hwmon: applesmc: specified number of bytes to read should match actual
[linux-2.6/verdex.git] / kernel / sched_rt.c
blobcdf5740ab03e8133c0a2b7713d6c77d2be1f07bf
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 if (!rq->online)
16 return;
18 cpu_set(rq->cpu, rq->rd->rto_mask);
20 * Make sure the mask is visible before we set
21 * the overload count. That is checked to determine
22 * if we should look at the mask. It would be a shame
23 * if we looked at the mask, but the mask was not
24 * updated yet.
26 wmb();
27 atomic_inc(&rq->rd->rto_count);
30 static inline void rt_clear_overload(struct rq *rq)
32 if (!rq->online)
33 return;
35 /* the order here really doesn't matter */
36 atomic_dec(&rq->rd->rto_count);
37 cpu_clear(rq->cpu, rq->rd->rto_mask);
40 static void update_rt_migration(struct rq *rq)
42 if (rq->rt.rt_nr_migratory && (rq->rt.rt_nr_running > 1)) {
43 if (!rq->rt.overloaded) {
44 rt_set_overload(rq);
45 rq->rt.overloaded = 1;
47 } else if (rq->rt.overloaded) {
48 rt_clear_overload(rq);
49 rq->rt.overloaded = 0;
52 #endif /* CONFIG_SMP */
54 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
56 return container_of(rt_se, struct task_struct, rt);
59 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
61 return !list_empty(&rt_se->run_list);
64 #ifdef CONFIG_RT_GROUP_SCHED
66 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
68 if (!rt_rq->tg)
69 return RUNTIME_INF;
71 return rt_rq->rt_runtime;
74 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
76 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
79 #define for_each_leaf_rt_rq(rt_rq, rq) \
80 list_for_each_entry(rt_rq, &rq->leaf_rt_rq_list, leaf_rt_rq_list)
82 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
84 return rt_rq->rq;
87 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
89 return rt_se->rt_rq;
92 #define for_each_sched_rt_entity(rt_se) \
93 for (; rt_se; rt_se = rt_se->parent)
95 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
97 return rt_se->my_q;
100 static void enqueue_rt_entity(struct sched_rt_entity *rt_se);
101 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
103 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
105 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
106 struct sched_rt_entity *rt_se = rt_rq->rt_se;
108 if (rt_rq->rt_nr_running) {
109 if (rt_se && !on_rt_rq(rt_se))
110 enqueue_rt_entity(rt_se);
111 if (rt_rq->highest_prio < curr->prio)
112 resched_task(curr);
116 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
118 struct sched_rt_entity *rt_se = rt_rq->rt_se;
120 if (rt_se && on_rt_rq(rt_se))
121 dequeue_rt_entity(rt_se);
124 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
126 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
129 static int rt_se_boosted(struct sched_rt_entity *rt_se)
131 struct rt_rq *rt_rq = group_rt_rq(rt_se);
132 struct task_struct *p;
134 if (rt_rq)
135 return !!rt_rq->rt_nr_boosted;
137 p = rt_task_of(rt_se);
138 return p->prio != p->normal_prio;
141 #ifdef CONFIG_SMP
142 static inline cpumask_t sched_rt_period_mask(void)
144 return cpu_rq(smp_processor_id())->rd->span;
146 #else
147 static inline cpumask_t sched_rt_period_mask(void)
149 return cpu_online_map;
151 #endif
153 static inline
154 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
156 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
159 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
161 return &rt_rq->tg->rt_bandwidth;
164 #else /* !CONFIG_RT_GROUP_SCHED */
166 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
168 return rt_rq->rt_runtime;
171 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
173 return ktime_to_ns(def_rt_bandwidth.rt_period);
176 #define for_each_leaf_rt_rq(rt_rq, rq) \
177 for (rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
179 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
181 return container_of(rt_rq, struct rq, rt);
184 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
186 struct task_struct *p = rt_task_of(rt_se);
187 struct rq *rq = task_rq(p);
189 return &rq->rt;
192 #define for_each_sched_rt_entity(rt_se) \
193 for (; rt_se; rt_se = NULL)
195 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
197 return NULL;
200 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
202 if (rt_rq->rt_nr_running)
203 resched_task(rq_of_rt_rq(rt_rq)->curr);
206 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
210 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
212 return rt_rq->rt_throttled;
215 static inline cpumask_t sched_rt_period_mask(void)
217 return cpu_online_map;
220 static inline
221 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
223 return &cpu_rq(cpu)->rt;
226 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
228 return &def_rt_bandwidth;
231 #endif /* CONFIG_RT_GROUP_SCHED */
233 #ifdef CONFIG_SMP
235 * We ran out of runtime, see if we can borrow some from our neighbours.
237 static int do_balance_runtime(struct rt_rq *rt_rq)
239 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
240 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
241 int i, weight, more = 0;
242 u64 rt_period;
244 weight = cpus_weight(rd->span);
246 spin_lock(&rt_b->rt_runtime_lock);
247 rt_period = ktime_to_ns(rt_b->rt_period);
248 for_each_cpu_mask_nr(i, rd->span) {
249 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
250 s64 diff;
252 if (iter == rt_rq)
253 continue;
255 spin_lock(&iter->rt_runtime_lock);
257 * Either all rqs have inf runtime and there's nothing to steal
258 * or __disable_runtime() below sets a specific rq to inf to
259 * indicate its been disabled and disalow stealing.
