0b4a47905575f2ee658d8a2f1e373aec71e935e5
[sfrench/cifs-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/rbtree.h>
13 #include <linux/ioprio.h>
14
15 /*
16  * tunables
17  */
18 static const int cfq_quantum = 4;               /* max queue in one round of service */
19 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
20 static const int cfq_back_max = 16 * 1024;      /* maximum backwards seek, in KiB */
21 static const int cfq_back_penalty = 2;          /* penalty of a backwards seek */
22
23 static const int cfq_slice_sync = HZ / 10;
24 static int cfq_slice_async = HZ / 25;
25 static const int cfq_slice_async_rq = 2;
26 static int cfq_slice_idle = HZ / 125;
27
28 /*
29  * grace period before allowing idle class to get disk access
30  */
31 #define CFQ_IDLE_GRACE          (HZ / 10)
32
33 /*
34  * below this threshold, we consider thinktime immediate
35  */
36 #define CFQ_MIN_TT              (2)
37
38 #define CFQ_SLICE_SCALE         (5)
39
40 #define RQ_CIC(rq)              ((struct cfq_io_context*)(rq)->elevator_private)
41 #define RQ_CFQQ(rq)             ((rq)->elevator_private2)
42
43 static struct kmem_cache *cfq_pool;
44 static struct kmem_cache *cfq_ioc_pool;
45
46 static DEFINE_PER_CPU(unsigned long, ioc_count);
47 static struct completion *ioc_gone;
48
49 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
50 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
51 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
52
53 #define ASYNC                   (0)
54 #define SYNC                    (1)
55
56 #define sample_valid(samples)   ((samples) > 80)
57
58 /*
59  * Most of our rbtree usage is for sorting with min extraction, so
60  * if we cache the leftmost node we don't have to walk down the tree
61  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
62  * move this into the elevator for the rq sorting as well.
63  */
64 struct cfq_rb_root {
65         struct rb_root rb;
66         struct rb_node *left;
67 };
68 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, }
69
70 /*
71  * Per block device queue structure
72  */
73 struct cfq_data {
74         struct request_queue *queue;
75
76         /*
77          * rr list of queues with requests and the count of them
78          */
79         struct cfq_rb_root service_tree;
80         unsigned int busy_queues;
81
82         int rq_in_driver;
83         int sync_flight;
84         int hw_tag;
85
86         /*
87          * idle window management
88          */
89         struct timer_list idle_slice_timer;
90         struct work_struct unplug_work;
91
92         struct cfq_queue *active_queue;
93         struct cfq_io_context *active_cic;
94
95         /*
96          * async queue for each priority case
97          */
98         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
99         struct cfq_queue *async_idle_cfqq;
100
101         struct timer_list idle_class_timer;
102
103         sector_t last_position;
104         unsigned long last_end_request;
105
106         /*
107          * tunables, see top of file
108          */
109         unsigned int cfq_quantum;
110         unsigned int cfq_fifo_expire[2];
111         unsigned int cfq_back_penalty;
112         unsigned int cfq_back_max;
113         unsigned int cfq_slice[2];
114         unsigned int cfq_slice_async_rq;
115         unsigned int cfq_slice_idle;
116
117         struct list_head cic_list;
118 };
119
120 /*
121  * Per process-grouping structure
122  */
123 struct cfq_queue {
124         /* reference count */
125         atomic_t ref;
126         /* parent cfq_data */
127         struct cfq_data *cfqd;
128         /* service_tree member */
129         struct rb_node rb_node;
130         /* service_tree key */
131         unsigned long rb_key;
132         /* sorted list of pending requests */
133         struct rb_root sort_list;
134         /* if fifo isn't expired, next request to serve */
135         struct request *next_rq;
136         /* requests queued in sort_list */
137         int queued[2];
138         /* currently allocated requests */
139         int allocated[2];
140         /* pending metadata requests */
141         int meta_pending;
142         /* fifo list of requests in sort_list */
143         struct list_head fifo;
144
145         unsigned long slice_end;
146         long slice_resid;
147
148         /* number of requests that are on the dispatch list or inside driver */
149         int dispatched;
150
151         /* io prio of this group */
152         unsigned short ioprio, org_ioprio;
153         unsigned short ioprio_class, org_ioprio_class;
154
155         /* various state flags, see below */
156         unsigned int flags;
157 };
158
159 enum cfqq_state_flags {
160         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
161         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
162         CFQ_CFQQ_FLAG_must_alloc,       /* must be allowed rq alloc */
163         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
164         CFQ_CFQQ_FLAG_must_dispatch,    /* must dispatch, even if expired */
165         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
166         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
167         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
168         CFQ_CFQQ_FLAG_queue_new,        /* queue never been serviced */
169         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
170         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
171 };
172
173 #define CFQ_CFQQ_FNS(name)                                              \
174 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
175 {                                                                       \
176         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
177 }                                                                       \
178 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
179 {                                                                       \
180         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
181 }                                                                       \
182 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
183 {                                                                       \
184         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
185 }
186
187 CFQ_CFQQ_FNS(on_rr);
188 CFQ_CFQQ_FNS(wait_request);
189 CFQ_CFQQ_FNS(must_alloc);
190 CFQ_CFQQ_FNS(must_alloc_slice);
191 CFQ_CFQQ_FNS(must_dispatch);
192 CFQ_CFQQ_FNS(fifo_expire);
193 CFQ_CFQQ_FNS(idle_window);
194 CFQ_CFQQ_FNS(prio_changed);
195 CFQ_CFQQ_FNS(queue_new);
196 CFQ_CFQQ_FNS(slice_new);
197 CFQ_CFQQ_FNS(sync);
198 #undef CFQ_CFQQ_FNS
199
200 static void cfq_dispatch_insert(struct request_queue *, struct request *);
201 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
202                                        struct task_struct *, gfp_t);
203 static struct cfq_io_context *cfq_cic_rb_lookup(struct cfq_data *,
204                                                 struct io_context *);
205
206 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
207                                             int is_sync)
208 {
209         return cic->cfqq[!!is_sync];
210 }
211
212 static inline void cic_set_cfqq(struct cfq_io_context *cic,
213                                 struct cfq_queue *cfqq, int is_sync)
214 {
215         cic->cfqq[!!is_sync] = cfqq;
216 }
217
218 /*
219  * We regard a request as SYNC, if it's either a read or has the SYNC bit
220  * set (in which case it could also be direct WRITE).
221  */
222 static inline int cfq_bio_sync(struct bio *bio)
223 {
224         if (bio_data_dir(bio) == READ || bio_sync(bio))
225                 return 1;
226
227         return 0;
228 }
229
230 /*
231  * scheduler run of queue, if there are requests pending and no one in the
232  * driver that will restart queueing
233  */
234 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
235 {
236         if (cfqd->busy_queues)
237                 kblockd_schedule_work(&cfqd->unplug_work);
238 }
239
240 static int cfq_queue_empty(struct request_queue *q)
241 {
242         struct cfq_data *cfqd = q->elevator->elevator_data;
243
244         return !cfqd->busy_queues;
245 }
246
247 /*
248  * Scale schedule slice based on io priority. Use the sync time slice only
249  * if a queue is marked sync and has sync io queued. A sync queue with async
250  * io only, should not get full sync slice length.
251  */
252 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
253                                  unsigned short prio)
254 {
255         const int base_slice = cfqd->cfq_slice[sync];
256
257         WARN_ON(prio >= IOPRIO_BE_NR);
258
259         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
260 }
261
262 static inline int
263 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
264 {
265         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
266 }
267
268 static inline void
269 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
270 {
271         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
272 }
273
274 /*
275  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
276  * isn't valid until the first request from the dispatch is activated
277  * and the slice time set.
278  */
279 static inline int cfq_slice_used(struct cfq_queue *cfqq)
280 {
281         if (cfq_cfqq_slice_new(cfqq))
282                 return 0;
283         if (time_before(jiffies, cfqq->slice_end))
284                 return 0;
285
286         return 1;
287 }
288
289 /*
290  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
291  * We choose the request that is closest to the head right now. Distance
292  * behind the head is penalized and only allowed to a certain extent.
