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 static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
21 return container_of(dl_se, struct task_struct, dl);
24 static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
26 return container_of(dl_rq, struct rq, dl);
29 static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
31 struct task_struct *p = dl_task_of(dl_se);
32 struct rq *rq = task_rq(p);
37 static inline int on_dl_rq(struct sched_dl_entity *dl_se)
39 return !RB_EMPTY_NODE(&dl_se->rb_node);
42 #ifdef CONFIG_RT_MUTEXES
43 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
48 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
50 return pi_of(dl_se) != dl_se;
53 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
58 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
65 static inline struct dl_bw *dl_bw_of(int i)
67 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
68 "sched RCU must be held");
69 return &cpu_rq(i)->rd->dl_bw;
72 static inline int dl_bw_cpus(int i)
74 struct root_domain *rd = cpu_rq(i)->rd;
77 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
78 "sched RCU must be held");
80 if (cpumask_subset(rd->span, cpu_active_mask))
81 return cpumask_weight(rd->span);
85 for_each_cpu_and(i, rd->span, cpu_active_mask)
91 static inline unsigned long __dl_bw_capacity(int i)
93 struct root_domain *rd = cpu_rq(i)->rd;
94 unsigned long cap = 0;
96 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
97 "sched RCU must be held");
99 for_each_cpu_and(i, rd->span, cpu_active_mask)
100 cap += capacity_orig_of(i);
106 * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
107 * of the CPU the task is running on rather rd's \Sum CPU capacity.
109 static inline unsigned long dl_bw_capacity(int i)
111 if (!static_branch_unlikely(&sched_asym_cpucapacity) &&
112 capacity_orig_of(i) == SCHED_CAPACITY_SCALE) {
113 return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT;
115 return __dl_bw_capacity(i);
119 static inline bool dl_bw_visited(int cpu, u64 gen)
121 struct root_domain *rd = cpu_rq(cpu)->rd;
123 if (rd->visit_gen == gen)
131 void __dl_update(struct dl_bw *dl_b, s64 bw)
133 struct root_domain *rd = container_of(dl_b, struct root_domain, dl_bw);
136 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
137 "sched RCU must be held");
138 for_each_cpu_and(i, rd->span, cpu_active_mask) {
139 struct rq *rq = cpu_rq(i);
141 rq->dl.extra_bw += bw;
145 static inline struct dl_bw *dl_bw_of(int i)
147 return &cpu_rq(i)->dl.dl_bw;
150 static inline int dl_bw_cpus(int i)
155 static inline unsigned long dl_bw_capacity(int i)
157 return SCHED_CAPACITY_SCALE;
160 static inline bool dl_bw_visited(int cpu, u64 gen)
166 void __dl_update(struct dl_bw *dl_b, s64 bw)
168 struct dl_rq *dl = container_of(dl_b, struct dl_rq, dl_bw);
175 void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
177 dl_b->total_bw -= tsk_bw;
178 __dl_update(dl_b, (s32)tsk_bw / cpus);
182 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, int cpus)
184 dl_b->total_bw += tsk_bw;
185 __dl_update(dl_b, -((s32)tsk_bw / cpus));
189 __dl_overflow(struct dl_bw *dl_b, unsigned long cap, u64 old_bw, u64 new_bw)
191 return dl_b->bw != -1 &&
192 cap_scale(dl_b->bw, cap) < dl_b->total_bw - old_bw + new_bw;
196 void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
198 u64 old = dl_rq->running_bw;
200 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
201 dl_rq->running_bw += dl_bw;
202 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
203 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
204 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
205 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
209 void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
211 u64 old = dl_rq->running_bw;
213 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
214 dl_rq->running_bw -= dl_bw;
215 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
216 if (dl_rq->running_bw > old)
217 dl_rq->running_bw = 0;
218 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
219 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
223 void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
225 u64 old = dl_rq->this_bw;
227 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
228 dl_rq->this_bw += dl_bw;
229 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
233 void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
235 u64 old = dl_rq->this_bw;
237 lockdep_assert_rq_held(rq_of_dl_rq(dl_rq));
238 dl_rq->this_bw -= dl_bw;
239 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
240 if (dl_rq->this_bw > old)
242 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
246 void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
248 if (!dl_entity_is_special(dl_se))
249 __add_rq_bw(dl_se->dl_bw, dl_rq);
253 void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
255 if (!dl_entity_is_special(dl_se))
256 __sub_rq_bw(dl_se->dl_bw, dl_rq);
260 void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
262 if (!dl_entity_is_special(dl_se))
263 __add_running_bw(dl_se->dl_bw, dl_rq);
267 void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
269 if (!dl_entity_is_special(dl_se))
270 __sub_running_bw(dl_se->dl_bw, dl_rq);
273 static void dl_change_utilization(struct task_struct *p, u64 new_bw)
277 BUG_ON(p->dl.flags & SCHED_FLAG_SUGOV);
279 if (task_on_rq_queued(p))
283 if (p->dl.dl_non_contending) {
284 sub_running_bw(&p->dl, &rq->dl);
285 p->dl.dl_non_contending = 0;
287 * If the timer handler is currently running and the
288 * timer cannot be canceled, inactive_task_timer()
289 * will see that dl_not_contending is not set, and
290 * will not touch the rq's active utilization,
291 * so we are still safe.
293 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
296 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
297 __add_rq_bw(new_bw, &rq->dl);
301 * The utilization of a task cannot be immediately removed from
302 * the rq active utilization (running_bw) when the task blocks.
303 * Instead, we have to wait for the so called "0-lag time".
305 * If a task blocks before the "0-lag time", a timer (the inactive
306 * timer) is armed, and running_bw is decreased when the timer
309 * If the task wakes up again before the inactive timer fires,
310 * the timer is canceled, whereas if the task wakes up after the
311 * inactive timer fired (and running_bw has been decreased) the
312 * task's utilization has to be added to running_bw again.
313 * A flag in the deadline scheduling entity (dl_non_contending)
314 * is used to avoid race conditions between the inactive timer handler
317 * The following diagram shows how running_bw is updated. A task is
318 * "ACTIVE" when its utilization contributes to running_bw; an
319 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
320 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
321 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
322 * time already passed, which does not contribute to running_bw anymore.
323 * +------------------+
325 * +------------------>+ contending |
326 * | add_running_bw | |
327 * | +----+------+------+
330 * +--------+-------+ | |
331 * | | t >= 0-lag | | wakeup
332 * | INACTIVE |<---------------+ |
333 * | | sub_running_bw | |
334 * +--------+-------+ | |
339 * | +----+------+------+
340 * | sub_running_bw | ACTIVE |
341 * +-------------------+ |
342 * inactive timer | non contending |
343 * fired +------------------+
345 * The task_non_contending() function is invoked when a task
346 * blocks, and checks if the 0-lag time already passed or
347 * not (in the first case, it directly updates running_bw;
348 * in the second case, it arms the inactive timer).
350 * The task_contending() function is invoked when a task wakes
351 * up, and checks if the task is still in the "ACTIVE non contending"
352 * state or not (in the second case, it updates running_bw).
354 static void task_non_contending(struct task_struct *p)
356 struct sched_dl_entity *dl_se = &p->dl;
357 struct hrtimer *timer = &dl_se->inactive_timer;
358 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
359 struct rq *rq = rq_of_dl_rq(dl_rq);
363 * If this is a non-deadline task that has been boosted,
366 if (dl_se->dl_runtime == 0)
369 if (dl_entity_is_special(dl_se))
372 WARN_ON(dl_se->dl_non_contending);
374 zerolag_time = dl_se->deadline -
375 div64_long((dl_se->runtime * dl_se->dl_period),
379 * Using relative times instead of the absolute "0-lag time"
380 * allows to simplify the code
382 zerolag_time -= rq_clock(rq);
385 * If the "0-lag time" already passed, decrease the active
386 * utilization now, instead of starting a timer
388 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
390 sub_running_bw(dl_se, dl_rq);
391 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
392 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
394 if (READ_ONCE(p->__state) == TASK_DEAD)
395 sub_rq_bw(&p->dl, &rq->dl);
396 raw_spin_lock(&dl_b->lock);
397 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
398 __dl_clear_params(p);
399 raw_spin_unlock(&dl_b->lock);
405 dl_se->dl_non_contending = 1;
407 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD);
410 static void task_contending(struct sched_dl_entity *dl_se, int flags)
412 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
415 * If this is a non-deadline task that has been boosted,
418 if (dl_se->dl_runtime == 0)
421 if (flags & ENQUEUE_MIGRATED)
422 add_rq_bw(dl_se, dl_rq);
424 if (dl_se->dl_non_contending) {
425 dl_se->dl_non_contending = 0;
427 * If the timer handler is currently running and the
428 * timer cannot be canceled, inactive_task_timer()
429 * will see that dl_not_contending is not set, and
430 * will not touch the rq's active utilization,
431 * so we are still safe.
433 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1)
434 put_task_struct(dl_task_of(dl_se));
437 * Since "dl_non_contending" is not set, the
438 * task's utilization has already been removed from
439 * active utilization (either when the task blocked,
440 * when the "inactive timer" fired).
