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 inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
59 return container_of(dl_se, struct task_struct, dl);
62 static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
64 return container_of(dl_rq, struct rq, dl);
67 static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
69 struct task_struct *p = dl_task_of(dl_se);
70 struct rq *rq = task_rq(p);
75 static inline int on_dl_rq(struct sched_dl_entity *dl_se)
77 return !RB_EMPTY_NODE(&dl_se->rb_node);
80 #ifdef CONFIG_RT_MUTEXES
81 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
86 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
88 return pi_of(dl_se) != dl_se;
91 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
96 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
103 static inline struct dl_bw *dl_bw_of(int i)
105 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
106 "sched RCU must be held");
107 return &cpu_rq(i)->rd->dl_bw;
110 static inline int dl_bw_cpus(int i)
112 struct root_domain *rd = cpu_rq(i)->rd;
115 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
116 "sched RCU must be held");
118 if (cpumask_subset(rd->span, cpu_active_mask))
119 return cpumask_weight(rd->span);
123 for_each_cpu_and(i, rd->span, cpu_active_mask)
129 static inline unsigned long __dl_bw_capacity(const struct cpumask *mask)
131 unsigned long cap = 0;
134 for_each_cpu_and(i, mask, cpu_active_mask)
135 cap += capacity_orig_of(i);
141 * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
142 * of the CPU the task is running on rather rd's \Sum CPU capacity.
144 static inline unsigned long dl_bw_capacity(int i)
146 if (!sched_asym_cpucap_active() &&
147 capacity_orig_of(i) == SCHED_CAPACITY_SCALE) {
148 return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT;
150 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
151 "sched RCU must be held");
153 return __dl_bw_capacity(cpu_rq(i)->rd->span);
157 static inline bool dl_bw_visited(int cpu, u64 gen)
159 struct root_domain *rd = cpu_rq(cpu)->rd;
161 if (rd->visit_gen == gen)
169 void __dl_update(struct dl_bw *dl_b, s64 bw)
171 struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw);
174 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
175 "sched RCU must be held");
176 for_each_cpu_and(i, rd->span, cpu_active_mask) {
177 struct rq *rq = cpu_rq(i);
179 rq->dl.extra_bw += bw;
183 static inline struct dl_bw *dl_bw_of(int i)
185 return &cpu_rq(i)->dl.dl_bw;
188 static inline int dl_bw_cpus(int i)
193 static inline unsigned long dl_bw_capacity(int i)
195 return SCHED_CAPACITY_SCALE;
198 static inline bool dl_bw_visited(int cpu, u64 gen)
204 void __dl_update(struct dl_bw *dl_b, s64 bw)
206 struct dl_rq *dl = container_of(dl_b, struct dl_rq, dl_bw);
213 void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
215 dl_b->total_bw -= tsk_bw;
216 __dl_update(dl_b, (s32)tsk_bw / cpus);
220 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
222 dl_b->total_bw += tsk_bw;
223 __dl_update(dl_b, -((s32)tsk_bw / cpus));
227 __dl_overflow(struct dl_bw *dl_b, unsigned long cap, u64 old_bw, u64 new_bw)
229 return dl_b->bw != -1 &&
230 cap_scale(dl_b->bw, cap) < dl_b->total_bw - old_bw + new_bw;
234 void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
236 u64 old = dl_rq->running_bw;
238 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
239 dl_rq->running_bw += dl_bw;
240 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
241 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
242 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
243 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
247 void __sub_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); /* underflow */
254 if (dl_rq->running_bw > old)
255 dl_rq->running_bw = 0;
256 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
257 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
261 void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
263 u64 old = dl_rq->this_bw;
265 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
266 dl_rq->this_bw += dl_bw;
267 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
271 void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
273 u64 old = dl_rq->this_bw;
275 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
276 dl_rq->this_bw -= dl_bw;
277 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
278 if (dl_rq->this_bw > old)
280 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
284 void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
286 if (!dl_entity_is_special(dl_se))
287 __add_rq_bw(dl_se->dl_bw, dl_rq);
291 void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
293 if (!dl_entity_is_special(dl_se))
294 __sub_rq_bw(dl_se->dl_bw, dl_rq);
298 void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
300 if (!dl_entity_is_special(dl_se))
301 __add_running_bw(dl_se->dl_bw, dl_rq);
305 void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
307 if (!dl_entity_is_special(dl_se))
308 __sub_running_bw(dl_se->dl_bw, dl_rq);
311 static void dl_change_utilization(struct task_struct *p, u64 new_bw)
315 WARN_ON_ONCE(p->dl.flags & SCHED_FLAG_SUGOV);
317 if (task_on_rq_queued(p))
321 if (p->dl.dl_non_contending) {
322 sub_running_bw(&p->dl, &rq->dl);
323 p->dl.dl_non_contending = 0;
325 * If the timer handler is currently running and the
326 * timer cannot be canceled, inactive_task_timer()
327 * will see that dl_not_contending is not set, and
328 * will not touch the rq's active utilization,
329 * so we are still safe.
331 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
334 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
335 __add_rq_bw(new_bw, &rq->dl);
339 * The utilization of a task cannot be immediately removed from
340 * the rq active utilization (running_bw) when the task blocks.
341 * Instead, we have to wait for the so called "0-lag time".
343 * If a task blocks before the "0-lag time", a timer (the inactive
344 * timer) is armed, and running_bw is decreased when the timer
347 * If the task wakes up again before the inactive timer fires,
348 * the timer is canceled, whereas if the task wakes up after the
349 * inactive timer fired (and running_bw has been decreased) the
350 * task's utilization has to be added to running_bw again.
351 * A flag in the deadline scheduling entity (dl_non_contending)
352 * is used to avoid race conditions between the inactive timer handler
355 * The following diagram shows how running_bw is updated. A task is
356 * "ACTIVE" when its utilization contributes to running_bw; an
357 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
358 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
359 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
360 * time already passed, which does not contribute to running_bw anymore.
361 * +------------------+
363 * +------------------>+ contending |
364 * | add_running_bw | |
365 * | +----+------+------+
368 * +--------+-------+ | |
369 * | | t >= 0-lag | | wakeup
370 * | INACTIVE |<---------------+ |
371 * | | sub_running_bw | |
372 * +--------+-------+ | |
377 * | +----+------+------+
378 * | sub_running_bw | ACTIVE |
379 * +-------------------+ |
380 * inactive timer | non contending |
381 * fired +------------------+
383 * The task_non_contending() function is invoked when a task
384 * blocks, and checks if the 0-lag time already passed or
385 * not (in the first case, it directly updates running_bw;
386 * in the second case, it arms the inactive timer).
388 * The task_contending() function is invoked when a task wakes
389 * up, and checks if the task is still in the "ACTIVE non contending"
390 * state or not (in the second case, it updates running_bw).
392 static void task_non_contending(struct task_struct *p)
394 struct sched_dl_entity *dl_se = &p->dl;
395 struct hrtimer *timer = &dl_se->inactive_timer;
396 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
397 struct rq *rq = rq_of_dl_rq(dl_rq);
401 * If this is a non-deadline task that has been boosted,
404 if (dl_se->dl_runtime == 0)
407 if (dl_entity_is_special(dl_se))
410 WARN_ON(dl_se->dl_non_contending);
412 zerolag_time = dl_se->deadline -
413 div64_long((dl_se->runtime * dl_se->dl_period),
417 * Using relative times instead of the absolute "0-lag time"
418 * allows to simplify the code
420 zerolag_time -= rq_clock(rq);
423 * If the "0-lag time" already passed, decrease the active
424 * utilization now, instead of starting a timer
426 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
428 sub_running_bw(dl_se, dl_rq);
429 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
430 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
432 if (READ_ONCE(p->__state) == TASK_DEAD)
433 sub_rq_bw(&p->dl, &rq->dl);
434 raw_spin_lock(&dl_b->lock);
435 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
436 raw_spin_unlock(&dl_b->lock);
437 __dl_clear_params(p);
443 dl_se->dl_non_contending = 1;
445 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD);
448 static void task_contending(struct sched_dl_entity *dl_se, int flags)
450 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
453 * If this is a non-deadline task that has been boosted,
456 if (dl_se->dl_runtime == 0)
459 if (flags & ENQUEUE_MIGRATED)
460 add_rq_bw(dl_se, dl_rq);
462 if (dl_se->dl_non_contending) {
463 dl_se->dl_non_contending = 0;
465 * If the timer handler is currently running and the
466 * timer cannot be canceled, inactive_task_timer()
467 * will see that dl_not_contending is not set, and
468 * will not touch the rq's active utilization,
469 * so we are still safe.
