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;
12 int sysctl_sched_rr_timeslice = (MSEC_PER_SEC / HZ) * RR_TIMESLICE;
14 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
16 struct rt_bandwidth def_rt_bandwidth;
18 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
20 struct rt_bandwidth *rt_b =
21 container_of(timer, struct rt_bandwidth, rt_period_timer);
25 raw_spin_lock(&rt_b->rt_runtime_lock);
27 overrun = hrtimer_forward_now(timer, rt_b->rt_period);
31 raw_spin_unlock(&rt_b->rt_runtime_lock);
32 idle = do_sched_rt_period_timer(rt_b, overrun);
33 raw_spin_lock(&rt_b->rt_runtime_lock);
36 rt_b->rt_period_active = 0;
37 raw_spin_unlock(&rt_b->rt_runtime_lock);
39 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
42 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
44 rt_b->rt_period = ns_to_ktime(period);
45 rt_b->rt_runtime = runtime;
47 raw_spin_lock_init(&rt_b->rt_runtime_lock);
49 hrtimer_init(&rt_b->rt_period_timer,
50 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
51 rt_b->rt_period_timer.function = sched_rt_period_timer;
54 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
56 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
59 raw_spin_lock(&rt_b->rt_runtime_lock);
60 if (!rt_b->rt_period_active) {
61 rt_b->rt_period_active = 1;
63 * SCHED_DEADLINE updates the bandwidth, as a run away
64 * RT task with a DL task could hog a CPU. But DL does
65 * not reset the period. If a deadline task was running
66 * without an RT task running, it can cause RT tasks to
67 * throttle when they start up. Kick the timer right away
68 * to update the period.
70 hrtimer_forward_now(&rt_b->rt_period_timer, ns_to_ktime(0));
71 hrtimer_start_expires(&rt_b->rt_period_timer, HRTIMER_MODE_ABS_PINNED);
73 raw_spin_unlock(&rt_b->rt_runtime_lock);
76 #if defined(CONFIG_SMP) && defined(HAVE_RT_PUSH_IPI)
77 static void push_irq_work_func(struct irq_work *work);
80 void init_rt_rq(struct rt_rq *rt_rq)
82 struct rt_prio_array *array;
85 array = &rt_rq->active;
86 for (i = 0; i < MAX_RT_PRIO; i++) {
87 INIT_LIST_HEAD(array->queue + i);
88 __clear_bit(i, array->bitmap);
90 /* delimiter for bitsearch: */
91 __set_bit(MAX_RT_PRIO, array->bitmap);
93 #if defined CONFIG_SMP
94 rt_rq->highest_prio.curr = MAX_RT_PRIO;
95 rt_rq->highest_prio.next = MAX_RT_PRIO;
96 rt_rq->rt_nr_migratory = 0;
97 rt_rq->overloaded = 0;
98 plist_head_init(&rt_rq->pushable_tasks);
100 #ifdef HAVE_RT_PUSH_IPI
101 rt_rq->push_flags = 0;
102 rt_rq->push_cpu = nr_cpu_ids;
103 raw_spin_lock_init(&rt_rq->push_lock);
104 init_irq_work(&rt_rq->push_work, push_irq_work_func);
106 #endif /* CONFIG_SMP */
107 /* We start is dequeued state, because no RT tasks are queued */
108 rt_rq->rt_queued = 0;
111 rt_rq->rt_throttled = 0;
112 rt_rq->rt_runtime = 0;
113 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
116 #ifdef CONFIG_RT_GROUP_SCHED
117 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
119 hrtimer_cancel(&rt_b->rt_period_timer);
122 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
124 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
126 #ifdef CONFIG_SCHED_DEBUG
127 WARN_ON_ONCE(!rt_entity_is_task(rt_se));
129 return container_of(rt_se, struct task_struct, rt);
132 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
137 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
142 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
144 struct rt_rq *rt_rq = rt_se->rt_rq;
149 void free_rt_sched_group(struct task_group *tg)
154 destroy_rt_bandwidth(&tg->rt_bandwidth);
156 for_each_possible_cpu(i) {
167 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
168 struct sched_rt_entity *rt_se, int cpu,
169 struct sched_rt_entity *parent)
171 struct rq *rq = cpu_rq(cpu);
173 rt_rq->highest_prio.curr = MAX_RT_PRIO;
174 rt_rq->rt_nr_boosted = 0;
178 tg->rt_rq[cpu] = rt_rq;
179 tg->rt_se[cpu] = rt_se;
185 rt_se->rt_rq = &rq->rt;
187 rt_se->rt_rq = parent->my_q;
190 rt_se->parent = parent;
191 INIT_LIST_HEAD(&rt_se->run_list);
194 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
197 struct sched_rt_entity *rt_se;
200 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
203 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
207 init_rt_bandwidth(&tg->rt_bandwidth,
208 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
210 for_each_possible_cpu(i) {
211 rt_rq = kzalloc_node(sizeof(struct rt_rq),
212 GFP_KERNEL, cpu_to_node(i));
216 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
217 GFP_KERNEL, cpu_to_node(i));
222 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
223 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
234 #else /* CONFIG_RT_GROUP_SCHED */
236 #define rt_entity_is_task(rt_se) (1)
238 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
240 return container_of(rt_se, struct task_struct, rt);
243 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
245 return container_of(rt_rq, struct rq, rt);
248 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
250 struct task_struct *p = rt_task_of(rt_se);
255 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
257 struct rq *rq = rq_of_rt_se(rt_se);
262 void free_rt_sched_group(struct task_group *tg) { }
264 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
268 #endif /* CONFIG_RT_GROUP_SCHED */
272 static void pull_rt_task(struct rq *this_rq);
274 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
276 /* Try to pull RT tasks here if we lower this rq's prio */
277 return rq->rt.highest_prio.curr > prev->prio;
280 static inline int rt_overloaded(struct rq *rq)
282 return atomic_read(&rq->rd->rto_count);
285 static inline void rt_set_overload(struct rq *rq)
290 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
292 * Make sure the mask is visible before we set
293 * the overload count. That is checked to determine
294 * if we should look at the mask. It would be a shame
295 * if we looked at the mask, but the mask was not
298 * Matched by the barrier in pull_rt_task().
