sched: fix cpupri hotplug support
[linux-2.6.git] / kernel / sched_rt.c
1 /*
2  * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
3  * policies)
4  */
5
6 #ifdef CONFIG_SMP
7
8 static inline int rt_overloaded(struct rq *rq)
9 {
10         return atomic_read(&rq->rd->rto_count);
11 }
12
13 static inline void rt_set_overload(struct rq *rq)
14 {
15         if (!rq->online)
16                 return;
17
18         cpu_set(rq->cpu, rq->rd->rto_mask);
19         /*
20          * Make sure the mask is visible before we set
21          * the overload count. That is checked to determine
22          * if we should look at the mask. It would be a shame
23          * if we looked at the mask, but the mask was not
24          * updated yet.
25          */
26         wmb();
27         atomic_inc(&rq->rd->rto_count);
28 }
29
30 static inline void rt_clear_overload(struct rq *rq)
31 {
32         if (!rq->online)
33                 return;
34
35         /* the order here really doesn't matter */
36         atomic_dec(&rq->rd->rto_count);
37         cpu_clear(rq->cpu, rq->rd->rto_mask);
38 }
39
40 static void update_rt_migration(struct rq *rq)
41 {
42         if (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1)) {
43                 if (!rq->rt.overloaded) {
44                         rt_set_overload(rq);
45                         rq->rt.overloaded = 1;
46                 }
47         } else if (rq->rt.overloaded) {
48                 rt_clear_overload(rq);
49                 rq->rt.overloaded = 0;
50         }
51 }
52 #endif /* CONFIG_SMP */
53
54 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
55 {
56         return container_of(rt_se, struct task_struct, rt);
57 }
58
59 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
60 {
61         return !list_empty(&rt_se->run_list);
62 }
63
64 #ifdef CONFIG_RT_GROUP_SCHED
65
66 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
67 {
68         if (!rt_rq->tg)
69                 return RUNTIME_INF;
70
71         return rt_rq->rt_runtime;
72 }
73
74 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
75 {
76         return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
77 }
78
79 #define for_each_leaf_rt_rq(rt_rq, rq) \
80         list_for_each_entry(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
81
82 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
83 {
84         return rt_rq->rq;
85 }
86
87 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
88 {
89         return rt_se->rt_rq;
90 }
91
92 #define for_each_sched_rt_entity(rt_se) \
93         for (; rt_se; rt_se = rt_se->parent)
94
95 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
96 {
97         return rt_se->my_q;
98 }
99
100 static void enqueue_rt_entity(struct sched_rt_entity *rt_se);
101 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
102
103 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
104 {
105         struct sched_rt_entity *rt_se = rt_rq->rt_se;
106
107         if (rt_se && !on_rt_rq(rt_se) && rt_rq->rt_nr_running) {
108                 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
109
110                 enqueue_rt_entity(rt_se);
111                 if (rt_rq->highest_prio < curr->prio)
112                         resched_task(curr);
113         }
114 }
115
116 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
117 {
118         struct sched_rt_entity *rt_se = rt_rq->rt_se;
119
120         if (rt_se && on_rt_rq(rt_se))
121                 dequeue_rt_entity(rt_se);
122 }
123
124 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
125 {
126         return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
127 }
128
129 static int rt_se_boosted(struct sched_rt_entity *rt_se)
130 {
131         struct rt_rq *rt_rq = group_rt_rq(rt_se);
132         struct task_struct *p;
133
134         if (rt_rq)
135                 return !!rt_rq->rt_nr_boosted;
136
137         p = rt_task_of(rt_se);
138         return p->prio != p->normal_prio;
139 }
140
141 #ifdef CONFIG_SMP
142 static inline cpumask_t sched_rt_period_mask(void)
143 {
144         return cpu_rq(smp_processor_id())->rd->span;
145 }
146 #else
147 static inline cpumask_t sched_rt_period_mask(void)
148 {
149         return cpu_online_map;
150 }
151 #endif
152
153 static inline
154 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
155 {
156         return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
157 }
158
159 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
160 {
161         return &rt_rq->tg->rt_bandwidth;
162 }
163
164 #else
165
166 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
167 {
168         return rt_rq->rt_runtime;
169 }
170
171 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
172 {
173         return ktime_to_ns(def_rt_bandwidth.rt_period);
174 }
175
176 #define for_each_leaf_rt_rq(rt_rq, rq) \
177         for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
178
179 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
180 {
181         return container_of(rt_rq, struct rq, rt);
182 }
183
184 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
185 {
186         struct task_struct *p = rt_task_of(rt_se);
187         struct rq *rq = task_rq(p);
188
189         return &rq->rt;
190 }
191
192 #define for_each_sched_rt_entity(rt_se) \
