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>
20 * Default limits for DL period; on the top end we guard against small util
21 * tasks still getting ridiculously long effective runtimes, on the bottom end we
22 * guard against timer DoS.
24 static unsigned int sysctl_sched_dl_period_max = 1 << 22; /* ~4 seconds */
25 static unsigned int sysctl_sched_dl_period_min = 100; /* 100 us */
27 static struct ctl_table sched_dl_sysctls[] = {
29 .procname = "sched_deadline_period_max_us",
30 .data = &sysctl_sched_dl_period_max,
31 .maxlen = sizeof(unsigned int),
33 .proc_handler = proc_dointvec,
36 .procname = "sched_deadline_period_min_us",
37 .data = &sysctl_sched_dl_period_min,
38 .maxlen = sizeof(unsigned int),
40 .proc_handler = proc_dointvec,
45 static int __init sched_dl_sysctl_init(void)
47 register_sysctl_init("kernel", sched_dl_sysctls);
50 late_initcall(sched_dl_sysctl_init);
53 static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
55 return container_of(dl_se, struct task_struct, dl);
58 static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
60 return container_of(dl_rq, struct rq, dl);
63 static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
65 struct task_struct *p = dl_task_of(dl_se);
66 struct rq *rq = task_rq(p);
71 static inline int on_dl_rq(struct sched_dl_entity *dl_se)
73 return !RB_EMPTY_NODE(&dl_se->rb_node);
76 #ifdef CONFIG_RT_MUTEXES
77 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
82 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
84 return pi_of(dl_se) != dl_se;
87 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
92 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
99 static inline struct dl_bw *dl_bw_of(int i)
101 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
102 "sched RCU must be held");
103 return &cpu_rq(i)->rd->dl_bw;
106 static inline int dl_bw_cpus(int i)
108 struct root_domain *rd = cpu_rq(i)->rd;
111 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
112 "sched RCU must be held");
114 if (cpumask_subset(rd->span, cpu_active_mask))
115 return cpumask_weight(rd->span);
119 for_each_cpu_and(i, rd->span, cpu_active_mask)
125 static inline unsigned long __dl_bw_capacity(int i)
127 struct root_domain *rd = cpu_rq(i)->rd;
128 unsigned long cap = 0;
130 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
131 "sched RCU must be held");
133 for_each_cpu_and(i, rd->span, cpu_active_mask)
134 cap += capacity_orig_of(i);
140 * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
141 * of the CPU the task is running on rather rd's \Sum CPU capacity.
143 static inline unsigned long dl_bw_capacity(int i)
145 if (!static_branch_unlikely(&sched_asym_cpucapacity) &&
146 capacity_orig_of(i) == SCHED_CAPACITY_SCALE) {
147 return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT;
149 return __dl_bw_capacity(i);
153 static inline bool dl_bw_visited(int cpu, u64 gen)
155 struct root_domain *rd = cpu_rq(cpu)->rd;
157 if (rd->visit_gen == gen)
165 void __dl_update(struct dl_bw *dl_b, s64 bw)
167 struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw);
170 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
171 "sched RCU must be held");
172 for_each_cpu_and(i, rd->span, cpu_active_mask) {
173 struct rq *rq = cpu_rq(i);
175 rq->dl.extra_bw += bw;
179 static inline struct dl_bw *dl_bw_of(int i)
181 return &cpu_rq(i)->dl.dl_bw;
184 static inline int dl_bw_cpus(int i)
189 static inline unsigned long dl_bw_capacity(int i)
191 return SCHED_CAPACITY_SCALE;
194 static inline bool dl_bw_visited(int cpu, u64 gen)
200 void __dl_update(struct dl_bw *dl_b, s64 bw)
202 struct dl_rq *dl = container_of(dl_b, struct dl_rq, dl_bw);
209 void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
211 dl_b->total_bw -= tsk_bw;
212 __dl_update(dl_b, (s32)tsk_bw / cpus);
216 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
218 dl_b->total_bw += tsk_bw;
219 __dl_update(dl_b, -((s32)tsk_bw / cpus));
223 __dl_overflow(struct dl_bw *dl_b, unsigned long cap, u64 old_bw, u64 new_bw)
225 return dl_b->bw != -1 &&
226 cap_scale(dl_b->bw, cap) < dl_b->total_bw - old_bw + new_bw;
230 void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
232 u64 old = dl_rq->running_bw;
234 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
235 dl_rq->running_bw += dl_bw;
236 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
237 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
238 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
239 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
243 void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
245 u64 old = dl_rq->running_bw;
247 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
248 dl_rq->running_bw -= dl_bw;
249 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
250 if (dl_rq->running_bw > old)
251 dl_rq->running_bw = 0;
252 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
253 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
257 void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
259 u64 old = dl_rq->this_bw;
261 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
262 dl_rq->this_bw += dl_bw;
263 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
267 void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
269 u64 old = dl_rq->this_bw;
271 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
272 dl_rq->this_bw -= dl_bw;
273 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
274 if (dl_rq->this_bw > old)
276 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
280 void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
282 if (!dl_entity_is_special(dl_se))
283 __add_rq_bw(dl_se->dl_bw, dl_rq);
287 void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
289 if (!dl_entity_is_special(dl_se))
290 __sub_rq_bw(dl_se->dl_bw, dl_rq);
294 void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
296 if (!dl_entity_is_special(dl_se))
297 __add_running_bw(dl_se->dl_bw, dl_rq);
301 void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
303 if (!dl_entity_is_special(dl_se))
304 __sub_running_bw(dl_se->dl_bw, dl_rq);
307 static void dl_change_utilization(struct task_struct *p, u64 new_bw)
311 BUG_ON(p->dl.flags & SCHED_FLAG_SUGOV);
313 if (task_on_rq_queued(p))
317 if (p->dl.dl_non_contending) {
318 sub_running_bw(&p->dl, &rq->dl);
319 p->dl.dl_non_contending = 0;
321 * If the timer handler is currently running and the
322 * timer cannot be canceled, inactive_task_timer()
323 * will see that dl_not_contending is not set, and
324 * will not touch the rq's active utilization,
325 * so we are still safe.
327 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
330 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
331 __add_rq_bw(new_bw, &rq->dl);
335 * The utilization of a task cannot be immediately removed from
336 * the rq active utilization (running_bw) when the task blocks.
337 * Instead, we have to wait for the so called "0-lag time".
339 * If a task blocks before the "0-lag time", a timer (the inactive
340 * timer) is armed, and running_bw is decreased when the timer
343 * If the task wakes up again before the inactive timer fires,
344 * the timer is canceled, whereas if the task wakes up after the
345 * inactive timer fired (and running_bw has been decreased) the
346 * task's utilization has to be added to running_bw again.
347 * A flag in the deadline scheduling entity (dl_non_contending)
348 * is used to avoid race conditions between the inactive timer handler
351 * The following diagram shows how running_bw is updated. A task is
352 * "ACTIVE" when its utilization contributes to running_bw; an
353 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
354 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
355 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
356 * time already passed, which does not contribute to running_bw anymore.
357 * +------------------+
359 * +------------------>+ contending |
360 * | add_running_bw | |
361 * | +----+------+------+
364 * +--------+-------+ | |
365 * | | t >= 0-lag | | wakeup
366 * | INACTIVE |<---------------+ |
367 * | | sub_running_bw | |
368 * +--------+-------+ | |
373 * | +----+------+------+
374 * | sub_running_bw | ACTIVE |
375 * +-------------------+ |
376 * inactive timer | non contending |
377 * fired +------------------+
379 * The task_non_contending() function is invoked when a task
380 * blocks, and checks if the 0-lag time already passed or
381 * not (in the first case, it directly updates running_bw;
382 * in the second case, it arms the inactive timer).
384 * The task_contending() function is invoked when a task wakes
385 * up, and checks if the task is still in the "ACTIVE non contending"
386 * state or not (in the second case, it updates running_bw).
388 static void task_non_contending(struct task_struct *p)
390 struct sched_dl_entity *dl_se = &p->dl;
391 struct hrtimer *timer = &dl_se->inactive_timer;
392 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
393 struct rq *rq = rq_of_dl_rq(dl_rq);
397 * If this is a non-deadline task that has been boosted,
400 if (dl_se->dl_runtime == 0)
403 if (dl_entity_is_special(dl_se))
406 WARN_ON(dl_se->dl_non_contending);
408 zerolag_time = dl_se->deadline -
409 div64_long((dl_se->runtime * dl_se->dl_period),
413 * Using relative times instead of the absolute "0-lag time"
414 * allows to simplify the code
416 zerolag_time -= rq_clock(rq);
419 * If the "0-lag time" already passed, decrease the active
420 * utilization now, instead of starting a timer
422 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
424 sub_running_bw(dl_se, dl_rq);
425 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
426 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
428 if (READ_ONCE(p->__state) == TASK_DEAD)
429 sub_rq_bw(&p->dl, &rq->dl);
430 raw_spin_lock(&dl_b->lock);
431 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
432 __dl_clear_params(p);
433 raw_spin_unlock(&dl_b->lock);
439 dl_se->dl_non_contending = 1;
441 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD);
444 static void task_contending(struct sched_dl_entity *dl_se, int flags)
446 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
449 * If this is a non-deadline task that has been boosted,
452 if (dl_se->dl_runtime == 0)
455 if (flags & ENQUEUE_MIGRATED)
456 add_rq_bw(dl_se, dl_rq);
458 if (dl_se->dl_non_contending) {
459 dl_se->dl_non_contending = 0;
461 * If the timer handler is currently running and the
462 * timer cannot be canceled, inactive_task_timer()
463 * will see that dl_not_contending is not set, and
464 * will not touch the rq's active utilization,
465 * so we are still safe.
