1 // SPDX-License-Identifier: GPL-2.0
3 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
10 int sched_rr_timeslice
= RR_TIMESLICE
;
11 int sysctl_sched_rr_timeslice
= (MSEC_PER_SEC
/ HZ
) * 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 void init_rt_rq(struct rt_rq
*rt_rq
)
77 struct rt_prio_array
*array
;
80 array
= &rt_rq
->active
;
81 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
82 INIT_LIST_HEAD(array
->queue
+ i
);
83 __clear_bit(i
, array
->bitmap
);
85 /* delimiter for bitsearch: */
86 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
88 #if defined CONFIG_SMP
89 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
90 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
91 rt_rq
->rt_nr_migratory
= 0;
92 rt_rq
->overloaded
= 0;
93 plist_head_init(&rt_rq
->pushable_tasks
);
94 #endif /* CONFIG_SMP */
95 /* We start is dequeued state, because no RT tasks are queued */
99 rt_rq
->rt_throttled
= 0;
100 rt_rq
->rt_runtime
= 0;
101 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
104 #ifdef CONFIG_RT_GROUP_SCHED
105 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
107 hrtimer_cancel(&rt_b
->rt_period_timer
);
110 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
112 static inline struct task_struct
*rt_task_of(struct sched_rt_entity
*rt_se
)
114 #ifdef CONFIG_SCHED_DEBUG
115 WARN_ON_ONCE(!rt_entity_is_task(rt_se
));
117 return container_of(rt_se
, struct task_struct
, rt
);
120 static inline struct rq
*rq_of_rt_rq(struct rt_rq
*rt_rq
)
125 static inline struct rt_rq
*rt_rq_of_se(struct sched_rt_entity
*rt_se
)
130 static inline struct rq
*rq_of_rt_se(struct sched_rt_entity
*rt_se
)
132 struct rt_rq
*rt_rq
= rt_se
->rt_rq
;
137 void free_rt_sched_group(struct task_group
*tg
)
142 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
144 for_each_possible_cpu(i
) {
155 void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
156 struct sched_rt_entity
*rt_se
, int cpu
,
157 struct sched_rt_entity
*parent
)
159 struct rq
*rq
= cpu_rq(cpu
);
161 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
162 rt_rq
->rt_nr_boosted
= 0;
166 tg
->rt_rq
[cpu
] = rt_rq
;
167 tg
->rt_se
[cpu
] = rt_se
;
173 rt_se
->rt_rq
= &rq
->rt
;
175 rt_se
->rt_rq
= parent
->my_q
;
178 rt_se
->parent
= parent
;
179 INIT_LIST_HEAD(&rt_se
->run_list
);
182 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
185 struct sched_rt_entity
*rt_se
;
188 tg
->rt_rq
= kcalloc(nr_cpu_ids
, sizeof(rt_rq
), GFP_KERNEL
);
191 tg
->rt_se
= kcalloc(nr_cpu_ids
, sizeof(rt_se
), GFP_KERNEL
);
195 init_rt_bandwidth(&tg
->rt_bandwidth
,
196 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
198 for_each_possible_cpu(i
) {
199 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
200 GFP_KERNEL
, cpu_to_node(i
));
204 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
205 GFP_KERNEL
, cpu_to_node(i
));
210 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
211 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, parent
->rt_se
[i
]);
222 #else /* CONFIG_RT_GROUP_SCHED */
224 #define rt_entity_is_task(rt_se) (1)
226 static inline struct task_struct
*rt_task_of(struct sched_rt_entity
*rt_se
)
228 return container_of(rt_se
, struct task_struct
, rt
);
231 static inline struct rq
*rq_of_rt_rq(struct rt_rq
*rt_rq
)
233 return container_of(rt_rq
, struct rq
, rt
);
236 static inline struct rq
*rq_of_rt_se(struct sched_rt_entity
*rt_se
)
238 struct task_struct
*p
= rt_task_of(rt_se
);
243 static inline struct rt_rq
*rt_rq_of_se(struct sched_rt_entity
*rt_se
)
245 struct rq
*rq
= rq_of_rt_se(rt_se
);
250 void free_rt_sched_group(struct task_group
*tg
) { }
252 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
256 #endif /* CONFIG_RT_GROUP_SCHED */
260 static void pull_rt_task(struct rq
*this_rq
);
262 static inline bool need_pull_rt_task(struct rq
*rq
, struct task_struct
*prev
)
264 /* Try to pull RT tasks here if we lower this rq's prio */
265 return rq
->rt
.highest_prio
.curr
> prev
->prio
;
268 static inline int rt_overloaded(struct rq
*rq
)
270 return atomic_read(&rq
->rd
->rto_count
);
273 static inline void rt_set_overload(struct rq
*rq
)
278 cpumask_set_cpu(rq
->cpu
, rq
->rd
->rto_mask
);
280 * Make sure the mask is visible before we set
281 * the overload count. That is checked to determine
282 * if we should look at the mask. It would be a shame
283 * if we looked at the mask, but the mask was not
286 * Matched by the barrier in pull_rt_task().
289 atomic_inc(&rq
->rd
->rto_count
);
292 static inline void rt_clear_overload(struct rq
*rq
)
297 /* the order here really doesn't matter */
298 atomic_dec(&rq
->rd
->rto_count
);
299 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->rto_mask
);
302 static void update_rt_migration(struct rt_rq
*rt_rq
)
304 if (rt_rq
->rt_nr_migratory
&& rt_rq
->rt_nr_total
> 1) {
305 if (!rt_rq
->overloaded
) {
306 rt_set_overload(rq_of_rt_rq(rt_rq
));
307 rt_rq
->overloaded
= 1;
309 } else if (rt_rq
->overloaded
) {
310 rt_clear_overload(rq_of_rt_rq(rt_rq
));
311 rt_rq
->overloaded
= 0;
315 static void inc_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
317 struct task_struct
*p
;
319 if (!rt_entity_is_task(rt_se
))
322 p
= rt_task_of(rt_se
);
323 rt_rq
= &rq_of_rt_rq(rt_rq
)->rt
;
325 rt_rq
->rt_nr_total
++;
326 if (p
->nr_cpus_allowed
> 1)
327 rt_rq
->rt_nr_migratory
++;
329 update_rt_migration(rt_rq
);
332 static void dec_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
334 struct task_struct
*p
;
336 if (!rt_entity_is_task(rt_se
))
339 p
= rt_task_of(rt_se
);
340 rt_rq
= &rq_of_rt_rq(rt_rq
)->rt
;
342 rt_rq
->rt_nr_total
--;
343 if (p
->nr_cpus_allowed
> 1)
344 rt_rq
->rt_nr_migratory
--;
346 update_rt_migration(rt_rq
);
349 static inline int has_pushable_tasks(struct rq
*rq
)
351 return !plist_head_empty(&rq
->rt
.pushable_tasks
);
354 static DEFINE_PER_CPU(struct callback_head
, rt_push_head
);
355 static DEFINE_PER_CPU(struct callback_head
, rt_pull_head
);
357 static void push_rt_tasks(struct rq
*);
358 static void pull_rt_task(struct rq
*);
360 static inline void rt_queue_push_tasks(struct rq
*rq
)
362 if (!has_pushable_tasks(rq
))
365 queue_balance_callback(rq
, &per_cpu(rt_push_head
, rq
->cpu
), push_rt_tasks
);
368 static inline void rt_queue_pull_task(struct rq
*rq
)
370 queue_balance_callback(rq
, &per_cpu(rt_pull_head
, rq
->cpu
), pull_rt_task
);
373 static void enqueue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
375 plist_del(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
376 plist_node_init(&p
->pushable_tasks
, p
->prio
);
377 plist_add(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
379 /* Update the highest prio pushable task */
380 if (p
->prio
< rq
->rt
.highest_prio
.next
)
381 rq
->rt
.highest_prio
.next
= p
->prio
;
384 static void dequeue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
386 plist_del(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
388 /* Update the new highest prio pushable task */
389 if (has_pushable_tasks(rq
)) {
390 p
= plist_first_entry(&rq
->rt
.pushable_tasks
,
391 struct task_struct
, pushable_tasks
);
392 rq
->rt
.highest_prio
.next
= p
->prio
;
394 rq
->rt
.highest_prio
.next
= MAX_RT_PRIO
;
399 static inline void enqueue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
403 static inline void dequeue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
408 void inc_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
413 void dec_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
417 static inline bool need_pull_rt_task(struct rq