443 add_running_bw(dl_se, dl_rq);
447 static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
449 struct sched_dl_entity *dl_se = &p->dl;
451 return rb_first_cached(&dl_rq->root) == &dl_se->rb_node;
454 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
456 void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
458 raw_spin_lock_init(&dl_b->dl_runtime_lock);
459 dl_b->dl_period = period;
460 dl_b->dl_runtime = runtime;
463 void init_dl_bw(struct dl_bw *dl_b)
465 raw_spin_lock_init(&dl_b->lock);
466 if (global_rt_runtime() == RUNTIME_INF)
469 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
473 void init_dl_rq(struct dl_rq *dl_rq)
475 dl_rq->root = RB_ROOT_CACHED;
478 /* zero means no -deadline tasks */
479 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
481 dl_rq->dl_nr_migratory = 0;
482 dl_rq->overloaded = 0;
483 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
485 init_dl_bw(&dl_rq->dl_bw);
488 dl_rq->running_bw = 0;
490 init_dl_rq_bw_ratio(dl_rq);
495 static inline int dl_overloaded(struct rq *rq)
497 return atomic_read(&rq->rd->dlo_count);
500 static inline void dl_set_overload(struct rq *rq)
505 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
507 * Must be visible before the overload count is
508 * set (as in sched_rt.c).
510 * Matched by the barrier in pull_dl_task().
513 atomic_inc(&rq->rd->dlo_count);
516 static inline void dl_clear_overload(struct rq *rq)
521 atomic_dec(&rq->rd->dlo_count);
522 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
525 static void update_dl_migration(struct dl_rq *dl_rq)
527 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
528 if (!dl_rq->overloaded) {
529 dl_set_overload(rq_of_dl_rq(dl_rq));
530 dl_rq->overloaded = 1;
532 } else if (dl_rq->overloaded) {
533 dl_clear_overload(rq_of_dl_rq(dl_rq));
534 dl_rq->overloaded = 0;
538 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
540 struct task_struct *p = dl_task_of(dl_se);
542 if (p->nr_cpus_allowed > 1)
543 dl_rq->dl_nr_migratory++;
545 update_dl_migration(dl_rq);
548 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
550 struct task_struct *p = dl_task_of(dl_se);
552 if (p->nr_cpus_allowed > 1)
553 dl_rq->dl_nr_migratory--;
555 update_dl_migration(dl_rq);
558 #define __node_2_pdl(node) \
559 rb_entry((node), struct task_struct, pushable_dl_tasks)
561 static inline bool __pushable_less(struct rb_node *a, const struct rb_node *b)
563 return dl_entity_preempt(&__node_2_pdl(a)->dl, &__node_2_pdl(b)->dl);
567 * The list of pushable -deadline task is not a plist, like in
568 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
570 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
572 struct rb_node *leftmost;
574 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
576 leftmost = rb_add_cached(&p->pushable_dl_tasks,
577 &rq->dl.pushable_dl_tasks_root,
580 rq->dl.earliest_dl.next = p->dl.deadline;
583 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
585 struct dl_rq *dl_rq = &rq->dl;
586 struct rb_root_cached *root = &dl_rq->pushable_dl_tasks_root;
587 struct rb_node *leftmost;
589 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
592 leftmost = rb_erase_cached(&p->pushable_dl_tasks, root);
594 dl_rq->earliest_dl.next = __node_2_pdl(leftmost)->dl.deadline;
596 RB_CLEAR_NODE(&p->pushable_dl_tasks);
599 static inline int has_pushable_dl_tasks(struct rq *rq)
601 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
604 static int push_dl_task(struct rq *rq);
606 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
608 return rq->online && dl_task(prev);
611 static DEFINE_PER_CPU(struct callback_head, dl_push_head);
612 static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
614 static void push_dl_tasks(struct rq *);
615 static void pull_dl_task(struct rq *);
617 static inline void deadline_queue_push_tasks(struct rq *rq)
619 if (!has_pushable_dl_tasks(rq))
622 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
625 static inline void deadline_queue_pull_task(struct rq *rq)
627 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
630 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
632 static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
634 struct rq *later_rq = NULL;
637 later_rq = find_lock_later_rq(p, rq);
642 * If we cannot preempt any rq, fall back to pick any
645 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
646 if (cpu >= nr_cpu_ids) {
648 * Failed to find any suitable CPU.
649 * The task will never come back!
651 BUG_ON(dl_bandwidth_enabled());
654 * If admission control is disabled we
655 * try a little harder to let the task
658 cpu = cpumask_any(cpu_active_mask);
660 later_rq = cpu_rq(cpu);
661 double_lock_balance(rq, later_rq);
664 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
666 * Inactive timer is armed (or callback is running, but
667 * waiting for us to release rq locks). In any case, when it
668 * will fire (or continue), it will see running_bw of this
669 * task migrated to later_rq (and correctly handle it).
671 sub_running_bw(&p->dl, &rq->dl);
672 sub_rq_bw(&p->dl, &rq->dl);
674 add_rq_bw(&p->dl, &later_rq->dl);
675 add_running_bw(&p->dl, &later_rq->dl);
677 sub_rq_bw(&p->dl, &rq->dl);
678 add_rq_bw(&p->dl, &later_rq->dl);
682 * And we finally need to fixup root_domain(s) bandwidth accounting,
683 * since p is still hanging out in the old (now moved to default) root
686 dl_b = &rq->rd->dl_bw;
687 raw_spin_lock(&dl_b->lock);
688 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
689 raw_spin_unlock(&dl_b->lock);
691 dl_b = &later_rq->rd->dl_bw;
692 raw_spin_lock(&dl_b->lock);
693 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
694 raw_spin_unlock(&dl_b->lock);
696 set_task_cpu(p, later_rq->cpu);
697 double_unlock_balance(later_rq, rq);
705 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
710 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
715 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
720 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
724 static inline void deadline_queue_push_tasks(struct rq *rq)
728 static inline void deadline_queue_pull_task(struct rq *rq)
731 #endif /* CONFIG_SMP */
733 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
734 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
735 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags);
738 * We are being explicitly informed that a new instance is starting,
739 * and this means that:
740 * - the absolute deadline of the entity has to be placed at
741 * current time + relative deadline;
742 * - the runtime of the entity has to be set to the maximum value.
744 * The capability of specifying such event is useful whenever a -deadline
745 * entity wants to (try to!) synchronize its behaviour with the scheduler's
746 * one, and to (try to!) reconcile itself with its own scheduling
749 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
751 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
752 struct rq *rq = rq_of_dl_rq(dl_rq);
754 WARN_ON(is_dl_boosted(dl_se));
755 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
758 * We are racing with the deadline timer. So, do nothing because
759 * the deadline timer handler will take care of properly recharging
760 * the runtime and postponing the deadline
762 if (dl_se->dl_throttled)
766 * We use the regular wall clock time to set deadlines in the
767 * future; in fact, we must consider execution overheads (time
768 * spent on hardirq context, etc.).
770 dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline;
771 dl_se->runtime = dl_se->dl_runtime;
775 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
776 * possibility of a entity lasting more than what it declared, and thus
777 * exhausting its runtime.
779 * Here we are interested in making runtime overrun possible, but we do
780 * not want a entity which is misbehaving to affect the scheduling of all
782 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
783 * is used, in order to confine each entity within its own bandwidth.
785 * This function deals exactly with that, and ensures that when the runtime
786 * of a entity is replenished, its deadline is also postponed. That ensures
787 * the overrunning entity can't interfere with other entity in the system and
788 * can't make them miss their deadlines. Reasons why this kind of overruns
789 * could happen are, typically, a entity voluntarily trying to overcome its
790 * runtime, or it just underestimated it during sched_setattr().
792 static void replenish_dl_entity(struct sched_dl_entity *dl_se)
794 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
795 struct rq *rq = rq_of_dl_rq(dl_rq);
797 BUG_ON(pi_of(dl_se)->dl_runtime <= 0);
800 * This could be the case for a !-dl task that is boosted.
801 * Just go with full inherited parameters.
803 if (dl_se->dl_deadline == 0) {
804 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
805 dl_se->runtime = pi_of(dl_se)->dl_runtime;
808 if (dl_se->dl_yielded && dl_se->runtime > 0)
812 * We keep moving the deadline away until we get some
813 * available runtime for the entity. This ensures correct
814 * handling of situations where the runtime overrun is
817 while (dl_se->runtime <= 0) {
818 dl_se->deadline += pi_of(dl_se)->dl_period;
819 dl_se->runtime += pi_of(dl_se)->dl_runtime;
823 * At this point, the deadline really should be "in
824 * the future" with respect to rq->clock. If it's
825 * not, we are, for some reason, lagging too much!
826 * Anyway, after having warn userspace abut that,
827 * we still try to keep the things running by
828 * resetting the deadline and the budget of the
831 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
832 printk_deferred_once("sched: DL replenish lagged too much\n");
833 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
834 dl_se->runtime = pi_of(dl_se)->dl_runtime;
837 if (dl_se->dl_yielded)
838 dl_se->dl_yielded = 0;
839 if (dl_se->dl_throttled)
840 dl_se->dl_throttled = 0;
844 * Here we check if --at time t-- an entity (which is probably being
845 * [re]activated or, in general, enqueued) can use its remaining runtime
846 * and its current deadline _without_ exceeding the bandwidth it is
847 * assigned (function returns true if it can't). We are in fact applying
848 * one of the CBS rules: when a task wakes up, if the residual runtime
849 * over residual deadline fits within the allocated bandwidth, then we
850 * can keep the current (absolute) deadline and residual budget without
851 * disrupting the schedulability of the system. Otherwise, we should
852 * refill the runtime and set the deadline a period in the future,
853 * because keeping the current (absolute) deadline of the task would
854 * result in breaking guarantees promised to other tasks (refer to
855 * Documentation/scheduler/sched-deadline.rst for more information).
