Merge branch 'for-linus' of ssh://master.kernel.org/pub/scm/linux/kernel/git/ieee1394...
[linux-2.6] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/hash.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15
16 /*
17  * tunables
18  */
19 static const int cfq_quantum = 4;               /* max queue in one round of service */
20 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
21 static const int cfq_back_max = 16 * 1024;      /* maximum backwards seek, in KiB */
22 static const int cfq_back_penalty = 2;          /* penalty of a backwards seek */
23
24 static const int cfq_slice_sync = HZ / 10;
25 static int cfq_slice_async = HZ / 25;
26 static const int cfq_slice_async_rq = 2;
27 static int cfq_slice_idle = HZ / 125;
28
29 #define CFQ_IDLE_GRACE          (HZ / 10)
30 #define CFQ_SLICE_SCALE         (5)
31
32 #define CFQ_KEY_ASYNC           (0)
33
34 /*
35  * for the hash of cfqq inside the cfqd
36  */
37 #define CFQ_QHASH_SHIFT         6
38 #define CFQ_QHASH_ENTRIES       (1 << CFQ_QHASH_SHIFT)
39 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
40
41 #define list_entry_cfqq(ptr)    list_entry((ptr), struct cfq_queue, cfq_list)
42
43 #define RQ_CIC(rq)              ((struct cfq_io_context*)(rq)->elevator_private)
44 #define RQ_CFQQ(rq)             ((rq)->elevator_private2)
45
46 static struct kmem_cache *cfq_pool;
47 static struct kmem_cache *cfq_ioc_pool;
48
49 static DEFINE_PER_CPU(unsigned long, ioc_count);
50 static struct completion *ioc_gone;
51
52 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
53 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
54 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
55
56 #define ASYNC                   (0)
57 #define SYNC                    (1)
58
59 #define cfq_cfqq_dispatched(cfqq)       \
60         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
61
62 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
63
64 #define cfq_cfqq_sync(cfqq)             \
65         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
66
67 #define sample_valid(samples)   ((samples) > 80)
68
69 /*
70  * Per block device queue structure
71  */
72 struct cfq_data {
73         request_queue_t *queue;
74
75         /*
76          * rr list of queues with requests and the count of them
77          */
78         struct list_head rr_list[CFQ_PRIO_LISTS];
79         struct list_head busy_rr;
80         struct list_head cur_rr;
81         struct list_head idle_rr;
82         unsigned int busy_queues;
83
84         /*
85          * cfqq lookup hash
86          */
87         struct hlist_head *cfq_hash;
88
89         int rq_in_driver;
90         int hw_tag;
91
92         /*
93          * idle window management
94          */
95         struct timer_list idle_slice_timer;
96         struct work_struct unplug_work;
97
98         struct cfq_queue *active_queue;
99         struct cfq_io_context *active_cic;
100         int cur_prio, cur_end_prio;
101         unsigned int dispatch_slice;
102
103         struct timer_list idle_class_timer;
104
105         sector_t last_sector;
106         unsigned long last_end_request;
107
108         /*
109          * tunables, see top of file
110          */
111         unsigned int cfq_quantum;
112         unsigned int cfq_fifo_expire[2];
113         unsigned int cfq_back_penalty;
114         unsigned int cfq_back_max;
115         unsigned int cfq_slice[2];
116         unsigned int cfq_slice_async_rq;
117         unsigned int cfq_slice_idle;
118
119         struct list_head cic_list;
120 };
121
122 /*
123  * Per process-grouping structure
124  */
125 struct cfq_queue {
126         /* reference count */
127         atomic_t ref;
128         /* parent cfq_data */
129         struct cfq_data *cfqd;
130         /* cfqq lookup hash */
131         struct hlist_node cfq_hash;
132         /* hash key */
133         unsigned int key;
134         /* member of the rr/busy/cur/idle cfqd list */
135         struct list_head cfq_list;
136         /* sorted list of pending requests */
137         struct rb_root sort_list;
138         /* if fifo isn't expired, next request to serve */
139         struct request *next_rq;
140         /* requests queued in sort_list */
141         int queued[2];
142         /* currently allocated requests */
143         int allocated[2];
144         /* pending metadata requests */
145         int meta_pending;
146         /* fifo list of requests in sort_list */
147         struct list_head fifo;
148
149         unsigned long slice_end;
150         unsigned long service_last;
151         long slice_resid;
152
153         /* number of requests that are on the dispatch list */
154         int on_dispatch[2];
155
156         /* io prio of this group */
157         unsigned short ioprio, org_ioprio;
158         unsigned short ioprio_class, org_ioprio_class;
159
160         /* various state flags, see below */
161         unsigned int flags;
162 };
163
164 enum cfqq_state_flags {
165         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
166         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
167         CFQ_CFQQ_FLAG_must_alloc,       /* must be allowed rq alloc */
168         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
169         CFQ_CFQQ_FLAG_must_dispatch,    /* must dispatch, even if expired */
170         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
171         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
172         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
173         CFQ_CFQQ_FLAG_queue_new,        /* queue never been serviced */
174         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
175 };
176
177 #define CFQ_CFQQ_FNS(name)                                              \
178 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
179 {                                                                       \
180         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
181 }                                                                       \
182 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
183 {                                                                       \
184         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
185 }                                                                       \
186 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
187 {                                                                       \
188         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
189 }
190
191 CFQ_CFQQ_FNS(on_rr);
192 CFQ_CFQQ_FNS(wait_request);
193 CFQ_CFQQ_FNS(must_alloc);
194 CFQ_CFQQ_FNS(must_alloc_slice);
195 CFQ_CFQQ_FNS(must_dispatch);
196 CFQ_CFQQ_FNS(fifo_expire);
197 CFQ_CFQQ_FNS(idle_window);
198 CFQ_CFQQ_FNS(prio_changed);
199 CFQ_CFQQ_FNS(queue_new);
200 CFQ_CFQQ_FNS(slice_new);
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, int is_sync)
225 {
226         /*
227          * Use the per-process queue, for read requests and syncronous writes
228          */
229         if (!(rw & REQ_RW) || is_sync)
230                 return task->pid;
231
232         return CFQ_KEY_ASYNC;
233 }
234
235 /*
236  * Scale schedule slice based on io priority. Use the sync time slice only
237  * if a queue is marked sync and has sync io queued. A sync queue with async
238  * io only, should not get full sync slice length.
239  */
240 static inline int
241 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
242 {
243         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
244
245         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
246
247         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
248 }
249
250 static inline void
251 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
252 {
253         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
254         cfqq->slice_end += cfqq->slice_resid;
255
256         /*
257          * Don't carry over residual for more than one slice, we only want
258          * to slightly correct the fairness. Carrying over forever would
259          * easily introduce oscillations.
260          */
261         cfqq->slice_resid = 0;
262 }
263
264 /*
265  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
266  * isn't valid until the first request from the dispatch is activated
267  * and the slice time set.
268  */
269 static inline int cfq_slice_used(struct cfq_queue *cfqq)
270 {
271         if (cfq_cfqq_slice_new(cfqq))
272                 return 0;
273         if (time_before(jiffies, cfqq->slice_end))
274                 return 0;
275
276         return 1;
277 }
278
279 /*
280  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
281  * We choose the request that is closest to the head right now. Distance
282  * behind the head is penalized and only allowed to a certain extent.
283  */
284 static struct request *
285 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
286 {
287         sector_t last, s1, s2, d1 = 0, d2 = 0;
288         unsigned long back_max;
289 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
290 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
291         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
292
293         if (rq1 == NULL || rq1 == rq2)
294                 return rq2;
295         if (rq2 == NULL)
296                 return rq1;
297
298         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
299                 return rq1;
300         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
301                 return rq2;
302         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
303                 return rq1;
304         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
305                 return rq2;
306
307         s1 = rq1->sector;
308         s2 = rq2->sector;
309
310         last = cfqd->last_sector;
311
312         /*
313          * by definition, 1KiB is 2 sectors
314          */
315         back_max = cfqd->cfq_back_max * 2;
316
317         /*
318          * Strict one way elevator _except_ in the case where we allow
319          * short backward seeks which are biased as twice the cost of a
320          * similar forward seek.
