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;
62 * SCHED_DEADLINE updates the bandwidth, as a run away
63 * RT task with a DL task could hog a CPU. But DL does
64 * not reset the period. If a deadline task was running
65 * without an RT task running, it can cause RT tasks to
66 * throttle when they start up. Kick the timer right away
67 * to update the period.
69 hrtimer_forward_now(&rt_b->rt_period_timer, ns_to_ktime(0));
70 hrtimer_start_expires(&rt_b->rt_period_timer, HRTIMER_MODE_ABS_PINNED);
72 raw_spin_unlock(&rt_b->rt_runtime_lock);
75 #if defined(CONFIG_SMP) && defined(HAVE_RT_PUSH_IPI)
76 static void push_irq_work_func(struct irq_work *work);
79 void init_rt_rq(struct rt_rq *rt_rq)
81 struct rt_prio_array *array;
84 array = &rt_rq->active;
85 for (i = 0; i < MAX_RT_PRIO; i++) {
86 INIT_LIST_HEAD(array->queue + i);
87 __clear_bit(i, array->bitmap);
89 /* delimiter for bitsearch: */
90 __set_bit(MAX_RT_PRIO, array->bitmap);
92 #if defined CONFIG_SMP
93 rt_rq->highest_prio.curr = MAX_RT_PRIO;
94 rt_rq->highest_prio.next = MAX_RT_PRIO;
95 rt_rq->rt_nr_migratory = 0;
96 rt_rq->overloaded = 0;
97 plist_head_init(&rt_rq->pushable_tasks);
99 #ifdef HAVE_RT_PUSH_IPI
100 rt_rq->push_flags = 0;
101 rt_rq->push_cpu = nr_cpu_ids;
102 raw_spin_lock_init(&rt_rq->push_lock);
103 init_irq_work(&rt_rq->push_work, push_irq_work_func);
105 #endif /* CONFIG_SMP */
106 /* We start is dequeued state, because no RT tasks are queued */
107 rt_rq->rt_queued = 0;
110 rt_rq->rt_throttled = 0;
111 rt_rq->rt_runtime = 0;
112 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
115 #ifdef CONFIG_RT_GROUP_SCHED
116 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
118 hrtimer_cancel(&rt_b->rt_period_timer);
121 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
123 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
125 #ifdef CONFIG_SCHED_DEBUG
126 WARN_ON_ONCE(!rt_entity_is_task(rt_se));
128 return container_of(rt_se, struct task_struct, rt);
131 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
136 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
141 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
143 struct rt_rq *rt_rq = rt_se->rt_rq;
148 void free_rt_sched_group(struct task_group *tg)
153 destroy_rt_bandwidth(&tg->rt_bandwidth);
155 for_each_possible_cpu(i) {
166 void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
167 struct sched_rt_entity *rt_se, int cpu,
168 struct sched_rt_entity *parent)
170 struct rq *rq = cpu_rq(cpu);
172 rt_rq->highest_prio.curr = MAX_RT_PRIO;
173 rt_rq->rt_nr_boosted = 0;
177 tg->rt_rq[cpu] = rt_rq;
178 tg->rt_se[cpu] = rt_se;
184 rt_se->rt_rq = &rq->rt;
186 rt_se->rt_rq = parent->my_q;
189 rt_se->parent = parent;
190 INIT_LIST_HEAD(&rt_se->run_list);
193 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
196 struct sched_rt_entity *rt_se;
199 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
202 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
206 init_rt_bandwidth(&tg->rt_bandwidth,
207 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
209 for_each_possible_cpu(i) {
210 rt_rq = kzalloc_node(sizeof(struct rt_rq),
211 GFP_KERNEL, cpu_to_node(i));
215 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
216 GFP_KERNEL, cpu_to_node(i));
221 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
222 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
233 #else /* CONFIG_RT_GROUP_SCHED */
235 #define rt_entity_is_task(rt_se) (1)
237 static inline struct task_struct *rt_task_of(struct sched_rt_entity *rt_se)
239 return container_of(rt_se, struct task_struct, rt);
242 static inline struct rq *rq_of_rt_rq(struct rt_rq *rt_rq)
244 return container_of(rt_rq, struct rq, rt);
247 static inline struct rq *rq_of_rt_se(struct sched_rt_entity *rt_se)
249 struct task_struct *p = rt_task_of(rt_se);
254 static inline struct rt_rq *rt_rq_of_se(struct sched_rt_entity *rt_se)
256 struct rq *rq = rq_of_rt_se(rt_se);
261 void free_rt_sched_group(struct task_group *tg) { }
263 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
267 #endif /* CONFIG_RT_GROUP_SCHED */
271 static void pull_rt_task(struct rq *this_rq);
273 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
275 /* Try to pull RT tasks here if we lower this rq's prio */
276 return rq->rt.highest_prio.curr > prev->prio;
279 static inline int rt_overloaded(struct rq *rq)
281 return atomic_read(&rq->rd->rto_count);
284 static inline void rt_set_overload(struct rq *rq)
289 cpumask_set_cpu(rq->cpu, rq->rd->rto_mask);
291 * Make sure the mask is visible before we set
292 * the overload count. That is checked to determine
293 * if we should look at the mask. It would be a shame
294 * if we looked at the mask, but the mask was not
297 * Matched by the barrier in pull_rt_task().
