Merge branch 'core/percpu' into percpu-cpumask-x86-for-linus-2
[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 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
14 {
15         return rt_rq->rq;
16 }
17
18 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
19 {
20         return rt_se->rt_rq;
21 }
22
23 #else /* CONFIG_RT_GROUP_SCHED */
24
25 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
26 {
27         return container_of(rt_rq, struct rq, rt);
28 }
29
30 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
31 {
32         struct task_struct *p = rt_task_of(rt_se);
33         struct rq *rq = task_rq(p);
34
35         return &rq->rt;
36 }
37
38 #endif /* CONFIG_RT_GROUP_SCHED */
39
40 #ifdef CONFIG_SMP
41
42 static inline int rt_overloaded(struct rq *rq)
43 {
44         return atomic_read(&rq->rd->rto_count);
45 }
46
47 static inline void rt_set_overload(struct rq *rq)
48 {
49         if (!rq->online)
50                 return;
51
52         cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
53         /*
54          * Make sure the mask is visible before we set
55          * the overload count. That is checked to determine
56          * if we should look at the mask. It would be a shame
57          * if we looked at the mask, but the mask was not
58          * updated yet.
59          */
60         wmb();
61         atomic_inc(&rq->rd->rto_count);
62 }
63
64 static inline void rt_clear_overload(struct rq *rq)
65 {
66         if (!rq->online)
67                 return;
68
69         /* the order here really doesn't matter */
70         atomic_dec(&rq->rd->rto_count);
71         cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
72 }
73
74 static void update_rt_migration(struct rt_rq *rt_rq)
75 {
76         if (rt_rq->rt_nr_migratory && (rt_rq->rt_nr_running > 1)) {
77                 if (!rt_rq->overloaded) {
78                         rt_set_overload(rq_of_rt_rq(rt_rq));
79                         rt_rq->overloaded = 1;
80                 }
81         } else if (rt_rq->overloaded) {
82                 rt_clear_overload(rq_of_rt_rq(rt_rq));
83                 rt_rq->overloaded = 0;
84         }
85 }
86
87 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
88 {
89         if (rt_se->nr_cpus_allowed > 1)
90                 rt_rq->rt_nr_migratory++;
91
92         update_rt_migration(rt_rq);
93 }
94
95 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
96 {
97         if (rt_se->nr_cpus_allowed > 1)
98                 rt_rq->rt_nr_migratory--;
99
100         update_rt_migration(rt_rq);
101 }
102
103 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
104 {
105         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
106         plist_node_init(&p->pushable_tasks, p->prio);
107         plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
108 }
109
110 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
111 {
112         plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
113 }
114
115 #else
116
117 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
118 {
119 }
120
121 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
122 {
123 }
124
125 static inline
126 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
127 {
128 }
129
130 static inline
131 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
132 {
133 }
134
135 #endif /* CONFIG_SMP */
136
137 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
138 {
139         return !list_empty(&rt_se->run_list);
140 }
141
142 #ifdef CONFIG_RT_GROUP_SCHED
143
144 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
145 {
146         if (!rt_rq->tg)
147                 return RUNTIME_INF;
148
149         return rt_rq->rt_runtime;
150 }
151
152 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
153 {
154         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
155 }
156
157 #define for_each_leaf_rt_rq(rt_rq, rq) \
158         list_for_each_entry_rcu(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
159
160 #define for_each_sched_rt_entity(rt_se) \
161         for (; rt_se; rt_se = rt_se->parent)
162
163 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
164 {
165         return rt_se->my_q;
166 }
167
168 static void enqueue_rt_entity(struct sched_rt_entity *rt_se);
169 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
170
171 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
172 {
173         struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
174         struct sched_rt_entity *rt_se = rt_rq->rt_se;
175
176         if (rt_rq->rt_nr_running) {
177                 if (rt_se && !on_rt_rq(rt_se))
178                         enqueue_rt_entity(rt_se);
179                 if (rt_rq->highest_prio.curr < curr->prio)
180                         resched_task(curr);
181         }
182 }
183
184 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
185 {
186         struct sched_rt_entity *rt_se = rt_rq->rt_se;
187
188         if (rt_se && on_rt_rq(rt_se))
189                 dequeue_rt_entity(rt_se);
190 }
191
192 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
193 {
194         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
195 }
196
197 static int rt_se_boosted(struct sched_rt_entity *rt_se)
198 {
199         struct rt_rq *rt_rq = group_rt_rq(rt_se);
200         struct task_struct *p;
201
202         if (rt_rq)
203                 return !!rt_rq->rt_nr_boosted;
204
205         p = rt_task_of(rt_se);
206         return p->prio != p->normal_prio;
207 }
208
209 #ifdef CONFIG_SMP
210 static inline const struct cpumask *sched_rt_period_mask(void)
211 {
212         return cpu_rq(smp_processor_id())->rd->span;
213 }
214 #else
215 static inline const struct cpumask *sched_rt_period_mask(void)
216 {
217         return cpu_online_mask;
218 }
219 #endif
220
221 static inline
222 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
223 {
224         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
225 }
226
227 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
228 {
229         return &rt_rq->tg->rt_bandwidth;
230 }
231
232 #else /* !CONFIG_RT_GROUP_SCHED */
233
234 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
235 {
236         return rt_rq->rt_runtime;
237 }
238
239 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
240 {
241         return ktime_to_ns(def_rt_bandwidth.rt_period);
242 }
243
244 #define for_each_leaf_rt_rq(rt_rq, rq) \
245         for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
246
247 #define for_each_sched_rt_entity(rt_se) \
248         for (; rt_se; rt_se = NULL)
249
250 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
251 {
252         return NULL;
253 }
254
255 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
256 {
257         if (rt_rq->rt_nr_running)
258                 resched_task(rq_of_rt_rq(rt_rq)->curr);
259 }
260
261 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
262 {
263 }
264
265 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
266 {
267         return rt_rq->rt_throttled;
268 }
269
270 static inline const struct cpumask *sched_rt_period_mask(void)
271 {
272         return cpu_online_mask;
273 }
274
275 static inline
276 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
277 {
278         return &cpu_rq(cpu)->rt;
279 }
280
281 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
282 {
283         return &def_rt_bandwidth;
284 }
285
286 #endif /* CONFIG_RT_GROUP_SCHED */
287
288 #ifdef CONFIG_SMP
289 /*
290  * We ran out of runtime, see if we can borrow some from our neighbours.
