sched: Select_task_rq_fair() must honour SD_LOAD_BALANCE
[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 #include <linux/sched.h>
25
26 /*
27  * Targeted preemption latency for CPU-bound tasks:
28  * (default: 5ms * (1 + ilog(ncpus)), units: nanoseconds)
29  *
30  * NOTE: this latency value is not the same as the concept of
31  * 'timeslice length' - timeslices in CFS are of variable length
32  * and have no persistent notion like in traditional, time-slice
33  * based scheduling concepts.
34  *
35  * (to see the precise effective timeslice length of your workload,
36  *  run vmstat and monitor the context-switches (cs) field)
37  */
38 unsigned int sysctl_sched_latency = 5000000ULL;
39 unsigned int normalized_sysctl_sched_latency = 5000000ULL;
40
41 /*
42  * The initial- and re-scaling of tunables is configurable
43  * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
44  *
45  * Options are:
46  * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47  * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48  * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
49  */
50 enum sched_tunable_scaling sysctl_sched_tunable_scaling
51         = SCHED_TUNABLESCALING_LOG;
52
53 /*
54  * Minimal preemption granularity for CPU-bound tasks:
55  * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
56  */
57 unsigned int sysctl_sched_min_granularity = 1000000ULL;
58 unsigned int normalized_sysctl_sched_min_granularity = 1000000ULL;
59
60 /*
61  * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
62  */
63 static unsigned int sched_nr_latency = 5;
64
65 /*
66  * After fork, child runs first. If set to 0 (default) then
67  * parent will (try to) run first.
68  */
69 unsigned int sysctl_sched_child_runs_first __read_mostly;
70
71 /*
72  * sys_sched_yield() compat mode
73  *
74  * This option switches the agressive yield implementation of the
75  * old scheduler back on.
76  */
77 unsigned int __read_mostly sysctl_sched_compat_yield;
78
79 /*
80  * SCHED_OTHER wake-up granularity.
81  * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
82  *
83  * This option delays the preemption effects of decoupled workloads
84  * and reduces their over-scheduling. Synchronous workloads will still
85  * have immediate wakeup/sleep latencies.
86  */
87 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
88 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
89
90 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
91
92 static const struct sched_class fair_sched_class;
93
94 /**************************************************************
95  * CFS operations on generic schedulable entities:
96  */
97
98 #ifdef CONFIG_FAIR_GROUP_SCHED
99
100 /* cpu runqueue to which this cfs_rq is attached */
101 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
102 {
103         return cfs_rq->rq;
104 }
105
106 /* An entity is a task if it doesn't "own" a runqueue */
107 #define entity_is_task(se)      (!se->my_q)
108
109 static inline struct task_struct *task_of(struct sched_entity *se)
110 {
111 #ifdef CONFIG_SCHED_DEBUG
112         WARN_ON_ONCE(!entity_is_task(se));
113 #endif
114         return container_of(se, struct task_struct, se);
115 }
116
117 /* Walk up scheduling entities hierarchy */
118 #define for_each_sched_entity(se) \
119                 for (; se; se = se->parent)
120
121 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
122 {
123         return p->se.cfs_rq;
124 }
125
126 /* runqueue on which this entity is (to be) queued */
127 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
128 {
129         return se->cfs_rq;
130 }
131
132 /* runqueue "owned" by this group */
133 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
134 {
135         return grp->my_q;
136 }
137
138 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
139  * another cpu ('this_cpu')
140  */
141 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
142 {
143         return cfs_rq->tg->cfs_rq[this_cpu];
144 }
145
146 /* Iterate thr' all leaf cfs_rq's on a runqueue */
147 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
148         list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
149
150 /* Do the two (enqueued) entities belong to the same group ? */
151 static inline int
152 is_same_group(struct sched_entity *se, struct sched_entity *pse)
153 {
154         if (se->cfs_rq == pse->cfs_rq)
155                 return 1;
156
157         return 0;
158 }
159
160 static inline struct sched_entity *parent_entity(struct sched_entity *se)
161 {
162         return se->parent;
163 }
164
165 /* return depth at which a sched entity is present in the hierarchy */
166 static inline int depth_se(struct sched_entity *se)
167 {
168         int depth = 0;
169
170         for_each_sched_entity(se)
171                 depth++;
172
173         return depth;
174 }
175
176 static void
177 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
178 {
179         int se_depth, pse_depth;
180
181         /*
182          * preemption test can be made between sibling entities who are in the
183          * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
184          * both tasks until we find their ancestors who are siblings of common
185          * parent.
186          */
187
188         /* First walk up until both entities are at same depth */
189         se_depth = depth_se(*se);
190         pse_depth = depth_se(*pse);
191
192         while (se_depth > pse_depth) {
193                 se_depth--;
194                 *se = parent_entity(*se);
195         }
196
197         while (pse_depth > se_depth) {
198                 pse_depth--;
199                 *pse = parent_entity(*pse);
200         }
201
202         while (!is_same_group(*se, *pse)) {
203                 *se = parent_entity(*se);
204                 *pse = parent_entity(*pse);
205         }
206 }
207
208 #else   /* !CONFIG_FAIR_GROUP_SCHED */
209
210 static inline struct task_struct *task_of(struct sched_entity *se)
211 {
212         return container_of(se, struct task_struct, se);
213 }
214
215 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
216 {
217         return container_of(cfs_rq, struct rq, cfs);
218 }
219
220 #define entity_is_task(se)      1
221
222 #define for_each_sched_entity(se) \
223                 for (; se; se = NULL)
224
225 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
226 {
227         return &task_rq(p)->cfs;
228 }
229
230 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
231 {
232         struct task_struct *p = task_of(se);
233         struct rq *rq = task_rq(p);
234
235         return &rq->cfs;
236 }
237
238 /* runqueue "owned" by this group */
239 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
240 {
241         return NULL;
242 }
243
244 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
245 {
246         return &cpu_rq(this_cpu)->cfs;
247 }
248
249 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
250                 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
251
252 static inline int
253 is_same_group(struct sched_entity *se, struct sched_entity *pse)
254 {
255         return 1;
256 }
257
258 static inline struct sched_entity *parent_entity(struct sched_entity *se)
259 {
260         return NULL;
261 }
262
263 static inline void
264 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
265 {
266 }
267
268 #endif  /* CONFIG_FAIR_GROUP_SCHED */
269
270
271 /**************************************************************
272  * Scheduling class tree data structure manipulation methods:
273  */
274
275 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
276 {
277         s64 delta = (s64)(vruntime - min_vruntime);
278         if (delta > 0)
279                 min_vruntime = vruntime;
280
281         return min_vruntime;
282 }
283
284 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
285 {
286         s64 delta = (s64)(vruntime - min_vruntime);
287         if (delta < 0)
288                 min_vruntime = vruntime;
289
290         return min_vruntime;
291 }
292
293 static inline int entity_before(struct sched_entity *a,
294                                 struct sched_entity *b)
295 {
296         return (s64)(a->vruntime - b->vruntime) < 0;
297 }
298
299 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
300 {
301         return se->vruntime - cfs_rq->min_vruntime;
302 }
303
304 static void update_min_vruntime(struct cfs_rq *cfs_rq)
305 {
306         u64 vruntime = cfs_rq->min_vruntime;
307
308         if (cfs_rq->curr)
309                 vruntime = cfs_rq->curr->vruntime;
310
311         if (cfs_rq->rb_leftmost) {
312                 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
313                                                    struct sched_entity,
314                                                    run_node);
315
316                 if (!cfs_rq->curr)
317                         vruntime = se->vruntime;
318                 else
319                         vruntime = min_vruntime(vruntime, se->vruntime);
320         }
321
322         cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
323 }
324
325 /*
326  * Enqueue an entity into the rb-tree:
327  */
328 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
329 {
330         struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
331         struct rb_node *parent = NULL;
332         struct sched_entity *entry;
333         s64 key = entity_key(cfs_rq, se);
334         int leftmost = 1;
335
336         /*
337          * Find the right place in the rbtree:
338          */
339         while (*link) {
340                 parent = *link;
341                 entry = rb_entry(parent, struct sched_entity, run_node);
342                 /*
343                  * We dont care about collisions. Nodes with
344                  * the same key stay together.
