2 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
8 #include <linux/slab.h>
9 #include <linux/irq_work.h>
11 int sched_rr_timeslice
= RR_TIMESLICE
;
12 int sysctl_sched_rr_timeslice
= (MSEC_PER_SEC
/ HZ
) * RR_TIMESLICE
;
14 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
16 struct rt_bandwidth def_rt_bandwidth
;
18 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
20 struct rt_bandwidth
*rt_b
=
21 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
25 raw_spin_lock(&rt_b
->rt_runtime_lock
);
27 overrun
= hrtimer_forward_now(timer
, rt_b
->rt_period
);
31 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
32 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
33 raw_spin_lock(&rt_b
->rt_runtime_lock
);
36 rt_b
->rt_period_active
= 0;
37 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
39 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
42 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
44 rt_b
->rt_period
= ns_to_ktime(period
);
45 rt_b
->rt_runtime
= runtime
;
47 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
49 hrtimer_init(&rt_b
->rt_period_timer
,
50 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
51 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
54 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
56 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
59 raw_spin_lock(&rt_b
->rt_runtime_lock
);
60 if (!rt_b
->rt_period_active
) {
61 rt_b
->rt_period_active
= 1;
63 * SCHED_DEADLINE updates the bandwidth, as a run away
64 * RT task with a DL task could hog a CPU. But DL does
65 * not reset the period. If a deadline task was running
66 * without an RT task running, it can cause RT tasks to
67 * throttle when they start up. Kick the timer right away
68 * to update the period.
70 hrtimer_forward_now(&rt_b
->rt_period_timer
, ns_to_ktime(0));
71 hrtimer_start_expires(&rt_b
->rt_period_timer
, HRTIMER_MODE_ABS_PINNED
);
73 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
76 #if defined(CONFIG_SMP) && defined(HAVE_RT_PUSH_IPI)
77 static void push_irq_work_func(struct irq_work
*work
);
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
);
100 #ifdef HAVE_RT_PUSH_IPI
101 rt_rq
->push_flags
= 0;
102 rt_rq
->push_cpu
= nr_cpu_ids
;
103 raw_spin_lock_init(&rt_rq
->push_lock
);
104 init_irq_work(&rt_rq
->push_work
, push_irq_work_func
);
106 #endif /* CONFIG_SMP */
107 /* We start is dequeued state, because no RT tasks are queued */
108 rt_rq
->rt_queued
= 0;
111 rt_rq
->rt_throttled
= 0;
112 rt_rq
->rt_runtime
= 0;
113 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
116 #ifdef CONFIG_RT_GROUP_SCHED
117 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
119 hrtimer_cancel(&rt_b
->rt_period_timer
);
122 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
124 static inline struct task_struct
*rt_task_of(struct sched_rt_entity
*rt_se
)
126 #ifdef CONFIG_SCHED_DEBUG
127 WARN_ON_ONCE(!rt_entity_is_task(rt_se
));
129 return container_of(rt_se
, struct task_struct
, rt
);
132 static inline struct rq
*rq_of_rt_rq(struct rt_rq
*rt_rq
)
137 static inline struct rt_rq
*rt_rq_of_se(struct sched_rt_entity
*rt_se
)
142 static inline struct rq
*rq_of_rt_se(struct sched_rt_entity
*rt_se
)
144 struct rt_rq
*rt_rq
= rt_se
->rt_rq
;
149 void free_rt_sched_group(struct task_group
*tg
)
154 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
156 for_each_possible_cpu(i
) {
167 void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
168 struct sched_rt_entity
*rt_se
, int cpu
,
169 struct sched_rt_entity
*parent
)
171 struct rq
*rq
= cpu_rq(cpu
);
173 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
174 rt_rq
->rt_nr_boosted
= 0;
178 tg
->rt_rq
[cpu
] = rt_rq
;
179 tg
->rt_se
[cpu
] = rt_se
;
185 rt_se
->rt_rq
= &rq
->rt
;
187 rt_se
->rt_rq
= parent
->my_q
;
190 rt_se
->parent
= parent
;
191 INIT_LIST_HEAD(&rt_se
->run_list
);
194 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
197 struct sched_rt_entity
*rt_se
;
200 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
203 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
207 init_rt_bandwidth(&tg
->rt_bandwidth
,
208 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
210 for_each_possible_cpu(i
) {
211 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
212 GFP_KERNEL
, cpu_to_node(i
));
216 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
217 GFP_KERNEL
, cpu_to_node(i
));
222 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
223 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, parent
->rt_se
[i
]);
234 #else /* CONFIG_RT_GROUP_SCHED */
236 #define rt_entity_is_task(rt_se) (1)
238 static inline struct task_struct
*rt_task_of(struct sched_rt_entity
*rt_se
)
240 return container_of(rt_se
, struct task_struct
, rt
);
243 static inline struct rq
*rq_of_rt_rq(struct rt_rq
*rt_rq
)
245 return container_of(rt_rq
, struct rq
, rt
);
248 static inline struct rq
*rq_of_rt_se(struct sched_rt_entity
*rt_se
)
250 struct task_struct
*p
= rt_task_of(rt_se
);
255 static inline struct rt_rq
*rt_rq_of_se(struct sched_rt_entity
*rt_se
)
257 struct rq
*rq
= rq_of_rt_se(rt_se
);
262 void free_rt_sched_group(struct task_group
*tg
) { }
264 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
268 #endif /* CONFIG_RT_GROUP_SCHED */
272 static void pull_rt_task(struct rq
*this_rq
);
274 static inline bool need_pull_rt_task(struct rq
*rq
, struct task_struct
*prev
)
276 /* Try to pull RT tasks here if we lower this rq's prio */
277 return rq
->rt
.highest_prio
.curr
> prev
->prio
;
280 static inline int rt_overloaded(struct rq
*rq
)
282 return atomic_read(&rq
->rd
->rto_count
);
285 static inline void rt_set_overload(struct rq
*rq
)
290 cpumask_set_cpu(rq
->cpu
, rq
->rd
->rto_mask
);
292 * Make sure the mask is visible before we set
293 * the overload count. That is checked to determine
294 * if we should look at the mask. It would be a shame
295 * if we looked at the mask, but the mask was not
298 * Matched by the barrier in pull_rt_task().
301 atomic_inc(&rq
->rd
->rto_count
);
304 static inline void rt_clear_overload(struct rq
*rq
)
309 /* the order here really doesn't matter */
310 atomic_dec(&rq
->rd
->rto_count
);
311 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->rto_mask
);
314 static void update_rt_migration(struct rt_rq
*rt_rq
)
316 if (rt_rq
->rt_nr_migratory
&& rt_rq
->rt_nr_total
> 1) {
317 if (!rt_rq
->overloaded
) {
318 rt_set_overload(rq_of_rt_rq(rt_rq
));
319 rt_rq
->overloaded
= 1;
321 } else if (rt_rq
->overloaded
) {
322 rt_clear_overload(rq_of_rt_rq(rt_rq
));
323 rt_rq
->overloaded
= 0;
327 static void inc_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
329 struct task_struct
*p
;
331 if (!rt_entity_is_task(rt_se
))
334 p
= rt_task_of(rt_se
);
335 rt_rq
= &rq_of_rt_rq(rt_rq
)->rt
;
337 rt_rq
->rt_nr_total
++;
338 if (p
->nr_cpus_allowed
> 1)
339 rt_rq
->rt_nr_migratory
++;
341 update_rt_migration(rt_rq
);
344 static void dec_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
346 struct task_struct
*p
;
348 if (!rt_entity_is_task(rt_se
))
351 p
= rt_task_of(rt_se
);
352 rt_rq
= &rq_of_rt_rq(rt_rq
)->rt
;
354 rt_rq
->rt_nr_total
--;
355 if (p
->nr_cpus_allowed
> 1)
356 rt_rq
->rt_nr_migratory
--;
358 update_rt_migration(rt_rq
);
361 static inline int has_pushable_tasks(struct rq
*rq
)
363 return !plist_head_empty(&rq
->rt
.pushable_tasks
);
366 static DEFINE_PER_CPU(struct callback_head
, rt_push_head
);
367 static DEFINE_PER_CPU(struct callback_head
, rt_pull_head
);
369 static void push_rt_tasks(struct rq
*);
370 static void pull_rt_task(struct rq
*);
372 static inline void queue_push_tasks(struct rq
*rq
)
374 if (!has_pushable_tasks(rq
))
377 queue_balance_callback(rq
, &per_cpu(rt_push_head
, rq
->cpu
), push_rt_tasks
);
380 static inline void queue_pull_task(struct rq
*rq
)
382 queue_balance_callback(rq
, &per_cpu(rt_pull_head
, rq
->cpu
), pull_rt_task
);
385 static void enqueue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
387 plist_del(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
388 plist_node_init(&p
->pushable_tasks
, p
->prio
);
389 plist_add(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
391 /* Update the highest prio pushable task */
392 if (p
->prio
< rq
->rt
.highest_prio
