1 // SPDX-License-Identifier: GPL-2.0
3 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
9 #include <linux/slab.h>
10 #include <linux/irq_work.h>
15 int sched_rr_timeslice
= RR_TIMESLICE
;
16 int sysctl_sched_rr_timeslice
= (MSEC_PER_SEC
/ HZ
) * RR_TIMESLICE
;
18 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
20 struct rt_bandwidth def_rt_bandwidth
;
22 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
24 struct rt_bandwidth
*rt_b
=
25 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
29 raw_spin_lock(&rt_b
->rt_runtime_lock
);
31 overrun
= hrtimer_forward_now(timer
, rt_b
->rt_period
);
35 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
36 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
37 raw_spin_lock(&rt_b
->rt_runtime_lock
);
40 rt_b
->rt_period_active
= 0;
41 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
43 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
46 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
48 rt_b
->rt_period
= ns_to_ktime(period
);
49 rt_b
->rt_runtime
= runtime
;
51 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
53 hrtimer_init(&rt_b
->rt_period_timer
,
54 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
55 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
58 static inline void do_start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
60 raw_spin_lock(&rt_b
->rt_runtime_lock
);
61 if (!rt_b
->rt_period_active
) {
62 rt_b
->rt_period_active
= 1;
64 * SCHED_DEADLINE updates the bandwidth, as a run away
65 * RT task with a DL task could hog a CPU. But DL does
66 * not reset the period. If a deadline task was running
67 * without an RT task running, it can cause RT tasks to
68 * throttle when they start up. Kick the timer right away
69 * to update the period.
71 hrtimer_forward_now(&rt_b
->rt_period_timer
, ns_to_ktime(0));
72 hrtimer_start_expires(&rt_b
->rt_period_timer
, HRTIMER_MODE_ABS_PINNED
);
74 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
77 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
79 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
82 do_start_rt_bandwidth(rt_b
);
85 void init_rt_rq(struct rt_rq
*rt_rq
)
87 struct rt_prio_array
*array
;
90 array
= &rt_rq
->active
;
91 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
92 INIT_LIST_HEAD(array
->queue
+ i
);
93 __clear_bit(i
, array
->bitmap
);
95 /* delimiter for bitsearch: */
96 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
98 #if defined CONFIG_SMP
99 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
100 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
101 rt_rq
->rt_nr_migratory
= 0;
102 rt_rq
->overloaded
= 0;
103 plist_head_init(&rt_rq
->pushable_tasks
);
104 #endif /* CONFIG_SMP */
105 /* We start is dequeued state, because no RT tasks are queued */
106 rt_rq
->rt_queued
= 0;
109 rt_rq
->rt_throttled
= 0;
110 rt_rq
->rt_runtime
= 0;
111 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
114 #ifdef CONFIG_RT_GROUP_SCHED
115 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
117 hrtimer_cancel(&rt_b
->rt_period_timer
);
120 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
122 static inline struct task_struct
*rt_task_of(struct sched_rt_entity
*rt_se
)
124 #ifdef CONFIG_SCHED_DEBUG
125 WARN_ON_ONCE(!rt_entity_is_task(rt_se
));
127 return container_of(rt_se
, struct task_struct
, rt
);
130 static inline struct rq
*rq_of_rt_rq(struct rt_rq
*rt_rq
)
135 static inline struct rt_rq
*rt_rq_of_se(struct sched_rt_entity
*rt_se
)
140 static inline struct rq
*rq_of_rt_se(struct sched_rt_entity
*rt_se
)
142 struct rt_rq
*rt_rq
= rt_se
->rt_rq
;
147 void free_rt_sched_group(struct task_group
*tg
)
152 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
154 for_each_possible_cpu(i
) {
165 void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
166 struct sched_rt_entity
*rt_se
, int cpu
,
167 struct sched_rt_entity
*parent
)
169 struct rq
*rq
= cpu_rq(cpu
);
171 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
172 rt_rq
->rt_nr_boosted
= 0;
176 tg
->rt_rq
[cpu
] = rt_rq
;
177 tg
->rt_se
[cpu
] = rt_se
;
183 rt_se
->rt_rq
= &rq
->rt
;
185 rt_se
->rt_rq
= parent
->my_q
;
188 rt_se
->parent
= parent
;
189 INIT_LIST_HEAD(&rt_se
->run_list
);
192 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
195 struct sched_rt_entity
*rt_se
;
198 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
201 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
205 init_rt_bandwidth(&tg
->rt_bandwidth
,
206 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
208 for_each_possible_cpu(i
) {
209 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
210 GFP_KERNEL
, cpu_to_node(i
));
214 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
215 GFP_KERNEL
, cpu_to_node(i
));
220 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
221 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, parent
->rt_se
[i
]);
232 #else /* CONFIG_RT_GROUP_SCHED */
234 #define rt_entity_is_task(rt_se) (1)
236 static inline struct task_struct
*rt_task_of(struct sched_rt_entity
*rt_se
)
238 return container_of(rt_se
, struct task_struct
, rt
);
241 static inline struct rq
*rq_of_rt_rq(struct rt_rq
*rt_rq
)
243 return container_of(rt_rq
, struct rq
, rt
);
246 static inline struct rq
*rq_of_rt_se(struct sched_rt_entity
*rt_se
)
248 struct task_struct
*p
= rt_task_of(rt_se
);
253 static inline struct rt_rq
*rt_rq_of_se(struct sched_rt_entity
*rt_se
)
255 struct rq
*rq
= rq_of_rt_se(rt_se
);
260 void free_rt_sched_group(struct task_group
*tg
) { }
262 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
266 #endif /* CONFIG_RT_GROUP_SCHED */
270 static void pull_rt_task(struct rq
*this_rq
);
272 static inline bool need_pull_rt_task(struct rq
*rq
, struct task_struct
*prev
)
274 /* Try to pull RT tasks here if we lower this rq's prio */
275 return rq
->rt
.highest_prio
.curr
> prev
->prio
;
278 static inline int rt_overloaded(struct rq
*rq
)
280 return atomic_read(&rq
->rd
->rto_count
);
283 static inline void rt_set_overload(struct rq
*rq
)
288 cpumask_set_cpu(rq
->cpu
, rq
->rd
->rto_mask
);
290 * Make sure the mask is visible before we set
291 * the overload count. That is checked to determine
292 * if we should look at the mask. It would be a shame
293 * if we looked at the mask, but the mask was not
296 * Matched by the barrier in pull_rt_task().
