Merge branch 'x86-cpu-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[pandora-kernel.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 inline int has_pushable_tasks(struct rq *rq)
128 {
129         return !plist_head_empty(&rq->rt.pushable_tasks);
130 }
131
132 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
133 {
134         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
135         plist_node_init(&p->pushable_tasks, p->prio);
136         plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
137
138         /* Update the highest prio pushable task */
139         if (p->prio < rq->rt.highest_prio.next)
140                 rq->rt.highest_prio.next = p->prio;
141 }
142
143 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
144 {
145         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
146
147         /* Update the new highest prio pushable task */
148         if (has_pushable_tasks(rq)) {
149                 p = plist_first_entry(&rq->rt.pushable_tasks,
150                                       struct task_struct, pushable_tasks);
151                 rq->rt.highest_prio.next = p->prio;
152         } else
153                 rq->rt.highest_prio.next = MAX_RT_PRIO;
154 }
155
156 #else
157
158 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
159 {
160 }
161
162 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
163 {
164 }
165
166 static inline
167 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
168 {
169 }
170
171 static inline
172 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
173 {
174 }
175
176 #endif /* CONFIG_SMP */
177
178 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
179 {
180         return !list_empty(&rt_se->run_list);
181 }
182
183 #ifdef CONFIG_RT_GROUP_SCHED
184
185 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
186 {
187         if (!rt_rq->tg)
188                 return RUNTIME_INF;
189
190         return rt_rq->rt_runtime;
191 }
192
193 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
194 {
195         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
196 }
197
198 typedef struct task_group *rt_rq_iter_t;
199
200 static inline struct task_group *next_task_group(struct task_group *tg)
201 {
202         do {
203                 tg = list_entry_rcu(tg->list.next,
204                         typeof(struct task_group), list);
205         } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
206
207         if (&tg->list == &task_groups)
208                 tg = NULL;
209
210         return tg;
211 }
212
213 #define for_each_rt_rq(rt_rq, iter, rq)                                 \
214         for (iter = container_of(&task_groups, typeof(*iter), list);    \
215                 (iter = next_task_group(iter)) &&                       \
216                 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
217
218 static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
219 {
220         list_add_rcu(&rt_rq->leaf_rt_rq_list,
221                         &rq_of_rt_rq(rt_rq)->leaf_rt_rq_list);
222 }
223
224 static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
225 {
226         list_del_rcu(&rt_rq->leaf_rt_rq_list);
227 }
228
229 #define for_each_leaf_rt_rq(rt_rq, rq) \
230         list_for_each_entry_rcu(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
231
232 #define for_each_sched_rt_entity(rt_se) \
233         for (; rt_se; rt_se = rt_se->parent)
234
235 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
236 {
237         return rt_se->my_q;
238 }
239
240 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
241 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
242
243 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
244 {
245         struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
246         struct sched_rt_entity *rt_se;
247
248         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
249
250         rt_se = rt_rq->tg->rt_se[cpu];
251
252         if (rt_rq->rt_nr_running) {
253                 if (rt_se && !on_rt_rq(rt_se))
254                         enqueue_rt_entity(rt_se, false);
255                 if (rt_rq->highest_prio.curr < curr->prio)
256                         resched_task(curr);
257         }
258 }
259
260 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
261 {
262         struct sched_rt_entity *rt_se;
263         int cpu = cpu_of(rq_of_rt_rq(rt_rq));
264
265         rt_se = rt_rq->tg->rt_se[cpu];
266
267         if (rt_se && on_rt_rq(rt_se))
268                 dequeue_rt_entity(rt_se);
269 }
270
271 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
272 {
273         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
274 }
275
276 static int rt_se_boosted(struct sched_rt_entity *rt_se)
277 {
278         struct rt_rq *rt_rq = group_rt_rq(rt_se);
279         struct task_struct *p;
280
281         if (rt_rq)
282                 return !!rt_rq->rt_nr_boosted;
283
284         p = rt_task_of(rt_se);
285         return p->prio != p->normal_prio;
286 }
287
288 #ifdef CONFIG_SMP
289 static inline const struct cpumask *sched_rt_period_mask(void)
290 {
291         return cpu_rq(smp_processor_id())->rd->span;
292 }
293 #else
294 static inline const struct cpumask *sched_rt_period_mask(void)
295 {
296         return cpu_online_mask;
297 }
298 #endif
299
300 static inline
301 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
302 {
303         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
304 }
305
306 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
307 {
308         return &rt_rq->tg->rt_bandwidth;
309 }
310
311 #else /* !CONFIG_RT_GROUP_SCHED */
312
313 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
314 {
315         return rt_rq->rt_runtime;
316 }
317
318 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
319 {
320         return ktime_to_ns(def_rt_bandwidth.rt_period);
321 }
322
323 typedef struct rt_rq *rt_rq_iter_t;
324
325 #define for_each_rt_rq(rt_rq, iter, rq) \
326         for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
327
328 static inline void list_add_leaf_rt_rq(struct rt_rq *rt_rq)
329 {
330 }
331
332 static inline void list_del_leaf_rt_rq(struct rt_rq *rt_rq)
333 {
334 }
335
336 #define for_each_leaf_rt_rq(rt_rq, rq) \
337         for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
338
339 #define for_each_sched_rt_entity(rt_se) \
340         for (; rt_se; rt_se = NULL)
341
342 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
343 {
344         return NULL;
345 }
346
347 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
348 {
349         if (rt_rq->rt_nr_running)
350                 resched_task(rq_of_rt_rq(rt_rq)->curr);
351 }
352
353 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
354 {
355 }
356
357 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
358 {
359         return rt_rq->rt_throttled;
360 }
361
362 static inline const struct cpumask *sched_rt_period_mask(void)
363 {
364         return cpu_online_mask;
365 }
366
367 static inline
368 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
369 {
370         return &cpu_rq(cpu)->rt;
371 }
372
373 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
374 {
375         return &def_rt_bandwidth;
376 }
377
378 #endif /* CONFIG_RT_GROUP_SCHED */
379
380 #ifdef CONFIG_SMP
381 /*
382  * We ran out of runtime, see if we can borrow some from our neighbours.
