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 void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
60 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
63 raw_spin_lock(&rt_b
->rt_runtime_lock
);
64 if (!rt_b
->rt_period_active
) {
65 rt_b
->rt_period_active
= 1;
67 * SCHED_DEADLINE updates the bandwidth, as a run away
68 * RT task with a DL task could hog a CPU. But DL does
69 * not reset the period. If a deadline task was running
70 * without an RT task running, it can cause RT tasks to
71 * throttle when they start up. Kick the timer right away
72 * to update the period.
74 hrtimer_forward_now(&rt_b
->rt_period_timer
, ns_to_ktime(0));
75 hrtimer_start_expires(&rt_b
->rt_period_timer
, HRTIMER_MODE_ABS_PINNED
);
77 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
80 void init_rt_rq(struct rt_rq
*rt_rq
)
82 struct rt_prio_array
*array
;
85 array
= &rt_rq
->active
;
86 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
87 INIT_LIST_HEAD(array
->queue
+ i
);
88 __clear_bit(i
, array
->bitmap
);
90 /* delimiter for bitsearch: */
91 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
93 #if defined CONFIG_SMP
94 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
95 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
96 rt_rq
->rt_nr_migratory
= 0;
97 rt_rq
->overloaded
= 0;
98 plist_head_init(&rt_rq
->pushable_tasks
);
99 #endif /* CONFIG_SMP */
100 /* We start is dequeued state, because no RT tasks are queued */
101 rt_rq
->rt_queued
= 0;
104 rt_rq
->rt_throttled
= 0;
105 rt_rq
->rt_runtime
= 0;
106 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
109 #ifdef CONFIG_RT_GROUP_SCHED
110 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
112 hrtimer_cancel(&rt_b
->rt_period_timer
);
115 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
117 static inline struct task_struct
*rt_task_of(struct sched_rt_entity
*rt_se
)
119 #ifdef CONFIG_SCHED_DEBUG
120 WARN_ON_ONCE(!rt_entity_is_task(rt_se
));
122 return container_of(rt_se
, struct task_struct
, rt
);
125 static inline struct rq
*rq_of_rt_rq(struct rt_rq
*rt_rq
)
130 static inline struct rt_rq
*rt_rq_of_se(struct sched_rt_entity
*rt_se
)
135 static inline struct rq
*rq_of_rt_se(struct sched_rt_entity
*rt_se
)
137 struct rt_rq
*rt_rq
= rt_se
->rt_rq
;
142 void free_rt_sched_group(struct task_group
*tg
)
147 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
149 for_each_possible_cpu(i
) {
160 void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
161 struct sched_rt_entity
*rt_se
, int cpu
,
162 struct sched_rt_entity
*parent
)
164 struct rq
*rq
= cpu_rq(cpu
);
166 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
167 rt_rq
->rt_nr_boosted
= 0;
171 tg
->rt_rq
[cpu
] = rt_rq
;
172 tg
->rt_se
[cpu
] = rt_se
;
178 rt_se
->rt_rq
= &rq
->rt
;
180 rt_se
->rt_rq
= parent
->my_q
;
183 rt_se
->parent
= parent
;
184 INIT_LIST_HEAD(&rt_se
->run_list
);
187 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
190 struct sched_rt_entity
*rt_se
;
193 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
196 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
200 init_rt_bandwidth(&tg
->rt_bandwidth
,
201 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
203 for_each_possible_cpu(i
) {
204 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
205 GFP_KERNEL
, cpu_to_node(i
));
209 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
210 GFP_KERNEL
, cpu_to_node(i
));
215 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
216 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, parent
->rt_se
[i
]);
227 #else /* CONFIG_RT_GROUP_SCHED */
229 #define rt_entity_is_task(rt_se) (1)
231 static inline struct task_struct
*rt_task_of(struct sched_rt_entity
*rt_se
)
233 return container_of(rt_se
, struct task_struct
, rt
);
236 static inline struct rq
*rq_of_rt_rq(struct rt_rq
*rt_rq
)
238 return container_of(rt_rq
, struct rq
, rt
);
241 static inline struct rq
*rq_of_rt_se(struct sched_rt_entity
*rt_se
)
243 struct task_struct
*p
= rt_task_of(rt_se
);
248 static inline struct rt_rq
*rt_rq_of_se(struct sched_rt_entity
*rt_se
)
250 struct rq
*rq
= rq_of_rt_se(rt_se
);
255 void free_rt_sched_group(struct task_group
*tg
) { }
257 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
261 #endif /* CONFIG_RT_GROUP_SCHED */
265 static void pull_rt_task(struct rq
*this_rq
);
267 static inline bool need_pull_rt_task(struct rq
*rq
, struct task_struct
*prev
)
269 /* Try to pull RT tasks here if we lower this rq's prio */
270 return rq
->rt
.highest_prio
.curr
> prev
->prio
;
273 static inline int rt_overloaded(struct rq
*rq
)
275 return atomic_read(&rq
->rd
->rto_count
);
278 static inline void rt_set_overload(struct rq
*rq
)
283 cpumask_set_cpu(rq
->cpu
, rq
->rd
->rto_mask
);
285 * Make sure the mask is visible before we set
286 * the overload count. That is checked to determine
287 * if we should look at the mask. It would be a shame
288 * if we looked at the mask, but the mask was not
291 * Matched by the barrier in pull_rt_task().
294 atomic_inc(&rq
->rd
->rto_count
);
297 static inline void rt_clear_overload(struct rq
*rq
)
302 /* the order here really doesn't matter */
303 atomic_dec(&rq
->rd
->rto_count
);
304 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->rto_mask
);
307 static void update_rt_migration(struct rt_rq
*rt_rq
)
309 if (rt_rq
->rt_nr_migratory
&& rt_rq
->rt_nr_total
> 1) {
310 if (!rt_rq
->overloaded
) {
311 rt_set_overload(rq_of_rt_rq(rt_rq
));
312 rt_rq
->overloaded
= 1;
314 } else if (rt_rq
->overloaded
) {
315 rt_clear_overload(rq_of_rt_rq(rt_rq
));
316 rt_rq
->overloaded
= 0;
320 static void inc_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
322 struct task_struct
*p
;
324 if (!rt_entity_is_task(rt_se
))
327 p
= rt_task_of(rt_se
);
328 rt_rq
= &rq_of_rt_rq(rt_rq
)->rt
;
330 rt_rq
->rt_nr_total
++;
331 if (p
->nr_cpus_allowed
> 1)
332 rt_rq
->rt_nr_migratory
++;
334 update_rt_migration(rt_rq
);
337 static void dec_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
339 struct task_struct
*p
;
341 if (!rt_entity_is_task(rt_se
))
344 p
= rt_task_of(rt_se
);
345 rt_rq
= &rq_of_rt_rq(rt_rq
)->rt
;
347 rt_rq
->rt_nr_total
--;
348 if (p
->nr_cpus_allowed
> 1)
349 rt_rq
->rt_nr_migratory
--;
351 update_rt_migration(rt_rq
);
354 static inline int has_pushable_tasks(struct rq
*rq
)
356 return !plist_head_empty(&rq
->rt
.pushable_tasks
);
359 static DEFINE_PER_CPU(struct callback_head
, rt_push_head
);
360 static DEFINE_PER_CPU(struct callback_head
, rt_pull_head
);
362 static void push_rt_tasks(struct rq
*);
363 static void pull_rt_task(struct rq
*);
365 static inline void queue_push_tasks(struct rq
*rq
)
367 if (!has_pushable_tasks(rq
))
370 queue_balance_callback(rq
, &per_cpu(rt_push_head
, rq
->cpu
), push_rt_tasks
);
373 static inline void queue_pull_task(struct rq
*rq
)
375 queue_balance_callback(rq
, &per_cpu(rt_pull_head
, rq
->cpu
), pull_rt_task
);
378 static void enqueue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
380 plist_del(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
381 plist_node_init(&p
->pushable_tasks
, p
->prio
);
382 plist_add(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
384 /* Update the highest prio pushable task */
385 if (p
->prio
< rq
->rt
.highest_prio
.next
)
386 rq
->rt
.highest_prio
.next
= p
->prio
;
389 static void dequeue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
391 plist_del(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
393 /* Update the new highest prio pushable task */
394 if (has_pushable_tasks(rq
)) {
395 p
= plist_first_entry(&rq
->rt
.pushable_tasks
,
396 struct task_struct
, pushable_tasks
);
397 rq
->rt
.highest_prio
.next
= p
->prio
;
399 rq
->rt
.highest_prio
.next
= MAX_RT_PRIO
;