299 atomic_inc(&rq
->rd
->rto_count
);
302 static inline void rt_clear_overload(struct rq
*rq
)
307 /* the order here really doesn't matter */
308 atomic_dec(&rq
->rd
->rto_count
);
309 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->rto_mask
);
312 static void update_rt_migration(struct rt_rq
*rt_rq
)
314 if (rt_rq
->rt_nr_migratory
&& rt_rq
->rt_nr_total
> 1) {
315 if (!rt_rq
->overloaded
) {
316 rt_set_overload(rq_of_rt_rq(rt_rq
));
317 rt_rq
->overloaded
= 1;
319 } else if (rt_rq
->overloaded
) {
320 rt_clear_overload(rq_of_rt_rq(rt_rq
));
321 rt_rq
->overloaded
= 0;
325 static void inc_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
327 struct task_struct
*p
;
329 if (!rt_entity_is_task(rt_se
))
332 p
= rt_task_of(rt_se
);
333 rt_rq
= &rq_of_rt_rq(rt_rq
)->rt
;
335 rt_rq
->rt_nr_total
++;
336 if (p
->nr_cpus_allowed
> 1)
337 rt_rq
->rt_nr_migratory
++;
339 update_rt_migration(rt_rq
);
342 static void dec_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
344 struct task_struct
*p
;
346 if (!rt_entity_is_task(rt_se
))
349 p
= rt_task_of(rt_se
);
350 rt_rq
= &rq_of_rt_rq(rt_rq
)->rt
;
352 rt_rq
->rt_nr_total
--;
353 if (p
->nr_cpus_allowed
> 1)
354 rt_rq
->rt_nr_migratory
--;
356 update_rt_migration(rt_rq
);
359 static inline int has_pushable_tasks(struct rq
*rq
)
361 return !plist_head_empty(&rq
->rt
.pushable_tasks
);
364 static DEFINE_PER_CPU(struct callback_head
, rt_push_head
);
365 static DEFINE_PER_CPU(struct callback_head
, rt_pull_head
);
367 static void push_rt_tasks(struct rq
*);
368 static void pull_rt_task(struct rq
*);
370 static inline void queue_push_tasks(struct rq
*rq
)
372 if (!has_pushable_tasks(rq
))
375 queue_balance_callback(rq
, &per_cpu(rt_push_head
, rq
->cpu
), push_rt_tasks
);
378 static inline void queue_pull_task(struct rq
*rq
)
380 queue_balance_callback(rq
, &per_cpu(rt_pull_head
, rq
->cpu
), pull_rt_task
);
383 static void enqueue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
385 plist_del(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
386 plist_node_init(&p
->pushable_tasks
, p
->prio
);
387 plist_add(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
389 /* Update the highest prio pushable task */
390 if (p
->prio
< rq
->rt
.highest_prio
.next
)
391 rq
->rt
.highest_prio
.next
= p
->prio
;
394 static void dequeue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
396 plist_del(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
398 /* Update the new highest prio pushable task */
399 if (has_pushable_tasks(rq
)) {
400 p
= plist_first_entry(&rq
->rt
.pushable_tasks
,
401 struct task_struct
, pushable_tasks
);
402 rq
->rt
.highest_prio
.next
= p
->prio
;
404 rq
->rt
.highest_prio
.next
= MAX_RT_PRIO
;
409 static inline void enqueue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
413 static inline void dequeue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
418 void inc_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
423 void dec_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
427 static inline bool need_pull_rt_task(struct rq
*rq
, struct task_struct
*prev
)
432 static inline void pull_rt_task(struct rq
*this_rq
)
436 static inline void queue_push_tasks(struct rq
*rq
)
439 #endif /* CONFIG_SMP */
441 static void enqueue_top_rt_rq(struct rt_rq
*rt_rq
);
442 static void dequeue_top_rt_rq(struct rt_rq
*rt_rq
);
444 static inline int on_rt_rq(struct sched_rt_entity
*rt_se
)
449 #ifdef CONFIG_RT_GROUP_SCHED
451 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
456 return rt_rq
->rt_runtime
;
459 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
461 return ktime_to_ns(rt_rq
->tg
->rt_bandwidth
.rt_period
);
464 typedef struct task_group
*rt_rq_iter_t
;
466 static inline struct task_group
*next_task_group(struct task_group
*tg
)
469 tg
= list_entry_rcu(tg
->list
.next
,
470 typeof(struct task_group
), list
);
471 } while (&tg
->list
!= &task_groups
&& task_group_is_autogroup(tg
));
473 if (&tg
->list
== &task_groups
)
479 #define for_each_rt_rq(rt_rq, iter, rq) \
480 for (iter = container_of(&task_groups, typeof(*iter), list); \
481 (iter = next_task_group(iter)) && \
482 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
484 #define for_each_sched_rt_entity(rt_se) \
485 for (; rt_se; rt_se = rt_se->parent)
487 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
492 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
);
493 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
);
495 static void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
497 struct task_struct
*curr
= rq_of_rt_rq(rt_rq
)->curr
;
498 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
499 struct sched_rt_entity
*rt_se
;
501 int cpu
= cpu_of(rq
);
503 rt_se
= rt_rq
->tg
->rt_se
[cpu
];
505 if (rt_rq
->rt_nr_running
) {
507 enqueue_top_rt_rq(rt_rq
);
508 else if (!on_rt_rq(rt_se
))
509 enqueue_rt_entity(rt_se
, 0);
511 if (rt_rq
->highest_prio
.curr
< curr
->prio
)
516 static void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
518 struct sched_rt_entity
*rt_se
;
519 int cpu
= cpu_of(rq_of_rt_rq(rt_rq
));
521 rt_se
= rt_rq
->tg
->rt_se
[cpu
];
524 dequeue_top_rt_rq(rt_rq
);
525 else if (on_rt_rq(rt_se
))
526 dequeue_rt_entity(rt_se
, 0);
529 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
531 return rt_rq
->rt_throttled
&& !rt_rq
->rt_nr_boosted
;
534 static int rt_se_boosted(struct sched_rt_entity
*rt_se
)
536 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
537 struct task_struct
*p
;
540 return !!rt_rq
->rt_nr_boosted
;
542 p
= rt_task_of(rt_se
);
543 return p
->prio
!= p
->normal_prio
;
547 static inline const struct cpumask
*sched_rt_period_mask(void)
549 return this_rq()->rd
->span
;
552 static inline const struct cpumask
*sched_rt_period_mask(void)
554 return cpu_online_mask
;
559 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
561 return container_of(rt_b
, struct task_group
, rt_bandwidth
)->rt_rq
[cpu
];
564 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
566 return &rt_rq
->tg
->rt_bandwidth
;
569 #else /* !CONFIG_RT_GROUP_SCHED */
571 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
573 return rt_rq
->rt_runtime
;
576 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
578 return ktime_to_ns(def_rt_bandwidth
.rt_period
);
581 typedef struct rt_rq
*rt_rq_iter_t
;
583 #define for_each_rt_rq(rt_rq, iter, rq) \
584 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
586 #define for_each_sched_rt_entity(rt_se) \
587 for (; rt_se; rt_se = NULL)
589 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
594 static inline void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
596 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
598 if (!rt_rq
->rt_nr_running
)
601 enqueue_top_rt_rq(rt_rq
);
605 static inline void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
607 dequeue_top_rt_rq(rt_rq
);
610 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
612 return rt_rq
->rt_throttled
;
615 static inline const struct cpumask
*sched_rt_period_mask(void)
617 return cpu_online_mask
;
621 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
623 return &cpu_rq(cpu
)->rt
;
626 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
628 return &def_rt_bandwidth
;
631 #endif /* CONFIG_RT_GROUP_SCHED */
633 bool sched_rt_bandwidth_account(struct rt_rq
*rt_rq
)
635 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
637 return (hrtimer_active(&rt_b
->rt_period_timer
) ||
638 rt_rq
->rt_time
< rt_b
->rt_runtime
);
643 * We ran out of runtime, see if we can borrow some from our neighbours.
645 static void do_balance_runtime(struct rt_rq
*rt_rq
)
647 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
648 struct root_domain
*rd
= rq_of_rt_rq(rt_rq
)->rd
;
652 weight
= cpumask_weight(rd
->span
);
654 raw_spin_lock(&rt_b
->rt_runtime_lock
);
655 rt_period
= ktime_to_ns(rt_b
->rt_period
);
656 for_each_cpu(i
, rd
->span
) {
657 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
663 raw_spin_lock(&iter
->rt_runtime_lock
);
665 * Either all rqs have inf runtime and there's nothing to steal
666 * or __disable_runtime() below sets a specific rq to inf to
667 * indicate its been disabled and disalow stealing.
669 if (iter
->rt_runtime
== RUNTIME_INF
)
673 * From runqueues with spare time, take 1/n part of their
674 * spare time, but no more than our period.
676 diff
= iter
->rt_runtime
- iter
->rt_time
;
678 diff
= div_u64((u64
)diff
, weight
);
679 if (rt_rq
->rt_runtime
+ diff
> rt_period
)
680 diff
= rt_period
- rt_rq
->rt_runtime
;
681 iter
->rt_runtime
-= diff
;
682 rt_rq
->rt_runtime
+= diff
;
683 if (rt_rq
->rt_runtime
== rt_period
) {
684 raw_spin_unlock(&iter
->rt_runtime_lock
);
689 raw_spin_unlock(&iter
->rt_runtime_lock
);
691 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
695 * Ensure this RQ takes back all the runtime it lend to its neighbours.
697 static void __disable_runtime(struct rq
*rq
)
699 struct root_domain
*rd
= rq
->rd
;
703 if (unlikely(!scheduler_running
))
706 for_each_rt_rq(rt_rq
, iter
, rq
) {
707 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
711 raw_spin_lock(&rt_b
->rt_runtime_lock
);
712 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
714 * Either we're all inf and nobody needs to borrow, or we're
715 * already disabled and thus have nothing to do, or we have
716 * exactly the right amount of runtime to take out.
718 if (rt_rq
->rt_runtime
== RUNTIME_INF
||
719 rt_rq
->rt_runtime
== rt_b
->rt_runtime
)
721 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
724 * Calculate the difference between what we started out with
725 * and what we current have, that's the amount of runtime
726 * we lend and now have to reclaim.
728 want
= rt_b
->rt_runtime
- rt_rq
->rt_runtime
;
731 * Greedy reclaim, take back as much as we can.
733 for_each_cpu(i
, rd
->span
) {
734 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
738 * Can't reclaim from ourselves or disabled runqueues.
740 if (iter
== rt_rq
|| iter
->rt_runtime
== RUNTIME_INF
)
743 raw_spin_lock(&iter
->rt_runtime_lock
);
745 diff
= min_t(s64
, iter
->rt_runtime
, want
);
746 iter
->rt_runtime
-= diff
;
749 iter
->rt_runtime
-= want
;
752 raw_spin_unlock(&iter
->rt_runtime_lock
);
758 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
760 * We cannot be left wanting - that would mean some runtime
761 * leaked out of the system.
