Merge branch 'master' of master.kernel.org:/pub/scm/linux/kernel/git/davem/net-2.6
[linux-2.6] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/rbtree.h>
13 #include <linux/ioprio.h>
14
15 /*
16  * tunables
17  */
18 /* max queue in one round of service */
19 static const int cfq_quantum = 4;
20 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
21 /* maximum backwards seek, in KiB */
22 static const int cfq_back_max = 16 * 1024;
23 /* penalty of a backwards seek */
24 static const int cfq_back_penalty = 2;
25 static const int cfq_slice_sync = HZ / 10;
26 static int cfq_slice_async = HZ / 25;
27 static const int cfq_slice_async_rq = 2;
28 static int cfq_slice_idle = HZ / 125;
29
30 /*
31  * offset from end of service tree
32  */
33 #define CFQ_IDLE_DELAY          (HZ / 5)
34
35 /*
36  * below this threshold, we consider thinktime immediate
37  */
38 #define CFQ_MIN_TT              (2)
39
40 #define CFQ_SLICE_SCALE         (5)
41
42 #define RQ_CIC(rq)              \
43         ((struct cfq_io_context *) (rq)->elevator_private)
44 #define RQ_CFQQ(rq)             ((rq)->elevator_private2)
45
46 static struct kmem_cache *cfq_pool;
47 static struct kmem_cache *cfq_ioc_pool;
48
49 static DEFINE_PER_CPU(unsigned long, ioc_count);
50 static struct completion *ioc_gone;
51
52 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
53 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
54 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
55
56 #define ASYNC                   (0)
57 #define SYNC                    (1)
58
59 #define sample_valid(samples)   ((samples) > 80)
60
61 /*
62  * Most of our rbtree usage is for sorting with min extraction, so
63  * if we cache the leftmost node we don't have to walk down the tree
64  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
65  * move this into the elevator for the rq sorting as well.
66  */
67 struct cfq_rb_root {
68         struct rb_root rb;
69         struct rb_node *left;
70 };
71 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, }
72
73 /*
74  * Per block device queue structure
75  */
76 struct cfq_data {
77         struct request_queue *queue;
78
79         /*
80          * rr list of queues with requests and the count of them
81          */
82         struct cfq_rb_root service_tree;
83         unsigned int busy_queues;
84
85         int rq_in_driver;
86         int sync_flight;
87         int hw_tag;
88
89         /*
90          * idle window management
91          */
92         struct timer_list idle_slice_timer;
93         struct work_struct unplug_work;
94
95         struct cfq_queue *active_queue;
96         struct cfq_io_context *active_cic;
97
98         /*
99          * async queue for each priority case
100          */
101         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
102         struct cfq_queue *async_idle_cfqq;
103
104         sector_t last_position;
105         unsigned long last_end_request;
106
107         /*
108          * tunables, see top of file
109          */
110         unsigned int cfq_quantum;
111         unsigned int cfq_fifo_expire[2];
112         unsigned int cfq_back_penalty;
113         unsigned int cfq_back_max;
114         unsigned int cfq_slice[2];
115         unsigned int cfq_slice_async_rq;
116         unsigned int cfq_slice_idle;
117
118         struct list_head cic_list;
119 };
120
121 /*
122  * Per process-grouping structure
123  */
124 struct cfq_queue {
125         /* reference count */
126         atomic_t ref;
127         /* various state flags, see below */
128         unsigned int flags;
129         /* parent cfq_data */
130         struct cfq_data *cfqd;
131         /* service_tree member */
132         struct rb_node rb_node;
133         /* service_tree key */
134         unsigned long rb_key;
135         /* sorted list of pending requests */
136         struct rb_root sort_list;
137         /* if fifo isn't expired, next request to serve */
138         struct request *next_rq;
139         /* requests queued in sort_list */
140         int queued[2];
141         /* currently allocated requests */
142         int allocated[2];
143         /* fifo list of requests in sort_list */
144         struct list_head fifo;
145
146         unsigned long slice_end;
147         long slice_resid;
148
149         /* pending metadata requests */
150         int meta_pending;
151         /* number of requests that are on the dispatch list or inside driver */
152         int dispatched;
153
154         /* io prio of this group */
155         unsigned short ioprio, org_ioprio;
156         unsigned short ioprio_class, org_ioprio_class;
157
158 };
159
160 enum cfqq_state_flags {
161         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
162         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
163         CFQ_CFQQ_FLAG_must_alloc,       /* must be allowed rq alloc */
164         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
165         CFQ_CFQQ_FLAG_must_dispatch,    /* must dispatch, even if expired */
166         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
167         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
168         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
169         CFQ_CFQQ_FLAG_queue_new,        /* queue never been serviced */
170         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
171         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
172 };
173
174 #define CFQ_CFQQ_FNS(name)                                              \
175 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
176 {                                                                       \
177         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
178 }                                                                       \
179 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
180 {                                                                       \
181         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
182 }                                                                       \
183 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
184 {                                                                       \
185         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
186 }
187
188 CFQ_CFQQ_FNS(on_rr);
189 CFQ_CFQQ_FNS(wait_request);
190 CFQ_CFQQ_FNS(must_alloc);
191 CFQ_CFQQ_FNS(must_alloc_slice);
192 CFQ_CFQQ_FNS(must_dispatch);
193 CFQ_CFQQ_FNS(fifo_expire);
194 CFQ_CFQQ_FNS(idle_window);
195 CFQ_CFQQ_FNS(prio_changed);
196 CFQ_CFQQ_FNS(queue_new);
197 CFQ_CFQQ_FNS(slice_new);
198 CFQ_CFQQ_FNS(sync);
199 #undef CFQ_CFQQ_FNS
200
201 static void cfq_dispatch_insert(struct request_queue *, struct request *);
202 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
203                                        struct io_context *, gfp_t);
204 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
205                                                 struct io_context *);
206
207 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
208                                             int is_sync)
209 {
210         return cic->cfqq[!!is_sync];
211 }
212
213 static inline void cic_set_cfqq(struct cfq_io_context *cic,
214                                 struct cfq_queue *cfqq, int is_sync)
215 {
216         cic->cfqq[!!is_sync] = cfqq;
217 }
218
219 /*
220  * We regard a request as SYNC, if it's either a read or has the SYNC bit
221  * set (in which case it could also be direct WRITE).
222  */
223 static inline int cfq_bio_sync(struct bio *bio)
224 {
225         if (bio_data_dir(bio) == READ || bio_sync(bio))
226                 return 1;
227
228         return 0;
229 }
230
231 /*
232  * scheduler run of queue, if there are requests pending and no one in the
233  * driver that will restart queueing
234  */
235 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
236 {
237         if (cfqd->busy_queues)
238                 kblockd_schedule_work(&cfqd->unplug_work);
239 }
240
241 static int cfq_queue_empty(struct request_queue *q)
242 {
243         struct cfq_data *cfqd = q->elevator->elevator_data;
244
245         return !cfqd->busy_queues;
246 }
247
248 /*
249  * Scale schedule slice based on io priority. Use the sync time slice only
250  * if a queue is marked sync and has sync io queued. A sync queue with async
251  * io only, should not get full sync slice length.
252  */
253 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
254                                  unsigned short prio)
255 {
256         const int base_slice = cfqd->cfq_slice[sync];
257
258         WARN_ON(prio >= IOPRIO_BE_NR);
259
260         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
261 }
262
263 static inline int
264 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
265 {
266         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
267 }
268
269 static inline void
270 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
271 {
272         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
273 }
274
275 /*
276  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
277  * isn't valid until the first request from the dispatch is activated
278  * and the slice time set.
279  */
280 static inline int cfq_slice_used(struct cfq_queue *cfqq)
281 {
282         if (cfq_cfqq_slice_new(cfqq))
283                 return 0;
284         if (time_before(jiffies, cfqq->slice_end))
285                 return 0;
286
287         return 1;
288 }
289
290 /*
291  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
292  * We choose the request that is closest to the head right now. Distance
293  * behind the head is penalized and only allowed to a certain extent.