321          */
322         if (s1 >= last)
323                 d1 = s1 - last;
324         else if (s1 + back_max >= last)
325                 d1 = (last - s1) * cfqd->cfq_back_penalty;
326         else
327                 wrap |= CFQ_RQ1_WRAP;
328
329         if (s2 >= last)
330                 d2 = s2 - last;
331         else if (s2 + back_max >= last)
332                 d2 = (last - s2) * cfqd->cfq_back_penalty;
333         else
334                 wrap |= CFQ_RQ2_WRAP;
335
336         /* Found required data */
337
338         /*
339          * By doing switch() on the bit mask "wrap" we avoid having to
340          * check two variables for all permutations: --> faster!
341          */
342         switch (wrap) {
343         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
344                 if (d1 < d2)
345                         return rq1;
346                 else if (d2 < d1)
347                         return rq2;
348                 else {
349                         if (s1 >= s2)
350                                 return rq1;
351                         else
352                                 return rq2;
353                 }
354
355         case CFQ_RQ2_WRAP:
356                 return rq1;
357         case CFQ_RQ1_WRAP:
358                 return rq2;
359         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
360         default:
361                 /*
362                  * Since both rqs are wrapped,
363                  * start with the one that's further behind head
364                  * (--> only *one* back seek required),
365                  * since back seek takes more time than forward.
366                  */
367                 if (s1 <= s2)
368                         return rq1;
369                 else
370                         return rq2;
371         }
372 }
373
374 /*
375  * would be nice to take fifo expire time into account as well
376  */
377 static struct request *
378 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
379                   struct request *last)
380 {
381         struct rb_node *rbnext = rb_next(&last->rb_node);
382         struct rb_node *rbprev = rb_prev(&last->rb_node);
383         struct request *next = NULL, *prev = NULL;
384
385         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
386
387         if (rbprev)
388                 prev = rb_entry_rq(rbprev);
389
390         if (rbnext)
391                 next = rb_entry_rq(rbnext);
392         else {
393                 rbnext = rb_first(&cfqq->sort_list);
394                 if (rbnext && rbnext != &last->rb_node)
395                         next = rb_entry_rq(rbnext);
396         }
397
398         return cfq_choose_req(cfqd, next, prev);
399 }
400
401 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
402 {
403         struct cfq_data *cfqd = cfqq->cfqd;
404         struct list_head *list, *n;
405         struct cfq_queue *__cfqq;
406
407         /*
408          * Resorting requires the cfqq to be on the RR list already.
409          */
410         if (!cfq_cfqq_on_rr(cfqq))
411                 return;
412
413         list_del(&cfqq->cfq_list);
414
415         if (cfq_class_rt(cfqq))
416                 list = &cfqd->cur_rr;
417         else if (cfq_class_idle(cfqq))
418                 list = &cfqd->idle_rr;
419         else {
420                 /*
421                  * if cfqq has requests in flight, don't allow it to be
422                  * found in cfq_set_active_queue before it has finished them.
423                  * this is done to increase fairness between a process that
424                  * has lots of io pending vs one that only generates one
425                  * sporadically or synchronously
426                  */
427                 if (cfq_cfqq_dispatched(cfqq))
428                         list = &cfqd->busy_rr;
429                 else
430                         list = &cfqd->rr_list[cfqq->ioprio];
431         }
432
433         if (preempted || cfq_cfqq_queue_new(cfqq)) {
434                 /*
435                  * If this queue was preempted or is new (never been serviced),
436                  * let it be added first for fairness but beind other new
437                  * queues.
438                  */
439                 n = list;
440                 while (n->next != list) {
441                         __cfqq = list_entry_cfqq(n->next);
442                         if (!cfq_cfqq_queue_new(__cfqq))
443                                 break;
444
445                         n = n->next;
446                 }
447                 list_add_tail(&cfqq->cfq_list, n);
448         } else if (!cfq_cfqq_class_sync(cfqq)) {
449                 /*
450                  * async queue always goes to the end. this wont be overly
451                  * unfair to writes, as the sort of the sync queue wont be
452                  * allowed to pass the async queue again.
453                  */
454                 list_add_tail(&cfqq->cfq_list, list);
455         } else {
456                 /*
457                  * sort by last service, but don't cross a new or async
458                  * queue. we don't cross a new queue because it hasn't been
459                  * service before, and we don't cross an async queue because
460                  * it gets added to the end on expire.
461                  */
462                 n = list;
463                 while ((n = n->prev) != list) {
464                         struct cfq_queue *__cfqq = list_entry_cfqq(n);
465
466                         if (!cfq_cfqq_class_sync(cfqq) || !__cfqq->service_last)
467                                 break;
468                         if (time_before(__cfqq->service_last, cfqq->service_last))
469                                 break;
470                 }
471                 list_add(&cfqq->cfq_list, n);
472         }
473 }
474
475 /*
476  * add to busy list of queues for service, trying to be fair in ordering
477  * the pending list according to last request service
478  */
479 static inline void
480 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
481 {
482         BUG_ON(cfq_cfqq_on_rr(cfqq));
483         cfq_mark_cfqq_on_rr(cfqq);
484         cfqd->busy_queues++;
485
486         cfq_resort_rr_list(cfqq, 0);
487 }
488
489 static inline void
490 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
491 {
492         BUG_ON(!cfq_cfqq_on_rr(cfqq));
493         cfq_clear_cfqq_on_rr(cfqq);
494         list_del_init(&cfqq->cfq_list);
495
496         BUG_ON(!cfqd->busy_queues);
497         cfqd->busy_queues--;
498 }
499
500 /*
501  * rb tree support functions
502  */
503 static inline void cfq_del_rq_rb(struct request *rq)
504 {
505         struct cfq_queue *cfqq = RQ_CFQQ(rq);
506         struct cfq_data *cfqd = cfqq->cfqd;
507         const int sync = rq_is_sync(rq);
508
509         BUG_ON(!cfqq->queued[sync]);
510         cfqq->queued[sync]--;
511
512         elv_rb_del(&cfqq->sort_list, rq);
513
514         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
515                 cfq_del_cfqq_rr(cfqd, cfqq);
516 }
517
518 static void cfq_add_rq_rb(struct request *rq)
519 {
520         struct cfq_queue *cfqq = RQ_CFQQ(rq);
521         struct cfq_data *cfqd = cfqq->cfqd;
522         struct request *__alias;
523
524         cfqq->queued[rq_is_sync(rq)]++;
525
526         /*
527          * looks a little odd, but the first insert might return an alias.
528          * if that happens, put the alias on the dispatch list
529          */
530         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
531                 cfq_dispatch_insert(cfqd->queue, __alias);
532
533         if (!cfq_cfqq_on_rr(cfqq))
534                 cfq_add_cfqq_rr(cfqd, cfqq);
535 }
536
537 static inline void
538 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
539 {
540         elv_rb_del(&cfqq->sort_list, rq);
541         cfqq->queued[rq_is_sync(rq)]--;
542         cfq_add_rq_rb(rq);
543 }
544
545 static struct request *
546 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
547 {
548         struct task_struct *tsk = current;
549         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio));
550         struct cfq_queue *cfqq;
551
552         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
553         if (cfqq) {
554                 sector_t sector = bio->bi_sector + bio_sectors(bio);
555
556                 return elv_rb_find(&cfqq->sort_list, sector);
557         }
558
559         return NULL;
560 }
561
562 static void cfq_activate_request(request_queue_t *q, struct request *rq)
563 {
564         struct cfq_data *cfqd = q->elevator->elevator_data;
565
566         cfqd->rq_in_driver++;
567
568         /*
569          * If the depth is larger 1, it really could be queueing. But lets
570          * make the mark a little higher - idling could still be good for
571          * low queueing, and a low queueing number could also just indicate
572          * a SCSI mid layer like behaviour where limit+1 is often seen.