383  */
384 static int do_balance_runtime(struct rt_rq *rt_rq)
385 {
386         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
387         struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
388         int i, weight, more = 0;
389         u64 rt_period;
390
391         weight = cpumask_weight(rd->span);
392
393         raw_spin_lock(&rt_b->rt_runtime_lock);
394         rt_period = ktime_to_ns(rt_b->rt_period);
395         for_each_cpu(i, rd->span) {
396                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
397                 s64 diff;
398
399                 if (iter == rt_rq)
400                         continue;
401
402                 raw_spin_lock(&iter->rt_runtime_lock);
403                 /*
404                  * Either all rqs have inf runtime and there's nothing to steal
405                  * or __disable_runtime() below sets a specific rq to inf to
406                  * indicate its been disabled and disalow stealing.
407                  */
408                 if (iter->rt_runtime == RUNTIME_INF)
409                         goto next;
410
411                 /*
412                  * From runqueues with spare time, take 1/n part of their
413                  * spare time, but no more than our period.
414                  */
415                 diff = iter->rt_runtime - iter->rt_time;
416                 if (diff > 0) {
417                         diff = div_u64((u64)diff, weight);
418                         if (rt_rq->rt_runtime + diff > rt_period)
419                                 diff = rt_period - rt_rq->rt_runtime;
420                         iter->rt_runtime -= diff;
421                         rt_rq->rt_runtime += diff;
422                         more = 1;
423                         if (rt_rq->rt_runtime == rt_period) {
424                                 raw_spin_unlock(&iter->rt_runtime_lock);
425                                 break;
426                         }
427                 }
428 next:
429                 raw_spin_unlock(&iter->rt_runtime_lock);
430         }
431         raw_spin_unlock(&rt_b->rt_runtime_lock);
432
433         return more;
434 }
435
436 /*
437  * Ensure this RQ takes back all the runtime it lend to its neighbours.
438  */
439 static void __disable_runtime(struct rq *rq)
440 {
441         struct root_domain *rd = rq->rd;
442         rt_rq_iter_t iter;
443         struct rt_rq *rt_rq;
444
445         if (unlikely(!scheduler_running))
446                 return;
447
448         for_each_rt_rq(rt_rq, iter, rq) {
449                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
450                 s64 want;
451                 int i;
452
453                 raw_spin_lock(&rt_b->rt_runtime_lock);
454                 raw_spin_lock(&rt_rq->rt_runtime_lock);
455                 /*
456                  * Either we're all inf and nobody needs to borrow, or we're
457                  * already disabled and thus have nothing to do, or we have
458                  * exactly the right amount of runtime to take out.
459                  */
460                 if (rt_rq->rt_runtime == RUNTIME_INF ||
461                                 rt_rq->rt_runtime == rt_b->rt_runtime)
462                         goto balanced;
463                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
464
465                 /*
466                  * Calculate the difference between what we started out with
467                  * and what we current have, that's the amount of runtime
468                  * we lend and now have to reclaim.
469                  */
470                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
471
472                 /*
473                  * Greedy reclaim, take back as much as we can.
474                  */
475                 for_each_cpu(i, rd->span) {
476                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
477                         s64 diff;
478
479                         /*
480                          * Can't reclaim from ourselves or disabled runqueues.
481                          */
482                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
483                                 continue;
484
485                         raw_spin_lock(&iter->rt_runtime_lock);
486                         if (want > 0) {
487                                 diff = min_t(s64, iter->rt_runtime, want);
488                                 iter->rt_runtime -= diff;
489                                 want -= diff;
490                         } else {
491                                 iter->rt_runtime -= want;
492                                 want -= want;
493                         }
494                         raw_spin_unlock(&iter->rt_runtime_lock);
495
496                         if (!want)
497                                 break;
498                 }
499
500                 raw_spin_lock(&rt_rq->rt_runtime_lock);
501                 /*
502                  * We cannot be left wanting - that would mean some runtime
503                  * leaked out of the system.
504                  */
505                 BUG_ON(want);
506 balanced:
507                 /*
508                  * Disable all the borrow logic by pretending we have inf
509                  * runtime - in which case borrowing doesn't make sense.
510                  */
511                 rt_rq->rt_runtime = RUNTIME_INF;
512                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
513                 raw_spin_unlock(&rt_b->rt_runtime_lock);
514         }
515 }
516
517 static void disable_runtime(struct rq *rq)
518 {
519         unsigned long flags;
520
521         raw_spin_lock_irqsave(&rq->lock, flags);
522         __disable_runtime(rq);
523         raw_spin_unlock_irqrestore(&rq->lock, flags);
524 }
525
526 static void __enable_runtime(struct rq *rq)
527 {
528         rt_rq_iter_t iter;
529         struct rt_rq *rt_rq;
530
531         if (unlikely(!scheduler_running))
532                 return;
533
534         /*
535          * Reset each runqueue's bandwidth settings
536          */
537         for_each_rt_rq(rt_rq, iter, rq) {
538                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
539
540                 raw_spin_lock(&rt_b->rt_runtime_lock);
541                 raw_spin_lock(&rt_rq->rt_runtime_lock);
542                 rt_rq->rt_runtime = rt_b->rt_runtime;
543                 rt_rq->rt_time = 0;
544                 rt_rq->rt_throttled = 0;
545                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
546                 raw_spin_unlock(&rt_b->rt_runtime_lock);
547         }
548 }
549
550 static void enable_runtime(struct rq *rq)
551 {
552         unsigned long flags;
553
554         raw_spin_lock_irqsave(&rq->lock, flags);
555         __enable_runtime(rq);
556         raw_spin_unlock_irqrestore(&rq->lock, flags);
557 }
558
559 static int balance_runtime(struct rt_rq *rt_rq)
560 {
561         int more = 0;
562
563         if (rt_rq->rt_time > rt_rq->rt_runtime) {
564                 raw_spin_unlock(&rt_rq->rt_runtime_lock);
565                 more = do_balance_runtime(rt_rq);
566                 raw_spin_lock(&rt_rq->rt_runtime_lock);
567         }
568
569         return more;
570 }
571 #else /* !CONFIG_SMP */
572 static inline int balance_runtime(struct rt_rq *rt_rq)
573 {
574         return 0;
575 }
576 #endif /* CONFIG_SMP */
577
578 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
579 {
580         int i, idle = 1;
581         const struct cpumask *span;
582
583         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
584                 return 1;
585
586         span = sched_rt_period_mask();
587         for_each_cpu(i, span) {
588                 int enqueue = 0;
589                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
590                 struct rq *rq = rq_of_rt_rq(rt_rq);
591
592                 raw_spin_lock(&rq->lock);
593                 if (rt_rq->rt_time) {
594                         u64 runtime;
595
596                         raw_spin_lock(&rt_rq->rt_runtime_lock);
597                         if (rt_rq->rt_throttled)
598                                 balance_runtime(rt_rq);
599                         runtime = rt_rq->rt_runtime;
600                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
601                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
602                                 rt_rq->rt_throttled = 0;
603                                 enqueue = 1;
604
605                                 /*
606                                  * Force a clock update if the CPU was idle,
607                                  * lest wakeup -> unthrottle time accumulate.
