
1049         kmem_cache_free(cfq_pool, cfqq);
1050 }
1051
1052 static inline struct cfq_queue *
1053 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1054                     const int hashval)
1055 {
1056         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1057         struct hlist_node *entry;
1058         struct cfq_queue *__cfqq;
1059
1060         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1061                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1062
1063                 if (__cfqq->key == key && (__p == prio || !prio))
1064                         return __cfqq;
1065         }
1066
1067         return NULL;
1068 }
1069
1070 static struct cfq_queue *
1071 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1072 {
1073         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1074 }
1075
1076 static void cfq_free_io_context(struct io_context *ioc)
1077 {
1078         struct cfq_io_context *__cic;
1079         struct rb_node *n;
1080         int freed = 0;
1081
1082         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1083                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1084                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1085                 kmem_cache_free(cfq_ioc_pool, __cic);
1086                 freed++;
1087         }
1088
1089         if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
1090                 complete(ioc_gone);
1091 }
1092
1093 static void cfq_trim(struct io_context *ioc)
1094 {
1095         ioc->set_ioprio = NULL;
1096         cfq_free_io_context(ioc);
1097 }
1098
1099 /*
1100  * Called with interrupts disabled
1101  */
1102 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1103 {
1104         struct cfq_data *cfqd = cic->key;
1105         request_queue_t *q;
1106
1107         if (!cfqd)
1108                 return;
1109
1110         q = cfqd->queue;
1111
1112         WARN_ON(!irqs_disabled());
1113
1114         spin_lock(q->queue_lock);
1115
1116         if (cic->cfqq[ASYNC]) {
1117                 if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue))
1118                         __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0);
1119                 cfq_put_queue(cic->cfqq[ASYNC]);
1120                 cic->cfqq[ASYNC] = NULL;
1121         }
1122
1123         if (cic->cfqq[SYNC]) {
1124                 if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue))
1125                         __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0);
1126                 cfq_put_queue(cic->cfqq[SYNC]);
1127                 cic->cfqq[SYNC] = NULL;
1128         }
1129
1130         cic->key = NULL;
1131         list_del_init(&cic->queue_list);
1132         spin_unlock(q->queue_lock);
1133 }
1134
1135 static void cfq_exit_io_context(struct io_context *ioc)
1136 {
1137         struct cfq_io_context *__cic;
1138         unsigned long flags;
1139         struct rb_node *n;
1140
1141         /*
1142          * put the reference this task is holding to the various queues
1143          */
1144         spin_lock_irqsave(&cfq_exit_lock, flags);
1145
1146         n = rb_first(&ioc->cic_root);
1147         while (n != NULL) {
1148                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1149
1150                 cfq_exit_single_io_context(__cic);
1151                 n = rb_next(n);
1152         }
1153
1154         spin_unlock_irqrestore(&cfq_exit_lock, flags);
1155 }
1156
1157 static struct cfq_io_context *
1158 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1159 {
1160         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1161
1162         if (cic) {
1163                 memset(cic, 0, sizeof(*cic));
1164                 cic->last_end_request = jiffies;
1165                 INIT_LIST_HEAD(&cic->queue_list);
1166                 cic->dtor = cfq_free_io_context;
1167                 cic->exit = cfq_exit_io_context;
1168                 atomic_inc(&ioc_count);
1169         }
1170
1171         return cic;
1172 }
1173
1174 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1175 {
1176         struct task_struct *tsk = current;
1177         int ioprio_class;
1178
1179         if (!cfq_cfqq_prio_changed(cfqq))
1180                 return;
1181
1182         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1183         switch (ioprio_class) {
1184                 default:
1185                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1186                 case IOPRIO_CLASS_NONE:
1187                         /*
1188                          * no prio set, place us in the middle of the BE classes
1189                          */
1190                         cfqq->ioprio = task_nice_ioprio(tsk);
1191                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1192                         break;
1193                 case IOPRIO_CLASS_RT:
1194                         cfqq->ioprio = task_ioprio(tsk);
1195                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1196                         break;
1197                 case IOPRIO_CLASS_BE:
1198                         cfqq->ioprio = task_ioprio(tsk);
1199                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1200                         break;
1201                 case IOPRIO_CLASS_IDLE:
1202                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1203                         cfqq->ioprio = 7;
