2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
8 #include <linux/slab.h>
9 #include <linux/irq_work.h>
11 int sched_rr_timeslice = RR_TIMESLICE;
13 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
15 struct rt_bandwidth def_rt_bandwidth;
17 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
19 struct rt_bandwidth *rt_b =
20 container_of(timer, struct rt_bandwidth, rt_period_timer);
24 raw_spin_lock(&rt_b->rt_runtime_lock);
26 overrun = hrtimer_forward_now(timer, rt_b->rt_period);
30 raw_spin_unlock(&rt_b->rt_runtime_lock);
31 idle = do_sched_rt_period_timer(rt_b, overrun);
32 raw_spin_lock(&rt_b->rt_runtime_lock);
35 rt_b->rt_period_active = 0;
36 raw_spin_unlock(&rt_b->rt_runtime_lock);
38 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
41 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
43 rt_b->rt_period = ns_to_ktime(period);
44 rt_b->rt_runtime = runtime;
46 raw_spin_lock_init(&rt_b->rt_runtime_lock);
48 hrtimer_init(&rt_b->rt_period_timer,
49 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
50 rt_b->rt_period_timer.function = sched_rt_period_timer;
53 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
55 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
58 raw_spin_lock(&rt_b->rt_runtime_lock);
59 if (!rt_b->rt_period_active) {
60 rt_b->rt_period_active = 1;
61 hrtimer_forward_now(&rt_b->rt_period_timer, rt_b->rt_period);
62 hrtimer_start_expires(&rt_b->rt_period_timer, HRTIMER_MODE_ABS_PINNED);
64 raw_spin_unlock(&rt_b->rt_runtime_lock);
67 #if defined(CONFIG_SMP) && defined(HAVE_RT_PUSH_IPI)
68 static void push_irq_work_func(struct irq_work *work);
71 void init_rt_rq(struct rt_rq *rt_rq)
73 struct rt_prio_array *array;
76 array = &rt_rq->active;
77 for (i = 0; i < MAX_RT_PRIO; i++) {
78 INIT_LIST_HEAD(array->queue + i);
79 __clear_bit(i, array->bitmap);
81 /* delimiter for bitsearch: */
82 __set_bit(MAX_RT_PRIO, array->bitmap);
84 #if defined CONFIG_SMP
85 rt_rq->highest_prio.curr = MAX_RT_PRIO;
86 rt_rq->highest_prio.next = MAX_RT_PRIO;
87 rt_rq->rt_nr_migratory = 0;
88 rt_rq->overloaded = 0;
89 plist_head_init(&rt_rq->pushable_tasks);
91 #ifdef HAVE_RT_PUSH_IPI
92 rt_rq->push_flags = 0;
93 rt_rq->push_cpu = nr_cpu_ids;
94 raw_spin_lock_init(&rt_rq->push_lock);
95 init_irq_work(&rt_rq->push_work, push_irq_work_func);
97 #endif /* CONFIG_SMP */
98 /* We start is dequeued state, because no RT tasks are queued */
102 rt_rq->rt_throttled = 0;
103 rt_rq->rt_runtime = 0;
104 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
107 #ifdef CONFIG_RT_GROUP_SCHED
108 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
110 hrtimer_cancel(&rt_b->rt_period_timer);
113 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
115 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
117 #ifdef CONFIG_SCHED_DEBUG
118 WARN_ON_ONCE(!rt_entity_is_task(rt_se));
120 return container_of(rt_se, struct task_struct, rt);
123 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
128 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
133 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
135 struct rt_rq *rt_rq = rt_se->rt_rq;
140 void free_rt_sched_group(struct task_group *tg)
145 destroy_rt_bandwidth(&tg->rt_bandwidth);
147 for_each_possible_cpu(i) {
158 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
159 struct sched_rt_entity *rt_se, int cpu,
160 struct sched_rt_entity *parent)
162 struct rq *rq = cpu_rq(cpu);
164 rt_rq->highest_prio.curr = MAX_RT_PRIO;
165 rt_rq->rt_nr_boosted = 0;
169 tg->rt_rq[cpu] = rt_rq;
170 tg->rt_se[cpu] = rt_se;
176 rt_se->rt_rq = &rq->rt;
178 rt_se->rt_rq = parent->my_q;
181 rt_se->parent = parent;
182 INIT_LIST_HEAD(&rt_se->run_list);
185 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
188 struct sched_rt_entity *rt_se;
191 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
194 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
198 init_rt_bandwidth(&tg->rt_bandwidth,
199 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
201 for_each_possible_cpu(i) {
202 rt_rq = kzalloc_node(sizeof(struct rt_rq),
203 GFP_KERNEL, cpu_to_node(i));
207 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
208 GFP_KERNEL, cpu_to_node(i));
213 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
214 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
225 #else /* CONFIG_RT_GROUP_SCHED */
227 #define rt_entity_is_task(rt_se) (1)
229 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
231 return container_of(rt_se, struct task_struct, rt);
234 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
236 return container_of(rt_rq, struct rq, rt);
239 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
241 struct task_struct *p = rt_task_of(rt_se);
246 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
248 struct rq *rq = rq_of_rt_se(rt_se);
253 void free_rt_sched_group(struct task_group *tg) { }
255 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
259 #endif /* CONFIG_RT_GROUP_SCHED */
263 static void pull_rt_task(struct rq *this_rq);
265 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
267 /* Try to pull RT tasks here if we lower this rq's prio */
268 return rq->rt.highest_prio.curr > prev->prio;
271 static inline int rt_overloaded(struct rq *rq)
273 return atomic_read(&rq->rd->rto_count);
276 static inline void rt_set_overload(struct rq *rq)
281 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
283 * Make sure the mask is visible before we set
284 * the overload count. That is checked to determine
285 * if we should look at the mask. It would be a shame
286 * if we looked at the mask, but the mask was not
289 * Matched by the barrier in pull_rt_task().
