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>
21 struct dl_bandwidth def_dl_bandwidth;
23 static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
25 return container_of(dl_se, struct task_struct, dl);
28 static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
30 return container_of(dl_rq, struct rq, dl);
33 static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
35 struct task_struct *p = dl_task_of(dl_se);
36 struct rq *rq = task_rq(p);
41 static inline int on_dl_rq(struct sched_dl_entity *dl_se)
43 return !RB_EMPTY_NODE(&dl_se->rb_node);
46 #ifdef CONFIG_RT_MUTEXES
47 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
52 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
54 return pi_of(dl_se) != dl_se;
57 static inline struct sched_dl_entity *pi_of(struct sched_dl_entity *dl_se)
62 static inline bool is_dl_boosted(struct sched_dl_entity *dl_se)
69 static inline struct dl_bw *dl_bw_of(int i)
71 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
72 "sched RCU must be held");
73 return &cpu_rq(i)->rd->dl_bw;
76 static inline int dl_bw_cpus(int i)
78 struct root_domain *rd = cpu_rq(i)->rd;
81 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
82 "sched RCU must be held");
84 if (cpumask_subset(rd->span, cpu_active_mask))
85 return cpumask_weight(rd->span);
89 for_each_cpu_and(i, rd->span, cpu_active_mask)
95 static inline unsigned long __dl_bw_capacity(int i)
97 struct root_domain *rd = cpu_rq(i)->rd;
98 unsigned long cap = 0;
100 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
101 "sched RCU must be held");
103 for_each_cpu_and(i, rd->span, cpu_active_mask)
104 cap += capacity_orig_of(i);
110 * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
111 * of the CPU the task is running on rather rd's \Sum CPU capacity.
113 static inline unsigned long dl_bw_capacity(int i)
115 if (!static_branch_unlikely(&sched_asym_cpucapacity) &&
116 capacity_orig_of(i) == SCHED_CAPACITY_SCALE) {
117 return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT;
119 return __dl_bw_capacity(i);
123 static inline bool dl_bw_visited(int cpu, u64 gen)
125 struct root_domain *rd = cpu_rq(cpu)->rd;
127 if (rd->visit_gen == gen)
134 static inline struct dl_bw *dl_bw_of(int i)
136 return &cpu_rq(i)->dl.dl_bw;
139 static inline int dl_bw_cpus(int i)
144 static inline unsigned long dl_bw_capacity(int i)
146 return SCHED_CAPACITY_SCALE;
149 static inline bool dl_bw_visited(int cpu, u64 gen)
156 void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
158 u64 old = dl_rq->running_bw;
160 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
161 dl_rq->running_bw += dl_bw;
162 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
163 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
164 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
165 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
169 void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
171 u64 old = dl_rq->running_bw;
173 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
174 dl_rq->running_bw -= dl_bw;
175 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
176 if (dl_rq->running_bw > old)
177 dl_rq->running_bw = 0;
178 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
179 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
183 void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
185 u64 old = dl_rq->this_bw;
187 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
188 dl_rq->this_bw += dl_bw;
189 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
193 void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
195 u64 old = dl_rq->this_bw;
197 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
198 dl_rq->this_bw -= dl_bw;
199 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
200 if (dl_rq->this_bw > old)
202 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
206 void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
208 if (!dl_entity_is_special(dl_se))
209 __add_rq_bw(dl_se->dl_bw, dl_rq);
213 void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
215 if (!dl_entity_is_special(dl_se))
216 __sub_rq_bw(dl_se->dl_bw, dl_rq);
220 void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
222 if (!dl_entity_is_special(dl_se))
223 __add_running_bw(dl_se->dl_bw, dl_rq);
227 void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
229 if (!dl_entity_is_special(dl_se))
230 __sub_running_bw(dl_se->dl_bw, dl_rq);
233 static void dl_change_utilization(struct task_struct *p, u64 new_bw)
237 BUG_ON(p->dl.flags & SCHED_FLAG_SUGOV);
239 if (task_on_rq_queued(p))
243 if (p->dl.dl_non_contending) {
244 sub_running_bw(&p->dl, &rq->dl);
245 p->dl.dl_non_contending = 0;
247 * If the timer handler is currently running and the
248 * timer cannot be cancelled, inactive_task_timer()
249 * will see that dl_not_contending is not set, and
250 * will not touch the rq's active utilization,
251 * so we are still safe.