*rq
, struct task_struct
*prev
)
422 static inline void pull_rt_task(struct rq
*this_rq
)
426 static inline void rt_queue_push_tasks(struct rq
*rq
)
429 #endif /* CONFIG_SMP */
431 static void enqueue_top_rt_rq(struct rt_rq
*rt_rq
);
432 static void dequeue_top_rt_rq(struct rt_rq
*rt_rq
);
434 static inline int on_rt_rq(struct sched_rt_entity
*rt_se
)
439 #ifdef CONFIG_RT_GROUP_SCHED
441 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
446 return rt_rq
->rt_runtime
;
449 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
451 return ktime_to_ns(rt_rq
->tg
->rt_bandwidth
.rt_period
);
454 typedef struct task_group
*rt_rq_iter_t
;
456 static inline struct task_group
*next_task_group(struct task_group
*tg
)
459 tg
= list_entry_rcu(tg
->list
.next
,
460 typeof(struct task_group
), list
);
461 } while (&tg
->list
!= &task_groups
&& task_group_is_autogroup(tg
));
463 if (&tg
->list
== &task_groups
)
469 #define for_each_rt_rq(rt_rq, iter, rq) \
470 for (iter = container_of(&task_groups, typeof(*iter), list); \
471 (iter = next_task_group(iter)) && \
472 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
474 #define for_each_sched_rt_entity(rt_se) \
475 for (; rt_se; rt_se = rt_se->parent)
477 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
482 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
);
483 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
);
485 static void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
487 struct task_struct
*curr
= rq_of_rt_rq(rt_rq
)->curr
;
488 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
489 struct sched_rt_entity
*rt_se
;
491 int cpu
= cpu_of(rq
);
493 rt_se
= rt_rq
->tg
->rt_se
[cpu
];
495 if (rt_rq
->rt_nr_running
) {
497 enqueue_top_rt_rq(rt_rq
);
498 else if (!on_rt_rq(rt_se
))
499 enqueue_rt_entity(rt_se
, 0);
501 if (rt_rq
->highest_prio
.curr
< curr
->prio
)
506 static void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
508 struct sched_rt_entity
*rt_se
;
509 int cpu
= cpu_of(rq_of_rt_rq(rt_rq
));
511 rt_se
= rt_rq
->tg
->rt_se
[cpu
];
514 dequeue_top_rt_rq(rt_rq
);
515 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
516 cpufreq_update_util(rq_of_rt_rq(rt_rq
), 0);
518 else if (on_rt_rq(rt_se
))
519 dequeue_rt_entity(rt_se
, 0);
522 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
524 return rt_rq
->rt_throttled
&& !rt_rq
->rt_nr_boosted
;
527 static int rt_se_boosted(struct sched_rt_entity
*rt_se
)
529 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
530 struct task_struct
*p
;
533 return !!rt_rq
->rt_nr_boosted
;
535 p
= rt_task_of(rt_se
);
536 return p
->prio
!= p
->normal_prio
;
540 static inline const struct cpumask
*sched_rt_period_mask(void)
542 return this_rq()->rd
->span
;
545 static inline const struct cpumask
*sched_rt_period_mask(void)
547 return cpu_online_mask
;
552 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
554 return container_of(rt_b
, struct task_group
, rt_bandwidth
)->rt_rq
[cpu
];
557 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
559 return &rt_rq
->tg
->rt_bandwidth
;
562 #else /* !CONFIG_RT_GROUP_SCHED */
564 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
566 return rt_rq
->rt_runtime
;
569 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
571 return ktime_to_ns(def_rt_bandwidth
.rt_period
);
574 typedef struct rt_rq
*rt_rq_iter_t
;
576 #define for_each_rt_rq(rt_rq, iter, rq) \
577 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
579 #define for_each_sched_rt_entity(rt_se) \
580 for (; rt_se; rt_se = NULL)
582 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
587 static inline void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
589 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
591 if (!rt_rq
->rt_nr_running
)
594 enqueue_top_rt_rq(rt_rq
);
598 static inline void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
600 dequeue_top_rt_rq(rt_rq
);
603 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
605 return rt_rq
->rt_throttled
;
608 static inline const struct cpumask
*sched_rt_period_mask(void)
610 return cpu_online_mask
;
614 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
616 return &cpu_rq(cpu
)->rt
;
619 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
621 return &def_rt_bandwidth
;
624 #endif /* CONFIG_RT_GROUP_SCHED */
626 bool sched_rt_bandwidth_account(struct rt_rq
*rt_rq
)
628 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
630 return (hrtimer_active(&rt_b
->rt_period_timer
) ||
631 rt_rq
->rt_time
< rt_b
->rt_runtime
);
636 * We ran out of runtime, see if we can borrow some from our neighbours.
638 static void do_balance_runtime(struct rt_rq
*rt_rq
)
640 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
641 struct root_domain
*rd
= rq_of_rt_rq(rt_rq
)->rd
;
645 weight
= cpumask_weight(rd
->span
);
647 raw_spin_lock(&rt_b
->rt_runtime_lock
);
648 rt_period
= ktime_to_ns(rt_b
->rt_period
);
649 for_each_cpu(i
, rd
->span
) {
650 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
656 raw_spin_lock(&iter
->rt_runtime_lock
);
658 * Either all rqs have inf runtime and there's nothing to steal
659 * or __disable_runtime() below sets a specific rq to inf to
660 * indicate its been disabled and disalow stealing.
662 if (iter
->rt_runtime
== RUNTIME_INF
)
666 * From runqueues with spare time, take 1/n part of their
667 * spare time, but no more than our period.
669 diff
= iter
->rt_runtime
- iter
->rt_time
;
671 diff
= div_u64((u64
)diff
, weight
);
672 if (rt_rq
->rt_runtime
+ diff
> rt_period
)
673 diff
= rt_period
- rt_rq
->rt_runtime
;
674 iter
->rt_runtime
-= diff
;
675 rt_rq
->rt_runtime
+= diff
;
676 if (rt_rq
->rt_runtime
== rt_period
) {
677 raw_spin_unlock(&iter
->rt_runtime_lock
);
682 raw_spin_unlock(&iter
->rt_runtime_lock
);
684 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
688 * Ensure this RQ takes back all the runtime it lend to its neighbours.
690 static void __disable_runtime(struct rq
*rq
)
692 struct root_domain
*rd
= rq
->rd
;
696 if (unlikely(!scheduler_running
))
699 for_each_rt_rq(rt_rq
, iter
, rq
) {
700 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
704 raw_spin_lock(&rt_b
->rt_runtime_lock
);
705 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
707 * Either we're all inf and nobody needs to borrow, or we're
708 * already disabled and thus have nothing to do, or we have
709 * exactly the right amount of runtime to take out.
711 if (rt_rq
->rt_runtime
== RUNTIME_INF
||
712 rt_rq
->rt_runtime
== rt_b
->rt_runtime
)
714 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
717 * Calculate the difference between what we started out with
718 * and what we current have, that's the amount of runtime
719 * we lend and now have to reclaim.
721 want
= rt_b
->rt_runtime
- rt_rq
->rt_runtime
;
724 * Greedy reclaim, take back as much as we can.
726 for_each_cpu(i
, rd
->span
) {
727 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
731 * Can't reclaim from ourselves or disabled runqueues.
733 if (iter
== rt_rq
|| iter
->rt_runtime
== RUNTIME_INF
)
736 raw_spin_lock(&iter
->rt_runtime_lock
);
738 diff
= min_t(s64
, iter
->rt_runtime
, want
);
739 iter
->rt_runtime
-= diff
;
742 iter
->rt_runtime
-= want
;
745 raw_spin_unlock(&iter
->rt_runtime_lock
);
751 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
753 * We cannot be left wanting - that would mean some runtime
754 * leaked out of the system.
759 * Disable all the borrow logic by pretending we have inf
760 * runtime - in which case borrowing doesn't make sense.