857 * This function returns true if:
859 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
861 * IOW we can't recycle current parameters.
863 * Notice that the bandwidth check is done against the deadline. For
864 * task with deadline equal to period this is the same of using
865 * dl_period instead of dl_deadline in the equation above.
867 static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t)
872 * left and right are the two sides of the equation above,
873 * after a bit of shuffling to use multiplications instead
876 * Note that none of the time values involved in the two
877 * multiplications are absolute: dl_deadline and dl_runtime
878 * are the relative deadline and the maximum runtime of each
879 * instance, runtime is the runtime left for the last instance
880 * and (deadline - t), since t is rq->clock, is the time left
881 * to the (absolute) deadline. Even if overflowing the u64 type
882 * is very unlikely to occur in both cases, here we scale down
883 * as we want to avoid that risk at all. Scaling down by 10
884 * means that we reduce granularity to 1us. We are fine with it,
885 * since this is only a true/false check and, anyway, thinking
886 * of anything below microseconds resolution is actually fiction
887 * (but still we want to give the user that illusion >;).
889 left = (pi_of(dl_se)->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
890 right = ((dl_se->deadline - t) >> DL_SCALE) *
891 (pi_of(dl_se)->dl_runtime >> DL_SCALE);
893 return dl_time_before(right, left);
897 * Revised wakeup rule [1]: For self-suspending tasks, rather then
898 * re-initializing task's runtime and deadline, the revised wakeup
899 * rule adjusts the task's runtime to avoid the task to overrun its
902 * Reasoning: a task may overrun the density if:
903 * runtime / (deadline - t) > dl_runtime / dl_deadline
905 * Therefore, runtime can be adjusted to:
906 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
908 * In such way that runtime will be equal to the maximum density
909 * the task can use without breaking any rule.
911 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
912 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
915 update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
917 u64 laxity = dl_se->deadline - rq_clock(rq);
920 * If the task has deadline < period, and the deadline is in the past,
921 * it should already be throttled before this check.
923 * See update_dl_entity() comments for further details.
925 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
927 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
931 * Regarding the deadline, a task with implicit deadline has a relative
932 * deadline == relative period. A task with constrained deadline has a
933 * relative deadline <= relative period.
935 * We support constrained deadline tasks. However, there are some restrictions
936 * applied only for tasks which do not have an implicit deadline. See
937 * update_dl_entity() to know more about such restrictions.
939 * The dl_is_implicit() returns true if the task has an implicit deadline.
941 static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
943 return dl_se->dl_deadline == dl_se->dl_period;
947 * When a deadline entity is placed in the runqueue, its runtime and deadline
948 * might need to be updated. This is done by a CBS wake up rule. There are two
949 * different rules: 1) the original CBS; and 2) the Revisited CBS.
951 * When the task is starting a new period, the Original CBS is used. In this
952 * case, the runtime is replenished and a new absolute deadline is set.
954 * When a task is queued before the begin of the next period, using the
955 * remaining runtime and deadline could make the entity to overflow, see
956 * dl_entity_overflow() to find more about runtime overflow. When such case
957 * is detected, the runtime and deadline need to be updated.
959 * If the task has an implicit deadline, i.e., deadline == period, the Original
960 * CBS is applied. the runtime is replenished and a new absolute deadline is
961 * set, as in the previous cases.
963 * However, the Original CBS does not work properly for tasks with
964 * deadline < period, which are said to have a constrained deadline. By
965 * applying the Original CBS, a constrained deadline task would be able to run
966 * runtime/deadline in a period. With deadline < period, the task would
967 * overrun the runtime/period allowed bandwidth, breaking the admission test.
969 * In order to prevent this misbehave, the Revisited CBS is used for
970 * constrained deadline tasks when a runtime overflow is detected. In the
971 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
972 * the remaining runtime of the task is reduced to avoid runtime overflow.
973 * Please refer to the comments update_dl_revised_wakeup() function to find
974 * more about the Revised CBS rule.
976 static void update_dl_entity(struct sched_dl_entity *dl_se)
978 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
979 struct rq *rq = rq_of_dl_rq(dl_rq);
981 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
982 dl_entity_overflow(dl_se, rq_clock(rq))) {
984 if (unlikely(!dl_is_implicit(dl_se) &&
985 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
986 !is_dl_boosted(dl_se))) {
987 update_dl_revised_wakeup(dl_se, rq);
991 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
992 dl_se->runtime = pi_of(dl_se)->dl_runtime;
996 static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
998 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
1002 * If the entity depleted all its runtime, and if we want it to sleep
1003 * while waiting for some new execution time to become available, we
1004 * set the bandwidth replenishment timer to the replenishment instant
1005 * and try to activate it.
1007 * Notice that it is important for the caller to know if the timer
1008 * actually started or not (i.e., the replenishment instant is in
1009 * the future or in the past).
1011 static int start_dl_timer(struct task_struct *p)
1013 struct sched_dl_entity *dl_se = &p->dl;
1014 struct hrtimer *timer = &dl_se->dl_timer;
1015 struct rq *rq = task_rq(p);
1019 lockdep_assert_rq_held(rq);
1022 * We want the timer to fire at the deadline, but considering
1023 * that it is actually coming from rq->clock and not from
1024 * hrtimer's time base reading.
1026 act = ns_to_ktime(dl_next_period(dl_se));
1027 now = hrtimer_cb_get_time(timer);
1028 delta = ktime_to_ns(now) - rq_clock(rq);
1029 act = ktime_add_ns(act, delta);
1032 * If the expiry time already passed, e.g., because the value
1033 * chosen as the deadline is too small, don't even try to
1034 * start the timer in the past!
1036 if (ktime_us_delta(act, now) < 0)
1040 * !enqueued will guarantee another callback; even if one is already in
1041 * progress. This ensures a balanced {get,put}_task_struct().
1043 * The race against __run_timer() clearing the enqueued state is
1044 * harmless because we're holding task_rq()->lock, therefore the timer
1045 * expiring after we've done the check will wait on its task_rq_lock()
1046 * and observe our state.
1048 if (!hrtimer_is_queued(timer)) {
1050 hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD);
1057 * This is the bandwidth enforcement timer callback. If here, we know
1058 * a task is not on its dl_rq, since the fact that the timer was running
1059 * means the task is throttled and needs a runtime replenishment.
1061 * However, what we actually do depends on the fact the task is active,
1062 * (it is on its rq) or has been removed from there by a call to
1063 * dequeue_task_dl(). In the former case we must issue the runtime
1064 * replenishment and add the task back to the dl_rq; in the latter, we just
1065 * do nothing but clearing dl_throttled, so that runtime and deadline
1066 * updating (and the queueing back to dl_rq) will be done by the
1067 * next call to enqueue_task_dl().
1069 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
1071 struct sched_dl_entity *dl_se = container_of(timer,
1072 struct sched_dl_entity,
1074 struct task_struct *p = dl_task_of(dl_se);
1078 rq = task_rq_lock(p, &rf);
1081 * The task might have changed its scheduling policy to something
1082 * different than SCHED_DEADLINE (through switched_from_dl()).
1088 * The task might have been boosted by someone else and might be in the
1089 * boosting/deboosting path, its not throttled.
1091 if (is_dl_boosted(dl_se))
1095 * Spurious timer due to start_dl_timer() race; or we already received
1096 * a replenishment from rt_mutex_setprio().
1098 if (!dl_se->dl_throttled)
1102 update_rq_clock(rq);
1105 * If the throttle happened during sched-out; like:
1112 * __dequeue_task_dl()
1115 * We can be both throttled and !queued. Replenish the counter
1116 * but do not enqueue -- wait for our wakeup to do that.
1118 if (!task_on_rq_queued(p)) {
1119 replenish_dl_entity(dl_se);
1124 if (unlikely(!rq->online)) {
1126 * If the runqueue is no longer available, migrate the
1127 * task elsewhere. This necessarily changes rq.
1129 lockdep_unpin_lock(__rq_lockp(rq), rf.cookie);
1130 rq = dl_task_offline_migration(rq, p);
1131 rf.cookie = lockdep_pin_lock(__rq_lockp(rq));
1132 update_rq_clock(rq);
1135 * Now that the task has been migrated to the new RQ and we
1136 * have that locked, proceed as normal and enqueue the task
1142 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1143 if (dl_task(rq->curr))
1144 check_preempt_curr_dl(rq, p, 0);
1150 * Queueing this task back might have overloaded rq, check if we need
1151 * to kick someone away.
1153 if (has_pushable_dl_tasks(rq)) {
1155 * Nothing relies on rq->lock after this, so its safe to drop
1158 rq_unpin_lock(rq, &rf);
1160 rq_repin_lock(rq, &rf);
1165 task_rq_unlock(rq, p, &rf);
1168 * This can free the task_struct, including this hrtimer, do not touch
1169 * anything related to that after this.