1204                         cfq_clear_cfqq_idle_window(cfqq);
1205                         break;
1206         }
1207
1208         /*
1209          * keep track of original prio settings in case we have to temporarily
1210          * elevate the priority of this queue
1211          */
1212         cfqq->org_ioprio = cfqq->ioprio;
1213         cfqq->org_ioprio_class = cfqq->ioprio_class;
1214
1215         if (cfq_cfqq_on_rr(cfqq))
1216                 cfq_resort_rr_list(cfqq, 0);
1217
1218         cfq_clear_cfqq_prio_changed(cfqq);
1219 }
1220
1221 static inline void changed_ioprio(struct cfq_io_context *cic)
1222 {
1223         struct cfq_data *cfqd = cic->key;
1224         struct cfq_queue *cfqq;
1225
1226         if (unlikely(!cfqd))
1227                 return;
1228
1229         spin_lock(cfqd->queue->queue_lock);
1230
1231         cfqq = cic->cfqq[ASYNC];
1232         if (cfqq) {
1233                 struct cfq_queue *new_cfqq;
1234                 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1235                                          GFP_ATOMIC);
1236                 if (new_cfqq) {
1237                         cic->cfqq[ASYNC] = new_cfqq;
1238                         cfq_put_queue(cfqq);
1239                 }
1240         }
1241
1242         cfqq = cic->cfqq[SYNC];
1243         if (cfqq)
1244                 cfq_mark_cfqq_prio_changed(cfqq);
1245
1246         spin_unlock(cfqd->queue->queue_lock);
1247 }
1248
1249 /*
1250  * callback from sys_ioprio_set, irqs are disabled
1251  */
1252 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1253 {
1254         struct cfq_io_context *cic;
1255         struct rb_node *n;
1256
1257         spin_lock(&cfq_exit_lock);
1258
1259         n = rb_first(&ioc->cic_root);
1260         while (n != NULL) {
1261                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1262
1263                 changed_ioprio(cic);
1264                 n = rb_next(n);
1265         }
1266
1267         spin_unlock(&cfq_exit_lock);
1268
1269         return 0;
1270 }
1271
1272 static struct cfq_queue *
1273 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1274               gfp_t gfp_mask)
1275 {
1276         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1277         struct cfq_queue *cfqq, *new_cfqq = NULL;
1278         unsigned short ioprio;
1279
1280 retry:
1281         ioprio = tsk->ioprio;
1282         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1283
1284         if (!cfqq) {
1285                 if (new_cfqq) {
1286                         cfqq = new_cfqq;
1287                         new_cfqq = NULL;
1288                 } else if (gfp_mask & __GFP_WAIT) {
1289                         spin_unlock_irq(cfqd->queue->queue_lock);
1290                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1291                         spin_lock_irq(cfqd->queue->queue_lock);
1292                         goto retry;
1293                 } else {
1294                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1295                         if (!cfqq)
1296                                 goto out;
1297                 }
1298
1299                 memset(cfqq, 0, sizeof(*cfqq));
1300
1301                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1302                 INIT_LIST_HEAD(&cfqq->cfq_list);
1303                 INIT_LIST_HEAD(&cfqq->fifo);
1304
1305                 cfqq->key = key;
1306                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1307                 atomic_set(&cfqq->ref, 0);
1308                 cfqq->cfqd = cfqd;
1309                 cfqq->service_last = 0;
1310                 /*
1311                  * set ->slice_left to allow preemption for a new process
1312                  */
1313                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1314                 cfq_mark_cfqq_idle_window(cfqq);
1315                 cfq_mark_cfqq_prio_changed(cfqq);
1316                 cfq_init_prio_data(cfqq);
1317         }
1318
1319         if (new_cfqq)
1320                 kmem_cache_free(cfq_pool, new_cfqq);
1321
1322         atomic_inc(&cfqq->ref);
1323 out:
1324         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1325         return cfqq;
1326 }
1327
1328 static void
1329 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1330 {
1331         spin_lock(&cfq_exit_lock);
1332         rb_erase(&cic->rb_node, &ioc->cic_root);
1333         list_del_init(&cic->queue_list);
1334         spin_unlock(&cfq_exit_lock);
1335         kmem_cache_free(cfq_ioc_pool, cic);
1336         atomic_dec(&ioc_count);
1337 }
1338
1339 static struct cfq_io_context *
1340 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1341 {
1342         struct rb_node *n;
1343         struct cfq_io_context *cic;
1344         void *k, *key = cfqd;
1345
1346 restart:
1347         n = ioc->cic_root.rb_node;
1348         while (n) {
1349                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1350                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1351                 k = cic->key;
1352                 if (unlikely(!k)) {
1353                         cfq_drop_dead_cic(ioc, cic);
1354                         goto restart;
1355                 }
1356
1357                 if (key < k)
1358                         n = n->rb_left;
1359                 else if (key > k)
1360                         n = n->rb_right;
1361                 else
1362                         return cic;