193         for (; rt_se; rt_se = NULL)
194
195 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
196 {
197         return NULL;
198 }
199
200 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
201 {
202 }
203
204 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
205 {
206 }
207
208 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
209 {
210         return rt_rq->rt_throttled;
211 }
212
213 static inline cpumask_t sched_rt_period_mask(void)
214 {
215         return cpu_online_map;
216 }
217
218 static inline
219 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
220 {
221         return &cpu_rq(cpu)->rt;
222 }
223
224 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
225 {
226         return &def_rt_bandwidth;
227 }
228
229 #endif
230
231 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
232 {
233         int i, idle = 1;
234         cpumask_t span;
235
236         if (rt_b->rt_runtime == RUNTIME_INF)
237                 return 1;
238
239         span = sched_rt_period_mask();
240         for_each_cpu_mask(i, span) {
241                 int enqueue = 0;
242                 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
243                 struct rq *rq = rq_of_rt_rq(rt_rq);
244
245                 spin_lock(&rq->lock);
246                 if (rt_rq->rt_time) {
247                         u64 runtime;
248
249                         spin_lock(&rt_rq->rt_runtime_lock);
250                         runtime = rt_rq->rt_runtime;
251                         rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
252                         if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
253                                 rt_rq->rt_throttled = 0;
254                                 enqueue = 1;
255                         }
256                         if (rt_rq->rt_time || rt_rq->rt_nr_running)
257                                 idle = 0;
258                         spin_unlock(&rt_rq->rt_runtime_lock);
259                 }
260
261                 if (enqueue)
262                         sched_rt_rq_enqueue(rt_rq);
263                 spin_unlock(&rq->lock);
264         }
265
266         return idle;
267 }
268
269 #ifdef CONFIG_SMP
270 static int balance_runtime(struct rt_rq *rt_rq)
271 {
272         struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
273         struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
274         int i, weight, more = 0;
275         u64 rt_period;
276
277         weight = cpus_weight(rd->span);
278
279         spin_lock(&rt_b->rt_runtime_lock);
280         rt_period = ktime_to_ns(rt_b->rt_period);
281         for_each_cpu_mask(i, rd->span) {
282                 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
283                 s64 diff;
284
285                 if (iter == rt_rq)
286                         continue;
287
288                 spin_lock(&iter->rt_runtime_lock);
289                 diff = iter->rt_runtime - iter->rt_time;
290                 if (diff > 0) {
291                         do_div(diff, weight);
292                         if (rt_rq->rt_runtime + diff > rt_period)
293                                 diff = rt_period - rt_rq->rt_runtime;
294                         iter->rt_runtime -= diff;
295                         rt_rq->rt_runtime += diff;
296                         more = 1;
297                         if (rt_rq->rt_runtime == rt_period) {
298                                 spin_unlock(&iter->rt_runtime_lock);
299                                 break;
300                         }
301                 }
302                 spin_unlock(&iter->rt_runtime_lock);
303         }
304         spin_unlock(&rt_b->rt_runtime_lock);
305
306         return more;
307 }
308 #endif
309
310 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
311 {
312 #ifdef CONFIG_RT_GROUP_SCHED
313         struct rt_rq *rt_rq = group_rt_rq(rt_se);
314
315         if (rt_rq)
316                 return rt_rq->highest_prio;
317 #endif
318
319         return rt_task_of(rt_se)->prio;
320 }
321
322 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
323 {
324         u64 runtime = sched_rt_runtime(rt_rq);
325
326         if (runtime == RUNTIME_INF)
327                 return 0;
328
329         if (rt_rq->rt_throttled)
330                 return rt_rq_throttled(rt_rq);
331
332         if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
333                 return 0;
334
335 #ifdef CONFIG_SMP
336         if (rt_rq->rt_time > runtime) {
337                 int more;
338
339                 spin_unlock(&rt_rq->rt_runtime_lock);
340                 more = balance_runtime(rt_rq);
341                 spin_lock(&rt_rq->rt_runtime_lock);
342
343                 if (more)
344                         runtime = sched_rt_runtime(rt_rq);
345         }
346 #endif
347
348         if (rt_rq->rt_time > runtime) {
349                 rt_rq->rt_throttled = 1;
350                 if (rt_rq_throttled(rt_rq)) {
351                         sched_rt_rq_dequeue(rt_rq);
352                         return 1;
353                 }
354         }
355
356         return 0;
357 }
358
359 /*
360  * Update the current task's runtime statistics. Skip current tasks that
361  * are not in our scheduling class.