345                  */
346                 if (key < entity_key(cfs_rq, entry)) {
347                         link = &parent->rb_left;
348                 } else {
349                         link = &parent->rb_right;
350                         leftmost = 0;
351                 }
352         }
353
354         /*
355          * Maintain a cache of leftmost tree entries (it is frequently
356          * used):
357          */
358         if (leftmost)
359                 cfs_rq->rb_leftmost = &se->run_node;
360
361         rb_link_node(&se->run_node, parent, link);
362         rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
363 }
364
365 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
366 {
367         if (cfs_rq->rb_leftmost == &se->run_node) {
368                 struct rb_node *next_node;
369
370                 next_node = rb_next(&se->run_node);
371                 cfs_rq->rb_leftmost = next_node;
372         }
373
374         rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
375 }
376
377 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
378 {
379         struct rb_node *left = cfs_rq->rb_leftmost;
380
381         if (!left)
382                 return NULL;
383
384         return rb_entry(left, struct sched_entity, run_node);
385 }
386
387 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
388 {
389         struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
390
391         if (!last)
392                 return NULL;
393
394         return rb_entry(last, struct sched_entity, run_node);
395 }
396
397 /**************************************************************
398  * Scheduling class statistics methods:
399  */
400
401 #ifdef CONFIG_SCHED_DEBUG
402 int sched_proc_update_handler(struct ctl_table *table, int write,
403                 void __user *buffer, size_t *lenp,
404                 loff_t *ppos)
405 {
406         int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
407         int factor = get_update_sysctl_factor();
408
409         if (ret || !write)
410                 return ret;
411
412         sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
413                                         sysctl_sched_min_granularity);
414
415 #define WRT_SYSCTL(name) \
416         (normalized_sysctl_##name = sysctl_##name / (factor))
417         WRT_SYSCTL(sched_min_granularity);
418         WRT_SYSCTL(sched_latency);
419         WRT_SYSCTL(sched_wakeup_granularity);
420         WRT_SYSCTL(sched_shares_ratelimit);
421 #undef WRT_SYSCTL
422
423         return 0;
424 }
425 #endif
426
427 /*
428  * delta /= w
429  */
430 static inline unsigned long
431 calc_delta_fair(unsigned long delta, struct sched_entity *se)
432 {
433         if (unlikely(se->load.weight != NICE_0_LOAD))
434                 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
435
436         return delta;
437 }
438
439 /*
440  * The idea is to set a period in which each task runs once.
441  *
442  * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
443  * this period because otherwise the slices get too small.
444  *
445  * p = (nr <= nl) ? l : l*nr/nl
446  */
447 static u64 __sched_period(unsigned long nr_running)
448 {
449         u64 period = sysctl_sched_latency;
450         unsigned long nr_latency = sched_nr_latency;
451
452         if (unlikely(nr_running > nr_latency)) {
453                 period = sysctl_sched_min_granularity;
454                 period *= nr_running;
455         }
456
457         return period;
458 }
459
460 /*
461  * We calculate the wall-time slice from the period by taking a part
462  * proportional to the weight.
463  *
464  * s = p*P[w/rw]
465  */
466 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
467 {
468         u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
469
470         for_each_sched_entity(se) {
471                 struct load_weight *load;
472                 struct load_weight lw;
473
474                 cfs_rq = cfs_rq_of(se);
475                 load = &cfs_rq->load;
476
477                 if (unlikely(!se->on_rq)) {
478                         lw = cfs_rq->load;
479
480                         update_load_add(&lw, se->load.weight);
481                         load = &lw;
482                 }
483                 slice = calc_delta_mine(slice, se->load.weight, load);
484         }
485         return slice;
486 }
487
488 /*
489  * We calculate the vruntime slice of a to be inserted task
490  *
491  * vs = s/w
492  */
493 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
494 {
495         return calc_delta_fair(sched_slice(cfs_rq, se), se);
496 }
497
498 /*
499  * Update the current task's runtime statistics. Skip current tasks that
500  * are not in our scheduling class.
501  */
502 static inline void
503 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
504               unsigned long delta_exec)
505 {
506         unsigned long delta_exec_weighted;
507
508         schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
509
510         curr->sum_exec_runtime += delta_exec;
511         schedstat_add(cfs_rq, exec_clock, delta_exec);
512         delta_exec_weighted = calc_delta_fair(delta_exec, curr);
513         curr->vruntime += delta_exec_weighted;
514         update_min_vruntime(cfs_rq);
515 }
516
517 static void update_curr(struct cfs_rq *cfs_rq)
518 {
519         struct sched_entity *curr = cfs_rq->curr;
520         u64 now = rq_of(cfs_rq)->clock;
521         unsigned long delta_exec;
522
523         if (unlikely(!curr))
524                 return;
525
526         /*
527          * Get the amount of time the current task was running
528          * since the last time we changed load (this cannot
529          * overflow on 32 bits):
530          */
531         delta_exec = (unsigned long)(now - curr->exec_start);
532         if (!delta_exec)
533                 return;
534
535         __update_curr(cfs_rq, curr, delta_exec);
536         curr->exec_start = now;
537
538         if (entity_is_task(curr)) {
539                 struct task_struct *curtask = task_of(curr);
540
541                 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
542                 cpuacct_charge(curtask, delta_exec);
543                 account_group_exec_runtime(curtask, delta_exec);
544         }
545 }
546
547 static inline void
548 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
549 {
550         schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
551 }
552
553 /*
554  * Task is being enqueued - update stats:
555  */
556 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
557 {
558         /*
559          * Are we enqueueing a waiting task? (for current tasks
560          * a dequeue/enqueue event is a NOP)
561          */
562         if (se != cfs_rq->curr)
563                 update_stats_wait_start(cfs_rq, se);
564 }
565
566 static void
567 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
568 {
569         schedstat_set(se->wait_max, max(se->wait_max,
570                         rq_of(cfs_rq)->clock - se->wait_start));
571         schedstat_set(se->wait_count, se->wait_count + 1);
572         schedstat_set(se->wait_sum, se->wait_sum +
573                         rq_of(cfs_rq)->clock - se->wait_start);
574 #ifdef CONFIG_SCHEDSTATS
575         if (entity_is_task(se)) {
576                 trace_sched_stat_wait(task_of(se),
577                         rq_of(cfs_rq)->clock - se->wait_start);
578         }
579 #endif
580         schedstat_set(se->wait_start, 0);
581 }
582
583 static inline void
584 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
585 {
586         /*
587          * Mark the end of the wait period if dequeueing a
588          * waiting task:
589          */
590         if (se != cfs_rq->curr)
591                 update_stats_wait_end(cfs_rq, se);
592 }
593
594 /*
595  * We are picking a new current task - update its stats:
596  */
597 static inline void
598 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
599 {
600         /*
601          * We are starting a new run period:
602          */
603         se->exec_start = rq_of(cfs_rq)->clock;
604 }
605
606 /**************************************************
607  * Scheduling class queueing methods:
608  */
609
610 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
611 static void
612 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
613 {
614         cfs_rq->task_weight += weight;
615 }
616 #else
617 static inline void
618 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
619 {
620 }
621 #endif
622
623 static void
624 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
625 {
626         update_load_add(&cfs_rq->load, se->load.weight);
627         if (!parent_entity(se))
628                 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
629         if (entity_is_task(se)) {
630                 add_cfs_task_weight(cfs_rq, se->load.weight);
