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