1 // SPDX-License-Identifier: GPL-2.0
3 * Deadline Scheduling Class (SCHED_DEADLINE)
5 * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS).
7 * Tasks that periodically executes their instances for less than their
8 * runtime won't miss any of their deadlines.
9 * Tasks that are not periodic or sporadic or that tries to execute more
10 * than their reserved bandwidth will be slowed down (and may potentially
11 * miss some of their deadlines), and won't affect any other task.
13 * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>,
14 * Juri Lelli <juri.lelli@gmail.com>,
15 * Michael Trimarchi <michael@amarulasolutions.com>,
16 * Fabio Checconi <fchecconi@gmail.com>
19 #include <linux/cpuset.h>
22 * Default limits for DL period; on the top end we guard against small util
23 * tasks still getting ridiculously long effective runtimes, on the bottom end we
24 * guard against timer DoS.
26 static unsigned int sysctl_sched_dl_period_max = 1 << 22; /* ~4 seconds */
27 static unsigned int sysctl_sched_dl_period_min = 100; /* 100 us */
29 static struct ctl_table sched_dl_sysctls[] = {
31 .procname = "sched_deadline_period_max_us",
32 .data = &sysctl_sched_dl_period_max,
33 .maxlen = sizeof(unsigned int),
35 .proc_handler = proc_douintvec_minmax,
36 .extra1 = (void *)&sysctl_sched_dl_period_min,
39 .procname = "sched_deadline_period_min_us",
40 .data = &sysctl_sched_dl_period_min,
41 .maxlen = sizeof(unsigned int),
43 .proc_handler = proc_douintvec_minmax,
44 .extra2 = (void *)&sysctl_sched_dl_period_max,
49 static int __init sched_dl_sysctl_init(void)
51 register_sysctl_init("kernel", sched_dl_sysctls);
54 late_initcall(sched_dl_sysctl_init);
57 static bool dl_server(struct sched_dl_entity *dl_se)
59 return dl_se->dl_server;
62 static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
64 BUG_ON(dl_server(dl_se));
65 return container_of(dl_se, struct task_struct, dl);
68 static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
70 return container_of(dl_rq, struct rq, dl);
73 static inline struct rq *rq_of_dl_se(struct sched_dl_entity *dl_se)
75 struct rq *rq = dl_se->rq;
77 if (!dl_server(dl_se))
78 rq = task_rq(dl_task_of(dl_se));
83 static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
85 return &rq_of_dl_se(dl_se)->dl;
88 static inline int on_dl_rq(struct sched_dl_entity *dl_se)
90 return !RB_EMPTY_NODE(&dl_se->rb_node);
93 #ifdef CONFIG_RT_MUTEXES
94 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
99 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
101 return pi_of(dl_se) != dl_se;
104 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
109 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
116 static inline struct dl_bw *dl_bw_of(int i)
118 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
119 "sched RCU must be held");
120 return &cpu_rq(i)->rd->dl_bw;
123 static inline int dl_bw_cpus(int i)
125 struct root_domain *rd = cpu_rq(i)->rd;
128 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
129 "sched RCU must be held");
131 if (cpumask_subset(rd->span, cpu_active_mask))
132 return cpumask_weight(rd->span);
136 for_each_cpu_and(i, rd->span, cpu_active_mask)
142 static inline unsigned long __dl_bw_capacity(const struct cpumask *mask)
144 unsigned long cap = 0;
147 for_each_cpu_and(i, mask, cpu_active_mask)
148 cap += arch_scale_cpu_capacity(i);
154 * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
155 * of the CPU the task is running on rather rd's \Sum CPU capacity.
157 static inline unsigned long dl_bw_capacity(int i)
159 if (!sched_asym_cpucap_active() &&
160 arch_scale_cpu_capacity(i) == SCHED_CAPACITY_SCALE) {
161 return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT;
163 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
164 "sched RCU must be held");
166 return __dl_bw_capacity(cpu_rq(i)->rd->span);
170 static inline bool dl_bw_visited(int cpu, u64 gen)
172 struct root_domain *rd = cpu_rq(cpu)->rd;
174 if (rd->visit_gen == gen)
182 void __dl_update(struct dl_bw *dl_b, s64 bw)
184 struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw);
187 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
188 "sched RCU must be held");
189 for_each_cpu_and(i, rd->span, cpu_active_mask) {
190 struct rq *rq = cpu_rq(i);
192 rq->dl.extra_bw += bw;
196 static inline struct dl_bw *dl_bw_of(int i)
198 return &cpu_rq(i)->dl.dl_bw;
201 static inline int dl_bw_cpus(int i)
206 static inline unsigned long dl_bw_capacity(int i)
208 return SCHED_CAPACITY_SCALE;
211 static inline bool dl_bw_visited(int cpu, u64 gen)
217 void __dl_update(struct dl_bw *dl_b, s64 bw)
219 struct dl_rq *dl = container_of(dl_b, struct dl_rq, dl_bw);
226 void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
228 dl_b->total_bw -= tsk_bw;
229 __dl_update(dl_b, (s32)tsk_bw / cpus);
233 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
235 dl_b->total_bw += tsk_bw;
236 __dl_update(dl_b, -((s32)tsk_bw / cpus));
240 __dl_overflow(struct dl_bw *dl_b, unsigned long cap, u64 old_bw, u64 new_bw)
242 return dl_b->bw != -1 &&
243 cap_scale(dl_b->bw, cap) < dl_b->total_bw - old_bw + new_bw;
247 void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
249 u64 old = dl_rq->running_bw;
251 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
252 dl_rq->running_bw += dl_bw;
253 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
254 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
255 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
256 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
260 void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
262 u64 old = dl_rq->running_bw;
264 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
265 dl_rq->running_bw -= dl_bw;
266 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
267 if (dl_rq->running_bw > old)
268 dl_rq->running_bw = 0;
269 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
270 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
274 void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
276 u64 old = dl_rq->this_bw;
278 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
279 dl_rq->this_bw += dl_bw;
280 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
284 void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
286 u64 old = dl_rq->this_bw;
288 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
289 dl_rq->this_bw -= dl_bw;
290 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
291 if (dl_rq->this_bw > old)
293 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
297 void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
299 if (!dl_entity_is_special(dl_se))
300 __add_rq_bw(dl_se->dl_bw, dl_rq);
304 void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
306 if (!dl_entity_is_special(dl_se))
307 __sub_rq_bw(dl_se->dl_bw, dl_rq);
311 void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
313 if (!dl_entity_is_special(dl_se))
314 __add_running_bw(dl_se->dl_bw, dl_rq);
318 void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
320 if (!dl_entity_is_special(dl_se))
321 __sub_running_bw(dl_se->dl_bw, dl_rq);
324 static void dl_change_utilization(struct task_struct *p, u64 new_bw)
328 WARN_ON_ONCE(p->dl.flags & SCHED_FLAG_SUGOV);
330 if (task_on_rq_queued(p))
334 if (p->dl.dl_non_contending) {
335 sub_running_bw(&p->dl, &rq->dl);
336 p->dl.dl_non_contending = 0;
338 * If the timer handler is currently running and the
339 * timer cannot be canceled, inactive_task_timer()
340 * will see that dl_not_contending is not set, and
341 * will not touch the rq's active utilization,
342 * so we are still safe.
344 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
347 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
348 __add_rq_bw(new_bw, &rq->dl);
351 static void __dl_clear_params(struct sched_dl_entity *dl_se);
354 * The utilization of a task cannot be immediately removed from
355 * the rq active utilization (running_bw) when the task blocks.
356 * Instead, we have to wait for the so called "0-lag time".
358 * If a task blocks before the "0-lag time", a timer (the inactive
359 * timer) is armed, and running_bw is decreased when the timer
362 * If the task wakes up again before the inactive timer fires,
363 * the timer is canceled, whereas if the task wakes up after the
364 * inactive timer fired (and running_bw has been decreased) the
365 * task's utilization has to be added to running_bw again.
366 * A flag in the deadline scheduling entity (dl_non_contending)
367 * is used to avoid race conditions between the inactive timer handler
370 * The following diagram shows how running_bw is updated. A task is
371 * "ACTIVE" when its utilization contributes to running_bw; an
372 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
373 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
374 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
375 * time already passed, which does not contribute to running_bw anymore.
376 * +------------------+
378 * +------------------>+ contending |
379 * | add_running_bw | |
380 * | +----+------+------+
383 * +--------+-------+ | |
384 * | | t >= 0-lag | | wakeup
385 * | INACTIVE |<---------------+ |
386 * | | sub_running_bw | |
387 * +--------+-------+ | |
392 * | +----+------+------+
393 * | sub_running_bw | ACTIVE |
394 * +-------------------+ |
395 * inactive timer | non contending |
396 * fired +------------------+
398 * The task_non_contending() function is invoked when a task
399 * blocks, and checks if the 0-lag time already passed or
400 * not (in the first case, it directly updates running_bw;
401 * in the second case, it arms the inactive timer).
403 * The task_contending() function is invoked when a task wakes
404 * up, and checks if the task is still in the "ACTIVE non contending"
405 * state or not (in the second case, it updates running_bw).