294  */
295 static struct request *
296 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
297 {
298         sector_t last, s1, s2, d1 = 0, d2 = 0;
299         unsigned long back_max;
300 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
301 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
302         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
303
304         if (rq1 == NULL || rq1 == rq2)
305                 return rq2;
306         if (rq2 == NULL)
307                 return rq1;
308
309         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
310                 return rq1;
311         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
312                 return rq2;
313         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
314                 return rq1;
315         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
316                 return rq2;
317
318         s1 = rq1->sector;
319         s2 = rq2->sector;
320
321         last = cfqd->last_position;
322
323         /*
324          * by definition, 1KiB is 2 sectors
325          */
326         back_max = cfqd->cfq_back_max * 2;
327
328         /*
329          * Strict one way elevator _except_ in the case where we allow
330          * short backward seeks which are biased as twice the cost of a
331          * similar forward seek.
332          */
333         if (s1 >= last)
334                 d1 = s1 - last;
335         else if (s1 + back_max >= last)
336                 d1 = (last - s1) * cfqd->cfq_back_penalty;
337         else
338                 wrap |= CFQ_RQ1_WRAP;
339
340         if (s2 >= last)
341                 d2 = s2 - last;
342         else if (s2 + back_max >= last)
343                 d2 = (last - s2) * cfqd->cfq_back_penalty;
344         else
345                 wrap |= CFQ_RQ2_WRAP;
346
347         /* Found required data */
348
349         /*
350          * By doing switch() on the bit mask "wrap" we avoid having to
351          * check two variables for all permutations: --> faster!
352          */
353         switch (wrap) {
354         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
355                 if (d1 < d2)
356                         return rq1;
357                 else if (d2 < d1)
358                         return rq2;
359                 else {
360                         if (s1 >= s2)
361                                 return rq1;
362                         else
363                                 return rq2;
364                 }
365
366         case CFQ_RQ2_WRAP:
367                 return rq1;
368         case CFQ_RQ1_WRAP:
369                 return rq2;
370         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
371         default:
372                 /*
373                  * Since both rqs are wrapped,
374                  * start with the one that's further behind head
375                  * (--> only *one* back seek required),
376                  * since back seek takes more time than forward.
377                  */
378                 if (s1 <= s2)
379                         return rq1;
380                 else
381                         return rq2;
382         }
383 }
384
385 /*
386  * The below is leftmost cache rbtree addon
387  */
388 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
389 {
390         if (!root->left)
391                 root->left = rb_first(&root->rb);
392
393         if (root->left)
394                 return rb_entry(root->left, struct cfq_queue, rb_node);
395
396         return NULL;
397 }
398
399 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
400 {
401         if (root->left == n)
402                 root->left = NULL;
403
404         rb_erase(n, &root->rb);
405         RB_CLEAR_NODE(n);
406 }
407
408 /*
409  * would be nice to take fifo expire time into account as well
410  */
411 static struct request *
412 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
413                   struct request *last)
414 {
415         struct rb_node *rbnext = rb_next(&last->rb_node);
416         struct rb_node *rbprev = rb_prev(&last->rb_node);
417         struct request *next = NULL, *prev = NULL;
418
419         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
420
421         if (rbprev)
422                 prev = rb_entry_rq(rbprev);
423
424         if (rbnext)
425                 next = rb_entry_rq(rbnext);
426         else {
427                 rbnext = rb_first(&cfqq->sort_list);
428                 if (rbnext && rbnext != &last->rb_node)
429                         next = rb_entry_rq(rbnext);
430         }
431
432         return cfq_choose_req(cfqd, next, prev);
433 }
434
435 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
436                                       struct cfq_queue *cfqq)
437 {
438         /*
439          * just an approximation, should be ok.
440          */
441         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
442                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
443 }
444
445 /*
446  * The cfqd->service_tree holds all pending cfq_queue's that have
447  * requests waiting to be processed. It is sorted in the order that
448  * we will service the queues.
449  */
450 static void cfq_service_tree_add(struct cfq_data *cfqd,
451                                     struct cfq_queue *cfqq, int add_front)
452 {
453         struct rb_node **p, *parent;
454         struct cfq_queue *__cfqq;
455         unsigned long rb_key;
456         int left;
457
458         if (cfq_class_idle(cfqq)) {
459                 rb_key = CFQ_IDLE_DELAY;
460                 parent = rb_last(&cfqd->service_tree.rb);
461                 if (parent && parent != &cfqq->rb_node) {
462                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
463                         rb_key += __cfqq->rb_key;
464                 } else
465                         rb_key += jiffies;
466         } else if (!add_front) {
467                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
468                 rb_key += cfqq->slice_resid;
469                 cfqq->slice_resid = 0;
470         } else
471                 rb_key = 0;
472
473         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
474                 /*
475                  * same position, nothing more to do
476                  */
477                 if (rb_key == cfqq->rb_key)
478                         return;
479
480                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
481         }
482
483         left = 1;
484         parent = NULL;
485         p = &cfqd->service_tree.rb.rb_node;
486         while (*p) {
487                 struct rb_node **n;
488
489                 parent = *p;
490                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
491
492                 /*
493                  * sort RT queues first, we always want to give
494                  * preference to them. IDLE queues goes to the back.
495                  * after that, sort on the next service time.
496                  */
497                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
498                         n = &(*p)->rb_left;
499                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
500                         n = &(*p)->rb_right;
501                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
502                         n = &(*p)->rb_left;
503                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
504                         n = &(*p)->rb_right;
505                 else if (rb_key < __cfqq->rb_key)
506                         n = &(*p)->rb_left;
507                 else
508                         n = &(*p)->rb_right;
509
510                 if (n == &(*p)->rb_right)
511                         left = 0;
512
513                 p = n;
514         }
515
516         if (left)
517                 cfqd->service_tree.left = &cfqq->rb_node;
518
519         cfqq->rb_key = rb_key;
520         rb_link_node(&cfqq->rb_node, parent, p);
521         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
522 }
523
524 /*
525  * Update cfqq's position in the service tree.
526  */
527 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
528 {
529         /*
530          * Resorting requires the cfqq to be on the RR list already.
531          */
532         if (cfq_cfqq_on_rr(cfqq))
533                 cfq_service_tree_add(cfqd, cfqq, 0);
534 }
535
536 /*
537  * add to busy list of queues for service, trying to be fair in ordering
538  * the pending list according to last request service
539  */
540 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
541 {
542         BUG_ON(cfq_cfqq_on_rr(cfqq));
543         cfq_mark_cfqq_on_rr(cfqq);
544         cfqd->busy_queues++;
545
546         cfq_resort_rr_list(cfqd, cfqq);
547 }
548
549 /*
550  * Called when the cfqq no longer has requests pending, remove it from
551  * the service tree.
552  */
553 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
554 {
555         BUG_ON(!cfq_cfqq_on_rr(cfqq));
556         cfq_clear_cfqq_on_rr(cfqq);
557
558         if (!RB_EMPTY_NODE(&cfqq->rb_node))
559                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
560
561         BUG_ON(!cfqd->busy_queues);
562         cfqd->busy_queues--;
563 }
564
565 /*
566  * rb tree support functions
567  */
568 static void cfq_del_rq_rb(struct request *rq)
569 {
570         struct cfq_queue *cfqq = RQ_CFQQ(rq);
571         struct cfq_data *cfqd = cfqq->cfqd;
572         const int sync = rq_is_sync(rq);
573
574         BUG_ON(!cfqq->queued[sync]);
575         cfqq->queued[sync]--;
576
577         elv_rb_del(&cfqq->sort_list, rq);
578
579         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
580                 cfq_del_cfqq_rr(cfqd, cfqq);
581 }
582
583 static void cfq_add_rq_rb(struct request *rq)
584 {
585         struct cfq_queue *cfqq = RQ_CFQQ(rq);
586         struct cfq_data *cfqd = cfqq->cfqd;
587         struct request *__alias;
588
589         cfqq->queued[rq_is_sync(rq)]++;
590
591         /*
592          * looks a little odd, but the first insert might return an alias.