762 rt_rq
->rt_runtime
= RUNTIME_INF
;
763 rt_rq
->rt_throttled
= 0;
764 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
765 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
767 /* Make rt_rq available for pick_next_task() */
768 sched_rt_rq_enqueue(rt_rq
);
772 static void __enable_runtime(struct rq
*rq
)
777 if (unlikely(!scheduler_running
))
781 * Reset each runqueue's bandwidth settings
783 for_each_rt_rq(rt_rq
, iter
, rq
) {
784 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
786 raw_spin_lock(&rt_b
->rt_runtime_lock
);
787 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
788 rt_rq
->rt_runtime
= rt_b
->rt_runtime
;
790 rt_rq
->rt_throttled
= 0;
791 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
792 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
796 static void balance_runtime(struct rt_rq
*rt_rq
)
798 if (!sched_feat(RT_RUNTIME_SHARE
))
801 if (rt_rq
->rt_time
> rt_rq
->rt_runtime
) {
802 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
803 do_balance_runtime(rt_rq
);
804 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
807 #else /* !CONFIG_SMP */
808 static inline void balance_runtime(struct rt_rq
*rt_rq
) {}
809 #endif /* CONFIG_SMP */
811 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
)
813 int i
, idle
= 1, throttled
= 0;
814 const struct cpumask
*span
;
816 span
= sched_rt_period_mask();
817 #ifdef CONFIG_RT_GROUP_SCHED
819 * FIXME: isolated CPUs should really leave the root task group,
820 * whether they are isolcpus or were isolated via cpusets, lest
821 * the timer run on a CPU which does not service all runqueues,
822 * potentially leaving other CPUs indefinitely throttled. If
823 * isolation is really required, the user will turn the throttle
824 * off to kill the perturbations it causes anyway. Meanwhile,
825 * this maintains functionality for boot and/or troubleshooting.
827 if (rt_b
== &root_task_group
.rt_bandwidth
)
828 span
= cpu_online_mask
;
830 for_each_cpu(i
, span
) {
832 struct rt_rq
*rt_rq
= sched_rt_period_rt_rq(rt_b
, i
);
833 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
837 * When span == cpu_online_mask, taking each rq->lock
838 * can be time-consuming. Try to avoid it when possible.
840 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
841 if (!sched_feat(RT_RUNTIME_SHARE
) && rt_rq
->rt_runtime
!= RUNTIME_INF
)
842 rt_rq
->rt_runtime
= rt_b
->rt_runtime
;
843 skip
= !rt_rq
->rt_time
&& !rt_rq
->rt_nr_running
;
844 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
848 raw_spin_lock(&rq
->lock
);
851 if (rt_rq
->rt_time
) {
854 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
855 if (rt_rq
->rt_throttled
)
856 balance_runtime(rt_rq
);
857 runtime
= rt_rq
->rt_runtime
;
858 rt_rq
->rt_time
-= min(rt_rq
->rt_time
, overrun
*runtime
);
859 if (rt_rq
->rt_throttled
&& rt_rq
->rt_time
< runtime
) {
860 rt_rq
->rt_throttled
= 0;
864 * When we're idle and a woken (rt) task is
865 * throttled check_preempt_curr() will set
866 * skip_update and the time between the wakeup
867 * and this unthrottle will get accounted as
870 if (rt_rq
->rt_nr_running
&& rq
->curr
== rq
->idle
)
871 rq_clock_cancel_skipupdate(rq
);
873 if (rt_rq
->rt_time
|| rt_rq
->rt_nr_running
)
875 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
876 } else if (rt_rq
->rt_nr_running
) {
878 if (!rt_rq_throttled(rt_rq
))
881 if (rt_rq
->rt_throttled
)
885 sched_rt_rq_enqueue(rt_rq
);
886 raw_spin_unlock(&rq
->lock
);
889 if (!throttled
&& (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
))
895 static inline int rt_se_prio(struct sched_rt_entity
*rt_se
)
897 #ifdef CONFIG_RT_GROUP_SCHED
898 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
901 return rt_rq
->highest_prio
.curr
;
904 return rt_task_of(rt_se
)->prio
;
907 static int sched_rt_runtime_exceeded(struct rt_rq
*rt_rq
)
909 u64 runtime
= sched_rt_runtime(rt_rq
);
911 if (rt_rq
->rt_throttled
)
912 return rt_rq_throttled(rt_rq
);
914 if (runtime
>= sched_rt_period(rt_rq
))
917 balance_runtime(rt_rq
);
918 runtime
= sched_rt_runtime(rt_rq
);
919 if (runtime
== RUNTIME_INF
)
922 if (rt_rq
->rt_time
> runtime
) {
923 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
926 * Don't actually throttle groups that have no runtime assigned
927 * but accrue some time due to boosting.
929 if (likely(rt_b
->rt_runtime
)) {
930 rt_rq
->rt_throttled
= 1;
931 printk_deferred_once("sched: RT throttling activated\n");
934 * In case we did anyway, make it go away,
935 * replenishment is a joke, since it will replenish us
941 if (rt_rq_throttled(rt_rq
)) {
942 sched_rt_rq_dequeue(rt_rq
);
951 * Update the current task's runtime statistics. Skip current tasks that
952 * are not in our scheduling class.
954 static void update_curr_rt(struct rq
*rq
)
956 struct task_struct
*curr
= rq
->curr
;
957 struct sched_rt_entity
*rt_se
= &curr
->rt
;
961 if (curr
->sched_class
!= &rt_sched_class
)
964 now
= rq_clock_task(rq
);
965 delta_exec
= now
- curr
->se
.exec_start
;
966 if (unlikely((s64
)delta_exec
<= 0))
969 schedstat_set(curr
->se
.statistics
.exec_max
,
970 max(curr
->se
.statistics
.exec_max
, delta_exec
));
972 curr
->se
.sum_exec_runtime
+= delta_exec
;
973 account_group_exec_runtime(curr
, delta_exec
);
975 curr
->se
.exec_start
= now
;
976 cgroup_account_cputime(curr
, delta_exec
);
978 if (!rt_bandwidth_enabled())
981 for_each_sched_rt_entity(rt_se
) {
982 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
984 if (sched_rt_runtime(rt_rq
) != RUNTIME_INF
) {
985 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
986 rt_rq
->rt_time
+= delta_exec
;
987 if (sched_rt_runtime_exceeded(rt_rq
))
989 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
995 dequeue_top_rt_rq(struct rt_rq
*rt_rq
)
997 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
999 BUG_ON(&rq
->rt
!= rt_rq
);
1001 if (!rt_rq
->rt_queued
)
1004 BUG_ON(!rq
->nr_running
);
1006 sub_nr_running(rq
, rt_rq
->rt_nr_running
);
1007 rt_rq
->rt_queued
= 0;
1012 enqueue_top_rt_rq(struct rt_rq
*rt_rq
)
1014 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1016 BUG_ON(&rq
->rt
!= rt_rq
);
1018 if (rt_rq
->rt_queued
)
1021 if (rt_rq_throttled(rt_rq
))
1024 if (rt_rq
->rt_nr_running
) {
1025 add_nr_running(rq
, rt_rq
->rt_nr_running
);
1026 rt_rq
->rt_queued
= 1;
1029 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
1030 cpufreq_update_util(rq
, 0);
1033 #if defined CONFIG_SMP
1036 inc_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
)
1038 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1040 #ifdef CONFIG_RT_GROUP_SCHED
1042 * Change rq's cpupri only if rt_rq is the top queue.
1044 if (&rq
->rt
!= rt_rq
)
1047 if (rq
->online
&& prio
< prev_prio
)
1048 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, prio
);
1052 dec_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
)
1054 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1056 #ifdef CONFIG_RT_GROUP_SCHED
1058 * Change rq's cpupri only if rt_rq is the top queue.