404 static inline void enqueue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
408 static inline void dequeue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
413 void inc_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
418 void dec_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
422 static inline bool need_pull_rt_task(struct rq
*rq
, struct task_struct
*prev
)
427 static inline void pull_rt_task(struct rq
*this_rq
)
431 static inline void queue_push_tasks(struct rq
*rq
)
434 #endif /* CONFIG_SMP */
436 static void enqueue_top_rt_rq(struct rt_rq
*rt_rq
);
437 static void dequeue_top_rt_rq(struct rt_rq
*rt_rq
);
439 static inline int on_rt_rq(struct sched_rt_entity
*rt_se
)
444 #ifdef CONFIG_RT_GROUP_SCHED
446 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
451 return rt_rq
->rt_runtime
;
454 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
456 return ktime_to_ns(rt_rq
->tg
->rt_bandwidth
.rt_period
);
459 typedef struct task_group
*rt_rq_iter_t
;
461 static inline struct task_group
*next_task_group(struct task_group
*tg
)
464 tg
= list_entry_rcu(tg
->list
.next
,
465 typeof(struct task_group
), list
);
466 } while (&tg
->list
!= &task_groups
&& task_group_is_autogroup(tg
));
468 if (&tg
->list
== &task_groups
)
474 #define for_each_rt_rq(rt_rq, iter, rq) \
475 for (iter = container_of(&task_groups, typeof(*iter), list); \
476 (iter = next_task_group(iter)) && \
477 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
479 #define for_each_sched_rt_entity(rt_se) \
480 for (; rt_se; rt_se = rt_se->parent)
482 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
487 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
);
488 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
);
490 static void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
492 struct task_struct
*curr
= rq_of_rt_rq(rt_rq
)->curr
;
493 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
494 struct sched_rt_entity
*rt_se
;
496 int cpu
= cpu_of(rq
);
498 rt_se
= rt_rq
->tg
->rt_se
[cpu
];
500 if (rt_rq
->rt_nr_running
) {
502 enqueue_top_rt_rq(rt_rq
);
503 else if (!on_rt_rq(rt_se
))
504 enqueue_rt_entity(rt_se
, 0);
506 if (rt_rq
->highest_prio
.curr
< curr
->prio
)
511 static void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
513 struct sched_rt_entity
*rt_se
;
514 int cpu
= cpu_of(rq_of_rt_rq(rt_rq
));
516 rt_se
= rt_rq
->tg
->rt_se
[cpu
];
519 dequeue_top_rt_rq(rt_rq
);
520 else if (on_rt_rq(rt_se
))
521 dequeue_rt_entity(rt_se
, 0);
524 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
526 return rt_rq
->rt_throttled
&& !rt_rq
->rt_nr_boosted
;
529 static int rt_se_boosted(struct sched_rt_entity
*rt_se
)
531 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
532 struct task_struct
*p
;
535 return !!rt_rq
->rt_nr_boosted
;
537 p
= rt_task_of(rt_se
);
538 return p
->prio
!= p
->normal_prio
;
542 static inline const struct cpumask
*sched_rt_period_mask(void)
544 return this_rq()->rd
->span
;
547 static inline const struct cpumask
*sched_rt_period_mask(void)
549 return cpu_online_mask
;
554 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
556 return container_of(rt_b
, struct task_group
, rt_bandwidth
)->rt_rq
[cpu
];
559 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
561 return &rt_rq
->tg
->rt_bandwidth
;
564 #else /* !CONFIG_RT_GROUP_SCHED */
566 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
568 return rt_rq
->rt_runtime
;
571 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
573 return ktime_to_ns(def_rt_bandwidth
.rt_period
);
576 typedef struct rt_rq
*rt_rq_iter_t
;
578 #define for_each_rt_rq(rt_rq, iter, rq) \
579 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
581 #define for_each_sched_rt_entity(rt_se) \
582 for (; rt_se; rt_se = NULL)
584 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
589 static inline void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
591 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
593 if (!rt_rq
->rt_nr_running
)
596 enqueue_top_rt_rq(rt_rq
);
600 static inline void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
602 dequeue_top_rt_rq(rt_rq
);
605 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
607 return rt_rq
->rt_throttled
;
610 static inline const struct cpumask
*sched_rt_period_mask(void)
612 return cpu_online_mask
;
616 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
618 return &cpu_rq(cpu
)->rt
;
621 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
623 return &def_rt_bandwidth
;
626 #endif /* CONFIG_RT_GROUP_SCHED */
628 bool sched_rt_bandwidth_account(struct rt_rq
*rt_rq
)
630 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
632 return (hrtimer_active(&rt_b
->rt_period_timer
) ||
633 rt_rq
->rt_time
< rt_b
->rt_runtime
);
638 * We ran out of runtime, see if we can borrow some from our neighbours.
640 static void do_balance_runtime(struct rt_rq
*rt_rq
)
642 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
643 struct root_domain
*rd
= rq_of_rt_rq(rt_rq
)->rd
;
647 weight
= cpumask_weight(rd
->span
);
649 raw_spin_lock(&rt_b
->rt_runtime_lock
);
650 rt_period
= ktime_to_ns(rt_b
->rt_period
);
651 for_each_cpu(i
, rd
->span
) {
652 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
658 raw_spin_lock(&iter
->rt_runtime_lock
);
660 * Either all rqs have inf runtime and there's nothing to steal
661 * or __disable_runtime() below sets a specific rq to inf to
662 * indicate its been disabled and disalow stealing.
664 if (iter
->rt_runtime
== RUNTIME_INF
)
668 * From runqueues with spare time, take 1/n part of their
669 * spare time, but no more than our period.
671 diff
= iter
->rt_runtime
- iter
->rt_time
;
673 diff
= div_u64((u64
)diff
, weight
);
674 if (rt_rq
->rt_runtime
+ diff
> rt_period
)
675 diff
= rt_period
- rt_rq
->rt_runtime
;
676 iter
->rt_runtime
-= diff
;
677 rt_rq
->rt_runtime
+= diff
;
678 if (rt_rq
->rt_runtime
== rt_period
) {
679 raw_spin_unlock(&iter
->rt_runtime_lock
);
684 raw_spin_unlock(&iter
->rt_runtime_lock
);
686 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
690 * Ensure this RQ takes back all the runtime it lend to its neighbours.
692 static void __disable_runtime(struct rq
*rq
)
694 struct root_domain
*rd
= rq
->rd
;
698 if (unlikely(!scheduler_running
))
701 for_each_rt_rq(rt_rq
, iter
, rq
) {
702 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
706 raw_spin_lock(&rt_b
->rt_runtime_lock
);
707 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
709 * Either we're all inf and nobody needs to borrow, or we're
710 * already disabled and thus have nothing to do, or we have
711 * exactly the right amount of runtime to take out.
713 if (rt_rq
->rt_runtime
== RUNTIME_INF
||
714 rt_rq
->rt_runtime
== rt_b
->rt_runtime
)
716 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
719 * Calculate the difference between what we started out with
720 * and what we current have, that's the amount of runtime
721 * we lend and now have to reclaim.
723 want
= rt_b
->rt_runtime
- rt_rq
->rt_runtime
;
726 * Greedy reclaim, take back as much as we can.
728 for_each_cpu(i
, rd
->span
) {
729 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
733 * Can't reclaim from ourselves or disabled runqueues.
735 if (iter
== rt_rq
|| iter
->rt_runtime
== RUNTIME_INF
)
738 raw_spin_lock(&iter
->rt_runtime_lock
);
740 diff
= min_t(s64
, iter
->rt_runtime
, want
);
741 iter
->rt_runtime
-= diff
;
744 iter
->rt_runtime
-= want
;
747 raw_spin_unlock(&iter
->rt_runtime_lock
);
753 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
755 * We cannot be left wanting - that would mean some runtime
756 * leaked out of the system.
761 * Disable all the borrow logic by pretending we have inf
762 * runtime - in which case borrowing doesn't make sense.
