arm: tegra: raise cpu floor when display is on
[linux-2.6.git] / kernel / sched_rt.c
1 /*
2  * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
3  * policies)
4  */
5
6 #ifdef CONFIG_RT_GROUP_SCHED
7
8 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
9
10 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
11 {
12 #ifdef CONFIG_SCHED_DEBUG
13         WARN_ON_ONCE(!rt_entity_is_task(rt_se));
14 #endif
15         return container_of(rt_se, struct task_struct, rt);
16 }
17
18 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
19 {
20         return rt_rq->rq;
21 }
22
23 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
24 {
25         return rt_se->rt_rq;
26 }
27
28 #else /* CONFIG_RT_GROUP_SCHED */
29
30 #define rt_entity_is_task(rt_se) (1)
31
32 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
33 {
34         return container_of(rt_se, struct task_struct, rt);
35 }
36
37 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
38 {
39         return container_of(rt_rq, struct rq, rt);
40 }
41
42 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
43 {
44         struct task_struct *p = rt_task_of(rt_se);
45         struct rq *rq = task_rq(p);
46
47         return &rq->rt;
48 }
49
50 #endif /* CONFIG_RT_GROUP_SCHED */
51
52 #ifdef CONFIG_SMP
53
54 static inline int rt_overloaded(struct rq *rq)
55 {
56         return atomic_read(&rq->rd->rto_count);
57 }
58
59 static inline void rt_set_overload(struct rq *rq)
60 {
61         if (!rq->online)
62                 return;
63
64         cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
65         /*
66          * Make sure the mask is visible before we set
67          * the overload count. That is checked to determine
68          * if we should look at the mask. It would be a shame
69          * if we looked at the mask, but the mask was not
70          * updated yet.
71          */
72         wmb();
73         atomic_inc(&rq->rd->rto_count);
74 }
75
76 static inline void rt_clear_overload(struct rq *rq)
77 {
78         if (!rq->online)
79                 return;
80
81         /* the order here really doesn't matter */
82         atomic_dec(&rq->rd->rto_count);
83         cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
84 }
85
86 static void update_rt_migration(struct rt_rq *rt_rq)
87 {
88         if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
89                 if (!rt_rq->overloaded) {
90                         rt_set_overload(rq_of_rt_rq(rt_rq));
91                         rt_rq->overloaded = 1;
92                 }
93         } else if (rt_rq->overloaded) {
94                 rt_clear_overload(rq_of_rt_rq(rt_rq));
95                 rt_rq->overloaded = 0;
96         }
97 }
98
99 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
100 {
101         if (!rt_entity_is_task(rt_se))
102                 return;
103
104         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
105
106         rt_rq->rt_nr_total++;
107         if (rt_se->nr_cpus_allowed > 1)
108                 rt_rq->rt_nr_migratory++;
109
110         update_rt_migration(rt_rq);
111 }
112
113 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
114 {
115         if (!rt_entity_is_task(rt_se))
116                 return;
117
118         rt_rq = &rq_of_rt_rq(rt_rq)->rt;
119
120         rt_rq->rt_nr_total--;
121         if (rt_se->nr_cpus_allowed > 1)
122                 rt_rq->rt_nr_migratory--;
123
124         update_rt_migration(rt_rq);
125 }
126
127 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
128 {
129         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
130         plist_node_init(&p->pushable_tasks, p->prio);
131         plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
132 }
133
134 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
135 {
136         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
137 }
138
139 static inline int has_pushable_tasks(struct rq *rq)
140 {
141         return !plist_head_empty(&rq->rt.pushable_tasks);
142 }
143
144 #else
145
146 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
147 {
148 }
149
150 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
151 {
152 }
153
154 static inline
155 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
156 {
157 }
158
159 static inline
160 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
161 {
162 }
163
164 #endif /* CONFIG_SMP */
165
166 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
167 {
168         return !list_empty(&rt_se->run_list);
169 }
170
171 #ifdef CONFIG_RT_GROUP_SCHED
172
173 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
174 {
175         if (!rt_rq->tg)
176                 return RUNTIME_INF;
177
178         return rt_rq->rt_runtime;
179 }
180
181 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
182 {
183         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
184 }
185
186 typedef struct task_group *rt_rq_iter_t;
187
188 static inline struct task_group *next_task_group(struct task_group *tg)
189 {
190         do {
191                 tg = list_entry_rcu(tg->list.next,
192                         typeof(struct task_group), list);
193         } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
194
195         if (&tg->list == &task_groups)
196                 tg = NULL;
197
198         return tg;
199 }
200
201 #define for_each_rt_rq(rt_rq, iter, rq)                                 \
202         for (iter = container_of(&task_groups, typeof(*iter), list);    \
203                 (iter = next_task_group(iter)) &&                       \
204                 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
205
206 static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
207 {
208         list_add_rcu(&rt_rq->leaf_rt_rq_list,
209                         &rq_of_rt_rq(rt_rq)->leaf_rt_rq_list);
210 }
211
212 static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
213 {
214         list_del_rcu(&rt_rq->leaf_rt_rq_list);
215 }
216
217 #define for_each_leaf_rt_rq(rt_rq, rq) \
218         list_for_each_entry_rcu(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
219
220 #define for_each_sched_rt_entity(rt_se) \
221         for (; rt_se; rt_se = rt_se->parent)
222
223 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
224 {
225         return rt_se->my_q;
226 }
227
228 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
229 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
230
231 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
232 {
233         struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
234         struct sched_rt_entity *rt_se;
235
236         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
237
238         rt_se = rt_rq->tg->rt_se[cpu];
239
240         if (rt_rq->rt_nr_running) {
241                 if (rt_se && !on_rt_rq(rt_se))
242                         enqueue_rt_entity(rt_se, false);
243                 if (rt_rq->highest_prio.curr < curr->prio)
244                         resched_task(curr);
245         }
246 }
247
248 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
249 {
250         struct sched_rt_entity *rt_se;
251         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
252
253         rt_se = rt_rq->tg->rt_se[cpu];
254
255         if (rt_se && on_rt_rq(rt_se))
256                 dequeue_rt_entity(rt_se);
257 }
258
259 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
260 {
261         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
262 }
263
264 static int rt_se_boosted(struct sched_rt_entity *rt_se)
265 {
266         struct rt_rq *rt_rq = group_rt_rq(rt_se);
267         struct task_struct *p;
268
269         if (rt_rq)
270                 return !!rt_rq->rt_nr_boosted;
271
272         p = rt_task_of(rt_se);
273         return p->prio != p->normal_prio;
274 }
275
276 #ifdef CONFIG_SMP
277 static inline const struct cpumask *sched_rt_period_mask(void)
278 {
279         return cpu_rq(smp_processor_id())->rd->span;
280 }
281 #else
282 static inline const struct cpumask *sched_rt_period_mask(void)
283 {
284         return cpu_online_mask;
285 }
286 #endif
287
288 static inline
289 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
290 {
291         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
292 }
293
294 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
295 {
296         return &rt_rq->tg->rt_bandwidth;
297 }
298
299 #else /* !CONFIG_RT_GROUP_SCHED */
300
301 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
302 {
303         return rt_rq->rt_runtime;
304 }
305
306 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
307 {
308         return ktime_to_ns(def_rt_bandwidth.rt_period);
309 }
310
311 typedef struct rt_rq *rt_rq_iter_t;
312
313 #define for_each_rt_rq(rt_rq, iter, rq) \
314         for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
315
316 static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
317 {
318 }
319
320 static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
321 {
322 }
323
324 #define for_each_leaf_rt_rq(rt_rq, rq) \
325         for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
326
327 #define for_each_sched_rt_entity(rt_se) \
328         for (; rt_se; rt_se = NULL)
329
330 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
331 {
332         return NULL;
333 }
334
335 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
336 {
337         if (rt_rq->rt_nr_running)
338                 resched_task(rq_of_rt_rq(rt_rq)->curr);
339 }
340
341 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
342 {
343 }
344
345 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
346 {
347         return rt_rq->rt_throttled;
348 }
349
350 static inline const struct cpumask *sched_rt_period_mask(void)
351 {
352         return cpu_online_mask;
353 }
354
355 static inline
356 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
357 {
358         return &cpu_rq(cpu)->rt;
359 }
360
361 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
362 {
363         return &def_rt_bandwidth;
364 }
365
366 #endif /* CONFIG_RT_GROUP_SCHED */
367
368 #ifdef CONFIG_SMP
369 /*
370  * We ran out of runtime, see if we can borrow some from our neighbours.