291  */
292 static int do_balance_runtime(struct rt_rq *rt_rq)
293 {
294         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
295         struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
296         int i, weight, more = 0;
297         u64 rt_period;
298
299         weight = cpumask_weight(rd->span);
300
301         spin_lock(&rt_b->rt_runtime_lock);
302         rt_period = ktime_to_ns(rt_b->rt_period);
303         for_each_cpu(i, rd->span) {
304                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
305                 s64 diff;
306
307                 if (iter == rt_rq)
308                         continue;
309
310                 spin_lock(&iter->rt_runtime_lock);
311                 /*
312                  * Either all rqs have inf runtime and there's nothing to steal
313                  * or __disable_runtime() below sets a specific rq to inf to
314                  * indicate its been disabled and disalow stealing.
315                  */
316                 if (iter->rt_runtime == RUNTIME_INF)
317                         goto next;
318
319                 /*
320                  * From runqueues with spare time, take 1/n part of their
321                  * spare time, but no more than our period.
322                  */
323                 diff = iter->rt_runtime - iter->rt_time;
324                 if (diff > 0) {
325                         diff = div_u64((u64)diff, weight);
326                         if (rt_rq->rt_runtime + diff > rt_period)
327                                 diff = rt_period - rt_rq->rt_runtime;
328                         iter->rt_runtime -= diff;
329                         rt_rq->rt_runtime += diff;
330                         more = 1;
331                         if (rt_rq->rt_runtime == rt_period) {
332                                 spin_unlock(&iter->rt_runtime_lock);
333                                 break;
334                         }
335                 }
336 next:
337                 spin_unlock(&iter->rt_runtime_lock);
338         }
339         spin_unlock(&rt_b->rt_runtime_lock);
340
341         return more;
342 }
343
344 /*
345  * Ensure this RQ takes back all the runtime it lend to its neighbours.
346  */
347 static void __disable_runtime(struct rq *rq)
348 {
349         struct root_domain *rd = rq->rd;
350         struct rt_rq *rt_rq;
351
352         if (unlikely(!scheduler_running))
353                 return;
354
355         for_each_leaf_rt_rq(rt_rq, rq) {
356                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
357                 s64 want;
358                 int i;
359
360                 spin_lock(&rt_b->rt_runtime_lock);
361                 spin_lock(&rt_rq->rt_runtime_lock);
362                 /*
363                  * Either we're all inf and nobody needs to borrow, or we're
364                  * already disabled and thus have nothing to do, or we have
365                  * exactly the right amount of runtime to take out.
366                  */
367                 if (rt_rq->rt_runtime == RUNTIME_INF ||
368                                 rt_rq->rt_runtime == rt_b->rt_runtime)
369                         goto balanced;
370                 spin_unlock(&rt_rq->rt_runtime_lock);
371
372                 /*
373                  * Calculate the difference between what we started out with
374                  * and what we current have, that's the amount of runtime
375                  * we lend and now have to reclaim.
376                  */
377                 want = rt_b->rt_runtime - rt_rq->rt_runtime;
378
379                 /*
380                  * Greedy reclaim, take back as much as we can.
381                  */
382                 for_each_cpu(i, rd->span) {
383                         struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
384                         s64 diff;
385
386                         /*
387                          * Can't reclaim from ourselves or disabled runqueues.
388                          */
389                         if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
390                                 continue;
391
392                         spin_lock(&iter->rt_runtime_lock);
393                         if (want > 0) {
394                                 diff = min_t(s64, iter->rt_runtime, want);
395                                 iter->rt_runtime -= diff;
396                                 want -= diff;
397                         } else {
398                                 iter->rt_runtime -= want;
399                                 want -= want;
400                         }
401                         spin_unlock(&iter->rt_runtime_lock);
402
403                         if (!want)
404                                 break;
405                 }
406
407                 spin_lock(&rt_rq->rt_runtime_lock);
408                 /*
409                  * We cannot be left wanting - that would mean some runtime
410                  * leaked out of the system.
411                  */
412                 BUG_ON(want);
413 balanced:
414                 /*
415                  * Disable all the borrow logic by pretending we have inf
416                  * runtime - in which case borrowing doesn't make sense.
417                  */
418                 rt_rq->rt_runtime = RUNTIME_INF;
419                 spin_unlock(&rt_rq->rt_runtime_lock);
420                 spin_unlock(&rt_b->rt_runtime_lock);
421         }
422 }
423
424 static void disable_runtime(struct rq *rq)
425 {
426         unsigned long flags;
427
428         spin_lock_irqsave(&rq->lock, flags);
429         __disable_runtime(rq);
430         spin_unlock_irqrestore(&rq->lock, flags);
431 }
432
433 static void __enable_runtime(struct rq *rq)
434 {
435         struct rt_rq *rt_rq;
436
437         if (unlikely(!scheduler_running))
438                 return;
439
440         /*
441          * Reset each runqueue's bandwidth settings
442          */
443         for_each_leaf_rt_rq(rt_rq, rq) {
444                 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
445
446                 spin_lock(&rt_b->rt_runtime_lock);
447                 spin_lock(&rt_rq->rt_runtime_lock);
448                 rt_rq->rt_runtime = rt_b->rt_runtime;
449                 rt_rq->rt_time = 0;
450                 rt_rq->rt_throttled = 0;
451                 spin_unlock(&rt_rq->rt_runtime_lock);
452                 spin_unlock(&rt_b->rt_runtime_lock);
453         }
454 }
455
456 static void enable_runtime(struct rq *rq)
457 {
458         unsigned long flags;
459
460         spin_lock_irqsave(&rq->lock, flags);
461         __enable_runtime(rq);
462         spin_unlock_irqrestore(&rq->lock, flags);
463 }
464
465 static int balance_runtime(struct rt_rq *rt_rq)
466 {
467         int more = 0;
468
469         if (rt_rq->rt_time > rt_rq->rt_runtime) {
470                 spin_unlock(&rt_rq->rt_runtime_lock);
471                 more = do_balance_runtime(rt_rq);
472                 spin_lock(&rt_rq->rt_runtime_lock);
473         }
474
475         return more;
476 }
477 #else /* !CONFIG_SMP */
478 static inline int balance_runtime(struct rt_rq *rt_rq)
479 {
480         return 0;
481 }
482 #endif /* CONFIG_SMP */
483
484 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
485 {