300 atomic_inc(&rq->rd->rto_count);
303 static inline void rt_clear_overload(struct rq *rq)
308 /* the order here really doesn't matter */
309 atomic_dec(&rq->rd->rto_count);
310 cpumask_clear_cpu(rq->cpu, rq->rd->rto_mask);
313 static void update_rt_migration(struct rt_rq *rt_rq)
315 if (rt_rq->rt_nr_migratory && rt_rq->rt_nr_total > 1) {
316 if (!rt_rq->overloaded) {
317 rt_set_overload(rq_of_rt_rq(rt_rq));
318 rt_rq->overloaded = 1;
320 } else if (rt_rq->overloaded) {
321 rt_clear_overload(rq_of_rt_rq(rt_rq));
322 rt_rq->overloaded = 0;
326 static void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
328 struct task_struct *p;
330 if (!rt_entity_is_task(rt_se))
333 p = rt_task_of(rt_se);
334 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
336 rt_rq->rt_nr_total++;
337 if (p->nr_cpus_allowed > 1)
338 rt_rq->rt_nr_migratory++;
340 update_rt_migration(rt_rq);
343 static void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
345 struct task_struct *p;
347 if (!rt_entity_is_task(rt_se))
350 p = rt_task_of(rt_se);
351 rt_rq = &rq_of_rt_rq(rt_rq)->rt;
353 rt_rq->rt_nr_total--;
354 if (p->nr_cpus_allowed > 1)
355 rt_rq->rt_nr_migratory--;
357 update_rt_migration(rt_rq);
360 static inline int has_pushable_tasks(struct rq *rq)
362 return !plist_head_empty(&rq->rt.pushable_tasks);
365 static DEFINE_PER_CPU(struct callback_head, rt_push_head);
366 static DEFINE_PER_CPU(struct callback_head, rt_pull_head);
368 static void push_rt_tasks(struct rq *);
369 static void pull_rt_task(struct rq *);
371 static inline void queue_push_tasks(struct rq *rq)
373 if (!has_pushable_tasks(rq))
376 queue_balance_callback(rq, &per_cpu(rt_push_head, rq->cpu), push_rt_tasks);
379 static inline void queue_pull_task(struct rq *rq)
381 queue_balance_callback(rq, &per_cpu(rt_pull_head, rq->cpu), pull_rt_task);
384 static void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
386 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
387 plist_node_init(&p->pushable_tasks, p->prio);
388 plist_add(&p->pushable_tasks, &rq->rt.pushable_tasks);
390 /* Update the highest prio pushable task */
391 if (p->prio < rq->rt.highest_prio.next)
392 rq->rt.highest_prio.next = p->prio;
395 static void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
397 plist_del(&p->pushable_tasks, &rq->rt.pushable_tasks);
399 /* Update the new highest prio pushable task */
400 if (has_pushable_tasks(rq)) {
401 p = plist_first_entry(&rq->rt.pushable_tasks,
402 struct task_struct, pushable_tasks);
403 rq->rt.highest_prio.next = p->prio;
405 rq->rt.highest_prio.next = MAX_RT_PRIO;
410 static inline void enqueue_pushable_task(struct rq *rq, struct task_struct *p)
414 static inline void dequeue_pushable_task(struct rq *rq, struct task_struct *p)
419 void inc_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
424 void dec_rt_migration(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
428 static inline bool need_pull_rt_task(struct rq *rq, struct task_struct *prev)
433 static inline void pull_rt_task(struct rq *this_rq)
437 static inline void queue_push_tasks(struct rq *rq)
440 #endif /* CONFIG_SMP */
442 static void enqueue_top_rt_rq(struct rt_rq *rt_rq);
443 static void dequeue_top_rt_rq(struct rt_rq *rt_rq);
445 static inline int on_rt_rq(struct sched_rt_entity *rt_se)
450 #ifdef CONFIG_RT_GROUP_SCHED
452 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
457 return rt_rq->rt_runtime;
460 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
462 return ktime_to_ns(rt_rq->tg->rt_bandwidth.rt_period);
465 typedef struct task_group *rt_rq_iter_t;
467 static inline struct task_group *next_task_group(struct task_group *tg)
470 tg = list_entry_rcu(tg->list.next,
471 typeof(struct task_group), list);
472 } while (&tg->list != &task_groups && task_group_is_autogroup(tg));
474 if (&tg->list == &task_groups)
480 #define for_each_rt_rq(rt_rq, iter, rq) \
481 for (iter = container_of(&task_groups, typeof(*iter), list); \
482 (iter = next_task_group(iter)) && \
483 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
485 #define for_each_sched_rt_entity(rt_se) \
486 for (; rt_se; rt_se = rt_se->parent)
488 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
493 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
494 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags);
496 static void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
498 struct task_struct *curr = rq_of_rt_rq(rt_rq)->curr;
499 struct rq *rq = rq_of_rt_rq(rt_rq);
500 struct sched_rt_entity *rt_se;
502 int cpu = cpu_of(rq);
504 rt_se = rt_rq->tg->rt_se[cpu];
506 if (rt_rq->rt_nr_running) {
508 enqueue_top_rt_rq(rt_rq);
509 else if (!on_rt_rq(rt_se))
510 enqueue_rt_entity(rt_se, 0);
512 if (rt_rq->highest_prio.curr < curr->prio)
517 static void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
519 struct sched_rt_entity *rt_se;
520 int cpu = cpu_of(rq_of_rt_rq(rt_rq));
522 rt_se = rt_rq->tg->rt_se[cpu];
525 dequeue_top_rt_rq(rt_rq);
526 else if (on_rt_rq(rt_se))
527 dequeue_rt_entity(rt_se, 0);
530 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
532 return rt_rq->rt_throttled && !rt_rq->rt_nr_boosted;
535 static int rt_se_boosted(struct sched_rt_entity *rt_se)
537 struct rt_rq *rt_rq = group_rt_rq(rt_se);
538 struct task_struct *p;
541 return !!rt_rq->rt_nr_boosted;
543 p = rt_task_of(rt_se);
544 return p->prio != p->normal_prio;
548 static inline const struct cpumask *sched_rt_period_mask(void)
550 return this_rq()->rd->span;
553 static inline const struct cpumask *sched_rt_period_mask(void)
555 return cpu_online_mask;
560 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
562 return container_of(rt_b, struct task_group, rt_bandwidth)->rt_rq[cpu];
565 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
567 return &rt_rq->tg->rt_bandwidth;
570 #else /* !CONFIG_RT_GROUP_SCHED */
572 static inline u64 sched_rt_runtime(struct rt_rq *rt_rq)
574 return rt_rq->rt_runtime;
577 static inline u64 sched_rt_period(struct rt_rq *rt_rq)
579 return ktime_to_ns(def_rt_bandwidth.rt_period);
582 typedef struct rt_rq *rt_rq_iter_t;
584 #define for_each_rt_rq(rt_rq, iter, rq) \
585 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
587 #define for_each_sched_rt_entity(rt_se) \
588 for (; rt_se; rt_se = NULL)
590 static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
595 static inline void sched_rt_rq_enqueue(struct rt_rq *rt_rq)
597 struct rq *rq = rq_of_rt_rq(rt_rq);
599 if (!rt_rq->rt_nr_running)
602 enqueue_top_rt_rq(rt_rq);
606 static inline void sched_rt_rq_dequeue(struct rt_rq *rt_rq)
608 dequeue_top_rt_rq(rt_rq);
611 static inline int rt_rq_throttled(struct rt_rq *rt_rq)
613 return rt_rq->rt_throttled;
616 static inline const struct cpumask *sched_rt_period_mask(void)
618 return cpu_online_mask;
622 struct rt_rq *sched_rt_period_rt_rq(struct rt_bandwidth *rt_b, int cpu)
624 return &cpu_rq(cpu)->rt;
627 static inline struct rt_bandwidth *sched_rt_bandwidth(struct rt_rq *rt_rq)
629 return &def_rt_bandwidth;
632 #endif /* CONFIG_RT_GROUP_SCHED */
634 bool sched_rt_bandwidth_account(struct rt_rq *rt_rq)
636 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
638 return (hrtimer_active(&rt_b->rt_period_timer) ||
639 rt_rq->rt_time < rt_b->rt_runtime);
644 * We ran out of runtime, see if we can borrow some from our neighbours.
646 static void do_balance_runtime(struct rt_rq *rt_rq)
648 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
649 struct root_domain *rd = rq_of_rt_rq(rt_rq)->rd;
653 weight = cpumask_weight(rd->span);
655 raw_spin_lock(&rt_b->rt_runtime_lock);
656 rt_period = ktime_to_ns(rt_b->rt_period);
657 for_each_cpu(i, rd->span) {
658 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
664 raw_spin_lock(&iter->rt_runtime_lock);
666 * Either all rqs have inf runtime and there's nothing to steal
667 * or __disable_runtime() below sets a specific rq to inf to
668 * indicate its been disabled and disalow stealing.
670 if (iter->rt_runtime == RUNTIME_INF)
674 * From runqueues with spare time, take 1/n part of their
675 * spare time, but no more than our period.