301 atomic_inc(&rq->rd->rto_count);
304 static inline void rt_clear_overload(struct rq *rq)
309 /* the order here really doesn't matter */
310 atomic_dec(&rq->rd->rto_count);
311 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
314 static void update_rt_migration(struct rt_rq *rt_rq)
316 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
317 if (!rt_rq->overloaded) {
318 rt_set_overload(rq_of_rt_rq(rt_rq));
319 rt_rq->overloaded = 1;
321 } else if (rt_rq->overloaded) {
322 rt_clear_overload(rq_of_rt_rq(rt_rq));
323 rt_rq->overloaded = 0;
327 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
329 struct task_struct *p;
331 if (!rt_entity_is_task(rt_se))
334 p = rt_task_of(rt_se);
335 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
337 rt_rq->rt_nr_total++;
338 if (tsk_nr_cpus_allowed(p) > 1)
339 rt_rq->rt_nr_migratory++;
341 update_rt_migration(rt_rq);
344 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
346 struct task_struct *p;
348 if (!rt_entity_is_task(rt_se))
351 p = rt_task_of(rt_se);
352 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
354 rt_rq->rt_nr_total--;
355 if (tsk_nr_cpus_allowed(p) > 1)
356 rt_rq->rt_nr_migratory--;
358 update_rt_migration(rt_rq);
361 static inline int has_pushable_tasks(struct rq *rq)
363 return !plist_head_empty(&rq->rt.pushable_tasks);
366 static DEFINE_PER_CPU(struct callback_head, rt_push_head);
367 static DEFINE_PER_CPU(struct callback_head, rt_pull_head);
369 static void push_rt_tasks(struct rq *);
370 static void pull_rt_task(struct rq *);
372 static inline void queue_push_tasks(struct rq *rq)
374 if (!has_pushable_tasks(rq))
377 queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks);
380 static inline void queue_pull_task(struct rq *rq)
382 queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task);
385 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
387 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
388 plist_node_init(&p->pushable_tasks, p->prio);
389 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
391 /* Update the highest prio pushable task */
392 if (p->prio < rq->rt.highest_prio.next)
393 rq->rt.highest_prio.next = p->prio;
396 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
398 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
400 /* Update the new highest prio pushable task */
401 if (has_pushable_tasks(rq)) {
402 p = plist_first_entry(&rq->rt.pushable_tasks,
403 struct task_struct, pushable_tasks);
404 rq->rt.highest_prio.next = p->prio;
406 rq->rt.highest_prio.next = MAX_RT_PRIO;
411 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
415 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
420 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
425 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
429 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
434 static inline void pull_rt_task(struct rq *this_rq)
438 static inline void queue_push_tasks(struct rq *rq)
441 #endif /* CONFIG_SMP */
443 static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
444 static void dequeue_top_rt_rq(struct rt_rq *rt_rq);
446 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
451 #ifdef CONFIG_RT_GROUP_SCHED
453 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
458 return rt_rq->rt_runtime;
461 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
463 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
466 typedef struct task_group *rt_rq_iter_t;
468 static inline struct task_group *next_task_group(struct task_group *tg)
471 tg = list_entry_rcu(tg->list.next,
472 typeof(struct task_group), list);
473 } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
475 if (&tg->list == &task_groups)
481 #define for_each_rt_rq(rt_rq, iter, rq) \
482 for (iter = container_of(&task_groups, typeof(*iter), list); \
483 (iter = next_task_group(iter)) && \
484 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
486 #define for_each_sched_rt_entity(rt_se) \
487 for (; rt_se; rt_se = rt_se->parent)
489 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
494 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
495 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
497 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
499 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
500 struct rq *rq = rq_of_rt_rq(rt_rq);
501 struct sched_rt_entity *rt_se;
503 int cpu = cpu_of(rq);
505 rt_se = rt_rq->tg->rt_se[cpu];
507 if (rt_rq->rt_nr_running) {
509 enqueue_top_rt_rq(rt_rq);
510 else if (!on_rt_rq(rt_se))
511 enqueue_rt_entity(rt_se, 0);
513 if (rt_rq->highest_prio.curr < curr->prio)
518 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
520 struct sched_rt_entity *rt_se;
521 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
523 rt_se = rt_rq->tg->rt_se[cpu];
526 dequeue_top_rt_rq(rt_rq);
527 else if (on_rt_rq(rt_se))
528 dequeue_rt_entity(rt_se, 0);
531 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
533 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
536 static int rt_se_boosted(struct sched_rt_entity *rt_se)
538 struct rt_rq *rt_rq = group_rt_rq(rt_se);
539 struct task_struct *p;
542 return !!rt_rq->rt_nr_boosted;
544 p = rt_task_of(rt_se);
545 return p->prio != p->normal_prio;
549 static inline const struct cpumask *sched_rt_period_mask(void)
551 return this_rq()->rd->span;
554 static inline const struct cpumask *sched_rt_period_mask(void)
556 return cpu_online_mask;
561 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
563 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
566 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
568 return &rt_rq->tg->rt_bandwidth;
571 #else /* !CONFIG_RT_GROUP_SCHED */
573 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
575 return rt_rq->rt_runtime;
578 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
580 return ktime_to_ns(def_rt_bandwidth.rt_period);
583 typedef struct rt_rq *rt_rq_iter_t;
585 #define for_each_rt_rq(rt_rq, iter, rq) \
586 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
588 #define for_each_sched_rt_entity(rt_se) \
589 for (; rt_se; rt_se = NULL)
591 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
596 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
598 struct rq *rq = rq_of_rt_rq(rt_rq);
600 if (!rt_rq->rt_nr_running)
603 enqueue_top_rt_rq(rt_rq);
607 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
609 dequeue_top_rt_rq(rt_rq);
612 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
614 return rt_rq->rt_throttled;
617 static inline const struct cpumask *sched_rt_period_mask(void)
619 return cpu_online_mask;
623 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
625 return &cpu_rq(cpu)->rt;
628 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
630 return &def_rt_bandwidth;
633 #endif /* CONFIG_RT_GROUP_SCHED */
635 bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
637 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
639 return (hrtimer_active(&rt_b->rt_period_timer) ||
640 rt_rq->rt_time < rt_b->rt_runtime);
645 * We ran out of runtime, see if we can borrow some from our neighbours.
647 static void do_balance_runtime(struct rt_rq *rt_rq)
649 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
650 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
654 weight = cpumask_weight(rd->span);
656 raw_spin_lock(&rt_b->rt_runtime_lock);
657 rt_period = ktime_to_ns(rt_b->rt_period);
658 for_each_cpu(i, rd->span) {
659 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
665 raw_spin_lock(&iter->rt_runtime_lock);
667 * Either all rqs have inf runtime and there's nothing to steal
668 * or __disable_runtime() below sets a specific rq to inf to
669 * indicate its been disabled and disalow stealing.
671 if (iter->rt_runtime == RUNTIME_INF)
675 * From runqueues with spare time, take 1/n part of their
676 * spare time, but no more than our period.