486         int i, idle = 1;
487         const struct cpumask *span;
488
489         if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
490                 return 1;
491
492         span = sched_rt_period_mask();
493         for_each_cpu(i, span) {
494                 int enqueue = 0;
495                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
496                 struct rq *rq = rq_of_rt_rq(rt_rq);
497
498                 spin_lock(&rq->lock);
499                 if (rt_rq->rt_time) {
500                         u64 runtime;
501
502                         spin_lock(&rt_rq->rt_runtime_lock);
503                         if (rt_rq->rt_throttled)
504                                 balance_runtime(rt_rq);
505                         runtime = rt_rq->rt_runtime;
506                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
507                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
508                                 rt_rq->rt_throttled = 0;
509                                 enqueue = 1;
510                         }
511                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
512                                 idle = 0;
513                         spin_unlock(&rt_rq->rt_runtime_lock);
514                 } else if (rt_rq->rt_nr_running)
515                         idle = 0;
516
517                 if (enqueue)
518                         sched_rt_rq_enqueue(rt_rq);
519                 spin_unlock(&rq->lock);
520         }
521
522         return idle;
523 }
524
525 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
526 {
527 #ifdef CONFIG_RT_GROUP_SCHED
528         struct rt_rq *rt_rq = group_rt_rq(rt_se);
529
530         if (rt_rq)
531                 return rt_rq->highest_prio.curr;
532 #endif
533
534         return rt_task_of(rt_se)->prio;
535 }
536
537 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
538 {
539         u64 runtime = sched_rt_runtime(rt_rq);
540
541         if (rt_rq->rt_throttled)
542                 return rt_rq_throttled(rt_rq);
543
544         if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
545                 return 0;
546
547         balance_runtime(rt_rq);
548         runtime = sched_rt_runtime(rt_rq);
549         if (runtime == RUNTIME_INF)
550                 return 0;
551
552         if (rt_rq->rt_time > runtime) {
553                 rt_rq->rt_throttled = 1;
554                 if (rt_rq_throttled(rt_rq)) {
555                         sched_rt_rq_dequeue(rt_rq);
556                         return 1;
557                 }
558         }
559
560         return 0;
561 }
562
563 /*
564  * Update the current task's runtime statistics. Skip current tasks that
565  * are not in our scheduling class.
566  */
567 static void update_curr_rt(struct rq *rq)
568 {
569         struct task_struct *curr = rq->curr;
570         struct sched_rt_entity *rt_se = &curr->rt;
571         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
572         u64 delta_exec;
573
574         if (!task_has_rt_policy(curr))
575                 return;
576
577         delta_exec = rq->clock - curr->se.exec_start;
578         if (unlikely((s64)delta_exec < 0))
579                 delta_exec = 0;
580
581         schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
582
583         curr->se.sum_exec_runtime += delta_exec;
584         account_group_exec_runtime(curr, delta_exec);
585
586         curr->se.exec_start = rq->clock;
587         cpuacct_charge(curr, delta_exec);
588
589         if (!rt_bandwidth_enabled())
590                 return;
591
592         for_each_sched_rt_entity(rt_se) {
593                 rt_rq = rt_rq_of_se(rt_se);
594
595                 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
596                         spin_lock(&rt_rq->rt_runtime_lock);
597                         rt_rq->rt_time += delta_exec;
598                         if (sched_rt_runtime_exceeded(rt_rq))
599                                 resched_task(curr);
600                         spin_unlock(&rt_rq->rt_runtime_lock);
601                 }
602         }
603 }
604
605 #if defined CONFIG_SMP
606
607 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu);
608
609 static inline int next_prio(struct rq *rq)
610 {
611         struct task_struct *next = pick_next_highest_task_rt(rq, rq->cpu);
612
613         if (next && rt_prio(next->prio))
614                 return next->prio;
615         else
616                 return MAX_RT_PRIO;
617 }
618
619 static void
620 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
621 {
622         struct rq *rq = rq_of_rt_rq(rt_rq);
623
624         if (prio < prev_prio) {
625
626                 /*
627                  * If the new task is higher in priority than anything on the
628                  * run-queue, we know that the previous high becomes our
629                  * next-highest.
630                  */
631                 rt_rq->highest_prio.next = prev_prio;
632
633                 if (rq->online)
634                         cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
635
636         } else if (prio == rt_rq->highest_prio.curr)
637                 /*
638                  * If the next task is equal in priority to the highest on
639                  * the run-queue, then we implicitly know that the next highest
640                  * task cannot be any lower than current
641                  */
642                 rt_rq->highest_prio.next = prio;
643         else if (prio < rt_rq->highest_prio.next)
644                 /*
645                  * Otherwise, we need to recompute next-highest
646                  */
647                 rt_rq->highest_prio.next = next_prio(rq);
648 }
649
650 static void
651 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
652 {
653         struct rq *rq = rq_of_rt_rq(rt_rq);
654
655         if (rt_rq->rt_nr_running && (prio <= rt_rq->highest_prio.next))
656                 rt_rq->highest_prio.next = next_prio(rq);
657
658         if (rq->online && rt_rq->highest_prio.curr != prev_prio)
659                 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
660 }
661
662 #else /* CONFIG_SMP */
663
664 static inline
665 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
666 static inline
667 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
668
669 #endif /* CONFIG_SMP */
670
671 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
672 static void
673 inc_rt_prio(struct rt_rq *rt_rq, int prio)
674 {
675         int prev_prio = rt_rq->highest_prio.curr;
676
677         if (prio < prev_prio)
678                 rt_rq->highest_prio.curr = prio;
679
680         inc_rt_prio_smp(rt_rq, prio, prev_prio);
681 }
682
683 static void
684 dec_rt_prio(struct rt_rq *rt_rq, int prio)
685 {
686         int prev_prio = rt_rq->highest_prio.curr;
687
688         if (rt_rq->rt_nr_running) {
689
690                 WARN_ON(prio < prev_prio);
691
692                 /*
693                  * This may have been our highest task, and therefore
694                  * we may have some recomputation to do
695                  */
696                 if (prio == prev_prio) {
697                         struct rt_prio_array *array = &rt_rq->active;
698
699                         rt_rq->highest_prio.curr =