593          * if that happens, put the alias on the dispatch list
594          */
595         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
596                 cfq_dispatch_insert(cfqd->queue, __alias);
597
598         if (!cfq_cfqq_on_rr(cfqq))
599                 cfq_add_cfqq_rr(cfqd, cfqq);
600
601         /*
602          * check if this request is a better next-serve candidate
603          */
604         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
605         BUG_ON(!cfqq->next_rq);
606 }
607
608 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
609 {
610         elv_rb_del(&cfqq->sort_list, rq);
611         cfqq->queued[rq_is_sync(rq)]--;
612         cfq_add_rq_rb(rq);
613 }
614
615 static struct request *
616 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
617 {
618         struct task_struct *tsk = current;
619         struct cfq_io_context *cic;
620         struct cfq_queue *cfqq;
621
622         cic = cfq_cic_lookup(cfqd, tsk->io_context);
623         if (!cic)
624                 return NULL;
625
626         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
627         if (cfqq) {
628                 sector_t sector = bio->bi_sector + bio_sectors(bio);
629
630                 return elv_rb_find(&cfqq->sort_list, sector);
631         }
632
633         return NULL;
634 }
635
636 static void cfq_activate_request(struct request_queue *q, struct request *rq)
637 {
638         struct cfq_data *cfqd = q->elevator->elevator_data;
639
640         cfqd->rq_in_driver++;
641
642         /*
643          * If the depth is larger 1, it really could be queueing. But lets
644          * make the mark a little higher - idling could still be good for
645          * low queueing, and a low queueing number could also just indicate
646          * a SCSI mid layer like behaviour where limit+1 is often seen.
647          */
648         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
649                 cfqd->hw_tag = 1;
650
651         cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
652 }
653
654 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
655 {
656         struct cfq_data *cfqd = q->elevator->elevator_data;
657
658         WARN_ON(!cfqd->rq_in_driver);
659         cfqd->rq_in_driver--;
660 }
661
662 static void cfq_remove_request(struct request *rq)
663 {
664         struct cfq_queue *cfqq = RQ_CFQQ(rq);
665
666         if (cfqq->next_rq == rq)
667                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
668
669         list_del_init(&rq->queuelist);
670         cfq_del_rq_rb(rq);
671
672         if (rq_is_meta(rq)) {
673                 WARN_ON(!cfqq->meta_pending);
674                 cfqq->meta_pending--;
675         }
676 }
677
678 static int cfq_merge(struct request_queue *q, struct request **req,
679                      struct bio *bio)
680 {
681         struct cfq_data *cfqd = q->elevator->elevator_data;
682         struct request *__rq;
683
684         __rq = cfq_find_rq_fmerge(cfqd, bio);
685         if (__rq && elv_rq_merge_ok(__rq, bio)) {
686                 *req = __rq;
687                 return ELEVATOR_FRONT_MERGE;
688         }
689
690         return ELEVATOR_NO_MERGE;
691 }
692
693 static void cfq_merged_request(struct request_queue *q, struct request *req,
694                                int type)
695 {
696         if (type == ELEVATOR_FRONT_MERGE) {
697                 struct cfq_queue *cfqq = RQ_CFQQ(req);
698
699                 cfq_reposition_rq_rb(cfqq, req);
700         }
701 }
702
703 static void
704 cfq_merged_requests(struct request_queue *q, struct request *rq,
705                     struct request *next)
706 {
707         /*
708          * reposition in fifo if next is older than rq
709          */
710         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
711             time_before(next->start_time, rq->start_time))
712                 list_move(&rq->queuelist, &next->queuelist);
713
714         cfq_remove_request(next);
715 }
716
717 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
718                            struct bio *bio)
719 {
720         struct cfq_data *cfqd = q->elevator->elevator_data;
721         struct cfq_io_context *cic;
722         struct cfq_queue *cfqq;
723
724         /*
725          * Disallow merge of a sync bio into an async request.
726          */
727         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
728                 return 0;
729
730         /*
731          * Lookup the cfqq that this bio will be queued with. Allow
732          * merge only if rq is queued there.
733          */
734         cic = cfq_cic_lookup(cfqd, current->io_context);
735         if (!cic)
736                 return 0;
737
738         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
739         if (cfqq == RQ_CFQQ(rq))
740                 return 1;
741
742         return 0;
743 }
744
745 static void __cfq_set_active_queue(struct cfq_data *cfqd,
746                                    struct cfq_queue *cfqq)
747 {
748         if (cfqq) {
749                 cfqq->slice_end = 0;
750                 cfq_clear_cfqq_must_alloc_slice(cfqq);
751                 cfq_clear_cfqq_fifo_expire(cfqq);
752                 cfq_mark_cfqq_slice_new(cfqq);
753                 cfq_clear_cfqq_queue_new(cfqq);
754         }
755
756         cfqd->active_queue = cfqq;
757 }
758
759 /*
760  * current cfqq expired its slice (or was too idle), select new one
761  */
762 static void
763 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
764                     int timed_out)
765 {
766         if (cfq_cfqq_wait_request(cfqq))
767                 del_timer(&cfqd->idle_slice_timer);
768
769         cfq_clear_cfqq_must_dispatch(cfqq);
770         cfq_clear_cfqq_wait_request(cfqq);
771
772         /*
773          * store what was left of this slice, if the queue idled/timed out
774          */
775         if (timed_out && !cfq_cfqq_slice_new(cfqq))
776                 cfqq->slice_resid = cfqq->slice_end - jiffies;
777
778         cfq_resort_rr_list(cfqd, cfqq);
779
780         if (cfqq == cfqd->active_queue)
781                 cfqd->active_queue = NULL;
782
783         if (cfqd->active_cic) {
784                 put_io_context(cfqd->active_cic->ioc);
785                 cfqd->active_cic = NULL;
786         }
787 }
788
789 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
790 {
791         struct cfq_queue *cfqq = cfqd->active_queue;
792
793         if (cfqq)
794                 __cfq_slice_expired(cfqd, cfqq, timed_out);
795 }
796
797 /*
798  * Get next queue for service. Unless we have a queue preemption,
799  * we'll simply select the first cfqq in the service tree.
800  */
801 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
802 {
803         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
804                 return NULL;
805
806         return cfq_rb_first(&cfqd->service_tree);
807 }
808
809 /*
810  * Get and set a new active queue for service.
811  */
812 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
813 {
814         struct cfq_queue *cfqq;
815
816         cfqq = cfq_get_next_queue(cfqd);
817         __cfq_set_active_queue(cfqd, cfqq);
818         return cfqq;
819 }
820
821 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
822                                           struct request *rq)
823 {
824         if (rq->sector >= cfqd->last_position)
825                 return rq->sector - cfqd->last_position;
826         else
827                 return cfqd->last_position - rq->sector;
828 }
829
830 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
831 {
832         struct cfq_io_context *cic = cfqd->active_cic;
833
834         if (!sample_valid(cic->seek_samples))
835                 return 0;
836
837         return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
838 }
839
840 static int cfq_close_cooperator(struct cfq_data *cfq_data,
841                                 struct cfq_queue *cfqq)
842 {
843         /*
844          * We should notice if some of the queues are cooperating, eg
845          * working closely on the same area of the disk. In that case,
846          * we can group them together and don't waste time idling.
847          */
848         return 0;
849 }
850
851 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
852
853 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
854 {
855         struct cfq_queue *cfqq = cfqd->active_queue;
856         struct cfq_io_context *cic;
857         unsigned long sl;
858
859         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
860         WARN_ON(cfq_cfqq_slice_new(cfqq));
861
862         /*
863          * idle is disabled, either manually or by past process history
864          */
865         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
866                 return;
867
868         /*
869          * task has exited, don't wait
870          */
871         cic = cfqd->active_cic;
872         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
873                 return;
874
875         /*
876          * See if this prio level has a good candidate
877          */
878         if (cfq_close_cooperator(cfqd, cfqq) &&
879             (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
880                 return;
881
882         cfq_mark_cfqq_must_dispatch(cfqq);
883         cfq_mark_cfqq_wait_request(cfqq);
884
885         /*
886          * we don't want to idle for seeks, but we do want to allow
887          * fair distribution of slice time for a process doing back-to-back
888          * seeks. so allow a little bit of time for him to submit a new rq
889          */
890         sl = cfqd->cfq_slice_idle;
891         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
892                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
893
894         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
895 }
896
897 /*
898  * Move request from internal lists to the request queue dispatch list.
899  */
900 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
901 {
902         struct cfq_data *cfqd = q->elevator->elevator_data;
903         struct cfq_queue *cfqq = RQ_CFQQ(rq);
904
905         cfq_remove_request(rq);
906         cfqq->dispatched++;
907         elv_dispatch_sort(q, rq);
908
909         if (cfq_cfqq_sync(cfqq))
910                 cfqd->sync_flight++;
911 }
912
913 /*
914  * return expired entry, or NULL to just start from scratch in rbtree
915  */
916 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
917 {
918         struct cfq_data *cfqd = cfqq->cfqd;
919         struct request *rq;
920         int fifo;
921
922         if (cfq_cfqq_fifo_expire(cfqq))
923                 return NULL;
924
925         cfq_mark_cfqq_fifo_expire(cfqq);
926
927         if (list_empty(&cfqq->fifo))
928                 return NULL;
929
930         fifo = cfq_cfqq_sync(cfqq);
931         rq = rq_entry_fifo(cfqq->fifo.next);
932
933         if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
934                 return NULL;
935
936         return rq;
937 }
938
939 static inline int
940 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
941 {
942         const int base_rq = cfqd->cfq_slice_async_rq;
943
944         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
945
946         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
947 }
948
949 /*
950  * Select a queue for service. If we have a current active queue,
951  * check whether to continue servicing it, or retrieve and set a new one.