678 diff = iter->rt_runtime - iter->rt_time;
680 diff = div_u64((u64)diff, weight);
681 if (rt_rq->rt_runtime + diff > rt_period)
682 diff = rt_period - rt_rq->rt_runtime;
683 iter->rt_runtime -= diff;
684 rt_rq->rt_runtime += diff;
685 if (rt_rq->rt_runtime == rt_period) {
686 raw_spin_unlock(&iter->rt_runtime_lock);
691 raw_spin_unlock(&iter->rt_runtime_lock);
693 raw_spin_unlock(&rt_b->rt_runtime_lock);
697 * Ensure this RQ takes back all the runtime it lend to its neighbours.
699 static void __disable_runtime(struct rq *rq)
701 struct root_domain *rd = rq->rd;
705 if (unlikely(!scheduler_running))
708 for_each_rt_rq(rt_rq, iter, rq) {
709 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
713 raw_spin_lock(&rt_b->rt_runtime_lock);
714 raw_spin_lock(&rt_rq->rt_runtime_lock);
716 * Either we're all inf and nobody needs to borrow, or we're
717 * already disabled and thus have nothing to do, or we have
718 * exactly the right amount of runtime to take out.
720 if (rt_rq->rt_runtime == RUNTIME_INF ||
721 rt_rq->rt_runtime == rt_b->rt_runtime)
723 raw_spin_unlock(&rt_rq->rt_runtime_lock);
726 * Calculate the difference between what we started out with
727 * and what we current have, that's the amount of runtime
728 * we lend and now have to reclaim.
730 want = rt_b->rt_runtime - rt_rq->rt_runtime;
733 * Greedy reclaim, take back as much as we can.
735 for_each_cpu(i, rd->span) {
736 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
740 * Can't reclaim from ourselves or disabled runqueues.
742 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
745 raw_spin_lock(&iter->rt_runtime_lock);
747 diff = min_t(s64, iter->rt_runtime, want);
748 iter->rt_runtime -= diff;
751 iter->rt_runtime -= want;
754 raw_spin_unlock(&iter->rt_runtime_lock);
760 raw_spin_lock(&rt_rq->rt_runtime_lock);
762 * We cannot be left wanting - that would mean some runtime
763 * leaked out of the system.
768 * Disable all the borrow logic by pretending we have inf
769 * runtime - in which case borrowing doesn't make sense.
771 rt_rq->rt_runtime = RUNTIME_INF;
772 rt_rq->rt_throttled = 0;
773 raw_spin_unlock(&rt_rq->rt_runtime_lock);
774 raw_spin_unlock(&rt_b->rt_runtime_lock);
776 /* Make rt_rq available for pick_next_task() */
777 sched_rt_rq_enqueue(rt_rq);
781 static void __enable_runtime(struct rq *rq)
786 if (unlikely(!scheduler_running))
790 * Reset each runqueue's bandwidth settings
792 for_each_rt_rq(rt_rq, iter, rq) {
793 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
795 raw_spin_lock(&rt_b->rt_runtime_lock);
796 raw_spin_lock(&rt_rq->rt_runtime_lock);
797 rt_rq->rt_runtime = rt_b->rt_runtime;
799 rt_rq->rt_throttled = 0;
800 raw_spin_unlock(&rt_rq->rt_runtime_lock);
801 raw_spin_unlock(&rt_b->rt_runtime_lock);
805 static void balance_runtime(struct rt_rq *rt_rq)
807 if (!sched_feat(RT_RUNTIME_SHARE))
810 if (rt_rq->rt_time > rt_rq->rt_runtime) {
811 raw_spin_unlock(&rt_rq->rt_runtime_lock);
812 do_balance_runtime(rt_rq);
813 raw_spin_lock(&rt_rq->rt_runtime_lock);
816 #else /* !CONFIG_SMP */
817 static inline void balance_runtime(struct rt_rq *rt_rq) {}
818 #endif /* CONFIG_SMP */
820 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
822 int i, idle = 1, throttled = 0;
823 const struct cpumask *span;
825 span = sched_rt_period_mask();
826 #ifdef CONFIG_RT_GROUP_SCHED
828 * FIXME: isolated CPUs should really leave the root task group,
829 * whether they are isolcpus or were isolated via cpusets, lest
830 * the timer run on a CPU which does not service all runqueues,
831 * potentially leaving other CPUs indefinitely throttled. If
832 * isolation is really required, the user will turn the throttle
833 * off to kill the perturbations it causes anyway. Meanwhile,
834 * this maintains functionality for boot and/or troubleshooting.
836 if (rt_b == &root_task_group.rt_bandwidth)
837 span = cpu_online_mask;
839 for_each_cpu(i, span) {
841 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
842 struct rq *rq = rq_of_rt_rq(rt_rq);
844 raw_spin_lock(&rq->lock);
845 if (rt_rq->rt_time) {
848 raw_spin_lock(&rt_rq->rt_runtime_lock);
849 if (rt_rq->rt_throttled)
850 balance_runtime(rt_rq);
851 runtime = rt_rq->rt_runtime;
852 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
853 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
854 rt_rq->rt_throttled = 0;
858 * When we're idle and a woken (rt) task is
859 * throttled check_preempt_curr() will set
860 * skip_update and the time between the wakeup
861 * and this unthrottle will get accounted as
864 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
865 rq_clock_skip_update(rq, false);
867 if (rt_rq->rt_time || rt_rq->rt_nr_running)
869 raw_spin_unlock(&rt_rq->rt_runtime_lock);
870 } else if (rt_rq->rt_nr_running) {
872 if (!rt_rq_throttled(rt_rq))
875 if (rt_rq->rt_throttled)
879 sched_rt_rq_enqueue(rt_rq);
880 raw_spin_unlock(&rq->lock);
883 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
889 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
891 #ifdef CONFIG_RT_GROUP_SCHED
892 struct rt_rq *rt_rq = group_rt_rq(rt_se);
895 return rt_rq->highest_prio.curr;
898 return rt_task_of(rt_se)->prio;
901 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
903 u64 runtime = sched_rt_runtime(rt_rq);
905 if (rt_rq->rt_throttled)
906 return rt_rq_throttled(rt_rq);
908 if (runtime >= sched_rt_period(rt_rq))
911 balance_runtime(rt_rq);
912 runtime = sched_rt_runtime(rt_rq);
913 if (runtime == RUNTIME_INF)
916 if (rt_rq->rt_time > runtime) {
917 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
920 * Don't actually throttle groups that have no runtime assigned
921 * but accrue some time due to boosting.