471 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1)
472 put_task_struct(dl_task_of(dl_se));
475 * Since "dl_non_contending" is not set, the
476 * task's utilization has already been removed from
477 * active utilization (either when the task blocked,
478 * when the "inactive timer" fired).
481 add_running_bw(dl_se, dl_rq);
485 static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
487 struct sched_dl_entity *dl_se = &p->dl;
489 return rb_first_cached(&dl_rq->root) == &dl_se->rb_node;
492 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
494 void init_dl_bw(struct dl_bw *dl_b)
496 raw_spin_lock_init(&dl_b->lock);
497 if (global_rt_runtime() == RUNTIME_INF)
500 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
504 void init_dl_rq(struct dl_rq *dl_rq)
506 dl_rq->root = RB_ROOT_CACHED;
509 /* zero means no -deadline tasks */
510 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
512 dl_rq->dl_nr_migratory = 0;
513 dl_rq->overloaded = 0;
514 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
516 init_dl_bw(&dl_rq->dl_bw);
519 dl_rq->running_bw = 0;
521 init_dl_rq_bw_ratio(dl_rq);
526 static inline int dl_overloaded(struct rq *rq)
528 return atomic_read(&rq->rd->dlo_count);
531 static inline void dl_set_overload(struct rq *rq)
536 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
538 * Must be visible before the overload count is
539 * set (as in sched_rt.c).
541 * Matched by the barrier in pull_dl_task().
544 atomic_inc(&rq->rd->dlo_count);
547 static inline void dl_clear_overload(struct rq *rq)
552 atomic_dec(&rq->rd->dlo_count);
553 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
556 static void update_dl_migration(struct dl_rq *dl_rq)
558 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
559 if (!dl_rq->overloaded) {
560 dl_set_overload(rq_of_dl_rq(dl_rq));
561 dl_rq->overloaded = 1;
563 } else if (dl_rq->overloaded) {
564 dl_clear_overload(rq_of_dl_rq(dl_rq));
565 dl_rq->overloaded = 0;
569 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
571 struct task_struct *p = dl_task_of(dl_se);
573 if (p->nr_cpus_allowed > 1)
574 dl_rq->dl_nr_migratory++;
576 update_dl_migration(dl_rq);
579 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
581 struct task_struct *p = dl_task_of(dl_se);
583 if (p->nr_cpus_allowed > 1)
584 dl_rq->dl_nr_migratory--;
586 update_dl_migration(dl_rq);
589 #define __node_2_pdl(node) \
590 rb_entry((node), struct task_struct, pushable_dl_tasks)
592 static inline bool __pushable_less(struct rb_node *a, const struct rb_node *b)
594 return dl_entity_preempt(&__node_2_pdl(a)->dl, &__node_2_pdl(b)->dl);
598 * The list of pushable -deadline task is not a plist, like in
599 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
601 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
603 struct rb_node *leftmost;
605 WARN_ON_ONCE(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
607 leftmost = rb_add_cached(&p->pushable_dl_tasks,
608 &rq->dl.pushable_dl_tasks_root,
611 rq->dl.earliest_dl.next = p->dl.deadline;
614 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
616 struct dl_rq *dl_rq = &rq->dl;
617 struct rb_root_cached *root = &dl_rq->pushable_dl_tasks_root;
618 struct rb_node *leftmost;
620 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
623 leftmost = rb_erase_cached(&p->pushable_dl_tasks, root);
625 dl_rq->earliest_dl.next = __node_2_pdl(leftmost)->dl.deadline;
627 RB_CLEAR_NODE(&p->pushable_dl_tasks);
630 static inline int has_pushable_dl_tasks(struct rq *rq)
632 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
635 static int push_dl_task(struct rq *rq);
637 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
639 return rq->online && dl_task(prev);
642 static DEFINE_PER_CPU(struct balance_callback, dl_push_head);
643 static DEFINE_PER_CPU(struct balance_callback, dl_pull_head);
645 static void push_dl_tasks(struct rq *);
646 static void pull_dl_task(struct rq *);
648 static inline void deadline_queue_push_tasks(struct rq *rq)
650 if (!has_pushable_dl_tasks(rq))
653 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
656 static inline void deadline_queue_pull_task(struct rq *rq)
658 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
661 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
663 static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
665 struct rq *later_rq = NULL;
668 later_rq = find_lock_later_rq(p, rq);
673 * If we cannot preempt any rq, fall back to pick any
676 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
677 if (cpu >= nr_cpu_ids) {
679 * Failed to find any suitable CPU.
680 * The task will never come back!
682 WARN_ON_ONCE(dl_bandwidth_enabled());
685 * If admission control is disabled we
686 * try a little harder to let the task
689 cpu = cpumask_any(cpu_active_mask);
691 later_rq = cpu_rq(cpu);
692 double_lock_balance(rq, later_rq);
695 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
697 * Inactive timer is armed (or callback is running, but
698 * waiting for us to release rq locks). In any case, when it
699 * will fire (or continue), it will see running_bw of this
700 * task migrated to later_rq (and correctly handle it).
702 sub_running_bw(&p->dl, &rq->dl);
703 sub_rq_bw(&p->dl, &rq->dl);
705 add_rq_bw(&p->dl, &later_rq->dl);
706 add_running_bw(&p->dl, &later_rq->dl);
708 sub_rq_bw(&p->dl, &rq->dl);
709 add_rq_bw(&p->dl, &later_rq->dl);
713 * And we finally need to fixup root_domain(s) bandwidth accounting,
714 * since p is still hanging out in the old (now moved to default) root
717 dl_b = &rq->rd->dl_bw;
718 raw_spin_lock(&dl_b->lock);
719 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
720 raw_spin_unlock(&dl_b->lock);
722 dl_b = &later_rq->rd->dl_bw;
723 raw_spin_lock(&dl_b->lock);
724 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
725 raw_spin_unlock(&dl_b->lock);
727 set_task_cpu(p, later_rq->cpu);
728 double_unlock_balance(later_rq, rq);
736 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
741 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
746 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
751 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
755 static inline void deadline_queue_push_tasks(struct rq *rq)
759 static inline void deadline_queue_pull_task(struct rq *rq)
762 #endif /* CONFIG_SMP */
764 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
765 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
766 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags);
768 static inline void replenish_dl_new_period(struct sched_dl_entity *dl_se,
771 /* for non-boosted task, pi_of(dl_se) == dl_se */
772 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
773 dl_se->runtime = pi_of(dl_se)->dl_runtime;
777 * We are being explicitly informed that a new instance is starting,
778 * and this means that:
779 * - the absolute deadline of the entity has to be placed at
780 * current time + relative deadline;
781 * - the runtime of the entity has to be set to the maximum value.
783 * The capability of specifying such event is useful whenever a -deadline
784 * entity wants to (try to!) synchronize its behaviour with the scheduler's
785 * one, and to (try to!) reconcile itself with its own scheduling
788 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
790 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
791 struct rq *rq = rq_of_dl_rq(dl_rq);
793 WARN_ON(is_dl_boosted(dl_se));
794 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
797 * We are racing with the deadline timer. So, do nothing because
798 * the deadline timer handler will take care of properly recharging
799 * the runtime and postponing the deadline
801 if (dl_se->dl_throttled)
805 * We use the regular wall clock time to set deadlines in the
806 * future; in fact, we must consider execution overheads (time
807 * spent on hardirq context, etc.).
809 replenish_dl_new_period(dl_se, rq);
813 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
814 * possibility of a entity lasting more than what it declared, and thus
815 * exhausting its runtime.
817 * Here we are interested in making runtime overrun possible, but we do
818 * not want a entity which is misbehaving to affect the scheduling of all
820 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
821 * is used, in order to confine each entity within its own bandwidth.
823 * This function deals exactly with that, and ensures that when the runtime
824 * of a entity is replenished, its deadline is also postponed. That ensures
825 * the overrunning entity can't interfere with other entity in the system and
826 * can't make them miss their deadlines. Reasons why this kind of overruns
827 * could happen are, typically, a entity voluntarily trying to overcome its
828 * runtime, or it just underestimated it during sched_setattr().
830 static void replenish_dl_entity(struct sched_dl_entity *dl_se)
832 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
833 struct rq *rq = rq_of_dl_rq(dl_rq);
835 WARN_ON_ONCE(pi_of(dl_se)->dl_runtime <= 0);
838 * This could be the case for a !-dl task that is boosted.
839 * Just go with full inherited parameters.