362  */
363 static void update_curr_rt(struct rq *rq)
364 {
365         struct task_struct *curr = rq->curr;
366         struct sched_rt_entity *rt_se = &curr->rt;
367         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
368         u64 delta_exec;
369
370         if (!task_has_rt_policy(curr))
371                 return;
372
373         delta_exec = rq->clock - curr->se.exec_start;
374         if (unlikely((s64)delta_exec < 0))
375                 delta_exec = 0;
376
377         schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
378
379         curr->se.sum_exec_runtime += delta_exec;
380         curr->se.exec_start = rq->clock;
381         cpuacct_charge(curr, delta_exec);
382
383         for_each_sched_rt_entity(rt_se) {
384                 rt_rq = rt_rq_of_se(rt_se);
385
386                 spin_lock(&rt_rq->rt_runtime_lock);
387                 rt_rq->rt_time += delta_exec;
388                 if (sched_rt_runtime_exceeded(rt_rq))
389                         resched_task(curr);
390                 spin_unlock(&rt_rq->rt_runtime_lock);
391         }
392 }
393
394 static inline
395 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
396 {
397         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
398         rt_rq->rt_nr_running++;
399 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
400         if (rt_se_prio(rt_se) < rt_rq->highest_prio) {
401                 struct rq *rq = rq_of_rt_rq(rt_rq);
402                 rt_rq->highest_prio = rt_se_prio(rt_se);
403
404                 if (rq->online)
405                         cpupri_set(&rq->rd->cpupri, rq->cpu,
406                                    rt_se_prio(rt_se));
407         }
408 #endif
409 #ifdef CONFIG_SMP
410         if (rt_se->nr_cpus_allowed > 1) {
411                 struct rq *rq = rq_of_rt_rq(rt_rq);
412                 rq->rt.rt_nr_migratory++;
413         }
414
415         update_rt_migration(rq_of_rt_rq(rt_rq));
416 #endif
417 #ifdef CONFIG_RT_GROUP_SCHED
418         if (rt_se_boosted(rt_se))
419                 rt_rq->rt_nr_boosted++;
420
421         if (rt_rq->tg)
422                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
423 #else
424         start_rt_bandwidth(&def_rt_bandwidth);
425 #endif
426 }
427
428 static inline
429 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
430 {
431 #ifdef CONFIG_SMP
432         int highest_prio = rt_rq->highest_prio;
433 #endif
434
435         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
436         WARN_ON(!rt_rq->rt_nr_running);
437         rt_rq->rt_nr_running--;
438 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
439         if (rt_rq->rt_nr_running) {
440                 struct rt_prio_array *array;
441
442                 WARN_ON(rt_se_prio(rt_se) < rt_rq->highest_prio);
443                 if (rt_se_prio(rt_se) == rt_rq->highest_prio) {
444                         /* recalculate */
445                         array = &rt_rq->active;
446                         rt_rq->highest_prio =
447                                 sched_find_first_bit(array->bitmap);
448                 } /* otherwise leave rq->highest prio alone */
449         } else
450                 rt_rq->highest_prio = MAX_RT_PRIO;
451 #endif
452 #ifdef CONFIG_SMP
453         if (rt_se->nr_cpus_allowed > 1) {
454                 struct rq *rq = rq_of_rt_rq(rt_rq);
455                 rq->rt.rt_nr_migratory--;
456         }
457
458         if (rt_rq->highest_prio != highest_prio) {
459                 struct rq *rq = rq_of_rt_rq(rt_rq);
460
461                 if (rq->online)
462                         cpupri_set(&rq->rd->cpupri, rq->cpu,
463                                    rt_rq->highest_prio);
464         }
465
466         update_rt_migration(rq_of_rt_rq(rt_rq));
467 #endif /* CONFIG_SMP */
468 #ifdef CONFIG_RT_GROUP_SCHED
469         if (rt_se_boosted(rt_se))
470                 rt_rq->rt_nr_boosted--;
471
472         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
473 #endif
474 }
475
476 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
477 {
478         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
479         struct rt_prio_array *array = &rt_rq->active;
480         struct rt_rq *group_rq = group_rt_rq(rt_se);
481
482         if (group_rq && rt_rq_throttled(group_rq))
483                 return;
484
485         if (rt_se->nr_cpus_allowed == 1)
486                 list_add_tail(&rt_se->run_list,
487                               array->xqueue + rt_se_prio(rt_se));
488         else
489                 list_add_tail(&rt_se->run_list,
490                               array->squeue + rt_se_prio(rt_se));
491
492         __set_bit(rt_se_prio(rt_se), array->bitmap);
493
494         inc_rt_tasks(rt_se, rt_rq);
495 }
496
497 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
498 {
499         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
500         struct rt_prio_array *array = &rt_rq->active;
501
502         list_del_init(&rt_se->run_list);
503         if (list_empty(array->squeue + rt_se_prio(rt_se))
504             && list_empty(array->xqueue + rt_se_prio(rt_se)))
505                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
506
507         dec_rt_tasks(rt_se, rt_rq);
508 }
509
510 /*
511  * Because the prio of an upper entry depends on the lower
512  * entries, we must remove entries top - down.
513  */
514 static void dequeue_rt_stack(struct task_struct *p)
515 {
516         struct sched_rt_entity *rt_se, *back = NULL;
517
518         rt_se = &p->rt;
519         for_each_sched_rt_entity(rt_se) {
520                 rt_se->back = back;
521                 back = rt_se;
522         }
523
524         for (rt_se = back; rt_se; rt_se = rt_se->back) {
525                 if (on_rt_rq(rt_se))
526                         dequeue_rt_entity(rt_se);
527         }
528 }
529
530 /*
531  * Adding/removing a task to/from a priority array:
532  */
533 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
534 {
535         struct sched_rt_entity *rt_se = &p->rt;
536
537         if (wakeup)
538                 rt_se->timeout = 0;
539
540         dequeue_rt_stack(p);
541
542         /*
543          * enqueue everybody, bottom - up.