.next
)
393 rq
->rt
.highest_prio
.next
= p
->prio
;
396 static void dequeue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
398 plist_del(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
400 /* Update the new highest prio pushable task */
401 if (has_pushable_tasks(rq
)) {
402 p
= plist_first_entry(&rq
->rt
.pushable_tasks
,
403 struct task_struct
, pushable_tasks
);
404 rq
->rt
.highest_prio
.next
= p
->prio
;
406 rq
->rt
.highest_prio
.next
= MAX_RT_PRIO
;
411 static inline void enqueue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
415 static inline void dequeue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
420 void inc_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
425 void dec_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
429 static inline bool need_pull_rt_task(struct rq
*rq
, struct task_struct
*prev
)
434 static inline void pull_rt_task(struct rq
*this_rq
)
438 static inline void queue_push_tasks(struct rq
*rq
)
441 #endif /* CONFIG_SMP */
443 static void enqueue_top_rt_rq(struct rt_rq
*rt_rq
);
444 static void dequeue_top_rt_rq(struct rt_rq
*rt_rq
);
446 static inline int on_rt_rq(struct sched_rt_entity
*rt_se
)
451 #ifdef CONFIG_RT_GROUP_SCHED
453 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
458 return rt_rq
->rt_runtime
;
461 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
463 return ktime_to_ns(rt_rq
->tg
->rt_bandwidth
.rt_period
);
466 typedef struct task_group
*rt_rq_iter_t
;
468 static inline struct task_group
*next_task_group(struct task_group
*tg
)
471 tg
= list_entry_rcu(tg
->list
.next
,
472 typeof(struct task_group
), list
);
473 } while (&tg
->list
!= &task_groups
&& task_group_is_autogroup(tg
));
475 if (&tg
->list
== &task_groups
)
481 #define for_each_rt_rq(rt_rq, iter, rq) \
482 for (iter = container_of(&task_groups, typeof(*iter), list); \
483 (iter = next_task_group(iter)) && \
484 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
486 #define for_each_sched_rt_entity(rt_se) \
487 for (; rt_se; rt_se = rt_se->parent)
489 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
494 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
);
495 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
);
497 static void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
499 struct task_struct
*curr
= rq_of_rt_rq(rt_rq
)->curr
;
500 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
501 struct sched_rt_entity
*rt_se
;
503 int cpu
= cpu_of(rq
);
505 rt_se
= rt_rq
->tg
->rt_se
[cpu
];
507 if (rt_rq
->rt_nr_running
) {
509 enqueue_top_rt_rq(rt_rq
);
510 else if (!on_rt_rq(rt_se
))
511 enqueue_rt_entity(rt_se
, 0);
513 if (rt_rq
->highest_prio
.curr
< curr
->prio
)
518 static void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
520 struct sched_rt_entity
*rt_se
;
521 int cpu
= cpu_of(rq_of_rt_rq(rt_rq
));
523 rt_se
= rt_rq
->tg
->rt_se
[cpu
];
526 dequeue_top_rt_rq(rt_rq
);
527 else if (on_rt_rq(rt_se
))
528 dequeue_rt_entity(rt_se
, 0);
531 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
533 return rt_rq
->rt_throttled
&& !rt_rq
->rt_nr_boosted
;
536 static int rt_se_boosted(struct sched_rt_entity
*rt_se
)
538 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
539 struct task_struct
*p
;
542 return !!rt_rq
->rt_nr_boosted
;
544 p
= rt_task_of(rt_se
);
545 return p
->prio
!= p
->normal_prio
;
549 static inline const struct cpumask
*sched_rt_period_mask(void)
551 return this_rq()->rd
->span
;
554 static inline const struct cpumask
*sched_rt_period_mask(void)
556 return cpu_online_mask
;
561 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
563 return container_of(rt_b
, struct task_group
, rt_bandwidth
)->rt_rq
[cpu
];
566 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
568 return &rt_rq
->tg
->rt_bandwidth
;
571 #else /* !CONFIG_RT_GROUP_SCHED */
573 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
575 return rt_rq
->rt_runtime
;
578 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
580 return ktime_to_ns(def_rt_bandwidth
.rt_period
);
583 typedef struct rt_rq
*rt_rq_iter_t
;
585 #define for_each_rt_rq(rt_rq, iter, rq) \
586 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
588 #define for_each_sched_rt_entity(rt_se) \
589 for (; rt_se; rt_se = NULL)
591 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
596 static inline void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
598 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
600 if (!rt_rq
->rt_nr_running
)
603 enqueue_top_rt_rq(rt_rq
);
607 static inline void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
609 dequeue_top_rt_rq(rt_rq
);
612 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
614 return rt_rq
->rt_throttled
;
617 static inline const struct cpumask
*sched_rt_period_mask(void)
619 return cpu_online_mask
;
623 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
625 return &cpu_rq(cpu
)->rt
;
628 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
630 return &def_rt_bandwidth
;
633 #endif /* CONFIG_RT_GROUP_SCHED */
635 bool sched_rt_bandwidth_account(struct rt_rq
*rt_rq
)
637 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
639 return (hrtimer_active(&rt_b
->rt_period_timer
) ||
640 rt_rq
->rt_time
< rt_b
->rt_runtime
);
645 * We ran out of runtime, see if we can borrow some from our neighbours.
647 static void do_balance_runtime(struct rt_rq
*rt_rq
)
649 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
650 struct root_domain
*rd
= rq_of_rt_rq(rt_rq
)->rd
;
654 weight
= cpumask_weight(rd
->span
);
656 raw_spin_lock(&rt_b
->rt_runtime_lock
);
657 rt_period
= ktime_to_ns(rt_b
->rt_period
);
658 for_each_cpu(i
, rd
->span
) {
659 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
665 raw_spin_lock(&iter
->rt_runtime_lock
);
667 * Either all rqs have inf runtime and there's nothing to steal
668 * or __disable_runtime() below sets a specific rq to inf to
669 * indicate its been disabled and disalow stealing.
671 if (iter
->rt_runtime
== RUNTIME_INF
)
675 * From runqueues with spare time, take 1/n part of their
676 * spare time, but no more than our period.
678 diff
= iter
->rt_runtime
- iter
->rt_time
;
680 diff
= div_u64((u64
)diff
, weight
);
681 if (rt_rq
->rt_runtime
+ diff
> rt_period
)
682 diff
= rt_period
- rt_rq
->rt_runtime
;
683 iter
->rt_runtime
-= diff
;
684 rt_rq
->rt_runtime
+= diff
;
685 if (rt_rq
->rt_runtime
== rt_period
) {
686 raw_spin_unlock(&iter
->rt_runtime_lock
);
691 raw_spin_unlock(&iter
->rt_runtime_lock
);
693 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
697 * Ensure this RQ takes back all the runtime it lend to its neighbours.
699 static void __disable_runtime(struct rq
*rq
)
701 struct root_domain
*rd
= rq
->rd
;
705 if (unlikely(!scheduler_running
))
708 for_each_rt_rq(rt_rq
, iter
, rq
) {
709 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
713 raw_spin_lock(&rt_b
->rt_runtime_lock
);
714 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
716 * Either we're all inf and nobody needs to borrow, or we're
717 * already disabled and thus have nothing to do, or we have
718 * exactly the right amount of runtime to take out.
720 if (rt_rq
->rt_runtime
== RUNTIME_INF
||
721 rt_rq
->rt_runtime
== rt_b
->rt_runtime
)
723 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
726 * Calculate the difference between what we started out with
727 * and what we current have, that's the amount of runtime
728 * we lend and now have to reclaim.
730 want
= rt_b
->rt_runtime
- rt_rq
->rt_runtime
;
733 * Greedy reclaim, take back as much as we can.
735 for_each_cpu(i
, rd
->span
) {
736 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
740 * Can't reclaim from ourselves or disabled runqueues.
742 if (iter
== rt_rq
|| iter
->rt_runtime
== RUNTIME_INF
)
745 raw_spin_lock(&iter
->rt_runtime_lock
);
747 diff
= min_t(s64
, iter
->rt_runtime
, want
);
748 iter
->rt_runtime
-= diff
;
751 iter
->rt_runtime
-= want
;
754 raw_spin_unlock(&iter
->rt_runtime_lock
);
760 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
762 * We cannot be left wanting - that would mean some runtime
763 * leaked out of the system.
768 * Disable all the borrow logic by pretending we have inf
769 * runtime - in which case borrowing doesn't make sense.