700                                 sched_find_first_bit(array->bitmap);
701                 }
702
703         } else
704                 rt_rq->highest_prio.curr = MAX_RT_PRIO;
705
706         dec_rt_prio_smp(rt_rq, prio, prev_prio);
707 }
708
709 #else
710
711 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
712 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
713
714 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
715
716 #ifdef CONFIG_RT_GROUP_SCHED
717
718 static void
719 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
720 {
721         if (rt_se_boosted(rt_se))
722                 rt_rq->rt_nr_boosted++;
723
724         if (rt_rq->tg)
725                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
726 }
727
728 static void
729 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
730 {
731         if (rt_se_boosted(rt_se))
732                 rt_rq->rt_nr_boosted--;
733
734         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
735 }
736
737 #else /* CONFIG_RT_GROUP_SCHED */
738
739 static void
740 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
741 {
742         start_rt_bandwidth(&def_rt_bandwidth);
743 }
744
745 static inline
746 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
747
748 #endif /* CONFIG_RT_GROUP_SCHED */
749
750 static inline
751 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
752 {
753         int prio = rt_se_prio(rt_se);
754
755         WARN_ON(!rt_prio(prio));
756         rt_rq->rt_nr_running++;
757
758         inc_rt_prio(rt_rq, prio);
759         inc_rt_migration(rt_se, rt_rq);
760         inc_rt_group(rt_se, rt_rq);
761 }
762
763 static inline
764 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
765 {
766         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
767         WARN_ON(!rt_rq->rt_nr_running);
768         rt_rq->rt_nr_running--;
769
770         dec_rt_prio(rt_rq, rt_se_prio(rt_se));
771         dec_rt_migration(rt_se, rt_rq);
772         dec_rt_group(rt_se, rt_rq);
773 }
774
775 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se)
776 {
777         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
778         struct rt_prio_array *array = &rt_rq->active;
779         struct rt_rq *group_rq = group_rt_rq(rt_se);
780         struct list_head *queue = array->queue + rt_se_prio(rt_se);
781
782         /*
783          * Don't enqueue the group if its throttled, or when empty.
784          * The latter is a consequence of the former when a child group
785          * get throttled and the current group doesn't have any other
786          * active members.
787          */
788         if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
789                 return;
790
791         list_add_tail(&rt_se->run_list, queue);
792         __set_bit(rt_se_prio(rt_se), array->bitmap);
793
794         inc_rt_tasks(rt_se, rt_rq);
795 }
796
797 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
798 {
799         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
800         struct rt_prio_array *array = &rt_rq->active;
801
802         list_del_init(&rt_se->run_list);
803         if (list_empty(array->queue + rt_se_prio(rt_se)))
804                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
805
806         dec_rt_tasks(rt_se, rt_rq);
807 }
808
809 /*
810  * Because the prio of an upper entry depends on the lower
811  * entries, we must remove entries top - down.
812  */
813 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
814 {
815         struct sched_rt_entity *back = NULL;
816
817         for_each_sched_rt_entity(rt_se) {
818                 rt_se->back = back;
819                 back = rt_se;
820         }
821
822         for (rt_se = back; rt_se; rt_se = rt_se->back) {
823                 if (on_rt_rq(rt_se))
824                         __dequeue_rt_entity(rt_se);
825         }
826 }
827
828 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
829 {
830         dequeue_rt_stack(rt_se);
831         for_each_sched_rt_entity(rt_se)
832                 __enqueue_rt_entity(rt_se);
833 }
834
835 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
836 {
837         dequeue_rt_stack(rt_se);
838
839         for_each_sched_rt_entity(rt_se) {
840                 struct rt_rq *rt_rq = group_rt_rq(rt_se);
841
842                 if (rt_rq && rt_rq->rt_nr_running)
843                         __enqueue_rt_entity(rt_se);
844         }
845 }
846
847 /*
848  * Adding/removing a task to/from a priority array:
849  */
850 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
851 {
852         struct sched_rt_entity *rt_se = &p->rt;
853
854         if (wakeup)
855                 rt_se->timeout = 0;
856
857         enqueue_rt_entity(rt_se);
858
859         if (!task_current(rq, p) && p->rt.nr_cpus_allowed > 1)
860                 enqueue_pushable_task(rq, p);
861
862         inc_cpu_load(rq, p->se.load.weight);
863 }
864
865 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
866 {
867         struct sched_rt_entity *rt_se = &p->rt;
868
869         update_curr_rt(rq);
870         dequeue_rt_entity(rt_se);
871
872         dequeue_pushable_task(rq, p);
873
874         dec_cpu_load(rq, p->se.load.weight);
875 }
876
877 /*
878  * Put task to the end of the run list without the overhead of dequeue
879  * followed by enqueue.
880  */
881 static void
882 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
883 {
884         if (on_rt_rq(rt_se)) {
885                 struct rt_prio_array *array = &rt_rq->active;
886                 struct list_head *queue = array->queue + rt_se_prio(rt_se);
887
888                 if (head)
889                         list_move(&rt_se->run_list, queue);
890                 else
891                         list_move_tail(&rt_se->run_list, queue);
892         }
893 }
894
895 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
896 {
897         struct sched_rt_entity *rt_se = &p->rt;
898         struct rt_rq *rt_rq;
899
900         for_each_sched_rt_entity(rt_se) {
901                 rt_rq = rt_rq_of_se(rt_se);
902                 requeue_rt_entity(rt_rq, rt_se, head);
903         }
904 }
905
906 static void yield_task_rt(struct rq *rq)
907 {
908         requeue_task_rt(rq, rq->curr, 0);
909 }
910
911 #ifdef CONFIG_SMP
912 static int find_lowest_rq(struct task_struct *task);
913
914 static int select_task_rq_rt(struct task_struct *p, int sync)
915 {
916         struct rq *rq = task_rq(p);
917
918         /*
919          * If the current task is an RT task, then
920          * try to see if we can wake this RT task up on another
921          * runqueue. Otherwise simply start this RT task
922          * on its current runqueue.
923          *
924          * We want to avoid overloading runqueues. Even if
925          * the RT task is of higher priority than the current RT task.
926          * RT tasks behave differently than other tasks. If
927          * one gets preempted, we try to push it off to another queue.
928          * So trying to keep a preempting RT task on the same
929          * cache hot CPU will force the running RT task to
930          * a cold CPU. So we waste all the cache for the lower
931          * RT task in hopes of saving some of a RT task
932          * that is just being woken and probably will have
933          * cold cache anyway.
