Pull sn2-mmio-writes into release branch
[linux-2.6] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@suse.de>
8  */
9 #include <linux/kernel.h>
10 #include <linux/fs.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/bio.h>
14 #include <linux/config.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/init.h>
18 #include <linux/compiler.h>
19 #include <linux/hash.h>
20 #include <linux/rbtree.h>
21 #include <linux/mempool.h>
22 #include <linux/ioprio.h>
23 #include <linux/writeback.h>
24
25 /*
26  * tunables
27  */
28 static const int cfq_quantum = 4;               /* max queue in one round of service */
29 static const int cfq_queued = 8;                /* minimum rq allocate limit per-queue*/
30 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
31 static const int cfq_back_max = 16 * 1024;      /* maximum backwards seek, in KiB */
32 static const int cfq_back_penalty = 2;          /* penalty of a backwards seek */
33
34 static const int cfq_slice_sync = HZ / 10;
35 static int cfq_slice_async = HZ / 25;
36 static const int cfq_slice_async_rq = 2;
37 static int cfq_slice_idle = HZ / 100;
38
39 #define CFQ_IDLE_GRACE          (HZ / 10)
40 #define CFQ_SLICE_SCALE         (5)
41
42 #define CFQ_KEY_ASYNC           (0)
43 #define CFQ_KEY_ANY             (0xffff)
44
45 /*
46  * disable queueing at the driver/hardware level
47  */
48 static const int cfq_max_depth = 2;
49
50 /*
51  * for the hash of cfqq inside the cfqd
52  */
53 #define CFQ_QHASH_SHIFT         6
54 #define CFQ_QHASH_ENTRIES       (1 << CFQ_QHASH_SHIFT)
55 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
56
57 /*
58  * for the hash of crq inside the cfqq
59  */
60 #define CFQ_MHASH_SHIFT         6
61 #define CFQ_MHASH_BLOCK(sec)    ((sec) >> 3)
62 #define CFQ_MHASH_ENTRIES       (1 << CFQ_MHASH_SHIFT)
63 #define CFQ_MHASH_FN(sec)       hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
64 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
65 #define list_entry_hash(ptr)    hlist_entry((ptr), struct cfq_rq, hash)
66
67 #define list_entry_cfqq(ptr)    list_entry((ptr), struct cfq_queue, cfq_list)
68 #define list_entry_fifo(ptr)    list_entry((ptr), struct request, queuelist)
69
70 #define RQ_DATA(rq)             (rq)->elevator_private
71
72 /*
73  * rb-tree defines
74  */
75 #define RB_NONE                 (2)
76 #define RB_EMPTY(node)          ((node)->rb_node == NULL)
77 #define RB_CLEAR_COLOR(node)    (node)->rb_color = RB_NONE
78 #define RB_CLEAR(node)          do {    \
79         (node)->rb_parent = NULL;       \
80         RB_CLEAR_COLOR((node));         \
81         (node)->rb_right = NULL;        \
82         (node)->rb_left = NULL;         \
83 } while (0)
84 #define RB_CLEAR_ROOT(root)     ((root)->rb_node = NULL)
85 #define rb_entry_crq(node)      rb_entry((node), struct cfq_rq, rb_node)
86 #define rq_rb_key(rq)           (rq)->sector
87
88 static kmem_cache_t *crq_pool;
89 static kmem_cache_t *cfq_pool;
90 static kmem_cache_t *cfq_ioc_pool;
91
92 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
93 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
94 #define cfq_class_be(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
95 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
96
97 #define ASYNC                   (0)
98 #define SYNC                    (1)
99
100 #define cfq_cfqq_dispatched(cfqq)       \
101         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
102
103 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
104
105 #define cfq_cfqq_sync(cfqq)             \
106         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
107
108 /*
109  * Per block device queue structure
110  */
111 struct cfq_data {
112         atomic_t ref;
113         request_queue_t *queue;
114
115         /*
116          * rr list of queues with requests and the count of them
117          */
118         struct list_head rr_list[CFQ_PRIO_LISTS];
119         struct list_head busy_rr;
120         struct list_head cur_rr;
121         struct list_head idle_rr;
122         unsigned int busy_queues;
123
124         /*
125          * non-ordered list of empty cfqq's
126          */
127         struct list_head empty_list;
128
129         /*
130          * cfqq lookup hash
131          */
132         struct hlist_head *cfq_hash;
133
134         /*
135          * global crq hash for all queues
136          */
137         struct hlist_head *crq_hash;
138
139         unsigned int max_queued;
140
141         mempool_t *crq_pool;
142
143         int rq_in_driver;
144
145         /*
146          * schedule slice state info
147          */
148         /*
149          * idle window management
150          */
151         struct timer_list idle_slice_timer;
152         struct work_struct unplug_work;
153
154         struct cfq_queue *active_queue;
155         struct cfq_io_context *active_cic;
156         int cur_prio, cur_end_prio;
157         unsigned int dispatch_slice;
158
159         struct timer_list idle_class_timer;
160
161         sector_t last_sector;
162         unsigned long last_end_request;
163
164         unsigned int rq_starved;
165
166         /*
167          * tunables, see top of file
168          */
169         unsigned int cfq_quantum;
170         unsigned int cfq_queued;
171         unsigned int cfq_fifo_expire[2];
172         unsigned int cfq_back_penalty;
173         unsigned int cfq_back_max;
174         unsigned int cfq_slice[2];
175         unsigned int cfq_slice_async_rq;
176         unsigned int cfq_slice_idle;
177         unsigned int cfq_max_depth;
178 };
179
180 /*
181  * Per process-grouping structure
182  */
183 struct cfq_queue {
184         /* reference count */
185         atomic_t ref;
186         /* parent cfq_data */
187         struct cfq_data *cfqd;
188         /* cfqq lookup hash */
189         struct hlist_node cfq_hash;
190         /* hash key */
191         unsigned int key;
192         /* on either rr or empty list of cfqd */
193         struct list_head cfq_list;
194         /* sorted list of pending requests */
195         struct rb_root sort_list;
196         /* if fifo isn't expired, next request to serve */
197         struct cfq_rq *next_crq;
198         /* requests queued in sort_list */
199         int queued[2];
200         /* currently allocated requests */
201         int allocated[2];
202         /* fifo list of requests in sort_list */
203         struct list_head fifo;
204
205         unsigned long slice_start;
206         unsigned long slice_end;
207         unsigned long slice_left;
208         unsigned long service_last;
209
210         /* number of requests that are on the dispatch list */
211         int on_dispatch[2];
212
213         /* io prio of this group */
214         unsigned short ioprio, org_ioprio;
215         unsigned short ioprio_class, org_ioprio_class;
216
217         /* various state flags, see below */
218         unsigned int flags;
219 };
220
221 struct cfq_rq {
222         struct rb_node rb_node;
223         sector_t rb_key;
224         struct request *request;
225         struct hlist_node hash;
226
227         struct cfq_queue *cfq_queue;
228         struct cfq_io_context *io_context;
229
230         unsigned int crq_flags;
231 };
232
233 enum cfqq_state_flags {
234         CFQ_CFQQ_FLAG_on_rr = 0,
235         CFQ_CFQQ_FLAG_wait_request,
236         CFQ_CFQQ_FLAG_must_alloc,
237         CFQ_CFQQ_FLAG_must_alloc_slice,
238         CFQ_CFQQ_FLAG_must_dispatch,
239         CFQ_CFQQ_FLAG_fifo_expire,
240         CFQ_CFQQ_FLAG_idle_window,
241         CFQ_CFQQ_FLAG_prio_changed,
242 };
243
244 #define CFQ_CFQQ_FNS(name)                                              \
245 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
246 {                                                                       \
247         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
248 }                                                                       \
249 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
250 {                                                                       \
251         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
252 }                                                                       \
253 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
254 {                                                                       \
255         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
256 }
257
258 CFQ_CFQQ_FNS(on_rr);
259 CFQ_CFQQ_FNS(wait_request);
260 CFQ_CFQQ_FNS(must_alloc);
261 CFQ_CFQQ_FNS(must_alloc_slice);
262 CFQ_CFQQ_FNS(must_dispatch);
263 CFQ_CFQQ_FNS(fifo_expire);
264 CFQ_CFQQ_FNS(idle_window);
265 CFQ_CFQQ_FNS(prio_changed);
266 #undef CFQ_CFQQ_FNS
267
268 enum cfq_rq_state_flags {
269         CFQ_CRQ_FLAG_is_sync = 0,
270 };
271
272 #define CFQ_CRQ_FNS(name)                                               \
273 static inline void cfq_mark_crq_##name(struct cfq_rq *crq)              \
274 {                                                                       \
275         crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name);                   \
276 }                                                                       \
277 static inline void cfq_clear_crq_##name(struct cfq_rq *crq)             \
278 {                                                                       \
279         crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name);                  \
280 }                                                                       \
281 static inline int cfq_crq_##name(const struct cfq_rq *crq)              \
282 {                                                                       \
283         return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0;      \
284 }
285
286 CFQ_CRQ_FNS(is_sync);
287 #undef CFQ_CRQ_FNS
288
289 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
290 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
291 static void cfq_put_cfqd(struct cfq_data *cfqd);
292
293 #define process_sync(tsk)       ((tsk)->flags & PF_SYNCWRITE)
294
295 /*
296  * lots of deadline iosched dupes, can be abstracted later...
