x86: always page-align per-cpu area start and size
[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 #include <linux/blktrace_api.h>
15
16 /*
17  * tunables
18  */
19 /* max queue in one round of service */
20 static const int cfq_quantum = 4;
21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
22 /* maximum backwards seek, in KiB */
23 static const int cfq_back_max = 16 * 1024;
24 /* penalty of a backwards seek */
25 static const int cfq_back_penalty = 2;
26 static const int cfq_slice_sync = HZ / 10;
27 static int cfq_slice_async = HZ / 25;
28 static const int cfq_slice_async_rq = 2;
29 static int cfq_slice_idle = HZ / 125;
30
31 /*
32  * offset from end of service tree
33  */
34 #define CFQ_IDLE_DELAY          (HZ / 5)
35
36 /*
37  * below this threshold, we consider thinktime immediate
38  */
39 #define CFQ_MIN_TT              (2)
40
41 #define CFQ_SLICE_SCALE         (5)
42 #define CFQ_HW_QUEUE_MIN        (5)
43
44 #define RQ_CIC(rq)              \
45         ((struct cfq_io_context *) (rq)->elevator_private)
46 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
47
48 static struct kmem_cache *cfq_pool;
49 static struct kmem_cache *cfq_ioc_pool;
50
51 static DEFINE_PER_CPU(unsigned long, ioc_count);
52 static struct completion *ioc_gone;
53 static DEFINE_SPINLOCK(ioc_gone_lock);
54
55 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
56 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
57 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
58
59 #define ASYNC                   (0)
60 #define SYNC                    (1)
61
62 #define sample_valid(samples)   ((samples) > 80)
63
64 /*
65  * Most of our rbtree usage is for sorting with min extraction, so
66  * if we cache the leftmost node we don't have to walk down the tree
67  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
68  * move this into the elevator for the rq sorting as well.
69  */
70 struct cfq_rb_root {
71         struct rb_root rb;
72         struct rb_node *left;
73 };
74 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, }
75
76 /*
77  * Per block device queue structure
78  */
79 struct cfq_data {
80         struct request_queue *queue;
81
82         /*
83          * rr list of queues with requests and the count of them
84          */
85         struct cfq_rb_root service_tree;
86         unsigned int busy_queues;
87
88         int rq_in_driver;
89         int sync_flight;
90
91         /*
92          * queue-depth detection
93          */
94         int rq_queued;
95         int hw_tag;
96         int hw_tag_samples;
97         int rq_in_driver_peak;
98
99         /*
100          * idle window management
101          */
102         struct timer_list idle_slice_timer;
103         struct work_struct unplug_work;
104
105         struct cfq_queue *active_queue;
106         struct cfq_io_context *active_cic;
107
108         /*
109          * async queue for each priority case
110          */
111         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
112         struct cfq_queue *async_idle_cfqq;
113
114         sector_t last_position;
115         unsigned long last_end_request;
116
117         /*
118          * tunables, see top of file
119          */
120         unsigned int cfq_quantum;
121         unsigned int cfq_fifo_expire[2];
122         unsigned int cfq_back_penalty;
123         unsigned int cfq_back_max;
124         unsigned int cfq_slice[2];
125         unsigned int cfq_slice_async_rq;
126         unsigned int cfq_slice_idle;
127
128         struct list_head cic_list;
129 };
130
131 /*
132  * Per process-grouping structure
133  */
134 struct cfq_queue {
135         /* reference count */
136         atomic_t ref;
137         /* various state flags, see below */
138         unsigned int flags;
139         /* parent cfq_data */
140         struct cfq_data *cfqd;
141         /* service_tree member */
142         struct rb_node rb_node;
143         /* service_tree key */
144         unsigned long rb_key;
145         /* sorted list of pending requests */
146         struct rb_root sort_list;
147         /* if fifo isn't expired, next request to serve */
148         struct request *next_rq;
149         /* requests queued in sort_list */
150         int queued[2];
151         /* currently allocated requests */
152         int allocated[2];
153         /* fifo list of requests in sort_list */
154         struct list_head fifo;
155
156         unsigned long slice_end;
157         long slice_resid;
158
159         /* pending metadata requests */
160         int meta_pending;
161         /* number of requests that are on the dispatch list or inside driver */
162         int dispatched;
163
164         /* io prio of this group */
165         unsigned short ioprio, org_ioprio;
166         unsigned short ioprio_class, org_ioprio_class;
167
168         pid_t pid;
169 };
170
171 enum cfqq_state_flags {
172         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
173         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
174         CFQ_CFQQ_FLAG_must_alloc,       /* must be allowed rq alloc */
175         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
176         CFQ_CFQQ_FLAG_must_dispatch,    /* must dispatch, even if expired */
177         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
178         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
179         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
180         CFQ_CFQQ_FLAG_queue_new,        /* queue never been serviced */
181         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
182         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
183 };
184
185 #define CFQ_CFQQ_FNS(name)                                              \
186 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
187 {                                                                       \
188         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
189 }                                                                       \
190 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
191 {                                                                       \
192         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
193 }                                                                       \
194 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
195 {                                                                       \
196         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
197 }
198
199 CFQ_CFQQ_FNS(on_rr);
200 CFQ_CFQQ_FNS(wait_request);
201 CFQ_CFQQ_FNS(must_alloc);
202 CFQ_CFQQ_FNS(must_alloc_slice);
203 CFQ_CFQQ_FNS(must_dispatch);
204 CFQ_CFQQ_FNS(fifo_expire);
205 CFQ_CFQQ_FNS(idle_window);
206 CFQ_CFQQ_FNS(prio_changed);
207 CFQ_CFQQ_FNS(queue_new);
208 CFQ_CFQQ_FNS(slice_new);
209 CFQ_CFQQ_FNS(sync);
210 #undef CFQ_CFQQ_FNS
211
212 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
213         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
214 #define cfq_log(cfqd, fmt, args...)     \
215         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
216
217 static void cfq_dispatch_insert(struct request_queue *, struct request *);
218 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
219                                        struct io_context *, gfp_t);
220 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
221                                                 struct io_context *);
222
223 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
224                                             int is_sync)
225 {
226         return cic->cfqq[!!is_sync];
227 }
228
229 static inline void cic_set_cfqq(struct cfq_io_context *cic,
230                                 struct cfq_queue *cfqq, int is_sync)
231 {
232         cic->cfqq[!!is_sync] = cfqq;
233 }
234
235 /*
236  * We regard a request as SYNC, if it's either a read or has the SYNC bit
237  * set (in which case it could also be direct WRITE).
238  */
239 static inline int cfq_bio_sync(struct bio *bio)
240 {
241         if (bio_data_dir(bio) == READ || bio_sync(bio))
242                 return 1;
243
244         return 0;
245 }
246
247 /*
248  * scheduler run of queue, if there are requests pending and no one in the
249  * driver that will restart queueing
250  */
251 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
252 {
253         if (cfqd->busy_queues) {
254                 cfq_log(cfqd, "schedule dispatch");
255                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
256         }
257 }
258
259 static int cfq_queue_empty(struct request_queue *q)
260 {
261         struct cfq_data *cfqd = q->elevator->elevator_data;
262
263         return !cfqd->busy_queues;
264 }
265
266 /*
267  * Scale schedule slice based on io priority. Use the sync time slice only
268  * if a queue is marked sync and has sync io queued. A sync queue with async
269  * io only, should not get full sync slice length.
270  */
271 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
272                                  unsigned short prio)
273 {
274         const int base_slice = cfqd->cfq_slice[sync];
275
276         WARN_ON(prio >= IOPRIO_BE_NR);
277
278         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
279 }
280
281 static inline int
282 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
283 {
284         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
285 }
286
287 static inline void
288 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
289 {
290         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
291         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
292 }
293
294 /*
295  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
296  * isn't valid until the first request from the dispatch is activated
297  * and the slice time set.
298  */
299 static inline int cfq_slice_used(struct cfq_queue *cfqq)
300 {
301         if (cfq_cfqq_slice_new(cfqq))
302                 return 0;
303         if (time_before(jiffies, cfqq->slice_end))
304                 return 0;
305
306         return 1;
307 }
308
309 /*
310  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
311  * We choose the request that is closest to the head right now. Distance
312  * behind the head is penalized and only allowed to a certain extent.
313  */
314 static struct request *
315 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
316 {
317         sector_t last, s1, s2, d1 = 0, d2 = 0;
318         unsigned long back_max;
319 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
320 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
321         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
322
323         if (rq1 == NULL || rq1 == rq2)
324                 return rq2;
325         if (rq2 == NULL)
326                 return rq1;
327
328         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
329                 return rq1;
330         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
331                 return rq2;
332         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
333                 return rq1;
334         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
335                 return rq2;
336
337         s1 = rq1->sector;
338         s2 = rq2->sector;
339
340         last = cfqd->last_position;
341
342         /*
343          * by definition, 1KiB is 2 sectors
344          */
345         back_max = cfqd->cfq_back_max * 2;
346
347         /*
348          * Strict one way elevator _except_ in the case where we allow
349          * short backward seeks which are biased as twice the cost of a
350          * similar forward seek.