261 if (iter->rt_runtime == RUNTIME_INF)
262 goto next;
265 * From runqueues with spare time, take 1/n part of their
266 * spare time, but no more than our period.
268 diff = iter->rt_runtime - iter->rt_time;
269 if (diff > 0) {
270 diff = div_u64((u64)diff, weight);
271 if (rt_rq->rt_runtime + diff > rt_period)
272 diff = rt_period - rt_rq->rt_runtime;
273 iter->rt_runtime -= diff;
274 rt_rq->rt_runtime += diff;
275 more = 1;
276 if (rt_rq->rt_runtime == rt_period) {
277 spin_unlock(&iter->rt_runtime_lock);
278 break;
281 next:
282 spin_unlock(&iter->rt_runtime_lock);
284 spin_unlock(&rt_b->rt_runtime_lock);
286 return more;
290 * Ensure this RQ takes back all the runtime it lend to its neighbours.
292 static void __disable_runtime(struct rq *rq)
294 struct root_domain *rd = rq->rd;
295 struct rt_rq *rt_rq;
297 if (unlikely(!scheduler_running))
298 return;
300 for_each_leaf_rt_rq(rt_rq, rq) {
301 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
302 s64 want;
303 int i;
305 spin_lock(&rt_b->rt_runtime_lock);
306 spin_lock(&rt_rq->rt_runtime_lock);
308 * Either we're all inf and nobody needs to borrow, or we're
309 * already disabled and thus have nothing to do, or we have
310 * exactly the right amount of runtime to take out.
312 if (rt_rq->rt_runtime == RUNTIME_INF ||
313 rt_rq->rt_runtime == rt_b->rt_runtime)
314 goto balanced;
315 spin_unlock(&rt_rq->rt_runtime_lock);
318 * Calculate the difference between what we started out with
319 * and what we current have, that's the amount of runtime
320 * we lend and now have to reclaim.
322 want = rt_b->rt_runtime - rt_rq->rt_runtime;
325 * Greedy reclaim, take back as much as we can.
327 for_each_cpu_mask(i, rd->span) {
328 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
329 s64 diff;
332 * Can't reclaim from ourselves or disabled runqueues.
334 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
335 continue;
337 spin_lock(&iter->rt_runtime_lock);
338 if (want > 0) {
339 diff = min_t(s64, iter->rt_runtime, want);
340 iter->rt_runtime -= diff;
341 want -= diff;
342 } else {
343 iter->rt_runtime -= want;
344 want -= want;
346 spin_unlock(&iter->rt_runtime_lock);
348 if (!want)
349 break;
352 spin_lock(&rt_rq->rt_runtime_lock);
354 * We cannot be left wanting - that would mean some runtime
355 * leaked out of the system.
357 BUG_ON(want);
358 balanced:
360 * Disable all the borrow logic by pretending we have inf
361 * runtime - in which case borrowing doesn't make sense.
363 rt_rq->rt_runtime = RUNTIME_INF;
364 spin_unlock(&rt_rq->rt_runtime_lock);
365 spin_unlock(&rt_b->rt_runtime_lock);
369 static void disable_runtime(struct rq *rq)
371 unsigned long flags;
373 spin_lock_irqsave(&rq->lock, flags);
374 __disable_runtime(rq);
375 spin_unlock_irqrestore(&rq->lock, flags);
378 static void __enable_runtime(struct rq *rq)
380 struct rt_rq *rt_rq;
382 if (unlikely(!scheduler_running))
383 return;
386 * Reset each runqueue's bandwidth settings
388 for_each_leaf_rt_rq(rt_rq, rq) {
389 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
391 spin_lock(&rt_b->rt_runtime_lock);
392 spin_lock(&rt_rq->rt_runtime_lock);
393 rt_rq->rt_runtime = rt_b->rt_runtime;
394 rt_rq->rt_time = 0;
395 rt_rq->rt_throttled = 0;
396 spin_unlock(&rt_rq->rt_runtime_lock);
397 spin_unlock(&rt_b->rt_runtime_lock);
401 static void enable_runtime(struct rq *rq)
403 unsigned long flags;
405 spin_lock_irqsave(&rq->lock, flags);
406 __enable_runtime(rq);
407 spin_unlock_irqrestore(&rq->lock, flags);
410 static int balance_runtime(struct rt_rq *rt_rq)
412 int more = 0;
414 if (rt_rq->rt_time > rt_rq->rt_runtime) {
415 spin_unlock(&rt_rq->rt_runtime_lock);
416 more = do_balance_runtime(rt_rq);
417 spin_lock(&rt_rq->rt_runtime_lock);
420 return more;
422 #else /* !CONFIG_SMP */
423 static inline int balance_runtime(struct rt_rq *rt_rq)
425 return 0;
427 #endif /* CONFIG_SMP */
429 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
431 int i, idle = 1;
432 cpumask_t span;
434 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
435 return 1;
437 span = sched_rt_period_mask();
438 for_each_cpu_mask(i, span) {
439 int enqueue = 0;
440 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
441 struct rq *rq = rq_of_rt_rq(rt_rq);
443 spin_lock(&rq->lock);
444 if (rt_rq->rt_time) {
445 u64 runtime;
447 spin_lock(&rt_rq->rt_runtime_lock);
448 if (rt_rq->rt_throttled)
449 balance_runtime(rt_rq);
450 runtime = rt_rq->rt_runtime;
451 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
452 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
453 rt_rq->rt_throttled = 0;
454 enqueue = 1;