631                 list_add(&se->group_node, &cfs_rq->tasks);
632         }
633         cfs_rq->nr_running++;
634         se->on_rq = 1;
635 }
636
637 static void
638 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
639 {
640         update_load_sub(&cfs_rq->load, se->load.weight);
641         if (!parent_entity(se))
642                 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
643         if (entity_is_task(se)) {
644                 add_cfs_task_weight(cfs_rq, -se->load.weight);
645                 list_del_init(&se->group_node);
646         }
647         cfs_rq->nr_running--;
648         se->on_rq = 0;
649 }
650
651 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
652 {
653 #ifdef CONFIG_SCHEDSTATS
654         struct task_struct *tsk = NULL;
655
656         if (entity_is_task(se))
657                 tsk = task_of(se);
658
659         if (se->sleep_start) {
660                 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
661
662                 if ((s64)delta < 0)
663                         delta = 0;
664
665                 if (unlikely(delta > se->sleep_max))
666                         se->sleep_max = delta;
667
668                 se->sleep_start = 0;
669                 se->sum_sleep_runtime += delta;
670
671                 if (tsk) {
672                         account_scheduler_latency(tsk, delta >> 10, 1);
673                         trace_sched_stat_sleep(tsk, delta);
674                 }
675         }
676         if (se->block_start) {
677                 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
678
679                 if ((s64)delta < 0)
680                         delta = 0;
681
682                 if (unlikely(delta > se->block_max))
683                         se->block_max = delta;
684
685                 se->block_start = 0;
686                 se->sum_sleep_runtime += delta;
687
688                 if (tsk) {
689                         if (tsk->in_iowait) {
690                                 se->iowait_sum += delta;
691                                 se->iowait_count++;
692                                 trace_sched_stat_iowait(tsk, delta);
693                         }
694
695                         /*
696                          * Blocking time is in units of nanosecs, so shift by
697                          * 20 to get a milliseconds-range estimation of the
698                          * amount of time that the task spent sleeping:
699                          */
700                         if (unlikely(prof_on == SLEEP_PROFILING)) {
701                                 profile_hits(SLEEP_PROFILING,
702                                                 (void *)get_wchan(tsk),
703                                                 delta >> 20);
704                         }
705                         account_scheduler_latency(tsk, delta >> 10, 0);
706                 }
707         }
708 #endif
709 }
710
711 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
712 {
713 #ifdef CONFIG_SCHED_DEBUG
714         s64 d = se->vruntime - cfs_rq->min_vruntime;
715
716         if (d < 0)
717                 d = -d;
718
719         if (d > 3*sysctl_sched_latency)
720                 schedstat_inc(cfs_rq, nr_spread_over);
721 #endif
722 }
723
724 static void
725 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
726 {
727         u64 vruntime = cfs_rq->min_vruntime;
728
729         /*
730          * The 'current' period is already promised to the current tasks,
731          * however the extra weight of the new task will slow them down a
732          * little, place the new task so that it fits in the slot that
733          * stays open at the end.
734          */
735         if (initial && sched_feat(START_DEBIT))
736                 vruntime += sched_vslice(cfs_rq, se);
737
738         /* sleeps up to a single latency don't count. */
739         if (!initial && sched_feat(FAIR_SLEEPERS)) {
740                 unsigned long thresh = sysctl_sched_latency;
741
742                 /*
743                  * Convert the sleeper threshold into virtual time.
744                  * SCHED_IDLE is a special sub-class.  We care about
745                  * fairness only relative to other SCHED_IDLE tasks,
746                  * all of which have the same weight.
747                  */
748                 if (sched_feat(NORMALIZED_SLEEPER) && (!entity_is_task(se) ||
749                                  task_of(se)->policy != SCHED_IDLE))
750                         thresh = calc_delta_fair(thresh, se);
751
752                 /*
753                  * Halve their sleep time's effect, to allow
754                  * for a gentler effect of sleepers:
755                  */
756                 if (sched_feat(GENTLE_FAIR_SLEEPERS))
757                         thresh >>= 1;
758
759                 vruntime -= thresh;
760         }
761
762         /* ensure we never gain time by being placed backwards. */
763         vruntime = max_vruntime(se->vruntime, vruntime);
764
765         se->vruntime = vruntime;
766 }
767
768 static void
769 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
770 {
771         /*
772          * Update run-time statistics of the 'current'.
773          */
774         update_curr(cfs_rq);
775         account_entity_enqueue(cfs_rq, se);
776
777         if (wakeup) {
778                 place_entity(cfs_rq, se, 0);
779                 enqueue_sleeper(cfs_rq, se);
780         }
781
782         update_stats_enqueue(cfs_rq, se);
783         check_spread(cfs_rq, se);
784         if (se != cfs_rq->curr)
785                 __enqueue_entity(cfs_rq, se);
786 }
787
788 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
789 {
790         if (!se || cfs_rq->last == se)
791                 cfs_rq->last = NULL;
792
793         if (!se || cfs_rq->next == se)
794                 cfs_rq->next = NULL;
795 }
796
797 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
798 {
799         for_each_sched_entity(se)
800                 __clear_buddies(cfs_rq_of(se), se);
801 }
802
803 static void
804 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
805 {
806         /*
807          * Update run-time statistics of the 'current'.
808          */
809         update_curr(cfs_rq);
810
811         update_stats_dequeue(cfs_rq, se);
812         if (sleep) {
813 #ifdef CONFIG_SCHEDSTATS
814                 if (entity_is_task(se)) {
815                         struct task_struct *tsk = task_of(se);
816
817                         if (tsk->state & TASK_INTERRUPTIBLE)
818                                 se->sleep_start = rq_of(cfs_rq)->clock;
819                         if (tsk->state & TASK_UNINTERRUPTIBLE)
820                                 se->block_start = rq_of(cfs_rq)->clock;
821                 }
822 #endif
823         }
824
825         clear_buddies(cfs_rq, se);
826
827         if (se != cfs_rq->curr)
828                 __dequeue_entity(cfs_rq, se);
829         account_entity_dequeue(cfs_rq, se);
830         update_min_vruntime(cfs_rq);
831 }
832
833 /*
834  * Preempt the current task with a newly woken task if needed:
835  */
836 static void
837 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
838 {
839         unsigned long ideal_runtime, delta_exec;
840
841         ideal_runtime = sched_slice(cfs_rq, curr);
842         delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
843         if (delta_exec > ideal_runtime) {
844                 resched_task(rq_of(cfs_rq)->curr);
845                 /*
846                  * The current task ran long enough, ensure it doesn't get
847                  * re-elected due to buddy favours.
848                  */
849                 clear_buddies(cfs_rq, curr);
850                 return;
851         }
852
853         /*
854          * Ensure that a task that missed wakeup preemption by a
855          * narrow margin doesn't have to wait for a full slice.
856          * This also mitigates buddy induced latencies under load.
857          */
858         if (!sched_feat(WAKEUP_PREEMPT))
859                 return;
860
861         if (delta_exec < sysctl_sched_min_granularity)
862                 return;
863
864         if (cfs_rq->nr_running > 1) {
865                 struct sched_entity *se = __pick_next_entity(cfs_rq);
866                 s64 delta = curr->vruntime - se->vruntime;
867
868                 if (delta > ideal_runtime)
869                         resched_task(rq_of(cfs_rq)->curr);
870         }
871 }
872
873 static void
874 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
875 {
876         /* 'current' is not kept within the tree. */
877         if (se->on_rq) {
878                 /*
879                  * Any task has to be enqueued before it get to execute on
880                  * a CPU. So account for the time it spent waiting on the
881                  * runqueue.