467 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1)
468 put_task_struct(dl_task_of(dl_se));
471 * Since "dl_non_contending" is not set, the
472 * task's utilization has already been removed from
473 * active utilization (either when the task blocked,
474 * when the "inactive timer" fired).
477 add_running_bw(dl_se, dl_rq);
481 static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
483 struct sched_dl_entity *dl_se = &p->dl;
485 return rb_first_cached(&dl_rq->root) == &dl_se->rb_node;
488 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
490 void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
492 raw_spin_lock_init(&dl_b->dl_runtime_lock);
493 dl_b->dl_period = period;
494 dl_b->dl_runtime = runtime;
497 void init_dl_bw(struct dl_bw *dl_b)
499 raw_spin_lock_init(&dl_b->lock);
500 if (global_rt_runtime() == RUNTIME_INF)
503 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
507 void init_dl_rq(struct dl_rq *dl_rq)
509 dl_rq->root = RB_ROOT_CACHED;
512 /* zero means no -deadline tasks */
513 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
515 dl_rq->dl_nr_migratory = 0;
516 dl_rq->overloaded = 0;
517 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
519 init_dl_bw(&dl_rq->dl_bw);
522 dl_rq->running_bw = 0;
524 init_dl_rq_bw_ratio(dl_rq);
529 static inline int dl_overloaded(struct rq *rq)
531 return atomic_read(&rq->rd->dlo_count);
534 static inline void dl_set_overload(struct rq *rq)
539 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
541 * Must be visible before the overload count is
542 * set (as in sched_rt.c).
544 * Matched by the barrier in pull_dl_task().
547 atomic_inc(&rq->rd->dlo_count);
550 static inline void dl_clear_overload(struct rq *rq)
555 atomic_dec(&rq->rd->dlo_count);
556 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
559 static void update_dl_migration(struct dl_rq *dl_rq)
561 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
562 if (!dl_rq->overloaded) {
563 dl_set_overload(rq_of_dl_rq(dl_rq));
564 dl_rq->overloaded = 1;
566 } else if (dl_rq->overloaded) {
567 dl_clear_overload(rq_of_dl_rq(dl_rq));
568 dl_rq->overloaded = 0;
572 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
574 struct task_struct *p = dl_task_of(dl_se);
576 if (p->nr_cpus_allowed > 1)
577 dl_rq->dl_nr_migratory++;
579 update_dl_migration(dl_rq);
582 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
584 struct task_struct *p = dl_task_of(dl_se);
586 if (p->nr_cpus_allowed > 1)
587 dl_rq->dl_nr_migratory--;
589 update_dl_migration(dl_rq);
592 #define __node_2_pdl(node) \
593 rb_entry((node), struct task_struct, pushable_dl_tasks)
595 static inline bool __pushable_less(struct rb_node *a, const struct rb_node *b)
597 return dl_entity_preempt(&__node_2_pdl(a)->dl, &__node_2_pdl(b)->dl);
601 * The list of pushable -deadline task is not a plist, like in
602 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
604 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
606 struct rb_node *leftmost;
608 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
610 leftmost = rb_add_cached(&p->pushable_dl_tasks,
611 &rq->dl.pushable_dl_tasks_root,
614 rq->dl.earliest_dl.next = p->dl.deadline;
617 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
619 struct dl_rq *dl_rq = &rq->dl;
620 struct rb_root_cached *root = &dl_rq->pushable_dl_tasks_root;
621 struct rb_node *leftmost;
623 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
626 leftmost = rb_erase_cached(&p->pushable_dl_tasks, root);
628 dl_rq->earliest_dl.next = __node_2_pdl(leftmost)->dl.deadline;
630 RB_CLEAR_NODE(&p->pushable_dl_tasks);
633 static inline int has_pushable_dl_tasks(struct rq *rq)
635 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
638 static int push_dl_task(struct rq *rq);
640 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
642 return rq->online && dl_task(prev);
645 static DEFINE_PER_CPU(struct callback_head, dl_push_head);
646 static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
648 static void push_dl_tasks(struct rq *);
649 static void pull_dl_task(struct rq *);
651 static inline void deadline_queue_push_tasks(struct rq *rq)
653 if (!has_pushable_dl_tasks(rq))
656 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
659 static inline void deadline_queue_pull_task(struct rq *rq)
661 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
664 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
666 static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
668 struct rq *later_rq = NULL;
671 later_rq = find_lock_later_rq(p, rq);
676 * If we cannot preempt any rq, fall back to pick any
679 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
680 if (cpu >= nr_cpu_ids) {
682 * Failed to find any suitable CPU.
683 * The task will never come back!
685 BUG_ON(dl_bandwidth_enabled());
688 * If admission control is disabled we
689 * try a little harder to let the task
692 cpu = cpumask_any(cpu_active_mask);
694 later_rq = cpu_rq(cpu);
695 double_lock_balance(rq, later_rq);
698 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
700 * Inactive timer is armed (or callback is running, but
701 * waiting for us to release rq locks). In any case, when it
702 * will fire (or continue), it will see running_bw of this
703 * task migrated to later_rq (and correctly handle it).
705 sub_running_bw(&p->dl, &rq->dl);
706 sub_rq_bw(&p->dl, &rq->dl);
708 add_rq_bw(&p->dl, &later_rq->dl);
709 add_running_bw(&p->dl, &later_rq->dl);
711 sub_rq_bw(&p->dl, &rq->dl);
712 add_rq_bw(&p->dl, &later_rq->dl);
716 * And we finally need to fixup root_domain(s) bandwidth accounting,
717 * since p is still hanging out in the old (now moved to default) root
720 dl_b = &rq->rd->dl_bw;
721 raw_spin_lock(&dl_b->lock);
722 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
723 raw_spin_unlock(&dl_b->lock);
725 dl_b = &later_rq->rd->dl_bw;
726 raw_spin_lock(&dl_b->lock);
727 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
728 raw_spin_unlock(&dl_b->lock);
730 set_task_cpu(p, later_rq->cpu);
731 double_unlock_balance(later_rq, rq);
739 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
744 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
749 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
754 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
758 static inline void deadline_queue_push_tasks(struct rq *rq)
762 static inline void deadline_queue_pull_task(struct rq *rq)
765 #endif /* CONFIG_SMP */
767 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
768 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
769 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags);
772 * We are being explicitly informed that a new instance is starting,
773 * and this means that:
774 * - the absolute deadline of the entity has to be placed at
775 * current time + relative deadline;
776 * - the runtime of the entity has to be set to the maximum value.
778 * The capability of specifying such event is useful whenever a -deadline
779 * entity wants to (try to!) synchronize its behaviour with the scheduler's
780 * one, and to (try to!) reconcile itself with its own scheduling
783 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
785 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
786 struct rq *rq = rq_of_dl_rq(dl_rq);
788 WARN_ON(is_dl_boosted(dl_se));
789 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
792 * We are racing with the deadline timer. So, do nothing because
793 * the deadline timer handler will take care of properly recharging
794 * the runtime and postponing the deadline
796 if (dl_se->dl_throttled)
800 * We use the regular wall clock time to set deadlines in the
801 * future; in fact, we must consider execution overheads (time
802 * spent on hardirq context, etc.).
804 dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline;
805 dl_se->runtime = dl_se->dl_runtime;
809 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
810 * possibility of a entity lasting more than what it declared, and thus
811 * exhausting its runtime.
813 * Here we are interested in making runtime overrun possible, but we do
814 * not want a entity which is misbehaving to affect the scheduling of all
816 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
817 * is used, in order to confine each entity within its own bandwidth.
819 * This function deals exactly with that, and ensures that when the runtime
820 * of a entity is replenished, its deadline is also postponed. That ensures
821 * the overrunning entity can't interfere with other entity in the system and
822 * can't make them miss their deadlines. Reasons why this kind of overruns
823 * could happen are, typically, a entity voluntarily trying to overcome its
824 * runtime, or it just underestimated it during sched_setattr().
826 static void replenish_dl_entity(struct sched_dl_entity *dl_se)
828 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
829 struct rq *rq = rq_of_dl_rq(dl_rq);
831 BUG_ON(pi_of(dl_se)->dl_runtime <= 0);
834 * This could be the case for a !-dl task that is boosted.
835 * Just go with full inherited parameters.
