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_douintvec_minmax,
34 .extra1 = (void *)&sysctl_sched_dl_period_min,
37 .procname = "sched_deadline_period_min_us",
38 .data = &sysctl_sched_dl_period_min,
39 .maxlen = sizeof(unsigned int),
41 .proc_handler = proc_douintvec_minmax,
42 .extra2 = (void *)&sysctl_sched_dl_period_max,
47 static int __init sched_dl_sysctl_init(void)
49 register_sysctl_init("kernel", sched_dl_sysctls);
52 late_initcall(sched_dl_sysctl_init);
55 static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
57 return container_of(dl_se, struct task_struct, dl);
60 static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
62 return container_of(dl_rq, struct rq, dl);
65 static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
67 struct task_struct *p = dl_task_of(dl_se);
68 struct rq *rq = task_rq(p);
73 static inline int on_dl_rq(struct sched_dl_entity *dl_se)
75 return !RB_EMPTY_NODE(&dl_se->rb_node);
78 #ifdef CONFIG_RT_MUTEXES
79 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
84 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
86 return pi_of(dl_se) != dl_se;
89 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
94 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
101 static inline struct dl_bw *dl_bw_of(int i)
103 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
104 "sched RCU must be held");
105 return &cpu_rq(i)->rd->dl_bw;
108 static inline int dl_bw_cpus(int i)
110 struct root_domain *rd = cpu_rq(i)->rd;
113 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
114 "sched RCU must be held");
116 if (cpumask_subset(rd->span, cpu_active_mask))
117 return cpumask_weight(rd->span);
121 for_each_cpu_and(i, rd->span, cpu_active_mask)
127 static inline unsigned long __dl_bw_capacity(const struct cpumask *mask)
129 unsigned long cap = 0;
132 for_each_cpu_and(i, mask, cpu_active_mask)
133 cap += capacity_orig_of(i);
139 * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
140 * of the CPU the task is running on rather rd's \Sum CPU capacity.
142 static inline unsigned long dl_bw_capacity(int i)
144 if (!sched_asym_cpucap_active() &&
145 capacity_orig_of(i) == SCHED_CAPACITY_SCALE) {
146 return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT;
148 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
149 "sched RCU must be held");
151 return __dl_bw_capacity(cpu_rq(i)->rd->span);
155 static inline bool dl_bw_visited(int cpu, u64 gen)
157 struct root_domain *rd = cpu_rq(cpu)->rd;
159 if (rd->visit_gen == gen)
167 void __dl_update(struct dl_bw *dl_b, s64 bw)
169 struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw);
172 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
173 "sched RCU must be held");
174 for_each_cpu_and(i, rd->span, cpu_active_mask) {
175 struct rq *rq = cpu_rq(i);
177 rq->dl.extra_bw += bw;
181 static inline struct dl_bw *dl_bw_of(int i)
183 return &cpu_rq(i)->dl.dl_bw;
186 static inline int dl_bw_cpus(int i)
191 static inline unsigned long dl_bw_capacity(int i)
193 return SCHED_CAPACITY_SCALE;
196 static inline bool dl_bw_visited(int cpu, u64 gen)
202 void __dl_update(struct dl_bw *dl_b, s64 bw)
204 struct dl_rq *dl = container_of(dl_b, struct dl_rq, dl_bw);
211 void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
213 dl_b->total_bw -= tsk_bw;
214 __dl_update(dl_b, (s32)tsk_bw / cpus);
218 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
220 dl_b->total_bw += tsk_bw;
221 __dl_update(dl_b, -((s32)tsk_bw / cpus));
225 __dl_overflow(struct dl_bw *dl_b, unsigned long cap, u64 old_bw, u64 new_bw)
227 return dl_b->bw != -1 &&
228 cap_scale(dl_b->bw, cap) < dl_b->total_bw - old_bw + new_bw;
232 void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
234 u64 old = dl_rq->running_bw;
236 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
237 dl_rq->running_bw += dl_bw;
238 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
239 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
240 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
241 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
245 void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
247 u64 old = dl_rq->running_bw;
249 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
250 dl_rq->running_bw -= dl_bw;
251 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
252 if (dl_rq->running_bw > old)
253 dl_rq->running_bw = 0;
254 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
255 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
259 void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
261 u64 old = dl_rq->this_bw;
263 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
264 dl_rq->this_bw += dl_bw;
265 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
269 void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
271 u64 old = dl_rq->this_bw;
273 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
274 dl_rq->this_bw -= dl_bw;
275 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
276 if (dl_rq->this_bw > old)
278 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
282 void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
284 if (!dl_entity_is_special(dl_se))
285 __add_rq_bw(dl_se->dl_bw, dl_rq);
289 void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
291 if (!dl_entity_is_special(dl_se))
292 __sub_rq_bw(dl_se->dl_bw, dl_rq);
296 void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
298 if (!dl_entity_is_special(dl_se))
299 __add_running_bw(dl_se->dl_bw, dl_rq);
303 void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
305 if (!dl_entity_is_special(dl_se))
306 __sub_running_bw(dl_se->dl_bw, dl_rq);
309 static void dl_change_utilization(struct task_struct *p, u64 new_bw)
313 WARN_ON_ONCE(p->dl.flags & SCHED_FLAG_SUGOV);
315 if (task_on_rq_queued(p))
319 if (p->dl.dl_non_contending) {
320 sub_running_bw(&p->dl, &rq->dl);
321 p->dl.dl_non_contending = 0;
323 * If the timer handler is currently running and the
324 * timer cannot be canceled, inactive_task_timer()
325 * will see that dl_not_contending is not set, and
326 * will not touch the rq's active utilization,
327 * so we are still safe.
329 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
332 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
333 __add_rq_bw(new_bw, &rq->dl);
337 * The utilization of a task cannot be immediately removed from
338 * the rq active utilization (running_bw) when the task blocks.
339 * Instead, we have to wait for the so called "0-lag time".
341 * If a task blocks before the "0-lag time", a timer (the inactive
342 * timer) is armed, and running_bw is decreased when the timer
345 * If the task wakes up again before the inactive timer fires,
346 * the timer is canceled, whereas if the task wakes up after the
347 * inactive timer fired (and running_bw has been decreased) the
348 * task's utilization has to be added to running_bw again.
349 * A flag in the deadline scheduling entity (dl_non_contending)
350 * is used to avoid race conditions between the inactive timer handler
353 * The following diagram shows how running_bw is updated. A task is
354 * "ACTIVE" when its utilization contributes to running_bw; an
355 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
356 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
357 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
358 * time already passed, which does not contribute to running_bw anymore.
359 * +------------------+
361 * +------------------>+ contending |
362 * | add_running_bw | |
363 * | +----+------+------+
366 * +--------+-------+ | |
367 * | | t >= 0-lag | | wakeup
368 * | INACTIVE |<---------------+ |
369 * | | sub_running_bw | |
370 * +--------+-------+ | |
375 * | +----+------+------+
376 * | sub_running_bw | ACTIVE |
377 * +-------------------+ |
378 * inactive timer | non contending |
379 * fired +------------------+
381 * The task_non_contending() function is invoked when a task
382 * blocks, and checks if the 0-lag time already passed or
383 * not (in the first case, it directly updates running_bw;
384 * in the second case, it arms the inactive timer).
386 * The task_contending() function is invoked when a task wakes
387 * up, and checks if the task is still in the "ACTIVE non contending"
388 * state or not (in the second case, it updates running_bw).
