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>
12 int sched_rr_timeslice
= RR_TIMESLICE
;
13 int sysctl_sched_rr_timeslice
= (MSEC_PER_SEC
/ HZ
) * RR_TIMESLICE
;
15 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
17 struct rt_bandwidth def_rt_bandwidth
;
19 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
21 struct rt_bandwidth
*rt_b
=
22 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
26 raw_spin_lock(&rt_b
->rt_runtime_lock
);
28 overrun
= hrtimer_forward_now(timer
, rt_b
->rt_period
);
32 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
33 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
34 raw_spin_lock(&rt_b
->rt_runtime_lock
);
37 rt_b
->rt_period_active
= 0;
38 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
40 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
43 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
45 rt_b
->rt_period
= ns_to_ktime(period
);
46 rt_b
->rt_runtime
= runtime
;
48 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
50 hrtimer_init(&rt_b
->rt_period_timer
,
51 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
52 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
55 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
57 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
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 void init_rt_rq(struct rt_rq
*rt_rq
)
79 struct rt_prio_array
*array
;
82 array
= &rt_rq
->active
;
83 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
84 INIT_LIST_HEAD(array
->queue
+ i
);
85 __clear_bit(i
, array
->bitmap
);
87 /* delimiter for bitsearch: */
88 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
90 #if defined CONFIG_SMP
91 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
92 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
93 rt_rq
->rt_nr_migratory
= 0;
94 rt_rq
->overloaded
= 0;
95 plist_head_init(&rt_rq
->pushable_tasks
);
96 #endif /* CONFIG_SMP */
97 /* We start is dequeued state, because no RT tasks are queued */
101 rt_rq
->rt_throttled
= 0;
102 rt_rq
->rt_runtime
= 0;
103 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
106 #ifdef CONFIG_RT_GROUP_SCHED
107 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
109 hrtimer_cancel(&rt_b
->rt_period_timer
);
112 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
114 static inline struct task_struct
*rt_task_of(struct sched_rt_entity
*rt_se
)
116 #ifdef CONFIG_SCHED_DEBUG
117 WARN_ON_ONCE(!rt_entity_is_task(rt_se
));
119 return container_of(rt_se
, struct task_struct
, rt
);
122 static inline struct rq
*rq_of_rt_rq(struct rt_rq
*rt_rq
)
127 static inline struct rt_rq
*rt_rq_of_se(struct sched_rt_entity
*rt_se
)
132 static inline struct rq
*rq_of_rt_se(struct sched_rt_entity
*rt_se
)
134 struct rt_rq
*rt_rq
= rt_se
->rt_rq
;
139 void free_rt_sched_group(struct task_group
*tg
)
144 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
146 for_each_possible_cpu(i
) {
157 void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
158 struct sched_rt_entity
*rt_se
, int cpu
,
159 struct sched_rt_entity
*parent
)
161 struct rq
*rq
= cpu_rq(cpu
);
163 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
164 rt_rq
->rt_nr_boosted
= 0;
168 tg
->rt_rq
[cpu
] = rt_rq
;
169 tg
->rt_se
[cpu
] = rt_se
;
175 rt_se
->rt_rq
= &rq
->rt
;
177 rt_se
->rt_rq
= parent
->my_q
;
180 rt_se
->parent
= parent
;
181 INIT_LIST_HEAD(&rt_se
->run_list
);
184 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
187 struct sched_rt_entity
*rt_se
;
190 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
193 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
197 init_rt_bandwidth(&tg
->rt_bandwidth
,
198 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
200 for_each_possible_cpu(i
) {
201 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
202 GFP_KERNEL
, cpu_to_node(i
));
206 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
207 GFP_KERNEL
, cpu_to_node(i
));
212 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
213 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, parent
->rt_se
[i
]);
224 #else /* CONFIG_RT_GROUP_SCHED */
226 #define rt_entity_is_task(rt_se) (1)
228 static inline struct task_struct
*rt_task_of(struct sched_rt_entity
*rt_se
)
230 return container_of(rt_se
, struct task_struct
, rt
);
233 static inline struct rq
*rq_of_rt_rq(struct rt_rq
*rt_rq
)
235 return container_of(rt_rq
, struct rq
, rt
);
238 static inline struct rq
*rq_of_rt_se(struct sched_rt_entity
*rt_se
)
240 struct task_struct
*p
= rt_task_of(rt_se
);
245 static inline struct rt_rq
*rt_rq_of_se(struct sched_rt_entity
*rt_se
)
247 struct rq
*rq
= rq_of_rt_se(rt_se
);
252 void free_rt_sched_group(struct task_group
*tg
) { }
254 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
258 #endif /* CONFIG_RT_GROUP_SCHED */
262 static void pull_rt_task(struct rq
*this_rq
);
264 static inline bool need_pull_rt_task(struct rq
*rq
, struct task_struct
*prev
)
266 /* Try to pull RT tasks here if we lower this rq's prio */
267 return rq
->rt
.highest_prio
.curr
> prev
->prio
;
270 static inline int rt_overloaded(struct rq
*rq
)
272 return atomic_read(&rq
->rd
->rto_count
);
275 static inline void rt_set_overload(struct rq
*rq
)
280 cpumask_set_cpu(rq
->cpu
, rq
->rd
->rto_mask
);
282 * Make sure the mask is visible before we set
283 * the overload count. That is checked to determine
284 * if we should look at the mask. It would be a shame
285 * if we looked at the mask, but the mask was not
288 * Matched by the barrier in pull_rt_task().
291 atomic_inc(&rq
->rd
->rto_count
);
294 static inline void rt_clear_overload(struct rq
*rq
)
299 /* the order here really doesn't matter */
300 atomic_dec(&rq
->rd
->rto_count
);
301 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->rto_mask
);
304 static void update_rt_migration(struct rt_rq
*rt_rq
)
306 if (rt_rq
->rt_nr_migratory
&& rt_rq
->rt_nr_total
> 1) {
307 if (!rt_rq
->overloaded
) {
308 rt_set_overload(rq_of_rt_rq(rt_rq
));
309 rt_rq
->overloaded
= 1;
311 } else if (rt_rq
->overloaded
) {
312 rt_clear_overload(rq_of_rt_rq(rt_rq
));
313 rt_rq
->overloaded
= 0;
317 static void inc_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
319 struct task_struct
*p
;
321 if (!rt_entity_is_task(rt_se
))
324 p
= rt_task_of(rt_se
);
325 rt_rq
= &rq_of_rt_rq(rt_rq
)->rt
;
327 rt_rq
->rt_nr_total
++;
328 if (p
->nr_cpus_allowed
> 1)
329 rt_rq
->rt_nr_migratory
++;
331 update_rt_migration(rt_rq
);
334 static void dec_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
336 struct task_struct
*p
;
338 if (!rt_entity_is_task(rt_se
))
341 p
= rt_task_of(rt_se
);
342 rt_rq
= &rq_of_rt_rq(rt_rq
)->rt
;
344 rt_rq
->rt_nr_total
--;
345 if (p
->nr_cpus_allowed
> 1)
346 rt_rq
->rt_nr_migratory
--;
348 update_rt_migration(rt_rq
);
351 static inline int has_pushable_tasks(struct rq
*rq
)
353 return !plist_head_empty(&rq
->rt
.pushable_tasks
);
356 static DEFINE_PER_CPU(struct callback_head
, rt_push_head
);
357 static DEFINE_PER_CPU(struct callback_head
, rt_pull_head
);
359 static void push_rt_tasks(struct rq
*);
360 static void pull_rt_task(struct rq
*);
362 static inline void queue_push_tasks(struct rq
*rq
)
364 if (!has_pushable_tasks(rq
))
367 queue_balance_callback(rq
, &per_cpu(rt_push_head
, rq
->cpu
), push_rt_tasks
);
370 static inline void queue_pull_task(struct rq
*rq
)
372 queue_balance_callback(rq
, &per_cpu(rt_pull_head
, rq
->cpu
), pull_rt_task
);
375 static void enqueue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
377 plist_del(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
378 plist_node_init(&p
->pushable_tasks
, p
->prio
);
379 plist_add(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
381 /* Update the highest prio pushable task */
382 if (p
->prio
< rq
->rt
.highest_prio
.next
)
383 rq
->rt
.highest_prio
.next
= p
->prio
;
386 static void dequeue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
388 plist_del(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
390 /* Update the new highest prio pushable task */