934          */
935         if (unlikely(rt_task(rq->curr)) &&
936             (p->rt.nr_cpus_allowed > 1)) {
937                 int cpu = find_lowest_rq(p);
938
939                 return (cpu == -1) ? task_cpu(p) : cpu;
940         }
941
942         /*
943          * Otherwise, just let it ride on the affined RQ and the
944          * post-schedule router will push the preempted task away
945          */
946         return task_cpu(p);
947 }
948
949 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
950 {
951         cpumask_var_t mask;
952
953         if (rq->curr->rt.nr_cpus_allowed == 1)
954                 return;
955
956         if (!alloc_cpumask_var(&mask, GFP_ATOMIC))
957                 return;
958
959         if (p->rt.nr_cpus_allowed != 1
960             && cpupri_find(&rq->rd->cpupri, p, mask))
961                 goto free;
962
963         if (!cpupri_find(&rq->rd->cpupri, rq->curr, mask))
964                 goto free;
965
966         /*
967          * There appears to be other cpus that can accept
968          * current and none to run 'p', so lets reschedule
969          * to try and push current away:
970          */
971         requeue_task_rt(rq, p, 1);
972         resched_task(rq->curr);
973 free:
974         free_cpumask_var(mask);
975 }
976
977 #endif /* CONFIG_SMP */
978
979 /*
980  * Preempt the current task with a newly woken task if needed:
981  */
982 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int sync)
983 {
984         if (p->prio < rq->curr->prio) {
985                 resched_task(rq->curr);
986                 return;
987         }
988
989 #ifdef CONFIG_SMP
990         /*
991          * If:
992          *
993          * - the newly woken task is of equal priority to the current task
994          * - the newly woken task is non-migratable while current is migratable
995          * - current will be preempted on the next reschedule
996          *
997          * we should check to see if current can readily move to a different
998          * cpu.  If so, we will reschedule to allow the push logic to try
999          * to move current somewhere else, making room for our non-migratable
1000          * task.
1001          */
1002         if (p->prio == rq->curr->prio && !need_resched())
1003                 check_preempt_equal_prio(rq, p);
1004 #endif
1005 }
1006
1007 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1008                                                    struct rt_rq *rt_rq)
1009 {
1010         struct rt_prio_array *array = &rt_rq->active;
1011         struct sched_rt_entity *next = NULL;
1012         struct list_head *queue;
1013         int idx;
1014
1015         idx = sched_find_first_bit(array->bitmap);
1016         BUG_ON(idx >= MAX_RT_PRIO);
1017
1018         queue = array->queue + idx;
1019         next = list_entry(queue->next, struct sched_rt_entity, run_list);
1020
1021         return next;
1022 }
1023
1024 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1025 {
1026         struct sched_rt_entity *rt_se;
1027         struct task_struct *p;
1028         struct rt_rq *rt_rq;
1029
1030         rt_rq = &rq->rt;
1031
1032         if (unlikely(!rt_rq->rt_nr_running))
1033                 return NULL;
1034
1035         if (rt_rq_throttled(rt_rq))
1036                 return NULL;
1037
1038         do {
1039                 rt_se = pick_next_rt_entity(rq, rt_rq);
1040                 BUG_ON(!rt_se);
1041                 rt_rq = group_rt_rq(rt_se);
1042         } while (rt_rq);
1043
1044         p = rt_task_of(rt_se);
1045         p->se.exec_start = rq->clock;
1046
1047         return p;
1048 }
1049
1050 static struct task_struct *pick_next_task_rt(struct rq *rq)
1051 {
1052         struct task_struct *p = _pick_next_task_rt(rq);
1053
1054         /* The running task is never eligible for pushing */
1055         if (p)
1056                 dequeue_pushable_task(rq, p);
1057
1058         return p;
1059 }
1060
1061 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1062 {
1063         update_curr_rt(rq);
1064         p->se.exec_start = 0;
1065
1066         /*
1067          * The previous task needs to be made eligible for pushing
1068          * if it is still active
1069          */
1070         if (p->se.on_rq && p->rt.nr_cpus_allowed > 1)
1071                 enqueue_pushable_task(rq, p);
1072 }
1073
1074 #ifdef CONFIG_SMP
1075
1076 /* Only try algorithms three times */
1077 #define RT_MAX_TRIES 3
1078
1079 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
1080
1081 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1082 {
1083         if (!task_running(rq, p) &&
1084             (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
1085             (p->rt.nr_cpus_allowed > 1))
1086                 return 1;
1087         return 0;
1088 }
1089
1090 /* Return the second highest RT task, NULL otherwise */
1091 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
1092 {
1093         struct task_struct *next = NULL;
1094         struct sched_rt_entity *rt_se;
1095         struct rt_prio_array *array;
1096         struct rt_rq *rt_rq;
1097         int idx;
1098
1099         for_each_leaf_rt_rq(rt_rq, rq) {
1100                 array = &rt_rq->active;
1101                 idx = sched_find_first_bit(array->bitmap);
1102  next_idx:
1103                 if (idx >= MAX_RT_PRIO)
1104                         continue;
1105                 if (next && next->prio < idx)
1106                         continue;
1107                 list_for_each_entry(rt_se, array->queue + idx, run_list) {
1108                         struct task_struct *p = rt_task_of(rt_se);
1109                         if (pick_rt_task(rq, p, cpu)) {
1110                                 next = p;
1111                                 break;
1112                         }
1113                 }
1114                 if (!next) {
1115                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
1116                         goto next_idx;
1117                 }
1118         }
1119
1120         return next;
1121 }
1122
1123 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1124
1125 static inline int pick_optimal_cpu(int this_cpu,
1126                                    const struct cpumask *mask)
1127 {
1128         int first;
1129
1130         /* "this_cpu" is cheaper to preempt than a remote processor */
1131         if ((this_cpu != -1) && cpumask_test_cpu(this_cpu, mask))
1132                 return this_cpu;
1133
1134         first = cpumask_first(mask);
1135         if (first < nr_cpu_ids)
1136                 return first;
1137
1138         return -1;
1139 }
1140
1141 static int find_lowest_rq(struct task_struct *task)
1142 {
1143         struct sched_domain *sd;
1144         struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1145         int this_cpu = smp_processor_id();
1146         int cpu      = task_cpu(task);
1147         cpumask_var_t domain_mask;
1148
1149         if (task->rt.nr_cpus_allowed == 1)
1150                 return -1; /* No other targets possible */
1151
1152         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1153                 return -1; /* No targets found */
1154
1155         /*
1156          * Only consider CPUs that are usable for migration.