841 if (dl_se->dl_deadline == 0)
842 replenish_dl_new_period(dl_se, rq);
844 if (dl_se->dl_yielded && dl_se->runtime > 0)
848 * We keep moving the deadline away until we get some
849 * available runtime for the entity. This ensures correct
850 * handling of situations where the runtime overrun is
853 while (dl_se->runtime <= 0) {
854 dl_se->deadline += pi_of(dl_se)->dl_period;
855 dl_se->runtime += pi_of(dl_se)->dl_runtime;
859 * At this point, the deadline really should be "in
860 * the future" with respect to rq->clock. If it's
861 * not, we are, for some reason, lagging too much!
862 * Anyway, after having warn userspace abut that,
863 * we still try to keep the things running by
864 * resetting the deadline and the budget of the
867 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
868 printk_deferred_once("sched: DL replenish lagged too much\n");
869 replenish_dl_new_period(dl_se, rq);
872 if (dl_se->dl_yielded)
873 dl_se->dl_yielded = 0;
874 if (dl_se->dl_throttled)
875 dl_se->dl_throttled = 0;
879 * Here we check if --at time t-- an entity (which is probably being
880 * [re]activated or, in general, enqueued) can use its remaining runtime
881 * and its current deadline _without_ exceeding the bandwidth it is
882 * assigned (function returns true if it can't). We are in fact applying
883 * one of the CBS rules: when a task wakes up, if the residual runtime
884 * over residual deadline fits within the allocated bandwidth, then we
885 * can keep the current (absolute) deadline and residual budget without
886 * disrupting the schedulability of the system. Otherwise, we should
887 * refill the runtime and set the deadline a period in the future,
888 * because keeping the current (absolute) deadline of the task would
889 * result in breaking guarantees promised to other tasks (refer to
890 * Documentation/scheduler/sched-deadline.rst for more information).
892 * This function returns true if:
894 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
896 * IOW we can't recycle current parameters.
898 * Notice that the bandwidth check is done against the deadline. For
899 * task with deadline equal to period this is the same of using
900 * dl_period instead of dl_deadline in the equation above.
902 static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t)
907 * left and right are the two sides of the equation above,
908 * after a bit of shuffling to use multiplications instead
911 * Note that none of the time values involved in the two
912 * multiplications are absolute: dl_deadline and dl_runtime
913 * are the relative deadline and the maximum runtime of each
914 * instance, runtime is the runtime left for the last instance
915 * and (deadline - t), since t is rq->clock, is the time left
916 * to the (absolute) deadline. Even if overflowing the u64 type
917 * is very unlikely to occur in both cases, here we scale down
918 * as we want to avoid that risk at all. Scaling down by 10
919 * means that we reduce granularity to 1us. We are fine with it,
920 * since this is only a true/false check and, anyway, thinking
921 * of anything below microseconds resolution is actually fiction
922 * (but still we want to give the user that illusion >;).
924 left = (pi_of(dl_se)->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
925 right = ((dl_se->deadline - t) >> DL_SCALE) *
926 (pi_of(dl_se)->dl_runtime >> DL_SCALE);
928 return dl_time_before(right, left);
932 * Revised wakeup rule [1]: For self-suspending tasks, rather then
933 * re-initializing task's runtime and deadline, the revised wakeup
934 * rule adjusts the task's runtime to avoid the task to overrun its
937 * Reasoning: a task may overrun the density if:
938 * runtime / (deadline - t) > dl_runtime / dl_deadline
940 * Therefore, runtime can be adjusted to:
941 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
943 * In such way that runtime will be equal to the maximum density
944 * the task can use without breaking any rule.
946 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
947 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
950 update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
952 u64 laxity = dl_se->deadline - rq_clock(rq);
955 * If the task has deadline < period, and the deadline is in the past,
956 * it should already be throttled before this check.
958 * See update_dl_entity() comments for further details.
960 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
962 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
966 * Regarding the deadline, a task with implicit deadline has a relative
967 * deadline == relative period. A task with constrained deadline has a
968 * relative deadline <= relative period.
970 * We support constrained deadline tasks. However, there are some restrictions
971 * applied only for tasks which do not have an implicit deadline. See
972 * update_dl_entity() to know more about such restrictions.
974 * The dl_is_implicit() returns true if the task has an implicit deadline.
976 static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
978 return dl_se->dl_deadline == dl_se->dl_period;
982 * When a deadline entity is placed in the runqueue, its runtime and deadline
983 * might need to be updated. This is done by a CBS wake up rule. There are two
984 * different rules: 1) the original CBS; and 2) the Revisited CBS.
986 * When the task is starting a new period, the Original CBS is used. In this
987 * case, the runtime is replenished and a new absolute deadline is set.
989 * When a task is queued before the begin of the next period, using the
990 * remaining runtime and deadline could make the entity to overflow, see
991 * dl_entity_overflow() to find more about runtime overflow. When such case
992 * is detected, the runtime and deadline need to be updated.
994 * If the task has an implicit deadline, i.e., deadline == period, the Original
995 * CBS is applied. the runtime is replenished and a new absolute deadline is
996 * set, as in the previous cases.
998 * However, the Original CBS does not work properly for tasks with
999 * deadline < period, which are said to have a constrained deadline. By
1000 * applying the Original CBS, a constrained deadline task would be able to run
1001 * runtime/deadline in a period. With deadline < period, the task would
1002 * overrun the runtime/period allowed bandwidth, breaking the admission test.
1004 * In order to prevent this misbehave, the Revisited CBS is used for
1005 * constrained deadline tasks when a runtime overflow is detected. In the
1006 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
1007 * the remaining runtime of the task is reduced to avoid runtime overflow.
1008 * Please refer to the comments update_dl_revised_wakeup() function to find
1009 * more about the Revised CBS rule.
1011 static void update_dl_entity(struct sched_dl_entity *dl_se)
1013 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1014 struct rq *rq = rq_of_dl_rq(dl_rq);
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 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. According to the
1259 * GRUB reclaiming algorithm, the runtime is not decreased as "dq = -dt",
1260 * but as "dq = -(max{u, (Umax - Uinact - Uextra)} / Umax) dt",
1261 * where u is the utilization of the task, Umax is the maximum reclaimable
1262 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1263 * as the difference between the "total runqueue utilization" and the
1264 * "runqueue active utilization", and Uextra is the (per runqueue) extra
1265 * reclaimable utilization.
1266 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations multiplied
1267 * by 2^BW_SHIFT, the result has to be shifted right by BW_SHIFT.
1268 * Since rq->dl.bw_ratio contains 1 / Umax multiplied by 2^RATIO_SHIFT, dl_bw
1269 * is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1270 * Since delta is a 64 bit variable, to have an overflow its value should be
1271 * larger than 2^(64 - 20 - 8), which is more than 64 seconds. So, overflow is
1272 * not an issue here.
1274 static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1277 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1280 * Instead of computing max{u, (u_max - u_inact - u_extra)}, we
1281 * compare u_inact + u_extra with u_max - u, because u_inact + u_extra
1282 * can be larger than u_max. So, u_max - u_inact - u_extra would be
1283 * negative leading to wrong results.
1285 if (u_inact + rq->dl.extra_bw > rq->dl.max_bw - dl_se->dl_bw)
1286 u_act = dl_se->dl_bw;
1288 u_act = rq->dl.max_bw - u_inact - rq->dl.extra_bw;
1290 u_act = (u_act * rq->dl.bw_ratio) >> RATIO_SHIFT;
1291 return (delta * u_act) >> BW_SHIFT;
1295 * Update the current task's runtime statistics (provided it is still
1296 * a -deadline task and has not been removed from the dl_rq).
1298 static void update_curr_dl(struct rq *rq)
1300 struct task_struct *curr = rq->curr;
1301 struct sched_dl_entity *dl_se = &curr->dl;
1302 u64 delta_exec, scaled_delta_exec;
1303 int cpu = cpu_of(rq);
1306 if (!dl_task(curr) || !on_dl_rq(dl_se))
1310 * Consumed budget is computed considering the time as
1311 * observed by schedulable tasks (excluding time spent
1312 * in hardirq context, etc.). Deadlines are instead
1313 * computed using hard walltime. This seems to be the more
1314 * natural solution, but the full ramifications of this
1315 * approach need further study.
1317 now = rq_clock_task(rq);
1318 delta_exec = now - curr->se.exec_start;
1319 if (unlikely((s64)delta_exec <= 0)) {
1320 if (unlikely(dl_se->dl_yielded))
1325 schedstat_set(curr->stats.exec_max,
1326 max(curr->stats.exec_max, delta_exec));
1328 trace_sched_stat_runtime(curr, delta_exec, 0);
1330 update_current_exec_runtime(curr, now, delta_exec);
1332 if (dl_entity_is_special(dl_se))
1336 * For tasks that participate in GRUB, we implement GRUB-PA: the
1337 * spare reclaimed bandwidth is used to clock down frequency.