407 static void task_non_contending(struct sched_dl_entity *dl_se)
409 struct hrtimer *timer = &dl_se->inactive_timer;
410 struct rq *rq = rq_of_dl_se(dl_se);
411 struct dl_rq *dl_rq = &rq->dl;
415 * If this is a non-deadline task that has been boosted,
418 if (dl_se->dl_runtime == 0)
421 if (dl_entity_is_special(dl_se))
424 WARN_ON(dl_se->dl_non_contending);
426 zerolag_time = dl_se->deadline -
427 div64_long((dl_se->runtime * dl_se->dl_period),
431 * Using relative times instead of the absolute "0-lag time"
432 * allows to simplify the code
434 zerolag_time -= rq_clock(rq);
437 * If the "0-lag time" already passed, decrease the active
438 * utilization now, instead of starting a timer
440 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
441 if (dl_server(dl_se)) {
442 sub_running_bw(dl_se, dl_rq);
444 struct task_struct *p = dl_task_of(dl_se);
447 sub_running_bw(dl_se, dl_rq);
449 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
450 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
452 if (READ_ONCE(p->__state) == TASK_DEAD)
453 sub_rq_bw(dl_se, &rq->dl);
454 raw_spin_lock(&dl_b->lock);
455 __dl_sub(dl_b, dl_se->dl_bw, dl_bw_cpus(task_cpu(p)));
456 raw_spin_unlock(&dl_b->lock);
457 __dl_clear_params(dl_se);
464 dl_se->dl_non_contending = 1;
465 if (!dl_server(dl_se))
466 get_task_struct(dl_task_of(dl_se));
468 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD);
471 static void task_contending(struct sched_dl_entity *dl_se, int flags)
473 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
476 * If this is a non-deadline task that has been boosted,
479 if (dl_se->dl_runtime == 0)
482 if (flags & ENQUEUE_MIGRATED)
483 add_rq_bw(dl_se, dl_rq);
485 if (dl_se->dl_non_contending) {
486 dl_se->dl_non_contending = 0;
488 * If the timer handler is currently running and the
489 * timer cannot be canceled, inactive_task_timer()
490 * will see that dl_not_contending is not set, and
491 * will not touch the rq's active utilization,
492 * so we are still safe.
494 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1) {
495 if (!dl_server(dl_se))
496 put_task_struct(dl_task_of(dl_se));
500 * Since "dl_non_contending" is not set, the
501 * task's utilization has already been removed from
502 * active utilization (either when the task blocked,
503 * when the "inactive timer" fired).
506 add_running_bw(dl_se, dl_rq);
510 static inline int is_leftmost(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
512 return rb_first_cached(&dl_rq->root) == &dl_se->rb_node;
515 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
517 void init_dl_bw(struct dl_bw *dl_b)
519 raw_spin_lock_init(&dl_b->lock);
520 if (global_rt_runtime() == RUNTIME_INF)
523 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
527 void init_dl_rq(struct dl_rq *dl_rq)
529 dl_rq->root = RB_ROOT_CACHED;
532 /* zero means no -deadline tasks */
533 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
535 dl_rq->overloaded = 0;
536 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
538 init_dl_bw(&dl_rq->dl_bw);
541 dl_rq->running_bw = 0;
543 init_dl_rq_bw_ratio(dl_rq);
548 static inline int dl_overloaded(struct rq *rq)
550 return atomic_read(&rq->rd->dlo_count);
553 static inline void dl_set_overload(struct rq *rq)
558 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
560 * Must be visible before the overload count is
561 * set (as in sched_rt.c).
563 * Matched by the barrier in pull_dl_task().
566 atomic_inc(&rq->rd->dlo_count);
569 static inline void dl_clear_overload(struct rq *rq)
574 atomic_dec(&rq->rd->dlo_count);
575 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
578 #define __node_2_pdl(node) \
579 rb_entry((node), struct task_struct, pushable_dl_tasks)
581 static inline bool __pushable_less(struct rb_node *a, const struct rb_node *b)
583 return dl_entity_preempt(&__node_2_pdl(a)->dl, &__node_2_pdl(b)->dl);
586 static inline int has_pushable_dl_tasks(struct rq *rq)
588 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
592 * The list of pushable -deadline task is not a plist, like in
593 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
595 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
597 struct rb_node *leftmost;
599 WARN_ON_ONCE(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
601 leftmost = rb_add_cached(&p->pushable_dl_tasks,
602 &rq->dl.pushable_dl_tasks_root,
605 rq->dl.earliest_dl.next = p->dl.deadline;
607 if (!rq->dl.overloaded) {
609 rq->dl.overloaded = 1;
613 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
615 struct dl_rq *dl_rq = &rq->dl;
616 struct rb_root_cached *root = &dl_rq->pushable_dl_tasks_root;
617 struct rb_node *leftmost;
619 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
622 leftmost = rb_erase_cached(&p->pushable_dl_tasks, root);
624 dl_rq->earliest_dl.next = __node_2_pdl(leftmost)->dl.deadline;
626 RB_CLEAR_NODE(&p->pushable_dl_tasks);
628 if (!has_pushable_dl_tasks(rq) && rq->dl.overloaded) {
629 dl_clear_overload(rq);
630 rq->dl.overloaded = 0;
634 static int push_dl_task(struct rq *rq);
636 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
638 return rq->online && dl_task(prev);
641 static DEFINE_PER_CPU(struct balance_callback, dl_push_head);
642 static DEFINE_PER_CPU(struct balance_callback, dl_pull_head);
644 static void push_dl_tasks(struct rq *);
645 static void pull_dl_task(struct rq *);
647 static inline void deadline_queue_push_tasks(struct rq *rq)
649 if (!has_pushable_dl_tasks(rq))
652 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
655 static inline void deadline_queue_pull_task(struct rq *rq)
657 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
660 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
662 static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
664 struct rq *later_rq = NULL;
667 later_rq = find_lock_later_rq(p, rq);
672 * If we cannot preempt any rq, fall back to pick any
675 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
676 if (cpu >= nr_cpu_ids) {
678 * Failed to find any suitable CPU.
679 * The task will never come back!
681 WARN_ON_ONCE(dl_bandwidth_enabled());
684 * If admission control is disabled we
685 * try a little harder to let the task
688 cpu = cpumask_any(cpu_active_mask);
690 later_rq = cpu_rq(cpu);
691 double_lock_balance(rq, later_rq);
694 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
696 * Inactive timer is armed (or callback is running, but
697 * waiting for us to release rq locks). In any case, when it
698 * will fire (or continue), it will see running_bw of this
699 * task migrated to later_rq (and correctly handle it).
701 sub_running_bw(&p->dl, &rq->dl);
702 sub_rq_bw(&p->dl, &rq->dl);
704 add_rq_bw(&p->dl, &later_rq->dl);
705 add_running_bw(&p->dl, &later_rq->dl);
707 sub_rq_bw(&p->dl, &rq->dl);
708 add_rq_bw(&p->dl, &later_rq->dl);
712 * And we finally need to fixup root_domain(s) bandwidth accounting,
713 * since p is still hanging out in the old (now moved to default) root
716 dl_b = &rq->rd->dl_bw;
717 raw_spin_lock(&dl_b->lock);
718 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
719 raw_spin_unlock(&dl_b->lock);
721 dl_b = &later_rq->rd->dl_bw;
722 raw_spin_lock(&dl_b->lock);
723 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
724 raw_spin_unlock(&dl_b->lock);
726 set_task_cpu(p, later_rq->cpu);
727 double_unlock_balance(later_rq, rq);
735 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
740 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
745 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
750 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
754 static inline void deadline_queue_push_tasks(struct rq *rq)
758 static inline void deadline_queue_pull_task(struct rq *rq)
761 #endif /* CONFIG_SMP */
764 enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags);
765 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
766 static void dequeue_dl_entity(struct sched_dl_entity *dl_se, int flags);
767 static void wakeup_preempt_dl(struct rq *rq, struct task_struct *p, int flags);
769 static inline void replenish_dl_new_period(struct sched_dl_entity *dl_se,
772 /* for non-boosted task, pi_of(dl_se) == dl_se */
773 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
774 dl_se->runtime = pi_of(dl_se)->dl_runtime;
778 * We are being explicitly informed that a new instance is starting,
779 * and this means that:
780 * - the absolute deadline of the entity has to be placed at
781 * current time + relative deadline;
782 * - the runtime of the entity has to be set to the maximum value.
784 * The capability of specifying such event is useful whenever a -deadline
785 * entity wants to (try to!) synchronize its behaviour with the scheduler's
786 * one, and to (try to!) reconcile itself with its own scheduling
789 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
791 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
792 struct rq *rq = rq_of_dl_rq(dl_rq);
794 WARN_ON(is_dl_boosted(dl_se));
795 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
798 * We are racing with the deadline timer. So, do nothing because
799 * the deadline timer handler will take care of properly recharging
800 * the runtime and postponing the deadline
802 if (dl_se->dl_throttled)
806 * We use the regular wall clock time to set deadlines in the
807 * future; in fact, we must consider execution overheads (time
808 * spent on hardirq context, etc.).
810 replenish_dl_new_period(dl_se, rq);
814 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
815 * possibility of a entity lasting more than what it declared, and thus
816 * exhausting its runtime.
818 * Here we are interested in making runtime overrun possible, but we do
819 * not want a entity which is misbehaving to affect the scheduling of all
821 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
822 * is used, in order to confine each entity within its own bandwidth.
824 * This function deals exactly with that, and ensures that when the runtime
825 * of a entity is replenished, its deadline is also postponed. That ensures
826 * the overrunning entity can't interfere with other entity in the system and
827 * can't make them miss their deadlines. Reasons why this kind of overruns
828 * could happen are, typically, a entity voluntarily trying to overcome its
829 * runtime, or it just underestimated it during sched_setattr().
831 static void replenish_dl_entity(struct sched_dl_entity *dl_se)
833 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
834 struct rq *rq = rq_of_dl_rq(dl_rq);
836 WARN_ON_ONCE(pi_of(dl_se)->dl_runtime <= 0);
839 * This could be the case for a !-dl task that is boosted.
840 * Just go with full inherited parameters.
842 if (dl_se->dl_deadline == 0)
843 replenish_dl_new_period(dl_se, rq);
845 if (dl_se->dl_yielded && dl_se->runtime > 0)
849 * We keep moving the deadline away until we get some
850 * available runtime for the entity. This ensures correct
851 * handling of situations where the runtime overrun is
854 while (dl_se->runtime <= 0) {
855 dl_se->deadline += pi_of(dl_se)->dl_period;
856 dl_se->runtime += pi_of(dl_se)->dl_runtime;
860 * At this point, the deadline really should be "in
861 * the future" with respect to rq->clock. If it's
862 * not, we are, for some reason, lagging too much!