253 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
256 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
257 __add_rq_bw(new_bw, &rq->dl);
261 * The utilization of a task cannot be immediately removed from
262 * the rq active utilization (running_bw) when the task blocks.
263 * Instead, we have to wait for the so called "0-lag time".
265 * If a task blocks before the "0-lag time", a timer (the inactive
266 * timer) is armed, and running_bw is decreased when the timer
269 * If the task wakes up again before the inactive timer fires,
270 * the timer is cancelled, whereas if the task wakes up after the
271 * inactive timer fired (and running_bw has been decreased) the
272 * task's utilization has to be added to running_bw again.
273 * A flag in the deadline scheduling entity (dl_non_contending)
274 * is used to avoid race conditions between the inactive timer handler
277 * The following diagram shows how running_bw is updated. A task is
278 * "ACTIVE" when its utilization contributes to running_bw; an
279 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
280 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
281 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
282 * time already passed, which does not contribute to running_bw anymore.
283 * +------------------+
285 * +------------------>+ contending |
286 * | add_running_bw | |
287 * | +----+------+------+
290 * +--------+-------+ | |
291 * | | t >= 0-lag | | wakeup
292 * | INACTIVE |<---------------+ |
293 * | | sub_running_bw | |
294 * +--------+-------+ | |
299 * | +----+------+------+
300 * | sub_running_bw | ACTIVE |
301 * +-------------------+ |
302 * inactive timer | non contending |
303 * fired +------------------+
305 * The task_non_contending() function is invoked when a task
306 * blocks, and checks if the 0-lag time already passed or
307 * not (in the first case, it directly updates running_bw;
308 * in the second case, it arms the inactive timer).
310 * The task_contending() function is invoked when a task wakes
311 * up, and checks if the task is still in the "ACTIVE non contending"
312 * state or not (in the second case, it updates running_bw).
314 static void task_non_contending(struct task_struct *p)
316 struct sched_dl_entity *dl_se = &p->dl;
317 struct hrtimer *timer = &dl_se->inactive_timer;
318 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
319 struct rq *rq = rq_of_dl_rq(dl_rq);
323 * If this is a non-deadline task that has been boosted,
326 if (dl_se->dl_runtime == 0)
329 if (dl_entity_is_special(dl_se))
332 WARN_ON(dl_se->dl_non_contending);
334 zerolag_time = dl_se->deadline -
335 div64_long((dl_se->runtime * dl_se->dl_period),
339 * Using relative times instead of the absolute "0-lag time"
340 * allows to simplify the code
342 zerolag_time -= rq_clock(rq);
345 * If the "0-lag time" already passed, decrease the active
346 * utilization now, instead of starting a timer
348 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
350 sub_running_bw(dl_se, dl_rq);
351 if (!dl_task(p) || p->state == TASK_DEAD) {
352 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
354 if (p->state == TASK_DEAD)
355 sub_rq_bw(&p->dl, &rq->dl);
356 raw_spin_lock(&dl_b->lock);
357 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
358 __dl_clear_params(p);
359 raw_spin_unlock(&dl_b->lock);
365 dl_se->dl_non_contending = 1;
367 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD);
370 static void task_contending(struct sched_dl_entity *dl_se, int flags)
372 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
375 * If this is a non-deadline task that has been boosted,
378 if (dl_se->dl_runtime == 0)
381 if (flags & ENQUEUE_MIGRATED)
382 add_rq_bw(dl_se, dl_rq);
384 if (dl_se->dl_non_contending) {
385 dl_se->dl_non_contending = 0;
387 * If the timer handler is currently running and the
388 * timer cannot be cancelled, inactive_task_timer()
389 * will see that dl_not_contending is not set, and
390 * will not touch the rq's active utilization,
391 * so we are still safe.
393 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1)
394 put_task_struct(dl_task_of(dl_se));
397 * Since "dl_non_contending" is not set, the
398 * task's utilization has already been removed from
399 * active utilization (either when the task blocked,
400 * when the "inactive timer" fired).