297  */
298 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
299 {
300         hlist_del_init(&crq->hash);
301 }
302
303 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
304 {
305         const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
306
307         hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
308 }
309
310 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
311 {
312         struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
313         struct hlist_node *entry, *next;
314
315         hlist_for_each_safe(entry, next, hash_list) {
316                 struct cfq_rq *crq = list_entry_hash(entry);
317                 struct request *__rq = crq->request;
318
319                 if (!rq_mergeable(__rq)) {
320                         cfq_del_crq_hash(crq);
321                         continue;
322                 }
323
324                 if (rq_hash_key(__rq) == offset)
325                         return __rq;
326         }
327
328         return NULL;
329 }
330
331 /*
332  * scheduler run of queue, if there are requests pending and no one in the
333  * driver that will restart queueing
334  */
335 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
336 {
337         if (cfqd->busy_queues)
338                 kblockd_schedule_work(&cfqd->unplug_work);
339 }
340
341 static int cfq_queue_empty(request_queue_t *q)
342 {
343         struct cfq_data *cfqd = q->elevator->elevator_data;
344
345         return !cfqd->busy_queues;
346 }
347
348 /*
349  * Lifted from AS - choose which of crq1 and crq2 that is best served now.
350  * We choose the request that is closest to the head right now. Distance
351  * behind the head are penalized and only allowed to a certain extent.
352  */
353 static struct cfq_rq *
354 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
355 {
356         sector_t last, s1, s2, d1 = 0, d2 = 0;
357         int r1_wrap = 0, r2_wrap = 0;   /* requests are behind the disk head */
358         unsigned long back_max;
359
360         if (crq1 == NULL || crq1 == crq2)
361                 return crq2;
362         if (crq2 == NULL)
363                 return crq1;
364
365         if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
366                 return crq1;
367         else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
368                 return crq2;
369
370         s1 = crq1->request->sector;
371         s2 = crq2->request->sector;
372
373         last = cfqd->last_sector;
374
375         /*
376          * by definition, 1KiB is 2 sectors
377          */
378         back_max = cfqd->cfq_back_max * 2;
379
380         /*
381          * Strict one way elevator _except_ in the case where we allow
382          * short backward seeks which are biased as twice the cost of a
383          * similar forward seek.
384          */
385         if (s1 >= last)
386                 d1 = s1 - last;
387         else if (s1 + back_max >= last)
388                 d1 = (last - s1) * cfqd->cfq_back_penalty;
389         else
390                 r1_wrap = 1;
391
392         if (s2 >= last)
393                 d2 = s2 - last;
394         else if (s2 + back_max >= last)
395                 d2 = (last - s2) * cfqd->cfq_back_penalty;
396         else
397                 r2_wrap = 1;
398
399         /* Found required data */
400         if (!r1_wrap && r2_wrap)
401                 return crq1;
402         else if (!r2_wrap && r1_wrap)
403                 return crq2;
404         else if (r1_wrap && r2_wrap) {
405                 /* both behind the head */
406                 if (s1 <= s2)
407                         return crq1;
408                 else
409                         return crq2;
410         }
411
412         /* Both requests in front of the head */
413         if (d1 < d2)
414                 return crq1;
415         else if (d2 < d1)
416                 return crq2;
417         else {
418                 if (s1 >= s2)
419                         return crq1;
420                 else
421                         return crq2;
422         }
423 }
424
425 /*
426  * would be nice to take fifo expire time into account as well
427  */
428 static struct cfq_rq *
429 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
430                   struct cfq_rq *last)
431 {
432         struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
433         struct rb_node *rbnext, *rbprev;
434
435         if (!(rbnext = rb_next(&last->rb_node))) {
436                 rbnext = rb_first(&cfqq->sort_list);
437                 if (rbnext == &last->rb_node)
438                         rbnext = NULL;
439         }
440
441         rbprev = rb_prev(&last->rb_node);
442
443         if (rbprev)
444                 crq_prev = rb_entry_crq(rbprev);
445         if (rbnext)
446                 crq_next = rb_entry_crq(rbnext);
447
448         return cfq_choose_req(cfqd, crq_next, crq_prev);
449 }
450
451 static void cfq_update_next_crq(struct cfq_rq *crq)
452 {
453         struct cfq_queue *cfqq = crq->cfq_queue;
454
455         if (cfqq->next_crq == crq)
456                 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
457 }
458
459 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
460 {
461         struct cfq_data *cfqd = cfqq->cfqd;
462         struct list_head *list, *entry;
463
464         BUG_ON(!cfq_cfqq_on_rr(cfqq));
465
466         list_del(&cfqq->cfq_list);
467
468         if (cfq_class_rt(cfqq))
469                 list = &cfqd->cur_rr;
470         else if (cfq_class_idle(cfqq))
471                 list = &cfqd->idle_rr;
472         else {
473                 /*
474                  * if cfqq has requests in flight, don't allow it to be
475                  * found in cfq_set_active_queue before it has finished them.
476                  * this is done to increase fairness between a process that
477                  * has lots of io pending vs one that only generates one
478                  * sporadically or synchronously
479                  */
480                 if (cfq_cfqq_dispatched(cfqq))
481                         list = &cfqd->busy_rr;
482                 else
483                         list = &cfqd->rr_list[cfqq->ioprio];
484         }
485
486         /*
487          * if queue was preempted, just add to front to be fair. busy_rr
488          * isn't sorted.
489          */
490         if (preempted || list == &cfqd->busy_rr) {
491                 list_add(&cfqq->cfq_list, list);
492                 return;
493         }
494
495         /*
496          * sort by when queue was last serviced
497          */
498         entry = list;
499         while ((entry = entry->prev) != list) {
500                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
501
502                 if (!__cfqq->service_last)
503                         break;
504                 if (time_before(__cfqq->service_last, cfqq->service_last))
505                         break;
506         }
507
508         list_add(&cfqq->cfq_list, entry);
509 }
510
511 /*
512  * add to busy list of queues for service, trying to be fair in ordering
513  * the pending list according to last request service
514  */
515 static inline void
516 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
517 {
518         BUG_ON(cfq_cfqq_on_rr(cfqq));
519         cfq_mark_cfqq_on_rr(cfqq);
520         cfqd->busy_queues++;
521
522         cfq_resort_rr_list(cfqq, 0);
523 }
524
525 static inline void
526 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
527 {
528         BUG_ON(!cfq_cfqq_on_rr(cfqq));
529         cfq_clear_cfqq_on_rr(cfqq);
530         list_move(&cfqq->cfq_list, &cfqd->empty_list);
531
532         BUG_ON(!cfqd->busy_queues);
533         cfqd->busy_queues--;
534 }
535
536 /*
537  * rb tree support functions
538  */
539 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
540 {
541         struct cfq_queue *cfqq = crq->cfq_queue;
542         struct cfq_data *cfqd = cfqq->cfqd;
543         const int sync = cfq_crq_is_sync(crq);
544
545         BUG_ON(!cfqq->queued[sync]);
546         cfqq->queued[sync]--;
547
548         cfq_update_next_crq(crq);
549
550         rb_erase(&crq->rb_node, &cfqq->sort_list);
551         RB_CLEAR_COLOR(&crq->rb_node);
552
553         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
554                 cfq_del_cfqq_rr(cfqd, cfqq);
555 }
556
557 static struct cfq_rq *
558 __cfq_add_crq_rb(struct cfq_rq *crq)
559 {
560         struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
561         struct rb_node *parent = NULL;
562         struct cfq_rq *__crq;
563
564         while (*p) {
565                 parent = *p;
566                 __crq = rb_entry_crq(parent);
567
568                 if (crq->rb_key < __crq->rb_key)
569                         p = &(*p)->rb_left;
570                 else if (crq->rb_key > __crq->rb_key)
571                         p = &(*p)->rb_right;
572                 else
573                         return __crq;
574         }
575
576         rb_link_node(&crq->rb_node, parent, p);
577         return NULL;
578 }
579
580 static void cfq_add_crq_rb(struct cfq_rq *crq)
581 {
582         struct cfq_queue *cfqq = crq->cfq_queue;
583         struct cfq_data *cfqd = cfqq->cfqd;
584         struct request *rq = crq->request;
585         struct cfq_rq *__alias;
586
587         crq->rb_key = rq_rb_key(rq);
588         cfqq->queued[cfq_crq_is_sync(crq)]++;
589
590         /*
591          * looks a little odd, but the first insert might return an alias.