293  */
294 static struct request *
295 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
296 {
297         sector_t last, s1, s2, d1 = 0, d2 = 0;
298         unsigned long back_max;
299 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
300 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
301         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
302
303         if (rq1 == NULL || rq1 == rq2)
304                 return rq2;
305         if (rq2 == NULL)
306                 return rq1;
307
308         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
309                 return rq1;
310         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
311                 return rq2;
312         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
313                 return rq1;
314         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
315                 return rq2;
316
317         s1 = rq1->sector;
318         s2 = rq2->sector;
319
320         last = cfqd->last_position;
321
322         /*
323          * by definition, 1KiB is 2 sectors
324          */
325         back_max = cfqd->cfq_back_max * 2;
326
327         /*
328          * Strict one way elevator _except_ in the case where we allow
329          * short backward seeks which are biased as twice the cost of a
330          * similar forward seek.
331          */
332         if (s1 >= last)
333                 d1 = s1 - last;
334         else if (s1 + back_max >= last)
335                 d1 = (last - s1) * cfqd->cfq_back_penalty;
336         else
337                 wrap |= CFQ_RQ1_WRAP;
338
339         if (s2 >= last)
340                 d2 = s2 - last;
341         else if (s2 + back_max >= last)
342                 d2 = (last - s2) * cfqd->cfq_back_penalty;
343         else
344                 wrap |= CFQ_RQ2_WRAP;
345
346         /* Found required data */
347
348         /*
349          * By doing switch() on the bit mask "wrap" we avoid having to
350          * check two variables for all permutations: --> faster!
351          */
352         switch (wrap) {
353         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
354                 if (d1 < d2)
355                         return rq1;
356                 else if (d2 < d1)
357                         return rq2;
358                 else {
359                         if (s1 >= s2)
360                                 return rq1;
361                         else
362                                 return rq2;
363                 }
364
365         case CFQ_RQ2_WRAP:
366                 return rq1;
367         case CFQ_RQ1_WRAP:
368                 return rq2;
369         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
370         default:
371                 /*
372                  * Since both rqs are wrapped,
373                  * start with the one that's further behind head
374                  * (--> only *one* back seek required),
375                  * since back seek takes more time than forward.
376                  */
377                 if (s1 <= s2)
378                         return rq1;
379                 else
380                         return rq2;
381         }
382 }
383
384 /*
385  * The below is leftmost cache rbtree addon
386  */
387 static struct rb_node *cfq_rb_first(struct cfq_rb_root *root)
388 {
389         if (!root->left)
390                 root->left = rb_first(&root->rb);
391
392         return root->left;
393 }
394
395 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
396 {
397         if (root->left == n)
398                 root->left = NULL;
399
400         rb_erase(n, &root->rb);
401         RB_CLEAR_NODE(n);
402 }
403
404 /*
405  * would be nice to take fifo expire time into account as well
406  */
407 static struct request *
408 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
409                   struct request *last)
410 {
411         struct rb_node *rbnext = rb_next(&last->rb_node);
412         struct rb_node *rbprev = rb_prev(&last->rb_node);
413         struct request *next = NULL, *prev = NULL;
414
415         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
416
417         if (rbprev)
418                 prev = rb_entry_rq(rbprev);
419
420         if (rbnext)
421                 next = rb_entry_rq(rbnext);
422         else {
423                 rbnext = rb_first(&cfqq->sort_list);
424                 if (rbnext && rbnext != &last->rb_node)
425                         next = rb_entry_rq(rbnext);
426         }
427
428         return cfq_choose_req(cfqd, next, prev);
429 }
430
431 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
432                                       struct cfq_queue *cfqq)
433 {
434         /*
435          * just an approximation, should be ok.
436          */
437         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
438                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
439 }
440
441 /*
442  * The cfqd->service_tree holds all pending cfq_queue's that have
443  * requests waiting to be processed. It is sorted in the order that
444  * we will service the queues.
445  */
446 static void cfq_service_tree_add(struct cfq_data *cfqd,
447                                     struct cfq_queue *cfqq, int add_front)
448 {
449         struct rb_node **p = &cfqd->service_tree.rb.rb_node;
450         struct rb_node *parent = NULL;
451         unsigned long rb_key;
452         int left;
453
454         if (!add_front) {
455                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
456                 rb_key += cfqq->slice_resid;
457                 cfqq->slice_resid = 0;
458         } else
459                 rb_key = 0;
460
461         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
462                 /*
463                  * same position, nothing more to do
464                  */
465                 if (rb_key == cfqq->rb_key)
466                         return;
467
468                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
469         }
470
471         left = 1;
472         while (*p) {
473                 struct cfq_queue *__cfqq;
474                 struct rb_node **n;
475
476                 parent = *p;
477                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
478
479                 /*
480                  * sort RT queues first, we always want to give
481                  * preference to them. IDLE queues goes to the back.
482                  * after that, sort on the next service time.
483                  */
484                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
485                         n = &(*p)->rb_left;
486                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
487                         n = &(*p)->rb_right;
488                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
489                         n = &(*p)->rb_left;
490                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
491                         n = &(*p)->rb_right;
492                 else if (rb_key < __cfqq->rb_key)
493                         n = &(*p)->rb_left;
494                 else
495                         n = &(*p)->rb_right;
496
497                 if (n == &(*p)->rb_right)
498                         left = 0;
499
500                 p = n;
501         }
502
503         if (left)
504                 cfqd->service_tree.left = &cfqq->rb_node;
505
506         cfqq->rb_key = rb_key;
507         rb_link_node(&cfqq->rb_node, parent, p);
508         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
509 }
510
511 /*
512  * Update cfqq's position in the service tree.
513  */
514 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
515 {
516         /*
517          * Resorting requires the cfqq to be on the RR list already.
518          */
519         if (cfq_cfqq_on_rr(cfqq))
520                 cfq_service_tree_add(cfqd, cfqq, 0);
521 }
522
523 /*
524  * add to busy list of queues for service, trying to be fair in ordering
525  * the pending list according to last request service
526  */
527 static inline void
528 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
529 {
530         BUG_ON(cfq_cfqq_on_rr(cfqq));
531         cfq_mark_cfqq_on_rr(cfqq);
532         cfqd->busy_queues++;
533
534         cfq_resort_rr_list(cfqd, cfqq);
535 }
536
537 /*
538  * Called when the cfqq no longer has requests pending, remove it from
539  * the service tree.
540  */
541 static inline void
542 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
543 {
544         BUG_ON(!cfq_cfqq_on_rr(cfqq));
545         cfq_clear_cfqq_on_rr(cfqq);
546
547         if (!RB_EMPTY_NODE(&cfqq->rb_node))
548                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
549
550         BUG_ON(!cfqd->busy_queues);
551         cfqd->busy_queues--;
552 }
553
554 /*
555  * rb tree support functions
556  */
557 static inline void cfq_del_rq_rb(struct request *rq)
558 {
559         struct cfq_queue *cfqq = RQ_CFQQ(rq);
560         struct cfq_data *cfqd = cfqq->cfqd;
561         const int sync = rq_is_sync(rq);
562
563         BUG_ON(!cfqq->queued[sync]);
564         cfqq->queued[sync]--;
565
566         elv_rb_del(&cfqq->sort_list, rq);
567
568         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
569                 cfq_del_cfqq_rr(cfqd, cfqq);
570 }
571
572 static void cfq_add_rq_rb(struct request *rq)
573 {
574         struct cfq_queue *cfqq = RQ_CFQQ(rq);
575         struct cfq_data *cfqd = cfqq->cfqd;
576         struct request *__alias;
577
578         cfqq->queued[rq_is_sync(rq)]++;
579
580         /*
581          * looks a little odd, but the first insert might return an alias.
582          * if that happens, put the alias on the dispatch list
583          */
584         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
585                 cfq_dispatch_insert(cfqd->queue, __alias);
586
587         if (!cfq_cfqq_on_rr(cfqq))
588                 cfq_add_cfqq_rr(cfqd, cfqq);
589
590         /*
591          * check if this request is a better next-serve candidate
592          */
593         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
594         BUG_ON(!cfqq->next_rq);
595 }
596
597 static inline void
598 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
599 {
600         elv_rb_del(&cfqq->sort_list, rq);
601         cfqq->queued[rq_is_sync(rq)]--;
602         cfq_add_rq_rb(rq);
603 }
604
605 static struct request *
606 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
607 {
608         struct task_struct *tsk = current;
609         struct cfq_io_context *cic;
610         struct cfq_queue *cfqq;
611
612         cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
613         if (!cic)
614                 return NULL;
615
616         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
617         if (cfqq) {
618                 sector_t sector = bio->bi_sector + bio_sectors(bio);
619
620                 return elv_rb_find(&cfqq->sort_list, sector);
621         }
622
623         return NULL;
624 }
625
626 static void cfq_activate_request(struct request_queue *q, struct request *rq)
627 {
628         struct cfq_data *cfqd = q->elevator->elevator_data;
629
630         cfqd->rq_in_driver++;
631
632         /*
633          * If the depth is larger 1, it really could be queueing. But lets
634          * make the mark a little higher - idling could still be good for
635          * low queueing, and a low queueing number could also just indicate
636          * a SCSI mid layer like behaviour where limit+1 is often seen.