863 * Anyway, after having warn userspace abut that,
864 * we still try to keep the things running by
865 * resetting the deadline and the budget of the
868 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
869 printk_deferred_once("sched: DL replenish lagged too much\n");
870 replenish_dl_new_period(dl_se, rq);
873 if (dl_se->dl_yielded)
874 dl_se->dl_yielded = 0;
875 if (dl_se->dl_throttled)
876 dl_se->dl_throttled = 0;
880 * Here we check if --at time t-- an entity (which is probably being
881 * [re]activated or, in general, enqueued) can use its remaining runtime
882 * and its current deadline _without_ exceeding the bandwidth it is
883 * assigned (function returns true if it can't). We are in fact applying
884 * one of the CBS rules: when a task wakes up, if the residual runtime
885 * over residual deadline fits within the allocated bandwidth, then we
886 * can keep the current (absolute) deadline and residual budget without
887 * disrupting the schedulability of the system. Otherwise, we should
888 * refill the runtime and set the deadline a period in the future,
889 * because keeping the current (absolute) deadline of the task would
890 * result in breaking guarantees promised to other tasks (refer to
891 * Documentation/scheduler/sched-deadline.rst for more information).
893 * This function returns true if:
895 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
897 * IOW we can't recycle current parameters.
899 * Notice that the bandwidth check is done against the deadline. For
900 * task with deadline equal to period this is the same of using
901 * dl_period instead of dl_deadline in the equation above.
903 static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t)
908 * left and right are the two sides of the equation above,
909 * after a bit of shuffling to use multiplications instead
912 * Note that none of the time values involved in the two
913 * multiplications are absolute: dl_deadline and dl_runtime
914 * are the relative deadline and the maximum runtime of each
915 * instance, runtime is the runtime left for the last instance
916 * and (deadline - t), since t is rq->clock, is the time left
917 * to the (absolute) deadline. Even if overflowing the u64 type
918 * is very unlikely to occur in both cases, here we scale down
919 * as we want to avoid that risk at all. Scaling down by 10
920 * means that we reduce granularity to 1us. We are fine with it,
921 * since this is only a true/false check and, anyway, thinking
922 * of anything below microseconds resolution is actually fiction
923 * (but still we want to give the user that illusion >;).
925 left = (pi_of(dl_se)->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
926 right = ((dl_se->deadline - t) >> DL_SCALE) *
927 (pi_of(dl_se)->dl_runtime >> DL_SCALE);
929 return dl_time_before(right, left);
933 * Revised wakeup rule [1]: For self-suspending tasks, rather then
934 * re-initializing task's runtime and deadline, the revised wakeup
935 * rule adjusts the task's runtime to avoid the task to overrun its
938 * Reasoning: a task may overrun the density if:
939 * runtime / (deadline - t) > dl_runtime / dl_deadline
941 * Therefore, runtime can be adjusted to:
942 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
944 * In such way that runtime will be equal to the maximum density
945 * the task can use without breaking any rule.
947 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
948 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
951 update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
953 u64 laxity = dl_se->deadline - rq_clock(rq);
956 * If the task has deadline < period, and the deadline is in the past,
957 * it should already be throttled before this check.
959 * See update_dl_entity() comments for further details.
961 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
963 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
967 * Regarding the deadline, a task with implicit deadline has a relative
968 * deadline == relative period. A task with constrained deadline has a
969 * relative deadline <= relative period.
971 * We support constrained deadline tasks. However, there are some restrictions
972 * applied only for tasks which do not have an implicit deadline. See
973 * update_dl_entity() to know more about such restrictions.
975 * The dl_is_implicit() returns true if the task has an implicit deadline.
977 static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
979 return dl_se->dl_deadline == dl_se->dl_period;
983 * When a deadline entity is placed in the runqueue, its runtime and deadline
984 * might need to be updated. This is done by a CBS wake up rule. There are two
985 * different rules: 1) the original CBS; and 2) the Revisited CBS.
987 * When the task is starting a new period, the Original CBS is used. In this
988 * case, the runtime is replenished and a new absolute deadline is set.
990 * When a task is queued before the begin of the next period, using the
991 * remaining runtime and deadline could make the entity to overflow, see
992 * dl_entity_overflow() to find more about runtime overflow. When such case
993 * is detected, the runtime and deadline need to be updated.
995 * If the task has an implicit deadline, i.e., deadline == period, the Original
996 * CBS is applied. the runtime is replenished and a new absolute deadline is
997 * set, as in the previous cases.
999 * However, the Original CBS does not work properly for tasks with
1000 * deadline < period, which are said to have a constrained deadline. By
1001 * applying the Original CBS, a constrained deadline task would be able to run
1002 * runtime/deadline in a period. With deadline < period, the task would
1003 * overrun the runtime/period allowed bandwidth, breaking the admission test.
1005 * In order to prevent this misbehave, the Revisited CBS is used for
1006 * constrained deadline tasks when a runtime overflow is detected. In the
1007 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
1008 * the remaining runtime of the task is reduced to avoid runtime overflow.
1009 * Please refer to the comments update_dl_revised_wakeup() function to find
1010 * more about the Revised CBS rule.
1012 static void update_dl_entity(struct sched_dl_entity *dl_se)
1014 struct rq *rq = rq_of_dl_se(dl_se);
1016 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
1017 dl_entity_overflow(dl_se, rq_clock(rq))) {
1019 if (unlikely(!dl_is_implicit(dl_se) &&
1020 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1021 !is_dl_boosted(dl_se))) {
1022 update_dl_revised_wakeup(dl_se, rq);
1026 replenish_dl_new_period(dl_se, rq);
1030 static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
1032 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
1036 * If the entity depleted all its runtime, and if we want it to sleep
1037 * while waiting for some new execution time to become available, we
1038 * set the bandwidth replenishment timer to the replenishment instant
1039 * and try to activate it.
1041 * Notice that it is important for the caller to know if the timer
1042 * actually started or not (i.e., the replenishment instant is in
1043 * the future or in the past).
1045 static int start_dl_timer(struct sched_dl_entity *dl_se)
1047 struct hrtimer *timer = &dl_se->dl_timer;
1048 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1049 struct rq *rq = rq_of_dl_rq(dl_rq);
1053 lockdep_assert_rq_held(rq);
1056 * We want the timer to fire at the deadline, but considering
1057 * that it is actually coming from rq->clock and not from
1058 * hrtimer's time base reading.
1060 act = ns_to_ktime(dl_next_period(dl_se));
1061 now = hrtimer_cb_get_time(timer);
1062 delta = ktime_to_ns(now) - rq_clock(rq);
1063 act = ktime_add_ns(act, delta);
1066 * If the expiry time already passed, e.g., because the value
1067 * chosen as the deadline is too small, don't even try to
1068 * start the timer in the past!
1070 if (ktime_us_delta(act, now) < 0)
1074 * !enqueued will guarantee another callback; even if one is already in
1075 * progress. This ensures a balanced {get,put}_task_struct().
1077 * The race against __run_timer() clearing the enqueued state is
1078 * harmless because we're holding task_rq()->lock, therefore the timer
1079 * expiring after we've done the check will wait on its task_rq_lock()
1080 * and observe our state.
1082 if (!hrtimer_is_queued(timer)) {
1083 if (!dl_server(dl_se))
1084 get_task_struct(dl_task_of(dl_se));
1085 hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD);
1091 static void __push_dl_task(struct rq *rq, struct rq_flags *rf)
1095 * Queueing this task back might have overloaded rq, check if we need
1096 * to kick someone away.
1098 if (has_pushable_dl_tasks(rq)) {
1100 * Nothing relies on rq->lock after this, so its safe to drop
1103 rq_unpin_lock(rq, rf);
1105 rq_repin_lock(rq, rf);
1111 * This is the bandwidth enforcement timer callback. If here, we know
1112 * a task is not on its dl_rq, since the fact that the timer was running
1113 * means the task is throttled and needs a runtime replenishment.
1115 * However, what we actually do depends on the fact the task is active,
1116 * (it is on its rq) or has been removed from there by a call to
1117 * dequeue_task_dl(). In the former case we must issue the runtime
1118 * replenishment and add the task back to the dl_rq; in the latter, we just
1119 * do nothing but clearing dl_throttled, so that runtime and deadline
1120 * updating (and the queueing back to dl_rq) will be done by the
1121 * next call to enqueue_task_dl().
1123 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
1125 struct sched_dl_entity *dl_se = container_of(timer,
1126 struct sched_dl_entity,
1128 struct task_struct *p;
1132 if (dl_server(dl_se)) {
1133 struct rq *rq = rq_of_dl_se(dl_se);
1137 if (dl_se->dl_throttled) {
1139 update_rq_clock(rq);
1141 if (dl_se->server_has_tasks(dl_se)) {
1142 enqueue_dl_entity(dl_se, ENQUEUE_REPLENISH);
1144 __push_dl_task(rq, &rf);
1146 replenish_dl_entity(dl_se);
1152 return HRTIMER_NORESTART;
1155 p = dl_task_of(dl_se);
1156 rq = task_rq_lock(p, &rf);
1159 * The task might have changed its scheduling policy to something
1160 * different than SCHED_DEADLINE (through switched_from_dl()).
1166 * The task might have been boosted by someone else and might be in the
1167 * boosting/deboosting path, its not throttled.
1169 if (is_dl_boosted(dl_se))
1173 * Spurious timer due to start_dl_timer() race; or we already received
1174 * a replenishment from rt_mutex_setprio().
1176 if (!dl_se->dl_throttled)
1180 update_rq_clock(rq);
1183 * If the throttle happened during sched-out; like:
1190 * __dequeue_task_dl()
1193 * We can be both throttled and !queued. Replenish the counter
1194 * but do not enqueue -- wait for our wakeup to do that.
1196 if (!task_on_rq_queued(p)) {
1197 replenish_dl_entity(dl_se);
1202 if (unlikely(!rq->online)) {
1204 * If the runqueue is no longer available, migrate the
1205 * task elsewhere. This necessarily changes rq.
1207 lockdep_unpin_lock(__rq_lockp(rq), rf.cookie);
1208 rq = dl_task_offline_migration(rq, p);
1209 rf.cookie = lockdep_pin_lock(__rq_lockp(rq));
1210 update_rq_clock(rq);
1213 * Now that the task has been migrated to the new RQ and we
1214 * have that locked, proceed as normal and enqueue the task
1220 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1221 if (dl_task(rq->curr))
1222 wakeup_preempt_dl(rq, p, 0);
1226 __push_dl_task(rq, &rf);
1229 task_rq_unlock(rq, p, &rf);
1232 * This can free the task_struct, including this hrtimer, do not touch
1233 * anything related to that after this.
