sched: Re-add lost cpu_allowed check to sched_fair.c::select_task_rq_fair()
[linux-2.6.git] / kernel / sched_fair.c
1 /*
2  * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
3  *
4  *  Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
5  *
6  *  Interactivity improvements by Mike Galbraith
7  *  (C) 2007 Mike Galbraith <efault@gmx.de>
8  *
9  *  Various enhancements by Dmitry Adamushko.
10  *  (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
11  *
12  *  Group scheduling enhancements by Srivatsa Vaddagiri
13  *  Copyright IBM Corporation, 2007
14  *  Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
15  *
16  *  Scaled math optimizations by Thomas Gleixner
17  *  Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
18  *
19  *  Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
21  */
22
23 #include <linux/latencytop.h>
24
25 /*
26  * Targeted preemption latency for CPU-bound tasks:
27  * (default: 5ms * (1 + ilog(ncpus)), units: nanoseconds)
28  *
29  * NOTE: this latency value is not the same as the concept of
30  * 'timeslice length' - timeslices in CFS are of variable length
31  * and have no persistent notion like in traditional, time-slice
32  * based scheduling concepts.
33  *
34  * (to see the precise effective timeslice length of your workload,
35  *  run vmstat and monitor the context-switches (cs) field)
36  */
37 unsigned int sysctl_sched_latency = 5000000ULL;
38
39 /*
40  * Minimal preemption granularity for CPU-bound tasks:
41  * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
42  */
43 unsigned int sysctl_sched_min_granularity = 1000000ULL;
44
45 /*
46  * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
47  */
48 static unsigned int sched_nr_latency = 5;
49
50 /*
51  * After fork, child runs first. If set to 0 (default) then
52  * parent will (try to) run first.
53  */
54 unsigned int sysctl_sched_child_runs_first __read_mostly;
55
56 /*
57  * sys_sched_yield() compat mode
58  *
59  * This option switches the agressive yield implementation of the
60  * old scheduler back on.
61  */
62 unsigned int __read_mostly sysctl_sched_compat_yield;
63
64 /*
65  * SCHED_OTHER wake-up granularity.
66  * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
67  *
68  * This option delays the preemption effects of decoupled workloads
69  * and reduces their over-scheduling. Synchronous workloads will still
70  * have immediate wakeup/sleep latencies.
71  */
72 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
73
74 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
75
76 static const struct sched_class fair_sched_class;
77
78 /**************************************************************
79  * CFS operations on generic schedulable entities:
80  */
81
82 #ifdef CONFIG_FAIR_GROUP_SCHED
83
84 /* cpu runqueue to which this cfs_rq is attached */
85 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
86 {
87         return cfs_rq->rq;
88 }
89
90 /* An entity is a task if it doesn't "own" a runqueue */
91 #define entity_is_task(se)      (!se->my_q)
92
93 static inline struct task_struct *task_of(struct sched_entity *se)
94 {
95 #ifdef CONFIG_SCHED_DEBUG
96         WARN_ON_ONCE(!entity_is_task(se));
97 #endif
98         return container_of(se, struct task_struct, se);
99 }
100
101 /* Walk up scheduling entities hierarchy */
102 #define for_each_sched_entity(se) \
103                 for (; se; se = se->parent)
104
105 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
106 {
107         return p->se.cfs_rq;
108 }
109
110 /* runqueue on which this entity is (to be) queued */
111 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
112 {
113         return se->cfs_rq;
114 }
115
116 /* runqueue "owned" by this group */
117 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
118 {
119         return grp->my_q;
120 }
121
122 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
123  * another cpu ('this_cpu')
124  */
125 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
126 {
127         return cfs_rq->tg->cfs_rq[this_cpu];
128 }
129
130 /* Iterate thr' all leaf cfs_rq's on a runqueue */
131 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
132         list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
133
134 /* Do the two (enqueued) entities belong to the same group ? */
135 static inline int
136 is_same_group(struct sched_entity *se, struct sched_entity *pse)
137 {
138         if (se->cfs_rq == pse->cfs_rq)
139                 return 1;
140
141         return 0;
142 }
143
144 static inline struct sched_entity *parent_entity(struct sched_entity *se)
145 {
146         return se->parent;
147 }
148
149 /* return depth at which a sched entity is present in the hierarchy */
150 static inline int depth_se(struct sched_entity *se)
151 {
152         int depth = 0;
153
154         for_each_sched_entity(se)
155                 depth++;
156
157         return depth;
158 }
159
160 static void
161 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
162 {
163         int se_depth, pse_depth;
164
165         /*
166          * preemption test can be made between sibling entities who are in the
167          * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
168          * both tasks until we find their ancestors who are siblings of common
169          * parent.
170          */
171
172         /* First walk up until both entities are at same depth */
173         se_depth = depth_se(*se);
174         pse_depth = depth_se(*pse);
175
176         while (se_depth > pse_depth) {
177                 se_depth--;
178                 *se = parent_entity(*se);
179         }
180
181         while (pse_depth > se_depth) {
182                 pse_depth--;
183                 *pse = parent_entity(*pse);
184         }
185
186         while (!is_same_group(*se, *pse)) {
187                 *se = parent_entity(*se);
188                 *pse = parent_entity(*pse);
189         }
190 }
191
192 #else   /* !CONFIG_FAIR_GROUP_SCHED */
193
194 static inline struct task_struct *task_of(struct sched_entity *se)
195 {
196         return container_of(se, struct task_struct, se);
197 }
198
199 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
200 {
201         return container_of(cfs_rq, struct rq, cfs);
202 }
203
204 #define entity_is_task(se)      1
205
206 #define for_each_sched_entity(se) \
207                 for (; se; se = NULL)
208
209 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
210 {
211         return &task_rq(p)->cfs;
212 }
213
214 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
215 {
216         struct task_struct *p = task_of(se);
217         struct rq *rq = task_rq(p);
218
219         return &rq->cfs;
220 }
221
222 /* runqueue "owned" by this group */
223 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
224 {
225         return NULL;
226 }
227
228 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
229 {
230         return &cpu_rq(this_cpu)->cfs;
231 }
232
233 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
234                 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
235
236 static inline int
237 is_same_group(struct sched_entity *se, struct sched_entity *pse)
238 {
239         return 1;
240 }
241
242 static inline struct sched_entity *parent_entity(struct sched_entity *se)
243 {
244         return NULL;
245 }
246
247 static inline void
248 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
249 {
250 }
251
252 #endif  /* CONFIG_FAIR_GROUP_SCHED */
253
254
255 /**************************************************************
256  * Scheduling class tree data structure manipulation methods:
257  */
258
259 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
260 {
261         s64 delta = (s64)(vruntime - min_vruntime);
262         if (delta > 0)
263                 min_vruntime = vruntime;
264
265         return min_vruntime;
266 }
267
268 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
269 {
270         s64 delta = (s64)(vruntime - min_vruntime);
271         if (delta < 0)
272                 min_vruntime = vruntime;
273
274         return min_vruntime;
275 }
276
277 static inline int entity_before(struct sched_entity *a,
278                                 struct sched_entity *b)
279 {
280         return (s64)(a->vruntime - b->vruntime) < 0;
281 }
282
283 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
284 {
285         return se->vruntime - cfs_rq->min_vruntime;
286 }
287
288 static void update_min_vruntime(struct cfs_rq *cfs_rq)
289 {
290         u64 vruntime = cfs_rq->min_vruntime;
291
292         if (cfs_rq->curr)
293                 vruntime = cfs_rq->curr->vruntime;
294
295         if (cfs_rq->rb_leftmost) {
296                 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
297                                                    struct sched_entity,
298                                                    run_node);
299
300                 if (!cfs_rq->curr)
301                         vruntime = se->vruntime;
302                 else
303                         vruntime = min_vruntime(vruntime, se->vruntime);
304         }
305
306         cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
307 }
308
309 /*
310  * Enqueue an entity into the rb-tree:
311  */
312 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
313 {
314         struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
315         struct rb_node *parent = NULL;
316         struct sched_entity *entry;
317         s64 key = entity_key(cfs_rq, se);
318         int leftmost = 1;
319
320         /*
321          * Find the right place in the rbtree:
322          */
323         while (*link) {
324                 parent = *link;
325                 entry = rb_entry(parent, struct sched_entity, run_node);
326                 /*
327                  * We dont care about collisions. Nodes with
328                  * the same key stay together.