677 diff = iter->rt_runtime - iter->rt_time;
679 diff = div_u64((u64)diff, weight);
680 if (rt_rq->rt_runtime + diff > rt_period)
681 diff = rt_period - rt_rq->rt_runtime;
682 iter->rt_runtime -= diff;
683 rt_rq->rt_runtime += diff;
684 if (rt_rq->rt_runtime == rt_period) {
685 raw_spin_unlock(&iter->rt_runtime_lock);
690 raw_spin_unlock(&iter->rt_runtime_lock);
692 raw_spin_unlock(&rt_b->rt_runtime_lock);
696 * Ensure this RQ takes back all the runtime it lend to its neighbours.
698 static void __disable_runtime(struct rq *rq)
700 struct root_domain *rd = rq->rd;
704 if (unlikely(!scheduler_running))
707 for_each_rt_rq(rt_rq, iter, rq) {
708 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
712 raw_spin_lock(&rt_b->rt_runtime_lock);
713 raw_spin_lock(&rt_rq->rt_runtime_lock);
715 * Either we're all inf and nobody needs to borrow, or we're
716 * already disabled and thus have nothing to do, or we have
717 * exactly the right amount of runtime to take out.
719 if (rt_rq->rt_runtime == RUNTIME_INF ||
720 rt_rq->rt_runtime == rt_b->rt_runtime)
722 raw_spin_unlock(&rt_rq->rt_runtime_lock);
725 * Calculate the difference between what we started out with
726 * and what we current have, that's the amount of runtime
727 * we lend and now have to reclaim.
729 want = rt_b->rt_runtime - rt_rq->rt_runtime;
732 * Greedy reclaim, take back as much as we can.
734 for_each_cpu(i, rd->span) {
735 struct rt_rq *iter = sched_rt_period_rt_rq(rt_b, i);
739 * Can't reclaim from ourselves or disabled runqueues.
741 if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
744 raw_spin_lock(&iter->rt_runtime_lock);
746 diff = min_t(s64, iter->rt_runtime, want);
747 iter->rt_runtime -= diff;
750 iter->rt_runtime -= want;
753 raw_spin_unlock(&iter->rt_runtime_lock);
759 raw_spin_lock(&rt_rq->rt_runtime_lock);
761 * We cannot be left wanting - that would mean some runtime
762 * leaked out of the system.
767 * Disable all the borrow logic by pretending we have inf
768 * runtime - in which case borrowing doesn't make sense.
770 rt_rq->rt_runtime = RUNTIME_INF;
771 rt_rq->rt_throttled = 0;
772 raw_spin_unlock(&rt_rq->rt_runtime_lock);
773 raw_spin_unlock(&rt_b->rt_runtime_lock);
775 /* Make rt_rq available for pick_next_task() */
776 sched_rt_rq_enqueue(rt_rq);
780 static void __enable_runtime(struct rq *rq)
785 if (unlikely(!scheduler_running))
789 * Reset each runqueue's bandwidth settings
791 for_each_rt_rq(rt_rq, iter, rq) {
792 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
794 raw_spin_lock(&rt_b->rt_runtime_lock);
795 raw_spin_lock(&rt_rq->rt_runtime_lock);
796 rt_rq->rt_runtime = rt_b->rt_runtime;
798 rt_rq->rt_throttled = 0;
799 raw_spin_unlock(&rt_rq->rt_runtime_lock);
800 raw_spin_unlock(&rt_b->rt_runtime_lock);
804 static void balance_runtime(struct rt_rq *rt_rq)
806 if (!sched_feat(RT_RUNTIME_SHARE))
809 if (rt_rq->rt_time > rt_rq->rt_runtime) {
810 raw_spin_unlock(&rt_rq->rt_runtime_lock);
811 do_balance_runtime(rt_rq);
812 raw_spin_lock(&rt_rq->rt_runtime_lock);
815 #else /* !CONFIG_SMP */
816 static inline void balance_runtime(struct rt_rq *rt_rq) {}
817 #endif /* CONFIG_SMP */
819 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
821 int i, idle = 1, throttled = 0;
822 const struct cpumask *span;
824 span = sched_rt_period_mask();
825 #ifdef CONFIG_RT_GROUP_SCHED
827 * FIXME: isolated CPUs should really leave the root task group,
828 * whether they are isolcpus or were isolated via cpusets, lest
829 * the timer run on a CPU which does not service all runqueues,
830 * potentially leaving other CPUs indefinitely throttled. If
831 * isolation is really required, the user will turn the throttle
832 * off to kill the perturbations it causes anyway. Meanwhile,
833 * this maintains functionality for boot and/or troubleshooting.
835 if (rt_b == &root_task_group.rt_bandwidth)
836 span = cpu_online_mask;
838 for_each_cpu(i, span) {
840 struct rt_rq *rt_rq = sched_rt_period_rt_rq(rt_b, i);
841 struct rq *rq = rq_of_rt_rq(rt_rq);
843 raw_spin_lock(&rq->lock);
844 if (rt_rq->rt_time) {
847 raw_spin_lock(&rt_rq->rt_runtime_lock);
848 if (rt_rq->rt_throttled)
849 balance_runtime(rt_rq);
850 runtime = rt_rq->rt_runtime;
851 rt_rq->rt_time -= min(rt_rq->rt_time, overrun*runtime);
852 if (rt_rq->rt_throttled && rt_rq->rt_time < runtime) {
853 rt_rq->rt_throttled = 0;
857 * When we're idle and a woken (rt) task is
858 * throttled check_preempt_curr() will set
859 * skip_update and the time between the wakeup
860 * and this unthrottle will get accounted as
863 if (rt_rq->rt_nr_running && rq->curr == rq->idle)
864 rq_clock_skip_update(rq, false);
866 if (rt_rq->rt_time || rt_rq->rt_nr_running)
868 raw_spin_unlock(&rt_rq->rt_runtime_lock);
869 } else if (rt_rq->rt_nr_running) {
871 if (!rt_rq_throttled(rt_rq))
874 if (rt_rq->rt_throttled)
878 sched_rt_rq_enqueue(rt_rq);
879 raw_spin_unlock(&rq->lock);
882 if (!throttled && (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF))
888 static inline int rt_se_prio(struct sched_rt_entity *rt_se)
890 #ifdef CONFIG_RT_GROUP_SCHED
891 struct rt_rq *rt_rq = group_rt_rq(rt_se);
894 return rt_rq->highest_prio.curr;
897 return rt_task_of(rt_se)->prio;
900 static int sched_rt_runtime_exceeded(struct rt_rq *rt_rq)
902 u64 runtime = sched_rt_runtime(rt_rq);
904 if (rt_rq->rt_throttled)
905 return rt_rq_throttled(rt_rq);
907 if (runtime >= sched_rt_period(rt_rq))
910 balance_runtime(rt_rq);
911 runtime = sched_rt_runtime(rt_rq);
912 if (runtime == RUNTIME_INF)
915 if (rt_rq->rt_time > runtime) {
916 struct rt_bandwidth *rt_b = sched_rt_bandwidth(rt_rq);
919 * Don't actually throttle groups that have no runtime assigned
920 * but accrue some time due to boosting.
922 if (likely(rt_b->rt_runtime)) {
923 rt_rq->rt_throttled = 1;
924 printk_deferred_once("sched: RT throttling activated\n");
927 * In case we did anyway, make it go away,
928 * replenishment is a joke, since it will replenish us
934 if (rt_rq_throttled(rt_rq)) {
935 sched_rt_rq_dequeue(rt_rq);
944 * Update the current task's runtime statistics. Skip current tasks that
945 * are not in our scheduling class.
