54dc05439009d09968572e6f12ce30d0fef151e8
[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 /*
793  * Get next queue for service. Unless we have a queue preemption,
794  * we'll simply select the first cfqq in the service tree.
795  */
796 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
797 {
798         struct cfq_queue *cfqq;
799         struct rb_node *n;
800
801         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
802                 return NULL;
803
804         n = cfq_rb_first(&cfqd->service_tree);
805         cfqq = rb_entry(n, struct cfq_queue, rb_node);
806
807         if (cfq_class_idle(cfqq)) {
808                 unsigned long end;
809
810                 /*
811                  * if we have idle queues and no rt or be queues had
812                  * pending requests, either allow immediate service if
813                  * the grace period has passed or arm the idle grace
814                  * timer
815                  */
816                 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
817                 if (time_before(jiffies, end)) {
818                         mod_timer(&cfqd->idle_class_timer, end);
819                         cfqq = NULL;
820                 }
821         }
822
823         return cfqq;
824 }
825
826 /*
827  * Get and set a new active queue for service.
828  */
829 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
830 {
831         struct cfq_queue *cfqq;
832
833         cfqq = cfq_get_next_queue(cfqd);
834         __cfq_set_active_queue(cfqd, cfqq);
835         return cfqq;
836 }
837
838 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
839                                           struct request *rq)
840 {
841         if (rq->sector >= cfqd->last_position)
842                 return rq->sector - cfqd->last_position;
843         else
844                 return cfqd->last_position - rq->sector;
845 }
846
847 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
848 {
849         struct cfq_io_context *cic = cfqd->active_cic;
850
851         if (!sample_valid(cic->seek_samples))
852                 return 0;
853
854         return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
855 }
856
857 static int cfq_close_cooperator(struct cfq_data *cfq_data,
858                                 struct cfq_queue *cfqq)
859 {
860         /*
861          * We should notice if some of the queues are cooperating, eg
862          * working closely on the same area of the disk. In that case,
863          * we can group them together and don't waste time idling.
864          */
865         return 0;
866 }
867
868 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
869
870 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
871 {
872         struct cfq_queue *cfqq = cfqd->active_queue;
873         struct cfq_io_context *cic;
874         unsigned long sl;
875
876         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
877         WARN_ON(cfq_cfqq_slice_new(cfqq));
878
879         /*
880          * idle is disabled, either manually or by past process history
881          */
882         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
883                 return;
884
885         /*
886          * task has exited, don't wait
887          */
888         cic = cfqd->active_cic;
889         if (!cic || !cic->ioc->task)
890                 return;
891
892         /*
893          * See if this prio level has a good candidate
894          */
895         if (cfq_close_cooperator(cfqd, cfqq) &&
896             (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
897                 return;
898
899         cfq_mark_cfqq_must_dispatch(cfqq);
900         cfq_mark_cfqq_wait_request(cfqq);
901
902         /*
903          * we don't want to idle for seeks, but we do want to allow
904          * fair distribution of slice time for a process doing back-to-back
905          * seeks. so allow a little bit of time for him to submit a new rq
906          */
907         sl = cfqd->cfq_slice_idle;
908         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
909                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
910
911         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
912 }
913
914 /*
915  * Move request from internal lists to the request queue dispatch list.
916  */
917 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
918 {
919         struct cfq_data *cfqd = q->elevator->elevator_data;
920         struct cfq_queue *cfqq = RQ_CFQQ(rq);
921
922         cfq_remove_request(rq);
923         cfqq->dispatched++;
924         elv_dispatch_sort(q, rq);
925
926         if (cfq_cfqq_sync(cfqq))
927                 cfqd->sync_flight++;
928 }
929
930 /*
931  * return expired entry, or NULL to just start from scratch in rbtree
932  */
933 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
934 {
935         struct cfq_data *cfqd = cfqq->cfqd;
936         struct request *rq;
937         int fifo;
938
939         if (cfq_cfqq_fifo_expire(cfqq))
940                 return NULL;
941
942         cfq_mark_cfqq_fifo_expire(cfqq);
943
944         if (list_empty(&cfqq->fifo))
945                 return NULL;
946
947         fifo = cfq_cfqq_sync(cfqq);
948         rq = rq_entry_fifo(cfqq->fifo.next);
949
950         if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
951                 return NULL;
952
953         return rq;
954 }
955
956 static inline int
957 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
958 {
959         const int base_rq = cfqd->cfq_slice_async_rq;
960
961         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
962
963         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
964 }
965
966 /*
967  * Select a queue for service. If we have a current active queue,
968  * check whether to continue servicing it, or retrieve and set a new one.
969  */
970 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
971 {
972         struct cfq_queue *cfqq;
973
974         cfqq = cfqd->active_queue;
975         if (!cfqq)
976                 goto new_queue;
977
978         /*
979          * The active queue has run out of time, expire it and select new.
980          */
981         if (cfq_slice_used(cfqq))
982                 goto expire;
983
984         /*
985          * The active queue has requests and isn't expired, allow it to
986          * dispatch.
987          */
988         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
989                 goto keep_queue;
990
991         /*
992          * No requests pending. If the active queue still has requests in
993          * flight or is idling for a new request, allow either of these
994          * conditions to happen (or time out) before selecting a new queue.
995          */
996         if (timer_pending(&cfqd->idle_slice_timer) ||
997             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
998                 cfqq = NULL;
999                 goto keep_queue;
1000         }
1001
1002 expire:
1003         cfq_slice_expired(cfqd, 0);
1004 new_queue:
1005         cfqq = cfq_set_active_queue(cfqd);
1006 keep_queue:
1007         return cfqq;
1008 }
1009
1010 /*
1011  * Dispatch some requests from cfqq, moving them to the request queue
1012  * dispatch list.