882                  */
883                 update_stats_wait_end(cfs_rq, se);
884                 __dequeue_entity(cfs_rq, se);
885         }
886
887         update_stats_curr_start(cfs_rq, se);
888         cfs_rq->curr = se;
889 #ifdef CONFIG_SCHEDSTATS
890         /*
891          * Track our maximum slice length, if the CPU's load is at
892          * least twice that of our own weight (i.e. dont track it
893          * when there are only lesser-weight tasks around):
894          */
895         if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
896                 se->slice_max = max(se->slice_max,
897                         se->sum_exec_runtime - se->prev_sum_exec_runtime);
898         }
899 #endif
900         se->prev_sum_exec_runtime = se->sum_exec_runtime;
901 }
902
903 static int
904 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
905
906 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
907 {
908         struct sched_entity *se = __pick_next_entity(cfs_rq);
909         struct sched_entity *left = se;
910
911         if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
912                 se = cfs_rq->next;
913
914         /*
915          * Prefer last buddy, try to return the CPU to a preempted task.
916          */
917         if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
918                 se = cfs_rq->last;
919
920         clear_buddies(cfs_rq, se);
921
922         return se;
923 }
924
925 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
926 {
927         /*
928          * If still on the runqueue then deactivate_task()
929          * was not called and update_curr() has to be done:
930          */
931         if (prev->on_rq)
932                 update_curr(cfs_rq);
933
934         check_spread(cfs_rq, prev);
935         if (prev->on_rq) {
936                 update_stats_wait_start(cfs_rq, prev);
937                 /* Put 'current' back into the tree. */
938                 __enqueue_entity(cfs_rq, prev);
939         }
940         cfs_rq->curr = NULL;
941 }
942
943 static void
944 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
945 {
946         /*
947          * Update run-time statistics of the 'current'.
948          */
949         update_curr(cfs_rq);
950
951 #ifdef CONFIG_SCHED_HRTICK
952         /*
953          * queued ticks are scheduled to match the slice, so don't bother
954          * validating it and just reschedule.
955          */
956         if (queued) {
957                 resched_task(rq_of(cfs_rq)->curr);
958                 return;
959         }
960         /*
961          * don't let the period tick interfere with the hrtick preemption
962          */
963         if (!sched_feat(DOUBLE_TICK) &&
964                         hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
965                 return;
966 #endif
967
968         if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
969                 check_preempt_tick(cfs_rq, curr);
970 }
971
972 /**************************************************
973  * CFS operations on tasks:
974  */
975
976 #ifdef CONFIG_SCHED_HRTICK
977 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
978 {
979         struct sched_entity *se = &p->se;
980         struct cfs_rq *cfs_rq = cfs_rq_of(se);
981
982         WARN_ON(task_rq(p) != rq);
983
984         if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
985                 u64 slice = sched_slice(cfs_rq, se);
986                 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
987                 s64 delta = slice - ran;
988
989                 if (delta < 0) {
990                         if (rq->curr == p)
991                                 resched_task(p);
992                         return;
993                 }
994
995                 /*
996                  * Don't schedule slices shorter than 10000ns, that just
997                  * doesn't make sense. Rely on vruntime for fairness.
998                  */
999                 if (rq->curr != p)
1000                         delta = max_t(s64, 10000LL, delta);
1001
1002                 hrtick_start(rq, delta);
1003         }
1004 }
1005
1006 /*
1007  * called from enqueue/dequeue and updates the hrtick when the
1008  * current task is from our class and nr_running is low enough
1009  * to matter.
1010  */
1011 static void hrtick_update(struct rq *rq)
1012 {
1013         struct task_struct *curr = rq->curr;
1014
1015         if (curr->sched_class != &fair_sched_class)
1016                 return;
1017
1018         if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1019                 hrtick_start_fair(rq, curr);
1020 }
1021 #else /* !CONFIG_SCHED_HRTICK */
1022 static inline void
1023 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1024 {
1025 }
1026
1027 static inline void hrtick_update(struct rq *rq)
1028 {
1029 }
1030 #endif
1031
1032 /*
1033  * The enqueue_task method is called before nr_running is
1034  * increased. Here we update the fair scheduling stats and
1035  * then put the task into the rbtree:
1036  */
1037 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
1038 {
1039         struct cfs_rq *cfs_rq;
1040         struct sched_entity *se = &p->se;
1041
1042         for_each_sched_entity(se) {
1043                 if (se->on_rq)
1044                         break;
1045                 cfs_rq = cfs_rq_of(se);
1046                 enqueue_entity(cfs_rq, se, wakeup);
1047                 wakeup = 1;
1048         }
1049
1050         hrtick_update(rq);
1051 }
1052
1053 /*
1054  * The dequeue_task method is called before nr_running is
1055  * decreased. We remove the task from the rbtree and
1056  * update the fair scheduling stats:
1057  */
1058 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
1059 {
1060         struct cfs_rq *cfs_rq;
1061         struct sched_entity *se = &p->se;
1062
1063         for_each_sched_entity(se) {
1064                 cfs_rq = cfs_rq_of(se);
1065                 dequeue_entity(cfs_rq, se, sleep);
1066                 /* Don't dequeue parent if it has other entities besides us */
1067                 if (cfs_rq->load.weight)
1068                         break;
1069                 sleep = 1;
1070         }
1071
1072         hrtick_update(rq);
1073 }
1074
1075 /*
1076  * sched_yield() support is very simple - we dequeue and enqueue.
1077  *
1078  * If compat_yield is turned on then we requeue to the end of the tree.
1079  */
1080 static void yield_task_fair(struct rq *rq)
1081 {
1082         struct task_struct *curr = rq->curr;
1083         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1084         struct sched_entity *rightmost, *se = &curr->se;
1085
1086         /*
1087          * Are we the only task in the tree?
1088          */
1089         if (unlikely(cfs_rq->nr_running == 1))
1090                 return;
1091
1092         clear_buddies(cfs_rq, se);
1093
1094         if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1095                 update_rq_clock(rq);
1096                 /*
1097                  * Update run-time statistics of the 'current'.
1098                  */
1099                 update_curr(cfs_rq);
1100
1101                 return;
1102         }
1103         /*
1104          * Find the rightmost entry in the rbtree:
1105          */
1106         rightmost = __pick_last_entity(cfs_rq);
1107         /*
1108          * Already in the rightmost position?
1109          */
1110         if (unlikely(!rightmost || entity_before(rightmost, se)))
1111                 return;
1112
1113         /*
1114          * Minimally necessary key value to be last in the tree:
1115          * Upon rescheduling, sched_class::put_prev_task() will place
1116          * 'current' within the tree based on its new key value.
1117          */
1118         se->vruntime = rightmost->vruntime + 1;
1119 }
1120
1121 #ifdef CONFIG_SMP
1122
1123 #ifdef CONFIG_FAIR_GROUP_SCHED
1124 /*
1125  * effective_load() calculates the load change as seen from the root_task_group
1126  *
1127  * Adding load to a group doesn't make a group heavier, but can cause movement
1128  * of group shares between cpus. Assuming the shares were perfectly aligned one
1129  * can calculate the shift in shares.
1130  *
1131  * The problem is that perfectly aligning the shares is rather expensive, hence
1132  * we try to avoid doing that too often - see update_shares(), which ratelimits
1133  * this change.
1134  *
1135  * We compensate this by not only taking the current delta into account, but
1136  * also considering the delta between when the shares were last adjusted and
1137  * now.