837 if (dl_se->dl_deadline == 0) {
838 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
839 dl_se->runtime = pi_of(dl_se)->dl_runtime;
842 if (dl_se->dl_yielded && dl_se->runtime > 0)
846 * We keep moving the deadline away until we get some
847 * available runtime for the entity. This ensures correct
848 * handling of situations where the runtime overrun is
851 while (dl_se->runtime <= 0) {
852 dl_se->deadline += pi_of(dl_se)->dl_period;
853 dl_se->runtime += pi_of(dl_se)->dl_runtime;
857 * At this point, the deadline really should be "in
858 * the future" with respect to rq->clock. If it's
859 * not, we are, for some reason, lagging too much!
860 * Anyway, after having warn userspace abut that,
861 * we still try to keep the things running by
862 * resetting the deadline and the budget of the
865 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
866 printk_deferred_once("sched: DL replenish lagged too much\n");
867 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
868 dl_se->runtime = pi_of(dl_se)->dl_runtime;
871 if (dl_se->dl_yielded)
872 dl_se->dl_yielded = 0;
873 if (dl_se->dl_throttled)
874 dl_se->dl_throttled = 0;
878 * Here we check if --at time t-- an entity (which is probably being
879 * [re]activated or, in general, enqueued) can use its remaining runtime
880 * and its current deadline _without_ exceeding the bandwidth it is
881 * assigned (function returns true if it can't). We are in fact applying
882 * one of the CBS rules: when a task wakes up, if the residual runtime
883 * over residual deadline fits within the allocated bandwidth, then we
884 * can keep the current (absolute) deadline and residual budget without
885 * disrupting the schedulability of the system. Otherwise, we should
886 * refill the runtime and set the deadline a period in the future,
887 * because keeping the current (absolute) deadline of the task would
888 * result in breaking guarantees promised to other tasks (refer to
889 * Documentation/scheduler/sched-deadline.rst for more information).
891 * This function returns true if:
893 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
895 * IOW we can't recycle current parameters.
897 * Notice that the bandwidth check is done against the deadline. For
898 * task with deadline equal to period this is the same of using
899 * dl_period instead of dl_deadline in the equation above.
901 static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t)
906 * left and right are the two sides of the equation above,
907 * after a bit of shuffling to use multiplications instead
910 * Note that none of the time values involved in the two
911 * multiplications are absolute: dl_deadline and dl_runtime
912 * are the relative deadline and the maximum runtime of each
913 * instance, runtime is the runtime left for the last instance
914 * and (deadline - t), since t is rq->clock, is the time left
915 * to the (absolute) deadline. Even if overflowing the u64 type
916 * is very unlikely to occur in both cases, here we scale down
917 * as we want to avoid that risk at all. Scaling down by 10
918 * means that we reduce granularity to 1us. We are fine with it,
919 * since this is only a true/false check and, anyway, thinking
920 * of anything below microseconds resolution is actually fiction
921 * (but still we want to give the user that illusion >;).
923 left = (pi_of(dl_se)->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
924 right = ((dl_se->deadline - t) >> DL_SCALE) *
925 (pi_of(dl_se)->dl_runtime >> DL_SCALE);
927 return dl_time_before(right, left);
931 * Revised wakeup rule [1]: For self-suspending tasks, rather then
932 * re-initializing task's runtime and deadline, the revised wakeup
933 * rule adjusts the task's runtime to avoid the task to overrun its
936 * Reasoning: a task may overrun the density if:
937 * runtime / (deadline - t) > dl_runtime / dl_deadline
939 * Therefore, runtime can be adjusted to:
940 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
942 * In such way that runtime will be equal to the maximum density
943 * the task can use without breaking any rule.
945 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
946 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
949 update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
951 u64 laxity = dl_se->deadline - rq_clock(rq);
954 * If the task has deadline < period, and the deadline is in the past,
955 * it should already be throttled before this check.
957 * See update_dl_entity() comments for further details.
959 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
961 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
965 * Regarding the deadline, a task with implicit deadline has a relative
966 * deadline == relative period. A task with constrained deadline has a
967 * relative deadline <= relative period.
969 * We support constrained deadline tasks. However, there are some restrictions
970 * applied only for tasks which do not have an implicit deadline. See
971 * update_dl_entity() to know more about such restrictions.
973 * The dl_is_implicit() returns true if the task has an implicit deadline.
975 static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
977 return dl_se->dl_deadline == dl_se->dl_period;
981 * When a deadline entity is placed in the runqueue, its runtime and deadline
982 * might need to be updated. This is done by a CBS wake up rule. There are two
983 * different rules: 1) the original CBS; and 2) the Revisited CBS.
985 * When the task is starting a new period, the Original CBS is used. In this
986 * case, the runtime is replenished and a new absolute deadline is set.
988 * When a task is queued before the begin of the next period, using the
989 * remaining runtime and deadline could make the entity to overflow, see
990 * dl_entity_overflow() to find more about runtime overflow. When such case
991 * is detected, the runtime and deadline need to be updated.
993 * If the task has an implicit deadline, i.e., deadline == period, the Original
994 * CBS is applied. the runtime is replenished and a new absolute deadline is
995 * set, as in the previous cases.
997 * However, the Original CBS does not work properly for tasks with
998 * deadline < period, which are said to have a constrained deadline. By
999 * applying the Original CBS, a constrained deadline task would be able to run
1000 * runtime/deadline in a period. With deadline < period, the task would
1001 * overrun the runtime/period allowed bandwidth, breaking the admission test.
1003 * In order to prevent this misbehave, the Revisited CBS is used for
1004 * constrained deadline tasks when a runtime overflow is detected. In the
1005 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
1006 * the remaining runtime of the task is reduced to avoid runtime overflow.
1007 * Please refer to the comments update_dl_revised_wakeup() function to find
1008 * more about the Revised CBS rule.
1010 static void update_dl_entity(struct sched_dl_entity *dl_se)
1012 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1013 struct rq *rq = rq_of_dl_rq(dl_rq);
1015 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
1016 dl_entity_overflow(dl_se, rq_clock(rq))) {
1018 if (unlikely(!dl_is_implicit(dl_se) &&
1019 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1020 !is_dl_boosted(dl_se))) {
1021 update_dl_revised_wakeup(dl_se, rq);
1025 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
1026 dl_se->runtime = pi_of(dl_se)->dl_runtime;
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 task_struct *p)
1047 struct sched_dl_entity *dl_se = &p->dl;
1048 struct hrtimer *timer = &dl_se->dl_timer;
1049 struct rq *rq = task_rq(p);
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)) {
1084 hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD);
1091 * This is the bandwidth enforcement timer callback. If here, we know
1092 * a task is not on its dl_rq, since the fact that the timer was running
1093 * means the task is throttled and needs a runtime replenishment.
1095 * However, what we actually do depends on the fact the task is active,
1096 * (it is on its rq) or has been removed from there by a call to
1097 * dequeue_task_dl(). In the former case we must issue the runtime
1098 * replenishment and add the task back to the dl_rq; in the latter, we just
1099 * do nothing but clearing dl_throttled, so that runtime and deadline
1100 * updating (and the queueing back to dl_rq) will be done by the
1101 * next call to enqueue_task_dl().
1103 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
1105 struct sched_dl_entity *dl_se = container_of(timer,
1106 struct sched_dl_entity,
1108 struct task_struct *p = dl_task_of(dl_se);
1112 rq = task_rq_lock(p, &rf);
1115 * The task might have changed its scheduling policy to something
1116 * different than SCHED_DEADLINE (through switched_from_dl()).
1122 * The task might have been boosted by someone else and might be in the
1123 * boosting/deboosting path, its not throttled.
1125 if (is_dl_boosted(dl_se))
1129 * Spurious timer due to start_dl_timer() race; or we already received
1130 * a replenishment from rt_mutex_setprio().
1132 if (!dl_se->dl_throttled)
1136 update_rq_clock(rq);
1139 * If the throttle happened during sched-out; like:
1146 * __dequeue_task_dl()
1149 * We can be both throttled and !queued. Replenish the counter
1150 * but do not enqueue -- wait for our wakeup to do that.
1152 if (!task_on_rq_queued(p)) {
1153 replenish_dl_entity(dl_se);
1158 if (unlikely(!rq->online)) {
1160 * If the runqueue is no longer available, migrate the
1161 * task elsewhere. This necessarily changes rq.
1163 lockdep_unpin_lock(__rq_lockp(rq), rf.cookie);
1164 rq = dl_task_offline_migration(rq, p);
1165 rf.cookie = lockdep_pin_lock(__rq_lockp(rq));
1166 update_rq_clock(rq);
1169 * Now that the task has been migrated to the new RQ and we
1170 * have that locked, proceed as normal and enqueue the task
1176 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1177 if (dl_task(rq->curr))
1178 check_preempt_curr_dl(rq, p, 0);
1184 * Queueing this task back might have overloaded rq, check if we need
1185 * to kick someone away.
1187 if (has_pushable_dl_tasks(rq)) {
1189 * Nothing relies on rq->lock after this, so its safe to drop
1192 rq_unpin_lock(rq, &rf);
1194 rq_repin_lock(rq, &rf);
1199 task_rq_unlock(rq, p, &rf);
1202 * This can free the task_struct, including this hrtimer, do not touch
1203 * anything related to that after this.