608                                  */
609                                 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
610                                         rq->skip_clock_update = -1;
611                         }
612                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
613                                 idle = 0;
614                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
615                 } else if (rt_rq->rt_nr_running) {
616                         idle = 0;
617                         if (!rt_rq_throttled(rt_rq))
618                                 enqueue = 1;
619                 }
620
621                 if (enqueue)
622                         sched_rt_rq_enqueue(rt_rq);
623                 raw_spin_unlock(&rq->lock);
624         }
625
626         return idle;
627 }
628
629 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
630 {
631 #ifdef CONFIG_RT_GROUP_SCHED
632         struct rt_rq *rt_rq = group_rt_rq(rt_se);
633
634         if (rt_rq)
635                 return rt_rq->highest_prio.curr;
636 #endif
637
638         return rt_task_of(rt_se)->prio;
639 }
640
641 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
642 {
643         u64 runtime = sched_rt_runtime(rt_rq);
644
645         if (rt_rq->rt_throttled)
646                 return rt_rq_throttled(rt_rq);
647
648         if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
649                 return 0;
650
651         balance_runtime(rt_rq);
652         runtime = sched_rt_runtime(rt_rq);
653         if (runtime == RUNTIME_INF)
654                 return 0;
655
656         if (rt_rq->rt_time > runtime) {
657                 rt_rq->rt_throttled = 1;
658                 printk_once(KERN_WARNING "sched: RT throttling activated\n");
659                 if (rt_rq_throttled(rt_rq)) {
660                         sched_rt_rq_dequeue(rt_rq);
661                         return 1;
662                 }
663         }
664
665         return 0;
666 }
667
668 /*
669  * Update the current task's runtime statistics. Skip current tasks that
670  * are not in our scheduling class.
671  */
672 static void update_curr_rt(struct rq *rq)
673 {
674         struct task_struct *curr = rq->curr;
675         struct sched_rt_entity *rt_se = &curr->rt;
676         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
677         u64 delta_exec;
678
679         if (curr->sched_class != &rt_sched_class)
680                 return;
681
682         delta_exec = rq->clock_task - curr->se.exec_start;
683         if (unlikely((s64)delta_exec < 0))
684                 delta_exec = 0;
685
686         schedstat_set(curr->se.statistics.exec_max, max(curr->se.statistics.exec_max, delta_exec));
687
688         curr->se.sum_exec_runtime += delta_exec;
689         account_group_exec_runtime(curr, delta_exec);
690
691         curr->se.exec_start = rq->clock_task;
692         cpuacct_charge(curr, delta_exec);
693
694         sched_rt_avg_update(rq, delta_exec);
695
696         if (!rt_bandwidth_enabled())
697                 return;
698
699         for_each_sched_rt_entity(rt_se) {
700                 rt_rq = rt_rq_of_se(rt_se);
701
702                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
703                         raw_spin_lock(&rt_rq->rt_runtime_lock);
704                         rt_rq->rt_time += delta_exec;
705                         if (sched_rt_runtime_exceeded(rt_rq))
706                                 resched_task(curr);
707                         raw_spin_unlock(&rt_rq->rt_runtime_lock);
708                 }
709         }
710 }
711
712 #if defined CONFIG_SMP
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 (rq->online && prio < prev_prio)
720                 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
721 }
722
723 static void
724 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
725 {
726         struct rq *rq = rq_of_rt_rq(rt_rq);
727
728         if (rq->online && rt_rq->highest_prio.curr != prev_prio)
729                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
730 }
731
732 #else /* CONFIG_SMP */
733
734 static inline
735 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
736 static inline
737 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
738
739 #endif /* CONFIG_SMP */
740
741 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
742 static void
743 inc_rt_prio(struct rt_rq *rt_rq, int prio)
744 {
745         int prev_prio = rt_rq->highest_prio.curr;
746
747         if (prio < prev_prio)
748                 rt_rq->highest_prio.curr = prio;
749
750         inc_rt_prio_smp(rt_rq, prio, prev_prio);
751 }
752
753 static void
754 dec_rt_prio(struct rt_rq *rt_rq, int prio)
755 {
756         int prev_prio = rt_rq->highest_prio.curr;
757
758         if (rt_rq->rt_nr_running) {
759
760                 WARN_ON(prio < prev_prio);
761
762                 /*
763                  * This may have been our highest task, and therefore
764                  * we may have some recomputation to do
765                  */
766                 if (prio == prev_prio) {
767                         struct rt_prio_array *array = &rt_rq->active;
768
769                         rt_rq->highest_prio.curr =
770                                 sched_find_first_bit(array->bitmap);
771                 }
772
773         } else
774                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
775
776         dec_rt_prio_smp(rt_rq, prio, prev_prio);
777 }
778
779 #else
780
781 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
782 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
783
784 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
785
786 #ifdef CONFIG_RT_GROUP_SCHED
787
788 static void
789 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
790 {
791         if (rt_se_boosted(rt_se))
792                 rt_rq->rt_nr_boosted++;
793
794         if (rt_rq->tg)
795                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
796 }
797
798 static void
799 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
800 {
801         if (rt_se_boosted(rt_se))
802                 rt_rq->rt_nr_boosted--;
803
804         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
805 }
806
807 #else /* CONFIG_RT_GROUP_SCHED */
808
809 static void
810 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
811 {
812         start_rt_bandwidth(&def_rt_bandwidth);
813 }
814
815 static inline
816 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
817
818 #endif /* CONFIG_RT_GROUP_SCHED */
819
820 static inline
821 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
822 {
823         int prio = rt_se_prio(rt_se);
824
825         WARN_ON(!rt_prio(prio));
826         rt_rq->rt_nr_running++;
827
828         inc_rt_prio(rt_rq, prio);
829         inc_rt_migration(rt_se, rt_rq);
830         inc_rt_group(rt_se, rt_rq);
831 }
832
833 static inline
834 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
835 {
836         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
837         WARN_ON(!rt_rq->rt_nr_running);
838         rt_rq->rt_nr_running--;
839
840         dec_rt_prio(rt_rq, rt_se_prio(rt_se));
841         dec_rt_migration(rt_se, rt_rq);
842         dec_rt_group(rt_se, rt_rq);
843 }
844
845 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
846 {
847         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
848         struct rt_prio_array *array = &rt_rq->active;
849         struct rt_rq *group_rq = group_rt_rq(rt_se);
850         struct list_head *queue = array->queue + rt_se_prio(rt_se);
851
852         /*
853          * Don't enqueue the group if its throttled, or when empty.