637          */
638         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
639                 cfqd->hw_tag = 1;
640
641         cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
642 }
643
644 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
645 {
646         struct cfq_data *cfqd = q->elevator->elevator_data;
647
648         WARN_ON(!cfqd->rq_in_driver);
649         cfqd->rq_in_driver--;
650 }
651
652 static void cfq_remove_request(struct request *rq)
653 {
654         struct cfq_queue *cfqq = RQ_CFQQ(rq);
655
656         if (cfqq->next_rq == rq)
657                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
658
659         list_del_init(&rq->queuelist);
660         cfq_del_rq_rb(rq);
661
662         if (rq_is_meta(rq)) {
663                 WARN_ON(!cfqq->meta_pending);
664                 cfqq->meta_pending--;
665         }
666 }
667
668 static int cfq_merge(struct request_queue *q, struct request **req,
669                      struct bio *bio)
670 {
671         struct cfq_data *cfqd = q->elevator->elevator_data;
672         struct request *__rq;
673
674         __rq = cfq_find_rq_fmerge(cfqd, bio);
675         if (__rq && elv_rq_merge_ok(__rq, bio)) {
676                 *req = __rq;
677                 return ELEVATOR_FRONT_MERGE;
678         }
679
680         return ELEVATOR_NO_MERGE;
681 }
682
683 static void cfq_merged_request(struct request_queue *q, struct request *req,
684                                int type)
685 {
686         if (type == ELEVATOR_FRONT_MERGE) {
687                 struct cfq_queue *cfqq = RQ_CFQQ(req);
688
689                 cfq_reposition_rq_rb(cfqq, req);
690         }
691 }
692
693 static void
694 cfq_merged_requests(struct request_queue *q, struct request *rq,
695                     struct request *next)
696 {
697         /*
698          * reposition in fifo if next is older than rq
699          */
700         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
701             time_before(next->start_time, rq->start_time))
702                 list_move(&rq->queuelist, &next->queuelist);
703
704         cfq_remove_request(next);
705 }
706
707 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
708                            struct bio *bio)
709 {
710         struct cfq_data *cfqd = q->elevator->elevator_data;
711         struct cfq_io_context *cic;
712         struct cfq_queue *cfqq;
713
714         /*
715          * Disallow merge of a sync bio into an async request.
716          */
717         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
718                 return 0;
719
720         /*
721          * Lookup the cfqq that this bio will be queued with. Allow
722          * merge only if rq is queued there.
723          */
724         cic = cfq_cic_rb_lookup(cfqd, current->io_context);
725         if (!cic)
726                 return 0;
727
728         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
729         if (cfqq == RQ_CFQQ(rq))
730                 return 1;
731
732         return 0;
733 }
734
735 static inline void
736 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
737 {
738         if (cfqq) {
739                 /*
740                  * stop potential idle class queues waiting service
741                  */
742                 del_timer(&cfqd->idle_class_timer);
743
744                 cfqq->slice_end = 0;
745                 cfq_clear_cfqq_must_alloc_slice(cfqq);
746                 cfq_clear_cfqq_fifo_expire(cfqq);
747                 cfq_mark_cfqq_slice_new(cfqq);
748                 cfq_clear_cfqq_queue_new(cfqq);
749         }
750
751         cfqd->active_queue = cfqq;
752 }
753
754 /*
755  * current cfqq expired its slice (or was too idle), select new one
756  */
757 static void
758 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
759                     int timed_out)
760 {
761         if (cfq_cfqq_wait_request(cfqq))
762                 del_timer(&cfqd->idle_slice_timer);
763
764         cfq_clear_cfqq_must_dispatch(cfqq);
765         cfq_clear_cfqq_wait_request(cfqq);
766
767         /*
768          * store what was left of this slice, if the queue idled/timed out
769          */
770         if (timed_out && !cfq_cfqq_slice_new(cfqq))
771                 cfqq->slice_resid = cfqq->slice_end - jiffies;
772
773         cfq_resort_rr_list(cfqd, cfqq);
774
775         if (cfqq == cfqd->active_queue)
776                 cfqd->active_queue = NULL;
777
778         if (cfqd->active_cic) {
779                 put_io_context(cfqd->active_cic->ioc);
780                 cfqd->active_cic = NULL;
781         }
782 }
783
784 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
785 {
786         struct cfq_queue *cfqq = cfqd->active_queue;
787
788         if (cfqq)
789                 __cfq_slice_expired(cfqd, cfqq, timed_out);
790 }
791
792 static int start_idle_class_timer(struct cfq_data *cfqd)
793 {
794         unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
795         unsigned long now = jiffies;
796
797         if (time_before(now, end) &&
798             time_after_eq(now, cfqd->last_end_request)) {
799                 mod_timer(&cfqd->idle_class_timer, end);
800                 return 1;
801         }
802
803         return 0;
804 }
805
806 /*
807  * Get next queue for service. Unless we have a queue preemption,
808  * we'll simply select the first cfqq in the service tree.
809  */
810 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
811 {
812         struct cfq_queue *cfqq;
813         struct rb_node *n;
814
815         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
816                 return NULL;
817
818         n = cfq_rb_first(&cfqd->service_tree);
819         cfqq = rb_entry(n, struct cfq_queue, rb_node);
820
821         if (cfq_class_idle(cfqq)) {
822                 /*
823                  * if we have idle queues and no rt or be queues had
824                  * pending requests, either allow immediate service if
825                  * the grace period has passed or arm the idle grace
826                  * timer
827                  */
828                 if (start_idle_class_timer(cfqd))
829                         cfqq = NULL;
830         }
831
832         return cfqq;
833 }
834
835 /*
836  * Get and set a new active queue for service.
837  */
838 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
839 {
840         struct cfq_queue *cfqq;
841
842         cfqq = cfq_get_next_queue(cfqd);
843         __cfq_set_active_queue(cfqd, cfqq);
844         return cfqq;
845 }
846
847 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
848                                           struct request *rq)
849 {
850         if (rq->sector >= cfqd->last_position)
851                 return rq->sector - cfqd->last_position;
852         else
853                 return cfqd->last_position - rq->sector;
854 }
855
856 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
857 {
858         struct cfq_io_context *cic = cfqd->active_cic;
859
860         if (!sample_valid(cic->seek_samples))
861                 return 0;
862
863         return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
864 }
865
866 static int cfq_close_cooperator(struct cfq_data *cfq_data,
867                                 struct cfq_queue *cfqq)
868 {
869         /*
870          * We should notice if some of the queues are cooperating, eg
871          * working closely on the same area of the disk. In that case,
872          * we can group them together and don't waste time idling.
873          */
874         return 0;
875 }
876
877 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
878
879 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
880 {
881         struct cfq_queue *cfqq = cfqd->active_queue;
882         struct cfq_io_context *cic;
883         unsigned long sl;
884
885         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
886         WARN_ON(cfq_cfqq_slice_new(cfqq));
887
888         /*
889          * idle is disabled, either manually or by past process history
890          */
891         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
892                 return;
893
894         /*
895          * task has exited, don't wait
896          */
897         cic = cfqd->active_cic;
898         if (!cic || !cic->ioc->task)
899                 return;
900
901         /*
902          * See if this prio level has a good candidate
903          */
904         if (cfq_close_cooperator(cfqd, cfqq) &&
905             (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
906                 return;
907
908         cfq_mark_cfqq_must_dispatch(cfqq);
909         cfq_mark_cfqq_wait_request(cfqq);
910
911         /*
912          * we don't want to idle for seeks, but we do want to allow
913          * fair distribution of slice time for a process doing back-to-back
914          * seeks. so allow a little bit of time for him to submit a new rq
915          */
916         sl = cfqd->cfq_slice_idle;
917         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
918                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
919
920         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
921 }
922
923 /*
924  * Move request from internal lists to the request queue dispatch list.