403 add_running_bw(dl_se, dl_rq);
407 static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
409 struct sched_dl_entity *dl_se = &p->dl;
411 return dl_rq->root.rb_leftmost == &dl_se->rb_node;
414 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
416 void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
418 raw_spin_lock_init(&dl_b->dl_runtime_lock);
419 dl_b->dl_period = period;
420 dl_b->dl_runtime = runtime;
423 void init_dl_bw(struct dl_bw *dl_b)
425 raw_spin_lock_init(&dl_b->lock);
426 raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock);
427 if (global_rt_runtime() == RUNTIME_INF)
430 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
431 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock);
435 void init_dl_rq(struct dl_rq *dl_rq)
437 dl_rq->root = RB_ROOT_CACHED;
440 /* zero means no -deadline tasks */
441 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
443 dl_rq->dl_nr_migratory = 0;
444 dl_rq->overloaded = 0;
445 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
447 init_dl_bw(&dl_rq->dl_bw);
450 dl_rq->running_bw = 0;
452 init_dl_rq_bw_ratio(dl_rq);
457 static inline int dl_overloaded(struct rq *rq)
459 return atomic_read(&rq->rd->dlo_count);
462 static inline void dl_set_overload(struct rq *rq)
467 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
469 * Must be visible before the overload count is
470 * set (as in sched_rt.c).
472 * Matched by the barrier in pull_dl_task().
475 atomic_inc(&rq->rd->dlo_count);
478 static inline void dl_clear_overload(struct rq *rq)
483 atomic_dec(&rq->rd->dlo_count);
484 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
487 static void update_dl_migration(struct dl_rq *dl_rq)
489 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
490 if (!dl_rq->overloaded) {
491 dl_set_overload(rq_of_dl_rq(dl_rq));
492 dl_rq->overloaded = 1;
494 } else if (dl_rq->overloaded) {
495 dl_clear_overload(rq_of_dl_rq(dl_rq));
496 dl_rq->overloaded = 0;
500 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
502 struct task_struct *p = dl_task_of(dl_se);
504 if (p->nr_cpus_allowed > 1)
505 dl_rq->dl_nr_migratory++;
507 update_dl_migration(dl_rq);
510 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
512 struct task_struct *p = dl_task_of(dl_se);
514 if (p->nr_cpus_allowed > 1)
515 dl_rq->dl_nr_migratory--;
517 update_dl_migration(dl_rq);
520 #define __node_2_pdl(node) \
521 rb_entry((node), struct task_struct, pushable_dl_tasks)
523 static inline bool __pushable_less(struct rb_node *a, const struct rb_node *b)
525 return dl_entity_preempt(&__node_2_pdl(a)->dl, &__node_2_pdl(b)->dl);
529 * The list of pushable -deadline task is not a plist, like in
530 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
532 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
534 struct rb_node *leftmost;
536 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
538 leftmost = rb_add_cached(&p->pushable_dl_tasks,
539 &rq->dl.pushable_dl_tasks_root,
542 rq->dl.earliest_dl.next = p->dl.deadline;
545 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
547 struct dl_rq *dl_rq = &rq->dl;
548 struct rb_root_cached *root = &dl_rq->pushable_dl_tasks_root;
549 struct rb_node *leftmost;
551 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
554 leftmost = rb_erase_cached(&p->pushable_dl_tasks, root);
556 dl_rq->earliest_dl.next = __node_2_pdl(leftmost)->dl.deadline;
558 RB_CLEAR_NODE(&p->pushable_dl_tasks);
561 static inline int has_pushable_dl_tasks(struct rq *rq)
563 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
566 static int push_dl_task(struct rq *rq);
568 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
570 return rq->online && dl_task(prev);
573 static DEFINE_PER_CPU(struct callback_head, dl_push_head);
574 static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
576 static void push_dl_tasks(struct rq *);
577 static void pull_dl_task(struct rq *);
579 static inline void deadline_queue_push_tasks(struct rq *rq)
581 if (!has_pushable_dl_tasks(rq))
584 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
587 static inline void deadline_queue_pull_task(struct rq *rq)
589 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
592 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
594 static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
596 struct rq *later_rq = NULL;
599 later_rq = find_lock_later_rq(p, rq);
604 * If we cannot preempt any rq, fall back to pick any
607 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
608 if (cpu >= nr_cpu_ids) {
610 * Failed to find any suitable CPU.