329                  */
330                 if (key < entity_key(cfs_rq, entry)) {
331                         link = &parent->rb_left;
332                 } else {
333                         link = &parent->rb_right;
334                         leftmost = 0;
335                 }
336         }
337
338         /*
339          * Maintain a cache of leftmost tree entries (it is frequently
340          * used):
341          */
342         if (leftmost)
343                 cfs_rq->rb_leftmost = &se->run_node;
344
345         rb_link_node(&se->run_node, parent, link);
346         rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
347 }
348
349 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
350 {
351         if (cfs_rq->rb_leftmost == &se->run_node) {
352                 struct rb_node *next_node;
353
354                 next_node = rb_next(&se->run_node);
355                 cfs_rq->rb_leftmost = next_node;
356         }
357
358         rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
359 }
360
361 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
362 {
363         struct rb_node *left = cfs_rq->rb_leftmost;
364
365         if (!left)
366                 return NULL;
367
368         return rb_entry(left, struct sched_entity, run_node);
369 }
370
371 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
372 {
373         struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
374
375         if (!last)
376                 return NULL;
377
378         return rb_entry(last, struct sched_entity, run_node);
379 }
380
381 /**************************************************************
382  * Scheduling class statistics methods:
383  */
384
385 #ifdef CONFIG_SCHED_DEBUG
386 int sched_nr_latency_handler(struct ctl_table *table, int write,
387                 struct file *filp, void __user *buffer, size_t *lenp,
388                 loff_t *ppos)
389 {
390         int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
391
392         if (ret || !write)
393                 return ret;
394
395         sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
396                                         sysctl_sched_min_granularity);
397
398         return 0;
399 }
400 #endif
401
402 /*
403  * delta /= w
404  */
405 static inline unsigned long
406 calc_delta_fair(unsigned long delta, struct sched_entity *se)
407 {
408         if (unlikely(se->load.weight != NICE_0_LOAD))
409                 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
410
411         return delta;
412 }
413
414 /*
415  * The idea is to set a period in which each task runs once.
416  *
417  * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
418  * this period because otherwise the slices get too small.
419  *
420  * p = (nr <= nl) ? l : l*nr/nl
421  */
422 static u64 __sched_period(unsigned long nr_running)
423 {
424         u64 period = sysctl_sched_latency;
425         unsigned long nr_latency = sched_nr_latency;
426
427         if (unlikely(nr_running > nr_latency)) {
428                 period = sysctl_sched_min_granularity;
429                 period *= nr_running;
430         }
431
432         return period;
433 }
434
435 /*
436  * We calculate the wall-time slice from the period by taking a part
437  * proportional to the weight.
438  *
439  * s = p*P[w/rw]
440  */
441 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
442 {
443         u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
444
445         for_each_sched_entity(se) {
446                 struct load_weight *load;
447                 struct load_weight lw;
448
449                 cfs_rq = cfs_rq_of(se);
450                 load = &cfs_rq->load;
451
452                 if (unlikely(!se->on_rq)) {
453                         lw = cfs_rq->load;
454
455                         update_load_add(&lw, se->load.weight);
456                         load = &lw;
457                 }
458                 slice = calc_delta_mine(slice, se->load.weight, load);
459         }
460         return slice;
461 }
462
463 /*
464  * We calculate the vruntime slice of a to be inserted task
465  *
466  * vs = s/w
467  */
468 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
469 {
470         return calc_delta_fair(sched_slice(cfs_rq, se), se);
471 }
472
473 /*
474  * Update the current task's runtime statistics. Skip current tasks that
475  * are not in our scheduling class.
476  */
477 static inline void
478 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
479               unsigned long delta_exec)
480 {
481         unsigned long delta_exec_weighted;
482
483         schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
484
485         curr->sum_exec_runtime += delta_exec;
486         schedstat_add(cfs_rq, exec_clock, delta_exec);
487         delta_exec_weighted = calc_delta_fair(delta_exec, curr);
488         curr->vruntime += delta_exec_weighted;
489         update_min_vruntime(cfs_rq);
490 }
491
492 static void update_curr(struct cfs_rq *cfs_rq)
493 {
494         struct sched_entity *curr = cfs_rq->curr;
495         u64 now = rq_of(cfs_rq)->clock;
496         unsigned long delta_exec;
497
498         if (unlikely(!curr))
499                 return;
500
501         /*
502          * Get the amount of time the current task was running
503          * since the last time we changed load (this cannot
504          * overflow on 32 bits):
505          */
506         delta_exec = (unsigned long)(now - curr->exec_start);
507         if (!delta_exec)
508                 return;
509
510         __update_curr(cfs_rq, curr, delta_exec);
511         curr->exec_start = now;
512
513         if (entity_is_task(curr)) {
514                 struct task_struct *curtask = task_of(curr);
515
516                 cpuacct_charge(curtask, delta_exec);
517                 account_group_exec_runtime(curtask, delta_exec);
518         }
519 }
520
521 static inline void
522 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
523 {
524         schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
525 }
526
527 /*
528  * Task is being enqueued - update stats:
529  */
530 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
531 {
532         /*
533          * Are we enqueueing a waiting task? (for current tasks
534          * a dequeue/enqueue event is a NOP)
535          */
536         if (se != cfs_rq->curr)
537                 update_stats_wait_start(cfs_rq, se);
538 }
539
540 static void
541 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
542 {
543         schedstat_set(se->wait_max, max(se->wait_max,
544                         rq_of(cfs_rq)->clock - se->wait_start));
545         schedstat_set(se->wait_count, se->wait_count + 1);
546         schedstat_set(se->wait_sum, se->wait_sum +
547                         rq_of(cfs_rq)->clock - se->wait_start);
548 #ifdef CONFIG_SCHEDSTATS
549         if (entity_is_task(se)) {
550                 trace_sched_stat_wait(task_of(se),
551                         rq_of(cfs_rq)->clock - se->wait_start);
552         }
553 #endif
554         schedstat_set(se->wait_start, 0);
555 }
556
557 static inline void
558 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
559 {
560         /*
561          * Mark the end of the wait period if dequeueing a
562          * waiting task:
563          */
564         if (se != cfs_rq->curr)
565                 update_stats_wait_end(cfs_rq, se);
566 }
567
568 /*
569  * We are picking a new current task - update its stats:
570  */
571 static inline void
572 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
573 {
574         /*
575          * We are starting a new run period:
576          */
577         se->exec_start = rq_of(cfs_rq)->clock;
578 }
579
580 /**************************************************
581  * Scheduling class queueing methods:
582  */
583
584 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
585 static void
586 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
587 {
588         cfs_rq->task_weight += weight;
589 }
590 #else
591 static inline void