764 rt_rq
->rt_runtime
= RUNTIME_INF
;
765 rt_rq
->rt_throttled
= 0;
766 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
767 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
769 /* Make rt_rq available for pick_next_task() */
770 sched_rt_rq_enqueue(rt_rq
);
774 static void __enable_runtime(struct rq
*rq
)
779 if (unlikely(!scheduler_running
))
783 * Reset each runqueue's bandwidth settings
785 for_each_rt_rq(rt_rq
, iter
, rq
) {
786 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
788 raw_spin_lock(&rt_b
->rt_runtime_lock
);
789 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
790 rt_rq
->rt_runtime
= rt_b
->rt_runtime
;
792 rt_rq
->rt_throttled
= 0;
793 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
794 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
798 static void balance_runtime(struct rt_rq
*rt_rq
)
800 if (!sched_feat(RT_RUNTIME_SHARE
))
803 if (rt_rq
->rt_time
> rt_rq
->rt_runtime
) {
804 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
805 do_balance_runtime(rt_rq
);
806 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
809 #else /* !CONFIG_SMP */
810 static inline void balance_runtime(struct rt_rq
*rt_rq
) {}
811 #endif /* CONFIG_SMP */
813 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
)
815 int i
, idle
= 1, throttled
= 0;
816 const struct cpumask
*span
;
818 span
= sched_rt_period_mask();
819 #ifdef CONFIG_RT_GROUP_SCHED
821 * FIXME: isolated CPUs should really leave the root task group,
822 * whether they are isolcpus or were isolated via cpusets, lest
823 * the timer run on a CPU which does not service all runqueues,
824 * potentially leaving other CPUs indefinitely throttled. If
825 * isolation is really required, the user will turn the throttle
826 * off to kill the perturbations it causes anyway. Meanwhile,
827 * this maintains functionality for boot and/or troubleshooting.
829 if (rt_b
== &root_task_group
.rt_bandwidth
)
830 span
= cpu_online_mask
;
832 for_each_cpu(i
, span
) {
834 struct rt_rq
*rt_rq
= sched_rt_period_rt_rq(rt_b
, i
);
835 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
839 * When span == cpu_online_mask, taking each rq->lock
840 * can be time-consuming. Try to avoid it when possible.
842 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
843 if (!sched_feat(RT_RUNTIME_SHARE
) && rt_rq
->rt_runtime
!= RUNTIME_INF
)
844 rt_rq
->rt_runtime
= rt_b
->rt_runtime
;
845 skip
= !rt_rq
->rt_time
&& !rt_rq
->rt_nr_running
;
846 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
850 raw_spin_lock(&rq
->lock
);
853 if (rt_rq
->rt_time
) {
856 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
857 if (rt_rq
->rt_throttled
)
858 balance_runtime(rt_rq
);
859 runtime
= rt_rq
->rt_runtime
;
860 rt_rq
->rt_time
-= min(rt_rq
->rt_time
, overrun
*runtime
);
861 if (rt_rq
->rt_throttled
&& rt_rq
->rt_time
< runtime
) {
862 rt_rq
->rt_throttled
= 0;
866 * When we're idle and a woken (rt) task is
867 * throttled check_preempt_curr() will set
868 * skip_update and the time between the wakeup
869 * and this unthrottle will get accounted as
872 if (rt_rq
->rt_nr_running
&& rq
->curr
== rq
->idle
)
873 rq_clock_skip_update(rq
, false);
875 if (rt_rq
->rt_time
|| rt_rq
->rt_nr_running
)
877 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
878 } else if (rt_rq
->rt_nr_running
) {
880 if (!rt_rq_throttled(rt_rq
))
883 if (rt_rq
->rt_throttled
)
887 sched_rt_rq_enqueue(rt_rq
);
888 raw_spin_unlock(&rq
->lock
);
891 if (!throttled
&& (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
))
897 static inline int rt_se_prio(struct sched_rt_entity
*rt_se
)
899 #ifdef CONFIG_RT_GROUP_SCHED
900 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
903 return rt_rq
->highest_prio
.curr
;
906 return rt_task_of(rt_se
)->prio
;
909 static int sched_rt_runtime_exceeded(struct rt_rq
*rt_rq
)
911 u64 runtime
= sched_rt_runtime(rt_rq
);
913 if (rt_rq
->rt_throttled
)
914 return rt_rq_throttled(rt_rq
);
916 if (runtime
>= sched_rt_period(rt_rq
))
919 balance_runtime(rt_rq
);
920 runtime
= sched_rt_runtime(rt_rq
);
921 if (runtime
== RUNTIME_INF
)
924 if (rt_rq
->rt_time
> runtime
) {
925 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
928 * Don't actually throttle groups that have no runtime assigned
929 * but accrue some time due to boosting.
931 if (likely(rt_b
->rt_runtime
)) {
932 rt_rq
->rt_throttled
= 1;
933 printk_deferred_once("sched: RT throttling activated\n");
936 * In case we did anyway, make it go away,
937 * replenishment is a joke, since it will replenish us
943 if (rt_rq_throttled(rt_rq
)) {
944 sched_rt_rq_dequeue(rt_rq
);
953 * Update the current task's runtime statistics. Skip current tasks that
954 * are not in our scheduling class.
956 static void update_curr_rt(struct rq
*rq
)
958 struct task_struct
*curr
= rq
->curr
;
959 struct sched_rt_entity
*rt_se
= &curr
->rt
;
962 if (curr
->sched_class
!= &rt_sched_class
)
965 delta_exec
= rq_clock_task(rq
) - curr
->se
.exec_start
;
966 if (unlikely((s64
)delta_exec
<= 0))
969 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
970 cpufreq_update_util(rq
, SCHED_CPUFREQ_RT
);
972 schedstat_set(curr
->se
.statistics
.exec_max
,
973 max(curr
->se
.statistics
.exec_max
, delta_exec
));
975 curr
->se
.sum_exec_runtime
+= delta_exec
;
976 account_group_exec_runtime(curr
, delta_exec
);
978 curr
->se
.exec_start
= rq_clock_task(rq
);
979 cpuacct_charge(curr
, delta_exec
);
981 sched_rt_avg_update(rq
, delta_exec
);
983 if (!rt_bandwidth_enabled())
986 for_each_sched_rt_entity(rt_se
) {
987 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
989 if (sched_rt_runtime(rt_rq
) != RUNTIME_INF
) {
990 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
991 rt_rq
->rt_time
+= delta_exec
;
992 if (sched_rt_runtime_exceeded(rt_rq
))
994 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
1000 dequeue_top_rt_rq(struct rt_rq
*rt_rq
)
1002 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1004 BUG_ON(&rq
->rt
!= rt_rq
);
1006 if (!rt_rq
->rt_queued
)
1009 BUG_ON(!rq
->nr_running
);
1011 sub_nr_running(rq
, rt_rq
->rt_nr_running
);
1012 rt_rq
->rt_queued
= 0;
1016 enqueue_top_rt_rq(struct rt_rq
*rt_rq
)
1018 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1020 BUG_ON(&rq
->rt
!= rt_rq
);
1022 if (rt_rq
->rt_queued
)
1024 if (rt_rq_throttled(rt_rq
) || !rt_rq
->rt_nr_running
)
1027 add_nr_running(rq
, rt_rq
->rt_nr_running
);
1028 rt_rq
->rt_queued
= 1;
1031 #if defined CONFIG_SMP
1034 inc_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
)
1036 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1038 #ifdef CONFIG_RT_GROUP_SCHED
1040 * Change rq's cpupri only if rt_rq is the top queue.
1042 if (&rq
->rt
!= rt_rq
)
1045 if (rq
->online
&& prio
< prev_prio
)
1046 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, prio
);
1050 dec_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
)
1052 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1054 #ifdef CONFIG_RT_GROUP_SCHED
1056 * Change rq's cpupri only if rt_rq is the top queue.