456 if (rt_rq->rt_time || rt_rq->rt_nr_running)
457 idle = 0;
458 spin_unlock(&rt_rq->rt_runtime_lock);
459 } else if (rt_rq->rt_nr_running)
460 idle = 0;
462 if (enqueue)
463 sched_rt_rq_enqueue(rt_rq);
464 spin_unlock(&rq->lock);
467 return idle;
470 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
472 #ifdef CONFIG_RT_GROUP_SCHED
473 struct rt_rq *rt_rq = group_rt_rq(rt_se);
475 if (rt_rq)
476 return rt_rq->highest_prio;
477 #endif
479 return rt_task_of(rt_se)->prio;
482 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
484 u64 runtime = sched_rt_runtime(rt_rq);
486 if (rt_rq->rt_throttled)
487 return rt_rq_throttled(rt_rq);
489 if (sched_rt_runtime(rt_rq) >= sched_rt_period(rt_rq))
490 return 0;
492 balance_runtime(rt_rq);
493 runtime = sched_rt_runtime(rt_rq);
494 if (runtime == RUNTIME_INF)
495 return 0;
497 if (rt_rq->rt_time > runtime) {
498 rt_rq->rt_throttled = 1;
499 if (rt_rq_throttled(rt_rq)) {
500 sched_rt_rq_dequeue(rt_rq);
501 return 1;
505 return 0;
509 * Update the current task's runtime statistics. Skip current tasks that
510 * are not in our scheduling class.
512 static void update_curr_rt(struct rq *rq)
514 struct task_struct *curr = rq->curr;
515 struct sched_rt_entity *rt_se = &curr->rt;
516 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
517 u64 delta_exec;
519 if (!task_has_rt_policy(curr))
520 return;
522 delta_exec = rq->clock - curr->se.exec_start;
523 if (unlikely((s64)delta_exec < 0))
524 delta_exec = 0;
526 schedstat_set(curr->se.exec_max, max(curr->se.exec_max, delta_exec));
528 curr->se.sum_exec_runtime += delta_exec;
529 curr->se.exec_start = rq->clock;
530 cpuacct_charge(curr, delta_exec);
532 if (!rt_bandwidth_enabled())
533 return;
535 for_each_sched_rt_entity(rt_se) {
536 rt_rq = rt_rq_of_se(rt_se);
538 spin_lock(&rt_rq->rt_runtime_lock);
539 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
540 rt_rq->rt_time += delta_exec;
541 if (sched_rt_runtime_exceeded(rt_rq))
542 resched_task(curr);
544 spin_unlock(&rt_rq->rt_runtime_lock);
548 static inline
549 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
551 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
552 rt_rq->rt_nr_running++;
553 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
554 if (rt_se_prio(rt_se) < rt_rq->highest_prio) {
555 #ifdef CONFIG_SMP
556 struct rq *rq = rq_of_rt_rq(rt_rq);
557 #endif
559 rt_rq->highest_prio = rt_se_prio(rt_se);
560 #ifdef CONFIG_SMP
561 if (rq->online)
562 cpupri_set(&rq->rd->cpupri, rq->cpu,
563 rt_se_prio(rt_se));
564 #endif
566 #endif
567 #ifdef CONFIG_SMP
568 if (rt_se->nr_cpus_allowed > 1) {
569 struct rq *rq = rq_of_rt_rq(rt_rq);
571 rq->rt.rt_nr_migratory++;
574 update_rt_migration(rq_of_rt_rq(rt_rq));
575 #endif
576 #ifdef CONFIG_RT_GROUP_SCHED
577 if (rt_se_boosted(rt_se))
578 rt_rq->rt_nr_boosted++;
580 if (rt_rq->tg)
581 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
582 #else
583 start_rt_bandwidth(&def_rt_bandwidth);
584 #endif
587 static inline
588 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
590 #ifdef CONFIG_SMP
591 int highest_prio = rt_rq->highest_prio;
592 #endif
594 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
595 WARN_ON(!rt_rq->rt_nr_running);
596 rt_rq->rt_nr_running--;
597 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
598 if (rt_rq->rt_nr_running) {
599 struct rt_prio_array *array;
601 WARN_ON(rt_se_prio(rt_se) < rt_rq->highest_prio);
602 if (rt_se_prio(rt_se) == rt_rq->highest_prio) {
603 /* recalculate */
604 array = &rt_rq->active;
605 rt_rq->highest_prio =
606 sched_find_first_bit(array->bitmap);
607 } /* otherwise leave rq->highest prio alone */
608 } else
609 rt_rq->highest_prio = MAX_RT_PRIO;
610 #endif
611 #ifdef CONFIG_SMP
612 if (rt_se->nr_cpus_allowed > 1) {
613 struct rq *rq = rq_of_rt_rq(rt_rq);
614 rq->rt.rt_nr_migratory--;
617 if (rt_rq->highest_prio != highest_prio) {
618 struct rq *rq = rq_of_rt_rq(rt_rq);
620 if (rq->online)
621 cpupri_set(&rq->rd->cpupri, rq->cpu,
622 rt_rq->highest_prio);
625 update_rt_migration(rq_of_rt_rq(rt_rq));
626 #endif /* CONFIG_SMP */
627 #ifdef CONFIG_RT_GROUP_SCHED
628 if (rt_se_boosted(rt_se))
629 rt_rq->rt_nr_boosted--;
631 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
632 #endif
635 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se)
637 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
638 struct rt_prio_array *array = &rt_rq->active;
639 struct rt_rq *group_rq = group_rt_rq(rt_se);
640 struct list_head *queue = array->queue + rt_se_prio(rt_se);
643 * Don't enqueue the group if its throttled, or when empty.