351          */
352         if (s1 >= last)
353                 d1 = s1 - last;
354         else if (s1 + back_max >= last)
355                 d1 = (last - s1) * cfqd->cfq_back_penalty;
356         else
357                 wrap |= CFQ_RQ1_WRAP;
358
359         if (s2 >= last)
360                 d2 = s2 - last;
361         else if (s2 + back_max >= last)
362                 d2 = (last - s2) * cfqd->cfq_back_penalty;
363         else
364                 wrap |= CFQ_RQ2_WRAP;
365
366         /* Found required data */
367
368         /*
369          * By doing switch() on the bit mask "wrap" we avoid having to
370          * check two variables for all permutations: --> faster!
371          */
372         switch (wrap) {
373         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
374                 if (d1 < d2)
375                         return rq1;
376                 else if (d2 < d1)
377                         return rq2;
378                 else {
379                         if (s1 >= s2)
380                                 return rq1;
381                         else
382                                 return rq2;
383                 }
384
385         case CFQ_RQ2_WRAP:
386                 return rq1;
387         case CFQ_RQ1_WRAP:
388                 return rq2;
389         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
390         default:
391                 /*
392                  * Since both rqs are wrapped,
393                  * start with the one that's further behind head
394                  * (--> only *one* back seek required),
395                  * since back seek takes more time than forward.
396                  */
397                 if (s1 <= s2)
398                         return rq1;
399                 else
400                         return rq2;
401         }
402 }
403
404 /*
405  * The below is leftmost cache rbtree addon
406  */
407 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
408 {
409         if (!root->left)
410                 root->left = rb_first(&root->rb);
411
412         if (root->left)
413                 return rb_entry(root->left, struct cfq_queue, rb_node);
414
415         return NULL;
416 }
417
418 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
419 {
420         if (root->left == n)
421                 root->left = NULL;
422
423         rb_erase(n, &root->rb);
424         RB_CLEAR_NODE(n);
425 }
426
427 /*
428  * would be nice to take fifo expire time into account as well
429  */
430 static struct request *
431 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
432                   struct request *last)
433 {
434         struct rb_node *rbnext = rb_next(&last->rb_node);
435         struct rb_node *rbprev = rb_prev(&last->rb_node);
436         struct request *next = NULL, *prev = NULL;
437
438         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
439
440         if (rbprev)
441                 prev = rb_entry_rq(rbprev);
442
443         if (rbnext)
444                 next = rb_entry_rq(rbnext);
445         else {
446                 rbnext = rb_first(&cfqq->sort_list);
447                 if (rbnext && rbnext != &last->rb_node)
448                         next = rb_entry_rq(rbnext);
449         }
450
451         return cfq_choose_req(cfqd, next, prev);
452 }
453
454 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
455                                       struct cfq_queue *cfqq)
456 {
457         /*
458          * just an approximation, should be ok.
459          */
460         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
461                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
462 }
463
464 /*
465  * The cfqd->service_tree holds all pending cfq_queue's that have
466  * requests waiting to be processed. It is sorted in the order that
467  * we will service the queues.
468  */
469 static void cfq_service_tree_add(struct cfq_data *cfqd,
470                                     struct cfq_queue *cfqq, int add_front)
471 {
472         struct rb_node **p, *parent;
473         struct cfq_queue *__cfqq;
474         unsigned long rb_key;
475         int left;
476
477         if (cfq_class_idle(cfqq)) {
478                 rb_key = CFQ_IDLE_DELAY;
479                 parent = rb_last(&cfqd->service_tree.rb);
480                 if (parent && parent != &cfqq->rb_node) {
481                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
482                         rb_key += __cfqq->rb_key;
483                 } else
484                         rb_key += jiffies;
485         } else if (!add_front) {
486                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
487                 rb_key += cfqq->slice_resid;
488                 cfqq->slice_resid = 0;
489         } else
490                 rb_key = 0;
491
492         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
493                 /*
494                  * same position, nothing more to do
495                  */
496                 if (rb_key == cfqq->rb_key)
497                         return;
498
499                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
500         }
501
502         left = 1;
503         parent = NULL;
504         p = &cfqd->service_tree.rb.rb_node;
505         while (*p) {
506                 struct rb_node **n;
507
508                 parent = *p;
509                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
510
511                 /*
512                  * sort RT queues first, we always want to give
513                  * preference to them. IDLE queues goes to the back.
514                  * after that, sort on the next service time.
515                  */
516                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
517                         n = &(*p)->rb_left;
518                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
519                         n = &(*p)->rb_right;
520                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
521                         n = &(*p)->rb_left;
522                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
523                         n = &(*p)->rb_right;
524                 else if (rb_key < __cfqq->rb_key)
525                         n = &(*p)->rb_left;
526                 else
527                         n = &(*p)->rb_right;
528
529                 if (n == &(*p)->rb_right)
530                         left = 0;
531
532                 p = n;
533         }
534
535         if (left)
536                 cfqd->service_tree.left = &cfqq->rb_node;
537
538         cfqq->rb_key = rb_key;
539         rb_link_node(&cfqq->rb_node, parent, p);
540         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
541 }
542
543 /*
544  * Update cfqq's position in the service tree.
545  */
546 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
547 {
548         /*
549          * Resorting requires the cfqq to be on the RR list already.
550          */
551         if (cfq_cfqq_on_rr(cfqq))
552                 cfq_service_tree_add(cfqd, cfqq, 0);
553 }
554
555 /*
556  * add to busy list of queues for service, trying to be fair in ordering
557  * the pending list according to last request service
558  */
559 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
560 {
561         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
562         BUG_ON(cfq_cfqq_on_rr(cfqq));
563         cfq_mark_cfqq_on_rr(cfqq);
564         cfqd->busy_queues++;
565
566         cfq_resort_rr_list(cfqd, cfqq);
567 }
568
569 /*
570  * Called when the cfqq no longer has requests pending, remove it from
571  * the service tree.
572  */
573 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
574 {
575         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
576         BUG_ON(!cfq_cfqq_on_rr(cfqq));
577         cfq_clear_cfqq_on_rr(cfqq);
578
579         if (!RB_EMPTY_NODE(&cfqq->rb_node))
580                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
581
582         BUG_ON(!cfqd->busy_queues);
583         cfqd->busy_queues--;
584 }
585
586 /*
587  * rb tree support functions
588  */
589 static void cfq_del_rq_rb(struct request *rq)
590 {
591         struct cfq_queue *cfqq = RQ_CFQQ(rq);
592         struct cfq_data *cfqd = cfqq->cfqd;
593         const int sync = rq_is_sync(rq);
594
595         BUG_ON(!cfqq->queued[sync]);
596         cfqq->queued[sync]--;
597
598         elv_rb_del(&cfqq->sort_list, rq);
599
600         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
601                 cfq_del_cfqq_rr(cfqd, cfqq);
602 }
603
604 static void cfq_add_rq_rb(struct request *rq)
605 {
606         struct cfq_queue *cfqq = RQ_CFQQ(rq);
607         struct cfq_data *cfqd = cfqq->cfqd;
608         struct request *__alias;
609
610         cfqq->queued[rq_is_sync(rq)]++;
611
612         /*
613          * looks a little odd, but the first insert might return an alias.