1058 if (&rq
->rt
!= rt_rq
)
1061 if (rq
->online
&& rt_rq
->highest_prio
.curr
!= prev_prio
)
1062 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rt_rq
->highest_prio
.curr
);
1065 #else /* CONFIG_SMP */
1068 void inc_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
) {}
1070 void dec_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
) {}
1072 #endif /* CONFIG_SMP */
1074 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1076 inc_rt_prio(struct rt_rq
*rt_rq
, int prio
)
1078 int prev_prio
= rt_rq
->highest_prio
.curr
;
1080 if (prio
< prev_prio
)
1081 rt_rq
->highest_prio
.curr
= prio
;
1083 inc_rt_prio_smp(rt_rq
, prio
, prev_prio
);
1087 dec_rt_prio(struct rt_rq
*rt_rq
, int prio
)
1089 int prev_prio
= rt_rq
->highest_prio
.curr
;
1091 if (rt_rq
->rt_nr_running
) {
1093 WARN_ON(prio
< prev_prio
);
1096 * This may have been our highest task, and therefore
1097 * we may have some recomputation to do
1099 if (prio
== prev_prio
) {
1100 struct rt_prio_array
*array
= &rt_rq
->active
;
1102 rt_rq
->highest_prio
.curr
=
1103 sched_find_first_bit(array
->bitmap
);
1107 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
1109 dec_rt_prio_smp(rt_rq
, prio
, prev_prio
);
1114 static inline void inc_rt_prio(struct rt_rq
*rt_rq
, int prio
) {}
1115 static inline void dec_rt_prio(struct rt_rq
*rt_rq
, int prio
) {}
1117 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1119 #ifdef CONFIG_RT_GROUP_SCHED
1122 inc_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1124 if (rt_se_boosted(rt_se
))
1125 rt_rq
->rt_nr_boosted
++;
1128 start_rt_bandwidth(&rt_rq
->tg
->rt_bandwidth
);
1132 dec_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1134 if (rt_se_boosted(rt_se
))
1135 rt_rq
->rt_nr_boosted
--;
1137 WARN_ON(!rt_rq
->rt_nr_running
&& rt_rq
->rt_nr_boosted
);
1140 #else /* CONFIG_RT_GROUP_SCHED */
1143 inc_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1145 start_rt_bandwidth(&def_rt_bandwidth
);
1149 void dec_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
) {}
1151 #endif /* CONFIG_RT_GROUP_SCHED */
1154 unsigned int rt_se_nr_running(struct sched_rt_entity
*rt_se
)
1156 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1159 return group_rq
->rt_nr_running
;
1165 unsigned int rt_se_rr_nr_running(struct sched_rt_entity
*rt_se
)
1167 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1168 struct task_struct
*tsk
;
1171 return group_rq
->rr_nr_running
;
1173 tsk
= rt_task_of(rt_se
);
1175 return (tsk
->policy
== SCHED_RR
) ? 1 : 0;
1179 void inc_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1181 int prio
= rt_se_prio(rt_se
);
1183 WARN_ON(!rt_prio(prio
));
1184 rt_rq
->rt_nr_running
+= rt_se_nr_running(rt_se
);
1185 rt_rq
->rr_nr_running
+= rt_se_rr_nr_running(rt_se
);
1187 inc_rt_prio(rt_rq
, prio
);
1188 inc_rt_migration(rt_se
, rt_rq
);
1189 inc_rt_group(rt_se
, rt_rq
);
1193 void dec_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1195 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
1196 WARN_ON(!rt_rq
->rt_nr_running
);
1197 rt_rq
->rt_nr_running
-= rt_se_nr_running(rt_se
);
1198 rt_rq
->rr_nr_running
-= rt_se_rr_nr_running(rt_se
);
1200 dec_rt_prio(rt_rq
, rt_se_prio(rt_se
));
1201 dec_rt_migration(rt_se
, rt_rq
);
1202 dec_rt_group(rt_se
, rt_rq
);
1206 * Change rt_se->run_list location unless SAVE && !MOVE
1208 * assumes ENQUEUE/DEQUEUE flags match
1210 static inline bool move_entity(unsigned int flags
)
1212 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) == DEQUEUE_SAVE
)
1218 static void __delist_rt_entity(struct sched_rt_entity
*rt_se
, struct rt_prio_array
*array
)
1220 list_del_init(&rt_se
->run_list
);
1222 if (list_empty(array
->queue
+ rt_se_prio(rt_se
)))
1223 __clear_bit(rt_se_prio(rt_se
), array
->bitmap
);
1228 static void __enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1230 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
1231 struct rt_prio_array
*array
= &rt_rq
->active
;
1232 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1233 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
1236 * Don't enqueue the group if its throttled, or when empty.
1237 * The latter is a consequence of the former when a child group
1238 * get throttled and the current group doesn't have any other
1241 if (group_rq
&& (rt_rq_throttled(group_rq
) || !group_rq
->rt_nr_running
)) {
1243 __delist_rt_entity(rt_se
, array
);
1247 if (move_entity(flags
)) {
1248 WARN_ON_ONCE(rt_se
->on_list
);
1249 if (flags
& ENQUEUE_HEAD
)
1250 list_add(&rt_se
->run_list
, queue
);
1252 list_add_tail(&rt_se
->run_list
, queue
);
1254 __set_bit(rt_se_prio(rt_se
), array
->bitmap
);
1259 inc_rt_tasks(rt_se
, rt_rq
);
1262 static void __dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1264 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
1265 struct rt_prio_array
*array
= &rt_rq
->active
;
1267 if (move_entity(flags
)) {
1268 WARN_ON_ONCE(!rt_se
->on_list
);
1269 __delist_rt_entity(rt_se
, array
);
1273 dec_rt_tasks(rt_se
, rt_rq
);
1277 * Because the prio of an upper entry depends on the lower
1278 * entries, we must remove entries top - down.
1280 static void dequeue_rt_stack(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1282 struct sched_rt_entity
*back
= NULL
;
1284 for_each_sched_rt_entity(rt_se
) {
1289 dequeue_top_rt_rq(rt_rq_of_se(back
));
1291 for (rt_se
= back
; rt_se
; rt_se
= rt_se
->back
) {
1292 if (on_rt_rq(rt_se
))
1293 __dequeue_rt_entity(rt_se
, flags
);
1297 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1299 struct rq
*rq
= rq_of_rt_se(rt_se
);
1301 dequeue_rt_stack(rt_se
, flags
);
1302 for_each_sched_rt_entity(rt_se
)
1303 __enqueue_rt_entity(rt_se
, flags
);
1304 enqueue_top_rt_rq(&rq
->rt
);
1307 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1309 struct rq
*rq
= rq_of_rt_se(rt_se
);
1311 dequeue_rt_stack(rt_se
, flags
);
1313 for_each_sched_rt_entity(rt_se
) {
1314 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
1316 if (rt_rq
&& rt_rq
->rt_nr_running
)
1317 __enqueue_rt_entity(rt_se
, flags
);
1319 enqueue_top_rt_rq(&rq
->rt
);
1323 * Adding/removing a task to/from a priority array:
1326 enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1328 struct sched_rt_entity
*rt_se
= &p
->rt
;
1330 schedtune_enqueue_task(p
, cpu_of(rq
));
1332 if (flags
& ENQUEUE_WAKEUP
)
1335 enqueue_rt_entity(rt_se
, flags
);
1336 walt_inc_cumulative_runnable_avg(rq
, p
);
1338 if (!task_current(rq
, p
) && p
->nr_cpus_allowed
> 1)
1339 enqueue_pushable_task(rq
, p
);
1342 static void dequeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1344 struct sched_rt_entity
*rt_se
= &p
->rt
;
1346 schedtune_dequeue_task(p
, cpu_of(rq
));
1349 dequeue_rt_entity(rt_se
, flags
);
1350 walt_dec_cumulative_runnable_avg(rq
, p
);
1352 dequeue_pushable_task(rq
, p
);
1356 * Put task to the head or the end of the run list without the overhead of
1357 * dequeue followed by enqueue.
1360 requeue_rt_entity(struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
, int head
)
1362 if (on_rt_rq(rt_se
)) {
1363 struct rt_prio_array
*array
= &rt_rq
->active
;
1364 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
1367 list_move(&rt_se
->run_list
, queue
);
1369 list_move_tail(&rt_se
->run_list
, queue
);
1373 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int head
)
1375 struct sched_rt_entity
*rt_se
= &p
->rt
;
1376 struct rt_rq
*rt_rq
;
1378 for_each_sched_rt_entity(rt_se
) {
1379 rt_rq
= rt_rq_of_se(rt_se
);
1380 requeue_rt_entity(rt_rq
, rt_se
, head
);
1384 static void yield_task_rt(struct rq
*rq
)
1386 requeue_task_rt(rq
, rq
->curr
, 0);
1390 static int find_lowest_rq(struct task_struct
*task
);
1393 select_task_rq_rt(struct task_struct
*p
, int cpu
, int sd_flag
, int flags
,
1394 int sibling_count_hint
)
1396 struct task_struct
*curr
;
1399 /* For anything but wake ups, just return the task_cpu */
1400 if (sd_flag
!= SD_BALANCE_WAKE
&& sd_flag
!= SD_BALANCE_FORK
)
1406 curr
= READ_ONCE(rq
->curr
); /* unlocked access */
1409 * If the current task on @p's runqueue is an RT task, then
1410 * try to see if we can wake this RT task up on another
1411 * runqueue. Otherwise simply start this RT task
1412 * on its current runqueue.
1414 * We want to avoid overloading runqueues. If the woken
1415 * task is a higher priority, then it will stay on this CPU
1416 * and the lower prio task should be moved to another CPU.
