Merge branch 'merge' of git://git.kernel.org/pub/scm/linux/kernel/git/paulus/powerpc
[linux-2.6/lfs.git] / kernel / sched_rt.c
blob060e87b0cb1c7e3ab7c12084da2061322e4a3f5a
1 /*
2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
3 * policies)
4 */
6 #ifdef CONFIG_SMP
8 static inline int rt_overloaded(struct rq *rq)
10 return atomic_read(&rq->rd->rto_count);
13 static inline void rt_set_overload(struct rq *rq)
15 cpu_set(rq->cpu, rq->rd->rto_mask);
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.
23 wmb();
24 atomic_inc(&rq->rd->rto_count);
27 static inline void rt_clear_overload(struct rq *rq)
29 /* the order here really doesn't matter */
30 atomic_dec(&rq->rd->rto_count);
31 cpu_clear(rq->cpu, rq->rd->rto_mask);
34 static void update_rt_migration(struct rq *rq)
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;
41 } else if (rq->rt.overloaded) {
42 rt_clear_overload(rq);
43 rq->rt.overloaded = 0;
46 #endif /* CONFIG_SMP */
48 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
50 return container_of(rt_se, struct task_struct, rt);
53 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
55 return !list_empty(&rt_se->run_list);
58 #ifdef CONFIG_RT_GROUP_SCHED
60 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
62 if (!rt_rq->tg)
63 return RUNTIME_INF;
65 return rt_rq->rt_runtime;
68 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
70 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
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)
76 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
78 return rt_rq->rq;
81 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
83 return rt_se->rt_rq;
86 #define for_each_sched_rt_entity(rt_se) \
87 for (; rt_se; rt_se = rt_se->parent)
89 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
91 return rt_se->my_q;
94 static void enqueue_rt_entity(struct sched_rt_entity *rt_se);
95 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
97 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
99 struct sched_rt_entity *rt_se = rt_rq->rt_se;
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;
104 enqueue_rt_entity(rt_se);
105 if (rt_rq->highest_prio < curr->prio)
106 resched_task(curr);
110 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
112 struct sched_rt_entity *rt_se = rt_rq->rt_se;
114 if (rt_se && on_rt_rq(rt_se))
115 dequeue_rt_entity(rt_se);
118 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
120 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
123 static int rt_se_boosted(struct sched_rt_entity *rt_se)
125 struct rt_rq *rt_rq = group_rt_rq(rt_se);
126 struct task_struct *p;
128 if (rt_rq)
129 return !!rt_rq->rt_nr_boosted;
131 p = rt_task_of(rt_se);
132 return p->prio != p->normal_prio;
135 #ifdef CONFIG_SMP
136 static inline cpumask_t sched_rt_period_mask(void)
138 return cpu_rq(smp_processor_id())->rd->span;
140 #else
141 static inline cpumask_t sched_rt_period_mask(void)
143 return cpu_online_map;
145 #endif
147 static inline
148 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
150 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
153 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
155 return &rt_rq->tg->rt_bandwidth;
158 #else
160 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
162 return rt_rq->rt_runtime;
165 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
167 return ktime_to_ns(def_rt_bandwidth.rt_period);
170 #define for_each_leaf_rt_rq(rt_rq, rq) \
171 for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
173 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
175 return container_of(rt_rq, struct rq, rt);
178 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
180 struct task_struct *p = rt_task_of(rt_se);
181 struct rq *rq = task_rq(p);
183 return &rq->rt;
186 #define for_each_sched_rt_entity(rt_se) \
187 for (; rt_se; rt_se = NULL)
189 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
191 return NULL;
194 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
198 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
202 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
204 return rt_rq->rt_throttled;
207 static inline cpumask_t sched_rt_period_mask(void)
209 return cpu_online_map;
212 static inline
213 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
215 return &cpu_rq(cpu)->rt;
218 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
220 return &def_rt_bandwidth;
223 #endif
225 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