1173 return HRTIMER_NORESTART;
1176 void init_dl_task_timer(struct sched_dl_entity *dl_se)
1178 struct hrtimer *timer = &dl_se->dl_timer;
1180 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1181 timer->function = dl_task_timer;
1185 * During the activation, CBS checks if it can reuse the current task's
1186 * runtime and period. If the deadline of the task is in the past, CBS
1187 * cannot use the runtime, and so it replenishes the task. This rule
1188 * works fine for implicit deadline tasks (deadline == period), and the
1189 * CBS was designed for implicit deadline tasks. However, a task with
1190 * constrained deadline (deadline < period) might be awakened after the
1191 * deadline, but before the next period. In this case, replenishing the
1192 * task would allow it to run for runtime / deadline. As in this case
1193 * deadline < period, CBS enables a task to run for more than the
1194 * runtime / period. In a very loaded system, this can cause a domino
1195 * effect, making other tasks miss their deadlines.
1197 * To avoid this problem, in the activation of a constrained deadline
1198 * task after the deadline but before the next period, throttle the
1199 * task and set the replenishing timer to the begin of the next period,
1200 * unless it is boosted.
1202 static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1204 struct task_struct *p = dl_task_of(dl_se);
1205 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1207 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1208 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1209 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(p)))
1211 dl_se->dl_throttled = 1;
1212 if (dl_se->runtime > 0)
1218 int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1220 return (dl_se->runtime <= 0);
1224 * This function implements the GRUB accounting rule:
1225 * according to the GRUB reclaiming algorithm, the runtime is
1226 * not decreased as "dq = -dt", but as
1227 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1228 * where u is the utilization of the task, Umax is the maximum reclaimable
1229 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1230 * as the difference between the "total runqueue utilization" and the
1231 * runqueue active utilization, and Uextra is the (per runqueue) extra
1232 * reclaimable utilization.
1233 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1234 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1236 * Since rq->dl.bw_ratio contains 1 / Umax multiplied by 2^RATIO_SHIFT,
1237 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1238 * Since delta is a 64 bit variable, to have an overflow its value
1239 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
1240 * So, overflow is not an issue here.
1242 static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1244 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1246 u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT;
1249 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1250 * we compare u_inact + rq->dl.extra_bw with
1251 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1252 * u_inact + rq->dl.extra_bw can be larger than
1253 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
1254 * leading to wrong results)
1256 if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min)
1259 u_act = BW_UNIT - u_inact - rq->dl.extra_bw;
1261 return (delta * u_act) >> BW_SHIFT;
1265 * Update the current task's runtime statistics (provided it is still
1266 * a -deadline task and has not been removed from the dl_rq).
1268 static void update_curr_dl(struct rq *rq)
1270 struct task_struct *curr = rq->curr;
1271 struct sched_dl_entity *dl_se = &curr->dl;
1272 u64 delta_exec, scaled_delta_exec;
1273 int cpu = cpu_of(rq);
1276 if (!dl_task(curr) || !on_dl_rq(dl_se))
1280 * Consumed budget is computed considering the time as
1281 * observed by schedulable tasks (excluding time spent
1282 * in hardirq context, etc.). Deadlines are instead
1283 * computed using hard walltime. This seems to be the more
1284 * natural solution, but the full ramifications of this
1285 * approach need further study.
1287 now = rq_clock_task(rq);
1288 delta_exec = now - curr->se.exec_start;
1289 if (unlikely((s64)delta_exec <= 0)) {
1290 if (unlikely(dl_se->dl_yielded))
1295 schedstat_set(curr->stats.exec_max,
1296 max(curr->stats.exec_max, delta_exec));
1298 trace_sched_stat_runtime(curr, delta_exec, 0);
1300 curr->se.sum_exec_runtime += delta_exec;
1301 account_group_exec_runtime(curr, delta_exec);
1303 curr->se.exec_start = now;
1304 cgroup_account_cputime(curr, delta_exec);
1306 if (dl_entity_is_special(dl_se))
1310 * For tasks that participate in GRUB, we implement GRUB-PA: the
1311 * spare reclaimed bandwidth is used to clock down frequency.
1313 * For the others, we still need to scale reservation parameters
1314 * according to current frequency and CPU maximum capacity.
1316 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1317 scaled_delta_exec = grub_reclaim(delta_exec,
1321 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1322 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1324 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1325 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1328 dl_se->runtime -= scaled_delta_exec;
1331 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1332 dl_se->dl_throttled = 1;
1334 /* If requested, inform the user about runtime overruns. */
1335 if (dl_runtime_exceeded(dl_se) &&
1336 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1337 dl_se->dl_overrun = 1;
1339 __dequeue_task_dl(rq, curr, 0);
1340 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(curr)))
1341 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1343 if (!is_leftmost(curr, &rq->dl))
1348 * Because -- for now -- we share the rt bandwidth, we need to
1349 * account our runtime there too, otherwise actual rt tasks
1350 * would be able to exceed the shared quota.
1352 * Account to the root rt group for now.
1354 * The solution we're working towards is having the RT groups scheduled
1355 * using deadline servers -- however there's a few nasties to figure
1356 * out before that can happen.
1358 if (rt_bandwidth_enabled()) {
1359 struct rt_rq *rt_rq = &rq->rt;
1361 raw_spin_lock(&rt_rq->rt_runtime_lock);
1363 * We'll let actual RT tasks worry about the overflow here, we
1364 * have our own CBS to keep us inline; only account when RT
1365 * bandwidth is relevant.
1367 if (sched_rt_bandwidth_account(rt_rq))
1368 rt_rq->rt_time += delta_exec;
1369 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1373 static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1375 struct sched_dl_entity *dl_se = container_of(timer,
1376 struct sched_dl_entity,
1378 struct task_struct *p = dl_task_of(dl_se);
1382 rq = task_rq_lock(p, &rf);
1385 update_rq_clock(rq);
1387 if (!dl_task(p) || READ_ONCE(p->__state) == TASK_DEAD) {
1388 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1390 if (READ_ONCE(p->__state) == TASK_DEAD && dl_se->dl_non_contending) {
1391 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1392 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1393 dl_se->dl_non_contending = 0;
1396 raw_spin_lock(&dl_b->lock);
1397 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1398 raw_spin_unlock(&dl_b->lock);
1399 __dl_clear_params(p);
1403 if (dl_se->dl_non_contending == 0)
1406 sub_running_bw(dl_se, &rq->dl);
1407 dl_se->dl_non_contending = 0;
1409 task_rq_unlock(rq, p, &rf);
1412 return HRTIMER_NORESTART;
1415 void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1417 struct hrtimer *timer = &dl_se->inactive_timer;
1419 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1420 timer->function = inactive_task_timer;
1423 #define __node_2_dle(node) \
1424 rb_entry((node), struct sched_dl_entity, rb_node)
1428 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1430 struct rq *rq = rq_of_dl_rq(dl_rq);
1432 if (dl_rq->earliest_dl.curr == 0 ||
1433 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1434 if (dl_rq->earliest_dl.curr == 0)
1435 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_HIGHER);
1436 dl_rq->earliest_dl.curr = deadline;
1437 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1441 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1443 struct rq *rq = rq_of_dl_rq(dl_rq);
1446 * Since we may have removed our earliest (and/or next earliest)
1447 * task we must recompute them.
