452538644bceaee55def80c3ae3f856a6c65da16
[linux-3.10.git] / block / cfq-iosched.c
1 /*
2  *  linux/drivers/block/cfq-iosched.c
3  *
4  *  CFQ, or complete fairness queueing, disk scheduler.
5  *
6  *  Based on ideas from a previously unfinished io
7  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
8  *
9  *  Copyright (C) 2003 Jens Axboe <axboe@suse.de>
10  */
11 #include <linux/kernel.h>
12 #include <linux/fs.h>
13 #include <linux/blkdev.h>
14 #include <linux/elevator.h>
15 #include <linux/bio.h>
16 #include <linux/config.h>
17 #include <linux/module.h>
18 #include <linux/slab.h>
19 #include <linux/init.h>
20 #include <linux/compiler.h>
21 #include <linux/hash.h>
22 #include <linux/rbtree.h>
23 #include <linux/mempool.h>
24 #include <linux/ioprio.h>
25 #include <linux/writeback.h>
26
27 /*
28  * tunables
29  */
30 static int cfq_quantum = 4;             /* max queue in one round of service */
31 static int cfq_queued = 8;              /* minimum rq allocate limit per-queue*/
32 static int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
33 static int cfq_back_max = 16 * 1024;    /* maximum backwards seek, in KiB */
34 static int cfq_back_penalty = 2;        /* penalty of a backwards seek */
35
36 static int cfq_slice_sync = HZ / 10;
37 static int cfq_slice_async = HZ / 25;
38 static int cfq_slice_async_rq = 2;
39 static int cfq_slice_idle = HZ / 100;
40
41 #define CFQ_IDLE_GRACE          (HZ / 10)
42 #define CFQ_SLICE_SCALE         (5)
43
44 #define CFQ_KEY_ASYNC           (0)
45 #define CFQ_KEY_ANY             (0xffff)
46
47 /*
48  * disable queueing at the driver/hardware level
49  */
50 static int cfq_max_depth = 2;
51
52 /*
53  * for the hash of cfqq inside the cfqd
54  */
55 #define CFQ_QHASH_SHIFT         6
56 #define CFQ_QHASH_ENTRIES       (1 << CFQ_QHASH_SHIFT)
57 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
58
59 /*
60  * for the hash of crq inside the cfqq
61  */
62 #define CFQ_MHASH_SHIFT         6
63 #define CFQ_MHASH_BLOCK(sec)    ((sec) >> 3)
64 #define CFQ_MHASH_ENTRIES       (1 << CFQ_MHASH_SHIFT)
65 #define CFQ_MHASH_FN(sec)       hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
66 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
67 #define list_entry_hash(ptr)    hlist_entry((ptr), struct cfq_rq, hash)
68
69 #define list_entry_cfqq(ptr)    list_entry((ptr), struct cfq_queue, cfq_list)
70 #define list_entry_fifo(ptr)    list_entry((ptr), struct request, queuelist)
71
72 #define RQ_DATA(rq)             (rq)->elevator_private
73
74 /*
75  * rb-tree defines
76  */
77 #define RB_NONE                 (2)
78 #define RB_EMPTY(node)          ((node)->rb_node == NULL)
79 #define RB_CLEAR_COLOR(node)    (node)->rb_color = RB_NONE
80 #define RB_CLEAR(node)          do {    \
81         (node)->rb_parent = NULL;       \
82         RB_CLEAR_COLOR((node));         \
83         (node)->rb_right = NULL;        \
84         (node)->rb_left = NULL;         \
85 } while (0)
86 #define RB_CLEAR_ROOT(root)     ((root)->rb_node = NULL)
87 #define rb_entry_crq(node)      rb_entry((node), struct cfq_rq, rb_node)
88 #define rq_rb_key(rq)           (rq)->sector
89
90 static kmem_cache_t *crq_pool;
91 static kmem_cache_t *cfq_pool;
92 static kmem_cache_t *cfq_ioc_pool;
93
94 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
95 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
96 #define cfq_class_be(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
97 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
98
99 #define ASYNC                   (0)
100 #define SYNC                    (1)
101
102 #define cfq_cfqq_dispatched(cfqq)       \
103         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
104
105 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
106
107 #define cfq_cfqq_sync(cfqq)             \
108         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
109
110 /*
111  * Per block device queue structure
112  */
113 struct cfq_data {
114         atomic_t ref;
115         request_queue_t *queue;
116
117         /*
118          * rr list of queues with requests and the count of them
119          */
120         struct list_head rr_list[CFQ_PRIO_LISTS];
121         struct list_head busy_rr;
122         struct list_head cur_rr;
123         struct list_head idle_rr;
124         unsigned int busy_queues;
125
126         /*
127          * non-ordered list of empty cfqq's
128          */
129         struct list_head empty_list;
130
131         /*
132          * cfqq lookup hash
133          */
134         struct hlist_head *cfq_hash;
135
136         /*
137          * global crq hash for all queues
138          */
139         struct hlist_head *crq_hash;
140
141         unsigned int max_queued;
142
143         mempool_t *crq_pool;
144
145         int rq_in_driver;
146
147         /*
148          * schedule slice state info
149          */
150         /*
151          * idle window management
152          */
153         struct timer_list idle_slice_timer;
154         struct work_struct unplug_work;
155
156         struct cfq_queue *active_queue;
157         struct cfq_io_context *active_cic;
158         int cur_prio, cur_end_prio;
159         unsigned int dispatch_slice;
160
161         struct timer_list idle_class_timer;
162
163         sector_t last_sector;
164         unsigned long last_end_request;
165
166         unsigned int rq_starved;
167
168         /*
169          * tunables, see top of file
170          */
171         unsigned int cfq_quantum;
172         unsigned int cfq_queued;
173         unsigned int cfq_fifo_expire[2];
174         unsigned int cfq_back_penalty;
175         unsigned int cfq_back_max;
176         unsigned int cfq_slice[2];
177         unsigned int cfq_slice_async_rq;
178         unsigned int cfq_slice_idle;
179         unsigned int cfq_max_depth;
180 };
181
182 /*
183  * Per process-grouping structure
184  */
185 struct cfq_queue {
186         /* reference count */
187         atomic_t ref;
188         /* parent cfq_data */
189         struct cfq_data *cfqd;
190         /* cfqq lookup hash */
191         struct hlist_node cfq_hash;
192         /* hash key */
193         unsigned int key;
194         /* on either rr or empty list of cfqd */
195         struct list_head cfq_list;
196         /* sorted list of pending requests */
197         struct rb_root sort_list;
198         /* if fifo isn't expired, next request to serve */
199         struct cfq_rq *next_crq;
200         /* requests queued in sort_list */
201         int queued[2];
202         /* currently allocated requests */
203         int allocated[2];
204         /* fifo list of requests in sort_list */
205         struct list_head fifo;
206
207         unsigned long slice_start;
208         unsigned long slice_end;
209         unsigned long slice_left;
210         unsigned long service_last;
211
212         /* number of requests that are on the dispatch list */
213         int on_dispatch[2];
214
215         /* io prio of this group */
216         unsigned short ioprio, org_ioprio;
217         unsigned short ioprio_class, org_ioprio_class;
218
219         /* various state flags, see below */
220         unsigned int flags;
221 };
222
223 struct cfq_rq {
224         struct rb_node rb_node;
225         sector_t rb_key;
226         struct request *request;
227         struct hlist_node hash;
228
229         struct cfq_queue *cfq_queue;
230         struct cfq_io_context *io_context;
231
232         unsigned int crq_flags;
233 };
234
235 enum cfqq_state_flags {
236         CFQ_CFQQ_FLAG_on_rr = 0,
237         CFQ_CFQQ_FLAG_wait_request,
238         CFQ_CFQQ_FLAG_must_alloc,
239         CFQ_CFQQ_FLAG_must_alloc_slice,
240         CFQ_CFQQ_FLAG_must_dispatch,
241         CFQ_CFQQ_FLAG_fifo_expire,
242         CFQ_CFQQ_FLAG_idle_window,
243         CFQ_CFQQ_FLAG_prio_changed,
244         CFQ_CFQQ_FLAG_expired,
245 };
246
247 #define CFQ_CFQQ_FNS(name)                                              \
248 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
249 {                                                                       \
250         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
251 }                                                                       \
252 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
253 {                                                                       \
254         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
255 }                                                                       \
256 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
257 {                                                                       \
258         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
259 }
260
261 CFQ_CFQQ_FNS(on_rr);
262 CFQ_CFQQ_FNS(wait_request);
263 CFQ_CFQQ_FNS(must_alloc);
264 CFQ_CFQQ_FNS(must_alloc_slice);
265 CFQ_CFQQ_FNS(must_dispatch);
266 CFQ_CFQQ_FNS(fifo_expire);
267 CFQ_CFQQ_FNS(idle_window);
268 CFQ_CFQQ_FNS(prio_changed);
269 CFQ_CFQQ_FNS(expired);
270 #undef CFQ_CFQQ_FNS
271
272 enum cfq_rq_state_flags {
273         CFQ_CRQ_FLAG_is_sync = 0,
274 };
275
276 #define CFQ_CRQ_FNS(name)                                               \
277 static inline void cfq_mark_crq_##name(struct cfq_rq *crq)              \
278 {                                                                       \
279         crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name);                   \
280 }                                                                       \
281 static inline void cfq_clear_crq_##name(struct cfq_rq *crq)             \
282 {                                                                       \
283         crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name);                  \
284 }                                                                       \
285 static inline int cfq_crq_##name(const struct cfq_rq *crq)              \
286 {                                                                       \
287         return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0;      \
288 }
289
290 CFQ_CRQ_FNS(is_sync);
291 #undef CFQ_CRQ_FNS
292
293 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
294 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
295 static void cfq_put_cfqd(struct cfq_data *cfqd);
296
297 #define process_sync(tsk)       ((tsk)->flags & PF_SYNCWRITE)
298
299 /*
300  * lots of deadline iosched dupes, can be abstracted later...
