sched: Implement a gentler fair-sleepers feature
[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         if (!initial) {
713                 /* sleeps upto a single latency don't count. */
714                 if (sched_feat(FAIR_SLEEPERS)) {
715                         unsigned long thresh = sysctl_sched_latency;
716
717                         /*
718                          * Convert the sleeper threshold into virtual time.
719                          * SCHED_IDLE is a special sub-class.  We care about
720                          * fairness only relative to other SCHED_IDLE tasks,
721                          * all of which have the same weight.
722                          */
723                         if (sched_feat(NORMALIZED_SLEEPER) &&
724                                         (!entity_is_task(se) ||
725                                          task_of(se)->policy != SCHED_IDLE))
726                                 thresh = calc_delta_fair(thresh, se);
727
728                         /*
729                          * Halve their sleep time's effect, to allow
730                          * for a gentler effect of sleepers:
731                          */
732                         if (sched_feat(GENTLE_FAIR_SLEEPERS))
733                                 thresh >>= 1;
734
735                         vruntime -= thresh;
736                 }
737         }
738
739         /* ensure we never gain time by being placed backwards. */
740         vruntime = max_vruntime(se->vruntime, vruntime);
741
742         se->vruntime = vruntime;
743 }
744
745 static void
746 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
747 {
748         /*
749          * Update run-time statistics of the 'current'.
750          */
751         update_curr(cfs_rq);
752         account_entity_enqueue(cfs_rq, se);
753
754         if (wakeup) {
755                 place_entity(cfs_rq, se, 0);
756                 enqueue_sleeper(cfs_rq, se);
757         }
758
759         update_stats_enqueue(cfs_rq, se);
760         check_spread(cfs_rq, se);
761         if (se != cfs_rq->curr)
762                 __enqueue_entity(cfs_rq, se);
763 }
764
765 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
766 {
767         if (cfs_rq->last == se)
768                 cfs_rq->last = NULL;
769
770         if (cfs_rq->next == se)
771                 cfs_rq->next = NULL;
772 }
773
774 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
775 {
776         for_each_sched_entity(se)
777                 __clear_buddies(cfs_rq_of(se), se);
778 }
779
780 static void
781 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
782 {
783         /*
784          * Update run-time statistics of the 'current'.
785          */
786         update_curr(cfs_rq);
787
788         update_stats_dequeue(cfs_rq, se);
789         if (sleep) {
790 #ifdef CONFIG_SCHEDSTATS
791                 if (entity_is_task(se)) {
792                         struct task_struct *tsk = task_of(se);
793
794                         if (tsk->state & TASK_INTERRUPTIBLE)
795                                 se->sleep_start = rq_of(cfs_rq)->clock;
796                         if (tsk->state & TASK_UNINTERRUPTIBLE)
797                                 se->block_start = rq_of(cfs_rq)->clock;
798                 }
799 #endif
800         }
801
802         clear_buddies(cfs_rq, se);
803
804         if (se != cfs_rq->curr)
805                 __dequeue_entity(cfs_rq, se);
806         account_entity_dequeue(cfs_rq, se);
807         update_min_vruntime(cfs_rq);
808 }
809
810 /*
811  * Preempt the current task with a newly woken task if needed:
812  */
813 static void
814 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
815 {
816         unsigned long ideal_runtime, delta_exec;
817
818         ideal_runtime = sched_slice(cfs_rq, curr);
819         delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
820         if (delta_exec > ideal_runtime) {
821                 resched_task(rq_of(cfs_rq)->curr);
822                 /*
823                  * The current task ran long enough, ensure it doesn't get
824                  * re-elected due to buddy favours.
825                  */
826                 clear_buddies(cfs_rq, curr);
827         }
828 }
829
830 static void
831 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
832 {
833         /* 'current' is not kept within the tree. */
834         if (se->on_rq) {
835                 /*
836                  * Any task has to be enqueued before it get to execute on
837                  * a CPU. So account for the time it spent waiting on the
838                  * runqueue.
839                  */
840                 update_stats_wait_end(cfs_rq, se);
841                 __dequeue_entity(cfs_rq, se);
842         }
843
844         update_stats_curr_start(cfs_rq, se);
845         cfs_rq->curr = se;
846 #ifdef CONFIG_SCHEDSTATS
847         /*
848          * Track our maximum slice length, if the CPU's load is at
849          * least twice that of our own weight (i.e. dont track it
850          * when there are only lesser-weight tasks around):
851          */
852         if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
853                 se->slice_max = max(se->slice_max,
854                         se->sum_exec_runtime - se->prev_sum_exec_runtime);
855         }
856 #endif
857         se->prev_sum_exec_runtime = se->sum_exec_runtime;
858 }
859
860 static int
861 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
862
863 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
864 {
865         struct sched_entity *se = __pick_next_entity(cfs_rq);
866
867         if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, se) < 1)
868                 return cfs_rq->next;
869
870         if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, se) < 1)
871                 return cfs_rq->last;
872
873         return se;
874 }
875
876 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
877 {
878         /*
879          * If still on the runqueue then deactivate_task()
880          * was not called and update_curr() has to be done:
881          */
882         if (prev->on_rq)
883                 update_curr(cfs_rq);
884
885         check_spread(cfs_rq, prev);
886         if (prev->on_rq) {
887                 update_stats_wait_start(cfs_rq, prev);
888                 /* Put 'current' back into the tree. */
889                 __enqueue_entity(cfs_rq, prev);
890         }
891         cfs_rq->curr = NULL;
892 }
893
894 static void
895 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
896 {
897         /*
898          * Update run-time statistics of the 'current'.