1339 * For the others, we still need to scale reservation parameters
1340 * according to current frequency and CPU maximum capacity.
1342 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1343 scaled_delta_exec = grub_reclaim(delta_exec,
1347 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1348 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1350 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1351 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1354 dl_se->runtime -= scaled_delta_exec;
1357 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1358 dl_se->dl_throttled = 1;
1360 /* If requested, inform the user about runtime overruns. */
1361 if (dl_runtime_exceeded(dl_se) &&
1362 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1363 dl_se->dl_overrun = 1;
1365 __dequeue_task_dl(rq, curr, 0);
1366 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(curr)))
1367 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1369 if (!is_leftmost(curr, &rq->dl))
1374 * Because -- for now -- we share the rt bandwidth, we need to
1375 * account our runtime there too, otherwise actual rt tasks
1376 * would be able to exceed the shared quota.
1378 * Account to the root rt group for now.
1380 * The solution we're working towards is having the RT groups scheduled
1381 * using deadline servers -- however there's a few nasties to figure
1382 * out before that can happen.
1384 if (rt_bandwidth_enabled()) {
1385 struct rt_rq *rt_rq = &rq->rt;
1387 raw_spin_lock(&rt_rq->rt_runtime_lock);
1389 * We'll let actual RT tasks worry about the overflow here, we
1390 * have our own CBS to keep us inline; only account when RT
1391 * bandwidth is relevant.
1393 if (sched_rt_bandwidth_account(rt_rq))
1394 rt_rq->rt_time += delta_exec;
1395 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1399 static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1401 struct sched_dl_entity *dl_se = container_of(timer,
1402 struct sched_dl_entity,
1404 struct task_struct *p = dl_task_of(dl_se);
1408 rq = task_rq_lock(p, &rf);
1411 update_rq_clock(rq);
1413 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
1414 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1416 if (READ_ONCE(p->__state) == TASK_DEAD && dl_se->dl_non_contending) {
1417 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1418 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1419 dl_se->dl_non_contending = 0;
1422 raw_spin_lock(&dl_b->lock);
1423 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1424 raw_spin_unlock(&dl_b->lock);
1425 __dl_clear_params(p);
1429 if (dl_se->dl_non_contending == 0)
1432 sub_running_bw(dl_se, &rq->dl);
1433 dl_se->dl_non_contending = 0;
1435 task_rq_unlock(rq, p, &rf);
1438 return HRTIMER_NORESTART;
1441 void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1443 struct hrtimer *timer = &dl_se->inactive_timer;
1445 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1446 timer->function = inactive_task_timer;
1449 #define __node_2_dle(node) \
1450 rb_entry((node), struct sched_dl_entity, rb_node)
1454 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1456 struct rq *rq = rq_of_dl_rq(dl_rq);
1458 if (dl_rq->earliest_dl.curr == 0 ||
1459 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1460 if (dl_rq->earliest_dl.curr == 0)
1461 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_HIGHER);
1462 dl_rq->earliest_dl.curr = deadline;
1463 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1467 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1469 struct rq *rq = rq_of_dl_rq(dl_rq);
1472 * Since we may have removed our earliest (and/or next earliest)
1473 * task we must recompute them.
1475 if (!dl_rq->dl_nr_running) {
1476 dl_rq->earliest_dl.curr = 0;
1477 dl_rq->earliest_dl.next = 0;
1478 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1479 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1481 struct rb_node *leftmost = rb_first_cached(&dl_rq->root);
1482 struct sched_dl_entity *entry = __node_2_dle(leftmost);
1484 dl_rq->earliest_dl.curr = entry->deadline;
1485 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1491 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1492 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1494 #endif /* CONFIG_SMP */
1497 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1499 int prio = dl_task_of(dl_se)->prio;
1500 u64 deadline = dl_se->deadline;
1502 WARN_ON(!dl_prio(prio));
1503 dl_rq->dl_nr_running++;
1504 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1506 inc_dl_deadline(dl_rq, deadline);
1507 inc_dl_migration(dl_se, dl_rq);
1511 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1513 int prio = dl_task_of(dl_se)->prio;
1515 WARN_ON(!dl_prio(prio));
1516 WARN_ON(!dl_rq->dl_nr_running);
1517 dl_rq->dl_nr_running--;
1518 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1520 dec_dl_deadline(dl_rq, dl_se->deadline);
1521 dec_dl_migration(dl_se, dl_rq);
1524 static inline bool __dl_less(struct rb_node *a, const struct rb_node *b)
1526 return dl_time_before(__node_2_dle(a)->deadline, __node_2_dle(b)->deadline);
1529 static inline struct sched_statistics *
1530 __schedstats_from_dl_se(struct sched_dl_entity *dl_se)
1532 return &dl_task_of(dl_se)->stats;
1536 update_stats_wait_start_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1538 struct sched_statistics *stats;
1540 if (!schedstat_enabled())
1543 stats = __schedstats_from_dl_se(dl_se);
1544 __update_stats_wait_start(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1548 update_stats_wait_end_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1550 struct sched_statistics *stats;
1552 if (!schedstat_enabled())
1555 stats = __schedstats_from_dl_se(dl_se);
1556 __update_stats_wait_end(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1560 update_stats_enqueue_sleeper_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1562 struct sched_statistics *stats;
1564 if (!schedstat_enabled())
1567 stats = __schedstats_from_dl_se(dl_se);
1568 __update_stats_enqueue_sleeper(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1572 update_stats_enqueue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1575 if (!schedstat_enabled())
1578 if (flags & ENQUEUE_WAKEUP)
1579 update_stats_enqueue_sleeper_dl(dl_rq, dl_se);
1583 update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1586 struct task_struct *p = dl_task_of(dl_se);
1588 if (!schedstat_enabled())
1591 if ((flags & DEQUEUE_SLEEP)) {
1594 state = READ_ONCE(p->__state);
1595 if (state & TASK_INTERRUPTIBLE)
1596 __schedstat_set(p->stats.sleep_start,
1597 rq_clock(rq_of_dl_rq(dl_rq)));
1599 if (state & TASK_UNINTERRUPTIBLE)
1600 __schedstat_set(p->stats.block_start,
1601 rq_clock(rq_of_dl_rq(dl_rq)));
1605 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1607 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1609 WARN_ON_ONCE(!RB_EMPTY_NODE(&dl_se->rb_node));
1611 rb_add_cached(&dl_se->rb_node, &dl_rq->root, __dl_less);
1613 inc_dl_tasks(dl_se, dl_rq);
1616 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1618 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1620 if (RB_EMPTY_NODE(&dl_se->rb_node))
1623 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1625 RB_CLEAR_NODE(&dl_se->rb_node);
1627 dec_dl_tasks(dl_se, dl_rq);
1631 enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags)
1633 WARN_ON_ONCE(on_dl_rq(dl_se));
1635 update_stats_enqueue_dl(dl_rq_of_se(dl_se), dl_se, flags);
1638 * If this is a wakeup or a new instance, the scheduling
1639 * parameters of the task might need updating. Otherwise,
1640 * we want a replenishment of its runtime.
1642 if (flags & ENQUEUE_WAKEUP) {
1643 task_contending(dl_se, flags);
1644 update_dl_entity(dl_se);
1645 } else if (flags & ENQUEUE_REPLENISH) {
1646 replenish_dl_entity(dl_se);
1647 } else if ((flags & ENQUEUE_RESTORE) &&
1648 dl_time_before(dl_se->deadline,
1649 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) {
1650 setup_new_dl_entity(dl_se);
1653 __enqueue_dl_entity(dl_se);
1656 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
1658 __dequeue_dl_entity(dl_se);
1661 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1663 if (is_dl_boosted(&p->dl)) {
1665 * Because of delays in the detection of the overrun of a
1666 * thread's runtime, it might be the case that a thread
1667 * goes to sleep in a rt mutex with negative runtime. As
1668 * a consequence, the thread will be throttled.
1670 * While waiting for the mutex, this thread can also be
1671 * boosted via PI, resulting in a thread that is throttled
1672 * and boosted at the same time.
1674 * In this case, the boost overrides the throttle.
1676 if (p->dl.dl_throttled) {
1678 * The replenish timer needs to be canceled. No
1679 * problem if it fires concurrently: boosted threads
1680 * are ignored in dl_task_timer().
