2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
8 #include <linux/slab.h>
10 int sched_rr_timeslice = RR_TIMESLICE;
12 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
14 struct rt_bandwidth def_rt_bandwidth;
16 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
18 struct rt_bandwidth *rt_b =
19 container_of(timer, struct rt_bandwidth, rt_period_timer);
25 now = hrtimer_cb_get_time(timer);
26 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
31 idle = do_sched_rt_period_timer(rt_b, overrun);
34 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
37 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
39 rt_b->rt_period = ns_to_ktime(period);
40 rt_b->rt_runtime = runtime;
42 raw_spin_lock_init(&rt_b->rt_runtime_lock);
44 hrtimer_init(&rt_b->rt_period_timer,
45 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
46 rt_b->rt_period_timer.function = sched_rt_period_timer;
49 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
51 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
54 if (hrtimer_active(&rt_b->rt_period_timer))
57 raw_spin_lock(&rt_b->rt_runtime_lock);
58 start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period);
59 raw_spin_unlock(&rt_b->rt_runtime_lock);
62 void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
64 struct rt_prio_array *array;
67 array = &rt_rq->active;
68 for (i = 0; i < MAX_RT_PRIO; i++) {
69 INIT_LIST_HEAD(array->queue + i);
70 __clear_bit(i, array->bitmap);
72 /* delimiter for bitsearch: */
73 __set_bit(MAX_RT_PRIO, array->bitmap);
75 #if defined CONFIG_SMP
76 rt_rq->highest_prio.curr = MAX_RT_PRIO;
77 rt_rq->highest_prio.next = MAX_RT_PRIO;
78 rt_rq->rt_nr_migratory = 0;
79 rt_rq->overloaded = 0;
80 plist_head_init(&rt_rq->pushable_tasks);
84 rt_rq->rt_throttled = 0;
85 rt_rq->rt_runtime = 0;
86 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
89 #ifdef CONFIG_RT_GROUP_SCHED
90 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
92 hrtimer_cancel(&rt_b->rt_period_timer);
95 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
97 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
99 #ifdef CONFIG_SCHED_DEBUG
100 WARN_ON_ONCE(!rt_entity_is_task(rt_se));
102 return container_of(rt_se, struct task_struct, rt);
105 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
110 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
115 void free_rt_sched_group(struct task_group *tg)
120 destroy_rt_bandwidth(&tg->rt_bandwidth);
122 for_each_possible_cpu(i) {
133 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
134 struct sched_rt_entity *rt_se, int cpu,
135 struct sched_rt_entity *parent)
137 struct rq *rq = cpu_rq(cpu);
139 rt_rq->highest_prio.curr = MAX_RT_PRIO;
140 rt_rq->rt_nr_boosted = 0;
144 tg->rt_rq[cpu] = rt_rq;
145 tg->rt_se[cpu] = rt_se;
151 rt_se->rt_rq = &rq->rt;
153 rt_se->rt_rq = parent->my_q;
156 rt_se->parent = parent;
157 INIT_LIST_HEAD(&rt_se->run_list);
160 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
163 struct sched_rt_entity *rt_se;
166 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
169 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
173 init_rt_bandwidth(&tg->rt_bandwidth,
174 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
176 for_each_possible_cpu(i) {
177 rt_rq = kzalloc_node(sizeof(struct rt_rq),
178 GFP_KERNEL, cpu_to_node(i));
182 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
183 GFP_KERNEL, cpu_to_node(i));
187 init_rt_rq(rt_rq, cpu_rq(i));
188 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
189 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
200 #else /* CONFIG_RT_GROUP_SCHED */
202 #define rt_entity_is_task(rt_se) (1)
204 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
206 return container_of(rt_se, struct task_struct, rt);
209 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
211 return container_of(rt_rq, struct rq, rt);
214 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
216 struct task_struct *p = rt_task_of(rt_se);
217 struct rq *rq = task_rq(p);
222 void free_rt_sched_group(struct task_group *tg) { }
224 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
228 #endif /* CONFIG_RT_GROUP_SCHED */
232 static int pull_rt_task(struct rq *this_rq);
234 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
236 /* Try to pull RT tasks here if we lower this rq's prio */
237 return rq->rt.highest_prio.curr > prev->prio;
240 static inline int rt_overloaded(struct rq *rq)
242 return atomic_read(&rq->rd->rto_count);
245 static inline void rt_set_overload(struct rq *rq)
250 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
252 * Make sure the mask is visible before we set
253 * the overload count. That is checked to determine
254 * if we should look at the mask. It would be a shame
255 * if we looked at the mask, but the mask was not
258 * Matched by the barrier in pull_rt_task().