1013  */
1014 static int
1015 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1016                         int max_dispatch)
1017 {
1018         int dispatched = 0;
1019
1020         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1021
1022         do {
1023                 struct request *rq;
1024
1025                 /*
1026                  * follow expired path, else get first next available
1027                  */
1028                 if ((rq = cfq_check_fifo(cfqq)) == NULL)
1029                         rq = cfqq->next_rq;
1030
1031                 /*
1032                  * finally, insert request into driver dispatch list
1033                  */
1034                 cfq_dispatch_insert(cfqd->queue, rq);
1035
1036                 dispatched++;
1037
1038                 if (!cfqd->active_cic) {
1039                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1040                         cfqd->active_cic = RQ_CIC(rq);
1041                 }
1042
1043                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
1044                         break;
1045
1046         } while (dispatched < max_dispatch);
1047
1048         /*
1049          * expire an async queue immediately if it has used up its slice. idle
1050          * queue always expire after 1 dispatch round.
1051          */
1052         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1053             dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1054             cfq_class_idle(cfqq))) {
1055                 cfqq->slice_end = jiffies + 1;
1056                 cfq_slice_expired(cfqd, 0);
1057         }
1058
1059         return dispatched;
1060 }
1061
1062 static inline int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1063 {
1064         int dispatched = 0;
1065
1066         while (cfqq->next_rq) {
1067                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1068                 dispatched++;
1069         }
1070
1071         BUG_ON(!list_empty(&cfqq->fifo));
1072         return dispatched;
1073 }
1074
1075 /*
1076  * Drain our current requests. Used for barriers and when switching
1077  * io schedulers on-the-fly.
1078  */
1079 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1080 {
1081         int dispatched = 0;
1082         struct rb_node *n;
1083
1084         while ((n = cfq_rb_first(&cfqd->service_tree)) != NULL) {
1085                 struct cfq_queue *cfqq = rb_entry(n, struct cfq_queue, rb_node);
1086
1087                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1088         }
1089
1090         cfq_slice_expired(cfqd, 0);
1091
1092         BUG_ON(cfqd->busy_queues);
1093
1094         return dispatched;
1095 }
1096
1097 static int cfq_dispatch_requests(struct request_queue *q, int force)
1098 {
1099         struct cfq_data *cfqd = q->elevator->elevator_data;
1100         struct cfq_queue *cfqq;
1101         int dispatched;
1102
1103         if (!cfqd->busy_queues)
1104                 return 0;
1105
1106         if (unlikely(force))
1107                 return cfq_forced_dispatch(cfqd);
1108
1109         dispatched = 0;
1110         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1111                 int max_dispatch;
1112
1113                 max_dispatch = cfqd->cfq_quantum;
1114                 if (cfq_class_idle(cfqq))
1115                         max_dispatch = 1;
1116
1117                 if (cfqq->dispatched >= max_dispatch) {
1118                         if (cfqd->busy_queues > 1)
1119                                 break;
1120                         if (cfqq->dispatched >= 4 * max_dispatch)
1121                                 break;
1122                 }
1123
1124                 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1125                         break;
1126
1127                 cfq_clear_cfqq_must_dispatch(cfqq);
1128                 cfq_clear_cfqq_wait_request(cfqq);
1129                 del_timer(&cfqd->idle_slice_timer);
1130
1131                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1132         }
1133
1134         return dispatched;
1135 }
1136
1137 /*
1138  * task holds one reference to the queue, dropped when task exits. each rq
1139  * in-flight on this queue also holds a reference, dropped when rq is freed.
1140  *
1141  * queue lock must be held here.
1142  */
1143 static void cfq_put_queue(struct cfq_queue *cfqq)
1144 {
1145         struct cfq_data *cfqd = cfqq->cfqd;
1146
1147         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1148
1149         if (!atomic_dec_and_test(&cfqq->ref))
1150                 return;
1151
1152         BUG_ON(rb_first(&cfqq->sort_list));
1153         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1154         BUG_ON(cfq_cfqq_on_rr(cfqq));
1155
1156         if (unlikely(cfqd->active_queue == cfqq)) {
1157                 __cfq_slice_expired(cfqd, cfqq, 0);
1158                 cfq_schedule_dispatch(cfqd);
1159         }
1160
1161         kmem_cache_free(cfq_pool, cfqq);
1162 }
1163
1164 static void cfq_free_io_context(struct io_context *ioc)
1165 {
1166         struct cfq_io_context *__cic;
1167         struct rb_node *n;
1168         int freed = 0;
1169
1170         ioc->ioc_data = NULL;
1171
1172         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1173                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1174                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1175                 kmem_cache_free(cfq_ioc_pool, __cic);
1176                 freed++;
1177         }
1178
1179         elv_ioc_count_mod(ioc_count, -freed);
1180
1181         if (ioc_gone && !elv_ioc_count_read(ioc_count))
1182                 complete(ioc_gone);
1183 }
1184
1185 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1186 {
1187         if (unlikely(cfqq == cfqd->active_queue)) {
1188                 __cfq_slice_expired(cfqd, cfqq, 0);
1189                 cfq_schedule_dispatch(cfqd);
1190         }
1191
1192         cfq_put_queue(cfqq);
1193 }
1194
1195 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1196                                          struct cfq_io_context *cic)
1197 {
1198         list_del_init(&cic->queue_list);
1199         smp_wmb();
1200         cic->key = NULL;
1201
1202         if (cic->cfqq[ASYNC]) {
1203                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1204                 cic->cfqq[ASYNC] = NULL;
1205         }
1206
1207         if (cic->cfqq[SYNC]) {
1208                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1209                 cic->cfqq[SYNC] = NULL;
1210         }
1211 }
1212
1213 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1214 {
1215         struct cfq_data *cfqd = cic->key;
1216
1217         if (cfqd) {
1218                 struct request_queue *q = cfqd->queue;
1219
1220                 spin_lock_irq(q->queue_lock);
1221                 __cfq_exit_single_io_context(cfqd, cic);
1222                 spin_unlock_irq(q->queue_lock);
1223         }
1224 }
1225
1226 /*
1227  * The process that ioc belongs to has exited, we need to clean up
1228  * and put the internal structures we have that belongs to that process.