592 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
593 {
594 }
595 #endif
596
597 static void
598 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
599 {
600         update_load_add(&cfs_rq->load, se->load.weight);
601         if (!parent_entity(se))
602                 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
603         if (entity_is_task(se)) {
604                 add_cfs_task_weight(cfs_rq, se->load.weight);
605                 list_add(&se->group_node, &cfs_rq->tasks);
606         }
607         cfs_rq->nr_running++;
608         se->on_rq = 1;
609 }
610
611 static void
612 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
613 {
614         update_load_sub(&cfs_rq->load, se->load.weight);
615         if (!parent_entity(se))
616                 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
617         if (entity_is_task(se)) {
618                 add_cfs_task_weight(cfs_rq, -se->load.weight);
619                 list_del_init(&se->group_node);
620         }
621         cfs_rq->nr_running--;
622         se->on_rq = 0;
623 }
624
625 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
626 {
627 #ifdef CONFIG_SCHEDSTATS
628         struct task_struct *tsk = NULL;
629
630         if (entity_is_task(se))
631                 tsk = task_of(se);
632
633         if (se->sleep_start) {
634                 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
635
636                 if ((s64)delta < 0)
637                         delta = 0;
638
639                 if (unlikely(delta > se->sleep_max))
640                         se->sleep_max = delta;
641
642                 se->sleep_start = 0;
643                 se->sum_sleep_runtime += delta;
644
645                 if (tsk) {
646                         account_scheduler_latency(tsk, delta >> 10, 1);
647                         trace_sched_stat_sleep(tsk, delta);
648                 }
649         }
650         if (se->block_start) {
651                 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
652
653                 if ((s64)delta < 0)
654                         delta = 0;
655
656                 if (unlikely(delta > se->block_max))
657                         se->block_max = delta;
658
659                 se->block_start = 0;
660                 se->sum_sleep_runtime += delta;
661
662                 if (tsk) {
663                         if (tsk->in_iowait) {
664                                 se->iowait_sum += delta;
665                                 se->iowait_count++;
666                                 trace_sched_stat_iowait(tsk, delta);
667                         }
668
669                         /*
670                          * Blocking time is in units of nanosecs, so shift by
671                          * 20 to get a milliseconds-range estimation of the
672                          * amount of time that the task spent sleeping:
673                          */
674                         if (unlikely(prof_on == SLEEP_PROFILING)) {
675                                 profile_hits(SLEEP_PROFILING,
676                                                 (void *)get_wchan(tsk),
677                                                 delta >> 20);
678                         }
679                         account_scheduler_latency(tsk, delta >> 10, 0);
680                 }
681         }
682 #endif
683 }
684
685 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
686 {
687 #ifdef CONFIG_SCHED_DEBUG
688         s64 d = se->vruntime - cfs_rq->min_vruntime;
689
690         if (d < 0)
691                 d = -d;
692
693         if (d > 3*sysctl_sched_latency)
694                 schedstat_inc(cfs_rq, nr_spread_over);
695 #endif
696 }
697
698 static void
699 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
700 {
701         u64 vruntime = cfs_rq->min_vruntime;
702
703         /*
704          * The 'current' period is already promised to the current tasks,
705          * however the extra weight of the new task will slow them down a
706          * little, place the new task so that it fits in the slot that
707          * stays open at the end.
708          */
709         if (initial && sched_feat(START_DEBIT))
710                 vruntime += sched_vslice(cfs_rq, se);
711
712         /* sleeps up to a single latency don't count. */
713         if (!initial && sched_feat(FAIR_SLEEPERS)) {
714                 unsigned long thresh = sysctl_sched_latency;
715
716                 /*
717                  * Convert the sleeper threshold into virtual time.
718                  * SCHED_IDLE is a special sub-class.  We care about
719                  * fairness only relative to other SCHED_IDLE tasks,
720                  * all of which have the same weight.
721                  */
722                 if (sched_feat(NORMALIZED_SLEEPER) && (!entity_is_task(se) ||
723                                  task_of(se)->policy != SCHED_IDLE))
724                         thresh = calc_delta_fair(thresh, se);
725
726                 /*
727                  * Halve their sleep time's effect, to allow
728                  * for a gentler effect of sleepers:
729                  */
730                 if (sched_feat(GENTLE_FAIR_SLEEPERS))
731                         thresh >>= 1;
732
733                 vruntime -= thresh;
734         }
735
736         /* ensure we never gain time by being placed backwards. */
737         vruntime = max_vruntime(se->vruntime, vruntime);
738
739         se->vruntime = vruntime;
740 }
741
742 static void
743 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
744 {
745         /*
746          * Update run-time statistics of the 'current'.
747          */
748         update_curr(cfs_rq);
749         account_entity_enqueue(cfs_rq, se);
750
751         if (wakeup) {
752                 place_entity(cfs_rq, se, 0);
753                 enqueue_sleeper(cfs_rq, se);
754         }
755
756         update_stats_enqueue(cfs_rq, se);
757         check_spread(cfs_rq, se);
758         if (se != cfs_rq->curr)
759                 __enqueue_entity(cfs_rq, se);
760 }
761
762 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
763 {
764         if (!se || cfs_rq->last == se)
765                 cfs_rq->last = NULL;
766
767         if (!se || cfs_rq->next == se)
768                 cfs_rq->next = NULL;
769 }
770
771 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
772 {
773         for_each_sched_entity(se)
774                 __clear_buddies(cfs_rq_of(se), se);
775 }
776
777 static void
778 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
779 {
780         /*
781          * Update run-time statistics of the 'current'.
782          */
783         update_curr(cfs_rq);
784
785         update_stats_dequeue(cfs_rq, se);
786         if (sleep) {
787 #ifdef CONFIG_SCHEDSTATS
788                 if (entity_is_task(se)) {
789                         struct task_struct *tsk = task_of(se);
790
791                         if (tsk->state & TASK_INTERRUPTIBLE)
792                                 se->sleep_start = rq_of(cfs_rq)->clock;
793                         if (tsk->state & TASK_UNINTERRUPTIBLE)
794                                 se->block_start = rq_of(cfs_rq)->clock;
795                 }
796 #endif
797         }
798
799         clear_buddies(cfs_rq, se);
800
801         if (se != cfs_rq->curr)
802                 __dequeue_entity(cfs_rq, se);
803         account_entity_dequeue(cfs_rq, se);
804         update_min_vruntime(cfs_rq);
805 }
806
807 /*
808  * Preempt the current task with a newly woken task if needed:
809  */
810 static void
811 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
812 {
813         unsigned long ideal_runtime, delta_exec;
814
815         ideal_runtime = sched_slice(cfs_rq, curr);
816         delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
817         if (delta_exec > ideal_runtime) {
818                 resched_task(rq_of(cfs_rq)->curr);
819                 /*
820                  * The current task ran long enough, ensure it doesn't get
821                  * re-elected due to buddy favours.