1157          * I guess we might want to change cpupri_find() to ignore those
1158          * in the first place.
1159          */
1160         cpumask_and(lowest_mask, lowest_mask, cpu_active_mask);
1161
1162         /*
1163          * At this point we have built a mask of cpus representing the
1164          * lowest priority tasks in the system.  Now we want to elect
1165          * the best one based on our affinity and topology.
1166          *
1167          * We prioritize the last cpu that the task executed on since
1168          * it is most likely cache-hot in that location.
1169          */
1170         if (cpumask_test_cpu(cpu, lowest_mask))
1171                 return cpu;
1172
1173         /*
1174          * Otherwise, we consult the sched_domains span maps to figure
1175          * out which cpu is logically closest to our hot cache data.
1176          */
1177         if (this_cpu == cpu)
1178                 this_cpu = -1; /* Skip this_cpu opt if the same */
1179
1180         if (alloc_cpumask_var(&domain_mask, GFP_ATOMIC)) {
1181                 for_each_domain(cpu, sd) {
1182                         if (sd->flags & SD_WAKE_AFFINE) {
1183                                 int best_cpu;
1184
1185                                 cpumask_and(domain_mask,
1186                                             sched_domain_span(sd),
1187                                             lowest_mask);
1188
1189                                 best_cpu = pick_optimal_cpu(this_cpu,
1190                                                             domain_mask);
1191
1192                                 if (best_cpu != -1) {
1193                                         free_cpumask_var(domain_mask);
1194                                         return best_cpu;
1195                                 }
1196                         }
1197                 }
1198                 free_cpumask_var(domain_mask);
1199         }
1200
1201         /*
1202          * And finally, if there were no matches within the domains
1203          * just give the caller *something* to work with from the compatible
1204          * locations.
1205          */
1206         return pick_optimal_cpu(this_cpu, lowest_mask);
1207 }
1208
1209 /* Will lock the rq it finds */
1210 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1211 {
1212         struct rq *lowest_rq = NULL;
1213         int tries;
1214         int cpu;
1215
1216         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1217                 cpu = find_lowest_rq(task);
1218
1219                 if ((cpu == -1) || (cpu == rq->cpu))
1220                         break;
1221
1222                 lowest_rq = cpu_rq(cpu);
1223
1224                 /* if the prio of this runqueue changed, try again */
1225                 if (double_lock_balance(rq, lowest_rq)) {
1226                         /*
1227                          * We had to unlock the run queue. In
1228                          * the mean time, task could have
1229                          * migrated already or had its affinity changed.
1230                          * Also make sure that it wasn't scheduled on its rq.
1231                          */
1232                         if (unlikely(task_rq(task) != rq ||
1233                                      !cpumask_test_cpu(lowest_rq->cpu,
1234                                                        &task->cpus_allowed) ||
1235                                      task_running(rq, task) ||
1236                                      !task->se.on_rq)) {
1237
1238                                 spin_unlock(&lowest_rq->lock);
1239                                 lowest_rq = NULL;
1240                                 break;
1241                         }
1242                 }
1243
1244                 /* If this rq is still suitable use it. */
1245                 if (lowest_rq->rt.highest_prio.curr > task->prio)
1246                         break;
1247
1248                 /* try again */
1249                 double_unlock_balance(rq, lowest_rq);
1250                 lowest_rq = NULL;
1251         }
1252
1253         return lowest_rq;
1254 }
1255
1256 static inline int has_pushable_tasks(struct rq *rq)
1257 {
1258         return !plist_head_empty(&rq->rt.pushable_tasks);
1259 }
1260
1261 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1262 {
1263         struct task_struct *p;
1264
1265         if (!has_pushable_tasks(rq))
1266                 return NULL;
1267
1268         p = plist_first_entry(&rq->rt.pushable_tasks,
1269                               struct task_struct, pushable_tasks);
1270
1271         BUG_ON(rq->cpu != task_cpu(p));
1272         BUG_ON(task_current(rq, p));
1273         BUG_ON(p->rt.nr_cpus_allowed <= 1);
1274
1275         BUG_ON(!p->se.on_rq);
1276         BUG_ON(!rt_task(p));
1277
1278         return p;
1279 }
1280
1281 /*
1282  * If the current CPU has more than one RT task, see if the non
1283  * running task can migrate over to a CPU that is running a task
1284  * of lesser priority.
1285  */
1286 static int push_rt_task(struct rq *rq)
1287 {
1288         struct task_struct *next_task;
1289         struct rq *lowest_rq;
1290
1291         if (!rq->rt.overloaded)
1292                 return 0;
1293
1294         next_task = pick_next_pushable_task(rq);
1295         if (!next_task)
1296                 return 0;
1297
1298  retry:
1299         if (unlikely(next_task == rq->curr)) {
1300                 WARN_ON(1);
1301                 return 0;
1302         }
1303
1304         /*
1305          * It's possible that the next_task slipped in of
1306          * higher priority than current. If that's the case
1307          * just reschedule current.
1308          */
1309         if (unlikely(next_task->prio < rq->curr->prio)) {
1310                 resched_task(rq->curr);
1311                 return 0;
1312         }
1313
1314         /* We might release rq lock */
1315         get_task_struct(next_task);
1316
1317         /* find_lock_lowest_rq locks the rq if found */
1318         lowest_rq = find_lock_lowest_rq(next_task, rq);
1319         if (!lowest_rq) {
1320                 struct task_struct *task;
1321                 /*
1322                  * find lock_lowest_rq releases rq->lock
1323                  * so it is possible that next_task has migrated.
1324                  *
1325                  * We need to make sure that the task is still on the same
1326                  * run-queue and is also still the next task eligible for
1327                  * pushing.
1328                  */
1329                 task = pick_next_pushable_task(rq);
1330                 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1331                         /*
1332                          * If we get here, the task hasnt moved at all, but
1333                          * it has failed to push.  We will not try again,
1334                          * since the other cpus will pull from us when they
1335                          * are ready.