592          * if that happens, put the alias on the dispatch list
593          */
594         while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
595                 cfq_dispatch_insert(cfqd->queue, __alias);
596
597         rb_insert_color(&crq->rb_node, &cfqq->sort_list);
598
599         if (!cfq_cfqq_on_rr(cfqq))
600                 cfq_add_cfqq_rr(cfqd, cfqq);
601
602         /*
603          * check if this request is a better next-serve candidate
604          */
605         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
606 }
607
608 static inline void
609 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
610 {
611         rb_erase(&crq->rb_node, &cfqq->sort_list);
612         cfqq->queued[cfq_crq_is_sync(crq)]--;
613
614         cfq_add_crq_rb(crq);
615 }
616
617 static struct request *cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector)
618
619 {
620         struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, current->pid, CFQ_KEY_ANY);
621         struct rb_node *n;
622
623         if (!cfqq)
624                 goto out;
625
626         n = cfqq->sort_list.rb_node;
627         while (n) {
628                 struct cfq_rq *crq = rb_entry_crq(n);
629
630                 if (sector < crq->rb_key)
631                         n = n->rb_left;
632                 else if (sector > crq->rb_key)
633                         n = n->rb_right;
634                 else
635                         return crq->request;
636         }
637
638 out:
639         return NULL;
640 }
641
642 static void cfq_activate_request(request_queue_t *q, struct request *rq)
643 {
644         struct cfq_data *cfqd = q->elevator->elevator_data;
645
646         cfqd->rq_in_driver++;
647 }
648
649 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
650 {
651         struct cfq_data *cfqd = q->elevator->elevator_data;
652
653         WARN_ON(!cfqd->rq_in_driver);
654         cfqd->rq_in_driver--;
655 }
656
657 static void cfq_remove_request(struct request *rq)
658 {
659         struct cfq_rq *crq = RQ_DATA(rq);
660
661         list_del_init(&rq->queuelist);
662         cfq_del_crq_rb(crq);
663         cfq_del_crq_hash(crq);
664 }
665
666 static int
667 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
668 {
669         struct cfq_data *cfqd = q->elevator->elevator_data;
670         struct request *__rq;
671         int ret;
672
673         __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
674         if (__rq && elv_rq_merge_ok(__rq, bio)) {
675                 ret = ELEVATOR_BACK_MERGE;
676                 goto out;
677         }
678
679         __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio));
680         if (__rq && elv_rq_merge_ok(__rq, bio)) {
681                 ret = ELEVATOR_FRONT_MERGE;
682                 goto out;
683         }
684
685         return ELEVATOR_NO_MERGE;
686 out:
687         *req = __rq;
688         return ret;
689 }
690
691 static void cfq_merged_request(request_queue_t *q, struct request *req)
692 {
693         struct cfq_data *cfqd = q->elevator->elevator_data;
694         struct cfq_rq *crq = RQ_DATA(req);
695
696         cfq_del_crq_hash(crq);
697         cfq_add_crq_hash(cfqd, crq);
698
699         if (rq_rb_key(req) != crq->rb_key) {
700                 struct cfq_queue *cfqq = crq->cfq_queue;
701
702                 cfq_update_next_crq(crq);
703                 cfq_reposition_crq_rb(cfqq, crq);
704         }
705 }
706
707 static void
708 cfq_merged_requests(request_queue_t *q, struct request *rq,
709                     struct request *next)
710 {
711         cfq_merged_request(q, rq);
712
713         /*
714          * reposition in fifo if next is older than rq
715          */
716         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
717             time_before(next->start_time, rq->start_time))
718                 list_move(&rq->queuelist, &next->queuelist);
719
720         cfq_remove_request(next);
721 }
722
723 static inline void
724 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
725 {
726         if (cfqq) {
727                 /*
728                  * stop potential idle class queues waiting service
729                  */
730                 del_timer(&cfqd->idle_class_timer);
731
732                 cfqq->slice_start = jiffies;
733                 cfqq->slice_end = 0;
734                 cfqq->slice_left = 0;
735                 cfq_clear_cfqq_must_alloc_slice(cfqq);
736                 cfq_clear_cfqq_fifo_expire(cfqq);
737         }
738
739         cfqd->active_queue = cfqq;
740 }
741
742 /*
743  * current cfqq expired its slice (or was too idle), select new one
744  */
745 static void
746 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
747                     int preempted)
748 {
749         unsigned long now = jiffies;
750
751         if (cfq_cfqq_wait_request(cfqq))
752                 del_timer(&cfqd->idle_slice_timer);
753
754         if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
755                 cfqq->service_last = now;
756                 cfq_schedule_dispatch(cfqd);
757         }
758
759         cfq_clear_cfqq_must_dispatch(cfqq);
760         cfq_clear_cfqq_wait_request(cfqq);
761
762         /*
763          * store what was left of this slice, if the queue idled out
764          * or was preempted
765          */
766         if (time_after(cfqq->slice_end, now))
767                 cfqq->slice_left = cfqq->slice_end - now;
768         else
769                 cfqq->slice_left = 0;
770
771         if (cfq_cfqq_on_rr(cfqq))
772                 cfq_resort_rr_list(cfqq, preempted);
773
774         if (cfqq == cfqd->active_queue)
775                 cfqd->active_queue = NULL;
776
777         if (cfqd->active_cic) {
778                 put_io_context(cfqd->active_cic->ioc);
779                 cfqd->active_cic = NULL;
780         }
781
782         cfqd->dispatch_slice = 0;
783 }
784
785 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
786 {
787         struct cfq_queue *cfqq = cfqd->active_queue;
788
789         if (cfqq)
790                 __cfq_slice_expired(cfqd, cfqq, preempted);
791 }
792
793 /*
794  * 0
795  * 0,1
796  * 0,1,2
797  * 0,1,2,3
798  * 0,1,2,3,4
799  * 0,1,2,3,4,5
800  * 0,1,2,3,4,5,6
801  * 0,1,2,3,4,5,6,7
802  */
803 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
804 {
805         int prio, wrap;
806
807         prio = -1;
808         wrap = 0;
809         do {
810                 int p;
811
812                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
813                         if (!list_empty(&cfqd->rr_list[p])) {
814                                 prio = p;
815                                 break;
816                         }
817                 }
818
819                 if (prio != -1)
820                         break;
821                 cfqd->cur_prio = 0;
822                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
823                         cfqd->cur_end_prio = 0;
824                         if (wrap)
825                                 break;
826                         wrap = 1;
827                 }
828         } while (1);
829
830         if (unlikely(prio == -1))
831                 return -1;
832
833         BUG_ON(prio >= CFQ_PRIO_LISTS);
834
835         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
836
837         cfqd->cur_prio = prio + 1;
838         if (cfqd->cur_prio > cfqd->cur_end_prio) {
839                 cfqd->cur_end_prio = cfqd->cur_prio;
840                 cfqd->cur_prio = 0;
841         }
842         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
843                 cfqd->cur_prio = 0;
844                 cfqd->cur_end_prio = 0;
845         }
846
847         return prio;
848 }
849
850 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
851 {
852         struct cfq_queue *cfqq = NULL;
853
854         /*
855          * if current list is non-empty, grab first entry. if it is empty,
856          * get next prio level and grab first entry then if any are spliced
857          */
858         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
859                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
860
861         /*
862          * if we have idle queues and no rt or be queues had pending
863          * requests, either allow immediate service if the grace period
864          * has passed or arm the idle grace timer
865          */
866         if (!cfqq && !list_empty(&cfqd->idle_rr)) {
867                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
868
869                 if (time_after_eq(jiffies, end))
870                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
871                 else
872                         mod_timer(&cfqd->idle_class_timer, end);
873         }
874
875         __cfq_set_active_queue(cfqd, cfqq);
876         return cfqq;
877 }
878
879 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
880
881 {
882         unsigned long sl;
883
884         WARN_ON(!RB_EMPTY(&cfqq->sort_list));
885         WARN_ON(cfqq != cfqd->active_queue);
886
887         /*
888          * idle is disabled, either manually or by past process history
889          */
890         if (!cfqd->cfq_slice_idle)
891                 return 0;
892         if (!cfq_cfqq_idle_window(cfqq))
893                 return 0;
894         /*
895          * task has exited, don't wait
896          */
897         if (cfqd->active_cic && !cfqd->active_cic->ioc->task)
898                 return 0;
899
900         cfq_mark_cfqq_must_dispatch(cfqq);
901         cfq_mark_cfqq_wait_request(cfqq);
902
903         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
904         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
905         return 1;
906 }
907
908 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
909 {
910         struct cfq_data *cfqd = q->elevator->elevator_data;
911         struct cfq_queue *cfqq = crq->cfq_queue;
912
913         cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
914         cfq_remove_request(crq->request);
915         cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
916         elv_dispatch_sort(q, crq->request);
917 }
918
919 /*
920  * return expired entry, or NULL to just start from scratch in rbtree
921  */
922 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
923 {
924         struct cfq_data *cfqd = cfqq->cfqd;
925         struct request *rq;
926         struct cfq_rq *crq;
927
928         if (cfq_cfqq_fifo_expire(cfqq))
929                 return NULL;
930
931         if (!list_empty(&cfqq->fifo)) {
932                 int fifo = cfq_cfqq_class_sync(cfqq);
933
934                 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
935                 rq = crq->request;
936                 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
937                         cfq_mark_cfqq_fifo_expire(cfqq);
938                         return crq;
939                 }
940         }
941
942         return NULL;
943 }
944
945 /*
946  * Scale schedule slice based on io priority. Use the sync time slice only
947  * if a queue is marked sync and has sync io queued. A sync queue with async
948  * io only, should not get full sync slice length.