925  */
926 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
927 {
928         struct cfq_data *cfqd = q->elevator->elevator_data;
929         struct cfq_queue *cfqq = RQ_CFQQ(rq);
930
931         cfq_remove_request(rq);
932         cfqq->dispatched++;
933         elv_dispatch_sort(q, rq);
934
935         if (cfq_cfqq_sync(cfqq))
936                 cfqd->sync_flight++;
937 }
938
939 /*
940  * return expired entry, or NULL to just start from scratch in rbtree
941  */
942 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
943 {
944         struct cfq_data *cfqd = cfqq->cfqd;
945         struct request *rq;
946         int fifo;
947
948         if (cfq_cfqq_fifo_expire(cfqq))
949                 return NULL;
950
951         cfq_mark_cfqq_fifo_expire(cfqq);
952
953         if (list_empty(&cfqq->fifo))
954                 return NULL;
955
956         fifo = cfq_cfqq_sync(cfqq);
957         rq = rq_entry_fifo(cfqq->fifo.next);
958
959         if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
960                 return NULL;
961
962         return rq;
963 }
964
965 static inline int
966 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
967 {
968         const int base_rq = cfqd->cfq_slice_async_rq;
969
970         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
971
972         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
973 }
974
975 /*
976  * Select a queue for service. If we have a current active queue,
977  * check whether to continue servicing it, or retrieve and set a new one.
978  */
979 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
980 {
981         struct cfq_queue *cfqq;
982
983         cfqq = cfqd->active_queue;
984         if (!cfqq)
985                 goto new_queue;
986
987         /*
988          * The active queue has run out of time, expire it and select new.
989          */
990         if (cfq_slice_used(cfqq))
991                 goto expire;
992
993         /*
994          * The active queue has requests and isn't expired, allow it to
995          * dispatch.
996          */
997         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
998                 goto keep_queue;
999
1000         /*
1001          * No requests pending. If the active queue still has requests in
1002          * flight or is idling for a new request, allow either of these
1003          * conditions to happen (or time out) before selecting a new queue.
1004          */
1005         if (timer_pending(&cfqd->idle_slice_timer) ||
1006             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1007                 cfqq = NULL;
1008                 goto keep_queue;
1009         }
1010
1011 expire:
1012         cfq_slice_expired(cfqd, 0);
1013 new_queue:
1014         cfqq = cfq_set_active_queue(cfqd);
1015 keep_queue:
1016         return cfqq;
1017 }
1018
1019 /*
1020  * Dispatch some requests from cfqq, moving them to the request queue
1021  * dispatch list.
1022  */
1023 static int
1024 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1025                         int max_dispatch)
1026 {
1027         int dispatched = 0;
1028
1029         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1030
1031         do {
1032                 struct request *rq;
1033
1034                 /*
1035                  * follow expired path, else get first next available
1036                  */
1037                 if ((rq = cfq_check_fifo(cfqq)) == NULL)
1038                         rq = cfqq->next_rq;
1039
1040                 /*
1041                  * finally, insert request into driver dispatch list
1042                  */
1043                 cfq_dispatch_insert(cfqd->queue, rq);
1044
1045                 dispatched++;
1046
1047                 if (!cfqd->active_cic) {
1048                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1049                         cfqd->active_cic = RQ_CIC(rq);
1050                 }
1051
1052                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
1053                         break;
1054
1055         } while (dispatched < max_dispatch);
1056
1057         /*
1058          * expire an async queue immediately if it has used up its slice. idle
1059          * queue always expire after 1 dispatch round.
1060          */
1061         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1062             dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1063             cfq_class_idle(cfqq))) {
1064                 cfqq->slice_end = jiffies + 1;
1065                 cfq_slice_expired(cfqd, 0);
1066         }
1067
1068         return dispatched;
1069 }
1070
1071 static inline int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1072 {
1073         int dispatched = 0;
1074
1075         while (cfqq->next_rq) {
1076                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1077                 dispatched++;
1078         }
1079
1080         BUG_ON(!list_empty(&cfqq->fifo));
1081         return dispatched;
1082 }
1083
1084 /*
1085  * Drain our current requests. Used for barriers and when switching
1086  * io schedulers on-the-fly.
1087  */
1088 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1089 {
1090         int dispatched = 0;
1091         struct rb_node *n;
1092
1093         while ((n = cfq_rb_first(&cfqd->service_tree)) != NULL) {
1094                 struct cfq_queue *cfqq = rb_entry(n, struct cfq_queue, rb_node);
1095
1096                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1097         }
1098
1099         cfq_slice_expired(cfqd, 0);
1100
1101         BUG_ON(cfqd->busy_queues);
1102
1103         return dispatched;
1104 }
1105
1106 static int cfq_dispatch_requests(struct request_queue *q, int force)
1107 {
1108         struct cfq_data *cfqd = q->elevator->elevator_data;
1109         struct cfq_queue *cfqq;
1110         int dispatched;
1111
1112         if (!cfqd->busy_queues)
1113                 return 0;
1114
1115         if (unlikely(force))
1116                 return cfq_forced_dispatch(cfqd);
1117
1118         dispatched = 0;
1119         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1120                 int max_dispatch;
1121
1122                 max_dispatch = cfqd->cfq_quantum;
1123                 if (cfq_class_idle(cfqq))
1124                         max_dispatch = 1;
1125
1126                 if (cfqq->dispatched >= max_dispatch) {
1127                         if (cfqd->busy_queues > 1)
1128                                 break;
1129                         if (cfqq->dispatched >= 4 * max_dispatch)
1130                                 break;
1131                 }
1132
1133                 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1134                         break;
1135
1136                 cfq_clear_cfqq_must_dispatch(cfqq);
1137                 cfq_clear_cfqq_wait_request(cfqq);
1138                 del_timer(&cfqd->idle_slice_timer);
1139
1140                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1141         }
1142
1143         return dispatched;
1144 }
1145
1146 /*
1147  * task holds one reference to the queue, dropped when task exits. each rq
1148  * in-flight on this queue also holds a reference, dropped when rq is freed.
1149  *
1150  * queue lock must be held here.
1151  */
1152 static void cfq_put_queue(struct cfq_queue *cfqq)
1153 {
1154         struct cfq_data *cfqd = cfqq->cfqd;
1155
1156         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1157
1158         if (!atomic_dec_and_test(&cfqq->ref))
1159                 return;
1160
1161         BUG_ON(rb_first(&cfqq->sort_list));
1162         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1163         BUG_ON(cfq_cfqq_on_rr(cfqq));
1164
1165         if (unlikely(cfqd->active_queue == cfqq)) {
1166                 __cfq_slice_expired(cfqd, cfqq, 0);
1167                 cfq_schedule_dispatch(cfqd);
1168         }
1169
1170         kmem_cache_free(cfq_pool, cfqq);
1171 }
1172
1173 static void cfq_free_io_context(struct io_context *ioc)
1174 {
1175         struct cfq_io_context *__cic;
1176         struct rb_node *n;
1177         int freed = 0;
1178
1179         ioc->ioc_data = NULL;
1180
1181         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1182                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1183                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1184                 kmem_cache_free(cfq_ioc_pool, __cic);
1185                 freed++;
1186         }
1187
1188         elv_ioc_count_mod(ioc_count, -freed);
1189
1190         if (ioc_gone && !elv_ioc_count_read(ioc_count))
1191                 complete(ioc_gone);
1192 }
1193
1194 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1195 {
1196         if (unlikely(cfqq == cfqd->active_queue)) {
1197                 __cfq_slice_expired(cfqd, cfqq, 0);
1198                 cfq_schedule_dispatch(cfqd);
1199         }
1200
1201         cfq_put_queue(cfqq);
1202 }
1203
1204 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1205                                          struct cfq_io_context *cic)
1206 {
1207         list_del_init(&cic->queue_list);
1208         smp_wmb();
1209         cic->key = NULL;
1210
1211         if (cic->cfqq[ASYNC]) {
1212                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1213                 cic->cfqq[ASYNC] = NULL;
1214         }
1215
1216         if (cic->cfqq[SYNC]) {
1217                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1218                 cic->cfqq[SYNC] = NULL;
1219         }
1220 }
1221
1222 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1223 {
1224         struct cfq_data *cfqd = cic->key;
1225
1226         if (cfqd) {
1227                 struct request_queue *q = cfqd->queue;
1228
1229                 spin_lock_irq(q->queue_lock);
1230                 __cfq_exit_single_io_context(cfqd, cic);
1231                 spin_unlock_irq(q->queue_lock);
1232         }
1233 }
1234
1235 /*
1236  * The process that ioc belongs to has exited, we need to clean up
1237  * and put the internal structures we have that belongs to that process.