1336                          */
1337                         dequeue_pushable_task(rq, next_task);
1338                         goto out;
1339                 }
1340
1341                 if (!task)
1342                         /* No more tasks, just exit */
1343                         goto out;
1344
1345                 /*
1346                  * Something has shifted, try again.
1347                  */
1348                 put_task_struct(next_task);
1349                 next_task = task;
1350                 goto retry;
1351         }
1352
1353         deactivate_task(rq, next_task, 0);
1354         set_task_cpu(next_task, lowest_rq->cpu);
1355         activate_task(lowest_rq, next_task, 0);
1356
1357         resched_task(lowest_rq->curr);
1358
1359         double_unlock_balance(rq, lowest_rq);
1360
1361 out:
1362         put_task_struct(next_task);
1363
1364         return 1;
1365 }
1366
1367 static void push_rt_tasks(struct rq *rq)
1368 {
1369         /* push_rt_task will return true if it moved an RT */
1370         while (push_rt_task(rq))
1371                 ;
1372 }
1373
1374 static int pull_rt_task(struct rq *this_rq)
1375 {
1376         int this_cpu = this_rq->cpu, ret = 0, cpu;
1377         struct task_struct *p;
1378         struct rq *src_rq;
1379
1380         if (likely(!rt_overloaded(this_rq)))
1381                 return 0;
1382
1383         for_each_cpu(cpu, this_rq->rd->rto_mask) {
1384                 if (this_cpu == cpu)
1385                         continue;
1386
1387                 src_rq = cpu_rq(cpu);
1388
1389                 /*
1390                  * Don't bother taking the src_rq->lock if the next highest
1391                  * task is known to be lower-priority than our current task.
1392                  * This may look racy, but if this value is about to go
1393                  * logically higher, the src_rq will push this task away.
1394                  * And if its going logically lower, we do not care
1395                  */
1396                 if (src_rq->rt.highest_prio.next >=
1397                     this_rq->rt.highest_prio.curr)
1398                         continue;
1399
1400                 /*
1401                  * We can potentially drop this_rq's lock in
1402                  * double_lock_balance, and another CPU could
1403                  * alter this_rq
1404                  */
1405                 double_lock_balance(this_rq, src_rq);
1406
1407                 /*
1408                  * Are there still pullable RT tasks?
1409                  */
1410                 if (src_rq->rt.rt_nr_running <= 1)
1411                         goto skip;
1412
1413                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1414
1415                 /*
1416                  * Do we have an RT task that preempts
1417                  * the to-be-scheduled task?
1418                  */
1419                 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1420                         WARN_ON(p == src_rq->curr);
1421                         WARN_ON(!p->se.on_rq);
1422
1423                         /*
1424                          * There's a chance that p is higher in priority
1425                          * than what's currently running on its cpu.
1426                          * This is just that p is wakeing up and hasn't
1427                          * had a chance to schedule. We only pull
1428                          * p if it is lower in priority than the
1429                          * current task on the run queue
1430                          */
1431                         if (p->prio < src_rq->curr->prio)
1432                                 goto skip;
1433
1434                         ret = 1;
1435
1436                         deactivate_task(src_rq, p, 0);
1437                         set_task_cpu(p, this_cpu);
1438                         activate_task(this_rq, p, 0);
1439                         /*
1440                          * We continue with the search, just in
1441                          * case there's an even higher prio task
1442                          * in another runqueue. (low likelyhood
1443                          * but possible)
1444                          */
1445                 }
1446  skip:
1447                 double_unlock_balance(this_rq, src_rq);
1448         }
1449
1450         return ret;
1451 }
1452
1453 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1454 {
1455         /* Try to pull RT tasks here if we lower this rq's prio */
1456         if (unlikely(rt_task(prev)) && rq->rt.highest_prio.curr > prev->prio)
1457                 pull_rt_task(rq);
1458 }
1459
1460 /*
1461  * assumes rq->lock is held
1462  */
1463 static int needs_post_schedule_rt(struct rq *rq)
1464 {
1465         return has_pushable_tasks(rq);
1466 }
1467
1468 static void post_schedule_rt(struct rq *rq)
1469 {
1470         /*
1471          * This is only called if needs_post_schedule_rt() indicates that
1472          * we need to push tasks away
1473          */
1474         spin_lock_irq(&rq->lock);
1475         push_rt_tasks(rq);
1476         spin_unlock_irq(&rq->lock);
1477 }
1478
1479 /*
1480  * If we are not running and we are not going to reschedule soon, we should
1481  * try to push tasks away now
1482  */
1483 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1484 {
1485         if (!task_running(rq, p) &&
1486             !test_tsk_need_resched(rq->curr) &&
1487             has_pushable_tasks(rq) &&
1488             p->rt.nr_cpus_allowed > 1)
1489                 push_rt_tasks(rq);
1490 }
1491
1492 static unsigned long
1493 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1494                 unsigned long max_load_move,
1495                 struct sched_domain *sd, enum cpu_idle_type idle,
1496                 int *all_pinned, int *this_best_prio)
1497 {
1498         /* don't touch RT tasks */
1499         return 0;
1500 }
1501
1502 static int
1503 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1504                  struct sched_domain *sd, enum cpu_idle_type idle)
1505 {
1506         /* don't touch RT tasks */
1507         return 0;
1508 }
1509
1510 static void set_cpus_allowed_rt(struct task_struct *p,
1511                                 const struct cpumask *new_mask)
1512 {
1513         int weight = cpumask_weight(new_mask);
1514
1515         BUG_ON(!rt_task(p));
1516
1517         /*
1518          * Update the migration status of the RQ if we have an RT task
1519          * which is running AND changing its weight value.
1520          */
1521         if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1522                 struct rq *rq = task_rq(p);
1523
1524                 if (!task_current(rq, p)) {
1525                         /*
1526                          * Make sure we dequeue this task from the pushable list
1527                          * before going further.  It will either remain off of
1528                          * the list because we are no longer pushable, or it
1529                          * will be requeued.