822                  */
823                 clear_buddies(cfs_rq, curr);
824         }
825 }
826
827 static void
828 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
829 {
830         /* 'current' is not kept within the tree. */
831         if (se->on_rq) {
832                 /*
833                  * Any task has to be enqueued before it get to execute on
834                  * a CPU. So account for the time it spent waiting on the
835                  * runqueue.
836                  */
837                 update_stats_wait_end(cfs_rq, se);
838                 __dequeue_entity(cfs_rq, se);
839         }
840
841         update_stats_curr_start(cfs_rq, se);
842         cfs_rq->curr = se;
843 #ifdef CONFIG_SCHEDSTATS
844         /*
845          * Track our maximum slice length, if the CPU's load is at
846          * least twice that of our own weight (i.e. dont track it
847          * when there are only lesser-weight tasks around):
848          */
849         if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
850                 se->slice_max = max(se->slice_max,
851                         se->sum_exec_runtime - se->prev_sum_exec_runtime);
852         }
853 #endif
854         se->prev_sum_exec_runtime = se->sum_exec_runtime;
855 }
856
857 static int
858 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
859
860 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
861 {
862         struct sched_entity *se = __pick_next_entity(cfs_rq);
863
864         if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, se) < 1)
865                 return cfs_rq->next;
866
867         if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, se) < 1)
868                 return cfs_rq->last;
869
870         return se;
871 }
872
873 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
874 {
875         /*
876          * If still on the runqueue then deactivate_task()
877          * was not called and update_curr() has to be done:
878          */
879         if (prev->on_rq)
880                 update_curr(cfs_rq);
881
882         check_spread(cfs_rq, prev);
883         if (prev->on_rq) {
884                 update_stats_wait_start(cfs_rq, prev);
885                 /* Put 'current' back into the tree. */
886                 __enqueue_entity(cfs_rq, prev);
887         }
888         cfs_rq->curr = NULL;
889 }
890
891 static void
892 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
893 {
894         /*
895          * Update run-time statistics of the 'current'.
896          */
897         update_curr(cfs_rq);
898
899 #ifdef CONFIG_SCHED_HRTICK
900         /*
901          * queued ticks are scheduled to match the slice, so don't bother
902          * validating it and just reschedule.
903          */
904         if (queued) {
905                 resched_task(rq_of(cfs_rq)->curr);
906                 return;
907         }
908         /*
909          * don't let the period tick interfere with the hrtick preemption
910          */
911         if (!sched_feat(DOUBLE_TICK) &&
912                         hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
913                 return;
914 #endif
915
916         if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
917                 check_preempt_tick(cfs_rq, curr);
918 }
919
920 /**************************************************
921  * CFS operations on tasks:
922  */
923
924 #ifdef CONFIG_SCHED_HRTICK
925 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
926 {
927         struct sched_entity *se = &p->se;
928         struct cfs_rq *cfs_rq = cfs_rq_of(se);
929
930         WARN_ON(task_rq(p) != rq);
931
932         if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
933                 u64 slice = sched_slice(cfs_rq, se);
934                 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
935                 s64 delta = slice - ran;
936
937                 if (delta < 0) {
938                         if (rq->curr == p)
939                                 resched_task(p);
940                         return;
941                 }
942
943                 /*
944                  * Don't schedule slices shorter than 10000ns, that just
945                  * doesn't make sense. Rely on vruntime for fairness.
946                  */
947                 if (rq->curr != p)
948                         delta = max_t(s64, 10000LL, delta);
949
950                 hrtick_start(rq, delta);
951         }
952 }
953
954 /*
955  * called from enqueue/dequeue and updates the hrtick when the
956  * current task is from our class and nr_running is low enough
957  * to matter.
958  */
959 static void hrtick_update(struct rq *rq)
960 {
961         struct task_struct *curr = rq->curr;
962
963         if (curr->sched_class != &fair_sched_class)
964                 return;
965
966         if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
967                 hrtick_start_fair(rq, curr);
968 }
969 #else /* !CONFIG_SCHED_HRTICK */
970 static inline void
971 hrtick_start_fair(struct rq *rq, struct task_struct *p)
972 {
973 }
974
975 static inline void hrtick_update(struct rq *rq)
976 {
977 }
978 #endif
979
980 /*
981  * The enqueue_task method is called before nr_running is
982  * increased. Here we update the fair scheduling stats and
983  * then put the task into the rbtree:
984  */
985 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
986 {
987         struct cfs_rq *cfs_rq;
988         struct sched_entity *se = &p->se;
989
990         for_each_sched_entity(se) {
991                 if (se->on_rq)
992                         break;
993                 cfs_rq = cfs_rq_of(se);
994                 enqueue_entity(cfs_rq, se, wakeup);
995                 wakeup = 1;
996         }
997
998         hrtick_update(rq);
999 }
1000
1001 /*
1002  * The dequeue_task method is called before nr_running is
1003  * decreased. We remove the task from the rbtree and
1004  * update the fair scheduling stats:
1005  */
1006 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
1007 {
1008         struct cfs_rq *cfs_rq;
1009         struct sched_entity *se = &p->se;
1010
1011         for_each_sched_entity(se) {
1012                 cfs_rq = cfs_rq_of(se);
1013                 dequeue_entity(cfs_rq, se, sleep);
1014                 /* Don't dequeue parent if it has other entities besides us */
1015                 if (cfs_rq->load.weight)
1016                         break;
1017                 sleep = 1;
1018         }
1019
1020         hrtick_update(rq);
1021 }
1022
1023 /*
1024  * sched_yield() support is very simple - we dequeue and enqueue.
1025  *
1026  * If compat_yield is turned on then we requeue to the end of the tree.
1027  */
1028 static void yield_task_fair(struct rq *rq)
1029 {
1030         struct task_struct *curr = rq->curr;
1031         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1032         struct sched_entity *rightmost, *se = &curr->se;
1033
1034         /*
1035          * Are we the only task in the tree?
1036          */
1037         if (unlikely(cfs_rq->nr_running == 1))
1038                 return;
1039
1040         clear_buddies(cfs_rq, se);
1041
1042         if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1043                 update_rq_clock(rq);
1044                 /*
1045                  * Update run-time statistics of the 'current'.
1046                  */
1047                 update_curr(cfs_rq);
1048
1049                 return;
1050         }
1051         /*
1052          * Find the rightmost entry in the rbtree:
1053          */
1054         rightmost = __pick_last_entity(cfs_rq);
1055         /*
1056          * Already in the rightmost position?
1057          */
1058         if (unlikely(!rightmost || entity_before(rightmost, se)))
1059                 return;
1060
1061         /*
1062          * Minimally necessary key value to be last in the tree:
1063          * Upon rescheduling, sched_class::put_prev_task() will place
1064          * 'current' within the tree based on its new key value.
1065          */
1066         se->vruntime = rightmost->vruntime + 1;
1067 }
1068
1069 #ifdef CONFIG_SMP
1070
1071 #ifdef CONFIG_FAIR_GROUP_SCHED
1072 /*
1073  * effective_load() calculates the load change as seen from the root_task_group
1074  *
1075  * Adding load to a group doesn't make a group heavier, but can cause movement
1076  * of group shares between cpus. Assuming the shares were perfectly aligned one
1077  * can calculate the shift in shares.