899          */
900         update_curr(cfs_rq);
901
902 #ifdef CONFIG_SCHED_HRTICK
903         /*
904          * queued ticks are scheduled to match the slice, so don't bother
905          * validating it and just reschedule.
906          */
907         if (queued) {
908                 resched_task(rq_of(cfs_rq)->curr);
909                 return;
910         }
911         /*
912          * don't let the period tick interfere with the hrtick preemption
913          */
914         if (!sched_feat(DOUBLE_TICK) &&
915                         hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
916                 return;
917 #endif
918
919         if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
920                 check_preempt_tick(cfs_rq, curr);
921 }
922
923 /**************************************************
924  * CFS operations on tasks:
925  */
926
927 #ifdef CONFIG_SCHED_HRTICK
928 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
929 {
930         struct sched_entity *se = &p->se;
931         struct cfs_rq *cfs_rq = cfs_rq_of(se);
932
933         WARN_ON(task_rq(p) != rq);
934
935         if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
936                 u64 slice = sched_slice(cfs_rq, se);
937                 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
938                 s64 delta = slice - ran;
939
940                 if (delta < 0) {
941                         if (rq->curr == p)
942                                 resched_task(p);
943                         return;
944                 }
945
946                 /*
947                  * Don't schedule slices shorter than 10000ns, that just
948                  * doesn't make sense. Rely on vruntime for fairness.
949                  */
950                 if (rq->curr != p)
951                         delta = max_t(s64, 10000LL, delta);
952
953                 hrtick_start(rq, delta);
954         }
955 }
956
957 /*
958  * called from enqueue/dequeue and updates the hrtick when the
959  * current task is from our class and nr_running is low enough
960  * to matter.
961  */
962 static void hrtick_update(struct rq *rq)
963 {
964         struct task_struct *curr = rq->curr;
965
966         if (curr->sched_class != &fair_sched_class)
967                 return;
968
969         if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
970                 hrtick_start_fair(rq, curr);
971 }
972 #else /* !CONFIG_SCHED_HRTICK */
973 static inline void
974 hrtick_start_fair(struct rq *rq, struct task_struct *p)
975 {
976 }
977
978 static inline void hrtick_update(struct rq *rq)
979 {
980 }
981 #endif
982
983 /*
984  * The enqueue_task method is called before nr_running is
985  * increased. Here we update the fair scheduling stats and
986  * then put the task into the rbtree:
987  */
988 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
989 {
990         struct cfs_rq *cfs_rq;
991         struct sched_entity *se = &p->se;
992
993         for_each_sched_entity(se) {
994                 if (se->on_rq)
995                         break;
996                 cfs_rq = cfs_rq_of(se);
997                 enqueue_entity(cfs_rq, se, wakeup);
998                 wakeup = 1;
999         }
1000
1001         hrtick_update(rq);
1002 }
1003
1004 /*
1005  * The dequeue_task method is called before nr_running is
1006  * decreased. We remove the task from the rbtree and
1007  * update the fair scheduling stats:
1008  */
1009 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
1010 {
1011         struct cfs_rq *cfs_rq;
1012         struct sched_entity *se = &p->se;
1013
1014         for_each_sched_entity(se) {
1015                 cfs_rq = cfs_rq_of(se);
1016                 dequeue_entity(cfs_rq, se, sleep);
1017                 /* Don't dequeue parent if it has other entities besides us */
1018                 if (cfs_rq->load.weight)
1019                         break;
1020                 sleep = 1;
1021         }
1022
1023         hrtick_update(rq);
1024 }
1025
1026 /*
1027  * sched_yield() support is very simple - we dequeue and enqueue.
1028  *
1029  * If compat_yield is turned on then we requeue to the end of the tree.
1030  */
1031 static void yield_task_fair(struct rq *rq)
1032 {
1033         struct task_struct *curr = rq->curr;
1034         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1035         struct sched_entity *rightmost, *se = &curr->se;
1036
1037         /*
1038          * Are we the only task in the tree?
1039          */
1040         if (unlikely(cfs_rq->nr_running == 1))
1041                 return;
1042
1043         clear_buddies(cfs_rq, se);
1044
1045         if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1046                 update_rq_clock(rq);
1047                 /*
1048                  * Update run-time statistics of the 'current'.
