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