1138  *
1139  * We still saw a performance dip, some tracing learned us that between
1140  * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1141  * significantly. Therefore try to bias the error in direction of failing
1142  * the affine wakeup.
1143  *
1144  */
1145 static long effective_load(struct task_group *tg, int cpu,
1146                 long wl, long wg)
1147 {
1148         struct sched_entity *se = tg->se[cpu];
1149
1150         if (!tg->parent)
1151                 return wl;
1152
1153         /*
1154          * By not taking the decrease of shares on the other cpu into
1155          * account our error leans towards reducing the affine wakeups.
1156          */
1157         if (!wl && sched_feat(ASYM_EFF_LOAD))
1158                 return wl;
1159
1160         for_each_sched_entity(se) {
1161                 long S, rw, s, a, b;
1162                 long more_w;
1163
1164                 /*
1165                  * Instead of using this increment, also add the difference
1166                  * between when the shares were last updated and now.
1167                  */
1168                 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1169                 wl += more_w;
1170                 wg += more_w;
1171
1172                 S = se->my_q->tg->shares;
1173                 s = se->my_q->shares;
1174                 rw = se->my_q->rq_weight;
1175
1176                 a = S*(rw + wl);
1177                 b = S*rw + s*wg;
1178
1179                 wl = s*(a-b);
1180
1181                 if (likely(b))
1182                         wl /= b;
1183
1184                 /*
1185                  * Assume the group is already running and will
1186                  * thus already be accounted for in the weight.
1187                  *
1188                  * That is, moving shares between CPUs, does not
1189                  * alter the group weight.
1190                  */
1191                 wg = 0;
1192         }
1193
1194         return wl;
1195 }
1196
1197 #else
1198
1199 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1200                 unsigned long wl, unsigned long wg)
1201 {
1202         return wl;
1203 }
1204
1205 #endif
1206
1207 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1208 {
1209         struct task_struct *curr = current;
1210         unsigned long this_load, load;
1211         int idx, this_cpu, prev_cpu;
1212         unsigned long tl_per_task;
1213         unsigned int imbalance;
1214         struct task_group *tg;
1215         unsigned long weight;
1216         int balanced;
1217
1218         idx       = sd->wake_idx;
1219         this_cpu  = smp_processor_id();
1220         prev_cpu  = task_cpu(p);
1221         load      = source_load(prev_cpu, idx);
1222         this_load = target_load(this_cpu, idx);
1223
1224         if (sync) {
1225                if (sched_feat(SYNC_LESS) &&
1226                    (curr->se.avg_overlap > sysctl_sched_migration_cost ||
1227                     p->se.avg_overlap > sysctl_sched_migration_cost))
1228                        sync = 0;
1229         } else {
1230                 if (sched_feat(SYNC_MORE) &&
1231                     (curr->se.avg_overlap < sysctl_sched_migration_cost &&
1232                      p->se.avg_overlap < sysctl_sched_migration_cost))
1233                         sync = 1;
1234         }
1235
1236         /*
1237          * If sync wakeup then subtract the (maximum possible)
1238          * effect of the currently running task from the load
1239          * of the current CPU:
1240          */
1241         if (sync) {
1242                 tg = task_group(current);
1243                 weight = current->se.load.weight;
1244
1245                 this_load += effective_load(tg, this_cpu, -weight, -weight);
1246                 load += effective_load(tg, prev_cpu, 0, -weight);
1247         }
1248
1249         tg = task_group(p);
1250         weight = p->se.load.weight;
1251
1252         imbalance = 100 + (sd->imbalance_pct - 100) / 2;
1253
1254         /*
1255          * In low-load situations, where prev_cpu is idle and this_cpu is idle
1256          * due to the sync cause above having dropped this_load to 0, we'll
1257          * always have an imbalance, but there's really nothing you can do
1258          * about that, so that's good too.
1259          *
1260          * Otherwise check if either cpus are near enough in load to allow this
1261          * task to be woken on this_cpu.
1262          */
1263         balanced = !this_load ||
1264                 100*(this_load + effective_load(tg, this_cpu, weight, weight)) <=
1265                 imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
1266
1267         /*
1268          * If the currently running task will sleep within
1269          * a reasonable amount of time then attract this newly
1270          * woken task:
1271          */
1272         if (sync && balanced)
1273                 return 1;
1274
1275         schedstat_inc(p, se.nr_wakeups_affine_attempts);
1276         tl_per_task = cpu_avg_load_per_task(this_cpu);
1277
1278         if (balanced ||
1279             (this_load <= load &&
1280              this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1281                 /*
1282                  * This domain has SD_WAKE_AFFINE and
1283                  * p is cache cold in this domain, and
1284                  * there is no bad imbalance.
1285                  */
1286                 schedstat_inc(sd, ttwu_move_affine);
1287                 schedstat_inc(p, se.nr_wakeups_affine);
1288
1289                 return 1;
1290         }
1291         return 0;
1292 }
1293
1294 /*
1295  * find_idlest_group finds and returns the least busy CPU group within the
1296  * domain.
1297  */
1298 static struct sched_group *
1299 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1300                   int this_cpu, int load_idx)
1301 {
1302         struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1303         unsigned long min_load = ULONG_MAX, this_load = 0;
1304         int imbalance = 100 + (sd->imbalance_pct-100)/2;
1305
1306         do {
1307                 unsigned long load, avg_load;
1308                 int local_group;
1309                 int i;
1310
1311                 /* Skip over this group if it has no CPUs allowed */
1312                 if (!cpumask_intersects(sched_group_cpus(group),
1313                                         &p->cpus_allowed))
1314                         continue;
1315
1316                 local_group = cpumask_test_cpu(this_cpu,
1317                                                sched_group_cpus(group));
1318
1319                 /* Tally up the load of all CPUs in the group */
1320                 avg_load = 0;
1321
1322                 for_each_cpu(i, sched_group_cpus(group)) {
1323                         /* Bias balancing toward cpus of our domain */
1324                         if (local_group)
1325                                 load = source_load(i, load_idx);
1326                         else
1327                                 load = target_load(i, load_idx);
1328
1329                         avg_load += load;
1330                 }
1331
1332                 /* Adjust by relative CPU power of the group */
1333                 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1334
1335                 if (local_group) {
1336                         this_load = avg_load;
1337                         this = group;
1338                 } else if (avg_load < min_load) {
1339                         min_load = avg_load;
1340                         idlest = group;
1341                 }
1342         } while (group = group->next, group != sd->groups);
1343
1344         if (!idlest || 100*this_load < imbalance*min_load)
1345                 return NULL;
1346         return idlest;
1347 }
1348
1349 /*
1350  * find_idlest_cpu - find the idlest cpu among the cpus in group.
1351  */
1352 static int
1353 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1354 {
1355         unsigned long load, min_load = ULONG_MAX;
1356         int idlest = -1;
1357         int i;
1358
1359         /* Traverse only the allowed CPUs */
1360         for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1361                 load = weighted_cpuload(i);
1362
1363                 if (load < min_load || (load == min_load && i == this_cpu)) {
1364                         min_load = load;
1365                         idlest = i;
1366                 }
1367         }
1368
1369         return idlest;
1370 }
1371
1372 /*
1373  * Try and locate an idle CPU in the sched_domain.
1374  */
1375 static int
1376 select_idle_sibling(struct task_struct *p, struct sched_domain *sd, int target)
1377 {
1378         int cpu = smp_processor_id();
1379         int prev_cpu = task_cpu(p);
1380         int i;
1381
1382         /*
1383          * If this domain spans both cpu and prev_cpu (see the SD_WAKE_AFFINE
1384          * test in select_task_rq_fair) and the prev_cpu is idle then that's
1385          * always a better target than the current cpu.