371  */
372 static int do_balance_runtime(struct rt_rq *rt_rq)
373 {
374         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
375         struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
376         int i, weight, more = 0;
377         u64 rt_period;
378
379         weight = cpumask_weight(rd->span);
380
381         raw_spin_lock(&rt_b->rt_runtime_lock);
382         rt_period = ktime_to_ns(rt_b->rt_period);
383         for_each_cpu(i, rd->span) {
384                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
385                 s64 diff;
386
387                 if (iter == rt_rq)
388                         continue;
389
390                 raw_spin_lock(&iter->rt_runtime_lock);
391                 /*
392                  * Either all rqs have inf runtime and there's nothing to steal
393                  * or __disable_runtime() below sets a specific rq to inf to
394                  * indicate its been disabled and disalow stealing.
395                  */
396                 if (iter->rt_runtime == RUNTIME_INF)
397                         goto next;
398
399                 /*
400                  * From runqueues with spare time, take 1/n part of their
401                  * spare time, but no more than our period.
402                  */
403                 diff = iter->rt_runtime - iter->rt_time;
404                 if (diff > 0) {
405                         diff = div_u64((u64)diff, weight);
406                         if (rt_rq->rt_runtime + diff > rt_period)
407                                 diff = rt_period - rt_rq->rt_runtime;
408                         iter->rt_runtime -= diff;
409                         rt_rq->rt_runtime += diff;
410                         more = 1;
411                         if (rt_rq->rt_runtime == rt_period) {
412                                 raw_spin_unlock(&iter->rt_runtime_lock);
413                                 break;
414                         }
415                 }
416 next:
417                 raw_spin_unlock(&iter->rt_runtime_lock);
418         }
419         raw_spin_unlock(&rt_b->rt_runtime_lock);
420
421         return more;
422 }
423
424 /*
425  * Ensure this RQ takes back all the runtime it lend to its neighbours.
426  */
427 static void __disable_runtime(struct rq *rq)
428 {
429         struct root_domain *rd = rq->rd;
430         rt_rq_iter_t iter;
431         struct rt_rq *rt_rq;
432
433         if (unlikely(!scheduler_running))
434                 return;
435
436         for_each_rt_rq(rt_rq, iter, rq) {
437                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
438                 s64 want;
439                 int i;
440
441                 raw_spin_lock(&rt_b->rt_runtime_lock);
442                 raw_spin_lock(&rt_rq->rt_runtime_lock);
443                 /*
444                  * Either we're all inf and nobody needs to borrow, or we're
445                  * already disabled and thus have nothing to do, or we have
446                  * exactly the right amount of runtime to take out.
447                  */
448                 if (rt_rq->rt_runtime == RUNTIME_INF ||
449                                 rt_rq->rt_runtime == rt_b->rt_runtime)
450                         goto balanced;
451                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
452
453                 /*
454                  * Calculate the difference between what we started out with
455                  * and what we current have, that's the amount of runtime
456                  * we lend and now have to reclaim.
457                  */
458                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
459
460                 /*
461                  * Greedy reclaim, take back as much as we can.
462                  */
463                 for_each_cpu(i, rd->span) {
464                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
465                         s64 diff;
466
467                         /*
468                          * Can't reclaim from ourselves or disabled runqueues.
469                          */
470                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
471                                 continue;
472
473                         raw_spin_lock(&iter->rt_runtime_lock);
474                         if (want > 0) {
475                                 diff = min_t(s64, iter->rt_runtime, want);
476                                 iter->rt_runtime -= diff;
477                                 want -= diff;
478                         } else {
479                                 iter->rt_runtime -= want;
480                                 want -= want;
481                         }
482                         raw_spin_unlock(&iter->rt_runtime_lock);
483
484                         if (!want)
485                                 break;
486                 }
487
488                 raw_spin_lock(&rt_rq->rt_runtime_lock);
489                 /*
490                  * We cannot be left wanting - that would mean some runtime
491                  * leaked out of the system.
492                  */
493                 BUG_ON(want);
494 balanced:
495                 /*
496                  * Disable all the borrow logic by pretending we have inf
497                  * runtime - in which case borrowing doesn't make sense.