292 atomic_inc(&rq->rd->rto_count);
295 static inline void rt_clear_overload(struct rq *rq)
300 /* the order here really doesn't matter */
301 atomic_dec(&rq->rd->rto_count);
302 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
305 static void update_rt_migration(struct rt_rq *rt_rq)
307 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
308 if (!rt_rq->overloaded) {
309 rt_set_overload(rq_of_rt_rq(rt_rq));
310 rt_rq->overloaded = 1;
312 } else if (rt_rq->overloaded) {
313 rt_clear_overload(rq_of_rt_rq(rt_rq));
314 rt_rq->overloaded = 0;
318 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
320 struct task_struct *p;
322 if (!rt_entity_is_task(rt_se))
325 p = rt_task_of(rt_se);
326 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
328 rt_rq->rt_nr_total++;
329 if (p->nr_cpus_allowed > 1)
330 rt_rq->rt_nr_migratory++;
332 update_rt_migration(rt_rq);
335 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
337 struct task_struct *p;
339 if (!rt_entity_is_task(rt_se))
342 p = rt_task_of(rt_se);
343 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
345 rt_rq->rt_nr_total--;
346 if (p->nr_cpus_allowed > 1)
347 rt_rq->rt_nr_migratory--;
349 update_rt_migration(rt_rq);
352 static inline int has_pushable_tasks(struct rq *rq)
354 return !plist_head_empty(&rq->rt.pushable_tasks);
357 static DEFINE_PER_CPU(struct callback_head, rt_push_head);
358 static DEFINE_PER_CPU(struct callback_head, rt_pull_head);
360 static void push_rt_tasks(struct rq *);
361 static void pull_rt_task(struct rq *);
363 static inline void queue_push_tasks(struct rq *rq)
365 if (!has_pushable_tasks(rq))
368 queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks);
371 static inline void queue_pull_task(struct rq *rq)
373 queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task);
376 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
378 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
379 plist_node_init(&p->pushable_tasks, p->prio);
380 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
382 /* Update the highest prio pushable task */
383 if (p->prio < rq->rt.highest_prio.next)
384 rq->rt.highest_prio.next = p->prio;
387 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
389 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
391 /* Update the new highest prio pushable task */
392 if (has_pushable_tasks(rq)) {
393 p = plist_first_entry(&rq->rt.pushable_tasks,
394 struct task_struct, pushable_tasks);
395 rq->rt.highest_prio.next = p->prio;
397 rq->rt.highest_prio.next = MAX_RT_PRIO;
402 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
406 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
411 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
416 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
420 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
425 static inline void pull_rt_task(struct rq *this_rq)
429 static inline void queue_push_tasks(struct rq *rq)
432 #endif /* CONFIG_SMP */
434 static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
435 static void dequeue_top_rt_rq(struct rt_rq *rt_rq);
437 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
439 return !list_empty(&rt_se->run_list);
442 #ifdef CONFIG_RT_GROUP_SCHED
444 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
449 return rt_rq->rt_runtime;
452 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
454 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
457 typedef struct task_group *rt_rq_iter_t;
459 static inline struct task_group *next_task_group(struct task_group *tg)
462 tg = list_entry_rcu(tg->list.next,
463 typeof(struct task_group), list);
464 } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
466 if (&tg->list == &task_groups)
472 #define for_each_rt_rq(rt_rq, iter, rq) \
473 for (iter = container_of(&task_groups, typeof(*iter), list); \
474 (iter = next_task_group(iter)) && \
475 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
477 #define for_each_sched_rt_entity(rt_se) \
478 for (; rt_se; rt_se = rt_se->parent)
480 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
485 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head);
486 static void dequeue_rt_entity(struct sched_rt_entity *rt_se);
488 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
490 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
491 struct rq *rq = rq_of_rt_rq(rt_rq);
492 struct sched_rt_entity *rt_se;
494 int cpu = cpu_of(rq);
496 rt_se = rt_rq->tg->rt_se[cpu];
498 if (rt_rq->rt_nr_running) {
500 enqueue_top_rt_rq(rt_rq);
501 else if (!on_rt_rq(rt_se))
502 enqueue_rt_entity(rt_se, false);
504 if (rt_rq->highest_prio.curr < curr->prio)
509 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
511 struct sched_rt_entity *rt_se;
512 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
514 rt_se = rt_rq->tg->rt_se[cpu];
517 dequeue_top_rt_rq(rt_rq);
518 else if (on_rt_rq(rt_se))
519 dequeue_rt_entity(rt_se);
522 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
524 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
527 static int rt_se_boosted(struct sched_rt_entity *rt_se)
529 struct rt_rq *rt_rq = group_rt_rq(rt_se);
530 struct task_struct *p;
533 return !!rt_rq->rt_nr_boosted;
535 p = rt_task_of(rt_se);
536 return p->prio != p->normal_prio;
540 static inline const struct cpumask *sched_rt_period_mask(void)
542 return this_rq()->rd->span;
545 static inline const struct cpumask *sched_rt_period_mask(void)
547 return cpu_online_mask;
552 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
554 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
557 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
559 return &rt_rq->tg->rt_bandwidth;
562 #else /* !CONFIG_RT_GROUP_SCHED */
564 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
566 return rt_rq->rt_runtime;
569 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
571 return ktime_to_ns(def_rt_bandwidth.rt_period);
574 typedef struct rt_rq *rt_rq_iter_t;
576 #define for_each_rt_rq(rt_rq, iter, rq) \
577 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
579 #define for_each_sched_rt_entity(rt_se) \
580 for (; rt_se; rt_se = NULL)
582 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
587 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
589 struct rq *rq = rq_of_rt_rq(rt_rq);
591 if (!rt_rq->rt_nr_running)
594 enqueue_top_rt_rq(rt_rq);
598 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
600 dequeue_top_rt_rq(rt_rq);
603 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
605 return rt_rq->rt_throttled;
608 static inline const struct cpumask *sched_rt_period_mask(void)
610 return cpu_online_mask;
614 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
616 return &cpu_rq(cpu)->rt;
619 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
621 return &def_rt_bandwidth;
624 #endif /* CONFIG_RT_GROUP_SCHED */
626 bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
628 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
630 return (hrtimer_active(&rt_b->rt_period_timer) ||
631 rt_rq->rt_time < rt_b->rt_runtime);
636 * We ran out of runtime, see if we can borrow some from our neighbours.
638 static void do_balance_runtime(struct rt_rq *rt_rq)
640 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
641 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
645 weight = cpumask_weight(rd->span);
647 raw_spin_lock(&rt_b->rt_runtime_lock);
648 rt_period = ktime_to_ns(rt_b->rt_period);
649 for_each_cpu(i, rd->span) {
650 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
656 raw_spin_lock(&iter->rt_runtime_lock);
658 * Either all rqs have inf runtime and there's nothing to steal
659 * or __disable_runtime() below sets a specific rq to inf to
660 * indicate its been disabled and disalow stealing.
662 if (iter->rt_runtime == RUNTIME_INF)
666 * From runqueues with spare time, take 1/n part of their
667 * spare time, but no more than our period.