1207 return HRTIMER_NORESTART;
1210 void init_dl_task_timer(struct sched_dl_entity *dl_se)
1212 struct hrtimer *timer = &dl_se->dl_timer;
1214 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1215 timer->function = dl_task_timer;
1219 * During the activation, CBS checks if it can reuse the current task's
1220 * runtime and period. If the deadline of the task is in the past, CBS
1221 * cannot use the runtime, and so it replenishes the task. This rule
1222 * works fine for implicit deadline tasks (deadline == period), and the
1223 * CBS was designed for implicit deadline tasks. However, a task with
1224 * constrained deadline (deadline < period) might be awakened after the
1225 * deadline, but before the next period. In this case, replenishing the
1226 * task would allow it to run for runtime / deadline. As in this case
1227 * deadline < period, CBS enables a task to run for more than the
1228 * runtime / period. In a very loaded system, this can cause a domino
1229 * effect, making other tasks miss their deadlines.
1231 * To avoid this problem, in the activation of a constrained deadline
1232 * task after the deadline but before the next period, throttle the
1233 * task and set the replenishing timer to the begin of the next period,
1234 * unless it is boosted.
1236 static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1238 struct task_struct *p = dl_task_of(dl_se);
1239 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1241 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1242 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1243 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(p)))
1245 dl_se->dl_throttled = 1;
1246 if (dl_se->runtime > 0)
1252 int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1254 return (dl_se->runtime <= 0);
1258 * This function implements the GRUB accounting rule:
1259 * according to the GRUB reclaiming algorithm, the runtime is
1260 * not decreased as "dq = -dt", but as
1261 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1262 * where u is the utilization of the task, Umax is the maximum reclaimable
1263 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1264 * as the difference between the "total runqueue utilization" and the
1265 * runqueue active utilization, and Uextra is the (per runqueue) extra
1266 * reclaimable utilization.
1267 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1268 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1270 * Since rq->dl.bw_ratio contains 1 / Umax multiplied by 2^RATIO_SHIFT,
1271 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1272 * Since delta is a 64 bit variable, to have an overflow its value
1273 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
1274 * So, overflow is not an issue here.
1276 static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1278 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1280 u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT;
1283 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1284 * we compare u_inact + rq->dl.extra_bw with
1285 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1286 * u_inact + rq->dl.extra_bw can be larger than
1287 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
1288 * leading to wrong results)
1290 if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min)
1293 u_act = BW_UNIT - u_inact - rq->dl.extra_bw;
1295 return (delta * u_act) >> BW_SHIFT;
1299 * Update the current task's runtime statistics (provided it is still
1300 * a -deadline task and has not been removed from the dl_rq).
1302 static void update_curr_dl(struct rq *rq)
1304 struct task_struct *curr = rq->curr;
1305 struct sched_dl_entity *dl_se = &curr->dl;
1306 u64 delta_exec, scaled_delta_exec;
1307 int cpu = cpu_of(rq);
1310 if (!dl_task(curr) || !on_dl_rq(dl_se))
1314 * Consumed budget is computed considering the time as
1315 * observed by schedulable tasks (excluding time spent
1316 * in hardirq context, etc.). Deadlines are instead
1317 * computed using hard walltime. This seems to be the more
1318 * natural solution, but the full ramifications of this
1319 * approach need further study.
1321 now = rq_clock_task(rq);
1322 delta_exec = now - curr->se.exec_start;
1323 if (unlikely((s64)delta_exec <= 0)) {
1324 if (unlikely(dl_se->dl_yielded))
1329 schedstat_set(curr->stats.exec_max,
1330 max(curr->stats.exec_max, delta_exec));
1332 trace_sched_stat_runtime(curr, delta_exec, 0);
1334 curr->se.sum_exec_runtime += delta_exec;
1335 account_group_exec_runtime(curr, delta_exec);
1337 curr->se.exec_start = now;
1338 cgroup_account_cputime(curr, delta_exec);
1340 if (dl_entity_is_special(dl_se))
1344 * For tasks that participate in GRUB, we implement GRUB-PA: the
1345 * spare reclaimed bandwidth is used to clock down frequency.
1347 * For the others, we still need to scale reservation parameters
1348 * according to current frequency and CPU maximum capacity.
1350 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1351 scaled_delta_exec = grub_reclaim(delta_exec,
1355 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1356 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1358 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1359 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1362 dl_se->runtime -= scaled_delta_exec;
1365 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1366 dl_se->dl_throttled = 1;
1368 /* If requested, inform the user about runtime overruns. */
1369 if (dl_runtime_exceeded(dl_se) &&
1370 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1371 dl_se->dl_overrun = 1;
1373 __dequeue_task_dl(rq, curr, 0);
1374 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(curr)))
1375 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1377 if (!is_leftmost(curr, &rq->dl))
1382 * Because -- for now -- we share the rt bandwidth, we need to
1383 * account our runtime there too, otherwise actual rt tasks
1384 * would be able to exceed the shared quota.
1386 * Account to the root rt group for now.
1388 * The solution we're working towards is having the RT groups scheduled
1389 * using deadline servers -- however there's a few nasties to figure
1390 * out before that can happen.
1392 if (rt_bandwidth_enabled()) {
1393 struct rt_rq *rt_rq = &rq->rt;
1395 raw_spin_lock(&rt_rq->rt_runtime_lock);
1397 * We'll let actual RT tasks worry about the overflow here, we
1398 * have our own CBS to keep us inline; only account when RT
1399 * bandwidth is relevant.
1401 if (sched_rt_bandwidth_account(rt_rq))
1402 rt_rq->rt_time += delta_exec;
1403 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1407 static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1409 struct sched_dl_entity *dl_se = container_of(timer,
1410 struct sched_dl_entity,
1412 struct task_struct *p = dl_task_of(dl_se);
1416 rq = task_rq_lock(p, &rf);
1419 update_rq_clock(rq);
1421 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
1422 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1424 if (READ_ONCE(p->__state) == TASK_DEAD && dl_se->dl_non_contending) {
1425 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1426 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1427 dl_se->dl_non_contending = 0;
1430 raw_spin_lock(&dl_b->lock);
1431 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1432 raw_spin_unlock(&dl_b->lock);
1433 __dl_clear_params(p);
1437 if (dl_se->dl_non_contending == 0)
1440 sub_running_bw(dl_se, &rq->dl);
1441 dl_se->dl_non_contending = 0;
1443 task_rq_unlock(rq, p, &rf);
1446 return HRTIMER_NORESTART;
1449 void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1451 struct hrtimer *timer = &dl_se->inactive_timer;
1453 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1454 timer->function = inactive_task_timer;
1457 #define __node_2_dle(node) \
1458 rb_entry((node), struct sched_dl_entity, rb_node)
1462 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1464 struct rq *rq = rq_of_dl_rq(dl_rq);
1466 if (dl_rq->earliest_dl.curr == 0 ||
1467 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1468 if (dl_rq->earliest_dl.curr == 0)
1469 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_HIGHER);
1470 dl_rq->earliest_dl.curr = deadline;
1471 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1475 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1477 struct rq *rq = rq_of_dl_rq(dl_rq);
1480 * Since we may have removed our earliest (and/or next earliest)
1481 * task we must recompute them.
1483 if (!dl_rq->dl_nr_running) {
1484 dl_rq->earliest_dl.curr = 0;
1485 dl_rq->earliest_dl.next = 0;
1486 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1487 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1489 struct rb_node *leftmost = rb_first_cached(&dl_rq->root);
1490 struct sched_dl_entity *entry = __node_2_dle(leftmost);
1492 dl_rq->earliest_dl.curr = entry->deadline;
1493 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1499 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1500 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1502 #endif /* CONFIG_SMP */
1505 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1507 int prio = dl_task_of(dl_se)->prio;
1508 u64 deadline = dl_se->deadline;
1510 WARN_ON(!dl_prio(prio));
1511 dl_rq->dl_nr_running++;
1512 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1514 inc_dl_deadline(dl_rq, deadline);
1515 inc_dl_migration(dl_se, dl_rq);
1519 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1521 int prio = dl_task_of(dl_se)->prio;
1523 WARN_ON(!dl_prio(prio));
1524 WARN_ON(!dl_rq->dl_nr_running);
1525 dl_rq->dl_nr_running--;
1526 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1528 dec_dl_deadline(dl_rq, dl_se->deadline);
1529 dec_dl_migration(dl_se, dl_rq);
1532 static inline bool __dl_less(struct rb_node *a, const struct rb_node *b)
1534 return dl_time_before(__node_2_dle(a)->deadline, __node_2_dle(b)->deadline);
1537 static inline struct sched_statistics *
1538 __schedstats_from_dl_se(struct sched_dl_entity *dl_se)
1540 return &dl_task_of(dl_se)->stats;
1544 update_stats_wait_start_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1546 struct sched_statistics *stats;
1548 if (!schedstat_enabled())
1551 stats = __schedstats_from_dl_se(dl_se);
1552 __update_stats_wait_start(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1556 update_stats_wait_end_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1558 struct sched_statistics *stats;
1560 if (!schedstat_enabled())
1563 stats = __schedstats_from_dl_se(dl_se);
1564 __update_stats_wait_end(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1568 update_stats_enqueue_sleeper_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1570 struct sched_statistics *stats;
1572 if (!schedstat_enabled())
1575 stats = __schedstats_from_dl_se(dl_se);
1576 __update_stats_enqueue_sleeper(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1580 update_stats_enqueue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1583 if (!schedstat_enabled())
1586 if (flags & ENQUEUE_WAKEUP)
1587 update_stats_enqueue_sleeper_dl(dl_rq, dl_se);
1591 update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1594 struct task_struct *p = dl_task_of(dl_se);
1596 if (!schedstat_enabled())
1599 if ((flags & DEQUEUE_SLEEP)) {
1602 state = READ_ONCE(p->__state);
1603 if (state & TASK_INTERRUPTIBLE)
1604 __schedstat_set(p->stats.sleep_start,
1605 rq_clock(rq_of_dl_rq(dl_rq)));
1607 if (state & TASK_UNINTERRUPTIBLE)
1608 __schedstat_set(p->stats.block_start,
1609 rq_clock(rq_of_dl_rq(dl_rq)));
1613 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1615 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1617 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
1619 rb_add_cached(&dl_se->rb_node, &dl_rq->root, __dl_less);
1621 inc_dl_tasks(dl_se, dl_rq);
1624 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1626 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1628 if (RB_EMPTY_NODE(&dl_se->rb_node))
1631 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1633 RB_CLEAR_NODE(&dl_se->rb_node);
1635 dec_dl_tasks(dl_se, dl_rq);
1639 enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags)
1641 BUG_ON(on_dl_rq(dl_se));
1643 update_stats_enqueue_dl(dl_rq_of_se(dl_se), dl_se, flags);
1646 * If this is a wakeup or a new instance, the scheduling
1647 * parameters of the task might need updating. Otherwise,
1648 * we want a replenishment of its runtime.