573          */
574         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
575                 cfqd->hw_tag = 1;
576 }
577
578 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
579 {
580         struct cfq_data *cfqd = q->elevator->elevator_data;
581
582         WARN_ON(!cfqd->rq_in_driver);
583         cfqd->rq_in_driver--;
584 }
585
586 static void cfq_remove_request(struct request *rq)
587 {
588         struct cfq_queue *cfqq = RQ_CFQQ(rq);
589
590         if (cfqq->next_rq == rq)
591                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
592
593         list_del_init(&rq->queuelist);
594         cfq_del_rq_rb(rq);
595
596         if (rq_is_meta(rq)) {
597                 WARN_ON(!cfqq->meta_pending);
598                 cfqq->meta_pending--;
599         }
600 }
601
602 static int
603 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
604 {
605         struct cfq_data *cfqd = q->elevator->elevator_data;
606         struct request *__rq;
607
608         __rq = cfq_find_rq_fmerge(cfqd, bio);
609         if (__rq && elv_rq_merge_ok(__rq, bio)) {
610                 *req = __rq;
611                 return ELEVATOR_FRONT_MERGE;
612         }
613
614         return ELEVATOR_NO_MERGE;
615 }
616
617 static void cfq_merged_request(request_queue_t *q, struct request *req,
618                                int type)
619 {
620         if (type == ELEVATOR_FRONT_MERGE) {
621                 struct cfq_queue *cfqq = RQ_CFQQ(req);
622
623                 cfq_reposition_rq_rb(cfqq, req);
624         }
625 }
626
627 static void
628 cfq_merged_requests(request_queue_t *q, struct request *rq,
629                     struct request *next)
630 {
631         /*
632          * reposition in fifo if next is older than rq
633          */
634         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
635             time_before(next->start_time, rq->start_time))
636                 list_move(&rq->queuelist, &next->queuelist);
637
638         cfq_remove_request(next);
639 }
640
641 static int cfq_allow_merge(request_queue_t *q, struct request *rq,
642                            struct bio *bio)
643 {
644         struct cfq_data *cfqd = q->elevator->elevator_data;
645         const int rw = bio_data_dir(bio);
646         struct cfq_queue *cfqq;
647         pid_t key;
648
649         /*
650          * Disallow merge of a sync bio into an async request.
651          */
652         if ((bio_data_dir(bio) == READ || bio_sync(bio)) && !rq_is_sync(rq))
653                 return 0;
654
655         /*
656          * Lookup the cfqq that this bio will be queued with. Allow
657          * merge only if rq is queued there.
658          */
659         key = cfq_queue_pid(current, rw, bio_sync(bio));
660         cfqq = cfq_find_cfq_hash(cfqd, key, current->ioprio);
661
662         if (cfqq == RQ_CFQQ(rq))
663                 return 1;
664
665         return 0;
666 }
667
668 static inline void
669 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
670 {
671         if (cfqq) {
672                 /*
673                  * stop potential idle class queues waiting service
674                  */
675                 del_timer(&cfqd->idle_class_timer);
676
677                 cfqq->slice_end = 0;
678                 cfq_clear_cfqq_must_alloc_slice(cfqq);
679                 cfq_clear_cfqq_fifo_expire(cfqq);
680                 cfq_mark_cfqq_slice_new(cfqq);
681         }
682
683         cfqd->active_queue = cfqq;
684 }
685
686 /*
687  * current cfqq expired its slice (or was too idle), select new one
688  */
689 static void
690 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
691                     int preempted, int timed_out)
692 {
693         if (cfq_cfqq_wait_request(cfqq))
694                 del_timer(&cfqd->idle_slice_timer);
695
696         cfq_clear_cfqq_must_dispatch(cfqq);
697         cfq_clear_cfqq_wait_request(cfqq);
698         cfq_clear_cfqq_queue_new(cfqq);
699
700         /*
701          * store what was left of this slice, if the queue idled out
702          * or was preempted
703          */
704         if (timed_out && !cfq_cfqq_slice_new(cfqq))
705                 cfqq->slice_resid = cfqq->slice_end - jiffies;
706
707         cfq_resort_rr_list(cfqq, preempted);
708
709         if (cfqq == cfqd->active_queue)
710                 cfqd->active_queue = NULL;
711
712         if (cfqd->active_cic) {
713                 put_io_context(cfqd->active_cic->ioc);
714                 cfqd->active_cic = NULL;
715         }
716
717         cfqd->dispatch_slice = 0;
718 }
719
720 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted,
721                                      int timed_out)
722 {
723         struct cfq_queue *cfqq = cfqd->active_queue;
724
725         if (cfqq)
726                 __cfq_slice_expired(cfqd, cfqq, preempted, timed_out);
727 }
728
729 /*
730  * 0
731  * 0,1
732  * 0,1,2
733  * 0,1,2,3
734  * 0,1,2,3,4
735  * 0,1,2,3,4,5
736  * 0,1,2,3,4,5,6
737  * 0,1,2,3,4,5,6,7
738  */
739 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
740 {
741         int prio, wrap;
742
743         prio = -1;
744         wrap = 0;
745         do {
746                 int p;
747
748                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
749                         if (!list_empty(&cfqd->rr_list[p])) {
750                                 prio = p;
751                                 break;
752                         }
753                 }
754
755                 if (prio != -1)
756                         break;
757                 cfqd->cur_prio = 0;
758                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
759                         cfqd->cur_end_prio = 0;
760                         if (wrap)
761                                 break;
762                         wrap = 1;
763                 }
764         } while (1);
765
766         if (unlikely(prio == -1))
767                 return -1;
768
769         BUG_ON(prio >= CFQ_PRIO_LISTS);
770
771         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
772
773         cfqd->cur_prio = prio + 1;
774         if (cfqd->cur_prio > cfqd->cur_end_prio) {
775                 cfqd->cur_end_prio = cfqd->cur_prio;
776                 cfqd->cur_prio = 0;
777         }
778         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
779                 cfqd->cur_prio = 0;
780                 cfqd->cur_end_prio = 0;
781         }
782
783         return prio;
784 }
785
786 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
787 {
788         struct cfq_queue *cfqq = NULL;
789
790         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
791                 /*
792                  * if current list is non-empty, grab first entry. if it is
793                  * empty, get next prio level and grab first entry then if any
794                  * are spliced
795                  */
796                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
797         } else if (!list_empty(&cfqd->busy_rr)) {
798                 /*
799                  * If no new queues are available, check if the busy list has
800                  * some before falling back to idle io.