1682 hrtimer_try_to_cancel(&p->dl.dl_timer);
1683 p->dl.dl_throttled = 0;
1685 } else if (!dl_prio(p->normal_prio)) {
1687 * Special case in which we have a !SCHED_DEADLINE task that is going
1688 * to be deboosted, but exceeds its runtime while doing so. No point in
1689 * replenishing it, as it's going to return back to its original
1690 * scheduling class after this. If it has been throttled, we need to
1691 * clear the flag, otherwise the task may wake up as throttled after
1692 * being boosted again with no means to replenish the runtime and clear
1695 p->dl.dl_throttled = 0;
1696 if (!(flags & ENQUEUE_REPLENISH))
1697 printk_deferred_once("sched: DL de-boosted task PID %d: REPLENISH flag missing\n",
1704 * Check if a constrained deadline task was activated
1705 * after the deadline but before the next period.
1706 * If that is the case, the task will be throttled and
1707 * the replenishment timer will be set to the next period.
1709 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
1710 dl_check_constrained_dl(&p->dl);
1712 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
1713 add_rq_bw(&p->dl, &rq->dl);
1714 add_running_bw(&p->dl, &rq->dl);
1718 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1719 * its budget it needs a replenishment and, since it now is on
1720 * its rq, the bandwidth timer callback (which clearly has not
1721 * run yet) will take care of this.
1722 * However, the active utilization does not depend on the fact
1723 * that the task is on the runqueue or not (but depends on the
1724 * task's state - in GRUB parlance, "inactive" vs "active contending").
1725 * In other words, even if a task is throttled its utilization must
1726 * be counted in the active utilization; hence, we need to call
1729 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1730 if (flags & ENQUEUE_WAKEUP)
1731 task_contending(&p->dl, flags);
1736 check_schedstat_required();
1737 update_stats_wait_start_dl(dl_rq_of_se(&p->dl), &p->dl);
1739 enqueue_dl_entity(&p->dl, flags);
1741 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1742 enqueue_pushable_dl_task(rq, p);
1745 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1747 update_stats_dequeue_dl(&rq->dl, &p->dl, flags);
1748 dequeue_dl_entity(&p->dl);
1749 dequeue_pushable_dl_task(rq, p);
1752 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1755 __dequeue_task_dl(rq, p, flags);
1757 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
1758 sub_running_bw(&p->dl, &rq->dl);
1759 sub_rq_bw(&p->dl, &rq->dl);
1763 * This check allows to start the inactive timer (or to immediately
1764 * decrease the active utilization, if needed) in two cases:
1765 * when the task blocks and when it is terminating
1766 * (p->state == TASK_DEAD). We can handle the two cases in the same
1767 * way, because from GRUB's point of view the same thing is happening
1768 * (the task moves from "active contending" to "active non contending"
1771 if (flags & DEQUEUE_SLEEP)
1772 task_non_contending(p);
1776 * Yield task semantic for -deadline tasks is:
1778 * get off from the CPU until our next instance, with
1779 * a new runtime. This is of little use now, since we
1780 * don't have a bandwidth reclaiming mechanism. Anyway,
1781 * bandwidth reclaiming is planned for the future, and
1782 * yield_task_dl will indicate that some spare budget
1783 * is available for other task instances to use it.
1785 static void yield_task_dl(struct rq *rq)
1788 * We make the task go to sleep until its current deadline by
1789 * forcing its runtime to zero. This way, update_curr_dl() stops
1790 * it and the bandwidth timer will wake it up and will give it
1791 * new scheduling parameters (thanks to dl_yielded=1).
1793 rq->curr->dl.dl_yielded = 1;
1795 update_rq_clock(rq);
1798 * Tell update_rq_clock() that we've just updated,
1799 * so we don't do microscopic update in schedule()
1800 * and double the fastpath cost.
1802 rq_clock_skip_update(rq);
1807 static inline bool dl_task_is_earliest_deadline(struct task_struct *p,
1810 return (!rq->dl.dl_nr_running ||
1811 dl_time_before(p->dl.deadline,
1812 rq->dl.earliest_dl.curr));
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 (sched_asym_cpucap_active())
1851 select_rq |= !dl_task_fits_capacity(p, cpu);
1854 int target = find_later_rq(p);
1857 dl_task_is_earliest_deadline(p, cpu_rq(target)))
1866 static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
1871 if (READ_ONCE(p->__state) != TASK_WAKING)
1876 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1877 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1878 * rq->lock is not... So, lock it
1881 if (p->dl.dl_non_contending) {
1882 update_rq_clock(rq);
1883 sub_running_bw(&p->dl, &rq->dl);
1884 p->dl.dl_non_contending = 0;
1886 * If the timer handler is currently running and the
1887 * timer cannot be canceled, inactive_task_timer()
1888 * will see that dl_not_contending is not set, and
1889 * will not touch the rq's active utilization,
1890 * so we are still safe.
1892 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
1895 sub_rq_bw(&p->dl, &rq->dl);
1899 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1902 * Current can't be migrated, useless to reschedule,
1903 * let's hope p can move out.
1905 if (rq->curr->nr_cpus_allowed == 1 ||
1906 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
1910 * p is migratable, so let's not schedule it and
1911 * see if it is pushed or pulled somewhere else.
1913 if (p->nr_cpus_allowed != 1 &&
1914 cpudl_find(&rq->rd->cpudl, p, NULL))
1920 static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1922 if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) {
1924 * This is OK, because current is on_cpu, which avoids it being
1925 * picked for load-balance and preemption/IRQs are still
1926 * disabled avoiding further scheduler activity on it and we've
1927 * not yet started the picking loop.
1929 rq_unpin_lock(rq, rf);
1931 rq_repin_lock(rq, rf);
1934 return sched_stop_runnable(rq) || sched_dl_runnable(rq);
1936 #endif /* CONFIG_SMP */
1939 * Only called when both the current and waking task are -deadline
1942 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1945 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
1952 * In the unlikely case current and p have the same deadline
1953 * let us try to decide what's the best thing to do...
1955 if ((p->dl.deadline == rq->curr->dl.deadline) &&
1956 !test_tsk_need_resched(rq->curr))
1957 check_preempt_equal_dl(rq, p);
1958 #endif /* CONFIG_SMP */
1961 #ifdef CONFIG_SCHED_HRTICK
1962 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1964 hrtick_start(rq, p->dl.runtime);
1966 #else /* !CONFIG_SCHED_HRTICK */
1967 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1972 static void set_next_task_dl(struct rq *rq, struct task_struct *p, bool first)
1974 struct sched_dl_entity *dl_se = &p->dl;
1975 struct dl_rq *dl_rq = &rq->dl;
1977 p->se.exec_start = rq_clock_task(rq);
1978 if (on_dl_rq(&p->dl))
1979 update_stats_wait_end_dl(dl_rq, dl_se);
1981 /* You can't push away the running task */
1982 dequeue_pushable_dl_task(rq, p);
1987 if (hrtick_enabled_dl(rq))
1988 start_hrtick_dl(rq, p);
1990 if (rq->curr->sched_class != &dl_sched_class)
1991 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
1993 deadline_queue_push_tasks(rq);
1996 static struct sched_dl_entity *pick_next_dl_entity(struct dl_rq *dl_rq)
1998 struct rb_node *left = rb_first_cached(&dl_rq->root);
2003 return __node_2_dle(left);
2006 static struct task_struct *pick_task_dl(struct rq *rq)
2008 struct sched_dl_entity *dl_se;
2009 struct dl_rq *dl_rq = &rq->dl;
2010 struct task_struct *p;
2012 if (!sched_dl_runnable(rq))
2015 dl_se = pick_next_dl_entity(dl_rq);
2016 WARN_ON_ONCE(!dl_se);
2017 p = dl_task_of(dl_se);
2022 static struct task_struct *pick_next_task_dl(struct rq *rq)
2024 struct task_struct *p;
2026 p = pick_task_dl(rq);
2028 set_next_task_dl(rq, p, true);
2033 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
2035 struct sched_dl_entity *dl_se = &p->dl;
2036 struct dl_rq *dl_rq = &rq->dl;
2038 if (on_dl_rq(&p->dl))
2039 update_stats_wait_start_dl(dl_rq, dl_se);
2043 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2044 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
2045 enqueue_pushable_dl_task(rq, p);
2049 * scheduler tick hitting a task of our scheduling class.