261 atomic_inc(&rq->rd->rto_count);
264 static inline void rt_clear_overload(struct rq *rq)
269 /* the order here really doesn't matter */
270 atomic_dec(&rq->rd->rto_count);
271 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
274 static void update_rt_migration(struct rt_rq *rt_rq)
276 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
277 if (!rt_rq->overloaded) {
278 rt_set_overload(rq_of_rt_rq(rt_rq));
279 rt_rq->overloaded = 1;
281 } else if (rt_rq->overloaded) {
282 rt_clear_overload(rq_of_rt_rq(rt_rq));
283 rt_rq->overloaded = 0;
287 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
289 struct task_struct *p;
291 if (!rt_entity_is_task(rt_se))
294 p = rt_task_of(rt_se);
295 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
297 rt_rq->rt_nr_total++;
298 if (p->nr_cpus_allowed > 1)
299 rt_rq->rt_nr_migratory++;
301 update_rt_migration(rt_rq);
304 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
306 struct task_struct *p;
308 if (!rt_entity_is_task(rt_se))
311 p = rt_task_of(rt_se);
312 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
314 rt_rq->rt_nr_total--;
315 if (p->nr_cpus_allowed > 1)
316 rt_rq->rt_nr_migratory--;
318 update_rt_migration(rt_rq);
321 static inline int has_pushable_tasks(struct rq *rq)
323 return !plist_head_empty(&rq->rt.pushable_tasks);
326 static inline void set_post_schedule(struct rq *rq)
329 * We detect this state here so that we can avoid taking the RQ
330 * lock again later if there is no need to push
332 rq->post_schedule = has_pushable_tasks(rq);
335 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
337 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
338 plist_node_init(&p->pushable_tasks, p->prio);
339 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
341 /* Update the highest prio pushable task */
342 if (p->prio < rq->rt.highest_prio.next)
343 rq->rt.highest_prio.next = p->prio;
346 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
348 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
350 /* Update the new highest prio pushable task */
351 if (has_pushable_tasks(rq)) {
352 p = plist_first_entry(&rq->rt.pushable_tasks,
353 struct task_struct, pushable_tasks);
354 rq->rt.highest_prio.next = p->prio;
356 rq->rt.highest_prio.next = MAX_RT_PRIO;
361 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
365 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
370 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
375 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
379 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
384 static inline int pull_rt_task(struct rq *this_rq)
389 static inline void set_post_schedule(struct rq *rq)
392 #endif /* CONFIG_SMP */
394 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
396 return !list_empty(&rt_se->run_list);
399 #ifdef CONFIG_RT_GROUP_SCHED
401 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
406 return rt_rq->rt_runtime;
409 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
411 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
414 typedef struct task_group *rt_rq_iter_t;
416 static inline struct task_group *next_task_group(struct task_group *tg)
419 tg = list_entry_rcu(tg->list.next,
420 typeof(struct task_group), list);
421 } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
423 if (&tg->list == &task_groups)
429 #define for_each_rt_rq(rt_rq, iter, rq) \
430 for (iter = container_of(&task_groups, typeof(*iter), list); \
431 (iter = next_task_group(iter)) && \
432 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
434 #define for_each_sched_rt_entity(rt_se) \
435 for (; rt_se; rt_se = rt_se->parent)
437 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
442 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
443 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
445 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
447 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
448 struct sched_rt_entity *rt_se;
450 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
452 rt_se = rt_rq->tg->rt_se[cpu];
454 if (rt_rq->rt_nr_running) {
455 if (rt_se && !on_rt_rq(rt_se))
456 enqueue_rt_entity(rt_se, false);
457 if (rt_rq->highest_prio.curr < curr->prio)
462 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
464 struct sched_rt_entity *rt_se;
465 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
467 rt_se = rt_rq->tg->rt_se[cpu];
469 if (rt_se && on_rt_rq(rt_se))
470 dequeue_rt_entity(rt_se);
473 static int rt_se_boosted(struct sched_rt_entity *rt_se)
475 struct rt_rq *rt_rq = group_rt_rq(rt_se);
476 struct task_struct *p;
479 return !!rt_rq->rt_nr_boosted;
481 p = rt_task_of(rt_se);
482 return p->prio != p->normal_prio;
486 static inline const struct cpumask *sched_rt_period_mask(void)
488 return this_rq()->rd->span;
491 static inline const struct cpumask *sched_rt_period_mask(void)
493 return cpu_online_mask;
498 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
500 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
503 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
505 return &rt_rq->tg->rt_bandwidth;
508 #else /* !CONFIG_RT_GROUP_SCHED */
510 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
512 return rt_rq->rt_runtime;
515 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
517 return ktime_to_ns(def_rt_bandwidth.rt_period);
520 typedef struct rt_rq *rt_rq_iter_t;
522 #define for_each_rt_rq(rt_rq, iter, rq) \
523 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
525 #define for_each_sched_rt_entity(rt_se) \
526 for (; rt_se; rt_se = NULL)
528 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
533 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
535 if (rt_rq->rt_nr_running)
536 resched_task(rq_of_rt_rq(rt_rq)->curr);
539 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
543 static inline const struct cpumask *sched_rt_period_mask(void)
545 return cpu_online_mask;
549 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
551 return &cpu_rq(cpu)->rt;
554 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
556 return &def_rt_bandwidth;
559 #endif /* CONFIG_RT_GROUP_SCHED */
561 bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
563 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
565 return (hrtimer_active(&rt_b->rt_period_timer) ||
566 rt_rq->rt_time < rt_b->rt_runtime);
571 * We ran out of runtime, see if we can borrow some from our neighbours.
573 static int do_balance_runtime(struct rt_rq *rt_rq)
575 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
576 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
577 int i, weight, more = 0;
580 weight = cpumask_weight(rd->span);
582 raw_spin_lock(&rt_b->rt_runtime_lock);
583 rt_period = ktime_to_ns(rt_b->rt_period);
584 for_each_cpu(i, rd->span) {
585 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
591 raw_spin_lock(&iter->rt_runtime_lock);
593 * Either all rqs have inf runtime and there's nothing to steal
594 * or __disable_runtime() below sets a specific rq to inf to
595 * indicate its been disabled and disalow stealing.
597 if (iter->rt_runtime == RUNTIME_INF)
601 * From runqueues with spare time, take 1/n part of their
602 * spare time, but no more than our period.