669 diff = iter->rt_runtime - iter->rt_time;
671 diff = div_u64((u64)diff, weight);
672 if (rt_rq->rt_runtime + diff > rt_period)
673 diff = rt_period - rt_rq->rt_runtime;
674 iter->rt_runtime -= diff;
675 rt_rq->rt_runtime += diff;
676 if (rt_rq->rt_runtime == rt_period) {
677 raw_spin_unlock(&iter->rt_runtime_lock);
682 raw_spin_unlock(&iter->rt_runtime_lock);
684 raw_spin_unlock(&rt_b->rt_runtime_lock);
688 * Ensure this RQ takes back all the runtime it lend to its neighbours.
690 static void __disable_runtime(struct rq *rq)
692 struct root_domain *rd = rq->rd;
696 if (unlikely(!scheduler_running))
699 for_each_rt_rq(rt_rq, iter, rq) {
700 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
704 raw_spin_lock(&rt_b->rt_runtime_lock);
705 raw_spin_lock(&rt_rq->rt_runtime_lock);
707 * Either we're all inf and nobody needs to borrow, or we're
708 * already disabled and thus have nothing to do, or we have
709 * exactly the right amount of runtime to take out.
711 if (rt_rq->rt_runtime == RUNTIME_INF ||
712 rt_rq->rt_runtime == rt_b->rt_runtime)
714 raw_spin_unlock(&rt_rq->rt_runtime_lock);
717 * Calculate the difference between what we started out with
718 * and what we current have, that's the amount of runtime
719 * we lend and now have to reclaim.
721 want = rt_b->rt_runtime - rt_rq->rt_runtime;
724 * Greedy reclaim, take back as much as we can.
726 for_each_cpu(i, rd->span) {
727 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
731 * Can't reclaim from ourselves or disabled runqueues.
733 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
736 raw_spin_lock(&iter->rt_runtime_lock);
738 diff = min_t(s64, iter->rt_runtime, want);
739 iter->rt_runtime -= diff;
742 iter->rt_runtime -= want;
745 raw_spin_unlock(&iter->rt_runtime_lock);
751 raw_spin_lock(&rt_rq->rt_runtime_lock);
753 * We cannot be left wanting - that would mean some runtime
754 * leaked out of the system.
759 * Disable all the borrow logic by pretending we have inf
760 * runtime - in which case borrowing doesn't make sense.
762 rt_rq->rt_runtime = RUNTIME_INF;
763 rt_rq->rt_throttled = 0;
764 raw_spin_unlock(&rt_rq->rt_runtime_lock);
765 raw_spin_unlock(&rt_b->rt_runtime_lock);
767 /* Make rt_rq available for pick_next_task() */
768 sched_rt_rq_enqueue(rt_rq);
772 static void __enable_runtime(struct rq *rq)
777 if (unlikely(!scheduler_running))
781 * Reset each runqueue's bandwidth settings
783 for_each_rt_rq(rt_rq, iter, rq) {
784 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
786 raw_spin_lock(&rt_b->rt_runtime_lock);
787 raw_spin_lock(&rt_rq->rt_runtime_lock);
788 rt_rq->rt_runtime = rt_b->rt_runtime;
790 rt_rq->rt_throttled = 0;
791 raw_spin_unlock(&rt_rq->rt_runtime_lock);
792 raw_spin_unlock(&rt_b->rt_runtime_lock);
796 static void balance_runtime(struct rt_rq *rt_rq)
798 if (!sched_feat(RT_RUNTIME_SHARE))
801 if (rt_rq->rt_time > rt_rq->rt_runtime) {
802 raw_spin_unlock(&rt_rq->rt_runtime_lock);
803 do_balance_runtime(rt_rq);
804 raw_spin_lock(&rt_rq->rt_runtime_lock);
807 #else /* !CONFIG_SMP */
808 static inline void balance_runtime(struct rt_rq *rt_rq) {}
809 #endif /* CONFIG_SMP */
811 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
813 int i, idle = 1, throttled = 0;
814 const struct cpumask *span;
816 span = sched_rt_period_mask();
817 #ifdef CONFIG_RT_GROUP_SCHED
819 * FIXME: isolated CPUs should really leave the root task group,
820 * whether they are isolcpus or were isolated via cpusets, lest
821 * the timer run on a CPU which does not service all runqueues,
822 * potentially leaving other CPUs indefinitely throttled. If
823 * isolation is really required, the user will turn the throttle
824 * off to kill the perturbations it causes anyway. Meanwhile,
825 * this maintains functionality for boot and/or troubleshooting.
827 if (rt_b == &root_task_group.rt_bandwidth)
828 span = cpu_online_mask;
830 for_each_cpu(i, span) {
832 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
833 struct rq *rq = rq_of_rt_rq(rt_rq);
835 raw_spin_lock(&rq->lock);
836 if (rt_rq->rt_time) {
839 raw_spin_lock(&rt_rq->rt_runtime_lock);
840 if (rt_rq->rt_throttled)
841 balance_runtime(rt_rq);
842 runtime = rt_rq->rt_runtime;
843 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
844 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
845 rt_rq->rt_throttled = 0;
849 * When we're idle and a woken (rt) task is
850 * throttled check_preempt_curr() will set
851 * skip_update and the time between the wakeup
852 * and this unthrottle will get accounted as
855 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
856 rq_clock_skip_update(rq, false);
858 if (rt_rq->rt_time || rt_rq->rt_nr_running)
860 raw_spin_unlock(&rt_rq->rt_runtime_lock);
861 } else if (rt_rq->rt_nr_running) {
863 if (!rt_rq_throttled(rt_rq))
866 if (rt_rq->rt_throttled)
870 sched_rt_rq_enqueue(rt_rq);
871 raw_spin_unlock(&rq->lock);
874 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
880 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
882 #ifdef CONFIG_RT_GROUP_SCHED
883 struct rt_rq *rt_rq = group_rt_rq(rt_se);
886 return rt_rq->highest_prio.curr;
889 return rt_task_of(rt_se)->prio;
892 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
894 u64 runtime = sched_rt_runtime(rt_rq);
896 if (rt_rq->rt_throttled)
897 return rt_rq_throttled(rt_rq);
899 if (runtime >= sched_rt_period(rt_rq))
902 balance_runtime(rt_rq);
903 runtime = sched_rt_runtime(rt_rq);
904 if (runtime == RUNTIME_INF)
907 if (rt_rq->rt_time > runtime) {
908 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
911 * Don't actually throttle groups that have no runtime assigned
912 * but accrue some time due to boosting.
914 if (likely(rt_b->rt_runtime)) {
915 rt_rq->rt_throttled = 1;
916 printk_deferred_once("sched: RT throttling activated\n");
919 * In case we did anyway, make it go away,
920 * replenishment is a joke, since it will replenish us
926 if (rt_rq_throttled(rt_rq)) {
927 sched_rt_rq_dequeue(rt_rq);
936 * Update the current task's runtime statistics. Skip current tasks that
937 * are not in our scheduling class.