614          * if that happens, put the alias on the dispatch list
615          */
616         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
617                 cfq_dispatch_insert(cfqd->queue, __alias);
618
619         if (!cfq_cfqq_on_rr(cfqq))
620                 cfq_add_cfqq_rr(cfqd, cfqq);
621
622         /*
623          * check if this request is a better next-serve candidate
624          */
625         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
626         BUG_ON(!cfqq->next_rq);
627 }
628
629 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
630 {
631         elv_rb_del(&cfqq->sort_list, rq);
632         cfqq->queued[rq_is_sync(rq)]--;
633         cfq_add_rq_rb(rq);
634 }
635
636 static struct request *
637 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
638 {
639         struct task_struct *tsk = current;
640         struct cfq_io_context *cic;
641         struct cfq_queue *cfqq;
642
643         cic = cfq_cic_lookup(cfqd, tsk->io_context);
644         if (!cic)
645                 return NULL;
646
647         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
648         if (cfqq) {
649                 sector_t sector = bio->bi_sector + bio_sectors(bio);
650
651                 return elv_rb_find(&cfqq->sort_list, sector);
652         }
653
654         return NULL;
655 }
656
657 static void cfq_activate_request(struct request_queue *q, struct request *rq)
658 {
659         struct cfq_data *cfqd = q->elevator->elevator_data;
660
661         cfqd->rq_in_driver++;
662         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
663                                                 cfqd->rq_in_driver);
664
665         cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
666 }
667
668 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
669 {
670         struct cfq_data *cfqd = q->elevator->elevator_data;
671
672         WARN_ON(!cfqd->rq_in_driver);
673         cfqd->rq_in_driver--;
674         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
675                                                 cfqd->rq_in_driver);
676 }
677
678 static void cfq_remove_request(struct request *rq)
679 {
680         struct cfq_queue *cfqq = RQ_CFQQ(rq);
681
682         if (cfqq->next_rq == rq)
683                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
684
685         list_del_init(&rq->queuelist);
686         cfq_del_rq_rb(rq);
687
688         cfqq->cfqd->rq_queued--;
689         if (rq_is_meta(rq)) {
690                 WARN_ON(!cfqq->meta_pending);
691                 cfqq->meta_pending--;
692         }
693 }
694
695 static int cfq_merge(struct request_queue *q, struct request **req,
696                      struct bio *bio)
697 {
698         struct cfq_data *cfqd = q->elevator->elevator_data;
699         struct request *__rq;
700
701         __rq = cfq_find_rq_fmerge(cfqd, bio);
702         if (__rq && elv_rq_merge_ok(__rq, bio)) {
703                 *req = __rq;
704                 return ELEVATOR_FRONT_MERGE;
705         }
706
707         return ELEVATOR_NO_MERGE;
708 }
709
710 static void cfq_merged_request(struct request_queue *q, struct request *req,
711                                int type)
712 {
713         if (type == ELEVATOR_FRONT_MERGE) {
714                 struct cfq_queue *cfqq = RQ_CFQQ(req);
715
716                 cfq_reposition_rq_rb(cfqq, req);
717         }
718 }
719
720 static void
721 cfq_merged_requests(struct request_queue *q, struct request *rq,
722                     struct request *next)
723 {
724         /*
725          * reposition in fifo if next is older than rq
726          */
727         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
728             time_before(next->start_time, rq->start_time))
729                 list_move(&rq->queuelist, &next->queuelist);
730
731         cfq_remove_request(next);
732 }
733
734 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
735                            struct bio *bio)
736 {
737         struct cfq_data *cfqd = q->elevator->elevator_data;
738         struct cfq_io_context *cic;
739         struct cfq_queue *cfqq;
740
741         /*
742          * Disallow merge of a sync bio into an async request.
743          */
744         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
745                 return 0;
746
747         /*
748          * Lookup the cfqq that this bio will be queued with. Allow
749          * merge only if rq is queued there.
750          */
751         cic = cfq_cic_lookup(cfqd, current->io_context);
752         if (!cic)
753                 return 0;
754
755         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
756         if (cfqq == RQ_CFQQ(rq))
757                 return 1;
758
759         return 0;
760 }
761
762 static void __cfq_set_active_queue(struct cfq_data *cfqd,
763                                    struct cfq_queue *cfqq)
764 {
765         if (cfqq) {
766                 cfq_log_cfqq(cfqd, cfqq, "set_active");
767                 cfqq->slice_end = 0;
768                 cfq_clear_cfqq_must_alloc_slice(cfqq);
769                 cfq_clear_cfqq_fifo_expire(cfqq);
770                 cfq_mark_cfqq_slice_new(cfqq);
771                 cfq_clear_cfqq_queue_new(cfqq);
772         }
773
774         cfqd->active_queue = cfqq;
775 }
776
777 /*
778  * current cfqq expired its slice (or was too idle), select new one
779  */
780 static void
781 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
782                     int timed_out)
783 {
784         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
785
786         if (cfq_cfqq_wait_request(cfqq))
787                 del_timer(&cfqd->idle_slice_timer);
788
789         cfq_clear_cfqq_must_dispatch(cfqq);
790         cfq_clear_cfqq_wait_request(cfqq);
791
792         /*
793          * store what was left of this slice, if the queue idled/timed out
794          */
795         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
796                 cfqq->slice_resid = cfqq->slice_end - jiffies;
797                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
798         }
799
800         cfq_resort_rr_list(cfqd, cfqq);
801
802         if (cfqq == cfqd->active_queue)
803                 cfqd->active_queue = NULL;
804
805         if (cfqd->active_cic) {
806                 put_io_context(cfqd->active_cic->ioc);
807                 cfqd->active_cic = NULL;
808         }
809 }
810
811 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
812 {
813         struct cfq_queue *cfqq = cfqd->active_queue;
814
815         if (cfqq)
816                 __cfq_slice_expired(cfqd, cfqq, timed_out);
817 }
818
819 /*
820  * Get next queue for service. Unless we have a queue preemption,
821  * we'll simply select the first cfqq in the service tree.
822  */
823 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
824 {
825         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
826                 return NULL;
827
828         return cfq_rb_first(&cfqd->service_tree);
829 }
830
831 /*
832  * Get and set a new active queue for service.
833  */
834 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
835 {
836         struct cfq_queue *cfqq;
837
838         cfqq = cfq_get_next_queue(cfqd);
839         __cfq_set_active_queue(cfqd, cfqq);
840         return cfqq;
841 }
842
843 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
844                                           struct request *rq)
845 {
846         if (rq->sector >= cfqd->last_position)
847                 return rq->sector - cfqd->last_position;
848         else
849                 return cfqd->last_position - rq->sector;
850 }
851
852 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
853 {
854         struct cfq_io_context *cic = cfqd->active_cic;
855
856         if (!sample_valid(cic->seek_samples))
857                 return 0;
858
859         return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
860 }
861
862 static int cfq_close_cooperator(struct cfq_data *cfq_data,
863                                 struct cfq_queue *cfqq)
864 {
865         /*
866          * We should notice if some of the queues are cooperating, eg
867          * working closely on the same area of the disk. In that case,
868          * we can group them together and don't waste time idling.
869          */
870         return 0;
871 }
872
873 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
874
875 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
876 {
877         struct cfq_queue *cfqq = cfqd->active_queue;
878         struct cfq_io_context *cic;
879         unsigned long sl;
880
881         /*
882          * SSD device without seek penalty, disable idling. But only do so
883          * for devices that support queuing, otherwise we still have a problem
884          * with sync vs async workloads.
885          */
886         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
887                 return;
888
889         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
890         WARN_ON(cfq_cfqq_slice_new(cfqq));
891
892         /*
893          * idle is disabled, either manually or by past process history
894          */
895         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
896                 return;
897
898         /*
899          * still requests with the driver, don't idle
900          */
901         if (cfqd->rq_in_driver)
902                 return;
903
904         /*
905          * task has exited, don't wait
906          */
907         cic = cfqd->active_cic;
908         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
909                 return;
910
911         /*
912          * See if this prio level has a good candidate
913          */
914         if (cfq_close_cooperator(cfqd, cfqq) &&
915             (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
916                 return;
917
918         cfq_mark_cfqq_must_dispatch(cfqq);
919         cfq_mark_cfqq_wait_request(cfqq);
920
921         /*
922          * we don't want to idle for seeks, but we do want to allow
923          * fair distribution of slice time for a process doing back-to-back
924          * seeks. so allow a little bit of time for him to submit a new rq
925          */
926         sl = cfqd->cfq_slice_idle;
927         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
928                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
929
930         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
931         cfq_log(cfqd, "arm_idle: %lu", sl);
932 }
933
934 /*
935  * Move request from internal lists to the request queue dispatch list.
936  */
937 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
938 {
939         struct cfq_data *cfqd = q->elevator->elevator_data;
940         struct cfq_queue *cfqq = RQ_CFQQ(rq);
941
942         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
943
944         cfq_remove_request(rq);
945         cfqq->dispatched++;
946         elv_dispatch_sort(q, rq);
947
948         if (cfq_cfqq_sync(cfqq))
949                 cfqd->sync_flight++;
950 }
951
952 /*
953  * return expired entry, or NULL to just start from scratch in rbtree
954  */
955 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
956 {
957         struct cfq_data *cfqd = cfqq->cfqd;
958         struct request *rq;
959         int fifo;
960
961         if (cfq_cfqq_fifo_expire(cfqq))
962                 return NULL;
963
964         cfq_mark_cfqq_fifo_expire(cfqq);
965
966         if (list_empty(&cfqq->fifo))
967                 return NULL;
968
969         fifo = cfq_cfqq_sync(cfqq);
970         rq = rq_entry_fifo(cfqq->fifo.next);
971
972         if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
973                 rq = NULL;
974
975         cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
976         return rq;
977 }
978
979 static inline int
980 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
981 {
982         const int base_rq = cfqd->cfq_slice_async_rq;
983
984         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
985
986         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
987 }
988
989 /*
990  * Select a queue for service. If we have a current active queue,
991  * check whether to continue servicing it, or retrieve and set a new one.