390 static void task_non_contending(struct task_struct *p)
392 struct sched_dl_entity *dl_se = &p->dl;
393 struct hrtimer *timer = &dl_se->inactive_timer;
394 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
395 struct rq *rq = rq_of_dl_rq(dl_rq);
399 * If this is a non-deadline task that has been boosted,
402 if (dl_se->dl_runtime == 0)
405 if (dl_entity_is_special(dl_se))
408 WARN_ON(dl_se->dl_non_contending);
410 zerolag_time = dl_se->deadline -
411 div64_long((dl_se->runtime * dl_se->dl_period),
415 * Using relative times instead of the absolute "0-lag time"
416 * allows to simplify the code
418 zerolag_time -= rq_clock(rq);
421 * If the "0-lag time" already passed, decrease the active
422 * utilization now, instead of starting a timer
424 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
426 sub_running_bw(dl_se, dl_rq);
427 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
428 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
430 if (READ_ONCE(p->__state) == TASK_DEAD)
431 sub_rq_bw(&p->dl, &rq->dl);
432 raw_spin_lock(&dl_b->lock);
433 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
434 __dl_clear_params(p);
435 raw_spin_unlock(&dl_b->lock);
441 dl_se->dl_non_contending = 1;
443 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD);
446 static void task_contending(struct sched_dl_entity *dl_se, int flags)
448 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
451 * If this is a non-deadline task that has been boosted,
454 if (dl_se->dl_runtime == 0)
457 if (flags & ENQUEUE_MIGRATED)
458 add_rq_bw(dl_se, dl_rq);
460 if (dl_se->dl_non_contending) {
461 dl_se->dl_non_contending = 0;
463 * If the timer handler is currently running and the
464 * timer cannot be canceled, inactive_task_timer()
465 * will see that dl_not_contending is not set, and
466 * will not touch the rq's active utilization,
467 * so we are still safe.
469 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1)
470 put_task_struct(dl_task_of(dl_se));
473 * Since "dl_non_contending" is not set, the
474 * task's utilization has already been removed from
475 * active utilization (either when the task blocked,
476 * when the "inactive timer" fired).
479 add_running_bw(dl_se, dl_rq);
483 static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
485 struct sched_dl_entity *dl_se = &p->dl;
487 return rb_first_cached(&dl_rq->root) == &dl_se->rb_node;
490 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
492 void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
494 raw_spin_lock_init(&dl_b->dl_runtime_lock);
495 dl_b->dl_period = period;
496 dl_b->dl_runtime = runtime;
499 void init_dl_bw(struct dl_bw *dl_b)
501 raw_spin_lock_init(&dl_b->lock);
502 if (global_rt_runtime() == RUNTIME_INF)
505 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
509 void init_dl_rq(struct dl_rq *dl_rq)
511 dl_rq->root = RB_ROOT_CACHED;
514 /* zero means no -deadline tasks */
515 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
517 dl_rq->dl_nr_migratory = 0;
518 dl_rq->overloaded = 0;
519 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
521 init_dl_bw(&dl_rq->dl_bw);
524 dl_rq->running_bw = 0;
526 init_dl_rq_bw_ratio(dl_rq);
531 static inline int dl_overloaded(struct rq *rq)
533 return atomic_read(&rq->rd->dlo_count);
536 static inline void dl_set_overload(struct rq *rq)
541 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
543 * Must be visible before the overload count is
544 * set (as in sched_rt.c).
546 * Matched by the barrier in pull_dl_task().
549 atomic_inc(&rq->rd->dlo_count);
552 static inline void dl_clear_overload(struct rq *rq)
557 atomic_dec(&rq->rd->dlo_count);
558 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
561 static void update_dl_migration(struct dl_rq *dl_rq)
563 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
564 if (!dl_rq->overloaded) {
565 dl_set_overload(rq_of_dl_rq(dl_rq));
566 dl_rq->overloaded = 1;
568 } else if (dl_rq->overloaded) {
569 dl_clear_overload(rq_of_dl_rq(dl_rq));
570 dl_rq->overloaded = 0;
574 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
576 struct task_struct *p = dl_task_of(dl_se);
578 if (p->nr_cpus_allowed > 1)
579 dl_rq->dl_nr_migratory++;
581 update_dl_migration(dl_rq);
584 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
586 struct task_struct *p = dl_task_of(dl_se);
588 if (p->nr_cpus_allowed > 1)
589 dl_rq->dl_nr_migratory--;
591 update_dl_migration(dl_rq);
594 #define __node_2_pdl(node) \
595 rb_entry((node), struct task_struct, pushable_dl_tasks)
597 static inline bool __pushable_less(struct rb_node *a, const struct rb_node *b)
599 return dl_entity_preempt(&__node_2_pdl(a)->dl, &__node_2_pdl(b)->dl);
603 * The list of pushable -deadline task is not a plist, like in
604 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
606 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
608 struct rb_node *leftmost;
610 WARN_ON_ONCE(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
612 leftmost = rb_add_cached(&p->pushable_dl_tasks,
613 &rq->dl.pushable_dl_tasks_root,
616 rq->dl.earliest_dl.next = p->dl.deadline;
619 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
621 struct dl_rq *dl_rq = &rq->dl;
622 struct rb_root_cached *root = &dl_rq->pushable_dl_tasks_root;
623 struct rb_node *leftmost;
625 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
628 leftmost = rb_erase_cached(&p->pushable_dl_tasks, root);
630 dl_rq->earliest_dl.next = __node_2_pdl(leftmost)->dl.deadline;
632 RB_CLEAR_NODE(&p->pushable_dl_tasks);
635 static inline int has_pushable_dl_tasks(struct rq *rq)
637 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
640 static int push_dl_task(struct rq *rq);
642 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
644 return rq->online && dl_task(prev);
647 static DEFINE_PER_CPU(struct callback_head, dl_push_head);
648 static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
650 static void push_dl_tasks(struct rq *);
651 static void pull_dl_task(struct rq *);
653 static inline void deadline_queue_push_tasks(struct rq *rq)
655 if (!has_pushable_dl_tasks(rq))
658 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
661 static inline void deadline_queue_pull_task(struct rq *rq)
663 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
666 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
668 static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
670 struct rq *later_rq = NULL;
673 later_rq = find_lock_later_rq(p, rq);
678 * If we cannot preempt any rq, fall back to pick any
681 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
682 if (cpu >= nr_cpu_ids) {
684 * Failed to find any suitable CPU.
685 * The task will never come back!
687 WARN_ON_ONCE(dl_bandwidth_enabled());
690 * If admission control is disabled we
691 * try a little harder to let the task
694 cpu = cpumask_any(cpu_active_mask);
696 later_rq = cpu_rq(cpu);
697 double_lock_balance(rq, later_rq);
700 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
702 * Inactive timer is armed (or callback is running, but
703 * waiting for us to release rq locks). In any case, when it
704 * will fire (or continue), it will see running_bw of this
705 * task migrated to later_rq (and correctly handle it).
707 sub_running_bw(&p->dl, &rq->dl);
708 sub_rq_bw(&p->dl, &rq->dl);
710 add_rq_bw(&p->dl, &later_rq->dl);
711 add_running_bw(&p->dl, &later_rq->dl);
713 sub_rq_bw(&p->dl, &rq->dl);
714 add_rq_bw(&p->dl, &later_rq->dl);
718 * And we finally need to fixup root_domain(s) bandwidth accounting,
719 * since p is still hanging out in the old (now moved to default) root
722 dl_b = &rq->rd->dl_bw;
723 raw_spin_lock(&dl_b->lock);
724 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
725 raw_spin_unlock(&dl_b->lock);
727 dl_b = &later_rq->rd->dl_bw;
728 raw_spin_lock(&dl_b->lock);
729 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
730 raw_spin_unlock(&dl_b->lock);
732 set_task_cpu(p, later_rq->cpu);
733 double_unlock_balance(later_rq, rq);
741 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
746 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
751 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
756 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
760 static inline void deadline_queue_push_tasks(struct rq *rq)
764 static inline void deadline_queue_pull_task(struct rq *rq)
767 #endif /* CONFIG_SMP */
769 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
770 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
771 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags);
774 * We are being explicitly informed that a new instance is starting,
775 * and this means that:
776 * - the absolute deadline of the entity has to be placed at
777 * current time + relative deadline;
778 * - the runtime of the entity has to be set to the maximum value.
780 * The capability of specifying such event is useful whenever a -deadline
781 * entity wants to (try to!) synchronize its behaviour with the scheduler's
782 * one, and to (try to!) reconcile itself with its own scheduling
785 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
787 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
788 struct rq *rq = rq_of_dl_rq(dl_rq);
790 WARN_ON(is_dl_boosted(dl_se));
791 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
794 * We are racing with the deadline timer. So, do nothing because
795 * the deadline timer handler will take care of properly recharging
796 * the runtime and postponing the deadline
798 if (dl_se->dl_throttled)
802 * We use the regular wall clock time to set deadlines in the
803 * future; in fact, we must consider execution overheads (time
804 * spent on hardirq context, etc.).