1238  */
1239 static void cfq_exit_io_context(struct io_context *ioc)
1240 {
1241         struct cfq_io_context *__cic;
1242         struct rb_node *n;
1243
1244         ioc->ioc_data = NULL;
1245
1246         /*
1247          * put the reference this task is holding to the various queues
1248          */
1249         n = rb_first(&ioc->cic_root);
1250         while (n != NULL) {
1251                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1252
1253                 cfq_exit_single_io_context(__cic);
1254                 n = rb_next(n);
1255         }
1256 }
1257
1258 static struct cfq_io_context *
1259 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1260 {
1261         struct cfq_io_context *cic;
1262
1263         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1264                                                         cfqd->queue->node);
1265         if (cic) {
1266                 cic->last_end_request = jiffies;
1267                 INIT_LIST_HEAD(&cic->queue_list);
1268                 cic->dtor = cfq_free_io_context;
1269                 cic->exit = cfq_exit_io_context;
1270                 elv_ioc_count_inc(ioc_count);
1271         }
1272
1273         return cic;
1274 }
1275
1276 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1277 {
1278         struct task_struct *tsk = current;
1279         int ioprio_class;
1280
1281         if (!cfq_cfqq_prio_changed(cfqq))
1282                 return;
1283
1284         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1285         switch (ioprio_class) {
1286                 default:
1287                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1288                 case IOPRIO_CLASS_NONE:
1289                         /*
1290                          * no prio set, place us in the middle of the BE classes
1291                          */
1292                         cfqq->ioprio = task_nice_ioprio(tsk);
1293                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1294                         break;
1295                 case IOPRIO_CLASS_RT:
1296                         cfqq->ioprio = task_ioprio(tsk);
1297                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1298                         break;
1299                 case IOPRIO_CLASS_BE:
1300                         cfqq->ioprio = task_ioprio(tsk);
1301                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1302                         break;
1303                 case IOPRIO_CLASS_IDLE:
1304                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1305                         cfqq->ioprio = 7;
1306                         cfq_clear_cfqq_idle_window(cfqq);
1307                         break;
1308         }
1309
1310         /*
1311          * keep track of original prio settings in case we have to temporarily
1312          * elevate the priority of this queue
1313          */
1314         cfqq->org_ioprio = cfqq->ioprio;
1315         cfqq->org_ioprio_class = cfqq->ioprio_class;
1316         cfq_clear_cfqq_prio_changed(cfqq);
1317 }
1318
1319 static inline void changed_ioprio(struct cfq_io_context *cic)
1320 {
1321         struct cfq_data *cfqd = cic->key;
1322         struct cfq_queue *cfqq;
1323         unsigned long flags;
1324
1325         if (unlikely(!cfqd))
1326                 return;
1327
1328         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1329
1330         cfqq = cic->cfqq[ASYNC];
1331         if (cfqq) {
1332                 struct cfq_queue *new_cfqq;
1333                 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc->task,
1334                                          GFP_ATOMIC);
1335                 if (new_cfqq) {
1336                         cic->cfqq[ASYNC] = new_cfqq;
1337                         cfq_put_queue(cfqq);
1338                 }
1339         }
1340
1341         cfqq = cic->cfqq[SYNC];
1342         if (cfqq)
1343                 cfq_mark_cfqq_prio_changed(cfqq);
1344
1345         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1346 }
1347
1348 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1349 {
1350         struct cfq_io_context *cic;
1351         struct rb_node *n;
1352
1353         ioc->ioprio_changed = 0;
1354
1355         n = rb_first(&ioc->cic_root);
1356         while (n != NULL) {
1357                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1358
1359                 changed_ioprio(cic);
1360                 n = rb_next(n);
1361         }
1362 }
1363
1364 static struct cfq_queue *
1365 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1366                      struct task_struct *tsk, gfp_t gfp_mask)
1367 {
1368         struct cfq_queue *cfqq, *new_cfqq = NULL;
1369         struct cfq_io_context *cic;
1370
1371 retry:
1372         cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
1373         /* cic always exists here */
1374         cfqq = cic_to_cfqq(cic, is_sync);
1375
1376         if (!cfqq) {
1377                 if (new_cfqq) {
1378                         cfqq = new_cfqq;
1379                         new_cfqq = NULL;
1380                 } else if (gfp_mask & __GFP_WAIT) {
1381                         /*
1382                          * Inform the allocator of the fact that we will
1383                          * just repeat this allocation if it fails, to allow
1384                          * the allocator to do whatever it needs to attempt to
1385                          * free memory.
1386                          */
1387                         spin_unlock_irq(cfqd->queue->queue_lock);
1388                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1389                                         gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1390                                         cfqd->queue->node);
1391                         spin_lock_irq(cfqd->queue->queue_lock);
1392                         goto retry;
1393                 } else {
1394                         cfqq = kmem_cache_alloc_node(cfq_pool,
1395                                         gfp_mask | __GFP_ZERO,
1396                                         cfqd->queue->node);
1397                         if (!cfqq)
1398                                 goto out;
1399                 }
1400
1401                 RB_CLEAR_NODE(&cfqq->rb_node);
1402                 INIT_LIST_HEAD(&cfqq->fifo);
1403
1404                 atomic_set(&cfqq->ref, 0);
1405                 cfqq->cfqd = cfqd;
1406
1407                 if (is_sync) {
1408                         cfq_mark_cfqq_idle_window(cfqq);
1409                         cfq_mark_cfqq_sync(cfqq);
1410                 }
1411
1412                 cfq_mark_cfqq_prio_changed(cfqq);
1413                 cfq_mark_cfqq_queue_new(cfqq);
1414
1415                 cfq_init_prio_data(cfqq);
1416         }
1417
1418         if (new_cfqq)
1419                 kmem_cache_free(cfq_pool, new_cfqq);
1420
1421 out:
1422         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1423         return cfqq;
1424 }
1425
1426 static struct cfq_queue **
1427 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1428 {
1429         switch(ioprio_class) {
1430         case IOPRIO_CLASS_RT:
1431                 return &cfqd->async_cfqq[0][ioprio];
1432         case IOPRIO_CLASS_BE:
1433                 return &cfqd->async_cfqq[1][ioprio];
1434         case IOPRIO_CLASS_IDLE:
1435                 return &cfqd->async_idle_cfqq;
1436         default:
1437                 BUG();
1438         }
1439 }
1440
1441 static struct cfq_queue *
1442 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct task_struct *tsk,
1443               gfp_t gfp_mask)
1444 {
1445         const int ioprio = task_ioprio(tsk);
1446         const int ioprio_class = task_ioprio_class(tsk);
1447         struct cfq_queue **async_cfqq = NULL;
1448         struct cfq_queue *cfqq = NULL;
1449
1450         if (!is_sync) {
1451                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1452                 cfqq = *async_cfqq;
1453         }
1454
1455         if (!cfqq) {
1456                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, tsk, gfp_mask);
1457                 if (!cfqq)
1458                         return NULL;
1459         }
1460
1461         /*
1462          * pin the queue now that it's allocated, scheduler exit will prune it
1463          */
1464         if (!is_sync && !(*async_cfqq)) {
1465                 atomic_inc(&cfqq->ref);
1466                 *async_cfqq = cfqq;
1467         }
1468
1469         atomic_inc(&cfqq->ref);
1470         return cfqq;
1471 }
1472
1473 /*
1474  * We drop cfq io contexts lazily, so we may find a dead one.