1530                          */
1531                         if (p->rt.nr_cpus_allowed > 1)
1532                                 dequeue_pushable_task(rq, p);
1533
1534                         /*
1535                          * Requeue if our weight is changing and still > 1
1536                          */
1537                         if (weight > 1)
1538                                 enqueue_pushable_task(rq, p);
1539
1540                 }
1541
1542                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1543                         rq->rt.rt_nr_migratory++;
1544                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1545                         BUG_ON(!rq->rt.rt_nr_migratory);
1546                         rq->rt.rt_nr_migratory--;
1547                 }
1548
1549                 update_rt_migration(&rq->rt);
1550         }
1551
1552         cpumask_copy(&p->cpus_allowed, new_mask);
1553         p->rt.nr_cpus_allowed = weight;
1554 }
1555
1556 /* Assumes rq->lock is held */
1557 static void rq_online_rt(struct rq *rq)
1558 {
1559         if (rq->rt.overloaded)
1560                 rt_set_overload(rq);
1561
1562         __enable_runtime(rq);
1563
1564         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1565 }
1566
1567 /* Assumes rq->lock is held */
1568 static void rq_offline_rt(struct rq *rq)
1569 {
1570         if (rq->rt.overloaded)
1571                 rt_clear_overload(rq);
1572
1573         __disable_runtime(rq);
1574
1575         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1576 }
1577
1578 /*
1579  * When switch from the rt queue, we bring ourselves to a position
1580  * that we might want to pull RT tasks from other runqueues.
1581  */
1582 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1583                            int running)
1584 {
1585         /*
1586          * If there are other RT tasks then we will reschedule
1587          * and the scheduling of the other RT tasks will handle
1588          * the balancing. But if we are the last RT task
1589          * we may need to handle the pulling of RT tasks
1590          * now.
1591          */
1592         if (!rq->rt.rt_nr_running)
1593                 pull_rt_task(rq);
1594 }
1595
1596 static inline void init_sched_rt_class(void)
1597 {
1598         unsigned int i;
1599
1600         for_each_possible_cpu(i)
1601                 alloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1602                                         GFP_KERNEL, cpu_to_node(i));
1603 }
1604 #endif /* CONFIG_SMP */
1605
1606 /*
1607  * When switching a task to RT, we may overload the runqueue
1608  * with RT tasks. In this case we try to push them off to
1609  * other runqueues.
1610  */
1611 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1612                            int running)
1613 {
1614         int check_resched = 1;
1615
1616         /*
1617          * If we are already running, then there's nothing
1618          * that needs to be done. But if we are not running
1619          * we may need to preempt the current running task.
1620          * If that current running task is also an RT task
1621          * then see if we can move to another run queue.
1622          */
1623         if (!running) {
1624 #ifdef CONFIG_SMP
1625                 if (rq->rt.overloaded && push_rt_task(rq) &&
1626                     /* Don't resched if we changed runqueues */
1627                     rq != task_rq(p))
1628                         check_resched = 0;
1629 #endif /* CONFIG_SMP */
1630                 if (check_resched && p->prio < rq->curr->prio)
1631                         resched_task(rq->curr);
1632         }
1633 }
1634
1635 /*
1636  * Priority of the task has changed. This may cause
1637  * us to initiate a push or pull.
1638  */
1639 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1640                             int oldprio, int running)
1641 {
1642         if (running) {
1643 #ifdef CONFIG_SMP
1644                 /*
1645                  * If our priority decreases while running, we
1646                  * may need to pull tasks to this runqueue.
1647                  */
1648                 if (oldprio < p->prio)
1649                         pull_rt_task(rq);
1650                 /*
1651                  * If there's a higher priority task waiting to run
1652                  * then reschedule. Note, the above pull_rt_task
1653                  * can release the rq lock and p could migrate.
1654                  * Only reschedule if p is still on the same runqueue.
1655                  */
1656                 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1657                         resched_task(p);
1658 #else
1659                 /* For UP simply resched on drop of prio */
1660                 if (oldprio < p->prio)
1661                         resched_task(p);
1662 #endif /* CONFIG_SMP */
1663         } else {
1664                 /*
1665                  * This task is not running, but if it is
1666                  * greater than the current running task
1667                  * then reschedule.
1668                  */
1669                 if (p->prio < rq->curr->prio)
1670                         resched_task(rq->curr);
1671         }
1672 }
1673
1674 static void watchdog(struct rq *rq, struct task_struct *p)
1675 {
1676         unsigned long soft, hard;
1677
1678         if (!p->signal)
1679                 return;
1680
1681         soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1682         hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1683
1684         if (soft != RLIM_INFINITY) {
1685                 unsigned long next;
1686
1687                 p->rt.timeout++;
1688                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1689                 if (p->rt.timeout > next)
1690                         p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1691         }
1692 }
1693
1694 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1695 {
1696         update_curr_rt(rq);
1697
1698         watchdog(rq, p);
1699
1700         /*
1701          * RR tasks need a special form of timeslice management.
1702          * FIFO tasks have no timeslices.
1703          */
1704         if (p->policy != SCHED_RR)
1705                 return;
1706
1707         if (--p->rt.time_slice)
1708                 return;
1709
1710         p->rt.time_slice = DEF_TIMESLICE;
1711
1712         /*
1713          * Requeue to the end of queue if we are not the only element
1714          * on the queue:
1715          */
1716         if (p->rt.run_list.prev != p->rt.run_list.next) {
1717                 requeue_task_rt(rq, p, 0);
1718                 set_tsk_need_resched(p);
1719         }
1720 }
1721
1722 static void set_curr_task_rt(struct rq *rq)
1723 {
1724         struct task_struct *p = rq->curr;
1725
1726         p->se.exec_start = rq->clock;
1727
1728         /* The running task is never eligible for pushing */
1729         dequeue_pushable_task(rq, p);
1730 }
1731
1732 static const struct sched_class rt_sched_class = {
1733         .next                   = &fair_sched_class,
1734         .enqueue_task           = enqueue_task_rt,
1735         .dequeue_task           = dequeue_task_rt,
1736         .yield_task             = yield_task_rt,
1737
1738         .check_preempt_curr     = check_preempt_curr_rt,
1739
1740         .pick_next_task         = pick_next_task_rt,
1741         .put_prev_task          = put_prev_task_rt,
1742
1743 #ifdef CONFIG_SMP
1744         .select_task_rq         = select_task_rq_rt,
1745
1746         .load_balance           = load_balance_rt,
1747         .move_one_task          = move_one_task_rt,
1748         .set_cpus_allowed       = set_cpus_allowed_rt,
1749         .rq_online              = rq_online_rt,
1750         .rq_offline             = rq_offline_rt,
1751         .pre_schedule           = pre_schedule_rt,
1752         .needs_post_schedule    = needs_post_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