923 if (likely(rt_b->rt_runtime)) {
924 rt_rq->rt_throttled = 1;
925 printk_deferred_once("sched: RT throttling activated\n");
928 * In case we did anyway, make it go away,
929 * replenishment is a joke, since it will replenish us
935 if (rt_rq_throttled(rt_rq)) {
936 sched_rt_rq_dequeue(rt_rq);
945 * Update the current task's runtime statistics. Skip current tasks that
946 * are not in our scheduling class.
948 static void update_curr_rt(struct rq *rq)
950 struct task_struct *curr = rq->curr;
951 struct sched_rt_entity *rt_se = &curr->rt;
954 if (curr->sched_class != &rt_sched_class)
957 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
958 if (unlikely((s64)delta_exec <= 0))
961 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
962 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_RT);
964 schedstat_set(curr->se.statistics.exec_max,
965 max(curr->se.statistics.exec_max, delta_exec));
967 curr->se.sum_exec_runtime += delta_exec;
968 account_group_exec_runtime(curr, delta_exec);
970 curr->se.exec_start = rq_clock_task(rq);
971 cpuacct_charge(curr, delta_exec);
973 sched_rt_avg_update(rq, delta_exec);
975 if (!rt_bandwidth_enabled())
978 for_each_sched_rt_entity(rt_se) {
979 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
981 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
982 raw_spin_lock(&rt_rq->rt_runtime_lock);
983 rt_rq->rt_time += delta_exec;
984 if (sched_rt_runtime_exceeded(rt_rq))
986 raw_spin_unlock(&rt_rq->rt_runtime_lock);
992 dequeue_top_rt_rq(struct rt_rq *rt_rq)
994 struct rq *rq = rq_of_rt_rq(rt_rq);
996 BUG_ON(&rq->rt != rt_rq);
998 if (!rt_rq->rt_queued)
1001 BUG_ON(!rq->nr_running);
1003 sub_nr_running(rq, rt_rq->rt_nr_running);
1004 rt_rq->rt_queued = 0;
1008 enqueue_top_rt_rq(struct rt_rq *rt_rq)
1010 struct rq *rq = rq_of_rt_rq(rt_rq);
1012 BUG_ON(&rq->rt != rt_rq);
1014 if (rt_rq->rt_queued)
1016 if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running)
1019 add_nr_running(rq, rt_rq->rt_nr_running);
1020 rt_rq->rt_queued = 1;
1023 #if defined CONFIG_SMP
1026 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1028 struct rq *rq = rq_of_rt_rq(rt_rq);
1030 #ifdef CONFIG_RT_GROUP_SCHED
1032 * Change rq's cpupri only if rt_rq is the top queue.
1034 if (&rq->rt != rt_rq)
1037 if (rq->online && prio < prev_prio)
1038 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1042 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1044 struct rq *rq = rq_of_rt_rq(rt_rq);
1046 #ifdef CONFIG_RT_GROUP_SCHED
1048 * Change rq's cpupri only if rt_rq is the top queue.
1050 if (&rq->rt != rt_rq)
1053 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1054 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1057 #else /* CONFIG_SMP */
1060 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1062 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1064 #endif /* CONFIG_SMP */
1066 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1068 inc_rt_prio(struct rt_rq *rt_rq, int prio)
1070 int prev_prio = rt_rq->highest_prio.curr;
1072 if (prio < prev_prio)
1073 rt_rq->highest_prio.curr = prio;
1075 inc_rt_prio_smp(rt_rq, prio, prev_prio);
1079 dec_rt_prio(struct rt_rq *rt_rq, int prio)
1081 int prev_prio = rt_rq->highest_prio.curr;
1083 if (rt_rq->rt_nr_running) {
1085 WARN_ON(prio < prev_prio);
1088 * This may have been our highest task, and therefore
1089 * we may have some recomputation to do
1091 if (prio == prev_prio) {
1092 struct rt_prio_array *array = &rt_rq->active;
1094 rt_rq->highest_prio.curr =
1095 sched_find_first_bit(array->bitmap);
1099 rt_rq->highest_prio.curr = MAX_RT_PRIO;
1101 dec_rt_prio_smp(rt_rq, prio, prev_prio);
1106 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1107 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1109 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1111 #ifdef CONFIG_RT_GROUP_SCHED
1114 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1116 if (rt_se_boosted(rt_se))
1117 rt_rq->rt_nr_boosted++;
1120 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1124 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1126 if (rt_se_boosted(rt_se))
1127 rt_rq->rt_nr_boosted--;
1129 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1132 #else /* CONFIG_RT_GROUP_SCHED */
1135 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1137 start_rt_bandwidth(&def_rt_bandwidth);
1141 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1143 #endif /* CONFIG_RT_GROUP_SCHED */
1146 unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1148 struct rt_rq *group_rq = group_rt_rq(rt_se);
1151 return group_rq->rt_nr_running;
1157 unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1159 struct rt_rq *group_rq = group_rt_rq(rt_se);
1160 struct task_struct *tsk;
1163 return group_rq->rr_nr_running;
1165 tsk = rt_task_of(rt_se);
1167 return (tsk->policy == SCHED_RR) ? 1 : 0;
1171 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1173 int prio = rt_se_prio(rt_se);
1175 WARN_ON(!rt_prio(prio));
1176 rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1177 rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
1179 inc_rt_prio(rt_rq, prio);
1180 inc_rt_migration(rt_se, rt_rq);
1181 inc_rt_group(rt_se, rt_rq);
1185 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1187 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1188 WARN_ON(!rt_rq->rt_nr_running);
1189 rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1190 rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
1192 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1193 dec_rt_migration(rt_se, rt_rq);
1194 dec_rt_group(rt_se, rt_rq);
1198 * Change rt_se->run_list location unless SAVE && !MOVE
1200 * assumes ENQUEUE/DEQUEUE flags match
1202 static inline bool move_entity(unsigned int flags)
1204 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1210 static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1212 list_del_init(&rt_se->run_list);
1214 if (list_empty(array->queue + rt_se_prio(rt_se)))
1215 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1220 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1222 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1223 struct rt_prio_array *array = &rt_rq->active;
1224 struct rt_rq *group_rq = group_rt_rq(rt_se);
1225 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1228 * Don't enqueue the group if its throttled, or when empty.