766 * Disable all the borrow logic by pretending we have inf
767 * runtime - in which case borrowing doesn't make sense.
769 rt_rq
->rt_runtime
= RUNTIME_INF
;
770 rt_rq
->rt_throttled
= 0;
771 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
772 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
774 /* Make rt_rq available for pick_next_task() */
775 sched_rt_rq_enqueue(rt_rq
);
779 static void __enable_runtime(struct rq
*rq
)
784 if (unlikely(!scheduler_running
))
788 * Reset each runqueue's bandwidth settings
790 for_each_rt_rq(rt_rq
, iter
, rq
) {
791 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
793 raw_spin_lock(&rt_b
->rt_runtime_lock
);
794 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
795 rt_rq
->rt_runtime
= rt_b
->rt_runtime
;
797 rt_rq
->rt_throttled
= 0;
798 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
799 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
803 static void balance_runtime(struct rt_rq
*rt_rq
)
805 if (!sched_feat(RT_RUNTIME_SHARE
))
808 if (rt_rq
->rt_time
> rt_rq
->rt_runtime
) {
809 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
810 do_balance_runtime(rt_rq
);
811 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
814 #else /* !CONFIG_SMP */
815 static inline void balance_runtime(struct rt_rq
*rt_rq
) {}
816 #endif /* CONFIG_SMP */
818 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
)
820 int i
, idle
= 1, throttled
= 0;
821 const struct cpumask
*span
;
823 span
= sched_rt_period_mask();
824 #ifdef CONFIG_RT_GROUP_SCHED
826 * FIXME: isolated CPUs should really leave the root task group,
827 * whether they are isolcpus or were isolated via cpusets, lest
828 * the timer run on a CPU which does not service all runqueues,
829 * potentially leaving other CPUs indefinitely throttled. If
830 * isolation is really required, the user will turn the throttle
831 * off to kill the perturbations it causes anyway. Meanwhile,
832 * this maintains functionality for boot and/or troubleshooting.
834 if (rt_b
== &root_task_group
.rt_bandwidth
)
835 span
= cpu_online_mask
;
837 for_each_cpu(i
, span
) {
839 struct rt_rq
*rt_rq
= sched_rt_period_rt_rq(rt_b
, i
);
840 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
844 * When span == cpu_online_mask, taking each rq->lock
845 * can be time-consuming. Try to avoid it when possible.
847 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
848 if (!sched_feat(RT_RUNTIME_SHARE
) && rt_rq
->rt_runtime
!= RUNTIME_INF
)
849 rt_rq
->rt_runtime
= rt_b
->rt_runtime
;
850 skip
= !rt_rq
->rt_time
&& !rt_rq
->rt_nr_running
;
851 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
855 raw_spin_lock(&rq
->lock
);
858 if (rt_rq
->rt_time
) {
861 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
862 if (rt_rq
->rt_throttled
)
863 balance_runtime(rt_rq
);
864 runtime
= rt_rq
->rt_runtime
;
865 rt_rq
->rt_time
-= min(rt_rq
->rt_time
, overrun
*runtime
);
866 if (rt_rq
->rt_throttled
&& rt_rq
->rt_time
< runtime
) {
867 rt_rq
->rt_throttled
= 0;
871 * When we're idle and a woken (rt) task is
872 * throttled check_preempt_curr() will set
873 * skip_update and the time between the wakeup
874 * and this unthrottle will get accounted as
877 if (rt_rq
->rt_nr_running
&& rq
->curr
== rq
->idle
)
878 rq_clock_skip_update(rq
, false);
880 if (rt_rq
->rt_time
|| rt_rq
->rt_nr_running
)
882 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
883 } else if (rt_rq
->rt_nr_running
) {
885 if (!rt_rq_throttled(rt_rq
))
888 if (rt_rq
->rt_throttled
)
892 sched_rt_rq_enqueue(rt_rq
);
893 raw_spin_unlock(&rq
->lock
);
896 if (!throttled
&& (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
))
902 static inline int rt_se_prio(struct sched_rt_entity
*rt_se
)
904 #ifdef CONFIG_RT_GROUP_SCHED
905 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
908 return rt_rq
->highest_prio
.curr
;
911 return rt_task_of(rt_se
)->prio
;
914 static int sched_rt_runtime_exceeded(struct rt_rq
*rt_rq
)
916 u64 runtime
= sched_rt_runtime(rt_rq
);
918 if (rt_rq
->rt_throttled
)
919 return rt_rq_throttled(rt_rq
);
921 if (runtime
>= sched_rt_period(rt_rq
))
924 balance_runtime(rt_rq
);
925 runtime
= sched_rt_runtime(rt_rq
);
926 if (runtime
== RUNTIME_INF
)
929 if (rt_rq
->rt_time
> runtime
) {
930 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
933 * Don't actually throttle groups that have no runtime assigned
934 * but accrue some time due to boosting.
936 if (likely(rt_b
->rt_runtime
)) {
937 rt_rq
->rt_throttled
= 1;
938 printk_deferred_once("sched: RT throttling activated\n");
941 * In case we did anyway, make it go away,
942 * replenishment is a joke, since it will replenish us
948 if (rt_rq_throttled(rt_rq
)) {
949 sched_rt_rq_dequeue(rt_rq
);
958 * Update the current task's runtime statistics. Skip current tasks that
959 * are not in our scheduling class.
961 static void update_curr_rt(struct rq
*rq
)
963 struct task_struct
*curr
= rq
->curr
;
964 struct sched_rt_entity
*rt_se
= &curr
->rt
;
967 if (curr
->sched_class
!= &rt_sched_class
)
970 delta_exec
= rq_clock_task(rq
) - curr
->se
.exec_start
;
971 if (unlikely((s64
)delta_exec
<= 0))
974 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
975 cpufreq_update_util(rq
, SCHED_CPUFREQ_RT
);
977 schedstat_set(curr
->se
.statistics
.exec_max
,
978 max(curr
->se
.statistics
.exec_max
, delta_exec
));
980 curr
->se
.sum_exec_runtime
+= delta_exec
;
981 account_group_exec_runtime(curr
, delta_exec
);
983 curr
->se
.exec_start
= rq_clock_task(rq
);
984 cpuacct_charge(curr
, delta_exec
);
986 sched_rt_avg_update(rq
, delta_exec
);
988 if (!rt_bandwidth_enabled())
991 for_each_sched_rt_entity(rt_se
) {
992 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
995 if (sched_rt_runtime(rt_rq
) != RUNTIME_INF
) {
996 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
997 rt_rq
->rt_time
+= delta_exec
;
998 exceeded
= sched_rt_runtime_exceeded(rt_rq
);
1001 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
1003 do_start_rt_bandwidth(sched_rt_bandwidth(rt_rq
));
1009 dequeue_top_rt_rq(struct rt_rq
*rt_rq
)
1011 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1013 BUG_ON(&rq
->rt
!= rt_rq
);
1015 if (!rt_rq
->rt_queued
)
1018 BUG_ON(!rq
->nr_running
);
1020 sub_nr_running(rq
, rt_rq
->rt_nr_running
);
1021 rt_rq
->rt_queued
= 0;
1025 enqueue_top_rt_rq(struct rt_rq
*rt_rq
)
1027 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1029 BUG_ON(&rq
->rt
!= rt_rq
);
1031 if (rt_rq
->rt_queued
)
1033 if (rt_rq_throttled(rt_rq
) || !rt_rq
->rt_nr_running
)
1036 add_nr_running(rq
, rt_rq
->rt_nr_running
);
1037 rt_rq
->rt_queued
= 1;
1040 #if defined CONFIG_SMP
1043 inc_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
)
1045 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1047 #ifdef CONFIG_RT_GROUP_SCHED
1049 * Change rq's cpupri only if rt_rq is the top queue.
1051 if (&rq
->rt
!= rt_rq
)
1054 if (rq
->online
&& prio
< prev_prio
)
1055 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, prio
);
1059 dec_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
)
1061 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1063 #ifdef CONFIG_RT_GROUP_SCHED
1065 * Change rq's cpupri only if rt_rq is the top queue.