1417 * Even though this will probably make the lower prio task
1418 * lose its cache, we do not want to bounce a higher task
1419 * around just because it gave up its CPU, perhaps for a
1422 * For equal prio tasks, we just let the scheduler sort it out.
1424 * Otherwise, just let it ride on the affined RQ and the
1425 * post-schedule router will push the preempted task away
1427 * This test is optimistic, if we get it wrong the load-balancer
1428 * will have to sort it out.
1430 if (curr
&& unlikely(rt_task(curr
)) &&
1431 (curr
->nr_cpus_allowed
< 2 ||
1432 curr
->prio
<= p
->prio
)) {
1433 int target
= find_lowest_rq(p
);
1436 * Don't bother moving it if the destination CPU is
1437 * not running a lower priority task.
1440 p
->prio
< cpu_rq(target
)->rt
.highest_prio
.curr
)
1449 static void check_preempt_equal_prio(struct rq
*rq
, struct task_struct
*p
)
1452 * Current can't be migrated, useless to reschedule,
1453 * let's hope p can move out.
1455 if (rq
->curr
->nr_cpus_allowed
== 1 ||
1456 !cpupri_find(&rq
->rd
->cpupri
, rq
->curr
, NULL
))
1460 * p is migratable, so let's not schedule it and
1461 * see if it is pushed or pulled somewhere else.
1463 if (p
->nr_cpus_allowed
!= 1
1464 && cpupri_find(&rq
->rd
->cpupri
, p
, NULL
))
1468 * There appears to be other cpus that can accept
1469 * current and none to run 'p', so lets reschedule
1470 * to try and push current away:
1472 requeue_task_rt(rq
, p
, 1);
1476 #endif /* CONFIG_SMP */
1479 * Preempt the current task with a newly woken task if needed:
1481 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1483 if (p
->prio
< rq
->curr
->prio
) {
1492 * - the newly woken task is of equal priority to the current task
1493 * - the newly woken task is non-migratable while current is migratable
1494 * - current will be preempted on the next reschedule
1496 * we should check to see if current can readily move to a different
1497 * cpu. If so, we will reschedule to allow the push logic to try
1498 * to move current somewhere else, making room for our non-migratable
1501 if (p
->prio
== rq
->curr
->prio
&& !test_tsk_need_resched(rq
->curr
))
1502 check_preempt_equal_prio(rq
, p
);
1506 static struct sched_rt_entity
*pick_next_rt_entity(struct rq
*rq
,
1507 struct rt_rq
*rt_rq
)
1509 struct rt_prio_array
*array
= &rt_rq
->active
;
1510 struct sched_rt_entity
*next
= NULL
;
1511 struct list_head
*queue
;
1514 idx
= sched_find_first_bit(array
->bitmap
);
1515 BUG_ON(idx
>= MAX_RT_PRIO
);
1517 queue
= array
->queue
+ idx
;
1518 next
= list_entry(queue
->next
, struct sched_rt_entity
, run_list
);
1523 static struct task_struct
*_pick_next_task_rt(struct rq
*rq
)
1525 struct sched_rt_entity
*rt_se
;
1526 struct task_struct
*p
;
1527 struct rt_rq
*rt_rq
= &rq
->rt
;
1530 rt_se
= pick_next_rt_entity(rq
, rt_rq
);
1532 rt_rq
= group_rt_rq(rt_se
);
1535 p
= rt_task_of(rt_se
);
1536 p
->se
.exec_start
= rq_clock_task(rq
);
1541 extern int update_rt_rq_load_avg(u64 now
, int cpu
, struct rt_rq
*rt_rq
, int running
);
1543 static struct task_struct
*
1544 pick_next_task_rt(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
1546 struct task_struct
*p
;
1547 struct rt_rq
*rt_rq
= &rq
->rt
;
1549 if (need_pull_rt_task(rq
, prev
)) {
1551 * This is OK, because current is on_cpu, which avoids it being
1552 * picked for load-balance and preemption/IRQs are still
1553 * disabled avoiding further scheduler activity on it and we're
1554 * being very careful to re-start the picking loop.
1556 rq_unpin_lock(rq
, rf
);
1558 rq_repin_lock(rq
, rf
);
1560 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1561 * means a dl or stop task can slip in, in which case we need
1562 * to re-start task selection.
1564 if (unlikely((rq
->stop
&& task_on_rq_queued(rq
->stop
)) ||
1565 rq
->dl
.dl_nr_running
))
1570 * We may dequeue prev's rt_rq in put_prev_task().
1571 * So, we update time before rt_nr_running check.
1573 if (prev
->sched_class
== &rt_sched_class
)
1576 if (!rt_rq
->rt_queued
)
1579 put_prev_task(rq
, prev
);
1581 p
= _pick_next_task_rt(rq
);
1583 /* The running task is never eligible for pushing */
1584 dequeue_pushable_task(rq
, p
);
1586 queue_push_tasks(rq
);
1589 update_rt_rq_load_avg(rq_clock_task(rq
), cpu_of(rq
), rt_rq
,
1590 rq
->curr
->sched_class
== &rt_sched_class
);
1595 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
1599 update_rt_rq_load_avg(rq_clock_task(rq
), cpu_of(rq
), &rq
->rt
, 1);
1602 * The previous task needs to be made eligible for pushing
1603 * if it is still active
1605 if (on_rt_rq(&p
->rt
) && p
->nr_cpus_allowed
> 1)
1606 enqueue_pushable_task(rq
, p
);
1611 /* Only try algorithms three times */
1612 #define RT_MAX_TRIES 3
1614 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
1616 if (!task_running(rq
, p
) &&
1617 cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
1623 * Return the highest pushable rq's task, which is suitable to be executed
1624 * on the cpu, NULL otherwise
1626 static struct task_struct
*pick_highest_pushable_task(struct rq
*rq
, int cpu
)
1628 struct plist_head
*head
= &rq
->rt
.pushable_tasks
;
1629 struct task_struct
*p
;
1631 if (!has_pushable_tasks(rq
))
1634 plist_for_each_entry(p
, head
, pushable_tasks
) {
1635 if (pick_rt_task(rq
, p
, cpu
))
1642 static DEFINE_PER_CPU(cpumask_var_t
, local_cpu_mask
);
1644 static int find_lowest_rq(struct task_struct
*task
)
1646 struct sched_domain
*sd
;
1647 struct cpumask
*lowest_mask
= this_cpu_cpumask_var_ptr(local_cpu_mask
);
1648 int this_cpu
= smp_processor_id();
1649 int cpu
= task_cpu(task
);
1651 /* Make sure the mask is initialized first */
1652 if (unlikely(!lowest_mask
))
1655 if (task
->nr_cpus_allowed
== 1)
1656 return -1; /* No other targets possible */
1658 if (!cpupri_find(&task_rq(task
)->rd
->cpupri
, task
, lowest_mask
))
1659 return -1; /* No targets found */
1662 * At this point we have built a mask of cpus representing the
1663 * lowest priority tasks in the system. Now we want to elect
1664 * the best one based on our affinity and topology.
1666 * We prioritize the last cpu that the task executed on since
1667 * it is most likely cache-hot in that location.
1669 if (cpumask_test_cpu(cpu
, lowest_mask
))
1673 * Otherwise, we consult the sched_domains span maps to figure
1674 * out which cpu is logically closest to our hot cache data.
1676 if (!cpumask_test_cpu(this_cpu
, lowest_mask
))
1677 this_cpu
= -1; /* Skip this_cpu opt if not among lowest */
1680 for_each_domain(cpu
, sd
) {
1681 if (sd
->flags
& SD_WAKE_AFFINE
) {
1685 * "this_cpu" is cheaper to preempt than a
1688 if (this_cpu
!= -1 &&
1689 cpumask_test_cpu(this_cpu
, sched_domain_span(sd
))) {
1694 best_cpu
= cpumask_first_and(lowest_mask
,
1695 sched_domain_span(sd
));
1696 if (best_cpu
< nr_cpu_ids
) {
1705 * And finally, if there were no matches within the domains
1706 * just give the caller *something* to work with from the compatible
1712 cpu
= cpumask_any(lowest_mask
);
1713 if (cpu
< nr_cpu_ids
)
1718 /* Will lock the rq it finds */
1719 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
1721 struct rq
*lowest_rq
= NULL
;
1725 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
1726 cpu
= find_lowest_rq(task
);
1728 if ((cpu
== -1) || (cpu
== rq
->cpu
))
1731 lowest_rq
= cpu_rq(cpu
);
1733 if (lowest_rq
->rt
.highest_prio
.curr
<= task
->prio
) {
1735 * Target rq has tasks of equal or higher priority,
1736 * retrying does not release any lock and is unlikely
1737 * to yield a different result.