498                  */
499                 rt_rq->rt_runtime = RUNTIME_INF;
500                 rt_rq->rt_throttled = 0;
501                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
502                 raw_spin_unlock(&rt_b->rt_runtime_lock);
503         }
504 }
505
506 static void disable_runtime(struct rq *rq)
507 {
508         unsigned long flags;
509
510         raw_spin_lock_irqsave(&rq->lock, flags);
511         __disable_runtime(rq);
512         raw_spin_unlock_irqrestore(&rq->lock, flags);
513 }
514
515 static void __enable_runtime(struct rq *rq)
516 {
517         rt_rq_iter_t iter;
518         struct rt_rq *rt_rq;
519
520         if (unlikely(!scheduler_running))
521                 return;
522
523         /*
524          * Reset each runqueue's bandwidth settings
525          */
526         for_each_rt_rq(rt_rq, iter, rq) {
527                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
528
529                 raw_spin_lock(&rt_b->rt_runtime_lock);
530                 raw_spin_lock(&rt_rq->rt_runtime_lock);
531                 rt_rq->rt_runtime = rt_b->rt_runtime;
532                 rt_rq->rt_time = 0;
533                 rt_rq->rt_throttled = 0;
534                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
535                 raw_spin_unlock(&rt_b->rt_runtime_lock);
536         }
537 }
538
539 static void enable_runtime(struct rq *rq)
540 {
541         unsigned long flags;
542
543         raw_spin_lock_irqsave(&rq->lock, flags);
544         __enable_runtime(rq);
545         raw_spin_unlock_irqrestore(&rq->lock, flags);
546 }
547
548 static int balance_runtime(struct rt_rq *rt_rq)
549 {
550         int more = 0;
551
552         if (rt_rq->rt_time > rt_rq->rt_runtime) {
553                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
554                 more = do_balance_runtime(rt_rq);
555                 raw_spin_lock(&rt_rq->rt_runtime_lock);
556         }
557
558         return more;
559 }
560 #else /* !CONFIG_SMP */
561 static inline int balance_runtime(struct rt_rq *rt_rq)
562 {
563         return 0;
564 }
565 #endif /* CONFIG_SMP */
566
567 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
568 {
569         int i, idle = 1;
570         const struct cpumask *span;
571
572         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
573                 return 1;
574
575         span = sched_rt_period_mask();
576         for_each_cpu(i, span) {
577                 int enqueue = 0;
578                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
579                 struct rq *rq = rq_of_rt_rq(rt_rq);
580
581                 raw_spin_lock(&rq->lock);
582                 if (rt_rq->rt_time) {
583                         u64 runtime;
584
585                         raw_spin_lock(&rt_rq->rt_runtime_lock);
586                         if (rt_rq->rt_throttled)
587                                 balance_runtime(rt_rq);
588                         runtime = rt_rq->rt_runtime;
589                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
590                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
591                                 rt_rq->rt_throttled = 0;
592                                 enqueue = 1;
593
594                                 /*
595                                  * Force a clock update if the CPU was idle,
596                                  * lest wakeup -> unthrottle time accumulate.
597                                  */
598                                 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
599                                         rq->skip_clock_update = -1;
600                         }
601                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
602                                 idle = 0;
603                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
604                 } else if (rt_rq->rt_nr_running) {
605                         idle = 0;
606                         if (!rt_rq_throttled(rt_rq))
607                                 enqueue = 1;
608                 }
609
610                 if (enqueue)
611                         sched_rt_rq_enqueue(rt_rq);
612                 raw_spin_unlock(&rq->lock);
613         }
614
615         return idle;
616 }
617
618 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
619 {
620 #ifdef CONFIG_RT_GROUP_SCHED
621         struct rt_rq *rt_rq = group_rt_rq(rt_se);
622
623         if (rt_rq)
624                 return rt_rq->highest_prio.curr;
625 #endif
626
627         return rt_task_of(rt_se)->prio;
628 }
629
630 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
631 {
632         u64 runtime = sched_rt_runtime(rt_rq);
633
634         if (rt_rq->rt_throttled)
635                 return rt_rq_throttled(rt_rq);
636
637         if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
638                 return 0;
639
640         balance_runtime(rt_rq);
641         runtime = sched_rt_runtime(rt_rq);
642         if (runtime == RUNTIME_INF)
643                 return 0;
644
645         if (rt_rq->rt_time > runtime) {
646                 rt_rq->rt_throttled = 1;
647                 if (rt_rq_throttled(rt_rq)) {
648                         sched_rt_rq_dequeue(rt_rq);
649                         return 1;
650                 }
651         }
652
653         return 0;
654 }
655
656 /*
657  * Update the current task's runtime statistics. Skip current tasks that
658  * are not in our scheduling class.
659  */
660 static void update_curr_rt(struct rq *rq)
661 {
662         struct task_struct *curr = rq->curr;
663         struct sched_rt_entity *rt_se = &curr->rt;
664         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
665         u64 delta_exec;
666
667         if (curr->sched_class != &rt_sched_class)
668                 return;
669
670         delta_exec = rq->clock_task - curr->se.exec_start;
671         if (unlikely((s64)delta_exec < 0))
672                 delta_exec = 0;
673
674         schedstat_set(curr->se.statistics.exec_max, max(curr->se.statistics.exec_max, delta_exec));
675
676         curr->se.sum_exec_runtime += delta_exec;
677         account_group_exec_runtime(curr, delta_exec);
678
679         curr->se.exec_start = rq->clock_task;
680         cpuacct_charge(curr, delta_exec);
681
682         sched_rt_avg_update(rq, delta_exec);
683
684         if (!rt_bandwidth_enabled())
685                 return;
686
687         for_each_sched_rt_entity(rt_se) {
688                 rt_rq = rt_rq_of_se(rt_se);
689
690                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
691                         raw_spin_lock(&rt_rq->rt_runtime_lock);
692                         rt_rq->rt_time += delta_exec;
693                         if (sched_rt_runtime_exceeded(rt_rq))
694                                 resched_task(curr);
695                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
696                 }
697         }
698 }
699
700 #if defined CONFIG_SMP
701
702 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu);
703
704 static inline int next_prio(struct rq *rq)
705 {
706         struct task_struct *next = pick_next_highest_task_rt(rq, rq->cpu);
707
708         if (next && rt_prio(next->prio))
709                 return next->prio;
710         else
711                 return MAX_RT_PRIO;
712 }
713
714 static void
715 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
716 {
717         struct rq *rq = rq_of_rt_rq(rt_rq);
718
719         if (prio < prev_prio) {
720
721                 /*
722                  * If the new task is higher in priority than anything on the
723                  * run-queue, we know that the previous high becomes our
724                  * next-highest.