1067 if (&rq
->rt
!= rt_rq
)
1070 if (rq
->online
&& rt_rq
->highest_prio
.curr
!= prev_prio
)
1071 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rt_rq
->highest_prio
.curr
);
1074 #else /* CONFIG_SMP */
1077 void inc_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
) {}
1079 void dec_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
) {}
1081 #endif /* CONFIG_SMP */
1083 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1085 inc_rt_prio(struct rt_rq
*rt_rq
, int prio
)
1087 int prev_prio
= rt_rq
->highest_prio
.curr
;
1089 if (prio
< prev_prio
)
1090 rt_rq
->highest_prio
.curr
= prio
;
1092 inc_rt_prio_smp(rt_rq
, prio
, prev_prio
);
1096 dec_rt_prio(struct rt_rq
*rt_rq
, int prio
)
1098 int prev_prio
= rt_rq
->highest_prio
.curr
;
1100 if (rt_rq
->rt_nr_running
) {
1102 WARN_ON(prio
< prev_prio
);
1105 * This may have been our highest task, and therefore
1106 * we may have some recomputation to do
1108 if (prio
== prev_prio
) {
1109 struct rt_prio_array
*array
= &rt_rq
->active
;
1111 rt_rq
->highest_prio
.curr
=
1112 sched_find_first_bit(array
->bitmap
);
1116 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
1118 dec_rt_prio_smp(rt_rq
, prio
, prev_prio
);
1123 static inline void inc_rt_prio(struct rt_rq
*rt_rq
, int prio
) {}
1124 static inline void dec_rt_prio(struct rt_rq
*rt_rq
, int prio
) {}
1126 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1128 #ifdef CONFIG_RT_GROUP_SCHED
1131 inc_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1133 if (rt_se_boosted(rt_se
))
1134 rt_rq
->rt_nr_boosted
++;
1137 start_rt_bandwidth(&rt_rq
->tg
->rt_bandwidth
);
1141 dec_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1143 if (rt_se_boosted(rt_se
))
1144 rt_rq
->rt_nr_boosted
--;
1146 WARN_ON(!rt_rq
->rt_nr_running
&& rt_rq
->rt_nr_boosted
);
1149 #else /* CONFIG_RT_GROUP_SCHED */
1152 inc_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1154 start_rt_bandwidth(&def_rt_bandwidth
);
1158 void dec_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
) {}
1160 #endif /* CONFIG_RT_GROUP_SCHED */
1163 unsigned int rt_se_nr_running(struct sched_rt_entity
*rt_se
)
1165 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1168 return group_rq
->rt_nr_running
;
1174 unsigned int rt_se_rr_nr_running(struct sched_rt_entity
*rt_se
)
1176 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1177 struct task_struct
*tsk
;
1180 return group_rq
->rr_nr_running
;
1182 tsk
= rt_task_of(rt_se
);
1184 return (tsk
->policy
== SCHED_RR
) ? 1 : 0;
1188 void inc_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1190 int prio
= rt_se_prio(rt_se
);
1192 WARN_ON(!rt_prio(prio
));
1193 rt_rq
->rt_nr_running
+= rt_se_nr_running(rt_se
);
1194 rt_rq
->rr_nr_running
+= rt_se_rr_nr_running(rt_se
);
1196 inc_rt_prio(rt_rq
, prio
);
1197 inc_rt_migration(rt_se
, rt_rq
);
1198 inc_rt_group(rt_se
, rt_rq
);
1202 void dec_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1204 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
1205 WARN_ON(!rt_rq
->rt_nr_running
);
1206 rt_rq
->rt_nr_running
-= rt_se_nr_running(rt_se
);
1207 rt_rq
->rr_nr_running
-= rt_se_rr_nr_running(rt_se
);
1209 dec_rt_prio(rt_rq
, rt_se_prio(rt_se
));
1210 dec_rt_migration(rt_se
, rt_rq
);
1211 dec_rt_group(rt_se
, rt_rq
);
1215 * Change rt_se->run_list location unless SAVE && !MOVE
1217 * assumes ENQUEUE/DEQUEUE flags match
1219 static inline bool move_entity(unsigned int flags
)
1221 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) == DEQUEUE_SAVE
)
1227 static void __delist_rt_entity(struct sched_rt_entity
*rt_se
, struct rt_prio_array
*array
)
1229 list_del_init(&rt_se
->run_list
);
1231 if (list_empty(array
->queue
+ rt_se_prio(rt_se
)))
1232 __clear_bit(rt_se_prio(rt_se
), array
->bitmap
);
1237 static void __enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1239 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
1240 struct rt_prio_array
*array
= &rt_rq
->active
;
1241 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1242 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
1245 * Don't enqueue the group if its throttled, or when empty.
1246 * The latter is a consequence of the former when a child group
1247 * get throttled and the current group doesn't have any other
1250 if (group_rq
&& (rt_rq_throttled(group_rq
) || !group_rq
->rt_nr_running
)) {
1252 __delist_rt_entity(rt_se
, array
);
1256 if (move_entity(flags
)) {
1257 WARN_ON_ONCE(rt_se
->on_list
);
1258 if (flags
& ENQUEUE_HEAD
)
1259 list_add(&rt_se
->run_list
, queue
);
1261 list_add_tail(&rt_se
->run_list
, queue
);
1263 __set_bit(rt_se_prio(rt_se
), array
->bitmap
);
1268 inc_rt_tasks(rt_se
, rt_rq
);
1271 static void __dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1273 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
1274 struct rt_prio_array
*array
= &rt_rq
->active
;
1276 if (move_entity(flags
)) {
1277 WARN_ON_ONCE(!rt_se
->on_list
);
1278 __delist_rt_entity(rt_se
, array
);
1282 dec_rt_tasks(rt_se
, rt_rq
);
1286 * Because the prio of an upper entry depends on the lower
1287 * entries, we must remove entries top - down.
1289 static void dequeue_rt_stack(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1291 struct sched_rt_entity
*back
= NULL
;
1293 for_each_sched_rt_entity(rt_se
) {
1298 dequeue_top_rt_rq(rt_rq_of_se(back
));
1300 for (rt_se
= back
; rt_se
; rt_se
= rt_se
->back
) {
1301 if (on_rt_rq(rt_se
))
1302 __dequeue_rt_entity(rt_se
, flags
);
1306 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1308 struct rq
*rq
= rq_of_rt_se(rt_se
);
1310 dequeue_rt_stack(rt_se
, flags
);
1311 for_each_sched_rt_entity(rt_se
)
1312 __enqueue_rt_entity(rt_se
, flags
);
1313 enqueue_top_rt_rq(&rq
->rt
);
1316 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1318 struct rq
*rq
= rq_of_rt_se(rt_se
);
1320 dequeue_rt_stack(rt_se
, flags
);
1322 for_each_sched_rt_entity(rt_se
) {
1323 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
1325 if (rt_rq
&& rt_rq
->rt_nr_running
)
1326 __enqueue_rt_entity(rt_se
, flags
);
1328 enqueue_top_rt_rq(&rq
->rt
);
1332 * Adding/removing a task to/from a priority array:
1335 enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1337 struct sched_rt_entity
*rt_se
= &p
->rt
;
1339 schedtune_enqueue_task(p
, cpu_of(rq
));
1341 if (flags
& ENQUEUE_WAKEUP
)
1344 enqueue_rt_entity(rt_se
, flags
);
1345 walt_inc_cumulative_runnable_avg(rq
, p
);
1347 if (!task_current(rq
, p
) && p
->nr_cpus_allowed
> 1)
1348 enqueue_pushable_task(rq
, p
);
1351 static void dequeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1353 struct sched_rt_entity
*rt_se
= &p
->rt
;
1355 schedtune_dequeue_task(p
, cpu_of(rq
));
1358 dequeue_rt_entity(rt_se
, flags
);
1359 walt_dec_cumulative_runnable_avg(rq
, p
);
1361 dequeue_pushable_task(rq
, p
);
1365 * Put task to the head or the end of the run list without the overhead of
1366 * dequeue followed by enqueue.
1369 requeue_rt_entity(struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
, int head
)
1371 if (on_rt_rq(rt_se
)) {
1372 struct rt_prio_array
*array
= &rt_rq
->active
;
1373 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
1376 list_move(&rt_se
->run_list
, queue
);
1378 list_move_tail(&rt_se
->run_list
, queue
);
1382 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int head
)
1384 struct sched_rt_entity
*rt_se
= &p
->rt
;
1385 struct rt_rq
*rt_rq
;
1387 for_each_sched_rt_entity(rt_se
) {
1388 rt_rq
= rt_rq_of_se(rt_se
);
1389 requeue_rt_entity(rt_rq
, rt_se
, head
);
1393 static void yield_task_rt(struct rq
*rq
)
1395 requeue_task_rt(rq
, rq
->curr
, 0);
1399 static int find_lowest_rq(struct task_struct
*task
);
1402 select_task_rq_rt(struct task_struct
*p
, int cpu
, int sd_flag
, int flags
,
1403 int sibling_count_hint
)
1405 struct task_struct
*curr
;
1408 /* For anything but wake ups, just return the task_cpu */
1409 if (sd_flag
!= SD_BALANCE_WAKE
&& sd_flag
!= SD_BALANCE_FORK
)
1415 curr
= READ_ONCE(rq
->curr
); /* unlocked access */
1418 * If the current task on @p's runqueue is an RT task, then
1419 * try to see if we can wake this RT task up on another
1420 * runqueue. Otherwise simply start this RT task
1421 * on its current runqueue.
