1 // SPDX-License-Identifier: GPL-2.0
3 * Real-Time Scheduling Class (mapped to the SCHED_FIFO and SCHED_RR
10 int sched_rr_timeslice
= RR_TIMESLICE
;
11 int sysctl_sched_rr_timeslice
= (MSEC_PER_SEC
/ HZ
) * RR_TIMESLICE
;
12 /* More than 4 hours if BW_SHIFT equals 20. */
13 static const u64 max_rt_runtime
= MAX_BW
;
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
, CLOCK_MONOTONIC
,
51 HRTIMER_MODE_REL_HARD
);
52 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
55 static inline void do_start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
57 raw_spin_lock(&rt_b
->rt_runtime_lock
);
58 if (!rt_b
->rt_period_active
) {
59 rt_b
->rt_period_active
= 1;
61 * SCHED_DEADLINE updates the bandwidth, as a run away
62 * RT task with a DL task could hog a CPU. But DL does
63 * not reset the period. If a deadline task was running
64 * without an RT task running, it can cause RT tasks to
65 * throttle when they start up. Kick the timer right away
66 * to update the period.
68 hrtimer_forward_now(&rt_b
->rt_period_timer
, ns_to_ktime(0));
69 hrtimer_start_expires(&rt_b
->rt_period_timer
,
70 HRTIMER_MODE_ABS_PINNED_HARD
);
72 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
75 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
77 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
80 do_start_rt_bandwidth(rt_b
);
83 void init_rt_rq(struct rt_rq
*rt_rq
)
85 struct rt_prio_array
*array
;
88 array
= &rt_rq
->active
;
89 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
90 INIT_LIST_HEAD(array
->queue
+ i
);
91 __clear_bit(i
, array
->bitmap
);
93 /* delimiter for bitsearch: */
94 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
96 #if defined CONFIG_SMP
97 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
-1;
98 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
-1;
99 rt_rq
->rt_nr_migratory
= 0;
100 rt_rq
->overloaded
= 0;
101 plist_head_init(&rt_rq
->pushable_tasks
);
102 #endif /* CONFIG_SMP */
103 /* We start is dequeued state, because no RT tasks are queued */
104 rt_rq
->rt_queued
= 0;
107 rt_rq
->rt_throttled
= 0;
108 rt_rq
->rt_runtime
= 0;
109 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
112 #ifdef CONFIG_RT_GROUP_SCHED
113 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
115 hrtimer_cancel(&rt_b
->rt_period_timer
);
118 #define rt_entity_is_task(rt_se) (!(rt_se)->my_q)
120 static inline struct task_struct
*rt_task_of(struct sched_rt_entity
*rt_se
)
122 #ifdef CONFIG_SCHED_DEBUG
123 WARN_ON_ONCE(!rt_entity_is_task(rt_se
));
125 return container_of(rt_se
, struct task_struct
, rt
);
128 static inline struct rq
*rq_of_rt_rq(struct rt_rq
*rt_rq
)
133 static inline struct rt_rq
*rt_rq_of_se(struct sched_rt_entity
*rt_se
)
138 static inline struct rq
*rq_of_rt_se(struct sched_rt_entity
*rt_se
)
140 struct rt_rq
*rt_rq
= rt_se
->rt_rq
;
145 void unregister_rt_sched_group(struct task_group
*tg
)
148 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
152 void free_rt_sched_group(struct task_group
*tg
)
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
-1;
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
= kcalloc(nr_cpu_ids
, sizeof(rt_rq
), GFP_KERNEL
);
203 tg
->rt_se
= kcalloc(nr_cpu_ids
, sizeof(rt_se
), 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 unregister_rt_sched_group(struct task_group
*tg
) { }
264 void free_rt_sched_group(struct task_group
*tg
) { }
266 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
270 #endif /* CONFIG_RT_GROUP_SCHED */
274 static void pull_rt_task(struct rq
*this_rq
);
276 static inline bool need_pull_rt_task(struct rq
*rq
, struct task_struct
*prev
)
278 /* Try to pull RT tasks here if we lower this rq's prio */
279 return rq
->online
&& rq
->rt
.highest_prio
.curr
> prev
->prio
;
282 static inline int rt_overloaded(struct rq
*rq
)
284 return atomic_read(&rq
->rd
->rto_count
);
287 static inline void rt_set_overload(struct rq
*rq
)
292 cpumask_set_cpu(rq
->cpu
, rq
->rd
->rto_mask
);
294 * Make sure the mask is visible before we set
295 * the overload count. That is checked to determine
296 * if we should look at the mask. It would be a shame
297 * if we looked at the mask, but the mask was not
300 * Matched by the barrier in pull_rt_task().
303 atomic_inc(&rq
->rd
->rto_count
);
306 static inline void rt_clear_overload(struct rq
*rq
)
311 /* the order here really doesn't matter */
312 atomic_dec(&rq
->rd
->rto_count
);
313 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->rto_mask
);
316 static void update_rt_migration(struct rt_rq
*rt_rq
)
318 if (rt_rq
->rt_nr_migratory
&& rt_rq
->rt_nr_total
> 1) {
319 if (!rt_rq
->overloaded
) {
320 rt_set_overload(rq_of_rt_rq(rt_rq
));
321 rt_rq
->overloaded
= 1;
323 } else if (rt_rq
->overloaded
) {
324 rt_clear_overload(rq_of_rt_rq(rt_rq
));
325 rt_rq
->overloaded
= 0;
329 static void inc_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
331 struct task_struct
*p
;
333 if (!rt_entity_is_task(rt_se
))
336 p
= rt_task_of(rt_se
);
337 rt_rq
= &rq_of_rt_rq(rt_rq
)->rt
;
339 rt_rq
->rt_nr_total
++;
340 if (p
->nr_cpus_allowed
> 1)
341 rt_rq
->rt_nr_migratory
++;
343 update_rt_migration(rt_rq
);
346 static void dec_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
348 struct task_struct
*p
;
350 if (!rt_entity_is_task(rt_se
))
353 p
= rt_task_of(rt_se
);
354 rt_rq
= &rq_of_rt_rq(rt_rq
)->rt
;
356 rt_rq
->rt_nr_total
--;
357 if (p
->nr_cpus_allowed
> 1)
358 rt_rq
->rt_nr_migratory
--;
360 update_rt_migration(rt_rq
);
363 static inline int has_pushable_tasks(struct rq
*rq
)
365 return !plist_head_empty(&rq
->rt
.pushable_tasks
);
368 static DEFINE_PER_CPU(struct callback_head
, rt_push_head
);
369 static DEFINE_PER_CPU(struct callback_head
, rt_pull_head
);
371 static void push_rt_tasks(struct rq
*);
372 static void pull_rt_task(struct rq
*);
374 static inline void rt_queue_push_tasks(struct rq
*rq
)
376 if (!has_pushable_tasks(rq
))
379 queue_balance_callback(rq
, &per_cpu(rt_push_head
, rq
->cpu
), push_rt_tasks
);
382 static inline void rt_queue_pull_task(struct rq
*rq
)
384 queue_balance_callback(rq
, &per_cpu(rt_pull_head
, rq
->cpu
), pull_rt_task
);
387 static void enqueue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
389 plist_del(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
390 plist_node_init(&p
->pushable_tasks
, p
->prio
);
391 plist_add(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
393 /* Update the highest prio pushable task */
394 if (p
->prio
< rq
->rt
.highest_prio
.next
)
395 rq
->rt
.highest_prio
.next
= p
->prio
;
398 static void dequeue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
400 plist_del(&p
->pushable_tasks
, &rq
->rt
.pushable_tasks
);
402 /* Update the new highest prio pushable task */
403 if (has_pushable_tasks(rq
)) {
404 p
= plist_first_entry(&rq
->rt
.pushable_tasks
,
405 struct task_struct
, pushable_tasks
);
406 rq
->rt
.highest_prio
.next
= p
->prio
;
408 rq
->rt
.highest_prio
.next
= MAX_RT_PRIO
-1;
414 static inline void enqueue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
418 static inline void dequeue_pushable_task(struct rq
*rq
, struct task_struct
*p
)
423 void inc_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
428 void dec_rt_migration(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
432 static inline bool need_pull_rt_task(struct rq
*rq
, struct task_struct
*prev
)
437 static inline void pull_rt_task(struct rq
*this_rq
)
441 static inline void rt_queue_push_tasks(struct rq
*rq
)
444 #endif /* CONFIG_SMP */
446 static void enqueue_top_rt_rq(struct rt_rq
*rt_rq
);
447 static void dequeue_top_rt_rq(struct rt_rq
*rt_rq
);
449 static inline int on_rt_rq(struct sched_rt_entity
*rt_se
)
454 #ifdef CONFIG_UCLAMP_TASK
456 * Verify the fitness of task @p to run on @cpu taking into account the uclamp
459 * This check is only important for heterogeneous systems where uclamp_min value
460 * is higher than the capacity of a @cpu. For non-heterogeneous system this
461 * function will always return true.
463 * The function will return true if the capacity of the @cpu is >= the
464 * uclamp_min and false otherwise.