947 static void update_curr_rt(struct rq *rq)
949 struct task_struct *curr = rq->curr;
950 struct sched_rt_entity *rt_se = &curr->rt;
953 if (curr->sched_class != &rt_sched_class)
956 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
957 if (unlikely((s64)delta_exec <= 0))
960 schedstat_set(curr->se.statistics.exec_max,
961 max(curr->se.statistics.exec_max, delta_exec));
963 curr->se.sum_exec_runtime += delta_exec;
964 account_group_exec_runtime(curr, delta_exec);
966 curr->se.exec_start = rq_clock_task(rq);
967 cpuacct_charge(curr, delta_exec);
969 sched_rt_avg_update(rq, delta_exec);
971 if (!rt_bandwidth_enabled())
974 for_each_sched_rt_entity(rt_se) {
975 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
977 if (sched_rt_runtime(rt_rq) != RUNTIME_INF) {
978 raw_spin_lock(&rt_rq->rt_runtime_lock);
979 rt_rq->rt_time += delta_exec;
980 if (sched_rt_runtime_exceeded(rt_rq))
982 raw_spin_unlock(&rt_rq->rt_runtime_lock);
988 dequeue_top_rt_rq(struct rt_rq *rt_rq)
990 struct rq *rq = rq_of_rt_rq(rt_rq);
992 BUG_ON(&rq->rt != rt_rq);
994 if (!rt_rq->rt_queued)
997 BUG_ON(!rq->nr_running);
999 sub_nr_running(rq, rt_rq->rt_nr_running);
1000 rt_rq->rt_queued = 0;
1004 enqueue_top_rt_rq(struct rt_rq *rt_rq)
1006 struct rq *rq = rq_of_rt_rq(rt_rq);
1008 BUG_ON(&rq->rt != rt_rq);
1010 if (rt_rq->rt_queued)
1012 if (rt_rq_throttled(rt_rq) || !rt_rq->rt_nr_running)
1015 add_nr_running(rq, rt_rq->rt_nr_running);
1016 rt_rq->rt_queued = 1;
1019 #if defined CONFIG_SMP
1022 inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1024 struct rq *rq = rq_of_rt_rq(rt_rq);
1026 #ifdef CONFIG_RT_GROUP_SCHED
1028 * Change rq's cpupri only if rt_rq is the top queue.
1030 if (&rq->rt != rt_rq)
1033 if (rq->online && prio < prev_prio)
1034 cpupri_set(&rq->rd->cpupri, rq->cpu, prio);
1038 dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio)
1040 struct rq *rq = rq_of_rt_rq(rt_rq);
1042 #ifdef CONFIG_RT_GROUP_SCHED
1044 * Change rq's cpupri only if rt_rq is the top queue.
1046 if (&rq->rt != rt_rq)
1049 if (rq->online && rt_rq->highest_prio.curr != prev_prio)
1050 cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio.curr);
1053 #else /* CONFIG_SMP */
1056 void inc_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1058 void dec_rt_prio_smp(struct rt_rq *rt_rq, int prio, int prev_prio) {}
1060 #endif /* CONFIG_SMP */
1062 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1064 inc_rt_prio(struct rt_rq *rt_rq, int prio)
1066 int prev_prio = rt_rq->highest_prio.curr;
1068 if (prio < prev_prio)
1069 rt_rq->highest_prio.curr = prio;
1071 inc_rt_prio_smp(rt_rq, prio, prev_prio);
1075 dec_rt_prio(struct rt_rq *rt_rq, int prio)
1077 int prev_prio = rt_rq->highest_prio.curr;
1079 if (rt_rq->rt_nr_running) {
1081 WARN_ON(prio < prev_prio);
1084 * This may have been our highest task, and therefore
1085 * we may have some recomputation to do
1087 if (prio == prev_prio) {
1088 struct rt_prio_array *array = &rt_rq->active;
1090 rt_rq->highest_prio.curr =
1091 sched_find_first_bit(array->bitmap);
1095 rt_rq->highest_prio.curr = MAX_RT_PRIO;
1097 dec_rt_prio_smp(rt_rq, prio, prev_prio);
1102 static inline void inc_rt_prio(struct rt_rq *rt_rq, int prio) {}
1103 static inline void dec_rt_prio(struct rt_rq *rt_rq, int prio) {}
1105 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1107 #ifdef CONFIG_RT_GROUP_SCHED
1110 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1112 if (rt_se_boosted(rt_se))
1113 rt_rq->rt_nr_boosted++;
1116 start_rt_bandwidth(&rt_rq->tg->rt_bandwidth);
1120 dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1122 if (rt_se_boosted(rt_se))
1123 rt_rq->rt_nr_boosted--;
1125 WARN_ON(!rt_rq->rt_nr_running && rt_rq->rt_nr_boosted);
1128 #else /* CONFIG_RT_GROUP_SCHED */
1131 inc_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1133 start_rt_bandwidth(&def_rt_bandwidth);
1137 void dec_rt_group(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq) {}
1139 #endif /* CONFIG_RT_GROUP_SCHED */
1142 unsigned int rt_se_nr_running(struct sched_rt_entity *rt_se)
1144 struct rt_rq *group_rq = group_rt_rq(rt_se);
1147 return group_rq->rt_nr_running;
1153 unsigned int rt_se_rr_nr_running(struct sched_rt_entity *rt_se)
1155 struct rt_rq *group_rq = group_rt_rq(rt_se);
1156 struct task_struct *tsk;
1159 return group_rq->rr_nr_running;
1161 tsk = rt_task_of(rt_se);
1163 return (tsk->policy == SCHED_RR) ? 1 : 0;
1167 void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1169 int prio = rt_se_prio(rt_se);
1171 WARN_ON(!rt_prio(prio));
1172 rt_rq->rt_nr_running += rt_se_nr_running(rt_se);
1173 rt_rq->rr_nr_running += rt_se_rr_nr_running(rt_se);
1175 inc_rt_prio(rt_rq, prio);
1176 inc_rt_migration(rt_se, rt_rq);
1177 inc_rt_group(rt_se, rt_rq);
1181 void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
1183 WARN_ON(!rt_prio(rt_se_prio(rt_se)));
1184 WARN_ON(!rt_rq->rt_nr_running);
1185 rt_rq->rt_nr_running -= rt_se_nr_running(rt_se);
1186 rt_rq->rr_nr_running -= rt_se_rr_nr_running(rt_se);
1188 dec_rt_prio(rt_rq, rt_se_prio(rt_se));
1189 dec_rt_migration(rt_se, rt_rq);
1190 dec_rt_group(rt_se, rt_rq);
1194 * Change rt_se->run_list location unless SAVE && !MOVE
1196 * assumes ENQUEUE/DEQUEUE flags match
1198 static inline bool move_entity(unsigned int flags)
1200 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
1206 static void __delist_rt_entity(struct sched_rt_entity *rt_se, struct rt_prio_array *array)
1208 list_del_init(&rt_se->run_list);
1210 if (list_empty(array->queue + rt_se_prio(rt_se)))
1211 __clear_bit(rt_se_prio(rt_se), array->bitmap);
1216 static void __enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1218 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1219 struct rt_prio_array *array = &rt_rq->active;
1220 struct rt_rq *group_rq = group_rt_rq(rt_se);
1221 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1224 * Don't enqueue the group if its throttled, or when empty.