1475  */
1476 static void
1477 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1478 {
1479         WARN_ON(!list_empty(&cic->queue_list));
1480
1481         if (ioc->ioc_data == cic)
1482                 ioc->ioc_data = NULL;
1483
1484         rb_erase(&cic->rb_node, &ioc->cic_root);
1485         kmem_cache_free(cfq_ioc_pool, cic);
1486         elv_ioc_count_dec(ioc_count);
1487 }
1488
1489 static struct cfq_io_context *
1490 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1491 {
1492         struct rb_node *n;
1493         struct cfq_io_context *cic;
1494         void *k, *key = cfqd;
1495
1496         if (unlikely(!ioc))
1497                 return NULL;
1498
1499         /*
1500          * we maintain a last-hit cache, to avoid browsing over the tree
1501          */
1502         cic = ioc->ioc_data;
1503         if (cic && cic->key == cfqd)
1504                 return cic;
1505
1506 restart:
1507         n = ioc->cic_root.rb_node;
1508         while (n) {
1509                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1510                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1511                 k = cic->key;
1512                 if (unlikely(!k)) {
1513                         cfq_drop_dead_cic(ioc, cic);
1514                         goto restart;
1515                 }
1516
1517                 if (key < k)
1518                         n = n->rb_left;
1519                 else if (key > k)
1520                         n = n->rb_right;
1521                 else {
1522                         ioc->ioc_data = cic;
1523                         return cic;
1524                 }
1525         }
1526
1527         return NULL;
1528 }
1529
1530 static inline void
1531 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1532              struct cfq_io_context *cic)
1533 {
1534         struct rb_node **p;
1535         struct rb_node *parent;
1536         struct cfq_io_context *__cic;
1537         unsigned long flags;
1538         void *k;
1539
1540         cic->ioc = ioc;
1541         cic->key = cfqd;
1542
1543 restart:
1544         parent = NULL;
1545         p = &ioc->cic_root.rb_node;
1546         while (*p) {
1547                 parent = *p;
1548                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1549                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1550                 k = __cic->key;
1551                 if (unlikely(!k)) {
1552                         cfq_drop_dead_cic(ioc, __cic);
1553                         goto restart;
1554                 }
1555
1556                 if (cic->key < k)
1557                         p = &(*p)->rb_left;
1558                 else if (cic->key > k)
1559                         p = &(*p)->rb_right;
1560                 else
1561                         BUG();
1562         }
1563
1564         rb_link_node(&cic->rb_node, parent, p);
1565         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1566
1567         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1568         list_add(&cic->queue_list, &cfqd->cic_list);
1569         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1570 }
1571
1572 /*
1573  * Setup general io context and cfq io context. There can be several cfq
1574  * io contexts per general io context, if this process is doing io to more
1575  * than one device managed by cfq.
1576  */
1577 static struct cfq_io_context *
1578 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1579 {
1580         struct io_context *ioc = NULL;
1581         struct cfq_io_context *cic;
1582
1583         might_sleep_if(gfp_mask & __GFP_WAIT);
1584
1585         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1586         if (!ioc)
1587                 return NULL;
1588
1589         cic = cfq_cic_rb_lookup(cfqd, ioc);
1590         if (cic)
1591                 goto out;
1592
1593         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1594         if (cic == NULL)
1595                 goto err;
1596
1597         cfq_cic_link(cfqd, ioc, cic);
1598 out:
1599         smp_read_barrier_depends();
1600         if (unlikely(ioc->ioprio_changed))
1601                 cfq_ioc_set_ioprio(ioc);
1602
1603         return cic;
1604 err:
1605         put_io_context(ioc);
1606         return NULL;
1607 }
1608
1609 static void
1610 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1611 {
1612         unsigned long elapsed = jiffies - cic->last_end_request;
1613         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1614
1615         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1616         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1617         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1618 }
1619
1620 static void
1621 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1622                        struct request *rq)
1623 {
1624         sector_t sdist;
1625         u64 total;
1626
1627         if (cic->last_request_pos < rq->sector)
1628                 sdist = rq->sector - cic->last_request_pos;
1629         else
1630                 sdist = cic->last_request_pos - rq->sector;
1631
1632         /*
1633          * Don't allow the seek distance to get too large from the
1634          * odd fragment, pagein, etc
1635          */
1636         if (cic->seek_samples <= 60) /* second&third seek */
1637                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1638         else
1639                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1640
1641         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1642         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1643         total = cic->seek_total + (cic->seek_samples/2);
1644         do_div(total, cic->seek_samples);
1645         cic->seek_mean = (sector_t)total;
1646 }
1647
1648 /*
1649  * Disable idle window if the process thinks too long or seeks so much that
1650  * it doesn't matter
1651  */
1652 static void
1653 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1654                        struct cfq_io_context *cic)
1655 {
1656         int enable_idle;
1657
1658         if (!cfq_cfqq_sync(cfqq))
1659                 return;
1660
1661         enable_idle = cfq_cfqq_idle_window(cfqq);
1662
1663         if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1664             (cfqd->hw_tag && CIC_SEEKY(cic)))
1665                 enable_idle = 0;
1666         else if (sample_valid(cic->ttime_samples)) {
1667                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1668                         enable_idle = 0;
1669                 else
1670                         enable_idle = 1;
1671         }
1672
1673         if (enable_idle)
1674                 cfq_mark_cfqq_idle_window(cfqq);
1675         else
1676                 cfq_clear_cfqq_idle_window(cfqq);
1677 }
1678
1679 /*
1680  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1681  * no or if we aren't sure, a 1 will cause a preempt.
1682  */
1683 static int
1684 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1685                    struct request *rq)
1686 {
1687         struct cfq_queue *cfqq;
1688
1689         cfqq = cfqd->active_queue;
1690         if (!cfqq)
1691                 return 0;
1692
1693         if (cfq_slice_used(cfqq))
1694                 return 1;
1695
1696         if (cfq_class_idle(new_cfqq))
1697                 return 0;
1698
1699         if (cfq_class_idle(cfqq))
1700                 return 1;
1701
1702         /*
1703          * if the new request is sync, but the currently running queue is
1704          * not, let the sync request have priority.
1705          */
1706         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1707                 return 1;
1708
1709         /*
1710          * So both queues are sync. Let the new request get disk time if
1711          * it's a metadata request and the current queue is doing regular IO.
1712          */
1713         if (rq_is_meta(rq) && !cfqq->meta_pending)
1714                 return 1;
1715
1716         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1717                 return 0;
1718
1719         /*
1720          * if this request is as-good as one we would expect from the
1721          * current cfqq, let it preempt
1722          */
1723         if (cfq_rq_close(cfqd, rq))
1724                 return 1;
1725
1726         return 0;
1727 }
1728
1729 /*
1730  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1731  * let it have half of its nominal slice.
1732  */
1733 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1734 {
1735         cfq_slice_expired(cfqd, 1);
1736
1737         /*
1738          * Put the new queue at the front of the of the current list,
1739          * so we know that it will be selected next.
1740          */
1741         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1742
1743         cfq_service_tree_add(cfqd, cfqq, 1);
1744
1745         cfqq->slice_end = 0;
1746         cfq_mark_cfqq_slice_new(cfqq);
1747 }
1748
1749 /*
1750  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1751  * something we should do about it
1752  */
1753 static void
1754 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1755                 struct request *rq)
1756 {
1757         struct cfq_io_context *cic = RQ_CIC(rq);
1758
1759         if (rq_is_meta(rq))
1760                 cfqq->meta_pending++;
1761
1762         cfq_update_io_thinktime(cfqd, cic);
1763         cfq_update_io_seektime(cfqd, cic, rq);
1764         cfq_update_idle_window(cfqd, cfqq, cic);
1765
1766         cic->last_request_pos = rq->sector + rq->nr_sectors;
1767
1768         if (cfqq == cfqd->active_queue) {
1769                 /*
1770                  * if we are waiting for a request for this queue, let it rip
1771                  * immediately and flag that we must not expire this queue
1772                  * just now
1773                  */
1774                 if (cfq_cfqq_wait_request(cfqq)) {
1775                         cfq_mark_cfqq_must_dispatch(cfqq);
1776                         del_timer(&cfqd->idle_slice_timer);
1777                         blk_start_queueing(cfqd->queue);
1778                 }
1779         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1780                 /*
1781                  * not the active queue - expire current slice if it is
1782                  * idle and has expired it's mean thinktime or this new queue
1783                  * has some old slice time left and is of higher priority
1784                  */
1785                 cfq_preempt_queue(cfqd, cfqq);
1786                 cfq_mark_cfqq_must_dispatch(cfqq);
1787                 blk_start_queueing(cfqd->queue);
1788         }
1789 }
1790
1791 static void cfq_insert_request(struct request_queue *q, struct request *rq)
1792 {
1793         struct cfq_data *cfqd = q->elevator->elevator_data;
1794         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1795
1796         cfq_init_prio_data(cfqq);
1797
1798         cfq_add_rq_rb(rq);
1799
1800         list_add_tail(&rq->queuelist, &cfqq->fifo);
1801
1802         cfq_rq_enqueued(cfqd, cfqq, rq);
1803 }
1804
1805 static void cfq_completed_request(struct request_queue *q, struct request *rq)
1806 {
1807         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1808         struct cfq_data *cfqd = cfqq->cfqd;
1809         const int sync = rq_is_sync(rq);
1810         unsigned long now;
1811
1812         now = jiffies;
1813
1814         WARN_ON(!cfqd->rq_in_driver);
1815         WARN_ON(!cfqq->dispatched);
1816         cfqd->rq_in_driver--;
1817         cfqq->dispatched--;
1818
1819         if (cfq_cfqq_sync(cfqq))
1820                 cfqd->sync_flight--;
1821
1822         if (!cfq_class_idle(cfqq))
1823                 cfqd->last_end_request = now;
1824
1825         if (sync)
1826                 RQ_CIC(rq)->last_end_request = now;
1827
1828         /*
1829          * If this is the active queue, check if it needs to be expired,
1830          * or if we want to idle in case it has no pending requests.