466 * Note that uclamp_min will be clamped to uclamp_max if uclamp_min
469 static inline bool rt_task_fits_capacity(struct task_struct
*p
, int cpu
)
471 unsigned int min_cap
;
472 unsigned int max_cap
;
473 unsigned int cpu_cap
;
475 /* Only heterogeneous systems can benefit from this check */
476 if (!static_branch_unlikely(&sched_asym_cpucapacity
))
479 min_cap
= uclamp_eff_value(p
, UCLAMP_MIN
);
480 max_cap
= uclamp_eff_value(p
, UCLAMP_MAX
);
482 cpu_cap
= capacity_orig_of(cpu
);
484 return cpu_cap
>= min(min_cap
, max_cap
);
487 static inline bool rt_task_fits_capacity(struct task_struct
*p
, int cpu
)
493 #ifdef CONFIG_RT_GROUP_SCHED
495 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
500 return rt_rq
->rt_runtime
;
503 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
505 return ktime_to_ns(rt_rq
->tg
->rt_bandwidth
.rt_period
);
508 typedef struct task_group
*rt_rq_iter_t
;
510 static inline struct task_group
*next_task_group(struct task_group
*tg
)
513 tg
= list_entry_rcu(tg
->list
.next
,
514 typeof(struct task_group
), list
);
515 } while (&tg
->list
!= &task_groups
&& task_group_is_autogroup(tg
));
517 if (&tg
->list
== &task_groups
)
523 #define for_each_rt_rq(rt_rq, iter, rq) \
524 for (iter = container_of(&task_groups, typeof(*iter), list); \
525 (iter = next_task_group(iter)) && \
526 (rt_rq = iter->rt_rq[cpu_of(rq)]);)
528 #define for_each_sched_rt_entity(rt_se) \
529 for (; rt_se; rt_se = rt_se->parent)
531 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
536 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
);
537 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
);
539 static void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
541 struct task_struct
*curr
= rq_of_rt_rq(rt_rq
)->curr
;
542 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
543 struct sched_rt_entity
*rt_se
;
545 int cpu
= cpu_of(rq
);
547 rt_se
= rt_rq
->tg
->rt_se
[cpu
];
549 if (rt_rq
->rt_nr_running
) {
551 enqueue_top_rt_rq(rt_rq
);
552 else if (!on_rt_rq(rt_se
))
553 enqueue_rt_entity(rt_se
, 0);
555 if (rt_rq
->highest_prio
.curr
< curr
->prio
)
560 static void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
562 struct sched_rt_entity
*rt_se
;
563 int cpu
= cpu_of(rq_of_rt_rq(rt_rq
));
565 rt_se
= rt_rq
->tg
->rt_se
[cpu
];
568 dequeue_top_rt_rq(rt_rq
);
569 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
570 cpufreq_update_util(rq_of_rt_rq(rt_rq
), 0);
572 else if (on_rt_rq(rt_se
))
573 dequeue_rt_entity(rt_se
, 0);
576 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
578 return rt_rq
->rt_throttled
&& !rt_rq
->rt_nr_boosted
;
581 static int rt_se_boosted(struct sched_rt_entity
*rt_se
)
583 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
584 struct task_struct
*p
;
587 return !!rt_rq
->rt_nr_boosted
;
589 p
= rt_task_of(rt_se
);
590 return p
->prio
!= p
->normal_prio
;
594 static inline const struct cpumask
*sched_rt_period_mask(void)
596 return this_rq()->rd
->span
;
599 static inline const struct cpumask
*sched_rt_period_mask(void)
601 return cpu_online_mask
;
606 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
608 return container_of(rt_b
, struct task_group
, rt_bandwidth
)->rt_rq
[cpu
];
611 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
613 return &rt_rq
->tg
->rt_bandwidth
;
616 #else /* !CONFIG_RT_GROUP_SCHED */
618 static inline u64
sched_rt_runtime(struct rt_rq
*rt_rq
)
620 return rt_rq
->rt_runtime
;
623 static inline u64
sched_rt_period(struct rt_rq
*rt_rq
)
625 return ktime_to_ns(def_rt_bandwidth
.rt_period
);
628 typedef struct rt_rq
*rt_rq_iter_t
;
630 #define for_each_rt_rq(rt_rq, iter, rq) \
631 for ((void) iter, rt_rq = &rq->rt; rt_rq; rt_rq = NULL)
633 #define for_each_sched_rt_entity(rt_se) \
634 for (; rt_se; rt_se = NULL)
636 static inline struct rt_rq
*group_rt_rq(struct sched_rt_entity
*rt_se
)
641 static inline void sched_rt_rq_enqueue(struct rt_rq
*rt_rq
)
643 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
645 if (!rt_rq
->rt_nr_running
)
648 enqueue_top_rt_rq(rt_rq
);
652 static inline void sched_rt_rq_dequeue(struct rt_rq
*rt_rq
)
654 dequeue_top_rt_rq(rt_rq
);
657 static inline int rt_rq_throttled(struct rt_rq
*rt_rq
)
659 return rt_rq
->rt_throttled
;
662 static inline const struct cpumask
*sched_rt_period_mask(void)
664 return cpu_online_mask
;
668 struct rt_rq
*sched_rt_period_rt_rq(struct rt_bandwidth
*rt_b
, int cpu
)
670 return &cpu_rq(cpu
)->rt
;
673 static inline struct rt_bandwidth
*sched_rt_bandwidth(struct rt_rq
*rt_rq
)
675 return &def_rt_bandwidth
;
678 #endif /* CONFIG_RT_GROUP_SCHED */
680 bool sched_rt_bandwidth_account(struct rt_rq
*rt_rq
)
682 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
684 return (hrtimer_active(&rt_b
->rt_period_timer
) ||
685 rt_rq
->rt_time
< rt_b
->rt_runtime
);
690 * We ran out of runtime, see if we can borrow some from our neighbours.
692 static void do_balance_runtime(struct rt_rq
*rt_rq
)
694 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
695 struct root_domain
*rd
= rq_of_rt_rq(rt_rq
)->rd
;
699 weight
= cpumask_weight(rd
->span
);
701 raw_spin_lock(&rt_b
->rt_runtime_lock
);
702 rt_period
= ktime_to_ns(rt_b
->rt_period
);
703 for_each_cpu(i
, rd
->span
) {
704 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
710 raw_spin_lock(&iter
->rt_runtime_lock
);
712 * Either all rqs have inf runtime and there's nothing to steal
713 * or __disable_runtime() below sets a specific rq to inf to
714 * indicate its been disabled and disallow stealing.
716 if (iter
->rt_runtime
== RUNTIME_INF
)
720 * From runqueues with spare time, take 1/n part of their
721 * spare time, but no more than our period.
723 diff
= iter
->rt_runtime
- iter
->rt_time
;
725 diff
= div_u64((u64
)diff
, weight
);
726 if (rt_rq
->rt_runtime
+ diff
> rt_period
)
727 diff
= rt_period
- rt_rq
->rt_runtime
;
728 iter
->rt_runtime
-= diff
;
729 rt_rq
->rt_runtime
+= diff
;
730 if (rt_rq
->rt_runtime
== rt_period
) {
731 raw_spin_unlock(&iter
->rt_runtime_lock
);
736 raw_spin_unlock(&iter
->rt_runtime_lock
);
738 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
742 * Ensure this RQ takes back all the runtime it lend to its neighbours.
744 static void __disable_runtime(struct rq
*rq
)
746 struct root_domain
*rd
= rq
->rd
;
750 if (unlikely(!scheduler_running
))
753 for_each_rt_rq(rt_rq
, iter
, rq
) {
754 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
758 raw_spin_lock(&rt_b
->rt_runtime_lock
);
759 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
761 * Either we're all inf and nobody needs to borrow, or we're
762 * already disabled and thus have nothing to do, or we have
763 * exactly the right amount of runtime to take out.
765 if (rt_rq
->rt_runtime
== RUNTIME_INF
||
766 rt_rq
->rt_runtime
== rt_b
->rt_runtime
)
768 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
771 * Calculate the difference between what we started out with
772 * and what we current have, that's the amount of runtime
773 * we lend and now have to reclaim.
775 want
= rt_b
->rt_runtime
- rt_rq
->rt_runtime
;
778 * Greedy reclaim, take back as much as we can.
780 for_each_cpu(i
, rd
->span
) {
781 struct rt_rq
*iter
= sched_rt_period_rt_rq(rt_b
, i
);
785 * Can't reclaim from ourselves or disabled runqueues.
787 if (iter
== rt_rq
|| iter
->rt_runtime
== RUNTIME_INF
)
790 raw_spin_lock(&iter
->rt_runtime_lock
);
792 diff
= min_t(s64
, iter
->rt_runtime
, want
);
793 iter
->rt_runtime
-= diff
;
796 iter
->rt_runtime
-= want
;
799 raw_spin_unlock(&iter
->rt_runtime_lock
);
805 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
807 * We cannot be left wanting - that would mean some runtime
808 * leaked out of the system.
813 * Disable all the borrow logic by pretending we have inf
814 * runtime - in which case borrowing doesn't make sense.