227 int i, idle = 1;
228 cpumask_t span;
230 if (rt_b->rt_runtime == RUNTIME_INF)
231 return 1;
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);
239 spin_lock(&rq->lock);
240 if (rt_rq->rt_time) {
241 u64 runtime;
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;
250 if (rt_rq->rt_time || rt_rq->rt_nr_running)
251 idle = 0;
252 spin_unlock(&rt_rq->rt_runtime_lock);
255 if (enqueue)
256 sched_rt_rq_enqueue(rt_rq);
257 spin_unlock(&rq->lock);
260 return idle;
263 #ifdef CONFIG_SMP
264 static int balance_runtime(struct rt_rq *rt_rq)
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;
271 weight = cpus_weight(rd->span);
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;
279 if (iter == rt_rq)
280 continue;
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;
296 spin_unlock(&iter->rt_runtime_lock);
298 spin_unlock(&rt_b->rt_runtime_lock);
300 return more;
302 #endif
304 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
306 #ifdef CONFIG_RT_GROUP_SCHED
307 struct rt_rq *rt_rq = group_rt_rq(rt_se);
309 if (rt_rq)
310 return rt_rq->highest_prio;
311 #endif
313 return rt_task_of(rt_se)->prio;
316 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
318 u64 runtime = sched_rt_runtime(rt_rq);
320 if (runtime == RUNTIME_INF)
321 return 0;
323 if (rt_rq->rt_throttled)
324 return rt_rq_throttled(rt_rq);
326 if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
327 return 0;
329 #ifdef CONFIG_SMP
330 if (rt_rq->rt_time > runtime) {
331 int more;
333 spin_unlock(&rt_rq->rt_runtime_lock);
334 more = balance_runtime(rt_rq);
335 spin_lock(&rt_rq->rt_runtime_lock);
337 if (more)
338 runtime = sched_rt_runtime(rt_rq);
340 #endif
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;
350 return 0;
354 * Update the current task's runtime statistics. Skip current tasks that
355 * are not in our scheduling class.
357 static void update_curr_rt(struct rq *rq)
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;
364 if (!task_has_rt_policy(curr))
365 return;
367 delta_exec = rq->clock - curr->se.exec_start;
368 if (unlikely((s64)delta_exec < 0))
369 delta_exec = 0;
371 schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
373 curr->se.sum_exec_runtime += delta_exec;
374 curr->se.exec_start = rq->clock;
375 cpuacct_charge(curr, delta_exec);
377 for_each_sched_rt_entity(rt_se) {
378 rt_rq = rt_rq_of_se(rt_se);
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);
388 static inline
389 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
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++;
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++;
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
416 static inline
417 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
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;
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--;
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--;
448 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
449 #endif
452 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
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);
458 if (group_rq && rt_rq_throttled(group_rq))
459 return;
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);
464 inc_rt_tasks(rt_se, rt_rq);
467 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
469 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
470 struct rt_prio_array *array = &rt_rq->active;
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);
476 dec_rt_tasks(rt_se, rt_rq);
480 * Because the prio of an upper entry depends on the lower
481 * entries, we must remove entries top - down.
483 static void dequeue_rt_stack(struct task_struct *p)
485 struct sched_rt_entity *rt_se, *back = NULL;
487 rt_se = &p->rt;
488 for_each_sched_rt_entity(rt_se) {
489 rt_se->back = back;
490 back = rt_se;
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);
500 * Adding/removing a task to/from a priority array:
502 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
504 struct sched_rt_entity *rt_se = &p->rt;
506 if (wakeup)
507 rt_se->timeout = 0;
509 dequeue_rt_stack(p);
512 * enqueue everybody, bottom - up.
514 for_each_sched_rt_entity(rt_se)
515 enqueue_rt_entity(rt_se);
517 inc_cpu_load(rq, p->se.load.weight);
520 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
522 struct sched_rt_entity *rt_se = &p->rt;
523 struct rt_rq *rt_rq;
525 update_curr_rt(rq);
527 dequeue_rt_stack(p);
530 * re-enqueue all non-empty rt_rq entities.
