sched: fix cpuprio build bug
[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
403                 rt_rq->highest_prio = rt_se_prio(rt_se);
404 #ifdef CONFIG_SMP
405                 if (rq->online)
406                         cpupri_set(&rq->rd->cpupri, rq->cpu,
407                                    rt_se_prio(rt_se));
408 #endif
409         }
410 #endif
411 #ifdef CONFIG_SMP
412         if (rt_se->nr_cpus_allowed > 1) {
413                 struct rq *rq = rq_of_rt_rq(rt_rq);
414
415                 rq->rt.rt_nr_migratory++;
416         }
417
418         update_rt_migration(rq_of_rt_rq(rt_rq));
419 #endif
420 #ifdef CONFIG_RT_GROUP_SCHED
421         if (rt_se_boosted(rt_se))
422                 rt_rq->rt_nr_boosted++;
423
424         if (rt_rq->tg)
425                 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
426 #else
427         start_rt_bandwidth(&def_rt_bandwidth);
428 #endif
429 }
430
431 static inline
432 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
433 {
434 #ifdef CONFIG_SMP
435         int highest_prio = rt_rq->highest_prio;
436 #endif
437
438         WARN_ON(!rt_prio(rt_se_prio(rt_se)));
439         WARN_ON(!rt_rq->rt_nr_running);
440         rt_rq->rt_nr_running--;
441 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
442         if (rt_rq->rt_nr_running) {
443                 struct rt_prio_array *array;
444
445                 WARN_ON(rt_se_prio(rt_se) < rt_rq->highest_prio);
446                 if (rt_se_prio(rt_se) == rt_rq->highest_prio) {
447                         /* recalculate */
448                         array = &rt_rq->active;
449                         rt_rq->highest_prio =
450                                 sched_find_first_bit(array->bitmap);
451                 } /* otherwise leave rq->highest prio alone */
452         } else
453                 rt_rq->highest_prio = MAX_RT_PRIO;
454 #endif
455 #ifdef CONFIG_SMP
456         if (rt_se->nr_cpus_allowed > 1) {
457                 struct rq *rq = rq_of_rt_rq(rt_rq);
458                 rq->rt.rt_nr_migratory--;
459         }
460
461         if (rt_rq->highest_prio != highest_prio) {
462                 struct rq *rq = rq_of_rt_rq(rt_rq);
463
464                 if (rq->online)
465                         cpupri_set(&rq->rd->cpupri, rq->cpu,
466                                    rt_rq->highest_prio);
467         }
468
469         update_rt_migration(rq_of_rt_rq(rt_rq));
470 #endif /* CONFIG_SMP */
471 #ifdef CONFIG_RT_GROUP_SCHED
472         if (rt_se_boosted(rt_se))
473                 rt_rq->rt_nr_boosted--;
474
475         WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
476 #endif
477 }
478
479 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
480 {
481         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
482         struct rt_prio_array *array = &rt_rq->active;
483         struct rt_rq *group_rq = group_rt_rq(rt_se);
484
485         if (group_rq && rt_rq_throttled(group_rq))
486                 return;
487
488         if (rt_se->nr_cpus_allowed == 1)
489                 list_add_tail(&rt_se->run_list,
490                               array->xqueue + rt_se_prio(rt_se));
491         else
492                 list_add_tail(&rt_se->run_list,
493                               array->squeue + rt_se_prio(rt_se));
494
495         __set_bit(rt_se_prio(rt_se), array->bitmap);
496
497         inc_rt_tasks(rt_se, rt_rq);
498 }
499
500 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
501 {
502         struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
503         struct rt_prio_array *array = &rt_rq->active;
504
505         list_del_init(&rt_se->run_list);
506         if (list_empty(array->squeue + rt_se_prio(rt_se))
507             && list_empty(array->xqueue + rt_se_prio(rt_se)))
508                 __clear_bit(rt_se_prio(rt_se), array->bitmap);
509
510         dec_rt_tasks(rt_se, rt_rq);
511 }
512
513 /*
514  * Because the prio of an upper entry depends on the lower
515  * entries, we must remove entries top - down.