611 * The task will never come back!
613 BUG_ON(dl_bandwidth_enabled());
616 * If admission control is disabled we
617 * try a little harder to let the task
620 cpu = cpumask_any(cpu_active_mask);
622 later_rq = cpu_rq(cpu);
623 double_lock_balance(rq, later_rq);
626 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
628 * Inactive timer is armed (or callback is running, but
629 * waiting for us to release rq locks). In any case, when it
630 * will fire (or continue), it will see running_bw of this
631 * task migrated to later_rq (and correctly handle it).
633 sub_running_bw(&p->dl, &rq->dl);
634 sub_rq_bw(&p->dl, &rq->dl);
636 add_rq_bw(&p->dl, &later_rq->dl);
637 add_running_bw(&p->dl, &later_rq->dl);
639 sub_rq_bw(&p->dl, &rq->dl);
640 add_rq_bw(&p->dl, &later_rq->dl);
644 * And we finally need to fixup root_domain(s) bandwidth accounting,
645 * since p is still hanging out in the old (now moved to default) root
648 dl_b = &rq->rd->dl_bw;
649 raw_spin_lock(&dl_b->lock);
650 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
651 raw_spin_unlock(&dl_b->lock);
653 dl_b = &later_rq->rd->dl_bw;
654 raw_spin_lock(&dl_b->lock);
655 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
656 raw_spin_unlock(&dl_b->lock);
658 set_task_cpu(p, later_rq->cpu);
659 double_unlock_balance(later_rq, rq);
667 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
672 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
677 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
682 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
686 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
691 static inline void pull_dl_task(struct rq *rq)
695 static inline void deadline_queue_push_tasks(struct rq *rq)
699 static inline void deadline_queue_pull_task(struct rq *rq)
702 #endif /* CONFIG_SMP */
704 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
705 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
706 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags);
709 * We are being explicitly informed that a new instance is starting,
710 * and this means that:
711 * - the absolute deadline of the entity has to be placed at
712 * current time + relative deadline;
713 * - the runtime of the entity has to be set to the maximum value.
715 * The capability of specifying such event is useful whenever a -deadline
716 * entity wants to (try to!) synchronize its behaviour with the scheduler's
717 * one, and to (try to!) reconcile itself with its own scheduling
720 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
722 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
723 struct rq *rq = rq_of_dl_rq(dl_rq);
725 WARN_ON(is_dl_boosted(dl_se));
726 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
729 * We are racing with the deadline timer. So, do nothing because
730 * the deadline timer handler will take care of properly recharging
731 * the runtime and postponing the deadline
733 if (dl_se->dl_throttled)
737 * We use the regular wall clock time to set deadlines in the
738 * future; in fact, we must consider execution overheads (time
739 * spent on hardirq context, etc.).
741 dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline;
742 dl_se->runtime = dl_se->dl_runtime;
746 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
747 * possibility of a entity lasting more than what it declared, and thus
748 * exhausting its runtime.
750 * Here we are interested in making runtime overrun possible, but we do
751 * not want a entity which is misbehaving to affect the scheduling of all
753 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
754 * is used, in order to confine each entity within its own bandwidth.
756 * This function deals exactly with that, and ensures that when the runtime
757 * of a entity is replenished, its deadline is also postponed. That ensures
758 * the overrunning entity can't interfere with other entity in the system and
759 * can't make them miss their deadlines. Reasons why this kind of overruns
760 * could happen are, typically, a entity voluntarily trying to overcome its
761 * runtime, or it just underestimated it during sched_setattr().
763 static void replenish_dl_entity(struct sched_dl_entity *dl_se)
765 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
766 struct rq *rq = rq_of_dl_rq(dl_rq);
768 BUG_ON(pi_of(dl_se)->dl_runtime <= 0);
771 * This could be the case for a !-dl task that is boosted.
772 * Just go with full inherited parameters.