806 dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline;
807 dl_se->runtime = dl_se->dl_runtime;
811 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
812 * possibility of a entity lasting more than what it declared, and thus
813 * exhausting its runtime.
815 * Here we are interested in making runtime overrun possible, but we do
816 * not want a entity which is misbehaving to affect the scheduling of all
818 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
819 * is used, in order to confine each entity within its own bandwidth.
821 * This function deals exactly with that, and ensures that when the runtime
822 * of a entity is replenished, its deadline is also postponed. That ensures
823 * the overrunning entity can't interfere with other entity in the system and
824 * can't make them miss their deadlines. Reasons why this kind of overruns
825 * could happen are, typically, a entity voluntarily trying to overcome its
826 * runtime, or it just underestimated it during sched_setattr().
828 static void replenish_dl_entity(struct sched_dl_entity *dl_se)
830 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
831 struct rq *rq = rq_of_dl_rq(dl_rq);
833 WARN_ON_ONCE(pi_of(dl_se)->dl_runtime <= 0);
836 * This could be the case for a !-dl task that is boosted.
837 * Just go with full inherited parameters.
839 if (dl_se->dl_deadline == 0) {
840 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
841 dl_se->runtime = pi_of(dl_se)->dl_runtime;
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 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
870 dl_se->runtime = pi_of(dl_se)->dl_runtime;
873 if (dl_se->dl_yielded)
874 dl_se->dl_yielded = 0;
875 if (dl_se->dl_throttled)
876 dl_se->dl_throttled = 0;
880 * Here we check if --at time t-- an entity (which is probably being
881 * [re]activated or, in general, enqueued) can use its remaining runtime
882 * and its current deadline _without_ exceeding the bandwidth it is
883 * assigned (function returns true if it can't). We are in fact applying
884 * one of the CBS rules: when a task wakes up, if the residual runtime
885 * over residual deadline fits within the allocated bandwidth, then we
886 * can keep the current (absolute) deadline and residual budget without
887 * disrupting the schedulability of the system. Otherwise, we should
888 * refill the runtime and set the deadline a period in the future,
889 * because keeping the current (absolute) deadline of the task would
890 * result in breaking guarantees promised to other tasks (refer to
891 * Documentation/scheduler/sched-deadline.rst for more information).
893 * This function returns true if:
895 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
897 * IOW we can't recycle current parameters.
899 * Notice that the bandwidth check is done against the deadline. For
900 * task with deadline equal to period this is the same of using
901 * dl_period instead of dl_deadline in the equation above.
903 static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t)
908 * left and right are the two sides of the equation above,
909 * after a bit of shuffling to use multiplications instead
912 * Note that none of the time values involved in the two
913 * multiplications are absolute: dl_deadline and dl_runtime
914 * are the relative deadline and the maximum runtime of each
915 * instance, runtime is the runtime left for the last instance
916 * and (deadline - t), since t is rq->clock, is the time left
917 * to the (absolute) deadline. Even if overflowing the u64 type
918 * is very unlikely to occur in both cases, here we scale down
919 * as we want to avoid that risk at all. Scaling down by 10
920 * means that we reduce granularity to 1us. We are fine with it,
921 * since this is only a true/false check and, anyway, thinking
922 * of anything below microseconds resolution is actually fiction
923 * (but still we want to give the user that illusion >;).
925 left = (pi_of(dl_se)->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
926 right = ((dl_se->deadline - t) >> DL_SCALE) *
927 (pi_of(dl_se)->dl_runtime >> DL_SCALE);
929 return dl_time_before(right, left);
933 * Revised wakeup rule [1]: For self-suspending tasks, rather then
934 * re-initializing task's runtime and deadline, the revised wakeup
935 * rule adjusts the task's runtime to avoid the task to overrun its
938 * Reasoning: a task may overrun the density if:
939 * runtime / (deadline - t) > dl_runtime / dl_deadline
941 * Therefore, runtime can be adjusted to:
942 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
944 * In such way that runtime will be equal to the maximum density
945 * the task can use without breaking any rule.
947 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
948 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
951 update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
953 u64 laxity = dl_se->deadline - rq_clock(rq);
956 * If the task has deadline < period, and the deadline is in the past,
957 * it should already be throttled before this check.
959 * See update_dl_entity() comments for further details.
961 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
963 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
967 * Regarding the deadline, a task with implicit deadline has a relative
968 * deadline == relative period. A task with constrained deadline has a
969 * relative deadline <= relative period.
971 * We support constrained deadline tasks. However, there are some restrictions
972 * applied only for tasks which do not have an implicit deadline. See
973 * update_dl_entity() to know more about such restrictions.
975 * The dl_is_implicit() returns true if the task has an implicit deadline.
977 static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
979 return dl_se->dl_deadline == dl_se->dl_period;
983 * When a deadline entity is placed in the runqueue, its runtime and deadline
984 * might need to be updated. This is done by a CBS wake up rule. There are two
985 * different rules: 1) the original CBS; and 2) the Revisited CBS.
987 * When the task is starting a new period, the Original CBS is used. In this
988 * case, the runtime is replenished and a new absolute deadline is set.
990 * When a task is queued before the begin of the next period, using the
991 * remaining runtime and deadline could make the entity to overflow, see
992 * dl_entity_overflow() to find more about runtime overflow. When such case
993 * is detected, the runtime and deadline need to be updated.
995 * If the task has an implicit deadline, i.e., deadline == period, the Original
996 * CBS is applied. the runtime is replenished and a new absolute deadline is
997 * set, as in the previous cases.
999 * However, the Original CBS does not work properly for tasks with
1000 * deadline < period, which are said to have a constrained deadline. By
1001 * applying the Original CBS, a constrained deadline task would be able to run
1002 * runtime/deadline in a period. With deadline < period, the task would
1003 * overrun the runtime/period allowed bandwidth, breaking the admission test.
1005 * In order to prevent this misbehave, the Revisited CBS is used for
1006 * constrained deadline tasks when a runtime overflow is detected. In the
1007 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
1008 * the remaining runtime of the task is reduced to avoid runtime overflow.
1009 * Please refer to the comments update_dl_revised_wakeup() function to find
1010 * more about the Revised CBS rule.
1012 static void update_dl_entity(struct sched_dl_entity *dl_se)
1014 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1015 struct rq *rq = rq_of_dl_rq(dl_rq);
1017 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
1018 dl_entity_overflow(dl_se, rq_clock(rq))) {
1020 if (unlikely(!dl_is_implicit(dl_se) &&
1021 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1022 !is_dl_boosted(dl_se))) {
1023 update_dl_revised_wakeup(dl_se, rq);
1027 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
1028 dl_se->runtime = pi_of(dl_se)->dl_runtime;
1032 static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
1034 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
1038 * If the entity depleted all its runtime, and if we want it to sleep
1039 * while waiting for some new execution time to become available, we
1040 * set the bandwidth replenishment timer to the replenishment instant
1041 * and try to activate it.
1043 * Notice that it is important for the caller to know if the timer
1044 * actually started or not (i.e., the replenishment instant is in
1045 * the future or in the past).
1047 static int start_dl_timer(struct task_struct *p)
1049 struct sched_dl_entity *dl_se = &p->dl;
1050 struct hrtimer *timer = &dl_se->dl_timer;
1051 struct rq *rq = task_rq(p);
1055 lockdep_assert_rq_held(rq);
1058 * We want the timer to fire at the deadline, but considering
1059 * that it is actually coming from rq->clock and not from
1060 * hrtimer's time base reading.
1062 act = ns_to_ktime(dl_next_period(dl_se));
1063 now = hrtimer_cb_get_time(timer);
1064 delta = ktime_to_ns(now) - rq_clock(rq);
1065 act = ktime_add_ns(act, delta);
1068 * If the expiry time already passed, e.g., because the value
1069 * chosen as the deadline is too small, don't even try to
1070 * start the timer in the past!