1449 if (!dl_rq->dl_nr_running) {
1450 dl_rq->earliest_dl.curr = 0;
1451 dl_rq->earliest_dl.next = 0;
1452 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1453 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1455 struct rb_node *leftmost = rb_first_cached(&dl_rq->root);
1456 struct sched_dl_entity *entry = __node_2_dle(leftmost);
1458 dl_rq->earliest_dl.curr = entry->deadline;
1459 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1465 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1466 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1468 #endif /* CONFIG_SMP */
1471 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1473 int prio = dl_task_of(dl_se)->prio;
1474 u64 deadline = dl_se->deadline;
1476 WARN_ON(!dl_prio(prio));
1477 dl_rq->dl_nr_running++;
1478 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1480 inc_dl_deadline(dl_rq, deadline);
1481 inc_dl_migration(dl_se, dl_rq);
1485 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1487 int prio = dl_task_of(dl_se)->prio;
1489 WARN_ON(!dl_prio(prio));
1490 WARN_ON(!dl_rq->dl_nr_running);
1491 dl_rq->dl_nr_running--;
1492 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1494 dec_dl_deadline(dl_rq, dl_se->deadline);
1495 dec_dl_migration(dl_se, dl_rq);
1498 static inline bool __dl_less(struct rb_node *a, const struct rb_node *b)
1500 return dl_time_before(__node_2_dle(a)->deadline, __node_2_dle(b)->deadline);
1503 static inline struct sched_statistics *
1504 __schedstats_from_dl_se(struct sched_dl_entity *dl_se)
1506 return &dl_task_of(dl_se)->stats;
1510 update_stats_wait_start_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1512 struct sched_statistics *stats;
1514 if (!schedstat_enabled())
1517 stats = __schedstats_from_dl_se(dl_se);
1518 __update_stats_wait_start(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1522 update_stats_wait_end_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1524 struct sched_statistics *stats;
1526 if (!schedstat_enabled())
1529 stats = __schedstats_from_dl_se(dl_se);
1530 __update_stats_wait_end(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1534 update_stats_enqueue_sleeper_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se)
1536 struct sched_statistics *stats;
1538 if (!schedstat_enabled())
1541 stats = __schedstats_from_dl_se(dl_se);
1542 __update_stats_enqueue_sleeper(rq_of_dl_rq(dl_rq), dl_task_of(dl_se), stats);
1546 update_stats_enqueue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1549 if (!schedstat_enabled())
1552 if (flags & ENQUEUE_WAKEUP)
1553 update_stats_enqueue_sleeper_dl(dl_rq, dl_se);
1557 update_stats_dequeue_dl(struct dl_rq *dl_rq, struct sched_dl_entity *dl_se,
1560 struct task_struct *p = dl_task_of(dl_se);
1562 if (!schedstat_enabled())
1565 if ((flags & DEQUEUE_SLEEP)) {
1568 state = READ_ONCE(p->__state);
1569 if (state & TASK_INTERRUPTIBLE)
1570 __schedstat_set(p->stats.sleep_start,
1571 rq_clock(rq_of_dl_rq(dl_rq)));
1573 if (state & TASK_UNINTERRUPTIBLE)
1574 __schedstat_set(p->stats.block_start,
1575 rq_clock(rq_of_dl_rq(dl_rq)));
1579 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1581 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1583 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
1585 rb_add_cached(&dl_se->rb_node, &dl_rq->root, __dl_less);
1587 inc_dl_tasks(dl_se, dl_rq);
1590 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1592 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1594 if (RB_EMPTY_NODE(&dl_se->rb_node))
1597 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1599 RB_CLEAR_NODE(&dl_se->rb_node);
1601 dec_dl_tasks(dl_se, dl_rq);
1605 enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags)
1607 BUG_ON(on_dl_rq(dl_se));
1609 update_stats_enqueue_dl(dl_rq_of_se(dl_se), dl_se, flags);
1612 * If this is a wakeup or a new instance, the scheduling
1613 * parameters of the task might need updating. Otherwise,
1614 * we want a replenishment of its runtime.
1616 if (flags & ENQUEUE_WAKEUP) {
1617 task_contending(dl_se, flags);
1618 update_dl_entity(dl_se);
1619 } else if (flags & ENQUEUE_REPLENISH) {
1620 replenish_dl_entity(dl_se);
1621 } else if ((flags & ENQUEUE_RESTORE) &&
1622 dl_time_before(dl_se->deadline,
1623 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) {
1624 setup_new_dl_entity(dl_se);
1627 __enqueue_dl_entity(dl_se);
1630 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
1632 __dequeue_dl_entity(dl_se);
1635 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1637 if (is_dl_boosted(&p->dl)) {
1639 * Because of delays in the detection of the overrun of a
1640 * thread's runtime, it might be the case that a thread
1641 * goes to sleep in a rt mutex with negative runtime. As
1642 * a consequence, the thread will be throttled.
1644 * While waiting for the mutex, this thread can also be
1645 * boosted via PI, resulting in a thread that is throttled
1646 * and boosted at the same time.
1648 * In this case, the boost overrides the throttle.
1650 if (p->dl.dl_throttled) {
1652 * The replenish timer needs to be canceled. No
1653 * problem if it fires concurrently: boosted threads
1654 * are ignored in dl_task_timer().
1656 hrtimer_try_to_cancel(&p->dl.dl_timer);
1657 p->dl.dl_throttled = 0;
1659 } else if (!dl_prio(p->normal_prio)) {
1661 * Special case in which we have a !SCHED_DEADLINE task that is going
1662 * to be deboosted, but exceeds its runtime while doing so. No point in
1663 * replenishing it, as it's going to return back to its original
1664 * scheduling class after this. If it has been throttled, we need to
1665 * clear the flag, otherwise the task may wake up as throttled after
1666 * being boosted again with no means to replenish the runtime and clear
1669 p->dl.dl_throttled = 0;
1670 BUG_ON(!is_dl_boosted(&p->dl) || flags != ENQUEUE_REPLENISH);
1675 * Check if a constrained deadline task was activated
1676 * after the deadline but before the next period.
1677 * If that is the case, the task will be throttled and
1678 * the replenishment timer will be set to the next period.
1680 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
1681 dl_check_constrained_dl(&p->dl);
1683 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
1684 add_rq_bw(&p->dl, &rq->dl);
1685 add_running_bw(&p->dl, &rq->dl);
1689 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1690 * its budget it needs a replenishment and, since it now is on
1691 * its rq, the bandwidth timer callback (which clearly has not
1692 * run yet) will take care of this.
1693 * However, the active utilization does not depend on the fact
1694 * that the task is on the runqueue or not (but depends on the
1695 * task's state - in GRUB parlance, "inactive" vs "active contending").
1696 * In other words, even if a task is throttled its utilization must
1697 * be counted in the active utilization; hence, we need to call
1700 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1701 if (flags & ENQUEUE_WAKEUP)
1702 task_contending(&p->dl, flags);
1707 check_schedstat_required();
1708 update_stats_wait_start_dl(dl_rq_of_se(&p->dl), &p->dl);
1710 enqueue_dl_entity(&p->dl, flags);
1712 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1713 enqueue_pushable_dl_task(rq, p);
1716 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1718 update_stats_dequeue_dl(&rq->dl, &p->dl, flags);
1719 dequeue_dl_entity(&p->dl);
1720 dequeue_pushable_dl_task(rq, p);
1723 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1726 __dequeue_task_dl(rq, p, flags);
1728 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
1729 sub_running_bw(&p->dl, &rq->dl);
1730 sub_rq_bw(&p->dl, &rq->dl);
1734 * This check allows to start the inactive timer (or to immediately
1735 * decrease the active utilization, if needed) in two cases:
1736 * when the task blocks and when it is terminating
1737 * (p->state == TASK_DEAD). We can handle the two cases in the same
1738 * way, because from GRUB's point of view the same thing is happening
1739 * (the task moves from "active contending" to "active non contending"
1742 if (flags & DEQUEUE_SLEEP)
1743 task_non_contending(p);
1747 * Yield task semantic for -deadline tasks is:
1749 * get off from the CPU until our next instance, with
1750 * a new runtime. This is of little use now, since we
1751 * don't have a bandwidth reclaiming mechanism. Anyway,
1752 * bandwidth reclaiming is planned for the future, and
1753 * yield_task_dl will indicate that some spare budget
1754 * is available for other task instances to use it.
1756 static void yield_task_dl(struct rq *rq)
1759 * We make the task go to sleep until its current deadline by
1760 * forcing its runtime to zero. This way, update_curr_dl() stops
1761 * it and the bandwidth timer will wake it up and will give it
1762 * new scheduling parameters (thanks to dl_yielded=1).
1764 rq->curr->dl.dl_yielded = 1;
1766 update_rq_clock(rq);
1769 * Tell update_rq_clock() that we've just updated,
1770 * so we don't do microscopic update in schedule()
1771 * and double the fastpath cost.
1773 rq_clock_skip_update(rq);
1778 static int find_later_rq(struct task_struct *task);
1781 select_task_rq_dl(struct task_struct *p, int cpu, int flags)
1783 struct task_struct *curr;
1787 if (!(flags & WF_TTWU))
1793 curr = READ_ONCE(rq->curr); /* unlocked access */
1796 * If we are dealing with a -deadline task, we must
1797 * decide where to wake it up.
1798 * If it has a later deadline and the current task
1799 * on this rq can't move (provided the waking task
1800 * can!) we prefer to send it somewhere else. On the
1801 * other hand, if it has a shorter deadline, we
1802 * try to make it stay here, it might be important.
1804 select_rq = unlikely(dl_task(curr)) &&
1805 (curr->nr_cpus_allowed < 2 ||
1806 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1807 p->nr_cpus_allowed > 1;
1810 * Take the capacity of the CPU into account to
1811 * ensure it fits the requirement of the task.
1813 if (static_branch_unlikely(&sched_asym_cpucapacity))
1814 select_rq |= !dl_task_fits_capacity(p, cpu);
1817 int target = find_later_rq(p);
1820 (dl_time_before(p->dl.deadline,
1821 cpu_rq(target)->dl.earliest_dl.curr) ||
1822 (cpu_rq(target)->dl.dl_nr_running == 0)))
1831 static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
1836 if (READ_ONCE(p->__state) != TASK_WAKING)
1841 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1842 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1843 * rq->lock is not... So, lock it
1846 if (p->dl.dl_non_contending) {
1847 update_rq_clock(rq);
1848 sub_running_bw(&p->dl, &rq->dl);
1849 p->dl.dl_non_contending = 0;
1851 * If the timer handler is currently running and the
1852 * timer cannot be canceled, inactive_task_timer()
1853 * will see that dl_not_contending is not set, and
1854 * will not touch the rq's active utilization,
1855 * so we are still safe.
1857 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
1860 sub_rq_bw(&p->dl, &rq->dl);
1864 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1867 * Current can't be migrated, useless to reschedule,
1868 * let's hope p can move out.
1870 if (rq->curr->nr_cpus_allowed == 1 ||
1871 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
1875 * p is migratable, so let's not schedule it and
1876 * see if it is pushed or pulled somewhere else.