1423 * We want to avoid overloading runqueues. If the woken
1424 * task is a higher priority, then it will stay on this CPU
1425 * and the lower prio task should be moved to another CPU.
1426 * Even though this will probably make the lower prio task
1427 * lose its cache, we do not want to bounce a higher task
1428 * around just because it gave up its CPU, perhaps for a
1431 * For equal prio tasks, we just let the scheduler sort it out.
1433 * Otherwise, just let it ride on the affined RQ and the
1434 * post-schedule router will push the preempted task away
1436 * This test is optimistic, if we get it wrong the load-balancer
1437 * will have to sort it out.
1439 if (curr
&& unlikely(rt_task(curr
)) &&
1440 (curr
->nr_cpus_allowed
< 2 ||
1441 curr
->prio
<= p
->prio
)) {
1442 int target
= find_lowest_rq(p
);
1445 * Don't bother moving it if the destination CPU is
1446 * not running a lower priority task.
1449 p
->prio
< cpu_rq(target
)->rt
.highest_prio
.curr
)
1458 static void check_preempt_equal_prio(struct rq
*rq
, struct task_struct
*p
)
1461 * Current can't be migrated, useless to reschedule,
1462 * let's hope p can move out.
1464 if (rq
->curr
->nr_cpus_allowed
== 1 ||
1465 !cpupri_find(&rq
->rd
->cpupri
, rq
->curr
, NULL
))
1469 * p is migratable, so let's not schedule it and
1470 * see if it is pushed or pulled somewhere else.
1472 if (p
->nr_cpus_allowed
!= 1
1473 && cpupri_find(&rq
->rd
->cpupri
, p
, NULL
))
1477 * There appears to be other cpus that can accept
1478 * current and none to run 'p', so lets reschedule
1479 * to try and push current away:
1481 requeue_task_rt(rq
, p
, 1);
1485 #endif /* CONFIG_SMP */
1488 * Preempt the current task with a newly woken task if needed:
1490 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1492 if (p
->prio
< rq
->curr
->prio
) {
1501 * - the newly woken task is of equal priority to the current task
1502 * - the newly woken task is non-migratable while current is migratable
1503 * - current will be preempted on the next reschedule
1505 * we should check to see if current can readily move to a different
1506 * cpu. If so, we will reschedule to allow the push logic to try
1507 * to move current somewhere else, making room for our non-migratable
1510 if (p
->prio
== rq
->curr
->prio
&& !test_tsk_need_resched(rq
->curr
))
1511 check_preempt_equal_prio(rq
, p
);
1515 static struct sched_rt_entity
*pick_next_rt_entity(struct rq
*rq
,
1516 struct rt_rq
*rt_rq
)
1518 struct rt_prio_array
*array
= &rt_rq
->active
;
1519 struct sched_rt_entity
*next
= NULL
;
1520 struct list_head
*queue
;
1523 idx
= sched_find_first_bit(array
->bitmap
);
1524 BUG_ON(idx
>= MAX_RT_PRIO
);
1526 queue
= array
->queue
+ idx
;
1527 next
= list_entry(queue
->next
, struct sched_rt_entity
, run_list
);
1532 static struct task_struct
*_pick_next_task_rt(struct rq
*rq
)
1534 struct sched_rt_entity
*rt_se
;
1535 struct task_struct
*p
;
1536 struct rt_rq
*rt_rq
= &rq
->rt
;
1539 rt_se
= pick_next_rt_entity(rq
, rt_rq
);
1541 rt_rq
= group_rt_rq(rt_se
);
1544 p
= rt_task_of(rt_se
);
1545 p
->se
.exec_start
= rq_clock_task(rq
);
1550 extern int update_rt_rq_load_avg(u64 now
, int cpu
, struct rt_rq
*rt_rq
, int running
);
1552 static struct task_struct
*
1553 pick_next_task_rt(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
1555 struct task_struct
*p
;
1556 struct rt_rq
*rt_rq
= &rq
->rt
;
1558 if (need_pull_rt_task(rq
, prev
)) {
1560 * This is OK, because current is on_cpu, which avoids it being
1561 * picked for load-balance and preemption/IRQs are still
1562 * disabled avoiding further scheduler activity on it and we're
1563 * being very careful to re-start the picking loop.
1565 rq_unpin_lock(rq
, rf
);
1567 rq_repin_lock(rq
, rf
);
1569 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1570 * means a dl or stop task can slip in, in which case we need
1571 * to re-start task selection.
1573 if (unlikely((rq
->stop
&& task_on_rq_queued(rq
->stop
)) ||
1574 rq
->dl
.dl_nr_running
))
1579 * We may dequeue prev's rt_rq in put_prev_task().
1580 * So, we update time before rt_nr_running check.
1582 if (prev
->sched_class
== &rt_sched_class
)
1585 if (!rt_rq
->rt_queued
)
1588 put_prev_task(rq
, prev
);
1590 p
= _pick_next_task_rt(rq
);
1592 /* The running task is never eligible for pushing */
1593 dequeue_pushable_task(rq
, p
);
1595 queue_push_tasks(rq
);
1598 update_rt_rq_load_avg(rq_clock_task(rq
), cpu_of(rq
), rt_rq
,
1599 rq
->curr
->sched_class
== &rt_sched_class
);
1604 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
1608 update_rt_rq_load_avg(rq_clock_task(rq
), cpu_of(rq
), &rq
->rt
, 1);
1611 * The previous task needs to be made eligible for pushing
1612 * if it is still active
1614 if (on_rt_rq(&p
->rt
) && p
->nr_cpus_allowed
> 1)
1615 enqueue_pushable_task(rq
, p
);
1620 /* Only try algorithms three times */
1621 #define RT_MAX_TRIES 3
1623 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
1625 if (!task_running(rq
, p
) &&
1626 cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
1632 * Return the highest pushable rq's task, which is suitable to be executed
1633 * on the cpu, NULL otherwise
1635 static struct task_struct
*pick_highest_pushable_task(struct rq
*rq
, int cpu
)
1637 struct plist_head
*head
= &rq
->rt
.pushable_tasks
;
1638 struct task_struct
*p
;
1640 if (!has_pushable_tasks(rq
))
1643 plist_for_each_entry(p
, head
, pushable_tasks
) {
1644 if (pick_rt_task(rq
, p
, cpu
))
1651 static DEFINE_PER_CPU(cpumask_var_t
, local_cpu_mask
);
1653 static int find_lowest_rq(struct task_struct
*task
)
1655 struct sched_domain
*sd
;
1656 struct cpumask
*lowest_mask
= this_cpu_cpumask_var_ptr(local_cpu_mask
);
1657 int this_cpu
= smp_processor_id();
1658 int cpu
= task_cpu(task
);
1660 /* Make sure the mask is initialized first */
1661 if (unlikely(!lowest_mask
))
1664 if (task
->nr_cpus_allowed
== 1)
1665 return -1; /* No other targets possible */
1667 if (!cpupri_find(&task_rq(task
)->rd
->cpupri
, task
, lowest_mask
))
1668 return -1; /* No targets found */
1671 * At this point we have built a mask of cpus representing the
1672 * lowest priority tasks in the system. Now we want to elect
1673 * the best one based on our affinity and topology.
1675 * We prioritize the last cpu that the task executed on since
1676 * it is most likely cache-hot in that location.
1678 if (cpumask_test_cpu(cpu
, lowest_mask
))
1682 * Otherwise, we consult the sched_domains span maps to figure
1683 * out which cpu is logically closest to our hot cache data.