391 if (has_pushable_tasks(rq
)) {
392 p
= plist_first_entry(&rq
->rt
.pushable_tasks
,
393 struct task_struct
, pushable_tasks
);
394 rq
->rt
.highest_prio
.next
= p
->prio
;
396 rq
->rt
.highest_prio
.next
= MAX_RT_PRIO
;
401 static inline void enqueue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
405 static inline void dequeue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
410 void inc_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
415 void dec_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
419 static inline bool need_pull_rt_task(struct rq
*rq
, struct task_struct
*prev
)
424 static inline void pull_rt_task(struct rq
*this_rq
)
428 static inline void queue_push_tasks(struct rq
*rq
)
431 #endif /* CONFIG_SMP */
433 static void enqueue_top_rt_rq(struct rt_rq
*rt_rq
);
434 static void dequeue_top_rt_rq(struct rt_rq
*rt_rq
);
436 static inline int on_rt_rq(struct sched_rt_entity
*rt_se
)
441 #ifdef CONFIG_RT_GROUP_SCHED
443 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
448 return rt_rq
->rt_runtime
;
451 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
453 return ktime_to_ns(rt_rq
->tg
->rt_bandwidth
.rt_period
);
456 typedef struct task_group
*rt_rq_iter_t
;
458 static inline struct task_group
*next_task_group(struct task_group
*tg
)
461 tg
= list_entry_rcu(tg
->list
.next
,
462 typeof(struct task_group
), list
);
463 } while (&tg
->list
!= &task_groups
&& task_group_is_autogroup(tg
));
465 if (&tg
->list
== &task_groups
)
471 #define for_each_rt_rq(rt_rq, iter, rq) \
472 for (iter = container_of(&task_groups, typeof(*iter), list); \
473 (iter = next_task_group(iter)) && \
474 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
476 #define for_each_sched_rt_entity(rt_se) \
477 for (; rt_se; rt_se = rt_se->parent)
479 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
484 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
);
485 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
);
487 static void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
489 struct task_struct
*curr
= rq_of_rt_rq(rt_rq
)->curr
;
490 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
491 struct sched_rt_entity
*rt_se
;
493 int cpu
= cpu_of(rq
);
495 rt_se
= rt_rq
->tg
->rt_se
[cpu
];
497 if (rt_rq
->rt_nr_running
) {
499 enqueue_top_rt_rq(rt_rq
);
500 else if (!on_rt_rq(rt_se
))
501 enqueue_rt_entity(rt_se
, 0);
503 if (rt_rq
->highest_prio
.curr
< curr
->prio
)
508 static void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
510 struct sched_rt_entity
*rt_se
;
511 int cpu
= cpu_of(rq_of_rt_rq(rt_rq
));
513 rt_se
= rt_rq
->tg
->rt_se
[cpu
];
516 dequeue_top_rt_rq(rt_rq
);
517 else if (on_rt_rq(rt_se
))
518 dequeue_rt_entity(rt_se
, 0);
521 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
523 return rt_rq
->rt_throttled
&& !rt_rq
->rt_nr_boosted
;
526 static int rt_se_boosted(struct sched_rt_entity
*rt_se
)
528 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
529 struct task_struct
*p
;
532 return !!rt_rq
->rt_nr_boosted
;
534 p
= rt_task_of(rt_se
);
535 return p
->prio
!= p
->normal_prio
;
539 static inline const struct cpumask
*sched_rt_period_mask(void)
541 return this_rq()->rd
->span
;
544 static inline const struct cpumask
*sched_rt_period_mask(void)
546 return cpu_online_mask
;
551 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
553 return container_of(rt_b
, struct task_group
, rt_bandwidth
)->rt_rq
[cpu
];
556 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
558 return &rt_rq
->tg
->rt_bandwidth
;
561 #else /* !CONFIG_RT_GROUP_SCHED */
563 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
565 return rt_rq
->rt_runtime
;
568 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
570 return ktime_to_ns(def_rt_bandwidth
.rt_period
);
573 typedef struct rt_rq
*rt_rq_iter_t
;
575 #define for_each_rt_rq(rt_rq, iter, rq) \
576 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
578 #define for_each_sched_rt_entity(rt_se) \
579 for (; rt_se; rt_se = NULL)
581 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
586 static inline void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
588 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
590 if (!rt_rq
->rt_nr_running
)
593 enqueue_top_rt_rq(rt_rq
);
597 static inline void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
599 dequeue_top_rt_rq(rt_rq
);
602 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
604 return rt_rq
->rt_throttled
;
607 static inline const struct cpumask
*sched_rt_period_mask(void)
609 return cpu_online_mask
;
613 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
615 return &cpu_rq(cpu
)->rt
;
618 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
620 return &def_rt_bandwidth
;
623 #endif /* CONFIG_RT_GROUP_SCHED */
625 bool sched_rt_bandwidth_account(struct rt_rq
*rt_rq
)
627 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
629 return (hrtimer_active(&rt_b
->rt_period_timer
) ||
630 rt_rq
->rt_time
< rt_b
->rt_runtime
);
635 * We ran out of runtime, see if we can borrow some from our neighbours.
637 static void do_balance_runtime(struct rt_rq
*rt_rq
)
639 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
640 struct root_domain
*rd
= rq_of_rt_rq(rt_rq
)->rd
;
644 weight
= cpumask_weight(rd
->span
);
646 raw_spin_lock(&rt_b
->rt_runtime_lock
);
647 rt_period
= ktime_to_ns(rt_b
->rt_period
);
648 for_each_cpu(i
, rd
->span
) {
649 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
655 raw_spin_lock(&iter
->rt_runtime_lock
);
657 * Either all rqs have inf runtime and there's nothing to steal
658 * or __disable_runtime() below sets a specific rq to inf to
659 * indicate its been disabled and disalow stealing.
661 if (iter
->rt_runtime
== RUNTIME_INF
)
665 * From runqueues with spare time, take 1/n part of their
666 * spare time, but no more than our period.
668 diff
= iter
->rt_runtime
- iter
->rt_time
;
670 diff
= div_u64((u64
)diff
, weight
);
671 if (rt_rq
->rt_runtime
+ diff
> rt_period
)
672 diff
= rt_period
- rt_rq
->rt_runtime
;
673 iter
->rt_runtime
-= diff
;
674 rt_rq
->rt_runtime
+= diff
;
675 if (rt_rq
->rt_runtime
== rt_period
) {
676 raw_spin_unlock(&iter
->rt_runtime_lock
);
681 raw_spin_unlock(&iter
->rt_runtime_lock
);
683 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
687 * Ensure this RQ takes back all the runtime it lend to its neighbours.
689 static void __disable_runtime(struct rq
*rq
)
691 struct root_domain
*rd
= rq
->rd
;
695 if (unlikely(!scheduler_running
))
698 for_each_rt_rq(rt_rq
, iter
, rq
) {
699 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
703 raw_spin_lock(&rt_b
->rt_runtime_lock
);
704 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
706 * Either we're all inf and nobody needs to borrow, or we're
707 * already disabled and thus have nothing to do, or we have
708 * exactly the right amount of runtime to take out.
710 if (rt_rq
->rt_runtime
== RUNTIME_INF
||
711 rt_rq
->rt_runtime
== rt_b
->rt_runtime
)
713 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
716 * Calculate the difference between what we started out with
717 * and what we current have, that's the amount of runtime
718 * we lend and now have to reclaim.
720 want
= rt_b
->rt_runtime
- rt_rq
->rt_runtime
;
723 * Greedy reclaim, take back as much as we can.
725 for_each_cpu(i
, rd
->span
) {
726 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
730 * Can't reclaim from ourselves or disabled runqueues.
732 if (iter
== rt_rq
|| iter
->rt_runtime
== RUNTIME_INF
)
735 raw_spin_lock(&iter
->rt_runtime_lock
);
737 diff
= min_t(s64
, iter
->rt_runtime
, want
);
738 iter
->rt_runtime
-= diff
;
741 iter
->rt_runtime
-= want
;
744 raw_spin_unlock(&iter
->rt_runtime_lock
);
750 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
752 * We cannot be left wanting - that would mean some runtime
753 * leaked out of the system.
758 * Disable all the borrow logic by pretending we have inf
759 * runtime - in which case borrowing doesn't make sense.