604 diff = iter->rt_runtime - iter->rt_time;
606 diff = div_u64((u64)diff, weight);
607 if (rt_rq->rt_runtime + diff > rt_period)
608 diff = rt_period - rt_rq->rt_runtime;
609 iter->rt_runtime -= diff;
610 rt_rq->rt_runtime += diff;
612 if (rt_rq->rt_runtime == rt_period) {
613 raw_spin_unlock(&iter->rt_runtime_lock);
618 raw_spin_unlock(&iter->rt_runtime_lock);
620 raw_spin_unlock(&rt_b->rt_runtime_lock);
626 * Ensure this RQ takes back all the runtime it lend to its neighbours.
628 static void __disable_runtime(struct rq *rq)
630 struct root_domain *rd = rq->rd;
634 if (unlikely(!scheduler_running))
637 for_each_rt_rq(rt_rq, iter, rq) {
638 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
642 raw_spin_lock(&rt_b->rt_runtime_lock);
643 raw_spin_lock(&rt_rq->rt_runtime_lock);
645 * Either we're all inf and nobody needs to borrow, or we're
646 * already disabled and thus have nothing to do, or we have
647 * exactly the right amount of runtime to take out.
649 if (rt_rq->rt_runtime == RUNTIME_INF ||
650 rt_rq->rt_runtime == rt_b->rt_runtime)
652 raw_spin_unlock(&rt_rq->rt_runtime_lock);
655 * Calculate the difference between what we started out with
656 * and what we current have, that's the amount of runtime
657 * we lend and now have to reclaim.
659 want = rt_b->rt_runtime - rt_rq->rt_runtime;
662 * Greedy reclaim, take back as much as we can.
664 for_each_cpu(i, rd->span) {
665 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
669 * Can't reclaim from ourselves or disabled runqueues.
671 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
674 raw_spin_lock(&iter->rt_runtime_lock);
676 diff = min_t(s64, iter->rt_runtime, want);
677 iter->rt_runtime -= diff;
680 iter->rt_runtime -= want;
683 raw_spin_unlock(&iter->rt_runtime_lock);
689 raw_spin_lock(&rt_rq->rt_runtime_lock);
691 * We cannot be left wanting - that would mean some runtime
692 * leaked out of the system.
697 * Disable all the borrow logic by pretending we have inf
698 * runtime - in which case borrowing doesn't make sense.
700 rt_rq->rt_runtime = RUNTIME_INF;
701 rt_rq->rt_throttled = 0;
702 raw_spin_unlock(&rt_rq->rt_runtime_lock);
703 raw_spin_unlock(&rt_b->rt_runtime_lock);
707 static void __enable_runtime(struct rq *rq)
712 if (unlikely(!scheduler_running))
716 * Reset each runqueue's bandwidth settings
718 for_each_rt_rq(rt_rq, iter, rq) {
719 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
721 raw_spin_lock(&rt_b->rt_runtime_lock);
722 raw_spin_lock(&rt_rq->rt_runtime_lock);
723 rt_rq->rt_runtime = rt_b->rt_runtime;
725 rt_rq->rt_throttled = 0;
726 raw_spin_unlock(&rt_rq->rt_runtime_lock);
727 raw_spin_unlock(&rt_b->rt_runtime_lock);
731 static int balance_runtime(struct rt_rq *rt_rq)
735 if (!sched_feat(RT_RUNTIME_SHARE))
738 if (rt_rq->rt_time > rt_rq->rt_runtime) {
739 raw_spin_unlock(&rt_rq->rt_runtime_lock);
740 more = do_balance_runtime(rt_rq);
741 raw_spin_lock(&rt_rq->rt_runtime_lock);
746 #else /* !CONFIG_SMP */
747 static inline int balance_runtime(struct rt_rq *rt_rq)
751 #endif /* CONFIG_SMP */
753 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
755 int i, idle = 1, throttled = 0;
756 const struct cpumask *span;
758 span = sched_rt_period_mask();
759 #ifdef CONFIG_RT_GROUP_SCHED
761 * FIXME: isolated CPUs should really leave the root task group,
762 * whether they are isolcpus or were isolated via cpusets, lest
763 * the timer run on a CPU which does not service all runqueues,
764 * potentially leaving other CPUs indefinitely throttled. If
765 * isolation is really required, the user will turn the throttle
766 * off to kill the perturbations it causes anyway. Meanwhile,
767 * this maintains functionality for boot and/or troubleshooting.
769 if (rt_b == &root_task_group.rt_bandwidth)
770 span = cpu_online_mask;
772 for_each_cpu(i, span) {
774 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
775 struct rq *rq = rq_of_rt_rq(rt_rq);
777 raw_spin_lock(&rq->lock);
778 if (rt_rq->rt_time) {
781 raw_spin_lock(&rt_rq->rt_runtime_lock);
782 if (rt_rq->rt_throttled)
783 balance_runtime(rt_rq);
784 runtime = rt_rq->rt_runtime;
785 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
786 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
787 rt_rq->rt_throttled = 0;
791 * Force a clock update if the CPU was idle,
792 * lest wakeup -> unthrottle time accumulate.