1363         }
1364
1365         return NULL;
1366 }
1367
1368 static inline void
1369 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1370              struct cfq_io_context *cic)
1371 {
1372         struct rb_node **p;
1373         struct rb_node *parent;
1374         struct cfq_io_context *__cic;
1375         void *k;
1376
1377         cic->ioc = ioc;
1378         cic->key = cfqd;
1379
1380         ioc->set_ioprio = cfq_ioc_set_ioprio;
1381 restart:
1382         parent = NULL;
1383         p = &ioc->cic_root.rb_node;
1384         while (*p) {
1385                 parent = *p;
1386                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1387                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1388                 k = __cic->key;
1389                 if (unlikely(!k)) {
1390                         cfq_drop_dead_cic(ioc, __cic);
1391                         goto restart;
1392                 }
1393
1394                 if (cic->key < k)
1395                         p = &(*p)->rb_left;
1396                 else if (cic->key > k)
1397                         p = &(*p)->rb_right;
1398                 else
1399                         BUG();
1400         }
1401
1402         spin_lock(&cfq_exit_lock);
1403         rb_link_node(&cic->rb_node, parent, p);
1404         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1405         list_add(&cic->queue_list, &cfqd->cic_list);
1406         spin_unlock(&cfq_exit_lock);
1407 }
1408
1409 /*
1410  * Setup general io context and cfq io context. There can be several cfq
1411  * io contexts per general io context, if this process is doing io to more
1412  * than one device managed by cfq.
1413  */
1414 static struct cfq_io_context *
1415 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1416 {
1417         struct io_context *ioc = NULL;
1418         struct cfq_io_context *cic;
1419
1420         might_sleep_if(gfp_mask & __GFP_WAIT);
1421
1422         ioc = get_io_context(gfp_mask);
1423         if (!ioc)
1424                 return NULL;
1425
1426         cic = cfq_cic_rb_lookup(cfqd, ioc);
1427         if (cic)
1428                 goto out;
1429
1430         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1431         if (cic == NULL)
1432                 goto err;
1433
1434         cfq_cic_link(cfqd, ioc, cic);
1435 out:
1436         return cic;
1437 err:
1438         put_io_context(ioc);
1439         return NULL;
1440 }
1441
1442 static void
1443 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1444 {
1445         unsigned long elapsed, ttime;
1446
1447         /*
1448          * if this context already has stuff queued, thinktime is from
1449          * last queue not last end
1450          */
1451 #if 0
1452         if (time_after(cic->last_end_request, cic->last_queue))
1453                 elapsed = jiffies - cic->last_end_request;
1454         else
1455                 elapsed = jiffies - cic->last_queue;
1456 #else
1457                 elapsed = jiffies - cic->last_end_request;
1458 #endif
1459
1460         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1461
1462         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1463         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1464         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1465 }
1466
1467 static void
1468 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1469                        struct request *rq)
1470 {
1471         sector_t sdist;
1472         u64 total;
1473
1474         if (cic->last_request_pos < rq->sector)
1475                 sdist = rq->sector - cic->last_request_pos;
1476         else
1477                 sdist = cic->last_request_pos - rq->sector;
1478
1479         /*
1480          * Don't allow the seek distance to get too large from the
1481          * odd fragment, pagein, etc
1482          */
1483         if (cic->seek_samples <= 60) /* second&third seek */
1484                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1485         else
1486                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1487
1488         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1489         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1490         total = cic->seek_total + (cic->seek_samples/2);
1491         do_div(total, cic->seek_samples);
1492         cic->seek_mean = (sector_t)total;
1493 }
1494
1495 /*
1496  * Disable idle window if the process thinks too long or seeks so much that
1497  * it doesn't matter
1498  */
1499 static void
1500 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1501                        struct cfq_io_context *cic)
1502 {
1503         int enable_idle = cfq_cfqq_idle_window(cfqq);
1504
1505         if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1506             (cfqd->hw_tag && CIC_SEEKY(cic)))
1507                 enable_idle = 0;
1508         else if (sample_valid(cic->ttime_samples)) {
1509                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1510                         enable_idle = 0;
1511                 else
1512                         enable_idle = 1;
1513         }
1514
1515         if (enable_idle)
1516                 cfq_mark_cfqq_idle_window(cfqq);
1517         else
1518                 cfq_clear_cfqq_idle_window(cfqq);
1519 }
1520
1521
1522 /*
1523  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1524  * no or if we aren't sure, a 1 will cause a preempt.