1878 if (p->nr_cpus_allowed != 1 &&
1879 cpudl_find(&rq->rd->cpudl, p, NULL))
1885 static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1887 if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) {
1889 * This is OK, because current is on_cpu, which avoids it being
1890 * picked for load-balance and preemption/IRQs are still
1891 * disabled avoiding further scheduler activity on it and we've
1892 * not yet started the picking loop.
1894 rq_unpin_lock(rq, rf);
1896 rq_repin_lock(rq, rf);
1899 return sched_stop_runnable(rq) || sched_dl_runnable(rq);
1901 #endif /* CONFIG_SMP */
1904 * Only called when both the current and waking task are -deadline
1907 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1910 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
1917 * In the unlikely case current and p have the same deadline
1918 * let us try to decide what's the best thing to do...
1920 if ((p->dl.deadline == rq->curr->dl.deadline) &&
1921 !test_tsk_need_resched(rq->curr))
1922 check_preempt_equal_dl(rq, p);
1923 #endif /* CONFIG_SMP */
1926 #ifdef CONFIG_SCHED_HRTICK
1927 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1929 hrtick_start(rq, p->dl.runtime);
1931 #else /* !CONFIG_SCHED_HRTICK */
1932 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1937 static void set_next_task_dl(struct rq *rq, struct task_struct *p, bool first)
1939 struct sched_dl_entity *dl_se = &p->dl;
1940 struct dl_rq *dl_rq = &rq->dl;
1942 p->se.exec_start = rq_clock_task(rq);
1943 if (on_dl_rq(&p->dl))
1944 update_stats_wait_end_dl(dl_rq, dl_se);
1946 /* You can't push away the running task */
1947 dequeue_pushable_dl_task(rq, p);
1952 if (hrtick_enabled_dl(rq))
1953 start_hrtick_dl(rq, p);
1955 if (rq->curr->sched_class != &dl_sched_class)
1956 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
1958 deadline_queue_push_tasks(rq);
1961 static struct sched_dl_entity *pick_next_dl_entity(struct dl_rq *dl_rq)
1963 struct rb_node *left = rb_first_cached(&dl_rq->root);
1968 return __node_2_dle(left);
1971 static struct task_struct *pick_task_dl(struct rq *rq)
1973 struct sched_dl_entity *dl_se;
1974 struct dl_rq *dl_rq = &rq->dl;
1975 struct task_struct *p;
1977 if (!sched_dl_runnable(rq))
1980 dl_se = pick_next_dl_entity(dl_rq);
1982 p = dl_task_of(dl_se);
1987 static struct task_struct *pick_next_task_dl(struct rq *rq)
1989 struct task_struct *p;
1991 p = pick_task_dl(rq);
1993 set_next_task_dl(rq, p, true);
1998 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
2000 struct sched_dl_entity *dl_se = &p->dl;
2001 struct dl_rq *dl_rq = &rq->dl;
2003 if (on_dl_rq(&p->dl))
2004 update_stats_wait_start_dl(dl_rq, dl_se);
2008 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2009 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
2010 enqueue_pushable_dl_task(rq, p);
2014 * scheduler tick hitting a task of our scheduling class.
2016 * NOTE: This function can be called remotely by the tick offload that
2017 * goes along full dynticks. Therefore no local assumption can be made
2018 * and everything must be accessed through the @rq and @curr passed in
2021 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
2025 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2027 * Even when we have runtime, update_curr_dl() might have resulted in us
2028 * not being the leftmost task anymore. In that case NEED_RESCHED will
2029 * be set and schedule() will start a new hrtick for the next task.
2031 if (hrtick_enabled_dl(rq) && queued && p->dl.runtime > 0 &&
2032 is_leftmost(p, &rq->dl))
2033 start_hrtick_dl(rq, p);
2036 static void task_fork_dl(struct task_struct *p)
2039 * SCHED_DEADLINE tasks cannot fork and this is achieved through
2046 /* Only try algorithms three times */
2047 #define DL_MAX_TRIES 3
2049 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
2051 if (!task_running(rq, p) &&
2052 cpumask_test_cpu(cpu, &p->cpus_mask))
2058 * Return the earliest pushable rq's task, which is suitable to be executed
2059 * on the CPU, NULL otherwise:
2061 static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
2063 struct task_struct *p = NULL;
2064 struct rb_node *next_node;
2066 if (!has_pushable_dl_tasks(rq))
2069 next_node = rb_first_cached(&rq->dl.pushable_dl_tasks_root);
2073 p = __node_2_pdl(next_node);
2075 if (pick_dl_task(rq, p, cpu))
2078 next_node = rb_next(next_node);
2085 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
2087 static int find_later_rq(struct task_struct *task)
2089 struct sched_domain *sd;
2090 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
2091 int this_cpu = smp_processor_id();
2092 int cpu = task_cpu(task);
2094 /* Make sure the mask is initialized first */
2095 if (unlikely(!later_mask))
2098 if (task->nr_cpus_allowed == 1)
2102 * We have to consider system topology and task affinity
2103 * first, then we can look for a suitable CPU.
2105 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
2109 * If we are here, some targets have been found, including
2110 * the most suitable which is, among the runqueues where the
2111 * current tasks have later deadlines than the task's one, the
2112 * rq with the latest possible one.
2114 * Now we check how well this matches with task's
2115 * affinity and system topology.
2117 * The last CPU where the task run is our first
2118 * guess, since it is most likely cache-hot there.
2120 if (cpumask_test_cpu(cpu, later_mask))
2123 * Check if this_cpu is to be skipped (i.e., it is
2124 * not in the mask) or not.
2126 if (!cpumask_test_cpu(this_cpu, later_mask))
2130 for_each_domain(cpu, sd) {
2131 if (sd->flags & SD_WAKE_AFFINE) {
2135 * If possible, preempting this_cpu is
2136 * cheaper than migrating.
2138 if (this_cpu != -1 &&
2139 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
2144 best_cpu = cpumask_any_and_distribute(later_mask,
2145 sched_domain_span(sd));
2147 * Last chance: if a CPU being in both later_mask
2148 * and current sd span is valid, that becomes our
2149 * choice. Of course, the latest possible CPU is
2150 * already under consideration through later_mask.
2152 if (best_cpu < nr_cpu_ids) {
2161 * At this point, all our guesses failed, we just return
2162 * 'something', and let the caller sort the things out.
2167 cpu = cpumask_any_distribute(later_mask);
2168 if (cpu < nr_cpu_ids)
2174 /* Locks the rq it finds */
2175 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
2177 struct rq *later_rq = NULL;
2181 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
2182 cpu = find_later_rq(task);
2184 if ((cpu == -1) || (cpu == rq->cpu))
2187 later_rq = cpu_rq(cpu);
2189 if (later_rq->dl.dl_nr_running &&
2190 !dl_time_before(task->dl.deadline,
2191 later_rq->dl.earliest_dl.curr)) {
2193 * Target rq has tasks of equal or earlier deadline,
2194 * retrying does not release any lock and is unlikely
2195 * to yield a different result.
2201 /* Retry if something changed. */
2202 if (double_lock_balance(rq, later_rq)) {
2203 if (unlikely(task_rq(task) != rq ||
2204 !cpumask_test_cpu(later_rq->cpu, &task->cpus_mask) ||
2205 task_running(rq, task) ||
2207 !task_on_rq_queued(task))) {
2208 double_unlock_balance(rq, later_rq);
2215 * If the rq we found has no -deadline task, or
2216 * its earliest one has a later deadline than our
2217 * task, the rq is a good one.
2219 if (!later_rq->dl.dl_nr_running ||
2220 dl_time_before(task->dl.deadline,
2221 later_rq->dl.earliest_dl.curr))
2224 /* Otherwise we try again. */
2225 double_unlock_balance(rq, later_rq);
2232 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2234 struct task_struct *p;
2236 if (!has_pushable_dl_tasks(rq))
2239 p = __node_2_pdl(rb_first_cached(&rq->dl.pushable_dl_tasks_root));
2241 BUG_ON(rq->cpu != task_cpu(p));
2242 BUG_ON(task_current(rq, p));
2243 BUG_ON(p->nr_cpus_allowed <= 1);
2245 BUG_ON(!task_on_rq_queued(p));
2246 BUG_ON(!dl_task(p));
2252 * See if the non running -deadline tasks on this rq
2253 * can be sent to some other CPU where they can preempt
2254 * and start executing.
2256 static int push_dl_task(struct rq *rq)
2258 struct task_struct *next_task;
2259 struct rq *later_rq;
2262 if (!rq->dl.overloaded)
2265 next_task = pick_next_pushable_dl_task(rq);
2271 * If next_task preempts rq->curr, and rq->curr
2272 * can move away, it makes sense to just reschedule
2273 * without going further in pushing next_task.
2275 if (dl_task(rq->curr) &&
2276 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2277 rq->curr->nr_cpus_allowed > 1) {
2282 if (is_migration_disabled(next_task))
2285 if (WARN_ON(next_task == rq->curr))
2288 /* We might release rq lock */
2289 get_task_struct(next_task);
2291 /* Will lock the rq it'll find */
2292 later_rq = find_lock_later_rq(next_task, rq);
2294 struct task_struct *task;
2297 * We must check all this again, since
2298 * find_lock_later_rq releases rq->lock and it is
2299 * then possible that next_task has migrated.