801                  */
802                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
803         } else if (!list_empty(&cfqd->idle_rr)) {
804                 /*
805                  * if we have idle queues and no rt or be queues had pending
806                  * requests, either allow immediate service if the grace period
807                  * has passed or arm the idle grace timer
808                  */
809                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
810
811                 if (time_after_eq(jiffies, end))
812                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
813                 else
814                         mod_timer(&cfqd->idle_class_timer, end);
815         }
816
817         __cfq_set_active_queue(cfqd, cfqq);
818         return cfqq;
819 }
820
821 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
822
823 static int cfq_arm_slice_timer(struct cfq_data *cfqd)
824 {
825         struct cfq_queue *cfqq = cfqd->active_queue;
826         struct cfq_io_context *cic;
827         unsigned long sl;
828
829         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
830
831         /*
832          * idle is disabled, either manually or by past process history
833          */
834         if (!cfqd->cfq_slice_idle)
835                 return 0;
836         if (!cfq_cfqq_idle_window(cfqq))
837                 return 0;
838         /*
839          * task has exited, don't wait
840          */
841         cic = cfqd->active_cic;
842         if (!cic || !cic->ioc->task)
843                 return 0;
844
845         cfq_mark_cfqq_must_dispatch(cfqq);
846         cfq_mark_cfqq_wait_request(cfqq);
847
848         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
849
850         /*
851          * we don't want to idle for seeks, but we do want to allow
852          * fair distribution of slice time for a process doing back-to-back
853          * seeks. so allow a little bit of time for him to submit a new rq
854          */
855         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
856                 sl = min(sl, msecs_to_jiffies(2));
857
858         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
859         return 1;
860 }
861
862 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
863 {
864         struct cfq_data *cfqd = q->elevator->elevator_data;
865         struct cfq_queue *cfqq = RQ_CFQQ(rq);
866
867         cfq_remove_request(rq);
868         cfqq->on_dispatch[rq_is_sync(rq)]++;
869         elv_dispatch_sort(q, rq);
870
871         rq = list_entry(q->queue_head.prev, struct request, queuelist);
872         cfqd->last_sector = rq->sector + rq->nr_sectors;
873 }
874
875 /*
876  * return expired entry, or NULL to just start from scratch in rbtree
877  */
878 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
879 {
880         struct cfq_data *cfqd = cfqq->cfqd;
881         struct request *rq;
882         int fifo;
883
884         if (cfq_cfqq_fifo_expire(cfqq))
885                 return NULL;
886
887         cfq_mark_cfqq_fifo_expire(cfqq);
888
889         if (list_empty(&cfqq->fifo))
890                 return NULL;
891
892         fifo = cfq_cfqq_class_sync(cfqq);
893         rq = rq_entry_fifo(cfqq->fifo.next);
894
895         if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
896                 return rq;
897
898         return NULL;
899 }
900
901 static inline int
902 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
903 {
904         const int base_rq = cfqd->cfq_slice_async_rq;
905
906         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
907
908         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
909 }
910
911 /*
912  * get next queue for service
913  */
914 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
915 {
916         struct cfq_queue *cfqq;
917
918         cfqq = cfqd->active_queue;
919         if (!cfqq)
920                 goto new_queue;
921
922         /*
923          * slice has expired
924          */
925         if (!cfq_cfqq_must_dispatch(cfqq) && cfq_slice_used(cfqq))
926                 goto expire;
927
928         /*
929          * if queue has requests, dispatch one. if not, check if
930          * enough slice is left to wait for one
931          */
932         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
933                 goto keep_queue;
934         else if (cfq_cfqq_slice_new(cfqq) || cfq_cfqq_dispatched(cfqq)) {
935                 cfqq = NULL;
936                 goto keep_queue;
937         } else if (cfq_cfqq_class_sync(cfqq)) {
938                 if (cfq_arm_slice_timer(cfqd))
939                         return NULL;
940         }
941
942 expire:
943         cfq_slice_expired(cfqd, 0, 0);
944 new_queue:
945         cfqq = cfq_set_active_queue(cfqd);
946 keep_queue:
947         return cfqq;
948 }
949
950 static int
951 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
952                         int max_dispatch)
953 {
954         int dispatched = 0;
955
956         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
957
958         do {
959                 struct request *rq;
960
961                 /*
962                  * follow expired path, else get first next available
963                  */
964                 if ((rq = cfq_check_fifo(cfqq)) == NULL)
965                         rq = cfqq->next_rq;
966
967                 /*
968                  * finally, insert request into driver dispatch list
969                  */
970                 cfq_dispatch_insert(cfqd->queue, rq);
971
972                 cfqd->dispatch_slice++;
973                 dispatched++;
974
975                 if (!cfqd->active_cic) {
976                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
977                         cfqd->active_cic = RQ_CIC(rq);
978                 }
979
980                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
981                         break;
982
983         } while (dispatched < max_dispatch);
984
985         /*
986          * expire an async queue immediately if it has used up its slice. idle
987          * queue always expire after 1 dispatch round.
988          */
989         if ((!cfq_cfqq_sync(cfqq) &&
990             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
991             cfq_class_idle(cfqq)) {
992                 cfqq->slice_end = jiffies + 1;
993                 cfq_slice_expired(cfqd, 0, 0);
994         }
995
996         return dispatched;
997 }
998
999 static int
1000 cfq_forced_dispatch_cfqqs(struct list_head *list)
1001 {
1002         struct cfq_queue *cfqq, *next;
1003         int dispatched;
1004
1005         dispatched = 0;
1006         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1007                 while (cfqq->next_rq) {
1008                         cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1009                         dispatched++;
1010                 }
1011                 BUG_ON(!list_empty(&cfqq->fifo));
1012         }
1013
1014         return dispatched;
1015 }
1016
1017 static int
1018 cfq_forced_dispatch(struct cfq_data *cfqd)
1019 {
1020         int i, dispatched = 0;
1021
1022         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1023                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1024
1025         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1026         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1027         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1028
1029         cfq_slice_expired(cfqd, 0, 0);
1030
1031         BUG_ON(cfqd->busy_queues);
1032
1033         return dispatched;
1034 }
1035
1036 static int
1037 cfq_dispatch_requests(request_queue_t *q, int force)
1038 {
1039         struct cfq_data *cfqd = q->elevator->elevator_data;
1040         struct cfq_queue *cfqq, *prev_cfqq;
1041         int dispatched;
1042
1043         if (!cfqd->busy_queues)
1044                 return 0;
1045
1046         if (unlikely(force))
1047                 return cfq_forced_dispatch(cfqd);
1048
1049         dispatched = 0;
1050         prev_cfqq = NULL;
1051         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1052                 int max_dispatch;
1053
1054                 /*
1055                  * Don't repeat dispatch from the previous queue.
1056                  */
1057                 if (prev_cfqq == cfqq)
1058                         break;
1059
1060                 /*
1061                  * So we have dispatched before in this round, if the
1062                  * next queue has idling enabled (must be sync), don't
1063                  * allow it service until the previous have continued.
1064                  */
1065                 if (cfqd->rq_in_driver && cfq_cfqq_idle_window(cfqq))
1066                         break;
1067
1068                 cfq_clear_cfqq_must_dispatch(cfqq);
1069                 cfq_clear_cfqq_wait_request(cfqq);
1070                 del_timer(&cfqd->idle_slice_timer);
1071
1072                 max_dispatch = cfqd->cfq_quantum;
1073                 if (cfq_class_idle(cfqq))
1074                         max_dispatch = 1;
1075
1076                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1077                 prev_cfqq = cfqq;
1078         }
1079
1080         return dispatched;
1081 }
1082
1083 /*
1084  * task holds one reference to the queue, dropped when task exits. each rq
1085  * in-flight on this queue also holds a reference, dropped when rq is freed.
1086  *
1087  * queue lock must be held here.