794 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
795 rq->skip_clock_update = -1;
797 if (rt_rq->rt_time || rt_rq->rt_nr_running)
799 raw_spin_unlock(&rt_rq->rt_runtime_lock);
800 } else if (rt_rq->rt_nr_running) {
802 if (!rt_rq_throttled(rt_rq))
805 if (rt_rq->rt_throttled)
809 sched_rt_rq_enqueue(rt_rq);
810 raw_spin_unlock(&rq->lock);
813 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
819 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
821 #ifdef CONFIG_RT_GROUP_SCHED
822 struct rt_rq *rt_rq = group_rt_rq(rt_se);
825 return rt_rq->highest_prio.curr;
828 return rt_task_of(rt_se)->prio;
831 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
833 u64 runtime = sched_rt_runtime(rt_rq);
835 if (rt_rq->rt_throttled)
836 return rt_rq_throttled(rt_rq);
838 if (runtime >= sched_rt_period(rt_rq))
841 balance_runtime(rt_rq);
842 runtime = sched_rt_runtime(rt_rq);
843 if (runtime == RUNTIME_INF)
846 if (rt_rq->rt_time > runtime) {
847 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
850 * Don't actually throttle groups that have no runtime assigned
851 * but accrue some time due to boosting.
853 if (likely(rt_b->rt_runtime)) {
854 static bool once = false;
856 rt_rq->rt_throttled = 1;
860 printk_sched("sched: RT throttling activated\n");
864 * In case we did anyway, make it go away,
865 * replenishment is a joke, since it will replenish us
871 if (rt_rq_throttled(rt_rq)) {
872 sched_rt_rq_dequeue(rt_rq);
881 * Update the current task's runtime statistics. Skip current tasks that
882 * are not in our scheduling class.
884 static void update_curr_rt(struct rq *rq)
886 struct task_struct *curr = rq->curr;
887 struct sched_rt_entity *rt_se = &curr->rt;
888 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
891 if (curr->sched_class != &rt_sched_class)
894 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
895 if (unlikely((s64)delta_exec <= 0))
898 schedstat_set(curr->se.statistics.exec_max,
899 max(curr->se.statistics.exec_max, delta_exec));
901 curr->se.sum_exec_runtime += delta_exec;
902 account_group_exec_runtime(curr, delta_exec);
904 curr->se.exec_start = rq_clock_task(rq);
905 cpuacct_charge(curr, delta_exec);
907 sched_rt_avg_update(rq, delta_exec);
909 if (!rt_bandwidth_enabled())
912 for_each_sched_rt_entity(rt_se) {
913 rt_rq = rt_rq_of_se(rt_se);
915 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
916 raw_spin_lock(&rt_rq->rt_runtime_lock);
917 rt_rq->rt_time += delta_exec;
918 if (sched_rt_runtime_exceeded(rt_rq))
920 raw_spin_unlock(&rt_rq->rt_runtime_lock);
925 #if defined CONFIG_SMP
928 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
930 struct rq *rq = rq_of_rt_rq(rt_rq);
932 #ifdef CONFIG_RT_GROUP_SCHED
934 * Change rq's cpupri only if rt_rq is the top queue.
936 if (&rq->rt != rt_rq)
939 if (rq->online && prio < prev_prio)
940 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
944 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
946 struct rq *rq = rq_of_rt_rq(rt_rq);
948 #ifdef CONFIG_RT_GROUP_SCHED
950 * Change rq's cpupri only if rt_rq is the top queue.
952 if (&rq->rt != rt_rq)
955 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
956 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
959 #else /* CONFIG_SMP */
962 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
964 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
966 #endif /* CONFIG_SMP */
968 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
970 inc_rt_prio(struct rt_rq *rt_rq, int prio)
972 int prev_prio = rt_rq->highest_prio.curr;
974 if (prio < prev_prio)
975 rt_rq->highest_prio.curr = prio;
977 inc_rt_prio_smp(rt_rq, prio, prev_prio);
981 dec_rt_prio(struct rt_rq *rt_rq, int prio)
983 int prev_prio = rt_rq->highest_prio.curr;
985 if (rt_rq->rt_nr_running) {
987 WARN_ON(prio < prev_prio);
990 * This may have been our highest task, and therefore
991 * we may have some recomputation to do
993 if (prio == prev_prio) {
994 struct rt_prio_array *array = &rt_rq->active;
996 rt_rq->highest_prio.curr =
997 sched_find_first_bit(array->bitmap);
1001 rt_rq->highest_prio.curr = MAX_RT_PRIO;
1003 dec_rt_prio_smp(rt_rq, prio, prev_prio);
1008 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1009 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1011 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1013 #ifdef CONFIG_RT_GROUP_SCHED
1016 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1018 if (rt_se_boosted(rt_se))
1019 rt_rq->rt_nr_boosted++;
1022 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1026 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1028 if (rt_se_boosted(rt_se))
1029 rt_rq->rt_nr_boosted--;
1031 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1034 #else /* CONFIG_RT_GROUP_SCHED */
1037 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1039 start_rt_bandwidth(&def_rt_bandwidth);
1043 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1045 #endif /* CONFIG_RT_GROUP_SCHED */
1048 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1050 int prio = rt_se_prio(rt_se);
1052 WARN_ON(!rt_prio(prio));
1053 rt_rq->rt_nr_running++;
1055 inc_rt_prio(rt_rq, prio);
1056 inc_rt_migration(rt_se, rt_rq);
1057 inc_rt_group(rt_se, rt_rq);
1061 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1063 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1064 WARN_ON(!rt_rq->rt_nr_running);
1065 rt_rq->rt_nr_running--;
1067 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1068 dec_rt_migration(rt_se, rt_rq);
1069 dec_rt_group(rt_se, rt_rq);
1072 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
1074 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1075 struct rt_prio_array *array = &rt_rq->active;
1076 struct rt_rq *group_rq = group_rt_rq(rt_se);
1077 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1080 * Don't enqueue the group if its throttled, or when empty.