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);
538 dec_cpu_load(rq, p->se.load.weight);
542 * Put task to the end of the run list without the overhead of dequeue
543 * followed by enqueue.
545 static
546 void requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se)
548 struct rt_prio_array *array = &rt_rq->active;
550 list_move_tail(&rt_se->run_list, array->queue + rt_se_prio(rt_se));
553 static void requeue_task_rt(struct rq *rq, struct task_struct *p)
555 struct sched_rt_entity *rt_se = &p->rt;
556 struct rt_rq *rt_rq;
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);
564 static void yield_task_rt(struct rq *rq)
566 requeue_task_rt(rq, rq->curr);
569 #ifdef CONFIG_SMP
570 static int find_lowest_rq(struct task_struct *task);
572 static int select_task_rq_rt(struct task_struct *p, int sync)
574 struct rq *rq = task_rq(p);
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.
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.
593 if (unlikely(rt_task(rq->curr)) &&
594 (p->rt.nr_cpus_allowed > 1)) {
595 int cpu = find_lowest_rq(p);
597 return (cpu == -1) ? task_cpu(p) : cpu;
601 * Otherwise, just let it ride on the affined RQ and the
602 * post-schedule router will push the preempted task away
604 return task_cpu(p);
606 #endif /* CONFIG_SMP */
609 * Preempt the current task with a newly woken task if needed:
611 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p)
613 if (p->prio < rq->curr->prio)
614 resched_task(rq->curr);
617 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
618 struct rt_rq *rt_rq)
620 struct rt_prio_array *array = &rt_rq->active;
621 struct sched_rt_entity *next = NULL;
622 struct list_head *queue;
623 int idx;
625 idx = sched_find_first_bit(array->bitmap);
626 BUG_ON(idx >= MAX_RT_PRIO);
628 queue = array->queue + idx;
629 next = list_entry(queue->next, struct sched_rt_entity, run_list);
631 return next;
634 static struct task_struct *pick_next_task_rt(struct rq *rq)
636 struct sched_rt_entity *rt_se;
637 struct task_struct *p;
638 struct rt_rq *rt_rq;
640 rt_rq = &rq->rt;
642 if (unlikely(!rt_rq->rt_nr_running))
643 return NULL;
645 if (rt_rq_throttled(rt_rq))
646 return NULL;
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);
654 p = rt_task_of(rt_se);
655 p->se.exec_start = rq->clock;
656 return p;
659 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
661 update_curr_rt(rq);
662 p->se.exec_start = 0;
665 #ifdef CONFIG_SMP
667 /* Only try algorithms three times */
668 #define RT_MAX_TRIES 3
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);
673 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
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;
682 /* Return the second highest RT task, NULL otherwise */
683 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
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;
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;
706 if (!next) {
707 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
708 goto next_idx;
712 return next;
715 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
717 static int find_lowest_cpus(struct task_struct *task, cpumask_t *lowest_mask)
719 int lowest_prio = -1;
720 int lowest_cpu = -1;
721 int count = 0;
722 int cpu;
724 cpus_and(*lowest_mask, task_rq(task)->rd->online, task->cpus_allowed);
727 * Scan each rq for the lowest prio.
729 for_each_cpu_mask(cpu, *lowest_mask) {
730 struct rq *rq = cpu_rq(cpu);
732 /* We look for lowest RT prio or non-rt CPU */
733 if (rq->rt.highest_prio >= MAX_RT_PRIO) {
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.
742 if (lowest_cpu != -1) {
743 cpus_clear(*lowest_mask);
744 cpu_set(rq->cpu, *lowest_mask);
746 return 1;
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;
758 count++;
759 } else
760 cpu_clear(cpu, *lowest_mask);
764 * Clear out all the set bits that represent
765 * runqueues that were of higher prio than
766 * the lowest_prio.
768 if (lowest_cpu > 0) {
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.