725                  */
726                 rt_rq->highest_prio.next = prev_prio;
727
728                 if (rq->online)
729                         cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
730
731         } else if (prio == rt_rq->highest_prio.curr)
732                 /*
733                  * If the next task is equal in priority to the highest on
734                  * the run-queue, then we implicitly know that the next highest
735                  * task cannot be any lower than current
736                  */
737                 rt_rq->highest_prio.next = prio;
738         else if (prio < rt_rq->highest_prio.next)
739                 /*
740                  * Otherwise, we need to recompute next-highest
741                  */
742                 rt_rq->highest_prio.next = next_prio(rq);
743 }
744
745 static void
746 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
747 {
748         struct rq *rq = rq_of_rt_rq(rt_rq);
749
750         if (rt_rq->rt_nr_running && (prio <= rt_rq->highest_prio.next))
751                 rt_rq->highest_prio.next = next_prio(rq);
752
753         if (rq->online && rt_rq->highest_prio.curr != prev_prio)
754                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
755 }
756
757 #else /* CONFIG_SMP */
758
759 static inline
760 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
761 static inline
762 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
763
764 #endif /* CONFIG_SMP */
765
766 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
767 static void
768 inc_rt_prio(struct rt_rq *rt_rq, int prio)
769 {
770         int prev_prio = rt_rq->highest_prio.curr;
771
772         if (prio < prev_prio)
773                 rt_rq->highest_prio.curr = prio;
774
775         inc_rt_prio_smp(rt_rq, prio, prev_prio);
776 }
777
778 static void
779 dec_rt_prio(struct rt_rq *rt_rq, int prio)
780 {
781         int prev_prio = rt_rq->highest_prio.curr;
782
783         if (rt_rq->rt_nr_running) {
784
785                 WARN_ON(prio < prev_prio);
786
787                 /*
788                  * This may have been our highest task, and therefore
789                  * we may have some recomputation to do
790                  */
791                 if (prio == prev_prio) {
792                         struct rt_prio_array *array = &rt_rq->active;
793
794                         rt_rq->highest_prio.curr =
795                                 sched_find_first_bit(array->bitmap);
796                 }
797
798         } else
799                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
800
801         dec_rt_prio_smp(rt_rq, prio, prev_prio);
802 }
803
804 #else
805
806 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
807 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
808
809 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
810
811 #ifdef CONFIG_RT_GROUP_SCHED
812
813 static void
814 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
815 {
816         if (rt_se_boosted(rt_se))
817                 rt_rq->rt_nr_boosted++;
818
819         if (rt_rq->tg)
820                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
821 }
822
823 static void
824 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
825 {
826         if (rt_se_boosted(rt_se))
827                 rt_rq->rt_nr_boosted--;
828
829         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
830 }
831
832 #else /* CONFIG_RT_GROUP_SCHED */
833
834 static void
835 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
836 {
837         start_rt_bandwidth(&def_rt_bandwidth);
838 }
839
840 static inline
841 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
842
843 #endif /* CONFIG_RT_GROUP_SCHED */
844
845 static inline
846 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
847 {
848         int prio = rt_se_prio(rt_se);
849
850         WARN_ON(!rt_prio(prio));
851         rt_rq->rt_nr_running++;
852
853         inc_rt_prio(rt_rq, prio);
854         inc_rt_migration(rt_se, rt_rq);
855         inc_rt_group(rt_se, rt_rq);
856 }
857
858 static inline
859 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
860 {
861         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
862         WARN_ON(!rt_rq->rt_nr_running);
863         rt_rq->rt_nr_running--;
864
865         dec_rt_prio(rt_rq, rt_se_prio(rt_se));
866         dec_rt_migration(rt_se, rt_rq);
867         dec_rt_group(rt_se, rt_rq);
868 }
869
870 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
871 {
872         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
873         struct rt_prio_array *array = &rt_rq->active;
874         struct rt_rq *group_rq = group_rt_rq(rt_se);
875         struct list_head *queue = array->queue + rt_se_prio(rt_se);
876
877         /*
878          * Don't enqueue the group if its throttled, or when empty.
879          * The latter is a consequence of the former when a child group
880          * get throttled and the current group doesn't have any other
881          * active members.
882          */
883         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
884                 return;
885
886         if (!rt_rq->rt_nr_running)
887                 list_add_leaf_rt_rq(rt_rq);
888
889         if (head)
890                 list_add(&rt_se->run_list, queue);
891         else
892                 list_add_tail(&rt_se->run_list, queue);
893         __set_bit(rt_se_prio(rt_se), array->bitmap);
894
895         inc_rt_tasks(rt_se, rt_rq);
896 }
897
898 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
899 {
900         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
901         struct rt_prio_array *array = &rt_rq->active;
902
903         list_del_init(&rt_se->run_list);
904         if (list_empty(array->queue + rt_se_prio(rt_se)))
905                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
906
907         dec_rt_tasks(rt_se, rt_rq);
908         if (!rt_rq->rt_nr_running)
909                 list_del_leaf_rt_rq(rt_rq);
910 }
911
912 /*
913  * Because the prio of an upper entry depends on the lower
914  * entries, we must remove entries top - down.
915  */
916 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
917 {
918         struct sched_rt_entity *back = NULL;
919
920         for_each_sched_rt_entity(rt_se) {
921                 rt_se->back = back;
922                 back = rt_se;
923         }
924
925         for (rt_se = back; rt_se; rt_se = rt_se->back) {
926                 if (on_rt_rq(rt_se))
927                         __dequeue_rt_entity(rt_se);
928         }
929 }
930
931 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
932 {
933         dequeue_rt_stack(rt_se);
934         for_each_sched_rt_entity(rt_se)
935                 __enqueue_rt_entity(rt_se, head);
936 }
937
938 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
939 {
940         dequeue_rt_stack(rt_se);
941
942         for_each_sched_rt_entity(rt_se) {
943                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
944
945                 if (rt_rq && rt_rq->rt_nr_running)
946                         __enqueue_rt_entity(rt_se, false);
947         }
948 }
949
950 /*
951  * Adding/removing a task to/from a priority array:
952  */
953 static void
954 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
955 {
956         struct sched_rt_entity *rt_se = &p->rt;
957
958         if (flags & ENQUEUE_WAKEUP)
959                 rt_se->timeout = 0;
960
961         enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
962
963         if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
964                 enqueue_pushable_task(rq, p);
965 }
966
967 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
968 {
969         struct sched_rt_entity *rt_se = &p->rt;
970
971         update_curr_rt(rq);
972         dequeue_rt_entity(rt_se);
973
974         dequeue_pushable_task(rq, p);
975 }
976
977 /*
978  * Put task to the end of the run list without the overhead of dequeue
979  * followed by enqueue.