1225 * The latter is a consequence of the former when a child group
1226 * get throttled and the current group doesn't have any other
1229 if (group_rq && (rt_rq_throttled(group_rq) || !group_rq->rt_nr_running)) {
1231 __delist_rt_entity(rt_se, array);
1235 if (move_entity(flags)) {
1236 WARN_ON_ONCE(rt_se->on_list);
1237 if (flags & ENQUEUE_HEAD)
1238 list_add(&rt_se->run_list, queue);
1240 list_add_tail(&rt_se->run_list, queue);
1242 __set_bit(rt_se_prio(rt_se), array->bitmap);
1247 inc_rt_tasks(rt_se, rt_rq);
1250 static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1252 struct rt_rq *rt_rq = rt_rq_of_se(rt_se);
1253 struct rt_prio_array *array = &rt_rq->active;
1255 if (move_entity(flags)) {
1256 WARN_ON_ONCE(!rt_se->on_list);
1257 __delist_rt_entity(rt_se, array);
1261 dec_rt_tasks(rt_se, rt_rq);
1265 * Because the prio of an upper entry depends on the lower
1266 * entries, we must remove entries top - down.
1268 static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
1270 struct sched_rt_entity *back = NULL;
1272 for_each_sched_rt_entity(rt_se) {
1277 dequeue_top_rt_rq(rt_rq_of_se(back));
1279 for (rt_se = back; rt_se; rt_se = rt_se->back) {
1280 if (on_rt_rq(rt_se))
1281 __dequeue_rt_entity(rt_se, flags);
1285 static void enqueue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1287 struct rq *rq = rq_of_rt_se(rt_se);
1289 dequeue_rt_stack(rt_se, flags);
1290 for_each_sched_rt_entity(rt_se)
1291 __enqueue_rt_entity(rt_se, flags);
1292 enqueue_top_rt_rq(&rq->rt);
1295 static void dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flags)
1297 struct rq *rq = rq_of_rt_se(rt_se);
1299 dequeue_rt_stack(rt_se, flags);
1301 for_each_sched_rt_entity(rt_se) {
1302 struct rt_rq *rt_rq = group_rt_rq(rt_se);
1304 if (rt_rq && rt_rq->rt_nr_running)
1305 __enqueue_rt_entity(rt_se, flags);
1307 enqueue_top_rt_rq(&rq->rt);
1311 * Adding/removing a task to/from a priority array:
1314 enqueue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1316 struct sched_rt_entity *rt_se = &p->rt;
1318 if (flags & ENQUEUE_WAKEUP)
1321 enqueue_rt_entity(rt_se, flags);
1323 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1324 enqueue_pushable_task(rq, p);
1327 static void dequeue_task_rt(struct rq *rq, struct task_struct *p, int flags)
1329 struct sched_rt_entity *rt_se = &p->rt;
1332 dequeue_rt_entity(rt_se, flags);
1334 dequeue_pushable_task(rq, p);
1338 * Put task to the head or the end of the run list without the overhead of
1339 * dequeue followed by enqueue.
1342 requeue_rt_entity(struct rt_rq *rt_rq, struct sched_rt_entity *rt_se, int head)
1344 if (on_rt_rq(rt_se)) {
1345 struct rt_prio_array *array = &rt_rq->active;
1346 struct list_head *queue = array->queue + rt_se_prio(rt_se);
1349 list_move(&rt_se->run_list, queue);
1351 list_move_tail(&rt_se->run_list, queue);
1355 static void requeue_task_rt(struct rq *rq, struct task_struct *p, int head)
1357 struct sched_rt_entity *rt_se = &p->rt;
1358 struct rt_rq *rt_rq;
1360 for_each_sched_rt_entity(rt_se) {
1361 rt_rq = rt_rq_of_se(rt_se);
1362 requeue_rt_entity(rt_rq, rt_se, head);
1366 static void yield_task_rt(struct rq *rq)
1368 requeue_task_rt(rq, rq->curr, 0);
1372 static int find_lowest_rq(struct task_struct *task);
1375 select_task_rq_rt(struct task_struct *p, int cpu, int sd_flag, int flags)
1377 struct task_struct *curr;
1380 /* For anything but wake ups, just return the task_cpu */
1381 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
1387 curr = READ_ONCE(rq->curr); /* unlocked access */
1390 * If the current task on @p's runqueue is an RT task, then
1391 * try to see if we can wake this RT task up on another
1392 * runqueue. Otherwise simply start this RT task
1393 * on its current runqueue.
1395 * We want to avoid overloading runqueues. If the woken
1396 * task is a higher priority, then it will stay on this CPU
1397 * and the lower prio task should be moved to another CPU.
1398 * Even though this will probably make the lower prio task
1399 * lose its cache, we do not want to bounce a higher task
1400 * around just because it gave up its CPU, perhaps for a
1403 * For equal prio tasks, we just let the scheduler sort it out.
1405 * Otherwise, just let it ride on the affined RQ and the
1406 * post-schedule router will push the preempted task away
1408 * This test is optimistic, if we get it wrong the load-balancer
1409 * will have to sort it out.
1411 if (curr && unlikely(rt_task(curr)) &&
1412 (curr->nr_cpus_allowed < 2 ||
1413 curr->prio <= p->prio)) {
1414 int target = find_lowest_rq(p);
1417 * Don't bother moving it if the destination CPU is
1418 * not running a lower priority task.
1421 p->prio < cpu_rq(target)->rt.highest_prio.curr)
1430 static void check_preempt_equal_prio(struct rq *rq, struct task_struct *p)
1433 * Current can't be migrated, useless to reschedule,
1434 * let's hope p can move out.
1436 if (rq->curr->nr_cpus_allowed == 1 ||
1437 !cpupri_find(&rq->rd->cpupri, rq->curr, NULL))
1441 * p is migratable, so let's not schedule it and
1442 * see if it is pushed or pulled somewhere else.
1444 if (p->nr_cpus_allowed != 1
1445 && cpupri_find(&rq->rd->cpupri, p, NULL))
1449 * There appears to be other cpus that can accept
1450 * current and none to run 'p', so lets reschedule
1451 * to try and push current away:
1453 requeue_task_rt(rq, p, 1);
1457 #endif /* CONFIG_SMP */
1460 * Preempt the current task with a newly woken task if needed:
1462 static void check_preempt_curr_rt(struct rq *rq, struct task_struct *p, int flags)
1464 if (p->prio < rq->curr->prio) {
1473 * - the newly woken task is of equal priority to the current task
1474 * - the newly woken task is non-migratable while current is migratable
1475 * - current will be preempted on the next reschedule
1477 * we should check to see if current can readily move to a different
1478 * cpu. If so, we will reschedule to allow the push logic to try
1479 * to move current somewhere else, making room for our non-migratable
1482 if (p->prio == rq->curr->prio && !test_tsk_need_resched(rq->curr))
1483 check_preempt_equal_prio(rq, p);
1487 static struct sched_rt_entity *pick_next_rt_entity(struct rq *rq,
1488 struct rt_rq *rt_rq)
1490 struct rt_prio_array *array = &rt_rq->active;
1491 struct sched_rt_entity *next = NULL;
1492 struct list_head *queue;
1495 idx = sched_find_first_bit(array->bitmap);
1496 BUG_ON(idx >= MAX_RT_PRIO);
1498 queue = array->queue + idx;
1499 next = list_entry(queue->next, struct sched_rt_entity, run_list);
1504 static struct task_struct *_pick_next_task_rt(struct rq *rq)
1506 struct sched_rt_entity *rt_se;
1507 struct task_struct *p;
1508 struct rt_rq *rt_rq = &rq->rt;
1511 rt_se = pick_next_rt_entity(rq, rt_rq);
1513 rt_rq = group_rt_rq(rt_se);
1516 p = rt_task_of(rt_se);
1517 p->se.exec_start = rq_clock_task(rq);
1522 static struct task_struct *
1523 pick_next_task_rt(struct rq *rq, struct task_struct *prev)
1525 struct task_struct *p;
1526 struct rt_rq *rt_rq = &rq->rt;
1528 if (need_pull_rt_task(rq, prev)) {
1530 * This is OK, because current is on_cpu, which avoids it being
1531 * picked for load-balance and preemption/IRQs are still
1532 * disabled avoiding further scheduler activity on it and we're
1533 * being very careful to re-start the picking loop.