1078  *
1079  * The problem is that perfectly aligning the shares is rather expensive, hence
1080  * we try to avoid doing that too often - see update_shares(), which ratelimits
1081  * this change.
1082  *
1083  * We compensate this by not only taking the current delta into account, but
1084  * also considering the delta between when the shares were last adjusted and
1085  * now.
1086  *
1087  * We still saw a performance dip, some tracing learned us that between
1088  * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1089  * significantly. Therefore try to bias the error in direction of failing
1090  * the affine wakeup.
1091  *
1092  */
1093 static long effective_load(struct task_group *tg, int cpu,
1094                 long wl, long wg)
1095 {
1096         struct sched_entity *se = tg->se[cpu];
1097
1098         if (!tg->parent)
1099                 return wl;
1100
1101         /*
1102          * By not taking the decrease of shares on the other cpu into
1103          * account our error leans towards reducing the affine wakeups.
1104          */
1105         if (!wl && sched_feat(ASYM_EFF_LOAD))
1106                 return wl;
1107
1108         for_each_sched_entity(se) {
1109                 long S, rw, s, a, b;
1110                 long more_w;
1111
1112                 /*
1113                  * Instead of using this increment, also add the difference
1114                  * between when the shares were last updated and now.
1115                  */
1116                 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1117                 wl += more_w;
1118                 wg += more_w;
1119
1120                 S = se->my_q->tg->shares;
1121                 s = se->my_q->shares;
1122                 rw = se->my_q->rq_weight;
1123
1124                 a = S*(rw + wl);
1125                 b = S*rw + s*wg;
1126
1127                 wl = s*(a-b);
1128
1129                 if (likely(b))
1130                         wl /= b;
1131
1132                 /*
1133                  * Assume the group is already running and will
1134                  * thus already be accounted for in the weight.
1135                  *
1136                  * That is, moving shares between CPUs, does not
1137                  * alter the group weight.
1138                  */
1139                 wg = 0;
1140         }
1141
1142         return wl;
1143 }
1144
1145 #else
1146
1147 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1148                 unsigned long wl, unsigned long wg)
1149 {
1150         return wl;
1151 }
1152
1153 #endif
1154
1155 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1156 {
1157         struct task_struct *curr = current;
1158         unsigned long this_load, load;
1159         int idx, this_cpu, prev_cpu;
1160         unsigned long tl_per_task;
1161         unsigned int imbalance;
1162         struct task_group *tg;
1163         unsigned long weight;
1164         int balanced;
1165
1166         idx       = sd->wake_idx;
1167         this_cpu  = smp_processor_id();
1168         prev_cpu  = task_cpu(p);
1169         load      = source_load(prev_cpu, idx);
1170         this_load = target_load(this_cpu, idx);
1171
1172         if (sync) {
1173                if (sched_feat(SYNC_LESS) &&
1174                    (curr->se.avg_overlap > sysctl_sched_migration_cost ||
1175                     p->se.avg_overlap > sysctl_sched_migration_cost))
1176                        sync = 0;
1177         } else {
1178                 if (sched_feat(SYNC_MORE) &&
1179                     (curr->se.avg_overlap < sysctl_sched_migration_cost &&
1180                      p->se.avg_overlap < sysctl_sched_migration_cost))
1181                         sync = 1;
1182         }
1183
1184         /*
1185          * If sync wakeup then subtract the (maximum possible)
1186          * effect of the currently running task from the load
1187          * of the current CPU:
1188          */
1189         if (sync) {
1190                 tg = task_group(current);
1191                 weight = current->se.load.weight;
1192
1193                 this_load += effective_load(tg, this_cpu, -weight, -weight);
1194                 load += effective_load(tg, prev_cpu, 0, -weight);
1195         }
1196
1197         tg = task_group(p);
1198         weight = p->se.load.weight;
1199
1200         imbalance = 100 + (sd->imbalance_pct - 100) / 2;
1201
1202         /*
1203          * In low-load situations, where prev_cpu is idle and this_cpu is idle
1204          * due to the sync cause above having dropped this_load to 0, we'll
1205          * always have an imbalance, but there's really nothing you can do
1206          * about that, so that's good too.
1207          *
1208          * Otherwise check if either cpus are near enough in load to allow this
1209          * task to be woken on this_cpu.
1210          */
1211         balanced = !this_load ||
1212                 100*(this_load + effective_load(tg, this_cpu, weight, weight)) <=
1213                 imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
1214
1215         /*
1216          * If the currently running task will sleep within
1217          * a reasonable amount of time then attract this newly
1218          * woken task:
1219          */
1220         if (sync && balanced)
1221                 return 1;
1222
1223         schedstat_inc(p, se.nr_wakeups_affine_attempts);
1224         tl_per_task = cpu_avg_load_per_task(this_cpu);
1225
1226         if (balanced ||
1227             (this_load <= load &&
1228              this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1229                 /*
1230                  * This domain has SD_WAKE_AFFINE and
1231                  * p is cache cold in this domain, and
1232                  * there is no bad imbalance.
1233                  */
1234                 schedstat_inc(sd, ttwu_move_affine);
1235                 schedstat_inc(p, se.nr_wakeups_affine);
1236
1237                 return 1;
1238         }
1239         return 0;
1240 }
1241
1242 /*
1243  * find_idlest_group finds and returns the least busy CPU group within the
1244  * domain.
1245  */
1246 static struct sched_group *
1247 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1248                   int this_cpu, int load_idx)
1249 {
1250         struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1251         unsigned long min_load = ULONG_MAX, this_load = 0;
1252         int imbalance = 100 + (sd->imbalance_pct-100)/2;
1253
1254         do {
1255                 unsigned long load, avg_load;
1256                 int local_group;
1257                 int i;
1258
1259                 /* Skip over this group if it has no CPUs allowed */
1260                 if (!cpumask_intersects(sched_group_cpus(group),
1261                                         &p->cpus_allowed))
1262                         continue;
1263
1264                 local_group = cpumask_test_cpu(this_cpu,
1265                                                sched_group_cpus(group));
1266
1267                 /* Tally up the load of all CPUs in the group */
1268                 avg_load = 0;
1269
1270                 for_each_cpu(i, sched_group_cpus(group)) {
1271                         /* Bias balancing toward cpus of our domain */
1272                         if (local_group)
1273                                 load = source_load(i, load_idx);
1274                         else
1275                                 load = target_load(i, load_idx);
1276
1277                         avg_load += load;
1278                 }
1279
1280                 /* Adjust by relative CPU power of the group */
1281                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1282
1283                 if (local_group) {
1284                         this_load = avg_load;
1285                         this = group;
1286                 } else if (avg_load < min_load) {
1287                         min_load = avg_load;
1288                         idlest = group;
1289                 }
1290         } while (group = group->next, group != sd->groups);
1291
1292         if (!idlest || 100*this_load < imbalance*min_load)
1293                 return NULL;
1294         return idlest;
1295 }
1296
1297 /*
1298  * find_idlest_cpu - find the idlest cpu among the cpus in group.