1088  */
1089 static void cfq_put_queue(struct cfq_queue *cfqq)
1090 {
1091         struct cfq_data *cfqd = cfqq->cfqd;
1092
1093         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1094
1095         if (!atomic_dec_and_test(&cfqq->ref))
1096                 return;
1097
1098         BUG_ON(rb_first(&cfqq->sort_list));
1099         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1100         BUG_ON(cfq_cfqq_on_rr(cfqq));
1101
1102         if (unlikely(cfqd->active_queue == cfqq)) {
1103                 __cfq_slice_expired(cfqd, cfqq, 0, 0);
1104                 cfq_schedule_dispatch(cfqd);
1105         }
1106
1107         /*
1108          * it's on the empty list and still hashed
1109          */
1110         list_del(&cfqq->cfq_list);
1111         hlist_del(&cfqq->cfq_hash);
1112         kmem_cache_free(cfq_pool, cfqq);
1113 }
1114
1115 static struct cfq_queue *
1116 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1117                     const int hashval)
1118 {
1119         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1120         struct hlist_node *entry;
1121         struct cfq_queue *__cfqq;
1122
1123         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1124                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1125
1126                 if (__cfqq->key == key && (__p == prio || !prio))
1127                         return __cfqq;
1128         }
1129
1130         return NULL;
1131 }
1132
1133 static struct cfq_queue *
1134 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1135 {
1136         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1137 }
1138
1139 static void cfq_free_io_context(struct io_context *ioc)
1140 {
1141         struct cfq_io_context *__cic;
1142         struct rb_node *n;
1143         int freed = 0;
1144
1145         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1146                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1147                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1148                 kmem_cache_free(cfq_ioc_pool, __cic);
1149                 freed++;
1150         }
1151
1152         elv_ioc_count_mod(ioc_count, -freed);
1153
1154         if (ioc_gone && !elv_ioc_count_read(ioc_count))
1155                 complete(ioc_gone);
1156 }
1157
1158 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1159 {
1160         if (unlikely(cfqq == cfqd->active_queue)) {
1161                 __cfq_slice_expired(cfqd, cfqq, 0, 0);
1162                 cfq_schedule_dispatch(cfqd);
1163         }
1164
1165         cfq_put_queue(cfqq);
1166 }
1167
1168 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1169                                          struct cfq_io_context *cic)
1170 {
1171         list_del_init(&cic->queue_list);
1172         smp_wmb();
1173         cic->key = NULL;
1174
1175         if (cic->cfqq[ASYNC]) {
1176                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1177                 cic->cfqq[ASYNC] = NULL;
1178         }
1179
1180         if (cic->cfqq[SYNC]) {
1181                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1182                 cic->cfqq[SYNC] = NULL;
1183         }
1184 }
1185
1186
1187 /*
1188  * Called with interrupts disabled
1189  */
1190 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1191 {
1192         struct cfq_data *cfqd = cic->key;
1193
1194         if (cfqd) {
1195                 request_queue_t *q = cfqd->queue;
1196
1197                 spin_lock_irq(q->queue_lock);
1198                 __cfq_exit_single_io_context(cfqd, cic);
1199                 spin_unlock_irq(q->queue_lock);
1200         }
1201 }
1202
1203 static void cfq_exit_io_context(struct io_context *ioc)
1204 {
1205         struct cfq_io_context *__cic;
1206         struct rb_node *n;
1207
1208         /*
1209          * put the reference this task is holding to the various queues
1210          */
1211
1212         n = rb_first(&ioc->cic_root);
1213         while (n != NULL) {
1214                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1215
1216                 cfq_exit_single_io_context(__cic);
1217                 n = rb_next(n);
1218         }
1219 }
1220
1221 static struct cfq_io_context *
1222 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1223 {
1224         struct cfq_io_context *cic;
1225
1226         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
1227         if (cic) {
1228                 memset(cic, 0, sizeof(*cic));
1229                 cic->last_end_request = jiffies;
1230                 INIT_LIST_HEAD(&cic->queue_list);
1231                 cic->dtor = cfq_free_io_context;
1232                 cic->exit = cfq_exit_io_context;
1233                 elv_ioc_count_inc(ioc_count);
1234         }
1235
1236         return cic;
1237 }
1238
1239 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1240 {
1241         struct task_struct *tsk = current;
1242         int ioprio_class;
1243
1244         if (!cfq_cfqq_prio_changed(cfqq))
1245                 return;
1246
1247         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1248         switch (ioprio_class) {
1249                 default:
1250                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1251                 case IOPRIO_CLASS_NONE:
1252                         /*
1253                          * no prio set, place us in the middle of the BE classes
1254                          */
1255                         cfqq->ioprio = task_nice_ioprio(tsk);
1256                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1257                         break;
1258                 case IOPRIO_CLASS_RT:
1259                         cfqq->ioprio = task_ioprio(tsk);
1260                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1261                         break;
1262                 case IOPRIO_CLASS_BE:
1263                         cfqq->ioprio = task_ioprio(tsk);
1264                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1265                         break;
1266                 case IOPRIO_CLASS_IDLE:
1267                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1268                         cfqq->ioprio = 7;
1269                         cfq_clear_cfqq_idle_window(cfqq);
1270                         break;
1271         }
1272
1273         /*
1274          * keep track of original prio settings in case we have to temporarily
1275          * elevate the priority of this queue
1276          */
1277         cfqq->org_ioprio = cfqq->ioprio;
1278         cfqq->org_ioprio_class = cfqq->ioprio_class;
1279
1280         cfq_resort_rr_list(cfqq, 0);
1281         cfq_clear_cfqq_prio_changed(cfqq);
1282 }
1283
1284 static inline void changed_ioprio(struct cfq_io_context *cic)
1285 {
1286         struct cfq_data *cfqd = cic->key;
1287         struct cfq_queue *cfqq;
1288         unsigned long flags;
1289
1290         if (unlikely(!cfqd))
1291                 return;
1292
1293         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1294
1295         cfqq = cic->cfqq[ASYNC];
1296         if (cfqq) {
1297                 struct cfq_queue *new_cfqq;
1298                 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1299                                          GFP_ATOMIC);
1300                 if (new_cfqq) {
1301                         cic->cfqq[ASYNC] = new_cfqq;
1302                         cfq_put_queue(cfqq);
1303                 }
1304         }
1305
1306         cfqq = cic->cfqq[SYNC];
1307         if (cfqq)
1308                 cfq_mark_cfqq_prio_changed(cfqq);
1309
1310         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1311 }
1312
1313 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1314 {
1315         struct cfq_io_context *cic;
1316         struct rb_node *n;
1317
1318         ioc->ioprio_changed = 0;
1319
1320         n = rb_first(&ioc->cic_root);
1321         while (n != NULL) {
1322                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1323
1324                 changed_ioprio(cic);
1325                 n = rb_next(n);
1326         }
1327 }
1328
1329 static struct cfq_queue *
1330 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1331               gfp_t gfp_mask)
1332 {
1333         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1334         struct cfq_queue *cfqq, *new_cfqq = NULL;
1335         unsigned short ioprio;
1336
1337 retry:
1338         ioprio = tsk->ioprio;
1339         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1340
1341         if (!cfqq) {
1342                 if (new_cfqq) {
1343                         cfqq = new_cfqq;
1344                         new_cfqq = NULL;
1345                 } else if (gfp_mask & __GFP_WAIT) {
1346                         /*
1347                          * Inform the allocator of the fact that we will
1348                          * just repeat this allocation if it fails, to allow
1349                          * the allocator to do whatever it needs to attempt to
1350                          * free memory.
1351                          */
1352                         spin_unlock_irq(cfqd->queue->queue_lock);
1353                         new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
1354                         spin_lock_irq(cfqd->queue->queue_lock);
1355                         goto retry;
1356                 } else {
1357                         cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
1358                         if (!cfqq)
1359                                 goto out;
1360                 }
1361
1362                 memset(cfqq, 0, sizeof(*cfqq));
1363
1364                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1365                 INIT_LIST_HEAD(&cfqq->cfq_list);
1366                 INIT_LIST_HEAD(&cfqq->fifo);
1367
1368                 cfqq->key = key;
1369                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1370                 atomic_set(&cfqq->ref, 0);
1371                 cfqq->cfqd = cfqd;
1372
1373                 cfq_mark_cfqq_idle_window(cfqq);
1374                 cfq_mark_cfqq_prio_changed(cfqq);
1375                 cfq_mark_cfqq_queue_new(cfqq);
1376                 cfq_init_prio_data(cfqq);
1377         }
1378
1379         if (new_cfqq)
1380                 kmem_cache_free(cfq_pool, new_cfqq);
1381
1382         atomic_inc(&cfqq->ref);
1383 out:
1384         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1385         return cfqq;
1386 }
1387
1388 static void
1389 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1390 {
1391         WARN_ON(!list_empty(&cic->queue_list));
1392         rb_erase(&cic->rb_node, &ioc->cic_root);
1393         kmem_cache_free(cfq_ioc_pool, cic);
1394         elv_ioc_count_dec(ioc_count);
1395 }
1396
1397 static struct cfq_io_context *
1398 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1399 {
1400         struct rb_node *n;
1401         struct cfq_io_context *cic;
1402         void *k, *key = cfqd;
1403
1404 restart:
1405         n = ioc->cic_root.rb_node;
1406         while (n) {
1407                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1408                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1409                 k = cic->key;
1410                 if (unlikely(!k)) {
1411                         cfq_drop_dead_cic(ioc, cic);
1412                         goto restart;
1413                 }
1414
1415                 if (key < k)
1416                         n = n->rb_left;
1417                 else if (key > k)
1418                         n = n->rb_right;
1419                 else
1420                         return cic;
1421         }
1422
1423         return NULL;
1424 }
1425
1426 static inline void
1427 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1428              struct cfq_io_context *cic)
1429 {
1430         struct rb_node **p;
1431         struct rb_node *parent;
1432         struct cfq_io_context *__cic;
1433         unsigned long flags;
1434         void *k;
1435
1436         cic->ioc = ioc;
1437         cic->key = cfqd;
1438
1439 restart:
1440         parent = NULL;
1441         p = &ioc->cic_root.rb_node;
1442         while (*p) {
1443                 parent = *p;
1444                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1445                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1446                 k = __cic->key;
1447                 if (unlikely(!k)) {
1448                         cfq_drop_dead_cic(ioc, __cic);
1449                         goto restart;
1450                 }
1451
1452                 if (cic->key < k)
1453                         p = &(*p)->rb_left;
1454                 else if (cic->key > k)
1455                         p = &(*p)->rb_right;
1456                 else
1457                         BUG();
1458         }
1459
1460         rb_link_node(&cic->rb_node, parent, p);
1461         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1462
1463         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1464         list_add(&cic->queue_list, &cfqd->cic_list);
1465         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1466 }
1467
1468 /*
1469  * Setup general io context and cfq io context. There can be several cfq
1470  * io contexts per general io context, if this process is doing io to more
1471  * than one device managed by cfq.