1049                  */
1050                 update_curr(cfs_rq);
1051
1052                 return;
1053         }
1054         /*
1055          * Find the rightmost entry in the rbtree:
1056          */
1057         rightmost = __pick_last_entity(cfs_rq);
1058         /*
1059          * Already in the rightmost position?
1060          */
1061         if (unlikely(!rightmost || entity_before(rightmost, se)))
1062                 return;
1063
1064         /*
1065          * Minimally necessary key value to be last in the tree:
1066          * Upon rescheduling, sched_class::put_prev_task() will place
1067          * 'current' within the tree based on its new key value.
1068          */
1069         se->vruntime = rightmost->vruntime + 1;
1070 }
1071
1072 #ifdef CONFIG_SMP
1073
1074 #ifdef CONFIG_FAIR_GROUP_SCHED
1075 /*
1076  * effective_load() calculates the load change as seen from the root_task_group
1077  *
1078  * Adding load to a group doesn't make a group heavier, but can cause movement
1079  * of group shares between cpus. Assuming the shares were perfectly aligned one
1080  * can calculate the shift in shares.
1081  *
1082  * The problem is that perfectly aligning the shares is rather expensive, hence
1083  * we try to avoid doing that too often - see update_shares(), which ratelimits
1084  * this change.
1085  *
1086  * We compensate this by not only taking the current delta into account, but
1087  * also considering the delta between when the shares were last adjusted and
1088  * now.
1089  *
1090  * We still saw a performance dip, some tracing learned us that between
1091  * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1092  * significantly. Therefore try to bias the error in direction of failing
1093  * the affine wakeup.
1094  *
1095  */
1096 static long effective_load(struct task_group *tg, int cpu,
1097                 long wl, long wg)
1098 {
1099         struct sched_entity *se = tg->se[cpu];
1100
1101         if (!tg->parent)
1102                 return wl;
1103
1104         /*
1105          * By not taking the decrease of shares on the other cpu into
1106          * account our error leans towards reducing the affine wakeups.
1107          */
1108         if (!wl && sched_feat(ASYM_EFF_LOAD))
1109                 return wl;
1110
1111         for_each_sched_entity(se) {
1112                 long S, rw, s, a, b;
1113                 long more_w;
1114
1115                 /*
1116                  * Instead of using this increment, also add the difference
1117                  * between when the shares were last updated and now.
1118                  */
1119                 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1120                 wl += more_w;
1121                 wg += more_w;
1122
1123                 S = se->my_q->tg->shares;
1124                 s = se->my_q->shares;
1125                 rw = se->my_q->rq_weight;
1126
1127                 a = S*(rw + wl);
1128                 b = S*rw + s*wg;
1129
1130                 wl = s*(a-b);
1131
1132                 if (likely(b))
1133                         wl /= b;
1134
1135                 /*
1136                  * Assume the group is already running and will
1137                  * thus already be accounted for in the weight.
1138                  *
1139                  * That is, moving shares between CPUs, does not
1140                  * alter the group weight.
1141                  */
1142                 wg = 0;
1143         }
1144
1145         return wl;
1146 }
1147
1148 #else
1149
1150 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1151                 unsigned long wl, unsigned long wg)
1152 {
1153         return wl;
1154 }
1155
1156 #endif
1157
1158 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1159 {
1160         struct task_struct *curr = current;
1161         unsigned long this_load, load;
1162         int idx, this_cpu, prev_cpu;
1163         unsigned long tl_per_task;
1164         unsigned int imbalance;
1165         struct task_group *tg;
1166         unsigned long weight;
1167         int balanced;
1168
1169         idx       = sd->wake_idx;
1170         this_cpu  = smp_processor_id();
1171         prev_cpu  = task_cpu(p);
1172         load      = source_load(prev_cpu, idx);
1173         this_load = target_load(this_cpu, idx);
1174
1175         if (sync) {
1176                if (sched_feat(SYNC_LESS) &&
1177                    (curr->se.avg_overlap > sysctl_sched_migration_cost ||
1178                     p->se.avg_overlap > sysctl_sched_migration_cost))
1179                        sync = 0;
1180         } else {
1181                 if (sched_feat(SYNC_MORE) &&
1182                     (curr->se.avg_overlap < sysctl_sched_migration_cost &&
1183                      p->se.avg_overlap < sysctl_sched_migration_cost))
1184                         sync = 1;
1185         }
1186
1187         /*
1188          * If sync wakeup then subtract the (maximum possible)
1189          * effect of the currently running task from the load
1190          * of the current CPU:
1191          */
1192         if (sync) {
1193                 tg = task_group(current);
1194                 weight = current->se.load.weight;
1195
1196                 this_load += effective_load(tg, this_cpu, -weight, -weight);
1197                 load += effective_load(tg, prev_cpu, 0, -weight);
1198         }
1199
1200         tg = task_group(p);
1201         weight = p->se.load.weight;
1202
1203         imbalance = 100 + (sd->imbalance_pct - 100) / 2;
1204
1205         /*
1206          * In low-load situations, where prev_cpu is idle and this_cpu is idle
1207          * due to the sync cause above having dropped this_load to 0, we'll
1208          * always have an imbalance, but there's really nothing you can do
1209          * about that, so that's good too.