1299  */
1300 static int
1301 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1302 {
1303         unsigned long load, min_load = ULONG_MAX;
1304         int idlest = -1;
1305         int i;
1306
1307         /* Traverse only the allowed CPUs */
1308         for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1309                 load = weighted_cpuload(i);
1310
1311                 if (load < min_load || (load == min_load && i == this_cpu)) {
1312                         min_load = load;
1313                         idlest = i;
1314                 }
1315         }
1316
1317         return idlest;
1318 }
1319
1320 /*
1321  * sched_balance_self: balance the current task (running on cpu) in domains
1322  * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1323  * SD_BALANCE_EXEC.
1324  *
1325  * Balance, ie. select the least loaded group.
1326  *
1327  * Returns the target CPU number, or the same CPU if no balancing is needed.
1328  *
1329  * preempt must be disabled.
1330  */
1331 static int select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
1332 {
1333         struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1334         int cpu = smp_processor_id();
1335         int prev_cpu = task_cpu(p);
1336         int new_cpu = cpu;
1337         int want_affine = 0;
1338         int want_sd = 1;
1339         int sync = wake_flags & WF_SYNC;
1340
1341         if (sd_flag & SD_BALANCE_WAKE) {
1342                 if (sched_feat(AFFINE_WAKEUPS) &&
1343                     cpumask_test_cpu(cpu, &p->cpus_allowed))
1344                         want_affine = 1;
1345                 new_cpu = prev_cpu;
1346         }
1347
1348         rcu_read_lock();
1349         for_each_domain(cpu, tmp) {
1350                 /*
1351                  * If power savings logic is enabled for a domain, see if we
1352                  * are not overloaded, if so, don't balance wider.
1353                  */
1354                 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1355                         unsigned long power = 0;
1356                         unsigned long nr_running = 0;
1357                         unsigned long capacity;
1358                         int i;
1359
1360                         for_each_cpu(i, sched_domain_span(tmp)) {
1361                                 power += power_of(i);
1362                                 nr_running += cpu_rq(i)->cfs.nr_running;
1363                         }
1364
1365                         capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1366
1367                         if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1368                                 nr_running /= 2;
1369
1370                         if (nr_running < capacity)
1371                                 want_sd = 0;
1372                 }
1373
1374                 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1375                     cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1376
1377                         affine_sd = tmp;
1378                         want_affine = 0;
1379                 }
1380
1381                 if (!want_sd && !want_affine)
1382                         break;
1383
1384                 if (!(tmp->flags & sd_flag))
1385                         continue;
1386
1387                 if (want_sd)
1388                         sd = tmp;
1389         }
1390
1391         if (sched_feat(LB_SHARES_UPDATE)) {
1392                 /*
1393                  * Pick the largest domain to update shares over
1394                  */
1395                 tmp = sd;
1396                 if (affine_sd && (!tmp ||
1397                                   cpumask_weight(sched_domain_span(affine_sd)) >
1398                                   cpumask_weight(sched_domain_span(sd))))
1399                         tmp = affine_sd;
1400
1401                 if (tmp)
1402                         update_shares(tmp);
1403         }
1404
1405         if (affine_sd && wake_affine(affine_sd, p, sync)) {
1406                 new_cpu = cpu;
1407                 goto out;
1408         }
1409
1410         while (sd) {
1411                 int load_idx = sd->forkexec_idx;
1412                 struct sched_group *group;
1413                 int weight;
1414
1415                 if (!(sd->flags & sd_flag)) {
1416                         sd = sd->child;
1417                         continue;
1418                 }
1419
1420                 if (sd_flag & SD_BALANCE_WAKE)
1421                         load_idx = sd->wake_idx;
1422
1423                 group = find_idlest_group(sd, p, cpu, load_idx);
1424                 if (!group) {
1425                         sd = sd->child;
1426                         continue;
1427                 }
1428
1429                 new_cpu = find_idlest_cpu(group, p, cpu);
1430                 if (new_cpu == -1 || new_cpu == cpu) {
1431                         /* Now try balancing at a lower domain level of cpu */
1432                         sd = sd->child;
1433                         continue;
1434                 }
1435
1436                 /* Now try balancing at a lower domain level of new_cpu */
1437                 cpu = new_cpu;
1438                 weight = cpumask_weight(sched_domain_span(sd));
1439                 sd = NULL;
1440                 for_each_domain(cpu, tmp) {
1441                         if (weight <= cpumask_weight(sched_domain_span(tmp)))
1442                                 break;
1443                         if (tmp->flags & sd_flag)
1444                                 sd = tmp;
1445                 }
1446                 /* while loop will break here if sd == NULL */
1447         }
1448
1449 out:
1450         rcu_read_unlock();
1451         return new_cpu;
1452 }
1453 #endif /* CONFIG_SMP */
1454
1455 /*
1456  * Adaptive granularity
1457  *
1458  * se->avg_wakeup gives the average time a task runs until it does a wakeup,
1459  * with the limit of wakeup_gran -- when it never does a wakeup.
1460  *
1461  * So the smaller avg_wakeup is the faster we want this task to preempt,
1462  * but we don't want to treat the preemptee unfairly and therefore allow it
1463  * to run for at least the amount of time we'd like to run.
1464  *
1465  * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one
1466  *
1467  * NOTE: we use *nr_running to scale with load, this nicely matches the
1468  *       degrading latency on load.
1469  */
1470 static unsigned long
1471 adaptive_gran(struct sched_entity *curr, struct sched_entity *se)
1472 {
1473         u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1474         u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running;
1475         u64 gran = 0;
1476
1477         if (this_run < expected_wakeup)
1478                 gran = expected_wakeup - this_run;
1479
1480         return min_t(s64, gran, sysctl_sched_wakeup_granularity);
1481 }
1482
1483 static unsigned long
1484 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1485 {
1486         unsigned long gran = sysctl_sched_wakeup_granularity;
1487
1488         if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN))
1489                 gran = adaptive_gran(curr, se);
1490
1491         /*
1492          * Since its curr running now, convert the gran from real-time
1493          * to virtual-time in his units.
1494          */
1495         if (sched_feat(ASYM_GRAN)) {
1496                 /*
1497                  * By using 'se' instead of 'curr' we penalize light tasks, so
1498                  * they get preempted easier. That is, if 'se' < 'curr' then
1499                  * the resulting gran will be larger, therefore penalizing the
1500                  * lighter, if otoh 'se' > 'curr' then the resulting gran will
1501                  * be smaller, again penalizing the lighter task.
1502                  *
1503                  * This is especially important for buddies when the leftmost
1504                  * task is higher priority than the buddy.
1505                  */
1506                 if (unlikely(se->load.weight != NICE_0_LOAD))
1507                         gran = calc_delta_fair(gran, se);
1508         } else {
1509                 if (unlikely(curr->load.weight != NICE_0_LOAD))
1510                         gran = calc_delta_fair(gran, curr);
1511         }
1512
1513         return gran;
1514 }
1515
1516 /*
1517  * Should 'se' preempt 'curr'.