774 for_each_cpu_mask(cpu, *lowest_mask) {
775 if (cpu >= lowest_cpu)
776 break;
777 cpu_clear(cpu, *lowest_mask);
781 return count;
784 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
786 int first;
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;
792 first = first_cpu(*mask);
793 if (first != NR_CPUS)
794 return first;
796 return -1;
799 static int find_lowest_rq(struct task_struct *task)
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);
807 if (!count)
808 return -1; /* No targets found */
811 * There is no sense in performing an optimal search if only one
812 * target is found.
814 if (count == 1)
815 return first_cpu(*lowest_mask);
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.
822 * We prioritize the last cpu that the task executed on since
823 * it is most likely cache-hot in that location.
825 if (cpu_isset(cpu, *lowest_mask))
826 return cpu;
829 * Otherwise, we consult the sched_domains span maps to figure
830 * out which cpu is logically closest to our hot cache data.
832 if (this_cpu == cpu)
833 this_cpu = -1; /* Skip this_cpu opt if the same */
835 for_each_domain(cpu, sd) {
836 if (sd->flags & SD_WAKE_AFFINE) {
837 cpumask_t domain_mask;
838 int best_cpu;
840 cpus_and(domain_mask, sd->span, *lowest_mask);
842 best_cpu = pick_optimal_cpu(this_cpu,
843 &domain_mask);
844 if (best_cpu != -1)
845 return best_cpu;
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.
854 return pick_optimal_cpu(this_cpu, lowest_mask);
857 /* Will lock the rq it finds */
858 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
860 struct rq *lowest_rq = NULL;
861 int tries;
862 int cpu;
864 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
865 cpu = find_lowest_rq(task);
867 if ((cpu == -1) || (cpu == rq->cpu))
868 break;
870 lowest_rq = cpu_rq(cpu);
872 /* if the prio of this runqueue changed, try again */
873 if (double_lock_balance(rq, lowest_rq)) {
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.
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)) {
886 spin_unlock(&lowest_rq->lock);
887 lowest_rq = NULL;
888 break;
892 /* If this rq is still suitable use it. */
893 if (lowest_rq->rt.highest_prio > task->prio)
894 break;
896 /* try again */
897 spin_unlock(&lowest_rq->lock);
898 lowest_rq = NULL;
901 return lowest_rq;
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.
909 static int push_rt_task(struct rq *rq)
911 struct task_struct *next_task;
912 struct rq *lowest_rq;
913 int ret = 0;
914 int paranoid = RT_MAX_TRIES;
916 if (!rq->rt.overloaded)
917 return 0;
919 next_task = pick_next_highest_task_rt(rq, -1);
920 if (!next_task)
921 return 0;
923 retry:
924 if (unlikely(next_task == rq->curr)) {
925 WARN_ON(1);
926 return 0;
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.
934 if (unlikely(next_task->prio < rq->curr->prio)) {
935 resched_task(rq->curr);
936 return 0;
939 /* We might release rq lock */
940 get_task_struct(next_task);
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;
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.
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;
957 goto out;
960 deactivate_task(rq, next_task, 0);
961 set_task_cpu(next_task, lowest_rq->cpu);
962 activate_task(lowest_rq, next_task, 0);
964 resched_task(lowest_rq->curr);
966 spin_unlock(&lowest_rq->lock);
968 ret = 1;
969 out:
970 put_task_struct(next_task);
972 return ret;
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!
985 static void push_rt_tasks(struct rq *rq)
987 /* push_rt_task will return true if it moved an RT */
988 while (push_rt_task(rq))
992 static int pull_rt_task(struct rq *this_rq)
994 int this_cpu = this_rq->cpu, ret = 0, cpu;
995 struct task_struct *p, *next;
996 struct rq *src_rq;
998 if (likely(!rt_overloaded(this_rq)))
999 return 0;
1001 next = pick_next_task_rt(this_rq);
1003 for_each_cpu_mask(cpu, this_rq->rd->rto_mask) {
1004 if (this_cpu == cpu)
1005 continue;
1007 src_rq = cpu_rq(cpu);
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:
1015 if (double_lock_balance(this_rq, src_rq)) {
1016 struct task_struct *old_next = next;
1018 next = pick_next_task_rt(this_rq);
1019 if (next != old_next)
1020 ret = 1;
1024 * Are there still pullable RT tasks?