301  */
302 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
303 {
304         hlist_del_init(&crq->hash);
305 }
306
307 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
308 {
309         const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
310
311         hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
312 }
313
314 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
315 {
316         struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
317         struct hlist_node *entry, *next;
318
319         hlist_for_each_safe(entry, next, hash_list) {
320                 struct cfq_rq *crq = list_entry_hash(entry);
321                 struct request *__rq = crq->request;
322
323                 if (!rq_mergeable(__rq)) {
324                         cfq_del_crq_hash(crq);
325                         continue;
326                 }
327
328                 if (rq_hash_key(__rq) == offset)
329                         return __rq;
330         }
331
332         return NULL;
333 }
334
335 /*
336  * scheduler run of queue, if there are requests pending and no one in the
337  * driver that will restart queueing
338  */
339 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
340 {
341         if (!cfqd->rq_in_driver && cfqd->busy_queues)
342                 kblockd_schedule_work(&cfqd->unplug_work);
343 }
344
345 static int cfq_queue_empty(request_queue_t *q)
346 {
347         struct cfq_data *cfqd = q->elevator->elevator_data;
348
349         return !cfqd->busy_queues;
350 }
351
352 /*
353  * Lifted from AS - choose which of crq1 and crq2 that is best served now.
354  * We choose the request that is closest to the head right now. Distance
355  * behind the head are penalized and only allowed to a certain extent.
356  */
357 static struct cfq_rq *
358 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
359 {
360         sector_t last, s1, s2, d1 = 0, d2 = 0;
361         int r1_wrap = 0, r2_wrap = 0;   /* requests are behind the disk head */
362         unsigned long back_max;
363
364         if (crq1 == NULL || crq1 == crq2)
365                 return crq2;
366         if (crq2 == NULL)
367                 return crq1;
368
369         if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
370                 return crq1;
371         else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
372                 return crq2;
373
374         s1 = crq1->request->sector;
375         s2 = crq2->request->sector;
376
377         last = cfqd->last_sector;
378
379         /*
380          * by definition, 1KiB is 2 sectors
381          */
382         back_max = cfqd->cfq_back_max * 2;
383
384         /*
385          * Strict one way elevator _except_ in the case where we allow
386          * short backward seeks which are biased as twice the cost of a
387          * similar forward seek.
388          */
389         if (s1 >= last)
390                 d1 = s1 - last;
391         else if (s1 + back_max >= last)
392                 d1 = (last - s1) * cfqd->cfq_back_penalty;
393         else
394                 r1_wrap = 1;
395
396         if (s2 >= last)
397                 d2 = s2 - last;
398         else if (s2 + back_max >= last)
399                 d2 = (last - s2) * cfqd->cfq_back_penalty;
400         else
401                 r2_wrap = 1;
402
403         /* Found required data */
404         if (!r1_wrap && r2_wrap)
405                 return crq1;
406         else if (!r2_wrap && r1_wrap)
407                 return crq2;
408         else if (r1_wrap && r2_wrap) {
409                 /* both behind the head */
410                 if (s1 <= s2)
411                         return crq1;
412                 else
413                         return crq2;
414         }
415
416         /* Both requests in front of the head */
417         if (d1 < d2)
418                 return crq1;
419         else if (d2 < d1)
420                 return crq2;
421         else {
422                 if (s1 >= s2)
423                         return crq1;
424                 else
425                         return crq2;
426         }
427 }
428
429 /*
430  * would be nice to take fifo expire time into account as well
431  */
432 static struct cfq_rq *
433 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
434                   struct cfq_rq *last)
435 {
436         struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
437         struct rb_node *rbnext, *rbprev;
438
439         if (!(rbnext = rb_next(&last->rb_node))) {
440                 rbnext = rb_first(&cfqq->sort_list);
441                 if (rbnext == &last->rb_node)
442                         rbnext = NULL;
443         }
444
445         rbprev = rb_prev(&last->rb_node);
446
447         if (rbprev)
448                 crq_prev = rb_entry_crq(rbprev);
449         if (rbnext)
450                 crq_next = rb_entry_crq(rbnext);
451
452         return cfq_choose_req(cfqd, crq_next, crq_prev);
453 }
454
455 static void cfq_update_next_crq(struct cfq_rq *crq)
456 {
457         struct cfq_queue *cfqq = crq->cfq_queue;
458
459         if (cfqq->next_crq == crq)
460                 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
461 }
462
463 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
464 {
465         struct cfq_data *cfqd = cfqq->cfqd;
466         struct list_head *list, *entry;
467
468         BUG_ON(!cfq_cfqq_on_rr(cfqq));
469
470         list_del(&cfqq->cfq_list);
471
472         if (cfq_class_rt(cfqq))
473                 list = &cfqd->cur_rr;
474         else if (cfq_class_idle(cfqq))
475                 list = &cfqd->idle_rr;
476         else {
477                 /*
478                  * if cfqq has requests in flight, don't allow it to be
479                  * found in cfq_set_active_queue before it has finished them.
480                  * this is done to increase fairness between a process that
481                  * has lots of io pending vs one that only generates one
482                  * sporadically or synchronously
483                  */
484                 if (cfq_cfqq_dispatched(cfqq))
485                         list = &cfqd->busy_rr;
486                 else
487                         list = &cfqd->rr_list[cfqq->ioprio];
488         }
489
490         /*
491          * if queue was preempted, just add to front to be fair. busy_rr
492          * isn't sorted.
493          */
494         if (preempted || list == &cfqd->busy_rr) {
495                 list_add(&cfqq->cfq_list, list);
496                 return;
497         }
498
499         /*
500          * sort by when queue was last serviced
501          */
502         entry = list;
503         while ((entry = entry->prev) != list) {
504                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
505
506                 if (!__cfqq->service_last)
507                         break;
508                 if (time_before(__cfqq->service_last, cfqq->service_last))
509                         break;
510         }
511
512         list_add(&cfqq->cfq_list, entry);
513 }
514
515 /*
516  * add to busy list of queues for service, trying to be fair in ordering
517  * the pending list according to last request service
518  */
519 static inline void
520 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
521 {
522         BUG_ON(cfq_cfqq_on_rr(cfqq));
523         cfq_mark_cfqq_on_rr(cfqq);
524         cfqd->busy_queues++;
525
526         cfq_resort_rr_list(cfqq, 0);
527 }
528
529 static inline void
530 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
531 {
532         BUG_ON(!cfq_cfqq_on_rr(cfqq));
533         cfq_clear_cfqq_on_rr(cfqq);
534         list_move(&cfqq->cfq_list, &cfqd->empty_list);
535
536         BUG_ON(!cfqd->busy_queues);
537         cfqd->busy_queues--;
538 }
539
540 /*
541  * rb tree support functions
542  */
543 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
544 {
545         struct cfq_queue *cfqq = crq->cfq_queue;
546         struct cfq_data *cfqd = cfqq->cfqd;
547         const int sync = cfq_crq_is_sync(crq);
548
549         BUG_ON(!cfqq->queued[sync]);
550         cfqq->queued[sync]--;
551
552         cfq_update_next_crq(crq);
553
554         rb_erase(&crq->rb_node, &cfqq->sort_list);
555         RB_CLEAR_COLOR(&crq->rb_node);
556
557         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
558                 cfq_del_cfqq_rr(cfqd, cfqq);
559 }
560
561 static struct cfq_rq *
562 __cfq_add_crq_rb(struct cfq_rq *crq)
563 {
564         struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
565         struct rb_node *parent = NULL;
566         struct cfq_rq *__crq;
567
568         while (*p) {
569                 parent = *p;
570                 __crq = rb_entry_crq(parent);
571
572                 if (crq->rb_key < __crq->rb_key)
573                         p = &(*p)->rb_left;
574                 else if (crq->rb_key > __crq->rb_key)
575                         p = &(*p)->rb_right;
576                 else
577                         return __crq;
578         }
579
580         rb_link_node(&crq->rb_node, parent, p);
581         return NULL;
582 }
583
584 static void cfq_add_crq_rb(struct cfq_rq *crq)
585 {
586         struct cfq_queue *cfqq = crq->cfq_queue;
587         struct cfq_data *cfqd = cfqq->cfqd;
588         struct request *rq = crq->request;
589         struct cfq_rq *__alias;
590
591         crq->rb_key = rq_rb_key(rq);
592         cfqq->queued[cfq_crq_is_sync(crq)]++;
593
594         /*
595          * looks a little odd, but the first insert might return an alias.