939 static void update_curr_rt(struct rq *rq)
941 struct task_struct *curr = rq->curr;
942 struct sched_rt_entity *rt_se = &curr->rt;
945 if (curr->sched_class != &rt_sched_class)
948 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
949 if (unlikely((s64)delta_exec <= 0))
952 schedstat_set(curr->se.statistics.exec_max,
953 max(curr->se.statistics.exec_max, delta_exec));
955 curr->se.sum_exec_runtime += delta_exec;
956 account_group_exec_runtime(curr, delta_exec);
958 curr->se.exec_start = rq_clock_task(rq);
959 cpuacct_charge(curr, delta_exec);
961 sched_rt_avg_update(rq, delta_exec);
963 if (!rt_bandwidth_enabled())
966 for_each_sched_rt_entity(rt_se) {
967 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
969 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
970 raw_spin_lock(&rt_rq->rt_runtime_lock);
971 rt_rq->rt_time += delta_exec;
972 if (sched_rt_runtime_exceeded(rt_rq))
974 raw_spin_unlock(&rt_rq->rt_runtime_lock);
980 dequeue_top_rt_rq(struct rt_rq *rt_rq)
982 struct rq *rq = rq_of_rt_rq(rt_rq);
984 BUG_ON(&rq->rt != rt_rq);
986 if (!rt_rq->rt_queued)
989 BUG_ON(!rq->nr_running);
991 sub_nr_running(rq, rt_rq->rt_nr_running);
992 rt_rq->rt_queued = 0;
996 enqueue_top_rt_rq(struct rt_rq *rt_rq)
998 struct rq *rq = rq_of_rt_rq(rt_rq);
1000 BUG_ON(&rq->rt != rt_rq);
1002 if (rt_rq->rt_queued)
1004 if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running)
1007 add_nr_running(rq, rt_rq->rt_nr_running);
1008 rt_rq->rt_queued = 1;
1011 #if defined CONFIG_SMP
1014 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1016 struct rq *rq = rq_of_rt_rq(rt_rq);
1018 #ifdef CONFIG_RT_GROUP_SCHED
1020 * Change rq's cpupri only if rt_rq is the top queue.
1022 if (&rq->rt != rt_rq)
1025 if (rq->online && prio < prev_prio)
1026 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1030 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1032 struct rq *rq = rq_of_rt_rq(rt_rq);
1034 #ifdef CONFIG_RT_GROUP_SCHED
1036 * Change rq's cpupri only if rt_rq is the top queue.
1038 if (&rq->rt != rt_rq)
1041 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1042 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1045 #else /* CONFIG_SMP */
1048 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1050 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1052 #endif /* CONFIG_SMP */
1054 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1056 inc_rt_prio(struct rt_rq *rt_rq, int prio)
1058 int prev_prio = rt_rq->highest_prio.curr;
1060 if (prio < prev_prio)
1061 rt_rq->highest_prio.curr = prio;
1063 inc_rt_prio_smp(rt_rq, prio, prev_prio);
1067 dec_rt_prio(struct rt_rq *rt_rq, int prio)
1069 int prev_prio = rt_rq->highest_prio.curr;
1071 if (rt_rq->rt_nr_running) {
1073 WARN_ON(prio < prev_prio);
1076 * This may have been our highest task, and therefore
1077 * we may have some recomputation to do
1079 if (prio == prev_prio) {
1080 struct rt_prio_array *array = &rt_rq->active;
1082 rt_rq->highest_prio.curr =
1083 sched_find_first_bit(array->bitmap);
1087 rt_rq->highest_prio.curr = MAX_RT_PRIO;
1089 dec_rt_prio_smp(rt_rq, prio, prev_prio);
1094 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1095 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1097 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1099 #ifdef CONFIG_RT_GROUP_SCHED
1102 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1104 if (rt_se_boosted(rt_se))
1105 rt_rq->rt_nr_boosted++;
1108 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1112 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1114 if (rt_se_boosted(rt_se))
1115 rt_rq->rt_nr_boosted--;
1117 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1120 #else /* CONFIG_RT_GROUP_SCHED */
1123 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1125 start_rt_bandwidth(&def_rt_bandwidth);
1129 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1131 #endif /* CONFIG_RT_GROUP_SCHED */
1134 unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1136 struct rt_rq *group_rq = group_rt_rq(rt_se);
1139 return group_rq->rt_nr_running;
1145 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1147 int prio = rt_se_prio(rt_se);
1149 WARN_ON(!rt_prio(prio));
1150 rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1152 inc_rt_prio(rt_rq, prio);
1153 inc_rt_migration(rt_se, rt_rq);
1154 inc_rt_group(rt_se, rt_rq);
1158 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1160 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1161 WARN_ON(!rt_rq->rt_nr_running);
1162 rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1164 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1165 dec_rt_migration(rt_se, rt_rq);
1166 dec_rt_group(rt_se, rt_rq);
1169 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
1171 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1172 struct rt_prio_array *array = &rt_rq->active;
1173 struct rt_rq *group_rq = group_rt_rq(rt_se);
1174 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1177 * Don't enqueue the group if its throttled, or when empty.
1178 * The latter is a consequence of the former when a child group
1179 * get throttled and the current group doesn't have any other
1182 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running))
1186 list_add(&rt_se->run_list, queue);
1188 list_add_tail(&rt_se->run_list, queue);
1189 __set_bit(rt_se_prio(rt_se), array->bitmap);
1191 inc_rt_tasks(rt_se, rt_rq);
1194 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se)
1196 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1197 struct rt_prio_array *array = &rt_rq->active;
1199 list_del_init(&rt_se->run_list);
1200 if (list_empty(array->queue + rt_se_prio(rt_se)))
1201 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1203 dec_rt_tasks(rt_se, rt_rq);
1207 * Because the prio of an upper entry depends on the lower
1208 * entries, we must remove entries top - down.