1472  */
1473 static struct cfq_io_context *
1474 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1475 {
1476         struct io_context *ioc = NULL;
1477         struct cfq_io_context *cic;
1478
1479         might_sleep_if(gfp_mask & __GFP_WAIT);
1480
1481         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1482         if (!ioc)
1483                 return NULL;
1484
1485         cic = cfq_cic_rb_lookup(cfqd, ioc);
1486         if (cic)
1487                 goto out;
1488
1489         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1490         if (cic == NULL)
1491                 goto err;
1492
1493         cfq_cic_link(cfqd, ioc, cic);
1494 out:
1495         smp_read_barrier_depends();
1496         if (unlikely(ioc->ioprio_changed))
1497                 cfq_ioc_set_ioprio(ioc);
1498
1499         return cic;
1500 err:
1501         put_io_context(ioc);
1502         return NULL;
1503 }
1504
1505 static void
1506 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1507 {
1508         unsigned long elapsed = jiffies - cic->last_end_request;
1509         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1510
1511         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1512         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1513         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1514 }
1515
1516 static void
1517 cfq_update_io_seektime(struct cfq_io_context *cic, struct request *rq)
1518 {
1519         sector_t sdist;
1520         u64 total;
1521
1522         if (cic->last_request_pos < rq->sector)
1523                 sdist = rq->sector - cic->last_request_pos;
1524         else
1525                 sdist = cic->last_request_pos - rq->sector;
1526
1527         /*
1528          * Don't allow the seek distance to get too large from the
1529          * odd fragment, pagein, etc
1530          */
1531         if (cic->seek_samples <= 60) /* second&third seek */
1532                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1533         else
1534                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1535
1536         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1537         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1538         total = cic->seek_total + (cic->seek_samples/2);
1539         do_div(total, cic->seek_samples);
1540         cic->seek_mean = (sector_t)total;
1541 }
1542
1543 /*
1544  * Disable idle window if the process thinks too long or seeks so much that
1545  * it doesn't matter
1546  */
1547 static void
1548 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1549                        struct cfq_io_context *cic)
1550 {
1551         int enable_idle = cfq_cfqq_idle_window(cfqq);
1552
1553         if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1554             (cfqd->hw_tag && CIC_SEEKY(cic)))
1555                 enable_idle = 0;
1556         else if (sample_valid(cic->ttime_samples)) {
1557                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1558                         enable_idle = 0;
1559                 else
1560                         enable_idle = 1;
1561         }
1562
1563         if (enable_idle)
1564                 cfq_mark_cfqq_idle_window(cfqq);
1565         else
1566                 cfq_clear_cfqq_idle_window(cfqq);
1567 }
1568
1569 /*
1570  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1571  * no or if we aren't sure, a 1 will cause a preempt.
1572  */
1573 static int
1574 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1575                    struct request *rq)
1576 {
1577         struct cfq_queue *cfqq = cfqd->active_queue;
1578
1579         if (cfq_class_idle(new_cfqq))
1580                 return 0;
1581
1582         if (!cfqq)
1583                 return 0;
1584
1585         if (cfq_class_idle(cfqq))
1586                 return 1;
1587         if (!cfq_cfqq_wait_request(new_cfqq))
1588                 return 0;
1589         /*
1590          * if the new request is sync, but the currently running queue is
1591          * not, let the sync request have priority.
1592          */
1593         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1594                 return 1;
1595         /*
1596          * So both queues are sync. Let the new request get disk time if
1597          * it's a metadata request and the current queue is doing regular IO.
1598          */
1599         if (rq_is_meta(rq) && !cfqq->meta_pending)
1600                 return 1;
1601
1602         return 0;
1603 }
1604
1605 /*
1606  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1607  * let it have half of its nominal slice.
1608  */
1609 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1610 {
1611         cfq_slice_expired(cfqd, 1, 1);
1612
1613         /*
1614          * Put the new queue at the front of the of the current list,
1615          * so we know that it will be selected next.
1616          */
1617         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1618         list_move(&cfqq->cfq_list, &cfqd->cur_rr);
1619
1620         cfqq->slice_end = 0;
1621         cfq_mark_cfqq_slice_new(cfqq);
1622 }
1623
1624 /*
1625  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1626  * something we should do about it
1627  */
1628 static void
1629 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1630                 struct request *rq)
1631 {
1632         struct cfq_io_context *cic = RQ_CIC(rq);
1633
1634         if (rq_is_meta(rq))
1635                 cfqq->meta_pending++;
1636
1637         /*
1638          * check if this request is a better next-serve candidate)) {
1639          */
1640         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
1641         BUG_ON(!cfqq->next_rq);
1642
1643         /*
1644          * we never wait for an async request and we don't allow preemption
1645          * of an async request. so just return early
1646          */
1647         if (!rq_is_sync(rq)) {
1648                 /*
1649                  * sync process issued an async request, if it's waiting
1650                  * then expire it and kick rq handling.
1651                  */
1652                 if (cic == cfqd->active_cic &&
1653                     del_timer(&cfqd->idle_slice_timer)) {
1654                         cfq_slice_expired(cfqd, 0, 0);
1655                         blk_start_queueing(cfqd->queue);
1656                 }
1657                 return;
1658         }
1659
1660         cfq_update_io_thinktime(cfqd, cic);
1661         cfq_update_io_seektime(cic, rq);
1662         cfq_update_idle_window(cfqd, cfqq, cic);
1663
1664         cic->last_request_pos = rq->sector + rq->nr_sectors;
1665
1666         if (cfqq == cfqd->active_queue) {
1667                 /*
1668                  * if we are waiting for a request for this queue, let it rip
1669                  * immediately and flag that we must not expire this queue
1670                  * just now
1671                  */
1672                 if (cfq_cfqq_wait_request(cfqq)) {
1673                         cfq_mark_cfqq_must_dispatch(cfqq);
1674                         del_timer(&cfqd->idle_slice_timer);
1675                         blk_start_queueing(cfqd->queue);
1676                 }
1677         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1678                 /*
1679                  * not the active queue - expire current slice if it is
1680                  * idle and has expired it's mean thinktime or this new queue
1681                  * has some old slice time left and is of higher priority
1682                  */
1683                 cfq_preempt_queue(cfqd, cfqq);
1684                 cfq_mark_cfqq_must_dispatch(cfqq);
1685                 blk_start_queueing(cfqd->queue);
1686         }
1687 }
1688
1689 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1690 {
1691         struct cfq_data *cfqd = q->elevator->elevator_data;
1692         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1693
1694         cfq_init_prio_data(cfqq);
1695
1696         cfq_add_rq_rb(rq);
1697
1698         list_add_tail(&rq->queuelist, &cfqq->fifo);
1699
1700         cfq_rq_enqueued(cfqd, cfqq, rq);
1701 }
1702
1703 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1704 {
1705         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1706         struct cfq_data *cfqd = cfqq->cfqd;
1707         const int sync = rq_is_sync(rq);
1708         unsigned long now;
1709
1710         now = jiffies;
1711
1712         WARN_ON(!cfqd->rq_in_driver);
1713         WARN_ON(!cfqq->on_dispatch[sync]);
1714         cfqd->rq_in_driver--;
1715         cfqq->on_dispatch[sync]--;
1716         cfqq->service_last = now;
1717
1718         if (!cfq_class_idle(cfqq))
1719                 cfqd->last_end_request = now;
1720
1721         cfq_resort_rr_list(cfqq, 0);
1722
1723         if (sync)
1724                 RQ_CIC(rq)->last_end_request = now;
1725
1726         /*
1727          * If this is the active queue, check if it needs to be expired,
1728          * or if we want to idle in case it has no pending requests.