949  */
950 static inline int
951 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
952 {
953         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
954
955         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
956
957         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
958 }
959
960 static inline void
961 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
962 {
963         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
964 }
965
966 static inline int
967 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
968 {
969         const int base_rq = cfqd->cfq_slice_async_rq;
970
971         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
972
973         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
974 }
975
976 /*
977  * get next queue for service
978  */
979 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
980 {
981         unsigned long now = jiffies;
982         struct cfq_queue *cfqq;
983
984         cfqq = cfqd->active_queue;
985         if (!cfqq)
986                 goto new_queue;
987
988         /*
989          * slice has expired
990          */
991         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
992                 goto expire;
993
994         /*
995          * if queue has requests, dispatch one. if not, check if
996          * enough slice is left to wait for one
997          */
998         if (!RB_EMPTY(&cfqq->sort_list))
999                 goto keep_queue;
1000         else if (cfq_cfqq_class_sync(cfqq) &&
1001                  time_before(now, cfqq->slice_end)) {
1002                 if (cfq_arm_slice_timer(cfqd, cfqq))
1003                         return NULL;
1004         }
1005
1006 expire:
1007         cfq_slice_expired(cfqd, 0);
1008 new_queue:
1009         cfqq = cfq_set_active_queue(cfqd);
1010 keep_queue:
1011         return cfqq;
1012 }
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(&cfqq->sort_list));
1021
1022         do {
1023                 struct cfq_rq *crq;
1024
1025                 /*
1026                  * follow expired path, else get first next available
1027                  */
1028                 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1029                         crq = cfqq->next_crq;
1030
1031                 /*
1032                  * finally, insert request into driver dispatch list
1033                  */
1034                 cfq_dispatch_insert(cfqd->queue, crq);
1035
1036                 cfqd->dispatch_slice++;
1037                 dispatched++;
1038
1039                 if (!cfqd->active_cic) {
1040                         atomic_inc(&crq->io_context->ioc->refcount);
1041                         cfqd->active_cic = crq->io_context;
1042                 }
1043
1044                 if (RB_EMPTY(&cfqq->sort_list))
1045                         break;
1046
1047         } while (dispatched < max_dispatch);
1048
1049         /*
1050          * if slice end isn't set yet, set it. if at least one request was
1051          * sync, use the sync time slice value
1052          */
1053         if (!cfqq->slice_end)
1054                 cfq_set_prio_slice(cfqd, cfqq);
1055
1056         /*
1057          * expire an async queue immediately if it has used up its slice. idle
1058          * queue always expire after 1 dispatch round.
1059          */
1060         if ((!cfq_cfqq_sync(cfqq) &&
1061             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1062             cfq_class_idle(cfqq))
1063                 cfq_slice_expired(cfqd, 0);
1064
1065         return dispatched;
1066 }
1067
1068 static int
1069 cfq_forced_dispatch_cfqqs(struct list_head *list)
1070 {
1071         int dispatched = 0;
1072         struct cfq_queue *cfqq, *next;
1073         struct cfq_rq *crq;
1074
1075         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1076                 while ((crq = cfqq->next_crq)) {
1077                         cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1078                         dispatched++;
1079                 }
1080                 BUG_ON(!list_empty(&cfqq->fifo));
1081         }
1082         return dispatched;
1083 }
1084
1085 static int
1086 cfq_forced_dispatch(struct cfq_data *cfqd)
1087 {
1088         int i, dispatched = 0;
1089
1090         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1091                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1092
1093         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1094         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1095         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1096
1097         cfq_slice_expired(cfqd, 0);
1098
1099         BUG_ON(cfqd->busy_queues);
1100
1101         return dispatched;
1102 }
1103
1104 static int
1105 cfq_dispatch_requests(request_queue_t *q, int force)
1106 {
1107         struct cfq_data *cfqd = q->elevator->elevator_data;
1108         struct cfq_queue *cfqq;
1109
1110         if (!cfqd->busy_queues)
1111                 return 0;
1112
1113         if (unlikely(force))
1114                 return cfq_forced_dispatch(cfqd);
1115
1116         cfqq = cfq_select_queue(cfqd);
1117         if (cfqq) {
1118                 int max_dispatch;
1119
1120                 /*
1121                  * if idle window is disabled, allow queue buildup
1122                  */
1123                 if (!cfq_cfqq_idle_window(cfqq) &&
1124                     cfqd->rq_in_driver >= cfqd->cfq_max_depth)
1125                         return 0;
1126
1127                 cfq_clear_cfqq_must_dispatch(cfqq);
1128                 cfq_clear_cfqq_wait_request(cfqq);
1129                 del_timer(&cfqd->idle_slice_timer);
1130
1131                 max_dispatch = cfqd->cfq_quantum;
1132                 if (cfq_class_idle(cfqq))
1133                         max_dispatch = 1;
1134
1135                 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1136         }
1137
1138         return 0;
1139 }
1140
1141 /*
1142  * task holds one reference to the queue, dropped when task exits. each crq
1143  * in-flight on this queue also holds a reference, dropped when crq is freed.
1144  *
1145  * queue lock must be held here.
1146  */
1147 static void cfq_put_queue(struct cfq_queue *cfqq)
1148 {
1149         struct cfq_data *cfqd = cfqq->cfqd;
1150
1151         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1152
1153         if (!atomic_dec_and_test(&cfqq->ref))
1154                 return;
1155
1156         BUG_ON(rb_first(&cfqq->sort_list));
1157         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1158         BUG_ON(cfq_cfqq_on_rr(cfqq));
1159
1160         if (unlikely(cfqd->active_queue == cfqq))
1161                 __cfq_slice_expired(cfqd, cfqq, 0);
1162
1163         cfq_put_cfqd(cfqq->cfqd);
1164
1165         /*
1166          * it's on the empty list and still hashed
1167          */
1168         list_del(&cfqq->cfq_list);
1169         hlist_del(&cfqq->cfq_hash);
1170         kmem_cache_free(cfq_pool, cfqq);
1171 }
1172
1173 static inline struct cfq_queue *
1174 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1175                     const int hashval)
1176 {
1177         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1178         struct hlist_node *entry, *next;
1179
1180         hlist_for_each_safe(entry, next, hash_list) {
1181                 struct cfq_queue *__cfqq = list_entry_qhash(entry);
1182                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->ioprio_class, __cfqq->ioprio);
1183
1184                 if (__cfqq->key == key && (__p == prio || prio == CFQ_KEY_ANY))
1185                         return __cfqq;
1186         }
1187
1188         return NULL;
1189 }
1190
1191 static struct cfq_queue *
1192 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1193 {
1194         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1195 }
1196
1197 static void cfq_free_io_context(struct cfq_io_context *cic)
1198 {
1199         struct cfq_io_context *__cic;
1200         struct list_head *entry, *next;
1201
1202         list_for_each_safe(entry, next, &cic->list) {
1203                 __cic = list_entry(entry, struct cfq_io_context, list);
1204                 kmem_cache_free(cfq_ioc_pool, __cic);
1205         }
1206
1207         kmem_cache_free(cfq_ioc_pool, cic);
1208 }
1209
1210 /*
1211  * Called with interrupts disabled
1212  */
1213 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1214 {
1215         struct cfq_data *cfqd = cic->cfqq->cfqd;
1216         request_queue_t *q = cfqd->queue;
1217
1218         WARN_ON(!irqs_disabled());
1219
1220         spin_lock(q->queue_lock);
1221
1222         if (unlikely(cic->cfqq == cfqd->active_queue))
1223                 __cfq_slice_expired(cfqd, cic->cfqq, 0);
1224
1225         cfq_put_queue(cic->cfqq);
1226         cic->cfqq = NULL;
1227         spin_unlock(q->queue_lock);
1228 }
1229
1230 /*
1231  * Another task may update the task cic list, if it is doing a queue lookup
1232  * on its behalf. cfq_cic_lock excludes such concurrent updates
1233  */
1234 static void cfq_exit_io_context(struct cfq_io_context *cic)
1235 {
1236         struct cfq_io_context *__cic;
1237         struct list_head *entry;
1238         unsigned long flags;
1239
1240         local_irq_save(flags);
1241
1242         /*
1243          * put the reference this task is holding to the various queues
1244          */
1245         list_for_each(entry, &cic->list) {
1246                 __cic = list_entry(entry, struct cfq_io_context, list);
1247                 cfq_exit_single_io_context(__cic);
1248         }
1249
1250         cfq_exit_single_io_context(cic);
1251         local_irq_restore(flags);
1252 }
1253
1254 static struct cfq_io_context *
1255 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1256 {
1257         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1258
1259         if (cic) {
1260                 INIT_LIST_HEAD(&cic->list);
1261                 cic->cfqq = NULL;
1262                 cic->key = NULL;
1263                 cic->last_end_request = jiffies;
1264                 cic->ttime_total = 0;
1265                 cic->ttime_samples = 0;
1266                 cic->ttime_mean = 0;
1267                 cic->dtor = cfq_free_io_context;
1268                 cic->exit = cfq_exit_io_context;
1269         }
1270
1271         return cic;
1272 }
1273
1274 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1275 {
1276         struct task_struct *tsk = current;
1277         int ioprio_class;
1278
1279         if (!cfq_cfqq_prio_changed(cfqq))
1280                 return;
1281