771 rt_rq
->rt_runtime
= RUNTIME_INF
;
772 rt_rq
->rt_throttled
= 0;
773 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
774 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
776 /* Make rt_rq available for pick_next_task() */
777 sched_rt_rq_enqueue(rt_rq
);
781 static void __enable_runtime(struct rq
*rq
)
786 if (unlikely(!scheduler_running
))
790 * Reset each runqueue's bandwidth settings
792 for_each_rt_rq(rt_rq
, iter
, rq
) {
793 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
795 raw_spin_lock(&rt_b
->rt_runtime_lock
);
796 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
797 rt_rq
->rt_runtime
= rt_b
->rt_runtime
;
799 rt_rq
->rt_throttled
= 0;
800 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
801 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
805 static void balance_runtime(struct rt_rq
*rt_rq
)
807 if (!sched_feat(RT_RUNTIME_SHARE
))
810 if (rt_rq
->rt_time
> rt_rq
->rt_runtime
) {
811 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
812 do_balance_runtime(rt_rq
);
813 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
816 #else /* !CONFIG_SMP */
817 static inline void balance_runtime(struct rt_rq
*rt_rq
) {}
818 #endif /* CONFIG_SMP */
820 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
)
822 int i
, idle
= 1, throttled
= 0;
823 const struct cpumask
*span
;
825 span
= sched_rt_period_mask();
826 #ifdef CONFIG_RT_GROUP_SCHED
828 * FIXME: isolated CPUs should really leave the root task group,
829 * whether they are isolcpus or were isolated via cpusets, lest
830 * the timer run on a CPU which does not service all runqueues,
831 * potentially leaving other CPUs indefinitely throttled. If
832 * isolation is really required, the user will turn the throttle
833 * off to kill the perturbations it causes anyway. Meanwhile,
834 * this maintains functionality for boot and/or troubleshooting.
836 if (rt_b
== &root_task_group
.rt_bandwidth
)
837 span
= cpu_online_mask
;
839 for_each_cpu(i
, span
) {
841 struct rt_rq
*rt_rq
= sched_rt_period_rt_rq(rt_b
, i
);
842 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
846 * When span == cpu_online_mask, taking each rq->lock
847 * can be time-consuming. Try to avoid it when possible.
849 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
850 skip
= !rt_rq
->rt_time
&& !rt_rq
->rt_nr_running
;
851 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
855 raw_spin_lock(&rq
->lock
);
856 if (rt_rq
->rt_time
) {
859 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
860 if (rt_rq
->rt_throttled
)
861 balance_runtime(rt_rq
);
862 runtime
= rt_rq
->rt_runtime
;
863 rt_rq
->rt_time
-= min(rt_rq
->rt_time
, overrun
*runtime
);
864 if (rt_rq
->rt_throttled
&& rt_rq
->rt_time
< runtime
) {
865 rt_rq
->rt_throttled
= 0;
869 * When we're idle and a woken (rt) task is
870 * throttled check_preempt_curr() will set
871 * skip_update and the time between the wakeup
872 * and this unthrottle will get accounted as
875 if (rt_rq
->rt_nr_running
&& rq
->curr
== rq
->idle
)
876 rq_clock_skip_update(rq
, false);
878 if (rt_rq
->rt_time
|| rt_rq
->rt_nr_running
)
880 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
881 } else if (rt_rq
->rt_nr_running
) {
883 if (!rt_rq_throttled(rt_rq
))
886 if (rt_rq
->rt_throttled
)
890 sched_rt_rq_enqueue(rt_rq
);
891 raw_spin_unlock(&rq
->lock
);
894 if (!throttled
&& (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
))
900 static inline int rt_se_prio(struct sched_rt_entity
*rt_se
)
902 #ifdef CONFIG_RT_GROUP_SCHED
903 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
906 return rt_rq
->highest_prio
.curr
;
909 return rt_task_of(rt_se
)->prio
;
912 static int sched_rt_runtime_exceeded(struct rt_rq
*rt_rq
)
914 u64 runtime
= sched_rt_runtime(rt_rq
);
916 if (rt_rq
->rt_throttled
)
917 return rt_rq_throttled(rt_rq
);
919 if (runtime
>= sched_rt_period(rt_rq
))
922 balance_runtime(rt_rq
);
923 runtime
= sched_rt_runtime(rt_rq
);
924 if (runtime
== RUNTIME_INF
)
927 if (rt_rq
->rt_time
> runtime
) {
928 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
931 * Don't actually throttle groups that have no runtime assigned
932 * but accrue some time due to boosting.
934 if (likely(rt_b
->rt_runtime
)) {
935 rt_rq
->rt_throttled
= 1;
936 printk_deferred_once("sched: RT throttling activated\n");
939 * In case we did anyway, make it go away,
940 * replenishment is a joke, since it will replenish us
946 if (rt_rq_throttled(rt_rq
)) {
947 sched_rt_rq_dequeue(rt_rq
);
956 * Update the current task's runtime statistics. Skip current tasks that
957 * are not in our scheduling class.
959 static void update_curr_rt(struct rq
*rq
)
961 struct task_struct
*curr
= rq
->curr
;
962 struct sched_rt_entity
*rt_se
= &curr
->rt
;
965 if (curr
->sched_class
!= &rt_sched_class
)
968 delta_exec
= rq_clock_task(rq
) - curr
->se
.exec_start
;
969 if (unlikely((s64
)delta_exec
<= 0))
972 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
973 cpufreq_update_this_cpu(rq
, SCHED_CPUFREQ_RT
);
975 schedstat_set(curr
->se
.statistics
.exec_max
,
976 max(curr
->se
.statistics
.exec_max
, delta_exec
));
978 curr
->se
.sum_exec_runtime
+= delta_exec
;
979 account_group_exec_runtime(curr
, delta_exec
);
981 curr
->se
.exec_start
= rq_clock_task(rq
);
982 cpuacct_charge(curr
, delta_exec
);
984 sched_rt_avg_update(rq
, delta_exec
);
986 if (!rt_bandwidth_enabled())
989 for_each_sched_rt_entity(rt_se
) {
990 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
992 if (sched_rt_runtime(rt_rq
) != RUNTIME_INF
) {
993 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
994 rt_rq
->rt_time
+= delta_exec
;
995 if (sched_rt_runtime_exceeded(rt_rq
))
997 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
1003 dequeue_top_rt_rq(struct rt_rq
*rt_rq
)
1005 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1007 BUG_ON(&rq
->rt
!= rt_rq
);
1009 if (!rt_rq
->rt_queued
)
1012 BUG_ON(!rq
->nr_running
);
1014 sub_nr_running(rq
, rt_rq
->rt_nr_running
);
1015 rt_rq
->rt_queued
= 0;
1019 enqueue_top_rt_rq(struct rt_rq
*rt_rq
)
1021 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1023 BUG_ON(&rq
->rt
!= rt_rq
);
1025 if (rt_rq
->rt_queued
)
1027 if (rt_rq_throttled(rt_rq
) || !rt_rq
->rt_nr_running
)
1030 add_nr_running(rq
, rt_rq
->rt_nr_running
);
1031 rt_rq
->rt_queued
= 1;
1034 #if defined CONFIG_SMP
1037 inc_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
)
1039 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1041 #ifdef CONFIG_RT_GROUP_SCHED
1043 * Change rq's cpupri only if rt_rq is the top queue.
1045 if (&rq
->rt
!= rt_rq
)
1048 if (rq
->online
&& prio
< prev_prio
)
1049 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, prio
);
1053 dec_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
)
1055 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1057 #ifdef CONFIG_RT_GROUP_SCHED
1059 * Change rq's cpupri only if rt_rq is the top queue.