854          * The latter is a consequence of the former when a child group
855          * get throttled and the current group doesn't have any other
856          * active members.
857          */
858         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
859                 return;
860
861         if (!rt_rq->rt_nr_running)
862                 list_add_leaf_rt_rq(rt_rq);
863
864         if (head)
865                 list_add(&rt_se->run_list, queue);
866         else
867                 list_add_tail(&rt_se->run_list, queue);
868         __set_bit(rt_se_prio(rt_se), array->bitmap);
869
870         inc_rt_tasks(rt_se, rt_rq);
871 }
872
873 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
874 {
875         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
876         struct rt_prio_array *array = &rt_rq->active;
877
878         list_del_init(&rt_se->run_list);
879         if (list_empty(array->queue + rt_se_prio(rt_se)))
880                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
881
882         dec_rt_tasks(rt_se, rt_rq);
883         if (!rt_rq->rt_nr_running)
884                 list_del_leaf_rt_rq(rt_rq);
885 }
886
887 /*
888  * Because the prio of an upper entry depends on the lower
889  * entries, we must remove entries top - down.
890  */
891 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
892 {
893         struct sched_rt_entity *back = NULL;
894
895         for_each_sched_rt_entity(rt_se) {
896                 rt_se->back = back;
897                 back = rt_se;
898         }
899
900         for (rt_se = back; rt_se; rt_se = rt_se->back) {
901                 if (on_rt_rq(rt_se))
902                         __dequeue_rt_entity(rt_se);
903         }
904 }
905
906 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
907 {
908         dequeue_rt_stack(rt_se);
909         for_each_sched_rt_entity(rt_se)
910                 __enqueue_rt_entity(rt_se, head);
911 }
912
913 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
914 {
915         dequeue_rt_stack(rt_se);
916
917         for_each_sched_rt_entity(rt_se) {
918                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
919
920                 if (rt_rq && rt_rq->rt_nr_running)
921                         __enqueue_rt_entity(rt_se, false);
922         }
923 }
924
925 /*
926  * Adding/removing a task to/from a priority array:
927  */
928 static void
929 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
930 {
931         struct sched_rt_entity *rt_se = &p->rt;
932
933         if (flags & ENQUEUE_WAKEUP)
934                 rt_se->timeout = 0;
935
936         enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
937
938         if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
939                 enqueue_pushable_task(rq, p);
940
941         inc_nr_running(rq);
942 }
943
944 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
945 {
946         struct sched_rt_entity *rt_se = &p->rt;
947
948         update_curr_rt(rq);
949         dequeue_rt_entity(rt_se);
950
951         dequeue_pushable_task(rq, p);
952
953         dec_nr_running(rq);
954 }
955
956 /*
957  * Put task to the end of the run list without the overhead of dequeue
958  * followed by enqueue.
959  */
960 static void
961 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
962 {
963         if (on_rt_rq(rt_se)) {
964                 struct rt_prio_array *array = &rt_rq->active;
965                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
966
967                 if (head)
968                         list_move(&rt_se->run_list, queue);
969                 else
970                         list_move_tail(&rt_se->run_list, queue);
971         }
972 }
973
974 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
975 {
976         struct sched_rt_entity *rt_se = &p->rt;
977         struct rt_rq *rt_rq;
978
979         for_each_sched_rt_entity(rt_se) {
980                 rt_rq = rt_rq_of_se(rt_se);
981                 requeue_rt_entity(rt_rq, rt_se, head);
982         }
983 }
984
985 static void yield_task_rt(struct rq *rq)
986 {
987         requeue_task_rt(rq, rq->curr, 0);
988 }
989
990 #ifdef CONFIG_SMP
991 static int find_lowest_rq(struct task_struct *task);
992
993 static int
994 select_task_rq_rt(struct task_struct *p, int sd_flag, int flags)
995 {
996         struct task_struct *curr;
997         struct rq *rq;
998         int cpu;
999
1000         cpu = task_cpu(p);
1001
1002         /* For anything but wake ups, just return the task_cpu */
1003         if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1004                 goto out;
1005
1006         rq = cpu_rq(cpu);
1007
1008         rcu_read_lock();
1009         curr = ACCESS_ONCE(rq->curr); /* unlocked access */
1010
1011         /*
1012          * If the current task on @p's runqueue is an RT task, then
1013          * try to see if we can wake this RT task up on another
1014          * runqueue. Otherwise simply start this RT task
1015          * on its current runqueue.
1016          *
1017          * We want to avoid overloading runqueues. If the woken
1018          * task is a higher priority, then it will stay on this CPU
1019          * and the lower prio task should be moved to another CPU.
1020          * Even though this will probably make the lower prio task
1021          * lose its cache, we do not want to bounce a higher task
1022          * around just because it gave up its CPU, perhaps for a
1023          * lock?
1024          *
1025          * For equal prio tasks, we just let the scheduler sort it out.
1026          *
1027          * Otherwise, just let it ride on the affined RQ and the
1028          * post-schedule router will push the preempted task away
1029          *
1030          * This test is optimistic, if we get it wrong the load-balancer
1031          * will have to sort it out.