1525  */
1526 static int
1527 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1528                    struct request *rq)
1529 {
1530         struct cfq_queue *cfqq = cfqd->active_queue;
1531
1532         if (cfq_class_idle(new_cfqq))
1533                 return 0;
1534
1535         if (!cfqq)
1536                 return 0;
1537
1538         if (cfq_class_idle(cfqq))
1539                 return 1;
1540         if (!cfq_cfqq_wait_request(new_cfqq))
1541                 return 0;
1542         /*
1543          * if it doesn't have slice left, forget it
1544          */
1545         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1546                 return 0;
1547         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1548                 return 1;
1549
1550         return 0;
1551 }
1552
1553 /*
1554  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1555  * let it have half of its nominal slice.
1556  */
1557 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1558 {
1559         struct cfq_queue *__cfqq, *next;
1560
1561         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1562                 cfq_resort_rr_list(__cfqq, 1);
1563
1564         if (!cfqq->slice_left)
1565                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1566
1567         cfqq->slice_end = cfqq->slice_left + jiffies;
1568         cfq_slice_expired(cfqd, 1);
1569         __cfq_set_active_queue(cfqd, cfqq);
1570 }
1571
1572 /*
1573  * should really be a ll_rw_blk.c helper
1574  */
1575 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1576 {
1577         request_queue_t *q = cfqd->queue;
1578
1579         if (!blk_queue_plugged(q))
1580                 q->request_fn(q);
1581         else
1582                 __generic_unplug_device(q);
1583 }
1584
1585 /*
1586  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1587  * something we should do about it
1588  */
1589 static void
1590 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1591                 struct request *rq)
1592 {
1593         struct cfq_io_context *cic = RQ_CIC(rq);
1594
1595         /*
1596          * check if this request is a better next-serve candidate)) {
1597          */
1598         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
1599         BUG_ON(!cfqq->next_rq);
1600
1601         /*
1602          * we never wait for an async request and we don't allow preemption
1603          * of an async request. so just return early
1604          */
1605         if (!rq_is_sync(rq)) {
1606                 /*
1607                  * sync process issued an async request, if it's waiting
1608                  * then expire it and kick rq handling.
1609                  */
1610                 if (cic == cfqd->active_cic &&
1611                     del_timer(&cfqd->idle_slice_timer)) {
1612                         cfq_slice_expired(cfqd, 0);
1613                         cfq_start_queueing(cfqd, cfqq);
1614                 }
1615                 return;
1616         }
1617
1618         cfq_update_io_thinktime(cfqd, cic);
1619         cfq_update_io_seektime(cfqd, cic, rq);
1620         cfq_update_idle_window(cfqd, cfqq, cic);
1621
1622         cic->last_queue = jiffies;
1623         cic->last_request_pos = rq->sector + rq->nr_sectors;
1624
1625         if (cfqq == cfqd->active_queue) {
1626                 /*
1627                  * if we are waiting for a request for this queue, let it rip
1628                  * immediately and flag that we must not expire this queue
1629                  * just now
1630                  */
1631                 if (cfq_cfqq_wait_request(cfqq)) {
1632                         cfq_mark_cfqq_must_dispatch(cfqq);
1633                         del_timer(&cfqd->idle_slice_timer);
1634                         cfq_start_queueing(cfqd, cfqq);
1635                 }
1636         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1637                 /*
1638                  * not the active queue - expire current slice if it is
1639                  * idle and has expired it's mean thinktime or this new queue
1640                  * has some old slice time left and is of higher priority
1641                  */
1642                 cfq_preempt_queue(cfqd, cfqq);
1643                 cfq_mark_cfqq_must_dispatch(cfqq);
1644                 cfq_start_queueing(cfqd, cfqq);
1645         }
1646 }
1647
1648 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1649 {
1650         struct cfq_data *cfqd = q->elevator->elevator_data;
1651         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1652
1653         cfq_init_prio_data(cfqq);
1654
1655         cfq_add_rq_rb(rq);
1656
1657         if (!cfq_cfqq_on_rr(cfqq))
1658                 cfq_add_cfqq_rr(cfqd, cfqq);
1659
1660         list_add_tail(&rq->queuelist, &cfqq->fifo);
1661
1662         cfq_rq_enqueued(cfqd, cfqq, rq);
1663 }
1664
1665 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1666 {
1667         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1668         struct cfq_data *cfqd = cfqq->cfqd;
1669         const int sync = rq_is_sync(rq);
1670         unsigned long now;
1671
1672         now = jiffies;
1673
1674         WARN_ON(!cfqd->rq_in_driver);
1675         WARN_ON(!cfqq->on_dispatch[sync]);
1676         cfqd->rq_in_driver--;
1677         cfqq->on_dispatch[sync]--;
1678
1679         if (!cfq_class_idle(cfqq))
1680                 cfqd->last_end_request = now;
1681
1682         if (!cfq_cfqq_dispatched(cfqq)) {
1683                 if (cfq_cfqq_on_rr(cfqq)) {
1684                         cfqq->service_last = now;
1685                         cfq_resort_rr_list(cfqq, 0);
1686                 }
1687         }
1688
1689         if (sync)
1690                 RQ_CIC(rq)->last_end_request = now;
1691
1692         /*
1693          * If this is the active queue, check if it needs to be expired,
1694          * or if we want to idle in case it has no pending requests.