816 rt_rq
->rt_runtime
= RUNTIME_INF
;
817 rt_rq
->rt_throttled
= 0;
818 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
819 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
821 /* Make rt_rq available for pick_next_task() */
822 sched_rt_rq_enqueue(rt_rq
);
826 static void __enable_runtime(struct rq
*rq
)
831 if (unlikely(!scheduler_running
))
835 * Reset each runqueue's bandwidth settings
837 for_each_rt_rq(rt_rq
, iter
, rq
) {
838 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
840 raw_spin_lock(&rt_b
->rt_runtime_lock
);
841 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
842 rt_rq
->rt_runtime
= rt_b
->rt_runtime
;
844 rt_rq
->rt_throttled
= 0;
845 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
846 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
850 static void balance_runtime(struct rt_rq
*rt_rq
)
852 if (!sched_feat(RT_RUNTIME_SHARE
))
855 if (rt_rq
->rt_time
> rt_rq
->rt_runtime
) {
856 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
857 do_balance_runtime(rt_rq
);
858 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
861 #else /* !CONFIG_SMP */
862 static inline void balance_runtime(struct rt_rq
*rt_rq
) {}
863 #endif /* CONFIG_SMP */
865 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
)
867 int i
, idle
= 1, throttled
= 0;
868 const struct cpumask
*span
;
870 span
= sched_rt_period_mask();
871 #ifdef CONFIG_RT_GROUP_SCHED
873 * FIXME: isolated CPUs should really leave the root task group,
874 * whether they are isolcpus or were isolated via cpusets, lest
875 * the timer run on a CPU which does not service all runqueues,
876 * potentially leaving other CPUs indefinitely throttled. If
877 * isolation is really required, the user will turn the throttle
878 * off to kill the perturbations it causes anyway. Meanwhile,
879 * this maintains functionality for boot and/or troubleshooting.
881 if (rt_b
== &root_task_group
.rt_bandwidth
)
882 span
= cpu_online_mask
;
884 for_each_cpu(i
, span
) {
886 struct rt_rq
*rt_rq
= sched_rt_period_rt_rq(rt_b
, i
);
887 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
891 * When span == cpu_online_mask, taking each rq->lock
892 * can be time-consuming. Try to avoid it when possible.
894 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
895 if (!sched_feat(RT_RUNTIME_SHARE
) && rt_rq
->rt_runtime
!= RUNTIME_INF
)
896 rt_rq
->rt_runtime
= rt_b
->rt_runtime
;
897 skip
= !rt_rq
->rt_time
&& !rt_rq
->rt_nr_running
;
898 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
902 raw_spin_rq_lock(rq
);
905 if (rt_rq
->rt_time
) {
908 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
909 if (rt_rq
->rt_throttled
)
910 balance_runtime(rt_rq
);
911 runtime
= rt_rq
->rt_runtime
;
912 rt_rq
->rt_time
-= min(rt_rq
->rt_time
, overrun
*runtime
);
913 if (rt_rq
->rt_throttled
&& rt_rq
->rt_time
< runtime
) {
914 rt_rq
->rt_throttled
= 0;
918 * When we're idle and a woken (rt) task is
919 * throttled check_preempt_curr() will set
920 * skip_update and the time between the wakeup
921 * and this unthrottle will get accounted as
924 if (rt_rq
->rt_nr_running
&& rq
->curr
== rq
->idle
)
925 rq_clock_cancel_skipupdate(rq
);
927 if (rt_rq
->rt_time
|| rt_rq
->rt_nr_running
)
929 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
930 } else if (rt_rq
->rt_nr_running
) {
932 if (!rt_rq_throttled(rt_rq
))
935 if (rt_rq
->rt_throttled
)
939 sched_rt_rq_enqueue(rt_rq
);
940 raw_spin_rq_unlock(rq
);
943 if (!throttled
&& (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
))
949 static inline int rt_se_prio(struct sched_rt_entity
*rt_se
)
951 #ifdef CONFIG_RT_GROUP_SCHED
952 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
955 return rt_rq
->highest_prio
.curr
;
958 return rt_task_of(rt_se
)->prio
;
961 static int sched_rt_runtime_exceeded(struct rt_rq
*rt_rq
)
963 u64 runtime
= sched_rt_runtime(rt_rq
);
965 if (rt_rq
->rt_throttled
)
966 return rt_rq_throttled(rt_rq
);
968 if (runtime
>= sched_rt_period(rt_rq
))
971 balance_runtime(rt_rq
);
972 runtime
= sched_rt_runtime(rt_rq
);
973 if (runtime
== RUNTIME_INF
)
976 if (rt_rq
->rt_time
> runtime
) {
977 struct rt_bandwidth
*rt_b
= sched_rt_bandwidth(rt_rq
);
980 * Don't actually throttle groups that have no runtime assigned
981 * but accrue some time due to boosting.
983 if (likely(rt_b
->rt_runtime
)) {
984 rt_rq
->rt_throttled
= 1;
985 printk_deferred_once("sched: RT throttling activated\n");
988 * In case we did anyway, make it go away,
989 * replenishment is a joke, since it will replenish us
995 if (rt_rq_throttled(rt_rq
)) {
996 sched_rt_rq_dequeue(rt_rq
);
1005 * Update the current task's runtime statistics. Skip current tasks that
1006 * are not in our scheduling class.
1008 static void update_curr_rt(struct rq
*rq
)
1010 struct task_struct
*curr
= rq
->curr
;
1011 struct sched_rt_entity
*rt_se
= &curr
->rt
;
1015 if (curr
->sched_class
!= &rt_sched_class
)
1018 now
= rq_clock_task(rq
);
1019 delta_exec
= now
- curr
->se
.exec_start
;
1020 if (unlikely((s64
)delta_exec
<= 0))
1023 schedstat_set(curr
->se
.statistics
.exec_max
,
1024 max(curr
->se
.statistics
.exec_max
, delta_exec
));
1026 curr
->se
.sum_exec_runtime
+= delta_exec
;
1027 account_group_exec_runtime(curr
, delta_exec
);
1029 curr
->se
.exec_start
= now
;
1030 cgroup_account_cputime(curr
, delta_exec
);
1032 if (!rt_bandwidth_enabled())
1035 for_each_sched_rt_entity(rt_se
) {
1036 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
1039 if (sched_rt_runtime(rt_rq
) != RUNTIME_INF
) {
1040 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
1041 rt_rq
->rt_time
+= delta_exec
;
1042 exceeded
= sched_rt_runtime_exceeded(rt_rq
);
1045 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
1047 do_start_rt_bandwidth(sched_rt_bandwidth(rt_rq
));
1053 dequeue_top_rt_rq(struct rt_rq
*rt_rq
)
1055 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1057 BUG_ON(&rq
->rt
!= rt_rq
);
1059 if (!rt_rq
->rt_queued
)
1062 BUG_ON(!rq
->nr_running
);
1064 sub_nr_running(rq
, rt_rq
->rt_nr_running
);
1065 rt_rq
->rt_queued
= 0;
1070 enqueue_top_rt_rq(struct rt_rq
*rt_rq
)
1072 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1074 BUG_ON(&rq
->rt
!= rt_rq
);
1076 if (rt_rq
->rt_queued
)
1079 if (rt_rq_throttled(rt_rq
))
1082 if (rt_rq
->rt_nr_running
) {
1083 add_nr_running(rq
, rt_rq
->rt_nr_running
);
1084 rt_rq
->rt_queued
= 1;
1087 /* Kick cpufreq (see the comment in kernel/sched/sched.h). */
1088 cpufreq_update_util(rq
, 0);
1091 #if defined CONFIG_SMP
1094 inc_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
)
1096 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1098 #ifdef CONFIG_RT_GROUP_SCHED
1100 * Change rq's cpupri only if rt_rq is the top queue.
1102 if (&rq
->rt
!= rt_rq
)
1105 if (rq
->online
&& prio
< prev_prio
)
1106 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, prio
);
1110 dec_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
)
1112 struct rq
*rq
= rq_of_rt_rq(rt_rq
);
1114 #ifdef CONFIG_RT_GROUP_SCHED
1116 * Change rq's cpupri only if rt_rq is the top queue.