1229  */
1230 static void cfq_exit_io_context(struct io_context *ioc)
1231 {
1232         struct cfq_io_context *__cic;
1233         struct rb_node *n;
1234
1235         ioc->ioc_data = NULL;
1236
1237         /*
1238          * put the reference this task is holding to the various queues
1239          */
1240         n = rb_first(&ioc->cic_root);
1241         while (n != NULL) {
1242                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1243
1244                 cfq_exit_single_io_context(__cic);
1245                 n = rb_next(n);
1246         }
1247 }
1248
1249 static struct cfq_io_context *
1250 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1251 {
1252         struct cfq_io_context *cic;
1253
1254         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1255                                                         cfqd->queue->node);
1256         if (cic) {
1257                 cic->last_end_request = jiffies;
1258                 INIT_LIST_HEAD(&cic->queue_list);
1259                 cic->dtor = cfq_free_io_context;
1260                 cic->exit = cfq_exit_io_context;
1261                 elv_ioc_count_inc(ioc_count);
1262         }
1263
1264         return cic;
1265 }
1266
1267 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1268 {
1269         struct task_struct *tsk = current;
1270         int ioprio_class;
1271
1272         if (!cfq_cfqq_prio_changed(cfqq))
1273                 return;
1274
1275         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1276         switch (ioprio_class) {
1277                 default:
1278                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1279                 case IOPRIO_CLASS_NONE:
1280                         /*
1281                          * no prio set, place us in the middle of the BE classes
1282                          */
1283                         cfqq->ioprio = task_nice_ioprio(tsk);
1284                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1285                         break;
1286                 case IOPRIO_CLASS_RT:
1287                         cfqq->ioprio = task_ioprio(tsk);
1288                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1289                         break;
1290                 case IOPRIO_CLASS_BE:
1291                         cfqq->ioprio = task_ioprio(tsk);
1292                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1293                         break;
1294                 case IOPRIO_CLASS_IDLE:
1295                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1296                         cfqq->ioprio = 7;
1297                         cfq_clear_cfqq_idle_window(cfqq);
1298                         break;
1299         }
1300
1301         /*
1302          * keep track of original prio settings in case we have to temporarily
1303          * elevate the priority of this queue
1304          */
1305         cfqq->org_ioprio = cfqq->ioprio;
1306         cfqq->org_ioprio_class = cfqq->ioprio_class;
1307         cfq_clear_cfqq_prio_changed(cfqq);
1308 }
1309
1310 static inline void changed_ioprio(struct cfq_io_context *cic)
1311 {
1312         struct cfq_data *cfqd = cic->key;
1313         struct cfq_queue *cfqq;
1314         unsigned long flags;
1315
1316         if (unlikely(!cfqd))
1317                 return;
1318
1319         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1320
1321         cfqq = cic->cfqq[ASYNC];
1322         if (cfqq) {
1323                 struct cfq_queue *new_cfqq;
1324                 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc->task,
1325                                          GFP_ATOMIC);
1326                 if (new_cfqq) {
1327                         cic->cfqq[ASYNC] = new_cfqq;
1328                         cfq_put_queue(cfqq);
1329                 }
1330         }
1331
1332         cfqq = cic->cfqq[SYNC];
1333         if (cfqq)
1334                 cfq_mark_cfqq_prio_changed(cfqq);
1335
1336         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1337 }
1338
1339 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1340 {
1341         struct cfq_io_context *cic;
1342         struct rb_node *n;
1343
1344         ioc->ioprio_changed = 0;
1345
1346         n = rb_first(&ioc->cic_root);
1347         while (n != NULL) {
1348                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1349
1350                 changed_ioprio(cic);
1351                 n = rb_next(n);
1352         }
1353 }
1354
1355 static struct cfq_queue *
1356 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1357                      struct task_struct *tsk, gfp_t gfp_mask)
1358 {
1359         struct cfq_queue *cfqq, *new_cfqq = NULL;
1360         struct cfq_io_context *cic;
1361
1362 retry:
1363         cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
1364         /* cic always exists here */
1365         cfqq = cic_to_cfqq(cic, is_sync);
1366
1367         if (!cfqq) {
1368                 if (new_cfqq) {
1369                         cfqq = new_cfqq;
1370                         new_cfqq = NULL;
1371                 } else if (gfp_mask & __GFP_WAIT) {
1372                         /*
1373                          * Inform the allocator of the fact that we will
1374                          * just repeat this allocation if it fails, to allow
1375                          * the allocator to do whatever it needs to attempt to
1376                          * free memory.