2301 task = pick_next_pushable_dl_task(rq);
2302 if (task == next_task) {
2304 * The task is still there. We don't try
2305 * again, some other CPU will pull it when ready.
2314 put_task_struct(next_task);
2319 deactivate_task(rq, next_task, 0);
2320 set_task_cpu(next_task, later_rq->cpu);
2321 activate_task(later_rq, next_task, 0);
2324 resched_curr(later_rq);
2326 double_unlock_balance(rq, later_rq);
2329 put_task_struct(next_task);
2334 static void push_dl_tasks(struct rq *rq)
2336 /* push_dl_task() will return true if it moved a -deadline task */
2337 while (push_dl_task(rq))
2341 static void pull_dl_task(struct rq *this_rq)
2343 int this_cpu = this_rq->cpu, cpu;
2344 struct task_struct *p, *push_task;
2345 bool resched = false;
2347 u64 dmin = LONG_MAX;
2349 if (likely(!dl_overloaded(this_rq)))
2353 * Match the barrier from dl_set_overloaded; this guarantees that if we
2354 * see overloaded we must also see the dlo_mask bit.
2358 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2359 if (this_cpu == cpu)
2362 src_rq = cpu_rq(cpu);
2365 * It looks racy, abd it is! However, as in sched_rt.c,
2366 * we are fine with this.
2368 if (this_rq->dl.dl_nr_running &&
2369 dl_time_before(this_rq->dl.earliest_dl.curr,
2370 src_rq->dl.earliest_dl.next))
2373 /* Might drop this_rq->lock */
2375 double_lock_balance(this_rq, src_rq);
2378 * If there are no more pullable tasks on the
2379 * rq, we're done with it.
2381 if (src_rq->dl.dl_nr_running <= 1)
2384 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2387 * We found a task to be pulled if:
2388 * - it preempts our current (if there's one),
2389 * - it will preempt the last one we pulled (if any).
2391 if (p && dl_time_before(p->dl.deadline, dmin) &&
2392 (!this_rq->dl.dl_nr_running ||
2393 dl_time_before(p->dl.deadline,
2394 this_rq->dl.earliest_dl.curr))) {
2395 WARN_ON(p == src_rq->curr);
2396 WARN_ON(!task_on_rq_queued(p));
2399 * Then we pull iff p has actually an earlier
2400 * deadline than the current task of its runqueue.
2402 if (dl_time_before(p->dl.deadline,
2403 src_rq->curr->dl.deadline))
2406 if (is_migration_disabled(p)) {
2407 push_task = get_push_task(src_rq);
2409 deactivate_task(src_rq, p, 0);
2410 set_task_cpu(p, this_cpu);
2411 activate_task(this_rq, p, 0);
2412 dmin = p->dl.deadline;
2416 /* Is there any other task even earlier? */
2419 double_unlock_balance(this_rq, src_rq);
2422 raw_spin_rq_unlock(this_rq);
2423 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2424 push_task, &src_rq->push_work);
2425 raw_spin_rq_lock(this_rq);
2430 resched_curr(this_rq);
2434 * Since the task is not running and a reschedule is not going to happen
2435 * anytime soon on its runqueue, we try pushing it away now.
2437 static void task_woken_dl(struct rq *rq, struct task_struct *p)
2439 if (!task_running(rq, p) &&
2440 !test_tsk_need_resched(rq->curr) &&
2441 p->nr_cpus_allowed > 1 &&
2442 dl_task(rq->curr) &&
2443 (rq->curr->nr_cpus_allowed < 2 ||
2444 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2449 static void set_cpus_allowed_dl(struct task_struct *p,
2450 const struct cpumask *new_mask,
2453 struct root_domain *src_rd;
2456 BUG_ON(!dl_task(p));
2461 * Migrating a SCHED_DEADLINE task between exclusive
2462 * cpusets (different root_domains) entails a bandwidth
2463 * update. We already made space for us in the destination
2464 * domain (see cpuset_can_attach()).
2466 if (!cpumask_intersects(src_rd->span, new_mask)) {
2467 struct dl_bw *src_dl_b;
2469 src_dl_b = dl_bw_of(cpu_of(rq));
2471 * We now free resources of the root_domain we are migrating
2472 * off. In the worst case, sched_setattr() may temporary fail
2473 * until we complete the update.
2475 raw_spin_lock(&src_dl_b->lock);
2476 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2477 raw_spin_unlock(&src_dl_b->lock);
2480 set_cpus_allowed_common(p, new_mask, flags);
2483 /* Assumes rq->lock is held */
2484 static void rq_online_dl(struct rq *rq)
2486 if (rq->dl.overloaded)
2487 dl_set_overload(rq);
2489 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2490 if (rq->dl.dl_nr_running > 0)
2491 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2494 /* Assumes rq->lock is held */
2495 static void rq_offline_dl(struct rq *rq)
2497 if (rq->dl.overloaded)
2498 dl_clear_overload(rq);
2500 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2501 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2504 void __init init_sched_dl_class(void)
2508 for_each_possible_cpu(i)
2509 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2510 GFP_KERNEL, cpu_to_node(i));
2513 void dl_add_task_root_domain(struct task_struct *p)
2519 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2521 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
2525 rq = __task_rq_lock(p, &rf);
2527 dl_b = &rq->rd->dl_bw;
2528 raw_spin_lock(&dl_b->lock);
2530 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2532 raw_spin_unlock(&dl_b->lock);
2534 task_rq_unlock(rq, p, &rf);
2537 void dl_clear_root_domain(struct root_domain *rd)
2539 unsigned long flags;
2541 raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
2542 rd->dl_bw.total_bw = 0;
2543 raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
2546 #endif /* CONFIG_SMP */
2548 static void switched_from_dl(struct rq *rq, struct task_struct *p)
2551 * task_non_contending() can start the "inactive timer" (if the 0-lag
2552 * time is in the future). If the task switches back to dl before
2553 * the "inactive timer" fires, it can continue to consume its current
2554 * runtime using its current deadline. If it stays outside of
2555 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2556 * will reset the task parameters.
2558 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2559 task_non_contending(p);
2561 if (!task_on_rq_queued(p)) {
2563 * Inactive timer is armed. However, p is leaving DEADLINE and
2564 * might migrate away from this rq while continuing to run on
2565 * some other class. We need to remove its contribution from
2566 * this rq running_bw now, or sub_rq_bw (below) will complain.
2568 if (p->dl.dl_non_contending)
2569 sub_running_bw(&p->dl, &rq->dl);
2570 sub_rq_bw(&p->dl, &rq->dl);
2574 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2575 * at the 0-lag time, because the task could have been migrated
2576 * while SCHED_OTHER in the meanwhile.
2578 if (p->dl.dl_non_contending)
2579 p->dl.dl_non_contending = 0;
2582 * Since this might be the only -deadline task on the rq,
2583 * this is the right place to try to pull some other one
2584 * from an overloaded CPU, if any.
2586 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
2589 deadline_queue_pull_task(rq);
2593 * When switching to -deadline, we may overload the rq, then
2594 * we try to push someone off, if possible.
2596 static void switched_to_dl(struct rq *rq, struct task_struct *p)
2598 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2601 /* If p is not queued we will update its parameters at next wakeup. */
2602 if (!task_on_rq_queued(p)) {
2603 add_rq_bw(&p->dl, &rq->dl);
2608 if (rq->curr != p) {
2610 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
2611 deadline_queue_push_tasks(rq);
2613 if (dl_task(rq->curr))
2614 check_preempt_curr_dl(rq, p, 0);
2618 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2623 * If the scheduling parameters of a -deadline task changed,
2624 * a push or pull operation might be needed.
2626 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2629 if (task_on_rq_queued(p) || task_current(rq, p)) {
2632 * This might be too much, but unfortunately
2633 * we don't have the old deadline value, and
2634 * we can't argue if the task is increasing
2635 * or lowering its prio, so...
2637 if (!rq->dl.overloaded)
2638 deadline_queue_pull_task(rq);
2641 * If we now have a earlier deadline task than p,
2642 * then reschedule, provided p is still on this
2645 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
2649 * Again, we don't know if p has a earlier
2650 * or later deadline, so let's blindly set a
2651 * (maybe not needed) rescheduling point.
2654 #endif /* CONFIG_SMP */
2658 DEFINE_SCHED_CLASS(dl) = {
2660 .enqueue_task = enqueue_task_dl,
2661 .dequeue_task = dequeue_task_dl,
2662 .yield_task = yield_task_dl,
2664 .check_preempt_curr = check_preempt_curr_dl,
2666 .pick_next_task = pick_next_task_dl,
2667 .put_prev_task = put_prev_task_dl,
2668 .set_next_task = set_next_task_dl,
2671 .balance = balance_dl,
2672 .pick_task = pick_task_dl,
2673 .select_task_rq = select_task_rq_dl,
2674 .migrate_task_rq = migrate_task_rq_dl,
2675 .set_cpus_allowed = set_cpus_allowed_dl,
2676 .rq_online = rq_online_dl,
2677 .rq_offline = rq_offline_dl,
2678 .task_woken = task_woken_dl,
2679 .find_lock_rq = find_lock_later_rq,
2682 .task_tick = task_tick_dl,
2683 .task_fork = task_fork_dl,
2685 .prio_changed = prio_changed_dl,
2686 .switched_from = switched_from_dl,
2687 .switched_to = switched_to_dl,
2689 .update_curr = update_curr_dl,
2692 /* Used for dl_bw check and update, used under sched_rt_handler()::mutex */
2693 static u64 dl_generation;
2695 int sched_dl_global_validate(void)
2697 u64 runtime = global_rt_runtime();
2698 u64 period = global_rt_period();
2699 u64 new_bw = to_ratio(period, runtime);
2700 u64 gen = ++dl_generation;
2702 int cpu, cpus, ret = 0;
2703 unsigned long flags;
2706 * Here we want to check the bandwidth not being set to some
2707 * value smaller than the currently allocated bandwidth in
2708 * any of the root_domains.