761 rt_rq
->rt_runtime
= RUNTIME_INF
;
762 rt_rq
->rt_throttled
= 0;
763 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
764 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
766 /* Make rt_rq available for pick_next_task() */
767 sched_rt_rq_enqueue(rt_rq
);
771 static void __enable_runtime(struct rq
*rq
)
776 if (unlikely(!scheduler_running
))
780 * Reset each runqueue's bandwidth settings
782 for_each_rt_rq(rt_rq
, iter
, rq
) {
783 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
785 raw_spin_lock(&rt_b
->rt_runtime_lock
);
786 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
787 rt_rq
->rt_runtime
= rt_b
->rt_runtime
;
789 rt_rq
->rt_throttled
= 0;
790 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
791 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
795 static void balance_runtime(struct rt_rq
*rt_rq
)
797 if (!sched_feat(RT_RUNTIME_SHARE
))
800 if (rt_rq
->rt_time
> rt_rq
->rt_runtime
) {
801 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
802 do_balance_runtime(rt_rq
);
803 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
806 #else /* !CONFIG_SMP */
807 static inline void balance_runtime(struct rt_rq
*rt_rq
) {}
808 #endif /* CONFIG_SMP */
810 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
)
812 int i
, idle
= 1, throttled
= 0;
813 const struct cpumask
*span
;
815 span
= sched_rt_period_mask();
816 #ifdef CONFIG_RT_GROUP_SCHED
818 * FIXME: isolated CPUs should really leave the root task group,
819 * whether they are isolcpus or were isolated via cpusets, lest
820 * the timer run on a CPU which does not service all runqueues,
821 * potentially leaving other CPUs indefinitely throttled. If
822 * isolation is really required, the user will turn the throttle
823 * off to kill the perturbations it causes anyway. Meanwhile,
824 * this maintains functionality for boot and/or troubleshooting.
826 if (rt_b
== &root_task_group
.rt_bandwidth
)
827 span
= cpu_online_mask
;
829 for_each_cpu(i
, span
) {
831 struct rt_rq
*rt_rq
= sched_rt_period_rt_rq(rt_b
, i
);
832 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
836 * When span == cpu_online_mask, taking each rq->lock
837 * can be time-consuming. Try to avoid it when possible.
839 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
840 skip
= !rt_rq
->rt_time
&& !rt_rq
->rt_nr_running
;
841 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
845 raw_spin_lock(&rq
->lock
);
846 if (rt_rq
->rt_time
) {
849 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
850 if (rt_rq
->rt_throttled
)
851 balance_runtime(rt_rq
);
852 runtime
= rt_rq
->rt_runtime
;
853 rt_rq
->rt_time
-= min(rt_rq
->rt_time
, overrun
*runtime
);
854 if (rt_rq
->rt_throttled
&& rt_rq
->rt_time
< runtime
) {
855 rt_rq
->rt_throttled
= 0;
859 * When we're idle and a woken (rt) task is
860 * throttled check_preempt_curr() will set
861 * skip_update and the time between the wakeup
862 * and this unthrottle will get accounted as
865 if (rt_rq
->rt_nr_running
&& rq
->curr
== rq
->idle
)
866 rq_clock_skip_update(rq
, false);
868 if (rt_rq
->rt_time
|| rt_rq
->rt_nr_running
)
870 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
871 } else if (rt_rq
->rt_nr_running
) {
873 if (!rt_rq_throttled(rt_rq
))
876 if (rt_rq
->rt_throttled
)
880 sched_rt_rq_enqueue(rt_rq
);
881 raw_spin_unlock(&rq
->lock
);
884 if (!throttled
&& (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
))
890 static inline int rt_se_prio(struct sched_rt_entity
*rt_se
)
892 #ifdef CONFIG_RT_GROUP_SCHED
893 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
896 return rt_rq
->highest_prio
.curr
;
899 return rt_task_of(rt_se
)->prio
;
902 static int sched_rt_runtime_exceeded(struct rt_rq
*rt_rq
)
904 u64 runtime
= sched_rt_runtime(rt_rq
);
906 if (rt_rq
->rt_throttled
)
907 return rt_rq_throttled(rt_rq
);
909 if (runtime
>= sched_rt_period(rt_rq
))
912 balance_runtime(rt_rq
);
913 runtime
= sched_rt_runtime(rt_rq
);
914 if (runtime
== RUNTIME_INF
)
917 if (rt_rq
->rt_time
> runtime
) {
918 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
921 * Don't actually throttle groups that have no runtime assigned
922 * but accrue some time due to boosting.
924 if (likely(rt_b
->rt_runtime
)) {
925 rt_rq
->rt_throttled
= 1;
926 printk_deferred_once("sched: RT throttling activated\n");
929 * In case we did anyway, make it go away,
930 * replenishment is a joke, since it will replenish us
936 if (rt_rq_throttled(rt_rq
)) {
937 sched_rt_rq_dequeue(rt_rq
);
946 * Update the current task's runtime statistics. Skip current tasks that
947 * are not in our scheduling class.
949 static void update_curr_rt(struct rq
*rq
)
951 struct task_struct
*curr
= rq
->curr
;
952 struct sched_rt_entity
*rt_se
= &curr
->rt
;
956 if (curr
->sched_class
!= &rt_sched_class
)
959 now
= rq_clock_task(rq
);
960 delta_exec
= now
- curr
->se
.exec_start
;
961 if (unlikely((s64
)delta_exec
<= 0))
964 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
965 cpufreq_update_util(rq
, SCHED_CPUFREQ_RT
);
967 schedstat_set(curr
->se
.statistics
.exec_max
,
968 max(curr
->se
.statistics
.exec_max
, delta_exec
));
970 curr
->se
.sum_exec_runtime
+= delta_exec
;
971 account_group_exec_runtime(curr
, delta_exec
);
973 curr
->se
.exec_start
= now
;
974 cgroup_account_cputime(curr
, delta_exec
);
976 sched_rt_avg_update(rq
, delta_exec
);
978 if (!rt_bandwidth_enabled())
981 for_each_sched_rt_entity(rt_se
) {
982 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
984 if (sched_rt_runtime(rt_rq
) != RUNTIME_INF
) {
985 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
986 rt_rq
->rt_time
+= delta_exec
;
987 if (sched_rt_runtime_exceeded(rt_rq
))
989 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
995 dequeue_top_rt_rq(struct rt_rq
*rt_rq
)
997 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
999 BUG_ON(&rq
->rt
!= rt_rq
);
1001 if (!rt_rq
->rt_queued
)
1004 BUG_ON(!rq
->nr_running
);
1006 sub_nr_running(rq
, rt_rq
->rt_nr_running
);
1007 rt_rq
->rt_queued
= 0;
1011 enqueue_top_rt_rq(struct rt_rq
*rt_rq
)
1013 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1015 BUG_ON(&rq
->rt
!= rt_rq
);
1017 if (rt_rq
->rt_queued
)
1019 if (rt_rq_throttled(rt_rq
) || !rt_rq
->rt_nr_running
)
1022 add_nr_running(rq
, rt_rq
->rt_nr_running
);
1023 rt_rq
->rt_queued
= 1;
1026 #if defined CONFIG_SMP
1029 inc_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
)
1031 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1033 #ifdef CONFIG_RT_GROUP_SCHED
1035 * Change rq's cpupri only if rt_rq is the top queue.
1037 if (&rq
->rt
!= rt_rq
)
1040 if (rq
->online
&& prio
< prev_prio
)
1041 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, prio
);
1045 dec_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
)
1047 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1049 #ifdef CONFIG_RT_GROUP_SCHED
1051 * Change rq's cpupri only if rt_rq is the top queue.