516  */
517 static void dequeue_rt_stack(struct task_struct *p)
518 {
519         struct sched_rt_entity *rt_se, *back = NULL;
520
521         rt_se = &p->rt;
522         for_each_sched_rt_entity(rt_se) {
523                 rt_se->back = back;
524                 back = rt_se;
525         }
526
527         for (rt_se = back; rt_se; rt_se = rt_se->back) {
528                 if (on_rt_rq(rt_se))
529                         dequeue_rt_entity(rt_se);
530         }
531 }
532
533 /*
534  * Adding/removing a task to/from a priority array:
535  */
536 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
537 {
538         struct sched_rt_entity *rt_se = &p->rt;
539
540         if (wakeup)
541                 rt_se->timeout = 0;
542
543         dequeue_rt_stack(p);
544
545         /*
546          * enqueue everybody, bottom - up.
547          */
548         for_each_sched_rt_entity(rt_se)
549                 enqueue_rt_entity(rt_se);
550 }
551
552 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
553 {
554         struct sched_rt_entity *rt_se = &p->rt;
555         struct rt_rq *rt_rq;
556
557         update_curr_rt(rq);
558
559         dequeue_rt_stack(p);
560
561         /*
562          * re-enqueue all non-empty rt_rq entities.
563          */
564         for_each_sched_rt_entity(rt_se) {
565                 rt_rq = group_rt_rq(rt_se);
566                 if (rt_rq && rt_rq->rt_nr_running)
567                         enqueue_rt_entity(rt_se);
568         }
569 }
570
571 /*
572  * Put task to the end of the run list without the overhead of dequeue
573  * followed by enqueue.
574  *
575  * Note: We always enqueue the task to the shared-queue, regardless of its
576  * previous position w.r.t. exclusive vs shared.  This is so that exclusive RR
577  * tasks fairly round-robin with all tasks on the runqueue, not just other
578  * exclusive tasks.
579  */
580 static
581 void requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
582 {
583         struct rt_prio_array *array = &rt_rq->active;
584
585         list_del_init(&rt_se->run_list);
586         list_add_tail(&rt_se->run_list, array->squeue + rt_se_prio(rt_se));
587 }
588
589 static void requeue_task_rt(struct rq *rq, struct task_struct *p)
590 {
591         struct sched_rt_entity *rt_se = &p->rt;
592         struct rt_rq *rt_rq;
593
594         for_each_sched_rt_entity(rt_se) {
595                 rt_rq = rt_rq_of_se(rt_se);
596                 requeue_rt_entity(rt_rq, rt_se);
597         }
598 }
599
600 static void yield_task_rt(struct rq *rq)
601 {
602         requeue_task_rt(rq, rq->curr);
603 }
604
605 #ifdef CONFIG_SMP
606 static int find_lowest_rq(struct task_struct *task);
607
608 static int select_task_rq_rt(struct task_struct *p, int sync)
609 {
610         struct rq *rq = task_rq(p);
611
612         /*
613          * If the current task is an RT task, then
614          * try to see if we can wake this RT task up on another
615          * runqueue. Otherwise simply start this RT task
616          * on its current runqueue.
617          *
618          * We want to avoid overloading runqueues. Even if
619          * the RT task is of higher priority than the current RT task.
620          * RT tasks behave differently than other tasks. If
621          * one gets preempted, we try to push it off to another queue.
622          * So trying to keep a preempting RT task on the same
623          * cache hot CPU will force the running RT task to
624          * a cold CPU. So we waste all the cache for the lower
625          * RT task in hopes of saving some of a RT task
626          * that is just being woken and probably will have
627          * cold cache anyway.
628          */
629         if (unlikely(rt_task(rq->curr)) &&
630             (p->rt.nr_cpus_allowed > 1)) {
631                 int cpu = find_lowest_rq(p);
632
633                 return (cpu == -1) ? task_cpu(p) : cpu;
634         }
635
636         /*
637          * Otherwise, just let it ride on the affined RQ and the
638          * post-schedule router will push the preempted task away
639          */
640         return task_cpu(p);
641 }
642 #endif /* CONFIG_SMP */
643
644 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
645                                                    struct rt_rq *rt_rq);
646
647 /*
648  * Preempt the current task with a newly woken task if needed:
649  */
650 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
651 {
652         if (p->prio < rq->curr->prio) {
653                 resched_task(rq->curr);
654                 return;
655         }
656
657 #ifdef CONFIG_SMP
658         /*
659          * If:
660          *
661          * - the newly woken task is of equal priority to the current task
662          * - the newly woken task is non-migratable while current is migratable
663          * - current will be preempted on the next reschedule
664          *
665          * we should check to see if current can readily move to a different
666          * cpu.  If so, we will reschedule to allow the push logic to try
667          * to move current somewhere else, making room for our non-migratable
668          * task.