1081 * The latter is a consequence of the former when a child group
1082 * get throttled and the current group doesn't have any other
1085 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
1089 list_add(&rt_se->run_list, queue);
1091 list_add_tail(&rt_se->run_list, queue);
1092 __set_bit(rt_se_prio(rt_se), array->bitmap);
1094 inc_rt_tasks(rt_se, rt_rq);
1097 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
1099 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1100 struct rt_prio_array *array = &rt_rq->active;
1102 list_del_init(&rt_se->run_list);
1103 if (list_empty(array->queue + rt_se_prio(rt_se)))
1104 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1106 dec_rt_tasks(rt_se, rt_rq);
1110 * Because the prio of an upper entry depends on the lower
1111 * entries, we must remove entries top - down.
1113 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
1115 struct sched_rt_entity *back = NULL;
1117 for_each_sched_rt_entity(rt_se) {
1122 for (rt_se = back; rt_se; rt_se = rt_se->back) {
1123 if (on_rt_rq(rt_se))
1124 __dequeue_rt_entity(rt_se);
1128 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
1130 dequeue_rt_stack(rt_se);
1131 for_each_sched_rt_entity(rt_se)
1132 __enqueue_rt_entity(rt_se, head);
1135 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
1137 dequeue_rt_stack(rt_se);
1139 for_each_sched_rt_entity(rt_se) {
1140 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1142 if (rt_rq && rt_rq->rt_nr_running)
1143 __enqueue_rt_entity(rt_se, false);
1148 * Adding/removing a task to/from a priority array:
1151 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1153 struct sched_rt_entity *rt_se = &p->rt;
1155 if (flags & ENQUEUE_WAKEUP)
1158 enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
1160 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1161 enqueue_pushable_task(rq, p);
1166 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1168 struct sched_rt_entity *rt_se = &p->rt;
1171 dequeue_rt_entity(rt_se);
1173 dequeue_pushable_task(rq, p);
1179 * Put task to the head or the end of the run list without the overhead of
1180 * dequeue followed by enqueue.
1183 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1185 if (on_rt_rq(rt_se)) {
1186 struct rt_prio_array *array = &rt_rq->active;
1187 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1190 list_move(&rt_se->run_list, queue);
1192 list_move_tail(&rt_se->run_list, queue);
1196 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1198 struct sched_rt_entity *rt_se = &p->rt;
1199 struct rt_rq *rt_rq;
1201 for_each_sched_rt_entity(rt_se) {
1202 rt_rq = rt_rq_of_se(rt_se);
1203 requeue_rt_entity(rt_rq, rt_se, head);
1207 static void yield_task_rt(struct rq *rq)
1209 requeue_task_rt(rq, rq->curr, 0);
1213 static int find_lowest_rq(struct task_struct *task);
1216 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1218 struct task_struct *curr;
1221 if (p->nr_cpus_allowed == 1)
1224 /* For anything but wake ups, just return the task_cpu */
1225 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1231 curr = ACCESS_ONCE(rq->curr); /* unlocked access */
1234 * If the current task on @p's runqueue is an RT task, then
1235 * try to see if we can wake this RT task up on another
1236 * runqueue. Otherwise simply start this RT task
1237 * on its current runqueue.
1239 * We want to avoid overloading runqueues. If the woken
1240 * task is a higher priority, then it will stay on this CPU
1241 * and the lower prio task should be moved to another CPU.
1242 * Even though this will probably make the lower prio task
1243 * lose its cache, we do not want to bounce a higher task
1244 * around just because it gave up its CPU, perhaps for a
1247 * For equal prio tasks, we just let the scheduler sort it out.
1249 * Otherwise, just let it ride on the affined RQ and the
1250 * post-schedule router will push the preempted task away
1252 * This test is optimistic, if we get it wrong the load-balancer
1253 * will have to sort it out.
1255 if (curr && unlikely(rt_task(curr)) &&
1256 (curr->nr_cpus_allowed < 2 ||
1257 curr->prio <= p->prio)) {
1258 int target = find_lowest_rq(p);
1269 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1271 if (rq->curr->nr_cpus_allowed == 1)
1274 if (p->nr_cpus_allowed != 1
1275 && cpupri_find(&rq->rd->cpupri, p, NULL))
1278 if (!cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1282 * There appears to be other cpus that can accept
1283 * current and none to run 'p', so lets reschedule
1284 * to try and push current away:
1286 requeue_task_rt(rq, p, 1);
1287 resched_task(rq->curr);
1290 #endif /* CONFIG_SMP */
1293 * Preempt the current task with a newly woken task if needed:
1295 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1297 if (p->prio < rq->curr->prio) {
1298 resched_task(rq->curr);
1306 * - the newly woken task is of equal priority to the current task
1307 * - the newly woken task is non-migratable while current is migratable
1308 * - current will be preempted on the next reschedule
1310 * we should check to see if current can readily move to a different
1311 * cpu. If so, we will reschedule to allow the push logic to try
1312 * to move current somewhere else, making room for our non-migratable
1315 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1316 check_preempt_equal_prio(rq, p);
1320 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1321 struct rt_rq *rt_rq)
1323 struct rt_prio_array *array = &rt_rq->active;
1324 struct sched_rt_entity *next = NULL;
1325 struct list_head *queue;
1328 idx = sched_find_first_bit(array->bitmap);
1329 BUG_ON(idx >= MAX_RT_PRIO);
1331 queue = array->queue + idx;
1332 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1337 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1339 struct sched_rt_entity *rt_se;
1340 struct task_struct *p;
1341 struct rt_rq *rt_rq = &rq->rt;
1344 rt_se = pick_next_rt_entity(rq, rt_rq);
1346 rt_rq = group_rt_rq(rt_se);
1349 p = rt_task_of(rt_se);
1350 p->se.exec_start = rq_clock_task(rq);
1355 static struct task_struct *
1356 pick_next_task_rt(struct rq *rq, struct task_struct *prev)
1358 struct task_struct *p;
1359 struct rt_rq *rt_rq = &rq->rt;
1361 if (need_pull_rt_task(rq, prev)) {
1364 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1365 * means a dl task can slip in, in which case we need to
1366 * re-start task selection.