992  */
993 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
994 {
995         struct cfq_queue *cfqq;
996
997         cfqq = cfqd->active_queue;
998         if (!cfqq)
999                 goto new_queue;
1000
1001         /*
1002          * The active queue has run out of time, expire it and select new.
1003          */
1004         if (cfq_slice_used(cfqq))
1005                 goto expire;
1006
1007         /*
1008          * The active queue has requests and isn't expired, allow it to
1009          * dispatch.
1010          */
1011         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1012                 goto keep_queue;
1013
1014         /*
1015          * No requests pending. If the active queue still has requests in
1016          * flight or is idling for a new request, allow either of these
1017          * conditions to happen (or time out) before selecting a new queue.
1018          */
1019         if (timer_pending(&cfqd->idle_slice_timer) ||
1020             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1021                 cfqq = NULL;
1022                 goto keep_queue;
1023         }
1024
1025 expire:
1026         cfq_slice_expired(cfqd, 0);
1027 new_queue:
1028         cfqq = cfq_set_active_queue(cfqd);
1029 keep_queue:
1030         return cfqq;
1031 }
1032
1033 /*
1034  * Dispatch some requests from cfqq, moving them to the request queue
1035  * dispatch list.
1036  */
1037 static int
1038 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1039                         int max_dispatch)
1040 {
1041         int dispatched = 0;
1042
1043         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1044
1045         do {
1046                 struct request *rq;
1047
1048                 /*
1049                  * follow expired path, else get first next available
1050                  */
1051                 rq = cfq_check_fifo(cfqq);
1052                 if (rq == NULL)
1053                         rq = cfqq->next_rq;
1054
1055                 /*
1056                  * finally, insert request into driver dispatch list
1057                  */
1058                 cfq_dispatch_insert(cfqd->queue, rq);
1059
1060                 dispatched++;
1061
1062                 if (!cfqd->active_cic) {
1063                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1064                         cfqd->active_cic = RQ_CIC(rq);
1065                 }
1066
1067                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
1068                         break;
1069
1070         } while (dispatched < max_dispatch);
1071
1072         /*
1073          * expire an async queue immediately if it has used up its slice. idle
1074          * queue always expire after 1 dispatch round.
1075          */
1076         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1077             dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1078             cfq_class_idle(cfqq))) {
1079                 cfqq->slice_end = jiffies + 1;
1080                 cfq_slice_expired(cfqd, 0);
1081         }
1082
1083         return dispatched;
1084 }
1085
1086 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1087 {
1088         int dispatched = 0;
1089
1090         while (cfqq->next_rq) {
1091                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1092                 dispatched++;
1093         }
1094
1095         BUG_ON(!list_empty(&cfqq->fifo));
1096         return dispatched;
1097 }
1098
1099 /*
1100  * Drain our current requests. Used for barriers and when switching
1101  * io schedulers on-the-fly.
1102  */
1103 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1104 {
1105         struct cfq_queue *cfqq;
1106         int dispatched = 0;
1107
1108         while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1109                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1110
1111         cfq_slice_expired(cfqd, 0);
1112
1113         BUG_ON(cfqd->busy_queues);
1114
1115         cfq_log(cfqd, "forced_dispatch=%d\n", dispatched);
1116         return dispatched;
1117 }
1118
1119 static int cfq_dispatch_requests(struct request_queue *q, int force)
1120 {
1121         struct cfq_data *cfqd = q->elevator->elevator_data;
1122         struct cfq_queue *cfqq;
1123         int dispatched;
1124
1125         if (!cfqd->busy_queues)
1126                 return 0;
1127
1128         if (unlikely(force))
1129                 return cfq_forced_dispatch(cfqd);
1130
1131         dispatched = 0;
1132         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1133                 int max_dispatch;
1134
1135                 max_dispatch = cfqd->cfq_quantum;
1136                 if (cfq_class_idle(cfqq))
1137                         max_dispatch = 1;
1138
1139                 if (cfqq->dispatched >= max_dispatch && cfqd->busy_queues > 1)
1140                         break;
1141
1142                 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1143                         break;
1144
1145                 cfq_clear_cfqq_must_dispatch(cfqq);
1146                 cfq_clear_cfqq_wait_request(cfqq);
1147                 del_timer(&cfqd->idle_slice_timer);
1148
1149                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1150         }
1151
1152         cfq_log(cfqd, "dispatched=%d", dispatched);
1153         return dispatched;
1154 }
1155
1156 /*
1157  * task holds one reference to the queue, dropped when task exits. each rq
1158  * in-flight on this queue also holds a reference, dropped when rq is freed.
1159  *
1160  * queue lock must be held here.
1161  */
1162 static void cfq_put_queue(struct cfq_queue *cfqq)
1163 {
1164         struct cfq_data *cfqd = cfqq->cfqd;
1165
1166         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1167
1168         if (!atomic_dec_and_test(&cfqq->ref))
1169                 return;
1170
1171         cfq_log_cfqq(cfqd, cfqq, "put_queue");
1172         BUG_ON(rb_first(&cfqq->sort_list));
1173         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1174         BUG_ON(cfq_cfqq_on_rr(cfqq));
1175
1176         if (unlikely(cfqd->active_queue == cfqq)) {
1177                 __cfq_slice_expired(cfqd, cfqq, 0);
1178                 cfq_schedule_dispatch(cfqd);
1179         }
1180
1181         kmem_cache_free(cfq_pool, cfqq);
1182 }
1183
1184 /*
1185  * Must always be called with the rcu_read_lock() held
1186  */
1187 static void
1188 __call_for_each_cic(struct io_context *ioc,
1189                     void (*func)(struct io_context *, struct cfq_io_context *))
1190 {
1191         struct cfq_io_context *cic;
1192         struct hlist_node *n;
1193
1194         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1195                 func(ioc, cic);
1196 }
1197
1198 /*
1199  * Call func for each cic attached to this ioc.
1200  */
1201 static void
1202 call_for_each_cic(struct io_context *ioc,
1203                   void (*func)(struct io_context *, struct cfq_io_context *))
1204 {
1205         rcu_read_lock();
1206         __call_for_each_cic(ioc, func);
1207         rcu_read_unlock();
1208 }
1209
1210 static void cfq_cic_free_rcu(struct rcu_head *head)
1211 {
1212         struct cfq_io_context *cic;
1213
1214         cic = container_of(head, struct cfq_io_context, rcu_head);
1215
1216         kmem_cache_free(cfq_ioc_pool, cic);
1217         elv_ioc_count_dec(ioc_count);
1218
1219         if (ioc_gone) {
1220                 /*
1221                  * CFQ scheduler is exiting, grab exit lock and check
1222                  * the pending io context count. If it hits zero,
1223                  * complete ioc_gone and set it back to NULL
1224                  */
1225                 spin_lock(&ioc_gone_lock);
1226                 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1227                         complete(ioc_gone);
1228                         ioc_gone = NULL;
1229                 }
1230                 spin_unlock(&ioc_gone_lock);
1231         }
1232 }
1233
1234 static void cfq_cic_free(struct cfq_io_context *cic)
1235 {
1236         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1237 }
1238
1239 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1240 {
1241         unsigned long flags;
1242
1243         BUG_ON(!cic->dead_key);
1244
1245         spin_lock_irqsave(&ioc->lock, flags);
1246         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1247         hlist_del_rcu(&cic->cic_list);
1248         spin_unlock_irqrestore(&ioc->lock, flags);
1249
1250         cfq_cic_free(cic);
1251 }
1252
1253 /*
1254  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1255  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1256  * and ->trim() which is called with the task lock held
1257  */
1258 static void cfq_free_io_context(struct io_context *ioc)
1259 {
1260         /*
1261          * ioc->refcount is zero here, or we are called from elv_unregister(),
1262          * so no more cic's are allowed to be linked into this ioc.  So it
1263          * should be ok to iterate over the known list, we will see all cic's
1264          * since no new ones are added.