669          */
670         if((p->prio == rq->curr->prio)
671            && p->rt.nr_cpus_allowed == 1
672            && rq->curr->rt.nr_cpus_allowed != 1
673            && pick_next_rt_entity(rq, &rq->rt) != &rq->curr->rt) {
674                 cpumask_t mask;
675
676                 if (cpupri_find(&rq->rd->cpupri, rq->curr, &mask))
677                         /*
678                          * There appears to be other cpus that can accept
679                          * current, so lets reschedule to try and push it away
680                          */
681                         resched_task(rq->curr);
682         }
683 #endif
684 }
685
686 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
687                                                    struct rt_rq *rt_rq)
688 {
689         struct rt_prio_array *array = &rt_rq->active;
690         struct sched_rt_entity *next = NULL;
691         struct list_head *queue;
692         int idx;
693
694         idx = sched_find_first_bit(array->bitmap);
695         BUG_ON(idx >= MAX_RT_PRIO);
696
697         queue = array->xqueue + idx;
698         if (!list_empty(queue))
699                 next = list_entry(queue->next, struct sched_rt_entity,
700                                   run_list);
701         else {
702                 queue = array->squeue + idx;
703                 next = list_entry(queue->next, struct sched_rt_entity,
704                                   run_list);
705         }
706
707         return next;
708 }
709
710 static struct task_struct *pick_next_task_rt(struct rq *rq)
711 {
712         struct sched_rt_entity *rt_se;
713         struct task_struct *p;
714         struct rt_rq *rt_rq;
715
716         rt_rq = &rq->rt;
717
718         if (unlikely(!rt_rq->rt_nr_running))
719                 return NULL;
720
721         if (rt_rq_throttled(rt_rq))
722                 return NULL;
723
724         do {
725                 rt_se = pick_next_rt_entity(rq, rt_rq);
726                 BUG_ON(!rt_se);
727                 rt_rq = group_rt_rq(rt_se);
728         } while (rt_rq);
729
730         p = rt_task_of(rt_se);
731         p->se.exec_start = rq->clock;
732         return p;
733 }
734
735 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
736 {
737         update_curr_rt(rq);
738         p->se.exec_start = 0;
739 }
740
741 #ifdef CONFIG_SMP
742
743 /* Only try algorithms three times */
744 #define RT_MAX_TRIES 3
745
746 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
747 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
748
749 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
750 {
751         if (!task_running(rq, p) &&
752             (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
753             (p->rt.nr_cpus_allowed > 1))
754                 return 1;
755         return 0;
756 }
757
758 /* Return the second highest RT task, NULL otherwise */
759 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
760 {
761         struct task_struct *next = NULL;
762         struct sched_rt_entity *rt_se;
763         struct rt_prio_array *array;
764         struct rt_rq *rt_rq;
765         int idx;
766
767         for_each_leaf_rt_rq(rt_rq, rq) {
768                 array = &rt_rq->active;
769                 idx = sched_find_first_bit(array->bitmap);
770  next_idx:
771                 if (idx >= MAX_RT_PRIO)
772                         continue;
773                 if (next && next->prio < idx)
774                         continue;
775                 list_for_each_entry(rt_se, array->squeue + idx, run_list) {
776                         struct task_struct *p = rt_task_of(rt_se);
777                         if (pick_rt_task(rq, p, cpu)) {
778                                 next = p;
779                                 break;
780                         }
781                 }
782                 if (!next) {
783                         idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
784                         goto next_idx;
785                 }
786         }
787
788         return next;
789 }
790
791 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
792
793 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
794 {
795         int first;
796
797         /* "this_cpu" is cheaper to preempt than a remote processor */
798         if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
799                 return this_cpu;
800
801         first = first_cpu(*mask);
802         if (first != NR_CPUS)
803                 return first;
804
805         return -1;
806 }
807
808 static int find_lowest_rq(struct task_struct *task)
809 {
810         struct sched_domain *sd;
811         cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
812         int this_cpu = smp_processor_id();
813         int cpu      = task_cpu(task);
814
815         if (task->rt.nr_cpus_allowed == 1)
816                 return -1; /* No other targets possible */
817
818         if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
819                 return -1; /* No targets found */
820
821         /*
822          * At this point we have built a mask of cpus representing the
823          * lowest priority tasks in the system.  Now we want to elect
824          * the best one based on our affinity and topology.