2051 * NOTE: This function can be called remotely by the tick offload that
2052 * goes along full dynticks. Therefore no local assumption can be made
2053 * and everything must be accessed through the @rq and @curr passed in
2056 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
2060 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2062 * Even when we have runtime, update_curr_dl() might have resulted in us
2063 * not being the leftmost task anymore. In that case NEED_RESCHED will
2064 * be set and schedule() will start a new hrtick for the next task.
2066 if (hrtick_enabled_dl(rq) && queued && p->dl.runtime > 0 &&
2067 is_leftmost(p, &rq->dl))
2068 start_hrtick_dl(rq, p);
2071 static void task_fork_dl(struct task_struct *p)
2074 * SCHED_DEADLINE tasks cannot fork and this is achieved through
2081 /* Only try algorithms three times */
2082 #define DL_MAX_TRIES 3
2084 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
2086 if (!task_on_cpu(rq, p) &&
2087 cpumask_test_cpu(cpu, &p->cpus_mask))
2093 * Return the earliest pushable rq's task, which is suitable to be executed
2094 * on the CPU, NULL otherwise:
2096 static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
2098 struct task_struct *p = NULL;
2099 struct rb_node *next_node;
2101 if (!has_pushable_dl_tasks(rq))
2104 next_node = rb_first_cached(&rq->dl.pushable_dl_tasks_root);
2108 p = __node_2_pdl(next_node);
2110 if (pick_dl_task(rq, p, cpu))
2113 next_node = rb_next(next_node);
2120 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
2122 static int find_later_rq(struct task_struct *task)
2124 struct sched_domain *sd;
2125 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
2126 int this_cpu = smp_processor_id();
2127 int cpu = task_cpu(task);
2129 /* Make sure the mask is initialized first */
2130 if (unlikely(!later_mask))
2133 if (task->nr_cpus_allowed == 1)
2137 * We have to consider system topology and task affinity
2138 * first, then we can look for a suitable CPU.
2140 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
2144 * If we are here, some targets have been found, including
2145 * the most suitable which is, among the runqueues where the
2146 * current tasks have later deadlines than the task's one, the
2147 * rq with the latest possible one.
2149 * Now we check how well this matches with task's
2150 * affinity and system topology.
2152 * The last CPU where the task run is our first
2153 * guess, since it is most likely cache-hot there.
2155 if (cpumask_test_cpu(cpu, later_mask))
2158 * Check if this_cpu is to be skipped (i.e., it is
2159 * not in the mask) or not.
2161 if (!cpumask_test_cpu(this_cpu, later_mask))
2165 for_each_domain(cpu, sd) {
2166 if (sd->flags & SD_WAKE_AFFINE) {
2170 * If possible, preempting this_cpu is
2171 * cheaper than migrating.
2173 if (this_cpu != -1 &&
2174 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
2179 best_cpu = cpumask_any_and_distribute(later_mask,
2180 sched_domain_span(sd));
2182 * Last chance: if a CPU being in both later_mask
2183 * and current sd span is valid, that becomes our
2184 * choice. Of course, the latest possible CPU is
2185 * already under consideration through later_mask.
2187 if (best_cpu < nr_cpu_ids) {
2196 * At this point, all our guesses failed, we just return
2197 * 'something', and let the caller sort the things out.
2202 cpu = cpumask_any_distribute(later_mask);
2203 if (cpu < nr_cpu_ids)
2209 /* Locks the rq it finds */
2210 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
2212 struct rq *later_rq = NULL;
2216 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
2217 cpu = find_later_rq(task);
2219 if ((cpu == -1) || (cpu == rq->cpu))
2222 later_rq = cpu_rq(cpu);
2224 if (!dl_task_is_earliest_deadline(task, later_rq)) {
2226 * Target rq has tasks of equal or earlier deadline,
2227 * retrying does not release any lock and is unlikely
2228 * to yield a different result.
2234 /* Retry if something changed. */
2235 if (double_lock_balance(rq, later_rq)) {
2236 if (unlikely(task_rq(task) != rq ||
2237 !cpumask_test_cpu(later_rq->cpu, &task->cpus_mask) ||
2238 task_on_cpu(rq, task) ||
2240 is_migration_disabled(task) ||
2241 !task_on_rq_queued(task))) {
2242 double_unlock_balance(rq, later_rq);
2249 * If the rq we found has no -deadline task, or
2250 * its earliest one has a later deadline than our
2251 * task, the rq is a good one.
2253 if (dl_task_is_earliest_deadline(task, later_rq))
2256 /* Otherwise we try again. */
2257 double_unlock_balance(rq, later_rq);
2264 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2266 struct task_struct *p;
2268 if (!has_pushable_dl_tasks(rq))
2271 p = __node_2_pdl(rb_first_cached(&rq->dl.pushable_dl_tasks_root));
2273 WARN_ON_ONCE(rq->cpu != task_cpu(p));
2274 WARN_ON_ONCE(task_current(rq, p));
2275 WARN_ON_ONCE(p->nr_cpus_allowed <= 1);
2277 WARN_ON_ONCE(!task_on_rq_queued(p));
2278 WARN_ON_ONCE(!dl_task(p));
2284 * See if the non running -deadline tasks on this rq
2285 * can be sent to some other CPU where they can preempt
2286 * and start executing.
2288 static int push_dl_task(struct rq *rq)
2290 struct task_struct *next_task;
2291 struct rq *later_rq;
2294 if (!rq->dl.overloaded)
2297 next_task = pick_next_pushable_dl_task(rq);
2303 * If next_task preempts rq->curr, and rq->curr
2304 * can move away, it makes sense to just reschedule
2305 * without going further in pushing next_task.
2307 if (dl_task(rq->curr) &&
2308 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2309 rq->curr->nr_cpus_allowed > 1) {
2314 if (is_migration_disabled(next_task))
2317 if (WARN_ON(next_task == rq->curr))
2320 /* We might release rq lock */
2321 get_task_struct(next_task);
2323 /* Will lock the rq it'll find */
2324 later_rq = find_lock_later_rq(next_task, rq);
2326 struct task_struct *task;
2329 * We must check all this again, since
2330 * find_lock_later_rq releases rq->lock and it is
2331 * then possible that next_task has migrated.
2333 task = pick_next_pushable_dl_task(rq);
2334 if (task == next_task) {
2336 * The task is still there. We don't try
2337 * again, some other CPU will pull it when ready.
2346 put_task_struct(next_task);
2351 deactivate_task(rq, next_task, 0);
2352 set_task_cpu(next_task, later_rq->cpu);
2353 activate_task(later_rq, next_task, 0);
2356 resched_curr(later_rq);
2358 double_unlock_balance(rq, later_rq);
2361 put_task_struct(next_task);
2366 static void push_dl_tasks(struct rq *rq)
2368 /* push_dl_task() will return true if it moved a -deadline task */
2369 while (push_dl_task(rq))
2373 static void pull_dl_task(struct rq *this_rq)
2375 int this_cpu = this_rq->cpu, cpu;
2376 struct task_struct *p, *push_task;
2377 bool resched = false;
2379 u64 dmin = LONG_MAX;
2381 if (likely(!dl_overloaded(this_rq)))
2385 * Match the barrier from dl_set_overloaded; this guarantees that if we
2386 * see overloaded we must also see the dlo_mask bit.
2390 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2391 if (this_cpu == cpu)
2394 src_rq = cpu_rq(cpu);
2397 * It looks racy, abd it is! However, as in sched_rt.c,
2398 * we are fine with this.
2400 if (this_rq->dl.dl_nr_running &&
2401 dl_time_before(this_rq->dl.earliest_dl.curr,
2402 src_rq->dl.earliest_dl.next))
2405 /* Might drop this_rq->lock */
2407 double_lock_balance(this_rq, src_rq);
2410 * If there are no more pullable tasks on the
2411 * rq, we're done with it.
2413 if (src_rq->dl.dl_nr_running <= 1)
2416 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2419 * We found a task to be pulled if:
2420 * - it preempts our current (if there's one),
2421 * - it will preempt the last one we pulled (if any).
2423 if (p && dl_time_before(p->dl.deadline, dmin) &&
2424 dl_task_is_earliest_deadline(p, this_rq)) {
2425 WARN_ON(p == src_rq->curr);
2426 WARN_ON(!task_on_rq_queued(p));
2429 * Then we pull iff p has actually an earlier
2430 * deadline than the current task of its runqueue.