1072 if (ktime_us_delta(act, now) < 0)
1076 * !enqueued will guarantee another callback; even if one is already in
1077 * progress. This ensures a balanced {get,put}_task_struct().
1079 * The race against __run_timer() clearing the enqueued state is
1080 * harmless because we're holding task_rq()->lock, therefore the timer
1081 * expiring after we've done the check will wait on its task_rq_lock()
1082 * and observe our state.
1084 if (!hrtimer_is_queued(timer)) {
1086 hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD);
1093 * This is the bandwidth enforcement timer callback. If here, we know
1094 * a task is not on its dl_rq, since the fact that the timer was running
1095 * means the task is throttled and needs a runtime replenishment.
1097 * However, what we actually do depends on the fact the task is active,
1098 * (it is on its rq) or has been removed from there by a call to
1099 * dequeue_task_dl(). In the former case we must issue the runtime
1100 * replenishment and add the task back to the dl_rq; in the latter, we just
1101 * do nothing but clearing dl_throttled, so that runtime and deadline
1102 * updating (and the queueing back to dl_rq) will be done by the
1103 * next call to enqueue_task_dl().
1105 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
1107 struct sched_dl_entity *dl_se = container_of(timer,
1108 struct sched_dl_entity,
1110 struct task_struct *p = dl_task_of(dl_se);
1114 rq = task_rq_lock(p, &rf);
1117 * The task might have changed its scheduling policy to something
1118 * different than SCHED_DEADLINE (through switched_from_dl()).
1124 * The task might have been boosted by someone else and might be in the
1125 * boosting/deboosting path, its not throttled.
1127 if (is_dl_boosted(dl_se))
1131 * Spurious timer due to start_dl_timer() race; or we already received
1132 * a replenishment from rt_mutex_setprio().
1134 if (!dl_se->dl_throttled)
1138 update_rq_clock(rq);
1141 * If the throttle happened during sched-out; like:
1148 * __dequeue_task_dl()
1151 * We can be both throttled and !queued. Replenish the counter
1152 * but do not enqueue -- wait for our wakeup to do that.
1154 if (!task_on_rq_queued(p)) {
1155 replenish_dl_entity(dl_se);
1160 if (unlikely(!rq->online)) {
1162 * If the runqueue is no longer available, migrate the
1163 * task elsewhere. This necessarily changes rq.
1165 lockdep_unpin_lock(__rq_lockp(rq), rf.cookie);
1166 rq = dl_task_offline_migration(rq, p);
1167 rf.cookie = lockdep_pin_lock(__rq_lockp(rq));
1168 update_rq_clock(rq);
1171 * Now that the task has been migrated to the new RQ and we
1172 * have that locked, proceed as normal and enqueue the task
1178 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1179 if (dl_task(rq->curr))
1180 check_preempt_curr_dl(rq, p, 0);
1186 * Queueing this task back might have overloaded rq, check if we need
1187 * to kick someone away.
1189 if (has_pushable_dl_tasks(rq)) {
1191 * Nothing relies on rq->lock after this, so its safe to drop
1194 rq_unpin_lock(rq, &rf);
1196 rq_repin_lock(rq, &rf);
1201 task_rq_unlock(rq, p, &rf);
1204 * This can free the task_struct, including this hrtimer, do not touch
1205 * anything related to that after this.
1209 return HRTIMER_NORESTART;
1212 void init_dl_task_timer(struct sched_dl_entity *dl_se)
1214 struct hrtimer *timer = &dl_se->dl_timer;
1216 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1217 timer->function = dl_task_timer;
1221 * During the activation, CBS checks if it can reuse the current task's
1222 * runtime and period. If the deadline of the task is in the past, CBS
1223 * cannot use the runtime, and so it replenishes the task. This rule
1224 * works fine for implicit deadline tasks (deadline == period), and the
1225 * CBS was designed for implicit deadline tasks. However, a task with
1226 * constrained deadline (deadline < period) might be awakened after the
1227 * deadline, but before the next period. In this case, replenishing the
1228 * task would allow it to run for runtime / deadline. As in this case
1229 * deadline < period, CBS enables a task to run for more than the
1230 * runtime / period. In a very loaded system, this can cause a domino
1231 * effect, making other tasks miss their deadlines.
1233 * To avoid this problem, in the activation of a constrained deadline
1234 * task after the deadline but before the next period, throttle the
1235 * task and set the replenishing timer to the begin of the next period,
1236 * unless it is boosted.
1238 static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1240 struct task_struct *p = dl_task_of(dl_se);
1241 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1243 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1244 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1245 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(p)))
1247 dl_se->dl_throttled = 1;
1248 if (dl_se->runtime > 0)
1254 int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1256 return (dl_se->runtime <= 0);
1260 * This function implements the GRUB accounting rule:
1261 * according to the GRUB reclaiming algorithm, the runtime is
1262 * not decreased as "dq = -dt", but as
1263 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1264 * where u is the utilization of the task, Umax is the maximum reclaimable
1265 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1266 * as the difference between the "total runqueue utilization" and the
1267 * runqueue active utilization, and Uextra is the (per runqueue) extra
1268 * reclaimable utilization.
1269 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1270 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1272 * Since rq->dl.bw_ratio contains 1 / Umax multiplied by 2^RATIO_SHIFT,
1273 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1274 * Since delta is a 64 bit variable, to have an overflow its value
1275 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
1276 * So, overflow is not an issue here.
1278 static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1280 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1282 u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT;
1285 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1286 * we compare u_inact + rq->dl.extra_bw with
1287 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1288 * u_inact + rq->dl.extra_bw can be larger than
1289 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
1290 * leading to wrong results)
1292 if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min)
1295 u_act = BW_UNIT - u_inact - rq->dl.extra_bw;
1297 return (delta * u_act) >> BW_SHIFT;
1301 * Update the current task's runtime statistics (provided it is still
1302 * a -deadline task and has not been removed from the dl_rq).
1304 static void update_curr_dl(struct rq *rq)
1306 struct task_struct *curr = rq->curr;
1307 struct sched_dl_entity *dl_se = &curr->dl;
1308 u64 delta_exec, scaled_delta_exec;
1309 int cpu = cpu_of(rq);
1312 if (!dl_task(curr) || !on_dl_rq(dl_se))
1316 * Consumed budget is computed considering the time as
1317 * observed by schedulable tasks (excluding time spent
1318 * in hardirq context, etc.). Deadlines are instead
1319 * computed using hard walltime. This seems to be the more
1320 * natural solution, but the full ramifications of this
1321 * approach need further study.
1323 now = rq_clock_task(rq);
1324 delta_exec = now - curr->se.exec_start;
1325 if (unlikely((s64)delta_exec <= 0)) {
1326 if (unlikely(dl_se->dl_yielded))
1331 schedstat_set(curr->stats.exec_max,
1332 max(curr->stats.exec_max, delta_exec));
1334 trace_sched_stat_runtime(curr, delta_exec, 0);
1336 update_current_exec_runtime(curr, now, delta_exec);
1338 if (dl_entity_is_special(dl_se))
1342 * For tasks that participate in GRUB, we implement GRUB-PA: the
1343 * spare reclaimed bandwidth is used to clock down frequency.
1345 * For the others, we still need to scale reservation parameters
1346 * according to current frequency and CPU maximum capacity.
1348 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1349 scaled_delta_exec = grub_reclaim(delta_exec,
1353 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1354 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1356 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1357 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1360 dl_se->runtime -= scaled_delta_exec;
1363 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1364 dl_se->dl_throttled = 1;
1366 /* If requested, inform the user about runtime overruns. */
1367 if (dl_runtime_exceeded(dl_se) &&
1368 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1369 dl_se->dl_overrun = 1;
1371 __dequeue_task_dl(rq, curr, 0);
1372 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(curr)))
1373 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1375 if (!is_leftmost(curr, &rq->dl))
1380 * Because -- for now -- we share the rt bandwidth, we need to
1381 * account our runtime there too, otherwise actual rt tasks
1382 * would be able to exceed the shared quota.