1229 * The latter is a consequence of the former when a child group
1230 * get throttled and the current group doesn't have any other
1233 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1235 __delist_rt_entity(rt_se, array);
1239 if (move_entity(flags)) {
1240 WARN_ON_ONCE(rt_se->on_list);
1241 if (flags & ENQUEUE_HEAD)
1242 list_add(&rt_se->run_list, queue);
1244 list_add_tail(&rt_se->run_list, queue);
1246 __set_bit(rt_se_prio(rt_se), array->bitmap);
1251 inc_rt_tasks(rt_se, rt_rq);
1254 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1256 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1257 struct rt_prio_array *array = &rt_rq->active;
1259 if (move_entity(flags)) {
1260 WARN_ON_ONCE(!rt_se->on_list);
1261 __delist_rt_entity(rt_se, array);
1265 dec_rt_tasks(rt_se, rt_rq);
1269 * Because the prio of an upper entry depends on the lower
1270 * entries, we must remove entries top - down.
1272 static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
1274 struct sched_rt_entity *back = NULL;
1276 for_each_sched_rt_entity(rt_se) {
1281 dequeue_top_rt_rq(rt_rq_of_se(back));
1283 for (rt_se = back; rt_se; rt_se = rt_se->back) {
1284 if (on_rt_rq(rt_se))
1285 __dequeue_rt_entity(rt_se, flags);
1289 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1291 struct rq *rq = rq_of_rt_se(rt_se);
1293 dequeue_rt_stack(rt_se, flags);
1294 for_each_sched_rt_entity(rt_se)
1295 __enqueue_rt_entity(rt_se, flags);
1296 enqueue_top_rt_rq(&rq->rt);
1299 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1301 struct rq *rq = rq_of_rt_se(rt_se);
1303 dequeue_rt_stack(rt_se, flags);
1305 for_each_sched_rt_entity(rt_se) {
1306 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1308 if (rt_rq && rt_rq->rt_nr_running)
1309 __enqueue_rt_entity(rt_se, flags);
1311 enqueue_top_rt_rq(&rq->rt);
1315 * Adding/removing a task to/from a priority array:
1318 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1320 struct sched_rt_entity *rt_se = &p->rt;
1322 if (flags & ENQUEUE_WAKEUP)
1325 enqueue_rt_entity(rt_se, flags);
1327 if (!task_current(rq, p) && tsk_nr_cpus_allowed(p) > 1)
1328 enqueue_pushable_task(rq, p);
1331 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1333 struct sched_rt_entity *rt_se = &p->rt;
1336 dequeue_rt_entity(rt_se, flags);
1338 dequeue_pushable_task(rq, p);
1342 * Put task to the head or the end of the run list without the overhead of
1343 * dequeue followed by enqueue.
1346 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1348 if (on_rt_rq(rt_se)) {
1349 struct rt_prio_array *array = &rt_rq->active;
1350 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1353 list_move(&rt_se->run_list, queue);
1355 list_move_tail(&rt_se->run_list, queue);
1359 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1361 struct sched_rt_entity *rt_se = &p->rt;
1362 struct rt_rq *rt_rq;
1364 for_each_sched_rt_entity(rt_se) {
1365 rt_rq = rt_rq_of_se(rt_se);
1366 requeue_rt_entity(rt_rq, rt_se, head);
1370 static void yield_task_rt(struct rq *rq)
1372 requeue_task_rt(rq, rq->curr, 0);
1376 static int find_lowest_rq(struct task_struct *task);
1379 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1381 struct task_struct *curr;
1384 /* For anything but wake ups, just return the task_cpu */
1385 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1391 curr = READ_ONCE(rq->curr); /* unlocked access */
1394 * If the current task on @p's runqueue is an RT task, then
1395 * try to see if we can wake this RT task up on another
1396 * runqueue. Otherwise simply start this RT task
1397 * on its current runqueue.
1399 * We want to avoid overloading runqueues. If the woken
1400 * task is a higher priority, then it will stay on this CPU
1401 * and the lower prio task should be moved to another CPU.
1402 * Even though this will probably make the lower prio task
1403 * lose its cache, we do not want to bounce a higher task
1404 * around just because it gave up its CPU, perhaps for a
1407 * For equal prio tasks, we just let the scheduler sort it out.
1409 * Otherwise, just let it ride on the affined RQ and the
1410 * post-schedule router will push the preempted task away
1412 * This test is optimistic, if we get it wrong the load-balancer
1413 * will have to sort it out.
1415 if (curr && unlikely(rt_task(curr)) &&
1416 (tsk_nr_cpus_allowed(curr) < 2 ||
1417 curr->prio <= p->prio)) {
1418 int target = find_lowest_rq(p);
1421 * Don't bother moving it if the destination CPU is
1422 * not running a lower priority task.
1425 p->prio < cpu_rq(target)->rt.highest_prio.curr)
1434 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1437 * Current can't be migrated, useless to reschedule,
1438 * let's hope p can move out.
1440 if (tsk_nr_cpus_allowed(rq->curr) == 1 ||
1441 !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1445 * p is migratable, so let's not schedule it and
1446 * see if it is pushed or pulled somewhere else.
1448 if (tsk_nr_cpus_allowed(p) != 1
1449 && cpupri_find(&rq->rd->cpupri, p, NULL))
1453 * There appears to be other cpus that can accept
1454 * current and none to run 'p', so lets reschedule
1455 * to try and push current away:
1457 requeue_task_rt(rq, p, 1);
1461 #endif /* CONFIG_SMP */
1464 * Preempt the current task with a newly woken task if needed:
1466 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1468 if (p->prio < rq->curr->prio) {
1477 * - the newly woken task is of equal priority to the current task
1478 * - the newly woken task is non-migratable while current is migratable
1479 * - current will be preempted on the next reschedule
1481 * we should check to see if current can readily move to a different
1482 * cpu. If so, we will reschedule to allow the push logic to try
1483 * to move current somewhere else, making room for our non-migratable
1486 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1487 check_preempt_equal_prio(rq, p);
1491 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1492 struct rt_rq *rt_rq)
1494 struct rt_prio_array *array = &rt_rq->active;
1495 struct sched_rt_entity *next = NULL;
1496 struct list_head *queue;
1499 idx = sched_find_first_bit(array->bitmap);
1500 BUG_ON(idx >= MAX_RT_PRIO);
1502 queue = array->queue + idx;
1503 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1508 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1510 struct sched_rt_entity *rt_se;
1511 struct task_struct *p;
1512 struct rt_rq *rt_rq = &rq->rt;
1515 rt_se = pick_next_rt_entity(rq, rt_rq);
1517 rt_rq = group_rt_rq(rt_se);
1520 p = rt_task_of(rt_se);
1521 p->se.exec_start = rq_clock_task(rq);
1526 static struct task_struct *
1527 pick_next_task_rt(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
1529 struct task_struct *p;
1530 struct rt_rq *rt_rq = &rq->rt;
1532 if (need_pull_rt_task(rq, prev)) {
1534 * This is OK, because current is on_cpu, which avoids it being
1535 * picked for load-balance and preemption/IRQs are still
1536 * disabled avoiding further scheduler activity on it and we're
1537 * being very careful to re-start the picking loop.