980  */
981 static void
982 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
983 {
984         if (on_rt_rq(rt_se)) {
985                 struct rt_prio_array *array = &rt_rq->active;
986                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
987
988                 if (head)
989                         list_move(&rt_se->run_list, queue);
990                 else
991                         list_move_tail(&rt_se->run_list, queue);
992         }
993 }
994
995 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
996 {
997         struct sched_rt_entity *rt_se = &p->rt;
998         struct rt_rq *rt_rq;
999
1000         for_each_sched_rt_entity(rt_se) {
1001                 rt_rq = rt_rq_of_se(rt_se);
1002                 requeue_rt_entity(rt_rq, rt_se, head);
1003         }
1004 }
1005
1006 static void yield_task_rt(struct rq *rq)
1007 {
1008         requeue_task_rt(rq, rq->curr, 0);
1009 }
1010
1011 #ifdef CONFIG_SMP
1012 static int find_lowest_rq(struct task_struct *task);
1013
1014 static int
1015 select_task_rq_rt(struct task_struct *p, int sd_flag, int flags)
1016 {
1017         struct task_struct *curr;
1018         struct rq *rq;
1019         int cpu;
1020
1021         if (sd_flag != SD_BALANCE_WAKE)
1022                 return smp_processor_id();
1023
1024         cpu = task_cpu(p);
1025         rq = cpu_rq(cpu);
1026
1027         rcu_read_lock();
1028         curr = ACCESS_ONCE(rq->curr); /* unlocked access */
1029
1030         /*
1031          * If the current task on @p's runqueue is an RT task, then
1032          * try to see if we can wake this RT task up on another
1033          * runqueue. Otherwise simply start this RT task
1034          * on its current runqueue.
1035          *
1036          * We want to avoid overloading runqueues. If the woken
1037          * task is a higher priority, then it will stay on this CPU
1038          * and the lower prio task should be moved to another CPU.
1039          * Even though this will probably make the lower prio task
1040          * lose its cache, we do not want to bounce a higher task
1041          * around just because it gave up its CPU, perhaps for a
1042          * lock?
1043          *
1044          * For equal prio tasks, we just let the scheduler sort it out.
1045          *
1046          * Otherwise, just let it ride on the affined RQ and the
1047          * post-schedule router will push the preempted task away
1048          *
1049          * This test is optimistic, if we get it wrong the load-balancer
1050          * will have to sort it out.
1051          */
1052         if (curr && unlikely(rt_task(curr)) &&
1053             (curr->rt.nr_cpus_allowed < 2 ||
1054              curr->prio <= p->prio) &&
1055             (p->rt.nr_cpus_allowed > 1)) {
1056                 int target = find_lowest_rq(p);
1057
1058                 if (target != -1)
1059                         cpu = target;
1060         }
1061         rcu_read_unlock();
1062
1063         return cpu;
1064 }
1065
1066 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1067 {
1068         if (rq->curr->rt.nr_cpus_allowed == 1)
1069                 return;
1070
1071         if (p->rt.nr_cpus_allowed != 1
1072             && cpupri_find(&rq->rd->cpupri, p, NULL))
1073                 return;
1074
1075         if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1076                 return;
1077
1078         /*
1079          * There appears to be other cpus that can accept
1080          * current and none to run 'p', so lets reschedule
1081          * to try and push current away:
1082          */
1083         requeue_task_rt(rq, p, 1);
1084         resched_task(rq->curr);
1085 }
1086
1087 #endif /* CONFIG_SMP */
1088
1089 /*
1090  * Preempt the current task with a newly woken task if needed:
1091  */
1092 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1093 {
1094         if (p->prio < rq->curr->prio) {
1095                 resched_task(rq->curr);
1096                 return;
1097         }
1098
1099 #ifdef CONFIG_SMP
1100         /*
1101          * If:
1102          *
1103          * - the newly woken task is of equal priority to the current task
1104          * - the newly woken task is non-migratable while current is migratable
1105          * - current will be preempted on the next reschedule
1106          *
1107          * we should check to see if current can readily move to a different
1108          * cpu.  If so, we will reschedule to allow the push logic to try
1109          * to move current somewhere else, making room for our non-migratable
1110          * task.
1111          */
1112         if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1113                 check_preempt_equal_prio(rq, p);
1114 #endif
1115 }
1116
1117 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1118                                                    struct rt_rq *rt_rq)
1119 {
1120         struct rt_prio_array *array = &rt_rq->active;
1121         struct sched_rt_entity *next = NULL;
1122         struct list_head *queue;
1123         int idx;
1124
1125         idx = sched_find_first_bit(array->bitmap);
1126         BUG_ON(idx >= MAX_RT_PRIO);
1127
1128         queue = array->queue + idx;
1129         next = list_entry(queue->next, struct sched_rt_entity, run_list);
1130
1131         return next;
1132 }
1133
1134 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1135 {
1136         struct sched_rt_entity *rt_se;
1137         struct task_struct *p;
1138         struct rt_rq *rt_rq;
1139
1140         rt_rq = &rq->rt;
1141
1142         if (!rt_rq->rt_nr_running)
1143                 return NULL;
1144
1145         if (rt_rq_throttled(rt_rq))
1146                 return NULL;
1147
1148         do {
1149                 rt_se = pick_next_rt_entity(rq, rt_rq);
1150                 BUG_ON(!rt_se);
1151                 rt_rq = group_rt_rq(rt_se);
1152         } while (rt_rq);
1153
1154         p = rt_task_of(rt_se);
1155         p->se.exec_start = rq->clock_task;
1156
1157         return p;
1158 }
1159
1160 static struct task_struct *pick_next_task_rt(struct rq *rq)
1161 {
1162         struct task_struct *p = _pick_next_task_rt(rq);
1163
1164         /* The running task is never eligible for pushing */
1165         if (p)
1166                 dequeue_pushable_task(rq, p);
1167
1168 #ifdef CONFIG_SMP
1169         /*
1170          * We detect this state here so that we can avoid taking the RQ
1171          * lock again later if there is no need to push
1172          */
1173         rq->post_schedule = has_pushable_tasks(rq);
1174 #endif
1175
1176         return p;
1177 }
1178
1179 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1180 {
1181         update_curr_rt(rq);
1182         p->se.exec_start = 0;
1183
1184         /*
1185          * The previous task needs to be made eligible for pushing
1186          * if it is still active
1187          */
1188         if (on_rt_rq(&p->rt) && p->rt.nr_cpus_allowed > 1)
1189                 enqueue_pushable_task(rq, p);
1190 }
1191
1192 #ifdef CONFIG_SMP
1193
1194 /* Only try algorithms three times */
1195 #define RT_MAX_TRIES 3
1196
1197 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
1198
1199 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1200 {
1201         if (!task_running(rq, p) &&
1202             (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
1203             (p->rt.nr_cpus_allowed > 1))
1204                 return 1;
1205         return 0;
1206 }
1207
1208 /* Return the second highest RT task, NULL otherwise */
1209 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
1210 {
1211         struct task_struct *next = NULL;
1212         struct sched_rt_entity *rt_se;
1213         struct rt_prio_array *array;
1214         struct rt_rq *rt_rq;
1215         int idx;
1216
1217         for_each_leaf_rt_rq(rt_rq, rq) {
1218                 array = &rt_rq->active;
1219                 idx = sched_find_first_bit(array->bitmap);
1220 next_idx:
1221                 if (idx >= MAX_RT_PRIO)
1222                         continue;
1223                 if (next && next->prio < idx)
1224                         continue;
1225                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1226                         struct task_struct *p;
1227
1228                         if (!rt_entity_is_task(rt_se))
1229                                 continue;
1230
1231                         p = rt_task_of(rt_se);
1232                         if (pick_rt_task(rq, p, cpu)) {
1233                                 next = p;
1234                                 break;
1235                         }
1236                 }
1237                 if (!next) {
1238                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1239                         goto next_idx;
1240                 }
1241         }
1242
1243         return next;
1244 }
1245
1246 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1247
1248 static int find_lowest_rq(struct task_struct *task)
1249 {
1250         struct sched_domain *sd;
1251         struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1252         int this_cpu = smp_processor_id();
1253         int cpu      = task_cpu(task);
1254
1255         /* Make sure the mask is initialized first */
1256         if (unlikely(!lowest_mask))
1257                 return -1;
1258
1259         if (task->rt.nr_cpus_allowed == 1)
1260                 return -1; /* No other targets possible */
1261
1262         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1263                 return -1; /* No targets found */
1264
1265         /*
1266          * At this point we have built a mask of cpus representing the
1267          * lowest priority tasks in the system.  Now we want to elect
1268          * the best one based on our affinity and topology.