1368 if (unlikely(rq->dl.dl_nr_running))
1373 * We may dequeue prev's rt_rq in put_prev_task().
1374 * So, we update time before rt_nr_running check.
1376 if (prev->sched_class == &rt_sched_class)
1379 if (!rt_rq->rt_nr_running)
1382 if (rt_rq_throttled(rt_rq))
1385 put_prev_task(rq, prev);
1387 p = _pick_next_task_rt(rq);
1389 /* The running task is never eligible for pushing */
1391 dequeue_pushable_task(rq, p);
1393 set_post_schedule(rq);
1398 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1403 * The previous task needs to be made eligible for pushing
1404 * if it is still active
1406 if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1407 enqueue_pushable_task(rq, p);
1412 /* Only try algorithms three times */
1413 #define RT_MAX_TRIES 3
1415 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1417 if (!task_running(rq, p) &&
1418 cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
1424 * Return the highest pushable rq's task, which is suitable to be executed
1425 * on the cpu, NULL otherwise
1427 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1429 struct plist_head *head = &rq->rt.pushable_tasks;
1430 struct task_struct *p;
1432 if (!has_pushable_tasks(rq))
1435 plist_for_each_entry(p, head, pushable_tasks) {
1436 if (pick_rt_task(rq, p, cpu))
1443 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1445 static int find_lowest_rq(struct task_struct *task)
1447 struct sched_domain *sd;
1448 struct cpumask *lowest_mask = __get_cpu_var(local_cpu_mask);
1449 int this_cpu = smp_processor_id();
1450 int cpu = task_cpu(task);
1452 /* Make sure the mask is initialized first */
1453 if (unlikely(!lowest_mask))
1456 if (task->nr_cpus_allowed == 1)
1457 return -1; /* No other targets possible */
1459 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1460 return -1; /* No targets found */
1463 * At this point we have built a mask of cpus representing the
1464 * lowest priority tasks in the system. Now we want to elect
1465 * the best one based on our affinity and topology.
1467 * We prioritize the last cpu that the task executed on since
1468 * it is most likely cache-hot in that location.
1470 if (cpumask_test_cpu(cpu, lowest_mask))
1474 * Otherwise, we consult the sched_domains span maps to figure
1475 * out which cpu is logically closest to our hot cache data.
1477 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1478 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1481 for_each_domain(cpu, sd) {
1482 if (sd->flags & SD_WAKE_AFFINE) {
1486 * "this_cpu" is cheaper to preempt than a
1489 if (this_cpu != -1 &&
1490 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1495 best_cpu = cpumask_first_and(lowest_mask,
1496 sched_domain_span(sd));
1497 if (best_cpu < nr_cpu_ids) {
1506 * And finally, if there were no matches within the domains
1507 * just give the caller *something* to work with from the compatible
1513 cpu = cpumask_any(lowest_mask);
1514 if (cpu < nr_cpu_ids)
1519 /* Will lock the rq it finds */
1520 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1522 struct rq *lowest_rq = NULL;
1526 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1527 cpu = find_lowest_rq(task);
1529 if ((cpu == -1) || (cpu == rq->cpu))
1532 lowest_rq = cpu_rq(cpu);
1534 /* if the prio of this runqueue changed, try again */
1535 if (double_lock_balance(rq, lowest_rq)) {
1537 * We had to unlock the run queue. In
1538 * the mean time, task could have
1539 * migrated already or had its affinity changed.
1540 * Also make sure that it wasn't scheduled on its rq.
1542 if (unlikely(task_rq(task) != rq ||
1543 !cpumask_test_cpu(lowest_rq->cpu,
1544 tsk_cpus_allowed(task)) ||
1545 task_running(rq, task) ||
1548 double_unlock_balance(rq, lowest_rq);
1554 /* If this rq is still suitable use it. */
1555 if (lowest_rq->rt.highest_prio.curr > task->prio)
1559 double_unlock_balance(rq, lowest_rq);
1566 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1568 struct task_struct *p;
1570 if (!has_pushable_tasks(rq))
1573 p = plist_first_entry(&rq->rt.pushable_tasks,
1574 struct task_struct, pushable_tasks);
1576 BUG_ON(rq->cpu != task_cpu(p));
1577 BUG_ON(task_current(rq, p));
1578 BUG_ON(p->nr_cpus_allowed <= 1);
1581 BUG_ON(!rt_task(p));
1587 * If the current CPU has more than one RT task, see if the non
1588 * running task can migrate over to a CPU that is running a task
1589 * of lesser priority.