1377                          */
1378                         spin_unlock_irq(cfqd->queue->queue_lock);
1379                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1380                                         gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1381                                         cfqd->queue->node);
1382                         spin_lock_irq(cfqd->queue->queue_lock);
1383                         goto retry;
1384                 } else {
1385                         cfqq = kmem_cache_alloc_node(cfq_pool,
1386                                         gfp_mask | __GFP_ZERO,
1387                                         cfqd->queue->node);
1388                         if (!cfqq)
1389                                 goto out;
1390                 }
1391
1392                 RB_CLEAR_NODE(&cfqq->rb_node);
1393                 INIT_LIST_HEAD(&cfqq->fifo);
1394
1395                 atomic_set(&cfqq->ref, 0);
1396                 cfqq->cfqd = cfqd;
1397
1398                 if (is_sync) {
1399                         cfq_mark_cfqq_idle_window(cfqq);
1400                         cfq_mark_cfqq_sync(cfqq);
1401                 }
1402
1403                 cfq_mark_cfqq_prio_changed(cfqq);
1404                 cfq_mark_cfqq_queue_new(cfqq);
1405
1406                 cfq_init_prio_data(cfqq);
1407         }
1408
1409         if (new_cfqq)
1410                 kmem_cache_free(cfq_pool, new_cfqq);
1411
1412 out:
1413         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1414         return cfqq;
1415 }
1416
1417 static struct cfq_queue **
1418 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1419 {
1420         switch(ioprio_class) {
1421         case IOPRIO_CLASS_RT:
1422                 return &cfqd->async_cfqq[0][ioprio];
1423         case IOPRIO_CLASS_BE:
1424                 return &cfqd->async_cfqq[1][ioprio];
1425         case IOPRIO_CLASS_IDLE:
1426                 return &cfqd->async_idle_cfqq;
1427         default:
1428                 BUG();
1429         }
1430 }
1431
1432 static struct cfq_queue *
1433 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct task_struct *tsk,
1434               gfp_t gfp_mask)
1435 {
1436         const int ioprio = task_ioprio(tsk);
1437         const int ioprio_class = task_ioprio_class(tsk);
1438         struct cfq_queue **async_cfqq = NULL;
1439         struct cfq_queue *cfqq = NULL;
1440
1441         if (!is_sync) {
1442                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1443                 cfqq = *async_cfqq;
1444         }
1445
1446         if (!cfqq)
1447                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, tsk, gfp_mask);
1448
1449         /*
1450          * pin the queue now that it's allocated, scheduler exit will prune it
1451          */
1452         if (!is_sync && !(*async_cfqq)) {
1453                 atomic_inc(&cfqq->ref);
1454                 *async_cfqq = cfqq;
1455         }
1456
1457         atomic_inc(&cfqq->ref);
1458         return cfqq;
1459 }
1460
1461 /*
1462  * We drop cfq io contexts lazily, so we may find a dead one.
1463  */
1464 static void
1465 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1466 {
1467         WARN_ON(!list_empty(&cic->queue_list));
1468
1469         if (ioc->ioc_data == cic)
1470                 ioc->ioc_data = NULL;
1471
1472         rb_erase(&cic->rb_node, &ioc->cic_root);
1473         kmem_cache_free(cfq_ioc_pool, cic);
1474         elv_ioc_count_dec(ioc_count);
1475 }
1476
1477 static struct cfq_io_context *
1478 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1479 {
1480         struct rb_node *n;
1481         struct cfq_io_context *cic;
1482         void *k, *key = cfqd;
1483
1484         if (unlikely(!ioc))
1485                 return NULL;
1486
1487         /*
1488          * we maintain a last-hit cache, to avoid browsing over the tree
1489          */
1490         cic = ioc->ioc_data;
1491         if (cic && cic->key == cfqd)
1492                 return cic;
1493
1494 restart:
1495         n = ioc->cic_root.rb_node;
1496         while (n) {
1497                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1498                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1499                 k = cic->key;
1500                 if (unlikely(!k)) {
1501                         cfq_drop_dead_cic(ioc, cic);
1502                         goto restart;
1503                 }
1504
1505                 if (key < k)
1506                         n = n->rb_left;
1507                 else if (key > k)
1508                         n = n->rb_right;
1509                 else {
1510                         ioc->ioc_data = cic;
1511                         return cic;
1512                 }
1513         }
1514
1515         return NULL;
1516 }
1517
1518 static inline void
1519 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1520              struct cfq_io_context *cic)
1521 {
1522         struct rb_node **p;
1523         struct rb_node *parent;
1524         struct cfq_io_context *__cic;
1525         unsigned long flags;
1526         void *k;
1527
1528         cic->ioc = ioc;
1529         cic->key = cfqd;
1530
1531 restart:
1532         parent = NULL;
1533         p = &ioc->cic_root.rb_node;
1534         while (*p) {
1535                 parent = *p;
1536                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1537                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1538                 k = __cic->key;
1539                 if (unlikely(!k)) {
1540                         cfq_drop_dead_cic(ioc, __cic);
1541                         goto restart;
1542                 }
1543
1544                 if (cic->key < k)
1545                         p = &(*p)->rb_left;
1546                 else if (cic->key > k)
1547                         p = &(*p)->rb_right;
1548                 else
1549                         BUG();
1550         }
1551
1552         rb_link_node(&cic->rb_node, parent, p);
1553         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1554
1555         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1556         list_add(&cic->queue_list, &cfqd->cic_list);
1557         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1558 }
1559
1560 /*
1561  * Setup general io context and cfq io context. There can be several cfq
1562  * io contexts per general io context, if this process is doing io to more
1563  * than one device managed by cfq.