1118 if (&rq
->rt
!= rt_rq
)
1121 if (rq
->online
&& rt_rq
->highest_prio
.curr
!= prev_prio
)
1122 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rt_rq
->highest_prio
.curr
);
1125 #else /* CONFIG_SMP */
1128 void inc_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
) {}
1130 void dec_rt_prio_smp(struct rt_rq
*rt_rq
, int prio
, int prev_prio
) {}
1132 #endif /* CONFIG_SMP */
1134 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
1136 inc_rt_prio(struct rt_rq
*rt_rq
, int prio
)
1138 int prev_prio
= rt_rq
->highest_prio
.curr
;
1140 if (prio
< prev_prio
)
1141 rt_rq
->highest_prio
.curr
= prio
;
1143 inc_rt_prio_smp(rt_rq
, prio
, prev_prio
);
1147 dec_rt_prio(struct rt_rq
*rt_rq
, int prio
)
1149 int prev_prio
= rt_rq
->highest_prio
.curr
;
1151 if (rt_rq
->rt_nr_running
) {
1153 WARN_ON(prio
< prev_prio
);
1156 * This may have been our highest task, and therefore
1157 * we may have some recomputation to do
1159 if (prio
== prev_prio
) {
1160 struct rt_prio_array
*array
= &rt_rq
->active
;
1162 rt_rq
->highest_prio
.curr
=
1163 sched_find_first_bit(array
->bitmap
);
1167 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
-1;
1170 dec_rt_prio_smp(rt_rq
, prio
, prev_prio
);
1175 static inline void inc_rt_prio(struct rt_rq
*rt_rq
, int prio
) {}
1176 static inline void dec_rt_prio(struct rt_rq
*rt_rq
, int prio
) {}
1178 #endif /* CONFIG_SMP || CONFIG_RT_GROUP_SCHED */
1180 #ifdef CONFIG_RT_GROUP_SCHED
1183 inc_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1185 if (rt_se_boosted(rt_se
))
1186 rt_rq
->rt_nr_boosted
++;
1189 start_rt_bandwidth(&rt_rq
->tg
->rt_bandwidth
);
1193 dec_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1195 if (rt_se_boosted(rt_se
))
1196 rt_rq
->rt_nr_boosted
--;
1198 WARN_ON(!rt_rq
->rt_nr_running
&& rt_rq
->rt_nr_boosted
);
1201 #else /* CONFIG_RT_GROUP_SCHED */
1204 inc_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1206 start_rt_bandwidth(&def_rt_bandwidth
);
1210 void dec_rt_group(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
) {}
1212 #endif /* CONFIG_RT_GROUP_SCHED */
1215 unsigned int rt_se_nr_running(struct sched_rt_entity
*rt_se
)
1217 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1220 return group_rq
->rt_nr_running
;
1226 unsigned int rt_se_rr_nr_running(struct sched_rt_entity
*rt_se
)
1228 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1229 struct task_struct
*tsk
;
1232 return group_rq
->rr_nr_running
;
1234 tsk
= rt_task_of(rt_se
);
1236 return (tsk
->policy
== SCHED_RR
) ? 1 : 0;
1240 void inc_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1242 int prio
= rt_se_prio(rt_se
);
1244 WARN_ON(!rt_prio(prio
));
1245 rt_rq
->rt_nr_running
+= rt_se_nr_running(rt_se
);
1246 rt_rq
->rr_nr_running
+= rt_se_rr_nr_running(rt_se
);
1248 inc_rt_prio(rt_rq
, prio
);
1249 inc_rt_migration(rt_se
, rt_rq
);
1250 inc_rt_group(rt_se
, rt_rq
);
1254 void dec_rt_tasks(struct sched_rt_entity
*rt_se
, struct rt_rq
*rt_rq
)
1256 WARN_ON(!rt_prio(rt_se_prio(rt_se
)));
1257 WARN_ON(!rt_rq
->rt_nr_running
);
1258 rt_rq
->rt_nr_running
-= rt_se_nr_running(rt_se
);
1259 rt_rq
->rr_nr_running
-= rt_se_rr_nr_running(rt_se
);
1261 dec_rt_prio(rt_rq
, rt_se_prio(rt_se
));
1262 dec_rt_migration(rt_se
, rt_rq
);
1263 dec_rt_group(rt_se
, rt_rq
);
1267 * Change rt_se->run_list location unless SAVE && !MOVE
1269 * assumes ENQUEUE/DEQUEUE flags match
1271 static inline bool move_entity(unsigned int flags
)
1273 if ((flags
& (DEQUEUE_SAVE
| DEQUEUE_MOVE
)) == DEQUEUE_SAVE
)
1279 static void __delist_rt_entity(struct sched_rt_entity
*rt_se
, struct rt_prio_array
*array
)
1281 list_del_init(&rt_se
->run_list
);
1283 if (list_empty(array
->queue
+ rt_se_prio(rt_se
)))
1284 __clear_bit(rt_se_prio(rt_se
), array
->bitmap
);
1289 static void __enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1291 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
1292 struct rt_prio_array
*array
= &rt_rq
->active
;
1293 struct rt_rq
*group_rq
= group_rt_rq(rt_se
);
1294 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
1297 * Don't enqueue the group if its throttled, or when empty.
1298 * The latter is a consequence of the former when a child group
1299 * get throttled and the current group doesn't have any other
1302 if (group_rq
&& (rt_rq_throttled(group_rq
) || !group_rq
->rt_nr_running
)) {
1304 __delist_rt_entity(rt_se
, array
);
1308 if (move_entity(flags
)) {
1309 WARN_ON_ONCE(rt_se
->on_list
);
1310 if (flags
& ENQUEUE_HEAD
)
1311 list_add(&rt_se
->run_list
, queue
);
1313 list_add_tail(&rt_se
->run_list
, queue
);
1315 __set_bit(rt_se_prio(rt_se
), array
->bitmap
);
1320 inc_rt_tasks(rt_se
, rt_rq
);
1323 static void __dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1325 struct rt_rq
*rt_rq
= rt_rq_of_se(rt_se
);
1326 struct rt_prio_array
*array
= &rt_rq
->active
;
1328 if (move_entity(flags
)) {
1329 WARN_ON_ONCE(!rt_se
->on_list
);
1330 __delist_rt_entity(rt_se
, array
);
1334 dec_rt_tasks(rt_se
, rt_rq
);
1338 * Because the prio of an upper entry depends on the lower
1339 * entries, we must remove entries top - down.
1341 static void dequeue_rt_stack(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1343 struct sched_rt_entity
*back
= NULL
;
1345 for_each_sched_rt_entity(rt_se
) {
1350 dequeue_top_rt_rq(rt_rq_of_se(back
));
1352 for (rt_se
= back
; rt_se
; rt_se
= rt_se
->back
) {
1353 if (on_rt_rq(rt_se
))
1354 __dequeue_rt_entity(rt_se
, flags
);
1358 static void enqueue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1360 struct rq
*rq
= rq_of_rt_se(rt_se
);
1362 dequeue_rt_stack(rt_se
, flags
);
1363 for_each_sched_rt_entity(rt_se
)
1364 __enqueue_rt_entity(rt_se
, flags
);
1365 enqueue_top_rt_rq(&rq
->rt
);
1368 static void dequeue_rt_entity(struct sched_rt_entity
*rt_se
, unsigned int flags
)
1370 struct rq
*rq
= rq_of_rt_se(rt_se
);
1372 dequeue_rt_stack(rt_se
, flags
);
1374 for_each_sched_rt_entity(rt_se
) {
1375 struct rt_rq
*rt_rq
= group_rt_rq(rt_se
);
1377 if (rt_rq
&& rt_rq
->rt_nr_running
)
1378 __enqueue_rt_entity(rt_se
, flags
);
1380 enqueue_top_rt_rq(&rq
->rt
);
1384 * Adding/removing a task to/from a priority array:
1387 enqueue_task_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1389 struct sched_rt_entity
*rt_se
= &p
->rt
;
1391 if (flags
& ENQUEUE_WAKEUP
)
1394 enqueue_rt_entity(rt_se
, flags
);
1396 if (!task_current(rq
, p
) && p
->nr_cpus_allowed
> 1)
1397 enqueue_pushable_task(rq
, p
);
1400 static void dequeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1402 struct sched_rt_entity
*rt_se
= &p
->rt
;
1405 dequeue_rt_entity(rt_se
, flags
);
1407 dequeue_pushable_task(rq
, p
);
1411 * Put task to the head or the end of the run list without the overhead of
1412 * dequeue followed by enqueue.
1415 requeue_rt_entity(struct rt_rq
*rt_rq
, struct sched_rt_entity
*rt_se
, int head
)
1417 if (on_rt_rq(rt_se
)) {
1418 struct rt_prio_array
*array
= &rt_rq
->active
;
1419 struct list_head
*queue
= array
->queue
+ rt_se_prio(rt_se
);
1422 list_move(&rt_se
->run_list
, queue
);
1424 list_move_tail(&rt_se
->run_list
, queue
);
1428 static void requeue_task_rt(struct rq
*rq
, struct task_struct
*p
, int head
)
1430 struct sched_rt_entity
*rt_se
= &p
->rt
;
1431 struct rt_rq
*rt_rq
;
1433 for_each_sched_rt_entity(rt_se
) {
1434 rt_rq
= rt_rq_of_se(rt_se
);
1435 requeue_rt_entity(rt_rq
, rt_se
, head
);
1439 static void yield_task_rt(struct rq
*rq
)
1441 requeue_task_rt(rq
, rq
->curr
, 0);
1445 static int find_lowest_rq(struct task_struct
*task
);
1448 select_task_rq_rt(struct task_struct
*p
, int cpu
, int flags
)
1450 struct task_struct
*curr
;
1454 /* For anything but wake ups, just return the task_cpu */
1455 if (!(flags
& (WF_TTWU
| WF_FORK
)))
1461 curr
= READ_ONCE(rq
->curr
); /* unlocked access */
1464 * If the current task on @p's runqueue is an RT task, then
1465 * try to see if we can wake this RT task up on another
1466 * runqueue. Otherwise simply start this RT task
1467 * on its current runqueue.
1469 * We want to avoid overloading runqueues. If the woken
1470 * task is a higher priority, then it will stay on this CPU
1471 * and the lower prio task should be moved to another CPU.
1472 * Even though this will probably make the lower prio task
1473 * lose its cache, we do not want to bounce a higher task
1474 * around just because it gave up its CPU, perhaps for a
1477 * For equal prio tasks, we just let the scheduler sort it out.
1479 * Otherwise, just let it ride on the affined RQ and the
1480 * post-schedule router will push the preempted task away
1482 * This test is optimistic, if we get it wrong the load-balancer
1483 * will have to sort it out.
1485 * We take into account the capacity of the CPU to ensure it fits the
1486 * requirement of the task - which is only important on heterogeneous
1487 * systems like big.LITTLE.