952  */
953 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
954 {
955         struct cfq_queue *cfqq;
956
957         cfqq = cfqd->active_queue;
958         if (!cfqq)
959                 goto new_queue;
960
961         /*
962          * The active queue has run out of time, expire it and select new.
963          */
964         if (cfq_slice_used(cfqq))
965                 goto expire;
966
967         /*
968          * The active queue has requests and isn't expired, allow it to
969          * dispatch.
970          */
971         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
972                 goto keep_queue;
973
974         /*
975          * No requests pending. If the active queue still has requests in
976          * flight or is idling for a new request, allow either of these
977          * conditions to happen (or time out) before selecting a new queue.
978          */
979         if (timer_pending(&cfqd->idle_slice_timer) ||
980             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
981                 cfqq = NULL;
982                 goto keep_queue;
983         }
984
985 expire:
986         cfq_slice_expired(cfqd, 0);
987 new_queue:
988         cfqq = cfq_set_active_queue(cfqd);
989 keep_queue:
990         return cfqq;
991 }
992
993 /*
994  * Dispatch some requests from cfqq, moving them to the request queue
995  * dispatch list.
996  */
997 static int
998 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
999                         int max_dispatch)
1000 {
1001         int dispatched = 0;
1002
1003         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1004
1005         do {
1006                 struct request *rq;
1007
1008                 /*
1009                  * follow expired path, else get first next available
1010                  */
1011                 rq = cfq_check_fifo(cfqq);
1012                 if (rq == NULL)
1013                         rq = cfqq->next_rq;
1014
1015                 /*
1016                  * finally, insert request into driver dispatch list
1017                  */
1018                 cfq_dispatch_insert(cfqd->queue, rq);
1019
1020                 dispatched++;
1021
1022                 if (!cfqd->active_cic) {
1023                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1024                         cfqd->active_cic = RQ_CIC(rq);
1025                 }
1026
1027                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
1028                         break;
1029
1030         } while (dispatched < max_dispatch);
1031
1032         /*
1033          * expire an async queue immediately if it has used up its slice. idle
1034          * queue always expire after 1 dispatch round.
1035          */
1036         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1037             dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1038             cfq_class_idle(cfqq))) {
1039                 cfqq->slice_end = jiffies + 1;
1040                 cfq_slice_expired(cfqd, 0);
1041         }
1042
1043         return dispatched;
1044 }
1045
1046 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1047 {
1048         int dispatched = 0;
1049
1050         while (cfqq->next_rq) {
1051                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1052                 dispatched++;
1053         }
1054
1055         BUG_ON(!list_empty(&cfqq->fifo));
1056         return dispatched;
1057 }
1058
1059 /*
1060  * Drain our current requests. Used for barriers and when switching
1061  * io schedulers on-the-fly.
1062  */
1063 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1064 {
1065         struct cfq_queue *cfqq;
1066         int dispatched = 0;
1067
1068         while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1069                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1070
1071         cfq_slice_expired(cfqd, 0);
1072
1073         BUG_ON(cfqd->busy_queues);
1074
1075         return dispatched;
1076 }
1077
1078 static int cfq_dispatch_requests(struct request_queue *q, int force)
1079 {
1080         struct cfq_data *cfqd = q->elevator->elevator_data;
1081         struct cfq_queue *cfqq;
1082         int dispatched;
1083
1084         if (!cfqd->busy_queues)
1085                 return 0;
1086
1087         if (unlikely(force))
1088                 return cfq_forced_dispatch(cfqd);
1089
1090         dispatched = 0;
1091         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1092                 int max_dispatch;
1093
1094                 max_dispatch = cfqd->cfq_quantum;
1095                 if (cfq_class_idle(cfqq))
1096                         max_dispatch = 1;
1097
1098                 if (cfqq->dispatched >= max_dispatch) {
1099                         if (cfqd->busy_queues > 1)
1100                                 break;
1101                         if (cfqq->dispatched >= 4 * max_dispatch)
1102                                 break;
1103                 }
1104
1105                 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1106                         break;
1107
1108                 cfq_clear_cfqq_must_dispatch(cfqq);
1109                 cfq_clear_cfqq_wait_request(cfqq);
1110                 del_timer(&cfqd->idle_slice_timer);
1111
1112                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1113         }
1114
1115         return dispatched;
1116 }
1117
1118 /*
1119  * task holds one reference to the queue, dropped when task exits. each rq
1120  * in-flight on this queue also holds a reference, dropped when rq is freed.
1121  *
1122  * queue lock must be held here.
1123  */
1124 static void cfq_put_queue(struct cfq_queue *cfqq)
1125 {
1126         struct cfq_data *cfqd = cfqq->cfqd;
1127
1128         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1129
1130         if (!atomic_dec_and_test(&cfqq->ref))
1131                 return;
1132
1133         BUG_ON(rb_first(&cfqq->sort_list));
1134         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1135         BUG_ON(cfq_cfqq_on_rr(cfqq));
1136
1137         if (unlikely(cfqd->active_queue == cfqq)) {
1138                 __cfq_slice_expired(cfqd, cfqq, 0);
1139                 cfq_schedule_dispatch(cfqd);
1140         }
1141
1142         kmem_cache_free(cfq_pool, cfqq);
1143 }
1144
1145 /*
1146  * Must always be called with the rcu_read_lock() held
1147  */
1148 static void
1149 __call_for_each_cic(struct io_context *ioc,
1150                     void (*func)(struct io_context *, struct cfq_io_context *))
1151 {
1152         struct cfq_io_context *cic;
1153         struct hlist_node *n;
1154
1155         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1156                 func(ioc, cic);
1157 }
1158
1159 /*
1160  * Call func for each cic attached to this ioc.
1161  */
1162 static void
1163 call_for_each_cic(struct io_context *ioc,
1164                   void (*func)(struct io_context *, struct cfq_io_context *))
1165 {
1166         rcu_read_lock();
1167         __call_for_each_cic(ioc, func);
1168         rcu_read_unlock();
1169 }
1170
1171 static void cfq_cic_free_rcu(struct rcu_head *head)
1172 {
1173         struct cfq_io_context *cic;
1174
1175         cic = container_of(head, struct cfq_io_context, rcu_head);
1176
1177         kmem_cache_free(cfq_ioc_pool, cic);
1178         elv_ioc_count_dec(ioc_count);
1179
1180         if (ioc_gone && !elv_ioc_count_read(ioc_count))
1181                 complete(ioc_gone);
1182 }
1183
1184 static void cfq_cic_free(struct cfq_io_context *cic)
1185 {
1186         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1187 }
1188
1189 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1190 {
1191         unsigned long flags;
1192
1193         BUG_ON(!cic->dead_key);
1194
1195         spin_lock_irqsave(&ioc->lock, flags);
1196         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1197         hlist_del_rcu(&cic->cic_list);
1198         spin_unlock_irqrestore(&ioc->lock, flags);
1199
1200         cfq_cic_free(cic);
1201 }
1202
1203 /*
1204  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1205  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1206  * and ->trim() which is called with the task lock held
1207  */
1208 static void cfq_free_io_context(struct io_context *ioc)
1209 {
1210         /*
1211          * ioc->refcount is zero here, or we are called from elv_unregister(),
1212          * so no more cic's are allowed to be linked into this ioc.  So it
1213          * should be ok to iterate over the known list, we will see all cic's
1214          * since no new ones are added.