774 if (dl_se->dl_deadline == 0) {
775 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
776 dl_se->runtime = pi_of(dl_se)->dl_runtime;
779 if (dl_se->dl_yielded && dl_se->runtime > 0)
783 * We keep moving the deadline away until we get some
784 * available runtime for the entity. This ensures correct
785 * handling of situations where the runtime overrun is
788 while (dl_se->runtime <= 0) {
789 dl_se->deadline += pi_of(dl_se)->dl_period;
790 dl_se->runtime += pi_of(dl_se)->dl_runtime;
794 * At this point, the deadline really should be "in
795 * the future" with respect to rq->clock. If it's
796 * not, we are, for some reason, lagging too much!
797 * Anyway, after having warn userspace abut that,
798 * we still try to keep the things running by
799 * resetting the deadline and the budget of the
802 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
803 printk_deferred_once("sched: DL replenish lagged too much\n");
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)
809 dl_se->dl_yielded = 0;
810 if (dl_se->dl_throttled)
811 dl_se->dl_throttled = 0;
815 * Here we check if --at time t-- an entity (which is probably being
816 * [re]activated or, in general, enqueued) can use its remaining runtime
817 * and its current deadline _without_ exceeding the bandwidth it is
818 * assigned (function returns true if it can't). We are in fact applying
819 * one of the CBS rules: when a task wakes up, if the residual runtime
820 * over residual deadline fits within the allocated bandwidth, then we
821 * can keep the current (absolute) deadline and residual budget without
822 * disrupting the schedulability of the system. Otherwise, we should
823 * refill the runtime and set the deadline a period in the future,
824 * because keeping the current (absolute) deadline of the task would
825 * result in breaking guarantees promised to other tasks (refer to
826 * Documentation/scheduler/sched-deadline.rst for more information).
828 * This function returns true if:
830 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
832 * IOW we can't recycle current parameters.
834 * Notice that the bandwidth check is done against the deadline. For
835 * task with deadline equal to period this is the same of using
836 * dl_period instead of dl_deadline in the equation above.
838 static bool dl_entity_overflow(struct sched_dl_entity *dl_se, u64 t)
843 * left and right are the two sides of the equation above,
844 * after a bit of shuffling to use multiplications instead
847 * Note that none of the time values involved in the two
848 * multiplications are absolute: dl_deadline and dl_runtime
849 * are the relative deadline and the maximum runtime of each
850 * instance, runtime is the runtime left for the last instance
851 * and (deadline - t), since t is rq->clock, is the time left
852 * to the (absolute) deadline. Even if overflowing the u64 type
853 * is very unlikely to occur in both cases, here we scale down
854 * as we want to avoid that risk at all. Scaling down by 10
855 * means that we reduce granularity to 1us. We are fine with it,
856 * since this is only a true/false check and, anyway, thinking
857 * of anything below microseconds resolution is actually fiction
858 * (but still we want to give the user that illusion >;).
860 left = (pi_of(dl_se)->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
861 right = ((dl_se->deadline - t) >> DL_SCALE) *
862 (pi_of(dl_se)->dl_runtime >> DL_SCALE);
864 return dl_time_before(right, left);
868 * Revised wakeup rule [1]: For self-suspending tasks, rather then
869 * re-initializing task's runtime and deadline, the revised wakeup
870 * rule adjusts the task's runtime to avoid the task to overrun its
873 * Reasoning: a task may overrun the density if:
874 * runtime / (deadline - t) > dl_runtime / dl_deadline
876 * Therefore, runtime can be adjusted to:
877 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
879 * In such way that runtime will be equal to the maximum density
880 * the task can use without breaking any rule.
882 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
883 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
886 update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
888 u64 laxity = dl_se->deadline - rq_clock(rq);
891 * If the task has deadline < period, and the deadline is in the past,
892 * it should already be throttled before this check.
894 * See update_dl_entity() comments for further details.
896 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
898 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
902 * Regarding the deadline, a task with implicit deadline has a relative
903 * deadline == relative period. A task with constrained deadline has a
904 * relative deadline <= relative period.
906 * We support constrained deadline tasks. However, there are some restrictions
907 * applied only for tasks which do not have an implicit deadline. See
908 * update_dl_entity() to know more about such restrictions.
910 * The dl_is_implicit() returns true if the task has an implicit deadline.