1026 if (src_rq->rt.rt_nr_running <= 1)
1027 goto skip;
1029 p = pick_next_highest_task_rt(src_rq, this_cpu);
1032 * Do we have an RT task that preempts
1033 * the to-be-scheduled task?
1035 if (p && (!next || (p->prio < next->prio))) {
1036 WARN_ON(p == src_rq->curr);
1037 WARN_ON(!p->se.on_rq);
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.
1049 if (p->prio < src_rq->curr->prio ||
1050 (next && next->prio < src_rq->curr->prio))
1051 goto skip;
1053 ret = 1;
1055 deactivate_task(src_rq, p, 0);
1056 set_task_cpu(p, this_cpu);
1057 activate_task(this_rq, p, 0);
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)
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.
1068 next = p;
1071 skip:
1072 spin_unlock(&src_rq->lock);
1075 return ret;
1078 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
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);
1085 static void post_schedule_rt(struct rq *rq)
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.
1094 if (unlikely(rq->rt.overloaded)) {
1095 spin_lock_irq(&rq->lock);
1096 push_rt_tasks(rq);
1097 spin_unlock_irq(&rq->lock);
1102 * If we are not running and we are not going to reschedule soon, we should
1103 * try to push tasks away now
1105 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1107 if (!task_running(rq, p) &&
1108 !test_tsk_need_resched(rq->curr) &&
1109 rq->rt.overloaded)
1110 push_rt_tasks(rq);
1113 static unsigned long
1114 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1115 unsigned long max_load_move,
1116 struct sched_domain *sd, enum cpu_idle_type idle,
1117 int *all_pinned, int *this_best_prio)
1119 /* don't touch RT tasks */
1120 return 0;
1123 static int
1124 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1125 struct sched_domain *sd, enum cpu_idle_type idle)
1127 /* don't touch RT tasks */
1128 return 0;
1131 static void set_cpus_allowed_rt(struct task_struct *p,
1132 const cpumask_t *new_mask)
1134 int weight = cpus_weight(*new_mask);
1136 BUG_ON(!rt_task(p));
1139 * Update the migration status of the RQ if we have an RT task
1140 * which is running AND changing its weight value.
1142 if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1143 struct rq *rq = task_rq(p);
1145 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1146 rq->rt.rt_nr_migratory++;
1147 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1148 BUG_ON(!rq->rt.rt_nr_migratory);
1149 rq->rt.rt_nr_migratory--;
1152 update_rt_migration(rq);
1155 p->cpus_allowed = *new_mask;
1156 p->rt.nr_cpus_allowed = weight;
1159 /* Assumes rq->lock is held */
1160 static void join_domain_rt(struct rq *rq)
1162 if (rq->rt.overloaded)
1163 rt_set_overload(rq);
1166 /* Assumes rq->lock is held */
1167 static void leave_domain_rt(struct rq *rq)
1169 if (rq->rt.overloaded)
1170 rt_clear_overload(rq);
1174 * When switch from the rt queue, we bring ourselves to a position
1175 * that we might want to pull RT tasks from other runqueues.
1177 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1178 int running)
1181 * If there are other RT tasks then we will reschedule
1182 * and the scheduling of the other RT tasks will handle
1183 * the balancing. But if we are the last RT task
1184 * we may need to handle the pulling of RT tasks
1185 * now.
1187 if (!rq->rt.rt_nr_running)
1188 pull_rt_task(rq);
1190 #endif /* CONFIG_SMP */
1193 * When switching a task to RT, we may overload the runqueue
1194 * with RT tasks. In this case we try to push them off to
1195 * other runqueues.
1197 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1198 int running)
1200 int check_resched = 1;
1203 * If we are already running, then there's nothing
1204 * that needs to be done. But if we are not running
1205 * we may need to preempt the current running task.