1215          */
1216         __call_for_each_cic(ioc, cic_free_func);
1217 }
1218
1219 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1220 {
1221         if (unlikely(cfqq == cfqd->active_queue)) {
1222                 __cfq_slice_expired(cfqd, cfqq, 0);
1223                 cfq_schedule_dispatch(cfqd);
1224         }
1225
1226         cfq_put_queue(cfqq);
1227 }
1228
1229 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1230                                          struct cfq_io_context *cic)
1231 {
1232         struct io_context *ioc = cic->ioc;
1233
1234         list_del_init(&cic->queue_list);
1235
1236         /*
1237          * Make sure key == NULL is seen for dead queues
1238          */
1239         smp_wmb();
1240         cic->dead_key = (unsigned long) cic->key;
1241         cic->key = NULL;
1242
1243         if (ioc->ioc_data == cic)
1244                 rcu_assign_pointer(ioc->ioc_data, NULL);
1245
1246         if (cic->cfqq[ASYNC]) {
1247                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1248                 cic->cfqq[ASYNC] = NULL;
1249         }
1250
1251         if (cic->cfqq[SYNC]) {
1252                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1253                 cic->cfqq[SYNC] = NULL;
1254         }
1255 }
1256
1257 static void cfq_exit_single_io_context(struct io_context *ioc,
1258                                        struct cfq_io_context *cic)
1259 {
1260         struct cfq_data *cfqd = cic->key;
1261
1262         if (cfqd) {
1263                 struct request_queue *q = cfqd->queue;
1264                 unsigned long flags;
1265
1266                 spin_lock_irqsave(q->queue_lock, flags);
1267                 __cfq_exit_single_io_context(cfqd, cic);
1268                 spin_unlock_irqrestore(q->queue_lock, flags);
1269         }
1270 }
1271
1272 /*
1273  * The process that ioc belongs to has exited, we need to clean up
1274  * and put the internal structures we have that belongs to that process.
1275  */
1276 static void cfq_exit_io_context(struct io_context *ioc)
1277 {
1278         call_for_each_cic(ioc, cfq_exit_single_io_context);
1279 }
1280
1281 static struct cfq_io_context *
1282 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1283 {
1284         struct cfq_io_context *cic;
1285
1286         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1287                                                         cfqd->queue->node);
1288         if (cic) {
1289                 cic->last_end_request = jiffies;
1290                 INIT_LIST_HEAD(&cic->queue_list);
1291                 INIT_HLIST_NODE(&cic->cic_list);
1292                 cic->dtor = cfq_free_io_context;
1293                 cic->exit = cfq_exit_io_context;
1294                 elv_ioc_count_inc(ioc_count);
1295         }
1296
1297         return cic;
1298 }
1299
1300 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1301 {
1302         struct task_struct *tsk = current;
1303         int ioprio_class;
1304
1305         if (!cfq_cfqq_prio_changed(cfqq))
1306                 return;
1307
1308         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1309         switch (ioprio_class) {
1310         default:
1311                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1312         case IOPRIO_CLASS_NONE:
1313                 /*
1314                  * no prio set, inherit CPU scheduling settings
1315                  */
1316                 cfqq->ioprio = task_nice_ioprio(tsk);
1317                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1318                 break;
1319         case IOPRIO_CLASS_RT:
1320                 cfqq->ioprio = task_ioprio(ioc);
1321                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1322                 break;
1323         case IOPRIO_CLASS_BE:
1324                 cfqq->ioprio = task_ioprio(ioc);
1325                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1326                 break;
1327         case IOPRIO_CLASS_IDLE:
1328                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1329                 cfqq->ioprio = 7;
1330                 cfq_clear_cfqq_idle_window(cfqq);
1331                 break;
1332         }
1333
1334         /*
1335          * keep track of original prio settings in case we have to temporarily
1336          * elevate the priority of this queue
1337          */
1338         cfqq->org_ioprio = cfqq->ioprio;
1339         cfqq->org_ioprio_class = cfqq->ioprio_class;
1340         cfq_clear_cfqq_prio_changed(cfqq);
1341 }
1342
1343 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1344 {
1345         struct cfq_data *cfqd = cic->key;
1346         struct cfq_queue *cfqq;
1347         unsigned long flags;
1348
1349         if (unlikely(!cfqd))
1350                 return;
1351
1352         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1353
1354         cfqq = cic->cfqq[ASYNC];
1355         if (cfqq) {
1356                 struct cfq_queue *new_cfqq;
1357                 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC);
1358                 if (new_cfqq) {
1359                         cic->cfqq[ASYNC] = new_cfqq;
1360                         cfq_put_queue(cfqq);
1361                 }
1362         }
1363
1364         cfqq = cic->cfqq[SYNC];
1365         if (cfqq)
1366                 cfq_mark_cfqq_prio_changed(cfqq);
1367
1368         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1369 }
1370
1371 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1372 {
1373         call_for_each_cic(ioc, changed_ioprio);
1374         ioc->ioprio_changed = 0;
1375 }
1376
1377 static struct cfq_queue *
1378 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1379                      struct io_context *ioc, gfp_t gfp_mask)
1380 {
1381         struct cfq_queue *cfqq, *new_cfqq = NULL;
1382         struct cfq_io_context *cic;
1383
1384 retry:
1385         cic = cfq_cic_lookup(cfqd, ioc);
1386         /* cic always exists here */
1387         cfqq = cic_to_cfqq(cic, is_sync);
1388
1389         if (!cfqq) {
1390                 if (new_cfqq) {
1391                         cfqq = new_cfqq;
1392                         new_cfqq = NULL;
1393                 } else if (gfp_mask & __GFP_WAIT) {
1394                         /*
1395                          * Inform the allocator of the fact that we will
1396                          * just repeat this allocation if it fails, to allow
1397                          * the allocator to do whatever it needs to attempt to
1398                          * free memory.
1399                          */
1400                         spin_unlock_irq(cfqd->queue->queue_lock);
1401                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1402                                         gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1403                                         cfqd->queue->node);
1404                         spin_lock_irq(cfqd->queue->queue_lock);
1405                         goto retry;
1406                 } else {
1407                         cfqq = kmem_cache_alloc_node(cfq_pool,
1408                                         gfp_mask | __GFP_ZERO,
1409                                         cfqd->queue->node);
1410                         if (!cfqq)
1411                                 goto out;
1412                 }
1413
1414                 RB_CLEAR_NODE(&cfqq->rb_node);
1415                 INIT_LIST_HEAD(&cfqq->fifo);
1416
1417                 atomic_set(&cfqq->ref, 0);
1418                 cfqq->cfqd = cfqd;
1419
1420                 cfq_mark_cfqq_prio_changed(cfqq);
1421                 cfq_mark_cfqq_queue_new(cfqq);
1422
1423                 cfq_init_prio_data(cfqq, ioc);
1424
1425                 if (is_sync) {
1426                         if (!cfq_class_idle(cfqq))
1427                                 cfq_mark_cfqq_idle_window(cfqq);
1428                         cfq_mark_cfqq_sync(cfqq);
1429                 }
1430         }
1431
1432         if (new_cfqq)
1433                 kmem_cache_free(cfq_pool, new_cfqq);
1434
1435 out:
1436         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1437         return cfqq;
1438 }
1439
1440 static struct cfq_queue **
1441 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1442 {
1443         switch (ioprio_class) {
1444         case IOPRIO_CLASS_RT:
1445                 return &cfqd->async_cfqq[0][ioprio];
1446         case IOPRIO_CLASS_BE:
1447                 return &cfqd->async_cfqq[1][ioprio];
1448         case IOPRIO_CLASS_IDLE:
1449                 return &cfqd->async_idle_cfqq;
1450         default:
1451                 BUG();
1452         }
1453 }
1454
1455 static struct cfq_queue *
1456 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1457               gfp_t gfp_mask)
1458 {
1459         const int ioprio = task_ioprio(ioc);
1460         const int ioprio_class = task_ioprio_class(ioc);
1461         struct cfq_queue **async_cfqq = NULL;
1462         struct cfq_queue *cfqq = NULL;
1463
1464         if (!is_sync) {
1465                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1466                 cfqq = *async_cfqq;
1467         }
1468
1469         if (!cfqq) {
1470                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1471                 if (!cfqq)
1472                         return NULL;
1473         }
1474
1475         /*
1476          * pin the queue now that it's allocated, scheduler exit will prune it
1477          */
1478         if (!is_sync && !(*async_cfqq)) {
1479                 atomic_inc(&cfqq->ref);
1480                 *async_cfqq = cfqq;
1481         }
1482
1483         atomic_inc(&cfqq->ref);
1484         return cfqq;
1485 }
1486
1487 /*
1488  * We drop cfq io contexts lazily, so we may find a dead one.