1386          */
1387         if (target == cpu && !cpu_rq(prev_cpu)->cfs.nr_running)
1388                 return prev_cpu;
1389
1390         /*
1391          * Otherwise, iterate the domain and find an elegible idle cpu.
1392          */
1393         for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1394                 if (!cpu_rq(i)->cfs.nr_running) {
1395                         target = i;
1396                         break;
1397                 }
1398         }
1399
1400         return target;
1401 }
1402
1403 /*
1404  * sched_balance_self: balance the current task (running on cpu) in domains
1405  * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1406  * SD_BALANCE_EXEC.
1407  *
1408  * Balance, ie. select the least loaded group.
1409  *
1410  * Returns the target CPU number, or the same CPU if no balancing is needed.
1411  *
1412  * preempt must be disabled.
1413  */
1414 static int select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
1415 {
1416         struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1417         int cpu = smp_processor_id();
1418         int prev_cpu = task_cpu(p);
1419         int new_cpu = cpu;
1420         int want_affine = 0;
1421         int want_sd = 1;
1422         int sync = wake_flags & WF_SYNC;
1423
1424         if (sd_flag & SD_BALANCE_WAKE) {
1425                 if (sched_feat(AFFINE_WAKEUPS) &&
1426                     cpumask_test_cpu(cpu, &p->cpus_allowed))
1427                         want_affine = 1;
1428                 new_cpu = prev_cpu;
1429         }
1430
1431         for_each_domain(cpu, tmp) {
1432                 if (!(tmp->flags & SD_LOAD_BALANCE))
1433                         continue;
1434
1435                 /*
1436                  * If power savings logic is enabled for a domain, see if we
1437                  * are not overloaded, if so, don't balance wider.
1438                  */
1439                 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1440                         unsigned long power = 0;
1441                         unsigned long nr_running = 0;
1442                         unsigned long capacity;
1443                         int i;
1444
1445                         for_each_cpu(i, sched_domain_span(tmp)) {
1446                                 power += power_of(i);
1447                                 nr_running += cpu_rq(i)->cfs.nr_running;
1448                         }
1449
1450                         capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1451
1452                         if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1453                                 nr_running /= 2;
1454
1455                         if (nr_running < capacity)
1456                                 want_sd = 0;
1457                 }
1458
1459                 /*
1460                  * While iterating the domains looking for a spanning
1461                  * WAKE_AFFINE domain, adjust the affine target to any idle cpu
1462                  * in cache sharing domains along the way.
1463                  */
1464                 if (want_affine) {
1465                         int target = -1;
1466
1467                         /*
1468                          * If both cpu and prev_cpu are part of this domain,
1469                          * cpu is a valid SD_WAKE_AFFINE target.
1470                          */
1471                         if (cpumask_test_cpu(prev_cpu, sched_domain_span(tmp)))
1472                                 target = cpu;
1473
1474                         /*
1475                          * If there's an idle sibling in this domain, make that
1476                          * the wake_affine target instead of the current cpu.
1477                          */
1478                         if (tmp->flags & SD_PREFER_SIBLING)
1479                                 target = select_idle_sibling(p, tmp, target);
1480
1481                         if (target >= 0) {
1482                                 if (tmp->flags & SD_WAKE_AFFINE) {
1483                                         affine_sd = tmp;
1484                                         want_affine = 0;
1485                                 }
1486                                 cpu = target;
1487                         }
1488                 }
1489
1490                 if (!want_sd && !want_affine)
1491                         break;
1492
1493                 if (!(tmp->flags & sd_flag))
1494                         continue;
1495
1496                 if (want_sd)
1497                         sd = tmp;
1498         }
1499
1500         if (sched_feat(LB_SHARES_UPDATE)) {
1501                 /*
1502                  * Pick the largest domain to update shares over
1503                  */
1504                 tmp = sd;
1505                 if (affine_sd && (!tmp ||
1506                                   cpumask_weight(sched_domain_span(affine_sd)) >
1507                                   cpumask_weight(sched_domain_span(sd))))
1508                         tmp = affine_sd;
1509
1510                 if (tmp)
1511                         update_shares(tmp);
1512         }
1513
1514         if (affine_sd && wake_affine(affine_sd, p, sync))
1515                 return cpu;
1516
1517         while (sd) {
1518                 int load_idx = sd->forkexec_idx;
1519                 struct sched_group *group;
1520                 int weight;
1521
1522                 if (!(sd->flags & sd_flag)) {
1523                         sd = sd->child;
1524                         continue;
1525                 }
1526
1527                 if (sd_flag & SD_BALANCE_WAKE)
1528                         load_idx = sd->wake_idx;
1529
1530                 group = find_idlest_group(sd, p, cpu, load_idx);
1531                 if (!group) {
1532                         sd = sd->child;
1533                         continue;
1534                 }
1535
1536                 new_cpu = find_idlest_cpu(group, p, cpu);
1537                 if (new_cpu == -1 || new_cpu == cpu) {
1538                         /* Now try balancing at a lower domain level of cpu */
1539                         sd = sd->child;
1540                         continue;
1541                 }
1542
1543                 /* Now try balancing at a lower domain level of new_cpu */
1544                 cpu = new_cpu;
1545                 weight = cpumask_weight(sched_domain_span(sd));
1546                 sd = NULL;
1547                 for_each_domain(cpu, tmp) {
1548                         if (weight <= cpumask_weight(sched_domain_span(tmp)))
1549                                 break;
1550                         if (tmp->flags & sd_flag)
1551                                 sd = tmp;
1552                 }
1553                 /* while loop will break here if sd == NULL */
1554         }
1555
1556         return new_cpu;
1557 }
1558 #endif /* CONFIG_SMP */
1559
1560 /*
1561  * Adaptive granularity
1562  *
1563  * se->avg_wakeup gives the average time a task runs until it does a wakeup,
1564  * with the limit of wakeup_gran -- when it never does a wakeup.
1565  *
1566  * So the smaller avg_wakeup is the faster we want this task to preempt,
1567  * but we don't want to treat the preemptee unfairly and therefore allow it
1568  * to run for at least the amount of time we'd like to run.
1569  *
1570  * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one
1571  *
1572  * NOTE: we use *nr_running to scale with load, this nicely matches the
1573  *       degrading latency on load.
1574  */
1575 static unsigned long
1576 adaptive_gran(struct sched_entity *curr, struct sched_entity *se)
1577 {
1578         u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1579         u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running;
1580         u64 gran = 0;
1581
1582         if (this_run < expected_wakeup)
1583                 gran = expected_wakeup - this_run;
1584
1585         return min_t(s64, gran, sysctl_sched_wakeup_granularity);
1586 }
1587
1588 static unsigned long
1589 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1590 {
1591         unsigned long gran = sysctl_sched_wakeup_granularity;
1592
1593         if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN))
1594                 gran = adaptive_gran(curr, se);
1595
1596         /*
1597          * Since its curr running now, convert the gran from real-time
1598          * to virtual-time in his units.
1599          */
1600         if (sched_feat(ASYM_GRAN)) {
1601                 /*
1602                  * By using 'se' instead of 'curr' we penalize light tasks, so
1603                  * they get preempted easier. That is, if 'se' < 'curr' then
1604                  * the resulting gran will be larger, therefore penalizing the
1605                  * lighter, if otoh 'se' > 'curr' then the resulting gran will
1606                  * be smaller, again penalizing the lighter task.
1607                  *
1608                  * This is especially important for buddies when the leftmost
1609                  * task is higher priority than the buddy.