1237 return HRTIMER_NORESTART;
1240 static void init_dl_task_timer(struct sched_dl_entity *dl_se)
1242 struct hrtimer *timer = &dl_se->dl_timer;
1244 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1245 timer->function = dl_task_timer;
1249 * During the activation, CBS checks if it can reuse the current task's
1250 * runtime and period. If the deadline of the task is in the past, CBS
1251 * cannot use the runtime, and so it replenishes the task. This rule
1252 * works fine for implicit deadline tasks (deadline == period), and the
1253 * CBS was designed for implicit deadline tasks. However, a task with
1254 * constrained deadline (deadline < period) might be awakened after the
1255 * deadline, but before the next period. In this case, replenishing the
1256 * task would allow it to run for runtime / deadline. As in this case
1257 * deadline < period, CBS enables a task to run for more than the
1258 * runtime / period. In a very loaded system, this can cause a domino
1259 * effect, making other tasks miss their deadlines.
1261 * To avoid this problem, in the activation of a constrained deadline
1262 * task after the deadline but before the next period, throttle the
1263 * task and set the replenishing timer to the begin of the next period,
1264 * unless it is boosted.
1266 static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1268 struct rq *rq = rq_of_dl_se(dl_se);
1270 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1271 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1272 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(dl_se)))
1274 dl_se->dl_throttled = 1;
1275 if (dl_se->runtime > 0)
1281 int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1283 return (dl_se->runtime <= 0);
1287 * This function implements the GRUB accounting rule. According to the
1288 * GRUB reclaiming algorithm, the runtime is not decreased as "dq = -dt",
1289 * but as "dq = -(max{u, (Umax - Uinact - Uextra)} / Umax) dt",
1290 * where u is the utilization of the task, Umax is the maximum reclaimable
1291 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1292 * as the difference between the "total runqueue utilization" and the
1293 * "runqueue active utilization", and Uextra is the (per runqueue) extra
1294 * reclaimable utilization.
1295 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations multiplied
1296 * by 2^BW_SHIFT, the result has to be shifted right by BW_SHIFT.
1297 * Since rq->dl.bw_ratio contains 1 / Umax multiplied by 2^RATIO_SHIFT, dl_bw
1298 * is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1299 * Since delta is a 64 bit variable, to have an overflow its value should be
1300 * larger than 2^(64 - 20 - 8), which is more than 64 seconds. So, overflow is
1301 * not an issue here.
1303 static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1306 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1309 * Instead of computing max{u, (u_max - u_inact - u_extra)}, we
1310 * compare u_inact + u_extra with u_max - u, because u_inact + u_extra
1311 * can be larger than u_max. So, u_max - u_inact - u_extra would be
1312 * negative leading to wrong results.
1314 if (u_inact + rq->dl.extra_bw > rq->dl.max_bw - dl_se->dl_bw)
1315 u_act = dl_se->dl_bw;
1317 u_act = rq->dl.max_bw - u_inact - rq->dl.extra_bw;
1319 u_act = (u_act * rq->dl.bw_ratio) >> RATIO_SHIFT;
1320 return (delta * u_act) >> BW_SHIFT;
1324 update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1326 static void update_curr_dl_se(struct rq *rq, struct sched_dl_entity *dl_se, s64 delta_exec)
1328 s64 scaled_delta_exec;
1330 if (unlikely(delta_exec <= 0)) {
1331 if (unlikely(dl_se->dl_yielded))
1336 if (dl_entity_is_special(dl_se))
1340 * For tasks that participate in GRUB, we implement GRUB-PA: the
1341 * spare reclaimed bandwidth is used to clock down frequency.
1343 * For the others, we still need to scale reservation parameters
1344 * according to current frequency and CPU maximum capacity.
1346 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1347 scaled_delta_exec = grub_reclaim(delta_exec, rq, dl_se);
1349 int cpu = cpu_of(rq);
1350 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1351 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1353 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1354 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1357 dl_se->runtime -= scaled_delta_exec;
1360 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1361 dl_se->dl_throttled = 1;
1363 /* If requested, inform the user about runtime overruns. */
1364 if (dl_runtime_exceeded(dl_se) &&
1365 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1366 dl_se->dl_overrun = 1;
1368 dequeue_dl_entity(dl_se, 0);
1369 if (!dl_server(dl_se)) {
1370 update_stats_dequeue_dl(&rq->dl, dl_se, 0);
1371 dequeue_pushable_dl_task(rq, dl_task_of(dl_se));
1374 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(dl_se))) {
1375 if (dl_server(dl_se))
1376 enqueue_dl_entity(dl_se, ENQUEUE_REPLENISH);
1378 enqueue_task_dl(rq, dl_task_of(dl_se), ENQUEUE_REPLENISH);
1381 if (!is_leftmost(dl_se, &rq->dl))
1386 * Because -- for now -- we share the rt bandwidth, we need to
1387 * account our runtime there too, otherwise actual rt tasks
1388 * would be able to exceed the shared quota.
1390 * Account to the root rt group for now.
1392 * The solution we're working towards is having the RT groups scheduled
1393 * using deadline servers -- however there's a few nasties to figure
1394 * out before that can happen.
1396 if (rt_bandwidth_enabled()) {
1397 struct rt_rq *rt_rq = &rq->rt;
1399 raw_spin_lock(&rt_rq->rt_runtime_lock);
1401 * We'll let actual RT tasks worry about the overflow here, we
1402 * have our own CBS to keep us inline; only account when RT
1403 * bandwidth is relevant.
1405 if (sched_rt_bandwidth_account(rt_rq))
1406 rt_rq->rt_time += delta_exec;
1407 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1411 void dl_server_update(struct sched_dl_entity *dl_se, s64 delta_exec)
1413 update_curr_dl_se(dl_se->rq, dl_se, delta_exec);
1416 void dl_server_start(struct sched_dl_entity *dl_se)
1418 if (!dl_server(dl_se)) {
1419 dl_se->dl_server = 1;
1420 setup_new_dl_entity(dl_se);
1422 enqueue_dl_entity(dl_se, ENQUEUE_WAKEUP);
1425 void dl_server_stop(struct sched_dl_entity *dl_se)
1427 dequeue_dl_entity(dl_se, DEQUEUE_SLEEP);
1430 void dl_server_init(struct sched_dl_entity *dl_se, struct rq *rq,
1431 dl_server_has_tasks_f has_tasks,
1432 dl_server_pick_f pick)
1435 dl_se->server_has_tasks = has_tasks;
1436 dl_se->server_pick = pick;
1440 * Update the current task's runtime statistics (provided it is still
1441 * a -deadline task and has not been removed from the dl_rq).
1443 static void update_curr_dl(struct rq *rq)
1445 struct task_struct *curr = rq->curr;
1446 struct sched_dl_entity *dl_se = &curr->dl;
1449 if (!dl_task(curr) || !on_dl_rq(dl_se))
1453 * Consumed budget is computed considering the time as
1454 * observed by schedulable tasks (excluding time spent
1455 * in hardirq context, etc.). Deadlines are instead
1456 * computed using hard walltime. This seems to be the more
1457 * natural solution, but the full ramifications of this
1458 * approach need further study.
1460 delta_exec = update_curr_common(rq);
1461 update_curr_dl_se(rq, dl_se, delta_exec);
1464 static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1466 struct sched_dl_entity *dl_se = container_of(timer,
1467 struct sched_dl_entity,
1469 struct task_struct *p = NULL;
1473 if (!dl_server(dl_se)) {
1474 p = dl_task_of(dl_se);
1475 rq = task_rq_lock(p, &rf);
1482 update_rq_clock(rq);
1484 if (dl_server(dl_se))
1487 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
1488 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1490 if (READ_ONCE(p->__state) == TASK_DEAD && dl_se->dl_non_contending) {
1491 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1492 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1493 dl_se->dl_non_contending = 0;
1496 raw_spin_lock(&dl_b->lock);
1497 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1498 raw_spin_unlock(&dl_b->lock);
1499 __dl_clear_params(dl_se);
1505 if (dl_se->dl_non_contending == 0)
1508 sub_running_bw(dl_se, &rq->dl);
1509 dl_se->dl_non_contending = 0;
1512 if (!dl_server(dl_se)) {
1513 task_rq_unlock(rq, p, &rf);
1519 return HRTIMER_NORESTART;
1522 static void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1524 struct hrtimer *timer = &dl_se->inactive_timer;
1526 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1527 timer->function = inactive_task_timer;
1530 #define __node_2_dle(node) \
1531 rb_entry((node), struct sched_dl_entity, rb_node)
1535 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1537 struct rq *rq = rq_of_dl_rq(dl_rq);
1539 if (dl_rq->earliest_dl.curr == 0 ||
1540 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1541 if (dl_rq->earliest_dl.curr == 0)
1542 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_HIGHER);
1543 dl_rq->earliest_dl.curr = deadline;
1544 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1548 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1550 struct rq *rq = rq_of_dl_rq(dl_rq);
1553 * Since we may have removed our earliest (and/or next earliest)
1554 * task we must recompute them.