1384 * Account to the root rt group for now.
1386 * The solution we're working towards is having the RT groups scheduled
1387 * using deadline servers -- however there's a few nasties to figure
1388 * out before that can happen.
1390 if (rt_bandwidth_enabled()) {
1391 struct rt_rq *rt_rq = &rq->rt;
1393 raw_spin_lock(&rt_rq->rt_runtime_lock);
1395 * We'll let actual RT tasks worry about the overflow here, we
1396 * have our own CBS to keep us inline; only account when RT
1397 * bandwidth is relevant.
1399 if (sched_rt_bandwidth_account(rt_rq))
1400 rt_rq->rt_time += delta_exec;
1401 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1405 static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1407 struct sched_dl_entity *dl_se = container_of(timer,
1408 struct sched_dl_entity,
1410 struct task_struct *p = dl_task_of(dl_se);
1414 rq = task_rq_lock(p, &rf);
1417 update_rq_clock(rq);
1419 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
1420 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1422 if (READ_ONCE(p->__state) == TASK_DEAD && dl_se->dl_non_contending) {
1423 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1424 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1425 dl_se->dl_non_contending = 0;
1428 raw_spin_lock(&dl_b->lock);
1429 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1430 raw_spin_unlock(&dl_b->lock);
1431 __dl_clear_params(p);
1435 if (dl_se->dl_non_contending == 0)
1438 sub_running_bw(dl_se, &rq->dl);
1439 dl_se->dl_non_contending = 0;
1441 task_rq_unlock(rq, p, &rf);
1444 return HRTIMER_NORESTART;
1447 void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1449 struct hrtimer *timer = &dl_se->inactive_timer;
1451 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1452 timer->function = inactive_task_timer;
1455 #define __node_2_dle(node) \
1456 rb_entry((node), struct sched_dl_entity, rb_node)
1460 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1462 struct rq *rq = rq_of_dl_rq(dl_rq);
1464 if (dl_rq->earliest_dl.curr == 0 ||
1465 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1466 if (dl_rq->earliest_dl.curr == 0)
1467 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_HIGHER);
1468 dl_rq->earliest_dl.curr = deadline;
1469 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1473 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1475 struct rq *rq = rq_of_dl_rq(dl_rq);
1478 * Since we may have removed our earliest (and/or next earliest)
1479 * task we must recompute them.
1481 if (!dl_rq->dl_nr_running) {
1482 dl_rq->earliest_dl.curr = 0;
1483 dl_rq->earliest_dl.next = 0;
1484 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1485 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1487 struct rb_node *leftmost = rb_first_cached(&dl_rq->root);
1488 struct sched_dl_entity *entry = __node_2_dle(leftmost);
1490 dl_rq->earliest_dl.curr = entry->deadline;
1491 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1497 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1498 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1500 #endif /* CONFIG_SMP */
1503 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1505 int prio = dl_task_of(dl_se)->prio;
1506 u64 deadline = dl_se->deadline;
1508 WARN_ON(!dl_prio(prio));
1509 dl_rq->dl_nr_running++;
1510 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1512 inc_dl_deadline(dl_rq, deadline);
1513 inc_dl_migration(dl_se, dl_rq);
1517 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1519 int prio = dl_task_of(dl_se)->prio;
1521 WARN_ON(!dl_prio(prio));
1522 WARN_ON(!dl_rq->dl_nr_running);
1523 dl_rq->dl_nr_running--;
1524 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1526 dec_dl_deadline(dl_rq, dl_se->deadline);
1527 dec_dl_migration(dl_se, dl_rq);
1530 static inline bool __dl_less(struct rb_node *a, const struct rb_node *b)
1532 return dl_time_before(__node_2_dle(a)->deadline, __node_2_dle(b)->deadline);
1535 static inline struct sched_statistics *
1536 __schedstats_from_dl_se(struct sched_dl_entity *dl_se)
1538 return &dl_task_of(dl_se)->stats;
1542 update_stats_wait_start_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1544 struct sched_statistics *stats;
1546 if (!schedstat_enabled())
1549 stats = __schedstats_from_dl_se(dl_se);
1550 __update_stats_wait_start(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1554 update_stats_wait_end_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1556 struct sched_statistics *stats;
1558 if (!schedstat_enabled())
1561 stats = __schedstats_from_dl_se(dl_se);
1562 __update_stats_wait_end(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1566 update_stats_enqueue_sleeper_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1568 struct sched_statistics *stats;
1570 if (!schedstat_enabled())
1573 stats = __schedstats_from_dl_se(dl_se);
1574 __update_stats_enqueue_sleeper(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1578 update_stats_enqueue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1581 if (!schedstat_enabled())
1584 if (flags & ENQUEUE_WAKEUP)
1585 update_stats_enqueue_sleeper_dl(dl_rq, dl_se);
1589 update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1592 struct task_struct *p = dl_task_of(dl_se);
1594 if (!schedstat_enabled())
1597 if ((flags & DEQUEUE_SLEEP)) {
1600 state = READ_ONCE(p->__state);
1601 if (state & TASK_INTERRUPTIBLE)
1602 __schedstat_set(p->stats.sleep_start,
1603 rq_clock(rq_of_dl_rq(dl_rq)));
1605 if (state & TASK_UNINTERRUPTIBLE)
1606 __schedstat_set(p->stats.block_start,
1607 rq_clock(rq_of_dl_rq(dl_rq)));
1611 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1613 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1615 WARN_ON_ONCE(!RB_EMPTY_NODE(&dl_se->rb_node));
1617 rb_add_cached(&dl_se->rb_node, &dl_rq->root, __dl_less);
1619 inc_dl_tasks(dl_se, dl_rq);
1622 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1624 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1626 if (RB_EMPTY_NODE(&dl_se->rb_node))
1629 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1631 RB_CLEAR_NODE(&dl_se->rb_node);
1633 dec_dl_tasks(dl_se, dl_rq);
1637 enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags)
1639 WARN_ON_ONCE(on_dl_rq(dl_se));
1641 update_stats_enqueue_dl(dl_rq_of_se(dl_se), dl_se, flags);
1644 * If this is a wakeup or a new instance, the scheduling
1645 * parameters of the task might need updating. Otherwise,
1646 * we want a replenishment of its runtime.
1648 if (flags & ENQUEUE_WAKEUP) {
1649 task_contending(dl_se, flags);
1650 update_dl_entity(dl_se);
1651 } else if (flags & ENQUEUE_REPLENISH) {
1652 replenish_dl_entity(dl_se);
1653 } else if ((flags & ENQUEUE_RESTORE) &&
1654 dl_time_before(dl_se->deadline,
1655 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) {
1656 setup_new_dl_entity(dl_se);
1659 __enqueue_dl_entity(dl_se);
1662 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
1664 __dequeue_dl_entity(dl_se);
1667 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1669 if (is_dl_boosted(&p->dl)) {
1671 * Because of delays in the detection of the overrun of a
1672 * thread's runtime, it might be the case that a thread
1673 * goes to sleep in a rt mutex with negative runtime. As
1674 * a consequence, the thread will be throttled.
1676 * While waiting for the mutex, this thread can also be
1677 * boosted via PI, resulting in a thread that is throttled
1678 * and boosted at the same time.
1680 * In this case, the boost overrides the throttle.
1682 if (p->dl.dl_throttled) {
1684 * The replenish timer needs to be canceled. No
1685 * problem if it fires concurrently: boosted threads
1686 * are ignored in dl_task_timer().
1688 hrtimer_try_to_cancel(&p->dl.dl_timer);
1689 p->dl.dl_throttled = 0;
1691 } else if (!dl_prio(p->normal_prio)) {
1693 * Special case in which we have a !SCHED_DEADLINE task that is going
1694 * to be deboosted, but exceeds its runtime while doing so. No point in
1695 * replenishing it, as it's going to return back to its original
1696 * scheduling class after this. If it has been throttled, we need to
1697 * clear the flag, otherwise the task may wake up as throttled after
1698 * being boosted again with no means to replenish the runtime and clear
1701 p->dl.dl_throttled = 0;
1702 if (!(flags & ENQUEUE_REPLENISH))
1703 printk_deferred_once("sched: DL de-boosted task PID %d: REPLENISH flag missing\n",
1710 * Check if a constrained deadline task was activated
1711 * after the deadline but before the next period.