1610                  */
1611                 if (unlikely(se->load.weight != NICE_0_LOAD))
1612                         gran = calc_delta_fair(gran, se);
1613         } else {
1614                 if (unlikely(curr->load.weight != NICE_0_LOAD))
1615                         gran = calc_delta_fair(gran, curr);
1616         }
1617
1618         return gran;
1619 }
1620
1621 /*
1622  * Should 'se' preempt 'curr'.
1623  *
1624  *             |s1
1625  *        |s2
1626  *   |s3
1627  *         g
1628  *      |<--->|c
1629  *
1630  *  w(c, s1) = -1
1631  *  w(c, s2) =  0
1632  *  w(c, s3) =  1
1633  *
1634  */
1635 static int
1636 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1637 {
1638         s64 gran, vdiff = curr->vruntime - se->vruntime;
1639
1640         if (vdiff <= 0)
1641                 return -1;
1642
1643         gran = wakeup_gran(curr, se);
1644         if (vdiff > gran)
1645                 return 1;
1646
1647         return 0;
1648 }
1649
1650 static void set_last_buddy(struct sched_entity *se)
1651 {
1652         if (likely(task_of(se)->policy != SCHED_IDLE)) {
1653                 for_each_sched_entity(se)
1654                         cfs_rq_of(se)->last = se;
1655         }
1656 }
1657
1658 static void set_next_buddy(struct sched_entity *se)
1659 {
1660         if (likely(task_of(se)->policy != SCHED_IDLE)) {
1661                 for_each_sched_entity(se)
1662                         cfs_rq_of(se)->next = se;
1663         }
1664 }
1665
1666 /*
1667  * Preempt the current task with a newly woken task if needed:
1668  */
1669 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1670 {
1671         struct task_struct *curr = rq->curr;
1672         struct sched_entity *se = &curr->se, *pse = &p->se;
1673         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1674         int sync = wake_flags & WF_SYNC;
1675         int scale = cfs_rq->nr_running >= sched_nr_latency;
1676
1677         if (unlikely(rt_prio(p->prio)))
1678                 goto preempt;
1679
1680         if (unlikely(p->sched_class != &fair_sched_class))
1681                 return;
1682
1683         if (unlikely(se == pse))
1684                 return;
1685
1686         if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK))
1687                 set_next_buddy(pse);
1688
1689         /*
1690          * We can come here with TIF_NEED_RESCHED already set from new task
1691          * wake up path.
1692          */
1693         if (test_tsk_need_resched(curr))
1694                 return;
1695
1696         /*
1697          * Batch and idle tasks do not preempt (their preemption is driven by
1698          * the tick):
1699          */
1700         if (unlikely(p->policy != SCHED_NORMAL))
1701                 return;
1702
1703         /* Idle tasks are by definition preempted by everybody. */
1704         if (unlikely(curr->policy == SCHED_IDLE))
1705                 goto preempt;
1706
1707         if (sched_feat(WAKEUP_SYNC) && sync)
1708                 goto preempt;
1709
1710         if (sched_feat(WAKEUP_OVERLAP) &&
1711                         se->avg_overlap < sysctl_sched_migration_cost &&
1712                         pse->avg_overlap < sysctl_sched_migration_cost)
1713                 goto preempt;
1714
1715         if (!sched_feat(WAKEUP_PREEMPT))
1716                 return;
1717
1718         update_curr(cfs_rq);
1719         find_matching_se(&se, &pse);
1720         BUG_ON(!pse);
1721         if (wakeup_preempt_entity(se, pse) == 1)
1722                 goto preempt;
1723
1724         return;
1725
1726 preempt:
1727         resched_task(curr);
1728         /*
1729          * Only set the backward buddy when the current task is still
1730          * on the rq. This can happen when a wakeup gets interleaved
1731          * with schedule on the ->pre_schedule() or idle_balance()
1732          * point, either of which can * drop the rq lock.
1733          *
1734          * Also, during early boot the idle thread is in the fair class,
1735          * for obvious reasons its a bad idea to schedule back to it.
1736          */
1737         if (unlikely(!se->on_rq || curr == rq->idle))
1738                 return;
1739
1740         if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1741                 set_last_buddy(se);
1742 }
1743
1744 static struct task_struct *pick_next_task_fair(struct rq *rq)
1745 {
1746         struct task_struct *p;
1747         struct cfs_rq *cfs_rq = &rq->cfs;
1748         struct sched_entity *se;
1749
1750         if (!cfs_rq->nr_running)
1751                 return NULL;
1752
1753         do {
1754                 se = pick_next_entity(cfs_rq);
1755                 set_next_entity(cfs_rq, se);
1756                 cfs_rq = group_cfs_rq(se);
1757         } while (cfs_rq);
1758
1759         p = task_of(se);
1760         hrtick_start_fair(rq, p);
1761
1762         return p;
1763 }
1764
1765 /*
1766  * Account for a descheduled task:
1767  */
1768 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1769 {
1770         struct sched_entity *se = &prev->se;
1771         struct cfs_rq *cfs_rq;
1772
1773         for_each_sched_entity(se) {
1774                 cfs_rq = cfs_rq_of(se);
1775                 put_prev_entity(cfs_rq, se);
1776         }
1777 }
1778
1779 #ifdef CONFIG_SMP
1780 /**************************************************
1781  * Fair scheduling class load-balancing methods:
1782  */
1783
1784 /*
1785  * Load-balancing iterator. Note: while the runqueue stays locked
1786  * during the whole iteration, the current task might be
1787  * dequeued so the iterator has to be dequeue-safe. Here we
1788  * achieve that by always pre-iterating before returning
1789  * the current task:
1790  */
1791 static struct task_struct *
1792 __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
1793 {
1794         struct task_struct *p = NULL;
1795         struct sched_entity *se;
1796
1797         if (next == &cfs_rq->tasks)
1798                 return NULL;
1799
1800         se = list_entry(next, struct sched_entity, group_node);
1801         p = task_of(se);
1802         cfs_rq->balance_iterator = next->next;
1803
1804         return p;
1805 }
1806
1807 static struct task_struct *load_balance_start_fair(void *arg)
1808 {
1809         struct cfs_rq *cfs_rq = arg;
1810
1811         return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
1812 }
1813
1814 static struct task_struct *load_balance_next_fair(void *arg)
1815 {
1816         struct cfs_rq *cfs_rq = arg;
1817
1818         return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
1819 }
1820
1821 static unsigned long
1822 __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1823                 unsigned long max_load_move, struct sched_domain *sd,
1824                 enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
1825                 struct cfs_rq *cfs_rq)
1826 {
1827         struct rq_iterator cfs_rq_iterator;
1828
1829         cfs_rq_iterator.start = load_balance_start_fair;
1830         cfs_rq_iterator.next = load_balance_next_fair;
1831         cfs_rq_iterator.arg = cfs_rq;
1832
1833         return balance_tasks(this_rq, this_cpu, busiest,
1834                         max_load_move, sd, idle, all_pinned,
1835                         this_best_prio, &cfs_rq_iterator);
1836 }
1837
1838 #ifdef CONFIG_FAIR_GROUP_SCHED
1839 static unsigned long
1840 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1841                   unsigned long max_load_move,
1842                   struct sched_domain *sd, enum cpu_idle_type idle,
1843                   int *all_pinned, int *this_best_prio)
1844 {
1845         long rem_load_move = max_load_move;
1846         int busiest_cpu = cpu_of(busiest);
1847         struct task_group *tg;
1848
1849         rcu_read_lock();
1850         update_h_load(busiest_cpu);
1851
1852         list_for_each_entry_rcu(tg, &task_groups, list) {
1853                 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1854                 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1855                 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1856                 u64 rem_load, moved_load;
1857
1858                 /*
1859                  * empty group
1860                  */
1861                 if (!busiest_cfs_rq->task_weight)