825          *
826          * We prioritize the last cpu that the task executed on since
827          * it is most likely cache-hot in that location.
828          */
829         if (cpu_isset(cpu, *lowest_mask))
830                 return cpu;
831
832         /*
833          * Otherwise, we consult the sched_domains span maps to figure
834          * out which cpu is logically closest to our hot cache data.
835          */
836         if (this_cpu == cpu)
837                 this_cpu = -1; /* Skip this_cpu opt if the same */
838
839         for_each_domain(cpu, sd) {
840                 if (sd->flags & SD_WAKE_AFFINE) {
841                         cpumask_t domain_mask;
842                         int       best_cpu;
843
844                         cpus_and(domain_mask, sd->span, *lowest_mask);
845
846                         best_cpu = pick_optimal_cpu(this_cpu,
847                                                     &domain_mask);
848                         if (best_cpu != -1)
849                                 return best_cpu;
850                 }
851         }
852
853         /*
854          * And finally, if there were no matches within the domains
855          * just give the caller *something* to work with from the compatible
856          * locations.
857          */
858         return pick_optimal_cpu(this_cpu, lowest_mask);
859 }
860
861 /* Will lock the rq it finds */
862 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
863 {
864         struct rq *lowest_rq = NULL;
865         int tries;
866         int cpu;
867
868         for (tries = 0; tries < RT_MAX_TRIES; tries++) {
869                 cpu = find_lowest_rq(task);
870
871                 if ((cpu == -1) || (cpu == rq->cpu))
872                         break;
873
874                 lowest_rq = cpu_rq(cpu);
875
876                 /* if the prio of this runqueue changed, try again */
877                 if (double_lock_balance(rq, lowest_rq)) {
878                         /*
879                          * We had to unlock the run queue. In
880                          * the mean time, task could have
881                          * migrated already or had its affinity changed.
882                          * Also make sure that it wasn't scheduled on its rq.
883                          */
884                         if (unlikely(task_rq(task) != rq ||
885                                      !cpu_isset(lowest_rq->cpu,
886                                                 task->cpus_allowed) ||
887                                      task_running(rq, task) ||
888                                      !task->se.on_rq)) {
889
890                                 spin_unlock(&lowest_rq->lock);
891                                 lowest_rq = NULL;
892                                 break;
893                         }
894                 }
895
896                 /* If this rq is still suitable use it. */
897                 if (lowest_rq->rt.highest_prio > task->prio)
898                         break;
899
900                 /* try again */
901                 spin_unlock(&lowest_rq->lock);
902                 lowest_rq = NULL;
903         }
904
905         return lowest_rq;
906 }
907
908 /*
909  * If the current CPU has more than one RT task, see if the non
910  * running task can migrate over to a CPU that is running a task
911  * of lesser priority.
912  */
913 static int push_rt_task(struct rq *rq)
914 {
915         struct task_struct *next_task;
916         struct rq *lowest_rq;
917         int ret = 0;
918         int paranoid = RT_MAX_TRIES;
919
920         if (!rq->rt.overloaded)
921                 return 0;
922
923         next_task = pick_next_highest_task_rt(rq, -1);
924         if (!next_task)
925                 return 0;
926
927  retry:
928         if (unlikely(next_task == rq->curr)) {
929                 WARN_ON(1);
930                 return 0;
931         }
932
933         /*
934          * It's possible that the next_task slipped in of
935          * higher priority than current. If that's the case
936          * just reschedule current.
937          */
938         if (unlikely(next_task->prio < rq->curr->prio)) {
939                 resched_task(rq->curr);
940                 return 0;
941         }
942
943         /* We might release rq lock */
944         get_task_struct(next_task);
945
946         /* find_lock_lowest_rq locks the rq if found */
947         lowest_rq = find_lock_lowest_rq(next_task, rq);
948         if (!lowest_rq) {
949                 struct task_struct *task;
950                 /*
951                  * find lock_lowest_rq releases rq->lock
952                  * so it is possible that next_task has changed.
953                  * If it has, then try again.