1831          */
1832         if (cfqd->active_queue == cfqq) {
1833                 if (cfq_cfqq_slice_new(cfqq)) {
1834                         cfq_set_prio_slice(cfqd, cfqq);
1835                         cfq_clear_cfqq_slice_new(cfqq);
1836                 }
1837                 if (cfq_slice_used(cfqq))
1838                         cfq_slice_expired(cfqd, 1);
1839                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1840                         cfq_arm_slice_timer(cfqd);
1841         }
1842
1843         if (!cfqd->rq_in_driver)
1844                 cfq_schedule_dispatch(cfqd);
1845 }
1846
1847 /*
1848  * we temporarily boost lower priority queues if they are holding fs exclusive
1849  * resources. they are boosted to normal prio (CLASS_BE/4)
1850  */
1851 static void cfq_prio_boost(struct cfq_queue *cfqq)
1852 {
1853         if (has_fs_excl()) {
1854                 /*
1855                  * boost idle prio on transactions that would lock out other
1856                  * users of the filesystem
1857                  */
1858                 if (cfq_class_idle(cfqq))
1859                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1860                 if (cfqq->ioprio > IOPRIO_NORM)
1861                         cfqq->ioprio = IOPRIO_NORM;
1862         } else {
1863                 /*
1864                  * check if we need to unboost the queue
1865                  */
1866                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1867                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1868                 if (cfqq->ioprio != cfqq->org_ioprio)
1869                         cfqq->ioprio = cfqq->org_ioprio;
1870         }
1871 }
1872
1873 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1874 {
1875         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1876             !cfq_cfqq_must_alloc_slice(cfqq)) {
1877                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1878                 return ELV_MQUEUE_MUST;
1879         }
1880
1881         return ELV_MQUEUE_MAY;
1882 }
1883
1884 static int cfq_may_queue(struct request_queue *q, int rw)
1885 {
1886         struct cfq_data *cfqd = q->elevator->elevator_data;
1887         struct task_struct *tsk = current;
1888         struct cfq_io_context *cic;
1889         struct cfq_queue *cfqq;
1890
1891         /*
1892          * don't force setup of a queue from here, as a call to may_queue
1893          * does not necessarily imply that a request actually will be queued.
1894          * so just lookup a possibly existing queue, or return 'may queue'
1895          * if that fails
1896          */
1897         cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
1898         if (!cic)
1899                 return ELV_MQUEUE_MAY;
1900
1901         cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
1902         if (cfqq) {
1903                 cfq_init_prio_data(cfqq);
1904                 cfq_prio_boost(cfqq);
1905
1906                 return __cfq_may_queue(cfqq);
1907         }
1908
1909         return ELV_MQUEUE_MAY;
1910 }
1911
1912 /*
1913  * queue lock held here
1914  */
1915 static void cfq_put_request(struct request *rq)
1916 {
1917         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1918
1919         if (cfqq) {
1920                 const int rw = rq_data_dir(rq);
1921
1922                 BUG_ON(!cfqq->allocated[rw]);
1923                 cfqq->allocated[rw]--;
1924
1925                 put_io_context(RQ_CIC(rq)->ioc);
1926
1927                 rq->elevator_private = NULL;
1928                 rq->elevator_private2 = NULL;
1929
1930                 cfq_put_queue(cfqq);
1931         }
1932 }
1933
1934 /*
1935  * Allocate cfq data structures associated with this request.
1936  */
1937 static int
1938 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
1939 {
1940         struct cfq_data *cfqd = q->elevator->elevator_data;
1941         struct task_struct *tsk = current;
1942         struct cfq_io_context *cic;
1943         const int rw = rq_data_dir(rq);
1944         const int is_sync = rq_is_sync(rq);
1945         struct cfq_queue *cfqq;
1946         unsigned long flags;
1947
1948         might_sleep_if(gfp_mask & __GFP_WAIT);
1949
1950         cic = cfq_get_io_context(cfqd, gfp_mask);
1951
1952         spin_lock_irqsave(q->queue_lock, flags);
1953
1954         if (!cic)
1955                 goto queue_fail;
1956
1957         cfqq = cic_to_cfqq(cic, is_sync);
1958         if (!cfqq) {
1959                 cfqq = cfq_get_queue(cfqd, is_sync, tsk, gfp_mask);
1960
1961                 if (!cfqq)
1962                         goto queue_fail;
1963
1964                 cic_set_cfqq(cic, cfqq, is_sync);
1965         }
1966
1967         cfqq->allocated[rw]++;
1968         cfq_clear_cfqq_must_alloc(cfqq);
1969         atomic_inc(&cfqq->ref);
1970
1971         spin_unlock_irqrestore(q->queue_lock, flags);
1972
1973         rq->elevator_private = cic;
1974         rq->elevator_private2 = cfqq;
1975         return 0;
1976
1977 queue_fail:
1978         if (cic)
1979                 put_io_context(cic->ioc);
1980
1981         cfq_schedule_dispatch(cfqd);
1982         spin_unlock_irqrestore(q->queue_lock, flags);
1983         return 1;
1984 }
1985
1986 static void cfq_kick_queue(struct work_struct *work)
1987 {
1988         struct cfq_data *cfqd =
1989                 container_of(work, struct cfq_data, unplug_work);
1990         struct request_queue *q = cfqd->queue;
1991         unsigned long flags;
1992
1993         spin_lock_irqsave(q->queue_lock, flags);
1994         blk_start_queueing(q);
1995         spin_unlock_irqrestore(q->queue_lock, flags);
1996 }
1997
1998 /*
1999  * Timer running if the active_queue is currently idling inside its time slice
2000  */
2001 static void cfq_idle_slice_timer(unsigned long data)
2002 {
2003         struct cfq_data *cfqd = (struct cfq_data *) data;
2004         struct cfq_queue *cfqq;
2005         unsigned long flags;
2006         int timed_out = 1;
2007
2008         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2009
2010         if ((cfqq = cfqd->active_queue) != NULL) {
2011                 timed_out = 0;
2012
2013                 /*
2014                  * expired
2015                  */
2016                 if (cfq_slice_used(cfqq))
2017                         goto expire;
2018
2019                 /*
2020                  * only expire and reinvoke request handler, if there are
2021                  * other queues with pending requests
2022                  */
2023                 if (!cfqd->busy_queues)
2024                         goto out_cont;
2025
2026                 /*
2027                  * not expired and it has a request pending, let it dispatch
2028                  */
2029                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
2030                         cfq_mark_cfqq_must_dispatch(cfqq);
2031                         goto out_kick;
2032                 }
2033         }
2034 expire:
2035         cfq_slice_expired(cfqd, timed_out);
2036 out_kick:
2037         cfq_schedule_dispatch(cfqd);
2038 out_cont:
2039         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2040 }
2041
2042 /*
2043  * Timer running if an idle class queue is waiting for service
2044  */
2045 static void cfq_idle_class_timer(unsigned long data)
2046 {
2047         struct cfq_data *cfqd = (struct cfq_data *) data;
2048         unsigned long flags;
2049
2050         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2051
2052         /*
2053          * race with a non-idle queue, reset timer
2054          */
2055         if (!start_idle_class_timer(cfqd))
2056                 cfq_schedule_dispatch(cfqd);
2057
2058         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2059 }
2060
2061 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2062 {
2063         del_timer_sync(&cfqd->idle_slice_timer);
2064         del_timer_sync(&cfqd->idle_class_timer);
2065         kblockd_flush_work(&cfqd->unplug_work);
2066 }
2067
2068 static void cfq_put_async_queues(struct cfq_data *cfqd)
2069 {
2070         int i;
2071
2072         for (i = 0; i < IOPRIO_BE_NR; i++) {
2073                 if (cfqd->async_cfqq[0][i])
2074                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2075                 if (cfqd->async_cfqq[1][i])
2076                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2077         }
2078
2079         if (cfqd->async_idle_cfqq)
2080                 cfq_put_queue(cfqd->async_idle_cfqq);
2081 }
2082
2083 static void cfq_exit_queue(elevator_t *e)
2084 {
2085         struct cfq_data *cfqd = e->elevator_data;
2086         struct request_queue *q = cfqd->queue;
2087
2088         cfq_shutdown_timer_wq(cfqd);
2089
2090         spin_lock_irq(q->queue_lock);
2091
2092         if (cfqd->active_queue)
2093                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2094
2095         while (!list_empty(&cfqd->cic_list)) {