1206 * If that current running task is also an RT task
1207 * then see if we can move to another run queue.
1209 if (!running) {
1210 #ifdef CONFIG_SMP
1211 if (rq->rt.overloaded && push_rt_task(rq) &&
1212 /* Don't resched if we changed runqueues */
1213 rq != task_rq(p))
1214 check_resched = 0;
1215 #endif /* CONFIG_SMP */
1216 if (check_resched && p->prio < rq->curr->prio)
1217 resched_task(rq->curr);
1222 * Priority of the task has changed. This may cause
1223 * us to initiate a push or pull.
1225 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1226 int oldprio, int running)
1228 if (running) {
1229 #ifdef CONFIG_SMP
1231 * If our priority decreases while running, we
1232 * may need to pull tasks to this runqueue.
1234 if (oldprio < p->prio)
1235 pull_rt_task(rq);
1237 * If there's a higher priority task waiting to run
1238 * then reschedule. Note, the above pull_rt_task
1239 * can release the rq lock and p could migrate.
1240 * Only reschedule if p is still on the same runqueue.
1242 if (p->prio > rq->rt.highest_prio && rq->curr == p)
1243 resched_task(p);
1244 #else
1245 /* For UP simply resched on drop of prio */
1246 if (oldprio < p->prio)
1247 resched_task(p);
1248 #endif /* CONFIG_SMP */
1249 } else {
1251 * This task is not running, but if it is
1252 * greater than the current running task
1253 * then reschedule.
1255 if (p->prio < rq->curr->prio)
1256 resched_task(rq->curr);
1260 static void watchdog(struct rq *rq, struct task_struct *p)
1262 unsigned long soft, hard;
1264 if (!p->signal)
1265 return;
1267 soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1268 hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1270 if (soft != RLIM_INFINITY) {
1271 unsigned long next;
1273 p->rt.timeout++;
1274 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1275 if (p->rt.timeout > next)
1276 p->it_sched_expires = p->se.sum_exec_runtime;
1280 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1282 update_curr_rt(rq);
1284 watchdog(rq, p);
1287 * RR tasks need a special form of timeslice management.
1288 * FIFO tasks have no timeslices.
1290 if (p->policy != SCHED_RR)
1291 return;
1293 if (--p->rt.time_slice)
1294 return;
1296 p->rt.time_slice = DEF_TIMESLICE;
1299 * Requeue to the end of queue if we are not the only element
1300 * on the queue:
1302 if (p->rt.run_list.prev != p->rt.run_list.next) {
1303 requeue_task_rt(rq, p);
1304 set_tsk_need_resched(p);
1308 static void set_curr_task_rt(struct rq *rq)
1310 struct task_struct *p = rq->curr;
1312 p->se.exec_start = rq->clock;
1315 static const struct sched_class rt_sched_class = {
1316 .next = &fair_sched_class,
1317 .enqueue_task = enqueue_task_rt,
1318 .dequeue_task = dequeue_task_rt,
1319 .yield_task = yield_task_rt,
1320 #ifdef CONFIG_SMP
1321 .select_task_rq = select_task_rq_rt,
1322 #endif /* CONFIG_SMP */
1324 .check_preempt_curr = check_preempt_curr_rt,
1326 .pick_next_task = pick_next_task_rt,
1327 .put_prev_task = put_prev_task_rt,
1329 #ifdef CONFIG_SMP
1330 .load_balance = load_balance_rt,
1331 .move_one_task = move_one_task_rt,
1332 .set_cpus_allowed = set_cpus_allowed_rt,
1333 .join_domain = join_domain_rt,
1334 .leave_domain = leave_domain_rt,
1335 .pre_schedule = pre_schedule_rt,
1336 .post_schedule = post_schedule_rt,
1337 .task_wake_up = task_wake_up_rt,
1338 .switched_from = switched_from_rt,
1339 #endif
1341 .set_curr_task = set_curr_task_rt,
1342 .task_tick = task_tick_rt,
1344 .prio_changed = prio_changed_rt,
1345 .switched_to = switched_to_rt,