1269          *
1270          * We prioritize the last cpu that the task executed on since
1271          * it is most likely cache-hot in that location.
1272          */
1273         if (cpumask_test_cpu(cpu, lowest_mask))
1274                 return cpu;
1275
1276         /*
1277          * Otherwise, we consult the sched_domains span maps to figure
1278          * out which cpu is logically closest to our hot cache data.
1279          */
1280         if (!cpumask_test_cpu(this_cpu, lowest_mask))
1281                 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1282
1283         rcu_read_lock();
1284         for_each_domain(cpu, sd) {
1285                 if (sd->flags & SD_WAKE_AFFINE) {
1286                         int best_cpu;
1287
1288                         /*
1289                          * "this_cpu" is cheaper to preempt than a
1290                          * remote processor.
1291                          */
1292                         if (this_cpu != -1 &&
1293                             cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1294                                 rcu_read_unlock();
1295                                 return this_cpu;
1296                         }
1297
1298                         best_cpu = cpumask_first_and(lowest_mask,
1299                                                      sched_domain_span(sd));
1300                         if (best_cpu < nr_cpu_ids) {
1301                                 rcu_read_unlock();
1302                                 return best_cpu;
1303                         }
1304                 }
1305         }
1306         rcu_read_unlock();
1307
1308         /*
1309          * And finally, if there were no matches within the domains
1310          * just give the caller *something* to work with from the compatible
1311          * locations.
1312          */
1313         if (this_cpu != -1)
1314                 return this_cpu;
1315
1316         cpu = cpumask_any(lowest_mask);
1317         if (cpu < nr_cpu_ids)
1318                 return cpu;
1319         return -1;
1320 }
1321
1322 /* Will lock the rq it finds */
1323 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1324 {
1325         struct rq *lowest_rq = NULL;
1326         int tries;
1327         int cpu;
1328
1329         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1330                 cpu = find_lowest_rq(task);
1331
1332                 if ((cpu == -1) || (cpu == rq->cpu))
1333                         break;
1334
1335                 lowest_rq = cpu_rq(cpu);
1336
1337                 /* if the prio of this runqueue changed, try again */
1338                 if (double_lock_balance(rq, lowest_rq)) {
1339                         /*
1340                          * We had to unlock the run queue. In
1341                          * the mean time, task could have
1342                          * migrated already or had its affinity changed.
1343                          * Also make sure that it wasn't scheduled on its rq.
1344                          */
1345                         if (unlikely(task_rq(task) != rq ||
1346                                      !cpumask_test_cpu(lowest_rq->cpu,
1347                                                        &task->cpus_allowed) ||
1348                                      task_running(rq, task) ||
1349                                      !task->on_rq)) {
1350
1351                                 raw_spin_unlock(&lowest_rq->lock);
1352                                 lowest_rq = NULL;
1353                                 break;
1354                         }
1355                 }
1356
1357                 /* If this rq is still suitable use it. */
1358                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1359                         break;
1360
1361                 /* try again */
1362                 double_unlock_balance(rq, lowest_rq);
1363                 lowest_rq = NULL;
1364         }
1365
1366         return lowest_rq;
1367 }
1368
1369 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1370 {
1371         struct task_struct *p;
1372
1373         if (!has_pushable_tasks(rq))
1374                 return NULL;
1375
1376         p = plist_first_entry(&rq->rt.pushable_tasks,
1377                               struct task_struct, pushable_tasks);
1378
1379         BUG_ON(rq->cpu != task_cpu(p));
1380         BUG_ON(task_current(rq, p));
1381         BUG_ON(p->rt.nr_cpus_allowed <= 1);
1382
1383         BUG_ON(!p->on_rq);
1384         BUG_ON(!rt_task(p));
1385
1386         return p;
1387 }
1388
1389 /*
1390  * If the current CPU has more than one RT task, see if the non
1391  * running task can migrate over to a CPU that is running a task
1392  * of lesser priority.
1393  */
1394 static int push_rt_task(struct rq *rq)
1395 {
1396         struct task_struct *next_task;
1397         struct rq *lowest_rq;
1398
1399         if (!rq->rt.overloaded)
1400                 return 0;
1401
1402         next_task = pick_next_pushable_task(rq);
1403         if (!next_task)
1404                 return 0;
1405
1406 retry:
1407         if (unlikely(next_task == rq->curr)) {
1408                 WARN_ON(1);
1409                 return 0;
1410         }
1411
1412         /*
1413          * It's possible that the next_task slipped in of
1414          * higher priority than current. If that's the case
1415          * just reschedule current.
1416          */
1417         if (unlikely(next_task->prio < rq->curr->prio)) {
1418                 resched_task(rq->curr);
1419                 return 0;
1420         }
1421
1422         /* We might release rq lock */
1423         get_task_struct(next_task);
1424
1425         /* find_lock_lowest_rq locks the rq if found */
1426         lowest_rq = find_lock_lowest_rq(next_task, rq);
1427         if (!lowest_rq) {
1428                 struct task_struct *task;
1429                 /*
1430                  * find lock_lowest_rq releases rq->lock
1431                  * so it is possible that next_task has migrated.
1432                  *
1433                  * We need to make sure that the task is still on the same
1434                  * run-queue and is also still the next task eligible for
1435                  * pushing.
1436                  */
1437                 task = pick_next_pushable_task(rq);
1438                 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1439                         /*
1440                          * If we get here, the task hasn't moved at all, but
1441                          * it has failed to push.  We will not try again,
1442                          * since the other cpus will pull from us when they
1443                          * are ready.