1053 if (&rq
->rt
!= rt_rq
)
1056 if (rq
->online
&& rt_rq
->highest_prio
.curr
!= prev_prio
)
1057 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rt_rq
->highest_prio
.curr
);
1060 #else /* CONFIG_SMP */
1063 void inc_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
) {}
1065 void dec_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
) {}
1067 #endif /* CONFIG_SMP */
1069 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1071 inc_rt_prio(struct rt_rq
*rt_rq
, int prio
)
1073 int prev_prio
= rt_rq
->highest_prio
.curr
;
1075 if (prio
< prev_prio
)
1076 rt_rq
->highest_prio
.curr
= prio
;
1078 inc_rt_prio_smp(rt_rq
, prio
, prev_prio
);
1082 dec_rt_prio(struct rt_rq
*rt_rq
, int prio
)
1084 int prev_prio
= rt_rq
->highest_prio
.curr
;
1086 if (rt_rq
->rt_nr_running
) {
1088 WARN_ON(prio
< prev_prio
);
1091 * This may have been our highest task, and therefore
1092 * we may have some recomputation to do
1094 if (prio
== prev_prio
) {
1095 struct rt_prio_array
*array
= &rt_rq
->active
;
1097 rt_rq
->highest_prio
.curr
=
1098 sched_find_first_bit(array
->bitmap
);
1102 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
1104 dec_rt_prio_smp(rt_rq
, prio
, prev_prio
);
1109 static inline void inc_rt_prio(struct rt_rq
*rt_rq
, int prio
) {}
1110 static inline void dec_rt_prio(struct rt_rq
*rt_rq
, int prio
) {}
1112 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1114 #ifdef CONFIG_RT_GROUP_SCHED
1117 inc_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1119 if (rt_se_boosted(rt_se
))
1120 rt_rq
->rt_nr_boosted
++;
1123 start_rt_bandwidth(&rt_rq
->tg
->rt_bandwidth
);
1127 dec_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1129 if (rt_se_boosted(rt_se
))
1130 rt_rq
->rt_nr_boosted
--;
1132 WARN_ON(!rt_rq
->rt_nr_running
&& rt_rq
->rt_nr_boosted
);
1135 #else /* CONFIG_RT_GROUP_SCHED */
1138 inc_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1140 start_rt_bandwidth(&def_rt_bandwidth
);
1144 void dec_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
) {}
1146 #endif /* CONFIG_RT_GROUP_SCHED */
1149 unsigned int rt_se_nr_running(struct sched_rt_entity
*rt_se
)
1151 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1154 return group_rq
->rt_nr_running
;
1160 unsigned int rt_se_rr_nr_running(struct sched_rt_entity
*rt_se
)
1162 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1163 struct task_struct
*tsk
;
1166 return group_rq
->rr_nr_running
;
1168 tsk
= rt_task_of(rt_se
);
1170 return (tsk
->policy
== SCHED_RR
) ? 1 : 0;
1174 void inc_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1176 int prio
= rt_se_prio(rt_se
);
1178 WARN_ON(!rt_prio(prio
));
1179 rt_rq
->rt_nr_running
+= rt_se_nr_running(rt_se
);
1180 rt_rq
->rr_nr_running
+= rt_se_rr_nr_running(rt_se
);
1182 inc_rt_prio(rt_rq
, prio
);
1183 inc_rt_migration(rt_se
, rt_rq
);
1184 inc_rt_group(rt_se
, rt_rq
);
1188 void dec_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1190 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
1191 WARN_ON(!rt_rq
->rt_nr_running
);
1192 rt_rq
->rt_nr_running
-= rt_se_nr_running(rt_se
);
1193 rt_rq
->rr_nr_running
-= rt_se_rr_nr_running(rt_se
);
1195 dec_rt_prio(rt_rq
, rt_se_prio(rt_se
));
1196 dec_rt_migration(rt_se
, rt_rq
);
1197 dec_rt_group(rt_se
, rt_rq
);
1201 * Change rt_se->run_list location unless SAVE && !MOVE
1203 * assumes ENQUEUE/DEQUEUE flags match
1205 static inline bool move_entity(unsigned int flags
)
1207 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) == DEQUEUE_SAVE
)
1213 static void __delist_rt_entity(struct sched_rt_entity
*rt_se
, struct rt_prio_array
*array
)
1215 list_del_init(&rt_se
->run_list
);
1217 if (list_empty(array
->queue
+ rt_se_prio(rt_se
)))
1218 __clear_bit(rt_se_prio(rt_se
), array
->bitmap
);
1223 static void __enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1225 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
1226 struct rt_prio_array
*array
= &rt_rq
->active
;
1227 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1228 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
1231 * Don't enqueue the group if its throttled, or when empty.
1232 * The latter is a consequence of the former when a child group
1233 * get throttled and the current group doesn't have any other
1236 if (group_rq
&& (rt_rq_throttled(group_rq
) || !group_rq
->rt_nr_running
)) {
1238 __delist_rt_entity(rt_se
, array
);
1242 if (move_entity(flags
)) {
1243 WARN_ON_ONCE(rt_se
->on_list
);
1244 if (flags
& ENQUEUE_HEAD
)
1245 list_add(&rt_se
->run_list
, queue
);
1247 list_add_tail(&rt_se
->run_list
, queue
);
1249 __set_bit(rt_se_prio(rt_se
), array
->bitmap
);
1254 inc_rt_tasks(rt_se
, rt_rq
);
1257 static void __dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1259 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
1260 struct rt_prio_array
*array
= &rt_rq
->active
;
1262 if (move_entity(flags
)) {
1263 WARN_ON_ONCE(!rt_se
->on_list
);
1264 __delist_rt_entity(rt_se
, array
);
1268 dec_rt_tasks(rt_se
, rt_rq
);
1272 * Because the prio of an upper entry depends on the lower
1273 * entries, we must remove entries top - down.
1275 static void dequeue_rt_stack(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1277 struct sched_rt_entity
*back
= NULL
;
1279 for_each_sched_rt_entity(rt_se
) {
1284 dequeue_top_rt_rq(rt_rq_of_se(back
));
1286 for (rt_se
= back
; rt_se
; rt_se
= rt_se
->back
) {
1287 if (on_rt_rq(rt_se
))
1288 __dequeue_rt_entity(rt_se
, flags
);
1292 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1294 struct rq
*rq
= rq_of_rt_se(rt_se
);
1296 dequeue_rt_stack(rt_se
, flags
);
1297 for_each_sched_rt_entity(rt_se
)
1298 __enqueue_rt_entity(rt_se
, flags
);
1299 enqueue_top_rt_rq(&rq
->rt
);
1302 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1304 struct rq
*rq
= rq_of_rt_se(rt_se
);
1306 dequeue_rt_stack(rt_se
, flags
);
1308 for_each_sched_rt_entity(rt_se
) {
1309 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
1311 if (rt_rq
&& rt_rq
->rt_nr_running
)
1312 __enqueue_rt_entity(rt_se
, flags
);
1314 enqueue_top_rt_rq(&rq
->rt
);
1318 * Adding/removing a task to/from a priority array:
1321 enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1323 struct sched_rt_entity
*rt_se
= &p
->rt
;
1325 if (flags
& ENQUEUE_WAKEUP
)
1328 enqueue_rt_entity(rt_se
, flags
);
1330 if (!task_current(rq
, p
) && p
->nr_cpus_allowed
> 1)
1331 enqueue_pushable_task(rq
, p
);
1334 static void dequeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1336 struct sched_rt_entity
*rt_se
= &p
->rt
;
1339 dequeue_rt_entity(rt_se
, flags
);
1341 dequeue_pushable_task(rq
, p
);
1345 * Put task to the head or the end of the run list without the overhead of
1346 * dequeue followed by enqueue.
1349 requeue_rt_entity(struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
, int head
)
1351 if (on_rt_rq(rt_se
)) {
1352 struct rt_prio_array
*array
= &rt_rq
->active
;
1353 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
1356 list_move(&rt_se
->run_list
, queue
);
1358 list_move_tail(&rt_se
->run_list
, queue
);
1362 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int head
)
1364 struct sched_rt_entity
*rt_se
= &p
->rt
;
1365 struct rt_rq
*rt_rq
;
1367 for_each_sched_rt_entity(rt_se
) {
1368 rt_rq
= rt_rq_of_se(rt_se
);
1369 requeue_rt_entity(rt_rq
, rt_se
, head
);
1373 static void yield_task_rt(struct rq
*rq
)
1375 requeue_task_rt(rq
, rq
->curr
, 0);
1379 static int find_lowest_rq(struct task_struct
*task
);
1382 select_task_rq_rt(struct task_struct
*p
, int cpu
, int sd_flag
, int flags
)
1384 struct task_struct
*curr
;
1387 /* For anything but wake ups, just return the task_cpu */
1388 if (sd_flag
!= SD_BALANCE_WAKE
&& sd_flag
!= SD_BALANCE_FORK
)
1394 curr
= READ_ONCE(rq
->curr
); /* unlocked access */
1397 * If the current task on @p's runqueue is an RT task, then
1398 * try to see if we can wake this RT task up on another
1399 * runqueue. Otherwise simply start this RT task
1400 * on its current runqueue.