596          * if that happens, put the alias on the dispatch list
597          */
598         while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
599                 cfq_dispatch_insert(cfqd->queue, __alias);
600
601         rb_insert_color(&crq->rb_node, &cfqq->sort_list);
602
603         if (!cfq_cfqq_on_rr(cfqq))
604                 cfq_add_cfqq_rr(cfqd, cfqq);
605
606         /*
607          * check if this request is a better next-serve candidate
608          */
609         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
610 }
611
612 static inline void
613 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
614 {
615         rb_erase(&crq->rb_node, &cfqq->sort_list);
616         cfqq->queued[cfq_crq_is_sync(crq)]--;
617
618         cfq_add_crq_rb(crq);
619 }
620
621 static struct request *cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector)
622
623 {
624         struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, current->pid, CFQ_KEY_ANY);
625         struct rb_node *n;
626
627         if (!cfqq)
628                 goto out;
629
630         n = cfqq->sort_list.rb_node;
631         while (n) {
632                 struct cfq_rq *crq = rb_entry_crq(n);
633
634                 if (sector < crq->rb_key)
635                         n = n->rb_left;
636                 else if (sector > crq->rb_key)
637                         n = n->rb_right;
638                 else
639                         return crq->request;
640         }
641
642 out:
643         return NULL;
644 }
645
646 static void cfq_activate_request(request_queue_t *q, struct request *rq)
647 {
648         struct cfq_data *cfqd = q->elevator->elevator_data;
649
650         cfqd->rq_in_driver++;
651 }
652
653 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
654 {
655         struct cfq_data *cfqd = q->elevator->elevator_data;
656
657         WARN_ON(!cfqd->rq_in_driver);
658         cfqd->rq_in_driver--;
659 }
660
661 static void cfq_remove_request(struct request *rq)
662 {
663         struct cfq_rq *crq = RQ_DATA(rq);
664
665         list_del_init(&rq->queuelist);
666         cfq_del_crq_rb(crq);
667         cfq_del_crq_hash(crq);
668 }
669
670 static int
671 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
672 {
673         struct cfq_data *cfqd = q->elevator->elevator_data;
674         struct request *__rq;
675         int ret;
676
677         __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
678         if (__rq && elv_rq_merge_ok(__rq, bio)) {
679                 ret = ELEVATOR_BACK_MERGE;
680                 goto out;
681         }
682
683         __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio));
684         if (__rq && elv_rq_merge_ok(__rq, bio)) {
685                 ret = ELEVATOR_FRONT_MERGE;
686                 goto out;
687         }
688
689         return ELEVATOR_NO_MERGE;
690 out:
691         *req = __rq;
692         return ret;
693 }
694
695 static void cfq_merged_request(request_queue_t *q, struct request *req)
696 {
697         struct cfq_data *cfqd = q->elevator->elevator_data;
698         struct cfq_rq *crq = RQ_DATA(req);
699
700         cfq_del_crq_hash(crq);
701         cfq_add_crq_hash(cfqd, crq);
702
703         if (rq_rb_key(req) != crq->rb_key) {
704                 struct cfq_queue *cfqq = crq->cfq_queue;
705
706                 cfq_update_next_crq(crq);
707                 cfq_reposition_crq_rb(cfqq, crq);
708         }
709 }
710
711 static void
712 cfq_merged_requests(request_queue_t *q, struct request *rq,
713                     struct request *next)
714 {
715         cfq_merged_request(q, rq);
716
717         /*
718          * reposition in fifo if next is older than rq
719          */
720         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
721             time_before(next->start_time, rq->start_time))
722                 list_move(&rq->queuelist, &next->queuelist);
723
724         cfq_remove_request(next);
725 }
726
727 static inline void
728 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
729 {
730         if (cfqq) {
731                 /*
732                  * stop potential idle class queues waiting service
733                  */
734                 del_timer(&cfqd->idle_class_timer);
735
736                 cfqq->slice_start = jiffies;
737                 cfqq->slice_end = 0;
738                 cfqq->slice_left = 0;
739                 cfq_clear_cfqq_must_alloc_slice(cfqq);
740                 cfq_clear_cfqq_fifo_expire(cfqq);
741                 cfq_clear_cfqq_expired(cfqq);
742         }
743
744         cfqd->active_queue = cfqq;
745 }
746
747 /*
748  * 0
749  * 0,1
750  * 0,1,2
751  * 0,1,2,3
752  * 0,1,2,3,4
753  * 0,1,2,3,4,5
754  * 0,1,2,3,4,5,6
755  * 0,1,2,3,4,5,6,7
756  */
757 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
758 {
759         int prio, wrap;
760
761         prio = -1;
762         wrap = 0;
763         do {
764                 int p;
765
766                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
767                         if (!list_empty(&cfqd->rr_list[p])) {
768                                 prio = p;
769                                 break;
770                         }
771                 }
772
773                 if (prio != -1)
774                         break;
775                 cfqd->cur_prio = 0;
776                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
777                         cfqd->cur_end_prio = 0;
778                         if (wrap)
779                                 break;
780                         wrap = 1;
781                 }
782         } while (1);
783
784         if (unlikely(prio == -1))
785                 return -1;
786
787         BUG_ON(prio >= CFQ_PRIO_LISTS);
788
789         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
790
791         cfqd->cur_prio = prio + 1;
792         if (cfqd->cur_prio > cfqd->cur_end_prio) {
793                 cfqd->cur_end_prio = cfqd->cur_prio;
794                 cfqd->cur_prio = 0;
795         }
796         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
797                 cfqd->cur_prio = 0;
798                 cfqd->cur_end_prio = 0;
799         }
800
801         return prio;
802 }
803
804 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
805 {
806         struct cfq_queue *cfqq;
807
808         /*
809          * if current queue is expired but not done with its requests yet,
810          * wait for that to happen
811          */
812         if ((cfqq = cfqd->active_queue) != NULL) {
813                 if (cfq_cfqq_expired(cfqq) && cfq_cfqq_dispatched(cfqq))
814                         return NULL;
815         }
816
817         /*
818          * if current list is non-empty, grab first entry. if it is empty,
819          * get next prio level and grab first entry then if any are spliced
820          */
821         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
822                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
823
824         /*
825          * if we have idle queues and no rt or be queues had pending
826          * requests, either allow immediate service if the grace period
827          * has passed or arm the idle grace timer
828          */
829         if (!cfqq && !list_empty(&cfqd->idle_rr)) {
830                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
831
832                 if (time_after_eq(jiffies, end))
833                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
834                 else
835                         mod_timer(&cfqd->idle_class_timer, end);
836         }
837
838         __cfq_set_active_queue(cfqd, cfqq);
839         return cfqq;
840 }
841
842 /*
843  * current cfqq expired its slice (or was too idle), select new one
844  */
845 static void
846 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
847                     int preempted)
848 {
849         unsigned long now = jiffies;
850
851         if (cfq_cfqq_wait_request(cfqq))
852                 del_timer(&cfqd->idle_slice_timer);
853
854         if (!preempted && !cfq_cfqq_dispatched(cfqq))
855                 cfqq->service_last = now;
856
857         cfq_clear_cfqq_must_dispatch(cfqq);
858         cfq_clear_cfqq_wait_request(cfqq);
859
860         /*
861          * store what was left of this slice, if the queue idled out
862          * or was preempted
863          */
864         if (time_after(now, cfqq->slice_end))
865                 cfqq->slice_left = now - cfqq->slice_end;
866         else
867                 cfqq->slice_left = 0;
868
869         if (cfq_cfqq_on_rr(cfqq))
870                 cfq_resort_rr_list(cfqq, preempted);
871
872         if (cfqq == cfqd->active_queue)
873                 cfqd->active_queue = NULL;
874
875         if (cfqd->active_cic) {
876                 put_io_context(cfqd->active_cic->ioc);
877                 cfqd->active_cic = NULL;
878         }
879
880         cfqd->dispatch_slice = 0;
881 }
882
883 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
884 {
885         struct cfq_queue *cfqq = cfqd->active_queue;
886
887         if (cfqq) {
888                 /*
889                  * use deferred expiry, if there are requests in progress as
890                  * not to disturb the slice of the next queue
891                  */
892                 if (cfq_cfqq_dispatched(cfqq))
893                         cfq_mark_cfqq_expired(cfqq);
894                 else
895                         __cfq_slice_expired(cfqd, cfqq, preempted);
896         }
897 }
898
899 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
900
901 {
902         WARN_ON(!RB_EMPTY(&cfqq->sort_list));
903         WARN_ON(cfqq != cfqd->active_queue);
904
905         /*
906          * idle is disabled, either manually or by past process history
907          */
908         if (!cfqd->cfq_slice_idle)
909                 return 0;
910         if (!cfq_cfqq_idle_window(cfqq))
911                 return 0;
912         /*
913          * task has exited, don't wait
914          */
915         if (cfqd->active_cic && !cfqd->active_cic->ioc->task)
916                 return 0;
917
918         cfq_mark_cfqq_must_dispatch(cfqq);
919         cfq_mark_cfqq_wait_request(cfqq);
920
921         if (!timer_pending(&cfqd->idle_slice_timer)) {
922                 unsigned long slice_left = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
923
924                 cfqd->idle_slice_timer.expires = jiffies + slice_left;
925                 add_timer(&cfqd->idle_slice_timer);
926         }
927
928         return 1;
929 }
930
931 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
932 {
933         struct cfq_data *cfqd = q->elevator->elevator_data;
934         struct cfq_queue *cfqq = crq->cfq_queue;
935
936         cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
937         cfq_remove_request(crq->request);
938         cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
939         elv_dispatch_sort(q, crq->request);
940 }
941
942 /*
943  * return expired entry, or NULL to just start from scratch in rbtree
944  */
945 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
946 {
947         struct cfq_data *cfqd = cfqq->cfqd;
948         struct request *rq;
949         struct cfq_rq *crq;
950
951         if (cfq_cfqq_fifo_expire(cfqq))
952                 return NULL;
953
954         if (!list_empty(&cfqq->fifo)) {
955                 int fifo = cfq_cfqq_class_sync(cfqq);
956
957                 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
958                 rq = crq->request;
959                 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
960                         cfq_mark_cfqq_fifo_expire(cfqq);
961                         return crq;
962                 }
963         }
964
965         return NULL;
966 }
967
968 /*
969  * Scale schedule slice based on io priority. Use the sync time slice only
970  * if a queue is marked sync and has sync io queued. A sync queue with async
971  * io only, should not get full sync slice length.