1490 unlikely(rt_task(curr
)) &&
1491 (curr
->nr_cpus_allowed
< 2 || curr
->prio
<= p
->prio
);
1493 if (test
|| !rt_task_fits_capacity(p
, cpu
)) {
1494 int target
= find_lowest_rq(p
);
1497 * Bail out if we were forcing a migration to find a better
1498 * fitting CPU but our search failed.
1500 if (!test
&& target
!= -1 && !rt_task_fits_capacity(p
, target
))
1504 * Don't bother moving it if the destination CPU is
1505 * not running a lower priority task.
1508 p
->prio
< cpu_rq(target
)->rt
.highest_prio
.curr
)
1519 static void check_preempt_equal_prio(struct rq
*rq
, struct task_struct
*p
)
1522 * Current can't be migrated, useless to reschedule,
1523 * let's hope p can move out.
1525 if (rq
->curr
->nr_cpus_allowed
== 1 ||
1526 !cpupri_find(&rq
->rd
->cpupri
, rq
->curr
, NULL
))
1530 * p is migratable, so let's not schedule it and
1531 * see if it is pushed or pulled somewhere else.
1533 if (p
->nr_cpus_allowed
!= 1 &&
1534 cpupri_find(&rq
->rd
->cpupri
, p
, NULL
))
1538 * There appear to be other CPUs that can accept
1539 * the current task but none can run 'p', so lets reschedule
1540 * to try and push the current task away:
1542 requeue_task_rt(rq
, p
, 1);
1546 static int balance_rt(struct rq
*rq
, struct task_struct
*p
, struct rq_flags
*rf
)
1548 if (!on_rt_rq(&p
->rt
) && need_pull_rt_task(rq
, p
)) {
1550 * This is OK, because current is on_cpu, which avoids it being
1551 * picked for load-balance and preemption/IRQs are still
1552 * disabled avoiding further scheduler activity on it and we've
1553 * not yet started the picking loop.
1555 rq_unpin_lock(rq
, rf
);
1557 rq_repin_lock(rq
, rf
);
1560 return sched_stop_runnable(rq
) || sched_dl_runnable(rq
) || sched_rt_runnable(rq
);
1562 #endif /* CONFIG_SMP */
1565 * Preempt the current task with a newly woken task if needed:
1567 static void check_preempt_curr_rt(struct rq
*rq
, struct task_struct
*p
, int flags
)
1569 if (p
->prio
< rq
->curr
->prio
) {
1578 * - the newly woken task is of equal priority to the current task
1579 * - the newly woken task is non-migratable while current is migratable
1580 * - current will be preempted on the next reschedule
1582 * we should check to see if current can readily move to a different
1583 * cpu. If so, we will reschedule to allow the push logic to try
1584 * to move current somewhere else, making room for our non-migratable
1587 if (p
->prio
== rq
->curr
->prio
&& !test_tsk_need_resched(rq
->curr
))
1588 check_preempt_equal_prio(rq
, p
);
1592 static inline void set_next_task_rt(struct rq
*rq
, struct task_struct
*p
, bool first
)
1594 p
->se
.exec_start
= rq_clock_task(rq
);
1596 /* The running task is never eligible for pushing */
1597 dequeue_pushable_task(rq
, p
);
1603 * If prev task was rt, put_prev_task() has already updated the
1604 * utilization. We only care of the case where we start to schedule a
1607 if (rq
->curr
->sched_class
!= &rt_sched_class
)
1608 update_rt_rq_load_avg(rq_clock_pelt(rq
), rq
, 0);
1610 rt_queue_push_tasks(rq
);
1613 static struct sched_rt_entity
*pick_next_rt_entity(struct rq
*rq
,
1614 struct rt_rq
*rt_rq
)
1616 struct rt_prio_array
*array
= &rt_rq
->active
;
1617 struct sched_rt_entity
*next
= NULL
;
1618 struct list_head
*queue
;
1621 idx
= sched_find_first_bit(array
->bitmap
);
1622 BUG_ON(idx
>= MAX_RT_PRIO
);
1624 queue
= array
->queue
+ idx
;
1625 next
= list_entry(queue
->next
, struct sched_rt_entity
, run_list
);
1630 static struct task_struct
*_pick_next_task_rt(struct rq
*rq
)
1632 struct sched_rt_entity
*rt_se
;
1633 struct rt_rq
*rt_rq
= &rq
->rt
;
1636 rt_se
= pick_next_rt_entity(rq
, rt_rq
);
1638 rt_rq
= group_rt_rq(rt_se
);
1641 return rt_task_of(rt_se
);
1644 static struct task_struct
*pick_task_rt(struct rq
*rq
)
1646 struct task_struct
*p
;
1648 if (!sched_rt_runnable(rq
))
1651 p
= _pick_next_task_rt(rq
);
1656 static struct task_struct
*pick_next_task_rt(struct rq
*rq
)
1658 struct task_struct
*p
= pick_task_rt(rq
);
1661 set_next_task_rt(rq
, p
, true);
1666 static void put_prev_task_rt(struct rq
*rq
, struct task_struct
*p
)
1670 update_rt_rq_load_avg(rq_clock_pelt(rq
), rq
, 1);
1673 * The previous task needs to be made eligible for pushing
1674 * if it is still active
1676 if (on_rt_rq(&p
->rt
) && p
->nr_cpus_allowed
> 1)
1677 enqueue_pushable_task(rq
, p
);
1682 /* Only try algorithms three times */
1683 #define RT_MAX_TRIES 3
1685 static int pick_rt_task(struct rq
*rq
, struct task_struct
*p
, int cpu
)
1687 if (!task_running(rq
, p
) &&
1688 cpumask_test_cpu(cpu
, &p
->cpus_mask
))
1695 * Return the highest pushable rq's task, which is suitable to be executed
1696 * on the CPU, NULL otherwise
1698 static struct task_struct
*pick_highest_pushable_task(struct rq
*rq
, int cpu
)
1700 struct plist_head
*head
= &rq
->rt
.pushable_tasks
;
1701 struct task_struct
*p
;
1703 if (!has_pushable_tasks(rq
))
1706 plist_for_each_entry(p
, head
, pushable_tasks
) {
1707 if (pick_rt_task(rq
, p
, cpu
))
1714 static DEFINE_PER_CPU(cpumask_var_t
, local_cpu_mask
);
1716 static int find_lowest_rq(struct task_struct
*task
)
1718 struct sched_domain
*sd
;
1719 struct cpumask
*lowest_mask
= this_cpu_cpumask_var_ptr(local_cpu_mask
);
1720 int this_cpu
= smp_processor_id();
1721 int cpu
= task_cpu(task
);
1724 /* Make sure the mask is initialized first */
1725 if (unlikely(!lowest_mask
))
1728 if (task
->nr_cpus_allowed
== 1)
1729 return -1; /* No other targets possible */
1732 * If we're on asym system ensure we consider the different capacities
1733 * of the CPUs when searching for the lowest_mask.
1735 if (static_branch_unlikely(&sched_asym_cpucapacity
)) {
1737 ret
= cpupri_find_fitness(&task_rq(task
)->rd
->cpupri
,
1739 rt_task_fits_capacity
);
1742 ret
= cpupri_find(&task_rq(task
)->rd
->cpupri
,
1747 return -1; /* No targets found */
1750 * At this point we have built a mask of CPUs representing the
1751 * lowest priority tasks in the system. Now we want to elect
1752 * the best one based on our affinity and topology.
1754 * We prioritize the last CPU that the task executed on since
1755 * it is most likely cache-hot in that location.
1757 if (cpumask_test_cpu(cpu
, lowest_mask
))
1761 * Otherwise, we consult the sched_domains span maps to figure
1762 * out which CPU is logically closest to our hot cache data.
1764 if (!cpumask_test_cpu(this_cpu
, lowest_mask
))
1765 this_cpu
= -1; /* Skip this_cpu opt if not among lowest */
1768 for_each_domain(cpu
, sd
) {
1769 if (sd
->flags
& SD_WAKE_AFFINE
) {
1773 * "this_cpu" is cheaper to preempt than a
1776 if (this_cpu
!= -1 &&
1777 cpumask_test_cpu(this_cpu
, sched_domain_span(sd
))) {
1782 best_cpu
= cpumask_any_and_distribute(lowest_mask
,
1783 sched_domain_span(sd
));
1784 if (best_cpu
< nr_cpu_ids
) {
1793 * And finally, if there were no matches within the domains
1794 * just give the caller *something* to work with from the compatible
1800 cpu
= cpumask_any_distribute(lowest_mask
);
1801 if (cpu
< nr_cpu_ids
)
1807 /* Will lock the rq it finds */
1808 static struct rq
*find_lock_lowest_rq(struct task_struct
*task
, struct rq
*rq
)
1810 struct rq
*lowest_rq
= NULL
;
1814 for (tries
= 0; tries
< RT_MAX_TRIES
; tries
++) {
1815 cpu
= find_lowest_rq(task
);
1817 if ((cpu
== -1) || (cpu
== rq
->cpu
))
1820 lowest_rq
= cpu_rq(cpu
);
1822 if (lowest_rq
->rt
.highest_prio
.curr
<= task
->prio
) {
1824 * Target rq has tasks of equal or higher priority,
1825 * retrying does not release any lock and is unlikely
1826 * to yield a different result.