544          */
545         for_each_sched_rt_entity(rt_se)
546                 enqueue_rt_entity(rt_se);
547 }
548
549 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
550 {
551         struct sched_rt_entity *rt_se = &p->rt;
552         struct rt_rq *rt_rq;
553
554         update_curr_rt(rq);
555
556         dequeue_rt_stack(p);
557
558         /*
559          * re-enqueue all non-empty rt_rq entities.
560          */
561         for_each_sched_rt_entity(rt_se) {
562                 rt_rq = group_rt_rq(rt_se);
563                 if (rt_rq && rt_rq->rt_nr_running)
564                         enqueue_rt_entity(rt_se);
565         }
566 }
567
568 /*
569  * Put task to the end of the run list without the overhead of dequeue
570  * followed by enqueue.
571  *
572  * Note: We always enqueue the task to the shared-queue, regardless of its
573  * previous position w.r.t. exclusive vs shared.  This is so that exclusive RR
574  * tasks fairly round-robin with all tasks on the runqueue, not just other
575  * exclusive tasks.
576  */
577 static
578 void requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
579 {
580         struct rt_prio_array *array = &rt_rq->active;
581
582         list_del_init(&rt_se->run_list);
583         list_add_tail(&rt_se->run_list, array->squeue + rt_se_prio(rt_se));
584 }
585
586 static void requeue_task_rt(struct rq *rq, struct task_struct *p)
587 {
588         struct sched_rt_entity *rt_se = &p->rt;
589         struct rt_rq *rt_rq;
590
591         for_each_sched_rt_entity(rt_se) {
592                 rt_rq = rt_rq_of_se(rt_se);
593                 requeue_rt_entity(rt_rq, rt_se);
594         }
595 }
596
597 static void yield_task_rt(struct rq *rq)
598 {
599         requeue_task_rt(rq, rq->curr);
600 }
601
602 #ifdef CONFIG_SMP
603 static int find_lowest_rq(struct task_struct *task);
604
605 static int select_task_rq_rt(struct task_struct *p, int sync)
606 {
607         struct rq *rq = task_rq(p);
608
609         /*
610          * If the current task is an RT task, then
611          * try to see if we can wake this RT task up on another
612          * runqueue. Otherwise simply start this RT task
613          * on its current runqueue.
614          *
615          * We want to avoid overloading runqueues. Even if
616          * the RT task is of higher priority than the current RT task.
617          * RT tasks behave differently than other tasks. If
618          * one gets preempted, we try to push it off to another queue.
619          * So trying to keep a preempting RT task on the same
620          * cache hot CPU will force the running RT task to
621          * a cold CPU. So we waste all the cache for the lower
622          * RT task in hopes of saving some of a RT task
623          * that is just being woken and probably will have
624          * cold cache anyway.
625          */
626         if (unlikely(rt_task(rq->curr)) &&
627             (p->rt.nr_cpus_allowed > 1)) {
628                 int cpu = find_lowest_rq(p);
629
630                 return (cpu == -1) ? task_cpu(p) : cpu;
631         }
632
633         /*
634          * Otherwise, just let it ride on the affined RQ and the
635          * post-schedule router will push the preempted task away
636          */
637         return task_cpu(p);
638 }
639 #endif /* CONFIG_SMP */
640
641 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
642                                                    struct rt_rq *rt_rq);
643
644 /*
645  * Preempt the current task with a newly woken task if needed:
646  */
647 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
648 {
649         if (p->prio < rq->curr->prio) {
650                 resched_task(rq->curr);
651                 return;
652         }
653
654 #ifdef CONFIG_SMP
655         /*
656          * If:
657          *
658          * - the newly woken task is of equal priority to the current task
659          * - the newly woken task is non-migratable while current is migratable
660          * - current will be preempted on the next reschedule
661          *
662          * we should check to see if current can readily move to a different
663          * cpu.  If so, we will reschedule to allow the push logic to try
664          * to move current somewhere else, making room for our non-migratable
665          * task.