644 * The latter is a consequence of the former when a child group
645 * get throttled and the current group doesn't have any other
646 * active members.
648 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
649 return;
651 list_add_tail(&rt_se->run_list, queue);
652 __set_bit(rt_se_prio(rt_se), array->bitmap);
654 inc_rt_tasks(rt_se, rt_rq);
657 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
659 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
660 struct rt_prio_array *array = &rt_rq->active;
662 list_del_init(&rt_se->run_list);
663 if (list_empty(array->queue + rt_se_prio(rt_se)))
664 __clear_bit(rt_se_prio(rt_se), array->bitmap);
666 dec_rt_tasks(rt_se, rt_rq);
670 * Because the prio of an upper entry depends on the lower
671 * entries, we must remove entries top - down.
673 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
675 struct sched_rt_entity *back = NULL;
677 for_each_sched_rt_entity(rt_se) {
678 rt_se->back = back;
679 back = rt_se;
682 for (rt_se = back; rt_se; rt_se = rt_se->back) {
683 if (on_rt_rq(rt_se))
684 __dequeue_rt_entity(rt_se);
688 static void enqueue_rt_entity(struct sched_rt_entity *rt_se)
690 dequeue_rt_stack(rt_se);
691 for_each_sched_rt_entity(rt_se)
692 __enqueue_rt_entity(rt_se);
695 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
697 dequeue_rt_stack(rt_se);
699 for_each_sched_rt_entity(rt_se) {
700 struct rt_rq *rt_rq = group_rt_rq(rt_se);
702 if (rt_rq && rt_rq->rt_nr_running)
703 __enqueue_rt_entity(rt_se);
708 * Adding/removing a task to/from a priority array:
710 static void enqueue_task_rt(struct rq *rq, struct task_struct *p, int wakeup)
712 struct sched_rt_entity *rt_se = &p->rt;
714 if (wakeup)
715 rt_se->timeout = 0;
717 enqueue_rt_entity(rt_se);
719 inc_cpu_load(rq, p->se.load.weight);
722 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int sleep)
724 struct sched_rt_entity *rt_se = &p->rt;
726 update_curr_rt(rq);
727 dequeue_rt_entity(rt_se);
729 dec_cpu_load(rq, p->se.load.weight);
733 * Put task to the end of the run list without the overhead of dequeue
734 * followed by enqueue.
736 static void
737 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
739 if (on_rt_rq(rt_se)) {
740 struct rt_prio_array *array = &rt_rq->active;
741 struct list_head *queue = array->queue + rt_se_prio(rt_se);
743 if (head)
744 list_move(&rt_se->run_list, queue);
745 else
746 list_move_tail(&rt_se->run_list, queue);
750 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
752 struct sched_rt_entity *rt_se = &p->rt;
753 struct rt_rq *rt_rq;
755 for_each_sched_rt_entity(rt_se) {
756 rt_rq = rt_rq_of_se(rt_se);
757 requeue_rt_entity(rt_rq, rt_se, head);
761 static void yield_task_rt(struct rq *rq)
763 requeue_task_rt(rq, rq->curr, 0);
766 #ifdef CONFIG_SMP
767 static int find_lowest_rq(struct task_struct *task);
769 static int select_task_rq_rt(struct task_struct *p, int sync)
771 struct rq *rq = task_rq(p);
774 * If the current task is an RT task, then
775 * try to see if we can wake this RT task up on another
776 * runqueue. Otherwise simply start this RT task
777 * on its current runqueue.
779 * We want to avoid overloading runqueues. Even if
780 * the RT task is of higher priority than the current RT task.
781 * RT tasks behave differently than other tasks. If
782 * one gets preempted, we try to push it off to another queue.
783 * So trying to keep a preempting RT task on the same
784 * cache hot CPU will force the running RT task to
785 * a cold CPU. So we waste all the cache for the lower
786 * RT task in hopes of saving some of a RT task
787 * that is just being woken and probably will have
788 * cold cache anyway.