1650 if (flags & ENQUEUE_WAKEUP) {
1651 task_contending(dl_se, flags);
1652 update_dl_entity(dl_se);
1653 } else if (flags & ENQUEUE_REPLENISH) {
1654 replenish_dl_entity(dl_se);
1655 } else if ((flags & ENQUEUE_RESTORE) &&
1656 dl_time_before(dl_se->deadline,
1657 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) {
1658 setup_new_dl_entity(dl_se);
1661 __enqueue_dl_entity(dl_se);
1664 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
1666 __dequeue_dl_entity(dl_se);
1669 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1671 if (is_dl_boosted(&p->dl)) {
1673 * Because of delays in the detection of the overrun of a
1674 * thread's runtime, it might be the case that a thread
1675 * goes to sleep in a rt mutex with negative runtime. As
1676 * a consequence, the thread will be throttled.
1678 * While waiting for the mutex, this thread can also be
1679 * boosted via PI, resulting in a thread that is throttled
1680 * and boosted at the same time.
1682 * In this case, the boost overrides the throttle.
1684 if (p->dl.dl_throttled) {
1686 * The replenish timer needs to be canceled. No
1687 * problem if it fires concurrently: boosted threads
1688 * are ignored in dl_task_timer().
1690 hrtimer_try_to_cancel(&p->dl.dl_timer);
1691 p->dl.dl_throttled = 0;
1693 } else if (!dl_prio(p->normal_prio)) {
1695 * Special case in which we have a !SCHED_DEADLINE task that is going
1696 * to be deboosted, but exceeds its runtime while doing so. No point in
1697 * replenishing it, as it's going to return back to its original
1698 * scheduling class after this. If it has been throttled, we need to
1699 * clear the flag, otherwise the task may wake up as throttled after
1700 * being boosted again with no means to replenish the runtime and clear
1703 p->dl.dl_throttled = 0;
1704 if (!(flags & ENQUEUE_REPLENISH))
1705 printk_deferred_once("sched: DL de-boosted task PID %d: REPLENISH flag missing\n",
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 (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
1718 dl_check_constrained_dl(&p->dl);
1720 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
1721 add_rq_bw(&p->dl, &rq->dl);
1722 add_running_bw(&p->dl, &rq->dl);
1726 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1727 * its budget it needs a replenishment and, since it now is on
1728 * its rq, the bandwidth timer callback (which clearly has not
1729 * run yet) will take care of this.
1730 * However, the active utilization does not depend on the fact
1731 * that the task is on the runqueue or not (but depends on the
1732 * task's state - in GRUB parlance, "inactive" vs "active contending").
1733 * In other words, even if a task is throttled its utilization must
1734 * be counted in the active utilization; hence, we need to call
1737 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1738 if (flags & ENQUEUE_WAKEUP)
1739 task_contending(&p->dl, flags);
1744 check_schedstat_required();
1745 update_stats_wait_start_dl(dl_rq_of_se(&p->dl), &p->dl);
1747 enqueue_dl_entity(&p->dl, flags);
1749 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1750 enqueue_pushable_dl_task(rq, p);
1753 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1755 update_stats_dequeue_dl(&rq->dl, &p->dl, flags);
1756 dequeue_dl_entity(&p->dl);
1757 dequeue_pushable_dl_task(rq, p);
1760 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1763 __dequeue_task_dl(rq, p, flags);
1765 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
1766 sub_running_bw(&p->dl, &rq->dl);
1767 sub_rq_bw(&p->dl, &rq->dl);
1771 * This check allows to start the inactive timer (or to immediately
1772 * decrease the active utilization, if needed) in two cases:
1773 * when the task blocks and when it is terminating
1774 * (p->state == TASK_DEAD). We can handle the two cases in the same
1775 * way, because from GRUB's point of view the same thing is happening
1776 * (the task moves from "active contending" to "active non contending"
1779 if (flags & DEQUEUE_SLEEP)
1780 task_non_contending(p);
1784 * Yield task semantic for -deadline tasks is:
1786 * get off from the CPU until our next instance, with
1787 * a new runtime. This is of little use now, since we
1788 * don't have a bandwidth reclaiming mechanism. Anyway,
1789 * bandwidth reclaiming is planned for the future, and
1790 * yield_task_dl will indicate that some spare budget
1791 * is available for other task instances to use it.
1793 static void yield_task_dl(struct rq *rq)
1796 * We make the task go to sleep until its current deadline by
1797 * forcing its runtime to zero. This way, update_curr_dl() stops
1798 * it and the bandwidth timer will wake it up and will give it
1799 * new scheduling parameters (thanks to dl_yielded=1).
1801 rq->curr->dl.dl_yielded = 1;
1803 update_rq_clock(rq);
1806 * Tell update_rq_clock() that we've just updated,
1807 * so we don't do microscopic update in schedule()
1808 * and double the fastpath cost.
1810 rq_clock_skip_update(rq);
1815 static int find_later_rq(struct task_struct *task);
1818 select_task_rq_dl(struct task_struct *p, int cpu, int flags)
1820 struct task_struct *curr;
1824 if (!(flags & WF_TTWU))
1830 curr = READ_ONCE(rq->curr); /* unlocked access */
1833 * If we are dealing with a -deadline task, we must
1834 * decide where to wake it up.
1835 * If it has a later deadline and the current task
1836 * on this rq can't move (provided the waking task
1837 * can!) we prefer to send it somewhere else. On the
1838 * other hand, if it has a shorter deadline, we
1839 * try to make it stay here, it might be important.
1841 select_rq = unlikely(dl_task(curr)) &&
1842 (curr->nr_cpus_allowed < 2 ||
1843 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1844 p->nr_cpus_allowed > 1;
1847 * Take the capacity of the CPU into account to
1848 * ensure it fits the requirement of the task.
1850 if (static_branch_unlikely(&sched_asym_cpucapacity))
1851 select_rq |= !dl_task_fits_capacity(p, cpu);
1854 int target = find_later_rq(p);
1857 (dl_time_before(p->dl.deadline,
1858 cpu_rq(target)->dl.earliest_dl.curr) ||
1859 (cpu_rq(target)->dl.dl_nr_running == 0)))
1868 static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
1873 if (READ_ONCE(p->__state) != TASK_WAKING)
1878 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1879 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1880 * rq->lock is not... So, lock it
1883 if (p->dl.dl_non_contending) {
1884 update_rq_clock(rq);
1885 sub_running_bw(&p->dl, &rq->dl);
1886 p->dl.dl_non_contending = 0;
1888 * If the timer handler is currently running and the
1889 * timer cannot be canceled, inactive_task_timer()
1890 * will see that dl_not_contending is not set, and
1891 * will not touch the rq's active utilization,
1892 * so we are still safe.
1894 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
1897 sub_rq_bw(&p->dl, &rq->dl);
1901 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1904 * Current can't be migrated, useless to reschedule,
1905 * let's hope p can move out.
1907 if (rq->curr->nr_cpus_allowed == 1 ||
1908 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
1912 * p is migratable, so let's not schedule it and
1913 * see if it is pushed or pulled somewhere else.
1915 if (p->nr_cpus_allowed != 1 &&
1916 cpudl_find(&rq->rd->cpudl, p, NULL))
1922 static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1924 if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) {
1926 * This is OK, because current is on_cpu, which avoids it being
1927 * picked for load-balance and preemption/IRQs are still
1928 * disabled avoiding further scheduler activity on it and we've
1929 * not yet started the picking loop.