666          */
667         if((p->prio == rq->curr->prio)
668            && p->rt.nr_cpus_allowed == 1
669            && rq->curr->rt.nr_cpus_allowed != 1
670            && pick_next_rt_entity(rq, &rq->rt) != &rq->curr->rt) {
671                 cpumask_t mask;
672
673                 if (cpupri_find(&rq->rd->cpupri, rq->curr, &mask))
674                         /*
675                          * There appears to be other cpus that can accept
676                          * current, so lets reschedule to try and push it away
677                          */
678                         resched_task(rq->curr);
679         }
680 #endif
681 }
682
683 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
684                                                    struct rt_rq *rt_rq)
685 {
686         struct rt_prio_array *array = &rt_rq->active;
687         struct sched_rt_entity *next = NULL;
688         struct list_head *queue;
689         int idx;
690
691         idx = sched_find_first_bit(array->bitmap);
692         BUG_ON(idx >= MAX_RT_PRIO);
693
694         queue = array->xqueue + idx;
695         if (!list_empty(queue))
696                 next = list_entry(queue->next, struct sched_rt_entity,
697                                   run_list);
698         else {
699                 queue = array->squeue + idx;
700                 next = list_entry(queue->next, struct sched_rt_entity,
701                                   run_list);
702         }
703
704         return next;
705 }
706
707 static struct task_struct *pick_next_task_rt(struct rq *rq)
708 {
709         struct sched_rt_entity *rt_se;
710         struct task_struct *p;
711         struct rt_rq *rt_rq;
712
713         rt_rq = &rq->rt;
714
715         if (unlikely(!rt_rq->rt_nr_running))
716                 return NULL;
717
718         if (rt_rq_throttled(rt_rq))
719                 return NULL;
720
721         do {
722                 rt_se = pick_next_rt_entity(rq, rt_rq);
723                 BUG_ON(!rt_se);
724                 rt_rq = group_rt_rq(rt_se);
725         } while (rt_rq);
726
727         p = rt_task_of(rt_se);
728         p->se.exec_start = rq->clock;
729         return p;
730 }
731
732 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
733 {
734         update_curr_rt(rq);
735         p->se.exec_start = 0;
736 }
737
738 #ifdef CONFIG_SMP
739
740 /* Only try algorithms three times */
741 #define RT_MAX_TRIES 3
742
743 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
744 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
745
746 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
747 {
748         if (!task_running(rq, p) &&
749             (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
750             (p->rt.nr_cpus_allowed > 1))
751                 return 1;
752         return 0;
753 }
754
755 /* Return the second highest RT task, NULL otherwise */
756 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
757 {
758         struct task_struct *next = NULL;
759         struct sched_rt_entity *rt_se;
760         struct rt_prio_array *array;
761         struct rt_rq *rt_rq;
762         int idx;
763
764         for_each_leaf_rt_rq(rt_rq, rq) {
765                 array = &rt_rq->active;
766                 idx = sched_find_first_bit(array->bitmap);
767  next_idx:
768                 if (idx >= MAX_RT_PRIO)
769                         continue;
770                 if (next && next->prio < idx)
771                         continue;
772                 list_for_each_entry(rt_se, array->squeue + idx, run_list) {
773                         struct task_struct *p = rt_task_of(rt_se);
774                         if (pick_rt_task(rq, p, cpu)) {
775                                 next = p;
776                                 break;
777                         }
778                 }
779                 if (!next) {
780                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
781                         goto next_idx;
782                 }
783         }
784
785         return next;
786 }
787
788 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
789
790 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
791 {
792         int first;
793
794         /* "this_cpu" is cheaper to preempt than a remote processor */
795         if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
796                 return this_cpu;
797
798         first = first_cpu(*mask);
799         if (first != NR_CPUS)
800                 return first;
801
802         return -1;
803 }
804
805 static int find_lowest_rq(struct task_struct *task)
806 {
807         struct sched_domain *sd;
808         cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
809         int this_cpu = smp_processor_id();
810         int cpu      = task_cpu(task);
811
812         if (task->rt.nr_cpus_allowed == 1)
813                 return -1; /* No other targets possible */
814
815         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
816                 return -1; /* No targets found */
817
818         /*
819          * At this point we have built a mask of cpus representing the
820          * lowest priority tasks in the system.  Now we want to elect
821          * the best one based on our affinity and topology.
822          *
823          * We prioritize the last cpu that the task executed on since
824          * it is most likely cache-hot in that location.
825          */
826         if (cpu_isset(cpu, *lowest_mask))
827                 return cpu;
828
829         /*
830          * Otherwise, we consult the sched_domains span maps to figure
831          * out which cpu is logically closest to our hot cache data.
832          */
833         if (this_cpu == cpu)
834                 this_cpu = -1; /* Skip this_cpu opt if the same */
835
836         for_each_domain(cpu, sd) {
837                 if (sd->flags & SD_WAKE_AFFINE) {
838                         cpumask_t domain_mask;
839                         int       best_cpu;
840
841                         cpus_and(domain_mask, sd->span, *lowest_mask);
842
843                         best_cpu = pick_optimal_cpu(this_cpu,
844                                                     &domain_mask);
845                         if (best_cpu != -1)
846                                 return best_cpu;
847                 }
848         }
849
850         /*
851          * And finally, if there were no matches within the domains
852          * just give the caller *something* to work with from the compatible
853          * locations.
854          */
855         return pick_optimal_cpu(this_cpu, lowest_mask);
856 }
857
858 /* Will lock the rq it finds */
859 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
860 {
861         struct rq *lowest_rq = NULL;
862         int tries;
863         int cpu;
864
865         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
866                 cpu = find_lowest_rq(task);
867
868                 if ((cpu == -1) || (cpu == rq->cpu))
869                         break;
870
871                 lowest_rq = cpu_rq(cpu);
872
873                 /* if the prio of this runqueue changed, try again */
874                 if (double_lock_balance(rq, lowest_rq)) {
875                         /*
876                          * We had to unlock the run queue. In
877                          * the mean time, task could have
878                          * migrated already or had its affinity changed.
879                          * Also make sure that it wasn't scheduled on its rq.