972  */
973 static inline int
974 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
975 {
976         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
977
978         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
979
980         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
981 }
982
983 static inline void
984 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
985 {
986         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
987 }
988
989 static inline int
990 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
991 {
992         const int base_rq = cfqd->cfq_slice_async_rq;
993
994         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
995
996         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
997 }
998
999 /*
1000  * get next queue for service
1001  */
1002 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1003 {
1004         unsigned long now = jiffies;
1005         struct cfq_queue *cfqq;
1006
1007         cfqq = cfqd->active_queue;
1008         if (!cfqq)
1009                 goto new_queue;
1010
1011         if (cfq_cfqq_expired(cfqq))
1012                 goto new_queue;
1013
1014         /*
1015          * slice has expired
1016          */
1017         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
1018                 goto expire;
1019
1020         /*
1021          * if queue has requests, dispatch one. if not, check if
1022          * enough slice is left to wait for one
1023          */
1024         if (!RB_EMPTY(&cfqq->sort_list))
1025                 goto keep_queue;
1026         else if (cfq_cfqq_class_sync(cfqq) &&
1027                  time_before(now, cfqq->slice_end)) {
1028                 if (cfq_arm_slice_timer(cfqd, cfqq))
1029                         return NULL;
1030         }
1031
1032 expire:
1033         cfq_slice_expired(cfqd, 0);
1034 new_queue:
1035         cfqq = cfq_set_active_queue(cfqd);
1036 keep_queue:
1037         return cfqq;
1038 }
1039
1040 static int
1041 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1042                         int max_dispatch)
1043 {
1044         int dispatched = 0;
1045
1046         BUG_ON(RB_EMPTY(&cfqq->sort_list));
1047
1048         do {
1049                 struct cfq_rq *crq;
1050
1051                 /*
1052                  * follow expired path, else get first next available
1053                  */
1054                 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1055                         crq = cfqq->next_crq;
1056
1057                 /*
1058                  * finally, insert request into driver dispatch list
1059                  */
1060                 cfq_dispatch_insert(cfqd->queue, crq);
1061
1062                 cfqd->dispatch_slice++;
1063                 dispatched++;
1064
1065                 if (!cfqd->active_cic) {
1066                         atomic_inc(&crq->io_context->ioc->refcount);
1067                         cfqd->active_cic = crq->io_context;
1068                 }
1069
1070                 if (RB_EMPTY(&cfqq->sort_list))
1071                         break;
1072
1073         } while (dispatched < max_dispatch);
1074
1075         /*
1076          * if slice end isn't set yet, set it. if at least one request was
1077          * sync, use the sync time slice value
1078          */
1079         if (!cfqq->slice_end)
1080                 cfq_set_prio_slice(cfqd, cfqq);
1081
1082         /*
1083          * expire an async queue immediately if it has used up its slice. idle
1084          * queue always expire after 1 dispatch round.
1085          */
1086         if ((!cfq_cfqq_sync(cfqq) &&
1087             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1088             cfq_class_idle(cfqq))
1089                 cfq_slice_expired(cfqd, 0);
1090
1091         return dispatched;
1092 }
1093
1094 static int
1095 cfq_forced_dispatch_cfqqs(struct list_head *list)
1096 {
1097         int dispatched = 0;
1098         struct cfq_queue *cfqq, *next;
1099         struct cfq_rq *crq;
1100
1101         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1102                 while ((crq = cfqq->next_crq)) {
1103                         cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1104                         dispatched++;
1105                 }
1106                 BUG_ON(!list_empty(&cfqq->fifo));
1107         }
1108         return dispatched;
1109 }
1110
1111 static int
1112 cfq_forced_dispatch(struct cfq_data *cfqd)
1113 {
1114         int i, dispatched = 0;
1115
1116         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1117                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1118
1119         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1120         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1121         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1122
1123         cfq_slice_expired(cfqd, 0);
1124
1125         BUG_ON(cfqd->busy_queues);
1126
1127         return dispatched;
1128 }
1129
1130 static int
1131 cfq_dispatch_requests(request_queue_t *q, int force)
1132 {
1133         struct cfq_data *cfqd = q->elevator->elevator_data;
1134         struct cfq_queue *cfqq;
1135
1136         if (!cfqd->busy_queues)
1137                 return 0;
1138
1139         if (unlikely(force))
1140                 return cfq_forced_dispatch(cfqd);
1141
1142         cfqq = cfq_select_queue(cfqd);
1143         if (cfqq) {
1144                 int max_dispatch;
1145
1146                 /*
1147                  * if idle window is disabled, allow queue buildup
1148                  */
1149                 if (!cfq_cfqq_idle_window(cfqq) &&
1150                     cfqd->rq_in_driver >= cfqd->cfq_max_depth)
1151                         return 0;
1152
1153                 cfq_clear_cfqq_must_dispatch(cfqq);
1154                 cfq_clear_cfqq_wait_request(cfqq);
1155                 del_timer(&cfqd->idle_slice_timer);
1156
1157                 max_dispatch = cfqd->cfq_quantum;
1158                 if (cfq_class_idle(cfqq))
1159                         max_dispatch = 1;
1160
1161                 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1162         }
1163
1164         return 0;
1165 }
1166
1167 /*
1168  * task holds one reference to the queue, dropped when task exits. each crq
1169  * in-flight on this queue also holds a reference, dropped when crq is freed.
1170  *
1171  * queue lock must be held here.
1172  */
1173 static void cfq_put_queue(struct cfq_queue *cfqq)
1174 {
1175         struct cfq_data *cfqd = cfqq->cfqd;
1176
1177         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1178
1179         if (!atomic_dec_and_test(&cfqq->ref))
1180                 return;
1181
1182         BUG_ON(rb_first(&cfqq->sort_list));
1183         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1184         BUG_ON(cfq_cfqq_on_rr(cfqq));
1185
1186         if (unlikely(cfqd->active_queue == cfqq)) {
1187                 __cfq_slice_expired(cfqd, cfqq, 0);
1188                 cfq_schedule_dispatch(cfqd);
1189         }
1190
1191         cfq_put_cfqd(cfqq->cfqd);
1192
1193         /*
1194          * it's on the empty list and still hashed
1195          */
1196         list_del(&cfqq->cfq_list);
1197         hlist_del(&cfqq->cfq_hash);
1198         kmem_cache_free(cfq_pool, cfqq);
1199 }
1200
1201 static inline struct cfq_queue *
1202 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1203                     const int hashval)
1204 {
1205         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1206         struct hlist_node *entry, *next;
1207
1208         hlist_for_each_safe(entry, next, hash_list) {
1209                 struct cfq_queue *__cfqq = list_entry_qhash(entry);
1210                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->ioprio_class, __cfqq->ioprio);
1211
1212                 if (__cfqq->key == key && (__p == prio || prio == CFQ_KEY_ANY))
1213                         return __cfqq;
1214         }
1215
1216         return NULL;
1217 }
1218
1219 static struct cfq_queue *
1220 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1221 {
1222         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1223 }
1224
1225 static void cfq_free_io_context(struct cfq_io_context *cic)
1226 {
1227         struct cfq_io_context *__cic;
1228         struct list_head *entry, *next;
1229
1230         list_for_each_safe(entry, next, &cic->list) {
1231                 __cic = list_entry(entry, struct cfq_io_context, list);
1232                 kmem_cache_free(cfq_ioc_pool, __cic);
1233         }
1234
1235         kmem_cache_free(cfq_ioc_pool, cic);
1236 }
1237
1238 /*
1239  * Called with interrupts disabled
1240  */
1241 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1242 {
1243         struct cfq_data *cfqd = cic->cfqq->cfqd;
1244         request_queue_t *q = cfqd->queue;
1245
1246         WARN_ON(!irqs_disabled());
1247
1248         spin_lock(q->queue_lock);
1249
1250         if (unlikely(cic->cfqq == cfqd->active_queue)) {
1251                 __cfq_slice_expired(cfqd, cic->cfqq, 0);
1252                 cfq_schedule_dispatch(cfqd);
1253         }
1254
1255         cfq_put_queue(cic->cfqq);
1256         cic->cfqq = NULL;
1257         spin_unlock(q->queue_lock);
1258 }
1259
1260 /*
1261  * Another task may update the task cic list, if it is doing a queue lookup
1262  * on its behalf. cfq_cic_lock excludes such concurrent updates
1263  */
1264 static void cfq_exit_io_context(struct cfq_io_context *cic)
1265 {
1266         struct cfq_io_context *__cic;
1267         struct list_head *entry;
1268         unsigned long flags;
1269
1270         local_irq_save(flags);
1271
1272         /*
1273          * put the reference this task is holding to the various queues
1274          */
1275         list_for_each(entry, &cic->list) {
1276                 __cic = list_entry(entry, struct cfq_io_context, list);
1277                 cfq_exit_single_io_context(__cic);
1278         }
1279
1280         cfq_exit_single_io_context(cic);
1281         local_irq_restore(flags);
1282 }
1283
1284 static struct cfq_io_context *
1285 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1286 {
1287         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1288
1289         if (cic) {
1290                 INIT_LIST_HEAD(&cic->list);
1291                 cic->cfqq = NULL;
1292                 cic->key = NULL;
1293                 cic->last_end_request = jiffies;
1294                 cic->ttime_total = 0;
1295                 cic->ttime_samples = 0;
1296                 cic->ttime_mean = 0;
1297                 cic->dtor = cfq_free_io_context;
1298                 cic->exit = cfq_exit_io_context;
1299         }
1300
1301         return cic;
1302 }
1303
1304 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1305 {