1489  */
1490 static void
1491 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1492                   struct cfq_io_context *cic)
1493 {
1494         unsigned long flags;
1495
1496         WARN_ON(!list_empty(&cic->queue_list));
1497
1498         spin_lock_irqsave(&ioc->lock, flags);
1499
1500         BUG_ON(ioc->ioc_data == cic);
1501
1502         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1503         hlist_del_rcu(&cic->cic_list);
1504         spin_unlock_irqrestore(&ioc->lock, flags);
1505
1506         cfq_cic_free(cic);
1507 }
1508
1509 static struct cfq_io_context *
1510 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1511 {
1512         struct cfq_io_context *cic;
1513         unsigned long flags;
1514         void *k;
1515
1516         if (unlikely(!ioc))
1517                 return NULL;
1518
1519         rcu_read_lock();
1520
1521         /*
1522          * we maintain a last-hit cache, to avoid browsing over the tree
1523          */
1524         cic = rcu_dereference(ioc->ioc_data);
1525         if (cic && cic->key == cfqd) {
1526                 rcu_read_unlock();
1527                 return cic;
1528         }
1529
1530         do {
1531                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1532                 rcu_read_unlock();
1533                 if (!cic)
1534                         break;
1535                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1536                 k = cic->key;
1537                 if (unlikely(!k)) {
1538                         cfq_drop_dead_cic(cfqd, ioc, cic);
1539                         rcu_read_lock();
1540                         continue;
1541                 }
1542
1543                 spin_lock_irqsave(&ioc->lock, flags);
1544                 rcu_assign_pointer(ioc->ioc_data, cic);
1545                 spin_unlock_irqrestore(&ioc->lock, flags);
1546                 break;
1547         } while (1);
1548
1549         return cic;
1550 }
1551
1552 /*
1553  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1554  * the process specific cfq io context when entered from the block layer.
1555  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1556  */
1557 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1558                         struct cfq_io_context *cic, gfp_t gfp_mask)
1559 {
1560         unsigned long flags;
1561         int ret;
1562
1563         ret = radix_tree_preload(gfp_mask);
1564         if (!ret) {
1565                 cic->ioc = ioc;
1566                 cic->key = cfqd;
1567
1568                 spin_lock_irqsave(&ioc->lock, flags);
1569                 ret = radix_tree_insert(&ioc->radix_root,
1570                                                 (unsigned long) cfqd, cic);
1571                 if (!ret)
1572                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1573                 spin_unlock_irqrestore(&ioc->lock, flags);
1574
1575                 radix_tree_preload_end();
1576
1577                 if (!ret) {
1578                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1579                         list_add(&cic->queue_list, &cfqd->cic_list);
1580                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1581                 }
1582         }
1583
1584         if (ret)
1585                 printk(KERN_ERR "cfq: cic link failed!\n");
1586
1587         return ret;
1588 }
1589
1590 /*
1591  * Setup general io context and cfq io context. There can be several cfq
1592  * io contexts per general io context, if this process is doing io to more
1593  * than one device managed by cfq.
1594  */
1595 static struct cfq_io_context *
1596 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1597 {
1598         struct io_context *ioc = NULL;
1599         struct cfq_io_context *cic;
1600
1601         might_sleep_if(gfp_mask & __GFP_WAIT);
1602
1603         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1604         if (!ioc)
1605                 return NULL;
1606
1607         cic = cfq_cic_lookup(cfqd, ioc);
1608         if (cic)
1609                 goto out;
1610
1611         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1612         if (cic == NULL)
1613                 goto err;
1614
1615         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1616                 goto err_free;
1617
1618 out:
1619         smp_read_barrier_depends();
1620         if (unlikely(ioc->ioprio_changed))
1621                 cfq_ioc_set_ioprio(ioc);
1622
1623         return cic;
1624 err_free:
1625         cfq_cic_free(cic);
1626 err:
1627         put_io_context(ioc);
1628         return NULL;
1629 }
1630
1631 static void
1632 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1633 {
1634         unsigned long elapsed = jiffies - cic->last_end_request;
1635         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1636
1637         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1638         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1639         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1640 }
1641
1642 static void
1643 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1644                        struct request *rq)
1645 {
1646         sector_t sdist;
1647         u64 total;
1648
1649         if (cic->last_request_pos < rq->sector)
1650                 sdist = rq->sector - cic->last_request_pos;
1651         else
1652                 sdist = cic->last_request_pos - rq->sector;
1653
1654         /*
1655          * Don't allow the seek distance to get too large from the
1656          * odd fragment, pagein, etc
1657          */
1658         if (cic->seek_samples <= 60) /* second&third seek */
1659                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1660         else
1661                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1662
1663         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1664         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1665         total = cic->seek_total + (cic->seek_samples/2);
1666         do_div(total, cic->seek_samples);
1667         cic->seek_mean = (sector_t)total;
1668 }
1669
1670 /*
1671  * Disable idle window if the process thinks too long or seeks so much that
1672  * it doesn't matter
1673  */
1674 static void
1675 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1676                        struct cfq_io_context *cic)
1677 {
1678         int enable_idle;
1679
1680         /*
1681          * Don't idle for async or idle io prio class
1682          */
1683         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1684                 return;
1685
1686         enable_idle = cfq_cfqq_idle_window(cfqq);
1687
1688         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1689             (cfqd->hw_tag && CIC_SEEKY(cic)))
1690                 enable_idle = 0;
1691         else if (sample_valid(cic->ttime_samples)) {
1692                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1693                         enable_idle = 0;
1694                 else
1695                         enable_idle = 1;
1696         }
1697
1698         if (enable_idle)
1699                 cfq_mark_cfqq_idle_window(cfqq);
1700         else
1701                 cfq_clear_cfqq_idle_window(cfqq);
1702 }
1703
1704 /*
1705  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1706  * no or if we aren't sure, a 1 will cause a preempt.
1707  */
1708 static int
1709 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1710                    struct request *rq)
1711 {
1712         struct cfq_queue *cfqq;
1713
1714         cfqq = cfqd->active_queue;
1715         if (!cfqq)
1716                 return 0;
1717
1718         if (cfq_slice_used(cfqq))
1719                 return 1;
1720
1721         if (cfq_class_idle(new_cfqq))
1722                 return 0;
1723
1724         if (cfq_class_idle(cfqq))
1725                 return 1;
1726
1727         /*
1728          * if the new request is sync, but the currently running queue is
1729          * not, let the sync request have priority.
1730          */
1731         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1732                 return 1;
1733
1734         /*
1735          * So both queues are sync. Let the new request get disk time if
1736          * it's a metadata request and the current queue is doing regular IO.
1737          */
1738         if (rq_is_meta(rq) && !cfqq->meta_pending)
1739                 return 1;
1740
1741         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1742                 return 0;
1743
1744         /*
1745          * if this request is as-good as one we would expect from the
1746          * current cfqq, let it preempt
1747          */
1748         if (cfq_rq_close(cfqd, rq))
1749                 return 1;
1750
1751         return 0;
1752 }
1753
1754 /*
1755  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1756  * let it have half of its nominal slice.
1757  */
1758 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1759 {
1760         cfq_slice_expired(cfqd, 1);
1761
1762         /*
1763          * Put the new queue at the front of the of the current list,
1764          * so we know that it will be selected next.