790 if (unlikely(rt_task(rq->curr)) &&
791 (p->rt.nr_cpus_allowed > 1)) {
792 int cpu = find_lowest_rq(p);
794 return (cpu == -1) ? task_cpu(p) : cpu;
798 * Otherwise, just let it ride on the affined RQ and the
799 * post-schedule router will push the preempted task away
801 return task_cpu(p);
804 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
806 cpumask_t mask;
808 if (rq->curr->rt.nr_cpus_allowed == 1)
809 return;
811 if (p->rt.nr_cpus_allowed != 1
812 && cpupri_find(&rq->rd->cpupri, p, &mask))
813 return;
815 if (!cpupri_find(&rq->rd->cpupri, rq->curr, &mask))
816 return;
819 * There appears to be other cpus that can accept
820 * current and none to run 'p', so lets reschedule
821 * to try and push current away:
823 requeue_task_rt(rq, p, 1);
824 resched_task(rq->curr);
827 #endif /* CONFIG_SMP */
830 * Preempt the current task with a newly woken task if needed:
832 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int sync)
834 if (p->prio < rq->curr->prio) {
835 resched_task(rq->curr);
836 return;
839 #ifdef CONFIG_SMP
841 * If:
843 * - the newly woken task is of equal priority to the current task
844 * - the newly woken task is non-migratable while current is migratable
845 * - current will be preempted on the next reschedule
847 * we should check to see if current can readily move to a different
848 * cpu. If so, we will reschedule to allow the push logic to try
849 * to move current somewhere else, making room for our non-migratable
850 * task.
852 if (p->prio == rq->curr->prio && !need_resched())
853 check_preempt_equal_prio(rq, p);
854 #endif
857 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
858 struct rt_rq *rt_rq)
860 struct rt_prio_array *array = &rt_rq->active;
861 struct sched_rt_entity *next = NULL;
862 struct list_head *queue;
863 int idx;
865 idx = sched_find_first_bit(array->bitmap);
866 BUG_ON(idx >= MAX_RT_PRIO);
868 queue = array->queue + idx;
869 next = list_entry(queue->next, struct sched_rt_entity, run_list);
871 return next;
874 static struct task_struct *pick_next_task_rt(struct rq *rq)
876 struct sched_rt_entity *rt_se;
877 struct task_struct *p;
878 struct rt_rq *rt_rq;
880 rt_rq = &rq->rt;
882 if (unlikely(!rt_rq->rt_nr_running))
883 return NULL;
885 if (rt_rq_throttled(rt_rq))
886 return NULL;
888 do {
889 rt_se = pick_next_rt_entity(rq, rt_rq);
890 BUG_ON(!rt_se);
891 rt_rq = group_rt_rq(rt_se);
892 } while (rt_rq);
894 p = rt_task_of(rt_se);
895 p->se.exec_start = rq->clock;
896 return p;
899 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
901 update_curr_rt(rq);
902 p->se.exec_start = 0;
905 #ifdef CONFIG_SMP
907 /* Only try algorithms three times */
908 #define RT_MAX_TRIES 3
910 static int double_lock_balance(struct rq *this_rq, struct rq *busiest);
911 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest);
913 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep);
915 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
917 if (!task_running(rq, p) &&
918 (cpu < 0 || cpu_isset(cpu, p->cpus_allowed)) &&
919 (p->rt.nr_cpus_allowed > 1))
920 return 1;
921 return 0;
924 /* Return the second highest RT task, NULL otherwise */
925 static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)
927 struct task_struct *next = NULL;
928 struct sched_rt_entity *rt_se;
929 struct rt_prio_array *array;
930 struct rt_rq *rt_rq;
931 int idx;
933 for_each_leaf_rt_rq(rt_rq, rq) {
934 array = &rt_rq->active;
935 idx = sched_find_first_bit(array->bitmap);
936 next_idx:
937 if (idx >= MAX_RT_PRIO)
938 continue;
939 if (next && next->prio < idx)
940 continue;
941 list_for_each_entry(rt_se, array->queue + idx, run_list) {
942 struct task_struct *p = rt_task_of(rt_se);
943 if (pick_rt_task(rq, p, cpu)) {
944 next = p;
945 break;
948 if (!next) {
949 idx = find_next_bit(array->bitmap, MAX_RT_PRIO, idx+1);
950 goto next_idx;
954 return next;
957 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
959 static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
961 int first;
963 /* "this_cpu" is cheaper to preempt than a remote processor */
964 if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
965 return this_cpu;
967 first = first_cpu(*mask);
968 if (first != NR_CPUS)
969 return first;
971 return -1;
974 static int find_lowest_rq(struct task_struct *task)
976 struct sched_domain *sd;
977 cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
978 int this_cpu = smp_processor_id();
979 int cpu = task_cpu(task);
981 if (task->rt.nr_cpus_allowed == 1)
982 return -1; /* No other targets possible */
984 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
985 return -1; /* No targets found */
988 * Only consider CPUs that are usable for migration.
989 * I guess we might want to change cpupri_find() to ignore those
990 * in the first place.
992 cpus_and(*lowest_mask, *lowest_mask, cpu_active_map);
995 * At this point we have built a mask of cpus representing the
996 * lowest priority tasks in the system. Now we want to elect
997 * the best one based on our affinity and topology.
999 * We prioritize the last cpu that the task executed on since
1000 * it is most likely cache-hot in that location.
1002 if (cpu_isset(cpu, *lowest_mask))
1003 return cpu;
1006 * Otherwise, we consult the sched_domains span maps to figure
1007 * out which cpu is logically closest to our hot cache data.