1060 if (&rq
->rt
!= rt_rq
)
1063 if (rq
->online
&& rt_rq
->highest_prio
.curr
!= prev_prio
)
1064 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rt_rq
->highest_prio
.curr
);
1067 #else /* CONFIG_SMP */
1070 void inc_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
) {}
1072 void dec_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
) {}
1074 #endif /* CONFIG_SMP */
1076 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1078 inc_rt_prio(struct rt_rq
*rt_rq
, int prio
)
1080 int prev_prio
= rt_rq
->highest_prio
.curr
;
1082 if (prio
< prev_prio
)
1083 rt_rq
->highest_prio
.curr
= prio
;
1085 inc_rt_prio_smp(rt_rq
, prio
, prev_prio
);
1089 dec_rt_prio(struct rt_rq
*rt_rq
, int prio
)
1091 int prev_prio
= rt_rq
->highest_prio
.curr
;
1093 if (rt_rq
->rt_nr_running
) {
1095 WARN_ON(prio
< prev_prio
);
1098 * This may have been our highest task, and therefore
1099 * we may have some recomputation to do
1101 if (prio
== prev_prio
) {
1102 struct rt_prio_array
*array
= &rt_rq
->active
;
1104 rt_rq
->highest_prio
.curr
=
1105 sched_find_first_bit(array
->bitmap
);
1109 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
1111 dec_rt_prio_smp(rt_rq
, prio
, prev_prio
);
1116 static inline void inc_rt_prio(struct rt_rq
*rt_rq
, int prio
) {}
1117 static inline void dec_rt_prio(struct rt_rq
*rt_rq
, int prio
) {}
1119 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1121 #ifdef CONFIG_RT_GROUP_SCHED
1124 inc_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
++;
1130 start_rt_bandwidth(&rt_rq
->tg
->rt_bandwidth
);
1134 dec_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1136 if (rt_se_boosted(rt_se
))
1137 rt_rq
->rt_nr_boosted
--;
1139 WARN_ON(!rt_rq
->rt_nr_running
&& rt_rq
->rt_nr_boosted
);
1142 #else /* CONFIG_RT_GROUP_SCHED */
1145 inc_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1147 start_rt_bandwidth(&def_rt_bandwidth
);
1151 void dec_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
) {}
1153 #endif /* CONFIG_RT_GROUP_SCHED */
1156 unsigned int rt_se_nr_running(struct sched_rt_entity
*rt_se
)
1158 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1161 return group_rq
->rt_nr_running
;
1167 unsigned int rt_se_rr_nr_running(struct sched_rt_entity
*rt_se
)
1169 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1170 struct task_struct
*tsk
;
1173 return group_rq
->rr_nr_running
;
1175 tsk
= rt_task_of(rt_se
);
1177 return (tsk
->policy
== SCHED_RR
) ? 1 : 0;
1181 void inc_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1183 int prio
= rt_se_prio(rt_se
);
1185 WARN_ON(!rt_prio(prio
));
1186 rt_rq
->rt_nr_running
+= rt_se_nr_running(rt_se
);
1187 rt_rq
->rr_nr_running
+= rt_se_rr_nr_running(rt_se
);
1189 inc_rt_prio(rt_rq
, prio
);
1190 inc_rt_migration(rt_se
, rt_rq
);
1191 inc_rt_group(rt_se
, rt_rq
);
1195 void dec_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1197 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
1198 WARN_ON(!rt_rq
->rt_nr_running
);
1199 rt_rq
->rt_nr_running
-= rt_se_nr_running(rt_se
);
1200 rt_rq
->rr_nr_running
-= rt_se_rr_nr_running(rt_se
);
1202 dec_rt_prio(rt_rq
, rt_se_prio(rt_se
));
1203 dec_rt_migration(rt_se
, rt_rq
);
1204 dec_rt_group(rt_se
, rt_rq
);
1208 * Change rt_se->run_list location unless SAVE && !MOVE
1210 * assumes ENQUEUE/DEQUEUE flags match
1212 static inline bool move_entity(unsigned int flags
)
1214 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) == DEQUEUE_SAVE
)
1220 static void __delist_rt_entity(struct sched_rt_entity
*rt_se
, struct rt_prio_array
*array
)
1222 list_del_init(&rt_se
->run_list
);
1224 if (list_empty(array
->queue
+ rt_se_prio(rt_se
)))
1225 __clear_bit(rt_se_prio(rt_se
), array
->bitmap
);
1230 static void __enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1232 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
1233 struct rt_prio_array
*array
= &rt_rq
->active
;
1234 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1235 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
1238 * Don't enqueue the group if its throttled, or when empty.
1239 * The latter is a consequence of the former when a child group
1240 * get throttled and the current group doesn't have any other
1243 if (group_rq
&& (rt_rq_throttled(group_rq
) || !group_rq
->rt_nr_running
)) {
1245 __delist_rt_entity(rt_se
, array
);
1249 if (move_entity(flags
)) {
1250 WARN_ON_ONCE(rt_se
->on_list
);
1251 if (flags
& ENQUEUE_HEAD
)
1252 list_add(&rt_se
->run_list
, queue
);
1254 list_add_tail(&rt_se
->run_list
, queue
);
1256 __set_bit(rt_se_prio(rt_se
), array
->bitmap
);
1261 inc_rt_tasks(rt_se
, rt_rq
);
1264 static void __dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1266 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
1267 struct rt_prio_array
*array
= &rt_rq
->active
;
1269 if (move_entity(flags
)) {
1270 WARN_ON_ONCE(!rt_se
->on_list
);
1271 __delist_rt_entity(rt_se
, array
);
1275 dec_rt_tasks(rt_se
, rt_rq
);
1279 * Because the prio of an upper entry depends on the lower
1280 * entries, we must remove entries top - down.
1282 static void dequeue_rt_stack(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1284 struct sched_rt_entity
*back
= NULL
;
1286 for_each_sched_rt_entity(rt_se
) {
1291 dequeue_top_rt_rq(rt_rq_of_se(back
));
1293 for (rt_se
= back
; rt_se
; rt_se
= rt_se
->back
) {
1294 if (on_rt_rq(rt_se
))
1295 __dequeue_rt_entity(rt_se
, flags
);
1299 static void enqueue_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
);
1304 for_each_sched_rt_entity(rt_se
)
1305 __enqueue_rt_entity(rt_se
, flags
);
1306 enqueue_top_rt_rq(&rq
->rt
);
1309 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1311 struct rq
*rq
= rq_of_rt_se(rt_se
);
1313 dequeue_rt_stack(rt_se
, flags
);
1315 for_each_sched_rt_entity(rt_se
) {
1316 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
1318 if (rt_rq
&& rt_rq
->rt_nr_running
)
1319 __enqueue_rt_entity(rt_se
, flags
);
1321 enqueue_top_rt_rq(&rq
->rt
);
1325 * Adding/removing a task to/from a priority array:
1328 enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1330 struct sched_rt_entity
*rt_se
= &p
->rt
;
1332 if (flags
& ENQUEUE_WAKEUP
)
1335 enqueue_rt_entity(rt_se
, flags
);
1337 if (!task_current(rq
, p
) && p
->nr_cpus_allowed
> 1)
1338 enqueue_pushable_task(rq
, p
);
1341 static void dequeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1343 struct sched_rt_entity
*rt_se
= &p
->rt
;
1346 dequeue_rt_entity(rt_se
, flags
);
1348 dequeue_pushable_task(rq
, p
);
1352 * Put task to the head or the end of the run list without the overhead of
1353 * dequeue followed by enqueue.
1356 requeue_rt_entity(struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
, int head
)
1358 if (on_rt_rq(rt_se
)) {
1359 struct rt_prio_array
*array
= &rt_rq
->active
;
1360 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
1363 list_move(&rt_se
->run_list
, queue
);
1365 list_move_tail(&rt_se
->run_list
, queue
);
1369 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int head
)
1371 struct sched_rt_entity
*rt_se
= &p
->rt
;
1372 struct rt_rq
*rt_rq
;
1374 for_each_sched_rt_entity(rt_se
) {
1375 rt_rq
= rt_rq_of_se(rt_se
);
1376 requeue_rt_entity(rt_rq
, rt_se
, head
);
1380 static void yield_task_rt(struct rq
*rq
)
1382 requeue_task_rt(rq
, rq
->curr
, 0);
1386 static int find_lowest_rq(struct task_struct
*task
);
1389 select_task_rq_rt(struct task_struct
*p
, int cpu
, int sd_flag
, int flags
)
1391 struct task_struct
*curr
;
1394 /* For anything but wake ups, just return the task_cpu */
1395 if (sd_flag
!= SD_BALANCE_WAKE
&& sd_flag
!= SD_BALANCE_FORK
)
1401 curr
= READ_ONCE(rq
->curr
); /* unlocked access */
1404 * If the current task on @p's runqueue is an RT task, then
1405 * try to see if we can wake this RT task up on another
1406 * runqueue. Otherwise simply start this RT task
1407 * on its current runqueue.
1409 * We want to avoid overloading runqueues. If the woken
1410 * task is a higher priority, then it will stay on this CPU
1411 * and the lower prio task should be moved to another CPU.
1412 * Even though this will probably make the lower prio task
1413 * lose its cache, we do not want to bounce a higher task
1414 * around just because it gave up its CPU, perhaps for a
1417 * For equal prio tasks, we just let the scheduler sort it out.