954                  */
955                 task = pick_next_highest_task_rt(rq, -1);
956                 if (unlikely(task != next_task) && task && paranoid--) {
957                         put_task_struct(next_task);
958                         next_task = task;
959                         goto retry;
960                 }
961                 goto out;
962         }
963
964         deactivate_task(rq, next_task, 0);
965         set_task_cpu(next_task, lowest_rq->cpu);
966         activate_task(lowest_rq, next_task, 0);
967
968         resched_task(lowest_rq->curr);
969
970         spin_unlock(&lowest_rq->lock);
971
972         ret = 1;
973 out:
974         put_task_struct(next_task);
975
976         return ret;
977 }
978
979 /*
980  * TODO: Currently we just use the second highest prio task on
981  *       the queue, and stop when it can't migrate (or there's
982  *       no more RT tasks).  There may be a case where a lower
983  *       priority RT task has a different affinity than the
984  *       higher RT task. In this case the lower RT task could
985  *       possibly be able to migrate where as the higher priority
986  *       RT task could not.  We currently ignore this issue.
987  *       Enhancements are welcome!
988  */
989 static void push_rt_tasks(struct rq *rq)
990 {
991         /* push_rt_task will return true if it moved an RT */
992         while (push_rt_task(rq))
993                 ;
994 }
995
996 static int pull_rt_task(struct rq *this_rq)
997 {
998         int this_cpu = this_rq->cpu, ret = 0, cpu;
999         struct task_struct *p, *next;
1000         struct rq *src_rq;
1001
1002         if (likely(!rt_overloaded(this_rq)))
1003                 return 0;
1004
1005         next = pick_next_task_rt(this_rq);
1006
1007         for_each_cpu_mask(cpu, this_rq->rd->rto_mask) {
1008                 if (this_cpu == cpu)
1009                         continue;
1010
1011                 src_rq = cpu_rq(cpu);
1012                 /*
1013                  * We can potentially drop this_rq's lock in
1014                  * double_lock_balance, and another CPU could
1015                  * steal our next task - hence we must cause
1016                  * the caller to recalculate the next task
1017                  * in that case:
1018                  */
1019                 if (double_lock_balance(this_rq, src_rq)) {
1020                         struct task_struct *old_next = next;
1021
1022                         next = pick_next_task_rt(this_rq);
1023                         if (next != old_next)
1024                                 ret = 1;
1025                 }
1026
1027                 /*
1028                  * Are there still pullable RT tasks?
1029                  */
1030                 if (src_rq->rt.rt_nr_running <= 1)
1031                         goto skip;
1032
1033                 p = pick_next_highest_task_rt(src_rq, this_cpu);
1034
1035                 /*
1036                  * Do we have an RT task that preempts
1037                  * the to-be-scheduled task?
1038                  */
1039                 if (p && (!next || (p->prio < next->prio))) {
1040                         WARN_ON(p == src_rq->curr);
1041                         WARN_ON(!p->se.on_rq);
1042
1043                         /*
1044                          * There's a chance that p is higher in priority
1045                          * than what's currently running on its cpu.
1046                          * This is just that p is wakeing up and hasn't
1047                          * had a chance to schedule. We only pull
1048                          * p if it is lower in priority than the
1049                          * current task on the run queue or
1050                          * this_rq next task is lower in prio than
1051                          * the current task on that rq.
1052                          */
1053                         if (p->prio < src_rq->curr->prio ||
1054                             (next && next->prio < src_rq->curr->prio))
1055                                 goto skip;
1056
1057                         ret = 1;
1058
1059                         deactivate_task(src_rq, p, 0);
1060                         set_task_cpu(p, this_cpu);
1061                         activate_task(this_rq, p, 0);
1062                         /*
1063                          * We continue with the search, just in
1064                          * case there's an even higher prio task
1065                          * in another runqueue. (low likelyhood
1066                          * but possible)
1067                          *
1068                          * Update next so that we won't pick a task
1069                          * on another cpu with a priority lower (or equal)
1070                          * than the one we just picked.
1071                          */
1072                         next = p;
1073
1074                 }
1075  skip:
1076                 spin_unlock(&src_rq->lock);
1077         }
1078
1079         return ret;
1080 }
1081
1082 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1083 {
1084         /* Try to pull RT tasks here if we lower this rq's prio */
1085         if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio)
1086                 pull_rt_task(rq);
1087 }
1088
1089 static void post_schedule_rt(struct rq *rq)
1090 {
1091         /*
1092          * If we have more than one rt_task queued, then
1093          * see if we can push the other rt_tasks off to other CPUS.
