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