1685 if (!cpumask_test_cpu(this_cpu
, lowest_mask
))
1686 this_cpu
= -1; /* Skip this_cpu opt if not among lowest */
1689 for_each_domain(cpu
, sd
) {
1690 if (sd
->flags
& SD_WAKE_AFFINE
) {
1694 * "this_cpu" is cheaper to preempt than a
1697 if (this_cpu
!= -1 &&
1698 cpumask_test_cpu(this_cpu
, sched_domain_span(sd
))) {
1703 best_cpu
= cpumask_first_and(lowest_mask
,
1704 sched_domain_span(sd
));
1705 if (best_cpu
< nr_cpu_ids
) {
1714 * And finally, if there were no matches within the domains
1715 * just give the caller *something* to work with from the compatible
1721 cpu
= cpumask_any(lowest_mask
);
1722 if (cpu
< nr_cpu_ids
)
1727 /* Will lock the rq it finds */
1728 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
1730 struct rq
*lowest_rq
= NULL
;
1734 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
1735 cpu
= find_lowest_rq(task
);
1737 if ((cpu
== -1) || (cpu
== rq
->cpu
))
1740 lowest_rq
= cpu_rq(cpu
);
1742 if (lowest_rq
->rt
.highest_prio
.curr
<= task
->prio
) {
1744 * Target rq has tasks of equal or higher priority,
1745 * retrying does not release any lock and is unlikely
1746 * to yield a different result.
1752 /* if the prio of this runqueue changed, try again */
1753 if (double_lock_balance(rq
, lowest_rq
)) {
1755 * We had to unlock the run queue. In
1756 * the mean time, task could have
1757 * migrated already or had its affinity changed.
1758 * Also make sure that it wasn't scheduled on its rq.
1760 if (unlikely(task_rq(task
) != rq
||
1761 !cpumask_test_cpu(lowest_rq
->cpu
, &task
->cpus_allowed
) ||
1762 task_running(rq
, task
) ||
1764 !task_on_rq_queued(task
))) {
1766 double_unlock_balance(rq
, lowest_rq
);
1772 /* If this rq is still suitable use it. */
1773 if (lowest_rq
->rt
.highest_prio
.curr
> task
->prio
)
1777 double_unlock_balance(rq
, lowest_rq
);
1784 static struct task_struct
*pick_next_pushable_task(struct rq
*rq
)
1786 struct task_struct
*p
;
1788 if (!has_pushable_tasks(rq
))
1791 p
= plist_first_entry(&rq
->rt
.pushable_tasks
,
1792 struct task_struct
, pushable_tasks
);
1794 BUG_ON(rq
->cpu
!= task_cpu(p
));
1795 BUG_ON(task_current(rq
, p
));
1796 BUG_ON(p
->nr_cpus_allowed
<= 1);
1798 BUG_ON(!task_on_rq_queued(p
));
1799 BUG_ON(!rt_task(p
));
1805 * If the current CPU has more than one RT task, see if the non
1806 * running task can migrate over to a CPU that is running a task
1807 * of lesser priority.
1809 static int push_rt_task(struct rq
*rq
)
1811 struct task_struct
*next_task
;
1812 struct rq
*lowest_rq
;
1815 if (!rq
->rt
.overloaded
)
1818 next_task
= pick_next_pushable_task(rq
);
1823 if (unlikely(next_task
== rq
->curr
)) {
1829 * It's possible that the next_task slipped in of
1830 * higher priority than current. If that's the case
1831 * just reschedule current.
1833 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
1838 /* We might release rq lock */
1839 get_task_struct(next_task
);
1841 /* find_lock_lowest_rq locks the rq if found */
1842 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
1844 struct task_struct
*task
;
1846 * find_lock_lowest_rq releases rq->lock
1847 * so it is possible that next_task has migrated.
1849 * We need to make sure that the task is still on the same
1850 * run-queue and is also still the next task eligible for
1853 task
= pick_next_pushable_task(rq
);
1854 if (task
== next_task
) {
1856 * The task hasn't migrated, and is still the next
1857 * eligible task, but we failed to find a run-queue
1858 * to push it to. Do not retry in this case, since
1859 * other cpus will pull from us when ready.
1865 /* No more tasks, just exit */
1869 * Something has shifted, try again.
1871 put_task_struct(next_task
);
1876 deactivate_task(rq
, next_task
, 0);
1877 next_task
->on_rq
= TASK_ON_RQ_MIGRATING
;
1878 set_task_cpu(next_task
, lowest_rq
->cpu
);
1879 next_task
->on_rq
= TASK_ON_RQ_QUEUED
;
1880 activate_task(lowest_rq
, next_task
, 0);
1883 resched_curr(lowest_rq
);
1885 double_unlock_balance(rq
, lowest_rq
);
1888 put_task_struct(next_task
);
1893 static void push_rt_tasks(struct rq
*rq
)
1895 /* push_rt_task will return true if it moved an RT */
1896 while (push_rt_task(rq
))
1900 #ifdef HAVE_RT_PUSH_IPI
1903 * When a high priority task schedules out from a CPU and a lower priority
1904 * task is scheduled in, a check is made to see if there's any RT tasks
1905 * on other CPUs that are waiting to run because a higher priority RT task
1906 * is currently running on its CPU. In this case, the CPU with multiple RT
1907 * tasks queued on it (overloaded) needs to be notified that a CPU has opened
1908 * up that may be able to run one of its non-running queued RT tasks.
1910 * All CPUs with overloaded RT tasks need to be notified as there is currently
1911 * no way to know which of these CPUs have the highest priority task waiting
1912 * to run. Instead of trying to take a spinlock on each of these CPUs,
1913 * which has shown to cause large latency when done on machines with many
1914 * CPUs, sending an IPI to the CPUs to have them push off the overloaded
1915 * RT tasks waiting to run.
1917 * Just sending an IPI to each of the CPUs is also an issue, as on large
1918 * count CPU machines, this can cause an IPI storm on a CPU, especially
1919 * if its the only CPU with multiple RT tasks queued, and a large number
1920 * of CPUs scheduling a lower priority task at the same time.
1922 * Each root domain has its own irq work function that can iterate over
1923 * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT
1924 * tassk must be checked if there's one or many CPUs that are lowering
1925 * their priority, there's a single irq work iterator that will try to
1926 * push off RT tasks that are waiting to run.
1928 * When a CPU schedules a lower priority task, it will kick off the
1929 * irq work iterator that will jump to each CPU with overloaded RT tasks.
1930 * As it only takes the first CPU that schedules a lower priority task
1931 * to start the process, the rto_start variable is incremented and if
1932 * the atomic result is one, then that CPU will try to take the rto_lock.
1933 * This prevents high contention on the lock as the process handles all
1934 * CPUs scheduling lower priority tasks.
1936 * All CPUs that are scheduling a lower priority task will increment the
1937 * rt_loop_next variable. This will make sure that the irq work iterator
1938 * checks all RT overloaded CPUs whenever a CPU schedules a new lower
1939 * priority task, even if the iterator is in the middle of a scan. Incrementing
1940 * the rt_loop_next will cause the iterator to perform another scan.
1943 static int rto_next_cpu(struct root_domain
*rd
)
1949 * When starting the IPI RT pushing, the rto_cpu is set to -1,
1950 * rt_next_cpu() will simply return the first CPU found in
1953 * If rto_next_cpu() is called with rto_cpu is a valid cpu, it
1954 * will return the next CPU found in the rto_mask.
1956 * If there are no more CPUs left in the rto_mask, then a check is made
1957 * against rto_loop and rto_loop_next. rto_loop is only updated with
1958 * the rto_lock held, but any CPU may increment the rto_loop_next
1959 * without any locking.
1963 /* When rto_cpu is -1 this acts like cpumask_first() */
1964 cpu
= cpumask_next(rd
->rto_cpu
, rd
->rto_mask
);
1968 if (cpu
< nr_cpu_ids
)
1974 * ACQUIRE ensures we see the @rto_mask changes
1975 * made prior to the @next value observed.
1977 * Matches WMB in rt_set_overload().
1979 next
= atomic_read_acquire(&rd
->rto_loop_next
);
1981 if (rd
->rto_loop
== next
)
1984 rd
->rto_loop
= next
;
1990 static inline bool rto_start_trylock(atomic_t
*v
)
1992 return !atomic_cmpxchg_acquire(v
, 0, 1);
1995 static inline void rto_start_unlock(atomic_t
*v
)
1997 atomic_set_release(v
, 0);
2000 static void tell_cpu_to_push(struct rq
*rq
)
2004 /* Keep the loop going if the IPI is currently active */
2005 atomic_inc(&rq
->rd
->rto_loop_next
);
2007 /* Only one CPU can initiate a loop at a time */
2008 if (!rto_start_trylock(&rq
->rd
->rto_loop_start
))
2011 raw_spin_lock(&rq
->rd
->rto_lock
);
2014 * The rto_cpu is updated under the lock, if it has a valid cpu
2015 * then the IPI is still running and will continue due to the
2016 * update to loop_next, and nothing needs to be done here.
2017 * Otherwise it is finishing up and an ipi needs to be sent.