1765          */
1766         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1767
1768         cfq_service_tree_add(cfqd, cfqq, 1);
1769
1770         cfqq->slice_end = 0;
1771         cfq_mark_cfqq_slice_new(cfqq);
1772 }
1773
1774 /*
1775  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1776  * something we should do about it
1777  */
1778 static void
1779 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1780                 struct request *rq)
1781 {
1782         struct cfq_io_context *cic = RQ_CIC(rq);
1783
1784         if (rq_is_meta(rq))
1785                 cfqq->meta_pending++;
1786
1787         cfq_update_io_thinktime(cfqd, cic);
1788         cfq_update_io_seektime(cfqd, cic, rq);
1789         cfq_update_idle_window(cfqd, cfqq, cic);
1790
1791         cic->last_request_pos = rq->sector + rq->nr_sectors;
1792
1793         if (cfqq == cfqd->active_queue) {
1794                 /*
1795                  * if we are waiting for a request for this queue, let it rip
1796                  * immediately and flag that we must not expire this queue
1797                  * just now
1798                  */
1799                 if (cfq_cfqq_wait_request(cfqq)) {
1800                         cfq_mark_cfqq_must_dispatch(cfqq);
1801                         del_timer(&cfqd->idle_slice_timer);
1802                         blk_start_queueing(cfqd->queue);
1803                 }
1804         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1805                 /*
1806                  * not the active queue - expire current slice if it is
1807                  * idle and has expired it's mean thinktime or this new queue
1808                  * has some old slice time left and is of higher priority
1809                  */
1810                 cfq_preempt_queue(cfqd, cfqq);
1811                 cfq_mark_cfqq_must_dispatch(cfqq);
1812                 blk_start_queueing(cfqd->queue);
1813         }
1814 }
1815
1816 static void cfq_insert_request(struct request_queue *q, struct request *rq)
1817 {
1818         struct cfq_data *cfqd = q->elevator->elevator_data;
1819         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1820
1821         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
1822
1823         cfq_add_rq_rb(rq);
1824
1825         list_add_tail(&rq->queuelist, &cfqq->fifo);
1826
1827         cfq_rq_enqueued(cfqd, cfqq, rq);
1828 }
1829
1830 static void cfq_completed_request(struct request_queue *q, struct request *rq)
1831 {
1832         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1833         struct cfq_data *cfqd = cfqq->cfqd;
1834         const int sync = rq_is_sync(rq);
1835         unsigned long now;
1836
1837         now = jiffies;
1838
1839         WARN_ON(!cfqd->rq_in_driver);
1840         WARN_ON(!cfqq->dispatched);
1841         cfqd->rq_in_driver--;
1842         cfqq->dispatched--;
1843
1844         if (cfq_cfqq_sync(cfqq))
1845                 cfqd->sync_flight--;
1846
1847         if (!cfq_class_idle(cfqq))
1848                 cfqd->last_end_request = now;
1849
1850         if (sync)
1851                 RQ_CIC(rq)->last_end_request = now;
1852
1853         /*
1854          * If this is the active queue, check if it needs to be expired,
1855          * or if we want to idle in case it has no pending requests.
1856          */
1857         if (cfqd->active_queue == cfqq) {
1858                 if (cfq_cfqq_slice_new(cfqq)) {
1859                         cfq_set_prio_slice(cfqd, cfqq);
1860                         cfq_clear_cfqq_slice_new(cfqq);
1861                 }
1862                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
1863                         cfq_slice_expired(cfqd, 1);
1864                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1865                         cfq_arm_slice_timer(cfqd);
1866         }
1867
1868         if (!cfqd->rq_in_driver)
1869                 cfq_schedule_dispatch(cfqd);
1870 }
1871
1872 /*
1873  * we temporarily boost lower priority queues if they are holding fs exclusive
1874  * resources. they are boosted to normal prio (CLASS_BE/4)
1875  */
1876 static void cfq_prio_boost(struct cfq_queue *cfqq)
1877 {
1878         if (has_fs_excl()) {
1879                 /*
1880                  * boost idle prio on transactions that would lock out other
1881                  * users of the filesystem
1882                  */
1883                 if (cfq_class_idle(cfqq))
1884                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1885                 if (cfqq->ioprio > IOPRIO_NORM)
1886                         cfqq->ioprio = IOPRIO_NORM;
1887         } else {
1888                 /*
1889                  * check if we need to unboost the queue
1890                  */
1891                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1892                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1893                 if (cfqq->ioprio != cfqq->org_ioprio)
1894                         cfqq->ioprio = cfqq->org_ioprio;
1895         }
1896 }
1897
1898 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1899 {
1900         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1901             !cfq_cfqq_must_alloc_slice(cfqq)) {
1902                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1903                 return ELV_MQUEUE_MUST;
1904         }
1905
1906         return ELV_MQUEUE_MAY;
1907 }
1908
1909 static int cfq_may_queue(struct request_queue *q, int rw)
1910 {
1911         struct cfq_data *cfqd = q->elevator->elevator_data;
1912         struct task_struct *tsk = current;
1913         struct cfq_io_context *cic;
1914         struct cfq_queue *cfqq;
1915
1916         /*
1917          * don't force setup of a queue from here, as a call to may_queue
1918          * does not necessarily imply that a request actually will be queued.
1919          * so just lookup a possibly existing queue, or return 'may queue'
1920          * if that fails
1921          */
1922         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1923         if (!cic)
1924                 return ELV_MQUEUE_MAY;
1925
1926         cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
1927         if (cfqq) {
1928                 cfq_init_prio_data(cfqq, cic->ioc);
1929                 cfq_prio_boost(cfqq);
1930
1931                 return __cfq_may_queue(cfqq);
1932         }
1933
1934         return ELV_MQUEUE_MAY;
1935 }
1936
1937 /*
1938  * queue lock held here
1939  */
1940 static void cfq_put_request(struct request *rq)
1941 {
1942         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1943
1944         if (cfqq) {
1945                 const int rw = rq_data_dir(rq);
1946
1947                 BUG_ON(!cfqq->allocated[rw]);
1948                 cfqq->allocated[rw]--;
1949
1950                 put_io_context(RQ_CIC(rq)->ioc);
1951
1952                 rq->elevator_private = NULL;
1953                 rq->elevator_private2 = NULL;
1954
1955                 cfq_put_queue(cfqq);
1956         }
1957 }
1958
1959 /*
1960  * Allocate cfq data structures associated with this request.
1961  */
1962 static int
1963 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
1964 {
1965         struct cfq_data *cfqd = q->elevator->elevator_data;
1966         struct cfq_io_context *cic;
1967         const int rw = rq_data_dir(rq);
1968         const int is_sync = rq_is_sync(rq);
1969         struct cfq_queue *cfqq;
1970         unsigned long flags;
1971
1972         might_sleep_if(gfp_mask & __GFP_WAIT);
1973
1974         cic = cfq_get_io_context(cfqd, gfp_mask);
1975
1976         spin_lock_irqsave(q->queue_lock, flags);
1977
1978         if (!cic)
1979                 goto queue_fail;
1980
1981         cfqq = cic_to_cfqq(cic, is_sync);
1982         if (!cfqq) {
1983                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
1984
1985                 if (!cfqq)
1986                         goto queue_fail;
1987
1988                 cic_set_cfqq(cic, cfqq, is_sync);
1989         }
1990
1991         cfqq->allocated[rw]++;
1992         cfq_clear_cfqq_must_alloc(cfqq);
1993         atomic_inc(&cfqq->ref);
1994
1995         spin_unlock_irqrestore(q->queue_lock, flags);
1996
1997         rq->elevator_private = cic;
1998         rq->elevator_private2 = cfqq;
1999         return 0;
2000
2001 queue_fail:
2002         if (cic)
2003                 put_io_context(cic->ioc);
2004
2005         cfq_schedule_dispatch(cfqd);
2006         spin_unlock_irqrestore(q->queue_lock, flags);
2007         return 1;
2008 }
2009
2010 static void cfq_kick_queue(struct work_struct *work)
2011 {
2012         struct cfq_data *cfqd =
2013                 container_of(work, struct cfq_data, unplug_work);
2014         struct request_queue *q = cfqd->queue;
2015         unsigned long flags;
2016
2017         spin_lock_irqsave(q->queue_lock, flags);
2018         blk_start_queueing(q);
2019         spin_unlock_irqrestore(q->queue_lock, flags);
2020 }
2021
2022 /*
2023  * Timer running if the active_queue is currently idling inside its time slice
2024  */
2025 static void cfq_idle_slice_timer(unsigned long data)
2026 {
2027         struct cfq_data *cfqd = (struct cfq_data *) data;
2028         struct cfq_queue *cfqq;
2029         unsigned long flags;
2030         int timed_out = 1;
2031
2032         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2033
2034         cfqq = cfqd->active_queue;
2035         if (cfqq) {
2036                 timed_out = 0;
2037
2038                 /*
2039                  * expired
2040                  */
2041                 if (cfq_slice_used(cfqq))
2042                         goto expire;
2043
2044                 /*
2045                  * only expire and reinvoke request handler, if there are
2046                  * other queues with pending requests
2047                  */
2048                 if (!cfqd->busy_queues)
2049                         goto out_cont;
2050
2051                 /*
2052                  * not expired and it has a request pending, let it dispatch
2053                  */
2054                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
2055                         cfq_mark_cfqq_must_dispatch(cfqq);
2056                         goto out_kick;
2057                 }
2058         }
2059 expire:
2060         cfq_slice_expired(cfqd, timed_out);
2061 out_kick:
2062         cfq_schedule_dispatch(cfqd);