1712 * If that is the case, the task will be throttled and
1713 * the replenishment timer will be set to the next period.
1715 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
1716 dl_check_constrained_dl(&p->dl);
1718 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
1719 add_rq_bw(&p->dl, &rq->dl);
1720 add_running_bw(&p->dl, &rq->dl);
1724 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1725 * its budget it needs a replenishment and, since it now is on
1726 * its rq, the bandwidth timer callback (which clearly has not
1727 * run yet) will take care of this.
1728 * However, the active utilization does not depend on the fact
1729 * that the task is on the runqueue or not (but depends on the
1730 * task's state - in GRUB parlance, "inactive" vs "active contending").
1731 * In other words, even if a task is throttled its utilization must
1732 * be counted in the active utilization; hence, we need to call
1735 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1736 if (flags & ENQUEUE_WAKEUP)
1737 task_contending(&p->dl, flags);
1742 check_schedstat_required();
1743 update_stats_wait_start_dl(dl_rq_of_se(&p->dl), &p->dl);
1745 enqueue_dl_entity(&p->dl, flags);
1747 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1748 enqueue_pushable_dl_task(rq, p);
1751 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1753 update_stats_dequeue_dl(&rq->dl, &p->dl, flags);
1754 dequeue_dl_entity(&p->dl);
1755 dequeue_pushable_dl_task(rq, p);
1758 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1761 __dequeue_task_dl(rq, p, flags);
1763 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
1764 sub_running_bw(&p->dl, &rq->dl);
1765 sub_rq_bw(&p->dl, &rq->dl);
1769 * This check allows to start the inactive timer (or to immediately
1770 * decrease the active utilization, if needed) in two cases:
1771 * when the task blocks and when it is terminating
1772 * (p->state == TASK_DEAD). We can handle the two cases in the same
1773 * way, because from GRUB's point of view the same thing is happening
1774 * (the task moves from "active contending" to "active non contending"
1777 if (flags & DEQUEUE_SLEEP)
1778 task_non_contending(p);
1782 * Yield task semantic for -deadline tasks is:
1784 * get off from the CPU until our next instance, with
1785 * a new runtime. This is of little use now, since we
1786 * don't have a bandwidth reclaiming mechanism. Anyway,
1787 * bandwidth reclaiming is planned for the future, and
1788 * yield_task_dl will indicate that some spare budget
1789 * is available for other task instances to use it.
1791 static void yield_task_dl(struct rq *rq)
1794 * We make the task go to sleep until its current deadline by
1795 * forcing its runtime to zero. This way, update_curr_dl() stops
1796 * it and the bandwidth timer will wake it up and will give it
1797 * new scheduling parameters (thanks to dl_yielded=1).
1799 rq->curr->dl.dl_yielded = 1;
1801 update_rq_clock(rq);
1804 * Tell update_rq_clock() that we've just updated,
1805 * so we don't do microscopic update in schedule()
1806 * and double the fastpath cost.
1808 rq_clock_skip_update(rq);
1813 static int find_later_rq(struct task_struct *task);
1816 select_task_rq_dl(struct task_struct *p, int cpu, int flags)
1818 struct task_struct *curr;
1822 if (!(flags & WF_TTWU))
1828 curr = READ_ONCE(rq->curr); /* unlocked access */
1831 * If we are dealing with a -deadline task, we must
1832 * decide where to wake it up.
1833 * If it has a later deadline and the current task
1834 * on this rq can't move (provided the waking task
1835 * can!) we prefer to send it somewhere else. On the
1836 * other hand, if it has a shorter deadline, we
1837 * try to make it stay here, it might be important.
1839 select_rq = unlikely(dl_task(curr)) &&
1840 (curr->nr_cpus_allowed < 2 ||
1841 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1842 p->nr_cpus_allowed > 1;
1845 * Take the capacity of the CPU into account to
1846 * ensure it fits the requirement of the task.
1848 if (sched_asym_cpucap_active())
1849 select_rq |= !dl_task_fits_capacity(p, cpu);
1852 int target = find_later_rq(p);
1855 (dl_time_before(p->dl.deadline,
1856 cpu_rq(target)->dl.earliest_dl.curr) ||
1857 (cpu_rq(target)->dl.dl_nr_running == 0)))
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_running(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 (later_rq->dl.dl_nr_running &&
2225 !dl_time_before(task->dl.deadline,
2226 later_rq->dl.earliest_dl.curr)) {
2228 * Target rq has tasks of equal or earlier deadline,
2229 * retrying does not release any lock and is unlikely
2230 * to yield a different result.
2236 /* Retry if something changed. */
2237 if (double_lock_balance(rq, later_rq)) {
2238 if (unlikely(task_rq(task) != rq ||
2239 !cpumask_test_cpu(later_rq->cpu, &task->cpus_mask) ||
2240 task_running(rq, task) ||
2242 !task_on_rq_queued(task))) {
2243 double_unlock_balance(rq, later_rq);
2250 * If the rq we found has no -deadline task, or
2251 * its earliest one has a later deadline than our
2252 * task, the rq is a good one.
2254 if (!later_rq->dl.dl_nr_running ||
2255 dl_time_before(task->dl.deadline,
2256 later_rq->dl.earliest_dl.curr))
2259 /* Otherwise we try again. */
2260 double_unlock_balance(rq, later_rq);
2267 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2269 struct task_struct *p;
2271 if (!has_pushable_dl_tasks(rq))
2274 p = __node_2_pdl(rb_first_cached(&rq->dl.pushable_dl_tasks_root));
2276 WARN_ON_ONCE(rq->cpu != task_cpu(p));
2277 WARN_ON_ONCE(task_current(rq, p));
2278 WARN_ON_ONCE(p->nr_cpus_allowed <= 1);
2280 WARN_ON_ONCE(!task_on_rq_queued(p));
2281 WARN_ON_ONCE(!dl_task(p));
2287 * See if the non running -deadline tasks on this rq
2288 * can be sent to some other CPU where they can preempt
2289 * and start executing.
2291 static int push_dl_task(struct rq *rq)
2293 struct task_struct *next_task;
2294 struct rq *later_rq;
2297 if (!rq->dl.overloaded)
2300 next_task = pick_next_pushable_dl_task(rq);
2306 * If next_task preempts rq->curr, and rq->curr
2307 * can move away, it makes sense to just reschedule
2308 * without going further in pushing next_task.
2310 if (dl_task(rq->curr) &&
2311 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2312 rq->curr->nr_cpus_allowed > 1) {
2317 if (is_migration_disabled(next_task))
2320 if (WARN_ON(next_task == rq->curr))
2323 /* We might release rq lock */
2324 get_task_struct(next_task);
2326 /* Will lock the rq it'll find */
2327 later_rq = find_lock_later_rq(next_task, rq);
2329 struct task_struct *task;
2332 * We must check all this again, since
2333 * find_lock_later_rq releases rq->lock and it is
2334 * then possible that next_task has migrated.
2336 task = pick_next_pushable_dl_task(rq);
2337 if (task == next_task) {
2339 * The task is still there. We don't try
2340 * again, some other CPU will pull it when ready.
2349 put_task_struct(next_task);
2354 deactivate_task(rq, next_task, 0);
2355 set_task_cpu(next_task, later_rq->cpu);
2356 activate_task(later_rq, next_task, 0);
2359 resched_curr(later_rq);
2361 double_unlock_balance(rq, later_rq);
2364 put_task_struct(next_task);
2369 static void push_dl_tasks(struct rq *rq)
2371 /* push_dl_task() will return true if it moved a -deadline task */
2372 while (push_dl_task(rq))
2376 static void pull_dl_task(struct rq *this_rq)
2378 int this_cpu = this_rq->cpu, cpu;
2379 struct task_struct *p, *push_task;
2380 bool resched = false;
2382 u64 dmin = LONG_MAX;
2384 if (likely(!dl_overloaded(this_rq)))
2388 * Match the barrier from dl_set_overloaded; this guarantees that if we
2389 * see overloaded we must also see the dlo_mask bit.