1402 * We want to avoid overloading runqueues. If the woken
1403 * task is a higher priority, then it will stay on this CPU
1404 * and the lower prio task should be moved to another CPU.
1405 * Even though this will probably make the lower prio task
1406 * lose its cache, we do not want to bounce a higher task
1407 * around just because it gave up its CPU, perhaps for a
1410 * For equal prio tasks, we just let the scheduler sort it out.
1412 * Otherwise, just let it ride on the affined RQ and the
1413 * post-schedule router will push the preempted task away
1415 * This test is optimistic, if we get it wrong the load-balancer
1416 * will have to sort it out.
1418 if (curr
&& unlikely(rt_task(curr
)) &&
1419 (curr
->nr_cpus_allowed
< 2 ||
1420 curr
->prio
<= p
->prio
)) {
1421 int target
= find_lowest_rq(p
);
1424 * Don't bother moving it if the destination CPU is
1425 * not running a lower priority task.
1428 p
->prio
< cpu_rq(target
)->rt
.highest_prio
.curr
)
1437 static void check_preempt_equal_prio(struct rq
*rq
, struct task_struct
*p
)
1440 * Current can't be migrated, useless to reschedule,
1441 * let's hope p can move out.
1443 if (rq
->curr
->nr_cpus_allowed
== 1 ||
1444 !cpupri_find(&rq
->rd
->cpupri
, rq
->curr
, NULL
))
1448 * p is migratable, so let's not schedule it and
1449 * see if it is pushed or pulled somewhere else.
1451 if (p
->nr_cpus_allowed
!= 1
1452 && cpupri_find(&rq
->rd
->cpupri
, p
, NULL
))
1456 * There appear to be other CPUs that can accept
1457 * the current task but none can run 'p', so lets reschedule
1458 * to try and push the current task away:
1460 requeue_task_rt(rq
, p
, 1);
1464 #endif /* CONFIG_SMP */
1467 * Preempt the current task with a newly woken task if needed:
1469 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1471 if (p
->prio
< rq
->curr
->prio
) {
1480 * - the newly woken task is of equal priority to the current task
1481 * - the newly woken task is non-migratable while current is migratable
1482 * - current will be preempted on the next reschedule
1484 * we should check to see if current can readily move to a different
1485 * cpu. If so, we will reschedule to allow the push logic to try
1486 * to move current somewhere else, making room for our non-migratable
1489 if (p
->prio
== rq
->curr
->prio
&& !test_tsk_need_resched(rq
->curr
))
1490 check_preempt_equal_prio(rq
, p
);
1494 static struct sched_rt_entity
*pick_next_rt_entity(struct rq
*rq
,
1495 struct rt_rq
*rt_rq
)
1497 struct rt_prio_array
*array
= &rt_rq
->active
;
1498 struct sched_rt_entity
*next
= NULL
;
1499 struct list_head
*queue
;
1502 idx
= sched_find_first_bit(array
->bitmap
);
1503 BUG_ON(idx
>= MAX_RT_PRIO
);
1505 queue
= array
->queue
+ idx
;
1506 next
= list_entry(queue
->next
, struct sched_rt_entity
, run_list
);
1511 static struct task_struct
*_pick_next_task_rt(struct rq
*rq
)
1513 struct sched_rt_entity
*rt_se
;
1514 struct task_struct
*p
;
1515 struct rt_rq
*rt_rq
= &rq
->rt
;
1518 rt_se
= pick_next_rt_entity(rq
, rt_rq
);
1520 rt_rq
= group_rt_rq(rt_se
);
1523 p
= rt_task_of(rt_se
);
1524 p
->se
.exec_start
= rq_clock_task(rq
);
1529 static struct task_struct
*
1530 pick_next_task_rt(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
1532 struct task_struct
*p
;
1533 struct rt_rq
*rt_rq
= &rq
->rt
;
1535 if (need_pull_rt_task(rq
, prev
)) {
1537 * This is OK, because current is on_cpu, which avoids it being
1538 * picked for load-balance and preemption/IRQs are still
1539 * disabled avoiding further scheduler activity on it and we're
1540 * being very careful to re-start the picking loop.
1542 rq_unpin_lock(rq
, rf
);
1544 rq_repin_lock(rq
, rf
);
1546 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1547 * means a dl or stop task can slip in, in which case we need
1548 * to re-start task selection.
1550 if (unlikely((rq
->stop
&& task_on_rq_queued(rq
->stop
)) ||
1551 rq
->dl
.dl_nr_running
))
1556 * We may dequeue prev's rt_rq in put_prev_task().
1557 * So, we update time before rt_nr_running check.
1559 if (prev
->sched_class
== &rt_sched_class
)
1562 if (!rt_rq
->rt_queued
)
1565 put_prev_task(rq
, prev
);
1567 p
= _pick_next_task_rt(rq
);
1569 /* The running task is never eligible for pushing */
1570 dequeue_pushable_task(rq
, p
);
1572 queue_push_tasks(rq
);
1577 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
1582 * The previous task needs to be made eligible for pushing
1583 * if it is still active
1585 if (on_rt_rq(&p
->rt
) && p
->nr_cpus_allowed
> 1)
1586 enqueue_pushable_task(rq
, p
);
1591 /* Only try algorithms three times */
1592 #define RT_MAX_TRIES 3
1594 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
1596 if (!task_running(rq
, p
) &&
1597 cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
1604 * Return the highest pushable rq's task, which is suitable to be executed
1605 * on the CPU, NULL otherwise
1607 static struct task_struct
*pick_highest_pushable_task(struct rq
*rq
, int cpu
)
1609 struct plist_head
*head
= &rq
->rt
.pushable_tasks
;
1610 struct task_struct
*p
;
1612 if (!has_pushable_tasks(rq
))
1615 plist_for_each_entry(p
, head
, pushable_tasks
) {
1616 if (pick_rt_task(rq
, p
, cpu
))
1623 static DEFINE_PER_CPU(cpumask_var_t
, local_cpu_mask
);
1625 static int find_lowest_rq(struct task_struct
*task
)
1627 struct sched_domain
*sd
;
1628 struct cpumask
*lowest_mask
= this_cpu_cpumask_var_ptr(local_cpu_mask
);
1629 int this_cpu
= smp_processor_id();
1630 int cpu
= task_cpu(task
);
1632 /* Make sure the mask is initialized first */
1633 if (unlikely(!lowest_mask
))
1636 if (task
->nr_cpus_allowed
== 1)
1637 return -1; /* No other targets possible */
1639 if (!cpupri_find(&task_rq(task
)->rd
->cpupri
, task
, lowest_mask
))
1640 return -1; /* No targets found */
1643 * At this point we have built a mask of CPUs representing the
1644 * lowest priority tasks in the system. Now we want to elect
1645 * the best one based on our affinity and topology.
1647 * We prioritize the last CPU that the task executed on since
1648 * it is most likely cache-hot in that location.
1650 if (cpumask_test_cpu(cpu
, lowest_mask
))
1654 * Otherwise, we consult the sched_domains span maps to figure
1655 * out which CPU is logically closest to our hot cache data.
1657 if (!cpumask_test_cpu(this_cpu
, lowest_mask
))
1658 this_cpu
= -1; /* Skip this_cpu opt if not among lowest */
1661 for_each_domain(cpu
, sd
) {
1662 if (sd
->flags
& SD_WAKE_AFFINE
) {
1666 * "this_cpu" is cheaper to preempt than a
1669 if (this_cpu
!= -1 &&
1670 cpumask_test_cpu(this_cpu
, sched_domain_span(sd
))) {
1675 best_cpu
= cpumask_first_and(lowest_mask
,
1676 sched_domain_span(sd
));
1677 if (best_cpu
< nr_cpu_ids
) {
1686 * And finally, if there were no matches within the domains
1687 * just give the caller *something* to work with from the compatible
1693 cpu
= cpumask_any(lowest_mask
);
1694 if (cpu
< nr_cpu_ids
)
1700 /* Will lock the rq it finds */
1701 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
1703 struct rq
*lowest_rq
= NULL
;
1707 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
1708 cpu
= find_lowest_rq(task
);
1710 if ((cpu
== -1) || (cpu
== rq
->cpu
))
1713 lowest_rq
= cpu_rq(cpu
);
1715 if (lowest_rq
->rt
.highest_prio
.curr
<= task
->prio
) {
1717 * Target rq has tasks of equal or higher priority,
1718 * retrying does not release any lock and is unlikely
1719 * to yield a different result.