912 static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
914 return dl_se->dl_deadline == dl_se->dl_period;
918 * When a deadline entity is placed in the runqueue, its runtime and deadline
919 * might need to be updated. This is done by a CBS wake up rule. There are two
920 * different rules: 1) the original CBS; and 2) the Revisited CBS.
922 * When the task is starting a new period, the Original CBS is used. In this
923 * case, the runtime is replenished and a new absolute deadline is set.
925 * When a task is queued before the begin of the next period, using the
926 * remaining runtime and deadline could make the entity to overflow, see
927 * dl_entity_overflow() to find more about runtime overflow. When such case
928 * is detected, the runtime and deadline need to be updated.
930 * If the task has an implicit deadline, i.e., deadline == period, the Original
931 * CBS is applied. the runtime is replenished and a new absolute deadline is
932 * set, as in the previous cases.
934 * However, the Original CBS does not work properly for tasks with
935 * deadline < period, which are said to have a constrained deadline. By
936 * applying the Original CBS, a constrained deadline task would be able to run
937 * runtime/deadline in a period. With deadline < period, the task would
938 * overrun the runtime/period allowed bandwidth, breaking the admission test.
940 * In order to prevent this misbehave, the Revisited CBS is used for
941 * constrained deadline tasks when a runtime overflow is detected. In the
942 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
943 * the remaining runtime of the task is reduced to avoid runtime overflow.
944 * Please refer to the comments update_dl_revised_wakeup() function to find
945 * more about the Revised CBS rule.
947 static void update_dl_entity(struct sched_dl_entity *dl_se)
949 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
950 struct rq *rq = rq_of_dl_rq(dl_rq);
952 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
953 dl_entity_overflow(dl_se, rq_clock(rq))) {
955 if (unlikely(!dl_is_implicit(dl_se) &&
956 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
957 !is_dl_boosted(dl_se))) {
958 update_dl_revised_wakeup(dl_se, rq);
962 dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
963 dl_se->runtime = pi_of(dl_se)->dl_runtime;
967 static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
969 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
973 * If the entity depleted all its runtime, and if we want it to sleep
974 * while waiting for some new execution time to become available, we
975 * set the bandwidth replenishment timer to the replenishment instant
976 * and try to activate it.
978 * Notice that it is important for the caller to know if the timer
979 * actually started or not (i.e., the replenishment instant is in
980 * the future or in the past).
982 static int start_dl_timer(struct task_struct *p)
984 struct sched_dl_entity *dl_se = &p->dl;
985 struct hrtimer *timer = &dl_se->dl_timer;
986 struct rq *rq = task_rq(p);
990 lockdep_assert_held(&rq->lock);
993 * We want the timer to fire at the deadline, but considering
994 * that it is actually coming from rq->clock and not from
995 * hrtimer's time base reading.
997 act = ns_to_ktime(dl_next_period(dl_se));
998 now = hrtimer_cb_get_time(timer);
999 delta = ktime_to_ns(now) - rq_clock(rq);
1000 act = ktime_add_ns(act, delta);
1003 * If the expiry time already passed, e.g., because the value
1004 * chosen as the deadline is too small, don't even try to
1005 * start the timer in the past!
1007 if (ktime_us_delta(act, now) < 0)
1011 * !enqueued will guarantee another callback; even if one is already in
1012 * progress. This ensures a balanced {get,put}_task_struct().
1014 * The race against __run_timer() clearing the enqueued state is
1015 * harmless because we're holding task_rq()->lock, therefore the timer
1016 * expiring after we've done the check will wait on its task_rq_lock()
1017 * and observe our state.
1019 if (!hrtimer_is_queued(timer)) {
1021 hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD);
1028 * This is the bandwidth enforcement timer callback. If here, we know
1029 * a task is not on its dl_rq, since the fact that the timer was running
1030 * means the task is throttled and needs a runtime replenishment.
1032 * However, what we actually do depends on the fact the task is active,
1033 * (it is on its rq) or has been removed from there by a call to
1034 * dequeue_task_dl(). In the former case we must issue the runtime
1035 * replenishment and add the task back to the dl_rq; in the latter, we just
1036 * do nothing but clearing dl_throttled, so that runtime and deadline
1037 * updating (and the queueing back to dl_rq) will be done by the
1038 * next call to enqueue_task_dl().