880                          */
881                         if (unlikely(task_rq(task) != rq ||
882                                      !cpu_isset(lowest_rq->cpu,
883                                                 task->cpus_allowed) ||
884                                      task_running(rq, task) ||
885                                      !task->se.on_rq)) {
886
887                                 spin_unlock(&lowest_rq->lock);
888                                 lowest_rq = NULL;
889                                 break;
890                         }
891                 }
892
893                 /* If this rq is still suitable use it. */
894                 if (lowest_rq->rt.highest_prio > task->prio)
895                         break;
896
897                 /* try again */
898                 spin_unlock(&lowest_rq->lock);
899                 lowest_rq = NULL;
900         }
901
902         return lowest_rq;
903 }
904
905 /*
906  * If the current CPU has more than one RT task, see if the non
907  * running task can migrate over to a CPU that is running a task
908  * of lesser priority.
909  */
910 static int push_rt_task(struct rq *rq)
911 {
912         struct task_struct *next_task;
913         struct rq *lowest_rq;
914         int ret = 0;
915         int paranoid = RT_MAX_TRIES;
916
917         if (!rq->rt.overloaded)
918                 return 0;
919
920         next_task = pick_next_highest_task_rt(rq, -1);
921         if (!next_task)
922                 return 0;
923
924  retry:
925         if (unlikely(next_task == rq->curr)) {
926                 WARN_ON(1);
927                 return 0;
928         }
929
930         /*
931          * It's possible that the next_task slipped in of
932          * higher priority than current. If that's the case
933          * just reschedule current.
934          */
935         if (unlikely(next_task->prio < rq->curr->prio)) {
936                 resched_task(rq->curr);
937                 return 0;
938         }
939
940         /* We might release rq lock */
941         get_task_struct(next_task);
942
943         /* find_lock_lowest_rq locks the rq if found */
944         lowest_rq = find_lock_lowest_rq(next_task, rq);
945         if (!lowest_rq) {
946                 struct task_struct *task;
947                 /*
948                  * find lock_lowest_rq releases rq->lock
949                  * so it is possible that next_task has changed.
950                  * If it has, then try again.
951                  */
952                 task = pick_next_highest_task_rt(rq, -1);
953                 if (unlikely(task != next_task) && task && paranoid--) {
954                         put_task_struct(next_task);
955                         next_task = task;
956                         goto retry;
957                 }
958                 goto out;
959         }
960
961         deactivate_task(rq, next_task, 0);
962         set_task_cpu(next_task, lowest_rq->cpu);
963         activate_task(lowest_rq, next_task, 0);
964
965         resched_task(lowest_rq->curr);
966
967         spin_unlock(&lowest_rq->lock);
968
969         ret = 1;
970 out:
971         put_task_struct(next_task);
972
973         return ret;
974 }
975
976 /*
977  * TODO: Currently we just use the second highest prio task on
978  *       the queue, and stop when it can't migrate (or there's
979  *       no more RT tasks).  There may be a case where a lower
980  *       priority RT task has a different affinity than the
981  *       higher RT task. In this case the lower RT task could
982  *       possibly be able to migrate where as the higher priority
983  *       RT task could not.  We currently ignore this issue.
984  *       Enhancements are welcome!
985  */
986 static void push_rt_tasks(struct rq *rq)
987 {
988         /* push_rt_task will return true if it moved an RT */
989         while (push_rt_task(rq))
990                 ;
991 }
992
993 static int pull_rt_task(struct rq *this_rq)
994 {
995         int this_cpu = this_rq->cpu, ret = 0, cpu;
996         struct task_struct *p, *next;
997         struct rq *src_rq;
998
999         if (likely(!rt_overloaded(this_rq)))
1000                 return 0;
1001
1002         next = pick_next_task_rt(this_rq);
1003
1004         for_each_cpu_mask(cpu, this_rq->rd->rto_mask) {
1005                 if (this_cpu == cpu)
1006                         continue;
1007
1008                 src_rq = cpu_rq(cpu);
1009                 /*
1010                  * We can potentially drop this_rq's lock in
1011                  * double_lock_balance, and another CPU could
1012                  * steal our next task - hence we must cause
1013                  * the caller to recalculate the next task
1014                  * in that case:
1015                  */
1016                 if (double_lock_balance(this_rq, src_rq)) {
1017                         struct task_struct *old_next = next;
1018
1019                         next = pick_next_task_rt(this_rq);
1020                         if (next != old_next)
1021                                 ret = 1;
1022                 }
1023
1024                 /*
1025                  * Are there still pullable RT tasks?
1026                  */
1027                 if (src_rq->rt.rt_nr_running <= 1)
1028                         goto skip;
1029
1030                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1031
1032                 /*
1033                  * Do we have an RT task that preempts
1034                  * the to-be-scheduled task?
1035                  */
1036                 if (p && (!next || (p->prio < next->prio))) {
1037                         WARN_ON(p == src_rq->curr);
1038                         WARN_ON(!p->se.on_rq);
1039
1040                         /*
1041                          * There's a chance that p is higher in priority
1042                          * than what's currently running on its cpu.