1265          */
1266         __call_for_each_cic(ioc, cic_free_func);
1267 }
1268
1269 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1270 {
1271         if (unlikely(cfqq == cfqd->active_queue)) {
1272                 __cfq_slice_expired(cfqd, cfqq, 0);
1273                 cfq_schedule_dispatch(cfqd);
1274         }
1275
1276         cfq_put_queue(cfqq);
1277 }
1278
1279 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1280                                          struct cfq_io_context *cic)
1281 {
1282         struct io_context *ioc = cic->ioc;
1283
1284         list_del_init(&cic->queue_list);
1285
1286         /*
1287          * Make sure key == NULL is seen for dead queues
1288          */
1289         smp_wmb();
1290         cic->dead_key = (unsigned long) cic->key;
1291         cic->key = NULL;
1292
1293         if (ioc->ioc_data == cic)
1294                 rcu_assign_pointer(ioc->ioc_data, NULL);
1295
1296         if (cic->cfqq[ASYNC]) {
1297                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1298                 cic->cfqq[ASYNC] = NULL;
1299         }
1300
1301         if (cic->cfqq[SYNC]) {
1302                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1303                 cic->cfqq[SYNC] = NULL;
1304         }
1305 }
1306
1307 static void cfq_exit_single_io_context(struct io_context *ioc,
1308                                        struct cfq_io_context *cic)
1309 {
1310         struct cfq_data *cfqd = cic->key;
1311
1312         if (cfqd) {
1313                 struct request_queue *q = cfqd->queue;
1314                 unsigned long flags;
1315
1316                 spin_lock_irqsave(q->queue_lock, flags);
1317
1318                 /*
1319                  * Ensure we get a fresh copy of the ->key to prevent
1320                  * race between exiting task and queue
1321                  */
1322                 smp_read_barrier_depends();
1323                 if (cic->key)
1324                         __cfq_exit_single_io_context(cfqd, cic);
1325
1326                 spin_unlock_irqrestore(q->queue_lock, flags);
1327         }
1328 }
1329
1330 /*
1331  * The process that ioc belongs to has exited, we need to clean up
1332  * and put the internal structures we have that belongs to that process.
1333  */
1334 static void cfq_exit_io_context(struct io_context *ioc)
1335 {
1336         call_for_each_cic(ioc, cfq_exit_single_io_context);
1337 }
1338
1339 static struct cfq_io_context *
1340 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1341 {
1342         struct cfq_io_context *cic;
1343
1344         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1345                                                         cfqd->queue->node);
1346         if (cic) {
1347                 cic->last_end_request = jiffies;
1348                 INIT_LIST_HEAD(&cic->queue_list);
1349                 INIT_HLIST_NODE(&cic->cic_list);
1350                 cic->dtor = cfq_free_io_context;
1351                 cic->exit = cfq_exit_io_context;
1352                 elv_ioc_count_inc(ioc_count);
1353         }
1354
1355         return cic;
1356 }
1357
1358 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1359 {
1360         struct task_struct *tsk = current;
1361         int ioprio_class;
1362
1363         if (!cfq_cfqq_prio_changed(cfqq))
1364                 return;
1365
1366         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1367         switch (ioprio_class) {
1368         default:
1369                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1370         case IOPRIO_CLASS_NONE:
1371                 /*
1372                  * no prio set, inherit CPU scheduling settings
1373                  */
1374                 cfqq->ioprio = task_nice_ioprio(tsk);
1375                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1376                 break;
1377         case IOPRIO_CLASS_RT:
1378                 cfqq->ioprio = task_ioprio(ioc);
1379                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1380                 break;
1381         case IOPRIO_CLASS_BE:
1382                 cfqq->ioprio = task_ioprio(ioc);
1383                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1384                 break;
1385         case IOPRIO_CLASS_IDLE:
1386                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1387                 cfqq->ioprio = 7;
1388                 cfq_clear_cfqq_idle_window(cfqq);
1389                 break;
1390         }
1391
1392         /*
1393          * keep track of original prio settings in case we have to temporarily
1394          * elevate the priority of this queue
1395          */
1396         cfqq->org_ioprio = cfqq->ioprio;
1397         cfqq->org_ioprio_class = cfqq->ioprio_class;
1398         cfq_clear_cfqq_prio_changed(cfqq);
1399 }
1400
1401 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1402 {
1403         struct cfq_data *cfqd = cic->key;
1404         struct cfq_queue *cfqq;
1405         unsigned long flags;
1406
1407         if (unlikely(!cfqd))
1408                 return;
1409
1410         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1411
1412         cfqq = cic->cfqq[ASYNC];
1413         if (cfqq) {
1414                 struct cfq_queue *new_cfqq;
1415                 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC);
1416                 if (new_cfqq) {
1417                         cic->cfqq[ASYNC] = new_cfqq;
1418                         cfq_put_queue(cfqq);
1419                 }
1420         }
1421
1422         cfqq = cic->cfqq[SYNC];
1423         if (cfqq)
1424                 cfq_mark_cfqq_prio_changed(cfqq);
1425
1426         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1427 }
1428
1429 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1430 {
1431         call_for_each_cic(ioc, changed_ioprio);
1432         ioc->ioprio_changed = 0;
1433 }
1434
1435 static struct cfq_queue *
1436 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1437                      struct io_context *ioc, gfp_t gfp_mask)
1438 {
1439         struct cfq_queue *cfqq, *new_cfqq = NULL;
1440         struct cfq_io_context *cic;
1441
1442 retry:
1443         cic = cfq_cic_lookup(cfqd, ioc);
1444         /* cic always exists here */
1445         cfqq = cic_to_cfqq(cic, is_sync);
1446
1447         if (!cfqq) {
1448                 if (new_cfqq) {
1449                         cfqq = new_cfqq;
1450                         new_cfqq = NULL;
1451                 } else if (gfp_mask & __GFP_WAIT) {
1452                         /*
1453                          * Inform the allocator of the fact that we will
1454                          * just repeat this allocation if it fails, to allow
1455                          * the allocator to do whatever it needs to attempt to
1456                          * free memory.
1457                          */
1458                         spin_unlock_irq(cfqd->queue->queue_lock);
1459                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1460                                         gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1461                                         cfqd->queue->node);
1462                         spin_lock_irq(cfqd->queue->queue_lock);
1463                         goto retry;
1464                 } else {
1465                         cfqq = kmem_cache_alloc_node(cfq_pool,
1466                                         gfp_mask | __GFP_ZERO,
1467                                         cfqd->queue->node);
1468                         if (!cfqq)
1469                                 goto out;
1470                 }
1471
1472                 RB_CLEAR_NODE(&cfqq->rb_node);
1473                 INIT_LIST_HEAD(&cfqq->fifo);
1474
1475                 atomic_set(&cfqq->ref, 0);
1476                 cfqq->cfqd = cfqd;
1477
1478                 cfq_mark_cfqq_prio_changed(cfqq);
1479                 cfq_mark_cfqq_queue_new(cfqq);
1480
1481                 cfq_init_prio_data(cfqq, ioc);
1482
1483                 if (is_sync) {
1484                         if (!cfq_class_idle(cfqq))
1485                                 cfq_mark_cfqq_idle_window(cfqq);
1486                         cfq_mark_cfqq_sync(cfqq);
1487                 }
1488                 cfqq->pid = current->pid;
1489                 cfq_log_cfqq(cfqd, cfqq, "alloced");
1490         }
1491
1492         if (new_cfqq)
1493                 kmem_cache_free(cfq_pool, new_cfqq);
1494
1495 out:
1496         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1497         return cfqq;
1498 }
1499
1500 static struct cfq_queue **
1501 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1502 {
1503         switch (ioprio_class) {
1504         case IOPRIO_CLASS_RT:
1505                 return &cfqd->async_cfqq[0][ioprio];
1506         case IOPRIO_CLASS_BE:
1507                 return &cfqd->async_cfqq[1][ioprio];
1508         case IOPRIO_CLASS_IDLE:
1509                 return &cfqd->async_idle_cfqq;
1510         default:
1511                 BUG();
1512         }
1513 }
1514
1515 static struct cfq_queue *
1516 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1517               gfp_t gfp_mask)
1518 {
1519         const int ioprio = task_ioprio(ioc);
1520         const int ioprio_class = task_ioprio_class(ioc);
1521         struct cfq_queue **async_cfqq = NULL;
1522         struct cfq_queue *cfqq = NULL;
1523
1524         if (!is_sync) {
1525                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1526                 cfqq = *async_cfqq;
1527         }
1528
1529         if (!cfqq) {
1530                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1531                 if (!cfqq)
1532                         return NULL;
1533         }
1534
1535         /*
1536          * pin the queue now that it's allocated, scheduler exit will prune it
1537          */
1538         if (!is_sync && !(*async_cfqq)) {
1539                 atomic_inc(&cfqq->ref);
1540                 *async_cfqq = cfqq;
1541         }
1542
1543         atomic_inc(&cfqq->ref);
1544         return cfqq;
1545 }
1546
1547 /*
1548  * We drop cfq io contexts lazily, so we may find a dead one.