1210 static void dequeue_rt_stack(struct sched_rt_entity *rt_se)
1212 struct sched_rt_entity *back = NULL;
1214 for_each_sched_rt_entity(rt_se) {
1219 dequeue_top_rt_rq(rt_rq_of_se(back));
1221 for (rt_se = back; rt_se; rt_se = rt_se->back) {
1222 if (on_rt_rq(rt_se))
1223 __dequeue_rt_entity(rt_se);
1227 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, bool head)
1229 struct rq *rq = rq_of_rt_se(rt_se);
1231 dequeue_rt_stack(rt_se);
1232 for_each_sched_rt_entity(rt_se)
1233 __enqueue_rt_entity(rt_se, head);
1234 enqueue_top_rt_rq(&rq->rt);
1237 static void dequeue_rt_entity(struct sched_rt_entity *rt_se)
1239 struct rq *rq = rq_of_rt_se(rt_se);
1241 dequeue_rt_stack(rt_se);
1243 for_each_sched_rt_entity(rt_se) {
1244 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1246 if (rt_rq && rt_rq->rt_nr_running)
1247 __enqueue_rt_entity(rt_se, false);
1249 enqueue_top_rt_rq(&rq->rt);
1253 * Adding/removing a task to/from a priority array:
1256 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1258 struct sched_rt_entity *rt_se = &p->rt;
1260 if (flags & ENQUEUE_WAKEUP)
1263 enqueue_rt_entity(rt_se, flags & ENQUEUE_HEAD);
1265 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1266 enqueue_pushable_task(rq, p);
1269 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1271 struct sched_rt_entity *rt_se = &p->rt;
1274 dequeue_rt_entity(rt_se);
1276 dequeue_pushable_task(rq, p);
1280 * Put task to the head or the end of the run list without the overhead of
1281 * dequeue followed by enqueue.
1284 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1286 if (on_rt_rq(rt_se)) {
1287 struct rt_prio_array *array = &rt_rq->active;
1288 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1291 list_move(&rt_se->run_list, queue);
1293 list_move_tail(&rt_se->run_list, queue);
1297 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1299 struct sched_rt_entity *rt_se = &p->rt;
1300 struct rt_rq *rt_rq;
1302 for_each_sched_rt_entity(rt_se) {
1303 rt_rq = rt_rq_of_se(rt_se);
1304 requeue_rt_entity(rt_rq, rt_se, head);
1308 static void yield_task_rt(struct rq *rq)
1310 requeue_task_rt(rq, rq->curr, 0);
1314 static int find_lowest_rq(struct task_struct *task);
1317 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1319 struct task_struct *curr;
1322 /* For anything but wake ups, just return the task_cpu */
1323 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1329 curr = READ_ONCE(rq->curr); /* unlocked access */
1332 * If the current task on @p's runqueue is an RT task, then
1333 * try to see if we can wake this RT task up on another
1334 * runqueue. Otherwise simply start this RT task
1335 * on its current runqueue.
1337 * We want to avoid overloading runqueues. If the woken
1338 * task is a higher priority, then it will stay on this CPU
1339 * and the lower prio task should be moved to another CPU.
1340 * Even though this will probably make the lower prio task
1341 * lose its cache, we do not want to bounce a higher task
1342 * around just because it gave up its CPU, perhaps for a
1345 * For equal prio tasks, we just let the scheduler sort it out.
1347 * Otherwise, just let it ride on the affined RQ and the
1348 * post-schedule router will push the preempted task away
1350 * This test is optimistic, if we get it wrong the load-balancer
1351 * will have to sort it out.
1353 if (curr && unlikely(rt_task(curr)) &&
1354 (curr->nr_cpus_allowed < 2 ||
1355 curr->prio <= p->prio)) {
1356 int target = find_lowest_rq(p);
1359 * Don't bother moving it if the destination CPU is
1360 * not running a lower priority task.
1363 p->prio < cpu_rq(target)->rt.highest_prio.curr)
1372 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1375 * Current can't be migrated, useless to reschedule,
1376 * let's hope p can move out.
1378 if (rq->curr->nr_cpus_allowed == 1 ||
1379 !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1383 * p is migratable, so let's not schedule it and
1384 * see if it is pushed or pulled somewhere else.
1386 if (p->nr_cpus_allowed != 1
1387 && cpupri_find(&rq->rd->cpupri, p, NULL))
1391 * There appears to be other cpus that can accept
1392 * current and none to run 'p', so lets reschedule
1393 * to try and push current away:
1395 requeue_task_rt(rq, p, 1);
1399 #endif /* CONFIG_SMP */
1402 * Preempt the current task with a newly woken task if needed:
1404 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1406 if (p->prio < rq->curr->prio) {
1415 * - the newly woken task is of equal priority to the current task
1416 * - the newly woken task is non-migratable while current is migratable
1417 * - current will be preempted on the next reschedule
1419 * we should check to see if current can readily move to a different
1420 * cpu. If so, we will reschedule to allow the push logic to try
1421 * to move current somewhere else, making room for our non-migratable
1424 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1425 check_preempt_equal_prio(rq, p);
1430 static void sched_rt_update_capacity_req(struct rq *rq)
1432 u64 total, used, age_stamp, avg;
1438 sched_avg_update(rq);
1440 * Since we're reading these variables without serialization make sure
1441 * we read them once before doing sanity checks on them.
1443 age_stamp = READ_ONCE(rq->age_stamp);
1444 avg = READ_ONCE(rq->rt_avg);
1445 delta = rq_clock(rq) - age_stamp;
1447 if (unlikely(delta < 0))
1450 total = sched_avg_period() + delta;
1452 used = div_u64(avg, total);
1453 if (unlikely(used > SCHED_CAPACITY_SCALE))
1454 used = SCHED_CAPACITY_SCALE;
1456 set_rt_cpu_capacity(rq->cpu, 1, (unsigned long)(used));
1459 static inline void sched_rt_update_capacity_req(struct rq *rq)
1464 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1465 struct rt_rq *rt_rq)
1467 struct rt_prio_array *array = &rt_rq->active;
1468 struct sched_rt_entity *next = NULL;
1469 struct list_head *queue;
1472 idx = sched_find_first_bit(array->bitmap);
1473 BUG_ON(idx >= MAX_RT_PRIO);
1475 queue = array->queue + idx;
1476 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1481 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1483 struct sched_rt_entity *rt_se;
1484 struct task_struct *p;
1485 struct rt_rq *rt_rq = &rq->rt;
1488 rt_se = pick_next_rt_entity(rq, rt_rq);
1490 rt_rq = group_rt_rq(rt_se);
1493 p = rt_task_of(rt_se);
1494 p->se.exec_start = rq_clock_task(rq);
1499 static struct task_struct *
1500 pick_next_task_rt(struct rq *rq, struct task_struct *prev)
1502 struct task_struct *p;
1503 struct rt_rq *rt_rq = &rq->rt;
1505 if (need_pull_rt_task(rq, prev)) {
1507 * This is OK, because current is on_cpu, which avoids it being
1508 * picked for load-balance and preemption/IRQs are still
1509 * disabled avoiding further scheduler activity on it and we're
1510 * being very careful to re-start the picking loop.