2393 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2394 if (this_cpu == cpu)
2397 src_rq = cpu_rq(cpu);
2400 * It looks racy, abd it is! However, as in sched_rt.c,
2401 * we are fine with this.
2403 if (this_rq->dl.dl_nr_running &&
2404 dl_time_before(this_rq->dl.earliest_dl.curr,
2405 src_rq->dl.earliest_dl.next))
2408 /* Might drop this_rq->lock */
2410 double_lock_balance(this_rq, src_rq);
2413 * If there are no more pullable tasks on the
2414 * rq, we're done with it.
2416 if (src_rq->dl.dl_nr_running <= 1)
2419 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2422 * We found a task to be pulled if:
2423 * - it preempts our current (if there's one),
2424 * - it will preempt the last one we pulled (if any).
2426 if (p && dl_time_before(p->dl.deadline, dmin) &&
2427 (!this_rq->dl.dl_nr_running ||
2428 dl_time_before(p->dl.deadline,
2429 this_rq->dl.earliest_dl.curr))) {
2430 WARN_ON(p == src_rq->curr);
2431 WARN_ON(!task_on_rq_queued(p));
2434 * Then we pull iff p has actually an earlier
2435 * deadline than the current task of its runqueue.
2437 if (dl_time_before(p->dl.deadline,
2438 src_rq->curr->dl.deadline))
2441 if (is_migration_disabled(p)) {
2442 push_task = get_push_task(src_rq);
2444 deactivate_task(src_rq, p, 0);
2445 set_task_cpu(p, this_cpu);
2446 activate_task(this_rq, p, 0);
2447 dmin = p->dl.deadline;
2451 /* Is there any other task even earlier? */
2454 double_unlock_balance(this_rq, src_rq);
2457 raw_spin_rq_unlock(this_rq);
2458 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2459 push_task, &src_rq->push_work);
2460 raw_spin_rq_lock(this_rq);
2465 resched_curr(this_rq);
2469 * Since the task is not running and a reschedule is not going to happen
2470 * anytime soon on its runqueue, we try pushing it away now.
2472 static void task_woken_dl(struct rq *rq, struct task_struct *p)
2474 if (!task_running(rq, p) &&
2475 !test_tsk_need_resched(rq->curr) &&
2476 p->nr_cpus_allowed > 1 &&
2477 dl_task(rq->curr) &&
2478 (rq->curr->nr_cpus_allowed < 2 ||
2479 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2484 static void set_cpus_allowed_dl(struct task_struct *p,
2485 const struct cpumask *new_mask,
2488 struct root_domain *src_rd;
2491 WARN_ON_ONCE(!dl_task(p));
2496 * Migrating a SCHED_DEADLINE task between exclusive
2497 * cpusets (different root_domains) entails a bandwidth
2498 * update. We already made space for us in the destination
2499 * domain (see cpuset_can_attach()).
2501 if (!cpumask_intersects(src_rd->span, new_mask)) {
2502 struct dl_bw *src_dl_b;
2504 src_dl_b = dl_bw_of(cpu_of(rq));
2506 * We now free resources of the root_domain we are migrating
2507 * off. In the worst case, sched_setattr() may temporary fail
2508 * until we complete the update.
2510 raw_spin_lock(&src_dl_b->lock);
2511 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2512 raw_spin_unlock(&src_dl_b->lock);
2515 set_cpus_allowed_common(p, new_mask, flags);
2518 /* Assumes rq->lock is held */
2519 static void rq_online_dl(struct rq *rq)
2521 if (rq->dl.overloaded)
2522 dl_set_overload(rq);
2524 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2525 if (rq->dl.dl_nr_running > 0)
2526 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2529 /* Assumes rq->lock is held */
2530 static void rq_offline_dl(struct rq *rq)
2532 if (rq->dl.overloaded)
2533 dl_clear_overload(rq);
2535 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2536 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2539 void __init init_sched_dl_class(void)
2543 for_each_possible_cpu(i)
2544 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2545 GFP_KERNEL, cpu_to_node(i));
2548 void dl_add_task_root_domain(struct task_struct *p)
2554 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2556 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
2560 rq = __task_rq_lock(p, &rf);
2562 dl_b = &rq->rd->dl_bw;
2563 raw_spin_lock(&dl_b->lock);
2565 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2567 raw_spin_unlock(&dl_b->lock);
2569 task_rq_unlock(rq, p, &rf);
2572 void dl_clear_root_domain(struct root_domain *rd)
2574 unsigned long flags;
2576 raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
2577 rd->dl_bw.total_bw = 0;
2578 raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
2581 #endif /* CONFIG_SMP */
2583 static void switched_from_dl(struct rq *rq, struct task_struct *p)
2586 * task_non_contending() can start the "inactive timer" (if the 0-lag
2587 * time is in the future). If the task switches back to dl before
2588 * the "inactive timer" fires, it can continue to consume its current
2589 * runtime using its current deadline. If it stays outside of
2590 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2591 * will reset the task parameters.
2593 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2594 task_non_contending(p);
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)
2636 /* If p is not queued we will update its parameters at next wakeup. */
2637 if (!task_on_rq_queued(p)) {
2638 add_rq_bw(&p->dl, &rq->dl);
2643 if (rq->curr != p) {
2645 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
2646 deadline_queue_push_tasks(rq);
2648 if (dl_task(rq->curr))
2649 check_preempt_curr_dl(rq, p, 0);
2653 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2658 * If the scheduling parameters of a -deadline task changed,
2659 * a push or pull operation might be needed.
2661 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2664 if (task_on_rq_queued(p) || task_current(rq, p)) {
2667 * This might be too much, but unfortunately
2668 * we don't have the old deadline value, and
2669 * we can't argue if the task is increasing
2670 * or lowering its prio, so...
2672 if (!rq->dl.overloaded)
2673 deadline_queue_pull_task(rq);
2676 * If we now have a earlier deadline task than p,
2677 * then reschedule, provided p is still on this
2680 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
2684 * Again, we don't know if p has a earlier
2685 * or later deadline, so let's blindly set a
2686 * (maybe not needed) rescheduling point.
2689 #endif /* CONFIG_SMP */
2693 DEFINE_SCHED_CLASS(dl) = {
2695 .enqueue_task = enqueue_task_dl,
2696 .dequeue_task = dequeue_task_dl,
2697 .yield_task = yield_task_dl,
2699 .check_preempt_curr = check_preempt_curr_dl,
2701 .pick_next_task = pick_next_task_dl,
2702 .put_prev_task = put_prev_task_dl,
2703 .set_next_task = set_next_task_dl,
2706 .balance = balance_dl,
2707 .pick_task = pick_task_dl,
2708 .select_task_rq = select_task_rq_dl,
2709 .migrate_task_rq = migrate_task_rq_dl,
2710 .set_cpus_allowed = set_cpus_allowed_dl,
2711 .rq_online = rq_online_dl,
2712 .rq_offline = rq_offline_dl,
2713 .task_woken = task_woken_dl,
2714 .find_lock_rq = find_lock_later_rq,
2717 .task_tick = task_tick_dl,
2718 .task_fork = task_fork_dl,
2720 .prio_changed = prio_changed_dl,
2721 .switched_from = switched_from_dl,
2722 .switched_to = switched_to_dl,
2724 .update_curr = update_curr_dl,
2727 /* Used for dl_bw check and update, used under sched_rt_handler()::mutex */
2728 static u64 dl_generation;
2730 int sched_dl_global_validate(void)
2732 u64 runtime = global_rt_runtime();
2733 u64 period = global_rt_period();
2734 u64 new_bw = to_ratio(period, runtime);
2735 u64 gen = ++dl_generation;
2737 int cpu, cpus, ret = 0;
2738 unsigned long flags;
2741 * Here we want to check the bandwidth not being set to some
2742 * value smaller than the currently allocated bandwidth in
2743 * any of the root_domains.