1094          * Note we may release the rq lock, and since
1095          * the lock was owned by prev, we need to release it
1096          * first via finish_lock_switch and then reaquire it here.
1097          */
1098         if (unlikely(rq->rt.overloaded)) {
1099                 spin_lock_irq(&rq->lock);
1100                 push_rt_tasks(rq);
1101                 spin_unlock_irq(&rq->lock);
1102         }
1103 }
1104
1105 /*
1106  * If we are not running and we are not going to reschedule soon, we should
1107  * try to push tasks away now
1108  */
1109 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1110 {
1111         if (!task_running(rq, p) &&
1112             !test_tsk_need_resched(rq->curr) &&
1113             rq->rt.overloaded)
1114                 push_rt_tasks(rq);
1115 }
1116
1117 static unsigned long
1118 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1119                 unsigned long max_load_move,
1120                 struct sched_domain *sd, enum cpu_idle_type idle,
1121                 int *all_pinned, int *this_best_prio)
1122 {
1123         /* don't touch RT tasks */
1124         return 0;
1125 }
1126
1127 static int
1128 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1129                  struct sched_domain *sd, enum cpu_idle_type idle)
1130 {
1131         /* don't touch RT tasks */
1132         return 0;
1133 }
1134
1135 static void set_cpus_allowed_rt(struct task_struct *p,
1136                                 const cpumask_t *new_mask)
1137 {
1138         int weight = cpus_weight(*new_mask);
1139
1140         BUG_ON(!rt_task(p));
1141
1142         /*
1143          * Update the migration status of the RQ if we have an RT task
1144          * which is running AND changing its weight value.
1145          */
1146         if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1147                 struct rq *rq = task_rq(p);
1148
1149                 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1150                         rq->rt.rt_nr_migratory++;
1151                 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1152                         BUG_ON(!rq->rt.rt_nr_migratory);
1153                         rq->rt.rt_nr_migratory--;
1154                 }
1155
1156                 update_rt_migration(rq);
1157
1158                 if (unlikely(weight == 1 || p->rt.nr_cpus_allowed == 1))
1159                         /*
1160                          * If either the new or old weight is a "1", we need
1161                          * to requeue to properly move between shared and
1162                          * exclusive queues.
1163                          */
1164                         requeue_task_rt(rq, p);
1165         }
1166
1167         p->cpus_allowed    = *new_mask;
1168         p->rt.nr_cpus_allowed = weight;
1169 }
1170
1171 /* Assumes rq->lock is held */
1172 static void rq_online_rt(struct rq *rq)
1173 {
1174         if (rq->rt.overloaded)
1175                 rt_set_overload(rq);
1176
1177         cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio);
1178 }
1179
1180 /* Assumes rq->lock is held */
1181 static void rq_offline_rt(struct rq *rq)
1182 {
1183         if (rq->rt.overloaded)
1184                 rt_clear_overload(rq);
1185
1186         cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1187 }
1188
1189 /*
1190  * When switch from the rt queue, we bring ourselves to a position
1191  * that we might want to pull RT tasks from other runqueues.
1192  */
1193 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1194                            int running)
1195 {
1196         /*
1197          * If there are other RT tasks then we will reschedule
1198          * and the scheduling of the other RT tasks will handle
1199          * the balancing. But if we are the last RT task
1200          * we may need to handle the pulling of RT tasks
1201          * now.
1202          */
1203         if (!rq->rt.rt_nr_running)
1204                 pull_rt_task(rq);
1205 }
1206 #endif /* CONFIG_SMP */
1207
1208 /*
1209  * When switching a task to RT, we may overload the runqueue
1210  * with RT tasks. In this case we try to push them off to
1211  * other runqueues.
1212  */
1213 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1214                            int running)
1215 {
1216         int check_resched = 1;
1217
1218         /*
1219          * If we are already running, then there's nothing
1220          * that needs to be done. But if we are not running
1221          * we may need to preempt the current running task.
1222          * If that current running task is also an RT task
1223          * then see if we can move to another run queue.