1210          *
1211          * Otherwise check if either cpus are near enough in load to allow this
1212          * task to be woken on this_cpu.
1213          */
1214         balanced = !this_load ||
1215                 100*(this_load + effective_load(tg, this_cpu, weight, weight)) <=
1216                 imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
1217
1218         /*
1219          * If the currently running task will sleep within
1220          * a reasonable amount of time then attract this newly
1221          * woken task:
1222          */
1223         if (sync && balanced)
1224                 return 1;
1225
1226         schedstat_inc(p, se.nr_wakeups_affine_attempts);
1227         tl_per_task = cpu_avg_load_per_task(this_cpu);
1228
1229         if (balanced ||
1230             (this_load <= load &&
1231              this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1232                 /*
1233                  * This domain has SD_WAKE_AFFINE and
1234                  * p is cache cold in this domain, and
1235                  * there is no bad imbalance.
1236                  */
1237                 schedstat_inc(sd, ttwu_move_affine);
1238                 schedstat_inc(p, se.nr_wakeups_affine);
1239
1240                 return 1;
1241         }
1242         return 0;
1243 }
1244
1245 /*
1246  * find_idlest_group finds and returns the least busy CPU group within the
1247  * domain.
1248  */
1249 static struct sched_group *
1250 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1251                   int this_cpu, int flag)
1252 {
1253         struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1254         unsigned long min_load = ULONG_MAX, this_load = 0;
1255         int imbalance = 100 + (sd->imbalance_pct-100)/2;
1256         int load_idx = 0;
1257
1258         switch (flag) {
1259         case SD_BALANCE_FORK:
1260         case SD_BALANCE_EXEC:
1261                 load_idx = sd->forkexec_idx;
1262                 break;
1263
1264         case SD_BALANCE_WAKE:
1265                 load_idx = sd->wake_idx;
1266                 break;
1267
1268         default:
1269                 break;
1270         }
1271
1272         do {
1273                 unsigned long load, avg_load;
1274                 int local_group;
1275                 int i;
1276
1277                 /* Skip over this group if it has no CPUs allowed */
1278                 if (!cpumask_intersects(sched_group_cpus(group),
1279                                         &p->cpus_allowed))
1280                         continue;
1281
1282                 local_group = cpumask_test_cpu(this_cpu,
1283                                                sched_group_cpus(group));
1284
1285                 /* Tally up the load of all CPUs in the group */
1286                 avg_load = 0;
1287
1288                 for_each_cpu(i, sched_group_cpus(group)) {
1289                         /* Bias balancing toward cpus of our domain */
1290                         if (local_group)
1291                                 load = source_load(i, load_idx);
1292                         else
1293                                 load = target_load(i, load_idx);
1294
1295                         avg_load += load;
1296                 }
1297
1298                 /* Adjust by relative CPU power of the group */
1299                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1300
1301                 if (local_group) {
1302                         this_load = avg_load;
1303                         this = group;
1304                 } else if (avg_load < min_load) {
1305                         min_load = avg_load;
1306                         idlest = group;
1307                 }
1308         } while (group = group->next, group != sd->groups);
1309
1310         if (!idlest || 100*this_load < imbalance*min_load)
1311                 return NULL;
1312         return idlest;
1313 }
1314
1315 /*
1316  * find_idlest_cpu - find the idlest cpu among the cpus in group.
1317  */
1318 static int
1319 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1320 {
1321         unsigned long load, min_load = ULONG_MAX;
1322         int idlest = -1;
1323         int i;
1324
1325         /* Traverse only the allowed CPUs */
1326         for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1327                 load = weighted_cpuload(i);
1328
1329                 if (load < min_load || (load == min_load && i == this_cpu)) {
1330                         min_load = load;
1331                         idlest = i;
1332                 }
1333         }
1334
1335         return idlest;
1336 }
1337
1338 /*
1339  * sched_balance_self: balance the current task (running on cpu) in domains
1340  * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1341  * SD_BALANCE_EXEC.
1342  *
1343  * Balance, ie. select the least loaded group.
1344  *
1345  * Returns the target CPU number, or the same CPU if no balancing is needed.
1346  *
1347  * preempt must be disabled.
1348  */
1349 static int select_task_rq_fair(struct task_struct *p, int sd_flag, int flags)
1350 {
1351         struct sched_domain *tmp, *sd = NULL;
1352         int cpu = smp_processor_id();
1353         int prev_cpu = task_cpu(p);
1354         int new_cpu = cpu;
1355         int want_affine = 0;
1356         int sync = flags & WF_SYNC;
1357
1358         if (sd_flag & SD_BALANCE_WAKE) {
1359                 if (sched_feat(AFFINE_WAKEUPS))
1360                         want_affine = 1;
1361                 new_cpu = prev_cpu;
1362         }
1363
1364         rcu_read_lock();
1365         for_each_domain(cpu, tmp) {
1366                 /*
1367                  * If power savings logic is enabled for a domain, see if we
1368                  * are not overloaded, if so, don't balance wider.