1743 /* if the prio of this runqueue changed, try again */
1744 if (double_lock_balance(rq
, lowest_rq
)) {
1746 * We had to unlock the run queue. In
1747 * the mean time, task could have
1748 * migrated already or had its affinity changed.
1749 * Also make sure that it wasn't scheduled on its rq.
1751 if (unlikely(task_rq(task
) != rq
||
1752 !cpumask_test_cpu(lowest_rq
->cpu
, &task
->cpus_allowed
) ||
1753 task_running(rq
, task
) ||
1755 !task_on_rq_queued(task
))) {
1757 double_unlock_balance(rq
, lowest_rq
);
1763 /* If this rq is still suitable use it. */
1764 if (lowest_rq
->rt
.highest_prio
.curr
> task
->prio
)
1768 double_unlock_balance(rq
, lowest_rq
);
1775 static struct task_struct
*pick_next_pushable_task(struct rq
*rq
)
1777 struct task_struct
*p
;
1779 if (!has_pushable_tasks(rq
))
1782 p
= plist_first_entry(&rq
->rt
.pushable_tasks
,
1783 struct task_struct
, pushable_tasks
);
1785 BUG_ON(rq
->cpu
!= task_cpu(p
));
1786 BUG_ON(task_current(rq
, p
));
1787 BUG_ON(p
->nr_cpus_allowed
<= 1);
1789 BUG_ON(!task_on_rq_queued(p
));
1790 BUG_ON(!rt_task(p
));
1796 * If the current CPU has more than one RT task, see if the non
1797 * running task can migrate over to a CPU that is running a task
1798 * of lesser priority.
1800 static int push_rt_task(struct rq
*rq
)
1802 struct task_struct
*next_task
;
1803 struct rq
*lowest_rq
;
1806 if (!rq
->rt
.overloaded
)
1809 next_task
= pick_next_pushable_task(rq
);
1814 if (unlikely(next_task
== rq
->curr
)) {
1820 * It's possible that the next_task slipped in of
1821 * higher priority than current. If that's the case
1822 * just reschedule current.
1824 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
1829 /* We might release rq lock */
1830 get_task_struct(next_task
);
1832 /* find_lock_lowest_rq locks the rq if found */
1833 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
1835 struct task_struct
*task
;
1837 * find_lock_lowest_rq releases rq->lock
1838 * so it is possible that next_task has migrated.
1840 * We need to make sure that the task is still on the same
1841 * run-queue and is also still the next task eligible for
1844 task
= pick_next_pushable_task(rq
);
1845 if (task
== next_task
) {
1847 * The task hasn't migrated, and is still the next
1848 * eligible task, but we failed to find a run-queue
1849 * to push it to. Do not retry in this case, since
1850 * other cpus will pull from us when ready.
1856 /* No more tasks, just exit */
1860 * Something has shifted, try again.
1862 put_task_struct(next_task
);
1867 deactivate_task(rq
, next_task
, 0);
1868 next_task
->on_rq
= TASK_ON_RQ_MIGRATING
;
1869 set_task_cpu(next_task
, lowest_rq
->cpu
);
1870 next_task
->on_rq
= TASK_ON_RQ_QUEUED
;
1871 activate_task(lowest_rq
, next_task
, 0);
1874 resched_curr(lowest_rq
);
1876 double_unlock_balance(rq
, lowest_rq
);
1879 put_task_struct(next_task
);
1884 static void push_rt_tasks(struct rq
*rq
)
1886 /* push_rt_task will return true if it moved an RT */
1887 while (push_rt_task(rq
))
1891 #ifdef HAVE_RT_PUSH_IPI
1894 * When a high priority task schedules out from a CPU and a lower priority
1895 * task is scheduled in, a check is made to see if there's any RT tasks
1896 * on other CPUs that are waiting to run because a higher priority RT task
1897 * is currently running on its CPU. In this case, the CPU with multiple RT
1898 * tasks queued on it (overloaded) needs to be notified that a CPU has opened
1899 * up that may be able to run one of its non-running queued RT tasks.
1901 * All CPUs with overloaded RT tasks need to be notified as there is currently
1902 * no way to know which of these CPUs have the highest priority task waiting
1903 * to run. Instead of trying to take a spinlock on each of these CPUs,
1904 * which has shown to cause large latency when done on machines with many
1905 * CPUs, sending an IPI to the CPUs to have them push off the overloaded
1906 * RT tasks waiting to run.
1908 * Just sending an IPI to each of the CPUs is also an issue, as on large
1909 * count CPU machines, this can cause an IPI storm on a CPU, especially
1910 * if its the only CPU with multiple RT tasks queued, and a large number
1911 * of CPUs scheduling a lower priority task at the same time.
1913 * Each root domain has its own irq work function that can iterate over
1914 * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT
1915 * tassk must be checked if there's one or many CPUs that are lowering
1916 * their priority, there's a single irq work iterator that will try to
1917 * push off RT tasks that are waiting to run.
1919 * When a CPU schedules a lower priority task, it will kick off the
1920 * irq work iterator that will jump to each CPU with overloaded RT tasks.
1921 * As it only takes the first CPU that schedules a lower priority task
1922 * to start the process, the rto_start variable is incremented and if
1923 * the atomic result is one, then that CPU will try to take the rto_lock.
1924 * This prevents high contention on the lock as the process handles all
1925 * CPUs scheduling lower priority tasks.
1927 * All CPUs that are scheduling a lower priority task will increment the
1928 * rt_loop_next variable. This will make sure that the irq work iterator
1929 * checks all RT overloaded CPUs whenever a CPU schedules a new lower
1930 * priority task, even if the iterator is in the middle of a scan. Incrementing
1931 * the rt_loop_next will cause the iterator to perform another scan.
1934 static int rto_next_cpu(struct root_domain
*rd
)
1940 * When starting the IPI RT pushing, the rto_cpu is set to -1,
1941 * rt_next_cpu() will simply return the first CPU found in
1944 * If rto_next_cpu() is called with rto_cpu is a valid cpu, it
1945 * will return the next CPU found in the rto_mask.
1947 * If there are no more CPUs left in the rto_mask, then a check is made
1948 * against rto_loop and rto_loop_next. rto_loop is only updated with
1949 * the rto_lock held, but any CPU may increment the rto_loop_next
1950 * without any locking.
1954 /* When rto_cpu is -1 this acts like cpumask_first() */
1955 cpu
= cpumask_next(rd
->rto_cpu
, rd
->rto_mask
);
1959 if (cpu
< nr_cpu_ids
)
1965 * ACQUIRE ensures we see the @rto_mask changes
1966 * made prior to the @next value observed.
1968 * Matches WMB in rt_set_overload().
1970 next
= atomic_read_acquire(&rd
->rto_loop_next
);
1972 if (rd
->rto_loop
== next
)
1975 rd
->rto_loop
= next
;
1981 static inline bool rto_start_trylock(atomic_t
*v
)
1983 return !atomic_cmpxchg_acquire(v
, 0, 1);
1986 static inline void rto_start_unlock(atomic_t
*v
)
1988 atomic_set_release(v
, 0);
1991 static void tell_cpu_to_push(struct rq
*rq
)
1995 /* Keep the loop going if the IPI is currently active */
1996 atomic_inc(&rq
->rd
->rto_loop_next
);
1998 /* Only one CPU can initiate a loop at a time */
1999 if (!rto_start_trylock(&rq
->rd
->rto_loop_start
))
2002 raw_spin_lock(&rq
->rd
->rto_lock
);
2005 * The rto_cpu is updated under the lock, if it has a valid cpu
2006 * then the IPI is still running and will continue due to the
2007 * update to loop_next, and nothing needs to be done here.
2008 * Otherwise it is finishing up and an ipi needs to be sent.
2010 if (rq
->rd
->rto_cpu
< 0)
2011 cpu
= rto_next_cpu(rq
->rd
);
2013 raw_spin_unlock(&rq
->rd
->rto_lock
);
2015 rto_start_unlock(&rq
->rd
->rto_loop_start
);
2018 /* Make sure the rd does not get freed while pushing */
2019 sched_get_rd(rq
->rd
);
2020 irq_work_queue_on(&rq
->rd
->rto_push_work
, cpu
);
2024 /* Called from hardirq context */
2025 void rto_push_irq_work_func(struct irq_work
*work
)
2027 struct root_domain
*rd
=
2028 container_of(work
, struct root_domain
, rto_push_work
);
2035 * We do not need to grab the lock to check for has_pushable_tasks.