1032          */
1033         if (curr && unlikely(rt_task(curr)) &&
1034             (curr->rt.nr_cpus_allowed < 2 ||
1035              curr->prio <= p->prio) &&
1036             (p->rt.nr_cpus_allowed > 1)) {
1037                 int target = find_lowest_rq(p);
1038
1039                 if (target != -1)
1040                         cpu = target;
1041         }
1042         rcu_read_unlock();
1043
1044 out:
1045         return cpu;
1046 }
1047
1048 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1049 {
1050         if (rq->curr->rt.nr_cpus_allowed == 1)
1051                 return;
1052
1053         if (p->rt.nr_cpus_allowed != 1
1054             && cpupri_find(&rq->rd->cpupri, p, NULL))
1055                 return;
1056
1057         if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1058                 return;
1059
1060         /*
1061          * There appears to be other cpus that can accept
1062          * current and none to run 'p', so lets reschedule
1063          * to try and push current away:
1064          */
1065         requeue_task_rt(rq, p, 1);
1066         resched_task(rq->curr);
1067 }
1068
1069 #endif /* CONFIG_SMP */
1070
1071 /*
1072  * Preempt the current task with a newly woken task if needed:
1073  */
1074 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1075 {
1076         if (p->prio < rq->curr->prio) {
1077                 resched_task(rq->curr);
1078                 return;
1079         }
1080
1081 #ifdef CONFIG_SMP
1082         /*
1083          * If:
1084          *
1085          * - the newly woken task is of equal priority to the current task
1086          * - the newly woken task is non-migratable while current is migratable
1087          * - current will be preempted on the next reschedule
1088          *
1089          * we should check to see if current can readily move to a different
1090          * cpu.  If so, we will reschedule to allow the push logic to try
1091          * to move current somewhere else, making room for our non-migratable
1092          * task.
1093          */
1094         if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1095                 check_preempt_equal_prio(rq, p);
1096 #endif
1097 }
1098
1099 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1100                                                    struct rt_rq *rt_rq)
1101 {
1102         struct rt_prio_array *array = &rt_rq->active;
1103         struct sched_rt_entity *next = NULL;
1104         struct list_head *queue;
1105         int idx;
1106
1107         idx = sched_find_first_bit(array->bitmap);
1108         BUG_ON(idx >= MAX_RT_PRIO);
1109
1110         queue = array->queue + idx;
1111         next = list_entry(queue->next, struct sched_rt_entity, run_list);
1112
1113         return next;
1114 }
1115
1116 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1117 {
1118         struct sched_rt_entity *rt_se;
1119         struct task_struct *p;
1120         struct rt_rq *rt_rq;
1121
1122         rt_rq = &rq->rt;
1123
1124         if (!rt_rq->rt_nr_running)
1125                 return NULL;
1126
1127         if (rt_rq_throttled(rt_rq))
1128                 return NULL;
1129
1130         do {
1131                 rt_se = pick_next_rt_entity(rq, rt_rq);
1132                 BUG_ON(!rt_se);
1133                 rt_rq = group_rt_rq(rt_se);
1134         } while (rt_rq);
1135
1136         p = rt_task_of(rt_se);
1137         p->se.exec_start = rq->clock_task;
1138
1139         return p;
1140 }
1141
1142 static struct task_struct *pick_next_task_rt(struct rq *rq)
1143 {
1144         struct task_struct *p = _pick_next_task_rt(rq);
1145
1146         /* The running task is never eligible for pushing */
1147         if (p)
1148                 dequeue_pushable_task(rq, p);
1149
1150 #ifdef CONFIG_SMP
1151         /*
1152          * We detect this state here so that we can avoid taking the RQ
1153          * lock again later if there is no need to push
1154          */
1155         rq->post_schedule = has_pushable_tasks(rq);
1156 #endif
1157
1158         return p;
1159 }
1160
1161 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1162 {
1163         update_curr_rt(rq);
1164
1165         /*
1166          * The previous task needs to be made eligible for pushing
1167          * if it is still active
1168          */
1169         if (on_rt_rq(&p->rt) && p->rt.nr_cpus_allowed > 1)
1170                 enqueue_pushable_task(rq, p);
1171 }
1172
1173 #ifdef CONFIG_SMP
1174
1175 /* Only try algorithms three times */
1176 #define RT_MAX_TRIES 3
1177
1178 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
1179
1180 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1181 {
1182         if (!task_running(rq, p) &&
1183             (cpu < 0 || cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) &&
1184             (p->rt.nr_cpus_allowed > 1))
1185                 return 1;
1186         return 0;
1187 }
1188
1189 /* Return the second highest RT task, NULL otherwise */
1190 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
1191 {
1192         struct task_struct *next = NULL;
1193         struct sched_rt_entity *rt_se;
1194         struct rt_prio_array *array;
1195         struct rt_rq *rt_rq;
1196         int idx;
1197
1198         for_each_leaf_rt_rq(rt_rq, rq) {
1199                 array = &rt_rq->active;
1200                 idx = sched_find_first_bit(array->bitmap);
1201 next_idx:
1202                 if (idx >= MAX_RT_PRIO)
1203                         continue;
1204                 if (next && next->prio < idx)
1205                         continue;
1206                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1207                         struct task_struct *p;
1208
1209                         if (!rt_entity_is_task(rt_se))
1210                                 continue;
1211
1212                         p = rt_task_of(rt_se);
1213                         if (pick_rt_task(rq, p, cpu)) {
1214                                 next = p;
1215                                 break;
1216                         }
1217                 }
1218                 if (!next) {
1219                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1220                         goto next_idx;
1221                 }
1222         }
1223
1224         return next;
1225 }
1226
1227 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1228
1229 static int find_lowest_rq(struct task_struct *task)
1230 {
1231         struct sched_domain *sd;
1232         struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1233         int this_cpu = smp_processor_id();
1234         int cpu      = task_cpu(task);
1235
1236         /* Make sure the mask is initialized first */
1237         if (unlikely(!lowest_mask))
1238                 return -1;
1239
1240         if (task->rt.nr_cpus_allowed == 1)
1241                 return -1; /* No other targets possible */
1242
1243         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1244                 return -1; /* No targets found */
1245
1246         /*
1247          * At this point we have built a mask of cpus representing the
1248          * lowest priority tasks in the system.  Now we want to elect
1249          * the best one based on our affinity and topology.
1250          *
1251          * We prioritize the last cpu that the task executed on since
1252          * it is most likely cache-hot in that location.