1539 rq_unpin_lock(rq, rf);
1541 rq_repin_lock(rq, rf);
1543 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1544 * means a dl or stop task can slip in, in which case we need
1545 * to re-start task selection.
1547 if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) ||
1548 rq->dl.dl_nr_running))
1553 * We may dequeue prev's rt_rq in put_prev_task().
1554 * So, we update time before rt_nr_running check.
1556 if (prev->sched_class == &rt_sched_class)
1559 if (!rt_rq->rt_queued)
1562 put_prev_task(rq, prev);
1564 p = _pick_next_task_rt(rq);
1566 /* The running task is never eligible for pushing */
1567 dequeue_pushable_task(rq, p);
1569 queue_push_tasks(rq);
1574 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1579 * The previous task needs to be made eligible for pushing
1580 * if it is still active
1582 if (on_rt_rq(&p->rt) && tsk_nr_cpus_allowed(p) > 1)
1583 enqueue_pushable_task(rq, p);
1588 /* Only try algorithms three times */
1589 #define RT_MAX_TRIES 3
1591 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1593 if (!task_running(rq, p) &&
1594 cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
1600 * Return the highest pushable rq's task, which is suitable to be executed
1601 * on the cpu, NULL otherwise
1603 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1605 struct plist_head *head = &rq->rt.pushable_tasks;
1606 struct task_struct *p;
1608 if (!has_pushable_tasks(rq))
1611 plist_for_each_entry(p, head, pushable_tasks) {
1612 if (pick_rt_task(rq, p, cpu))
1619 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1621 static int find_lowest_rq(struct task_struct *task)
1623 struct sched_domain *sd;
1624 struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1625 int this_cpu = smp_processor_id();
1626 int cpu = task_cpu(task);
1628 /* Make sure the mask is initialized first */
1629 if (unlikely(!lowest_mask))
1632 if (tsk_nr_cpus_allowed(task) == 1)
1633 return -1; /* No other targets possible */
1635 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1636 return -1; /* No targets found */
1639 * At this point we have built a mask of cpus representing the
1640 * lowest priority tasks in the system. Now we want to elect
1641 * the best one based on our affinity and topology.
1643 * We prioritize the last cpu that the task executed on since
1644 * it is most likely cache-hot in that location.
1646 if (cpumask_test_cpu(cpu, lowest_mask))
1650 * Otherwise, we consult the sched_domains span maps to figure
1651 * out which cpu is logically closest to our hot cache data.
1653 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1654 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1657 for_each_domain(cpu, sd) {
1658 if (sd->flags & SD_WAKE_AFFINE) {
1662 * "this_cpu" is cheaper to preempt than a
1665 if (this_cpu != -1 &&
1666 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1671 best_cpu = cpumask_first_and(lowest_mask,
1672 sched_domain_span(sd));
1673 if (best_cpu < nr_cpu_ids) {
1682 * And finally, if there were no matches within the domains
1683 * just give the caller *something* to work with from the compatible
1689 cpu = cpumask_any(lowest_mask);
1690 if (cpu < nr_cpu_ids)
1695 /* Will lock the rq it finds */
1696 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1698 struct rq *lowest_rq = NULL;
1702 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1703 cpu = find_lowest_rq(task);
1705 if ((cpu == -1) || (cpu == rq->cpu))
1708 lowest_rq = cpu_rq(cpu);
1710 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1712 * Target rq has tasks of equal or higher priority,
1713 * retrying does not release any lock and is unlikely
1714 * to yield a different result.
1720 /* if the prio of this runqueue changed, try again */
1721 if (double_lock_balance(rq, lowest_rq)) {
1723 * We had to unlock the run queue. In
1724 * the mean time, task could have
1725 * migrated already or had its affinity changed.
1726 * Also make sure that it wasn't scheduled on its rq.
1728 if (unlikely(task_rq(task) != rq ||
1729 !cpumask_test_cpu(lowest_rq->cpu,
1730 tsk_cpus_allowed(task)) ||
1731 task_running(rq, task) ||
1733 !task_on_rq_queued(task))) {
1735 double_unlock_balance(rq, lowest_rq);
1741 /* If this rq is still suitable use it. */
1742 if (lowest_rq->rt.highest_prio.curr > task->prio)
1746 double_unlock_balance(rq, lowest_rq);
1753 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1755 struct task_struct *p;
1757 if (!has_pushable_tasks(rq))
1760 p = plist_first_entry(&rq->rt.pushable_tasks,
1761 struct task_struct, pushable_tasks);
1763 BUG_ON(rq->cpu != task_cpu(p));
1764 BUG_ON(task_current(rq, p));
1765 BUG_ON(tsk_nr_cpus_allowed(p) <= 1);
1767 BUG_ON(!task_on_rq_queued(p));
1768 BUG_ON(!rt_task(p));
1774 * If the current CPU has more than one RT task, see if the non
1775 * running task can migrate over to a CPU that is running a task
1776 * of lesser priority.
1778 static int push_rt_task(struct rq *rq)
1780 struct task_struct *next_task;
1781 struct rq *lowest_rq;
1784 if (!rq->rt.overloaded)
1787 next_task = pick_next_pushable_task(rq);
1792 if (unlikely(next_task == rq->curr)) {
1798 * It's possible that the next_task slipped in of
1799 * higher priority than current. If that's the case
1800 * just reschedule current.
1802 if (unlikely(next_task->prio < rq->curr->prio)) {
1807 /* We might release rq lock */
1808 get_task_struct(next_task);
1810 /* find_lock_lowest_rq locks the rq if found */
1811 lowest_rq = find_lock_lowest_rq(next_task, rq);
1813 struct task_struct *task;
1815 * find_lock_lowest_rq releases rq->lock
1816 * so it is possible that next_task has migrated.
1818 * We need to make sure that the task is still on the same
1819 * run-queue and is also still the next task eligible for
1822 task = pick_next_pushable_task(rq);
1823 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1825 * The task hasn't migrated, and is still the next
1826 * eligible task, but we failed to find a run-queue
1827 * to push it to. Do not retry in this case, since
1828 * other cpus will pull from us when ready.
1834 /* No more tasks, just exit */
1838 * Something has shifted, try again.