2432 if (dl_time_before(p->dl.deadline,
2433 src_rq->curr->dl.deadline))
2436 if (is_migration_disabled(p)) {
2437 push_task = get_push_task(src_rq);
2439 deactivate_task(src_rq, p, 0);
2440 set_task_cpu(p, this_cpu);
2441 activate_task(this_rq, p, 0);
2442 dmin = p->dl.deadline;
2446 /* Is there any other task even earlier? */
2449 double_unlock_balance(this_rq, src_rq);
2452 raw_spin_rq_unlock(this_rq);
2453 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2454 push_task, &src_rq->push_work);
2455 raw_spin_rq_lock(this_rq);
2460 resched_curr(this_rq);
2464 * Since the task is not running and a reschedule is not going to happen
2465 * anytime soon on its runqueue, we try pushing it away now.
2467 static void task_woken_dl(struct rq *rq, struct task_struct *p)
2469 if (!task_on_cpu(rq, p) &&
2470 !test_tsk_need_resched(rq->curr) &&
2471 p->nr_cpus_allowed > 1 &&
2472 dl_task(rq->curr) &&
2473 (rq->curr->nr_cpus_allowed < 2 ||
2474 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2479 static void set_cpus_allowed_dl(struct task_struct *p,
2480 struct affinity_context *ctx)
2482 struct root_domain *src_rd;
2485 WARN_ON_ONCE(!dl_task(p));
2490 * Migrating a SCHED_DEADLINE task between exclusive
2491 * cpusets (different root_domains) entails a bandwidth
2492 * update. We already made space for us in the destination
2493 * domain (see cpuset_can_attach()).
2495 if (!cpumask_intersects(src_rd->span, ctx->new_mask)) {
2496 struct dl_bw *src_dl_b;
2498 src_dl_b = dl_bw_of(cpu_of(rq));
2500 * We now free resources of the root_domain we are migrating
2501 * off. In the worst case, sched_setattr() may temporary fail
2502 * until we complete the update.
2504 raw_spin_lock(&src_dl_b->lock);
2505 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2506 raw_spin_unlock(&src_dl_b->lock);
2509 set_cpus_allowed_common(p, ctx);
2512 /* Assumes rq->lock is held */
2513 static void rq_online_dl(struct rq *rq)
2515 if (rq->dl.overloaded)
2516 dl_set_overload(rq);
2518 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2519 if (rq->dl.dl_nr_running > 0)
2520 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2523 /* Assumes rq->lock is held */
2524 static void rq_offline_dl(struct rq *rq)
2526 if (rq->dl.overloaded)
2527 dl_clear_overload(rq);
2529 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2530 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2533 void __init init_sched_dl_class(void)
2537 for_each_possible_cpu(i)
2538 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2539 GFP_KERNEL, cpu_to_node(i));
2542 void dl_add_task_root_domain(struct task_struct *p)
2548 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2550 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
2554 rq = __task_rq_lock(p, &rf);
2556 dl_b = &rq->rd->dl_bw;
2557 raw_spin_lock(&dl_b->lock);
2559 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2561 raw_spin_unlock(&dl_b->lock);
2563 task_rq_unlock(rq, p, &rf);
2566 void dl_clear_root_domain(struct root_domain *rd)
2568 unsigned long flags;
2570 raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
2571 rd->dl_bw.total_bw = 0;
2572 raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
2575 #endif /* CONFIG_SMP */
2577 static void switched_from_dl(struct rq *rq, struct task_struct *p)
2580 * task_non_contending() can start the "inactive timer" (if the 0-lag
2581 * time is in the future). If the task switches back to dl before
2582 * the "inactive timer" fires, it can continue to consume its current
2583 * runtime using its current deadline. If it stays outside of
2584 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2585 * will reset the task parameters.
2587 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2588 task_non_contending(p);
2591 * In case a task is setscheduled out from SCHED_DEADLINE we need to
2592 * keep track of that on its cpuset (for correct bandwidth tracking).
2596 if (!task_on_rq_queued(p)) {
2598 * Inactive timer is armed. However, p is leaving DEADLINE and
2599 * might migrate away from this rq while continuing to run on
2600 * some other class. We need to remove its contribution from
2601 * this rq running_bw now, or sub_rq_bw (below) will complain.
2603 if (p->dl.dl_non_contending)
2604 sub_running_bw(&p->dl, &rq->dl);
2605 sub_rq_bw(&p->dl, &rq->dl);
2609 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2610 * at the 0-lag time, because the task could have been migrated
2611 * while SCHED_OTHER in the meanwhile.
2613 if (p->dl.dl_non_contending)
2614 p->dl.dl_non_contending = 0;
2617 * Since this might be the only -deadline task on the rq,
2618 * this is the right place to try to pull some other one
2619 * from an overloaded CPU, if any.
2621 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
2624 deadline_queue_pull_task(rq);
2628 * When switching to -deadline, we may overload the rq, then
2629 * we try to push someone off, if possible.
2631 static void switched_to_dl(struct rq *rq, struct task_struct *p)
2633 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2637 * In case a task is setscheduled to SCHED_DEADLINE we need to keep
2638 * track of that on its cpuset (for correct bandwidth tracking).
2642 /* If p is not queued we will update its parameters at next wakeup. */
2643 if (!task_on_rq_queued(p)) {
2644 add_rq_bw(&p->dl, &rq->dl);
2649 if (rq->curr != p) {
2651 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
2652 deadline_queue_push_tasks(rq);
2654 if (dl_task(rq->curr))
2655 check_preempt_curr_dl(rq, p, 0);
2659 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2664 * If the scheduling parameters of a -deadline task changed,
2665 * a push or pull operation might be needed.
2667 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2670 if (!task_on_rq_queued(p))
2675 * This might be too much, but unfortunately
2676 * we don't have the old deadline value, and
2677 * we can't argue if the task is increasing
2678 * or lowering its prio, so...
2680 if (!rq->dl.overloaded)
2681 deadline_queue_pull_task(rq);
2683 if (task_current(rq, p)) {
2685 * If we now have a earlier deadline task than p,
2686 * then reschedule, provided p is still on this
2689 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
2693 * Current may not be deadline in case p was throttled but we
2694 * have just replenished it (e.g. rt_mutex_setprio()).
2696 * Otherwise, if p was given an earlier deadline, reschedule.
2698 if (!dl_task(rq->curr) ||
2699 dl_time_before(p->dl.deadline, rq->curr->dl.deadline))
2704 * We don't know if p has a earlier or later deadline, so let's blindly
2705 * set a (maybe not needed) rescheduling point.
2711 #ifdef CONFIG_SCHED_CORE
2712 static int task_is_throttled_dl(struct task_struct *p, int cpu)
2714 return p->dl.dl_throttled;
2718 DEFINE_SCHED_CLASS(dl) = {
2720 .enqueue_task = enqueue_task_dl,
2721 .dequeue_task = dequeue_task_dl,
2722 .yield_task = yield_task_dl,
2724 .check_preempt_curr = check_preempt_curr_dl,
2726 .pick_next_task = pick_next_task_dl,
2727 .put_prev_task = put_prev_task_dl,
2728 .set_next_task = set_next_task_dl,
2731 .balance = balance_dl,
2732 .pick_task = pick_task_dl,
2733 .select_task_rq = select_task_rq_dl,
2734 .migrate_task_rq = migrate_task_rq_dl,
2735 .set_cpus_allowed = set_cpus_allowed_dl,
2736 .rq_online = rq_online_dl,
2737 .rq_offline = rq_offline_dl,
2738 .task_woken = task_woken_dl,
2739 .find_lock_rq = find_lock_later_rq,
2742 .task_tick = task_tick_dl,
2743 .task_fork = task_fork_dl,
2745 .prio_changed = prio_changed_dl,
2746 .switched_from = switched_from_dl,
2747 .switched_to = switched_to_dl,
2749 .update_curr = update_curr_dl,
2750 #ifdef CONFIG_SCHED_CORE
2751 .task_is_throttled = task_is_throttled_dl,
2755 /* Used for dl_bw check and update, used under sched_rt_handler()::mutex */
2756 static u64 dl_generation;
2758 int sched_dl_global_validate(void)
2760 u64 runtime = global_rt_runtime();
2761 u64 period = global_rt_period();
2762 u64 new_bw = to_ratio(period, runtime);
2763 u64 gen = ++dl_generation;
2765 int cpu, cpus, ret = 0;
2766 unsigned long flags;
2769 * Here we want to check the bandwidth not being set to some
2770 * value smaller than the currently allocated bandwidth in
2771 * any of the root_domains.