1695          */
1696         if (cfqd->active_queue == cfqq) {
1697                 if (time_after(now, cfqq->slice_end))
1698                         cfq_slice_expired(cfqd, 0);
1699                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1700                         if (!cfq_arm_slice_timer(cfqd, cfqq))
1701                                 cfq_schedule_dispatch(cfqd);
1702                 }
1703         }
1704 }
1705
1706 /*
1707  * we temporarily boost lower priority queues if they are holding fs exclusive
1708  * resources. they are boosted to normal prio (CLASS_BE/4)
1709  */
1710 static void cfq_prio_boost(struct cfq_queue *cfqq)
1711 {
1712         const int ioprio_class = cfqq->ioprio_class;
1713         const int ioprio = cfqq->ioprio;
1714
1715         if (has_fs_excl()) {
1716                 /*
1717                  * boost idle prio on transactions that would lock out other
1718                  * users of the filesystem
1719                  */
1720                 if (cfq_class_idle(cfqq))
1721                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1722                 if (cfqq->ioprio > IOPRIO_NORM)
1723                         cfqq->ioprio = IOPRIO_NORM;
1724         } else {
1725                 /*
1726                  * check if we need to unboost the queue
1727                  */
1728                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1729                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1730                 if (cfqq->ioprio != cfqq->org_ioprio)
1731                         cfqq->ioprio = cfqq->org_ioprio;
1732         }
1733
1734         /*
1735          * refile between round-robin lists if we moved the priority class
1736          */
1737         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1738             cfq_cfqq_on_rr(cfqq))
1739                 cfq_resort_rr_list(cfqq, 0);
1740 }
1741
1742 static inline int
1743 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1744                 struct task_struct *task, int rw)
1745 {
1746         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1747             !cfq_cfqq_must_alloc_slice(cfqq)) {
1748                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1749                 return ELV_MQUEUE_MUST;
1750         }
1751
1752         return ELV_MQUEUE_MAY;
1753 }
1754
1755 static int cfq_may_queue(request_queue_t *q, int rw)
1756 {
1757         struct cfq_data *cfqd = q->elevator->elevator_data;
1758         struct task_struct *tsk = current;
1759         struct cfq_queue *cfqq;
1760
1761         /*
1762          * don't force setup of a queue from here, as a call to may_queue
1763          * does not necessarily imply that a request actually will be queued.
1764          * so just lookup a possibly existing queue, or return 'may queue'
1765          * if that fails
1766          */
1767         cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1768         if (cfqq) {
1769                 cfq_init_prio_data(cfqq);
1770                 cfq_prio_boost(cfqq);
1771
1772                 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1773         }
1774
1775         return ELV_MQUEUE_MAY;
1776 }
1777
1778 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1779 {
1780         struct cfq_data *cfqd = q->elevator->elevator_data;
1781
1782         if (unlikely(cfqd->rq_starved)) {
1783                 struct request_list *rl = &q->rq;
1784
1785                 smp_mb();
1786                 if (waitqueue_active(&rl->wait[READ]))
1787                         wake_up(&rl->wait[READ]);
1788                 if (waitqueue_active(&rl->wait[WRITE]))
1789                         wake_up(&rl->wait[WRITE]);
1790         }
1791 }
1792
1793 /*
1794  * queue lock held here
1795  */
1796 static void cfq_put_request(request_queue_t *q, struct request *rq)
1797 {
1798         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1799
1800         if (cfqq) {
1801                 const int rw = rq_data_dir(rq);
1802
1803                 BUG_ON(!cfqq->allocated[rw]);
1804                 cfqq->allocated[rw]--;
1805
1806                 put_io_context(RQ_CIC(rq)->ioc);
1807
1808                 rq->elevator_private = NULL;
1809                 rq->elevator_private2 = NULL;
1810
1811                 cfq_check_waiters(q, cfqq);
1812                 cfq_put_queue(cfqq);
1813         }
1814 }
1815
1816 /*
1817  * Allocate cfq data structures associated with this request.