1040 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
1042 struct sched_dl_entity *dl_se = container_of(timer,
1043 struct sched_dl_entity,
1045 struct task_struct *p = dl_task_of(dl_se);
1049 rq = task_rq_lock(p, &rf);
1052 * The task might have changed its scheduling policy to something
1053 * different than SCHED_DEADLINE (through switched_from_dl()).
1059 * The task might have been boosted by someone else and might be in the
1060 * boosting/deboosting path, its not throttled.
1062 if (is_dl_boosted(dl_se))
1066 * Spurious timer due to start_dl_timer() race; or we already received
1067 * a replenishment from rt_mutex_setprio().
1069 if (!dl_se->dl_throttled)
1073 update_rq_clock(rq);
1076 * If the throttle happened during sched-out; like:
1083 * __dequeue_task_dl()
1086 * We can be both throttled and !queued. Replenish the counter
1087 * but do not enqueue -- wait for our wakeup to do that.
1089 if (!task_on_rq_queued(p)) {
1090 replenish_dl_entity(dl_se);
1095 if (unlikely(!rq->online)) {
1097 * If the runqueue is no longer available, migrate the
1098 * task elsewhere. This necessarily changes rq.
1100 lockdep_unpin_lock(&rq->lock, rf.cookie);
1101 rq = dl_task_offline_migration(rq, p);
1102 rf.cookie = lockdep_pin_lock(&rq->lock);
1103 update_rq_clock(rq);
1106 * Now that the task has been migrated to the new RQ and we
1107 * have that locked, proceed as normal and enqueue the task
1113 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1114 if (dl_task(rq->curr))
1115 check_preempt_curr_dl(rq, p, 0);
1121 * Queueing this task back might have overloaded rq, check if we need
1122 * to kick someone away.
1124 if (has_pushable_dl_tasks(rq)) {
1126 * Nothing relies on rq->lock after this, so its safe to drop
1129 rq_unpin_lock(rq, &rf);
1131 rq_repin_lock(rq, &rf);
1136 task_rq_unlock(rq, p, &rf);
1139 * This can free the task_struct, including this hrtimer, do not touch
1140 * anything related to that after this.
1144 return HRTIMER_NORESTART;
1147 void init_dl_task_timer(struct sched_dl_entity *dl_se)
1149 struct hrtimer *timer = &dl_se->dl_timer;
1151 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1152 timer->function = dl_task_timer;
1156 * During the activation, CBS checks if it can reuse the current task's
1157 * runtime and period. If the deadline of the task is in the past, CBS
1158 * cannot use the runtime, and so it replenishes the task. This rule
1159 * works fine for implicit deadline tasks (deadline == period), and the
1160 * CBS was designed for implicit deadline tasks. However, a task with
1161 * constrained deadline (deadline < period) might be awakened after the
1162 * deadline, but before the next period. In this case, replenishing the
1163 * task would allow it to run for runtime / deadline. As in this case
1164 * deadline < period, CBS enables a task to run for more than the
1165 * runtime / period. In a very loaded system, this can cause a domino
1166 * effect, making other tasks miss their deadlines.
1168 * To avoid this problem, in the activation of a constrained deadline
1169 * task after the deadline but before the next period, throttle the
1170 * task and set the replenishing timer to the begin of the next period,
1171 * unless it is boosted.
1173 static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1175 struct task_struct *p = dl_task_of(dl_se);
1176 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1178 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1179 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1180 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(p)))
1182 dl_se->dl_throttled = 1;
1183 if (dl_se->runtime > 0)
1189 int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1191 return (dl_se->runtime <= 0);
1194 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
1197 * This function implements the GRUB accounting rule:
1198 * according to the GRUB reclaiming algorithm, the runtime is
1199 * not decreased as "dq = -dt", but as
1200 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1201 * where u is the utilization of the task, Umax is the maximum reclaimable
1202 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1203 * as the difference between the "total runqueue utilization" and the
1204 * runqueue active utilization, and Uextra is the (per runqueue) extra
1205 * reclaimable utilization.