1009 if (this_cpu == cpu)
1010 this_cpu = -1; /* Skip this_cpu opt if the same */
1012 for_each_domain(cpu, sd) {
1013 if (sd->flags & SD_WAKE_AFFINE) {
1014 cpumask_t domain_mask;
1015 int best_cpu;
1017 cpus_and(domain_mask, sd->span, *lowest_mask);
1019 best_cpu = pick_optimal_cpu(this_cpu,
1020 &domain_mask);
1021 if (best_cpu != -1)
1022 return best_cpu;
1027 * And finally, if there were no matches within the domains
1028 * just give the caller *something* to work with from the compatible
1029 * locations.
1031 return pick_optimal_cpu(this_cpu, lowest_mask);
1034 /* Will lock the rq it finds */
1035 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1037 struct rq *lowest_rq = NULL;
1038 int tries;
1039 int cpu;
1041 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1042 cpu = find_lowest_rq(task);
1044 if ((cpu == -1) || (cpu == rq->cpu))
1045 break;
1047 lowest_rq = cpu_rq(cpu);
1049 /* if the prio of this runqueue changed, try again */
1050 if (double_lock_balance(rq, lowest_rq)) {
1052 * We had to unlock the run queue. In
1053 * the mean time, task could have
1054 * migrated already or had its affinity changed.
1055 * Also make sure that it wasn't scheduled on its rq.
1057 if (unlikely(task_rq(task) != rq ||
1058 !cpu_isset(lowest_rq->cpu,
1059 task->cpus_allowed) ||
1060 task_running(rq, task) ||
1061 !task->se.on_rq)) {
1063 spin_unlock(&lowest_rq->lock);
1064 lowest_rq = NULL;
1065 break;
1069 /* If this rq is still suitable use it. */
1070 if (lowest_rq->rt.highest_prio > task->prio)
1071 break;
1073 /* try again */
1074 double_unlock_balance(rq, lowest_rq);
1075 lowest_rq = NULL;
1078 return lowest_rq;
1082 * If the current CPU has more than one RT task, see if the non
1083 * running task can migrate over to a CPU that is running a task
1084 * of lesser priority.
1086 static int push_rt_task(struct rq *rq)
1088 struct task_struct *next_task;
1089 struct rq *lowest_rq;
1090 int ret = 0;
1091 int paranoid = RT_MAX_TRIES;
1093 if (!rq->rt.overloaded)
1094 return 0;
1096 next_task = pick_next_highest_task_rt(rq, -1);
1097 if (!next_task)
1098 return 0;
1100 retry:
1101 if (unlikely(next_task == rq->curr)) {
1102 WARN_ON(1);
1103 return 0;
1107 * It's possible that the next_task slipped in of
1108 * higher priority than current. If that's the case
1109 * just reschedule current.
1111 if (unlikely(next_task->prio < rq->curr->prio)) {
1112 resched_task(rq->curr);
1113 return 0;
1116 /* We might release rq lock */
1117 get_task_struct(next_task);
1119 /* find_lock_lowest_rq locks the rq if found */
1120 lowest_rq = find_lock_lowest_rq(next_task, rq);
1121 if (!lowest_rq) {
1122 struct task_struct *task;
1124 * find lock_lowest_rq releases rq->lock
1125 * so it is possible that next_task has changed.
1126 * If it has, then try again.
1128 task = pick_next_highest_task_rt(rq, -1);
1129 if (unlikely(task != next_task) && task && paranoid--) {
1130 put_task_struct(next_task);
1131 next_task = task;
1132 goto retry;
1134 goto out;
1137 deactivate_task(rq, next_task, 0);
1138 set_task_cpu(next_task, lowest_rq->cpu);
1139 activate_task(lowest_rq, next_task, 0);
1141 resched_task(lowest_rq->curr);
1143 double_unlock_balance(rq, lowest_rq);
1145 ret = 1;
1146 out:
1147 put_task_struct(next_task);
1149 return ret;
1153 * TODO: Currently we just use the second highest prio task on
1154 * the queue, and stop when it can't migrate (or there's
1155 * no more RT tasks). There may be a case where a lower
1156 * priority RT task has a different affinity than the
1157 * higher RT task. In this case the lower RT task could
1158 * possibly be able to migrate where as the higher priority
1159 * RT task could not. We currently ignore this issue.
1160 * Enhancements are welcome!
1162 static void push_rt_tasks(struct rq *rq)
1164 /* push_rt_task will return true if it moved an RT */
1165 while (push_rt_task(rq))
1169 static int pull_rt_task(struct rq *this_rq)
1171 int this_cpu = this_rq->cpu, ret = 0, cpu;
1172 struct task_struct *p, *next;
1173 struct rq *src_rq;
1175 if (likely(!rt_overloaded(this_rq)))
1176 return 0;
1178 next = pick_next_task_rt(this_rq);
1180 for_each_cpu_mask_nr(cpu, this_rq->rd->rto_mask) {
1181 if (this_cpu == cpu)
1182 continue;
1184 src_rq = cpu_rq(cpu);
1186 * We can potentially drop this_rq's lock in
1187 * double_lock_balance, and another CPU could
1188 * steal our next task - hence we must cause
1189 * the caller to recalculate the next task
1190 * in that case:
1192 if (double_lock_balance(this_rq, src_rq)) {
1193 struct task_struct *old_next = next;
1195 next = pick_next_task_rt(this_rq);
1196 if (next != old_next)
1197 ret = 1;
1201 * Are there still pullable RT tasks?