1419 * Otherwise, just let it ride on the affined RQ and the
1420 * post-schedule router will push the preempted task away
1422 * This test is optimistic, if we get it wrong the load-balancer
1423 * will have to sort it out.
1425 if (curr
&& unlikely(rt_task(curr
)) &&
1426 (curr
->nr_cpus_allowed
< 2 ||
1427 curr
->prio
<= p
->prio
)) {
1428 int target
= find_lowest_rq(p
);
1431 * Don't bother moving it if the destination CPU is
1432 * not running a lower priority task.
1435 p
->prio
< cpu_rq(target
)->rt
.highest_prio
.curr
)
1444 static void check_preempt_equal_prio(struct rq
*rq
, struct task_struct
*p
)
1447 * Current can't be migrated, useless to reschedule,
1448 * let's hope p can move out.
1450 if (rq
->curr
->nr_cpus_allowed
== 1 ||
1451 !cpupri_find(&rq
->rd
->cpupri
, rq
->curr
, NULL
))
1455 * p is migratable, so let's not schedule it and
1456 * see if it is pushed or pulled somewhere else.
1458 if (p
->nr_cpus_allowed
!= 1
1459 && cpupri_find(&rq
->rd
->cpupri
, p
, NULL
))
1463 * There appear to be other CPUs that can accept
1464 * the current task but none can run 'p', so lets reschedule
1465 * to try and push the current task away:
1467 requeue_task_rt(rq
, p
, 1);
1471 #endif /* CONFIG_SMP */
1474 * Preempt the current task with a newly woken task if needed:
1476 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1478 if (p
->prio
< rq
->curr
->prio
) {
1487 * - the newly woken task is of equal priority to the current task
1488 * - the newly woken task is non-migratable while current is migratable
1489 * - current will be preempted on the next reschedule
1491 * we should check to see if current can readily move to a different
1492 * cpu. If so, we will reschedule to allow the push logic to try
1493 * to move current somewhere else, making room for our non-migratable
1496 if (p
->prio
== rq
->curr
->prio
&& !test_tsk_need_resched(rq
->curr
))
1497 check_preempt_equal_prio(rq
, p
);
1501 static struct sched_rt_entity
*pick_next_rt_entity(struct rq
*rq
,
1502 struct rt_rq
*rt_rq
)
1504 struct rt_prio_array
*array
= &rt_rq
->active
;
1505 struct sched_rt_entity
*next
= NULL
;
1506 struct list_head
*queue
;
1509 idx
= sched_find_first_bit(array
->bitmap
);
1510 BUG_ON(idx
>= MAX_RT_PRIO
);
1512 queue
= array
->queue
+ idx
;
1513 next
= list_entry(queue
->next
, struct sched_rt_entity
, run_list
);
1518 static struct task_struct
*_pick_next_task_rt(struct rq
*rq
)
1520 struct sched_rt_entity
*rt_se
;
1521 struct task_struct
*p
;
1522 struct rt_rq
*rt_rq
= &rq
->rt
;
1525 rt_se
= pick_next_rt_entity(rq
, rt_rq
);
1527 rt_rq
= group_rt_rq(rt_se
);
1530 p
= rt_task_of(rt_se
);
1531 p
->se
.exec_start
= rq_clock_task(rq
);
1536 static struct task_struct
*
1537 pick_next_task_rt(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
1539 struct task_struct
*p
;
1540 struct rt_rq
*rt_rq
= &rq
->rt
;
1542 if (need_pull_rt_task(rq
, prev
)) {
1544 * This is OK, because current is on_cpu, which avoids it being
1545 * picked for load-balance and preemption/IRQs are still
1546 * disabled avoiding further scheduler activity on it and we're
1547 * being very careful to re-start the picking loop.
1549 rq_unpin_lock(rq
, rf
);
1551 rq_repin_lock(rq
, rf
);
1553 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1554 * means a dl or stop task can slip in, in which case we need
1555 * to re-start task selection.
1557 if (unlikely((rq
->stop
&& task_on_rq_queued(rq
->stop
)) ||
1558 rq
->dl
.dl_nr_running
))
1563 * We may dequeue prev's rt_rq in put_prev_task().
1564 * So, we update time before rt_queued check.
1566 if (prev
->sched_class
== &rt_sched_class
)
1569 if (!rt_rq
->rt_queued
)
1572 put_prev_task(rq
, prev
);
1574 p
= _pick_next_task_rt(rq
);
1576 /* The running task is never eligible for pushing */
1577 dequeue_pushable_task(rq
, p
);
1579 rt_queue_push_tasks(rq
);
1582 * If prev task was rt, put_prev_task() has already updated the
1583 * utilization. We only care of the case where we start to schedule a
1586 if (rq
->curr
->sched_class
!= &rt_sched_class
)
1587 update_rt_rq_load_avg(rq_clock_task(rq
), rq
, 0);
1592 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
1596 update_rt_rq_load_avg(rq_clock_task(rq
), rq
, 1);
1599 * The previous task needs to be made eligible for pushing
1600 * if it is still active
1602 if (on_rt_rq(&p
->rt
) && p
->nr_cpus_allowed
> 1)
1603 enqueue_pushable_task(rq
, p
);
1608 /* Only try algorithms three times */
1609 #define RT_MAX_TRIES 3
1611 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
1613 if (!task_running(rq
, p
) &&
1614 cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
1621 * Return the highest pushable rq's task, which is suitable to be executed
1622 * on the CPU, NULL otherwise
1624 static struct task_struct
*pick_highest_pushable_task(struct rq
*rq
, int cpu
)
1626 struct plist_head
*head
= &rq
->rt
.pushable_tasks
;
1627 struct task_struct
*p
;
1629 if (!has_pushable_tasks(rq
))
1632 plist_for_each_entry(p
, head
, pushable_tasks
) {
1633 if (pick_rt_task(rq
, p
, cpu
))
1640 static DEFINE_PER_CPU(cpumask_var_t
, local_cpu_mask
);
1642 static int find_lowest_rq(struct task_struct
*task
)
1644 struct sched_domain
*sd
;
1645 struct cpumask
*lowest_mask
= this_cpu_cpumask_var_ptr(local_cpu_mask
);
1646 int this_cpu
= smp_processor_id();
1647 int cpu
= task_cpu(task
);
1649 /* Make sure the mask is initialized first */
1650 if (unlikely(!lowest_mask
))
1653 if (task
->nr_cpus_allowed
== 1)
1654 return -1; /* No other targets possible */
1656 if (!cpupri_find(&task_rq(task
)->rd
->cpupri
, task
, lowest_mask
))
1657 return -1; /* No targets found */
1660 * At this point we have built a mask of CPUs representing the
1661 * lowest priority tasks in the system. Now we want to elect
1662 * the best one based on our affinity and topology.
1664 * We prioritize the last CPU that the task executed on since
1665 * it is most likely cache-hot in that location.
1667 if (cpumask_test_cpu(cpu
, lowest_mask
))
1671 * Otherwise, we consult the sched_domains span maps to figure
1672 * out which CPU is logically closest to our hot cache data.
1674 if (!cpumask_test_cpu(this_cpu
, lowest_mask
))
1675 this_cpu
= -1; /* Skip this_cpu opt if not among lowest */
1678 for_each_domain(cpu
, sd
) {
1679 if (sd
->flags
& SD_WAKE_AFFINE
) {
1683 * "this_cpu" is cheaper to preempt than a
1686 if (this_cpu
!= -1 &&
1687 cpumask_test_cpu(this_cpu
, sched_domain_span(sd
))) {
1692 best_cpu
= cpumask_first_and(lowest_mask
,
1693 sched_domain_span(sd
));
1694 if (best_cpu
< nr_cpu_ids
) {
1703 * And finally, if there were no matches within the domains
1704 * just give the caller *something* to work with from the compatible
1710 cpu
= cpumask_any(lowest_mask
);
1711 if (cpu
< nr_cpu_ids
)
1717 /* Will lock the rq it finds */
1718 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
1720 struct rq
*lowest_rq
= NULL
;
1724 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
1725 cpu
= find_lowest_rq(task
);
1727 if ((cpu
== -1) || (cpu
== rq
->cpu
))
1730 lowest_rq
= cpu_rq(cpu
);
1732 if (lowest_rq
->rt
.highest_prio
.curr
<= task
->prio
) {
1734 * Target rq has tasks of equal or higher priority,
1735 * retrying does not release any lock and is unlikely
1736 * to yield a different result.