1725 /* if the prio of this runqueue changed, try again */
1726 if (double_lock_balance(rq
, lowest_rq
)) {
1728 * We had to unlock the run queue. In
1729 * the mean time, task could have
1730 * migrated already or had its affinity changed.
1731 * Also make sure that it wasn't scheduled on its rq.
1733 if (unlikely(task_rq(task
) != rq
||
1734 !cpumask_test_cpu(lowest_rq
->cpu
, &task
->cpus_allowed
) ||
1735 task_running(rq
, task
) ||
1737 !task_on_rq_queued(task
))) {
1739 double_unlock_balance(rq
, lowest_rq
);
1745 /* If this rq is still suitable use it. */
1746 if (lowest_rq
->rt
.highest_prio
.curr
> task
->prio
)
1750 double_unlock_balance(rq
, lowest_rq
);
1757 static struct task_struct
*pick_next_pushable_task(struct rq
*rq
)
1759 struct task_struct
*p
;
1761 if (!has_pushable_tasks(rq
))
1764 p
= plist_first_entry(&rq
->rt
.pushable_tasks
,
1765 struct task_struct
, pushable_tasks
);
1767 BUG_ON(rq
->cpu
!= task_cpu(p
));
1768 BUG_ON(task_current(rq
, p
));
1769 BUG_ON(p
->nr_cpus_allowed
<= 1);
1771 BUG_ON(!task_on_rq_queued(p
));
1772 BUG_ON(!rt_task(p
));
1778 * If the current CPU has more than one RT task, see if the non
1779 * running task can migrate over to a CPU that is running a task
1780 * of lesser priority.
1782 static int push_rt_task(struct rq
*rq
)
1784 struct task_struct
*next_task
;
1785 struct rq
*lowest_rq
;
1788 if (!rq
->rt
.overloaded
)
1791 next_task
= pick_next_pushable_task(rq
);
1796 if (unlikely(next_task
== rq
->curr
)) {
1802 * It's possible that the next_task slipped in of
1803 * higher priority than current. If that's the case
1804 * just reschedule current.
1806 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
1811 /* We might release rq lock */
1812 get_task_struct(next_task
);
1814 /* find_lock_lowest_rq locks the rq if found */
1815 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
1817 struct task_struct
*task
;
1819 * find_lock_lowest_rq releases rq->lock
1820 * so it is possible that next_task has migrated.
1822 * We need to make sure that the task is still on the same
1823 * run-queue and is also still the next task eligible for
1826 task
= pick_next_pushable_task(rq
);
1827 if (task
== next_task
) {
1829 * The task hasn't migrated, and is still the next
1830 * eligible task, but we failed to find a run-queue
1831 * to push it to. Do not retry in this case, since
1832 * other CPUs will pull from us when ready.
1838 /* No more tasks, just exit */
1842 * Something has shifted, try again.
1844 put_task_struct(next_task
);
1849 deactivate_task(rq
, next_task
, 0);
1850 set_task_cpu(next_task
, lowest_rq
->cpu
);
1851 activate_task(lowest_rq
, next_task
, 0);
1854 resched_curr(lowest_rq
);
1856 double_unlock_balance(rq
, lowest_rq
);
1859 put_task_struct(next_task
);
1864 static void push_rt_tasks(struct rq
*rq
)
1866 /* push_rt_task will return true if it moved an RT */
1867 while (push_rt_task(rq
))
1871 #ifdef HAVE_RT_PUSH_IPI
1874 * When a high priority task schedules out from a CPU and a lower priority
1875 * task is scheduled in, a check is made to see if there's any RT tasks
1876 * on other CPUs that are waiting to run because a higher priority RT task
1877 * is currently running on its CPU. In this case, the CPU with multiple RT
1878 * tasks queued on it (overloaded) needs to be notified that a CPU has opened
1879 * up that may be able to run one of its non-running queued RT tasks.
1881 * All CPUs with overloaded RT tasks need to be notified as there is currently
1882 * no way to know which of these CPUs have the highest priority task waiting
1883 * to run. Instead of trying to take a spinlock on each of these CPUs,
1884 * which has shown to cause large latency when done on machines with many
1885 * CPUs, sending an IPI to the CPUs to have them push off the overloaded
1886 * RT tasks waiting to run.
1888 * Just sending an IPI to each of the CPUs is also an issue, as on large
1889 * count CPU machines, this can cause an IPI storm on a CPU, especially
1890 * if its the only CPU with multiple RT tasks queued, and a large number
1891 * of CPUs scheduling a lower priority task at the same time.
1893 * Each root domain has its own irq work function that can iterate over
1894 * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT
1895 * tassk must be checked if there's one or many CPUs that are lowering
1896 * their priority, there's a single irq work iterator that will try to
1897 * push off RT tasks that are waiting to run.
1899 * When a CPU schedules a lower priority task, it will kick off the
1900 * irq work iterator that will jump to each CPU with overloaded RT tasks.
1901 * As it only takes the first CPU that schedules a lower priority task
1902 * to start the process, the rto_start variable is incremented and if
1903 * the atomic result is one, then that CPU will try to take the rto_lock.
1904 * This prevents high contention on the lock as the process handles all
1905 * CPUs scheduling lower priority tasks.
1907 * All CPUs that are scheduling a lower priority task will increment the
1908 * rt_loop_next variable. This will make sure that the irq work iterator
1909 * checks all RT overloaded CPUs whenever a CPU schedules a new lower
1910 * priority task, even if the iterator is in the middle of a scan. Incrementing
1911 * the rt_loop_next will cause the iterator to perform another scan.
1914 static int rto_next_cpu(struct root_domain
*rd
)
1920 * When starting the IPI RT pushing, the rto_cpu is set to -1,
1921 * rt_next_cpu() will simply return the first CPU found in
1924 * If rto_next_cpu() is called with rto_cpu is a valid CPU, it
1925 * will return the next CPU found in the rto_mask.
1927 * If there are no more CPUs left in the rto_mask, then a check is made
1928 * against rto_loop and rto_loop_next. rto_loop is only updated with
1929 * the rto_lock held, but any CPU may increment the rto_loop_next
1930 * without any locking.
1934 /* When rto_cpu is -1 this acts like cpumask_first() */
1935 cpu
= cpumask_next(rd
->rto_cpu
, rd
->rto_mask
);
1939 if (cpu
< nr_cpu_ids
)
1945 * ACQUIRE ensures we see the @rto_mask changes
1946 * made prior to the @next value observed.
1948 * Matches WMB in rt_set_overload().
1950 next
= atomic_read_acquire(&rd
->rto_loop_next
);
1952 if (rd
->rto_loop
== next
)
1955 rd
->rto_loop
= next
;
1961 static inline bool rto_start_trylock(atomic_t
*v
)
1963 return !atomic_cmpxchg_acquire(v
, 0, 1);
1966 static inline void rto_start_unlock(atomic_t
*v
)
1968 atomic_set_release(v
, 0);
1971 static void tell_cpu_to_push(struct rq
*rq
)
1975 /* Keep the loop going if the IPI is currently active */
1976 atomic_inc(&rq
->rd
->rto_loop_next
);
1978 /* Only one CPU can initiate a loop at a time */
1979 if (!rto_start_trylock(&rq
->rd
->rto_loop_start
))
1982 raw_spin_lock(&rq
->rd
->rto_lock
);
1985 * The rto_cpu is updated under the lock, if it has a valid CPU
1986 * then the IPI is still running and will continue due to the
1987 * update to loop_next, and nothing needs to be done here.
1988 * Otherwise it is finishing up and an ipi needs to be sent.
1990 if (rq
->rd
->rto_cpu
< 0)
1991 cpu
= rto_next_cpu(rq
->rd
);
1993 raw_spin_unlock(&rq
->rd
->rto_lock
);
1995 rto_start_unlock(&rq
->rd
->rto_loop_start
);
1998 /* Make sure the rd does not get freed while pushing */
1999 sched_get_rd(rq
->rd
);
2000 irq_work_queue_on(&rq
->rd
->rto_push_work
, cpu
);
2004 /* Called from hardirq context */
2005 void rto_push_irq_work_func(struct irq_work
*work
)
2007 struct root_domain
*rd
=
2008 container_of(work
, struct root_domain
, rto_push_work
);
2015 * We do not need to grab the lock to check for has_pushable_tasks.