1369                  */
1370                 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1371                         unsigned long power = 0;
1372                         unsigned long nr_running = 0;
1373                         unsigned long capacity;
1374                         int i;
1375
1376                         for_each_cpu(i, sched_domain_span(tmp)) {
1377                                 power += power_of(i);
1378                                 nr_running += cpu_rq(i)->cfs.nr_running;
1379                         }
1380
1381                         capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1382
1383                         if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1384                                 nr_running /= 2;
1385
1386                         if (nr_running < capacity)
1387                                 break;
1388                 }
1389
1390                 switch (sd_flag) {
1391                 case SD_BALANCE_WAKE:
1392                         if (!sched_feat(LB_WAKEUP_UPDATE))
1393                                 break;
1394                 case SD_BALANCE_FORK:
1395                 case SD_BALANCE_EXEC:
1396                         if (root_task_group_empty())
1397                                 break;
1398                         update_shares(tmp);
1399                 default:
1400                         break;
1401                 }
1402
1403                 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1404                     cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1405
1406                         if (wake_affine(tmp, p, sync)) {
1407                                 new_cpu = cpu;
1408                                 goto out;
1409                         }
1410
1411                         want_affine = 0;
1412                 }
1413
1414                 if (!(tmp->flags & sd_flag))
1415                         continue;
1416
1417                 sd = tmp;
1418         }
1419
1420         while (sd) {
1421                 struct sched_group *group;
1422                 int weight;
1423
1424                 if (!(sd->flags & sd_flag)) {
1425                         sd = sd->child;
1426                         continue;
1427                 }
1428
1429                 group = find_idlest_group(sd, p, cpu, sd_flag);
1430                 if (!group) {
1431                         sd = sd->child;
1432                         continue;
1433                 }
1434
1435                 new_cpu = find_idlest_cpu(group, p, cpu);
1436                 if (new_cpu == -1 || new_cpu == cpu) {
1437                         /* Now try balancing at a lower domain level of cpu */
1438                         sd = sd->child;
1439                         continue;
1440                 }
1441
1442                 /* Now try balancing at a lower domain level of new_cpu */
1443                 cpu = new_cpu;
1444                 weight = cpumask_weight(sched_domain_span(sd));
1445                 sd = NULL;
1446                 for_each_domain(cpu, tmp) {
1447                         if (weight <= cpumask_weight(sched_domain_span(tmp)))
1448                                 break;
1449                         if (tmp->flags & sd_flag)
1450                                 sd = tmp;
1451                 }
1452                 /* while loop will break here if sd == NULL */
1453         }
1454
1455 out:
1456         rcu_read_unlock();
1457         return new_cpu;
1458 }
1459 #endif /* CONFIG_SMP */
1460
1461 /*
1462  * Adaptive granularity
1463  *
1464  * se->avg_wakeup gives the average time a task runs until it does a wakeup,
1465  * with the limit of wakeup_gran -- when it never does a wakeup.
1466  *
1467  * So the smaller avg_wakeup is the faster we want this task to preempt,
1468  * but we don't want to treat the preemptee unfairly and therefore allow it
1469  * to run for at least the amount of time we'd like to run.
1470  *
1471  * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one
1472  *
1473  * NOTE: we use *nr_running to scale with load, this nicely matches the
1474  *       degrading latency on load.
1475  */
1476 static unsigned long
1477 adaptive_gran(struct sched_entity *curr, struct sched_entity *se)
1478 {
1479         u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1480         u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running;
1481         u64 gran = 0;
1482
1483         if (this_run < expected_wakeup)
1484                 gran = expected_wakeup - this_run;
1485
1486         return min_t(s64, gran, sysctl_sched_wakeup_granularity);
1487 }
1488
1489 static unsigned long
1490 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1491 {
1492         unsigned long gran = sysctl_sched_wakeup_granularity;
1493
1494         if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN))
1495                 gran = adaptive_gran(curr, se);
1496
1497         /*
1498          * Since its curr running now, convert the gran from real-time
1499          * to virtual-time in his units.
1500          */
1501         if (sched_feat(ASYM_GRAN)) {
1502                 /*
1503                  * By using 'se' instead of 'curr' we penalize light tasks, so
1504                  * they get preempted easier. That is, if 'se' < 'curr' then
1505                  * the resulting gran will be larger, therefore penalizing the
1506                  * lighter, if otoh 'se' > 'curr' then the resulting gran will
1507                  * be smaller, again penalizing the lighter task.
1508                  *
1509                  * This is especially important for buddies when the leftmost
1510                  * task is higher priority than the buddy.