1742 /* if the prio of this runqueue changed, try again */
1743 if (double_lock_balance(rq
, lowest_rq
)) {
1745 * We had to unlock the run queue. In
1746 * the mean time, task could have
1747 * migrated already or had its affinity changed.
1748 * Also make sure that it wasn't scheduled on its rq.
1750 if (unlikely(task_rq(task
) != rq
||
1751 !cpumask_test_cpu(lowest_rq
->cpu
, &task
->cpus_allowed
) ||
1752 task_running(rq
, task
) ||
1754 !task_on_rq_queued(task
))) {
1756 double_unlock_balance(rq
, lowest_rq
);
1762 /* If this rq is still suitable use it. */
1763 if (lowest_rq
->rt
.highest_prio
.curr
> task
->prio
)
1767 double_unlock_balance(rq
, lowest_rq
);
1774 static struct task_struct
*pick_next_pushable_task(struct rq
*rq
)
1776 struct task_struct
*p
;
1778 if (!has_pushable_tasks(rq
))
1781 p
= plist_first_entry(&rq
->rt
.pushable_tasks
,
1782 struct task_struct
, pushable_tasks
);
1784 BUG_ON(rq
->cpu
!= task_cpu(p
));
1785 BUG_ON(task_current(rq
, p
));
1786 BUG_ON(p
->nr_cpus_allowed
<= 1);
1788 BUG_ON(!task_on_rq_queued(p
));
1789 BUG_ON(!rt_task(p
));
1795 * If the current CPU has more than one RT task, see if the non
1796 * running task can migrate over to a CPU that is running a task
1797 * of lesser priority.
1799 static int push_rt_task(struct rq
*rq
)
1801 struct task_struct
*next_task
;
1802 struct rq
*lowest_rq
;
1805 if (!rq
->rt
.overloaded
)
1808 next_task
= pick_next_pushable_task(rq
);
1813 if (unlikely(next_task
== rq
->curr
)) {
1819 * It's possible that the next_task slipped in of
1820 * higher priority than current. If that's the case
1821 * just reschedule current.
1823 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
1828 /* We might release rq lock */
1829 get_task_struct(next_task
);
1831 /* find_lock_lowest_rq locks the rq if found */
1832 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
1834 struct task_struct
*task
;
1836 * find_lock_lowest_rq releases rq->lock
1837 * so it is possible that next_task has migrated.
1839 * We need to make sure that the task is still on the same
1840 * run-queue and is also still the next task eligible for
1843 task
= pick_next_pushable_task(rq
);
1844 if (task
== next_task
) {
1846 * The task hasn't migrated, and is still the next
1847 * eligible task, but we failed to find a run-queue
1848 * to push it to. Do not retry in this case, since
1849 * other CPUs will pull from us when ready.
1855 /* No more tasks, just exit */
1859 * Something has shifted, try again.
1861 put_task_struct(next_task
);
1866 deactivate_task(rq
, next_task
, 0);
1867 set_task_cpu(next_task
, lowest_rq
->cpu
);
1868 activate_task(lowest_rq
, next_task
, 0);
1871 resched_curr(lowest_rq
);
1873 double_unlock_balance(rq
, lowest_rq
);
1876 put_task_struct(next_task
);
1881 static void push_rt_tasks(struct rq
*rq
)
1883 /* push_rt_task will return true if it moved an RT */
1884 while (push_rt_task(rq
))
1888 #ifdef HAVE_RT_PUSH_IPI
1891 * When a high priority task schedules out from a CPU and a lower priority
1892 * task is scheduled in, a check is made to see if there's any RT tasks
1893 * on other CPUs that are waiting to run because a higher priority RT task
1894 * is currently running on its CPU. In this case, the CPU with multiple RT
1895 * tasks queued on it (overloaded) needs to be notified that a CPU has opened
1896 * up that may be able to run one of its non-running queued RT tasks.
1898 * All CPUs with overloaded RT tasks need to be notified as there is currently
1899 * no way to know which of these CPUs have the highest priority task waiting
1900 * to run. Instead of trying to take a spinlock on each of these CPUs,
1901 * which has shown to cause large latency when done on machines with many
1902 * CPUs, sending an IPI to the CPUs to have them push off the overloaded
1903 * RT tasks waiting to run.
1905 * Just sending an IPI to each of the CPUs is also an issue, as on large
1906 * count CPU machines, this can cause an IPI storm on a CPU, especially
1907 * if its the only CPU with multiple RT tasks queued, and a large number
1908 * of CPUs scheduling a lower priority task at the same time.
1910 * Each root domain has its own irq work function that can iterate over
1911 * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT
1912 * tassk must be checked if there's one or many CPUs that are lowering
1913 * their priority, there's a single irq work iterator that will try to
1914 * push off RT tasks that are waiting to run.
1916 * When a CPU schedules a lower priority task, it will kick off the
1917 * irq work iterator that will jump to each CPU with overloaded RT tasks.
1918 * As it only takes the first CPU that schedules a lower priority task
1919 * to start the process, the rto_start variable is incremented and if
1920 * the atomic result is one, then that CPU will try to take the rto_lock.
1921 * This prevents high contention on the lock as the process handles all
1922 * CPUs scheduling lower priority tasks.
1924 * All CPUs that are scheduling a lower priority task will increment the
1925 * rt_loop_next variable. This will make sure that the irq work iterator
1926 * checks all RT overloaded CPUs whenever a CPU schedules a new lower
1927 * priority task, even if the iterator is in the middle of a scan. Incrementing
1928 * the rt_loop_next will cause the iterator to perform another scan.
1931 static int rto_next_cpu(struct root_domain
*rd
)
1937 * When starting the IPI RT pushing, the rto_cpu is set to -1,
1938 * rt_next_cpu() will simply return the first CPU found in
1941 * If rto_next_cpu() is called with rto_cpu is a valid CPU, it
1942 * will return the next CPU found in the rto_mask.
1944 * If there are no more CPUs left in the rto_mask, then a check is made
1945 * against rto_loop and rto_loop_next. rto_loop is only updated with
1946 * the rto_lock held, but any CPU may increment the rto_loop_next
1947 * without any locking.
1951 /* When rto_cpu is -1 this acts like cpumask_first() */
1952 cpu
= cpumask_next(rd
->rto_cpu
, rd
->rto_mask
);
1956 if (cpu
< nr_cpu_ids
)
1962 * ACQUIRE ensures we see the @rto_mask changes
1963 * made prior to the @next value observed.
1965 * Matches WMB in rt_set_overload().
1967 next
= atomic_read_acquire(&rd
->rto_loop_next
);
1969 if (rd
->rto_loop
== next
)
1972 rd
->rto_loop
= next
;
1978 static inline bool rto_start_trylock(atomic_t
*v
)
1980 return !atomic_cmpxchg_acquire(v
, 0, 1);
1983 static inline void rto_start_unlock(atomic_t
*v
)
1985 atomic_set_release(v
, 0);
1988 static void tell_cpu_to_push(struct rq
*rq
)
1992 /* Keep the loop going if the IPI is currently active */
1993 atomic_inc(&rq
->rd
->rto_loop_next
);
1995 /* Only one CPU can initiate a loop at a time */
1996 if (!rto_start_trylock(&rq
->rd
->rto_loop_start
))
1999 raw_spin_lock(&rq
->rd
->rto_lock
);
2002 * The rto_cpu is updated under the lock, if it has a valid CPU
2003 * then the IPI is still running and will continue due to the
2004 * update to loop_next, and nothing needs to be done here.
2005 * Otherwise it is finishing up and an ipi needs to be sent.
2007 if (rq
->rd
->rto_cpu
< 0)
2008 cpu
= rto_next_cpu(rq
->rd
);
2010 raw_spin_unlock(&rq
->rd
->rto_lock
);
2012 rto_start_unlock(&rq
->rd
->rto_loop_start
);
2015 /* Make sure the rd does not get freed while pushing */
2016 sched_get_rd(rq
->rd
);
2017 irq_work_queue_on(&rq
->rd
->rto_push_work
, cpu
);
2021 /* Called from hardirq context */
2022 void rto_push_irq_work_func(struct irq_work
*work
)
2024 struct root_domain
*rd
=
2025 container_of(work
, struct root_domain
, rto_push_work
);
2032 * We do not need to grab the lock to check for has_pushable_tasks.