2019 if (rq
->rd
->rto_cpu
< 0)
2020 cpu
= rto_next_cpu(rq
->rd
);
2022 raw_spin_unlock(&rq
->rd
->rto_lock
);
2024 rto_start_unlock(&rq
->rd
->rto_loop_start
);
2027 /* Make sure the rd does not get freed while pushing */
2028 sched_get_rd(rq
->rd
);
2029 irq_work_queue_on(&rq
->rd
->rto_push_work
, cpu
);
2033 /* Called from hardirq context */
2034 void rto_push_irq_work_func(struct irq_work
*work
)
2036 struct root_domain
*rd
=
2037 container_of(work
, struct root_domain
, rto_push_work
);
2044 * We do not need to grab the lock to check for has_pushable_tasks.
2045 * When it gets updated, a check is made if a push is possible.
2047 if (has_pushable_tasks(rq
)) {
2048 raw_spin_lock(&rq
->lock
);
2050 raw_spin_unlock(&rq
->lock
);
2053 raw_spin_lock(&rd
->rto_lock
);
2055 /* Pass the IPI to the next rt overloaded queue */
2056 cpu
= rto_next_cpu(rd
);
2058 raw_spin_unlock(&rd
->rto_lock
);
2065 /* Try the next RT overloaded CPU */
2066 irq_work_queue_on(&rd
->rto_push_work
, cpu
);
2068 #endif /* HAVE_RT_PUSH_IPI */
2070 static void pull_rt_task(struct rq
*this_rq
)
2072 int this_cpu
= this_rq
->cpu
, cpu
;
2073 bool resched
= false;
2074 struct task_struct
*p
;
2076 int rt_overload_count
= rt_overloaded(this_rq
);
2078 if (likely(!rt_overload_count
))
2082 * Match the barrier from rt_set_overloaded; this guarantees that if we
2083 * see overloaded we must also see the rto_mask bit.
2087 /* If we are the only overloaded CPU do nothing */
2088 if (rt_overload_count
== 1 &&
2089 cpumask_test_cpu(this_rq
->cpu
, this_rq
->rd
->rto_mask
))
2092 #ifdef HAVE_RT_PUSH_IPI
2093 if (sched_feat(RT_PUSH_IPI
)) {
2094 tell_cpu_to_push(this_rq
);
2099 for_each_cpu(cpu
, this_rq
->rd
->rto_mask
) {
2100 if (this_cpu
== cpu
)
2103 src_rq
= cpu_rq(cpu
);
2106 * Don't bother taking the src_rq->lock if the next highest
2107 * task is known to be lower-priority than our current task.
2108 * This may look racy, but if this value is about to go
2109 * logically higher, the src_rq will push this task away.
2110 * And if its going logically lower, we do not care
2112 if (src_rq
->rt
.highest_prio
.next
>=
2113 this_rq
->rt
.highest_prio
.curr
)
2117 * We can potentially drop this_rq's lock in
2118 * double_lock_balance, and another CPU could
2121 double_lock_balance(this_rq
, src_rq
);
2124 * We can pull only a task, which is pushable
2125 * on its rq, and no others.
2127 p
= pick_highest_pushable_task(src_rq
, this_cpu
);
2130 * Do we have an RT task that preempts
2131 * the to-be-scheduled task?
2133 if (p
&& (p
->prio
< this_rq
->rt
.highest_prio
.curr
)) {
2134 WARN_ON(p
== src_rq
->curr
);
2135 WARN_ON(!task_on_rq_queued(p
));
2138 * There's a chance that p is higher in priority
2139 * than what's currently running on its cpu.
2140 * This is just that p is wakeing up and hasn't
2141 * had a chance to schedule. We only pull
2142 * p if it is lower in priority than the
2143 * current task on the run queue
2145 if (p
->prio
< src_rq
->curr
->prio
)
2150 deactivate_task(src_rq
, p
, 0);
2151 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
2152 set_task_cpu(p
, this_cpu
);
2153 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2154 activate_task(this_rq
, p
, 0);
2156 * We continue with the search, just in
2157 * case there's an even higher prio task
2158 * in another runqueue. (low likelihood
2163 double_unlock_balance(this_rq
, src_rq
);
2167 resched_curr(this_rq
);
2171 * If we are not running and we are not going to reschedule soon, we should
2172 * try to push tasks away now
2174 static void task_woken_rt(struct rq
*rq
, struct task_struct
*p
)
2176 if (!task_running(rq
, p
) &&
2177 !test_tsk_need_resched(rq
->curr
) &&
2178 p
->nr_cpus_allowed
> 1 &&
2179 (dl_task(rq
->curr
) || rt_task(rq
->curr
)) &&
2180 (rq
->curr
->nr_cpus_allowed
< 2 ||
2181 rq
->curr
->prio
<= p
->prio
))
2185 /* Assumes rq->lock is held */
2186 static void rq_online_rt(struct rq
*rq
)
2188 if (rq
->rt
.overloaded
)
2189 rt_set_overload(rq
);
2191 __enable_runtime(rq
);
2193 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rq
->rt
.highest_prio
.curr
);
2196 /* Assumes rq->lock is held */
2197 static void rq_offline_rt(struct rq
*rq
)
2199 if (rq
->rt
.overloaded
)
2200 rt_clear_overload(rq
);
2202 __disable_runtime(rq
);
2204 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, CPUPRI_INVALID
);
2208 * When switch from the rt queue, we bring ourselves to a position
2209 * that we might want to pull RT tasks from other runqueues.
2211 static void switched_from_rt(struct rq
*rq
, struct task_struct
*p
)
2214 * If there are other RT tasks then we will reschedule
2215 * and the scheduling of the other RT tasks will handle
2216 * the balancing. But if we are the last RT task
2217 * we may need to handle the pulling of RT tasks
2220 if (!task_on_rq_queued(p
) || rq
->rt
.rt_nr_running
)
2223 queue_pull_task(rq
);
2226 void __init
init_sched_rt_class(void)
2230 for_each_possible_cpu(i
) {
2231 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask
, i
),
2232 GFP_KERNEL
, cpu_to_node(i
));
2235 #endif /* CONFIG_SMP */
2238 * When switching a task to RT, we may overload the runqueue
2239 * with RT tasks. In this case we try to push them off to
2242 static void switched_to_rt(struct rq
*rq
, struct task_struct
*p
)
2245 * If we are already running, then there's nothing
2246 * that needs to be done. But if we are not running
2247 * we may need to preempt the current running task.
2248 * If that current running task is also an RT task
2249 * then see if we can move to another run queue.
2251 if (task_on_rq_queued(p
) && rq
->curr
!= p
) {
2253 if (p
->nr_cpus_allowed
> 1 && rq
->rt
.overloaded
)
2254 queue_push_tasks(rq
);
2255 #endif /* CONFIG_SMP */
2256 if (p
->prio
< rq
->curr
->prio
&& cpu_online(cpu_of(rq
)))
2262 * Priority of the task has changed. This may cause
2263 * us to initiate a push or pull.
2266 prio_changed_rt(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
2268 if (!task_on_rq_queued(p
))
2271 if (rq
->curr
== p
) {
2274 * If our priority decreases while running, we
2275 * may need to pull tasks to this runqueue.
2277 if (oldprio
< p
->prio
)
2278 queue_pull_task(rq
);
2281 * If there's a higher priority task waiting to run
2284 if (p
->prio
> rq
->rt
.highest_prio
.curr
)
2287 /* For UP simply resched on drop of prio */
2288 if (oldprio
< p
->prio
)
2290 #endif /* CONFIG_SMP */
2293 * This task is not running, but if it is
2294 * greater than the current running task
2297 if (p
->prio
< rq
->curr
->prio
)
2302 #ifdef CONFIG_POSIX_TIMERS
2303 static void watchdog(struct rq
*rq
, struct task_struct
*p
)
2305 unsigned long soft
, hard
;
2307 /* max may change after cur was read, this will be fixed next tick */
2308 soft
= task_rlimit(p
, RLIMIT_RTTIME
);
2309 hard
= task_rlimit_max(p
, RLIMIT_RTTIME
);
2311 if (soft
!= RLIM_INFINITY
) {
2314 if (p
->rt
.watchdog_stamp
!= jiffies
) {
2316 p
->rt
.watchdog_stamp
= jiffies
;
2319 next
= DIV_ROUND_UP(min(soft
, hard
), USEC_PER_SEC
/HZ
);
2320 if (p
->rt
.timeout
> next
)
2321 p
->cputime_expires
.sched_exp
= p
->se
.sum_exec_runtime
;
2325 static inline void watchdog(struct rq
*rq
, struct task_struct
*p
) { }
2328 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
, int queued
)
2330 struct sched_rt_entity
*rt_se
= &p
->rt
;
2333 update_rt_rq_load_avg(rq_clock_task(rq
), cpu_of(rq
), &rq
->rt
, 1);
2338 * RR tasks need a special form of timeslice management.