1556 if (!dl_rq->dl_nr_running) {
1557 dl_rq->earliest_dl.curr = 0;
1558 dl_rq->earliest_dl.next = 0;
1559 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1560 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1562 struct rb_node *leftmost = rb_first_cached(&dl_rq->root);
1563 struct sched_dl_entity *entry = __node_2_dle(leftmost);
1565 dl_rq->earliest_dl.curr = entry->deadline;
1566 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1572 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1573 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1575 #endif /* CONFIG_SMP */
1578 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1580 u64 deadline = dl_se->deadline;
1582 dl_rq->dl_nr_running++;
1583 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1585 inc_dl_deadline(dl_rq, deadline);
1589 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1591 WARN_ON(!dl_rq->dl_nr_running);
1592 dl_rq->dl_nr_running--;
1593 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1595 dec_dl_deadline(dl_rq, dl_se->deadline);
1598 static inline bool __dl_less(struct rb_node *a, const struct rb_node *b)
1600 return dl_time_before(__node_2_dle(a)->deadline, __node_2_dle(b)->deadline);
1603 static inline struct sched_statistics *
1604 __schedstats_from_dl_se(struct sched_dl_entity *dl_se)
1606 return &dl_task_of(dl_se)->stats;
1610 update_stats_wait_start_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1612 struct sched_statistics *stats;
1614 if (!schedstat_enabled())
1617 stats = __schedstats_from_dl_se(dl_se);
1618 __update_stats_wait_start(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1622 update_stats_wait_end_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1624 struct sched_statistics *stats;
1626 if (!schedstat_enabled())
1629 stats = __schedstats_from_dl_se(dl_se);
1630 __update_stats_wait_end(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1634 update_stats_enqueue_sleeper_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1636 struct sched_statistics *stats;
1638 if (!schedstat_enabled())
1641 stats = __schedstats_from_dl_se(dl_se);
1642 __update_stats_enqueue_sleeper(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1646 update_stats_enqueue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1649 if (!schedstat_enabled())
1652 if (flags & ENQUEUE_WAKEUP)
1653 update_stats_enqueue_sleeper_dl(dl_rq, dl_se);
1657 update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1660 struct task_struct *p = dl_task_of(dl_se);
1662 if (!schedstat_enabled())
1665 if ((flags & DEQUEUE_SLEEP)) {
1668 state = READ_ONCE(p->__state);
1669 if (state & TASK_INTERRUPTIBLE)
1670 __schedstat_set(p->stats.sleep_start,
1671 rq_clock(rq_of_dl_rq(dl_rq)));
1673 if (state & TASK_UNINTERRUPTIBLE)
1674 __schedstat_set(p->stats.block_start,
1675 rq_clock(rq_of_dl_rq(dl_rq)));
1679 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1681 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1683 WARN_ON_ONCE(!RB_EMPTY_NODE(&dl_se->rb_node));
1685 rb_add_cached(&dl_se->rb_node, &dl_rq->root, __dl_less);
1687 inc_dl_tasks(dl_se, dl_rq);
1690 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1692 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1694 if (RB_EMPTY_NODE(&dl_se->rb_node))
1697 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1699 RB_CLEAR_NODE(&dl_se->rb_node);
1701 dec_dl_tasks(dl_se, dl_rq);
1705 enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags)
1707 WARN_ON_ONCE(on_dl_rq(dl_se));
1709 update_stats_enqueue_dl(dl_rq_of_se(dl_se), dl_se, flags);
1712 * Check if a constrained deadline task was activated
1713 * after the deadline but before the next period.
1714 * If that is the case, the task will be throttled and
1715 * the replenishment timer will be set to the next period.
1717 if (!dl_se->dl_throttled && !dl_is_implicit(dl_se))
1718 dl_check_constrained_dl(dl_se);
1720 if (flags & (ENQUEUE_RESTORE|ENQUEUE_MIGRATING)) {
1721 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1723 add_rq_bw(dl_se, dl_rq);
1724 add_running_bw(dl_se, dl_rq);
1728 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1729 * its budget it needs a replenishment and, since it now is on
1730 * its rq, the bandwidth timer callback (which clearly has not
1731 * run yet) will take care of this.
1732 * However, the active utilization does not depend on the fact
1733 * that the task is on the runqueue or not (but depends on the
1734 * task's state - in GRUB parlance, "inactive" vs "active contending").
1735 * In other words, even if a task is throttled its utilization must
1736 * be counted in the active utilization; hence, we need to call
1739 if (dl_se->dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1740 if (flags & ENQUEUE_WAKEUP)
1741 task_contending(dl_se, flags);
1747 * If this is a wakeup or a new instance, the scheduling
1748 * parameters of the task might need updating. Otherwise,
1749 * we want a replenishment of its runtime.
1751 if (flags & ENQUEUE_WAKEUP) {
1752 task_contending(dl_se, flags);
1753 update_dl_entity(dl_se);
1754 } else if (flags & ENQUEUE_REPLENISH) {
1755 replenish_dl_entity(dl_se);
1756 } else if ((flags & ENQUEUE_RESTORE) &&
1757 dl_time_before(dl_se->deadline, rq_clock(rq_of_dl_se(dl_se)))) {
1758 setup_new_dl_entity(dl_se);
1761 __enqueue_dl_entity(dl_se);
1764 static void dequeue_dl_entity(struct sched_dl_entity *dl_se, int flags)
1766 __dequeue_dl_entity(dl_se);
1768 if (flags & (DEQUEUE_SAVE|DEQUEUE_MIGRATING)) {
1769 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1771 sub_running_bw(dl_se, dl_rq);
1772 sub_rq_bw(dl_se, dl_rq);
1776 * This check allows to start the inactive timer (or to immediately
1777 * decrease the active utilization, if needed) in two cases:
1778 * when the task blocks and when it is terminating
1779 * (p->state == TASK_DEAD). We can handle the two cases in the same
1780 * way, because from GRUB's point of view the same thing is happening
1781 * (the task moves from "active contending" to "active non contending"
1784 if (flags & DEQUEUE_SLEEP)
1785 task_non_contending(dl_se);
1788 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1790 if (is_dl_boosted(&p->dl)) {
1792 * Because of delays in the detection of the overrun of a
1793 * thread's runtime, it might be the case that a thread
1794 * goes to sleep in a rt mutex with negative runtime. As
1795 * a consequence, the thread will be throttled.
1797 * While waiting for the mutex, this thread can also be
1798 * boosted via PI, resulting in a thread that is throttled
1799 * and boosted at the same time.
1801 * In this case, the boost overrides the throttle.
1803 if (p->dl.dl_throttled) {
1805 * The replenish timer needs to be canceled. No
1806 * problem if it fires concurrently: boosted threads
1807 * are ignored in dl_task_timer().
1809 hrtimer_try_to_cancel(&p->dl.dl_timer);
1810 p->dl.dl_throttled = 0;
1812 } else if (!dl_prio(p->normal_prio)) {
1814 * Special case in which we have a !SCHED_DEADLINE task that is going
1815 * to be deboosted, but exceeds its runtime while doing so. No point in
1816 * replenishing it, as it's going to return back to its original
1817 * scheduling class after this. If it has been throttled, we need to
1818 * clear the flag, otherwise the task may wake up as throttled after
1819 * being boosted again with no means to replenish the runtime and clear
1822 p->dl.dl_throttled = 0;
1823 if (!(flags & ENQUEUE_REPLENISH))
1824 printk_deferred_once("sched: DL de-boosted task PID %d: REPLENISH flag missing\n",
1830 check_schedstat_required();
1831 update_stats_wait_start_dl(dl_rq_of_se(&p->dl), &p->dl);
1833 if (p->on_rq == TASK_ON_RQ_MIGRATING)
1834 flags |= ENQUEUE_MIGRATING;
1836 enqueue_dl_entity(&p->dl, flags);
1838 if (dl_server(&p->dl))
1841 if (!task_current(rq, p) && !p->dl.dl_throttled && p->nr_cpus_allowed > 1)
1842 enqueue_pushable_dl_task(rq, p);
1845 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1849 if (p->on_rq == TASK_ON_RQ_MIGRATING)
1850 flags |= DEQUEUE_MIGRATING;
1852 dequeue_dl_entity(&p->dl, flags);
1853 if (!p->dl.dl_throttled && !dl_server(&p->dl))
1854 dequeue_pushable_dl_task(rq, p);
1858 * Yield task semantic for -deadline tasks is:
1860 * get off from the CPU until our next instance, with
1861 * a new runtime. This is of little use now, since we
1862 * don't have a bandwidth reclaiming mechanism. Anyway,
1863 * bandwidth reclaiming is planned for the future, and
1864 * yield_task_dl will indicate that some spare budget
1865 * is available for other task instances to use it.
1867 static void yield_task_dl(struct rq *rq)
1870 * We make the task go to sleep until its current deadline by
1871 * forcing its runtime to zero. This way, update_curr_dl() stops
1872 * it and the bandwidth timer will wake it up and will give it
1873 * new scheduling parameters (thanks to dl_yielded=1).
1875 rq->curr->dl.dl_yielded = 1;
1877 update_rq_clock(rq);
1880 * Tell update_rq_clock() that we've just updated,
1881 * so we don't do microscopic update in schedule()
1882 * and double the fastpath cost.
1884 rq_clock_skip_update(rq);
1889 static inline bool dl_task_is_earliest_deadline(struct task_struct *p,
1892 return (!rq->dl.dl_nr_running ||
1893 dl_time_before(p->dl.deadline,
1894 rq->dl.earliest_dl.curr));
1897 static int find_later_rq(struct task_struct *task);
1900 select_task_rq_dl(struct task_struct *p, int cpu, int flags)
1902 struct task_struct *curr;
1906 if (!(flags & WF_TTWU))
1912 curr = READ_ONCE(rq->curr); /* unlocked access */
1915 * If we are dealing with a -deadline task, we must
1916 * decide where to wake it up.
1917 * If it has a later deadline and the current task
1918 * on this rq can't move (provided the waking task
1919 * can!) we prefer to send it somewhere else. On the
1920 * other hand, if it has a shorter deadline, we
1921 * try to make it stay here, it might be important.
1923 select_rq = unlikely(dl_task(curr)) &&
1924 (curr->nr_cpus_allowed < 2 ||
1925 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1926 p->nr_cpus_allowed > 1;
1929 * Take the capacity of the CPU into account to
1930 * ensure it fits the requirement of the task.
1932 if (sched_asym_cpucap_active())
1933 select_rq |= !dl_task_fits_capacity(p, cpu);
1936 int target = find_later_rq(p);
1939 dl_task_is_earliest_deadline(p, cpu_rq(target)))
1948 static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
1953 if (READ_ONCE(p->__state) != TASK_WAKING)
1958 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1959 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1960 * rq->lock is not... So, lock it
1963 if (p->dl.dl_non_contending) {
1964 update_rq_clock(rq);
1965 sub_running_bw(&p->dl, &rq->dl);
1966 p->dl.dl_non_contending = 0;
1968 * If the timer handler is currently running and the
1969 * timer cannot be canceled, inactive_task_timer()
1970 * will see that dl_not_contending is not set, and
1971 * will not touch the rq's active utilization,
1972 * so we are still safe.