1253          */
1254         if (cpumask_test_cpu(cpu, lowest_mask))
1255                 return cpu;
1256
1257         /*
1258          * Otherwise, we consult the sched_domains span maps to figure
1259          * out which cpu is logically closest to our hot cache data.
1260          */
1261         if (!cpumask_test_cpu(this_cpu, lowest_mask))
1262                 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1263
1264         rcu_read_lock();
1265         for_each_domain(cpu, sd) {
1266                 if (sd->flags & SD_WAKE_AFFINE) {
1267                         int best_cpu;
1268
1269                         /*
1270                          * "this_cpu" is cheaper to preempt than a
1271                          * remote processor.
1272                          */
1273                         if (this_cpu != -1 &&
1274                             cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1275                                 rcu_read_unlock();
1276                                 return this_cpu;
1277                         }
1278
1279                         best_cpu = cpumask_first_and(lowest_mask,
1280                                                      sched_domain_span(sd));
1281                         if (best_cpu < nr_cpu_ids) {
1282                                 rcu_read_unlock();
1283                                 return best_cpu;
1284                         }
1285                 }
1286         }
1287         rcu_read_unlock();
1288
1289         /*
1290          * And finally, if there were no matches within the domains
1291          * just give the caller *something* to work with from the compatible
1292          * locations.
1293          */
1294         if (this_cpu != -1)
1295                 return this_cpu;
1296
1297         cpu = cpumask_any(lowest_mask);
1298         if (cpu < nr_cpu_ids)
1299                 return cpu;
1300         return -1;
1301 }
1302
1303 /* Will lock the rq it finds */
1304 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1305 {
1306         struct rq *lowest_rq = NULL;
1307         int tries;
1308         int cpu;
1309
1310         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1311                 cpu = find_lowest_rq(task);
1312
1313                 if ((cpu == -1) || (cpu == rq->cpu))
1314                         break;
1315
1316                 lowest_rq = cpu_rq(cpu);
1317
1318                 /* if the prio of this runqueue changed, try again */
1319                 if (double_lock_balance(rq, lowest_rq)) {
1320                         /*
1321                          * We had to unlock the run queue. In
1322                          * the mean time, task could have
1323                          * migrated already or had its affinity changed.
1324                          * Also make sure that it wasn't scheduled on its rq.
1325                          */
1326                         if (unlikely(task_rq(task) != rq ||
1327                                      !cpumask_test_cpu(lowest_rq->cpu,
1328                                                        tsk_cpus_allowed(task)) ||
1329                                      task_running(rq, task) ||
1330                                      !task->on_rq)) {
1331
1332                                 raw_spin_unlock(&lowest_rq->lock);
1333                                 lowest_rq = NULL;
1334                                 break;
1335                         }
1336                 }
1337
1338                 /* If this rq is still suitable use it. */
1339                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1340                         break;
1341
1342                 /* try again */
1343                 double_unlock_balance(rq, lowest_rq);
1344                 lowest_rq = NULL;
1345         }
1346
1347         return lowest_rq;
1348 }
1349
1350 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1351 {
1352         struct task_struct *p;
1353
1354         if (!has_pushable_tasks(rq))
1355                 return NULL;
1356
1357         p = plist_first_entry(&rq->rt.pushable_tasks,
1358                               struct task_struct, pushable_tasks);
1359
1360         BUG_ON(rq->cpu != task_cpu(p));
1361         BUG_ON(task_current(rq, p));
1362         BUG_ON(p->rt.nr_cpus_allowed <= 1);
1363
1364         BUG_ON(!p->on_rq);
1365         BUG_ON(!rt_task(p));
1366
1367         return p;
1368 }
1369
1370 /*
1371  * If the current CPU has more than one RT task, see if the non
1372  * running task can migrate over to a CPU that is running a task
1373  * of lesser priority.
1374  */
1375 static int push_rt_task(struct rq *rq)
1376 {
1377         struct task_struct *next_task;
1378         struct rq *lowest_rq;
1379         int ret = 0;
1380
1381         if (!rq->rt.overloaded)
1382                 return 0;
1383
1384         next_task = pick_next_pushable_task(rq);
1385         if (!next_task)
1386                 return 0;
1387
1388 retry:
1389         if (unlikely(next_task == rq->curr)) {
1390                 WARN_ON(1);
1391                 return 0;
1392         }
1393
1394         /*
1395          * It's possible that the next_task slipped in of
1396          * higher priority than current. If that's the case
1397          * just reschedule current.
1398          */
1399         if (unlikely(next_task->prio < rq->curr->prio)) {
1400                 resched_task(rq->curr);
1401                 return 0;
1402         }
1403
1404         /* We might release rq lock */
1405         get_task_struct(next_task);
1406
1407         /* find_lock_lowest_rq locks the rq if found */
1408         lowest_rq = find_lock_lowest_rq(next_task, rq);
1409         if (!lowest_rq) {
1410                 struct task_struct *task;
1411                 /*
1412                  * find_lock_lowest_rq releases rq->lock
1413                  * so it is possible that next_task has migrated.
1414                  *
1415                  * We need to make sure that the task is still on the same
1416                  * run-queue and is also still the next task eligible for
1417                  * pushing.
1418                  */
1419                 task = pick_next_pushable_task(rq);
1420                 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1421                         /*
1422                          * The task hasn't migrated, and is still the next
1423                          * eligible task, but we failed to find a run-queue
1424                          * to push it to.  Do not retry in this case, since
1425                          * other cpus will pull from us when ready.
1426                          */
1427                         goto out;
1428                 }
1429
1430                 if (!task)
1431                         /* No more tasks, just exit */
1432                         goto out;
1433
1434                 /*
1435                  * Something has shifted, try again.