1535 lockdep_unpin_lock(&rq->lock);
1537 lockdep_pin_lock(&rq->lock);
1539 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1540 * means a dl or stop task can slip in, in which case we need
1541 * to re-start task selection.
1543 if (unlikely((rq->stop && task_on_rq_queued(rq->stop)) ||
1544 rq->dl.dl_nr_running))
1549 * We may dequeue prev's rt_rq in put_prev_task().
1550 * So, we update time before rt_nr_running check.
1552 if (prev->sched_class == &rt_sched_class)
1555 if (!rt_rq->rt_queued)
1558 put_prev_task(rq, prev);
1560 p = _pick_next_task_rt(rq);
1562 /* The running task is never eligible for pushing */
1563 dequeue_pushable_task(rq, p);
1565 queue_push_tasks(rq);
1570 static void put_prev_task_rt(struct rq *rq, struct task_struct *p)
1575 * The previous task needs to be made eligible for pushing
1576 * if it is still active
1578 if (on_rt_rq(&p->rt) && p->nr_cpus_allowed > 1)
1579 enqueue_pushable_task(rq, p);
1584 /* Only try algorithms three times */
1585 #define RT_MAX_TRIES 3
1587 static int pick_rt_task(struct rq *rq, struct task_struct *p, int cpu)
1589 if (!task_running(rq, p) &&
1590 cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
1596 * Return the highest pushable rq's task, which is suitable to be executed
1597 * on the cpu, NULL otherwise
1599 static struct task_struct *pick_highest_pushable_task(struct rq *rq, int cpu)
1601 struct plist_head *head = &rq->rt.pushable_tasks;
1602 struct task_struct *p;
1604 if (!has_pushable_tasks(rq))
1607 plist_for_each_entry(p, head, pushable_tasks) {
1608 if (pick_rt_task(rq, p, cpu))
1615 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask);
1617 static int find_lowest_rq(struct task_struct *task)
1619 struct sched_domain *sd;
1620 struct cpumask *lowest_mask = this_cpu_cpumask_var_ptr(local_cpu_mask);
1621 int this_cpu = smp_processor_id();
1622 int cpu = task_cpu(task);
1624 /* Make sure the mask is initialized first */
1625 if (unlikely(!lowest_mask))
1628 if (task->nr_cpus_allowed == 1)
1629 return -1; /* No other targets possible */
1631 if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
1632 return -1; /* No targets found */
1635 * At this point we have built a mask of cpus representing the
1636 * lowest priority tasks in the system. Now we want to elect
1637 * the best one based on our affinity and topology.
1639 * We prioritize the last cpu that the task executed on since
1640 * it is most likely cache-hot in that location.
1642 if (cpumask_test_cpu(cpu, lowest_mask))
1646 * Otherwise, we consult the sched_domains span maps to figure
1647 * out which cpu is logically closest to our hot cache data.
1649 if (!cpumask_test_cpu(this_cpu, lowest_mask))
1650 this_cpu = -1; /* Skip this_cpu opt if not among lowest */
1653 for_each_domain(cpu, sd) {
1654 if (sd->flags & SD_WAKE_AFFINE) {
1658 * "this_cpu" is cheaper to preempt than a
1661 if (this_cpu != -1 &&
1662 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1667 best_cpu = cpumask_first_and(lowest_mask,
1668 sched_domain_span(sd));
1669 if (best_cpu < nr_cpu_ids) {
1678 * And finally, if there were no matches within the domains
1679 * just give the caller *something* to work with from the compatible
1685 cpu = cpumask_any(lowest_mask);
1686 if (cpu < nr_cpu_ids)
1691 /* Will lock the rq it finds */
1692 static struct rq *find_lock_lowest_rq(struct task_struct *task, struct rq *rq)
1694 struct rq *lowest_rq = NULL;
1698 for (tries = 0; tries < RT_MAX_TRIES; tries++) {
1699 cpu = find_lowest_rq(task);
1701 if ((cpu == -1) || (cpu == rq->cpu))
1704 lowest_rq = cpu_rq(cpu);
1706 if (lowest_rq->rt.highest_prio.curr <= task->prio) {
1708 * Target rq has tasks of equal or higher priority,
1709 * retrying does not release any lock and is unlikely
1710 * to yield a different result.
1716 /* if the prio of this runqueue changed, try again */
1717 if (double_lock_balance(rq, lowest_rq)) {
1719 * We had to unlock the run queue. In
1720 * the mean time, task could have
1721 * migrated already or had its affinity changed.
1722 * Also make sure that it wasn't scheduled on its rq.
1724 if (unlikely(task_rq(task) != rq ||
1725 !cpumask_test_cpu(lowest_rq->cpu,
1726 tsk_cpus_allowed(task)) ||
1727 task_running(rq, task) ||
1728 !task_on_rq_queued(task))) {
1730 double_unlock_balance(rq, lowest_rq);
1736 /* If this rq is still suitable use it. */
1737 if (lowest_rq->rt.highest_prio.curr > task->prio)
1741 double_unlock_balance(rq, lowest_rq);
1748 static struct task_struct *pick_next_pushable_task(struct rq *rq)
1750 struct task_struct *p;
1752 if (!has_pushable_tasks(rq))
1755 p = plist_first_entry(&rq->rt.pushable_tasks,
1756 struct task_struct, pushable_tasks);
1758 BUG_ON(rq->cpu != task_cpu(p));
1759 BUG_ON(task_current(rq, p));
1760 BUG_ON(p->nr_cpus_allowed <= 1);
1762 BUG_ON(!task_on_rq_queued(p));
1763 BUG_ON(!rt_task(p));
1769 * If the current CPU has more than one RT task, see if the non
1770 * running task can migrate over to a CPU that is running a task
1771 * of lesser priority.
1773 static int push_rt_task(struct rq *rq)
1775 struct task_struct *next_task;
1776 struct rq *lowest_rq;
1779 if (!rq->rt.overloaded)
1782 next_task = pick_next_pushable_task(rq);
1787 if (unlikely(next_task == rq->curr)) {
1793 * It's possible that the next_task slipped in of
1794 * higher priority than current. If that's the case
1795 * just reschedule current.