1832 /* if the prio of this runqueue changed, try again */
1833 if (double_lock_balance(rq
, lowest_rq
)) {
1835 * We had to unlock the run queue. In
1836 * the mean time, task could have
1837 * migrated already or had its affinity changed.
1838 * Also make sure that it wasn't scheduled on its rq.
1840 if (unlikely(task_rq(task
) != rq
||
1841 !cpumask_test_cpu(lowest_rq
->cpu
, &task
->cpus_mask
) ||
1842 task_running(rq
, task
) ||
1844 !task_on_rq_queued(task
))) {
1846 double_unlock_balance(rq
, lowest_rq
);
1852 /* If this rq is still suitable use it. */
1853 if (lowest_rq
->rt
.highest_prio
.curr
> task
->prio
)
1857 double_unlock_balance(rq
, lowest_rq
);
1864 static struct task_struct
*pick_next_pushable_task(struct rq
*rq
)
1866 struct task_struct
*p
;
1868 if (!has_pushable_tasks(rq
))
1871 p
= plist_first_entry(&rq
->rt
.pushable_tasks
,
1872 struct task_struct
, pushable_tasks
);
1874 BUG_ON(rq
->cpu
!= task_cpu(p
));
1875 BUG_ON(task_current(rq
, p
));
1876 BUG_ON(p
->nr_cpus_allowed
<= 1);
1878 BUG_ON(!task_on_rq_queued(p
));
1879 BUG_ON(!rt_task(p
));
1885 * If the current CPU has more than one RT task, see if the non
1886 * running task can migrate over to a CPU that is running a task
1887 * of lesser priority.
1889 static int push_rt_task(struct rq
*rq
, bool pull
)
1891 struct task_struct
*next_task
;
1892 struct rq
*lowest_rq
;
1895 if (!rq
->rt
.overloaded
)
1898 next_task
= pick_next_pushable_task(rq
);
1904 * It's possible that the next_task slipped in of
1905 * higher priority than current. If that's the case
1906 * just reschedule current.
1908 if (unlikely(next_task
->prio
< rq
->curr
->prio
)) {
1913 if (is_migration_disabled(next_task
)) {
1914 struct task_struct
*push_task
= NULL
;
1917 if (!pull
|| rq
->push_busy
)
1921 * Invoking find_lowest_rq() on anything but an RT task doesn't
1922 * make sense. Per the above priority check, curr has to
1923 * be of higher priority than next_task, so no need to
1924 * reschedule when bailing out.
1926 * Note that the stoppers are masqueraded as SCHED_FIFO
1927 * (cf. sched_set_stop_task()), so we can't rely on rt_task().
1929 if (rq
->curr
->sched_class
!= &rt_sched_class
)
1932 cpu
= find_lowest_rq(rq
->curr
);
1933 if (cpu
== -1 || cpu
== rq
->cpu
)
1937 * Given we found a CPU with lower priority than @next_task,
1938 * therefore it should be running. However we cannot migrate it
1939 * to this other CPU, instead attempt to push the current
1940 * running task on this CPU away.
1942 push_task
= get_push_task(rq
);
1944 raw_spin_rq_unlock(rq
);
1945 stop_one_cpu_nowait(rq
->cpu
, push_cpu_stop
,
1946 push_task
, &rq
->push_work
);
1947 raw_spin_rq_lock(rq
);
1953 if (WARN_ON(next_task
== rq
->curr
))
1956 /* We might release rq lock */
1957 get_task_struct(next_task
);
1959 /* find_lock_lowest_rq locks the rq if found */
1960 lowest_rq
= find_lock_lowest_rq(next_task
, rq
);
1962 struct task_struct
*task
;
1964 * find_lock_lowest_rq releases rq->lock
1965 * so it is possible that next_task has migrated.
1967 * We need to make sure that the task is still on the same
1968 * run-queue and is also still the next task eligible for
1971 task
= pick_next_pushable_task(rq
);
1972 if (task
== next_task
) {
1974 * The task hasn't migrated, and is still the next
1975 * eligible task, but we failed to find a run-queue
1976 * to push it to. Do not retry in this case, since
1977 * other CPUs will pull from us when ready.
1983 /* No more tasks, just exit */
1987 * Something has shifted, try again.
1989 put_task_struct(next_task
);
1994 deactivate_task(rq
, next_task
, 0);
1995 set_task_cpu(next_task
, lowest_rq
->cpu
);
1996 activate_task(lowest_rq
, next_task
, 0);
1997 resched_curr(lowest_rq
);
2000 double_unlock_balance(rq
, lowest_rq
);
2002 put_task_struct(next_task
);
2007 static void push_rt_tasks(struct rq
*rq
)
2009 /* push_rt_task will return true if it moved an RT */
2010 while (push_rt_task(rq
, false))
2014 #ifdef HAVE_RT_PUSH_IPI
2017 * When a high priority task schedules out from a CPU and a lower priority
2018 * task is scheduled in, a check is made to see if there's any RT tasks
2019 * on other CPUs that are waiting to run because a higher priority RT task
2020 * is currently running on its CPU. In this case, the CPU with multiple RT
2021 * tasks queued on it (overloaded) needs to be notified that a CPU has opened
2022 * up that may be able to run one of its non-running queued RT tasks.
2024 * All CPUs with overloaded RT tasks need to be notified as there is currently
2025 * no way to know which of these CPUs have the highest priority task waiting
2026 * to run. Instead of trying to take a spinlock on each of these CPUs,
2027 * which has shown to cause large latency when done on machines with many
2028 * CPUs, sending an IPI to the CPUs to have them push off the overloaded
2029 * RT tasks waiting to run.
2031 * Just sending an IPI to each of the CPUs is also an issue, as on large
2032 * count CPU machines, this can cause an IPI storm on a CPU, especially
2033 * if its the only CPU with multiple RT tasks queued, and a large number
2034 * of CPUs scheduling a lower priority task at the same time.
2036 * Each root domain has its own irq work function that can iterate over
2037 * all CPUs with RT overloaded tasks. Since all CPUs with overloaded RT
2038 * task must be checked if there's one or many CPUs that are lowering
2039 * their priority, there's a single irq work iterator that will try to
2040 * push off RT tasks that are waiting to run.
2042 * When a CPU schedules a lower priority task, it will kick off the
2043 * irq work iterator that will jump to each CPU with overloaded RT tasks.
2044 * As it only takes the first CPU that schedules a lower priority task
2045 * to start the process, the rto_start variable is incremented and if
2046 * the atomic result is one, then that CPU will try to take the rto_lock.
2047 * This prevents high contention on the lock as the process handles all
2048 * CPUs scheduling lower priority tasks.
2050 * All CPUs that are scheduling a lower priority task will increment the
2051 * rt_loop_next variable. This will make sure that the irq work iterator
2052 * checks all RT overloaded CPUs whenever a CPU schedules a new lower
2053 * priority task, even if the iterator is in the middle of a scan. Incrementing
2054 * the rt_loop_next will cause the iterator to perform another scan.
2057 static int rto_next_cpu(struct root_domain
*rd
)
2063 * When starting the IPI RT pushing, the rto_cpu is set to -1,
2064 * rt_next_cpu() will simply return the first CPU found in
2067 * If rto_next_cpu() is called with rto_cpu is a valid CPU, it
2068 * will return the next CPU found in the rto_mask.
2070 * If there are no more CPUs left in the rto_mask, then a check is made
2071 * against rto_loop and rto_loop_next. rto_loop is only updated with
2072 * the rto_lock held, but any CPU may increment the rto_loop_next
2073 * without any locking.
2077 /* When rto_cpu is -1 this acts like cpumask_first() */
2078 cpu
= cpumask_next(rd
->rto_cpu
, rd
->rto_mask
);
2082 if (cpu
< nr_cpu_ids
)
2088 * ACQUIRE ensures we see the @rto_mask changes
2089 * made prior to the @next value observed.
2091 * Matches WMB in rt_set_overload().
2093 next
= atomic_read_acquire(&rd
->rto_loop_next
);
2095 if (rd
->rto_loop
== next
)
2098 rd
->rto_loop
= next
;
2104 static inline bool rto_start_trylock(atomic_t
*v
)
2106 return !atomic_cmpxchg_acquire(v
, 0, 1);
2109 static inline void rto_start_unlock(atomic_t
*v
)
2111 atomic_set_release(v
, 0);
2114 static void tell_cpu_to_push(struct rq
*rq
)
2118 /* Keep the loop going if the IPI is currently active */
2119 atomic_inc(&rq
->rd
->rto_loop_next
);
2121 /* Only one CPU can initiate a loop at a time */
2122 if (!rto_start_trylock(&rq
->rd
->rto_loop_start
))
2125 raw_spin_lock(&rq
->rd
->rto_lock
);
2128 * The rto_cpu is updated under the lock, if it has a valid CPU
2129 * then the IPI is still running and will continue due to the
2130 * update to loop_next, and nothing needs to be done here.
2131 * Otherwise it is finishing up and an ipi needs to be sent.