1840 put_task_struct(next_task);
1845 deactivate_task(rq, next_task, 0);
1846 set_task_cpu(next_task, lowest_rq->cpu);
1847 activate_task(lowest_rq, next_task, 0);
1850 resched_curr(lowest_rq);
1852 double_unlock_balance(rq, lowest_rq);
1855 put_task_struct(next_task);
1860 static void push_rt_tasks(struct rq *rq)
1862 /* push_rt_task will return true if it moved an RT */
1863 while (push_rt_task(rq))
1867 #ifdef HAVE_RT_PUSH_IPI
1869 * The search for the next cpu always starts at rq->cpu and ends
1870 * when we reach rq->cpu again. It will never return rq->cpu.
1871 * This returns the next cpu to check, or nr_cpu_ids if the loop
1874 * rq->rt.push_cpu holds the last cpu returned by this function,
1875 * or if this is the first instance, it must hold rq->cpu.
1877 static int rto_next_cpu(struct rq *rq)
1879 int prev_cpu = rq->rt.push_cpu;
1882 cpu = cpumask_next(prev_cpu, rq->rd->rto_mask);
1885 * If the previous cpu is less than the rq's CPU, then it already
1886 * passed the end of the mask, and has started from the beginning.
1887 * We end if the next CPU is greater or equal to rq's CPU.
1889 if (prev_cpu < rq->cpu) {
1893 } else if (cpu >= nr_cpu_ids) {
1895 * We passed the end of the mask, start at the beginning.
1896 * If the result is greater or equal to the rq's CPU, then
1897 * the loop is finished.
1899 cpu = cpumask_first(rq->rd->rto_mask);
1903 rq->rt.push_cpu = cpu;
1905 /* Return cpu to let the caller know if the loop is finished or not */
1909 static int find_next_push_cpu(struct rq *rq)
1915 cpu = rto_next_cpu(rq);
1916 if (cpu >= nr_cpu_ids)
1918 next_rq = cpu_rq(cpu);
1920 /* Make sure the next rq can push to this rq */
1921 if (next_rq->rt.highest_prio.next < rq->rt.highest_prio.curr)
1928 #define RT_PUSH_IPI_EXECUTING 1
1929 #define RT_PUSH_IPI_RESTART 2
1931 static void tell_cpu_to_push(struct rq *rq)
1935 if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
1936 raw_spin_lock(&rq->rt.push_lock);
1937 /* Make sure it's still executing */
1938 if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
1940 * Tell the IPI to restart the loop as things have
1941 * changed since it started.
1943 rq->rt.push_flags |= RT_PUSH_IPI_RESTART;
1944 raw_spin_unlock(&rq->rt.push_lock);
1947 raw_spin_unlock(&rq->rt.push_lock);
1950 /* When here, there's no IPI going around */
1952 rq->rt.push_cpu = rq->cpu;
1953 cpu = find_next_push_cpu(rq);
1954 if (cpu >= nr_cpu_ids)
1957 rq->rt.push_flags = RT_PUSH_IPI_EXECUTING;
1959 irq_work_queue_on(&rq->rt.push_work, cpu);
1962 /* Called from hardirq context */
1963 static void try_to_push_tasks(void *arg)
1965 struct rt_rq *rt_rq = arg;
1966 struct rq *rq, *src_rq;
1970 this_cpu = rt_rq->push_cpu;
1972 /* Paranoid check */
1973 BUG_ON(this_cpu != smp_processor_id());
1975 rq = cpu_rq(this_cpu);
1976 src_rq = rq_of_rt_rq(rt_rq);
1979 if (has_pushable_tasks(rq)) {
1980 raw_spin_lock(&rq->lock);
1982 raw_spin_unlock(&rq->lock);
1985 /* Pass the IPI to the next rt overloaded queue */
1986 raw_spin_lock(&rt_rq->push_lock);
1988 * If the source queue changed since the IPI went out,
1989 * we need to restart the search from that CPU again.
1991 if (rt_rq->push_flags & RT_PUSH_IPI_RESTART) {
1992 rt_rq->push_flags &= ~RT_PUSH_IPI_RESTART;
1993 rt_rq->push_cpu = src_rq->cpu;
1996 cpu = find_next_push_cpu(src_rq);
1998 if (cpu >= nr_cpu_ids)
1999 rt_rq->push_flags &= ~RT_PUSH_IPI_EXECUTING;
2000 raw_spin_unlock(&rt_rq->push_lock);
2002 if (cpu >= nr_cpu_ids)
2006 * It is possible that a restart caused this CPU to be
2007 * chosen again. Don't bother with an IPI, just see if we
2008 * have more to push.
2010 if (unlikely(cpu == rq->cpu))
2013 /* Try the next RT overloaded CPU */
2014 irq_work_queue_on(&rt_rq->push_work, cpu);
2017 static void push_irq_work_func(struct irq_work *work)
2019 struct rt_rq *rt_rq = container_of(work, struct rt_rq, push_work);
2021 try_to_push_tasks(rt_rq);
2023 #endif /* HAVE_RT_PUSH_IPI */
2025 static void pull_rt_task(struct rq *this_rq)
2027 int this_cpu = this_rq->cpu, cpu;
2028 bool resched = false;
2029 struct task_struct *p;
2032 if (likely(!rt_overloaded(this_rq)))
2036 * Match the barrier from rt_set_overloaded; this guarantees that if we
2037 * see overloaded we must also see the rto_mask bit.
2041 #ifdef HAVE_RT_PUSH_IPI
2042 if (sched_feat(RT_PUSH_IPI)) {
2043 tell_cpu_to_push(this_rq);
2048 for_each_cpu(cpu, this_rq->rd->rto_mask) {
2049 if (this_cpu == cpu)
2052 src_rq = cpu_rq(cpu);
2055 * Don't bother taking the src_rq->lock if the next highest
2056 * task is known to be lower-priority than our current task.
2057 * This may look racy, but if this value is about to go
2058 * logically higher, the src_rq will push this task away.
2059 * And if its going logically lower, we do not care
2061 if (src_rq->rt.highest_prio.next >=
2062 this_rq->rt.highest_prio.curr)
2066 * We can potentially drop this_rq's lock in
2067 * double_lock_balance, and another CPU could
2070 double_lock_balance(this_rq, src_rq);
2073 * We can pull only a task, which is pushable
2074 * on its rq, and no others.
2076 p = pick_highest_pushable_task(src_rq, this_cpu);
2079 * Do we have an RT task that preempts
2080 * the to-be-scheduled task?
2082 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2083 WARN_ON(p == src_rq->curr);
2084 WARN_ON(!task_on_rq_queued(p));
2087 * There's a chance that p is higher in priority
2088 * than what's currently running on its cpu.