1511                  */
1512                 if (unlikely(se->load.weight != NICE_0_LOAD))
1513                         gran = calc_delta_fair(gran, se);
1514         } else {
1515                 if (unlikely(curr->load.weight != NICE_0_LOAD))
1516                         gran = calc_delta_fair(gran, curr);
1517         }
1518
1519         return gran;
1520 }
1521
1522 /*
1523  * Should 'se' preempt 'curr'.
1524  *
1525  *             |s1
1526  *        |s2
1527  *   |s3
1528  *         g
1529  *      |<--->|c
1530  *
1531  *  w(c, s1) = -1
1532  *  w(c, s2) =  0
1533  *  w(c, s3) =  1
1534  *
1535  */
1536 static int
1537 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1538 {
1539         s64 gran, vdiff = curr->vruntime - se->vruntime;
1540
1541         if (vdiff <= 0)
1542                 return -1;
1543
1544         gran = wakeup_gran(curr, se);
1545         if (vdiff > gran)
1546                 return 1;
1547
1548         return 0;
1549 }
1550
1551 static void set_last_buddy(struct sched_entity *se)
1552 {
1553         if (likely(task_of(se)->policy != SCHED_IDLE)) {
1554                 for_each_sched_entity(se)
1555                         cfs_rq_of(se)->last = se;
1556         }
1557 }
1558
1559 static void set_next_buddy(struct sched_entity *se)
1560 {
1561         if (likely(task_of(se)->policy != SCHED_IDLE)) {
1562                 for_each_sched_entity(se)
1563                         cfs_rq_of(se)->next = se;
1564         }
1565 }
1566
1567 /*
1568  * Preempt the current task with a newly woken task if needed:
1569  */
1570 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int flags)
1571 {
1572         struct task_struct *curr = rq->curr;
1573         struct sched_entity *se = &curr->se, *pse = &p->se;
1574         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1575         int sync = flags & WF_SYNC;
1576
1577         update_curr(cfs_rq);
1578
1579         if (unlikely(rt_prio(p->prio))) {
1580                 resched_task(curr);
1581                 return;
1582         }
1583
1584         if (unlikely(p->sched_class != &fair_sched_class))
1585                 return;
1586
1587         if (unlikely(se == pse))
1588                 return;
1589
1590         /*
1591          * Only set the backward buddy when the current task is still on the
1592          * rq. This can happen when a wakeup gets interleaved with schedule on
1593          * the ->pre_schedule() or idle_balance() point, either of which can
1594          * drop the rq lock.
1595          *
1596          * Also, during early boot the idle thread is in the fair class, for
1597          * obvious reasons its a bad idea to schedule back to the idle thread.
1598          */
1599         if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle))
1600                 set_last_buddy(se);
1601         if (sched_feat(NEXT_BUDDY) && !(flags & WF_FORK))
1602                 set_next_buddy(pse);
1603
1604         /*
1605          * We can come here with TIF_NEED_RESCHED already set from new task
1606          * wake up path.
1607          */
1608         if (test_tsk_need_resched(curr))
1609                 return;
1610
1611         /*
1612          * Batch and idle tasks do not preempt (their preemption is driven by
1613          * the tick):
1614          */
1615         if (unlikely(p->policy != SCHED_NORMAL))
1616                 return;
1617
1618         /* Idle tasks are by definition preempted by everybody. */
1619         if (unlikely(curr->policy == SCHED_IDLE)) {
1620                 resched_task(curr);
1621                 return;
1622         }
1623
1624         if (!sched_feat(WAKEUP_PREEMPT))
1625                 return;
1626
1627         if ((sched_feat(WAKEUP_SYNC) && sync) ||
1628             (sched_feat(WAKEUP_OVERLAP) &&
1629              (se->avg_overlap < sysctl_sched_migration_cost &&
1630               pse->avg_overlap < sysctl_sched_migration_cost))) {
1631                 resched_task(curr);
1632                 return;
1633         }
1634
1635         find_matching_se(&se, &pse);
1636
1637         BUG_ON(!pse);
1638
1639         if (wakeup_preempt_entity(se, pse) == 1)
1640                 resched_task(curr);
1641 }
1642
1643 static struct task_struct *pick_next_task_fair(struct rq *rq)
1644 {
1645         struct task_struct *p;
1646         struct cfs_rq *cfs_rq = &rq->cfs;
1647         struct sched_entity *se;
1648
1649         if (unlikely(!cfs_rq->nr_running))
1650                 return NULL;
1651
1652         do {
1653                 se = pick_next_entity(cfs_rq);
1654                 /*
1655                  * If se was a buddy, clear it so that it will have to earn
1656                  * the favour again.