1444                          */
1445                         dequeue_pushable_task(rq, next_task);
1446                         goto out;
1447                 }
1448
1449                 if (!task)
1450                         /* No more tasks, just exit */
1451                         goto out;
1452
1453                 /*
1454                  * Something has shifted, try again.
1455                  */
1456                 put_task_struct(next_task);
1457                 next_task = task;
1458                 goto retry;
1459         }
1460
1461         deactivate_task(rq, next_task, 0);
1462         set_task_cpu(next_task, lowest_rq->cpu);
1463         activate_task(lowest_rq, next_task, 0);
1464
1465         resched_task(lowest_rq->curr);
1466
1467         double_unlock_balance(rq, lowest_rq);
1468
1469 out:
1470         put_task_struct(next_task);
1471
1472         return 1;
1473 }
1474
1475 static void push_rt_tasks(struct rq *rq)
1476 {
1477         /* push_rt_task will return true if it moved an RT */
1478         while (push_rt_task(rq))
1479                 ;
1480 }
1481
1482 static int pull_rt_task(struct rq *this_rq)
1483 {
1484         int this_cpu = this_rq->cpu, ret = 0, cpu;
1485         struct task_struct *p;
1486         struct rq *src_rq;
1487
1488         if (likely(!rt_overloaded(this_rq)))
1489                 return 0;
1490
1491         for_each_cpu(cpu, this_rq->rd->rto_mask) {
1492                 if (this_cpu == cpu)
1493                         continue;
1494
1495                 src_rq = cpu_rq(cpu);
1496
1497                 /*
1498                  * Don't bother taking the src_rq->lock if the next highest
1499                  * task is known to be lower-priority than our current task.
1500                  * This may look racy, but if this value is about to go
1501                  * logically higher, the src_rq will push this task away.
1502                  * And if its going logically lower, we do not care
1503                  */
1504                 if (src_rq->rt.highest_prio.next >=
1505                     this_rq->rt.highest_prio.curr)
1506                         continue;
1507
1508                 /*
1509                  * We can potentially drop this_rq's lock in
1510                  * double_lock_balance, and another CPU could
1511                  * alter this_rq
1512                  */
1513                 double_lock_balance(this_rq, src_rq);
1514
1515                 /*
1516                  * Are there still pullable RT tasks?
1517                  */
1518                 if (src_rq->rt.rt_nr_running <= 1)
1519                         goto skip;
1520
1521                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1522
1523                 /*
1524                  * Do we have an RT task that preempts
1525                  * the to-be-scheduled task?
1526                  */
1527                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1528                         WARN_ON(p == src_rq->curr);
1529                         WARN_ON(!p->on_rq);
1530
1531                         /*
1532                          * There's a chance that p is higher in priority
1533                          * than what's currently running on its cpu.
1534                          * This is just that p is wakeing up and hasn't
1535                          * had a chance to schedule. We only pull
1536                          * p if it is lower in priority than the
1537                          * current task on the run queue
1538                          */
1539                         if (p->prio < src_rq->curr->prio)
1540                                 goto skip;
1541
1542                         ret = 1;
1543
1544                         deactivate_task(src_rq, p, 0);
1545                         set_task_cpu(p, this_cpu);
1546                         activate_task(this_rq, p, 0);
1547                         /*
1548                          * We continue with the search, just in
1549                          * case there's an even higher prio task
1550                          * in another runqueue. (low likelihood
1551                          * but possible)
1552                          */
1553                 }
1554 skip:
1555                 double_unlock_balance(this_rq, src_rq);
1556         }
1557
1558         return ret;
1559 }
1560
1561 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1562 {
1563         /* Try to pull RT tasks here if we lower this rq's prio */
1564         if (rq->rt.highest_prio.curr > prev->prio)
1565                 pull_rt_task(rq);
1566 }
1567
1568 static void post_schedule_rt(struct rq *rq)
1569 {
1570         push_rt_tasks(rq);
1571 }
1572
1573 /*
1574  * If we are not running and we are not going to reschedule soon, we should
1575  * try to push tasks away now
1576  */
1577 static void task_woken_rt(struct rq *rq, struct task_struct *p)
1578 {
1579         if (!task_running(rq, p) &&
1580             !test_tsk_need_resched(rq->curr) &&
1581             has_pushable_tasks(rq) &&
1582             p->rt.nr_cpus_allowed > 1 &&
1583             rt_task(rq->curr) &&
1584             (rq->curr->rt.nr_cpus_allowed < 2 ||
1585              rq->curr->prio <= p->prio))
1586                 push_rt_tasks(rq);
1587 }
1588
1589 static void set_cpus_allowed_rt(struct task_struct *p,
1590                                 const struct cpumask *new_mask)
1591 {
1592         int weight = cpumask_weight(new_mask);
1593
1594         BUG_ON(!rt_task(p));
1595
1596         /*
1597          * Update the migration status of the RQ if we have an RT task
1598          * which is running AND changing its weight value.
1599          */
1600         if (p->on_rq && (weight != p->rt.nr_cpus_allowed)) {
1601                 struct rq *rq = task_rq(p);
1602
1603                 if (!task_current(rq, p)) {
1604                         /*
1605                          * Make sure we dequeue this task from the pushable list
1606                          * before going further.  It will either remain off of
1607                          * the list because we are no longer pushable, or it
1608                          * will be requeued.
1609                          */
1610                         if (p->rt.nr_cpus_allowed > 1)
1611                                 dequeue_pushable_task(rq, p);
1612
1613                         /*
1614                          * Requeue if our weight is changing and still > 1
1615                          */
1616                         if (weight > 1)
1617                                 enqueue_pushable_task(rq, p);
1618
1619                 }
1620
1621                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1622                         rq->rt.rt_nr_migratory++;
1623                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1624                         BUG_ON(!rq->rt.rt_nr_migratory);
1625                         rq->rt.rt_nr_migratory--;
1626                 }
1627
1628                 update_rt_migration(&rq->rt);
1629         }
1630
1631         cpumask_copy(&p->cpus_allowed, new_mask);
1632         p->rt.nr_cpus_allowed = weight;
1633 }
1634
1635 /* Assumes rq->lock is held */
1636 static void rq_online_rt(struct rq *rq)
1637 {
1638         if (rq->rt.overloaded)
1639                 rt_set_overload(rq);
1640
1641         __enable_runtime(rq);
1642
1643         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1644 }
1645
1646 /* Assumes rq->lock is held */
1647 static void rq_offline_rt(struct rq *rq)
1648 {
1649         if (rq->rt.overloaded)
1650                 rt_clear_overload(rq);
1651
1652         __disable_runtime(rq);
1653
1654         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1655 }
1656
1657 /*
1658  * When switch from the rt queue, we bring ourselves to a position
1659  * that we might want to pull RT tasks from other runqueues.