1818  */
1819 static int
1820 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1821 {
1822         struct cfq_data *cfqd = q->elevator->elevator_data;
1823         struct task_struct *tsk = current;
1824         struct cfq_io_context *cic;
1825         const int rw = rq_data_dir(rq);
1826         pid_t key = cfq_queue_pid(tsk, rw);
1827         struct cfq_queue *cfqq;
1828         unsigned long flags;
1829         int is_sync = key != CFQ_KEY_ASYNC;
1830
1831         might_sleep_if(gfp_mask & __GFP_WAIT);
1832
1833         cic = cfq_get_io_context(cfqd, gfp_mask);
1834
1835         spin_lock_irqsave(q->queue_lock, flags);
1836
1837         if (!cic)
1838                 goto queue_fail;
1839
1840         if (!cic->cfqq[is_sync]) {
1841                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1842                 if (!cfqq)
1843                         goto queue_fail;
1844
1845                 cic->cfqq[is_sync] = cfqq;
1846         } else
1847                 cfqq = cic->cfqq[is_sync];
1848
1849         cfqq->allocated[rw]++;
1850         cfq_clear_cfqq_must_alloc(cfqq);
1851         cfqd->rq_starved = 0;
1852         atomic_inc(&cfqq->ref);
1853
1854         spin_unlock_irqrestore(q->queue_lock, flags);
1855
1856         rq->elevator_private = cic;
1857         rq->elevator_private2 = cfqq;
1858         return 0;
1859
1860 queue_fail:
1861         if (cic)
1862                 put_io_context(cic->ioc);
1863         /*
1864          * mark us rq allocation starved. we need to kickstart the process
1865          * ourselves if there are no pending requests that can do it for us.
1866          * that would be an extremely rare OOM situation
1867          */
1868         cfqd->rq_starved = 1;
1869         cfq_schedule_dispatch(cfqd);
1870         spin_unlock_irqrestore(q->queue_lock, flags);
1871         return 1;
1872 }
1873
1874 static void cfq_kick_queue(void *data)
1875 {
1876         request_queue_t *q = data;
1877         struct cfq_data *cfqd = q->elevator->elevator_data;
1878         unsigned long flags;
1879
1880         spin_lock_irqsave(q->queue_lock, flags);
1881
1882         if (cfqd->rq_starved) {
1883                 struct request_list *rl = &q->rq;
1884
1885                 /*
1886                  * we aren't guaranteed to get a request after this, but we
1887                  * have to be opportunistic
1888                  */
1889                 smp_mb();
1890                 if (waitqueue_active(&rl->wait[READ]))
1891                         wake_up(&rl->wait[READ]);
1892                 if (waitqueue_active(&rl->wait[WRITE]))
1893                         wake_up(&rl->wait[WRITE]);
1894         }
1895
1896         blk_remove_plug(q);
1897         q->request_fn(q);
1898         spin_unlock_irqrestore(q->queue_lock, flags);
1899 }
1900
1901 /*
1902  * Timer running if the active_queue is currently idling inside its time slice
1903  */
1904 static void cfq_idle_slice_timer(unsigned long data)
1905 {
1906         struct cfq_data *cfqd = (struct cfq_data *) data;
1907         struct cfq_queue *cfqq;
1908         unsigned long flags;
1909
1910         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1911
1912         if ((cfqq = cfqd->active_queue) != NULL) {
1913                 unsigned long now = jiffies;
1914
1915                 /*
1916                  * expired
1917                  */
1918                 if (time_after(now, cfqq->slice_end))
1919                         goto expire;
1920
1921                 /*
1922                  * only expire and reinvoke request handler, if there are
1923                  * other queues with pending requests
1924                  */
1925                 if (!cfqd->busy_queues)
1926                         goto out_cont;
1927
1928                 /*
1929                  * not expired and it has a request pending, let it dispatch
1930                  */
1931                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1932                         cfq_mark_cfqq_must_dispatch(cfqq);
1933                         goto out_kick;
1934                 }
1935         }
1936 expire:
1937         cfq_slice_expired(cfqd, 0);
1938 out_kick:
1939         cfq_schedule_dispatch(cfqd);
1940 out_cont:
1941         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1942 }
1943
1944 /*
1945  * Timer running if an idle class queue is waiting for service
1946  */
1947 static void cfq_idle_class_timer(unsigned long data)
1948 {
1949         struct cfq_data *cfqd = (struct cfq_data *) data;
1950         unsigned long flags, end;
1951
1952         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1953
1954         /*
1955          * race with a non-idle queue, reset timer
1956          */
1957         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
1958         if (!time_after_eq(jiffies, end))
1959                 mod_timer(&cfqd->idle_class_timer, end);
1960         else
1961                 cfq_schedule_dispatch(cfqd);
1962
1963         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1964 }
1965
1966 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1967 {
1968         del_timer_sync(&cfqd->idle_slice_timer);
1969         del_timer_sync(&cfqd->idle_class_timer);
1970         blk_sync_queue(cfqd->queue);
1971 }
1972
1973 static void cfq_exit_queue(elevator_t *e)
1974 {