1657                  */
1658                 __clear_buddies(cfs_rq, se);
1659                 set_next_entity(cfs_rq, se);
1660                 cfs_rq = group_cfs_rq(se);
1661         } while (cfs_rq);
1662
1663         p = task_of(se);
1664         hrtick_start_fair(rq, p);
1665
1666         return p;
1667 }
1668
1669 /*
1670  * Account for a descheduled task:
1671  */
1672 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1673 {
1674         struct sched_entity *se = &prev->se;
1675         struct cfs_rq *cfs_rq;
1676
1677         for_each_sched_entity(se) {
1678                 cfs_rq = cfs_rq_of(se);
1679                 put_prev_entity(cfs_rq, se);
1680         }
1681 }
1682
1683 #ifdef CONFIG_SMP
1684 /**************************************************
1685  * Fair scheduling class load-balancing methods:
1686  */
1687
1688 /*
1689  * Load-balancing iterator. Note: while the runqueue stays locked
1690  * during the whole iteration, the current task might be
1691  * dequeued so the iterator has to be dequeue-safe. Here we
1692  * achieve that by always pre-iterating before returning
1693  * the current task:
1694  */
1695 static struct task_struct *
1696 __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
1697 {
1698         struct task_struct *p = NULL;
1699         struct sched_entity *se;
1700
1701         if (next == &cfs_rq->tasks)
1702                 return NULL;
1703
1704         se = list_entry(next, struct sched_entity, group_node);
1705         p = task_of(se);
1706         cfs_rq->balance_iterator = next->next;
1707
1708         return p;
1709 }
1710
1711 static struct task_struct *load_balance_start_fair(void *arg)
1712 {
1713         struct cfs_rq *cfs_rq = arg;
1714
1715         return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
1716 }
1717
1718 static struct task_struct *load_balance_next_fair(void *arg)
1719 {
1720         struct cfs_rq *cfs_rq = arg;
1721
1722         return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
1723 }
1724
1725 static unsigned long
1726 __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1727                 unsigned long max_load_move, struct sched_domain *sd,
1728                 enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
1729                 struct cfs_rq *cfs_rq)
1730 {
1731         struct rq_iterator cfs_rq_iterator;
1732
1733         cfs_rq_iterator.start = load_balance_start_fair;
1734         cfs_rq_iterator.next = load_balance_next_fair;
1735         cfs_rq_iterator.arg = cfs_rq;
1736
1737         return balance_tasks(this_rq, this_cpu, busiest,
1738                         max_load_move, sd, idle, all_pinned,
1739                         this_best_prio, &cfs_rq_iterator);
1740 }
1741
1742 #ifdef CONFIG_FAIR_GROUP_SCHED
1743 static unsigned long
1744 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1745                   unsigned long max_load_move,
1746                   struct sched_domain *sd, enum cpu_idle_type idle,
1747                   int *all_pinned, int *this_best_prio)
1748 {
1749         long rem_load_move = max_load_move;
1750         int busiest_cpu = cpu_of(busiest);
1751         struct task_group *tg;
1752
1753         rcu_read_lock();
1754         update_h_load(busiest_cpu);
1755
1756         list_for_each_entry_rcu(tg, &task_groups, list) {
1757                 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1758                 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1759                 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1760                 u64 rem_load, moved_load;
1761
1762                 /*
1763                  * empty group
1764                  */
1765                 if (!busiest_cfs_rq->task_weight)
1766                         continue;
1767
1768                 rem_load = (u64)rem_load_move * busiest_weight;
1769                 rem_load = div_u64(rem_load, busiest_h_load + 1);
1770
1771                 moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
1772                                 rem_load, sd, idle, all_pinned, this_best_prio,
1773                                 tg->cfs_rq[busiest_cpu]);
1774
1775                 if (!moved_load)
1776                         continue;
1777
1778                 moved_load *= busiest_h_load;
1779                 moved_load = div_u64(moved_load, busiest_weight + 1);
1780
1781                 rem_load_move -= moved_load;
1782                 if (rem_load_move < 0)
1783                         break;
1784         }
1785         rcu_read_unlock();
1786
1787         return max_load_move - rem_load_move;
1788 }
1789 #else
1790 static unsigned long
1791 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1792                   unsigned long max_load_move,
1793                   struct sched_domain *sd, enum cpu_idle_type idle,
1794                   int *all_pinned, int *this_best_prio)
1795 {
1796         return __load_balance_fair(this_rq, this_cpu, busiest,
1797                         max_load_move, sd, idle, all_pinned,
1798                         this_best_prio, &busiest->cfs);
1799 }
1800 #endif
1801
1802 static int
1803 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1804                    struct sched_domain *sd, enum cpu_idle_type idle)
1805 {
1806         struct cfs_rq *busy_cfs_rq;
1807         struct rq_iterator cfs_rq_iterator;
1808
1809         cfs_rq_iterator.start = load_balance_start_fair;
1810         cfs_rq_iterator.next = load_balance_next_fair;
1811
1812         for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1813                 /*
1814                  * pass busy_cfs_rq argument into
1815                  * load_balance_[start|next]_fair iterators
1816                  */
1817                 cfs_rq_iterator.arg = busy_cfs_rq;
1818                 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1819                                        &cfs_rq_iterator))
1820                     return 1;
1821         }
1822
1823         return 0;
1824 }
1825 #endif /* CONFIG_SMP */
1826
1827 /*
1828  * scheduler tick hitting a task of our scheduling class:
1829  */
1830 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1831 {
1832         struct cfs_rq *cfs_rq;
1833         struct sched_entity *se = &curr->se;
1834
1835         for_each_sched_entity(se) {
1836                 cfs_rq = cfs_rq_of(se);
1837                 entity_tick(cfs_rq, se, queued);
1838         }
1839 }
1840
1841 /*
1842  * Share the fairness runtime between parent and child, thus the
1843  * total amount of pressure for CPU stays equal - new tasks
1844  * get a chance to run but frequent forkers are not allowed to
1845  * monopolize the CPU. Note: the parent runqueue is locked,
1846  * the child is not running yet.