1306         struct task_struct *tsk = current;
1307         int ioprio_class;
1308
1309         if (!cfq_cfqq_prio_changed(cfqq))
1310                 return;
1311
1312         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1313         switch (ioprio_class) {
1314                 default:
1315                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1316                 case IOPRIO_CLASS_NONE:
1317                         /*
1318                          * no prio set, place us in the middle of the BE classes
1319                          */
1320                         cfqq->ioprio = task_nice_ioprio(tsk);
1321                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1322                         break;
1323                 case IOPRIO_CLASS_RT:
1324                         cfqq->ioprio = task_ioprio(tsk);
1325                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1326                         break;
1327                 case IOPRIO_CLASS_BE:
1328                         cfqq->ioprio = task_ioprio(tsk);
1329                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1330                         break;
1331                 case IOPRIO_CLASS_IDLE:
1332                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1333                         cfqq->ioprio = 7;
1334                         cfq_clear_cfqq_idle_window(cfqq);
1335                         break;
1336         }
1337
1338         /*
1339          * keep track of original prio settings in case we have to temporarily
1340          * elevate the priority of this queue
1341          */
1342         cfqq->org_ioprio = cfqq->ioprio;
1343         cfqq->org_ioprio_class = cfqq->ioprio_class;
1344
1345         if (cfq_cfqq_on_rr(cfqq))
1346                 cfq_resort_rr_list(cfqq, 0);
1347
1348         cfq_clear_cfqq_prio_changed(cfqq);
1349 }
1350
1351 static inline void changed_ioprio(struct cfq_queue *cfqq)
1352 {
1353         if (cfqq) {
1354                 struct cfq_data *cfqd = cfqq->cfqd;
1355
1356                 spin_lock(cfqd->queue->queue_lock);
1357                 cfq_mark_cfqq_prio_changed(cfqq);
1358                 cfq_init_prio_data(cfqq);
1359                 spin_unlock(cfqd->queue->queue_lock);
1360         }
1361 }
1362
1363 /*
1364  * callback from sys_ioprio_set, irqs are disabled
1365  */
1366 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1367 {
1368         struct cfq_io_context *cic = ioc->cic;
1369
1370         changed_ioprio(cic->cfqq);
1371
1372         list_for_each_entry(cic, &cic->list, list)
1373                 changed_ioprio(cic->cfqq);
1374
1375         return 0;
1376 }
1377
1378 static struct cfq_queue *
1379 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, unsigned short ioprio,
1380               gfp_t gfp_mask)
1381 {
1382         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1383         struct cfq_queue *cfqq, *new_cfqq = NULL;
1384
1385 retry:
1386         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1387
1388         if (!cfqq) {
1389                 if (new_cfqq) {
1390                         cfqq = new_cfqq;
1391                         new_cfqq = NULL;
1392                 } else if (gfp_mask & __GFP_WAIT) {
1393                         spin_unlock_irq(cfqd->queue->queue_lock);
1394                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1395                         spin_lock_irq(cfqd->queue->queue_lock);
1396                         goto retry;
1397                 } else {
1398                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1399                         if (!cfqq)
1400                                 goto out;
1401                 }
1402
1403                 memset(cfqq, 0, sizeof(*cfqq));
1404
1405                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1406                 INIT_LIST_HEAD(&cfqq->cfq_list);
1407                 RB_CLEAR_ROOT(&cfqq->sort_list);
1408                 INIT_LIST_HEAD(&cfqq->fifo);
1409
1410                 cfqq->key = key;
1411                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1412                 atomic_set(&cfqq->ref, 0);
1413                 cfqq->cfqd = cfqd;
1414                 atomic_inc(&cfqd->ref);
1415                 cfqq->service_last = 0;
1416                 /*
1417                  * set ->slice_left to allow preemption for a new process
1418                  */
1419                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1420                 cfq_mark_cfqq_idle_window(cfqq);
1421                 cfq_mark_cfqq_prio_changed(cfqq);
1422                 cfq_init_prio_data(cfqq);
1423         }
1424
1425         if (new_cfqq)
1426                 kmem_cache_free(cfq_pool, new_cfqq);
1427
1428         atomic_inc(&cfqq->ref);
1429 out:
1430         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1431         return cfqq;
1432 }
1433
1434 /*
1435  * Setup general io context and cfq io context. There can be several cfq
1436  * io contexts per general io context, if this process is doing io to more
1437  * than one device managed by cfq. Note that caller is holding a reference to
1438  * cfqq, so we don't need to worry about it disappearing
1439  */
1440 static struct cfq_io_context *
1441 cfq_get_io_context(struct cfq_data *cfqd, pid_t pid, gfp_t gfp_mask)
1442 {
1443         struct io_context *ioc = NULL;
1444         struct cfq_io_context *cic;
1445
1446         might_sleep_if(gfp_mask & __GFP_WAIT);
1447
1448         ioc = get_io_context(gfp_mask);
1449         if (!ioc)
1450                 return NULL;
1451
1452         if ((cic = ioc->cic) == NULL) {
1453                 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1454
1455                 if (cic == NULL)
1456                         goto err;
1457
1458                 /*
1459                  * manually increment generic io_context usage count, it
1460                  * cannot go away since we are already holding one ref to it
1461                  */
1462                 ioc->cic = cic;
1463                 ioc->set_ioprio = cfq_ioc_set_ioprio;
1464                 cic->ioc = ioc;
1465                 cic->key = cfqd;
1466                 atomic_inc(&cfqd->ref);
1467         } else {
1468                 struct cfq_io_context *__cic;
1469
1470                 /*
1471                  * the first cic on the list is actually the head itself
1472                  */
1473                 if (cic->key == cfqd)
1474                         goto out;
1475
1476                 /*
1477                  * cic exists, check if we already are there. linear search
1478                  * should be ok here, the list will usually not be more than
1479                  * 1 or a few entries long
1480                  */
1481                 list_for_each_entry(__cic, &cic->list, list) {
1482                         /*
1483                          * this process is already holding a reference to
1484                          * this queue, so no need to get one more
1485                          */
1486                         if (__cic->key == cfqd) {
1487                                 cic = __cic;
1488                                 goto out;
1489                         }
1490                 }
1491
1492                 /*
1493                  * nope, process doesn't have a cic assoicated with this
1494                  * cfqq yet. get a new one and add to list
1495                  */
1496                 __cic = cfq_alloc_io_context(cfqd, gfp_mask);
1497                 if (__cic == NULL)
1498                         goto err;
1499
1500                 __cic->ioc = ioc;
1501                 __cic->key = cfqd;
1502                 atomic_inc(&cfqd->ref);
1503                 list_add(&__cic->list, &cic->list);
1504                 cic = __cic;
1505         }
1506
1507 out:
1508         return cic;
1509 err:
1510         put_io_context(ioc);
1511         return NULL;
1512 }
1513
1514 static void
1515 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1516 {
1517         unsigned long elapsed, ttime;
1518
1519         /*
1520          * if this context already has stuff queued, thinktime is from
1521          * last queue not last end
1522          */
1523 #if 0
1524         if (time_after(cic->last_end_request, cic->last_queue))
1525                 elapsed = jiffies - cic->last_end_request;
1526         else
1527                 elapsed = jiffies - cic->last_queue;
1528 #else
1529                 elapsed = jiffies - cic->last_end_request;
1530 #endif
1531
1532         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1533
1534         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1535         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1536         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1537 }
1538
1539 #define sample_valid(samples)   ((samples) > 80)
1540
1541 /*
1542  * Disable idle window if the process thinks too long or seeks so much that
1543  * it doesn't matter
1544  */
1545 static void
1546 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1547                        struct cfq_io_context *cic)
1548 {
1549         int enable_idle = cfq_cfqq_idle_window(cfqq);
1550
1551         if (!cic->ioc->task || !cfqd->cfq_slice_idle)
1552                 enable_idle = 0;
1553         else if (sample_valid(cic->ttime_samples)) {
1554                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1555                         enable_idle = 0;
1556                 else
1557                         enable_idle = 1;
1558         }
1559
1560         if (enable_idle)
1561                 cfq_mark_cfqq_idle_window(cfqq);
1562         else
1563                 cfq_clear_cfqq_idle_window(cfqq);
1564 }
1565
1566
1567 /*
1568  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1569  * no or if we aren't sure, a 1 will cause a preempt.
1570  */
1571 static int
1572 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1573                    struct cfq_rq *crq)
1574 {
1575         struct cfq_queue *cfqq = cfqd->active_queue;
1576
1577         if (cfq_class_idle(new_cfqq))
1578                 return 0;
1579
1580         if (!cfqq)
1581                 return 1;
1582
1583         if (cfq_class_idle(cfqq))
1584                 return 1;
1585         if (!cfq_cfqq_wait_request(new_cfqq))
1586                 return 0;
1587         /*
1588          * if it doesn't have slice left, forget it
1589          */
1590         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1591                 return 0;
1592         if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1593                 return 1;
1594
1595         return 0;
1596 }
1597
1598 /*
1599  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1600  * let it have half of its nominal slice.