1729          */
1730         if (cfqd->active_queue == cfqq) {
1731                 if (cfq_cfqq_slice_new(cfqq)) {
1732                         cfq_set_prio_slice(cfqd, cfqq);
1733                         cfq_clear_cfqq_slice_new(cfqq);
1734                 }
1735                 if (cfq_slice_used(cfqq))
1736                         cfq_slice_expired(cfqd, 0, 1);
1737                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1738                         if (!cfq_arm_slice_timer(cfqd))
1739                                 cfq_schedule_dispatch(cfqd);
1740                 }
1741         }
1742 }
1743
1744 /*
1745  * we temporarily boost lower priority queues if they are holding fs exclusive
1746  * resources. they are boosted to normal prio (CLASS_BE/4)
1747  */
1748 static void cfq_prio_boost(struct cfq_queue *cfqq)
1749 {
1750         const int ioprio_class = cfqq->ioprio_class;
1751         const int ioprio = cfqq->ioprio;
1752
1753         if (has_fs_excl()) {
1754                 /*
1755                  * boost idle prio on transactions that would lock out other
1756                  * users of the filesystem
1757                  */
1758                 if (cfq_class_idle(cfqq))
1759                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1760                 if (cfqq->ioprio > IOPRIO_NORM)
1761                         cfqq->ioprio = IOPRIO_NORM;
1762         } else {
1763                 /*
1764                  * check if we need to unboost the queue
1765                  */
1766                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1767                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1768                 if (cfqq->ioprio != cfqq->org_ioprio)
1769                         cfqq->ioprio = cfqq->org_ioprio;
1770         }
1771
1772         /*
1773          * refile between round-robin lists if we moved the priority class
1774          */
1775         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio))
1776                 cfq_resort_rr_list(cfqq, 0);
1777 }
1778
1779 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1780 {
1781         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1782             !cfq_cfqq_must_alloc_slice(cfqq)) {
1783                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1784                 return ELV_MQUEUE_MUST;
1785         }
1786
1787         return ELV_MQUEUE_MAY;
1788 }
1789
1790 static int cfq_may_queue(request_queue_t *q, int rw)
1791 {
1792         struct cfq_data *cfqd = q->elevator->elevator_data;
1793         struct task_struct *tsk = current;
1794         struct cfq_queue *cfqq;
1795         unsigned int key;
1796
1797         key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC);
1798
1799         /*
1800          * don't force setup of a queue from here, as a call to may_queue
1801          * does not necessarily imply that a request actually will be queued.
1802          * so just lookup a possibly existing queue, or return 'may queue'
1803          * if that fails
1804          */
1805         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
1806         if (cfqq) {
1807                 cfq_init_prio_data(cfqq);
1808                 cfq_prio_boost(cfqq);
1809
1810                 return __cfq_may_queue(cfqq);
1811         }
1812
1813         return ELV_MQUEUE_MAY;
1814 }
1815
1816 /*
1817  * queue lock held here
1818  */
1819 static void cfq_put_request(struct request *rq)
1820 {
1821         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1822
1823         if (cfqq) {
1824                 const int rw = rq_data_dir(rq);
1825
1826                 BUG_ON(!cfqq->allocated[rw]);
1827                 cfqq->allocated[rw]--;
1828
1829                 put_io_context(RQ_CIC(rq)->ioc);
1830
1831                 rq->elevator_private = NULL;
1832                 rq->elevator_private2 = NULL;
1833
1834                 cfq_put_queue(cfqq);
1835         }
1836 }
1837
1838 /*
1839  * Allocate cfq data structures associated with this request.
1840  */
1841 static int
1842 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1843 {
1844         struct cfq_data *cfqd = q->elevator->elevator_data;
1845         struct task_struct *tsk = current;
1846         struct cfq_io_context *cic;
1847         const int rw = rq_data_dir(rq);
1848         const int is_sync = rq_is_sync(rq);
1849         pid_t key = cfq_queue_pid(tsk, rw, is_sync);
1850         struct cfq_queue *cfqq;
1851         unsigned long flags;
1852
1853         might_sleep_if(gfp_mask & __GFP_WAIT);
1854
1855         cic = cfq_get_io_context(cfqd, gfp_mask);
1856
1857         spin_lock_irqsave(q->queue_lock, flags);
1858
1859         if (!cic)
1860                 goto queue_fail;
1861
1862         if (!cic->cfqq[is_sync]) {
1863                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1864                 if (!cfqq)
1865                         goto queue_fail;
1866
1867                 cic->cfqq[is_sync] = cfqq;
1868         } else
1869                 cfqq = cic->cfqq[is_sync];
1870
1871         cfqq->allocated[rw]++;
1872         cfq_clear_cfqq_must_alloc(cfqq);
1873         atomic_inc(&cfqq->ref);
1874
1875         spin_unlock_irqrestore(q->queue_lock, flags);
1876
1877         rq->elevator_private = cic;
1878         rq->elevator_private2 = cfqq;
1879         return 0;
1880
1881 queue_fail:
1882         if (cic)
1883                 put_io_context(cic->ioc);
1884
1885         cfq_schedule_dispatch(cfqd);
1886         spin_unlock_irqrestore(q->queue_lock, flags);
1887         return 1;
1888 }
1889
1890 static void cfq_kick_queue(struct work_struct *work)
1891 {
1892         struct cfq_data *cfqd =
1893                 container_of(work, struct cfq_data, unplug_work);
1894         request_queue_t *q = cfqd->queue;
1895         unsigned long flags;
1896
1897         spin_lock_irqsave(q->queue_lock, flags);
1898         blk_start_queueing(q);
1899         spin_unlock_irqrestore(q->queue_lock, flags);
1900 }
1901
1902 /*
1903  * Timer running if the active_queue is currently idling inside its time slice
1904  */
1905 static void cfq_idle_slice_timer(unsigned long data)
1906 {
1907         struct cfq_data *cfqd = (struct cfq_data *) data;
1908         struct cfq_queue *cfqq;
1909         unsigned long flags;
1910         int timed_out = 1;
1911
1912         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1913
1914         if ((cfqq = cfqd->active_queue) != NULL) {
1915                 timed_out = 0;
1916
1917                 /*
1918                  * expired
1919                  */
1920                 if (cfq_slice_used(cfqq))
1921                         goto expire;
1922
1923                 /*
1924                  * only expire and reinvoke request handler, if there are
1925                  * other queues with pending requests
1926                  */
1927                 if (!cfqd->busy_queues)
1928                         goto out_cont;
1929
1930                 /*
1931                  * not expired and it has a request pending, let it dispatch
1932                  */
1933                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1934                         cfq_mark_cfqq_must_dispatch(cfqq);
1935                         goto out_kick;
1936                 }
1937         }
1938 expire:
1939         cfq_slice_expired(cfqd, 0, timed_out);
1940 out_kick:
1941         cfq_schedule_dispatch(cfqd);
1942 out_cont:
1943         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1944 }
1945
1946 /*
1947  * Timer running if an idle class queue is waiting for service
1948  */
1949 static void cfq_idle_class_timer(unsigned long data)
1950 {
1951         struct cfq_data *cfqd = (struct cfq_data *) data;
1952         unsigned long flags, end;
1953
1954         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1955
1956         /*
1957          * race with a non-idle queue, reset timer
1958          */
1959         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
1960         if (!time_after_eq(jiffies, end))
1961                 mod_timer(&cfqd->idle_class_timer, end);
1962         else
1963                 cfq_schedule_dispatch(cfqd);
1964
1965         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1966 }
1967
1968 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1969 {
1970         del_timer_sync(&cfqd->idle_slice_timer);
1971         del_timer_sync(&cfqd->idle_class_timer);
1972         blk_sync_queue(cfqd->queue);
1973 }
1974
1975 static void cfq_exit_queue(elevator_t *e)
1976 {
1977         struct cfq_data *cfqd = e->elevator_data;
1978         request_queue_t *q = cfqd->queue;
1979
1980         cfq_shutdown_timer_wq(cfqd);
1981
1982         spin_lock_irq(q->queue_lock);
1983
1984         if (cfqd->active_queue)
1985                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0, 0);
1986
1987         while (!list_empty(&cfqd->cic_list)) {
1988                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1989                                                         struct cfq_io_context,
1990                                                         queue_list);
1991
1992                 __cfq_exit_single_io_context(cfqd, cic);
1993         }
1994
1995         spin_unlock_irq(q->queue_lock);
1996
1997         cfq_shutdown_timer_wq(cfqd);