1975         struct cfq_data *cfqd = e->elevator_data;
1976         request_queue_t *q = cfqd->queue;
1977
1978         cfq_shutdown_timer_wq(cfqd);
1979
1980         spin_lock(&cfq_exit_lock);
1981         spin_lock_irq(q->queue_lock);
1982
1983         if (cfqd->active_queue)
1984                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
1985
1986         while (!list_empty(&cfqd->cic_list)) {
1987                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1988                                                         struct cfq_io_context,
1989                                                         queue_list);
1990                 if (cic->cfqq[ASYNC]) {
1991                         cfq_put_queue(cic->cfqq[ASYNC]);
1992                         cic->cfqq[ASYNC] = NULL;
1993                 }
1994                 if (cic->cfqq[SYNC]) {
1995                         cfq_put_queue(cic->cfqq[SYNC]);
1996                         cic->cfqq[SYNC] = NULL;
1997                 }
1998                 cic->key = NULL;
1999                 list_del_init(&cic->queue_list);
2000         }
2001
2002         spin_unlock_irq(q->queue_lock);
2003         spin_unlock(&cfq_exit_lock);
2004
2005         cfq_shutdown_timer_wq(cfqd);
2006
2007         kfree(cfqd->cfq_hash);
2008         kfree(cfqd);
2009 }
2010
2011 static void *cfq_init_queue(request_queue_t *q, elevator_t *e)
2012 {
2013         struct cfq_data *cfqd;
2014         int i;
2015
2016         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2017         if (!cfqd)
2018                 return NULL;
2019
2020         memset(cfqd, 0, sizeof(*cfqd));
2021
2022         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2023                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2024
2025         INIT_LIST_HEAD(&cfqd->busy_rr);
2026         INIT_LIST_HEAD(&cfqd->cur_rr);
2027         INIT_LIST_HEAD(&cfqd->idle_rr);
2028         INIT_LIST_HEAD(&cfqd->empty_list);
2029         INIT_LIST_HEAD(&cfqd->cic_list);
2030
2031         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2032         if (!cfqd->cfq_hash)
2033                 goto out_free;
2034
2035         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2036                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2037
2038         cfqd->queue = q;
2039
2040         init_timer(&cfqd->idle_slice_timer);
2041         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2042         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2043
2044         init_timer(&cfqd->idle_class_timer);
2045         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2046         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2047
2048         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2049
2050         cfqd->cfq_queued = cfq_queued;
2051         cfqd->cfq_quantum = cfq_quantum;
2052         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2053         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2054         cfqd->cfq_back_max = cfq_back_max;
2055         cfqd->cfq_back_penalty = cfq_back_penalty;
2056         cfqd->cfq_slice[0] = cfq_slice_async;
2057         cfqd->cfq_slice[1] = cfq_slice_sync;
2058         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2059         cfqd->cfq_slice_idle = cfq_slice_idle;
2060
2061         return cfqd;
2062 out_free:
2063         kfree(cfqd);
2064         return NULL;
2065 }
2066
2067 static void cfq_slab_kill(void)
2068 {
2069         if (cfq_pool)
2070                 kmem_cache_destroy(cfq_pool);
2071         if (cfq_ioc_pool)
2072                 kmem_cache_destroy(cfq_ioc_pool);
2073 }
2074
2075 static int __init cfq_slab_setup(void)
2076 {
2077         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2078                                         NULL, NULL);
2079         if (!cfq_pool)
2080                 goto fail;
2081
2082         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2083                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2084         if (!cfq_ioc_pool)
2085                 goto fail;
2086
2087         return 0;
2088 fail:
2089         cfq_slab_kill();
2090         return -ENOMEM;
2091 }
2092
2093 /*
2094  * sysfs parts below -->
2095  */
2096
2097 static ssize_t
2098 cfq_var_show(unsigned int var, char *page)
2099 {
2100         return sprintf(page, "%d\n", var);
2101 }
2102
2103 static ssize_t
2104 cfq_var_store(unsigned int *var, const char *page, size_t count)
2105 {
2106         char *p = (char *) page;
2107
2108         *var = simple_strtoul(p, &p, 10);
2109         return count;
2110 }
2111
2112 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2113 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2114 {                                                                       \
2115         struct cfq_data *cfqd = e->elevator_data;                       \
2116         unsigned int __data = __VAR;                                    \
2117         if (__CONV)                                                     \
2118                 __data = jiffies_to_msecs(__data);                      \
2119         return cfq_var_show(__data, (page));                            \
2120 }