2016 * When it gets updated, a check is made if a push is possible.
2018 if (has_pushable_tasks(rq
)) {
2019 raw_spin_lock(&rq
->lock
);
2021 raw_spin_unlock(&rq
->lock
);
2024 raw_spin_lock(&rd
->rto_lock
);
2026 /* Pass the IPI to the next rt overloaded queue */
2027 cpu
= rto_next_cpu(rd
);
2029 raw_spin_unlock(&rd
->rto_lock
);
2036 /* Try the next RT overloaded CPU */
2037 irq_work_queue_on(&rd
->rto_push_work
, cpu
);
2039 #endif /* HAVE_RT_PUSH_IPI */
2041 static void pull_rt_task(struct rq
*this_rq
)
2043 int this_cpu
= this_rq
->cpu
, cpu
;
2044 bool resched
= false;
2045 struct task_struct
*p
;
2047 int rt_overload_count
= rt_overloaded(this_rq
);
2049 if (likely(!rt_overload_count
))
2053 * Match the barrier from rt_set_overloaded; this guarantees that if we
2054 * see overloaded we must also see the rto_mask bit.
2058 /* If we are the only overloaded CPU do nothing */
2059 if (rt_overload_count
== 1 &&
2060 cpumask_test_cpu(this_rq
->cpu
, this_rq
->rd
->rto_mask
))
2063 #ifdef HAVE_RT_PUSH_IPI
2064 if (sched_feat(RT_PUSH_IPI
)) {
2065 tell_cpu_to_push(this_rq
);
2070 for_each_cpu(cpu
, this_rq
->rd
->rto_mask
) {
2071 if (this_cpu
== cpu
)
2074 src_rq
= cpu_rq(cpu
);
2077 * Don't bother taking the src_rq->lock if the next highest
2078 * task is known to be lower-priority than our current task.
2079 * This may look racy, but if this value is about to go
2080 * logically higher, the src_rq will push this task away.
2081 * And if its going logically lower, we do not care
2083 if (src_rq
->rt
.highest_prio
.next
>=
2084 this_rq
->rt
.highest_prio
.curr
)
2088 * We can potentially drop this_rq's lock in
2089 * double_lock_balance, and another CPU could
2092 double_lock_balance(this_rq
, src_rq
);
2095 * We can pull only a task, which is pushable
2096 * on its rq, and no others.
2098 p
= pick_highest_pushable_task(src_rq
, this_cpu
);
2101 * Do we have an RT task that preempts
2102 * the to-be-scheduled task?
2104 if (p
&& (p
->prio
< this_rq
->rt
.highest_prio
.curr
)) {
2105 WARN_ON(p
== src_rq
->curr
);
2106 WARN_ON(!task_on_rq_queued(p
));
2109 * There's a chance that p is higher in priority
2110 * than what's currently running on its CPU.
2111 * This is just that p is wakeing up and hasn't
2112 * had a chance to schedule. We only pull
2113 * p if it is lower in priority than the
2114 * current task on the run queue
2116 if (p
->prio
< src_rq
->curr
->prio
)
2121 deactivate_task(src_rq
, p
, 0);
2122 set_task_cpu(p
, this_cpu
);
2123 activate_task(this_rq
, p
, 0);
2125 * We continue with the search, just in
2126 * case there's an even higher prio task
2127 * in another runqueue. (low likelihood
2132 double_unlock_balance(this_rq
, src_rq
);
2136 resched_curr(this_rq
);
2140 * If we are not running and we are not going to reschedule soon, we should
2141 * try to push tasks away now
2143 static void task_woken_rt(struct rq
*rq
, struct task_struct
*p
)
2145 if (!task_running(rq
, p
) &&
2146 !test_tsk_need_resched(rq
->curr
) &&
2147 p
->nr_cpus_allowed
> 1 &&
2148 (dl_task(rq
->curr
) || rt_task(rq
->curr
)) &&
2149 (rq
->curr
->nr_cpus_allowed
< 2 ||
2150 rq
->curr
->prio
<= p
->prio
))
2154 /* Assumes rq->lock is held */
2155 static void rq_online_rt(struct rq
*rq
)
2157 if (rq
->rt
.overloaded
)
2158 rt_set_overload(rq
);
2160 __enable_runtime(rq
);
2162 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rq
->rt
.highest_prio
.curr
);
2165 /* Assumes rq->lock is held */
2166 static void rq_offline_rt(struct rq
*rq
)
2168 if (rq
->rt
.overloaded
)
2169 rt_clear_overload(rq
);
2171 __disable_runtime(rq
);
2173 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, CPUPRI_INVALID
);
2177 * When switch from the rt queue, we bring ourselves to a position
2178 * that we might want to pull RT tasks from other runqueues.
2180 static void switched_from_rt(struct rq
*rq
, struct task_struct
*p
)
2183 * If there are other RT tasks then we will reschedule
2184 * and the scheduling of the other RT tasks will handle
2185 * the balancing. But if we are the last RT task
2186 * we may need to handle the pulling of RT tasks
2189 if (!task_on_rq_queued(p
) || rq
->rt
.rt_nr_running
)
2192 queue_pull_task(rq
);
2195 void __init
init_sched_rt_class(void)
2199 for_each_possible_cpu(i
) {
2200 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask
, i
),
2201 GFP_KERNEL
, cpu_to_node(i
));
2204 #endif /* CONFIG_SMP */
2207 * When switching a task to RT, we may overload the runqueue
2208 * with RT tasks. In this case we try to push them off to
2211 static void switched_to_rt(struct rq
*rq
, struct task_struct
*p
)
2214 * If we are already running, then there's nothing
2215 * that needs to be done. But if we are not running
2216 * we may need to preempt the current running task.
2217 * If that current running task is also an RT task
2218 * then see if we can move to another run queue.
2220 if (task_on_rq_queued(p
) && rq
->curr
!= p
) {
2222 if (p
->nr_cpus_allowed
> 1 && rq
->rt
.overloaded
)
2223 queue_push_tasks(rq
);
2224 #endif /* CONFIG_SMP */
2225 if (p
->prio
< rq
->curr
->prio
&& cpu_online(cpu_of(rq
)))
2231 * Priority of the task has changed. This may cause
2232 * us to initiate a push or pull.
2235 prio_changed_rt(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
2237 if (!task_on_rq_queued(p
))
2240 if (rq
->curr
== p
) {
2243 * If our priority decreases while running, we
2244 * may need to pull tasks to this runqueue.
2246 if (oldprio
< p
->prio
)
2247 queue_pull_task(rq
);
2250 * If there's a higher priority task waiting to run
2253 if (p
->prio
> rq
->rt
.highest_prio
.curr
)
2256 /* For UP simply resched on drop of prio */
2257 if (oldprio
< p
->prio
)
2259 #endif /* CONFIG_SMP */
2262 * This task is not running, but if it is
2263 * greater than the current running task
2266 if (p
->prio
< rq
->curr
->prio
)
2271 #ifdef CONFIG_POSIX_TIMERS
2272 static void watchdog(struct rq
*rq
, struct task_struct
*p
)
2274 unsigned long soft
, hard
;
2276 /* max may change after cur was read, this will be fixed next tick */
2277 soft
= task_rlimit(p
, RLIMIT_RTTIME
);
2278 hard
= task_rlimit_max(p
, RLIMIT_RTTIME
);
2280 if (soft
!= RLIM_INFINITY
) {
2283 if (p
->rt
.watchdog_stamp
!= jiffies
) {
2285 p
->rt
.watchdog_stamp
= jiffies
;
2288 next
= DIV_ROUND_UP(min(soft
, hard
), USEC_PER_SEC
/HZ
);
2289 if (p
->rt
.timeout
> next
)
2290 p
->cputime_expires
.sched_exp
= p
->se
.sum_exec_runtime
;
2294 static inline void watchdog(struct rq
*rq
, struct task_struct
*p
) { }
2298 * scheduler tick hitting a task of our scheduling class.
2300 * NOTE: This function can be called remotely by the tick offload that
2301 * goes along full dynticks. Therefore no local assumption can be made
2302 * and everything must be accessed through the @rq and @curr passed in
2305 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
, int queued
)
2307 struct sched_rt_entity
*rt_se
= &p
->rt
;
2314 * RR tasks need a special form of timeslice management.
2315 * FIFO tasks have no timeslices.