1660  */
1661 static void switched_from_rt(struct rq *rq, struct task_struct *p)
1662 {
1663         /*
1664          * If there are other RT tasks then we will reschedule
1665          * and the scheduling of the other RT tasks will handle
1666          * the balancing. But if we are the last RT task
1667          * we may need to handle the pulling of RT tasks
1668          * now.
1669          */
1670         if (p->on_rq && !rq->rt.rt_nr_running)
1671                 pull_rt_task(rq);
1672 }
1673
1674 static inline void init_sched_rt_class(void)
1675 {
1676         unsigned int i;
1677
1678         for_each_possible_cpu(i)
1679                 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1680                                         GFP_KERNEL, cpu_to_node(i));
1681 }
1682 #endif /* CONFIG_SMP */
1683
1684 /*
1685  * When switching a task to RT, we may overload the runqueue
1686  * with RT tasks. In this case we try to push them off to
1687  * other runqueues.
1688  */
1689 static void switched_to_rt(struct rq *rq, struct task_struct *p)
1690 {
1691         int check_resched = 1;
1692
1693         /*
1694          * If we are already running, then there's nothing
1695          * that needs to be done. But if we are not running
1696          * we may need to preempt the current running task.
1697          * If that current running task is also an RT task
1698          * then see if we can move to another run queue.
1699          */
1700         if (p->on_rq && rq->curr != p) {
1701 #ifdef CONFIG_SMP
1702                 if (rq->rt.overloaded && push_rt_task(rq) &&
1703                     /* Don't resched if we changed runqueues */
1704                     rq != task_rq(p))
1705                         check_resched = 0;
1706 #endif /* CONFIG_SMP */
1707                 if (check_resched && p->prio < rq->curr->prio)
1708                         resched_task(rq->curr);
1709         }
1710 }
1711
1712 /*
1713  * Priority of the task has changed. This may cause
1714  * us to initiate a push or pull.
1715  */
1716 static void
1717 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
1718 {
1719         if (!p->on_rq)
1720                 return;
1721
1722         if (rq->curr == p) {
1723 #ifdef CONFIG_SMP
1724                 /*
1725                  * If our priority decreases while running, we
1726                  * may need to pull tasks to this runqueue.
1727                  */
1728                 if (oldprio < p->prio)
1729                         pull_rt_task(rq);
1730                 /*
1731                  * If there's a higher priority task waiting to run
1732                  * then reschedule. Note, the above pull_rt_task
1733                  * can release the rq lock and p could migrate.
1734                  * Only reschedule if p is still on the same runqueue.
1735                  */
1736                 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1737                         resched_task(p);
1738 #else
1739                 /* For UP simply resched on drop of prio */
1740                 if (oldprio < p->prio)
1741                         resched_task(p);
1742 #endif /* CONFIG_SMP */
1743         } else {
1744                 /*
1745                  * This task is not running, but if it is
1746                  * greater than the current running task
1747                  * then reschedule.
1748                  */
1749                 if (p->prio < rq->curr->prio)
1750                         resched_task(rq->curr);
1751         }
1752 }
1753
1754 static void watchdog(struct rq *rq, struct task_struct *p)
1755 {
1756         unsigned long soft, hard;
1757
1758         /* max may change after cur was read, this will be fixed next tick */
1759         soft = task_rlimit(p, RLIMIT_RTTIME);
1760         hard = task_rlimit_max(p, RLIMIT_RTTIME);
1761
1762         if (soft != RLIM_INFINITY) {
1763                 unsigned long next;
1764
1765                 p->rt.timeout++;
1766                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1767                 if (p->rt.timeout > next)
1768                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1769         }
1770 }
1771
1772 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1773 {
1774         update_curr_rt(rq);
1775
1776         watchdog(rq, p);
1777
1778         /*
1779          * RR tasks need a special form of timeslice management.
1780          * FIFO tasks have no timeslices.
1781          */
1782         if (p->policy != SCHED_RR)
1783                 return;
1784
1785         if (--p->rt.time_slice)
1786                 return;
1787
1788         p->rt.time_slice = DEF_TIMESLICE;
1789
1790         /*
1791          * Requeue to the end of queue if we are not the only element
1792          * on the queue:
1793          */
1794         if (p->rt.run_list.prev != p->rt.run_list.next) {
1795                 requeue_task_rt(rq, p, 0);
1796                 set_tsk_need_resched(p);
1797         }
1798 }
1799
1800 static void set_curr_task_rt(struct rq *rq)
1801 {
1802         struct task_struct *p = rq->curr;
1803
1804         p->se.exec_start = rq->clock_task;
1805
1806         /* The running task is never eligible for pushing */
1807         dequeue_pushable_task(rq, p);
1808 }
1809
1810 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
1811 {
1812         /*
1813          * Time slice is 0 for SCHED_FIFO tasks
1814          */
1815         if (task->policy == SCHED_RR)
1816                 return DEF_TIMESLICE;
1817         else
1818                 return 0;
1819 }
1820
1821 static const struct sched_class rt_sched_class = {
1822         .next                   = &fair_sched_class,
1823         .enqueue_task           = enqueue_task_rt,
1824         .dequeue_task           = dequeue_task_rt,
1825         .yield_task             = yield_task_rt,
1826
1827         .check_preempt_curr     = check_preempt_curr_rt,
1828
1829         .pick_next_task         = pick_next_task_rt,
1830         .put_prev_task          = put_prev_task_rt,
1831
1832 #ifdef CONFIG_SMP
1833         .select_task_rq         = select_task_rq_rt,
1834
1835         .set_cpus_allowed       = set_cpus_allowed_rt,
1836         .rq_online              = rq_online_rt,
1837         .rq_offline             = rq_offline_rt,
1838         .pre_schedule           = pre_schedule_rt,
1839         .post_schedule          = post_schedule_rt,
1840         .task_woken             = task_woken_rt,
1841         .switched_from          = switched_from_rt,
1842 #endif
1843
1844         .set_curr_task          = set_curr_task_rt,
1845         .task_tick              = task_tick_rt,
1846
1847         .get_rr_interval        = get_rr_interval_rt,
1848
1849         .prio_changed           = prio_changed_rt,
1850         .switched_to            = switched_to_rt,
1851 };
1852
1853 #ifdef CONFIG_SCHED_DEBUG
1854 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1855
1856 static void print_rt_stats(struct seq_file *m, int cpu)
1857 {
1858         rt_rq_iter_t iter;
1859         struct rt_rq *rt_rq;
1860
1861         rcu_read_lock();
1862         for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
1863                 print_rt_rq(m, cpu, rt_rq);
1864         rcu_read_unlock();
1865 }
1866 #endif /* CONFIG_SCHED_DEBUG */
1867