2063 out_cont:
2064         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2065 }
2066
2067 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2068 {
2069         del_timer_sync(&cfqd->idle_slice_timer);
2070         kblockd_flush_work(&cfqd->unplug_work);
2071 }
2072
2073 static void cfq_put_async_queues(struct cfq_data *cfqd)
2074 {
2075         int i;
2076
2077         for (i = 0; i < IOPRIO_BE_NR; i++) {
2078                 if (cfqd->async_cfqq[0][i])
2079                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2080                 if (cfqd->async_cfqq[1][i])
2081                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2082         }
2083
2084         if (cfqd->async_idle_cfqq)
2085                 cfq_put_queue(cfqd->async_idle_cfqq);
2086 }
2087
2088 static void cfq_exit_queue(elevator_t *e)
2089 {
2090         struct cfq_data *cfqd = e->elevator_data;
2091         struct request_queue *q = cfqd->queue;
2092
2093         cfq_shutdown_timer_wq(cfqd);
2094
2095         spin_lock_irq(q->queue_lock);
2096
2097         if (cfqd->active_queue)
2098                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2099
2100         while (!list_empty(&cfqd->cic_list)) {
2101                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2102                                                         struct cfq_io_context,
2103                                                         queue_list);
2104
2105                 __cfq_exit_single_io_context(cfqd, cic);
2106         }
2107
2108         cfq_put_async_queues(cfqd);
2109
2110         spin_unlock_irq(q->queue_lock);
2111
2112         cfq_shutdown_timer_wq(cfqd);
2113
2114         kfree(cfqd);
2115 }
2116
2117 static void *cfq_init_queue(struct request_queue *q)
2118 {
2119         struct cfq_data *cfqd;
2120
2121         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2122         if (!cfqd)
2123                 return NULL;
2124
2125         cfqd->service_tree = CFQ_RB_ROOT;
2126         INIT_LIST_HEAD(&cfqd->cic_list);
2127
2128         cfqd->queue = q;
2129
2130         init_timer(&cfqd->idle_slice_timer);
2131         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2132         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2133
2134         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2135
2136         cfqd->last_end_request = jiffies;
2137         cfqd->cfq_quantum = cfq_quantum;
2138         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2139         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2140         cfqd->cfq_back_max = cfq_back_max;
2141         cfqd->cfq_back_penalty = cfq_back_penalty;
2142         cfqd->cfq_slice[0] = cfq_slice_async;
2143         cfqd->cfq_slice[1] = cfq_slice_sync;
2144         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2145         cfqd->cfq_slice_idle = cfq_slice_idle;
2146
2147         return cfqd;
2148 }
2149
2150 static void cfq_slab_kill(void)
2151 {
2152         /*
2153          * Caller already ensured that pending RCU callbacks are completed,
2154          * so we should have no busy allocations at this point.
2155          */
2156         if (cfq_pool)
2157                 kmem_cache_destroy(cfq_pool);
2158         if (cfq_ioc_pool)
2159                 kmem_cache_destroy(cfq_ioc_pool);
2160 }
2161
2162 static int __init cfq_slab_setup(void)
2163 {
2164         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2165         if (!cfq_pool)
2166                 goto fail;
2167
2168         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2169         if (!cfq_ioc_pool)
2170                 goto fail;
2171
2172         return 0;
2173 fail:
2174         cfq_slab_kill();
2175         return -ENOMEM;
2176 }
2177
2178 /*
2179  * sysfs parts below -->
2180  */
2181 static ssize_t
2182 cfq_var_show(unsigned int var, char *page)
2183 {
2184         return sprintf(page, "%d\n", var);
2185 }
2186
2187 static ssize_t
2188 cfq_var_store(unsigned int *var, const char *page, size_t count)
2189 {
2190         char *p = (char *) page;
2191
2192         *var = simple_strtoul(p, &p, 10);
2193         return count;
2194 }
2195
2196 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2197 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2198 {                                                                       \
2199         struct cfq_data *cfqd = e->elevator_data;                       \
2200         unsigned int __data = __VAR;                                    \
2201         if (__CONV)                                                     \
2202                 __data = jiffies_to_msecs(__data);                      \
2203         return cfq_var_show(__data, (page));                            \
2204 }
2205 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2206 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2207 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2208 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2209 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2210 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2211 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2212 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2213 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2214 #undef SHOW_FUNCTION
2215
2216 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2217 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2218 {                                                                       \
2219         struct cfq_data *cfqd = e->elevator_data;                       \
2220         unsigned int __data;                                            \
2221         int ret = cfq_var_store(&__data, (page), count);                \
2222         if (__data < (MIN))                                             \
2223                 __data = (MIN);                                         \
2224         else if (__data > (MAX))                                        \
2225                 __data = (MAX);                                         \
2226         if (__CONV)                                                     \
2227                 *(__PTR) = msecs_to_jiffies(__data);                    \
2228         else                                                            \
2229                 *(__PTR) = __data;                                      \
2230         return ret;                                                     \
2231 }
2232 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2233 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2234                 UINT_MAX, 1);
2235 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2236                 UINT_MAX, 1);
2237 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2238 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2239                 UINT_MAX, 0);
2240 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2241 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2242 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2243 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2244                 UINT_MAX, 0);
2245 #undef STORE_FUNCTION
2246
2247 #define CFQ_ATTR(name) \
2248         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2249
2250 static struct elv_fs_entry cfq_attrs[] = {
2251         CFQ_ATTR(quantum),
2252         CFQ_ATTR(fifo_expire_sync),
2253         CFQ_ATTR(fifo_expire_async),
2254         CFQ_ATTR(back_seek_max),
2255         CFQ_ATTR(back_seek_penalty),
2256         CFQ_ATTR(slice_sync),
2257         CFQ_ATTR(slice_async),
2258         CFQ_ATTR(slice_async_rq),
2259         CFQ_ATTR(slice_idle),
2260         __ATTR_NULL
2261 };
2262
2263 static struct elevator_type iosched_cfq = {
2264         .ops = {
2265                 .elevator_merge_fn =            cfq_merge,
2266                 .elevator_merged_fn =           cfq_merged_request,
2267                 .elevator_merge_req_fn =        cfq_merged_requests,
2268                 .elevator_allow_merge_fn =      cfq_allow_merge,
2269                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2270                 .elevator_add_req_fn =          cfq_insert_request,
2271                 .elevator_activate_req_fn =     cfq_activate_request,
2272                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2273                 .elevator_queue_empty_fn =      cfq_queue_empty,
2274                 .elevator_completed_req_fn =    cfq_completed_request,
2275                 .elevator_former_req_fn =       elv_rb_former_request,
2276                 .elevator_latter_req_fn =       elv_rb_latter_request,
2277                 .elevator_set_req_fn =          cfq_set_request,
2278                 .elevator_put_req_fn =          cfq_put_request,
2279                 .elevator_may_queue_fn =        cfq_may_queue,
2280                 .elevator_init_fn =             cfq_init_queue,
2281                 .elevator_exit_fn =             cfq_exit_queue,
2282                 .trim =                         cfq_free_io_context,
2283         },
2284         .elevator_attrs =       cfq_attrs,
2285         .elevator_name =        "cfq",
2286         .elevator_owner =       THIS_MODULE,
2287 };
2288
2289 static int __init cfq_init(void)
2290 {
2291         /*
2292          * could be 0 on HZ < 1000 setups
2293          */
2294         if (!cfq_slice_async)
2295                 cfq_slice_async = 1;
2296         if (!cfq_slice_idle)
2297                 cfq_slice_idle = 1;
2298
2299         if (cfq_slab_setup())
2300                 return -ENOMEM;
2301
2302         elv_register(&iosched_cfq);
2303
2304         return 0;
2305 }
2306
2307 static void __exit cfq_exit(void)
2308 {
2309         DECLARE_COMPLETION_ONSTACK(all_gone);
2310         elv_unregister(&iosched_cfq);
2311         ioc_gone = &all_gone;
2312         /* ioc_gone's update must be visible before reading ioc_count */
2313         smp_wmb();
2314
2315         /*
2316          * this also protects us from entering cfq_slab_kill() with
2317          * pending RCU callbacks
2318          */
2319         if (elv_ioc_count_read(ioc_count))
2320                 wait_for_completion(ioc_gone);
2321         cfq_slab_kill();
2322 }
2323
2324 module_init(cfq_init);
2325 module_exit(cfq_exit);
2326
2327 MODULE_AUTHOR("Jens Axboe");
2328 MODULE_LICENSE("GPL");
2329 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");