1061 if (&rq
->rt
!= rt_rq
)
1064 if (rq
->online
&& rt_rq
->highest_prio
.curr
!= prev_prio
)
1065 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rt_rq
->highest_prio
.curr
);
1068 #else /* CONFIG_SMP */
1071 void inc_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
) {}
1073 void dec_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
) {}
1075 #endif /* CONFIG_SMP */
1077 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1079 inc_rt_prio(struct rt_rq
*rt_rq
, int prio
)
1081 int prev_prio
= rt_rq
->highest_prio
.curr
;
1083 if (prio
< prev_prio
)
1084 rt_rq
->highest_prio
.curr
= prio
;
1086 inc_rt_prio_smp(rt_rq
, prio
, prev_prio
);
1090 dec_rt_prio(struct rt_rq
*rt_rq
, int prio
)
1092 int prev_prio
= rt_rq
->highest_prio
.curr
;
1094 if (rt_rq
->rt_nr_running
) {
1096 WARN_ON(prio
< prev_prio
);
1099 * This may have been our highest task, and therefore
1100 * we may have some recomputation to do
1102 if (prio
== prev_prio
) {
1103 struct rt_prio_array
*array
= &rt_rq
->active
;
1105 rt_rq
->highest_prio
.curr
=
1106 sched_find_first_bit(array
->bitmap
);
1110 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
1112 dec_rt_prio_smp(rt_rq
, prio
, prev_prio
);
1117 static inline void inc_rt_prio(struct rt_rq
*rt_rq
, int prio
) {}
1118 static inline void dec_rt_prio(struct rt_rq
*rt_rq
, int prio
) {}
1120 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1122 #ifdef CONFIG_RT_GROUP_SCHED
1125 inc_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1127 if (rt_se_boosted(rt_se
))
1128 rt_rq
->rt_nr_boosted
++;
1131 start_rt_bandwidth(&rt_rq
->tg
->rt_bandwidth
);
1135 dec_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1137 if (rt_se_boosted(rt_se
))
1138 rt_rq
->rt_nr_boosted
--;
1140 WARN_ON(!rt_rq
->rt_nr_running
&& rt_rq
->rt_nr_boosted
);
1143 #else /* CONFIG_RT_GROUP_SCHED */
1146 inc_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1148 start_rt_bandwidth(&def_rt_bandwidth
);
1152 void dec_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
) {}
1154 #endif /* CONFIG_RT_GROUP_SCHED */
1157 unsigned int rt_se_nr_running(struct sched_rt_entity
*rt_se
)
1159 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1162 return group_rq
->rt_nr_running
;
1168 unsigned int rt_se_rr_nr_running(struct sched_rt_entity
*rt_se
)
1170 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1171 struct task_struct
*tsk
;
1174 return group_rq
->rr_nr_running
;
1176 tsk
= rt_task_of(rt_se
);
1178 return (tsk
->policy
== SCHED_RR
) ? 1 : 0;
1182 void inc_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1184 int prio
= rt_se_prio(rt_se
);
1186 WARN_ON(!rt_prio(prio
));
1187 rt_rq
->rt_nr_running
+= rt_se_nr_running(rt_se
);
1188 rt_rq
->rr_nr_running
+= rt_se_rr_nr_running(rt_se
);
1190 inc_rt_prio(rt_rq
, prio
);
1191 inc_rt_migration(rt_se
, rt_rq
);
1192 inc_rt_group(rt_se
, rt_rq
);
1196 void dec_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1198 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
1199 WARN_ON(!rt_rq
->rt_nr_running
);
1200 rt_rq
->rt_nr_running
-= rt_se_nr_running(rt_se
);
1201 rt_rq
->rr_nr_running
-= rt_se_rr_nr_running(rt_se
);
1203 dec_rt_prio(rt_rq
, rt_se_prio(rt_se
));
1204 dec_rt_migration(rt_se
, rt_rq
);
1205 dec_rt_group(rt_se
, rt_rq
);
1209 * Change rt_se->run_list location unless SAVE && !MOVE
1211 * assumes ENQUEUE/DEQUEUE flags match
1213 static inline bool move_entity(unsigned int flags
)
1215 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) == DEQUEUE_SAVE
)
1221 static void __delist_rt_entity(struct sched_rt_entity
*rt_se
, struct rt_prio_array
*array
)
1223 list_del_init(&rt_se
->run_list
);
1225 if (list_empty(array
->queue
+ rt_se_prio(rt_se
)))
1226 __clear_bit(rt_se_prio(rt_se
), array
->bitmap
);
1231 static void __enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1233 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
1234 struct rt_prio_array
*array
= &rt_rq
->active
;
1235 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1236 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
1239 * Don't enqueue the group if its throttled, or when empty.
1240 * The latter is a consequence of the former when a child group
1241 * get throttled and the current group doesn't have any other
1244 if (group_rq
&& (rt_rq_throttled(group_rq
) || !group_rq
->rt_nr_running
)) {
1246 __delist_rt_entity(rt_se
, array
);
1250 if (move_entity(flags
)) {
1251 WARN_ON_ONCE(rt_se
->on_list
);
1252 if (flags
& ENQUEUE_HEAD
)
1253 list_add(&rt_se
->run_list
, queue
);
1255 list_add_tail(&rt_se
->run_list
, queue
);
1257 __set_bit(rt_se_prio(rt_se
), array
->bitmap
);
1262 inc_rt_tasks(rt_se
, rt_rq
);
1265 static void __dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1267 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
1268 struct rt_prio_array
*array
= &rt_rq
->active
;
1270 if (move_entity(flags
)) {
1271 WARN_ON_ONCE(!rt_se
->on_list
);
1272 __delist_rt_entity(rt_se
, array
);
1276 dec_rt_tasks(rt_se
, rt_rq
);
1280 * Because the prio of an upper entry depends on the lower
1281 * entries, we must remove entries top - down.
1283 static void dequeue_rt_stack(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1285 struct sched_rt_entity
*back
= NULL
;
1287 for_each_sched_rt_entity(rt_se
) {
1292 dequeue_top_rt_rq(rt_rq_of_se(back
));
1294 for (rt_se
= back
; rt_se
; rt_se
= rt_se
->back
) {
1295 if (on_rt_rq(rt_se
))
1296 __dequeue_rt_entity(rt_se
, flags
);
1300 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1302 struct rq
*rq
= rq_of_rt_se(rt_se
);
1304 dequeue_rt_stack(rt_se
, flags
);
1305 for_each_sched_rt_entity(rt_se
)
1306 __enqueue_rt_entity(rt_se
, flags
);
1307 enqueue_top_rt_rq(&rq
->rt
);
1310 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1312 struct rq
*rq
= rq_of_rt_se(rt_se
);
1314 dequeue_rt_stack(rt_se
, flags
);
1316 for_each_sched_rt_entity(rt_se
) {
1317 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
1319 if (rt_rq
&& rt_rq
->rt_nr_running
)
1320 __enqueue_rt_entity(rt_se
, flags
);
1322 enqueue_top_rt_rq(&rq
->rt
);
1326 * Adding/removing a task to/from a priority array:
1329 enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1331 struct sched_rt_entity
*rt_se
= &p
->rt
;
1333 if (flags
& ENQUEUE_WAKEUP
)
1336 enqueue_rt_entity(rt_se
, flags
);
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
;
1347 dequeue_rt_entity(rt_se
, flags
);
1349 dequeue_pushable_task(rq
, p
);
1353 * Put task to the head or the end of the run list without the overhead of
1354 * dequeue followed by enqueue.
1357 requeue_rt_entity(struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
, int head
)
1359 if (on_rt_rq(rt_se
)) {
1360 struct rt_prio_array
*array
= &rt_rq
->active
;
1361 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
1364 list_move(&rt_se
->run_list
, queue
);
1366 list_move_tail(&rt_se
->run_list
, queue
);
1370 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int head
)
1372 struct sched_rt_entity
*rt_se
= &p
->rt
;
1373 struct rt_rq
*rt_rq
;
1375 for_each_sched_rt_entity(rt_se
) {
1376 rt_rq
= rt_rq_of_se(rt_se
);
1377 requeue_rt_entity(rt_rq
, rt_se
, head
);
1381 static void yield_task_rt(struct rq
*rq
)
1383 requeue_task_rt(rq
, rq
->curr
, 0);
1387 static int find_lowest_rq(struct task_struct
*task
);
1390 select_task_rq_rt(struct task_struct
*p
, int cpu
, int sd_flag
, int flags
)
1392 struct task_struct
*curr
;
1395 /* For anything but wake ups, just return the task_cpu */
1396 if (sd_flag
!= SD_BALANCE_WAKE
&& sd_flag
!= SD_BALANCE_FORK
)
1402 curr
= READ_ONCE(rq
->curr
); /* unlocked access */
1405 * If the current task on @p's runqueue is an RT task, then
1406 * try to see if we can wake this RT task up on another
1407 * runqueue. Otherwise simply start this RT task
1408 * on its current runqueue.
1410 * We want to avoid overloading runqueues. If the woken
1411 * task is a higher priority, then it will stay on this CPU
1412 * and the lower prio task should be moved to another CPU.
1413 * Even though this will probably make the lower prio task
1414 * lose its cache, we do not want to bounce a higher task
1415 * around just because it gave up its CPU, perhaps for a
1418 * For equal prio tasks, we just let the scheduler sort it out.
1420 * Otherwise, just let it ride on the affined RQ and the
1421 * post-schedule router will push the preempted task away
1423 * This test is optimistic, if we get it wrong the load-balancer
1424 * will have to sort it out.
1426 if (curr
&& unlikely(rt_task(curr
)) &&
1427 (curr
->nr_cpus_allowed
< 2 ||
1428 curr
->prio
<= p
->prio
)) {
1429 int target
= find_lowest_rq(p
);
1432 * Don't bother moving it if the destination CPU is
1433 * not running a lower priority task.
1436 p
->prio
< cpu_rq(target
)->rt
.highest_prio
.curr
)
1445 static void check_preempt_equal_prio(struct rq
*rq
, struct task_struct
*p
)
1448 * Current can't be migrated, useless to reschedule,
1449 * let's hope p can move out.
1451 if (rq
->curr
->nr_cpus_allowed
== 1 ||
1452 !cpupri_find(&rq
->rd
->cpupri
, rq
->curr
, NULL
))
1456 * p is migratable, so let's not schedule it and
1457 * see if it is pushed or pulled somewhere else.