1797 if (unlikely(next_task->prio < rq->curr->prio)) {
1802 /* We might release rq lock */
1803 get_task_struct(next_task);
1805 /* find_lock_lowest_rq locks the rq if found */
1806 lowest_rq = find_lock_lowest_rq(next_task, rq);
1808 struct task_struct *task;
1810 * find_lock_lowest_rq releases rq->lock
1811 * so it is possible that next_task has migrated.
1813 * We need to make sure that the task is still on the same
1814 * run-queue and is also still the next task eligible for
1817 task = pick_next_pushable_task(rq);
1818 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1820 * The task hasn't migrated, and is still the next
1821 * eligible task, but we failed to find a run-queue
1822 * to push it to. Do not retry in this case, since
1823 * other cpus will pull from us when ready.
1829 /* No more tasks, just exit */
1833 * Something has shifted, try again.
1835 put_task_struct(next_task);
1840 deactivate_task(rq, next_task, 0);
1841 set_task_cpu(next_task, lowest_rq->cpu);
1842 activate_task(lowest_rq, next_task, 0);
1845 resched_curr(lowest_rq);
1847 double_unlock_balance(rq, lowest_rq);
1850 put_task_struct(next_task);
1855 static void push_rt_tasks(struct rq *rq)
1857 /* push_rt_task will return true if it moved an RT */
1858 while (push_rt_task(rq))
1862 #ifdef HAVE_RT_PUSH_IPI
1864 * The search for the next cpu always starts at rq->cpu and ends
1865 * when we reach rq->cpu again. It will never return rq->cpu.
1866 * This returns the next cpu to check, or nr_cpu_ids if the loop
1869 * rq->rt.push_cpu holds the last cpu returned by this function,
1870 * or if this is the first instance, it must hold rq->cpu.
1872 static int rto_next_cpu(struct rq *rq)
1874 int prev_cpu = rq->rt.push_cpu;
1877 cpu = cpumask_next(prev_cpu, rq->rd->rto_mask);
1880 * If the previous cpu is less than the rq's CPU, then it already
1881 * passed the end of the mask, and has started from the beginning.
1882 * We end if the next CPU is greater or equal to rq's CPU.
1884 if (prev_cpu < rq->cpu) {
1888 } else if (cpu >= nr_cpu_ids) {
1890 * We passed the end of the mask, start at the beginning.
1891 * If the result is greater or equal to the rq's CPU, then
1892 * the loop is finished.
1894 cpu = cpumask_first(rq->rd->rto_mask);
1898 rq->rt.push_cpu = cpu;
1900 /* Return cpu to let the caller know if the loop is finished or not */
1904 static int find_next_push_cpu(struct rq *rq)
1910 cpu = rto_next_cpu(rq);
1911 if (cpu >= nr_cpu_ids)
1913 next_rq = cpu_rq(cpu);
1915 /* Make sure the next rq can push to this rq */
1916 if (next_rq->rt.highest_prio.next < rq->rt.highest_prio.curr)
1923 #define RT_PUSH_IPI_EXECUTING 1
1924 #define RT_PUSH_IPI_RESTART 2
1926 static void tell_cpu_to_push(struct rq *rq)
1930 if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
1931 raw_spin_lock(&rq->rt.push_lock);
1932 /* Make sure it's still executing */
1933 if (rq->rt.push_flags & RT_PUSH_IPI_EXECUTING) {
1935 * Tell the IPI to restart the loop as things have
1936 * changed since it started.
1938 rq->rt.push_flags |= RT_PUSH_IPI_RESTART;
1939 raw_spin_unlock(&rq->rt.push_lock);
1942 raw_spin_unlock(&rq->rt.push_lock);
1945 /* When here, there's no IPI going around */
1947 rq->rt.push_cpu = rq->cpu;
1948 cpu = find_next_push_cpu(rq);
1949 if (cpu >= nr_cpu_ids)
1952 rq->rt.push_flags = RT_PUSH_IPI_EXECUTING;
1954 irq_work_queue_on(&rq->rt.push_work, cpu);
1957 /* Called from hardirq context */
1958 static void try_to_push_tasks(void *arg)
1960 struct rt_rq *rt_rq = arg;
1961 struct rq *rq, *src_rq;
1965 this_cpu = rt_rq->push_cpu;
1967 /* Paranoid check */
1968 BUG_ON(this_cpu != smp_processor_id());
1970 rq = cpu_rq(this_cpu);
1971 src_rq = rq_of_rt_rq(rt_rq);
1974 if (has_pushable_tasks(rq)) {
1975 raw_spin_lock(&rq->lock);
1977 raw_spin_unlock(&rq->lock);
1980 /* Pass the IPI to the next rt overloaded queue */
1981 raw_spin_lock(&rt_rq->push_lock);
1983 * If the source queue changed since the IPI went out,
1984 * we need to restart the search from that CPU again.
1986 if (rt_rq->push_flags & RT_PUSH_IPI_RESTART) {
1987 rt_rq->push_flags &= ~RT_PUSH_IPI_RESTART;
1988 rt_rq->push_cpu = src_rq->cpu;
1991 cpu = find_next_push_cpu(src_rq);
1993 if (cpu >= nr_cpu_ids)
1994 rt_rq->push_flags &= ~RT_PUSH_IPI_EXECUTING;
1995 raw_spin_unlock(&rt_rq->push_lock);
1997 if (cpu >= nr_cpu_ids)
2001 * It is possible that a restart caused this CPU to be
2002 * chosen again. Don't bother with an IPI, just see if we
2003 * have more to push.
2005 if (unlikely(cpu == rq->cpu))
2008 /* Try the next RT overloaded CPU */
2009 irq_work_queue_on(&rt_rq->push_work, cpu);
2012 static void push_irq_work_func(struct irq_work *work)
2014 struct rt_rq *rt_rq = container_of(work, struct rt_rq, push_work);
2016 try_to_push_tasks(rt_rq);
2018 #endif /* HAVE_RT_PUSH_IPI */
2020 static void pull_rt_task(struct rq *this_rq)
2022 int this_cpu = this_rq->cpu, cpu;
2023 bool resched = false;
2024 struct task_struct *p;
2027 if (likely(!rt_overloaded(this_rq)))
2031 * Match the barrier from rt_set_overloaded; this guarantees that if we
2032 * see overloaded we must also see the rto_mask bit.
2036 #ifdef HAVE_RT_PUSH_IPI
2037 if (sched_feat(RT_PUSH_IPI)) {
2038 tell_cpu_to_push(this_rq);
2043 for_each_cpu(cpu, this_rq->rd->rto_mask) {
2044 if (this_cpu == cpu)
2047 src_rq = cpu_rq(cpu);
2050 * Don't bother taking the src_rq->lock if the next highest
2051 * task is known to be lower-priority than our current task.
2052 * This may look racy, but if this value is about to go
2053 * logically higher, the src_rq will push this task away.
2054 * And if its going logically lower, we do not care
2056 if (src_rq->rt.highest_prio.next >=
2057 this_rq->rt.highest_prio.curr)
2061 * We can potentially drop this_rq's lock in
2062 * double_lock_balance, and another CPU could
2065 double_lock_balance(this_rq, src_rq);
2068 * We can pull only a task, which is pushable
2069 * on its rq, and no others.
2071 p = pick_highest_pushable_task(src_rq, this_cpu);
2074 * Do we have an RT task that preempts
2075 * the to-be-scheduled task?