2133 if (rq
->rd
->rto_cpu
< 0)
2134 cpu
= rto_next_cpu(rq
->rd
);
2136 raw_spin_unlock(&rq
->rd
->rto_lock
);
2138 rto_start_unlock(&rq
->rd
->rto_loop_start
);
2141 /* Make sure the rd does not get freed while pushing */
2142 sched_get_rd(rq
->rd
);
2143 irq_work_queue_on(&rq
->rd
->rto_push_work
, cpu
);
2147 /* Called from hardirq context */
2148 void rto_push_irq_work_func(struct irq_work
*work
)
2150 struct root_domain
*rd
=
2151 container_of(work
, struct root_domain
, rto_push_work
);
2158 * We do not need to grab the lock to check for has_pushable_tasks.
2159 * When it gets updated, a check is made if a push is possible.
2161 if (has_pushable_tasks(rq
)) {
2162 raw_spin_rq_lock(rq
);
2163 while (push_rt_task(rq
, true))
2165 raw_spin_rq_unlock(rq
);
2168 raw_spin_lock(&rd
->rto_lock
);
2170 /* Pass the IPI to the next rt overloaded queue */
2171 cpu
= rto_next_cpu(rd
);
2173 raw_spin_unlock(&rd
->rto_lock
);
2180 /* Try the next RT overloaded CPU */
2181 irq_work_queue_on(&rd
->rto_push_work
, cpu
);
2183 #endif /* HAVE_RT_PUSH_IPI */
2185 static void pull_rt_task(struct rq
*this_rq
)
2187 int this_cpu
= this_rq
->cpu
, cpu
;
2188 bool resched
= false;
2189 struct task_struct
*p
, *push_task
;
2191 int rt_overload_count
= rt_overloaded(this_rq
);
2193 if (likely(!rt_overload_count
))
2197 * Match the barrier from rt_set_overloaded; this guarantees that if we
2198 * see overloaded we must also see the rto_mask bit.
2202 /* If we are the only overloaded CPU do nothing */
2203 if (rt_overload_count
== 1 &&
2204 cpumask_test_cpu(this_rq
->cpu
, this_rq
->rd
->rto_mask
))
2207 #ifdef HAVE_RT_PUSH_IPI
2208 if (sched_feat(RT_PUSH_IPI
)) {
2209 tell_cpu_to_push(this_rq
);
2214 for_each_cpu(cpu
, this_rq
->rd
->rto_mask
) {
2215 if (this_cpu
== cpu
)
2218 src_rq
= cpu_rq(cpu
);
2221 * Don't bother taking the src_rq->lock if the next highest
2222 * task is known to be lower-priority than our current task.
2223 * This may look racy, but if this value is about to go
2224 * logically higher, the src_rq will push this task away.
2225 * And if its going logically lower, we do not care
2227 if (src_rq
->rt
.highest_prio
.next
>=
2228 this_rq
->rt
.highest_prio
.curr
)
2232 * We can potentially drop this_rq's lock in
2233 * double_lock_balance, and another CPU could
2237 double_lock_balance(this_rq
, src_rq
);
2240 * We can pull only a task, which is pushable
2241 * on its rq, and no others.
2243 p
= pick_highest_pushable_task(src_rq
, this_cpu
);
2246 * Do we have an RT task that preempts
2247 * the to-be-scheduled task?
2249 if (p
&& (p
->prio
< this_rq
->rt
.highest_prio
.curr
)) {
2250 WARN_ON(p
== src_rq
->curr
);
2251 WARN_ON(!task_on_rq_queued(p
));
2254 * There's a chance that p is higher in priority
2255 * than what's currently running on its CPU.
2256 * This is just that p is waking up and hasn't
2257 * had a chance to schedule. We only pull
2258 * p if it is lower in priority than the
2259 * current task on the run queue
2261 if (p
->prio
< src_rq
->curr
->prio
)
2264 if (is_migration_disabled(p
)) {
2265 push_task
= get_push_task(src_rq
);
2267 deactivate_task(src_rq
, p
, 0);
2268 set_task_cpu(p
, this_cpu
);
2269 activate_task(this_rq
, p
, 0);
2273 * We continue with the search, just in
2274 * case there's an even higher prio task
2275 * in another runqueue. (low likelihood
2280 double_unlock_balance(this_rq
, src_rq
);
2283 raw_spin_rq_unlock(this_rq
);
2284 stop_one_cpu_nowait(src_rq
->cpu
, push_cpu_stop
,
2285 push_task
, &src_rq
->push_work
);
2286 raw_spin_rq_lock(this_rq
);
2291 resched_curr(this_rq
);
2295 * If we are not running and we are not going to reschedule soon, we should
2296 * try to push tasks away now
2298 static void task_woken_rt(struct rq
*rq
, struct task_struct
*p
)
2300 bool need_to_push
= !task_running(rq
, p
) &&
2301 !test_tsk_need_resched(rq
->curr
) &&
2302 p
->nr_cpus_allowed
> 1 &&
2303 (dl_task(rq
->curr
) || rt_task(rq
->curr
)) &&
2304 (rq
->curr
->nr_cpus_allowed
< 2 ||
2305 rq
->curr
->prio
<= p
->prio
);
2311 /* Assumes rq->lock is held */
2312 static void rq_online_rt(struct rq
*rq
)
2314 if (rq
->rt
.overloaded
)
2315 rt_set_overload(rq
);
2317 __enable_runtime(rq
);
2319 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, rq
->rt
.highest_prio
.curr
);
2322 /* Assumes rq->lock is held */
2323 static void rq_offline_rt(struct rq
*rq
)
2325 if (rq
->rt
.overloaded
)
2326 rt_clear_overload(rq
);
2328 __disable_runtime(rq
);
2330 cpupri_set(&rq
->rd
->cpupri
, rq
->cpu
, CPUPRI_INVALID
);
2334 * When switch from the rt queue, we bring ourselves to a position
2335 * that we might want to pull RT tasks from other runqueues.
2337 static void switched_from_rt(struct rq
*rq
, struct task_struct
*p
)
2340 * If there are other RT tasks then we will reschedule
2341 * and the scheduling of the other RT tasks will handle
2342 * the balancing. But if we are the last RT task
2343 * we may need to handle the pulling of RT tasks
2346 if (!task_on_rq_queued(p
) || rq
->rt
.rt_nr_running
)
2349 rt_queue_pull_task(rq
);
2352 void __init
init_sched_rt_class(void)
2356 for_each_possible_cpu(i
) {
2357 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask
, i
),
2358 GFP_KERNEL
, cpu_to_node(i
));
2361 #endif /* CONFIG_SMP */
2364 * When switching a task to RT, we may overload the runqueue
2365 * with RT tasks. In this case we try to push them off to
2368 static void switched_to_rt(struct rq
*rq
, struct task_struct
*p
)
2371 * If we are running, update the avg_rt tracking, as the running time
2372 * will now on be accounted into the latter.
2374 if (task_current(rq
, p
)) {
2375 update_rt_rq_load_avg(rq_clock_pelt(rq
), rq
, 0);
2380 * If we are not running we may need to preempt the current
2381 * running task. If that current running task is also an RT task
2382 * then see if we can move to another run queue.
2384 if (task_on_rq_queued(p
)) {
2386 if (p
->nr_cpus_allowed
> 1 && rq
->rt
.overloaded
)
2387 rt_queue_push_tasks(rq
);
2388 #endif /* CONFIG_SMP */
2389 if (p
->prio
< rq
->curr
->prio
&& cpu_online(cpu_of(rq
)))
2395 * Priority of the task has changed. This may cause
2396 * us to initiate a push or pull.
2399 prio_changed_rt(struct rq
*rq
, struct task_struct
*p
, int oldprio
)
2401 if (!task_on_rq_queued(p
))
2404 if (task_current(rq
, p
)) {
2407 * If our priority decreases while running, we
2408 * may need to pull tasks to this runqueue.
2410 if (oldprio
< p
->prio
)
2411 rt_queue_pull_task(rq
);
2414 * If there's a higher priority task waiting to run
2417 if (p
->prio
> rq
->rt
.highest_prio
.curr
)
2420 /* For UP simply resched on drop of prio */
2421 if (oldprio
< p
->prio
)
2423 #endif /* CONFIG_SMP */
2426 * This task is not running, but if it is
2427 * greater than the current running task
2430 if (p
->prio
< rq
->curr
->prio
)
2435 #ifdef CONFIG_POSIX_TIMERS
2436 static void watchdog(struct rq
*rq
, struct task_struct
*p
)
2438 unsigned long soft
, hard
;
2440 /* max may change after cur was read, this will be fixed next tick */
2441 soft
= task_rlimit(p
, RLIMIT_RTTIME
);
2442 hard
= task_rlimit_max(p
, RLIMIT_RTTIME
);
2444 if (soft
!= RLIM_INFINITY
) {
2447 if (p
->rt
.watchdog_stamp
!= jiffies
) {
2449 p
->rt
.watchdog_stamp
= jiffies
;
2452 next
= DIV_ROUND_UP(min(soft
, hard
), USEC_PER_SEC
/HZ
);
2453 if (p
->rt
.timeout
> next
) {
2454 posix_cputimers_rt_watchdog(&p
->posix_cputimers
,
2455 p
->se
.sum_exec_runtime
);
2460 static inline void watchdog(struct rq
*rq
, struct task_struct
*p
) { }
2464 * scheduler tick hitting a task of our scheduling class.