1043                          * This is just that p is wakeing up and hasn't
1044                          * had a chance to schedule. We only pull
1045                          * p if it is lower in priority than the
1046                          * current task on the run queue or
1047                          * this_rq next task is lower in prio than
1048                          * the current task on that rq.
1049                          */
1050                         if (p->prio < src_rq->curr->prio ||
1051                             (next && next->prio < src_rq->curr->prio))
1052                                 goto skip;
1053
1054                         ret = 1;
1055
1056                         deactivate_task(src_rq, p, 0);
1057                         set_task_cpu(p, this_cpu);
1058                         activate_task(this_rq, p, 0);
1059                         /*
1060                          * We continue with the search, just in
1061                          * case there's an even higher prio task
1062                          * in another runqueue. (low likelyhood
1063                          * but possible)
1064                          *
1065                          * Update next so that we won't pick a task
1066                          * on another cpu with a priority lower (or equal)
1067                          * than the one we just picked.
1068                          */
1069                         next = p;
1070
1071                 }
1072  skip:
1073                 spin_unlock(&src_rq->lock);
1074         }
1075
1076         return ret;
1077 }
1078
1079 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1080 {
1081         /* Try to pull RT tasks here if we lower this rq's prio */
1082         if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio)
1083                 pull_rt_task(rq);
1084 }
1085
1086 static void post_schedule_rt(struct rq *rq)
1087 {
1088         /*
1089          * If we have more than one rt_task queued, then
1090          * see if we can push the other rt_tasks off to other CPUS.
1091          * Note we may release the rq lock, and since
1092          * the lock was owned by prev, we need to release it
1093          * first via finish_lock_switch and then reaquire it here.
1094          */
1095         if (unlikely(rq->rt.overloaded)) {
1096                 spin_lock_irq(&rq->lock);
1097                 push_rt_tasks(rq);
1098                 spin_unlock_irq(&rq->lock);
1099         }
1100 }
1101
1102 /*
1103  * If we are not running and we are not going to reschedule soon, we should
1104  * try to push tasks away now
1105  */
1106 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1107 {
1108         if (!task_running(rq, p) &&
1109             !test_tsk_need_resched(rq->curr) &&
1110             rq->rt.overloaded)
1111                 push_rt_tasks(rq);
1112 }
1113
1114 static unsigned long
1115 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1116                 unsigned long max_load_move,
1117                 struct sched_domain *sd, enum cpu_idle_type idle,
1118                 int *all_pinned, int *this_best_prio)
1119 {
1120         /* don't touch RT tasks */
1121         return 0;
1122 }
1123
1124 static int
1125 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1126                  struct sched_domain *sd, enum cpu_idle_type idle)
1127 {
1128         /* don't touch RT tasks */
1129         return 0;
1130 }
1131
1132 static void set_cpus_allowed_rt(struct task_struct *p,
1133                                 const cpumask_t *new_mask)
1134 {
1135         int weight = cpus_weight(*new_mask);
1136
1137         BUG_ON(!rt_task(p));
1138
1139         /*
1140          * Update the migration status of the RQ if we have an RT task
1141          * which is running AND changing its weight value.
1142          */
1143         if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1144                 struct rq *rq = task_rq(p);
1145
1146                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1147                         rq->rt.rt_nr_migratory++;
1148                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1149                         BUG_ON(!rq->rt.rt_nr_migratory);
1150                         rq->rt.rt_nr_migratory--;
1151                 }
1152
1153                 update_rt_migration(rq);
1154
1155                 if (unlikely(weight == 1 || p->rt.nr_cpus_allowed == 1))
1156                         /*
1157                          * If either the new or old weight is a "1", we need
1158                          * to requeue to properly move between shared and
1159                          * exclusive queues.
1160                          */
1161                         requeue_task_rt(rq, p);
1162         }
1163
1164         p->cpus_allowed    = *new_mask;
1165         p->rt.nr_cpus_allowed = weight;
1166 }
1167
1168 /* Assumes rq->lock is held */
1169 static void rq_online_rt(struct rq *rq)
1170 {
1171         if (rq->rt.overloaded)
1172                 rt_set_overload(rq);
1173
1174         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio);
1175 }
1176
1177 /* Assumes rq->lock is held */
1178 static void rq_offline_rt(struct rq *rq)
1179 {
1180         if (rq->rt.overloaded)
1181                 rt_clear_overload(rq);
1182
1183         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1184 }
1185
1186 /*
1187  * When switch from the rt queue, we bring ourselves to a position
1188  * that we might want to pull RT tasks from other runqueues.
1189  */
1190 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1191                            int running)
1192 {
1193         /*
1194          * If there are other RT tasks then we will reschedule
1195          * and the scheduling of the other RT tasks will handle
1196          * the balancing. But if we are the last RT task
1197          * we may need to handle the pulling of RT tasks
1198          * now.
1199          */
1200         if (!rq->rt.rt_nr_running)
1201                 pull_rt_task(rq);
1202 }
1203 #endif /* CONFIG_SMP */
1204
1205 /*
1206  * When switching a task to RT, we may overload the runqueue
1207  * with RT tasks. In this case we try to push them off to
1208  * other runqueues.