1436                  */
1437                 put_task_struct(next_task);
1438                 next_task = task;
1439                 goto retry;
1440         }
1441
1442         deactivate_task(rq, next_task, 0);
1443         set_task_cpu(next_task, lowest_rq->cpu);
1444         activate_task(lowest_rq, next_task, 0);
1445         ret = 1;
1446
1447         resched_task(lowest_rq->curr);
1448
1449         double_unlock_balance(rq, lowest_rq);
1450
1451 out:
1452         put_task_struct(next_task);
1453
1454         return ret;
1455 }
1456
1457 static void push_rt_tasks(struct rq *rq)
1458 {
1459         /* push_rt_task will return true if it moved an RT */
1460         while (push_rt_task(rq))
1461                 ;
1462 }
1463
1464 static int pull_rt_task(struct rq *this_rq)
1465 {
1466         int this_cpu = this_rq->cpu, ret = 0, cpu;
1467         struct task_struct *p;
1468         struct rq *src_rq;
1469
1470         if (likely(!rt_overloaded(this_rq)))
1471                 return 0;
1472
1473         for_each_cpu(cpu, this_rq->rd->rto_mask) {
1474                 if (this_cpu == cpu)
1475                         continue;
1476
1477                 src_rq = cpu_rq(cpu);
1478
1479                 /*
1480                  * Don't bother taking the src_rq->lock if the next highest
1481                  * task is known to be lower-priority than our current task.
1482                  * This may look racy, but if this value is about to go
1483                  * logically higher, the src_rq will push this task away.
1484                  * And if its going logically lower, we do not care
1485                  */
1486                 if (src_rq->rt.highest_prio.next >=
1487                     this_rq->rt.highest_prio.curr)
1488                         continue;
1489
1490                 /*
1491                  * We can potentially drop this_rq's lock in
1492                  * double_lock_balance, and another CPU could
1493                  * alter this_rq
1494                  */
1495                 double_lock_balance(this_rq, src_rq);
1496
1497                 /*
1498                  * Are there still pullable RT tasks?
1499                  */
1500                 if (src_rq->rt.rt_nr_running <= 1)
1501                         goto skip;
1502
1503                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1504
1505                 /*
1506                  * Do we have an RT task that preempts
1507                  * the to-be-scheduled task?
1508                  */
1509                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1510                         WARN_ON(p == src_rq->curr);
1511                         WARN_ON(!p->on_rq);
1512
1513                         /*
1514                          * There's a chance that p is higher in priority
1515                          * than what's currently running on its cpu.
1516                          * This is just that p is wakeing up and hasn't
1517                          * had a chance to schedule. We only pull
1518                          * p if it is lower in priority than the
1519                          * current task on the run queue
1520                          */
1521                         if (p->prio < src_rq->curr->prio)
1522                                 goto skip;
1523
1524                         ret = 1;
1525
1526                         deactivate_task(src_rq, p, 0);
1527                         set_task_cpu(p, this_cpu);
1528                         activate_task(this_rq, p, 0);
1529                         /*
1530                          * We continue with the search, just in
1531                          * case there's an even higher prio task
1532                          * in another runqueue. (low likelihood
1533                          * but possible)
1534                          */
1535                 }
1536 skip:
1537                 double_unlock_balance(this_rq, src_rq);
1538         }
1539
1540         return ret;
1541 }
1542
1543 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1544 {
1545         /* Try to pull RT tasks here if we lower this rq's prio */
1546         if (rq->rt.highest_prio.curr > prev->prio)
1547                 pull_rt_task(rq);
1548 }
1549
1550 static void post_schedule_rt(struct rq *rq)
1551 {
1552         push_rt_tasks(rq);
1553 }
1554
1555 /*
1556  * If we are not running and we are not going to reschedule soon, we should
1557  * try to push tasks away now
1558  */
1559 static void task_woken_rt(struct rq *rq, struct task_struct *p)
1560 {
1561         if (!task_running(rq, p) &&
1562             !test_tsk_need_resched(rq->curr) &&
1563             has_pushable_tasks(rq) &&
1564             p->rt.nr_cpus_allowed > 1 &&
1565             rt_task(rq->curr) &&
1566             (rq->curr->rt.nr_cpus_allowed < 2 ||
1567              rq->curr->prio <= p->prio))
1568                 push_rt_tasks(rq);
1569 }
1570
1571 static void set_cpus_allowed_rt(struct task_struct *p,
1572                                 const struct cpumask *new_mask)
1573 {
1574         int weight = cpumask_weight(new_mask);
1575
1576         BUG_ON(!rt_task(p));
1577
1578         /*
1579          * Update the migration status of the RQ if we have an RT task
1580          * which is running AND changing its weight value.
1581          */
1582         if (p->on_rq && (weight != p->rt.nr_cpus_allowed)) {
1583                 struct rq *rq = task_rq(p);
1584
1585                 if (!task_current(rq, p)) {
1586                         /*
1587                          * Make sure we dequeue this task from the pushable list
1588                          * before going further.  It will either remain off of
1589                          * the list because we are no longer pushable, or it
1590                          * will be requeued.
1591                          */
1592                         if (p->rt.nr_cpus_allowed > 1)
1593                                 dequeue_pushable_task(rq, p);
1594
1595                         /*
1596                          * Requeue if our weight is changing and still > 1
1597                          */
1598                         if (weight > 1)
1599                                 enqueue_pushable_task(rq, p);
1600
1601                 }
1602
1603                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1604                         rq->rt.rt_nr_migratory++;
1605                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1606                         BUG_ON(!rq->rt.rt_nr_migratory);
1607                         rq->rt.rt_nr_migratory--;
1608                 }
1609
1610                 update_rt_migration(&rq->rt);
1611         }
1612 }
1613
1614 /* Assumes rq->lock is held */
1615 static void rq_online_rt(struct rq *rq)
1616 {
1617         if (rq->rt.overloaded)
1618                 rt_set_overload(rq);
1619
1620         __enable_runtime(rq);
1621
1622         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1623 }
1624
1625 /* Assumes rq->lock is held */
1626 static void rq_offline_rt(struct rq *rq)
1627 {
1628         if (rq->rt.overloaded)
1629                 rt_clear_overload(rq);
1630
1631         __disable_runtime(rq);
1632
1633         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1634 }
1635
1636 /*
1637  * When switch from the rt queue, we bring ourselves to a position
1638  * that we might want to pull RT tasks from other runqueues.
1639  */
1640 static void switched_from_rt(struct rq *rq, struct task_struct *p)
1641 {
1642         /*
1643          * If there are other RT tasks then we will reschedule
1644          * and the scheduling of the other RT tasks will handle
1645          * the balancing. But if we are the last RT task
1646          * we may need to handle the pulling of RT tasks
1647          * now.