1931 rq_unpin_lock(rq, rf);
1933 rq_repin_lock(rq, rf);
1936 return sched_stop_runnable(rq) || sched_dl_runnable(rq);
1938 #endif /* CONFIG_SMP */
1941 * Only called when both the current and waking task are -deadline
1944 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1947 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
1954 * In the unlikely case current and p have the same deadline
1955 * let us try to decide what's the best thing to do...
1957 if ((p->dl.deadline == rq->curr->dl.deadline) &&
1958 !test_tsk_need_resched(rq->curr))
1959 check_preempt_equal_dl(rq, p);
1960 #endif /* CONFIG_SMP */
1963 #ifdef CONFIG_SCHED_HRTICK
1964 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1966 hrtick_start(rq, p->dl.runtime);
1968 #else /* !CONFIG_SCHED_HRTICK */
1969 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1974 static void set_next_task_dl(struct rq *rq, struct task_struct *p, bool first)
1976 struct sched_dl_entity *dl_se = &p->dl;
1977 struct dl_rq *dl_rq = &rq->dl;
1979 p->se.exec_start = rq_clock_task(rq);
1980 if (on_dl_rq(&p->dl))
1981 update_stats_wait_end_dl(dl_rq, dl_se);
1983 /* You can't push away the running task */
1984 dequeue_pushable_dl_task(rq, p);
1989 if (hrtick_enabled_dl(rq))
1990 start_hrtick_dl(rq, p);
1992 if (rq->curr->sched_class != &dl_sched_class)
1993 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
1995 deadline_queue_push_tasks(rq);
1998 static struct sched_dl_entity *pick_next_dl_entity(struct dl_rq *dl_rq)
2000 struct rb_node *left = rb_first_cached(&dl_rq->root);
2005 return __node_2_dle(left);
2008 static struct task_struct *pick_task_dl(struct rq *rq)
2010 struct sched_dl_entity *dl_se;
2011 struct dl_rq *dl_rq = &rq->dl;
2012 struct task_struct *p;
2014 if (!sched_dl_runnable(rq))
2017 dl_se = pick_next_dl_entity(dl_rq);
2019 p = dl_task_of(dl_se);
2024 static struct task_struct *pick_next_task_dl(struct rq *rq)
2026 struct task_struct *p;
2028 p = pick_task_dl(rq);
2030 set_next_task_dl(rq, p, true);
2035 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
2037 struct sched_dl_entity *dl_se = &p->dl;
2038 struct dl_rq *dl_rq = &rq->dl;
2040 if (on_dl_rq(&p->dl))
2041 update_stats_wait_start_dl(dl_rq, dl_se);
2045 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2046 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
2047 enqueue_pushable_dl_task(rq, p);
2051 * scheduler tick hitting a task of our scheduling class.
2053 * NOTE: This function can be called remotely by the tick offload that
2054 * goes along full dynticks. Therefore no local assumption can be made
2055 * and everything must be accessed through the @rq and @curr passed in
2058 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
2062 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2064 * Even when we have runtime, update_curr_dl() might have resulted in us
2065 * not being the leftmost task anymore. In that case NEED_RESCHED will
2066 * be set and schedule() will start a new hrtick for the next task.
2068 if (hrtick_enabled_dl(rq) && queued && p->dl.runtime > 0 &&
2069 is_leftmost(p, &rq->dl))
2070 start_hrtick_dl(rq, p);
2073 static void task_fork_dl(struct task_struct *p)
2076 * SCHED_DEADLINE tasks cannot fork and this is achieved through
2083 /* Only try algorithms three times */
2084 #define DL_MAX_TRIES 3
2086 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
2088 if (!task_running(rq, p) &&
2089 cpumask_test_cpu(cpu, &p->cpus_mask))
2095 * Return the earliest pushable rq's task, which is suitable to be executed
2096 * on the CPU, NULL otherwise:
2098 static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
2100 struct task_struct *p = NULL;
2101 struct rb_node *next_node;
2103 if (!has_pushable_dl_tasks(rq))
2106 next_node = rb_first_cached(&rq->dl.pushable_dl_tasks_root);
2110 p = __node_2_pdl(next_node);
2112 if (pick_dl_task(rq, p, cpu))
2115 next_node = rb_next(next_node);
2122 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
2124 static int find_later_rq(struct task_struct *task)
2126 struct sched_domain *sd;
2127 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
2128 int this_cpu = smp_processor_id();
2129 int cpu = task_cpu(task);
2131 /* Make sure the mask is initialized first */
2132 if (unlikely(!later_mask))
2135 if (task->nr_cpus_allowed == 1)
2139 * We have to consider system topology and task affinity
2140 * first, then we can look for a suitable CPU.
2142 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
2146 * If we are here, some targets have been found, including
2147 * the most suitable which is, among the runqueues where the
2148 * current tasks have later deadlines than the task's one, the
2149 * rq with the latest possible one.
2151 * Now we check how well this matches with task's
2152 * affinity and system topology.
2154 * The last CPU where the task run is our first
2155 * guess, since it is most likely cache-hot there.
2157 if (cpumask_test_cpu(cpu, later_mask))
2160 * Check if this_cpu is to be skipped (i.e., it is
2161 * not in the mask) or not.
2163 if (!cpumask_test_cpu(this_cpu, later_mask))
2167 for_each_domain(cpu, sd) {
2168 if (sd->flags & SD_WAKE_AFFINE) {
2172 * If possible, preempting this_cpu is
2173 * cheaper than migrating.
2175 if (this_cpu != -1 &&
2176 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
2181 best_cpu = cpumask_any_and_distribute(later_mask,
2182 sched_domain_span(sd));
2184 * Last chance: if a CPU being in both later_mask
2185 * and current sd span is valid, that becomes our
2186 * choice. Of course, the latest possible CPU is
2187 * already under consideration through later_mask.
2189 if (best_cpu < nr_cpu_ids) {
2198 * At this point, all our guesses failed, we just return
2199 * 'something', and let the caller sort the things out.
2204 cpu = cpumask_any_distribute(later_mask);
2205 if (cpu < nr_cpu_ids)
2211 /* Locks the rq it finds */
2212 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
2214 struct rq *later_rq = NULL;
2218 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
2219 cpu = find_later_rq(task);
2221 if ((cpu == -1) || (cpu == rq->cpu))
2224 later_rq = cpu_rq(cpu);
2226 if (later_rq->dl.dl_nr_running &&
2227 !dl_time_before(task->dl.deadline,
2228 later_rq->dl.earliest_dl.curr)) {
2230 * Target rq has tasks of equal or earlier deadline,
2231 * retrying does not release any lock and is unlikely
2232 * to yield a different result.
2238 /* Retry if something changed. */
2239 if (double_lock_balance(rq, later_rq)) {
2240 if (unlikely(task_rq(task) != rq ||
2241 !cpumask_test_cpu(later_rq->cpu, &task->cpus_mask) ||
2242 task_running(rq, task) ||
2244 !task_on_rq_queued(task))) {
2245 double_unlock_balance(rq, later_rq);
2252 * If the rq we found has no -deadline task, or
2253 * its earliest one has a later deadline than our
2254 * task, the rq is a good one.
2256 if (!later_rq->dl.dl_nr_running ||
2257 dl_time_before(task->dl.deadline,
2258 later_rq->dl.earliest_dl.curr))
2261 /* Otherwise we try again. */
2262 double_unlock_balance(rq, later_rq);
2269 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2271 struct task_struct *p;
2273 if (!has_pushable_dl_tasks(rq))
2276 p = __node_2_pdl(rb_first_cached(&rq->dl.pushable_dl_tasks_root));
2278 BUG_ON(rq->cpu != task_cpu(p));
2279 BUG_ON(task_current(rq, p));
2280 BUG_ON(p->nr_cpus_allowed <= 1);
2282 BUG_ON(!task_on_rq_queued(p));
2283 BUG_ON(!dl_task(p));
2289 * See if the non running -deadline tasks on this rq
2290 * can be sent to some other CPU where they can preempt
2291 * and start executing.
2293 static int push_dl_task(struct rq *rq)
2295 struct task_struct *next_task;
2296 struct rq *later_rq;
2299 if (!rq->dl.overloaded)
2302 next_task = pick_next_pushable_dl_task(rq);
2308 * If next_task preempts rq->curr, and rq->curr
2309 * can move away, it makes sense to just reschedule
2310 * without going further in pushing next_task.
2312 if (dl_task(rq->curr) &&
2313 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2314 rq->curr->nr_cpus_allowed > 1) {
2319 if (is_migration_disabled(next_task))
2322 if (WARN_ON(next_task == rq->curr))
2325 /* We might release rq lock */
2326 get_task_struct(next_task);
2328 /* Will lock the rq it'll find */
2329 later_rq = find_lock_later_rq(next_task, rq);
2331 struct task_struct *task;
2334 * We must check all this again, since
2335 * find_lock_later_rq releases rq->lock and it is
2336 * then possible that next_task has migrated.