1601  */
1602 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1603 {
1604         struct cfq_queue *__cfqq, *next;
1605
1606         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1607                 cfq_resort_rr_list(__cfqq, 1);
1608
1609         if (!cfqq->slice_left)
1610                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1611
1612         cfqq->slice_end = cfqq->slice_left + jiffies;
1613         __cfq_slice_expired(cfqd, cfqq, 1);
1614         __cfq_set_active_queue(cfqd, cfqq);
1615 }
1616
1617 /*
1618  * should really be a ll_rw_blk.c helper
1619  */
1620 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1621 {
1622         request_queue_t *q = cfqd->queue;
1623
1624         if (!blk_queue_plugged(q))
1625                 q->request_fn(q);
1626         else
1627                 __generic_unplug_device(q);
1628 }
1629
1630 /*
1631  * Called when a new fs request (crq) is added (to cfqq). Check if there's
1632  * something we should do about it
1633  */
1634 static void
1635 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1636                  struct cfq_rq *crq)
1637 {
1638         struct cfq_io_context *cic;
1639
1640         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1641
1642         /*
1643          * we never wait for an async request and we don't allow preemption
1644          * of an async request. so just return early
1645          */
1646         if (!cfq_crq_is_sync(crq))
1647                 return;
1648
1649         cic = crq->io_context;
1650
1651         cfq_update_io_thinktime(cfqd, cic);
1652         cfq_update_idle_window(cfqd, cfqq, cic);
1653
1654         cic->last_queue = jiffies;
1655
1656         if (cfqq == cfqd->active_queue) {
1657                 /*
1658                  * if we are waiting for a request for this queue, let it rip
1659                  * immediately and flag that we must not expire this queue
1660                  * just now
1661                  */
1662                 if (cfq_cfqq_wait_request(cfqq)) {
1663                         cfq_mark_cfqq_must_dispatch(cfqq);
1664                         del_timer(&cfqd->idle_slice_timer);
1665                         cfq_start_queueing(cfqd, cfqq);
1666                 }
1667         } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1668                 /*
1669                  * not the active queue - expire current slice if it is
1670                  * idle and has expired it's mean thinktime or this new queue
1671                  * has some old slice time left and is of higher priority
1672                  */
1673                 cfq_preempt_queue(cfqd, cfqq);
1674                 cfq_mark_cfqq_must_dispatch(cfqq);
1675                 cfq_start_queueing(cfqd, cfqq);
1676         }
1677 }
1678
1679 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1680 {
1681         struct cfq_data *cfqd = q->elevator->elevator_data;
1682         struct cfq_rq *crq = RQ_DATA(rq);
1683         struct cfq_queue *cfqq = crq->cfq_queue;
1684
1685         cfq_init_prio_data(cfqq);
1686
1687         cfq_add_crq_rb(crq);
1688
1689         list_add_tail(&rq->queuelist, &cfqq->fifo);
1690
1691         if (rq_mergeable(rq))
1692                 cfq_add_crq_hash(cfqd, crq);
1693
1694         cfq_crq_enqueued(cfqd, cfqq, crq);
1695 }
1696
1697 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1698 {
1699         struct cfq_rq *crq = RQ_DATA(rq);
1700         struct cfq_queue *cfqq = crq->cfq_queue;
1701         struct cfq_data *cfqd = cfqq->cfqd;
1702         const int sync = cfq_crq_is_sync(crq);
1703         unsigned long now;
1704
1705         now = jiffies;
1706
1707         WARN_ON(!cfqd->rq_in_driver);
1708         WARN_ON(!cfqq->on_dispatch[sync]);
1709         cfqd->rq_in_driver--;
1710         cfqq->on_dispatch[sync]--;
1711
1712         if (!cfq_class_idle(cfqq))
1713                 cfqd->last_end_request = now;
1714
1715         if (!cfq_cfqq_dispatched(cfqq)) {
1716                 if (cfq_cfqq_on_rr(cfqq)) {
1717                         cfqq->service_last = now;
1718                         cfq_resort_rr_list(cfqq, 0);
1719                 }
1720                 if (cfq_cfqq_expired(cfqq)) {
1721                         __cfq_slice_expired(cfqd, cfqq, 0);
1722                         cfq_schedule_dispatch(cfqd);
1723                 }
1724         }
1725
1726         if (cfq_crq_is_sync(crq))
1727                 crq->io_context->last_end_request = now;
1728 }
1729
1730 static struct request *
1731 cfq_former_request(request_queue_t *q, struct request *rq)
1732 {
1733         struct cfq_rq *crq = RQ_DATA(rq);
1734         struct rb_node *rbprev = rb_prev(&crq->rb_node);
1735
1736         if (rbprev)
1737                 return rb_entry_crq(rbprev)->request;
1738
1739         return NULL;
1740 }
1741
1742 static struct request *
1743 cfq_latter_request(request_queue_t *q, struct request *rq)
1744 {
1745         struct cfq_rq *crq = RQ_DATA(rq);
1746         struct rb_node *rbnext = rb_next(&crq->rb_node);
1747
1748         if (rbnext)
1749                 return rb_entry_crq(rbnext)->request;
1750
1751         return NULL;
1752 }
1753
1754 /*
1755  * we temporarily boost lower priority queues if they are holding fs exclusive
1756  * resources. they are boosted to normal prio (CLASS_BE/4)
1757  */
1758 static void cfq_prio_boost(struct cfq_queue *cfqq)
1759 {
1760         const int ioprio_class = cfqq->ioprio_class;
1761         const int ioprio = cfqq->ioprio;
1762
1763         if (has_fs_excl()) {
1764                 /*
1765                  * boost idle prio on transactions that would lock out other
1766                  * users of the filesystem
1767                  */
1768                 if (cfq_class_idle(cfqq))
1769                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1770                 if (cfqq->ioprio > IOPRIO_NORM)
1771                         cfqq->ioprio = IOPRIO_NORM;
1772         } else {
1773                 /*
1774                  * check if we need to unboost the queue
1775                  */
1776                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1777                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1778                 if (cfqq->ioprio != cfqq->org_ioprio)
1779                         cfqq->ioprio = cfqq->org_ioprio;
1780         }
1781
1782         /*
1783          * refile between round-robin lists if we moved the priority class
1784          */
1785         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1786             cfq_cfqq_on_rr(cfqq))
1787                 cfq_resort_rr_list(cfqq, 0);
1788 }
1789
1790 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
1791 {
1792         if (rw == READ || process_sync(task))
1793                 return task->pid;
1794
1795         return CFQ_KEY_ASYNC;
1796 }
1797
1798 static inline int
1799 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1800                 struct task_struct *task, int rw)
1801 {
1802 #if 1
1803         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1804             !cfq_cfqq_must_alloc_slice(cfqq)) {
1805                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1806                 return ELV_MQUEUE_MUST;
1807         }
1808
1809         return ELV_MQUEUE_MAY;
1810 #else
1811         if (!cfqq || task->flags & PF_MEMALLOC)
1812                 return ELV_MQUEUE_MAY;
1813         if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1814                 if (cfq_cfqq_wait_request(cfqq))
1815                         return ELV_MQUEUE_MUST;
1816
1817                 /*
1818                  * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1819                  * can quickly flood the queue with writes from a single task
1820                  */
1821                 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
1822                         cfq_mark_cfqq_must_alloc_slice(cfqq);
1823                         return ELV_MQUEUE_MUST;
1824                 }
1825
1826                 return ELV_MQUEUE_MAY;
1827         }
1828         if (cfq_class_idle(cfqq))
1829                 return ELV_MQUEUE_NO;
1830         if (cfqq->allocated[rw] >= cfqd->max_queued) {
1831                 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1832                 int ret = ELV_MQUEUE_NO;
1833
1834                 if (ioc && ioc->nr_batch_requests)
1835                         ret = ELV_MQUEUE_MAY;
1836
1837                 put_io_context(ioc);
1838                 return ret;
1839         }
1840
1841         return ELV_MQUEUE_MAY;
1842 #endif
1843 }
1844
1845 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1846 {
1847         struct cfq_data *cfqd = q->elevator->elevator_data;
1848         struct task_struct *tsk = current;
1849         struct cfq_queue *cfqq;
1850
1851         /*
1852          * don't force setup of a queue from here, as a call to may_queue
1853          * does not necessarily imply that a request actually will be queued.
1854          * so just lookup a possibly existing queue, or return 'may queue'
1855          * if that fails
1856          */
1857         cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1858         if (cfqq) {
1859                 cfq_init_prio_data(cfqq);
1860                 cfq_prio_boost(cfqq);
1861
1862                 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1863         }
1864
1865         return ELV_MQUEUE_MAY;
1866 }
1867
1868 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1869 {
1870         struct cfq_data *cfqd = q->elevator->elevator_data;
1871         struct request_list *rl = &q->rq;
1872
1873         if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1874                 smp_mb();
1875                 if (waitqueue_active(&rl->wait[READ]))
1876                         wake_up(&rl->wait[READ]);
1877         }
1878
1879         if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
1880                 smp_mb();
1881                 if (waitqueue_active(&rl->wait[WRITE]))
1882                         wake_up(&rl->wait[WRITE]);
1883         }
1884 }
1885
1886 /*
1887  * queue lock held here
1888  */
1889 static void cfq_put_request(request_queue_t *q, struct request *rq)
1890 {
1891         struct cfq_data *cfqd = q->elevator->elevator_data;
1892         struct cfq_rq *crq = RQ_DATA(rq);
1893
1894         if (crq) {
1895                 struct cfq_queue *cfqq = crq->cfq_queue;
1896                 const int rw = rq_data_dir(rq);
1897
1898                 BUG_ON(!cfqq->allocated[rw]);
1899                 cfqq->allocated[rw]--;
1900
1901                 put_io_context(crq->io_context->ioc);
1902
1903                 mempool_free(crq, cfqd->crq_pool);
1904                 rq->elevator_private = NULL;
1905
1906                 cfq_check_waiters(q, cfqq);
1907                 cfq_put_queue(cfqq);
1908         }
1909 }
1910
1911 /*
1912  * Allocate cfq data structures associated with this request.