1459 if (p
->nr_cpus_allowed
!= 1
1460 && cpupri_find(&rq
->rd
->cpupri
, p
, NULL
))
1464 * There appears to be other cpus that can accept
1465 * current and none to run 'p', so lets reschedule
1466 * to try and push current away:
1468 requeue_task_rt(rq
, p
, 1);
1472 #endif /* CONFIG_SMP */
1475 * Preempt the current task with a newly woken task if needed:
1477 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1479 if (p
->prio
< rq
->curr
->prio
) {
1488 * - the newly woken task is of equal priority to the current task
1489 * - the newly woken task is non-migratable while current is migratable
1490 * - current will be preempted on the next reschedule
1492 * we should check to see if current can readily move to a different
1493 * cpu. If so, we will reschedule to allow the push logic to try
1494 * to move current somewhere else, making room for our non-migratable
1497 if (p
->prio
== rq
->curr
->prio
&& !test_tsk_need_resched(rq
->curr
))
1498 check_preempt_equal_prio(rq
, p
);
1502 static struct sched_rt_entity
*pick_next_rt_entity(struct rq
*rq
,
1503 struct rt_rq
*rt_rq
)
1505 struct rt_prio_array
*array
= &rt_rq
->active
;
1506 struct sched_rt_entity
*next
= NULL
;
1507 struct list_head
*queue
;
1510 idx
= sched_find_first_bit(array
->bitmap
);
1511 BUG_ON(idx
>= MAX_RT_PRIO
);
1513 queue
= array
->queue
+ idx
;
1514 next
= list_entry(queue
->next
, struct sched_rt_entity
, run_list
);
1519 static struct task_struct
*_pick_next_task_rt(struct rq
*rq
)
1521 struct sched_rt_entity
*rt_se
;
1522 struct task_struct
*p
;
1523 struct rt_rq
*rt_rq
= &rq
->rt
;
1526 rt_se
= pick_next_rt_entity(rq
, rt_rq
);
1528 rt_rq
= group_rt_rq(rt_se
);
1531 p
= rt_task_of(rt_se
);
1532 p
->se
.exec_start
= rq_clock_task(rq
);
1537 static struct task_struct
*
1538 pick_next_task_rt(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
1540 struct task_struct
*p
;
1541 struct rt_rq
*rt_rq
= &rq
->rt
;
1543 if (need_pull_rt_task(rq
, prev
)) {
1545 * This is OK, because current is on_cpu, which avoids it being
1546 * picked for load-balance and preemption/IRQs are still
1547 * disabled avoiding further scheduler activity on it and we're
1548 * being very careful to re-start the picking loop.
1550 rq_unpin_lock(rq
, rf
);
1552 rq_repin_lock(rq
, rf
);
1554 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1555 * means a dl or stop task can slip in, in which case we need
1556 * to re-start task selection.
1558 if (unlikely((rq
->stop
&& task_on_rq_queued(rq
->stop
)) ||
1559 rq
->dl
.dl_nr_running
))
1564 * We may dequeue prev's rt_rq in put_prev_task().
1565 * So, we update time before rt_nr_running check.
1567 if (prev
->sched_class
== &rt_sched_class
)
1570 if (!rt_rq
->rt_queued
)
1573 put_prev_task(rq
, prev
);
1575 p
= _pick_next_task_rt(rq
);
1577 /* The running task is never eligible for pushing */
1578 dequeue_pushable_task(rq
, p
);
1580 queue_push_tasks(rq
);
1585 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
1590 * The previous task needs to be made eligible for pushing
1591 * if it is still active
1593 if (on_rt_rq(&p
->rt
) && p
->nr_cpus_allowed
> 1)
1594 enqueue_pushable_task(rq
, p
);
1599 /* Only try algorithms three times */
1600 #define RT_MAX_TRIES 3
1602 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
1604 if (!task_running(rq
, p
) &&
1605 cpumask_test_cpu(cpu
, &p
->cpus_allowed
))
1611 * Return the highest pushable rq's task, which is suitable to be executed
1612 * on the cpu, NULL otherwise
1614 static struct task_struct
*pick_highest_pushable_task(struct rq
*rq
, int cpu
)
1616 struct plist_head
*head
= &rq
->rt
.pushable_tasks
;
1617 struct task_struct
*p
;
1619 if (!has_pushable_tasks(rq
))
1622 plist_for_each_entry(p
, head
, pushable_tasks
) {
1623 if (pick_rt_task(rq
, p
, cpu
))
1630 static DEFINE_PER_CPU(cpumask_var_t
, local_cpu_mask
);
1632 static int find_lowest_rq(struct task_struct
*task
)
1634 struct sched_domain
*sd
;
1635 struct cpumask
*lowest_mask
= this_cpu_cpumask_var_ptr(local_cpu_mask
);
1636 int this_cpu
= smp_processor_id();
1637 int cpu
= task_cpu(task
);
1639 /* Make sure the mask is initialized first */
1640 if (unlikely(!lowest_mask
))
1643 if (task
->nr_cpus_allowed
== 1)
1644 return -1; /* No other targets possible */
1646 if (!cpupri_find(&task_rq(task
)->rd
->cpupri
, task
, lowest_mask
))
1647 return -1; /* No targets found */
1650 * At this point we have built a mask of cpus representing the
1651 * lowest priority tasks in the system. Now we want to elect
1652 * the best one based on our affinity and topology.
1654 * We prioritize the last cpu that the task executed on since
1655 * it is most likely cache-hot in that location.
1657 if (cpumask_test_cpu(cpu
, lowest_mask
))
1661 * Otherwise, we consult the sched_domains span maps to figure
1662 * out which cpu is logically closest to our hot cache data.
1664 if (!cpumask_test_cpu(this_cpu
, lowest_mask
))
1665 this_cpu
= -1; /* Skip this_cpu opt if not among lowest */
1668 for_each_domain(cpu
, sd
) {
1669 if (sd
->flags
& SD_WAKE_AFFINE
) {
1673 * "this_cpu" is cheaper to preempt than a
1676 if (this_cpu
!= -1 &&
1677 cpumask_test_cpu(this_cpu
, sched_domain_span(sd
))) {
1682 best_cpu
= cpumask_first_and(lowest_mask
,
1683 sched_domain_span(sd
));
1684 if (best_cpu
< nr_cpu_ids
) {
1693 * And finally, if there were no matches within the domains
1694 * just give the caller *something* to work with from the compatible
1700 cpu
= cpumask_any(lowest_mask
);
1701 if (cpu
< nr_cpu_ids
)
1706 /* Will lock the rq it finds */
1707 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
1709 struct rq
*lowest_rq
= NULL
;
1713 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
1714 cpu
= find_lowest_rq(task
);
1716 if ((cpu
== -1) || (cpu
== rq
->cpu
))
1719 lowest_rq
= cpu_rq(cpu
);
1721 if (lowest_rq
->rt
.highest_prio
.curr
<= task
->prio
) {
1723 * Target rq has tasks of equal or higher priority,
1724 * retrying does not release any lock and is unlikely
1725 * to yield a different result.
1731 /* if the prio of this runqueue changed, try again */
1732 if (double_lock_balance(rq
, lowest_rq
)) {
1734 * We had to unlock the run queue. In
1735 * the mean time, task could have
1736 * migrated already or had its affinity changed.
1737 * Also make sure that it wasn't scheduled on its rq.
1739 if (unlikely(task_rq(task
) != rq
||
1740 !cpumask_test_cpu(lowest_rq
->cpu
, &task
->cpus_allowed
) ||
1741 task_running(rq
, task
) ||
1743 !task_on_rq_queued(task
))) {
1745 double_unlock_balance(rq
, lowest_rq
);
1751 /* If this rq is still suitable use it. */
1752 if (lowest_rq
->rt
.highest_prio
.curr
> task
->prio
)
1756 double_unlock_balance(rq
, lowest_rq
);
1763 static struct task_struct
*pick_next_pushable_task(struct rq
*rq
)
1765 struct task_struct
*p
;
1767 if (!has_pushable_tasks(rq
))
1770 p
= plist_first_entry(&rq
->rt
.pushable_tasks
,
1771 struct task_struct
, pushable_tasks
);
1773 BUG_ON(rq
->cpu
!= task_cpu(p
));
1774 BUG_ON(task_current(rq
, p
));
1775 BUG_ON(p
->nr_cpus_allowed
<= 1);
1777 BUG_ON(!task_on_rq_queued(p
));
1778 BUG_ON(!rt_task(p
));
1784 * If the current CPU has more than one RT task, see if the non
1785 * running task can migrate over to a CPU that is running a task
1786 * of lesser priority.
1788 static int push_rt_task(struct rq
*rq
)
1790 struct task_struct
*next_task
;
1791 struct rq
*lowest_rq
;
1794 if (!rq
->rt
.overloaded
)
1797 next_task
= pick_next_pushable_task(rq
);
1802 if (unlikely(next_task
== rq
->curr
)) {
1808 * It's possible that the next_task slipped in of
1809 * higher priority than current. If that's the case
1810 * just reschedule current.