1862                         continue;
1863
1864                 rem_load = (u64)rem_load_move * busiest_weight;
1865                 rem_load = div_u64(rem_load, busiest_h_load + 1);
1866
1867                 moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
1868                                 rem_load, sd, idle, all_pinned, this_best_prio,
1869                                 tg->cfs_rq[busiest_cpu]);
1870
1871                 if (!moved_load)
1872                         continue;
1873
1874                 moved_load *= busiest_h_load;
1875                 moved_load = div_u64(moved_load, busiest_weight + 1);
1876
1877                 rem_load_move -= moved_load;
1878                 if (rem_load_move < 0)
1879                         break;
1880         }
1881         rcu_read_unlock();
1882
1883         return max_load_move - rem_load_move;
1884 }
1885 #else
1886 static unsigned long
1887 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1888                   unsigned long max_load_move,
1889                   struct sched_domain *sd, enum cpu_idle_type idle,
1890                   int *all_pinned, int *this_best_prio)
1891 {
1892         return __load_balance_fair(this_rq, this_cpu, busiest,
1893                         max_load_move, sd, idle, all_pinned,
1894                         this_best_prio, &busiest->cfs);
1895 }
1896 #endif
1897
1898 static int
1899 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1900                    struct sched_domain *sd, enum cpu_idle_type idle)
1901 {
1902         struct cfs_rq *busy_cfs_rq;
1903         struct rq_iterator cfs_rq_iterator;
1904
1905         cfs_rq_iterator.start = load_balance_start_fair;
1906         cfs_rq_iterator.next = load_balance_next_fair;
1907
1908         for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1909                 /*
1910                  * pass busy_cfs_rq argument into
1911                  * load_balance_[start|next]_fair iterators
1912                  */
1913                 cfs_rq_iterator.arg = busy_cfs_rq;
1914                 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1915                                        &cfs_rq_iterator))
1916                     return 1;
1917         }
1918
1919         return 0;
1920 }
1921
1922 static void rq_online_fair(struct rq *rq)
1923 {
1924         update_sysctl();
1925 }
1926
1927 static void rq_offline_fair(struct rq *rq)
1928 {
1929         update_sysctl();
1930 }
1931
1932 #endif /* CONFIG_SMP */
1933
1934 /*
1935  * scheduler tick hitting a task of our scheduling class:
1936  */
1937 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1938 {
1939         struct cfs_rq *cfs_rq;
1940         struct sched_entity *se = &curr->se;
1941
1942         for_each_sched_entity(se) {
1943                 cfs_rq = cfs_rq_of(se);
1944                 entity_tick(cfs_rq, se, queued);
1945         }
1946 }
1947
1948 /*
1949  * called on fork with the child task as argument from the parent's context
1950  *  - child not yet on the tasklist
1951  *  - preemption disabled
1952  */
1953 static void task_fork_fair(struct task_struct *p)
1954 {
1955         struct cfs_rq *cfs_rq = task_cfs_rq(current);
1956         struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1957         int this_cpu = smp_processor_id();
1958         struct rq *rq = this_rq();
1959         unsigned long flags;
1960
1961         raw_spin_lock_irqsave(&rq->lock, flags);
1962
1963         if (unlikely(task_cpu(p) != this_cpu))
1964                 __set_task_cpu(p, this_cpu);
1965
1966         update_curr(cfs_rq);
1967
1968         if (curr)
1969                 se->vruntime = curr->vruntime;
1970         place_entity(cfs_rq, se, 1);
1971
1972         if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
1973                 /*
1974                  * Upon rescheduling, sched_class::put_prev_task() will place
1975                  * 'current' within the tree based on its new key value.
1976                  */
1977                 swap(curr->vruntime, se->vruntime);
1978                 resched_task(rq->curr);
1979         }
1980
1981         raw_spin_unlock_irqrestore(&rq->lock, flags);
1982 }
1983
1984 /*
1985  * Priority of the task has changed. Check to see if we preempt
1986  * the current task.
1987  */
1988 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1989                               int oldprio, int running)
1990 {
1991         /*
1992          * Reschedule if we are currently running on this runqueue and
1993          * our priority decreased, or if we are not currently running on
1994          * this runqueue and our priority is higher than the current's
1995          */
1996         if (running) {
1997                 if (p->prio > oldprio)
1998                         resched_task(rq->curr);
1999         } else
2000                 check_preempt_curr(rq, p, 0);
2001 }
2002
2003 /*
2004  * We switched to the sched_fair class.
2005  */
2006 static void switched_to_fair(struct rq *rq, struct task_struct *p,
2007                              int running)
2008 {
2009         /*
2010          * We were most likely switched from sched_rt, so
2011          * kick off the schedule if running, otherwise just see
2012          * if we can still preempt the current task.
2013          */
2014         if (running)
2015                 resched_task(rq->curr);
2016         else
2017                 check_preempt_curr(rq, p, 0);
2018 }
2019
2020 /* Account for a task changing its policy or group.
2021  *
2022  * This routine is mostly called to set cfs_rq->curr field when a task
2023  * migrates between groups/classes.
2024  */
2025 static void set_curr_task_fair(struct rq *rq)
2026 {
2027         struct sched_entity *se = &rq->curr->se;
2028
2029         for_each_sched_entity(se)
2030                 set_next_entity(cfs_rq_of(se), se);
2031 }
2032
2033 #ifdef CONFIG_FAIR_GROUP_SCHED
2034 static void moved_group_fair(struct task_struct *p)
2035 {
2036         struct cfs_rq *cfs_rq = task_cfs_rq(p);
2037
2038         update_curr(cfs_rq);
2039         place_entity(cfs_rq, &p->se, 1);
2040 }
2041 #endif
2042
2043 unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
2044 {
2045         struct sched_entity *se = &task->se;
2046         unsigned int rr_interval = 0;
2047
2048         /*
2049          * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
2050          * idle runqueue:
2051          */
2052         if (rq->cfs.load.weight)
2053                 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
2054
2055         return rr_interval;
2056 }
2057
2058 /*
2059  * All the scheduling class methods:
2060  */
2061 static const struct sched_class fair_sched_class = {
2062         .next                   = &idle_sched_class,
2063         .enqueue_task           = enqueue_task_fair,
2064         .dequeue_task           = dequeue_task_fair,
2065         .yield_task             = yield_task_fair,
2066
2067         .check_preempt_curr     = check_preempt_wakeup,
2068
2069         .pick_next_task         = pick_next_task_fair,
2070         .put_prev_task          = put_prev_task_fair,
2071
2072 #ifdef CONFIG_SMP
2073         .select_task_rq         = select_task_rq_fair,
2074
2075         .load_balance           = load_balance_fair,
2076         .move_one_task          = move_one_task_fair,
2077         .rq_online              = rq_online_fair,
2078         .rq_offline             = rq_offline_fair,
2079 #endif
2080
2081         .set_curr_task          = set_curr_task_fair,
2082         .task_tick              = task_tick_fair,
2083         .task_fork              = task_fork_fair,
2084
2085         .prio_changed           = prio_changed_fair,
2086         .switched_to            = switched_to_fair,
2087
2088         .get_rr_interval        = get_rr_interval_fair,
2089
2090 #ifdef CONFIG_FAIR_GROUP_SCHED
2091         .moved_group            = moved_group_fair,
2092 #endif
2093 };
2094
2095 #ifdef CONFIG_SCHED_DEBUG
2096 static void print_cfs_stats(struct seq_file *m, int cpu)
2097 {
2098         struct cfs_rq *cfs_rq;
2099
2100         rcu_read_lock();
2101         for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
2102                 print_cfs_rq(m, cpu, cfs_rq);
2103         rcu_read_unlock();
2104 }
2105 #endif