1913  */
1914 static int
1915 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
1916                 gfp_t gfp_mask)
1917 {
1918         struct cfq_data *cfqd = q->elevator->elevator_data;
1919         struct task_struct *tsk = current;
1920         struct cfq_io_context *cic;
1921         const int rw = rq_data_dir(rq);
1922         pid_t key = cfq_queue_pid(tsk, rw);
1923         struct cfq_queue *cfqq;
1924         struct cfq_rq *crq;
1925         unsigned long flags;
1926
1927         might_sleep_if(gfp_mask & __GFP_WAIT);
1928
1929         cic = cfq_get_io_context(cfqd, key, gfp_mask);
1930
1931         spin_lock_irqsave(q->queue_lock, flags);
1932
1933         if (!cic)
1934                 goto queue_fail;
1935
1936         if (!cic->cfqq) {
1937                 cfqq = cfq_get_queue(cfqd, key, tsk->ioprio, gfp_mask);
1938                 if (!cfqq)
1939                         goto queue_fail;
1940
1941                 cic->cfqq = cfqq;
1942         } else
1943                 cfqq = cic->cfqq;
1944
1945         cfqq->allocated[rw]++;
1946         cfq_clear_cfqq_must_alloc(cfqq);
1947         cfqd->rq_starved = 0;
1948         atomic_inc(&cfqq->ref);
1949         spin_unlock_irqrestore(q->queue_lock, flags);
1950
1951         crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
1952         if (crq) {
1953                 RB_CLEAR(&crq->rb_node);
1954                 crq->rb_key = 0;
1955                 crq->request = rq;
1956                 INIT_HLIST_NODE(&crq->hash);
1957                 crq->cfq_queue = cfqq;
1958                 crq->io_context = cic;
1959
1960                 if (rw == READ || process_sync(tsk))
1961                         cfq_mark_crq_is_sync(crq);
1962                 else
1963                         cfq_clear_crq_is_sync(crq);
1964
1965                 rq->elevator_private = crq;
1966                 return 0;
1967         }
1968
1969         spin_lock_irqsave(q->queue_lock, flags);
1970         cfqq->allocated[rw]--;
1971         if (!(cfqq->allocated[0] + cfqq->allocated[1]))
1972                 cfq_mark_cfqq_must_alloc(cfqq);
1973         cfq_put_queue(cfqq);
1974 queue_fail:
1975         if (cic)
1976                 put_io_context(cic->ioc);
1977         /*
1978          * mark us rq allocation starved. we need to kickstart the process
1979          * ourselves if there are no pending requests that can do it for us.
1980          * that would be an extremely rare OOM situation
1981          */
1982         cfqd->rq_starved = 1;
1983         cfq_schedule_dispatch(cfqd);
1984         spin_unlock_irqrestore(q->queue_lock, flags);
1985         return 1;
1986 }
1987
1988 static void cfq_kick_queue(void *data)
1989 {
1990         request_queue_t *q = data;
1991         struct cfq_data *cfqd = q->elevator->elevator_data;
1992         unsigned long flags;
1993
1994         spin_lock_irqsave(q->queue_lock, flags);
1995
1996         if (cfqd->rq_starved) {
1997                 struct request_list *rl = &q->rq;
1998
1999                 /*
2000                  * we aren't guaranteed to get a request after this, but we
2001                  * have to be opportunistic
2002                  */
2003                 smp_mb();
2004                 if (waitqueue_active(&rl->wait[READ]))
2005                         wake_up(&rl->wait[READ]);
2006                 if (waitqueue_active(&rl->wait[WRITE]))
2007                         wake_up(&rl->wait[WRITE]);
2008         }
2009
2010         blk_remove_plug(q);
2011         q->request_fn(q);
2012         spin_unlock_irqrestore(q->queue_lock, flags);
2013 }
2014
2015 /*
2016  * Timer running if the active_queue is currently idling inside its time slice
2017  */
2018 static void cfq_idle_slice_timer(unsigned long data)
2019 {
2020         struct cfq_data *cfqd = (struct cfq_data *) data;
2021         struct cfq_queue *cfqq;
2022         unsigned long flags;
2023
2024         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2025
2026         if ((cfqq = cfqd->active_queue) != NULL) {
2027                 unsigned long now = jiffies;
2028
2029                 /*
2030                  * expired
2031                  */
2032                 if (time_after(now, cfqq->slice_end))
2033                         goto expire;
2034
2035                 /*
2036                  * only expire and reinvoke request handler, if there are
2037                  * other queues with pending requests
2038                  */
2039                 if (!cfqd->busy_queues) {
2040                         cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2041                         add_timer(&cfqd->idle_slice_timer);
2042                         goto out_cont;
2043                 }
2044
2045                 /*
2046                  * not expired and it has a request pending, let it dispatch
2047                  */
2048                 if (!RB_EMPTY(&cfqq->sort_list)) {
2049                         cfq_mark_cfqq_must_dispatch(cfqq);
2050                         goto out_kick;
2051                 }
2052         }
2053 expire:
2054         cfq_slice_expired(cfqd, 0);
2055 out_kick:
2056         cfq_schedule_dispatch(cfqd);
2057 out_cont:
2058         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2059 }
2060
2061 /*
2062  * Timer running if an idle class queue is waiting for service
2063  */
2064 static void cfq_idle_class_timer(unsigned long data)
2065 {
2066         struct cfq_data *cfqd = (struct cfq_data *) data;
2067         unsigned long flags, end;
2068
2069         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2070
2071         /*
2072          * race with a non-idle queue, reset timer
2073          */
2074         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2075         if (!time_after_eq(jiffies, end)) {
2076                 cfqd->idle_class_timer.expires = end;
2077                 add_timer(&cfqd->idle_class_timer);
2078         } else
2079                 cfq_schedule_dispatch(cfqd);
2080
2081         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2082 }
2083
2084 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2085 {
2086         del_timer_sync(&cfqd->idle_slice_timer);
2087         del_timer_sync(&cfqd->idle_class_timer);
2088         blk_sync_queue(cfqd->queue);
2089 }
2090
2091 static void cfq_put_cfqd(struct cfq_data *cfqd)
2092 {
2093         request_queue_t *q = cfqd->queue;
2094
2095         if (!atomic_dec_and_test(&cfqd->ref))
2096                 return;
2097
2098         cfq_shutdown_timer_wq(cfqd);
2099         blk_put_queue(q);
2100
2101         mempool_destroy(cfqd->crq_pool);
2102         kfree(cfqd->crq_hash);
2103         kfree(cfqd->cfq_hash);
2104         kfree(cfqd);
2105 }
2106
2107 static void cfq_exit_queue(elevator_t *e)
2108 {
2109         struct cfq_data *cfqd = e->elevator_data;
2110
2111         cfq_shutdown_timer_wq(cfqd);
2112         cfq_put_cfqd(cfqd);
2113 }
2114
2115 static int cfq_init_queue(request_queue_t *q, elevator_t *e)
2116 {
2117         struct cfq_data *cfqd;
2118         int i;
2119
2120         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2121         if (!cfqd)
2122                 return -ENOMEM;
2123
2124         memset(cfqd, 0, sizeof(*cfqd));
2125
2126         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2127                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2128
2129         INIT_LIST_HEAD(&cfqd->busy_rr);
2130         INIT_LIST_HEAD(&cfqd->cur_rr);
2131         INIT_LIST_HEAD(&cfqd->idle_rr);
2132         INIT_LIST_HEAD(&cfqd->empty_list);
2133
2134         cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2135         if (!cfqd->crq_hash)
2136                 goto out_crqhash;
2137
2138         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2139         if (!cfqd->cfq_hash)
2140                 goto out_cfqhash;
2141
2142         cfqd->crq_pool = mempool_create(BLKDEV_MIN_RQ, mempool_alloc_slab, mempool_free_slab, crq_pool);
2143         if (!cfqd->crq_pool)
2144                 goto out_crqpool;
2145
2146         for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2147                 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2148         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2149                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2150
2151         e->elevator_data = cfqd;
2152
2153         cfqd->queue = q;
2154         atomic_inc(&q->refcnt);
2155
2156         cfqd->max_queued = q->nr_requests / 4;
2157         q->nr_batching = cfq_queued;
2158
2159         init_timer(&cfqd->idle_slice_timer);
2160         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2161         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2162
2163         init_timer(&cfqd->idle_class_timer);
2164         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2165         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2166
2167         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2168
2169         atomic_set(&cfqd->ref, 1);
2170
2171         cfqd->cfq_queued = cfq_queued;
2172         cfqd->cfq_quantum = cfq_quantum;
2173         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2174         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2175         cfqd->cfq_back_max = cfq_back_max;
2176         cfqd->cfq_back_penalty = cfq_back_penalty;
2177         cfqd->cfq_slice[0] = cfq_slice_async;
2178         cfqd->cfq_slice[1] = cfq_slice_sync;
2179         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2180         cfqd->cfq_slice_idle = cfq_slice_idle;
2181         cfqd->cfq_max_depth = cfq_max_depth;
2182
2183         return 0;
2184 out_crqpool:
2185         kfree(cfqd->cfq_hash);
2186 out_cfqhash:
2187         kfree(cfqd->crq_hash);
2188 out_crqhash:
2189         kfree(cfqd);
2190         return -ENOMEM;
2191 }
2192
2193 static void cfq_slab_kill(void)
2194 {
2195         if (crq_pool)
2196                 kmem_cache_destroy(crq_pool);
2197         if (cfq_pool)
2198                 kmem_cache_destroy(cfq_pool);
2199         if (cfq_ioc_pool)
2200                 kmem_cache_destroy(cfq_ioc_pool);
2201 }
2202
2203 static int __init cfq_slab_setup(void)
2204 {
2205         crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2206                                         NULL, NULL);
2207         if (!crq_pool)
2208                 goto fail;
2209
2210         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2211                                         NULL, NULL);