1812 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
1817 /* We might release rq lock */
1818 get_task_struct(next_task
);
1820 /* find_lock_lowest_rq locks the rq if found */
1821 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
1823 struct task_struct
*task
;
1825 * find_lock_lowest_rq releases rq->lock
1826 * so it is possible that next_task has migrated.
1828 * We need to make sure that the task is still on the same
1829 * run-queue and is also still the next task eligible for
1832 task
= pick_next_pushable_task(rq
);
1833 if (task
== next_task
) {
1835 * The task hasn't migrated, and is still the next
1836 * eligible task, but we failed to find a run-queue
1837 * to push it to. Do not retry in this case, since
1838 * other cpus will pull from us when ready.
1844 /* No more tasks, just exit */
1848 * Something has shifted, try again.
1850 put_task_struct(next_task
);
1855 deactivate_task(rq
, next_task
, 0);
1856 set_task_cpu(next_task
, lowest_rq
->cpu
);
1857 activate_task(lowest_rq
, next_task
, 0);
1860 resched_curr(lowest_rq
);
1862 double_unlock_balance(rq
, lowest_rq
);
1865 put_task_struct(next_task
);
1870 static void push_rt_tasks(struct rq
*rq
)
1872 /* push_rt_task will return true if it moved an RT */
1873 while (push_rt_task(rq
))
1877 #ifdef HAVE_RT_PUSH_IPI
1879 * The search for the next cpu always starts at rq->cpu and ends
1880 * when we reach rq->cpu again. It will never return rq->cpu.
1881 * This returns the next cpu to check, or nr_cpu_ids if the loop
1884 * rq->rt.push_cpu holds the last cpu returned by this function,
1885 * or if this is the first instance, it must hold rq->cpu.
1887 static int rto_next_cpu(struct rq
*rq
)
1889 int prev_cpu
= rq
->rt
.push_cpu
;
1892 cpu
= cpumask_next(prev_cpu
, rq
->rd
->rto_mask
);
1895 * If the previous cpu is less than the rq's CPU, then it already
1896 * passed the end of the mask, and has started from the beginning.
1897 * We end if the next CPU is greater or equal to rq's CPU.
1899 if (prev_cpu
< rq
->cpu
) {
1903 } else if (cpu
>= nr_cpu_ids
) {
1905 * We passed the end of the mask, start at the beginning.
1906 * If the result is greater or equal to the rq's CPU, then
1907 * the loop is finished.
1909 cpu
= cpumask_first(rq
->rd
->rto_mask
);
1913 rq
->rt
.push_cpu
= cpu
;
1915 /* Return cpu to let the caller know if the loop is finished or not */
1919 static int find_next_push_cpu(struct rq
*rq
)
1925 cpu
= rto_next_cpu(rq
);
1926 if (cpu
>= nr_cpu_ids
)
1928 next_rq
= cpu_rq(cpu
);
1930 /* Make sure the next rq can push to this rq */
1931 if (next_rq
->rt
.highest_prio
.next
< rq
->rt
.highest_prio
.curr
)
1938 #define RT_PUSH_IPI_EXECUTING 1
1939 #define RT_PUSH_IPI_RESTART 2
1942 * When a high priority task schedules out from a CPU and a lower priority
1943 * task is scheduled in, a check is made to see if there's any RT tasks
1944 * on other CPUs that are waiting to run because a higher priority RT task
1945 * is currently running on its CPU. In this case, the CPU with multiple RT
1946 * tasks queued on it (overloaded) needs to be notified that a CPU has opened
1947 * up that may be able to run one of its non-running queued RT tasks.
1949 * On large CPU boxes, there's the case that several CPUs could schedule
1950 * a lower priority task at the same time, in which case it will look for
1951 * any overloaded CPUs that it could pull a task from. To do this, the runqueue
1952 * lock must be taken from that overloaded CPU. Having 10s of CPUs all fighting
1953 * for a single overloaded CPU's runqueue lock can produce a large latency.
1954 * (This has actually been observed on large boxes running cyclictest).
1955 * Instead of taking the runqueue lock of the overloaded CPU, each of the
1956 * CPUs that scheduled a lower priority task simply sends an IPI to the
1957 * overloaded CPU. An IPI is much cheaper than taking an runqueue lock with
1958 * lots of contention. The overloaded CPU will look to push its non-running
1959 * RT task off, and if it does, it can then ignore the other IPIs coming
1960 * in, and just pass those IPIs off to any other overloaded CPU.
1962 * When a CPU schedules a lower priority task, it only sends an IPI to
1963 * the "next" CPU that has overloaded RT tasks. This prevents IPI storms,
1964 * as having 10 CPUs scheduling lower priority tasks and 10 CPUs with
1965 * RT overloaded tasks, would cause 100 IPIs to go out at once.
1967 * The overloaded RT CPU, when receiving an IPI, will try to push off its
1968 * overloaded RT tasks and then send an IPI to the next CPU that has
1969 * overloaded RT tasks. This stops when all CPUs with overloaded RT tasks
1970 * have completed. Just because a CPU may have pushed off its own overloaded
1971 * RT task does not mean it should stop sending the IPI around to other
1972 * overloaded CPUs. There may be another RT task waiting to run on one of
1973 * those CPUs that are of higher priority than the one that was just
1976 * An optimization that could possibly be made is to make a CPU array similar
1977 * to the cpupri array mask of all running RT tasks, but for the overloaded
1978 * case, then the IPI could be sent to only the CPU with the highest priority
1979 * RT task waiting, and that CPU could send off further IPIs to the CPU with
1980 * the next highest waiting task. Since the overloaded case is much less likely
1981 * to happen, the complexity of this implementation may not be worth it.
1982 * Instead, just send an IPI around to all overloaded CPUs.
1984 * The rq->rt.push_flags holds the status of the IPI that is going around.
1985 * A run queue can only send out a single IPI at a time. The possible flags
1986 * for rq->rt.push_flags are:
1988 * (None or zero): No IPI is going around for the current rq
1989 * RT_PUSH_IPI_EXECUTING: An IPI for the rq is being passed around
1990 * RT_PUSH_IPI_RESTART: The priority of the running task for the rq
1991 * has changed, and the IPI should restart
1992 * circulating the overloaded CPUs again.
1994 * rq->rt.push_cpu contains the CPU that is being sent the IPI. It is updated
1995 * before sending to the next CPU.
1997 * Instead of having all CPUs that schedule a lower priority task send
1998 * an IPI to the same "first" CPU in the RT overload mask, they send it
1999 * to the next overloaded CPU after their own CPU. This helps distribute
2000 * the work when there's more than one overloaded CPU and multiple CPUs
2001 * scheduling in lower priority tasks.
2003 * When a rq schedules a lower priority task than what was currently
2004 * running, the next CPU with overloaded RT tasks is examined first.
2005 * That is, if CPU 1 and 5 are overloaded, and CPU 3 schedules a lower
2006 * priority task, it will send an IPI first to CPU 5, then CPU 5 will
2007 * send to CPU 1 if it is still overloaded. CPU 1 will clear the
2008 * rq->rt.push_flags if RT_PUSH_IPI_RESTART is not set.
2010 * The first CPU to notice IPI_RESTART is set, will clear that flag and then
2011 * send an IPI to the next overloaded CPU after the rq->cpu and not the next
2012 * CPU after push_cpu. That is, if CPU 1, 4 and 5 are overloaded when CPU 3
2013 * schedules a lower priority task, and the IPI_RESTART gets set while the
2014 * handling is being done on CPU 5, it will clear the flag and send it back to
2015 * CPU 4 instead of CPU 1.
2017 * Note, the above logic can be disabled by turning off the sched_feature
2018 * RT_PUSH_IPI. Then the rq lock of the overloaded CPU will simply be
2019 * taken by the CPU requesting a pull and the waiting RT task will be pulled
2020 * by that CPU. This may be fine for machines with few CPUs.
2022 static void tell_cpu_to_push(struct rq
*rq
)
2026 if (rq
->rt
.push_flags
& RT_PUSH_IPI_EXECUTING
) {
2027 raw_spin_lock(&rq
->rt
.push_lock
);
2028 /* Make sure it's still executing */
2029 if (rq
->rt
.push_flags
& RT_PUSH_IPI_EXECUTING
) {
2031 * Tell the IPI to restart the loop as things have
2032 * changed since it started.