1998
1999         kfree(cfqd->cfq_hash);
2000         kfree(cfqd);
2001 }
2002
2003 static void *cfq_init_queue(request_queue_t *q)
2004 {
2005         struct cfq_data *cfqd;
2006         int i;
2007
2008         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
2009         if (!cfqd)
2010                 return NULL;
2011
2012         memset(cfqd, 0, sizeof(*cfqd));
2013
2014         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2015                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2016
2017         INIT_LIST_HEAD(&cfqd->busy_rr);
2018         INIT_LIST_HEAD(&cfqd->cur_rr);
2019         INIT_LIST_HEAD(&cfqd->idle_rr);
2020         INIT_LIST_HEAD(&cfqd->cic_list);
2021
2022         cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
2023         if (!cfqd->cfq_hash)
2024                 goto out_free;
2025
2026         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2027                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2028
2029         cfqd->queue = q;
2030
2031         init_timer(&cfqd->idle_slice_timer);
2032         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2033         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2034
2035         init_timer(&cfqd->idle_class_timer);
2036         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2037         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2038
2039         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2040
2041         cfqd->cfq_quantum = cfq_quantum;
2042         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2043         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2044         cfqd->cfq_back_max = cfq_back_max;
2045         cfqd->cfq_back_penalty = cfq_back_penalty;
2046         cfqd->cfq_slice[0] = cfq_slice_async;
2047         cfqd->cfq_slice[1] = cfq_slice_sync;
2048         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2049         cfqd->cfq_slice_idle = cfq_slice_idle;
2050
2051         return cfqd;
2052 out_free:
2053         kfree(cfqd);
2054         return NULL;
2055 }
2056
2057 static void cfq_slab_kill(void)
2058 {
2059         if (cfq_pool)
2060                 kmem_cache_destroy(cfq_pool);
2061         if (cfq_ioc_pool)
2062                 kmem_cache_destroy(cfq_ioc_pool);
2063 }
2064
2065 static int __init cfq_slab_setup(void)
2066 {
2067         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2068                                         NULL, NULL);
2069         if (!cfq_pool)
2070                 goto fail;
2071
2072         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2073                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2074         if (!cfq_ioc_pool)
2075                 goto fail;
2076
2077         return 0;
2078 fail:
2079         cfq_slab_kill();
2080         return -ENOMEM;
2081 }
2082
2083 /*
2084  * sysfs parts below -->
2085  */
2086
2087 static ssize_t
2088 cfq_var_show(unsigned int var, char *page)
2089 {
2090         return sprintf(page, "%d\n", var);
2091 }
2092
2093 static ssize_t
2094 cfq_var_store(unsigned int *var, const char *page, size_t count)
2095 {
2096         char *p = (char *) page;
2097
2098         *var = simple_strtoul(p, &p, 10);
2099         return count;
2100 }
2101
2102 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2103 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2104 {                                                                       \
2105         struct cfq_data *cfqd = e->elevator_data;                       \
2106         unsigned int __data = __VAR;                                    \
2107         if (__CONV)                                                     \
2108                 __data = jiffies_to_msecs(__data);                      \
2109         return cfq_var_show(__data, (page));                            \
2110 }
2111 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2112 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2113 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2114 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2115 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2116 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2117 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2118 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2119 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2120 #undef SHOW_FUNCTION
2121
2122 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2123 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2124 {                                                                       \
2125         struct cfq_data *cfqd = e->elevator_data;                       \
2126         unsigned int __data;                                            \
2127         int ret = cfq_var_store(&__data, (page), count);                \
2128         if (__data < (MIN))                                             \
2129                 __data = (MIN);                                         \
2130         else if (__data > (MAX))                                        \
2131                 __data = (MAX);                                         \
2132         if (__CONV)                                                     \
2133                 *(__PTR) = msecs_to_jiffies(__data);                    \
2134         else                                                            \
2135                 *(__PTR) = __data;                                      \
2136         return ret;                                                     \
2137 }
2138 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2139 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2140 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2141 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2142 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2143 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2144 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2145 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2146 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2147 #undef STORE_FUNCTION
2148
2149 #define CFQ_ATTR(name) \
2150         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2151
2152 static struct elv_fs_entry cfq_attrs[] = {
2153         CFQ_ATTR(quantum),
2154         CFQ_ATTR(fifo_expire_sync),
2155         CFQ_ATTR(fifo_expire_async),
2156         CFQ_ATTR(back_seek_max),
2157         CFQ_ATTR(back_seek_penalty),
2158         CFQ_ATTR(slice_sync),
2159         CFQ_ATTR(slice_async),
2160         CFQ_ATTR(slice_async_rq),
2161         CFQ_ATTR(slice_idle),
2162         __ATTR_NULL
2163 };
2164
2165 static struct elevator_type iosched_cfq = {
2166         .ops = {
2167                 .elevator_merge_fn =            cfq_merge,
2168                 .elevator_merged_fn =           cfq_merged_request,
2169                 .elevator_merge_req_fn =        cfq_merged_requests,
2170                 .elevator_allow_merge_fn =      cfq_allow_merge,
2171                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2172                 .elevator_add_req_fn =          cfq_insert_request,
2173                 .elevator_activate_req_fn =     cfq_activate_request,
2174                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2175                 .elevator_queue_empty_fn =      cfq_queue_empty,
2176                 .elevator_completed_req_fn =    cfq_completed_request,
2177                 .elevator_former_req_fn =       elv_rb_former_request,
2178                 .elevator_latter_req_fn =       elv_rb_latter_request,
2179                 .elevator_set_req_fn =          cfq_set_request,
2180                 .elevator_put_req_fn =          cfq_put_request,
2181                 .elevator_may_queue_fn =        cfq_may_queue,
2182                 .elevator_init_fn =             cfq_init_queue,
2183                 .elevator_exit_fn =             cfq_exit_queue,
2184                 .trim =                         cfq_free_io_context,
2185         },
2186         .elevator_attrs =       cfq_attrs,
2187         .elevator_name =        "cfq",
2188         .elevator_owner =       THIS_MODULE,
2189 };
2190
2191 static int __init cfq_init(void)
2192 {
2193         int ret;
2194
2195         /*
2196          * could be 0 on HZ < 1000 setups
2197          */
2198         if (!cfq_slice_async)
2199                 cfq_slice_async = 1;
2200         if (!cfq_slice_idle)
2201                 cfq_slice_idle = 1;
2202
2203         if (cfq_slab_setup())
2204                 return -ENOMEM;
2205
2206         ret = elv_register(&iosched_cfq);
2207         if (ret)
2208                 cfq_slab_kill();
2209
2210         return ret;
2211 }
2212
2213 static void __exit cfq_exit(void)
2214 {
2215         DECLARE_COMPLETION_ONSTACK(all_gone);
2216         elv_unregister(&iosched_cfq);
2217         ioc_gone = &all_gone;
2218         /* ioc_gone's update must be visible before reading ioc_count */
2219         smp_wmb();
2220         if (elv_ioc_count_read(ioc_count))
2221                 wait_for_completion(ioc_gone);
2222         synchronize_rcu();
2223         cfq_slab_kill();
2224 }
2225
2226 module_init(cfq_init);
2227 module_exit(cfq_exit);
2228
2229 MODULE_AUTHOR("Jens Axboe");
2230 MODULE_LICENSE("GPL");
2231 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");