2033 * When it gets updated, a check is made if a push is possible.
2035 if (has_pushable_tasks(rq
)) {
2036 raw_spin_lock(&rq
->lock
);
2038 raw_spin_unlock(&rq
->lock
);
2041 raw_spin_lock(&rd
->rto_lock
);
2043 /* Pass the IPI to the next rt overloaded queue */
2044 cpu
= rto_next_cpu(rd
);
2046 raw_spin_unlock(&rd
->rto_lock
);
2053 /* Try the next RT overloaded CPU */
2054 irq_work_queue_on(&rd
->rto_push_work
, cpu
);
2056 #endif /* HAVE_RT_PUSH_IPI */
2058 static void pull_rt_task(struct rq
*this_rq
)
2060 int this_cpu
= this_rq
->cpu
, cpu
;
2061 bool resched
= false;
2062 struct task_struct
*p
;
2064 int rt_overload_count
= rt_overloaded(this_rq
);
2066 if (likely(!rt_overload_count
))
2070 * Match the barrier from rt_set_overloaded; this guarantees that if we
2071 * see overloaded we must also see the rto_mask bit.
2075 /* If we are the only overloaded CPU do nothing */
2076 if (rt_overload_count
== 1 &&
2077 cpumask_test_cpu(this_rq
->cpu
, this_rq
->rd
->rto_mask
))
2080 #ifdef HAVE_RT_PUSH_IPI
2081 if (sched_feat(RT_PUSH_IPI
)) {
2082 tell_cpu_to_push(this_rq
);
2087 for_each_cpu(cpu
, this_rq
->rd
->rto_mask
) {
2088 if (this_cpu
== cpu
)
2091 src_rq
= cpu_rq(cpu
);
2094 * Don't bother taking the src_rq->lock if the next highest
2095 * task is known to be lower-priority than our current task.
2096 * This may look racy, but if this value is about to go
2097 * logically higher, the src_rq will push this task away.
2098 * And if its going logically lower, we do not care
2100 if (src_rq
->rt
.highest_prio
.next
>=
2101 this_rq
->rt
.highest_prio
.curr
)
2105 * We can potentially drop this_rq's lock in
2106 * double_lock_balance, and another CPU could
2109 double_lock_balance(this_rq
, src_rq
);
2112 * We can pull only a task, which is pushable
2113 * on its rq, and no others.
2115 p
= pick_highest_pushable_task(src_rq
, this_cpu
);
2118 * Do we have an RT task that preempts
2119 * the to-be-scheduled task?
2121 if (p
&& (p
->prio
< this_rq
->rt
.highest_prio
.curr
)) {
2122 WARN_ON(p
== src_rq
->curr
);
2123 WARN_ON(!task_on_rq_queued(p
));
2126 * There's a chance that p is higher in priority
2127 * than what's currently running on its CPU.
2128 * This is just that p is wakeing up and hasn't
2129 * had a chance to schedule. We only pull
2130 * p if it is lower in priority than the
2131 * current task on the run queue
2133 if (p
->prio
< src_rq
->curr
->prio
)
2138 deactivate_task(src_rq
, p
, 0);
2139 set_task_cpu(p
, this_cpu
);
2140 activate_task(this_rq
, p
, 0);
2142 * We continue with the search, just in
2143 * case there's an even higher prio task
2144 * in another runqueue. (low likelihood
2149 double_unlock_balance(this_rq
, src_rq
);
2153 resched_curr(this_rq
);
2157 * If we are not running and we are not going to reschedule soon, we should
2158 * try to push tasks away now
2160 static void task_woken_rt(struct rq
*rq
, struct task_struct
*p
)
2162 if (!task_running(rq
, p
) &&
2163 !test_tsk_need_resched(rq
->curr
) &&
2164 p
->nr_cpus_allowed
> 1 &&
2165 (dl_task(rq
->curr
) || rt_task(rq
->curr
)) &&
2166 (rq
->curr
->nr_cpus_allowed
< 2 ||
2167 rq
->curr
->prio
<= p
->prio
))
2171 /* Assumes rq->lock is held */
2172 static void rq_online_rt(struct rq
*rq
)
2174 if (rq
->rt
.overloaded
)
2175 rt_set_overload(rq
);
2177 __enable_runtime(rq
);
2179 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rq
->rt
.highest_prio
.curr
);
2182 /* Assumes rq->lock is held */
2183 static void rq_offline_rt(struct rq
*rq
)
2185 if (rq
->rt
.overloaded
)
2186 rt_clear_overload(rq
);
2188 __disable_runtime(rq
);
2190 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, CPUPRI_INVALID
);
2194 * When switch from the rt queue, we bring ourselves to a position
2195 * that we might want to pull RT tasks from other runqueues.
2197 static void switched_from_rt(struct rq
*rq
, struct task_struct
*p
)
2200 * If there are other RT tasks then we will reschedule
2201 * and the scheduling of the other RT tasks will handle
2202 * the balancing. But if we are the last RT task
2203 * we may need to handle the pulling of RT tasks
2206 if (!task_on_rq_queued(p
) || rq
->rt
.rt_nr_running
)
2209 rt_queue_pull_task(rq
);
2212 void __init
init_sched_rt_class(void)
2216 for_each_possible_cpu(i
) {
2217 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask
, i
),
2218 GFP_KERNEL
, cpu_to_node(i
));
2221 #endif /* CONFIG_SMP */
2224 * When switching a task to RT, we may overload the runqueue
2225 * with RT tasks. In this case we try to push them off to
2228 static void switched_to_rt(struct rq
*rq
, struct task_struct
*p
)
2231 * If we are already running, then there's nothing
2232 * that needs to be done. But if we are not running
2233 * we may need to preempt the current running task.
2234 * If that current running task is also an RT task
2235 * then see if we can move to another run queue.
2237 if (task_on_rq_queued(p
) && rq
->curr
!= p
) {
2239 if (p
->nr_cpus_allowed
> 1 && rq
->rt
.overloaded
)
2240 rt_queue_push_tasks(rq
);
2241 #endif /* CONFIG_SMP */
2242 if (p
->prio
< rq
->curr
->prio
&& cpu_online(cpu_of(rq
)))
2248 * Priority of the task has changed. This may cause
2249 * us to initiate a push or pull.
2252 prio_changed_rt(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
2254 if (!task_on_rq_queued(p
))
2257 if (rq
->curr
== p
) {
2260 * If our priority decreases while running, we
2261 * may need to pull tasks to this runqueue.
2263 if (oldprio
< p
->prio
)
2264 rt_queue_pull_task(rq
);
2267 * If there's a higher priority task waiting to run
2270 if (p
->prio
> rq
->rt
.highest_prio
.curr
)
2273 /* For UP simply resched on drop of prio */
2274 if (oldprio
< p
->prio
)
2276 #endif /* CONFIG_SMP */
2279 * This task is not running, but if it is
2280 * greater than the current running task
2283 if (p
->prio
< rq
->curr
->prio
)
2288 #ifdef CONFIG_POSIX_TIMERS
2289 static void watchdog(struct rq
*rq
, struct task_struct
*p
)
2291 unsigned long soft
, hard
;
2293 /* max may change after cur was read, this will be fixed next tick */
2294 soft
= task_rlimit(p
, RLIMIT_RTTIME
);
2295 hard
= task_rlimit_max(p
, RLIMIT_RTTIME
);
2297 if (soft
!= RLIM_INFINITY
) {
2300 if (p
->rt
.watchdog_stamp
!= jiffies
) {
2302 p
->rt
.watchdog_stamp
= jiffies
;
2305 next
= DIV_ROUND_UP(min(soft
, hard
), USEC_PER_SEC
/HZ
);
2306 if (p
->rt
.timeout
> next
)
2307 p
->cputime_expires
.sched_exp
= p
->se
.sum_exec_runtime
;
2311 static inline void watchdog(struct rq
*rq
, struct task_struct
*p
) { }
2315 * scheduler tick hitting a task of our scheduling class.
2317 * NOTE: This function can be called remotely by the tick offload that
2318 * goes along full dynticks. Therefore no local assumption can be made
2319 * and everything must be accessed through the @rq and @curr passed in
2322 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
, int queued
)
2324 struct sched_rt_entity
*rt_se
= &p
->rt
;
2327 update_rt_rq_load_avg(rq_clock_task(rq
), rq
, 1);
2332 * RR tasks need a special form of timeslice management.