1282         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1283         switch (ioprio_class) {
1284                 default:
1285                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1286                 case IOPRIO_CLASS_NONE:
1287                         /*
1288                          * no prio set, place us in the middle of the BE classes
1289                          */
1290                         cfqq->ioprio = task_nice_ioprio(tsk);
1291                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1292                         break;
1293                 case IOPRIO_CLASS_RT:
1294                         cfqq->ioprio = task_ioprio(tsk);
1295                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1296                         break;
1297                 case IOPRIO_CLASS_BE:
1298                         cfqq->ioprio = task_ioprio(tsk);
1299                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1300                         break;
1301                 case IOPRIO_CLASS_IDLE:
1302                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1303                         cfqq->ioprio = 7;
1304                         cfq_clear_cfqq_idle_window(cfqq);
1305                         break;
1306         }
1307
1308         /*
1309          * keep track of original prio settings in case we have to temporarily
1310          * elevate the priority of this queue
1311          */
1312         cfqq->org_ioprio = cfqq->ioprio;
1313         cfqq->org_ioprio_class = cfqq->ioprio_class;
1314
1315         if (cfq_cfqq_on_rr(cfqq))
1316                 cfq_resort_rr_list(cfqq, 0);
1317
1318         cfq_clear_cfqq_prio_changed(cfqq);
1319 }
1320
1321 static inline void changed_ioprio(struct cfq_queue *cfqq)
1322 {
1323         if (cfqq) {
1324                 struct cfq_data *cfqd = cfqq->cfqd;
1325
1326                 spin_lock(cfqd->queue->queue_lock);
1327                 cfq_mark_cfqq_prio_changed(cfqq);
1328                 cfq_init_prio_data(cfqq);
1329                 spin_unlock(cfqd->queue->queue_lock);
1330         }
1331 }
1332
1333 /*
1334  * callback from sys_ioprio_set, irqs are disabled
1335  */
1336 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1337 {
1338         struct cfq_io_context *cic = ioc->cic;
1339
1340         changed_ioprio(cic->cfqq);
1341
1342         list_for_each_entry(cic, &cic->list, list)
1343                 changed_ioprio(cic->cfqq);
1344
1345         return 0;
1346 }
1347
1348 static struct cfq_queue *
1349 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, unsigned short ioprio,
1350               gfp_t gfp_mask)
1351 {
1352         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1353         struct cfq_queue *cfqq, *new_cfqq = NULL;
1354
1355 retry:
1356         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1357
1358         if (!cfqq) {
1359                 if (new_cfqq) {
1360                         cfqq = new_cfqq;
1361                         new_cfqq = NULL;
1362                 } else if (gfp_mask & __GFP_WAIT) {
1363                         spin_unlock_irq(cfqd->queue->queue_lock);
1364                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1365                         spin_lock_irq(cfqd->queue->queue_lock);
1366                         goto retry;
1367                 } else {
1368                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1369                         if (!cfqq)
1370                                 goto out;
1371                 }
1372
1373                 memset(cfqq, 0, sizeof(*cfqq));
1374
1375                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1376                 INIT_LIST_HEAD(&cfqq->cfq_list);
1377                 RB_CLEAR_ROOT(&cfqq->sort_list);
1378                 INIT_LIST_HEAD(&cfqq->fifo);
1379
1380                 cfqq->key = key;
1381                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1382                 atomic_set(&cfqq->ref, 0);
1383                 cfqq->cfqd = cfqd;
1384                 atomic_inc(&cfqd->ref);
1385                 cfqq->service_last = 0;
1386                 /*
1387                  * set ->slice_left to allow preemption for a new process
1388                  */
1389                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1390                 cfq_mark_cfqq_idle_window(cfqq);
1391                 cfq_mark_cfqq_prio_changed(cfqq);
1392                 cfq_init_prio_data(cfqq);
1393         }
1394
1395         if (new_cfqq)
1396                 kmem_cache_free(cfq_pool, new_cfqq);
1397
1398         atomic_inc(&cfqq->ref);
1399 out:
1400         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1401         return cfqq;
1402 }
1403
1404 /*
1405  * Setup general io context and cfq io context. There can be several cfq
1406  * io contexts per general io context, if this process is doing io to more
1407  * than one device managed by cfq. Note that caller is holding a reference to
1408  * cfqq, so we don't need to worry about it disappearing
1409  */
1410 static struct cfq_io_context *
1411 cfq_get_io_context(struct cfq_data *cfqd, pid_t pid, gfp_t gfp_mask)
1412 {
1413         struct io_context *ioc = NULL;
1414         struct cfq_io_context *cic;
1415
1416         might_sleep_if(gfp_mask & __GFP_WAIT);
1417
1418         ioc = get_io_context(gfp_mask);
1419         if (!ioc)
1420                 return NULL;
1421
1422         if ((cic = ioc->cic) == NULL) {
1423                 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1424
1425                 if (cic == NULL)
1426                         goto err;
1427
1428                 /*
1429                  * manually increment generic io_context usage count, it
1430                  * cannot go away since we are already holding one ref to it
1431                  */
1432                 ioc->cic = cic;
1433                 ioc->set_ioprio = cfq_ioc_set_ioprio;
1434                 cic->ioc = ioc;
1435                 cic->key = cfqd;
1436                 atomic_inc(&cfqd->ref);
1437         } else {
1438                 struct cfq_io_context *__cic;
1439
1440                 /*
1441                  * the first cic on the list is actually the head itself
1442                  */
1443                 if (cic->key == cfqd)
1444                         goto out;
1445
1446                 /*
1447                  * cic exists, check if we already are there. linear search
1448                  * should be ok here, the list will usually not be more than
1449                  * 1 or a few entries long
1450                  */
1451                 list_for_each_entry(__cic, &cic->list, list) {
1452                         /*
1453                          * this process is already holding a reference to
1454                          * this queue, so no need to get one more
1455                          */
1456                         if (__cic->key == cfqd) {
1457                                 cic = __cic;
1458                                 goto out;
1459                         }
1460                 }
1461
1462                 /*
1463                  * nope, process doesn't have a cic assoicated with this
1464                  * cfqq yet. get a new one and add to list
1465                  */
1466                 __cic = cfq_alloc_io_context(cfqd, gfp_mask);
1467                 if (__cic == NULL)
1468                         goto err;
1469
1470                 __cic->ioc = ioc;
1471                 __cic->key = cfqd;
1472                 atomic_inc(&cfqd->ref);
1473                 list_add(&__cic->list, &cic->list);
1474                 cic = __cic;
1475         }
1476
1477 out:
1478         return cic;
1479 err:
1480         put_io_context(ioc);
1481         return NULL;
1482 }
1483
1484 static void
1485 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1486 {
1487         unsigned long elapsed, ttime;
1488
1489         /*
1490          * if this context already has stuff queued, thinktime is from
1491          * last queue not last end
1492          */
1493 #if 0
1494         if (time_after(cic->last_end_request, cic->last_queue))
1495                 elapsed = jiffies - cic->last_end_request;
1496         else
1497                 elapsed = jiffies - cic->last_queue;
1498 #else
1499                 elapsed = jiffies - cic->last_end_request;
1500 #endif
1501
1502         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1503
1504         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1505         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1506         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1507 }
1508
1509 #define sample_valid(samples)   ((samples) > 80)
1510
1511 /*
1512  * Disable idle window if the process thinks too long or seeks so much that
1513  * it doesn't matter
1514  */
1515 static void
1516 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1517                        struct cfq_io_context *cic)
1518 {
1519         int enable_idle = cfq_cfqq_idle_window(cfqq);
1520
1521         if (!cic->ioc->task || !cfqd->cfq_slice_idle)
1522                 enable_idle = 0;
1523         else if (sample_valid(cic->ttime_samples)) {
1524                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1525                         enable_idle = 0;
1526                 else
1527                         enable_idle = 1;
1528         }
1529
1530         if (enable_idle)
1531                 cfq_mark_cfqq_idle_window(cfqq);
1532         else
1533                 cfq_clear_cfqq_idle_window(cfqq);
1534 }
1535
1536
1537 /*
1538  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1539  * no or if we aren't sure, a 1 will cause a preempt.
1540  */
1541 static int
1542 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1543                    struct cfq_rq *crq)
1544 {
1545         struct cfq_queue *cfqq = cfqd->active_queue;
1546
1547         if (cfq_class_idle(new_cfqq))
1548                 return 0;
1549
1550         if (!cfqq)
1551                 return 1;
1552
1553         if (cfq_class_idle(cfqq))
1554                 return 1;
1555         if (!cfq_cfqq_wait_request(new_cfqq))
1556                 return 0;
1557         /*
1558          * if it doesn't have slice left, forget it
1559          */
1560         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1561                 return 0;
1562         if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1563                 return 1;
1564
1565         return 0;
1566 }
1567
1568 /*
1569  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1570  * let it have half of its nominal slice.