2212         if (!cfq_pool)
2213                 goto fail;
2214
2215         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2216                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2217         if (!cfq_ioc_pool)
2218                 goto fail;
2219
2220         return 0;
2221 fail:
2222         cfq_slab_kill();
2223         return -ENOMEM;
2224 }
2225
2226 /*
2227  * sysfs parts below -->
2228  */
2229 struct cfq_fs_entry {
2230         struct attribute attr;
2231         ssize_t (*show)(struct cfq_data *, char *);
2232         ssize_t (*store)(struct cfq_data *, const char *, size_t);
2233 };
2234
2235 static ssize_t
2236 cfq_var_show(unsigned int var, char *page)
2237 {
2238         return sprintf(page, "%d\n", var);
2239 }
2240
2241 static ssize_t
2242 cfq_var_store(unsigned int *var, const char *page, size_t count)
2243 {
2244         char *p = (char *) page;
2245
2246         *var = simple_strtoul(p, &p, 10);
2247         return count;
2248 }
2249
2250 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2251 static ssize_t __FUNC(struct cfq_data *cfqd, char *page)                \
2252 {                                                                       \
2253         unsigned int __data = __VAR;                                    \
2254         if (__CONV)                                                     \
2255                 __data = jiffies_to_msecs(__data);                      \
2256         return cfq_var_show(__data, (page));                            \
2257 }
2258 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2259 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2260 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2261 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2262 SHOW_FUNCTION(cfq_back_max_show, cfqd->cfq_back_max, 0);
2263 SHOW_FUNCTION(cfq_back_penalty_show, cfqd->cfq_back_penalty, 0);
2264 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2265 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2266 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2267 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2268 SHOW_FUNCTION(cfq_max_depth_show, cfqd->cfq_max_depth, 0);
2269 #undef SHOW_FUNCTION
2270
2271 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2272 static ssize_t __FUNC(struct cfq_data *cfqd, const char *page, size_t count)    \
2273 {                                                                       \
2274         unsigned int __data;                                            \
2275         int ret = cfq_var_store(&__data, (page), count);                \
2276         if (__data < (MIN))                                             \
2277                 __data = (MIN);                                         \
2278         else if (__data > (MAX))                                        \
2279                 __data = (MAX);                                         \
2280         if (__CONV)                                                     \
2281                 *(__PTR) = msecs_to_jiffies(__data);                    \
2282         else                                                            \
2283                 *(__PTR) = __data;                                      \
2284         return ret;                                                     \
2285 }
2286 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2287 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2288 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2289 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2290 STORE_FUNCTION(cfq_back_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2291 STORE_FUNCTION(cfq_back_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2292 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2293 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2294 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2295 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2296 STORE_FUNCTION(cfq_max_depth_store, &cfqd->cfq_max_depth, 1, UINT_MAX, 0);
2297 #undef STORE_FUNCTION
2298
2299 static struct cfq_fs_entry cfq_quantum_entry = {
2300         .attr = {.name = "quantum", .mode = S_IRUGO | S_IWUSR },
2301         .show = cfq_quantum_show,
2302         .store = cfq_quantum_store,
2303 };
2304 static struct cfq_fs_entry cfq_queued_entry = {
2305         .attr = {.name = "queued", .mode = S_IRUGO | S_IWUSR },
2306         .show = cfq_queued_show,
2307         .store = cfq_queued_store,
2308 };
2309 static struct cfq_fs_entry cfq_fifo_expire_sync_entry = {
2310         .attr = {.name = "fifo_expire_sync", .mode = S_IRUGO | S_IWUSR },
2311         .show = cfq_fifo_expire_sync_show,
2312         .store = cfq_fifo_expire_sync_store,
2313 };
2314 static struct cfq_fs_entry cfq_fifo_expire_async_entry = {
2315         .attr = {.name = "fifo_expire_async", .mode = S_IRUGO | S_IWUSR },
2316         .show = cfq_fifo_expire_async_show,
2317         .store = cfq_fifo_expire_async_store,
2318 };
2319 static struct cfq_fs_entry cfq_back_max_entry = {
2320         .attr = {.name = "back_seek_max", .mode = S_IRUGO | S_IWUSR },
2321         .show = cfq_back_max_show,
2322         .store = cfq_back_max_store,
2323 };
2324 static struct cfq_fs_entry cfq_back_penalty_entry = {
2325         .attr = {.name = "back_seek_penalty", .mode = S_IRUGO | S_IWUSR },
2326         .show = cfq_back_penalty_show,
2327         .store = cfq_back_penalty_store,
2328 };
2329 static struct cfq_fs_entry cfq_slice_sync_entry = {
2330         .attr = {.name = "slice_sync", .mode = S_IRUGO | S_IWUSR },
2331         .show = cfq_slice_sync_show,
2332         .store = cfq_slice_sync_store,
2333 };
2334 static struct cfq_fs_entry cfq_slice_async_entry = {
2335         .attr = {.name = "slice_async", .mode = S_IRUGO | S_IWUSR },
2336         .show = cfq_slice_async_show,
2337         .store = cfq_slice_async_store,
2338 };
2339 static struct cfq_fs_entry cfq_slice_async_rq_entry = {
2340         .attr = {.name = "slice_async_rq", .mode = S_IRUGO | S_IWUSR },
2341         .show = cfq_slice_async_rq_show,
2342         .store = cfq_slice_async_rq_store,
2343 };
2344 static struct cfq_fs_entry cfq_slice_idle_entry = {
2345         .attr = {.name = "slice_idle", .mode = S_IRUGO | S_IWUSR },
2346         .show = cfq_slice_idle_show,
2347         .store = cfq_slice_idle_store,
2348 };
2349 static struct cfq_fs_entry cfq_max_depth_entry = {
2350         .attr = {.name = "max_depth", .mode = S_IRUGO | S_IWUSR },
2351         .show = cfq_max_depth_show,
2352         .store = cfq_max_depth_store,
2353 };
2354
2355 static struct attribute *default_attrs[] = {
2356         &cfq_quantum_entry.attr,
2357         &cfq_queued_entry.attr,
2358         &cfq_fifo_expire_sync_entry.attr,
2359         &cfq_fifo_expire_async_entry.attr,
2360         &cfq_back_max_entry.attr,
2361         &cfq_back_penalty_entry.attr,
2362         &cfq_slice_sync_entry.attr,
2363         &cfq_slice_async_entry.attr,
2364         &cfq_slice_async_rq_entry.attr,
2365         &cfq_slice_idle_entry.attr,
2366         &cfq_max_depth_entry.attr,
2367         NULL,
2368 };
2369
2370 #define to_cfq(atr) container_of((atr), struct cfq_fs_entry, attr)
2371
2372 static ssize_t
2373 cfq_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
2374 {
2375         elevator_t *e = container_of(kobj, elevator_t, kobj);
2376         struct cfq_fs_entry *entry = to_cfq(attr);
2377
2378         if (!entry->show)
2379                 return -EIO;
2380
2381         return entry->show(e->elevator_data, page);
2382 }
2383
2384 static ssize_t
2385 cfq_attr_store(struct kobject *kobj, struct attribute *attr,
2386                const char *page, size_t length)
2387 {
2388         elevator_t *e = container_of(kobj, elevator_t, kobj);
2389         struct cfq_fs_entry *entry = to_cfq(attr);
2390
2391         if (!entry->store)
2392                 return -EIO;
2393
2394         return entry->store(e->elevator_data, page, length);
2395 }
2396
2397 static struct sysfs_ops cfq_sysfs_ops = {
2398         .show   = cfq_attr_show,
2399         .store  = cfq_attr_store,
2400 };
2401
2402 static struct kobj_type cfq_ktype = {
2403         .sysfs_ops      = &cfq_sysfs_ops,
2404         .default_attrs  = default_attrs,
2405 };
2406
2407 static struct elevator_type iosched_cfq = {
2408         .ops = {
2409                 .elevator_merge_fn =            cfq_merge,
2410                 .elevator_merged_fn =           cfq_merged_request,
2411                 .elevator_merge_req_fn =        cfq_merged_requests,
2412                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2413                 .elevator_add_req_fn =          cfq_insert_request,
2414                 .elevator_activate_req_fn =     cfq_activate_request,
2415                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2416                 .elevator_queue_empty_fn =      cfq_queue_empty,
2417                 .elevator_completed_req_fn =    cfq_completed_request,
2418                 .elevator_former_req_fn =       cfq_former_request,
2419                 .elevator_latter_req_fn =       cfq_latter_request,
2420                 .elevator_set_req_fn =          cfq_set_request,
2421                 .elevator_put_req_fn =          cfq_put_request,
2422                 .elevator_may_queue_fn =        cfq_may_queue,
2423                 .elevator_init_fn =             cfq_init_queue,
2424                 .elevator_exit_fn =             cfq_exit_queue,
2425         },
2426         .elevator_ktype =       &cfq_ktype,
2427         .elevator_name =        "cfq",
2428         .elevator_owner =       THIS_MODULE,
2429 };
2430
2431 static int __init cfq_init(void)
2432 {
2433         int ret;
2434
2435         /*
2436          * could be 0 on HZ < 1000 setups
2437          */
2438         if (!cfq_slice_async)
2439                 cfq_slice_async = 1;
2440         if (!cfq_slice_idle)
2441                 cfq_slice_idle = 1;
2442
2443         if (cfq_slab_setup())
2444                 return -ENOMEM;
2445
2446         ret = elv_register(&iosched_cfq);
2447         if (ret)
2448                 cfq_slab_kill();
2449
2450         return ret;
2451 }
2452
2453 static void __exit cfq_exit(void)
2454 {
2455         elv_unregister(&iosched_cfq);
2456         cfq_slab_kill();
2457 }
2458
2459 module_init(cfq_init);
2460 module_exit(cfq_exit);
2461
2462 MODULE_AUTHOR("Jens Axboe");
2463 MODULE_LICENSE("GPL");
2464 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");