1549  */
1550 static void
1551 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1552                   struct cfq_io_context *cic)
1553 {
1554         unsigned long flags;
1555
1556         WARN_ON(!list_empty(&cic->queue_list));
1557
1558         spin_lock_irqsave(&ioc->lock, flags);
1559
1560         BUG_ON(ioc->ioc_data == cic);
1561
1562         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1563         hlist_del_rcu(&cic->cic_list);
1564         spin_unlock_irqrestore(&ioc->lock, flags);
1565
1566         cfq_cic_free(cic);
1567 }
1568
1569 static struct cfq_io_context *
1570 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1571 {
1572         struct cfq_io_context *cic;
1573         unsigned long flags;
1574         void *k;
1575
1576         if (unlikely(!ioc))
1577                 return NULL;
1578
1579         rcu_read_lock();
1580
1581         /*
1582          * we maintain a last-hit cache, to avoid browsing over the tree
1583          */
1584         cic = rcu_dereference(ioc->ioc_data);
1585         if (cic && cic->key == cfqd) {
1586                 rcu_read_unlock();
1587                 return cic;
1588         }
1589
1590         do {
1591                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1592                 rcu_read_unlock();
1593                 if (!cic)
1594                         break;
1595                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1596                 k = cic->key;
1597                 if (unlikely(!k)) {
1598                         cfq_drop_dead_cic(cfqd, ioc, cic);
1599                         rcu_read_lock();
1600                         continue;
1601                 }
1602
1603                 spin_lock_irqsave(&ioc->lock, flags);
1604                 rcu_assign_pointer(ioc->ioc_data, cic);
1605                 spin_unlock_irqrestore(&ioc->lock, flags);
1606                 break;
1607         } while (1);
1608
1609         return cic;
1610 }
1611
1612 /*
1613  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1614  * the process specific cfq io context when entered from the block layer.
1615  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1616  */
1617 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1618                         struct cfq_io_context *cic, gfp_t gfp_mask)
1619 {
1620         unsigned long flags;
1621         int ret;
1622
1623         ret = radix_tree_preload(gfp_mask);
1624         if (!ret) {
1625                 cic->ioc = ioc;
1626                 cic->key = cfqd;
1627
1628                 spin_lock_irqsave(&ioc->lock, flags);
1629                 ret = radix_tree_insert(&ioc->radix_root,
1630                                                 (unsigned long) cfqd, cic);
1631                 if (!ret)
1632                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1633                 spin_unlock_irqrestore(&ioc->lock, flags);
1634
1635                 radix_tree_preload_end();
1636
1637                 if (!ret) {
1638                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1639                         list_add(&cic->queue_list, &cfqd->cic_list);
1640                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1641                 }
1642         }
1643
1644         if (ret)
1645                 printk(KERN_ERR "cfq: cic link failed!\n");
1646
1647         return ret;
1648 }
1649
1650 /*
1651  * Setup general io context and cfq io context. There can be several cfq
1652  * io contexts per general io context, if this process is doing io to more
1653  * than one device managed by cfq.
1654  */
1655 static struct cfq_io_context *
1656 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1657 {
1658         struct io_context *ioc = NULL;
1659         struct cfq_io_context *cic;
1660
1661         might_sleep_if(gfp_mask & __GFP_WAIT);
1662
1663         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1664         if (!ioc)
1665                 return NULL;
1666
1667         cic = cfq_cic_lookup(cfqd, ioc);
1668         if (cic)
1669                 goto out;
1670
1671         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1672         if (cic == NULL)
1673                 goto err;
1674
1675         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1676                 goto err_free;
1677
1678 out:
1679         smp_read_barrier_depends();
1680         if (unlikely(ioc->ioprio_changed))
1681                 cfq_ioc_set_ioprio(ioc);
1682
1683         return cic;
1684 err_free:
1685         cfq_cic_free(cic);
1686 err:
1687         put_io_context(ioc);
1688         return NULL;
1689 }
1690
1691 static void
1692 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1693 {
1694         unsigned long elapsed = jiffies - cic->last_end_request;
1695         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1696
1697         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1698         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1699         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1700 }
1701
1702 static void
1703 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1704                        struct request *rq)
1705 {
1706         sector_t sdist;
1707         u64 total;
1708
1709         if (cic->last_request_pos < rq->sector)
1710                 sdist = rq->sector - cic->last_request_pos;
1711         else
1712                 sdist = cic->last_request_pos - rq->sector;
1713
1714         /*
1715          * Don't allow the seek distance to get too large from the
1716          * odd fragment, pagein, etc
1717          */
1718         if (cic->seek_samples <= 60) /* second&third seek */
1719                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1720         else
1721                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1722
1723         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1724         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1725         total = cic->seek_total + (cic->seek_samples/2);
1726         do_div(total, cic->seek_samples);
1727         cic->seek_mean = (sector_t)total;
1728 }
1729
1730 /*
1731  * Disable idle window if the process thinks too long or seeks so much that
1732  * it doesn't matter
1733  */
1734 static void
1735 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1736                        struct cfq_io_context *cic)
1737 {
1738         int old_idle, enable_idle;
1739
1740         /*
1741          * Don't idle for async or idle io prio class
1742          */
1743         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1744                 return;
1745
1746         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1747
1748         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1749             (cfqd->hw_tag && CIC_SEEKY(cic)))
1750                 enable_idle = 0;
1751         else if (sample_valid(cic->ttime_samples)) {
1752                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1753                         enable_idle = 0;
1754                 else
1755                         enable_idle = 1;
1756         }
1757
1758         if (old_idle != enable_idle) {
1759                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1760                 if (enable_idle)
1761                         cfq_mark_cfqq_idle_window(cfqq);
1762                 else
1763                         cfq_clear_cfqq_idle_window(cfqq);
1764         }
1765 }
1766
1767 /*
1768  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1769  * no or if we aren't sure, a 1 will cause a preempt.
1770  */
1771 static int
1772 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1773                    struct request *rq)
1774 {
1775         struct cfq_queue *cfqq;
1776
1777         cfqq = cfqd->active_queue;
1778         if (!cfqq)
1779                 return 0;
1780
1781         if (cfq_slice_used(cfqq))
1782                 return 1;
1783
1784         if (cfq_class_idle(new_cfqq))
1785                 return 0;
1786
1787         if (cfq_class_idle(cfqq))
1788                 return 1;
1789
1790         /*
1791          * if the new request is sync, but the currently running queue is
1792          * not, let the sync request have priority.
1793          */
1794         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1795                 return 1;
1796
1797         /*
1798          * So both queues are sync. Let the new request get disk time if
1799          * it's a metadata request and the current queue is doing regular IO.
1800          */
1801         if (rq_is_meta(rq) && !cfqq->meta_pending)
1802                 return 1;
1803
1804         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1805                 return 0;
1806
1807         /*
1808          * if this request is as-good as one we would expect from the
1809          * current cfqq, let it preempt
1810          */
1811         if (cfq_rq_close(cfqd, rq))
1812                 return 1;
1813
1814         return 0;
1815 }
1816
1817 /*
1818  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1819  * let it have half of its nominal slice.
1820  */
1821 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1822 {
1823         cfq_log_cfqq(cfqd, cfqq, "preempt");
1824         cfq_slice_expired(cfqd, 1);
1825
1826         /*
1827          * Put the new queue at the front of the of the current list,
1828          * so we know that it will be selected next.
1829          */
1830         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1831
1832         cfq_service_tree_add(cfqd, cfqq, 1);
1833
1834         cfqq->slice_end = 0;
1835         cfq_mark_cfqq_slice_new(cfqq);
1836 }
1837
1838 /*
1839  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1840  * something we should do about it
1841  */
1842 static void
1843 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1844                 struct request *rq)
1845 {
1846         struct cfq_io_context *cic = RQ_CIC(rq);
1847
1848         cfqd->rq_queued++;
1849         if (rq_is_meta(rq))
1850                 cfqq->meta_pending++;
1851
1852         cfq_update_io_thinktime(cfqd, cic);
1853         cfq_update_io_seektime(cfqd, cic, rq);
1854         cfq_update_idle_window(cfqd, cfqq, cic);
1855
1856         cic->last_request_pos = rq->sector + rq->nr_sectors;
1857
1858         if (cfqq == cfqd->active_queue) {
1859                 /*
1860                  * if we are waiting for a request for this queue, let it rip
1861                  * immediately and flag that we must not expire this queue
1862                  * just now
1863                  */
1864                 if (cfq_cfqq_wait_request(cfqq)) {
1865                         cfq_mark_cfqq_must_dispatch(cfqq);
1866                         del_timer(&cfqd->idle_slice_timer);
1867                         blk_start_queueing(cfqd->queue);
1868                 }
1869         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1870                 /*
1871                  * not the active queue - expire current slice if it is
1872                  * idle and has expired it's mean thinktime or this new queue
1873                  * has some old slice time left and is of higher priority
1874                  */
1875                 cfq_preempt_queue(cfqd, cfqq);
1876                 cfq_mark_cfqq_must_dispatch(cfqq);
1877                 blk_start_queueing(cfqd->queue);
1878         }
1879 }
1880
1881 static void cfq_insert_request(struct request_queue *q, struct request *rq)
1882 {
1883         struct cfq_data *cfqd = q->elevator->elevator_data;
1884         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1885
1886         cfq_log_cfqq(cfqd, cfqq, "insert_request");
1887         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
1888
1889         cfq_add_rq_rb(rq);
1890
1891         list_add_tail(&rq->queuelist, &cfqq->fifo);
1892
1893         cfq_rq_enqueued(cfqd, cfqq, rq);
1894 }
1895
1896 /*
1897  * Update hw_tag based on peak queue depth over 50 samples under
1898  * sufficient load.