1591 static int push_rt_task(struct rq *rq)
1593 struct task_struct *next_task;
1594 struct rq *lowest_rq;
1597 if (!rq->rt.overloaded)
1600 next_task = pick_next_pushable_task(rq);
1605 if (unlikely(next_task == rq->curr)) {
1611 * It's possible that the next_task slipped in of
1612 * higher priority than current. If that's the case
1613 * just reschedule current.
1615 if (unlikely(next_task->prio < rq->curr->prio)) {
1616 resched_task(rq->curr);
1620 /* We might release rq lock */
1621 get_task_struct(next_task);
1623 /* find_lock_lowest_rq locks the rq if found */
1624 lowest_rq = find_lock_lowest_rq(next_task, rq);
1626 struct task_struct *task;
1628 * find_lock_lowest_rq releases rq->lock
1629 * so it is possible that next_task has migrated.
1631 * We need to make sure that the task is still on the same
1632 * run-queue and is also still the next task eligible for
1635 task = pick_next_pushable_task(rq);
1636 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1638 * The task hasn't migrated, and is still the next
1639 * eligible task, but we failed to find a run-queue
1640 * to push it to. Do not retry in this case, since
1641 * other cpus will pull from us when ready.
1647 /* No more tasks, just exit */
1651 * Something has shifted, try again.
1653 put_task_struct(next_task);
1658 deactivate_task(rq, next_task, 0);
1659 set_task_cpu(next_task, lowest_rq->cpu);
1660 activate_task(lowest_rq, next_task, 0);
1663 resched_task(lowest_rq->curr);
1665 double_unlock_balance(rq, lowest_rq);
1668 put_task_struct(next_task);
1673 static void push_rt_tasks(struct rq *rq)
1675 /* push_rt_task will return true if it moved an RT */
1676 while (push_rt_task(rq))
1680 static int pull_rt_task(struct rq *this_rq)
1682 int this_cpu = this_rq->cpu, ret = 0, cpu;
1683 struct task_struct *p;
1686 if (likely(!rt_overloaded(this_rq)))
1690 * Match the barrier from rt_set_overloaded; this guarantees that if we
1691 * see overloaded we must also see the rto_mask bit.
1695 for_each_cpu(cpu, this_rq->rd->rto_mask) {
1696 if (this_cpu == cpu)
1699 src_rq = cpu_rq(cpu);
1702 * Don't bother taking the src_rq->lock if the next highest
1703 * task is known to be lower-priority than our current task.
1704 * This may look racy, but if this value is about to go
1705 * logically higher, the src_rq will push this task away.
1706 * And if its going logically lower, we do not care
1708 if (src_rq->rt.highest_prio.next >=
1709 this_rq->rt.highest_prio.curr)
1713 * We can potentially drop this_rq's lock in
1714 * double_lock_balance, and another CPU could
1717 double_lock_balance(this_rq, src_rq);
1720 * We can pull only a task, which is pushable
1721 * on its rq, and no others.
1723 p = pick_highest_pushable_task(src_rq, this_cpu);
1726 * Do we have an RT task that preempts
1727 * the to-be-scheduled task?
1729 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
1730 WARN_ON(p == src_rq->curr);
1734 * There's a chance that p is higher in priority
1735 * than what's currently running on its cpu.
1736 * This is just that p is wakeing up and hasn't
1737 * had a chance to schedule. We only pull
1738 * p if it is lower in priority than the
1739 * current task on the run queue
1741 if (p->prio < src_rq->curr->prio)
1746 deactivate_task(src_rq, p, 0);
1747 set_task_cpu(p, this_cpu);
1748 activate_task(this_rq, p, 0);
1750 * We continue with the search, just in
1751 * case there's an even higher prio task
1752 * in another runqueue. (low likelihood
1757 double_unlock_balance(this_rq, src_rq);
1763 static void post_schedule_rt(struct rq *rq)
1769 * If we are not running and we are not going to reschedule soon, we should
1770 * try to push tasks away now
1772 static void task_woken_rt(struct rq *rq, struct task_struct *p)
1774 if (!task_running(rq, p) &&
1775 !test_tsk_need_resched(rq->curr) &&
1776 has_pushable_tasks(rq) &&
1777 p->nr_cpus_allowed > 1 &&
1778 (dl_task(rq->curr) || rt_task(rq->curr)) &&
1779 (rq->curr->nr_cpus_allowed < 2 ||
1780 rq->curr->prio <= p->prio))
1784 static void set_cpus_allowed_rt(struct task_struct *p,
1785 const struct cpumask *new_mask)
1790 BUG_ON(!rt_task(p));
1795 weight = cpumask_weight(new_mask);
1798 * Only update if the process changes its state from whether it
1799 * can migrate or not.
1801 if ((p->nr_cpus_allowed > 1) == (weight > 1))
1807 * The process used to be able to migrate OR it can now migrate
1810 if (!task_current(rq, p))
1811 dequeue_pushable_task(rq, p);
1812 BUG_ON(!rq->rt.rt_nr_migratory);
1813 rq->rt.rt_nr_migratory--;
1815 if (!task_current(rq, p))
1816 enqueue_pushable_task(rq, p);
1817 rq->rt.rt_nr_migratory++;
1820 update_rt_migration(&rq->rt);
1823 /* Assumes rq->lock is held */
1824 static void rq_online_rt(struct rq *rq)
1826 if (rq->rt.overloaded)
1827 rt_set_overload(rq);
1829 __enable_runtime(rq);
1831 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1834 /* Assumes rq->lock is held */
1835 static void rq_offline_rt(struct rq *rq)
1837 if (rq->rt.overloaded)
1838 rt_clear_overload(rq);
1840 __disable_runtime(rq);
1842 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
1846 * When switch from the rt queue, we bring ourselves to a position
1847 * that we might want to pull RT tasks from other runqueues.