1564  */
1565 static struct cfq_io_context *
1566 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1567 {
1568         struct io_context *ioc = NULL;
1569         struct cfq_io_context *cic;
1570
1571         might_sleep_if(gfp_mask & __GFP_WAIT);
1572
1573         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1574         if (!ioc)
1575                 return NULL;
1576
1577         cic = cfq_cic_rb_lookup(cfqd, ioc);
1578         if (cic)
1579                 goto out;
1580
1581         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1582         if (cic == NULL)
1583                 goto err;
1584
1585         cfq_cic_link(cfqd, ioc, cic);
1586 out:
1587         smp_read_barrier_depends();
1588         if (unlikely(ioc->ioprio_changed))
1589                 cfq_ioc_set_ioprio(ioc);
1590
1591         return cic;
1592 err:
1593         put_io_context(ioc);
1594         return NULL;
1595 }
1596
1597 static void
1598 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1599 {
1600         unsigned long elapsed = jiffies - cic->last_end_request;
1601         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1602
1603         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1604         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1605         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1606 }
1607
1608 static void
1609 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1610                        struct request *rq)
1611 {
1612         sector_t sdist;
1613         u64 total;
1614
1615         if (cic->last_request_pos < rq->sector)
1616                 sdist = rq->sector - cic->last_request_pos;
1617         else
1618                 sdist = cic->last_request_pos - rq->sector;
1619
1620         /*
1621          * Don't allow the seek distance to get too large from the
1622          * odd fragment, pagein, etc
1623          */
1624         if (cic->seek_samples <= 60) /* second&third seek */
1625                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1626         else
1627                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1628
1629         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1630         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1631         total = cic->seek_total + (cic->seek_samples/2);
1632         do_div(total, cic->seek_samples);
1633         cic->seek_mean = (sector_t)total;
1634 }
1635
1636 /*
1637  * Disable idle window if the process thinks too long or seeks so much that
1638  * it doesn't matter
1639  */
1640 static void
1641 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1642                        struct cfq_io_context *cic)
1643 {
1644         int enable_idle;
1645
1646         if (!cfq_cfqq_sync(cfqq))
1647                 return;
1648
1649         enable_idle = cfq_cfqq_idle_window(cfqq);
1650
1651         if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1652             (cfqd->hw_tag && CIC_SEEKY(cic)))
1653                 enable_idle = 0;
1654         else if (sample_valid(cic->ttime_samples)) {
1655                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1656                         enable_idle = 0;
1657                 else
1658                         enable_idle = 1;
1659         }
1660
1661         if (enable_idle)
1662                 cfq_mark_cfqq_idle_window(cfqq);
1663         else
1664                 cfq_clear_cfqq_idle_window(cfqq);
1665 }
1666
1667 /*
1668  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1669  * no or if we aren't sure, a 1 will cause a preempt.
1670  */
1671 static int
1672 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1673                    struct request *rq)
1674 {
1675         struct cfq_queue *cfqq;
1676
1677         cfqq = cfqd->active_queue;
1678         if (!cfqq)
1679                 return 0;
1680
1681         if (cfq_slice_used(cfqq))
1682                 return 1;
1683
1684         if (cfq_class_idle(new_cfqq))
1685                 return 0;
1686
1687         if (cfq_class_idle(cfqq))
1688                 return 1;
1689
1690         /*
1691          * if the new request is sync, but the currently running queue is
1692          * not, let the sync request have priority.
1693          */
1694         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1695                 return 1;
1696
1697         /*
1698          * So both queues are sync. Let the new request get disk time if
1699          * it's a metadata request and the current queue is doing regular IO.
1700          */
1701         if (rq_is_meta(rq) && !cfqq->meta_pending)
1702                 return 1;
1703
1704         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1705                 return 0;
1706
1707         /*
1708          * if this request is as-good as one we would expect from the
1709          * current cfqq, let it preempt
1710          */
1711         if (cfq_rq_close(cfqd, rq))
1712                 return 1;
1713
1714         return 0;
1715 }
1716
1717 /*
1718  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1719  * let it have half of its nominal slice.
1720  */
1721 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1722 {
1723         cfq_slice_expired(cfqd, 1);
1724
1725         /*
1726          * Put the new queue at the front of the of the current list,
1727          * so we know that it will be selected next.
1728          */
1729         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1730
1731         cfq_service_tree_add(cfqd, cfqq, 1);
1732
1733         cfqq->slice_end = 0;
1734         cfq_mark_cfqq_slice_new(cfqq);
1735 }
1736
1737 /*
1738  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1739  * something we should do about it
1740  */
1741 static void
1742 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1743                 struct request *rq)
1744 {
1745         struct cfq_io_context *cic = RQ_CIC(rq);
1746
1747         if (rq_is_meta(rq))
1748                 cfqq->meta_pending++;
1749
1750         cfq_update_io_thinktime(cfqd, cic);
1751         cfq_update_io_seektime(cfqd, cic, rq);
1752         cfq_update_idle_window(cfqd, cfqq, cic);
1753
1754         cic->last_request_pos = rq->sector + rq->nr_sectors;
1755
1756         if (cfqq == cfqd->active_queue) {
1757                 /*
1758                  * if we are waiting for a request for this queue, let it rip
1759                  * immediately and flag that we must not expire this queue
1760                  * just now
1761                  */
1762                 if (cfq_cfqq_wait_request(cfqq)) {
1763                         cfq_mark_cfqq_must_dispatch(cfqq);
1764                         del_timer(&cfqd->idle_slice_timer);
1765                         blk_start_queueing(cfqd->queue);
1766                 }
1767         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1768                 /*
1769                  * not the active queue - expire current slice if it is
1770                  * idle and has expired it's mean thinktime or this new queue
1771                  * has some old slice time left and is of higher priority
1772                  */
1773                 cfq_preempt_queue(cfqd, cfqq);
1774                 cfq_mark_cfqq_must_dispatch(cfqq);
1775                 blk_start_queueing(cfqd->queue);
1776         }
1777 }
1778
1779 static void cfq_insert_request(struct request_queue *q, struct request *rq)
1780 {
1781         struct cfq_data *cfqd = q->elevator->elevator_data;
1782         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1783
1784         cfq_init_prio_data(cfqq);
1785
1786         cfq_add_rq_rb(rq);
1787
1788         list_add_tail(&rq->queuelist, &cfqq->fifo);
1789
1790         cfq_rq_enqueued(cfqd, cfqq, rq);
1791 }
1792
1793 static void cfq_completed_request(struct request_queue *q, struct request *rq)
1794 {
1795         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1796         struct cfq_data *cfqd = cfqq->cfqd;
1797         const int sync = rq_is_sync(rq);
1798         unsigned long now;
1799
1800         now = jiffies;
1801
1802         WARN_ON(!cfqd->rq_in_driver);
1803         WARN_ON(!cfqq->dispatched);
1804         cfqd->rq_in_driver--;
1805         cfqq->dispatched--;
1806
1807         if (cfq_cfqq_sync(cfqq))
1808                 cfqd->sync_flight--;
1809
1810         if (!cfq_class_idle(cfqq))
1811                 cfqd->last_end_request = now;
1812
1813         if (sync)
1814                 RQ_CIC(rq)->last_end_request = now;
1815
1816         /*
1817          * If this is the active queue, check if it needs to be expired,
1818          * or if we want to idle in case it has no pending requests.