1518  *
1519  *             |s1
1520  *        |s2
1521  *   |s3
1522  *         g
1523  *      |<--->|c
1524  *
1525  *  w(c, s1) = -1
1526  *  w(c, s2) =  0
1527  *  w(c, s3) =  1
1528  *
1529  */
1530 static int
1531 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1532 {
1533         s64 gran, vdiff = curr->vruntime - se->vruntime;
1534
1535         if (vdiff <= 0)
1536                 return -1;
1537
1538         gran = wakeup_gran(curr, se);
1539         if (vdiff > gran)
1540                 return 1;
1541
1542         return 0;
1543 }
1544
1545 static void set_last_buddy(struct sched_entity *se)
1546 {
1547         if (likely(task_of(se)->policy != SCHED_IDLE)) {
1548                 for_each_sched_entity(se)
1549                         cfs_rq_of(se)->last = se;
1550         }
1551 }
1552
1553 static void set_next_buddy(struct sched_entity *se)
1554 {
1555         if (likely(task_of(se)->policy != SCHED_IDLE)) {
1556                 for_each_sched_entity(se)
1557                         cfs_rq_of(se)->next = se;
1558         }
1559 }
1560
1561 /*
1562  * Preempt the current task with a newly woken task if needed:
1563  */
1564 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1565 {
1566         struct task_struct *curr = rq->curr;
1567         struct sched_entity *se = &curr->se, *pse = &p->se;
1568         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1569         int sync = wake_flags & WF_SYNC;
1570
1571         update_curr(cfs_rq);
1572
1573         if (unlikely(rt_prio(p->prio))) {
1574                 resched_task(curr);
1575                 return;
1576         }
1577
1578         if (unlikely(p->sched_class != &fair_sched_class))
1579                 return;
1580
1581         if (unlikely(se == pse))
1582                 return;
1583
1584         /*
1585          * Only set the backward buddy when the current task is still on the
1586          * rq. This can happen when a wakeup gets interleaved with schedule on
1587          * the ->pre_schedule() or idle_balance() point, either of which can
1588          * drop the rq lock.
1589          *
1590          * Also, during early boot the idle thread is in the fair class, for
1591          * obvious reasons its a bad idea to schedule back to the idle thread.
1592          */
1593         if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle))
1594                 set_last_buddy(se);
1595         if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK))
1596                 set_next_buddy(pse);
1597
1598         /*
1599          * We can come here with TIF_NEED_RESCHED already set from new task
1600          * wake up path.
1601          */
1602         if (test_tsk_need_resched(curr))
1603                 return;
1604
1605         /*
1606          * Batch and idle tasks do not preempt (their preemption is driven by
1607          * the tick):
1608          */
1609         if (unlikely(p->policy != SCHED_NORMAL))
1610                 return;
1611
1612         /* Idle tasks are by definition preempted by everybody. */
1613         if (unlikely(curr->policy == SCHED_IDLE)) {
1614                 resched_task(curr);
1615                 return;
1616         }
1617
1618         if ((sched_feat(WAKEUP_SYNC) && sync) ||
1619             (sched_feat(WAKEUP_OVERLAP) &&
1620              (se->avg_overlap < sysctl_sched_migration_cost &&
1621               pse->avg_overlap < sysctl_sched_migration_cost))) {
1622                 resched_task(curr);
1623                 return;
1624         }
1625
1626         if (sched_feat(WAKEUP_RUNNING)) {
1627                 if (pse->avg_running < se->avg_running) {
1628                         set_next_buddy(pse);
1629                         resched_task(curr);
1630                         return;
1631                 }
1632         }
1633
1634         if (!sched_feat(WAKEUP_PREEMPT))
1635                 return;
1636
1637         find_matching_se(&se, &pse);
1638
1639         BUG_ON(!pse);
1640
1641         if (wakeup_preempt_entity(se, pse) == 1)
1642                 resched_task(curr);
1643 }
1644
1645 static struct task_struct *pick_next_task_fair(struct rq *rq)
1646 {
1647         struct task_struct *p;
1648         struct cfs_rq *cfs_rq = &rq->cfs;
1649         struct sched_entity *se;
1650
1651         if (unlikely(!cfs_rq->nr_running))
1652                 return NULL;
1653
1654         do {
1655                 se = pick_next_entity(cfs_rq);
1656                 /*
1657                  * If se was a buddy, clear it so that it will have to earn
1658                  * the favour again.
1659                  *
1660                  * If se was not a buddy, clear the buddies because neither
1661                  * was elegible to run, let them earn it again.
1662                  *
1663                  * IOW. unconditionally clear buddies.
1664                  */
1665                 __clear_buddies(cfs_rq, NULL);
1666                 set_next_entity(cfs_rq, se);
1667                 cfs_rq = group_cfs_rq(se);
1668         } while (cfs_rq);
1669
1670         p = task_of(se);
1671         hrtick_start_fair(rq, p);
1672
1673         return p;
1674 }
1675
1676 /*
1677  * Account for a descheduled task:
1678  */
1679 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1680 {
1681         struct sched_entity *se = &prev->se;
1682         struct cfs_rq *cfs_rq;
1683
1684         for_each_sched_entity(se) {
1685                 cfs_rq = cfs_rq_of(se);
1686                 put_prev_entity(cfs_rq, se);
1687         }
1688 }
1689
1690 #ifdef CONFIG_SMP
1691 /**************************************************
1692  * Fair scheduling class load-balancing methods:
1693  */
1694
1695 /*
1696  * Load-balancing iterator. Note: while the runqueue stays locked
1697  * during the whole iteration, the current task might be
1698  * dequeued so the iterator has to be dequeue-safe. Here we
1699  * achieve that by always pre-iterating before returning
1700  * the current task:
1701  */
1702 static struct task_struct *
1703 __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
1704 {
1705         struct task_struct *p = NULL;
1706         struct sched_entity *se;
1707
1708         if (next == &cfs_rq->tasks)
1709                 return NULL;
1710
1711         se = list_entry(next, struct sched_entity, group_node);
1712         p = task_of(se);
1713         cfs_rq->balance_iterator = next->next;
1714
1715         return p;
1716 }
1717
1718 static struct task_struct *load_balance_start_fair(void *arg)
1719 {
1720         struct cfs_rq *cfs_rq = arg;
1721
1722         return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
1723 }
1724
1725 static struct task_struct *load_balance_next_fair(void *arg)
1726 {
1727         struct cfs_rq *cfs_rq = arg;
1728
1729         return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
1730 }
1731
1732 static unsigned long
1733 __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1734                 unsigned long max_load_move, struct sched_domain *sd,
1735                 enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
1736                 struct cfs_rq *cfs_rq)
1737 {
1738         struct rq_iterator cfs_rq_iterator;
1739
1740         cfs_rq_iterator.start = load_balance_start_fair;
1741         cfs_rq_iterator.next = load_balance_next_fair;
1742         cfs_rq_iterator.arg = cfs_rq;
1743
1744         return balance_tasks(this_rq, this_cpu, busiest,
1745                         max_load_move, sd, idle, all_pinned,
1746                         this_best_prio, &cfs_rq_iterator);
1747 }
1748
1749 #ifdef CONFIG_FAIR_GROUP_SCHED
1750 static unsigned long
1751 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1752                   unsigned long max_load_move,
1753                   struct sched_domain *sd, enum cpu_idle_type idle,
1754                   int *all_pinned, int *this_best_prio)
1755 {
1756         long rem_load_move = max_load_move;
1757         int busiest_cpu = cpu_of(busiest);
1758         struct task_group *tg;
1759
1760         rcu_read_lock();