1847  */
1848 static void task_new_fair(struct rq *rq, struct task_struct *p)
1849 {
1850         struct cfs_rq *cfs_rq = task_cfs_rq(p);
1851         struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1852         int this_cpu = smp_processor_id();
1853
1854         sched_info_queued(p);
1855
1856         update_curr(cfs_rq);
1857         if (curr)
1858                 se->vruntime = curr->vruntime;
1859         place_entity(cfs_rq, se, 1);
1860
1861         /* 'curr' will be NULL if the child belongs to a different group */
1862         if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1863                         curr && entity_before(curr, se)) {
1864                 /*
1865                  * Upon rescheduling, sched_class::put_prev_task() will place
1866                  * 'current' within the tree based on its new key value.
1867                  */
1868                 swap(curr->vruntime, se->vruntime);
1869                 resched_task(rq->curr);
1870         }
1871
1872         enqueue_task_fair(rq, p, 0);
1873 }
1874
1875 /*
1876  * Priority of the task has changed. Check to see if we preempt
1877  * the current task.
1878  */
1879 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1880                               int oldprio, int running)
1881 {
1882         /*
1883          * Reschedule if we are currently running on this runqueue and
1884          * our priority decreased, or if we are not currently running on
1885          * this runqueue and our priority is higher than the current's
1886          */
1887         if (running) {
1888                 if (p->prio > oldprio)
1889                         resched_task(rq->curr);
1890         } else
1891                 check_preempt_curr(rq, p, 0);
1892 }
1893
1894 /*
1895  * We switched to the sched_fair class.
1896  */
1897 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1898                              int running)
1899 {
1900         /*
1901          * We were most likely switched from sched_rt, so
1902          * kick off the schedule if running, otherwise just see
1903          * if we can still preempt the current task.
1904          */
1905         if (running)
1906                 resched_task(rq->curr);
1907         else
1908                 check_preempt_curr(rq, p, 0);
1909 }
1910
1911 /* Account for a task changing its policy or group.
1912  *
1913  * This routine is mostly called to set cfs_rq->curr field when a task
1914  * migrates between groups/classes.
1915  */
1916 static void set_curr_task_fair(struct rq *rq)
1917 {
1918         struct sched_entity *se = &rq->curr->se;
1919
1920         for_each_sched_entity(se)
1921                 set_next_entity(cfs_rq_of(se), se);
1922 }
1923
1924 #ifdef CONFIG_FAIR_GROUP_SCHED
1925 static void moved_group_fair(struct task_struct *p)
1926 {
1927         struct cfs_rq *cfs_rq = task_cfs_rq(p);
1928
1929         update_curr(cfs_rq);
1930         place_entity(cfs_rq, &p->se, 1);
1931 }
1932 #endif
1933
1934 /*
1935  * All the scheduling class methods:
1936  */
1937 static const struct sched_class fair_sched_class = {
1938         .next                   = &idle_sched_class,
1939         .enqueue_task           = enqueue_task_fair,
1940         .dequeue_task           = dequeue_task_fair,
1941         .yield_task             = yield_task_fair,
1942
1943         .check_preempt_curr     = check_preempt_wakeup,
1944
1945         .pick_next_task         = pick_next_task_fair,
1946         .put_prev_task          = put_prev_task_fair,
1947
1948 #ifdef CONFIG_SMP
1949         .select_task_rq         = select_task_rq_fair,
1950
1951         .load_balance           = load_balance_fair,
1952         .move_one_task          = move_one_task_fair,
1953 #endif
1954
1955         .set_curr_task          = set_curr_task_fair,
1956         .task_tick              = task_tick_fair,
1957         .task_new               = task_new_fair,
1958
1959         .prio_changed           = prio_changed_fair,
1960         .switched_to            = switched_to_fair,
1961
1962 #ifdef CONFIG_FAIR_GROUP_SCHED
1963         .moved_group            = moved_group_fair,
1964 #endif
1965 };
1966
1967 #ifdef CONFIG_SCHED_DEBUG
1968 static void print_cfs_stats(struct seq_file *m, int cpu)
1969 {
1970         struct cfs_rq *cfs_rq;
1971
1972         rcu_read_lock();
1973         for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1974                 print_cfs_rq(m, cpu, cfs_rq);
1975         rcu_read_unlock();
1976 }
1977 #endif