1224          */
1225         if (!running) {
1226 #ifdef CONFIG_SMP
1227                 if (rq->rt.overloaded && push_rt_task(rq) &&
1228                     /* Don't resched if we changed runqueues */
1229                     rq != task_rq(p))
1230                         check_resched = 0;
1231 #endif /* CONFIG_SMP */
1232                 if (check_resched && p->prio < rq->curr->prio)
1233                         resched_task(rq->curr);
1234         }
1235 }
1236
1237 /*
1238  * Priority of the task has changed. This may cause
1239  * us to initiate a push or pull.
1240  */
1241 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1242                             int oldprio, int running)
1243 {
1244         if (running) {
1245 #ifdef CONFIG_SMP
1246                 /*
1247                  * If our priority decreases while running, we
1248                  * may need to pull tasks to this runqueue.
1249                  */
1250                 if (oldprio < p->prio)
1251                         pull_rt_task(rq);
1252                 /*
1253                  * If there's a higher priority task waiting to run
1254                  * then reschedule. Note, the above pull_rt_task
1255                  * can release the rq lock and p could migrate.
1256                  * Only reschedule if p is still on the same runqueue.
1257                  */
1258                 if (p->prio > rq->rt.highest_prio && rq->curr == p)
1259                         resched_task(p);
1260 #else
1261                 /* For UP simply resched on drop of prio */
1262                 if (oldprio < p->prio)
1263                         resched_task(p);
1264 #endif /* CONFIG_SMP */
1265         } else {
1266                 /*
1267                  * This task is not running, but if it is
1268                  * greater than the current running task
1269                  * then reschedule.
1270                  */
1271                 if (p->prio < rq->curr->prio)
1272                         resched_task(rq->curr);
1273         }
1274 }
1275
1276 static void watchdog(struct rq *rq, struct task_struct *p)
1277 {
1278         unsigned long soft, hard;
1279
1280         if (!p->signal)
1281                 return;
1282
1283         soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1284         hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1285
1286         if (soft != RLIM_INFINITY) {
1287                 unsigned long next;
1288
1289                 p->rt.timeout++;
1290                 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1291                 if (p->rt.timeout > next)
1292                         p->it_sched_expires = p->se.sum_exec_runtime;
1293         }
1294 }
1295
1296 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1297 {
1298         update_curr_rt(rq);
1299
1300         watchdog(rq, p);
1301
1302         /*
1303          * RR tasks need a special form of timeslice management.
1304          * FIFO tasks have no timeslices.
1305          */
1306         if (p->policy != SCHED_RR)
1307                 return;
1308
1309         if (--p->rt.time_slice)
1310                 return;
1311
1312         p->rt.time_slice = DEF_TIMESLICE;
1313
1314         /*
1315          * Requeue to the end of queue if we are not the only element
1316          * on the queue:
1317          */
1318         if (p->rt.run_list.prev != p->rt.run_list.next) {
1319                 requeue_task_rt(rq, p);
1320                 set_tsk_need_resched(p);
1321         }
1322 }
1323
1324 static void set_curr_task_rt(struct rq *rq)
1325 {
1326         struct task_struct *p = rq->curr;
1327
1328         p->se.exec_start = rq->clock;
1329 }
1330
1331 static const struct sched_class rt_sched_class = {
1332         .next                   = &fair_sched_class,
1333         .enqueue_task           = enqueue_task_rt,
1334         .dequeue_task           = dequeue_task_rt,
1335         .yield_task             = yield_task_rt,
1336 #ifdef CONFIG_SMP
1337         .select_task_rq         = select_task_rq_rt,
1338 #endif /* CONFIG_SMP */
1339
1340         .check_preempt_curr     = check_preempt_curr_rt,
1341
1342         .pick_next_task         = pick_next_task_rt,
1343         .put_prev_task          = put_prev_task_rt,
1344
1345 #ifdef CONFIG_SMP
1346         .load_balance           = load_balance_rt,
1347         .move_one_task          = move_one_task_rt,
1348         .set_cpus_allowed       = set_cpus_allowed_rt,
1349         .rq_online              = rq_online_rt,
1350         .rq_offline             = rq_offline_rt,
1351         .pre_schedule           = pre_schedule_rt,
1352         .post_schedule          = post_schedule_rt,
1353         .task_wake_up           = task_wake_up_rt,
1354         .switched_from          = switched_from_rt,
1355 #endif
1356
1357         .set_curr_task          = set_curr_task_rt,
1358         .task_tick              = task_tick_rt,
1359
1360         .prio_changed           = prio_changed_rt,
1361         .switched_to            = switched_to_rt,
1362 };