1571  */
1572 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1573 {
1574         struct cfq_queue *__cfqq, *next;
1575
1576         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1577                 cfq_resort_rr_list(__cfqq, 1);
1578
1579         if (!cfqq->slice_left)
1580                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1581
1582         cfqq->slice_end = cfqq->slice_left + jiffies;
1583         __cfq_slice_expired(cfqd, cfqq, 1);
1584         __cfq_set_active_queue(cfqd, cfqq);
1585 }
1586
1587 /*
1588  * should really be a ll_rw_blk.c helper
1589  */
1590 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1591 {
1592         request_queue_t *q = cfqd->queue;
1593
1594         if (!blk_queue_plugged(q))
1595                 q->request_fn(q);
1596         else
1597                 __generic_unplug_device(q);
1598 }
1599
1600 /*
1601  * Called when a new fs request (crq) is added (to cfqq). Check if there's
1602  * something we should do about it
1603  */
1604 static void
1605 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1606                  struct cfq_rq *crq)
1607 {
1608         struct cfq_io_context *cic;
1609
1610         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1611
1612         /*
1613          * we never wait for an async request and we don't allow preemption
1614          * of an async request. so just return early
1615          */
1616         if (!cfq_crq_is_sync(crq))
1617                 return;
1618
1619         cic = crq->io_context;
1620
1621         cfq_update_io_thinktime(cfqd, cic);
1622         cfq_update_idle_window(cfqd, cfqq, cic);
1623
1624         cic->last_queue = jiffies;
1625
1626         if (cfqq == cfqd->active_queue) {
1627                 /*
1628                  * if we are waiting for a request for this queue, let it rip
1629                  * immediately and flag that we must not expire this queue
1630                  * just now
1631                  */
1632                 if (cfq_cfqq_wait_request(cfqq)) {
1633                         cfq_mark_cfqq_must_dispatch(cfqq);
1634                         del_timer(&cfqd->idle_slice_timer);
1635                         cfq_start_queueing(cfqd, cfqq);
1636                 }
1637         } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1638                 /*
1639                  * not the active queue - expire current slice if it is
1640                  * idle and has expired it's mean thinktime or this new queue
1641                  * has some old slice time left and is of higher priority
1642                  */
1643                 cfq_preempt_queue(cfqd, cfqq);
1644                 cfq_mark_cfqq_must_dispatch(cfqq);
1645                 cfq_start_queueing(cfqd, cfqq);
1646         }
1647 }
1648
1649 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1650 {
1651         struct cfq_data *cfqd = q->elevator->elevator_data;
1652         struct cfq_rq *crq = RQ_DATA(rq);
1653         struct cfq_queue *cfqq = crq->cfq_queue;
1654
1655         cfq_init_prio_data(cfqq);
1656
1657         cfq_add_crq_rb(crq);
1658
1659         list_add_tail(&rq->queuelist, &cfqq->fifo);
1660
1661         if (rq_mergeable(rq))
1662                 cfq_add_crq_hash(cfqd, crq);
1663
1664         cfq_crq_enqueued(cfqd, cfqq, crq);
1665 }
1666
1667 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1668 {
1669         struct cfq_rq *crq = RQ_DATA(rq);
1670         struct cfq_queue *cfqq = crq->cfq_queue;
1671         struct cfq_data *cfqd = cfqq->cfqd;
1672         const int sync = cfq_crq_is_sync(crq);
1673         unsigned long now;
1674
1675         now = jiffies;
1676
1677         WARN_ON(!cfqd->rq_in_driver);
1678         WARN_ON(!cfqq->on_dispatch[sync]);
1679         cfqd->rq_in_driver--;
1680         cfqq->on_dispatch[sync]--;
1681
1682         if (!cfq_class_idle(cfqq))
1683                 cfqd->last_end_request = now;
1684
1685         if (!cfq_cfqq_dispatched(cfqq)) {
1686                 if (cfq_cfqq_on_rr(cfqq)) {
1687                         cfqq->service_last = now;
1688                         cfq_resort_rr_list(cfqq, 0);
1689                 }
1690                 cfq_schedule_dispatch(cfqd);
1691         }
1692
1693         if (cfq_crq_is_sync(crq))
1694                 crq->io_context->last_end_request = now;
1695 }
1696
1697 static struct request *
1698 cfq_former_request(request_queue_t *q, struct request *rq)
1699 {
1700         struct cfq_rq *crq = RQ_DATA(rq);
1701         struct rb_node *rbprev = rb_prev(&crq->rb_node);
1702
1703         if (rbprev)
1704                 return rb_entry_crq(rbprev)->request;
1705
1706         return NULL;
1707 }
1708
1709 static struct request *
1710 cfq_latter_request(request_queue_t *q, struct request *rq)
1711 {
1712         struct cfq_rq *crq = RQ_DATA(rq);
1713         struct rb_node *rbnext = rb_next(&crq->rb_node);
1714
1715         if (rbnext)
1716                 return rb_entry_crq(rbnext)->request;
1717
1718         return NULL;
1719 }
1720
1721 /*
1722  * we temporarily boost lower priority queues if they are holding fs exclusive
1723  * resources. they are boosted to normal prio (CLASS_BE/4)
1724  */
1725 static void cfq_prio_boost(struct cfq_queue *cfqq)
1726 {
1727         const int ioprio_class = cfqq->ioprio_class;
1728         const int ioprio = cfqq->ioprio;
1729
1730         if (has_fs_excl()) {
1731                 /*
1732                  * boost idle prio on transactions that would lock out other
1733                  * users of the filesystem
1734                  */
1735                 if (cfq_class_idle(cfqq))
1736                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1737                 if (cfqq->ioprio > IOPRIO_NORM)
1738                         cfqq->ioprio = IOPRIO_NORM;
1739         } else {
1740                 /*
1741                  * check if we need to unboost the queue
1742                  */
1743                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1744                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1745                 if (cfqq->ioprio != cfqq->org_ioprio)
1746                         cfqq->ioprio = cfqq->org_ioprio;
1747         }
1748
1749         /*
1750          * refile between round-robin lists if we moved the priority class
1751          */
1752         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1753             cfq_cfqq_on_rr(cfqq))
1754                 cfq_resort_rr_list(cfqq, 0);
1755 }
1756
1757 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
1758 {
1759         if (rw == READ || process_sync(task))
1760                 return task->pid;
1761
1762         return CFQ_KEY_ASYNC;
1763 }
1764
1765 static inline int
1766 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1767                 struct task_struct *task, int rw)
1768 {
1769 #if 1
1770         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1771             !cfq_cfqq_must_alloc_slice(cfqq)) {
1772                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1773                 return ELV_MQUEUE_MUST;
1774         }
1775
1776         return ELV_MQUEUE_MAY;
1777 #else
1778         if (!cfqq || task->flags & PF_MEMALLOC)
1779                 return ELV_MQUEUE_MAY;
1780         if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1781                 if (cfq_cfqq_wait_request(cfqq))
1782                         return ELV_MQUEUE_MUST;
1783
1784                 /*
1785                  * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1786                  * can quickly flood the queue with writes from a single task
1787                  */
1788                 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
1789                         cfq_mark_cfqq_must_alloc_slice(cfqq);
1790                         return ELV_MQUEUE_MUST;
1791                 }
1792
1793                 return ELV_MQUEUE_MAY;
1794         }
1795         if (cfq_class_idle(cfqq))
1796                 return ELV_MQUEUE_NO;
1797         if (cfqq->allocated[rw] >= cfqd->max_queued) {
1798                 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1799                 int ret = ELV_MQUEUE_NO;
1800
1801                 if (ioc && ioc->nr_batch_requests)
1802                         ret = ELV_MQUEUE_MAY;
1803
1804                 put_io_context(ioc);
1805                 return ret;
1806         }
1807
1808         return ELV_MQUEUE_MAY;
1809 #endif
1810 }
1811
1812 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1813 {
1814         struct cfq_data *cfqd = q->elevator->elevator_data;
1815         struct task_struct *tsk = current;
1816         struct cfq_queue *cfqq;
1817
1818         /*
1819          * don't force setup of a queue from here, as a call to may_queue
1820          * does not necessarily imply that a request actually will be queued.
1821          * so just lookup a possibly existing queue, or return 'may queue'
1822          * if that fails
1823          */
1824         cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1825         if (cfqq) {
1826                 cfq_init_prio_data(cfqq);
1827                 cfq_prio_boost(cfqq);
1828
1829                 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1830         }
1831
1832         return ELV_MQUEUE_MAY;
1833 }
1834
1835 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1836 {
1837         struct cfq_data *cfqd = q->elevator->elevator_data;
1838         struct request_list *rl = &q->rq;
1839
1840         if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1841                 smp_mb();
1842                 if (waitqueue_active(&rl->wait[READ]))
1843                         wake_up(&rl->wait[READ]);
1844         }
1845
1846         if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
1847                 smp_mb();
1848                 if (waitqueue_active(&rl->wait[WRITE]))
1849                         wake_up(&rl->wait[WRITE]);
1850         }
1851 }
1852
1853 /*
1854  * queue lock held here
1855  */
1856 static void cfq_put_request(request_queue_t *q, struct request *rq)
1857 {
1858         struct cfq_data *cfqd = q->elevator->elevator_data;
1859         struct cfq_rq *crq = RQ_DATA(rq);
1860
1861         if (crq) {
1862                 struct cfq_queue *cfqq = crq->cfq_queue;
1863                 const int rw = rq_data_dir(rq);
1864
1865                 BUG_ON(!cfqq->allocated[rw]);
1866                 cfqq->allocated[rw]--;
1867
1868                 put_io_context(crq->io_context->ioc);
1869
1870                 mempool_free(crq, cfqd->crq_pool);
1871                 rq->elevator_private = NULL;
1872
1873                 cfq_check_waiters(q, cfqq);
1874                 cfq_put_queue(cfqq);
1875         }
1876 }
1877
1878 /*
1879  * Allocate cfq data structures associated with this request.