2338 task = pick_next_pushable_dl_task(rq);
2339 if (task == next_task) {
2341 * The task is still there. We don't try
2342 * again, some other CPU will pull it when ready.
2351 put_task_struct(next_task);
2356 deactivate_task(rq, next_task, 0);
2357 set_task_cpu(next_task, later_rq->cpu);
2358 activate_task(later_rq, next_task, 0);
2361 resched_curr(later_rq);
2363 double_unlock_balance(rq, later_rq);
2366 put_task_struct(next_task);
2371 static void push_dl_tasks(struct rq *rq)
2373 /* push_dl_task() will return true if it moved a -deadline task */
2374 while (push_dl_task(rq))
2378 static void pull_dl_task(struct rq *this_rq)
2380 int this_cpu = this_rq->cpu, cpu;
2381 struct task_struct *p, *push_task;
2382 bool resched = false;
2384 u64 dmin = LONG_MAX;
2386 if (likely(!dl_overloaded(this_rq)))
2390 * Match the barrier from dl_set_overloaded; this guarantees that if we
2391 * see overloaded we must also see the dlo_mask bit.
2395 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2396 if (this_cpu == cpu)
2399 src_rq = cpu_rq(cpu);
2402 * It looks racy, abd it is! However, as in sched_rt.c,
2403 * we are fine with this.
2405 if (this_rq->dl.dl_nr_running &&
2406 dl_time_before(this_rq->dl.earliest_dl.curr,
2407 src_rq->dl.earliest_dl.next))
2410 /* Might drop this_rq->lock */
2412 double_lock_balance(this_rq, src_rq);
2415 * If there are no more pullable tasks on the
2416 * rq, we're done with it.
2418 if (src_rq->dl.dl_nr_running <= 1)
2421 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2424 * We found a task to be pulled if:
2425 * - it preempts our current (if there's one),
2426 * - it will preempt the last one we pulled (if any).
2428 if (p && dl_time_before(p->dl.deadline, dmin) &&
2429 (!this_rq->dl.dl_nr_running ||
2430 dl_time_before(p->dl.deadline,
2431 this_rq->dl.earliest_dl.curr))) {
2432 WARN_ON(p == src_rq->curr);
2433 WARN_ON(!task_on_rq_queued(p));
2436 * Then we pull iff p has actually an earlier
2437 * deadline than the current task of its runqueue.
2439 if (dl_time_before(p->dl.deadline,
2440 src_rq->curr->dl.deadline))
2443 if (is_migration_disabled(p)) {
2444 push_task = get_push_task(src_rq);
2446 deactivate_task(src_rq, p, 0);
2447 set_task_cpu(p, this_cpu);
2448 activate_task(this_rq, p, 0);
2449 dmin = p->dl.deadline;
2453 /* Is there any other task even earlier? */
2456 double_unlock_balance(this_rq, src_rq);
2459 raw_spin_rq_unlock(this_rq);
2460 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2461 push_task, &src_rq->push_work);
2462 raw_spin_rq_lock(this_rq);
2467 resched_curr(this_rq);
2471 * Since the task is not running and a reschedule is not going to happen
2472 * anytime soon on its runqueue, we try pushing it away now.
2474 static void task_woken_dl(struct rq *rq, struct task_struct *p)
2476 if (!task_running(rq, p) &&
2477 !test_tsk_need_resched(rq->curr) &&
2478 p->nr_cpus_allowed > 1 &&
2479 dl_task(rq->curr) &&
2480 (rq->curr->nr_cpus_allowed < 2 ||
2481 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2486 static void set_cpus_allowed_dl(struct task_struct *p,
2487 const struct cpumask *new_mask,
2490 struct root_domain *src_rd;
2493 BUG_ON(!dl_task(p));
2498 * Migrating a SCHED_DEADLINE task between exclusive
2499 * cpusets (different root_domains) entails a bandwidth
2500 * update. We already made space for us in the destination
2501 * domain (see cpuset_can_attach()).
2503 if (!cpumask_intersects(src_rd->span, new_mask)) {
2504 struct dl_bw *src_dl_b;
2506 src_dl_b = dl_bw_of(cpu_of(rq));
2508 * We now free resources of the root_domain we are migrating
2509 * off. In the worst case, sched_setattr() may temporary fail
2510 * until we complete the update.
2512 raw_spin_lock(&src_dl_b->lock);
2513 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2514 raw_spin_unlock(&src_dl_b->lock);
2517 set_cpus_allowed_common(p, new_mask, flags);
2520 /* Assumes rq->lock is held */
2521 static void rq_online_dl(struct rq *rq)
2523 if (rq->dl.overloaded)
2524 dl_set_overload(rq);
2526 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2527 if (rq->dl.dl_nr_running > 0)
2528 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2531 /* Assumes rq->lock is held */
2532 static void rq_offline_dl(struct rq *rq)
2534 if (rq->dl.overloaded)
2535 dl_clear_overload(rq);
2537 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2538 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2541 void __init init_sched_dl_class(void)
2545 for_each_possible_cpu(i)
2546 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2547 GFP_KERNEL, cpu_to_node(i));
2550 void dl_add_task_root_domain(struct task_struct *p)
2556 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2558 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
2562 rq = __task_rq_lock(p, &rf);
2564 dl_b = &rq->rd->dl_bw;
2565 raw_spin_lock(&dl_b->lock);
2567 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2569 raw_spin_unlock(&dl_b->lock);
2571 task_rq_unlock(rq, p, &rf);
2574 void dl_clear_root_domain(struct root_domain *rd)
2576 unsigned long flags;
2578 raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
2579 rd->dl_bw.total_bw = 0;
2580 raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
2583 #endif /* CONFIG_SMP */
2585 static void switched_from_dl(struct rq *rq, struct task_struct *p)
2588 * task_non_contending() can start the "inactive timer" (if the 0-lag
2589 * time is in the future). If the task switches back to dl before
2590 * the "inactive timer" fires, it can continue to consume its current
2591 * runtime using its current deadline. If it stays outside of
2592 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2593 * will reset the task parameters.
2595 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2596 task_non_contending(p);
2598 if (!task_on_rq_queued(p)) {
2600 * Inactive timer is armed. However, p is leaving DEADLINE and
2601 * might migrate away from this rq while continuing to run on
2602 * some other class. We need to remove its contribution from
2603 * this rq running_bw now, or sub_rq_bw (below) will complain.
2605 if (p->dl.dl_non_contending)
2606 sub_running_bw(&p->dl, &rq->dl);
2607 sub_rq_bw(&p->dl, &rq->dl);
2611 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2612 * at the 0-lag time, because the task could have been migrated
2613 * while SCHED_OTHER in the meanwhile.
2615 if (p->dl.dl_non_contending)
2616 p->dl.dl_non_contending = 0;
2619 * Since this might be the only -deadline task on the rq,
2620 * this is the right place to try to pull some other one
2621 * from an overloaded CPU, if any.
2623 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
2626 deadline_queue_pull_task(rq);
2630 * When switching to -deadline, we may overload the rq, then
2631 * we try to push someone off, if possible.
2633 static void switched_to_dl(struct rq *rq, struct task_struct *p)
2635 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2638 /* If p is not queued we will update its parameters at next wakeup. */
2639 if (!task_on_rq_queued(p)) {
2640 add_rq_bw(&p->dl, &rq->dl);
2645 if (rq->curr != p) {
2647 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
2648 deadline_queue_push_tasks(rq);
2650 if (dl_task(rq->curr))
2651 check_preempt_curr_dl(rq, p, 0);
2655 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2660 * If the scheduling parameters of a -deadline task changed,
2661 * a push or pull operation might be needed.
2663 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2666 if (task_on_rq_queued(p) || task_current(rq, p)) {
2669 * This might be too much, but unfortunately
2670 * we don't have the old deadline value, and
2671 * we can't argue if the task is increasing
2672 * or lowering its prio, so...
2674 if (!rq->dl.overloaded)
2675 deadline_queue_pull_task(rq);
2678 * If we now have a earlier deadline task than p,
2679 * then reschedule, provided p is still on this
2682 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
2686 * Again, we don't know if p has a earlier
2687 * or later deadline, so let's blindly set a
2688 * (maybe not needed) rescheduling point.