1203 if (src_rq->rt.rt_nr_running <= 1)
1204 goto skip;
1206 p = pick_next_highest_task_rt(src_rq, this_cpu);
1209 * Do we have an RT task that preempts
1210 * the to-be-scheduled task?
1212 if (p && (!next || (p->prio < next->prio))) {
1213 WARN_ON(p == src_rq->curr);
1214 WARN_ON(!p->se.on_rq);
1217 * There's a chance that p is higher in priority
1218 * than what's currently running on its cpu.
1219 * This is just that p is wakeing up and hasn't
1220 * had a chance to schedule. We only pull
1221 * p if it is lower in priority than the
1222 * current task on the run queue or
1223 * this_rq next task is lower in prio than
1224 * the current task on that rq.
1226 if (p->prio < src_rq->curr->prio ||
1227 (next && next->prio < src_rq->curr->prio))
1228 goto skip;
1230 ret = 1;
1232 deactivate_task(src_rq, p, 0);
1233 set_task_cpu(p, this_cpu);
1234 activate_task(this_rq, p, 0);
1236 * We continue with the search, just in
1237 * case there's an even higher prio task
1238 * in another runqueue. (low likelyhood
1239 * but possible)
1241 * Update next so that we won't pick a task
1242 * on another cpu with a priority lower (or equal)
1243 * than the one we just picked.
1245 next = p;
1248 skip:
1249 double_unlock_balance(this_rq, src_rq);
1252 return ret;
1255 static void pre_schedule_rt(struct rq *rq, struct task_struct *prev)
1257 /* Try to pull RT tasks here if we lower this rq's prio */
1258 if (unlikely(rt_task(prev)) && rq->rt.highest_prio > prev->prio)
1259 pull_rt_task(rq);
1262 static void post_schedule_rt(struct rq *rq)
1265 * If we have more than one rt_task queued, then
1266 * see if we can push the other rt_tasks off to other CPUS.
1267 * Note we may release the rq lock, and since
1268 * the lock was owned by prev, we need to release it
1269 * first via finish_lock_switch and then reaquire it here.
1271 if (unlikely(rq->rt.overloaded)) {
1272 spin_lock_irq(&rq->lock);
1273 push_rt_tasks(rq);
1274 spin_unlock_irq(&rq->lock);
1279 * If we are not running and we are not going to reschedule soon, we should
1280 * try to push tasks away now
1282 static void task_wake_up_rt(struct rq *rq, struct task_struct *p)
1284 if (!task_running(rq, p) &&
1285 !test_tsk_need_resched(rq->curr) &&
1286 rq->rt.overloaded)
1287 push_rt_tasks(rq);
1290 static unsigned long
1291 load_balance_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1292 unsigned long max_load_move,
1293 struct sched_domain *sd, enum cpu_idle_type idle,
1294 int *all_pinned, int *this_best_prio)
1296 /* don't touch RT tasks */
1297 return 0;
1300 static int
1301 move_one_task_rt(struct rq *this_rq, int this_cpu, struct rq *busiest,
1302 struct sched_domain *sd, enum cpu_idle_type idle)
1304 /* don't touch RT tasks */
1305 return 0;
1308 static void set_cpus_allowed_rt(struct task_struct *p,
1309 const cpumask_t *new_mask)
1311 int weight = cpus_weight(*new_mask);
1313 BUG_ON(!rt_task(p));
1316 * Update the migration status of the RQ if we have an RT task
1317 * which is running AND changing its weight value.
1319 if (p->se.on_rq && (weight != p->rt.nr_cpus_allowed)) {
1320 struct rq *rq = task_rq(p);
1322 if ((p->rt.nr_cpus_allowed <= 1) && (weight > 1)) {
1323 rq->rt.rt_nr_migratory++;
1324 } else if ((p->rt.nr_cpus_allowed > 1) && (weight <= 1)) {
1325 BUG_ON(!rq->rt.rt_nr_migratory);
1326 rq->rt.rt_nr_migratory--;
1329 update_rt_migration(rq);
1332 p->cpus_allowed = *new_mask;
1333 p->rt.nr_cpus_allowed = weight;
1336 /* Assumes rq->lock is held */
1337 static void rq_online_rt(struct rq *rq)
1339 if (rq->rt.overloaded)
1340 rt_set_overload(rq);
1342 __enable_runtime(rq);
1344 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio);
1347 /* Assumes rq->lock is held */
1348 static void rq_offline_rt(struct rq *rq)
1350 if (rq->rt.overloaded)
1351 rt_clear_overload(rq);
1353 __disable_runtime(rq);
1355 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1359 * When switch from the rt queue, we bring ourselves to a position
1360 * that we might want to pull RT tasks from other runqueues.
1362 static void switched_from_rt(struct rq *rq, struct task_struct *p,
1363 int running)
1366 * If there are other RT tasks then we will reschedule
1367 * and the scheduling of the other RT tasks will handle
1368 * the balancing. But if we are the last RT task
1369 * we may need to handle the pulling of RT tasks
1370 * now.