1880  */
1881 static int
1882 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
1883                 gfp_t gfp_mask)
1884 {
1885         struct cfq_data *cfqd = q->elevator->elevator_data;
1886         struct task_struct *tsk = current;
1887         struct cfq_io_context *cic;
1888         const int rw = rq_data_dir(rq);
1889         pid_t key = cfq_queue_pid(tsk, rw);
1890         struct cfq_queue *cfqq;
1891         struct cfq_rq *crq;
1892         unsigned long flags;
1893
1894         might_sleep_if(gfp_mask & __GFP_WAIT);
1895
1896         cic = cfq_get_io_context(cfqd, key, gfp_mask);
1897
1898         spin_lock_irqsave(q->queue_lock, flags);
1899
1900         if (!cic)
1901                 goto queue_fail;
1902
1903         if (!cic->cfqq) {
1904                 cfqq = cfq_get_queue(cfqd, key, tsk->ioprio, gfp_mask);
1905                 if (!cfqq)
1906                         goto queue_fail;
1907
1908                 cic->cfqq = cfqq;
1909         } else
1910                 cfqq = cic->cfqq;
1911
1912         cfqq->allocated[rw]++;
1913         cfq_clear_cfqq_must_alloc(cfqq);
1914         cfqd->rq_starved = 0;
1915         atomic_inc(&cfqq->ref);
1916         spin_unlock_irqrestore(q->queue_lock, flags);
1917
1918         crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
1919         if (crq) {
1920                 RB_CLEAR(&crq->rb_node);
1921                 crq->rb_key = 0;
1922                 crq->request = rq;
1923                 INIT_HLIST_NODE(&crq->hash);
1924                 crq->cfq_queue = cfqq;
1925                 crq->io_context = cic;
1926
1927                 if (rw == READ || process_sync(tsk))
1928                         cfq_mark_crq_is_sync(crq);
1929                 else
1930                         cfq_clear_crq_is_sync(crq);
1931
1932                 rq->elevator_private = crq;
1933                 return 0;
1934         }
1935
1936         spin_lock_irqsave(q->queue_lock, flags);
1937         cfqq->allocated[rw]--;
1938         if (!(cfqq->allocated[0] + cfqq->allocated[1]))
1939                 cfq_mark_cfqq_must_alloc(cfqq);
1940         cfq_put_queue(cfqq);
1941 queue_fail:
1942         if (cic)
1943                 put_io_context(cic->ioc);
1944         /*
1945          * mark us rq allocation starved. we need to kickstart the process
1946          * ourselves if there are no pending requests that can do it for us.
1947          * that would be an extremely rare OOM situation
1948          */
1949         cfqd->rq_starved = 1;
1950         cfq_schedule_dispatch(cfqd);
1951         spin_unlock_irqrestore(q->queue_lock, flags);
1952         return 1;
1953 }
1954
1955 static void cfq_kick_queue(void *data)
1956 {
1957         request_queue_t *q = data;
1958         struct cfq_data *cfqd = q->elevator->elevator_data;
1959         unsigned long flags;
1960
1961         spin_lock_irqsave(q->queue_lock, flags);
1962
1963         if (cfqd->rq_starved) {
1964                 struct request_list *rl = &q->rq;
1965
1966                 /*
1967                  * we aren't guaranteed to get a request after this, but we
1968                  * have to be opportunistic
1969                  */
1970                 smp_mb();
1971                 if (waitqueue_active(&rl->wait[READ]))
1972                         wake_up(&rl->wait[READ]);
1973                 if (waitqueue_active(&rl->wait[WRITE]))
1974                         wake_up(&rl->wait[WRITE]);
1975         }
1976
1977         blk_remove_plug(q);
1978         q->request_fn(q);
1979         spin_unlock_irqrestore(q->queue_lock, flags);
1980 }
1981
1982 /*
1983  * Timer running if the active_queue is currently idling inside its time slice
1984  */
1985 static void cfq_idle_slice_timer(unsigned long data)
1986 {
1987         struct cfq_data *cfqd = (struct cfq_data *) data;
1988         struct cfq_queue *cfqq;
1989         unsigned long flags;
1990
1991         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1992
1993         if ((cfqq = cfqd->active_queue) != NULL) {
1994                 unsigned long now = jiffies;
1995
1996                 /*
1997                  * expired
1998                  */
1999                 if (time_after(now, cfqq->slice_end))
2000                         goto expire;
2001
2002                 /*
2003                  * only expire and reinvoke request handler, if there are
2004                  * other queues with pending requests
2005                  */
2006                 if (!cfqd->busy_queues) {
2007                         cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2008                         add_timer(&cfqd->idle_slice_timer);
2009                         goto out_cont;
2010                 }
2011
2012                 /*
2013                  * not expired and it has a request pending, let it dispatch
2014                  */
2015                 if (!RB_EMPTY(&cfqq->sort_list)) {
2016                         cfq_mark_cfqq_must_dispatch(cfqq);
2017                         goto out_kick;
2018                 }
2019         }
2020 expire:
2021         cfq_slice_expired(cfqd, 0);
2022 out_kick:
2023         cfq_schedule_dispatch(cfqd);
2024 out_cont:
2025         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2026 }
2027
2028 /*
2029  * Timer running if an idle class queue is waiting for service
2030  */
2031 static void cfq_idle_class_timer(unsigned long data)
2032 {
2033         struct cfq_data *cfqd = (struct cfq_data *) data;
2034         unsigned long flags, end;
2035
2036         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2037
2038         /*
2039          * race with a non-idle queue, reset timer
2040          */
2041         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2042         if (!time_after_eq(jiffies, end)) {
2043                 cfqd->idle_class_timer.expires = end;
2044                 add_timer(&cfqd->idle_class_timer);
2045         } else
2046                 cfq_schedule_dispatch(cfqd);
2047
2048         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2049 }
2050
2051 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2052 {
2053         del_timer_sync(&cfqd->idle_slice_timer);
2054         del_timer_sync(&cfqd->idle_class_timer);
2055         blk_sync_queue(cfqd->queue);
2056 }
2057
2058 static void cfq_put_cfqd(struct cfq_data *cfqd)
2059 {
2060         request_queue_t *q = cfqd->queue;
2061
2062         if (!atomic_dec_and_test(&cfqd->ref))
2063                 return;
2064
2065         cfq_shutdown_timer_wq(cfqd);
2066         blk_put_queue(q);
2067
2068         mempool_destroy(cfqd->crq_pool);
2069         kfree(cfqd->crq_hash);
2070         kfree(cfqd->cfq_hash);
2071         kfree(cfqd);
2072 }
2073
2074 static void cfq_exit_queue(elevator_t *e)
2075 {
2076         struct cfq_data *cfqd = e->elevator_data;
2077
2078         cfq_shutdown_timer_wq(cfqd);
2079         cfq_put_cfqd(cfqd);
2080 }
2081
2082 static int cfq_init_queue(request_queue_t *q, elevator_t *e)
2083 {
2084         struct cfq_data *cfqd;
2085         int i;
2086
2087         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2088         if (!cfqd)
2089                 return -ENOMEM;
2090
2091         memset(cfqd, 0, sizeof(*cfqd));
2092
2093         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2094                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2095
2096         INIT_LIST_HEAD(&cfqd->busy_rr);
2097         INIT_LIST_HEAD(&cfqd->cur_rr);
2098         INIT_LIST_HEAD(&cfqd->idle_rr);
2099         INIT_LIST_HEAD(&cfqd->empty_list);
2100
2101         cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2102         if (!cfqd->crq_hash)
2103                 goto out_crqhash;
2104
2105         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2106         if (!cfqd->cfq_hash)
2107                 goto out_cfqhash;
2108
2109         cfqd->crq_pool = mempool_create(BLKDEV_MIN_RQ, mempool_alloc_slab, mempool_free_slab, crq_pool);
2110         if (!cfqd->crq_pool)
2111                 goto out_crqpool;
2112
2113         for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2114                 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2115         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2116                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2117
2118         e->elevator_data = cfqd;
2119
2120         cfqd->queue = q;
2121         atomic_inc(&q->refcnt);
2122
2123         cfqd->max_queued = q->nr_requests / 4;
2124         q->nr_batching = cfq_queued;
2125
2126         init_timer(&cfqd->idle_slice_timer);
2127         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2128         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2129
2130         init_timer(&cfqd->idle_class_timer);
2131         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2132         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2133
2134         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2135
2136         atomic_set(&cfqd->ref, 1);
2137
2138         cfqd->cfq_queued = cfq_queued;
2139         cfqd->cfq_quantum = cfq_quantum;
2140         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2141         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2142         cfqd->cfq_back_max = cfq_back_max;
2143         cfqd->cfq_back_penalty = cfq_back_penalty;
2144         cfqd->cfq_slice[0] = cfq_slice_async;
2145         cfqd->cfq_slice[1] = cfq_slice_sync;
2146         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2147         cfqd->cfq_slice_idle = cfq_slice_idle;
2148         cfqd->cfq_max_depth = cfq_max_depth;
2149
2150         return 0;
2151 out_crqpool:
2152         kfree(cfqd->cfq_hash);
2153 out_cfqhash:
2154         kfree(cfqd->crq_hash);
2155 out_crqhash:
2156         kfree(cfqd);
2157         return -ENOMEM;
2158 }
2159
2160 static void cfq_slab_kill(void)
2161 {
2162         if (crq_pool)
2163                 kmem_cache_destroy(crq_pool);
2164         if (cfq_pool)
2165                 kmem_cache_destroy(cfq_pool);
2166         if (cfq_ioc_pool)
2167                 kmem_cache_destroy(cfq_ioc_pool);
2168 }
2169
2170 static int __init cfq_slab_setup(void)
2171 {
2172         crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2173                                         NULL, NULL);
2174         if (!crq_pool)
2175                 goto fail;
2176
2177         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2178                                         NULL, NULL);