2317 if (p
->policy
!= SCHED_RR
)
2320 if (--p
->rt
.time_slice
)
2323 p
->rt
.time_slice
= sched_rr_timeslice
;
2326 * Requeue to the end of queue if we (and all of our ancestors) are not
2327 * the only element on the queue
2329 for_each_sched_rt_entity(rt_se
) {
2330 if (rt_se
->run_list
.prev
!= rt_se
->run_list
.next
) {
2331 requeue_task_rt(rq
, p
, 0);
2338 static void set_curr_task_rt(struct rq
*rq
)
2340 struct task_struct
*p
= rq
->curr
;
2342 p
->se
.exec_start
= rq_clock_task(rq
);
2344 /* The running task is never eligible for pushing */
2345 dequeue_pushable_task(rq
, p
);
2348 static unsigned int get_rr_interval_rt(struct rq
*rq
, struct task_struct
*task
)
2351 * Time slice is 0 for SCHED_FIFO tasks
2353 if (task
->policy
== SCHED_RR
)
2354 return sched_rr_timeslice
;
2359 const struct sched_class rt_sched_class
= {
2360 .next
= &fair_sched_class
,
2361 .enqueue_task
= enqueue_task_rt
,
2362 .dequeue_task
= dequeue_task_rt
,
2363 .yield_task
= yield_task_rt
,
2365 .check_preempt_curr
= check_preempt_curr_rt
,
2367 .pick_next_task
= pick_next_task_rt
,
2368 .put_prev_task
= put_prev_task_rt
,
2371 .select_task_rq
= select_task_rq_rt
,
2373 .set_cpus_allowed
= set_cpus_allowed_common
,
2374 .rq_online
= rq_online_rt
,
2375 .rq_offline
= rq_offline_rt
,
2376 .task_woken
= task_woken_rt
,
2377 .switched_from
= switched_from_rt
,
2380 .set_curr_task
= set_curr_task_rt
,
2381 .task_tick
= task_tick_rt
,
2383 .get_rr_interval
= get_rr_interval_rt
,
2385 .prio_changed
= prio_changed_rt
,
2386 .switched_to
= switched_to_rt
,
2388 .update_curr
= update_curr_rt
,
2391 #ifdef CONFIG_RT_GROUP_SCHED
2393 * Ensure that the real time constraints are schedulable.
2395 static DEFINE_MUTEX(rt_constraints_mutex
);
2397 /* Must be called with tasklist_lock held */
2398 static inline int tg_has_rt_tasks(struct task_group
*tg
)
2400 struct task_struct
*g
, *p
;
2403 * Autogroups do not have RT tasks; see autogroup_create().
2405 if (task_group_is_autogroup(tg
))
2408 for_each_process_thread(g
, p
) {
2409 if (rt_task(p
) && task_group(p
) == tg
)
2416 struct rt_schedulable_data
{
2417 struct task_group
*tg
;
2422 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
2424 struct rt_schedulable_data
*d
= data
;
2425 struct task_group
*child
;
2426 unsigned long total
, sum
= 0;
2427 u64 period
, runtime
;
2429 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2430 runtime
= tg
->rt_bandwidth
.rt_runtime
;
2433 period
= d
->rt_period
;
2434 runtime
= d
->rt_runtime
;
2438 * Cannot have more runtime than the period.
2440 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
2444 * Ensure we don't starve existing RT tasks.
2446 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
2449 total
= to_ratio(period
, runtime
);
2452 * Nobody can have more than the global setting allows.
2454 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
2458 * The sum of our children's runtime should not exceed our own.
2460 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
2461 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
2462 runtime
= child
->rt_bandwidth
.rt_runtime
;
2464 if (child
== d
->tg
) {
2465 period
= d
->rt_period
;
2466 runtime
= d
->rt_runtime
;
2469 sum
+= to_ratio(period
, runtime
);
2478 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
2482 struct rt_schedulable_data data
= {
2484 .rt_period
= period
,
2485 .rt_runtime
= runtime
,
2489 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
2495 static int tg_set_rt_bandwidth(struct task_group
*tg
,
2496 u64 rt_period
, u64 rt_runtime
)
2501 * Disallowing the root group RT runtime is BAD, it would disallow the
2502 * kernel creating (and or operating) RT threads.
2504 if (tg
== &root_task_group
&& rt_runtime
== 0)
2507 /* No period doesn't make any sense. */
2511 mutex_lock(&rt_constraints_mutex
);
2512 read_lock(&tasklist_lock
);
2513 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
2517 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
2518 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
2519 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
2521 for_each_possible_cpu(i
) {
2522 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
2524 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
2525 rt_rq
->rt_runtime
= rt_runtime
;
2526 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
2528 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
2530 read_unlock(&tasklist_lock
);
2531 mutex_unlock(&rt_constraints_mutex
);
2536 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
2538 u64 rt_runtime
, rt_period
;
2540 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2541 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
2542 if (rt_runtime_us
< 0)
2543 rt_runtime
= RUNTIME_INF
;
2545 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
2548 long sched_group_rt_runtime(struct task_group
*tg
)
2552 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
2555 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
2556 do_div(rt_runtime_us
, NSEC_PER_USEC
);
2557 return rt_runtime_us
;
2560 int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
2562 u64 rt_runtime
, rt_period
;
2564 rt_period
= rt_period_us
* NSEC_PER_USEC
;
2565 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
2567 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
2570 long sched_group_rt_period(struct task_group
*tg
)
2574 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2575 do_div(rt_period_us
, NSEC_PER_USEC
);
2576 return rt_period_us
;
2579 static int sched_rt_global_constraints(void)
2583 mutex_lock(&rt_constraints_mutex
);
2584 read_lock(&tasklist_lock
);
2585 ret
= __rt_schedulable(NULL
, 0, 0);
2586 read_unlock(&tasklist_lock
);
2587 mutex_unlock(&rt_constraints_mutex
);
2592 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
2594 /* Don't accept realtime tasks when there is no way for them to run */
2595 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
2601 #else /* !CONFIG_RT_GROUP_SCHED */
2602 static int sched_rt_global_constraints(void)
2604 unsigned long flags
;
2607 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2608 for_each_possible_cpu(i
) {
2609 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
2611 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
2612 rt_rq
->rt_runtime
= global_rt_runtime();
2613 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
2615 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2619 #endif /* CONFIG_RT_GROUP_SCHED */
2621 static int sched_rt_global_validate(void)
2623 if (sysctl_sched_rt_period
<= 0)
2626 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
2627 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
2633 static void sched_rt_do_global(void)
2635 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
2636 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
2639 int sched_rt_handler(struct ctl_table
*table
, int write
,
2640 void __user
*buffer
, size_t *lenp
,
2643 int old_period
, old_runtime
;
2644 static DEFINE_MUTEX(mutex
);
2648 old_period
= sysctl_sched_rt_period
;
2649 old_runtime
= sysctl_sched_rt_runtime
;
2651 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
2653 if (!ret
&& write
) {
2654 ret
= sched_rt_global_validate();
2658 ret
= sched_dl_global_validate();
2662 ret
= sched_rt_global_constraints();
2666 sched_rt_do_global();
2667 sched_dl_do_global();
2671 sysctl_sched_rt_period
= old_period
;
2672 sysctl_sched_rt_runtime
= old_runtime
;
2674 mutex_unlock(&mutex
);
2679 int sched_rr_handler(struct ctl_table
*table
, int write
,
2680 void __user
*buffer
, size_t *lenp
,
2684 static DEFINE_MUTEX(mutex
);
2687 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
2689 * Make sure that internally we keep jiffies.
2690 * Also, writing zero resets the timeslice to default:
2692 if (!ret
&& write
) {
2693 sched_rr_timeslice
=
2694 sysctl_sched_rr_timeslice
<= 0 ? RR_TIMESLICE
:
2695 msecs_to_jiffies(sysctl_sched_rr_timeslice
);
2697 mutex_unlock(&mutex
);
2702 #ifdef CONFIG_SCHED_DEBUG
2703 extern void print_rt_rq(struct seq_file
*m
, int cpu
, struct rt_rq
*rt_rq
);
2705 void print_rt_stats(struct seq_file
*m
, int cpu
)
2708 struct rt_rq
*rt_rq
;
2711 for_each_rt_rq(rt_rq
, iter
, cpu_rq(cpu
))
2712 print_rt_rq(m
, cpu
, rt_rq
);
2715 #endif /* CONFIG_SCHED_DEBUG */