2710 for_each_possible_cpu(cpu) {
2711 rcu_read_lock_sched();
2713 if (dl_bw_visited(cpu, gen))
2716 dl_b = dl_bw_of(cpu);
2717 cpus = dl_bw_cpus(cpu);
2719 raw_spin_lock_irqsave(&dl_b->lock, flags);
2720 if (new_bw * cpus < dl_b->total_bw)
2722 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2725 rcu_read_unlock_sched();
2734 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
2736 if (global_rt_runtime() == RUNTIME_INF) {
2737 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2738 dl_rq->extra_bw = 1 << BW_SHIFT;
2740 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2741 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2742 dl_rq->extra_bw = to_ratio(global_rt_period(),
2743 global_rt_runtime());
2747 void sched_dl_do_global(void)
2750 u64 gen = ++dl_generation;
2753 unsigned long flags;
2755 if (global_rt_runtime() != RUNTIME_INF)
2756 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2758 for_each_possible_cpu(cpu) {
2759 rcu_read_lock_sched();
2761 if (dl_bw_visited(cpu, gen)) {
2762 rcu_read_unlock_sched();
2766 dl_b = dl_bw_of(cpu);
2768 raw_spin_lock_irqsave(&dl_b->lock, flags);
2770 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2772 rcu_read_unlock_sched();
2773 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2778 * We must be sure that accepting a new task (or allowing changing the
2779 * parameters of an existing one) is consistent with the bandwidth
2780 * constraints. If yes, this function also accordingly updates the currently
2781 * allocated bandwidth to reflect the new situation.
2783 * This function is called while holding p's rq->lock.
2785 int sched_dl_overflow(struct task_struct *p, int policy,
2786 const struct sched_attr *attr)
2788 u64 period = attr->sched_period ?: attr->sched_deadline;
2789 u64 runtime = attr->sched_runtime;
2790 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2791 int cpus, err = -1, cpu = task_cpu(p);
2792 struct dl_bw *dl_b = dl_bw_of(cpu);
2795 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2798 /* !deadline task may carry old deadline bandwidth */
2799 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2803 * Either if a task, enters, leave, or stays -deadline but changes
2804 * its parameters, we may need to update accordingly the total
2805 * allocated bandwidth of the container.
2807 raw_spin_lock(&dl_b->lock);
2808 cpus = dl_bw_cpus(cpu);
2809 cap = dl_bw_capacity(cpu);
2811 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2812 !__dl_overflow(dl_b, cap, 0, new_bw)) {
2813 if (hrtimer_active(&p->dl.inactive_timer))
2814 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2815 __dl_add(dl_b, new_bw, cpus);
2817 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2818 !__dl_overflow(dl_b, cap, p->dl.dl_bw, new_bw)) {
2820 * XXX this is slightly incorrect: when the task
2821 * utilization decreases, we should delay the total
2822 * utilization change until the task's 0-lag point.
2823 * But this would require to set the task's "inactive
2824 * timer" when the task is not inactive.
2826 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2827 __dl_add(dl_b, new_bw, cpus);
2828 dl_change_utilization(p, new_bw);
2830 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2832 * Do not decrease the total deadline utilization here,
2833 * switched_from_dl() will take care to do it at the correct
2838 raw_spin_unlock(&dl_b->lock);
2844 * This function initializes the sched_dl_entity of a newly becoming
2845 * SCHED_DEADLINE task.
2847 * Only the static values are considered here, the actual runtime and the
2848 * absolute deadline will be properly calculated when the task is enqueued
2849 * for the first time with its new policy.
2851 void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
2853 struct sched_dl_entity *dl_se = &p->dl;
2855 dl_se->dl_runtime = attr->sched_runtime;
2856 dl_se->dl_deadline = attr->sched_deadline;
2857 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
2858 dl_se->flags = attr->sched_flags & SCHED_DL_FLAGS;
2859 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
2860 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
2863 void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
2865 struct sched_dl_entity *dl_se = &p->dl;
2867 attr->sched_priority = p->rt_priority;
2868 attr->sched_runtime = dl_se->dl_runtime;
2869 attr->sched_deadline = dl_se->dl_deadline;
2870 attr->sched_period = dl_se->dl_period;
2871 attr->sched_flags &= ~SCHED_DL_FLAGS;
2872 attr->sched_flags |= dl_se->flags;
2876 * Default limits for DL period; on the top end we guard against small util
2877 * tasks still getting ridiculously long effective runtimes, on the bottom end we
2878 * guard against timer DoS.
2880 unsigned int sysctl_sched_dl_period_max = 1 << 22; /* ~4 seconds */
2881 unsigned int sysctl_sched_dl_period_min = 100; /* 100 us */
2884 * This function validates the new parameters of a -deadline task.
2885 * We ask for the deadline not being zero, and greater or equal
2886 * than the runtime, as well as the period of being zero or
2887 * greater than deadline. Furthermore, we have to be sure that
2888 * user parameters are above the internal resolution of 1us (we
2889 * check sched_runtime only since it is always the smaller one) and
2890 * below 2^63 ns (we have to check both sched_deadline and
2891 * sched_period, as the latter can be zero).
2893 bool __checkparam_dl(const struct sched_attr *attr)
2895 u64 period, max, min;
2897 /* special dl tasks don't actually use any parameter */
2898 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2902 if (attr->sched_deadline == 0)
2906 * Since we truncate DL_SCALE bits, make sure we're at least
2909 if (attr->sched_runtime < (1ULL << DL_SCALE))
2913 * Since we use the MSB for wrap-around and sign issues, make
2914 * sure it's not set (mind that period can be equal to zero).
2916 if (attr->sched_deadline & (1ULL << 63) ||
2917 attr->sched_period & (1ULL << 63))
2920 period = attr->sched_period;
2922 period = attr->sched_deadline;
2924 /* runtime <= deadline <= period (if period != 0) */
2925 if (period < attr->sched_deadline ||
2926 attr->sched_deadline < attr->sched_runtime)
2929 max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
2930 min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
2932 if (period < min || period > max)
2939 * This function clears the sched_dl_entity static params.
2941 void __dl_clear_params(struct task_struct *p)
2943 struct sched_dl_entity *dl_se = &p->dl;
2945 dl_se->dl_runtime = 0;
2946 dl_se->dl_deadline = 0;
2947 dl_se->dl_period = 0;
2950 dl_se->dl_density = 0;
2952 dl_se->dl_throttled = 0;
2953 dl_se->dl_yielded = 0;
2954 dl_se->dl_non_contending = 0;
2955 dl_se->dl_overrun = 0;
2957 #ifdef CONFIG_RT_MUTEXES
2958 dl_se->pi_se = dl_se;
2962 bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
2964 struct sched_dl_entity *dl_se = &p->dl;
2966 if (dl_se->dl_runtime != attr->sched_runtime ||
2967 dl_se->dl_deadline != attr->sched_deadline ||
2968 dl_se->dl_period != attr->sched_period ||
2969 dl_se->flags != (attr->sched_flags & SCHED_DL_FLAGS))
2976 int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
2977 const struct cpumask *trial)
2979 int ret = 1, trial_cpus;
2980 struct dl_bw *cur_dl_b;
2981 unsigned long flags;
2983 rcu_read_lock_sched();
2984 cur_dl_b = dl_bw_of(cpumask_any(cur));
2985 trial_cpus = cpumask_weight(trial);
2987 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
2988 if (cur_dl_b->bw != -1 &&
2989 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
2991 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
2992 rcu_read_unlock_sched();
2997 int dl_cpu_busy(int cpu, struct task_struct *p)
2999 unsigned long flags, cap;
3003 rcu_read_lock_sched();
3004 dl_b = dl_bw_of(cpu);
3005 raw_spin_lock_irqsave(&dl_b->lock, flags);
3006 cap = dl_bw_capacity(cpu);
3007 overflow = __dl_overflow(dl_b, cap, 0, p ? p->dl.dl_bw : 0);
3009 if (!overflow && p) {
3011 * We reserve space for this task in the destination
3012 * root_domain, as we can't fail after this point.
3013 * We will free resources in the source root_domain
3014 * later on (see set_cpus_allowed_dl()).
3016 __dl_add(dl_b, p->dl.dl_bw, dl_bw_cpus(cpu));
3019 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3020 rcu_read_unlock_sched();
3022 return overflow ? -EBUSY : 0;
3026 #ifdef CONFIG_SCHED_DEBUG
3027 void print_dl_stats(struct seq_file *m, int cpu)
3029 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
3031 #endif /* CONFIG_SCHED_DEBUG */