2179         if (!cfq_pool)
2180                 goto fail;
2181
2182         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2183                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2184         if (!cfq_ioc_pool)
2185                 goto fail;
2186
2187         return 0;
2188 fail:
2189         cfq_slab_kill();
2190         return -ENOMEM;
2191 }
2192
2193 /*
2194  * sysfs parts below -->
2195  */
2196 struct cfq_fs_entry {
2197         struct attribute attr;
2198         ssize_t (*show)(struct cfq_data *, char *);
2199         ssize_t (*store)(struct cfq_data *, const char *, size_t);
2200 };
2201
2202 static ssize_t
2203 cfq_var_show(unsigned int var, char *page)
2204 {
2205         return sprintf(page, "%d\n", var);
2206 }
2207
2208 static ssize_t
2209 cfq_var_store(unsigned int *var, const char *page, size_t count)
2210 {
2211         char *p = (char *) page;
2212
2213         *var = simple_strtoul(p, &p, 10);
2214         return count;
2215 }
2216
2217 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2218 static ssize_t __FUNC(struct cfq_data *cfqd, char *page)                \
2219 {                                                                       \
2220         unsigned int __data = __VAR;                                    \
2221         if (__CONV)                                                     \
2222                 __data = jiffies_to_msecs(__data);                      \
2223         return cfq_var_show(__data, (page));                            \
2224 }
2225 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2226 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2227 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2228 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2229 SHOW_FUNCTION(cfq_back_max_show, cfqd->cfq_back_max, 0);
2230 SHOW_FUNCTION(cfq_back_penalty_show, cfqd->cfq_back_penalty, 0);
2231 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2232 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2233 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2234 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2235 SHOW_FUNCTION(cfq_max_depth_show, cfqd->cfq_max_depth, 0);
2236 #undef SHOW_FUNCTION
2237
2238 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2239 static ssize_t __FUNC(struct cfq_data *cfqd, const char *page, size_t count)    \
2240 {                                                                       \
2241         unsigned int __data;                                            \
2242         int ret = cfq_var_store(&__data, (page), count);                \
2243         if (__data < (MIN))                                             \
2244                 __data = (MIN);                                         \
2245         else if (__data > (MAX))                                        \
2246                 __data = (MAX);                                         \
2247         if (__CONV)                                                     \
2248                 *(__PTR) = msecs_to_jiffies(__data);                    \
2249         else                                                            \
2250                 *(__PTR) = __data;                                      \
2251         return ret;                                                     \
2252 }
2253 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2254 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2255 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2256 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2257 STORE_FUNCTION(cfq_back_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2258 STORE_FUNCTION(cfq_back_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2259 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2260 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2261 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2262 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2263 STORE_FUNCTION(cfq_max_depth_store, &cfqd->cfq_max_depth, 1, UINT_MAX, 0);
2264 #undef STORE_FUNCTION
2265
2266 static struct cfq_fs_entry cfq_quantum_entry = {
2267         .attr = {.name = "quantum", .mode = S_IRUGO | S_IWUSR },
2268         .show = cfq_quantum_show,
2269         .store = cfq_quantum_store,
2270 };
2271 static struct cfq_fs_entry cfq_queued_entry = {
2272         .attr = {.name = "queued", .mode = S_IRUGO | S_IWUSR },
2273         .show = cfq_queued_show,
2274         .store = cfq_queued_store,
2275 };
2276 static struct cfq_fs_entry cfq_fifo_expire_sync_entry = {
2277         .attr = {.name = "fifo_expire_sync", .mode = S_IRUGO | S_IWUSR },
2278         .show = cfq_fifo_expire_sync_show,
2279         .store = cfq_fifo_expire_sync_store,
2280 };
2281 static struct cfq_fs_entry cfq_fifo_expire_async_entry = {
2282         .attr = {.name = "fifo_expire_async", .mode = S_IRUGO | S_IWUSR },
2283         .show = cfq_fifo_expire_async_show,
2284         .store = cfq_fifo_expire_async_store,
2285 };
2286 static struct cfq_fs_entry cfq_back_max_entry = {
2287         .attr = {.name = "back_seek_max", .mode = S_IRUGO | S_IWUSR },
2288         .show = cfq_back_max_show,
2289         .store = cfq_back_max_store,
2290 };
2291 static struct cfq_fs_entry cfq_back_penalty_entry = {
2292         .attr = {.name = "back_seek_penalty", .mode = S_IRUGO | S_IWUSR },
2293         .show = cfq_back_penalty_show,
2294         .store = cfq_back_penalty_store,
2295 };
2296 static struct cfq_fs_entry cfq_slice_sync_entry = {
2297         .attr = {.name = "slice_sync", .mode = S_IRUGO | S_IWUSR },
2298         .show = cfq_slice_sync_show,
2299         .store = cfq_slice_sync_store,
2300 };
2301 static struct cfq_fs_entry cfq_slice_async_entry = {
2302         .attr = {.name = "slice_async", .mode = S_IRUGO | S_IWUSR },
2303         .show = cfq_slice_async_show,
2304         .store = cfq_slice_async_store,
2305 };
2306 static struct cfq_fs_entry cfq_slice_async_rq_entry = {
2307         .attr = {.name = "slice_async_rq", .mode = S_IRUGO | S_IWUSR },
2308         .show = cfq_slice_async_rq_show,
2309         .store = cfq_slice_async_rq_store,
2310 };
2311 static struct cfq_fs_entry cfq_slice_idle_entry = {
2312         .attr = {.name = "slice_idle", .mode = S_IRUGO | S_IWUSR },
2313         .show = cfq_slice_idle_show,
2314         .store = cfq_slice_idle_store,
2315 };
2316 static struct cfq_fs_entry cfq_max_depth_entry = {
2317         .attr = {.name = "max_depth", .mode = S_IRUGO | S_IWUSR },
2318         .show = cfq_max_depth_show,
2319         .store = cfq_max_depth_store,
2320 };
2321
2322 static struct attribute *default_attrs[] = {
2323         &cfq_quantum_entry.attr,
2324         &cfq_queued_entry.attr,
2325         &cfq_fifo_expire_sync_entry.attr,
2326         &cfq_fifo_expire_async_entry.attr,
2327         &cfq_back_max_entry.attr,
2328         &cfq_back_penalty_entry.attr,
2329         &cfq_slice_sync_entry.attr,
2330         &cfq_slice_async_entry.attr,
2331         &cfq_slice_async_rq_entry.attr,
2332         &cfq_slice_idle_entry.attr,
2333         &cfq_max_depth_entry.attr,
2334         NULL,
2335 };
2336
2337 #define to_cfq(atr) container_of((atr), struct cfq_fs_entry, attr)
2338
2339 static ssize_t
2340 cfq_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
2341 {
2342         elevator_t *e = container_of(kobj, elevator_t, kobj);
2343         struct cfq_fs_entry *entry = to_cfq(attr);
2344
2345         if (!entry->show)
2346                 return -EIO;
2347
2348         return entry->show(e->elevator_data, page);
2349 }
2350
2351 static ssize_t
2352 cfq_attr_store(struct kobject *kobj, struct attribute *attr,
2353                const char *page, size_t length)
2354 {
2355         elevator_t *e = container_of(kobj, elevator_t, kobj);
2356         struct cfq_fs_entry *entry = to_cfq(attr);
2357
2358         if (!entry->store)
2359                 return -EIO;
2360
2361         return entry->store(e->elevator_data, page, length);
2362 }
2363
2364 static struct sysfs_ops cfq_sysfs_ops = {
2365         .show   = cfq_attr_show,
2366         .store  = cfq_attr_store,
2367 };
2368
2369 static struct kobj_type cfq_ktype = {
2370         .sysfs_ops      = &cfq_sysfs_ops,
2371         .default_attrs  = default_attrs,
2372 };
2373
2374 static struct elevator_type iosched_cfq = {
2375         .ops = {
2376                 .elevator_merge_fn =            cfq_merge,
2377                 .elevator_merged_fn =           cfq_merged_request,
2378                 .elevator_merge_req_fn =        cfq_merged_requests,
2379                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2380                 .elevator_add_req_fn =          cfq_insert_request,
2381                 .elevator_activate_req_fn =     cfq_activate_request,
2382                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2383                 .elevator_queue_empty_fn =      cfq_queue_empty,
2384                 .elevator_completed_req_fn =    cfq_completed_request,
2385                 .elevator_former_req_fn =       cfq_former_request,
2386                 .elevator_latter_req_fn =       cfq_latter_request,
2387                 .elevator_set_req_fn =          cfq_set_request,
2388                 .elevator_put_req_fn =          cfq_put_request,
2389                 .elevator_may_queue_fn =        cfq_may_queue,
2390                 .elevator_init_fn =             cfq_init_queue,
2391                 .elevator_exit_fn =             cfq_exit_queue,
2392         },
2393         .elevator_ktype =       &cfq_ktype,
2394         .elevator_name =        "cfq",
2395         .elevator_owner =       THIS_MODULE,
2396 };
2397
2398 static int __init cfq_init(void)
2399 {
2400         int ret;
2401
2402         /*
2403          * could be 0 on HZ < 1000 setups
2404          */
2405         if (!cfq_slice_async)
2406                 cfq_slice_async = 1;
2407         if (!cfq_slice_idle)
2408                 cfq_slice_idle = 1;
2409
2410         if (cfq_slab_setup())
2411                 return -ENOMEM;
2412
2413         ret = elv_register(&iosched_cfq);
2414         if (ret)
2415                 cfq_slab_kill();
2416
2417         return ret;
2418 }
2419
2420 static void __exit cfq_exit(void)
2421 {
2422         elv_unregister(&iosched_cfq);
2423         cfq_slab_kill();
2424 }
2425
2426 module_init(cfq_init);
2427 module_exit(cfq_exit);
2428
2429 MODULE_AUTHOR("Jens Axboe");
2430 MODULE_LICENSE("GPL");
2431 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");