1849 static void switched_from_rt(struct rq *rq, struct task_struct *p)
1852 * If there are other RT tasks then we will reschedule
1853 * and the scheduling of the other RT tasks will handle
1854 * the balancing. But if we are the last RT task
1855 * we may need to handle the pulling of RT tasks
1858 if (!p->on_rq || rq->rt.rt_nr_running)
1861 if (pull_rt_task(rq))
1862 resched_task(rq->curr);
1865 void __init init_sched_rt_class(void)
1869 for_each_possible_cpu(i) {
1870 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
1871 GFP_KERNEL, cpu_to_node(i));
1874 #endif /* CONFIG_SMP */
1877 * When switching a task to RT, we may overload the runqueue
1878 * with RT tasks. In this case we try to push them off to
1881 static void switched_to_rt(struct rq *rq, struct task_struct *p)
1883 int check_resched = 1;
1886 * If we are already running, then there's nothing
1887 * that needs to be done. But if we are not running
1888 * we may need to preempt the current running task.
1889 * If that current running task is also an RT task
1890 * then see if we can move to another run queue.
1892 if (p->on_rq && rq->curr != p) {
1894 if (rq->rt.overloaded && push_rt_task(rq) &&
1895 /* Don't resched if we changed runqueues */
1898 #endif /* CONFIG_SMP */
1899 if (check_resched && p->prio < rq->curr->prio)
1900 resched_task(rq->curr);
1905 * Priority of the task has changed. This may cause
1906 * us to initiate a push or pull.
1909 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
1914 if (rq->curr == p) {
1917 * If our priority decreases while running, we
1918 * may need to pull tasks to this runqueue.
1920 if (oldprio < p->prio)
1923 * If there's a higher priority task waiting to run
1924 * then reschedule. Note, the above pull_rt_task
1925 * can release the rq lock and p could migrate.
1926 * Only reschedule if p is still on the same runqueue.
1928 if (p->prio > rq->rt.highest_prio.curr && rq->curr == p)
1931 /* For UP simply resched on drop of prio */
1932 if (oldprio < p->prio)
1934 #endif /* CONFIG_SMP */
1937 * This task is not running, but if it is
1938 * greater than the current running task
1941 if (p->prio < rq->curr->prio)
1942 resched_task(rq->curr);
1946 static void watchdog(struct rq *rq, struct task_struct *p)
1948 unsigned long soft, hard;
1950 /* max may change after cur was read, this will be fixed next tick */
1951 soft = task_rlimit(p, RLIMIT_RTTIME);
1952 hard = task_rlimit_max(p, RLIMIT_RTTIME);
1954 if (soft != RLIM_INFINITY) {
1957 if (p->rt.watchdog_stamp != jiffies) {
1959 p->rt.watchdog_stamp = jiffies;
1962 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
1963 if (p->rt.timeout > next)
1964 p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
1968 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
1970 struct sched_rt_entity *rt_se = &p->rt;
1977 * RR tasks need a special form of timeslice management.
1978 * FIFO tasks have no timeslices.
1980 if (p->policy != SCHED_RR)
1983 if (--p->rt.time_slice)
1986 p->rt.time_slice = sched_rr_timeslice;
1989 * Requeue to the end of queue if we (and all of our ancestors) are not
1990 * the only element on the queue
1992 for_each_sched_rt_entity(rt_se) {
1993 if (rt_se->run_list.prev != rt_se->run_list.next) {
1994 requeue_task_rt(rq, p, 0);
1995 set_tsk_need_resched(p);
2001 static void set_curr_task_rt(struct rq *rq)
2003 struct task_struct *p = rq->curr;
2005 p->se.exec_start = rq_clock_task(rq);
2007 /* The running task is never eligible for pushing */
2008 dequeue_pushable_task(rq, p);
2011 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2014 * Time slice is 0 for SCHED_FIFO tasks
2016 if (task->policy == SCHED_RR)
2017 return sched_rr_timeslice;
2022 const struct sched_class rt_sched_class = {
2023 .next = &fair_sched_class,
2024 .enqueue_task = enqueue_task_rt,
2025 .dequeue_task = dequeue_task_rt,
2026 .yield_task = yield_task_rt,
2028 .check_preempt_curr = check_preempt_curr_rt,
2030 .pick_next_task = pick_next_task_rt,
2031 .put_prev_task = put_prev_task_rt,
2034 .select_task_rq = select_task_rq_rt,
2036 .set_cpus_allowed = set_cpus_allowed_rt,
2037 .rq_online = rq_online_rt,
2038 .rq_offline = rq_offline_rt,
2039 .post_schedule = post_schedule_rt,
2040 .task_woken = task_woken_rt,
2041 .switched_from = switched_from_rt,
2044 .set_curr_task = set_curr_task_rt,
2045 .task_tick = task_tick_rt,
2047 .get_rr_interval = get_rr_interval_rt,
2049 .prio_changed = prio_changed_rt,
2050 .switched_to = switched_to_rt,
2053 #ifdef CONFIG_SCHED_DEBUG
2054 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2056 void print_rt_stats(struct seq_file *m, int cpu)
2059 struct rt_rq *rt_rq;
2062 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2063 print_rt_rq(m, cpu, rt_rq);
2066 #endif /* CONFIG_SCHED_DEBUG */