1974 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
1977 sub_rq_bw(&p->dl, &rq->dl);
1981 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1984 * Current can't be migrated, useless to reschedule,
1985 * let's hope p can move out.
1987 if (rq->curr->nr_cpus_allowed == 1 ||
1988 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
1992 * p is migratable, so let's not schedule it and
1993 * see if it is pushed or pulled somewhere else.
1995 if (p->nr_cpus_allowed != 1 &&
1996 cpudl_find(&rq->rd->cpudl, p, NULL))
2002 static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
2004 if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) {
2006 * This is OK, because current is on_cpu, which avoids it being
2007 * picked for load-balance and preemption/IRQs are still
2008 * disabled avoiding further scheduler activity on it and we've
2009 * not yet started the picking loop.
2011 rq_unpin_lock(rq, rf);
2013 rq_repin_lock(rq, rf);
2016 return sched_stop_runnable(rq) || sched_dl_runnable(rq);
2018 #endif /* CONFIG_SMP */
2021 * Only called when both the current and waking task are -deadline
2024 static void wakeup_preempt_dl(struct rq *rq, struct task_struct *p,
2027 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
2034 * In the unlikely case current and p have the same deadline
2035 * let us try to decide what's the best thing to do...
2037 if ((p->dl.deadline == rq->curr->dl.deadline) &&
2038 !test_tsk_need_resched(rq->curr))
2039 check_preempt_equal_dl(rq, p);
2040 #endif /* CONFIG_SMP */
2043 #ifdef CONFIG_SCHED_HRTICK
2044 static void start_hrtick_dl(struct rq *rq, struct sched_dl_entity *dl_se)
2046 hrtick_start(rq, dl_se->runtime);
2048 #else /* !CONFIG_SCHED_HRTICK */
2049 static void start_hrtick_dl(struct rq *rq, struct sched_dl_entity *dl_se)
2054 static void set_next_task_dl(struct rq *rq, struct task_struct *p, bool first)
2056 struct sched_dl_entity *dl_se = &p->dl;
2057 struct dl_rq *dl_rq = &rq->dl;
2059 p->se.exec_start = rq_clock_task(rq);
2060 if (on_dl_rq(&p->dl))
2061 update_stats_wait_end_dl(dl_rq, dl_se);
2063 /* You can't push away the running task */
2064 dequeue_pushable_dl_task(rq, p);
2069 if (rq->curr->sched_class != &dl_sched_class)
2070 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2072 deadline_queue_push_tasks(rq);
2075 static struct sched_dl_entity *pick_next_dl_entity(struct dl_rq *dl_rq)
2077 struct rb_node *left = rb_first_cached(&dl_rq->root);
2082 return __node_2_dle(left);
2085 static struct task_struct *pick_task_dl(struct rq *rq)
2087 struct sched_dl_entity *dl_se;
2088 struct dl_rq *dl_rq = &rq->dl;
2089 struct task_struct *p;
2092 if (!sched_dl_runnable(rq))
2095 dl_se = pick_next_dl_entity(dl_rq);
2096 WARN_ON_ONCE(!dl_se);
2098 if (dl_server(dl_se)) {
2099 p = dl_se->server_pick(dl_se);
2102 dl_se->dl_yielded = 1;
2103 update_curr_dl_se(rq, dl_se, 0);
2106 p->dl_server = dl_se;
2108 p = dl_task_of(dl_se);
2114 static struct task_struct *pick_next_task_dl(struct rq *rq)
2116 struct task_struct *p;
2118 p = pick_task_dl(rq);
2123 set_next_task_dl(rq, p, true);
2125 if (hrtick_enabled(rq))
2126 start_hrtick_dl(rq, &p->dl);
2131 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
2133 struct sched_dl_entity *dl_se = &p->dl;
2134 struct dl_rq *dl_rq = &rq->dl;
2136 if (on_dl_rq(&p->dl))
2137 update_stats_wait_start_dl(dl_rq, dl_se);
2141 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2142 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
2143 enqueue_pushable_dl_task(rq, p);
2147 * scheduler tick hitting a task of our scheduling class.
2149 * NOTE: This function can be called remotely by the tick offload that
2150 * goes along full dynticks. Therefore no local assumption can be made
2151 * and everything must be accessed through the @rq and @curr passed in
2154 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
2158 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2160 * Even when we have runtime, update_curr_dl() might have resulted in us
2161 * not being the leftmost task anymore. In that case NEED_RESCHED will
2162 * be set and schedule() will start a new hrtick for the next task.
2164 if (hrtick_enabled_dl(rq) && queued && p->dl.runtime > 0 &&
2165 is_leftmost(&p->dl, &rq->dl))
2166 start_hrtick_dl(rq, &p->dl);
2169 static void task_fork_dl(struct task_struct *p)
2172 * SCHED_DEADLINE tasks cannot fork and this is achieved through
2179 /* Only try algorithms three times */
2180 #define DL_MAX_TRIES 3
2182 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
2184 if (!task_on_cpu(rq, p) &&
2185 cpumask_test_cpu(cpu, &p->cpus_mask))
2191 * Return the earliest pushable rq's task, which is suitable to be executed
2192 * on the CPU, NULL otherwise:
2194 static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
2196 struct task_struct *p = NULL;
2197 struct rb_node *next_node;
2199 if (!has_pushable_dl_tasks(rq))
2202 next_node = rb_first_cached(&rq->dl.pushable_dl_tasks_root);
2206 p = __node_2_pdl(next_node);
2208 if (pick_dl_task(rq, p, cpu))
2211 next_node = rb_next(next_node);
2218 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
2220 static int find_later_rq(struct task_struct *task)
2222 struct sched_domain *sd;
2223 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
2224 int this_cpu = smp_processor_id();
2225 int cpu = task_cpu(task);
2227 /* Make sure the mask is initialized first */
2228 if (unlikely(!later_mask))
2231 if (task->nr_cpus_allowed == 1)
2235 * We have to consider system topology and task affinity
2236 * first, then we can look for a suitable CPU.
2238 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
2242 * If we are here, some targets have been found, including
2243 * the most suitable which is, among the runqueues where the
2244 * current tasks have later deadlines than the task's one, the
2245 * rq with the latest possible one.
2247 * Now we check how well this matches with task's
2248 * affinity and system topology.
2250 * The last CPU where the task run is our first
2251 * guess, since it is most likely cache-hot there.
2253 if (cpumask_test_cpu(cpu, later_mask))
2256 * Check if this_cpu is to be skipped (i.e., it is
2257 * not in the mask) or not.
2259 if (!cpumask_test_cpu(this_cpu, later_mask))
2263 for_each_domain(cpu, sd) {
2264 if (sd->flags & SD_WAKE_AFFINE) {
2268 * If possible, preempting this_cpu is
2269 * cheaper than migrating.
2271 if (this_cpu != -1 &&
2272 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
2277 best_cpu = cpumask_any_and_distribute(later_mask,
2278 sched_domain_span(sd));
2280 * Last chance: if a CPU being in both later_mask
2281 * and current sd span is valid, that becomes our
2282 * choice. Of course, the latest possible CPU is
2283 * already under consideration through later_mask.
2285 if (best_cpu < nr_cpu_ids) {
2294 * At this point, all our guesses failed, we just return
2295 * 'something', and let the caller sort the things out.
2300 cpu = cpumask_any_distribute(later_mask);
2301 if (cpu < nr_cpu_ids)
2307 /* Locks the rq it finds */
2308 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
2310 struct rq *later_rq = NULL;
2314 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
2315 cpu = find_later_rq(task);
2317 if ((cpu == -1) || (cpu == rq->cpu))
2320 later_rq = cpu_rq(cpu);
2322 if (!dl_task_is_earliest_deadline(task, later_rq)) {
2324 * Target rq has tasks of equal or earlier deadline,
2325 * retrying does not release any lock and is unlikely
2326 * to yield a different result.
2332 /* Retry if something changed. */
2333 if (double_lock_balance(rq, later_rq)) {
2334 if (unlikely(task_rq(task) != rq ||
2335 !cpumask_test_cpu(later_rq->cpu, &task->cpus_mask) ||
2336 task_on_cpu(rq, task) ||
2338 is_migration_disabled(task) ||
2339 !task_on_rq_queued(task))) {
2340 double_unlock_balance(rq, later_rq);
2347 * If the rq we found has no -deadline task, or
2348 * its earliest one has a later deadline than our
2349 * task, the rq is a good one.
2351 if (dl_task_is_earliest_deadline(task, later_rq))
2354 /* Otherwise we try again. */
2355 double_unlock_balance(rq, later_rq);
2362 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2364 struct task_struct *p;
2366 if (!has_pushable_dl_tasks(rq))
2369 p = __node_2_pdl(rb_first_cached(&rq->dl.pushable_dl_tasks_root));
2371 WARN_ON_ONCE(rq->cpu != task_cpu(p));
2372 WARN_ON_ONCE(task_current(rq, p));
2373 WARN_ON_ONCE(p->nr_cpus_allowed <= 1);
2375 WARN_ON_ONCE(!task_on_rq_queued(p));
2376 WARN_ON_ONCE(!dl_task(p));
2382 * See if the non running -deadline tasks on this rq
2383 * can be sent to some other CPU where they can preempt
2384 * and start executing.
2386 static int push_dl_task(struct rq *rq)
2388 struct task_struct *next_task;
2389 struct rq *later_rq;
2392 next_task = pick_next_pushable_dl_task(rq);
2398 * If next_task preempts rq->curr, and rq->curr
2399 * can move away, it makes sense to just reschedule
2400 * without going further in pushing next_task.
2402 if (dl_task(rq->curr) &&
2403 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2404 rq->curr->nr_cpus_allowed > 1) {
2409 if (is_migration_disabled(next_task))
2412 if (WARN_ON(next_task == rq->curr))
2415 /* We might release rq lock */
2416 get_task_struct(next_task);
2418 /* Will lock the rq it'll find */
2419 later_rq = find_lock_later_rq(next_task, rq);
2421 struct task_struct *task;
2424 * We must check all this again, since
2425 * find_lock_later_rq releases rq->lock and it is
2426 * then possible that next_task has migrated.