1899  */
1900 static void cfq_update_hw_tag(struct cfq_data *cfqd)
1901 {
1902         if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
1903                 cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
1904
1905         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
1906             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
1907                 return;
1908
1909         if (cfqd->hw_tag_samples++ < 50)
1910                 return;
1911
1912         if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
1913                 cfqd->hw_tag = 1;
1914         else
1915                 cfqd->hw_tag = 0;
1916
1917         cfqd->hw_tag_samples = 0;
1918         cfqd->rq_in_driver_peak = 0;
1919 }
1920
1921 static void cfq_completed_request(struct request_queue *q, struct request *rq)
1922 {
1923         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1924         struct cfq_data *cfqd = cfqq->cfqd;
1925         const int sync = rq_is_sync(rq);
1926         unsigned long now;
1927
1928         now = jiffies;
1929         cfq_log_cfqq(cfqd, cfqq, "complete");
1930
1931         cfq_update_hw_tag(cfqd);
1932
1933         WARN_ON(!cfqd->rq_in_driver);
1934         WARN_ON(!cfqq->dispatched);
1935         cfqd->rq_in_driver--;
1936         cfqq->dispatched--;
1937
1938         if (cfq_cfqq_sync(cfqq))
1939                 cfqd->sync_flight--;
1940
1941         if (!cfq_class_idle(cfqq))
1942                 cfqd->last_end_request = now;
1943
1944         if (sync)
1945                 RQ_CIC(rq)->last_end_request = now;
1946
1947         /*
1948          * If this is the active queue, check if it needs to be expired,
1949          * or if we want to idle in case it has no pending requests.
1950          */
1951         if (cfqd->active_queue == cfqq) {
1952                 if (cfq_cfqq_slice_new(cfqq)) {
1953                         cfq_set_prio_slice(cfqd, cfqq);
1954                         cfq_clear_cfqq_slice_new(cfqq);
1955                 }
1956                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
1957                         cfq_slice_expired(cfqd, 1);
1958                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1959                         cfq_arm_slice_timer(cfqd);
1960         }
1961
1962         if (!cfqd->rq_in_driver)
1963                 cfq_schedule_dispatch(cfqd);
1964 }
1965
1966 /*
1967  * we temporarily boost lower priority queues if they are holding fs exclusive
1968  * resources. they are boosted to normal prio (CLASS_BE/4)
1969  */
1970 static void cfq_prio_boost(struct cfq_queue *cfqq)
1971 {
1972         if (has_fs_excl()) {
1973                 /*
1974                  * boost idle prio on transactions that would lock out other
1975                  * users of the filesystem
1976                  */
1977                 if (cfq_class_idle(cfqq))
1978                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1979                 if (cfqq->ioprio > IOPRIO_NORM)
1980                         cfqq->ioprio = IOPRIO_NORM;
1981         } else {
1982                 /*
1983                  * check if we need to unboost the queue
1984                  */
1985                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1986                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1987                 if (cfqq->ioprio != cfqq->org_ioprio)
1988                         cfqq->ioprio = cfqq->org_ioprio;
1989         }
1990 }
1991
1992 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1993 {
1994         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1995             !cfq_cfqq_must_alloc_slice(cfqq)) {
1996                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1997                 return ELV_MQUEUE_MUST;
1998         }
1999
2000         return ELV_MQUEUE_MAY;
2001 }
2002
2003 static int cfq_may_queue(struct request_queue *q, int rw)
2004 {
2005         struct cfq_data *cfqd = q->elevator->elevator_data;
2006         struct task_struct *tsk = current;
2007         struct cfq_io_context *cic;
2008         struct cfq_queue *cfqq;
2009
2010         /*
2011          * don't force setup of a queue from here, as a call to may_queue
2012          * does not necessarily imply that a request actually will be queued.
2013          * so just lookup a possibly existing queue, or return 'may queue'
2014          * if that fails
2015          */
2016         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2017         if (!cic)
2018                 return ELV_MQUEUE_MAY;
2019
2020         cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
2021         if (cfqq) {
2022                 cfq_init_prio_data(cfqq, cic->ioc);
2023                 cfq_prio_boost(cfqq);
2024
2025                 return __cfq_may_queue(cfqq);
2026         }
2027
2028         return ELV_MQUEUE_MAY;
2029 }
2030
2031 /*
2032  * queue lock held here
2033  */
2034 static void cfq_put_request(struct request *rq)
2035 {
2036         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2037
2038         if (cfqq) {
2039                 const int rw = rq_data_dir(rq);
2040
2041                 BUG_ON(!cfqq->allocated[rw]);
2042                 cfqq->allocated[rw]--;
2043
2044                 put_io_context(RQ_CIC(rq)->ioc);
2045
2046                 rq->elevator_private = NULL;
2047                 rq->elevator_private2 = NULL;
2048
2049                 cfq_put_queue(cfqq);
2050         }
2051 }
2052
2053 /*
2054  * Allocate cfq data structures associated with this request.
2055  */
2056 static int
2057 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2058 {
2059         struct cfq_data *cfqd = q->elevator->elevator_data;
2060         struct cfq_io_context *cic;
2061         const int rw = rq_data_dir(rq);
2062         const int is_sync = rq_is_sync(rq);
2063         struct cfq_queue *cfqq;
2064         unsigned long flags;
2065
2066         might_sleep_if(gfp_mask & __GFP_WAIT);
2067
2068         cic = cfq_get_io_context(cfqd, gfp_mask);
2069
2070         spin_lock_irqsave(q->queue_lock, flags);
2071
2072         if (!cic)
2073                 goto queue_fail;
2074
2075         cfqq = cic_to_cfqq(cic, is_sync);
2076         if (!cfqq) {
2077                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2078
2079                 if (!cfqq)
2080                         goto queue_fail;
2081
2082                 cic_set_cfqq(cic, cfqq, is_sync);
2083         }
2084
2085         cfqq->allocated[rw]++;
2086         cfq_clear_cfqq_must_alloc(cfqq);
2087         atomic_inc(&cfqq->ref);
2088
2089         spin_unlock_irqrestore(q->queue_lock, flags);
2090
2091         rq->elevator_private = cic;
2092         rq->elevator_private2 = cfqq;
2093         return 0;
2094
2095 queue_fail:
2096         if (cic)
2097                 put_io_context(cic->ioc);
2098
2099         cfq_schedule_dispatch(cfqd);
2100         spin_unlock_irqrestore(q->queue_lock, flags);
2101         cfq_log(cfqd, "set_request fail");
2102         return 1;
2103 }
2104
2105 static void cfq_kick_queue(struct work_struct *work)
2106 {
2107         struct cfq_data *cfqd =
2108                 container_of(work, struct cfq_data, unplug_work);
2109         struct request_queue *q = cfqd->queue;
2110         unsigned long flags;
2111
2112         spin_lock_irqsave(q->queue_lock, flags);
2113         blk_start_queueing(q);
2114         spin_unlock_irqrestore(q->queue_lock, flags);
2115 }
2116
2117 /*
2118  * Timer running if the active_queue is currently idling inside its time slice
2119  */
2120 static void cfq_idle_slice_timer(unsigned long data)
2121 {
2122         struct cfq_data *cfqd = (struct cfq_data *) data;
2123         struct cfq_queue *cfqq;
2124         unsigned long flags;
2125         int timed_out = 1;
2126
2127         cfq_log(cfqd, "idle timer fired");
2128
2129         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2130
2131         cfqq = cfqd->active_queue;
2132         if (cfqq) {
2133                 timed_out = 0;
2134
2135                 /*
2136                  * expired
2137                  */
2138                 if (cfq_slice_used(cfqq))
2139                         goto expire;
2140
2141                 /*
2142                  * only expire and reinvoke request handler, if there are
2143                  * other queues with pending requests
2144                  */
2145                 if (!cfqd->busy_queues)
2146                         goto out_cont;
2147
2148                 /*
2149                  * not expired and it has a request pending, let it dispatch
2150                  */
2151                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
2152                         cfq_mark_cfqq_must_dispatch(cfqq);
2153                         goto out_kick;
2154                 }
2155         }
2156 expire:
2157         cfq_slice_expired(cfqd, timed_out);
2158 out_kick:
2159         cfq_schedule_dispatch(cfqd);
2160 out_cont:
2161         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2162 }
2163
2164 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2165 {
2166         del_timer_sync(&cfqd->idle_slice_timer);
2167         cancel_work_sync(&cfqd->unplug_work);
2168 }
2169
2170 static void cfq_put_async_queues(struct cfq_data *cfqd)
2171 {
2172         int i;
2173
2174         for (i = 0; i < IOPRIO_BE_NR; i++) {
2175                 if (cfqd->async_cfqq[0][i])
2176                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2177                 if (cfqd->async_cfqq[1][i])
2178                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2179         }
2180
2181         if (cfqd->async_idle_cfqq)