1819          */
1820         if (cfqd->active_queue == cfqq) {
1821                 if (cfq_cfqq_slice_new(cfqq)) {
1822                         cfq_set_prio_slice(cfqd, cfqq);
1823                         cfq_clear_cfqq_slice_new(cfqq);
1824                 }
1825                 if (cfq_slice_used(cfqq))
1826                         cfq_slice_expired(cfqd, 1);
1827                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1828                         cfq_arm_slice_timer(cfqd);
1829         }
1830
1831         if (!cfqd->rq_in_driver)
1832                 cfq_schedule_dispatch(cfqd);
1833 }
1834
1835 /*
1836  * we temporarily boost lower priority queues if they are holding fs exclusive
1837  * resources. they are boosted to normal prio (CLASS_BE/4)
1838  */
1839 static void cfq_prio_boost(struct cfq_queue *cfqq)
1840 {
1841         if (has_fs_excl()) {
1842                 /*
1843                  * boost idle prio on transactions that would lock out other
1844                  * users of the filesystem
1845                  */
1846                 if (cfq_class_idle(cfqq))
1847                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1848                 if (cfqq->ioprio > IOPRIO_NORM)
1849                         cfqq->ioprio = IOPRIO_NORM;
1850         } else {
1851                 /*
1852                  * check if we need to unboost the queue
1853                  */
1854                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1855                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1856                 if (cfqq->ioprio != cfqq->org_ioprio)
1857                         cfqq->ioprio = cfqq->org_ioprio;
1858         }
1859 }
1860
1861 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1862 {
1863         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1864             !cfq_cfqq_must_alloc_slice(cfqq)) {
1865                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1866                 return ELV_MQUEUE_MUST;
1867         }
1868
1869         return ELV_MQUEUE_MAY;
1870 }
1871
1872 static int cfq_may_queue(struct request_queue *q, int rw)
1873 {
1874         struct cfq_data *cfqd = q->elevator->elevator_data;
1875         struct task_struct *tsk = current;
1876         struct cfq_io_context *cic;
1877         struct cfq_queue *cfqq;
1878
1879         /*
1880          * don't force setup of a queue from here, as a call to may_queue
1881          * does not necessarily imply that a request actually will be queued.
1882          * so just lookup a possibly existing queue, or return 'may queue'
1883          * if that fails
1884          */
1885         cic = cfq_cic_rb_lookup(cfqd, tsk->io_context);
1886         if (!cic)
1887                 return ELV_MQUEUE_MAY;
1888
1889         cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
1890         if (cfqq) {
1891                 cfq_init_prio_data(cfqq);
1892                 cfq_prio_boost(cfqq);
1893
1894                 return __cfq_may_queue(cfqq);
1895         }
1896
1897         return ELV_MQUEUE_MAY;
1898 }
1899
1900 /*
1901  * queue lock held here
1902  */
1903 static void cfq_put_request(struct request *rq)
1904 {
1905         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1906
1907         if (cfqq) {
1908                 const int rw = rq_data_dir(rq);
1909
1910                 BUG_ON(!cfqq->allocated[rw]);
1911                 cfqq->allocated[rw]--;
1912
1913                 put_io_context(RQ_CIC(rq)->ioc);
1914
1915                 rq->elevator_private = NULL;
1916                 rq->elevator_private2 = NULL;
1917
1918                 cfq_put_queue(cfqq);
1919         }
1920 }
1921
1922 /*
1923  * Allocate cfq data structures associated with this request.