2077 if (p && (p->prio < this_rq->rt.highest_prio.curr)) {
2078 WARN_ON(p == src_rq->curr);
2079 WARN_ON(!task_on_rq_queued(p));
2082 * There's a chance that p is higher in priority
2083 * than what's currently running on its cpu.
2084 * This is just that p is wakeing up and hasn't
2085 * had a chance to schedule. We only pull
2086 * p if it is lower in priority than the
2087 * current task on the run queue
2089 if (p->prio < src_rq->curr->prio)
2094 deactivate_task(src_rq, p, 0);
2095 set_task_cpu(p, this_cpu);
2096 activate_task(this_rq, p, 0);
2098 * We continue with the search, just in
2099 * case there's an even higher prio task
2100 * in another runqueue. (low likelihood
2105 double_unlock_balance(this_rq, src_rq);
2109 resched_curr(this_rq);
2113 * If we are not running and we are not going to reschedule soon, we should
2114 * try to push tasks away now
2116 static void task_woken_rt(struct rq *rq, struct task_struct *p)
2118 if (!task_running(rq, p) &&
2119 !test_tsk_need_resched(rq->curr) &&
2120 p->nr_cpus_allowed > 1 &&
2121 (dl_task(rq->curr) || rt_task(rq->curr)) &&
2122 (rq->curr->nr_cpus_allowed < 2 ||
2123 rq->curr->prio <= p->prio))
2127 /* Assumes rq->lock is held */
2128 static void rq_online_rt(struct rq *rq)
2130 if (rq->rt.overloaded)
2131 rt_set_overload(rq);
2133 __enable_runtime(rq);
2135 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
2138 /* Assumes rq->lock is held */
2139 static void rq_offline_rt(struct rq *rq)
2141 if (rq->rt.overloaded)
2142 rt_clear_overload(rq);
2144 __disable_runtime(rq);
2146 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
2150 * When switch from the rt queue, we bring ourselves to a position
2151 * that we might want to pull RT tasks from other runqueues.
2153 static void switched_from_rt(struct rq *rq, struct task_struct *p)
2156 * If there are other RT tasks then we will reschedule
2157 * and the scheduling of the other RT tasks will handle
2158 * the balancing. But if we are the last RT task
2159 * we may need to handle the pulling of RT tasks
2162 if (!task_on_rq_queued(p) || rq->rt.rt_nr_running)
2165 queue_pull_task(rq);
2168 void __init init_sched_rt_class(void)
2172 for_each_possible_cpu(i) {
2173 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
2174 GFP_KERNEL, cpu_to_node(i));
2177 #endif /* CONFIG_SMP */
2180 * When switching a task to RT, we may overload the runqueue
2181 * with RT tasks. In this case we try to push them off to
2184 static void switched_to_rt(struct rq *rq, struct task_struct *p)
2187 * If we are already running, then there's nothing
2188 * that needs to be done. But if we are not running
2189 * we may need to preempt the current running task.
2190 * If that current running task is also an RT task
2191 * then see if we can move to another run queue.
2193 if (task_on_rq_queued(p) && rq->curr != p) {
2195 if (p->nr_cpus_allowed > 1 && rq->rt.overloaded)
2196 queue_push_tasks(rq);
2198 if (p->prio < rq->curr->prio)
2200 #endif /* CONFIG_SMP */
2205 * Priority of the task has changed. This may cause
2206 * us to initiate a push or pull.
2209 prio_changed_rt(struct rq *rq, struct task_struct *p, int oldprio)
2211 if (!task_on_rq_queued(p))
2214 if (rq->curr == p) {
2217 * If our priority decreases while running, we
2218 * may need to pull tasks to this runqueue.
2220 if (oldprio < p->prio)
2221 queue_pull_task(rq);
2224 * If there's a higher priority task waiting to run
2227 if (p->prio > rq->rt.highest_prio.curr)
2230 /* For UP simply resched on drop of prio */
2231 if (oldprio < p->prio)
2233 #endif /* CONFIG_SMP */
2236 * This task is not running, but if it is
2237 * greater than the current running task
2240 if (p->prio < rq->curr->prio)
2245 static void watchdog(struct rq *rq, struct task_struct *p)
2247 unsigned long soft, hard;
2249 /* max may change after cur was read, this will be fixed next tick */
2250 soft = task_rlimit(p, RLIMIT_RTTIME);
2251 hard = task_rlimit_max(p, RLIMIT_RTTIME);
2253 if (soft != RLIM_INFINITY) {
2256 if (p->rt.watchdog_stamp != jiffies) {
2258 p->rt.watchdog_stamp = jiffies;
2261 next = DIV_ROUND_UP(min(soft, hard), USEC_PER_SEC/HZ);
2262 if (p->rt.timeout > next)
2263 p->cputime_expires.sched_exp = p->se.sum_exec_runtime;
2267 static void task_tick_rt(struct rq *rq, struct task_struct *p, int queued)
2269 struct sched_rt_entity *rt_se = &p->rt;
2276 * RR tasks need a special form of timeslice management.
2277 * FIFO tasks have no timeslices.
2279 if (p->policy != SCHED_RR)
2282 if (--p->rt.time_slice)
2285 p->rt.time_slice = sched_rr_timeslice;
2288 * Requeue to the end of queue if we (and all of our ancestors) are not
2289 * the only element on the queue
2291 for_each_sched_rt_entity(rt_se) {
2292 if (rt_se->run_list.prev != rt_se->run_list.next) {
2293 requeue_task_rt(rq, p, 0);
2300 static void set_curr_task_rt(struct rq *rq)
2302 struct task_struct *p = rq->curr;
2304 p->se.exec_start = rq_clock_task(rq);
2306 /* The running task is never eligible for pushing */
2307 dequeue_pushable_task(rq, p);
2310 static unsigned int get_rr_interval_rt(struct rq *rq, struct task_struct *task)
2313 * Time slice is 0 for SCHED_FIFO tasks
2315 if (task->policy == SCHED_RR)
2316 return sched_rr_timeslice;
2321 const struct sched_class rt_sched_class = {
2322 .next = &fair_sched_class,
2323 .enqueue_task = enqueue_task_rt,
2324 .dequeue_task = dequeue_task_rt,
2325 .yield_task = yield_task_rt,
2327 .check_preempt_curr = check_preempt_curr_rt,
2329 .pick_next_task = pick_next_task_rt,
2330 .put_prev_task = put_prev_task_rt,
2333 .select_task_rq = select_task_rq_rt,
2335 .set_cpus_allowed = set_cpus_allowed_common,
2336 .rq_online = rq_online_rt,
2337 .rq_offline = rq_offline_rt,
2338 .task_woken = task_woken_rt,
2339 .switched_from = switched_from_rt,
2342 .set_curr_task = set_curr_task_rt,
2343 .task_tick = task_tick_rt,
2345 .get_rr_interval = get_rr_interval_rt,
2347 .prio_changed = prio_changed_rt,
2348 .switched_to = switched_to_rt,
2350 .update_curr = update_curr_rt,
2353 #ifdef CONFIG_SCHED_DEBUG
2354 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2356 void print_rt_stats(struct seq_file *m, int cpu)
2359 struct rt_rq *rt_rq;
2362 for_each_rt_rq(rt_rq, iter, cpu_rq(cpu))
2363 print_rt_rq(m, cpu, rt_rq);
2366 #endif /* CONFIG_SCHED_DEBUG */