2034 rq
->rt
.push_flags
|= RT_PUSH_IPI_RESTART
;
2035 raw_spin_unlock(&rq
->rt
.push_lock
);
2038 raw_spin_unlock(&rq
->rt
.push_lock
);
2041 /* When here, there's no IPI going around */
2043 rq
->rt
.push_cpu
= rq
->cpu
;
2044 cpu
= find_next_push_cpu(rq
);
2045 if (cpu
>= nr_cpu_ids
)
2048 rq
->rt
.push_flags
= RT_PUSH_IPI_EXECUTING
;
2050 irq_work_queue_on(&rq
->rt
.push_work
, cpu
);
2053 /* Called from hardirq context */
2054 static void try_to_push_tasks(void *arg
)
2056 struct rt_rq
*rt_rq
= arg
;
2057 struct rq
*rq
, *src_rq
;
2061 this_cpu
= rt_rq
->push_cpu
;
2063 /* Paranoid check */
2064 BUG_ON(this_cpu
!= smp_processor_id());
2066 rq
= cpu_rq(this_cpu
);
2067 src_rq
= rq_of_rt_rq(rt_rq
);
2070 if (has_pushable_tasks(rq
)) {
2071 raw_spin_lock(&rq
->lock
);
2073 raw_spin_unlock(&rq
->lock
);
2076 /* Pass the IPI to the next rt overloaded queue */
2077 raw_spin_lock(&rt_rq
->push_lock
);
2079 * If the source queue changed since the IPI went out,
2080 * we need to restart the search from that CPU again.
2082 if (rt_rq
->push_flags
& RT_PUSH_IPI_RESTART
) {
2083 rt_rq
->push_flags
&= ~RT_PUSH_IPI_RESTART
;
2084 rt_rq
->push_cpu
= src_rq
->cpu
;
2087 cpu
= find_next_push_cpu(src_rq
);
2089 if (cpu
>= nr_cpu_ids
)
2090 rt_rq
->push_flags
&= ~RT_PUSH_IPI_EXECUTING
;
2091 raw_spin_unlock(&rt_rq
->push_lock
);
2093 if (cpu
>= nr_cpu_ids
)
2097 * It is possible that a restart caused this CPU to be
2098 * chosen again. Don't bother with an IPI, just see if we
2099 * have more to push.
2101 if (unlikely(cpu
== rq
->cpu
))
2104 /* Try the next RT overloaded CPU */
2105 irq_work_queue_on(&rt_rq
->push_work
, cpu
);
2108 static void push_irq_work_func(struct irq_work
*work
)
2110 struct rt_rq
*rt_rq
= container_of(work
, struct rt_rq
, push_work
);
2112 try_to_push_tasks(rt_rq
);
2114 #endif /* HAVE_RT_PUSH_IPI */
2116 static void pull_rt_task(struct rq
*this_rq
)
2118 int this_cpu
= this_rq
->cpu
, cpu
;
2119 bool resched
= false;
2120 struct task_struct
*p
;
2123 if (likely(!rt_overloaded(this_rq
)))
2127 * Match the barrier from rt_set_overloaded; this guarantees that if we
2128 * see overloaded we must also see the rto_mask bit.
2132 #ifdef HAVE_RT_PUSH_IPI
2133 if (sched_feat(RT_PUSH_IPI
)) {
2134 tell_cpu_to_push(this_rq
);
2139 for_each_cpu(cpu
, this_rq
->rd
->rto_mask
) {
2140 if (this_cpu
== cpu
)
2143 src_rq
= cpu_rq(cpu
);
2146 * Don't bother taking the src_rq->lock if the next highest
2147 * task is known to be lower-priority than our current task.
2148 * This may look racy, but if this value is about to go
2149 * logically higher, the src_rq will push this task away.
2150 * And if its going logically lower, we do not care
2152 if (src_rq
->rt
.highest_prio
.next
>=
2153 this_rq
->rt
.highest_prio
.curr
)
2157 * We can potentially drop this_rq's lock in
2158 * double_lock_balance, and another CPU could
2161 double_lock_balance(this_rq
, src_rq
);
2164 * We can pull only a task, which is pushable
2165 * on its rq, and no others.
2167 p
= pick_highest_pushable_task(src_rq
, this_cpu
);
2170 * Do we have an RT task that preempts
2171 * the to-be-scheduled task?
2173 if (p
&& (p
->prio
< this_rq
->rt
.highest_prio
.curr
)) {
2174 WARN_ON(p
== src_rq
->curr
);
2175 WARN_ON(!task_on_rq_queued(p
));
2178 * There's a chance that p is higher in priority
2179 * than what's currently running on its cpu.
2180 * This is just that p is wakeing up and hasn't
2181 * had a chance to schedule. We only pull
2182 * p if it is lower in priority than the
2183 * current task on the run queue
2185 if (p
->prio
< src_rq
->curr
->prio
)
2190 deactivate_task(src_rq
, p
, 0);
2191 set_task_cpu(p
, this_cpu
);
2192 activate_task(this_rq
, p
, 0);
2194 * We continue with the search, just in
2195 * case there's an even higher prio task
2196 * in another runqueue. (low likelihood
2201 double_unlock_balance(this_rq
, src_rq
);
2205 resched_curr(this_rq
);
2209 * If we are not running and we are not going to reschedule soon, we should
2210 * try to push tasks away now
2212 static void task_woken_rt(struct rq
*rq
, struct task_struct
*p
)
2214 if (!task_running(rq
, p
) &&
2215 !test_tsk_need_resched(rq
->curr
) &&
2216 p
->nr_cpus_allowed
> 1 &&
2217 (dl_task(rq
->curr
) || rt_task(rq
->curr
)) &&
2218 (rq
->curr
->nr_cpus_allowed
< 2 ||
2219 rq
->curr
->prio
<= p
->prio
))
2223 /* Assumes rq->lock is held */
2224 static void rq_online_rt(struct rq
*rq
)
2226 if (rq
->rt
.overloaded
)
2227 rt_set_overload(rq
);
2229 __enable_runtime(rq
);
2231 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rq
->rt
.highest_prio
.curr
);
2234 /* Assumes rq->lock is held */
2235 static void rq_offline_rt(struct rq
*rq
)
2237 if (rq
->rt
.overloaded
)
2238 rt_clear_overload(rq
);
2240 __disable_runtime(rq
);
2242 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, CPUPRI_INVALID
);
2246 * When switch from the rt queue, we bring ourselves to a position
2247 * that we might want to pull RT tasks from other runqueues.
2249 static void switched_from_rt(struct rq
*rq
, struct task_struct
*p
)
2252 * If there are other RT tasks then we will reschedule
2253 * and the scheduling of the other RT tasks will handle
2254 * the balancing. But if we are the last RT task
2255 * we may need to handle the pulling of RT tasks
2258 if (!task_on_rq_queued(p
) || rq
->rt
.rt_nr_running
)
2261 queue_pull_task(rq
);
2264 void __init
init_sched_rt_class(void)
2268 for_each_possible_cpu(i
) {
2269 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask
, i
),
2270 GFP_KERNEL
, cpu_to_node(i
));
2273 #endif /* CONFIG_SMP */
2276 * When switching a task to RT, we may overload the runqueue
2277 * with RT tasks. In this case we try to push them off to
2280 static void switched_to_rt(struct rq
*rq
, struct task_struct
*p
)
2283 * If we are already running, then there's nothing
2284 * that needs to be done. But if we are not running
2285 * we may need to preempt the current running task.
2286 * If that current running task is also an RT task
2287 * then see if we can move to another run queue.
2289 if (task_on_rq_queued(p
) && rq
->curr
!= p
) {
2291 if (p
->nr_cpus_allowed
> 1 && rq
->rt
.overloaded
)
2292 queue_push_tasks(rq
);
2293 #endif /* CONFIG_SMP */
2294 if (p
->prio
< rq
->curr
->prio
)
2300 * Priority of the task has changed. This may cause
2301 * us to initiate a push or pull.
2304 prio_changed_rt(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
2306 if (!task_on_rq_queued(p
))
2309 if (rq
->curr
== p
) {
2312 * If our priority decreases while running, we
2313 * may need to pull tasks to this runqueue.
2315 if (oldprio
< p
->prio
)
2316 queue_pull_task(rq
);
2319 * If there's a higher priority task waiting to run
2322 if (p
->prio
> rq
->rt
.highest_prio
.curr
)
2325 /* For UP simply resched on drop of prio */
2326 if (oldprio
< p
->prio
)
2328 #endif /* CONFIG_SMP */
2331 * This task is not running, but if it is
2332 * greater than the current running task
2335 if (p
->prio
< rq
->curr
->prio
)
2340 #ifdef CONFIG_POSIX_TIMERS
2341 static void watchdog(struct rq
*rq
, struct task_struct
*p
)
2343 unsigned long soft
, hard
;
2345 /* max may change after cur was read, this will be fixed next tick */
2346 soft
= task_rlimit(p
, RLIMIT_RTTIME
);
2347 hard
= task_rlimit_max(p
, RLIMIT_RTTIME
);
2349 if (soft
!= RLIM_INFINITY
) {
2352 if (p
->rt
.watchdog_stamp
!= jiffies
) {
2354 p
->rt
.watchdog_stamp
= jiffies
;
2357 next
= DIV_ROUND_UP(min(soft
, hard
), USEC_PER_SEC
/HZ
);
2358 if (p
->rt
.timeout
> next
)
2359 p
->cputime_expires
.sched_exp
= p
->se
.sum_exec_runtime
;
2363 static inline void watchdog(struct rq
*rq
, struct task_struct
*p
) { }
2366 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
, int queued
)
2368 struct sched_rt_entity
*rt_se
= &p
->rt
;
2375 * RR tasks need a special form of timeslice management.