1924  */
1925 static int
1926 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
1927 {
1928         struct cfq_data *cfqd = q->elevator->elevator_data;
1929         struct task_struct *tsk = current;
1930         struct cfq_io_context *cic;
1931         const int rw = rq_data_dir(rq);
1932         const int is_sync = rq_is_sync(rq);
1933         struct cfq_queue *cfqq;
1934         unsigned long flags;
1935
1936         might_sleep_if(gfp_mask & __GFP_WAIT);
1937
1938         cic = cfq_get_io_context(cfqd, gfp_mask);
1939
1940         spin_lock_irqsave(q->queue_lock, flags);
1941
1942         if (!cic)
1943                 goto queue_fail;
1944
1945         cfqq = cic_to_cfqq(cic, is_sync);
1946         if (!cfqq) {
1947                 cfqq = cfq_get_queue(cfqd, is_sync, tsk, gfp_mask);
1948
1949                 if (!cfqq)
1950                         goto queue_fail;
1951
1952                 cic_set_cfqq(cic, cfqq, is_sync);
1953         }
1954
1955         cfqq->allocated[rw]++;
1956         cfq_clear_cfqq_must_alloc(cfqq);
1957         atomic_inc(&cfqq->ref);
1958
1959         spin_unlock_irqrestore(q->queue_lock, flags);
1960
1961         rq->elevator_private = cic;
1962         rq->elevator_private2 = cfqq;
1963         return 0;
1964
1965 queue_fail:
1966         if (cic)
1967                 put_io_context(cic->ioc);
1968
1969         cfq_schedule_dispatch(cfqd);
1970         spin_unlock_irqrestore(q->queue_lock, flags);
1971         return 1;
1972 }
1973
1974 static void cfq_kick_queue(struct work_struct *work)
1975 {
1976         struct cfq_data *cfqd =
1977                 container_of(work, struct cfq_data, unplug_work);
1978         struct request_queue *q = cfqd->queue;
1979         unsigned long flags;
1980
1981         spin_lock_irqsave(q->queue_lock, flags);
1982         blk_start_queueing(q);
1983         spin_unlock_irqrestore(q->queue_lock, flags);
1984 }
1985
1986 /*
1987  * Timer running if the active_queue is currently idling inside its time slice
1988  */
1989 static void cfq_idle_slice_timer(unsigned long data)
1990 {
1991         struct cfq_data *cfqd = (struct cfq_data *) data;
1992         struct cfq_queue *cfqq;
1993         unsigned long flags;
1994         int timed_out = 1;
1995
1996         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1997
1998         if ((cfqq = cfqd->active_queue) != NULL) {
1999                 timed_out = 0;
2000
2001                 /*
2002                  * expired
2003                  */
2004                 if (cfq_slice_used(cfqq))
2005                         goto expire;
2006
2007                 /*
2008                  * only expire and reinvoke request handler, if there are
2009                  * other queues with pending requests
2010                  */
2011                 if (!cfqd->busy_queues)
2012                         goto out_cont;
2013
2014                 /*
2015                  * not expired and it has a request pending, let it dispatch
2016                  */
2017                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
2018                         cfq_mark_cfqq_must_dispatch(cfqq);
2019                         goto out_kick;
2020                 }
2021         }
2022 expire:
2023         cfq_slice_expired(cfqd, timed_out);
2024 out_kick:
2025         cfq_schedule_dispatch(cfqd);
2026 out_cont:
2027         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2028 }
2029
2030 /*
2031  * Timer running if an idle class queue is waiting for service
2032  */
2033 static void cfq_idle_class_timer(unsigned long data)
2034 {
2035         struct cfq_data *cfqd = (struct cfq_data *) data;
2036         unsigned long flags, end;
2037
2038         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2039
2040         /*
2041          * race with a non-idle queue, reset timer
2042          */
2043         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2044         if (!time_after_eq(jiffies, end))
2045                 mod_timer(&cfqd->idle_class_timer, end);
2046         else
2047                 cfq_schedule_dispatch(cfqd);
2048
2049         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2050 }
2051
2052 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2053 {
2054         del_timer_sync(&cfqd->idle_slice_timer);
2055         del_timer_sync(&cfqd->idle_class_timer);
2056         blk_sync_queue(cfqd->queue);
2057 }
2058
2059 static void cfq_put_async_queues(struct cfq_data *cfqd)
2060 {
2061         int i;
2062
2063         for (i = 0; i < IOPRIO_BE_NR; i++) {
2064                 if (cfqd->async_cfqq[0][i])
2065                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2066                 if (cfqd->async_cfqq[1][i])
2067                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2068                 if (cfqd->async_idle_cfqq)
2069                         cfq_put_queue(cfqd->async_idle_cfqq);
2070         }
2071 }
2072
2073 static void cfq_exit_queue(elevator_t *e)
2074 {
2075         struct cfq_data *cfqd = e->elevator_data;
2076         struct request_queue *q = cfqd->queue;
2077
2078         cfq_shutdown_timer_wq(cfqd);
2079
2080         spin_lock_irq(q->queue_lock);
2081
2082         if (cfqd->active_queue)
2083                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2084
2085         while (!list_empty(&cfqd->cic_list)) {
2086                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2087                                                         struct cfq_io_context,
2088                                                         queue_list);
2089
2090                 __cfq_exit_single_io_context(cfqd, cic);
2091         }
2092
2093         cfq_put_async_queues(cfqd);
2094
2095         spin_unlock_irq(q->queue_lock);
2096
2097         cfq_shutdown_timer_wq(cfqd);
2098
2099         kfree(cfqd);
2100 }
2101
2102 static void *cfq_init_queue(struct request_queue *q)
2103 {
2104         struct cfq_data *cfqd;
2105
2106         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2107         if (!cfqd)
2108                 return NULL;
2109
2110         cfqd->service_tree = CFQ_RB_ROOT;
2111         INIT_LIST_HEAD(&cfqd->cic_list);
2112
2113         cfqd->queue = q;
2114
2115         init_timer(&cfqd->idle_slice_timer);
2116         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2117         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2118
2119         init_timer(&cfqd->idle_class_timer);
2120         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2121         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2122
2123         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2124
2125         cfqd->cfq_quantum = cfq_quantum;
2126         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2127         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2128         cfqd->cfq_back_max = cfq_back_max;
2129         cfqd->cfq_back_penalty = cfq_back_penalty;
2130         cfqd->cfq_slice[0] = cfq_slice_async;
2131         cfqd->cfq_slice[1] = cfq_slice_sync;
2132         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2133         cfqd->cfq_slice_idle = cfq_slice_idle;