2089 * This is just that p is wakeing up and hasn't
2090 * had a chance to schedule. We only pull
2091 * p if it is lower in priority than the
2092 * current task on the run queue
2094 if (p->prio < src_rq->curr->prio)
2099 deactivate_task(src_rq, p, 0);
2100 set_task_cpu(p, this_cpu);
2101 activate_task(this_rq, p, 0);
2103 * We continue with the search, just in
2104 * case there's an even higher prio task
2105 * in another runqueue. (low likelihood
2110 double_unlock_balance(this_rq, src_rq);
2114 resched_curr(this_rq);
2118 * If we are not running and we are not going to reschedule soon, we should
2119 * try to push tasks away now
2121 static void task_woken_rt(struct rq *rq, struct task_struct *p)
2123 if (!task_running(rq, p) &&
2124 !test_tsk_need_resched(rq->curr) &&
2125 tsk_nr_cpus_allowed(p) > 1 &&
2126 (dl_task(rq->curr) || rt_task(rq->curr)) &&
2127 (tsk_nr_cpus_allowed(rq->curr) < 2 ||
2128 rq->curr->prio <= p->prio))
2132 /* Assumes rq->lock is held */
2133 static void rq_online_rt(struct rq *rq)
2135 if (rq->rt.overloaded)
2136 rt_set_overload(rq);
2138 __enable_runtime(rq);
2140 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2143 /* Assumes rq->lock is held */
2144 static void rq_offline_rt(struct rq *rq)
2146 if (rq->rt.overloaded)
2147 rt_clear_overload(rq);
2149 __disable_runtime(rq);
2151 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2155 * When switch from the rt queue, we bring ourselves to a position
2156 * that we might want to pull RT tasks from other runqueues.
2158 static void switched_from_rt(struct rq *rq, struct task_struct *p)
2161 * If there are other RT tasks then we will reschedule
2162 * and the scheduling of the other RT tasks will handle
2163 * the balancing. But if we are the last RT task
2164 * we may need to handle the pulling of RT tasks
2167 if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2170 queue_pull_task(rq);
2173 void __init init_sched_rt_class(void)
2177 for_each_possible_cpu(i) {
2178 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2179 GFP_KERNEL, cpu_to_node(i));
2182 #endif /* CONFIG_SMP */
2185 * When switching a task to RT, we may overload the runqueue
2186 * with RT tasks. In this case we try to push them off to
2189 static void switched_to_rt(struct rq *rq, struct task_struct *p)
2192 * If we are already running, then there's nothing
2193 * that needs to be done. But if we are not running
2194 * we may need to preempt the current running task.
2195 * If that current running task is also an RT task
2196 * then see if we can move to another run queue.
2198 if (task_on_rq_queued(p) && rq->curr != p) {
2200 if (tsk_nr_cpus_allowed(p) > 1 && rq->rt.overloaded)
2201 queue_push_tasks(rq);
2202 #endif /* CONFIG_SMP */
2203 if (p->prio < rq->curr->prio)
2209 * Priority of the task has changed. This may cause
2210 * us to initiate a push or pull.
2213 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2215 if (!task_on_rq_queued(p))
2218 if (rq->curr == p) {
2221 * If our priority decreases while running, we
2222 * may need to pull tasks to this runqueue.
2224 if (oldprio < p->prio)
2225 queue_pull_task(rq);
2228 * If there's a higher priority task waiting to run
2231 if (p->prio > rq->rt.highest_prio.curr)
2234 /* For UP simply resched on drop of prio */
2235 if (oldprio < p->prio)
2237 #endif /* CONFIG_SMP */
2240 * This task is not running, but if it is
2241 * greater than the current running task
2244 if (p->prio < rq->curr->prio)
2249 #ifdef CONFIG_POSIX_TIMERS
2250 static void watchdog(struct rq *rq, struct task_struct *p)
2252 unsigned long soft, hard;
2254 /* max may change after cur was read, this will be fixed next tick */
2255 soft = task_rlimit(p, RLIMIT_RTTIME);
2256 hard = task_rlimit_max(p, RLIMIT_RTTIME);
2258 if (soft != RLIM_INFINITY) {
2261 if (p->rt.watchdog_stamp != jiffies) {
2263 p->rt.watchdog_stamp = jiffies;
2266 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2267 if (p->rt.timeout > next)
2268 p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
2272 static inline void watchdog(struct rq *rq, struct task_struct *p) { }
2275 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2277 struct sched_rt_entity *rt_se = &p->rt;
2284 * RR tasks need a special form of timeslice management.
2285 * FIFO tasks have no timeslices.
2287 if (p->policy != SCHED_RR)
2290 if (--p->rt.time_slice)
2293 p->rt.time_slice = sched_rr_timeslice;
2296 * Requeue to the end of queue if we (and all of our ancestors) are not
2297 * the only element on the queue
2299 for_each_sched_rt_entity(rt_se) {
2300 if (rt_se->run_list.prev != rt_se->run_list.next) {
2301 requeue_task_rt(rq, p, 0);
2308 static void set_curr_task_rt(struct rq *rq)
2310 struct task_struct *p = rq->curr;
2312 p->se.exec_start = rq_clock_task(rq);
2314 /* The running task is never eligible for pushing */
2315 dequeue_pushable_task(rq, p);
2318 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2321 * Time slice is 0 for SCHED_FIFO tasks
2323 if (task->policy == SCHED_RR)
2324 return sched_rr_timeslice;
2329 const struct sched_class rt_sched_class = {
2330 .next = &fair_sched_class,
2331 .enqueue_task = enqueue_task_rt,
2332 .dequeue_task = dequeue_task_rt,
2333 .yield_task = yield_task_rt,
2335 .check_preempt_curr = check_preempt_curr_rt,
2337 .pick_next_task = pick_next_task_rt,
2338 .put_prev_task = put_prev_task_rt,
2341 .select_task_rq = select_task_rq_rt,
2343 .set_cpus_allowed = set_cpus_allowed_common,
2344 .rq_online = rq_online_rt,
2345 .rq_offline = rq_offline_rt,
2346 .task_woken = task_woken_rt,
2347 .switched_from = switched_from_rt,
2350 .set_curr_task = set_curr_task_rt,
2351 .task_tick = task_tick_rt,
2353 .get_rr_interval = get_rr_interval_rt,
2355 .prio_changed = prio_changed_rt,
2356 .switched_to = switched_to_rt,
2358 .update_curr = update_curr_rt,
2361 #ifdef CONFIG_SCHED_DEBUG
2362 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2364 void print_rt_stats(struct seq_file *m, int cpu)
2367 struct rt_rq *rt_rq;
2370 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2371 print_rt_rq(m, cpu, rt_rq);
2374 #endif /* CONFIG_SCHED_DEBUG */