2376 * FIFO tasks have no timeslices.
2378 if (p
->policy
!= SCHED_RR
)
2381 if (--p
->rt
.time_slice
)
2384 p
->rt
.time_slice
= sched_rr_timeslice
;
2387 * Requeue to the end of queue if we (and all of our ancestors) are not
2388 * the only element on the queue
2390 for_each_sched_rt_entity(rt_se
) {
2391 if (rt_se
->run_list
.prev
!= rt_se
->run_list
.next
) {
2392 requeue_task_rt(rq
, p
, 0);
2399 static void set_curr_task_rt(struct rq
*rq
)
2401 struct task_struct
*p
= rq
->curr
;
2403 p
->se
.exec_start
= rq_clock_task(rq
);
2405 /* The running task is never eligible for pushing */
2406 dequeue_pushable_task(rq
, p
);
2409 static unsigned int get_rr_interval_rt(struct rq
*rq
, struct task_struct
*task
)
2412 * Time slice is 0 for SCHED_FIFO tasks
2414 if (task
->policy
== SCHED_RR
)
2415 return sched_rr_timeslice
;
2420 const struct sched_class rt_sched_class
= {
2421 .next
= &fair_sched_class
,
2422 .enqueue_task
= enqueue_task_rt
,
2423 .dequeue_task
= dequeue_task_rt
,
2424 .yield_task
= yield_task_rt
,
2426 .check_preempt_curr
= check_preempt_curr_rt
,
2428 .pick_next_task
= pick_next_task_rt
,
2429 .put_prev_task
= put_prev_task_rt
,
2432 .select_task_rq
= select_task_rq_rt
,
2434 .set_cpus_allowed
= set_cpus_allowed_common
,
2435 .rq_online
= rq_online_rt
,
2436 .rq_offline
= rq_offline_rt
,
2437 .task_woken
= task_woken_rt
,
2438 .switched_from
= switched_from_rt
,
2441 .set_curr_task
= set_curr_task_rt
,
2442 .task_tick
= task_tick_rt
,
2444 .get_rr_interval
= get_rr_interval_rt
,
2446 .prio_changed
= prio_changed_rt
,
2447 .switched_to
= switched_to_rt
,
2449 .update_curr
= update_curr_rt
,
2452 #ifdef CONFIG_RT_GROUP_SCHED
2454 * Ensure that the real time constraints are schedulable.
2456 static DEFINE_MUTEX(rt_constraints_mutex
);
2458 /* Must be called with tasklist_lock held */
2459 static inline int tg_has_rt_tasks(struct task_group
*tg
)
2461 struct task_struct
*g
, *p
;
2464 * Autogroups do not have RT tasks; see autogroup_create().
2466 if (task_group_is_autogroup(tg
))
2469 for_each_process_thread(g
, p
) {
2470 if (rt_task(p
) && task_group(p
) == tg
)
2477 struct rt_schedulable_data
{
2478 struct task_group
*tg
;
2483 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
2485 struct rt_schedulable_data
*d
= data
;
2486 struct task_group
*child
;
2487 unsigned long total
, sum
= 0;
2488 u64 period
, runtime
;
2490 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2491 runtime
= tg
->rt_bandwidth
.rt_runtime
;
2494 period
= d
->rt_period
;
2495 runtime
= d
->rt_runtime
;
2499 * Cannot have more runtime than the period.
2501 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
2505 * Ensure we don't starve existing RT tasks.
2507 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
2510 total
= to_ratio(period
, runtime
);
2513 * Nobody can have more than the global setting allows.
2515 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
2519 * The sum of our children's runtime should not exceed our own.
2521 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
2522 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
2523 runtime
= child
->rt_bandwidth
.rt_runtime
;
2525 if (child
== d
->tg
) {
2526 period
= d
->rt_period
;
2527 runtime
= d
->rt_runtime
;
2530 sum
+= to_ratio(period
, runtime
);
2539 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
2543 struct rt_schedulable_data data
= {
2545 .rt_period
= period
,
2546 .rt_runtime
= runtime
,
2550 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
2556 static int tg_set_rt_bandwidth(struct task_group
*tg
,
2557 u64 rt_period
, u64 rt_runtime
)
2562 * Disallowing the root group RT runtime is BAD, it would disallow the
2563 * kernel creating (and or operating) RT threads.
2565 if (tg
== &root_task_group
&& rt_runtime
== 0)
2568 /* No period doesn't make any sense. */
2572 mutex_lock(&rt_constraints_mutex
);
2573 read_lock(&tasklist_lock
);
2574 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
2578 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
2579 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
2580 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
2582 for_each_possible_cpu(i
) {
2583 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
2585 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
2586 rt_rq
->rt_runtime
= rt_runtime
;
2587 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
2589 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
2591 read_unlock(&tasklist_lock
);
2592 mutex_unlock(&rt_constraints_mutex
);
2597 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
2599 u64 rt_runtime
, rt_period
;
2601 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2602 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
2603 if (rt_runtime_us
< 0)
2604 rt_runtime
= RUNTIME_INF
;
2606 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
2609 long sched_group_rt_runtime(struct task_group
*tg
)
2613 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
2616 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
2617 do_div(rt_runtime_us
, NSEC_PER_USEC
);
2618 return rt_runtime_us
;
2621 int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
2623 u64 rt_runtime
, rt_period
;
2625 rt_period
= rt_period_us
* NSEC_PER_USEC
;
2626 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
2628 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
2631 long sched_group_rt_period(struct task_group
*tg
)
2635 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2636 do_div(rt_period_us
, NSEC_PER_USEC
);
2637 return rt_period_us
;
2640 static int sched_rt_global_constraints(void)
2644 mutex_lock(&rt_constraints_mutex
);
2645 read_lock(&tasklist_lock
);
2646 ret
= __rt_schedulable(NULL
, 0, 0);
2647 read_unlock(&tasklist_lock
);
2648 mutex_unlock(&rt_constraints_mutex
);
2653 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
2655 /* Don't accept realtime tasks when there is no way for them to run */
2656 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
2662 #else /* !CONFIG_RT_GROUP_SCHED */
2663 static int sched_rt_global_constraints(void)
2665 unsigned long flags
;
2668 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2669 for_each_possible_cpu(i
) {
2670 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
2672 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
2673 rt_rq
->rt_runtime
= global_rt_runtime();
2674 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
2676 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2680 #endif /* CONFIG_RT_GROUP_SCHED */
2682 static int sched_rt_global_validate(void)
2684 if (sysctl_sched_rt_period
<= 0)
2687 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
2688 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
2694 static void sched_rt_do_global(void)
2696 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
2697 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
2700 int sched_rt_handler(struct ctl_table
*table
, int write
,
2701 void __user
*buffer
, size_t *lenp
,
2704 int old_period
, old_runtime
;
2705 static DEFINE_MUTEX(mutex
);
2709 old_period
= sysctl_sched_rt_period
;
2710 old_runtime
= sysctl_sched_rt_runtime
;
2712 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
2714 if (!ret
&& write
) {
2715 ret
= sched_rt_global_validate();
2719 ret
= sched_dl_global_validate();
2723 ret
= sched_rt_global_constraints();
2727 sched_rt_do_global();
2728 sched_dl_do_global();
2732 sysctl_sched_rt_period
= old_period
;
2733 sysctl_sched_rt_runtime
= old_runtime
;
2735 mutex_unlock(&mutex
);
2740 int sched_rr_handler(struct ctl_table
*table
, int write
,
2741 void __user
*buffer
, size_t *lenp
,
2745 static DEFINE_MUTEX(mutex
);
2748 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
2750 * Make sure that internally we keep jiffies.
2751 * Also, writing zero resets the timeslice to default:
2753 if (!ret
&& write
) {
2754 sched_rr_timeslice
=
2755 sysctl_sched_rr_timeslice
<= 0 ? RR_TIMESLICE
:
2756 msecs_to_jiffies(sysctl_sched_rr_timeslice
);
2758 mutex_unlock(&mutex
);
2762 #ifdef CONFIG_SCHED_DEBUG
2763 extern void print_rt_rq(struct seq_file
*m
, int cpu
, struct rt_rq
*rt_rq
);
2765 void print_rt_stats(struct seq_file
*m
, int cpu
)
2768 struct rt_rq
*rt_rq
;
2771 for_each_rt_rq(rt_rq
, iter
, cpu_rq(cpu
))
2772 print_rt_rq(m
, cpu
, rt_rq
);
2775 #endif /* CONFIG_SCHED_DEBUG */