1209  */
1210 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1211                            int running)
1212 {
1213         int check_resched = 1;
1214
1215         /*
1216          * If we are already running, then there's nothing
1217          * that needs to be done. But if we are not running
1218          * we may need to preempt the current running task.
1219          * If that current running task is also an RT task
1220          * then see if we can move to another run queue.
1221          */
1222         if (!running) {
1223 #ifdef CONFIG_SMP
1224                 if (rq->rt.overloaded && push_rt_task(rq) &&
1225                     /* Don't resched if we changed runqueues */
1226                     rq != task_rq(p))
1227                         check_resched = 0;
1228 #endif /* CONFIG_SMP */
1229                 if (check_resched && p->prio < rq->curr->prio)
1230                         resched_task(rq->curr);
1231         }
1232 }
1233
1234 /*
1235  * Priority of the task has changed. This may cause
1236  * us to initiate a push or pull.
1237  */
1238 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1239                             int oldprio, int running)
1240 {
1241         if (running) {
1242 #ifdef CONFIG_SMP
1243                 /*
1244                  * If our priority decreases while running, we
1245                  * may need to pull tasks to this runqueue.
1246                  */
1247                 if (oldprio < p->prio)
1248                         pull_rt_task(rq);
1249                 /*
1250                  * If there's a higher priority task waiting to run
1251                  * then reschedule. Note, the above pull_rt_task
1252                  * can release the rq lock and p could migrate.
1253                  * Only reschedule if p is still on the same runqueue.
1254                  */
1255                 if (p->prio > rq->rt.highest_prio && rq->curr == p)
1256                         resched_task(p);
1257 #else
1258                 /* For UP simply resched on drop of prio */
1259                 if (oldprio < p->prio)
1260                         resched_task(p);
1261 #endif /* CONFIG_SMP */
1262         } else {
1263                 /*
1264                  * This task is not running, but if it is
1265                  * greater than the current running task
1266                  * then reschedule.
1267                  */
1268                 if (p->prio < rq->curr->prio)
1269                         resched_task(rq->curr);
1270         }
1271 }
1272
1273 static void watchdog(struct rq *rq, struct task_struct *p)
1274 {
1275         unsigned long soft, hard;
1276
1277         if (!p->signal)
1278                 return;
1279
1280         soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1281         hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1282
1283         if (soft != RLIM_INFINITY) {
1284                 unsigned long next;
1285
1286                 p->rt.timeout++;
1287                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1288                 if (p->rt.timeout > next)
1289                         p->it_sched_expires = p->se.sum_exec_runtime;
1290         }
1291 }
1292
1293 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1294 {
1295         update_curr_rt(rq);
1296
1297         watchdog(rq, p);
1298
1299         /*
1300          * RR tasks need a special form of timeslice management.
1301          * FIFO tasks have no timeslices.
1302          */
1303         if (p->policy != SCHED_RR)
1304                 return;
1305
1306         if (--p->rt.time_slice)
1307                 return;
1308
1309         p->rt.time_slice = DEF_TIMESLICE;
1310
1311         /*
1312          * Requeue to the end of queue if we are not the only element
1313          * on the queue:
1314          */
1315         if (p->rt.run_list.prev != p->rt.run_list.next) {
1316                 requeue_task_rt(rq, p);
1317                 set_tsk_need_resched(p);
1318         }
1319 }
1320
1321 static void set_curr_task_rt(struct rq *rq)
1322 {
1323         struct task_struct *p = rq->curr;
1324
1325         p->se.exec_start = rq->clock;
1326 }
1327
1328 static const struct sched_class rt_sched_class = {
1329         .next                   = &fair_sched_class,
1330         .enqueue_task           = enqueue_task_rt,
1331         .dequeue_task           = dequeue_task_rt,
1332         .yield_task             = yield_task_rt,
1333 #ifdef CONFIG_SMP
1334         .select_task_rq         = select_task_rq_rt,
1335 #endif /* CONFIG_SMP */
1336
1337         .check_preempt_curr     = check_preempt_curr_rt,
1338
1339         .pick_next_task         = pick_next_task_rt,
1340         .put_prev_task          = put_prev_task_rt,
1341
1342 #ifdef CONFIG_SMP
1343         .load_balance           = load_balance_rt,
1344         .move_one_task          = move_one_task_rt,
1345         .set_cpus_allowed       = set_cpus_allowed_rt,
1346         .rq_online              = rq_online_rt,
1347         .rq_offline             = rq_offline_rt,
1348         .pre_schedule           = pre_schedule_rt,
1349         .post_schedule          = post_schedule_rt,
1350         .task_wake_up           = task_wake_up_rt,
1351         .switched_from          = switched_from_rt,
1352 #endif
1353
1354         .set_curr_task          = set_curr_task_rt,
1355         .task_tick              = task_tick_rt,
1356
1357         .prio_changed           = prio_changed_rt,
1358         .switched_to            = switched_to_rt,
1359 };