2036 * When it gets updated, a check is made if a push is possible.
2038 if (has_pushable_tasks(rq
)) {
2039 raw_spin_lock(&rq
->lock
);
2041 raw_spin_unlock(&rq
->lock
);
2044 raw_spin_lock(&rd
->rto_lock
);
2046 /* Pass the IPI to the next rt overloaded queue */
2047 cpu
= rto_next_cpu(rd
);
2049 raw_spin_unlock(&rd
->rto_lock
);
2056 /* Try the next RT overloaded CPU */
2057 irq_work_queue_on(&rd
->rto_push_work
, cpu
);
2059 #endif /* HAVE_RT_PUSH_IPI */
2061 static void pull_rt_task(struct rq
*this_rq
)
2063 int this_cpu
= this_rq
->cpu
, cpu
;
2064 bool resched
= false;
2065 struct task_struct
*p
;
2067 int rt_overload_count
= rt_overloaded(this_rq
);
2069 if (likely(!rt_overload_count
))
2073 * Match the barrier from rt_set_overloaded; this guarantees that if we
2074 * see overloaded we must also see the rto_mask bit.
2078 /* If we are the only overloaded CPU do nothing */
2079 if (rt_overload_count
== 1 &&
2080 cpumask_test_cpu(this_rq
->cpu
, this_rq
->rd
->rto_mask
))
2083 #ifdef HAVE_RT_PUSH_IPI
2084 if (sched_feat(RT_PUSH_IPI
)) {
2085 tell_cpu_to_push(this_rq
);
2090 for_each_cpu(cpu
, this_rq
->rd
->rto_mask
) {
2091 if (this_cpu
== cpu
)
2094 src_rq
= cpu_rq(cpu
);
2097 * Don't bother taking the src_rq->lock if the next highest
2098 * task is known to be lower-priority than our current task.
2099 * This may look racy, but if this value is about to go
2100 * logically higher, the src_rq will push this task away.
2101 * And if its going logically lower, we do not care
2103 if (src_rq
->rt
.highest_prio
.next
>=
2104 this_rq
->rt
.highest_prio
.curr
)
2108 * We can potentially drop this_rq's lock in
2109 * double_lock_balance, and another CPU could
2112 double_lock_balance(this_rq
, src_rq
);
2115 * We can pull only a task, which is pushable
2116 * on its rq, and no others.
2118 p
= pick_highest_pushable_task(src_rq
, this_cpu
);
2121 * Do we have an RT task that preempts
2122 * the to-be-scheduled task?
2124 if (p
&& (p
->prio
< this_rq
->rt
.highest_prio
.curr
)) {
2125 WARN_ON(p
== src_rq
->curr
);
2126 WARN_ON(!task_on_rq_queued(p
));
2129 * There's a chance that p is higher in priority
2130 * than what's currently running on its cpu.
2131 * This is just that p is wakeing up and hasn't
2132 * had a chance to schedule. We only pull
2133 * p if it is lower in priority than the
2134 * current task on the run queue
2136 if (p
->prio
< src_rq
->curr
->prio
)
2141 deactivate_task(src_rq
, p
, 0);
2142 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
2143 set_task_cpu(p
, this_cpu
);
2144 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2145 activate_task(this_rq
, p
, 0);
2147 * We continue with the search, just in
2148 * case there's an even higher prio task
2149 * in another runqueue. (low likelihood
2154 double_unlock_balance(this_rq
, src_rq
);
2158 resched_curr(this_rq
);
2162 * If we are not running and we are not going to reschedule soon, we should
2163 * try to push tasks away now
2165 static void task_woken_rt(struct rq
*rq
, struct task_struct
*p
)
2167 if (!task_running(rq
, p
) &&
2168 !test_tsk_need_resched(rq
->curr
) &&
2169 p
->nr_cpus_allowed
> 1 &&
2170 (dl_task(rq
->curr
) || rt_task(rq
->curr
)) &&
2171 (rq
->curr
->nr_cpus_allowed
< 2 ||
2172 rq
->curr
->prio
<= p
->prio
))
2176 /* Assumes rq->lock is held */
2177 static void rq_online_rt(struct rq
*rq
)
2179 if (rq
->rt
.overloaded
)
2180 rt_set_overload(rq
);
2182 __enable_runtime(rq
);
2184 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rq
->rt
.highest_prio
.curr
);
2187 /* Assumes rq->lock is held */
2188 static void rq_offline_rt(struct rq
*rq
)
2190 if (rq
->rt
.overloaded
)
2191 rt_clear_overload(rq
);
2193 __disable_runtime(rq
);
2195 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, CPUPRI_INVALID
);
2199 * When switch from the rt queue, we bring ourselves to a position
2200 * that we might want to pull RT tasks from other runqueues.
2202 static void switched_from_rt(struct rq
*rq
, struct task_struct
*p
)
2205 * If there are other RT tasks then we will reschedule
2206 * and the scheduling of the other RT tasks will handle
2207 * the balancing. But if we are the last RT task
2208 * we may need to handle the pulling of RT tasks
2211 if (!task_on_rq_queued(p
) || rq
->rt
.rt_nr_running
)
2214 queue_pull_task(rq
);
2217 void __init
init_sched_rt_class(void)
2221 for_each_possible_cpu(i
) {
2222 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask
, i
),
2223 GFP_KERNEL
, cpu_to_node(i
));
2226 #endif /* CONFIG_SMP */
2229 * When switching a task to RT, we may overload the runqueue
2230 * with RT tasks. In this case we try to push them off to
2233 static void switched_to_rt(struct rq
*rq
, struct task_struct
*p
)
2236 * If we are already running, then there's nothing
2237 * that needs to be done. But if we are not running
2238 * we may need to preempt the current running task.
2239 * If that current running task is also an RT task
2240 * then see if we can move to another run queue.
2242 if (task_on_rq_queued(p
) && rq
->curr
!= p
) {
2244 if (p
->nr_cpus_allowed
> 1 && rq
->rt
.overloaded
)
2245 queue_push_tasks(rq
);
2246 #endif /* CONFIG_SMP */
2247 if (p
->prio
< rq
->curr
->prio
&& cpu_online(cpu_of(rq
)))
2253 * Priority of the task has changed. This may cause
2254 * us to initiate a push or pull.
2257 prio_changed_rt(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
2259 if (!task_on_rq_queued(p
))
2262 if (rq
->curr
== p
) {
2265 * If our priority decreases while running, we
2266 * may need to pull tasks to this runqueue.
2268 if (oldprio
< p
->prio
)
2269 queue_pull_task(rq
);
2272 * If there's a higher priority task waiting to run
2275 if (p
->prio
> rq
->rt
.highest_prio
.curr
)
2278 /* For UP simply resched on drop of prio */
2279 if (oldprio
< p
->prio
)
2281 #endif /* CONFIG_SMP */
2284 * This task is not running, but if it is
2285 * greater than the current running task
2288 if (p
->prio
< rq
->curr
->prio
)
2293 #ifdef CONFIG_POSIX_TIMERS
2294 static void watchdog(struct rq
*rq
, struct task_struct
*p
)
2296 unsigned long soft
, hard
;
2298 /* max may change after cur was read, this will be fixed next tick */
2299 soft
= task_rlimit(p
, RLIMIT_RTTIME
);
2300 hard
= task_rlimit_max(p
, RLIMIT_RTTIME
);
2302 if (soft
!= RLIM_INFINITY
) {
2305 if (p
->rt
.watchdog_stamp
!= jiffies
) {
2307 p
->rt
.watchdog_stamp
= jiffies
;
2310 next
= DIV_ROUND_UP(min(soft
, hard
), USEC_PER_SEC
/HZ
);
2311 if (p
->rt
.timeout
> next
)
2312 p
->cputime_expires
.sched_exp
= p
->se
.sum_exec_runtime
;
2316 static inline void watchdog(struct rq
*rq
, struct task_struct
*p
) { }
2319 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
, int queued
)
2321 struct sched_rt_entity
*rt_se
= &p
->rt
;
2324 update_rt_rq_load_avg(rq_clock_task(rq
), cpu_of(rq
), &rq
->rt
, 1);
2329 * RR tasks need a special form of timeslice management.
2330 * FIFO tasks have no timeslices.