2121 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2122 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2123 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2124 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2125 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2126 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2127 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2128 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2129 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2130 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2131 #undef SHOW_FUNCTION
2132
2133 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2134 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2135 {                                                                       \
2136         struct cfq_data *cfqd = e->elevator_data;                       \
2137         unsigned int __data;                                            \
2138         int ret = cfq_var_store(&__data, (page), count);                \
2139         if (__data < (MIN))                                             \
2140                 __data = (MIN);                                         \
2141         else if (__data > (MAX))                                        \
2142                 __data = (MAX);                                         \
2143         if (__CONV)                                                     \
2144                 *(__PTR) = msecs_to_jiffies(__data);                    \
2145         else                                                            \
2146                 *(__PTR) = __data;                                      \
2147         return ret;                                                     \
2148 }
2149 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2150 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2151 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2152 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2153 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2154 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2155 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2156 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2157 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2158 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2159 #undef STORE_FUNCTION
2160
2161 #define CFQ_ATTR(name) \
2162         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2163
2164 static struct elv_fs_entry cfq_attrs[] = {
2165         CFQ_ATTR(quantum),
2166         CFQ_ATTR(queued),
2167         CFQ_ATTR(fifo_expire_sync),
2168         CFQ_ATTR(fifo_expire_async),
2169         CFQ_ATTR(back_seek_max),
2170         CFQ_ATTR(back_seek_penalty),
2171         CFQ_ATTR(slice_sync),
2172         CFQ_ATTR(slice_async),
2173         CFQ_ATTR(slice_async_rq),
2174         CFQ_ATTR(slice_idle),
2175         __ATTR_NULL
2176 };
2177
2178 static struct elevator_type iosched_cfq = {
2179         .ops = {
2180                 .elevator_merge_fn =            cfq_merge,
2181                 .elevator_merged_fn =           cfq_merged_request,
2182                 .elevator_merge_req_fn =        cfq_merged_requests,
2183                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2184                 .elevator_add_req_fn =          cfq_insert_request,
2185                 .elevator_activate_req_fn =     cfq_activate_request,
2186                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2187                 .elevator_queue_empty_fn =      cfq_queue_empty,
2188                 .elevator_completed_req_fn =    cfq_completed_request,
2189                 .elevator_former_req_fn =       elv_rb_former_request,
2190                 .elevator_latter_req_fn =       elv_rb_latter_request,
2191                 .elevator_set_req_fn =          cfq_set_request,
2192                 .elevator_put_req_fn =          cfq_put_request,
2193                 .elevator_may_queue_fn =        cfq_may_queue,
2194                 .elevator_init_fn =             cfq_init_queue,
2195                 .elevator_exit_fn =             cfq_exit_queue,
2196                 .trim =                         cfq_trim,
2197         },
2198         .elevator_attrs =       cfq_attrs,
2199         .elevator_name =        "cfq",
2200         .elevator_owner =       THIS_MODULE,
2201 };
2202
2203 static int __init cfq_init(void)
2204 {
2205         int ret;
2206
2207         /*
2208          * could be 0 on HZ < 1000 setups
2209          */
2210         if (!cfq_slice_async)
2211                 cfq_slice_async = 1;
2212         if (!cfq_slice_idle)
2213                 cfq_slice_idle = 1;
2214
2215         if (cfq_slab_setup())
2216                 return -ENOMEM;
2217
2218         ret = elv_register(&iosched_cfq);
2219         if (ret)
2220                 cfq_slab_kill();
2221
2222         return ret;
2223 }
2224
2225 static void __exit cfq_exit(void)
2226 {
2227         DECLARE_COMPLETION(all_gone);
2228         elv_unregister(&iosched_cfq);
2229         ioc_gone = &all_gone;
2230         /* ioc_gone's update must be visible before reading ioc_count */
2231         smp_wmb();
2232         if (atomic_read(&ioc_count))
2233                 wait_for_completion(ioc_gone);
2234         synchronize_rcu();
2235         cfq_slab_kill();
2236 }
2237
2238 module_init(cfq_init);
2239 module_exit(cfq_exit);
2240
2241 MODULE_AUTHOR("Jens Axboe");
2242 MODULE_LICENSE("GPL");
2243 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");