2428 task = pick_next_pushable_dl_task(rq);
2429 if (task == next_task) {
2431 * The task is still there. We don't try
2432 * again, some other CPU will pull it when ready.
2441 put_task_struct(next_task);
2446 deactivate_task(rq, next_task, 0);
2447 set_task_cpu(next_task, later_rq->cpu);
2448 activate_task(later_rq, next_task, 0);
2451 resched_curr(later_rq);
2453 double_unlock_balance(rq, later_rq);
2456 put_task_struct(next_task);
2461 static void push_dl_tasks(struct rq *rq)
2463 /* push_dl_task() will return true if it moved a -deadline task */
2464 while (push_dl_task(rq))
2468 static void pull_dl_task(struct rq *this_rq)
2470 int this_cpu = this_rq->cpu, cpu;
2471 struct task_struct *p, *push_task;
2472 bool resched = false;
2474 u64 dmin = LONG_MAX;
2476 if (likely(!dl_overloaded(this_rq)))
2480 * Match the barrier from dl_set_overloaded; this guarantees that if we
2481 * see overloaded we must also see the dlo_mask bit.
2485 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2486 if (this_cpu == cpu)
2489 src_rq = cpu_rq(cpu);
2492 * It looks racy, abd it is! However, as in sched_rt.c,
2493 * we are fine with this.
2495 if (this_rq->dl.dl_nr_running &&
2496 dl_time_before(this_rq->dl.earliest_dl.curr,
2497 src_rq->dl.earliest_dl.next))
2500 /* Might drop this_rq->lock */
2502 double_lock_balance(this_rq, src_rq);
2505 * If there are no more pullable tasks on the
2506 * rq, we're done with it.
2508 if (src_rq->dl.dl_nr_running <= 1)
2511 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2514 * We found a task to be pulled if:
2515 * - it preempts our current (if there's one),
2516 * - it will preempt the last one we pulled (if any).
2518 if (p && dl_time_before(p->dl.deadline, dmin) &&
2519 dl_task_is_earliest_deadline(p, this_rq)) {
2520 WARN_ON(p == src_rq->curr);
2521 WARN_ON(!task_on_rq_queued(p));
2524 * Then we pull iff p has actually an earlier
2525 * deadline than the current task of its runqueue.
2527 if (dl_time_before(p->dl.deadline,
2528 src_rq->curr->dl.deadline))
2531 if (is_migration_disabled(p)) {
2532 push_task = get_push_task(src_rq);
2534 deactivate_task(src_rq, p, 0);
2535 set_task_cpu(p, this_cpu);
2536 activate_task(this_rq, p, 0);
2537 dmin = p->dl.deadline;
2541 /* Is there any other task even earlier? */
2544 double_unlock_balance(this_rq, src_rq);
2548 raw_spin_rq_unlock(this_rq);
2549 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2550 push_task, &src_rq->push_work);
2552 raw_spin_rq_lock(this_rq);
2557 resched_curr(this_rq);
2561 * Since the task is not running and a reschedule is not going to happen
2562 * anytime soon on its runqueue, we try pushing it away now.
2564 static void task_woken_dl(struct rq *rq, struct task_struct *p)
2566 if (!task_on_cpu(rq, p) &&
2567 !test_tsk_need_resched(rq->curr) &&
2568 p->nr_cpus_allowed > 1 &&
2569 dl_task(rq->curr) &&
2570 (rq->curr->nr_cpus_allowed < 2 ||
2571 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2576 static void set_cpus_allowed_dl(struct task_struct *p,
2577 struct affinity_context *ctx)
2579 struct root_domain *src_rd;
2582 WARN_ON_ONCE(!dl_task(p));
2587 * Migrating a SCHED_DEADLINE task between exclusive
2588 * cpusets (different root_domains) entails a bandwidth
2589 * update. We already made space for us in the destination
2590 * domain (see cpuset_can_attach()).
2592 if (!cpumask_intersects(src_rd->span, ctx->new_mask)) {
2593 struct dl_bw *src_dl_b;
2595 src_dl_b = dl_bw_of(cpu_of(rq));
2597 * We now free resources of the root_domain we are migrating
2598 * off. In the worst case, sched_setattr() may temporary fail
2599 * until we complete the update.
2601 raw_spin_lock(&src_dl_b->lock);
2602 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2603 raw_spin_unlock(&src_dl_b->lock);
2606 set_cpus_allowed_common(p, ctx);
2609 /* Assumes rq->lock is held */
2610 static void rq_online_dl(struct rq *rq)
2612 if (rq->dl.overloaded)
2613 dl_set_overload(rq);
2615 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2616 if (rq->dl.dl_nr_running > 0)
2617 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2620 /* Assumes rq->lock is held */
2621 static void rq_offline_dl(struct rq *rq)
2623 if (rq->dl.overloaded)
2624 dl_clear_overload(rq);
2626 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2627 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2630 void __init init_sched_dl_class(void)
2634 for_each_possible_cpu(i)
2635 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2636 GFP_KERNEL, cpu_to_node(i));
2639 void dl_add_task_root_domain(struct task_struct *p)
2645 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2647 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
2651 rq = __task_rq_lock(p, &rf);
2653 dl_b = &rq->rd->dl_bw;
2654 raw_spin_lock(&dl_b->lock);
2656 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2658 raw_spin_unlock(&dl_b->lock);
2660 task_rq_unlock(rq, p, &rf);
2663 void dl_clear_root_domain(struct root_domain *rd)
2665 unsigned long flags;
2667 raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
2668 rd->dl_bw.total_bw = 0;
2669 raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
2672 #endif /* CONFIG_SMP */
2674 static void switched_from_dl(struct rq *rq, struct task_struct *p)
2677 * task_non_contending() can start the "inactive timer" (if the 0-lag
2678 * time is in the future). If the task switches back to dl before
2679 * the "inactive timer" fires, it can continue to consume its current
2680 * runtime using its current deadline. If it stays outside of
2681 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2682 * will reset the task parameters.
2684 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2685 task_non_contending(&p->dl);
2688 * In case a task is setscheduled out from SCHED_DEADLINE we need to
2689 * keep track of that on its cpuset (for correct bandwidth tracking).
2693 if (!task_on_rq_queued(p)) {
2695 * Inactive timer is armed. However, p is leaving DEADLINE and
2696 * might migrate away from this rq while continuing to run on
2697 * some other class. We need to remove its contribution from
2698 * this rq running_bw now, or sub_rq_bw (below) will complain.
2700 if (p->dl.dl_non_contending)
2701 sub_running_bw(&p->dl, &rq->dl);
2702 sub_rq_bw(&p->dl, &rq->dl);
2706 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2707 * at the 0-lag time, because the task could have been migrated
2708 * while SCHED_OTHER in the meanwhile.
2710 if (p->dl.dl_non_contending)
2711 p->dl.dl_non_contending = 0;
2714 * Since this might be the only -deadline task on the rq,
2715 * this is the right place to try to pull some other one
2716 * from an overloaded CPU, if any.
2718 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
2721 deadline_queue_pull_task(rq);
2725 * When switching to -deadline, we may overload the rq, then
2726 * we try to push someone off, if possible.
2728 static void switched_to_dl(struct rq *rq, struct task_struct *p)
2730 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2734 * In case a task is setscheduled to SCHED_DEADLINE we need to keep
2735 * track of that on its cpuset (for correct bandwidth tracking).
2739 /* If p is not queued we will update its parameters at next wakeup. */
2740 if (!task_on_rq_queued(p)) {
2741 add_rq_bw(&p->dl, &rq->dl);
2746 if (rq->curr != p) {
2748 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
2749 deadline_queue_push_tasks(rq);
2751 if (dl_task(rq->curr))
2752 wakeup_preempt_dl(rq, p, 0);
2756 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2761 * If the scheduling parameters of a -deadline task changed,
2762 * a push or pull operation might be needed.
2764 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2767 if (!task_on_rq_queued(p))
2772 * This might be too much, but unfortunately
2773 * we don't have the old deadline value, and
2774 * we can't argue if the task is increasing
2775 * or lowering its prio, so...
2777 if (!rq->dl.overloaded)
2778 deadline_queue_pull_task(rq);
2780 if (task_current(rq, p)) {
2782 * If we now have a earlier deadline task than p,
2783 * then reschedule, provided p is still on this
2786 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
2790 * Current may not be deadline in case p was throttled but we
2791 * have just replenished it (e.g. rt_mutex_setprio()).
2793 * Otherwise, if p was given an earlier deadline, reschedule.
2795 if (!dl_task(rq->curr) ||
2796 dl_time_before(p->dl.deadline, rq->curr->dl.deadline))
2801 * We don't know if p has a earlier or later deadline, so let's blindly
2802 * set a (maybe not needed) rescheduling point.
2808 #ifdef CONFIG_SCHED_CORE
2809 static int task_is_throttled_dl(struct task_struct *p, int cpu)
2811 return p->dl.dl_throttled;
2815 DEFINE_SCHED_CLASS(dl) = {
2817 .enqueue_task = enqueue_task_dl,
2818 .dequeue_task = dequeue_task_dl,
2819 .yield_task = yield_task_dl,
2821 .wakeup_preempt = wakeup_preempt_dl,
2823 .pick_next_task = pick_next_task_dl,
2824 .put_prev_task = put_prev_task_dl,
2825 .set_next_task = set_next_task_dl,
2828 .balance = balance_dl,
2829 .pick_task = pick_task_dl,
2830 .select_task_rq = select_task_rq_dl,
2831 .migrate_task_rq = migrate_task_rq_dl,
2832 .set_cpus_allowed = set_cpus_allowed_dl,
2833 .rq_online = rq_online_dl,
2834 .rq_offline = rq_offline_dl,
2835 .task_woken = task_woken_dl,
2836 .find_lock_rq = find_lock_later_rq,
2839 .task_tick = task_tick_dl,
2840 .task_fork = task_fork_dl,
2842 .prio_changed = prio_changed_dl,
2843 .switched_from = switched_from_dl,
2844 .switched_to = switched_to_dl,
2846 .update_curr = update_curr_dl,
2847 #ifdef CONFIG_SCHED_CORE
2848 .task_is_throttled = task_is_throttled_dl,
2852 /* Used for dl_bw check and update, used under sched_rt_handler()::mutex */
2853 static u64 dl_generation;
2855 int sched_dl_global_validate(void)
2857 u64 runtime = global_rt_runtime();
2858 u64 period = global_rt_period();
2859 u64 new_bw = to_ratio(period, runtime);
2860 u64 gen = ++dl_generation;
2862 int cpu, cpus, ret = 0;
2863 unsigned long flags;
2866 * Here we want to check the bandwidth not being set to some
2867 * value smaller than the currently allocated bandwidth in
2868 * any of the root_domains.