1372 if (!rq->rt.rt_nr_running)
1373 pull_rt_task(rq);
1375 #endif /* CONFIG_SMP */
1378 * When switching a task to RT, we may overload the runqueue
1379 * with RT tasks. In this case we try to push them off to
1380 * other runqueues.
1382 static void switched_to_rt(struct rq *rq, struct task_struct *p,
1383 int running)
1385 int check_resched = 1;
1388 * If we are already running, then there's nothing
1389 * that needs to be done. But if we are not running
1390 * we may need to preempt the current running task.
1391 * If that current running task is also an RT task
1392 * then see if we can move to another run queue.
1394 if (!running) {
1395 #ifdef CONFIG_SMP
1396 if (rq->rt.overloaded && push_rt_task(rq) &&
1397 /* Don't resched if we changed runqueues */
1398 rq != task_rq(p))
1399 check_resched = 0;
1400 #endif /* CONFIG_SMP */
1401 if (check_resched && p->prio < rq->curr->prio)
1402 resched_task(rq->curr);
1407 * Priority of the task has changed. This may cause
1408 * us to initiate a push or pull.
1410 static void prio_changed_rt(struct rq *rq, struct task_struct *p,
1411 int oldprio, int running)
1413 if (running) {
1414 #ifdef CONFIG_SMP
1416 * If our priority decreases while running, we
1417 * may need to pull tasks to this runqueue.
1419 if (oldprio < p->prio)
1420 pull_rt_task(rq);
1422 * If there's a higher priority task waiting to run
1423 * then reschedule. Note, the above pull_rt_task
1424 * can release the rq lock and p could migrate.
1425 * Only reschedule if p is still on the same runqueue.
1427 if (p->prio > rq->rt.highest_prio && rq->curr == p)
1428 resched_task(p);
1429 #else
1430 /* For UP simply resched on drop of prio */
1431 if (oldprio < p->prio)
1432 resched_task(p);
1433 #endif /* CONFIG_SMP */
1434 } else {
1436 * This task is not running, but if it is
1437 * greater than the current running task
1438 * then reschedule.
1440 if (p->prio < rq->curr->prio)
1441 resched_task(rq->curr);
1445 static void watchdog(struct rq *rq, struct task_struct *p)
1447 unsigned long soft, hard;
1449 if (!p->signal)
1450 return;
1452 soft = p->signal->rlim[RLIMIT_RTTIME].rlim_cur;
1453 hard = p->signal->rlim[RLIMIT_RTTIME].rlim_max;
1455 if (soft != RLIM_INFINITY) {
1456 unsigned long next;
1458 p->rt.timeout++;
1459 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1460 if (p->rt.timeout > next)
1461 p->it_sched_expires = p->se.sum_exec_runtime;
1465 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1467 update_curr_rt(rq);
1469 watchdog(rq, p);
1472 * RR tasks need a special form of timeslice management.
1473 * FIFO tasks have no timeslices.
1475 if (p->policy != SCHED_RR)
1476 return;
1478 if (--p->rt.time_slice)
1479 return;
1481 p->rt.time_slice = DEF_TIMESLICE;
1484 * Requeue to the end of queue if we are not the only element
1485 * on the queue:
1487 if (p->rt.run_list.prev != p->rt.run_list.next) {
1488 requeue_task_rt(rq, p, 0);
1489 set_tsk_need_resched(p);
1493 static void set_curr_task_rt(struct rq *rq)
1495 struct task_struct *p = rq->curr;
1497 p->se.exec_start = rq->clock;
1500 static const struct sched_class rt_sched_class = {
1501 .next = &fair_sched_class,
1502 .enqueue_task = enqueue_task_rt,
1503 .dequeue_task = dequeue_task_rt,
1504 .yield_task = yield_task_rt,
1505 #ifdef CONFIG_SMP
1506 .select_task_rq = select_task_rq_rt,
1507 #endif /* CONFIG_SMP */
1509 .check_preempt_curr = check_preempt_curr_rt,
1511 .pick_next_task = pick_next_task_rt,
1512 .put_prev_task = put_prev_task_rt,
1514 #ifdef CONFIG_SMP
1515 .load_balance = load_balance_rt,
1516 .move_one_task = move_one_task_rt,
1517 .set_cpus_allowed = set_cpus_allowed_rt,
1518 .rq_online = rq_online_rt,
1519 .rq_offline = rq_offline_rt,
1520 .pre_schedule = pre_schedule_rt,
1521 .post_schedule = post_schedule_rt,
1522 .task_wake_up = task_wake_up_rt,
1523 .switched_from = switched_from_rt,
1524 #endif
1526 .set_curr_task = set_curr_task_rt,
1527 .task_tick = task_tick_rt,
1529 .prio_changed = prio_changed_rt,
1530 .switched_to = switched_to_rt,
1533 #ifdef CONFIG_SCHED_DEBUG
1534 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
1536 static void print_rt_stats(struct seq_file *m, int cpu)
1538 struct rt_rq *rt_rq;
1540 rcu_read_lock();
1541 for_each_leaf_rt_rq(rt_rq, cpu_rq(cpu))
1542 print_rt_rq(m, cpu, rt_rq);
1543 rcu_read_unlock();
1545 #endif /* CONFIG_SCHED_DEBUG */