2466 * NOTE: This function can be called remotely by the tick offload that
2467 * goes along full dynticks. Therefore no local assumption can be made
2468 * and everything must be accessed through the @rq and @curr passed in
2471 static void task_tick_rt(struct rq
*rq
, struct task_struct
*p
, int queued
)
2473 struct sched_rt_entity
*rt_se
= &p
->rt
;
2476 update_rt_rq_load_avg(rq_clock_pelt(rq
), rq
, 1);
2481 * RR tasks need a special form of timeslice management.
2482 * FIFO tasks have no timeslices.
2484 if (p
->policy
!= SCHED_RR
)
2487 if (--p
->rt
.time_slice
)
2490 p
->rt
.time_slice
= sched_rr_timeslice
;
2493 * Requeue to the end of queue if we (and all of our ancestors) are not
2494 * the only element on the queue
2496 for_each_sched_rt_entity(rt_se
) {
2497 if (rt_se
->run_list
.prev
!= rt_se
->run_list
.next
) {
2498 requeue_task_rt(rq
, p
, 0);
2505 static unsigned int get_rr_interval_rt(struct rq
*rq
, struct task_struct
*task
)
2508 * Time slice is 0 for SCHED_FIFO tasks
2510 if (task
->policy
== SCHED_RR
)
2511 return sched_rr_timeslice
;
2516 DEFINE_SCHED_CLASS(rt
) = {
2518 .enqueue_task
= enqueue_task_rt
,
2519 .dequeue_task
= dequeue_task_rt
,
2520 .yield_task
= yield_task_rt
,
2522 .check_preempt_curr
= check_preempt_curr_rt
,
2524 .pick_next_task
= pick_next_task_rt
,
2525 .put_prev_task
= put_prev_task_rt
,
2526 .set_next_task
= set_next_task_rt
,
2529 .balance
= balance_rt
,
2530 .pick_task
= pick_task_rt
,
2531 .select_task_rq
= select_task_rq_rt
,
2532 .set_cpus_allowed
= set_cpus_allowed_common
,
2533 .rq_online
= rq_online_rt
,
2534 .rq_offline
= rq_offline_rt
,
2535 .task_woken
= task_woken_rt
,
2536 .switched_from
= switched_from_rt
,
2537 .find_lock_rq
= find_lock_lowest_rq
,
2540 .task_tick
= task_tick_rt
,
2542 .get_rr_interval
= get_rr_interval_rt
,
2544 .prio_changed
= prio_changed_rt
,
2545 .switched_to
= switched_to_rt
,
2547 .update_curr
= update_curr_rt
,
2549 #ifdef CONFIG_UCLAMP_TASK
2550 .uclamp_enabled
= 1,
2554 #ifdef CONFIG_RT_GROUP_SCHED
2556 * Ensure that the real time constraints are schedulable.
2558 static DEFINE_MUTEX(rt_constraints_mutex
);
2560 static inline int tg_has_rt_tasks(struct task_group
*tg
)
2562 struct task_struct
*task
;
2563 struct css_task_iter it
;
2567 * Autogroups do not have RT tasks; see autogroup_create().
2569 if (task_group_is_autogroup(tg
))
2572 css_task_iter_start(&tg
->css
, 0, &it
);
2573 while (!ret
&& (task
= css_task_iter_next(&it
)))
2574 ret
|= rt_task(task
);
2575 css_task_iter_end(&it
);
2580 struct rt_schedulable_data
{
2581 struct task_group
*tg
;
2586 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
2588 struct rt_schedulable_data
*d
= data
;
2589 struct task_group
*child
;
2590 unsigned long total
, sum
= 0;
2591 u64 period
, runtime
;
2593 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2594 runtime
= tg
->rt_bandwidth
.rt_runtime
;
2597 period
= d
->rt_period
;
2598 runtime
= d
->rt_runtime
;
2602 * Cannot have more runtime than the period.
2604 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
2608 * Ensure we don't starve existing RT tasks if runtime turns zero.
2610 if (rt_bandwidth_enabled() && !runtime
&&
2611 tg
->rt_bandwidth
.rt_runtime
&& tg_has_rt_tasks(tg
))
2614 total
= to_ratio(period
, runtime
);
2617 * Nobody can have more than the global setting allows.
2619 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
2623 * The sum of our children's runtime should not exceed our own.
2625 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
2626 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
2627 runtime
= child
->rt_bandwidth
.rt_runtime
;
2629 if (child
== d
->tg
) {
2630 period
= d
->rt_period
;
2631 runtime
= d
->rt_runtime
;
2634 sum
+= to_ratio(period
, runtime
);
2643 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
2647 struct rt_schedulable_data data
= {
2649 .rt_period
= period
,
2650 .rt_runtime
= runtime
,
2654 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
2660 static int tg_set_rt_bandwidth(struct task_group
*tg
,
2661 u64 rt_period
, u64 rt_runtime
)
2666 * Disallowing the root group RT runtime is BAD, it would disallow the
2667 * kernel creating (and or operating) RT threads.
2669 if (tg
== &root_task_group
&& rt_runtime
== 0)
2672 /* No period doesn't make any sense. */
2677 * Bound quota to defend quota against overflow during bandwidth shift.
2679 if (rt_runtime
!= RUNTIME_INF
&& rt_runtime
> max_rt_runtime
)
2682 mutex_lock(&rt_constraints_mutex
);
2683 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
2687 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
2688 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
2689 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
2691 for_each_possible_cpu(i
) {
2692 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
2694 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
2695 rt_rq
->rt_runtime
= rt_runtime
;
2696 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
2698 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
2700 mutex_unlock(&rt_constraints_mutex
);
2705 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
2707 u64 rt_runtime
, rt_period
;
2709 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2710 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
2711 if (rt_runtime_us
< 0)
2712 rt_runtime
= RUNTIME_INF
;
2713 else if ((u64
)rt_runtime_us
> U64_MAX
/ NSEC_PER_USEC
)
2716 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
2719 long sched_group_rt_runtime(struct task_group
*tg
)
2723 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
2726 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
2727 do_div(rt_runtime_us
, NSEC_PER_USEC
);
2728 return rt_runtime_us
;
2731 int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
2733 u64 rt_runtime
, rt_period
;
2735 if (rt_period_us
> U64_MAX
/ NSEC_PER_USEC
)
2738 rt_period
= rt_period_us
* NSEC_PER_USEC
;
2739 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
2741 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
2744 long sched_group_rt_period(struct task_group
*tg
)
2748 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
2749 do_div(rt_period_us
, NSEC_PER_USEC
);
2750 return rt_period_us
;
2753 static int sched_rt_global_constraints(void)
2757 mutex_lock(&rt_constraints_mutex
);
2758 ret
= __rt_schedulable(NULL
, 0, 0);
2759 mutex_unlock(&rt_constraints_mutex
);
2764 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
2766 /* Don't accept realtime tasks when there is no way for them to run */
2767 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
2773 #else /* !CONFIG_RT_GROUP_SCHED */
2774 static int sched_rt_global_constraints(void)
2776 unsigned long flags
;
2779 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2780 for_each_possible_cpu(i
) {
2781 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
2783 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
2784 rt_rq
->rt_runtime
= global_rt_runtime();
2785 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
2787 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2791 #endif /* CONFIG_RT_GROUP_SCHED */
2793 static int sched_rt_global_validate(void)
2795 if (sysctl_sched_rt_period
<= 0)
2798 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
2799 ((sysctl_sched_rt_runtime
> sysctl_sched_rt_period
) ||
2800 ((u64
)sysctl_sched_rt_runtime
*
2801 NSEC_PER_USEC
> max_rt_runtime
)))
2807 static void sched_rt_do_global(void)
2809 unsigned long flags
;
2811 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2812 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
2813 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
2814 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
2817 int sched_rt_handler(struct ctl_table
*table
, int write
, void *buffer
,
2818 size_t *lenp
, loff_t
*ppos
)
2820 int old_period
, old_runtime
;
2821 static DEFINE_MUTEX(mutex
);
2825 old_period
= sysctl_sched_rt_period
;
2826 old_runtime
= sysctl_sched_rt_runtime
;
2828 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
2830 if (!ret
&& write
) {
2831 ret
= sched_rt_global_validate();
2835 ret
= sched_dl_global_validate();
2839 ret
= sched_rt_global_constraints();
2843 sched_rt_do_global();
2844 sched_dl_do_global();
2848 sysctl_sched_rt_period
= old_period
;
2849 sysctl_sched_rt_runtime
= old_runtime
;
2851 mutex_unlock(&mutex
);
2856 int sched_rr_handler(struct ctl_table
*table
, int write
, void *buffer
,
2857 size_t *lenp
, loff_t
*ppos
)
2860 static DEFINE_MUTEX(mutex
);
2863 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
2865 * Make sure that internally we keep jiffies.
2866 * Also, writing zero resets the timeslice to default:
2868 if (!ret
&& write
) {
2869 sched_rr_timeslice
=
2870 sysctl_sched_rr_timeslice
<= 0 ? RR_TIMESLICE
:
2871 msecs_to_jiffies(sysctl_sched_rr_timeslice
);
2873 mutex_unlock(&mutex
);
2878 #ifdef CONFIG_SCHED_DEBUG
2879 void print_rt_stats(struct seq_file
*m
, int cpu
)
2882 struct rt_rq
*rt_rq
;
2885 for_each_rt_rq(rt_rq
, iter
, cpu_rq(cpu
))
2886 print_rt_rq(m
, cpu
, rt_rq
);
2889 #endif /* CONFIG_SCHED_DEBUG */