2332 if (p
->policy
!= SCHED_RR
)
2335 if (--p
->rt
.time_slice
)
2338 p
->rt
.time_slice
= sched_rr_timeslice
;
2341 * Requeue to the end of queue if we (and all of our ancestors) are not
2342 * the only element on the queue
2344 for_each_sched_rt_entity(rt_se
) {
2345 if (rt_se
->run_list
.prev
!= rt_se
->run_list
.next
) {
2346 requeue_task_rt(rq
, p
, 0);
2353 static void set_curr_task_rt(struct rq
*rq
)
2355 struct task_struct
*p
= rq
->curr
;
2357 p
->se
.exec_start
= rq_clock_task(rq
);
2359 /* The running task is never eligible for pushing */
2360 dequeue_pushable_task(rq
, p
);
2363 static unsigned int get_rr_interval_rt(struct rq
*rq
, struct task_struct
*task
)
2366 * Time slice is 0 for SCHED_FIFO tasks
2368 if (task
->policy
== SCHED_RR
)
2369 return sched_rr_timeslice
;
2374 const struct sched_class rt_sched_class
= {
2375 .next
= &fair_sched_class
,
2376 .enqueue_task
= enqueue_task_rt
,
2377 .dequeue_task
= dequeue_task_rt
,
2378 .yield_task
= yield_task_rt
,
2380 .check_preempt_curr
= check_preempt_curr_rt
,
2382 .pick_next_task
= pick_next_task_rt
,
2383 .put_prev_task
= put_prev_task_rt
,
2386 .select_task_rq
= select_task_rq_rt
,
2388 .set_cpus_allowed
= set_cpus_allowed_common
,
2389 .rq_online
= rq_online_rt
,
2390 .rq_offline
= rq_offline_rt
,
2391 .task_woken
= task_woken_rt
,
2392 .switched_from
= switched_from_rt
,
2395 .set_curr_task
= set_curr_task_rt
,
2396 .task_tick
= task_tick_rt
,
2398 .get_rr_interval
= get_rr_interval_rt
,
2400 .prio_changed
= prio_changed_rt
,
2401 .switched_to
= switched_to_rt
,
2403 .update_curr
= update_curr_rt
,
2406 #ifdef CONFIG_RT_GROUP_SCHED
2408 * Ensure that the real time constraints are schedulable.
2410 static DEFINE_MUTEX(rt_constraints_mutex
);
2412 /* Must be called with tasklist_lock held */
2413 static inline int tg_has_rt_tasks(struct task_group
*tg
)
2415 struct task_struct
*g
, *p
;
2418 * Autogroups do not have RT tasks; see autogroup_create().
2420 if (task_group_is_autogroup(tg
))
2423 for_each_process_thread(g
, p
) {
2424 if (rt_task(p
) && task_group(p
) == tg
)
2431 struct rt_schedulable_data
{
2432 struct task_group
*tg
;
2437 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
2439 struct rt_schedulable_data
*d
= data
;
2440 struct task_group
*child
;
2441 unsigned long total
, sum
= 0;
2442 u64 period
, runtime
;
2444 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2445 runtime
= tg
->rt_bandwidth
.rt_runtime
;
2448 period
= d
->rt_period
;
2449 runtime
= d
->rt_runtime
;
2453 * Cannot have more runtime than the period.
2455 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
2459 * Ensure we don't starve existing RT tasks.
2461 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
2464 total
= to_ratio(period
, runtime
);
2467 * Nobody can have more than the global setting allows.
2469 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
2473 * The sum of our children's runtime should not exceed our own.
2475 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
2476 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
2477 runtime
= child
->rt_bandwidth
.rt_runtime
;
2479 if (child
== d
->tg
) {
2480 period
= d
->rt_period
;
2481 runtime
= d
->rt_runtime
;
2484 sum
+= to_ratio(period
, runtime
);
2493 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
2497 struct rt_schedulable_data data
= {
2499 .rt_period
= period
,
2500 .rt_runtime
= runtime
,
2504 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
2510 static int tg_set_rt_bandwidth(struct task_group
*tg
,
2511 u64 rt_period
, u64 rt_runtime
)
2516 * Disallowing the root group RT runtime is BAD, it would disallow the
2517 * kernel creating (and or operating) RT threads.
2519 if (tg
== &root_task_group
&& rt_runtime
== 0)
2522 /* No period doesn't make any sense. */
2526 mutex_lock(&rt_constraints_mutex
);
2527 read_lock(&tasklist_lock
);
2528 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
2532 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
2533 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
2534 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
2536 for_each_possible_cpu(i
) {
2537 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
2539 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
2540 rt_rq
->rt_runtime
= rt_runtime
;
2541 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
2543 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
2545 read_unlock(&tasklist_lock
);
2546 mutex_unlock(&rt_constraints_mutex
);
2551 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
2553 u64 rt_runtime
, rt_period
;
2555 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2556 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
2557 if (rt_runtime_us
< 0)
2558 rt_runtime
= RUNTIME_INF
;
2559 else if ((u64
)rt_runtime_us
> U64_MAX
/ NSEC_PER_USEC
)
2562 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
2565 long sched_group_rt_runtime(struct task_group
*tg
)
2569 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
2572 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
2573 do_div(rt_runtime_us
, NSEC_PER_USEC
);
2574 return rt_runtime_us
;
2577 int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
2579 u64 rt_runtime
, rt_period
;
2581 if (rt_period_us
> U64_MAX
/ NSEC_PER_USEC
)
2584 rt_period
= rt_period_us
* NSEC_PER_USEC
;
2585 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
2587 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
2590 long sched_group_rt_period(struct task_group
*tg
)
2594 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2595 do_div(rt_period_us
, NSEC_PER_USEC
);
2596 return rt_period_us
;
2599 static int sched_rt_global_constraints(void)
2603 mutex_lock(&rt_constraints_mutex
);
2604 read_lock(&tasklist_lock
);
2605 ret
= __rt_schedulable(NULL
, 0, 0);
2606 read_unlock(&tasklist_lock
);
2607 mutex_unlock(&rt_constraints_mutex
);
2612 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
2614 /* Don't accept realtime tasks when there is no way for them to run */
2615 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
2621 #else /* !CONFIG_RT_GROUP_SCHED */
2622 static int sched_rt_global_constraints(void)
2624 unsigned long flags
;
2627 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2628 for_each_possible_cpu(i
) {
2629 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
2631 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
2632 rt_rq
->rt_runtime
= global_rt_runtime();
2633 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
2635 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2639 #endif /* CONFIG_RT_GROUP_SCHED */
2641 static int sched_rt_global_validate(void)
2643 if (sysctl_sched_rt_period
<= 0)
2646 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
2647 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
2653 static void sched_rt_do_global(void)
2655 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
2656 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
2659 int sched_rt_handler(struct ctl_table
*table
, int write
,
2660 void __user
*buffer
, size_t *lenp
,
2663 int old_period
, old_runtime
;
2664 static DEFINE_MUTEX(mutex
);
2668 old_period
= sysctl_sched_rt_period
;
2669 old_runtime
= sysctl_sched_rt_runtime
;
2671 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
2673 if (!ret
&& write
) {
2674 ret
= sched_rt_global_validate();
2678 ret
= sched_dl_global_validate();
2682 ret
= sched_rt_global_constraints();
2686 sched_rt_do_global();
2687 sched_dl_do_global();
2691 sysctl_sched_rt_period
= old_period
;
2692 sysctl_sched_rt_runtime
= old_runtime
;
2694 mutex_unlock(&mutex
);
2699 int sched_rr_handler(struct ctl_table
*table
, int write
,
2700 void __user
*buffer
, size_t *lenp
,
2704 static DEFINE_MUTEX(mutex
);
2707 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
2709 * Make sure that internally we keep jiffies.
2710 * Also, writing zero resets the timeslice to default:
2712 if (!ret
&& write
) {
2713 sched_rr_timeslice
=
2714 sysctl_sched_rr_timeslice
<= 0 ? RR_TIMESLICE
:
2715 msecs_to_jiffies(sysctl_sched_rr_timeslice
);
2717 mutex_unlock(&mutex
);
2721 #ifdef CONFIG_SCHED_DEBUG
2722 void print_rt_stats(struct seq_file
*m
, int cpu
)
2725 struct rt_rq
*rt_rq
;
2728 for_each_rt_rq(rt_rq
, iter
, cpu_rq(cpu
))
2729 print_rt_rq(m
, cpu
, rt_rq
);
2732 #endif /* CONFIG_SCHED_DEBUG */