2182                 cfq_put_queue(cfqd->async_idle_cfqq);
2183 }
2184
2185 static void cfq_exit_queue(struct elevator_queue *e)
2186 {
2187         struct cfq_data *cfqd = e->elevator_data;
2188         struct request_queue *q = cfqd->queue;
2189
2190         cfq_shutdown_timer_wq(cfqd);
2191
2192         spin_lock_irq(q->queue_lock);
2193
2194         if (cfqd->active_queue)
2195                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2196
2197         while (!list_empty(&cfqd->cic_list)) {
2198                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2199                                                         struct cfq_io_context,
2200                                                         queue_list);
2201
2202                 __cfq_exit_single_io_context(cfqd, cic);
2203         }
2204
2205         cfq_put_async_queues(cfqd);
2206
2207         spin_unlock_irq(q->queue_lock);
2208
2209         cfq_shutdown_timer_wq(cfqd);
2210
2211         kfree(cfqd);
2212 }
2213
2214 static void *cfq_init_queue(struct request_queue *q)
2215 {
2216         struct cfq_data *cfqd;
2217
2218         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2219         if (!cfqd)
2220                 return NULL;
2221
2222         cfqd->service_tree = CFQ_RB_ROOT;
2223         INIT_LIST_HEAD(&cfqd->cic_list);
2224
2225         cfqd->queue = q;
2226
2227         init_timer(&cfqd->idle_slice_timer);
2228         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2229         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2230
2231         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2232
2233         cfqd->last_end_request = jiffies;
2234         cfqd->cfq_quantum = cfq_quantum;
2235         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2236         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2237         cfqd->cfq_back_max = cfq_back_max;
2238         cfqd->cfq_back_penalty = cfq_back_penalty;
2239         cfqd->cfq_slice[0] = cfq_slice_async;
2240         cfqd->cfq_slice[1] = cfq_slice_sync;
2241         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2242         cfqd->cfq_slice_idle = cfq_slice_idle;
2243         cfqd->hw_tag = 1;
2244
2245         return cfqd;
2246 }
2247
2248 static void cfq_slab_kill(void)
2249 {
2250         /*
2251          * Caller already ensured that pending RCU callbacks are completed,
2252          * so we should have no busy allocations at this point.
2253          */
2254         if (cfq_pool)
2255                 kmem_cache_destroy(cfq_pool);
2256         if (cfq_ioc_pool)
2257                 kmem_cache_destroy(cfq_ioc_pool);
2258 }
2259
2260 static int __init cfq_slab_setup(void)
2261 {
2262         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2263         if (!cfq_pool)
2264                 goto fail;
2265
2266         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2267         if (!cfq_ioc_pool)
2268                 goto fail;
2269
2270         return 0;
2271 fail:
2272         cfq_slab_kill();
2273         return -ENOMEM;
2274 }
2275
2276 /*
2277  * sysfs parts below -->
2278  */
2279 static ssize_t
2280 cfq_var_show(unsigned int var, char *page)
2281 {
2282         return sprintf(page, "%d\n", var);
2283 }
2284
2285 static ssize_t
2286 cfq_var_store(unsigned int *var, const char *page, size_t count)
2287 {
2288         char *p = (char *) page;
2289
2290         *var = simple_strtoul(p, &p, 10);
2291         return count;
2292 }
2293
2294 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2295 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
2296 {                                                                       \
2297         struct cfq_data *cfqd = e->elevator_data;                       \
2298         unsigned int __data = __VAR;                                    \
2299         if (__CONV)                                                     \
2300                 __data = jiffies_to_msecs(__data);                      \
2301         return cfq_var_show(__data, (page));                            \
2302 }
2303 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2304 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2305 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2306 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2307 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2308 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2309 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2310 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2311 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2312 #undef SHOW_FUNCTION
2313
2314 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2315 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2316 {                                                                       \
2317         struct cfq_data *cfqd = e->elevator_data;                       \
2318         unsigned int __data;                                            \
2319         int ret = cfq_var_store(&__data, (page), count);                \
2320         if (__data < (MIN))                                             \
2321                 __data = (MIN);                                         \
2322         else if (__data > (MAX))                                        \
2323                 __data = (MAX);                                         \
2324         if (__CONV)                                                     \
2325                 *(__PTR) = msecs_to_jiffies(__data);                    \
2326         else                                                            \
2327                 *(__PTR) = __data;                                      \
2328         return ret;                                                     \
2329 }
2330 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2331 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2332                 UINT_MAX, 1);
2333 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2334                 UINT_MAX, 1);
2335 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2336 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2337                 UINT_MAX, 0);
2338 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2339 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2340 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2341 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2342                 UINT_MAX, 0);
2343 #undef STORE_FUNCTION
2344
2345 #define CFQ_ATTR(name) \
2346         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2347
2348 static struct elv_fs_entry cfq_attrs[] = {
2349         CFQ_ATTR(quantum),
2350         CFQ_ATTR(fifo_expire_sync),
2351         CFQ_ATTR(fifo_expire_async),
2352         CFQ_ATTR(back_seek_max),
2353         CFQ_ATTR(back_seek_penalty),
2354         CFQ_ATTR(slice_sync),
2355         CFQ_ATTR(slice_async),
2356         CFQ_ATTR(slice_async_rq),
2357         CFQ_ATTR(slice_idle),
2358         __ATTR_NULL
2359 };
2360
2361 static struct elevator_type iosched_cfq = {
2362         .ops = {
2363                 .elevator_merge_fn =            cfq_merge,
2364                 .elevator_merged_fn =           cfq_merged_request,
2365                 .elevator_merge_req_fn =        cfq_merged_requests,
2366                 .elevator_allow_merge_fn =      cfq_allow_merge,
2367                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2368                 .elevator_add_req_fn =          cfq_insert_request,
2369                 .elevator_activate_req_fn =     cfq_activate_request,
2370                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2371                 .elevator_queue_empty_fn =      cfq_queue_empty,
2372                 .elevator_completed_req_fn =    cfq_completed_request,
2373                 .elevator_former_req_fn =       elv_rb_former_request,
2374                 .elevator_latter_req_fn =       elv_rb_latter_request,
2375                 .elevator_set_req_fn =          cfq_set_request,
2376                 .elevator_put_req_fn =          cfq_put_request,
2377                 .elevator_may_queue_fn =        cfq_may_queue,
2378                 .elevator_init_fn =             cfq_init_queue,
2379                 .elevator_exit_fn =             cfq_exit_queue,
2380                 .trim =                         cfq_free_io_context,
2381         },
2382         .elevator_attrs =       cfq_attrs,
2383         .elevator_name =        "cfq",
2384         .elevator_owner =       THIS_MODULE,
2385 };
2386
2387 static int __init cfq_init(void)
2388 {
2389         /*
2390          * could be 0 on HZ < 1000 setups
2391          */
2392         if (!cfq_slice_async)
2393                 cfq_slice_async = 1;
2394         if (!cfq_slice_idle)
2395                 cfq_slice_idle = 1;
2396
2397         if (cfq_slab_setup())
2398                 return -ENOMEM;
2399
2400         elv_register(&iosched_cfq);
2401
2402         return 0;
2403 }
2404
2405 static void __exit cfq_exit(void)
2406 {
2407         DECLARE_COMPLETION_ONSTACK(all_gone);
2408         elv_unregister(&iosched_cfq);
2409         ioc_gone = &all_gone;
2410         /* ioc_gone's update must be visible before reading ioc_count */
2411         smp_wmb();
2412
2413         /*
2414          * this also protects us from entering cfq_slab_kill() with
2415          * pending RCU callbacks
2416          */
2417         if (elv_ioc_count_read(ioc_count))
2418                 wait_for_completion(&all_gone);
2419         cfq_slab_kill();
2420 }
2421
2422 module_init(cfq_init);
2423 module_exit(cfq_exit);
2424
2425 MODULE_AUTHOR("Jens Axboe");
2426 MODULE_LICENSE("GPL");
2427 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");