2773 for_each_possible_cpu(cpu) {
2774 rcu_read_lock_sched();
2776 if (dl_bw_visited(cpu, gen))
2779 dl_b = dl_bw_of(cpu);
2780 cpus = dl_bw_cpus(cpu);
2782 raw_spin_lock_irqsave(&dl_b->lock, flags);
2783 if (new_bw * cpus < dl_b->total_bw)
2785 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2788 rcu_read_unlock_sched();
2797 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
2799 if (global_rt_runtime() == RUNTIME_INF) {
2800 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2801 dl_rq->max_bw = dl_rq->extra_bw = 1 << BW_SHIFT;
2803 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2804 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2805 dl_rq->max_bw = dl_rq->extra_bw =
2806 to_ratio(global_rt_period(), global_rt_runtime());
2810 void sched_dl_do_global(void)
2813 u64 gen = ++dl_generation;
2816 unsigned long flags;
2818 if (global_rt_runtime() != RUNTIME_INF)
2819 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2821 for_each_possible_cpu(cpu) {
2822 rcu_read_lock_sched();
2824 if (dl_bw_visited(cpu, gen)) {
2825 rcu_read_unlock_sched();
2829 dl_b = dl_bw_of(cpu);
2831 raw_spin_lock_irqsave(&dl_b->lock, flags);
2833 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2835 rcu_read_unlock_sched();
2836 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2841 * We must be sure that accepting a new task (or allowing changing the
2842 * parameters of an existing one) is consistent with the bandwidth
2843 * constraints. If yes, this function also accordingly updates the currently
2844 * allocated bandwidth to reflect the new situation.
2846 * This function is called while holding p's rq->lock.
2848 int sched_dl_overflow(struct task_struct *p, int policy,
2849 const struct sched_attr *attr)
2851 u64 period = attr->sched_period ?: attr->sched_deadline;
2852 u64 runtime = attr->sched_runtime;
2853 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2854 int cpus, err = -1, cpu = task_cpu(p);
2855 struct dl_bw *dl_b = dl_bw_of(cpu);
2858 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2861 /* !deadline task may carry old deadline bandwidth */
2862 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2866 * Either if a task, enters, leave, or stays -deadline but changes
2867 * its parameters, we may need to update accordingly the total
2868 * allocated bandwidth of the container.
2870 raw_spin_lock(&dl_b->lock);
2871 cpus = dl_bw_cpus(cpu);
2872 cap = dl_bw_capacity(cpu);
2874 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2875 !__dl_overflow(dl_b, cap, 0, new_bw)) {
2876 if (hrtimer_active(&p->dl.inactive_timer))
2877 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2878 __dl_add(dl_b, new_bw, cpus);
2880 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2881 !__dl_overflow(dl_b, cap, p->dl.dl_bw, new_bw)) {
2883 * XXX this is slightly incorrect: when the task
2884 * utilization decreases, we should delay the total
2885 * utilization change until the task's 0-lag point.
2886 * But this would require to set the task's "inactive
2887 * timer" when the task is not inactive.
2889 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2890 __dl_add(dl_b, new_bw, cpus);
2891 dl_change_utilization(p, new_bw);
2893 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2895 * Do not decrease the total deadline utilization here,
2896 * switched_from_dl() will take care to do it at the correct
2901 raw_spin_unlock(&dl_b->lock);
2907 * This function initializes the sched_dl_entity of a newly becoming
2908 * SCHED_DEADLINE task.
2910 * Only the static values are considered here, the actual runtime and the
2911 * absolute deadline will be properly calculated when the task is enqueued
2912 * for the first time with its new policy.
2914 void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
2916 struct sched_dl_entity *dl_se = &p->dl;
2918 dl_se->dl_runtime = attr->sched_runtime;
2919 dl_se->dl_deadline = attr->sched_deadline;
2920 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
2921 dl_se->flags = attr->sched_flags & SCHED_DL_FLAGS;
2922 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
2923 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
2926 void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
2928 struct sched_dl_entity *dl_se = &p->dl;
2930 attr->sched_priority = p->rt_priority;
2931 attr->sched_runtime = dl_se->dl_runtime;
2932 attr->sched_deadline = dl_se->dl_deadline;
2933 attr->sched_period = dl_se->dl_period;
2934 attr->sched_flags &= ~SCHED_DL_FLAGS;
2935 attr->sched_flags |= dl_se->flags;
2939 * This function validates the new parameters of a -deadline task.
2940 * We ask for the deadline not being zero, and greater or equal
2941 * than the runtime, as well as the period of being zero or
2942 * greater than deadline. Furthermore, we have to be sure that
2943 * user parameters are above the internal resolution of 1us (we
2944 * check sched_runtime only since it is always the smaller one) and
2945 * below 2^63 ns (we have to check both sched_deadline and
2946 * sched_period, as the latter can be zero).
2948 bool __checkparam_dl(const struct sched_attr *attr)
2950 u64 period, max, min;
2952 /* special dl tasks don't actually use any parameter */
2953 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2957 if (attr->sched_deadline == 0)
2961 * Since we truncate DL_SCALE bits, make sure we're at least
2964 if (attr->sched_runtime < (1ULL << DL_SCALE))
2968 * Since we use the MSB for wrap-around and sign issues, make
2969 * sure it's not set (mind that period can be equal to zero).
2971 if (attr->sched_deadline & (1ULL << 63) ||
2972 attr->sched_period & (1ULL << 63))
2975 period = attr->sched_period;
2977 period = attr->sched_deadline;
2979 /* runtime <= deadline <= period (if period != 0) */
2980 if (period < attr->sched_deadline ||
2981 attr->sched_deadline < attr->sched_runtime)
2984 max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
2985 min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
2987 if (period < min || period > max)
2994 * This function clears the sched_dl_entity static params.
2996 void __dl_clear_params(struct task_struct *p)
2998 struct sched_dl_entity *dl_se = &p->dl;
3000 dl_se->dl_runtime = 0;
3001 dl_se->dl_deadline = 0;
3002 dl_se->dl_period = 0;
3005 dl_se->dl_density = 0;
3007 dl_se->dl_throttled = 0;
3008 dl_se->dl_yielded = 0;
3009 dl_se->dl_non_contending = 0;
3010 dl_se->dl_overrun = 0;
3012 #ifdef CONFIG_RT_MUTEXES
3013 dl_se->pi_se = dl_se;
3017 bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
3019 struct sched_dl_entity *dl_se = &p->dl;
3021 if (dl_se->dl_runtime != attr->sched_runtime ||
3022 dl_se->dl_deadline != attr->sched_deadline ||
3023 dl_se->dl_period != attr->sched_period ||
3024 dl_se->flags != (attr->sched_flags & SCHED_DL_FLAGS))
3031 int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
3032 const struct cpumask *trial)
3034 unsigned long flags, cap;
3035 struct dl_bw *cur_dl_b;
3038 rcu_read_lock_sched();
3039 cur_dl_b = dl_bw_of(cpumask_any(cur));
3040 cap = __dl_bw_capacity(trial);
3041 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
3042 if (__dl_overflow(cur_dl_b, cap, 0, 0))
3044 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
3045 rcu_read_unlock_sched();
3050 enum dl_bw_request {
3051 dl_bw_req_check_overflow = 0,
3056 static int dl_bw_manage(enum dl_bw_request req, int cpu, u64 dl_bw)
3058 unsigned long flags;
3062 rcu_read_lock_sched();
3063 dl_b = dl_bw_of(cpu);
3064 raw_spin_lock_irqsave(&dl_b->lock, flags);
3066 if (req == dl_bw_req_free) {
3067 __dl_sub(dl_b, dl_bw, dl_bw_cpus(cpu));
3069 unsigned long cap = dl_bw_capacity(cpu);
3071 overflow = __dl_overflow(dl_b, cap, 0, dl_bw);
3073 if (req == dl_bw_req_alloc && !overflow) {
3075 * We reserve space in the destination
3076 * root_domain, as we can't fail after this point.
3077 * We will free resources in the source root_domain
3078 * later on (see set_cpus_allowed_dl()).
3080 __dl_add(dl_b, dl_bw, dl_bw_cpus(cpu));
3084 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3085 rcu_read_unlock_sched();
3087 return overflow ? -EBUSY : 0;
3090 int dl_bw_check_overflow(int cpu)
3092 return dl_bw_manage(dl_bw_req_check_overflow, cpu, 0);
3095 int dl_bw_alloc(int cpu, u64 dl_bw)
3097 return dl_bw_manage(dl_bw_req_alloc, cpu, dl_bw);
3100 void dl_bw_free(int cpu, u64 dl_bw)
3102 dl_bw_manage(dl_bw_req_free, cpu, dl_bw);
3106 #ifdef CONFIG_SCHED_DEBUG
3107 void print_dl_stats(struct seq_file *m, int cpu)
3109 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
3111 #endif /* CONFIG_SCHED_DEBUG */