2870 for_each_possible_cpu(cpu) {
2871 rcu_read_lock_sched();
2873 if (dl_bw_visited(cpu, gen))
2876 dl_b = dl_bw_of(cpu);
2877 cpus = dl_bw_cpus(cpu);
2879 raw_spin_lock_irqsave(&dl_b->lock, flags);
2880 if (new_bw * cpus < dl_b->total_bw)
2882 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2885 rcu_read_unlock_sched();
2894 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
2896 if (global_rt_runtime() == RUNTIME_INF) {
2897 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2898 dl_rq->max_bw = dl_rq->extra_bw = 1 << BW_SHIFT;
2900 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2901 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2902 dl_rq->max_bw = dl_rq->extra_bw =
2903 to_ratio(global_rt_period(), global_rt_runtime());
2907 void sched_dl_do_global(void)
2910 u64 gen = ++dl_generation;
2913 unsigned long flags;
2915 if (global_rt_runtime() != RUNTIME_INF)
2916 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2918 for_each_possible_cpu(cpu) {
2919 rcu_read_lock_sched();
2921 if (dl_bw_visited(cpu, gen)) {
2922 rcu_read_unlock_sched();
2926 dl_b = dl_bw_of(cpu);
2928 raw_spin_lock_irqsave(&dl_b->lock, flags);
2930 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2932 rcu_read_unlock_sched();
2933 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2938 * We must be sure that accepting a new task (or allowing changing the
2939 * parameters of an existing one) is consistent with the bandwidth
2940 * constraints. If yes, this function also accordingly updates the currently
2941 * allocated bandwidth to reflect the new situation.
2943 * This function is called while holding p's rq->lock.
2945 int sched_dl_overflow(struct task_struct *p, int policy,
2946 const struct sched_attr *attr)
2948 u64 period = attr->sched_period ?: attr->sched_deadline;
2949 u64 runtime = attr->sched_runtime;
2950 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2951 int cpus, err = -1, cpu = task_cpu(p);
2952 struct dl_bw *dl_b = dl_bw_of(cpu);
2955 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2958 /* !deadline task may carry old deadline bandwidth */
2959 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2963 * Either if a task, enters, leave, or stays -deadline but changes
2964 * its parameters, we may need to update accordingly the total
2965 * allocated bandwidth of the container.
2967 raw_spin_lock(&dl_b->lock);
2968 cpus = dl_bw_cpus(cpu);
2969 cap = dl_bw_capacity(cpu);
2971 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2972 !__dl_overflow(dl_b, cap, 0, new_bw)) {
2973 if (hrtimer_active(&p->dl.inactive_timer))
2974 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2975 __dl_add(dl_b, new_bw, cpus);
2977 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2978 !__dl_overflow(dl_b, cap, p->dl.dl_bw, new_bw)) {
2980 * XXX this is slightly incorrect: when the task
2981 * utilization decreases, we should delay the total
2982 * utilization change until the task's 0-lag point.
2983 * But this would require to set the task's "inactive
2984 * timer" when the task is not inactive.
2986 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2987 __dl_add(dl_b, new_bw, cpus);
2988 dl_change_utilization(p, new_bw);
2990 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2992 * Do not decrease the total deadline utilization here,
2993 * switched_from_dl() will take care to do it at the correct
2998 raw_spin_unlock(&dl_b->lock);
3004 * This function initializes the sched_dl_entity of a newly becoming
3005 * SCHED_DEADLINE task.
3007 * Only the static values are considered here, the actual runtime and the
3008 * absolute deadline will be properly calculated when the task is enqueued
3009 * for the first time with its new policy.
3011 void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3013 struct sched_dl_entity *dl_se = &p->dl;
3015 dl_se->dl_runtime = attr->sched_runtime;
3016 dl_se->dl_deadline = attr->sched_deadline;
3017 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3018 dl_se->flags = attr->sched_flags & SCHED_DL_FLAGS;
3019 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3020 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
3023 void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3025 struct sched_dl_entity *dl_se = &p->dl;
3027 attr->sched_priority = p->rt_priority;
3028 attr->sched_runtime = dl_se->dl_runtime;
3029 attr->sched_deadline = dl_se->dl_deadline;
3030 attr->sched_period = dl_se->dl_period;
3031 attr->sched_flags &= ~SCHED_DL_FLAGS;
3032 attr->sched_flags |= dl_se->flags;
3036 * This function validates the new parameters of a -deadline task.
3037 * We ask for the deadline not being zero, and greater or equal
3038 * than the runtime, as well as the period of being zero or
3039 * greater than deadline. Furthermore, we have to be sure that
3040 * user parameters are above the internal resolution of 1us (we
3041 * check sched_runtime only since it is always the smaller one) and
3042 * below 2^63 ns (we have to check both sched_deadline and
3043 * sched_period, as the latter can be zero).
3045 bool __checkparam_dl(const struct sched_attr *attr)
3047 u64 period, max, min;
3049 /* special dl tasks don't actually use any parameter */
3050 if (attr->sched_flags & SCHED_FLAG_SUGOV)
3054 if (attr->sched_deadline == 0)
3058 * Since we truncate DL_SCALE bits, make sure we're at least
3061 if (attr->sched_runtime < (1ULL << DL_SCALE))
3065 * Since we use the MSB for wrap-around and sign issues, make
3066 * sure it's not set (mind that period can be equal to zero).
3068 if (attr->sched_deadline & (1ULL << 63) ||
3069 attr->sched_period & (1ULL << 63))
3072 period = attr->sched_period;
3074 period = attr->sched_deadline;
3076 /* runtime <= deadline <= period (if period != 0) */
3077 if (period < attr->sched_deadline ||
3078 attr->sched_deadline < attr->sched_runtime)
3081 max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
3082 min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
3084 if (period < min || period > max)
3091 * This function clears the sched_dl_entity static params.
3093 static void __dl_clear_params(struct sched_dl_entity *dl_se)
3095 dl_se->dl_runtime = 0;
3096 dl_se->dl_deadline = 0;
3097 dl_se->dl_period = 0;
3100 dl_se->dl_density = 0;
3102 dl_se->dl_throttled = 0;
3103 dl_se->dl_yielded = 0;
3104 dl_se->dl_non_contending = 0;
3105 dl_se->dl_overrun = 0;
3106 dl_se->dl_server = 0;
3108 #ifdef CONFIG_RT_MUTEXES
3109 dl_se->pi_se = dl_se;
3113 void init_dl_entity(struct sched_dl_entity *dl_se)
3115 RB_CLEAR_NODE(&dl_se->rb_node);
3116 init_dl_task_timer(dl_se);
3117 init_dl_inactive_task_timer(dl_se);
3118 __dl_clear_params(dl_se);
3121 bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
3123 struct sched_dl_entity *dl_se = &p->dl;
3125 if (dl_se->dl_runtime != attr->sched_runtime ||
3126 dl_se->dl_deadline != attr->sched_deadline ||
3127 dl_se->dl_period != attr->sched_period ||
3128 dl_se->flags != (attr->sched_flags & SCHED_DL_FLAGS))
3135 int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
3136 const struct cpumask *trial)
3138 unsigned long flags, cap;
3139 struct dl_bw *cur_dl_b;
3142 rcu_read_lock_sched();
3143 cur_dl_b = dl_bw_of(cpumask_any(cur));
3144 cap = __dl_bw_capacity(trial);
3145 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
3146 if (__dl_overflow(cur_dl_b, cap, 0, 0))
3148 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
3149 rcu_read_unlock_sched();
3154 enum dl_bw_request {
3155 dl_bw_req_check_overflow = 0,
3160 static int dl_bw_manage(enum dl_bw_request req, int cpu, u64 dl_bw)
3162 unsigned long flags;
3166 rcu_read_lock_sched();
3167 dl_b = dl_bw_of(cpu);
3168 raw_spin_lock_irqsave(&dl_b->lock, flags);
3170 if (req == dl_bw_req_free) {
3171 __dl_sub(dl_b, dl_bw, dl_bw_cpus(cpu));
3173 unsigned long cap = dl_bw_capacity(cpu);
3175 overflow = __dl_overflow(dl_b, cap, 0, dl_bw);
3177 if (req == dl_bw_req_alloc && !overflow) {
3179 * We reserve space in the destination
3180 * root_domain, as we can't fail after this point.
3181 * We will free resources in the source root_domain
3182 * later on (see set_cpus_allowed_dl()).
3184 __dl_add(dl_b, dl_bw, dl_bw_cpus(cpu));
3188 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3189 rcu_read_unlock_sched();
3191 return overflow ? -EBUSY : 0;
3194 int dl_bw_check_overflow(int cpu)
3196 return dl_bw_manage(dl_bw_req_check_overflow, cpu, 0);
3199 int dl_bw_alloc(int cpu, u64 dl_bw)
3201 return dl_bw_manage(dl_bw_req_alloc, cpu, dl_bw);
3204 void dl_bw_free(int cpu, u64 dl_bw)
3206 dl_bw_manage(dl_bw_req_free, cpu, dl_bw);
3210 #ifdef CONFIG_SCHED_DEBUG
3211 void print_dl_stats(struct seq_file *m, int cpu)
3213 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
3215 #endif /* CONFIG_SCHED_DEBUG */