1512 lockdep_unpin_lock(&rq->lock);
1514 lockdep_pin_lock(&rq->lock);
1516 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1517 * means a dl or stop task can slip in, in which case we need
1518 * to re-start task selection.
1520 if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) ||
1521 rq->dl.dl_nr_running))
1526 * We may dequeue prev's rt_rq in put_prev_task().
1527 * So, we update time before rt_nr_running check.
1529 if (prev->sched_class == &rt_sched_class)
1532 if (!rt_rq->rt_queued) {
1534 * The next task to be picked on this rq will have a lower
1535 * priority than rt tasks so we can spend some time to update
1536 * the capacity used by rt tasks based on the last activity.
1537 * This value will be the used as an estimation of the next
1540 sched_rt_update_capacity_req(rq);
1544 put_prev_task(rq, prev);
1546 p = _pick_next_task_rt(rq);
1548 /* The running task is never eligible for pushing */
1549 dequeue_pushable_task(rq, p);
1551 queue_push_tasks(rq);
1556 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1561 * The previous task needs to be made eligible for pushing
1562 * if it is still active
1564 if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1565 enqueue_pushable_task(rq, p);
1570 /* Only try algorithms three times */
1571 #define RT_MAX_TRIES 3
1573 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1575 if (!task_running(rq, p) &&
1576 cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
1582 * Return the highest pushable rq's task, which is suitable to be executed
1583 * on the cpu, NULL otherwise
1585 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1587 struct plist_head *head = &rq->rt.pushable_tasks;
1588 struct task_struct *p;
1590 if (!has_pushable_tasks(rq))
1593 plist_for_each_entry(p, head, pushable_tasks) {
1594 if (pick_rt_task(rq, p, cpu))
1601 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1603 static int find_lowest_rq(struct task_struct *task)
1605 struct sched_domain *sd;
1606 struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1607 int this_cpu = smp_processor_id();
1608 int cpu = task_cpu(task);
1610 /* Make sure the mask is initialized first */
1611 if (unlikely(!lowest_mask))
1614 if (task->nr_cpus_allowed == 1)
1615 return -1; /* No other targets possible */
1617 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1618 return -1; /* No targets found */
1621 * At this point we have built a mask of cpus representing the
1622 * lowest priority tasks in the system. Now we want to elect
1623 * the best one based on our affinity and topology.
1625 * We prioritize the last cpu that the task executed on since
1626 * it is most likely cache-hot in that location.
1628 if (cpumask_test_cpu(cpu, lowest_mask))
1632 * Otherwise, we consult the sched_domains span maps to figure
1633 * out which cpu is logically closest to our hot cache data.
1635 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1636 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1639 for_each_domain(cpu, sd) {
1640 if (sd->flags & SD_WAKE_AFFINE) {
1644 * "this_cpu" is cheaper to preempt than a
1647 if (this_cpu != -1 &&
1648 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1653 best_cpu = cpumask_first_and(lowest_mask,
1654 sched_domain_span(sd));
1655 if (best_cpu < nr_cpu_ids) {
1664 * And finally, if there were no matches within the domains
1665 * just give the caller *something* to work with from the compatible
1671 cpu = cpumask_any(lowest_mask);
1672 if (cpu < nr_cpu_ids)
1677 /* Will lock the rq it finds */
1678 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1680 struct rq *lowest_rq = NULL;
1684 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1685 cpu = find_lowest_rq(task);
1687 if ((cpu == -1) || (cpu == rq->cpu))
1690 lowest_rq = cpu_rq(cpu);
1692 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1694 * Target rq has tasks of equal or higher priority,
1695 * retrying does not release any lock and is unlikely
1696 * to yield a different result.
1702 /* if the prio of this runqueue changed, try again */
1703 if (double_lock_balance(rq, lowest_rq)) {
1705 * We had to unlock the run queue. In
1706 * the mean time, task could have
1707 * migrated already or had its affinity changed.
1708 * Also make sure that it wasn't scheduled on its rq.
1710 if (unlikely(task_rq(task) != rq ||
1711 !cpumask_test_cpu(lowest_rq->cpu,
1712 tsk_cpus_allowed(task)) ||
1713 task_running(rq, task) ||
1714 !task_on_rq_queued(task))) {
1716 double_unlock_balance(rq, lowest_rq);
1722 /* If this rq is still suitable use it. */
1723 if (lowest_rq->rt.highest_prio.curr > task->prio)
1727 double_unlock_balance(rq, lowest_rq);
1734 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1736 struct task_struct *p;
1738 if (!has_pushable_tasks(rq))
1741 p = plist_first_entry(&rq->rt.pushable_tasks,
1742 struct task_struct, pushable_tasks);
1744 BUG_ON(rq->cpu != task_cpu(p));
1745 BUG_ON(task_current(rq, p));
1746 BUG_ON(p->nr_cpus_allowed <= 1);
1748 BUG_ON(!task_on_rq_queued(p));
1749 BUG_ON(!rt_task(p));
1755 * If the current CPU has more than one RT task, see if the non
1756 * running task can migrate over to a CPU that is running a task
1757 * of lesser priority.
1759 static int push_rt_task(struct rq *rq)
1761 struct task_struct *next_task;
1762 struct rq *lowest_rq;
1765 if (!rq->rt.overloaded)
1768 next_task = pick_next_pushable_task(rq);
1773 if (unlikely(next_task == rq->curr)) {
1779 * It's possible that the next_task slipped in of
1780 * higher priority than current. If that's the case
1781 * just reschedule current.
1783 if (unlikely(next_task->prio < rq->curr->prio)) {
1788 /* We might release rq lock */
1789 get_task_struct(next_task);
1791 /* find_lock_lowest_rq locks the rq if found */
1792 lowest_rq = find_lock_lowest_rq(next_task, rq);
1794 struct task_struct *task;
1796 * find_lock_lowest_rq releases rq->lock
1797 * so it is possible that next_task has migrated.
1799 * We need to make sure that the task is still on the same
1800 * run-queue and is also still the next task eligible for
1803 task = pick_next_pushable_task(rq);
1804 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1806 * The task hasn't migrated, and is still the next
1807 * eligible task, but we failed to find a run-queue
1808 * to push it to. Do not retry in this case, since
1809 * other cpus will pull from us when ready.
1815 /* No more tasks, just exit */
1819 * Something has shifted, try again.