2339 * FIFO tasks have no timeslices.
2341 if (p
->policy
!= SCHED_RR
)
2344 if (--p
->rt
.time_slice
)
2347 p
->rt
.time_slice
= sched_rr_timeslice
;
2350 * Requeue to the end of queue if we (and all of our ancestors) are not
2351 * the only element on the queue
2353 for_each_sched_rt_entity(rt_se
) {
2354 if (rt_se
->run_list
.prev
!= rt_se
->run_list
.next
) {
2355 requeue_task_rt(rq
, p
, 0);
2362 static void set_curr_task_rt(struct rq
*rq
)
2364 struct task_struct
*p
= rq
->curr
;
2366 p
->se
.exec_start
= rq_clock_task(rq
);
2368 /* The running task is never eligible for pushing */
2369 dequeue_pushable_task(rq
, p
);
2372 static unsigned int get_rr_interval_rt(struct rq
*rq
, struct task_struct
*task
)
2375 * Time slice is 0 for SCHED_FIFO tasks
2377 if (task
->policy
== SCHED_RR
)
2378 return sched_rr_timeslice
;
2383 const struct sched_class rt_sched_class
= {
2384 .next
= &fair_sched_class
,
2385 .enqueue_task
= enqueue_task_rt
,
2386 .dequeue_task
= dequeue_task_rt
,
2387 .yield_task
= yield_task_rt
,
2389 .check_preempt_curr
= check_preempt_curr_rt
,
2391 .pick_next_task
= pick_next_task_rt
,
2392 .put_prev_task
= put_prev_task_rt
,
2395 .select_task_rq
= select_task_rq_rt
,
2397 .set_cpus_allowed
= set_cpus_allowed_common
,
2398 .rq_online
= rq_online_rt
,
2399 .rq_offline
= rq_offline_rt
,
2400 .task_woken
= task_woken_rt
,
2401 .switched_from
= switched_from_rt
,
2404 .set_curr_task
= set_curr_task_rt
,
2405 .task_tick
= task_tick_rt
,
2407 .get_rr_interval
= get_rr_interval_rt
,
2409 .prio_changed
= prio_changed_rt
,
2410 .switched_to
= switched_to_rt
,
2412 .update_curr
= update_curr_rt
,
2415 #ifdef CONFIG_RT_GROUP_SCHED
2417 * Ensure that the real time constraints are schedulable.
2419 static DEFINE_MUTEX(rt_constraints_mutex
);
2421 /* Must be called with tasklist_lock held */
2422 static inline int tg_has_rt_tasks(struct task_group
*tg
)
2424 struct task_struct
*g
, *p
;
2427 * Autogroups do not have RT tasks; see autogroup_create().
2429 if (task_group_is_autogroup(tg
))
2432 for_each_process_thread(g
, p
) {
2433 if (rt_task(p
) && task_group(p
) == tg
)
2440 struct rt_schedulable_data
{
2441 struct task_group
*tg
;
2446 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
2448 struct rt_schedulable_data
*d
= data
;
2449 struct task_group
*child
;
2450 unsigned long total
, sum
= 0;
2451 u64 period
, runtime
;
2453 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2454 runtime
= tg
->rt_bandwidth
.rt_runtime
;
2457 period
= d
->rt_period
;
2458 runtime
= d
->rt_runtime
;
2462 * Cannot have more runtime than the period.
2464 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
2468 * Ensure we don't starve existing RT tasks.
2470 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
2473 total
= to_ratio(period
, runtime
);
2476 * Nobody can have more than the global setting allows.
2478 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
2482 * The sum of our children's runtime should not exceed our own.
2484 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
2485 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
2486 runtime
= child
->rt_bandwidth
.rt_runtime
;
2488 if (child
== d
->tg
) {
2489 period
= d
->rt_period
;
2490 runtime
= d
->rt_runtime
;
2493 sum
+= to_ratio(period
, runtime
);
2502 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
2506 struct rt_schedulable_data data
= {
2508 .rt_period
= period
,
2509 .rt_runtime
= runtime
,
2513 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
2519 static int tg_set_rt_bandwidth(struct task_group
*tg
,
2520 u64 rt_period
, u64 rt_runtime
)
2525 * Disallowing the root group RT runtime is BAD, it would disallow the
2526 * kernel creating (and or operating) RT threads.
2528 if (tg
== &root_task_group
&& rt_runtime
== 0)
2531 /* No period doesn't make any sense. */
2535 mutex_lock(&rt_constraints_mutex
);
2536 read_lock(&tasklist_lock
);
2537 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
2541 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
2542 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
2543 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
2545 for_each_possible_cpu(i
) {
2546 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
2548 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
2549 rt_rq
->rt_runtime
= rt_runtime
;
2550 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
2552 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
2554 read_unlock(&tasklist_lock
);
2555 mutex_unlock(&rt_constraints_mutex
);
2560 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
2562 u64 rt_runtime
, rt_period
;
2564 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2565 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
2566 if (rt_runtime_us
< 0)
2567 rt_runtime
= RUNTIME_INF
;
2568 else if ((u64
)rt_runtime_us
> U64_MAX
/ NSEC_PER_USEC
)
2571 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
2574 long sched_group_rt_runtime(struct task_group
*tg
)
2578 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
2581 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
2582 do_div(rt_runtime_us
, NSEC_PER_USEC
);
2583 return rt_runtime_us
;
2586 int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
2588 u64 rt_runtime
, rt_period
;
2590 if (rt_period_us
> U64_MAX
/ NSEC_PER_USEC
)
2593 rt_period
= rt_period_us
* NSEC_PER_USEC
;
2594 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
2596 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
2599 long sched_group_rt_period(struct task_group
*tg
)
2603 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2604 do_div(rt_period_us
, NSEC_PER_USEC
);
2605 return rt_period_us
;
2608 static int sched_rt_global_constraints(void)
2612 mutex_lock(&rt_constraints_mutex
);
2613 read_lock(&tasklist_lock
);
2614 ret
= __rt_schedulable(NULL
, 0, 0);
2615 read_unlock(&tasklist_lock
);
2616 mutex_unlock(&rt_constraints_mutex
);
2621 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
2623 /* Don't accept realtime tasks when there is no way for them to run */
2624 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
2630 #else /* !CONFIG_RT_GROUP_SCHED */
2631 static int sched_rt_global_constraints(void)
2633 unsigned long flags
;
2636 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2637 for_each_possible_cpu(i
) {
2638 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
2640 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
2641 rt_rq
->rt_runtime
= global_rt_runtime();
2642 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
2644 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2648 #endif /* CONFIG_RT_GROUP_SCHED */
2650 static int sched_rt_global_validate(void)
2652 if (sysctl_sched_rt_period
<= 0)
2655 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
2656 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
2662 static void sched_rt_do_global(void)
2664 unsigned long flags
;
2666 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2667 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
2668 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
2669 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2672 int sched_rt_handler(struct ctl_table
*table
, int write
,
2673 void __user
*buffer
, size_t *lenp
,
2676 int old_period
, old_runtime
;
2677 static DEFINE_MUTEX(mutex
);
2681 old_period
= sysctl_sched_rt_period
;
2682 old_runtime
= sysctl_sched_rt_runtime
;
2684 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
2686 if (!ret
&& write
) {
2687 ret
= sched_rt_global_validate();
2691 ret
= sched_dl_global_validate();
2695 ret
= sched_rt_global_constraints();
2699 sched_rt_do_global();
2700 sched_dl_do_global();
2704 sysctl_sched_rt_period
= old_period
;
2705 sysctl_sched_rt_runtime
= old_runtime
;
2707 mutex_unlock(&mutex
);
2712 int sched_rr_handler(struct ctl_table
*table
, int write
,
2713 void __user
*buffer
, size_t *lenp
,
2717 static DEFINE_MUTEX(mutex
);
2720 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
2722 * Make sure that internally we keep jiffies.
2723 * Also, writing zero resets the timeslice to default:
2725 if (!ret
&& write
) {
2726 sched_rr_timeslice
=
2727 sysctl_sched_rr_timeslice
<= 0 ? RR_TIMESLICE
:
2728 msecs_to_jiffies(sysctl_sched_rr_timeslice
);
2730 mutex_unlock(&mutex
);
2734 #ifdef CONFIG_SCHED_DEBUG
2735 void print_rt_stats(struct seq_file
*m
, int cpu
)
2738 struct rt_rq
*rt_rq
;
2741 for_each_rt_rq(rt_rq
, iter
, cpu_rq(cpu
))
2742 print_rt_rq(m
, cpu
, rt_rq
);
2745 #endif /* CONFIG_SCHED_DEBUG */