2691 #endif /* CONFIG_SMP */
2695 DEFINE_SCHED_CLASS(dl) = {
2697 .enqueue_task = enqueue_task_dl,
2698 .dequeue_task = dequeue_task_dl,
2699 .yield_task = yield_task_dl,
2701 .check_preempt_curr = check_preempt_curr_dl,
2703 .pick_next_task = pick_next_task_dl,
2704 .put_prev_task = put_prev_task_dl,
2705 .set_next_task = set_next_task_dl,
2708 .balance = balance_dl,
2709 .pick_task = pick_task_dl,
2710 .select_task_rq = select_task_rq_dl,
2711 .migrate_task_rq = migrate_task_rq_dl,
2712 .set_cpus_allowed = set_cpus_allowed_dl,
2713 .rq_online = rq_online_dl,
2714 .rq_offline = rq_offline_dl,
2715 .task_woken = task_woken_dl,
2716 .find_lock_rq = find_lock_later_rq,
2719 .task_tick = task_tick_dl,
2720 .task_fork = task_fork_dl,
2722 .prio_changed = prio_changed_dl,
2723 .switched_from = switched_from_dl,
2724 .switched_to = switched_to_dl,
2726 .update_curr = update_curr_dl,
2729 /* Used for dl_bw check and update, used under sched_rt_handler()::mutex */
2730 static u64 dl_generation;
2732 int sched_dl_global_validate(void)
2734 u64 runtime = global_rt_runtime();
2735 u64 period = global_rt_period();
2736 u64 new_bw = to_ratio(period, runtime);
2737 u64 gen = ++dl_generation;
2739 int cpu, cpus, ret = 0;
2740 unsigned long flags;
2743 * Here we want to check the bandwidth not being set to some
2744 * value smaller than the currently allocated bandwidth in
2745 * any of the root_domains.
2747 for_each_possible_cpu(cpu) {
2748 rcu_read_lock_sched();
2750 if (dl_bw_visited(cpu, gen))
2753 dl_b = dl_bw_of(cpu);
2754 cpus = dl_bw_cpus(cpu);
2756 raw_spin_lock_irqsave(&dl_b->lock, flags);
2757 if (new_bw * cpus < dl_b->total_bw)
2759 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2762 rcu_read_unlock_sched();
2771 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
2773 if (global_rt_runtime() == RUNTIME_INF) {
2774 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2775 dl_rq->extra_bw = 1 << BW_SHIFT;
2777 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2778 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2779 dl_rq->extra_bw = to_ratio(global_rt_period(),
2780 global_rt_runtime());
2784 void sched_dl_do_global(void)
2787 u64 gen = ++dl_generation;
2790 unsigned long flags;
2792 if (global_rt_runtime() != RUNTIME_INF)
2793 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2795 for_each_possible_cpu(cpu) {
2796 rcu_read_lock_sched();
2798 if (dl_bw_visited(cpu, gen)) {
2799 rcu_read_unlock_sched();
2803 dl_b = dl_bw_of(cpu);
2805 raw_spin_lock_irqsave(&dl_b->lock, flags);
2807 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2809 rcu_read_unlock_sched();
2810 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2815 * We must be sure that accepting a new task (or allowing changing the
2816 * parameters of an existing one) is consistent with the bandwidth
2817 * constraints. If yes, this function also accordingly updates the currently
2818 * allocated bandwidth to reflect the new situation.
2820 * This function is called while holding p's rq->lock.
2822 int sched_dl_overflow(struct task_struct *p, int policy,
2823 const struct sched_attr *attr)
2825 u64 period = attr->sched_period ?: attr->sched_deadline;
2826 u64 runtime = attr->sched_runtime;
2827 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2828 int cpus, err = -1, cpu = task_cpu(p);
2829 struct dl_bw *dl_b = dl_bw_of(cpu);
2832 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2835 /* !deadline task may carry old deadline bandwidth */
2836 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2840 * Either if a task, enters, leave, or stays -deadline but changes
2841 * its parameters, we may need to update accordingly the total
2842 * allocated bandwidth of the container.
2844 raw_spin_lock(&dl_b->lock);
2845 cpus = dl_bw_cpus(cpu);
2846 cap = dl_bw_capacity(cpu);
2848 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2849 !__dl_overflow(dl_b, cap, 0, new_bw)) {
2850 if (hrtimer_active(&p->dl.inactive_timer))
2851 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2852 __dl_add(dl_b, new_bw, cpus);
2854 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2855 !__dl_overflow(dl_b, cap, p->dl.dl_bw, new_bw)) {
2857 * XXX this is slightly incorrect: when the task
2858 * utilization decreases, we should delay the total
2859 * utilization change until the task's 0-lag point.
2860 * But this would require to set the task's "inactive
2861 * timer" when the task is not inactive.
2863 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2864 __dl_add(dl_b, new_bw, cpus);
2865 dl_change_utilization(p, new_bw);
2867 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2869 * Do not decrease the total deadline utilization here,
2870 * switched_from_dl() will take care to do it at the correct
2875 raw_spin_unlock(&dl_b->lock);
2881 * This function initializes the sched_dl_entity of a newly becoming
2882 * SCHED_DEADLINE task.
2884 * Only the static values are considered here, the actual runtime and the
2885 * absolute deadline will be properly calculated when the task is enqueued
2886 * for the first time with its new policy.
2888 void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
2890 struct sched_dl_entity *dl_se = &p->dl;
2892 dl_se->dl_runtime = attr->sched_runtime;
2893 dl_se->dl_deadline = attr->sched_deadline;
2894 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
2895 dl_se->flags = attr->sched_flags & SCHED_DL_FLAGS;
2896 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
2897 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
2900 void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
2902 struct sched_dl_entity *dl_se = &p->dl;
2904 attr->sched_priority = p->rt_priority;
2905 attr->sched_runtime = dl_se->dl_runtime;
2906 attr->sched_deadline = dl_se->dl_deadline;
2907 attr->sched_period = dl_se->dl_period;
2908 attr->sched_flags &= ~SCHED_DL_FLAGS;
2909 attr->sched_flags |= dl_se->flags;
2913 * This function validates the new parameters of a -deadline task.
2914 * We ask for the deadline not being zero, and greater or equal
2915 * than the runtime, as well as the period of being zero or
2916 * greater than deadline. Furthermore, we have to be sure that
2917 * user parameters are above the internal resolution of 1us (we
2918 * check sched_runtime only since it is always the smaller one) and
2919 * below 2^63 ns (we have to check both sched_deadline and
2920 * sched_period, as the latter can be zero).
2922 bool __checkparam_dl(const struct sched_attr *attr)
2924 u64 period, max, min;
2926 /* special dl tasks don't actually use any parameter */
2927 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2931 if (attr->sched_deadline == 0)
2935 * Since we truncate DL_SCALE bits, make sure we're at least
2938 if (attr->sched_runtime < (1ULL << DL_SCALE))
2942 * Since we use the MSB for wrap-around and sign issues, make
2943 * sure it's not set (mind that period can be equal to zero).
2945 if (attr->sched_deadline & (1ULL << 63) ||
2946 attr->sched_period & (1ULL << 63))
2949 period = attr->sched_period;
2951 period = attr->sched_deadline;
2953 /* runtime <= deadline <= period (if period != 0) */
2954 if (period < attr->sched_deadline ||
2955 attr->sched_deadline < attr->sched_runtime)
2958 max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
2959 min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
2961 if (period < min || period > max)
2968 * This function clears the sched_dl_entity static params.
2970 void __dl_clear_params(struct task_struct *p)
2972 struct sched_dl_entity *dl_se = &p->dl;
2974 dl_se->dl_runtime = 0;
2975 dl_se->dl_deadline = 0;
2976 dl_se->dl_period = 0;
2979 dl_se->dl_density = 0;
2981 dl_se->dl_throttled = 0;
2982 dl_se->dl_yielded = 0;
2983 dl_se->dl_non_contending = 0;
2984 dl_se->dl_overrun = 0;
2986 #ifdef CONFIG_RT_MUTEXES
2987 dl_se->pi_se = dl_se;
2991 bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
2993 struct sched_dl_entity *dl_se = &p->dl;
2995 if (dl_se->dl_runtime != attr->sched_runtime ||
2996 dl_se->dl_deadline != attr->sched_deadline ||
2997 dl_se->dl_period != attr->sched_period ||
2998 dl_se->flags != (attr->sched_flags & SCHED_DL_FLAGS))
3005 int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
3006 const struct cpumask *trial)
3008 int ret = 1, trial_cpus;
3009 struct dl_bw *cur_dl_b;
3010 unsigned long flags;
3012 rcu_read_lock_sched();
3013 cur_dl_b = dl_bw_of(cpumask_any(cur));
3014 trial_cpus = cpumask_weight(trial);
3016 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
3017 if (cur_dl_b->bw != -1 &&
3018 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
3020 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
3021 rcu_read_unlock_sched();
3026 int dl_cpu_busy(int cpu, struct task_struct *p)
3028 unsigned long flags, cap;
3032 rcu_read_lock_sched();
3033 dl_b = dl_bw_of(cpu);
3034 raw_spin_lock_irqsave(&dl_b->lock, flags);
3035 cap = dl_bw_capacity(cpu);
3036 overflow = __dl_overflow(dl_b, cap, 0, p ? p->dl.dl_bw : 0);
3038 if (!overflow && p) {
3040 * We reserve space for this task in the destination
3041 * root_domain, as we can't fail after this point.
3042 * We will free resources in the source root_domain
3043 * later on (see set_cpus_allowed_dl()).
3045 __dl_add(dl_b, p->dl.dl_bw, dl_bw_cpus(cpu));
3048 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3049 rcu_read_unlock_sched();
3051 return overflow ? -EBUSY : 0;
3055 #ifdef CONFIG_SCHED_DEBUG
3056 void print_dl_stats(struct seq_file *m, int cpu)
3058 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
3060 #endif /* CONFIG_SCHED_DEBUG */