2096                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2097                                                         struct cfq_io_context,
2098                                                         queue_list);
2099
2100                 __cfq_exit_single_io_context(cfqd, cic);
2101         }
2102
2103         cfq_put_async_queues(cfqd);
2104
2105         spin_unlock_irq(q->queue_lock);
2106
2107         cfq_shutdown_timer_wq(cfqd);
2108
2109         kfree(cfqd);
2110 }
2111
2112 static void *cfq_init_queue(struct request_queue *q)
2113 {
2114         struct cfq_data *cfqd;
2115
2116         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2117         if (!cfqd)
2118                 return NULL;
2119
2120         cfqd->service_tree = CFQ_RB_ROOT;
2121         INIT_LIST_HEAD(&cfqd->cic_list);
2122
2123         cfqd->queue = q;
2124
2125         init_timer(&cfqd->idle_slice_timer);
2126         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2127         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2128
2129         init_timer(&cfqd->idle_class_timer);
2130         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2131         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2132
2133         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2134
2135         cfqd->last_end_request = jiffies;
2136         cfqd->cfq_quantum = cfq_quantum;
2137         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2138         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2139         cfqd->cfq_back_max = cfq_back_max;
2140         cfqd->cfq_back_penalty = cfq_back_penalty;
2141         cfqd->cfq_slice[0] = cfq_slice_async;
2142         cfqd->cfq_slice[1] = cfq_slice_sync;
2143         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2144         cfqd->cfq_slice_idle = cfq_slice_idle;
2145
2146         return cfqd;
2147 }
2148
2149 static void cfq_slab_kill(void)
2150 {
2151         if (cfq_pool)
2152                 kmem_cache_destroy(cfq_pool);
2153         if (cfq_ioc_pool)
2154                 kmem_cache_destroy(cfq_ioc_pool);
2155 }
2156
2157 static int __init cfq_slab_setup(void)
2158 {
2159         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2160         if (!cfq_pool)
2161                 goto fail;
2162
2163         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2164         if (!cfq_ioc_pool)
2165                 goto fail;
2166
2167         return 0;
2168 fail:
2169         cfq_slab_kill();
2170         return -ENOMEM;
2171 }
2172
2173 /*
2174  * sysfs parts below -->
2175  */
2176 static ssize_t
2177 cfq_var_show(unsigned int var, char *page)
2178 {
2179         return sprintf(page, "%d\n", var);
2180 }
2181
2182 static ssize_t
2183 cfq_var_store(unsigned int *var, const char *page, size_t count)
2184 {
2185         char *p = (char *) page;
2186
2187         *var = simple_strtoul(p, &p, 10);
2188         return count;
2189 }
2190
2191 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2192 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2193 {                                                                       \
2194         struct cfq_data *cfqd = e->elevator_data;                       \
2195         unsigned int __data = __VAR;                                    \
2196         if (__CONV)                                                     \
2197                 __data = jiffies_to_msecs(__data);                      \
2198         return cfq_var_show(__data, (page));                            \
2199 }
2200 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2201 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2202 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2203 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2204 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2205 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2206 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2207 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2208 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2209 #undef SHOW_FUNCTION
2210
2211 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2212 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2213 {                                                                       \
2214         struct cfq_data *cfqd = e->elevator_data;                       \
2215         unsigned int __data;                                            \
2216         int ret = cfq_var_store(&__data, (page), count);                \
2217         if (__data < (MIN))                                             \
2218                 __data = (MIN);                                         \
2219         else if (__data > (MAX))                                        \
2220                 __data = (MAX);                                         \
2221         if (__CONV)                                                     \
2222                 *(__PTR) = msecs_to_jiffies(__data);                    \
2223         else                                                            \
2224                 *(__PTR) = __data;                                      \
2225         return ret;                                                     \
2226 }
2227 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2228 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2229 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2230 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2231 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2232 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2233 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2234 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2235 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2236 #undef STORE_FUNCTION
2237
2238 #define CFQ_ATTR(name) \
2239         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2240
2241 static struct elv_fs_entry cfq_attrs[] = {
2242         CFQ_ATTR(quantum),
2243         CFQ_ATTR(fifo_expire_sync),
2244         CFQ_ATTR(fifo_expire_async),
2245         CFQ_ATTR(back_seek_max),
2246         CFQ_ATTR(back_seek_penalty),
2247         CFQ_ATTR(slice_sync),
2248         CFQ_ATTR(slice_async),
2249         CFQ_ATTR(slice_async_rq),
2250         CFQ_ATTR(slice_idle),
2251         __ATTR_NULL
2252 };
2253
2254 static struct elevator_type iosched_cfq = {
2255         .ops = {
2256                 .elevator_merge_fn =            cfq_merge,
2257                 .elevator_merged_fn =           cfq_merged_request,
2258                 .elevator_merge_req_fn =        cfq_merged_requests,
2259                 .elevator_allow_merge_fn =      cfq_allow_merge,
2260                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2261                 .elevator_add_req_fn =          cfq_insert_request,
2262                 .elevator_activate_req_fn =     cfq_activate_request,
2263                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2264                 .elevator_queue_empty_fn =      cfq_queue_empty,
2265                 .elevator_completed_req_fn =    cfq_completed_request,
2266                 .elevator_former_req_fn =       elv_rb_former_request,
2267                 .elevator_latter_req_fn =       elv_rb_latter_request,
2268                 .elevator_set_req_fn =          cfq_set_request,
2269                 .elevator_put_req_fn =          cfq_put_request,
2270                 .elevator_may_queue_fn =        cfq_may_queue,
2271                 .elevator_init_fn =             cfq_init_queue,
2272                 .elevator_exit_fn =             cfq_exit_queue,
2273                 .trim =                         cfq_free_io_context,
2274         },
2275         .elevator_attrs =       cfq_attrs,
2276         .elevator_name =        "cfq",
2277         .elevator_owner =       THIS_MODULE,
2278 };
2279
2280 static int __init cfq_init(void)
2281 {
2282         int ret;
2283
2284         /*
2285          * could be 0 on HZ < 1000 setups
2286          */
2287         if (!cfq_slice_async)
2288                 cfq_slice_async = 1;
2289         if (!cfq_slice_idle)
2290                 cfq_slice_idle = 1;
2291
2292         if (cfq_slab_setup())
2293                 return -ENOMEM;
2294
2295         ret = elv_register(&iosched_cfq);
2296         if (ret)
2297                 cfq_slab_kill();
2298
2299         return ret;
2300 }
2301
2302 static void __exit cfq_exit(void)
2303 {
2304         DECLARE_COMPLETION_ONSTACK(all_gone);
2305         elv_unregister(&iosched_cfq);
2306         ioc_gone = &all_gone;
2307         /* ioc_gone's update must be visible before reading ioc_count */
2308         smp_wmb();
2309         if (elv_ioc_count_read(ioc_count))
2310                 wait_for_completion(ioc_gone);
2311         synchronize_rcu();
2312         cfq_slab_kill();
2313 }
2314
2315 module_init(cfq_init);
2316 module_exit(cfq_exit);
2317
2318 MODULE_AUTHOR("Jens Axboe");
2319 MODULE_LICENSE("GPL");
2320 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");