2745 for_each_possible_cpu(cpu) {
2746 rcu_read_lock_sched();
2748 if (dl_bw_visited(cpu, gen))
2751 dl_b = dl_bw_of(cpu);
2752 cpus = dl_bw_cpus(cpu);
2754 raw_spin_lock_irqsave(&dl_b->lock, flags);
2755 if (new_bw * cpus < dl_b->total_bw)
2757 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2760 rcu_read_unlock_sched();
2769 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
2771 if (global_rt_runtime() == RUNTIME_INF) {
2772 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2773 dl_rq->extra_bw = 1 << BW_SHIFT;
2775 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2776 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2777 dl_rq->extra_bw = to_ratio(global_rt_period(),
2778 global_rt_runtime());
2782 void sched_dl_do_global(void)
2785 u64 gen = ++dl_generation;
2788 unsigned long flags;
2790 if (global_rt_runtime() != RUNTIME_INF)
2791 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2793 for_each_possible_cpu(cpu) {
2794 rcu_read_lock_sched();
2796 if (dl_bw_visited(cpu, gen)) {
2797 rcu_read_unlock_sched();
2801 dl_b = dl_bw_of(cpu);
2803 raw_spin_lock_irqsave(&dl_b->lock, flags);
2805 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2807 rcu_read_unlock_sched();
2808 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2813 * We must be sure that accepting a new task (or allowing changing the
2814 * parameters of an existing one) is consistent with the bandwidth
2815 * constraints. If yes, this function also accordingly updates the currently
2816 * allocated bandwidth to reflect the new situation.
2818 * This function is called while holding p's rq->lock.
2820 int sched_dl_overflow(struct task_struct *p, int policy,
2821 const struct sched_attr *attr)
2823 u64 period = attr->sched_period ?: attr->sched_deadline;
2824 u64 runtime = attr->sched_runtime;
2825 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2826 int cpus, err = -1, cpu = task_cpu(p);
2827 struct dl_bw *dl_b = dl_bw_of(cpu);
2830 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2833 /* !deadline task may carry old deadline bandwidth */
2834 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2838 * Either if a task, enters, leave, or stays -deadline but changes
2839 * its parameters, we may need to update accordingly the total
2840 * allocated bandwidth of the container.
2842 raw_spin_lock(&dl_b->lock);
2843 cpus = dl_bw_cpus(cpu);
2844 cap = dl_bw_capacity(cpu);
2846 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2847 !__dl_overflow(dl_b, cap, 0, new_bw)) {
2848 if (hrtimer_active(&p->dl.inactive_timer))
2849 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2850 __dl_add(dl_b, new_bw, cpus);
2852 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2853 !__dl_overflow(dl_b, cap, p->dl.dl_bw, new_bw)) {
2855 * XXX this is slightly incorrect: when the task
2856 * utilization decreases, we should delay the total
2857 * utilization change until the task's 0-lag point.
2858 * But this would require to set the task's "inactive
2859 * timer" when the task is not inactive.
2861 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2862 __dl_add(dl_b, new_bw, cpus);
2863 dl_change_utilization(p, new_bw);
2865 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2867 * Do not decrease the total deadline utilization here,
2868 * switched_from_dl() will take care to do it at the correct
2873 raw_spin_unlock(&dl_b->lock);
2879 * This function initializes the sched_dl_entity of a newly becoming
2880 * SCHED_DEADLINE task.
2882 * Only the static values are considered here, the actual runtime and the
2883 * absolute deadline will be properly calculated when the task is enqueued
2884 * for the first time with its new policy.
2886 void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
2888 struct sched_dl_entity *dl_se = &p->dl;
2890 dl_se->dl_runtime = attr->sched_runtime;
2891 dl_se->dl_deadline = attr->sched_deadline;
2892 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
2893 dl_se->flags = attr->sched_flags & SCHED_DL_FLAGS;
2894 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
2895 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
2898 void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
2900 struct sched_dl_entity *dl_se = &p->dl;
2902 attr->sched_priority = p->rt_priority;
2903 attr->sched_runtime = dl_se->dl_runtime;
2904 attr->sched_deadline = dl_se->dl_deadline;
2905 attr->sched_period = dl_se->dl_period;
2906 attr->sched_flags &= ~SCHED_DL_FLAGS;
2907 attr->sched_flags |= dl_se->flags;
2911 * This function validates the new parameters of a -deadline task.
2912 * We ask for the deadline not being zero, and greater or equal
2913 * than the runtime, as well as the period of being zero or
2914 * greater than deadline. Furthermore, we have to be sure that
2915 * user parameters are above the internal resolution of 1us (we
2916 * check sched_runtime only since it is always the smaller one) and
2917 * below 2^63 ns (we have to check both sched_deadline and
2918 * sched_period, as the latter can be zero).
2920 bool __checkparam_dl(const struct sched_attr *attr)
2922 u64 period, max, min;
2924 /* special dl tasks don't actually use any parameter */
2925 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2929 if (attr->sched_deadline == 0)
2933 * Since we truncate DL_SCALE bits, make sure we're at least
2936 if (attr->sched_runtime < (1ULL << DL_SCALE))
2940 * Since we use the MSB for wrap-around and sign issues, make
2941 * sure it's not set (mind that period can be equal to zero).
2943 if (attr->sched_deadline & (1ULL << 63) ||
2944 attr->sched_period & (1ULL << 63))
2947 period = attr->sched_period;
2949 period = attr->sched_deadline;
2951 /* runtime <= deadline <= period (if period != 0) */
2952 if (period < attr->sched_deadline ||
2953 attr->sched_deadline < attr->sched_runtime)
2956 max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
2957 min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
2959 if (period < min || period > max)
2966 * This function clears the sched_dl_entity static params.
2968 void __dl_clear_params(struct task_struct *p)
2970 struct sched_dl_entity *dl_se = &p->dl;
2972 dl_se->dl_runtime = 0;
2973 dl_se->dl_deadline = 0;
2974 dl_se->dl_period = 0;
2977 dl_se->dl_density = 0;
2979 dl_se->dl_throttled = 0;
2980 dl_se->dl_yielded = 0;
2981 dl_se->dl_non_contending = 0;
2982 dl_se->dl_overrun = 0;
2984 #ifdef CONFIG_RT_MUTEXES
2985 dl_se->pi_se = dl_se;
2989 bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
2991 struct sched_dl_entity *dl_se = &p->dl;
2993 if (dl_se->dl_runtime != attr->sched_runtime ||
2994 dl_se->dl_deadline != attr->sched_deadline ||
2995 dl_se->dl_period != attr->sched_period ||
2996 dl_se->flags != (attr->sched_flags & SCHED_DL_FLAGS))
3003 int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
3004 const struct cpumask *trial)
3006 unsigned long flags, cap;
3007 struct dl_bw *cur_dl_b;
3010 rcu_read_lock_sched();
3011 cur_dl_b = dl_bw_of(cpumask_any(cur));
3012 cap = __dl_bw_capacity(trial);
3013 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
3014 if (__dl_overflow(cur_dl_b, cap, 0, 0))
3016 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
3017 rcu_read_unlock_sched();
3022 int dl_cpu_busy(int cpu, struct task_struct *p)
3024 unsigned long flags, cap;
3028 rcu_read_lock_sched();
3029 dl_b = dl_bw_of(cpu);
3030 raw_spin_lock_irqsave(&dl_b->lock, flags);
3031 cap = dl_bw_capacity(cpu);
3032 overflow = __dl_overflow(dl_b, cap, 0, p ? p->dl.dl_bw : 0);
3034 if (!overflow && p) {
3036 * We reserve space for this task in the destination
3037 * root_domain, as we can't fail after this point.
3038 * We will free resources in the source root_domain
3039 * later on (see set_cpus_allowed_dl()).
3041 __dl_add(dl_b, p->dl.dl_bw, dl_bw_cpus(cpu));
3044 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3045 rcu_read_unlock_sched();
3047 return overflow ? -EBUSY : 0;
3051 #ifdef CONFIG_SCHED_DEBUG
3052 void print_dl_stats(struct seq_file *m, int cpu)
3054 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
3056 #endif /* CONFIG_SCHED_DEBUG */