1648          */
1649         if (p->on_rq && !rq->rt.rt_nr_running)
1650                 pull_rt_task(rq);
1651 }
1652
1653 static inline void init_sched_rt_class(void)
1654 {
1655         unsigned int i;
1656
1657         for_each_possible_cpu(i)
1658                 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1659                                         GFP_KERNEL, cpu_to_node(i));
1660 }
1661 #endif /* CONFIG_SMP */
1662
1663 /*
1664  * When switching a task to RT, we may overload the runqueue
1665  * with RT tasks. In this case we try to push them off to
1666  * other runqueues.
1667  */
1668 static void switched_to_rt(struct rq *rq, struct task_struct *p)
1669 {
1670         int check_resched = 1;
1671
1672         /*
1673          * If we are already running, then there's nothing
1674          * that needs to be done. But if we are not running
1675          * we may need to preempt the current running task.
1676          * If that current running task is also an RT task
1677          * then see if we can move to another run queue.
1678          */
1679         if (p->on_rq && rq->curr != p) {
1680 #ifdef CONFIG_SMP
1681                 if (rq->rt.overloaded && push_rt_task(rq) &&
1682                     /* Don't resched if we changed runqueues */
1683                     rq != task_rq(p))
1684                         check_resched = 0;
1685 #endif /* CONFIG_SMP */
1686                 if (check_resched && p->prio < rq->curr->prio)
1687                         resched_task(rq->curr);
1688         }
1689 }
1690
1691 /*
1692  * Priority of the task has changed. This may cause
1693  * us to initiate a push or pull.
1694  */
1695 static void
1696 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
1697 {
1698         if (!p->on_rq)
1699                 return;
1700
1701         if (rq->curr == p) {
1702 #ifdef CONFIG_SMP
1703                 /*
1704                  * If our priority decreases while running, we
1705                  * may need to pull tasks to this runqueue.
1706                  */
1707                 if (oldprio < p->prio)
1708                         pull_rt_task(rq);
1709                 /*
1710                  * If there's a higher priority task waiting to run
1711                  * then reschedule. Note, the above pull_rt_task
1712                  * can release the rq lock and p could migrate.
1713                  * Only reschedule if p is still on the same runqueue.
1714                  */
1715                 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1716                         resched_task(p);
1717 #else
1718                 /* For UP simply resched on drop of prio */
1719                 if (oldprio < p->prio)
1720                         resched_task(p);
1721 #endif /* CONFIG_SMP */
1722         } else {
1723                 /*
1724                  * This task is not running, but if it is
1725                  * greater than the current running task
1726                  * then reschedule.
1727                  */
1728                 if (p->prio < rq->curr->prio)
1729                         resched_task(rq->curr);
1730         }
1731 }
1732
1733 static void watchdog(struct rq *rq, struct task_struct *p)
1734 {
1735         unsigned long soft, hard;
1736
1737         /* max may change after cur was read, this will be fixed next tick */
1738         soft = task_rlimit(p, RLIMIT_RTTIME);
1739         hard = task_rlimit_max(p, RLIMIT_RTTIME);
1740
1741         if (soft != RLIM_INFINITY) {
1742                 unsigned long next;
1743
1744                 p->rt.timeout++;
1745                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1746                 if (p->rt.timeout > next)
1747                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1748         }
1749 }
1750
1751 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1752 {
1753         update_curr_rt(rq);
1754
1755         watchdog(rq, p);
1756
1757         /*
1758          * RR tasks need a special form of timeslice management.
1759          * FIFO tasks have no timeslices.
1760          */
1761         if (p->policy != SCHED_RR)
1762                 return;
1763
1764         if (--p->rt.time_slice)
1765                 return;
1766
1767         p->rt.time_slice = DEF_TIMESLICE;
1768
1769         /*
1770          * Requeue to the end of queue if we are not the only element
1771          * on the queue:
1772          */
1773         if (p->rt.run_list.prev != p->rt.run_list.next) {
1774                 requeue_task_rt(rq, p, 0);
1775                 set_tsk_need_resched(p);
1776         }
1777 }
1778
1779 static void set_curr_task_rt(struct rq *rq)
1780 {
1781         struct task_struct *p = rq->curr;
1782
1783         p->se.exec_start = rq->clock_task;
1784
1785         /* The running task is never eligible for pushing */
1786         dequeue_pushable_task(rq, p);
1787 }
1788
1789 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
1790 {
1791         /*
1792          * Time slice is 0 for SCHED_FIFO tasks
1793          */
1794         if (task->policy == SCHED_RR)
1795                 return DEF_TIMESLICE;
1796         else
1797                 return 0;
1798 }
1799
1800 static const struct sched_class rt_sched_class = {
1801         .next                   = &fair_sched_class,
1802         .enqueue_task           = enqueue_task_rt,
1803         .dequeue_task           = dequeue_task_rt,
1804         .yield_task             = yield_task_rt,
1805
1806         .check_preempt_curr     = check_preempt_curr_rt,
1807
1808         .pick_next_task         = pick_next_task_rt,
1809         .put_prev_task          = put_prev_task_rt,
1810
1811 #ifdef CONFIG_SMP
1812         .select_task_rq         = select_task_rq_rt,
1813
1814         .set_cpus_allowed       = set_cpus_allowed_rt,
1815         .rq_online              = rq_online_rt,
1816         .rq_offline             = rq_offline_rt,
1817         .pre_schedule           = pre_schedule_rt,
1818         .post_schedule          = post_schedule_rt,
1819         .task_woken             = task_woken_rt,
1820         .switched_from          = switched_from_rt,
1821 #endif
1822
1823         .set_curr_task          = set_curr_task_rt,
1824         .task_tick              = task_tick_rt,
1825
1826         .get_rr_interval        = get_rr_interval_rt,
1827
1828         .prio_changed           = prio_changed_rt,
1829         .switched_to            = switched_to_rt,
1830 };
1831
1832 #ifdef CONFIG_SCHED_DEBUG
1833 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1834
1835 static void print_rt_stats(struct seq_file *m, int cpu)
1836 {
1837         rt_rq_iter_t iter;
1838         struct rt_rq *rt_rq;
1839
1840         rcu_read_lock();
1841         for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
1842                 print_rt_rq(m, cpu, rt_rq);
1843         rcu_read_unlock();
1844 }
1845 #endif /* CONFIG_SCHED_DEBUG */