2333 * FIFO tasks have no timeslices.
2335 if (p
->policy
!= SCHED_RR
)
2338 if (--p
->rt
.time_slice
)
2341 p
->rt
.time_slice
= sched_rr_timeslice
;
2344 * Requeue to the end of queue if we (and all of our ancestors) are not
2345 * the only element on the queue
2347 for_each_sched_rt_entity(rt_se
) {
2348 if (rt_se
->run_list
.prev
!= rt_se
->run_list
.next
) {
2349 requeue_task_rt(rq
, p
, 0);
2356 static void set_curr_task_rt(struct rq
*rq
)
2358 struct task_struct
*p
= rq
->curr
;
2360 p
->se
.exec_start
= rq_clock_task(rq
);
2362 /* The running task is never eligible for pushing */
2363 dequeue_pushable_task(rq
, p
);
2366 static unsigned int get_rr_interval_rt(struct rq
*rq
, struct task_struct
*task
)
2369 * Time slice is 0 for SCHED_FIFO tasks
2371 if (task
->policy
== SCHED_RR
)
2372 return sched_rr_timeslice
;
2377 const struct sched_class rt_sched_class
= {
2378 .next
= &fair_sched_class
,
2379 .enqueue_task
= enqueue_task_rt
,
2380 .dequeue_task
= dequeue_task_rt
,
2381 .yield_task
= yield_task_rt
,
2383 .check_preempt_curr
= check_preempt_curr_rt
,
2385 .pick_next_task
= pick_next_task_rt
,
2386 .put_prev_task
= put_prev_task_rt
,
2389 .select_task_rq
= select_task_rq_rt
,
2391 .set_cpus_allowed
= set_cpus_allowed_common
,
2392 .rq_online
= rq_online_rt
,
2393 .rq_offline
= rq_offline_rt
,
2394 .task_woken
= task_woken_rt
,
2395 .switched_from
= switched_from_rt
,
2398 .set_curr_task
= set_curr_task_rt
,
2399 .task_tick
= task_tick_rt
,
2401 .get_rr_interval
= get_rr_interval_rt
,
2403 .prio_changed
= prio_changed_rt
,
2404 .switched_to
= switched_to_rt
,
2406 .update_curr
= update_curr_rt
,
2409 #ifdef CONFIG_RT_GROUP_SCHED
2411 * Ensure that the real time constraints are schedulable.
2413 static DEFINE_MUTEX(rt_constraints_mutex
);
2415 /* Must be called with tasklist_lock held */
2416 static inline int tg_has_rt_tasks(struct task_group
*tg
)
2418 struct task_struct
*g
, *p
;
2421 * Autogroups do not have RT tasks; see autogroup_create().
2423 if (task_group_is_autogroup(tg
))
2426 for_each_process_thread(g
, p
) {
2427 if (rt_task(p
) && task_group(p
) == tg
)
2434 struct rt_schedulable_data
{
2435 struct task_group
*tg
;
2440 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
2442 struct rt_schedulable_data
*d
= data
;
2443 struct task_group
*child
;
2444 unsigned long total
, sum
= 0;
2445 u64 period
, runtime
;
2447 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2448 runtime
= tg
->rt_bandwidth
.rt_runtime
;
2451 period
= d
->rt_period
;
2452 runtime
= d
->rt_runtime
;
2456 * Cannot have more runtime than the period.
2458 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
2462 * Ensure we don't starve existing RT tasks.
2464 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
2467 total
= to_ratio(period
, runtime
);
2470 * Nobody can have more than the global setting allows.
2472 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
2476 * The sum of our children's runtime should not exceed our own.
2478 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
2479 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
2480 runtime
= child
->rt_bandwidth
.rt_runtime
;
2482 if (child
== d
->tg
) {
2483 period
= d
->rt_period
;
2484 runtime
= d
->rt_runtime
;
2487 sum
+= to_ratio(period
, runtime
);
2496 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
2500 struct rt_schedulable_data data
= {
2502 .rt_period
= period
,
2503 .rt_runtime
= runtime
,
2507 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
2513 static int tg_set_rt_bandwidth(struct task_group
*tg
,
2514 u64 rt_period
, u64 rt_runtime
)
2519 * Disallowing the root group RT runtime is BAD, it would disallow the
2520 * kernel creating (and or operating) RT threads.
2522 if (tg
== &root_task_group
&& rt_runtime
== 0)
2525 /* No period doesn't make any sense. */
2529 mutex_lock(&rt_constraints_mutex
);
2530 read_lock(&tasklist_lock
);
2531 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
2535 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
2536 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
2537 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
2539 for_each_possible_cpu(i
) {
2540 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
2542 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
2543 rt_rq
->rt_runtime
= rt_runtime
;
2544 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
2546 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
2548 read_unlock(&tasklist_lock
);
2549 mutex_unlock(&rt_constraints_mutex
);
2554 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
2556 u64 rt_runtime
, rt_period
;
2558 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2559 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
2560 if (rt_runtime_us
< 0)
2561 rt_runtime
= RUNTIME_INF
;
2563 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
2566 long sched_group_rt_runtime(struct task_group
*tg
)
2570 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
2573 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
2574 do_div(rt_runtime_us
, NSEC_PER_USEC
);
2575 return rt_runtime_us
;
2578 int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
2580 u64 rt_runtime
, rt_period
;
2582 rt_period
= rt_period_us
* NSEC_PER_USEC
;
2583 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
2585 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
2588 long sched_group_rt_period(struct task_group
*tg
)
2592 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2593 do_div(rt_period_us
, NSEC_PER_USEC
);
2594 return rt_period_us
;
2597 static int sched_rt_global_constraints(void)
2601 mutex_lock(&rt_constraints_mutex
);
2602 read_lock(&tasklist_lock
);
2603 ret
= __rt_schedulable(NULL
, 0, 0);
2604 read_unlock(&tasklist_lock
);
2605 mutex_unlock(&rt_constraints_mutex
);
2610 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
2612 /* Don't accept realtime tasks when there is no way for them to run */
2613 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
2619 #else /* !CONFIG_RT_GROUP_SCHED */
2620 static int sched_rt_global_constraints(void)
2622 unsigned long flags
;
2625 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2626 for_each_possible_cpu(i
) {
2627 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
2629 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
2630 rt_rq
->rt_runtime
= global_rt_runtime();
2631 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
2633 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2637 #endif /* CONFIG_RT_GROUP_SCHED */
2639 static int sched_rt_global_validate(void)
2641 if (sysctl_sched_rt_period
<= 0)
2644 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
2645 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
2651 static void sched_rt_do_global(void)
2653 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
2654 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
2657 int sched_rt_handler(struct ctl_table
*table
, int write
,
2658 void __user
*buffer
, size_t *lenp
,
2661 int old_period
, old_runtime
;
2662 static DEFINE_MUTEX(mutex
);
2666 old_period
= sysctl_sched_rt_period
;
2667 old_runtime
= sysctl_sched_rt_runtime
;
2669 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
2671 if (!ret
&& write
) {
2672 ret
= sched_rt_global_validate();
2676 ret
= sched_dl_global_validate();
2680 ret
= sched_rt_global_constraints();
2684 sched_rt_do_global();
2685 sched_dl_do_global();
2689 sysctl_sched_rt_period
= old_period
;
2690 sysctl_sched_rt_runtime
= old_runtime
;
2692 mutex_unlock(&mutex
);
2697 int sched_rr_handler(struct ctl_table
*table
, int write
,
2698 void __user
*buffer
, size_t *lenp
,
2702 static DEFINE_MUTEX(mutex
);
2705 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
2707 * Make sure that internally we keep jiffies.
2708 * Also, writing zero resets the timeslice to default:
2710 if (!ret
&& write
) {
2711 sched_rr_timeslice
=
2712 sysctl_sched_rr_timeslice
<= 0 ? RR_TIMESLICE
:
2713 msecs_to_jiffies(sysctl_sched_rr_timeslice
);
2715 mutex_unlock(&mutex
);
2720 #ifdef CONFIG_SCHED_DEBUG
2721 void print_rt_stats(struct seq_file
*m
, int cpu
)
2724 struct rt_rq
*rt_rq
;
2727 for_each_rt_rq(rt_rq
, iter
, cpu_rq(cpu
))
2728 print_rt_rq(m
, cpu
, rt_rq
);
2731 #endif /* CONFIG_SCHED_DEBUG */