1821 put_task_struct(next_task);
1826 deactivate_task(rq, next_task, 0);
1827 set_task_cpu(next_task, lowest_rq->cpu);
1828 activate_task(lowest_rq, next_task, 0);
1831 resched_curr(lowest_rq);
1833 double_unlock_balance(rq, lowest_rq);
1836 put_task_struct(next_task);
1841 static void push_rt_tasks(struct rq *rq)
1843 /* push_rt_task will return true if it moved an RT */
1844 while (push_rt_task(rq))
1848 #ifdef HAVE_RT_PUSH_IPI
1850 * The search for the next cpu always starts at rq->cpu and ends
1851 * when we reach rq->cpu again. It will never return rq->cpu.
1852 * This returns the next cpu to check, or nr_cpu_ids if the loop
1855 * rq->rt.push_cpu holds the last cpu returned by this function,
1856 * or if this is the first instance, it must hold rq->cpu.
1858 static int rto_next_cpu(struct rq *rq)
1860 int prev_cpu = rq->rt.push_cpu;
1863 cpu = cpumask_next(prev_cpu, rq->rd->rto_mask);
1866 * If the previous cpu is less than the rq's CPU, then it already
1867 * passed the end of the mask, and has started from the beginning.
1868 * We end if the next CPU is greater or equal to rq's CPU.
1870 if (prev_cpu < rq->cpu) {
1874 } else if (cpu >= nr_cpu_ids) {
1876 * We passed the end of the mask, start at the beginning.
1877 * If the result is greater or equal to the rq's CPU, then
1878 * the loop is finished.
1880 cpu = cpumask_first(rq->rd->rto_mask);
1884 rq->rt.push_cpu = cpu;
1886 /* Return cpu to let the caller know if the loop is finished or not */
1890 static int find_next_push_cpu(struct rq *rq)
1896 cpu = rto_next_cpu(rq);
1897 if (cpu >= nr_cpu_ids)
1899 next_rq = cpu_rq(cpu);
1901 /* Make sure the next rq can push to this rq */
1902 if (next_rq->rt.highest_prio.next < rq->rt.highest_prio.curr)
1909 #define RT_PUSH_IPI_EXECUTING 1
1910 #define RT_PUSH_IPI_RESTART 2
1912 static void tell_cpu_to_push(struct rq *rq)
1916 if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
1917 raw_spin_lock(&rq->rt.push_lock);
1918 /* Make sure it's still executing */
1919 if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
1921 * Tell the IPI to restart the loop as things have
1922 * changed since it started.
1924 rq->rt.push_flags |= RT_PUSH_IPI_RESTART;
1925 raw_spin_unlock(&rq->rt.push_lock);
1928 raw_spin_unlock(&rq->rt.push_lock);
1931 /* When here, there's no IPI going around */
1933 rq->rt.push_cpu = rq->cpu;
1934 cpu = find_next_push_cpu(rq);
1935 if (cpu >= nr_cpu_ids)
1938 rq->rt.push_flags = RT_PUSH_IPI_EXECUTING;
1940 irq_work_queue_on(&rq->rt.push_work, cpu);
1943 /* Called from hardirq context */
1944 static void try_to_push_tasks(void *arg)
1946 struct rt_rq *rt_rq = arg;
1947 struct rq *rq, *src_rq;
1951 this_cpu = rt_rq->push_cpu;
1953 /* Paranoid check */
1954 BUG_ON(this_cpu != smp_processor_id());
1956 rq = cpu_rq(this_cpu);
1957 src_rq = rq_of_rt_rq(rt_rq);
1960 if (has_pushable_tasks(rq)) {
1961 raw_spin_lock(&rq->lock);
1963 raw_spin_unlock(&rq->lock);
1966 /* Pass the IPI to the next rt overloaded queue */
1967 raw_spin_lock(&rt_rq->push_lock);
1969 * If the source queue changed since the IPI went out,
1970 * we need to restart the search from that CPU again.
1972 if (rt_rq->push_flags & RT_PUSH_IPI_RESTART) {
1973 rt_rq->push_flags &= ~RT_PUSH_IPI_RESTART;
1974 rt_rq->push_cpu = src_rq->cpu;
1977 cpu = find_next_push_cpu(src_rq);
1979 if (cpu >= nr_cpu_ids)
1980 rt_rq->push_flags &= ~RT_PUSH_IPI_EXECUTING;
1981 raw_spin_unlock(&rt_rq->push_lock);
1983 if (cpu >= nr_cpu_ids)
1987 * It is possible that a restart caused this CPU to be
1988 * chosen again. Don't bother with an IPI, just see if we
1989 * have more to push.
1991 if (unlikely(cpu == rq->cpu))
1994 /* Try the next RT overloaded CPU */
1995 irq_work_queue_on(&rt_rq->push_work, cpu);
1998 static void push_irq_work_func(struct irq_work *work)
2000 struct rt_rq *rt_rq = container_of(work, struct rt_rq, push_work);
2002 try_to_push_tasks(rt_rq);
2004 #endif /* HAVE_RT_PUSH_IPI */
2006 static void pull_rt_task(struct rq *this_rq)
2008 int this_cpu = this_rq->cpu, cpu;
2009 bool resched = false;
2010 struct task_struct *p;
2013 if (likely(!rt_overloaded(this_rq)))
2017 * Match the barrier from rt_set_overloaded; this guarantees that if we
2018 * see overloaded we must also see the rto_mask bit.
2022 #ifdef HAVE_RT_PUSH_IPI
2023 if (sched_feat(RT_PUSH_IPI)) {
2024 tell_cpu_to_push(this_rq);
2029 for_each_cpu(cpu, this_rq->rd->rto_mask) {
2030 if (this_cpu == cpu)
2033 src_rq = cpu_rq(cpu);
2036 * Don't bother taking the src_rq->lock if the next highest
2037 * task is known to be lower-priority than our current task.
2038 * This may look racy, but if this value is about to go
2039 * logically higher, the src_rq will push this task away.
2040 * And if its going logically lower, we do not care
2042 if (src_rq->rt.highest_prio.next >=
2043 this_rq->rt.highest_prio.curr)
2047 * We can potentially drop this_rq's lock in
2048 * double_lock_balance, and another CPU could
2051 double_lock_balance(this_rq, src_rq);
2054 * We can pull only a task, which is pushable
2055 * on its rq, and no others.
2057 p = pick_highest_pushable_task(src_rq, this_cpu);
2060 * Do we have an RT task that preempts
2061 * the to-be-scheduled task?
2063 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2064 WARN_ON(p == src_rq->curr);
2065 WARN_ON(!task_on_rq_queued(p));
2068 * There's a chance that p is higher in priority
2069 * than what's currently running on its cpu.