1761         update_h_load(busiest_cpu);
1762
1763         list_for_each_entry_rcu(tg, &task_groups, list) {
1764                 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1765                 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1766                 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1767                 u64 rem_load, moved_load;
1768
1769                 /*
1770                  * empty group
1771                  */
1772                 if (!busiest_cfs_rq->task_weight)
1773                         continue;
1774
1775                 rem_load = (u64)rem_load_move * busiest_weight;
1776                 rem_load = div_u64(rem_load, busiest_h_load + 1);
1777
1778                 moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
1779                                 rem_load, sd, idle, all_pinned, this_best_prio,
1780                                 tg->cfs_rq[busiest_cpu]);
1781
1782                 if (!moved_load)
1783                         continue;
1784
1785                 moved_load *= busiest_h_load;
1786                 moved_load = div_u64(moved_load, busiest_weight + 1);
1787
1788                 rem_load_move -= moved_load;
1789                 if (rem_load_move < 0)
1790                         break;
1791         }
1792         rcu_read_unlock();
1793
1794         return max_load_move - rem_load_move;
1795 }
1796 #else
1797 static unsigned long
1798 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1799                   unsigned long max_load_move,
1800                   struct sched_domain *sd, enum cpu_idle_type idle,
1801                   int *all_pinned, int *this_best_prio)
1802 {
1803         return __load_balance_fair(this_rq, this_cpu, busiest,
1804                         max_load_move, sd, idle, all_pinned,
1805                         this_best_prio, &busiest->cfs);
1806 }
1807 #endif
1808
1809 static int
1810 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1811                    struct sched_domain *sd, enum cpu_idle_type idle)
1812 {
1813         struct cfs_rq *busy_cfs_rq;
1814         struct rq_iterator cfs_rq_iterator;
1815
1816         cfs_rq_iterator.start = load_balance_start_fair;
1817         cfs_rq_iterator.next = load_balance_next_fair;
1818
1819         for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1820                 /*
1821                  * pass busy_cfs_rq argument into
1822                  * load_balance_[start|next]_fair iterators
1823                  */
1824                 cfs_rq_iterator.arg = busy_cfs_rq;
1825                 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1826                                        &cfs_rq_iterator))
1827                     return 1;
1828         }
1829
1830         return 0;
1831 }
1832 #endif /* CONFIG_SMP */
1833
1834 /*
1835  * scheduler tick hitting a task of our scheduling class:
1836  */
1837 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1838 {
1839         struct cfs_rq *cfs_rq;
1840         struct sched_entity *se = &curr->se;
1841
1842         for_each_sched_entity(se) {
1843                 cfs_rq = cfs_rq_of(se);
1844                 entity_tick(cfs_rq, se, queued);
1845         }
1846 }
1847
1848 /*
1849  * Share the fairness runtime between parent and child, thus the
1850  * total amount of pressure for CPU stays equal - new tasks
1851  * get a chance to run but frequent forkers are not allowed to
1852  * monopolize the CPU. Note: the parent runqueue is locked,
1853  * the child is not running yet.
1854  */
1855 static void task_new_fair(struct rq *rq, struct task_struct *p)
1856 {
1857         struct cfs_rq *cfs_rq = task_cfs_rq(p);
1858         struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1859         int this_cpu = smp_processor_id();
1860
1861         sched_info_queued(p);
1862
1863         update_curr(cfs_rq);
1864         if (curr)
1865                 se->vruntime = curr->vruntime;
1866         place_entity(cfs_rq, se, 1);
1867
1868         /* 'curr' will be NULL if the child belongs to a different group */
1869         if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1870                         curr && entity_before(curr, se)) {
1871                 /*
1872                  * Upon rescheduling, sched_class::put_prev_task() will place
1873                  * 'current' within the tree based on its new key value.
1874                  */
1875                 swap(curr->vruntime, se->vruntime);
1876                 resched_task(rq->curr);
1877         }
1878
1879         enqueue_task_fair(rq, p, 0);
1880 }
1881
1882 /*
1883  * Priority of the task has changed. Check to see if we preempt
1884  * the current task.
1885  */
1886 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1887                               int oldprio, int running)
1888 {
1889         /*
1890          * Reschedule if we are currently running on this runqueue and
1891          * our priority decreased, or if we are not currently running on
1892          * this runqueue and our priority is higher than the current's
1893          */
1894         if (running) {
1895                 if (p->prio > oldprio)
1896                         resched_task(rq->curr);
1897         } else
1898                 check_preempt_curr(rq, p, 0);
1899 }
1900
1901 /*
1902  * We switched to the sched_fair class.
1903  */
1904 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1905                              int running)
1906 {
1907         /*
1908          * We were most likely switched from sched_rt, so
1909          * kick off the schedule if running, otherwise just see
1910          * if we can still preempt the current task.
1911          */
1912         if (running)
1913                 resched_task(rq->curr);
1914         else
1915                 check_preempt_curr(rq, p, 0);
1916 }
1917
1918 /* Account for a task changing its policy or group.
1919  *
1920  * This routine is mostly called to set cfs_rq->curr field when a task
1921  * migrates between groups/classes.
1922  */
1923 static void set_curr_task_fair(struct rq *rq)
1924 {
1925         struct sched_entity *se = &rq->curr->se;
1926
1927         for_each_sched_entity(se)
1928                 set_next_entity(cfs_rq_of(se), se);
1929 }
1930
1931 #ifdef CONFIG_FAIR_GROUP_SCHED
1932 static void moved_group_fair(struct task_struct *p)
1933 {
1934         struct cfs_rq *cfs_rq = task_cfs_rq(p);
1935
1936         update_curr(cfs_rq);
1937         place_entity(cfs_rq, &p->se, 1);
1938 }
1939 #endif
1940
1941 /*
1942  * All the scheduling class methods:
1943  */
1944 static const struct sched_class fair_sched_class = {
1945         .next                   = &idle_sched_class,
1946         .enqueue_task           = enqueue_task_fair,
1947         .dequeue_task           = dequeue_task_fair,
1948         .yield_task             = yield_task_fair,
1949
1950         .check_preempt_curr     = check_preempt_wakeup,
1951
1952         .pick_next_task         = pick_next_task_fair,
1953         .put_prev_task          = put_prev_task_fair,
1954
1955 #ifdef CONFIG_SMP
1956         .select_task_rq         = select_task_rq_fair,
1957
1958         .load_balance           = load_balance_fair,
1959         .move_one_task          = move_one_task_fair,
1960 #endif
1961
1962         .set_curr_task          = set_curr_task_fair,
1963         .task_tick              = task_tick_fair,
1964         .task_new               = task_new_fair,
1965
1966         .prio_changed           = prio_changed_fair,
1967         .switched_to            = switched_to_fair,
1968
1969 #ifdef CONFIG_FAIR_GROUP_SCHED
1970         .moved_group            = moved_group_fair,
1971 #endif
1972 };
1973
1974 #ifdef CONFIG_SCHED_DEBUG
1975 static void print_cfs_stats(struct seq_file *m, int cpu)
1976 {
1977         struct cfs_rq *cfs_rq;
1978
1979         rcu_read_lock();
1980         for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1981                 print_cfs_rq(m, cpu, cfs_rq);
1982         rcu_read_unlock();
1983 }
1984 #endif