2134
2135         return cfqd;
2136 }
2137
2138 static void cfq_slab_kill(void)
2139 {
2140         if (cfq_pool)
2141                 kmem_cache_destroy(cfq_pool);
2142         if (cfq_ioc_pool)
2143                 kmem_cache_destroy(cfq_ioc_pool);
2144 }
2145
2146 static int __init cfq_slab_setup(void)
2147 {
2148         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2149         if (!cfq_pool)
2150                 goto fail;
2151
2152         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2153         if (!cfq_ioc_pool)
2154                 goto fail;
2155
2156         return 0;
2157 fail:
2158         cfq_slab_kill();
2159         return -ENOMEM;
2160 }
2161
2162 /*
2163  * sysfs parts below -->
2164  */
2165 static ssize_t
2166 cfq_var_show(unsigned int var, char *page)
2167 {
2168         return sprintf(page, "%d\n", var);
2169 }
2170
2171 static ssize_t
2172 cfq_var_store(unsigned int *var, const char *page, size_t count)
2173 {
2174         char *p = (char *) page;
2175
2176         *var = simple_strtoul(p, &p, 10);
2177         return count;
2178 }
2179
2180 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2181 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2182 {                                                                       \
2183         struct cfq_data *cfqd = e->elevator_data;                       \
2184         unsigned int __data = __VAR;                                    \
2185         if (__CONV)                                                     \
2186                 __data = jiffies_to_msecs(__data);                      \
2187         return cfq_var_show(__data, (page));                            \
2188 }
2189 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2190 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2191 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2192 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2193 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2194 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2195 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2196 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2197 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2198 #undef SHOW_FUNCTION
2199
2200 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2201 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2202 {                                                                       \
2203         struct cfq_data *cfqd = e->elevator_data;                       \
2204         unsigned int __data;                                            \
2205         int ret = cfq_var_store(&__data, (page), count);                \
2206         if (__data < (MIN))                                             \
2207                 __data = (MIN);                                         \
2208         else if (__data > (MAX))                                        \
2209                 __data = (MAX);                                         \
2210         if (__CONV)                                                     \
2211                 *(__PTR) = msecs_to_jiffies(__data);                    \
2212         else                                                            \
2213                 *(__PTR) = __data;                                      \
2214         return ret;                                                     \
2215 }
2216 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2217 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2218 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2219 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2220 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2221 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2222 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2223 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2224 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2225 #undef STORE_FUNCTION
2226
2227 #define CFQ_ATTR(name) \
2228         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2229
2230 static struct elv_fs_entry cfq_attrs[] = {
2231         CFQ_ATTR(quantum),
2232         CFQ_ATTR(fifo_expire_sync),
2233         CFQ_ATTR(fifo_expire_async),
2234         CFQ_ATTR(back_seek_max),
2235         CFQ_ATTR(back_seek_penalty),
2236         CFQ_ATTR(slice_sync),
2237         CFQ_ATTR(slice_async),
2238         CFQ_ATTR(slice_async_rq),
2239         CFQ_ATTR(slice_idle),
2240         __ATTR_NULL
2241 };
2242
2243 static struct elevator_type iosched_cfq = {
2244         .ops = {
2245                 .elevator_merge_fn =            cfq_merge,
2246                 .elevator_merged_fn =           cfq_merged_request,
2247                 .elevator_merge_req_fn =        cfq_merged_requests,
2248                 .elevator_allow_merge_fn =      cfq_allow_merge,
2249                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2250                 .elevator_add_req_fn =          cfq_insert_request,
2251                 .elevator_activate_req_fn =     cfq_activate_request,
2252                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2253                 .elevator_queue_empty_fn =      cfq_queue_empty,
2254                 .elevator_completed_req_fn =    cfq_completed_request,
2255                 .elevator_former_req_fn =       elv_rb_former_request,
2256                 .elevator_latter_req_fn =       elv_rb_latter_request,
2257                 .elevator_set_req_fn =          cfq_set_request,
2258                 .elevator_put_req_fn =          cfq_put_request,
2259                 .elevator_may_queue_fn =        cfq_may_queue,
2260                 .elevator_init_fn =             cfq_init_queue,
2261                 .elevator_exit_fn =             cfq_exit_queue,
2262                 .trim =                         cfq_free_io_context,
2263         },
2264         .elevator_attrs =       cfq_attrs,
2265         .elevator_name =        "cfq",
2266         .elevator_owner =       THIS_MODULE,
2267 };
2268
2269 static int __init cfq_init(void)
2270 {
2271         int ret;
2272
2273         /*
2274          * could be 0 on HZ < 1000 setups
2275          */
2276         if (!cfq_slice_async)
2277                 cfq_slice_async = 1;
2278         if (!cfq_slice_idle)
2279                 cfq_slice_idle = 1;
2280
2281         if (cfq_slab_setup())
2282                 return -ENOMEM;
2283
2284         ret = elv_register(&iosched_cfq);
2285         if (ret)
2286                 cfq_slab_kill();
2287
2288         return ret;
2289 }
2290
2291 static void __exit cfq_exit(void)
2292 {
2293         DECLARE_COMPLETION_ONSTACK(all_gone);
2294         elv_unregister(&iosched_cfq);
2295         ioc_gone = &all_gone;
2296         /* ioc_gone's update must be visible before reading ioc_count */
2297         smp_wmb();
2298         if (elv_ioc_count_read(ioc_count))
2299                 wait_for_completion(ioc_gone);
2300         synchronize_rcu();
2301         cfq_slab_kill();
2302 }
2303
2304 module_init(cfq_init);
2305 module_exit(cfq_exit);
2306
2307 MODULE_AUTHOR("Jens Axboe");
2308 MODULE_LICENSE("GPL");
2309 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");