2070 * This is just that p is wakeing up and hasn't
2071 * had a chance to schedule. We only pull
2072 * p if it is lower in priority than the
2073 * current task on the run queue
2075 if (p->prio < src_rq->curr->prio)
2080 deactivate_task(src_rq, p, 0);
2081 set_task_cpu(p, this_cpu);
2082 activate_task(this_rq, p, 0);
2084 * We continue with the search, just in
2085 * case there's an even higher prio task
2086 * in another runqueue. (low likelihood
2091 double_unlock_balance(this_rq, src_rq);
2095 resched_curr(this_rq);
2099 * If we are not running and we are not going to reschedule soon, we should
2100 * try to push tasks away now
2102 static void task_woken_rt(struct rq *rq, struct task_struct *p)
2104 if (!task_running(rq, p) &&
2105 !test_tsk_need_resched(rq->curr) &&
2106 p->nr_cpus_allowed > 1 &&
2107 (dl_task(rq->curr) || rt_task(rq->curr)) &&
2108 (rq->curr->nr_cpus_allowed < 2 ||
2109 rq->curr->prio <= p->prio))
2113 /* Assumes rq->lock is held */
2114 static void rq_online_rt(struct rq *rq)
2116 if (rq->rt.overloaded)
2117 rt_set_overload(rq);
2119 __enable_runtime(rq);
2121 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2124 /* Assumes rq->lock is held */
2125 static void rq_offline_rt(struct rq *rq)
2127 if (rq->rt.overloaded)
2128 rt_clear_overload(rq);
2130 __disable_runtime(rq);
2132 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2136 * When switch from the rt queue, we bring ourselves to a position
2137 * that we might want to pull RT tasks from other runqueues.
2139 static void switched_from_rt(struct rq *rq, struct task_struct *p)
2142 * If there are other RT tasks then we will reschedule
2143 * and the scheduling of the other RT tasks will handle
2144 * the balancing. But if we are the last RT task
2145 * we may need to handle the pulling of RT tasks
2148 if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2151 queue_pull_task(rq);
2154 void __init init_sched_rt_class(void)
2158 for_each_possible_cpu(i) {
2159 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2160 GFP_KERNEL, cpu_to_node(i));
2163 #endif /* CONFIG_SMP */
2166 * When switching a task to RT, we may overload the runqueue
2167 * with RT tasks. In this case we try to push them off to
2170 static void switched_to_rt(struct rq *rq, struct task_struct *p)
2173 * If we are already running, then there's nothing
2174 * that needs to be done. But if we are not running
2175 * we may need to preempt the current running task.
2176 * If that current running task is also an RT task
2177 * then see if we can move to another run queue.
2179 if (task_on_rq_queued(p) && rq->curr != p) {
2181 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded)
2182 queue_push_tasks(rq);
2184 if (p->prio < rq->curr->prio)
2186 #endif /* CONFIG_SMP */
2191 * Priority of the task has changed. This may cause
2192 * us to initiate a push or pull.
2195 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2197 if (!task_on_rq_queued(p))
2200 if (rq->curr == p) {
2203 * If our priority decreases while running, we
2204 * may need to pull tasks to this runqueue.
2206 if (oldprio < p->prio)
2207 queue_pull_task(rq);
2210 * If there's a higher priority task waiting to run
2213 if (p->prio > rq->rt.highest_prio.curr)
2216 /* For UP simply resched on drop of prio */
2217 if (oldprio < p->prio)
2219 #endif /* CONFIG_SMP */
2222 * This task is not running, but if it is
2223 * greater than the current running task
2226 if (p->prio < rq->curr->prio)
2231 static void watchdog(struct rq *rq, struct task_struct *p)
2233 unsigned long soft, hard;
2235 /* max may change after cur was read, this will be fixed next tick */
2236 soft = task_rlimit(p, RLIMIT_RTTIME);
2237 hard = task_rlimit_max(p, RLIMIT_RTTIME);
2239 if (soft != RLIM_INFINITY) {
2242 if (p->rt.watchdog_stamp != jiffies) {
2244 p->rt.watchdog_stamp = jiffies;
2247 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2248 if (p->rt.timeout > next)
2249 p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
2253 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2255 struct sched_rt_entity *rt_se = &p->rt;
2259 if (rq->rt.rt_nr_running)
2260 sched_rt_update_capacity_req(rq);
2265 * RR tasks need a special form of timeslice management.
2266 * FIFO tasks have no timeslices.
2268 if (p->policy != SCHED_RR)
2271 if (--p->rt.time_slice)
2274 p->rt.time_slice = sched_rr_timeslice;
2277 * Requeue to the end of queue if we (and all of our ancestors) are not
2278 * the only element on the queue
2280 for_each_sched_rt_entity(rt_se) {
2281 if (rt_se->run_list.prev != rt_se->run_list.next) {
2282 requeue_task_rt(rq, p, 0);
2289 static void set_curr_task_rt(struct rq *rq)
2291 struct task_struct *p = rq->curr;
2293 p->se.exec_start = rq_clock_task(rq);
2295 /* The running task is never eligible for pushing */
2296 dequeue_pushable_task(rq, p);
2299 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2302 * Time slice is 0 for SCHED_FIFO tasks
2304 if (task->policy == SCHED_RR)
2305 return sched_rr_timeslice;
2310 const struct sched_class rt_sched_class = {
2311 .next = &fair_sched_class,
2312 .enqueue_task = enqueue_task_rt,
2313 .dequeue_task = dequeue_task_rt,
2314 .yield_task = yield_task_rt,
2316 .check_preempt_curr = check_preempt_curr_rt,
2318 .pick_next_task = pick_next_task_rt,
2319 .put_prev_task = put_prev_task_rt,
2322 .select_task_rq = select_task_rq_rt,
2324 .set_cpus_allowed = set_cpus_allowed_common,
2325 .rq_online = rq_online_rt,
2326 .rq_offline = rq_offline_rt,
2327 .task_woken = task_woken_rt,
2328 .switched_from = switched_from_rt,
2331 .set_curr_task = set_curr_task_rt,
2332 .task_tick = task_tick_rt,
2334 .get_rr_interval = get_rr_interval_rt,
2336 .prio_changed = prio_changed_rt,
2337 .switched_to = switched_to_rt,
2339 .update_curr = update_curr_rt,
2342 #ifdef CONFIG_SCHED_DEBUG
2343 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2345 void print_rt_stats(struct seq_file *m, int cpu)
2348 struct rt_rq *rt_rq;
2351 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2352 print_rt_rq(m, cpu, rt_rq);
2355 #endif /* CONFIG_SCHED_DEBUG */