1206 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1207 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1209 * Since rq->dl.bw_ratio contains 1 / Umax multipled by 2^RATIO_SHIFT,
1210 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1211 * Since delta is a 64 bit variable, to have an overflow its value
1212 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
1213 * So, overflow is not an issue here.
1215 static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1217 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
1219 u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT;
1222 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1223 * we compare u_inact + rq->dl.extra_bw with
1224 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1225 * u_inact + rq->dl.extra_bw can be larger than
1226 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
1227 * leading to wrong results)
1229 if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min)
1232 u_act = BW_UNIT - u_inact - rq->dl.extra_bw;
1234 return (delta * u_act) >> BW_SHIFT;
1238 * Update the current task's runtime statistics (provided it is still
1239 * a -deadline task and has not been removed from the dl_rq).
1241 static void update_curr_dl(struct rq *rq)
1243 struct task_struct *curr = rq->curr;
1244 struct sched_dl_entity *dl_se = &curr->dl;
1245 u64 delta_exec, scaled_delta_exec;
1246 int cpu = cpu_of(rq);
1249 if (!dl_task(curr) || !on_dl_rq(dl_se))
1253 * Consumed budget is computed considering the time as
1254 * observed by schedulable tasks (excluding time spent
1255 * in hardirq context, etc.). Deadlines are instead
1256 * computed using hard walltime. This seems to be the more
1257 * natural solution, but the full ramifications of this
1258 * approach need further study.
1260 now = rq_clock_task(rq);
1261 delta_exec = now - curr->se.exec_start;
1262 if (unlikely((s64)delta_exec <= 0)) {
1263 if (unlikely(dl_se->dl_yielded))
1268 schedstat_set(curr->se.statistics.exec_max,
1269 max(curr->se.statistics.exec_max, delta_exec));
1271 curr->se.sum_exec_runtime += delta_exec;
1272 account_group_exec_runtime(curr, delta_exec);
1274 curr->se.exec_start = now;
1275 cgroup_account_cputime(curr, delta_exec);
1277 if (dl_entity_is_special(dl_se))
1281 * For tasks that participate in GRUB, we implement GRUB-PA: the
1282 * spare reclaimed bandwidth is used to clock down frequency.
1284 * For the others, we still need to scale reservation parameters
1285 * according to current frequency and CPU maximum capacity.
1287 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1288 scaled_delta_exec = grub_reclaim(delta_exec,
1292 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1293 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1295 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1296 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1299 dl_se->runtime -= scaled_delta_exec;
1302 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1303 dl_se->dl_throttled = 1;
1305 /* If requested, inform the user about runtime overruns. */
1306 if (dl_runtime_exceeded(dl_se) &&
1307 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1308 dl_se->dl_overrun = 1;
1310 __dequeue_task_dl(rq, curr, 0);
1311 if (unlikely(is_dl_boosted(dl_se) || !start_dl_timer(curr)))
1312 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1314 if (!is_leftmost(curr, &rq->dl))
1319 * Because -- for now -- we share the rt bandwidth, we need to
1320 * account our runtime there too, otherwise actual rt tasks
1321 * would be able to exceed the shared quota.
1323 * Account to the root rt group for now.
1325 * The solution we're working towards is having the RT groups scheduled
1326 * using deadline servers -- however there's a few nasties to figure
1327 * out before that can happen.
1329 if (rt_bandwidth_enabled()) {
1330 struct rt_rq *rt_rq = &rq->rt;
1332 raw_spin_lock(&rt_rq->rt_runtime_lock);
1334 * We'll let actual RT tasks worry about the overflow here, we
1335 * have our own CBS to keep us inline; only account when RT
1336 * bandwidth is relevant.
1338 if (sched_rt_bandwidth_account(rt_rq))
1339 rt_rq->rt_time += delta_exec;
1340 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1344 static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1346 struct sched_dl_entity *dl_se = container_of(timer,
1347 struct sched_dl_entity,
1349 struct task_struct *p = dl_task_of(dl_se);
1353 rq = task_rq_lock(p, &rf);
1356 update_rq_clock(rq);
1358 if (!dl_task(p) || p->state == TASK_DEAD) {
1359 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1361 if (p->state == TASK_DEAD && dl_se->dl_non_contending) {
1362 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1363 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1364 dl_se->dl_non_contending = 0;
1367 raw_spin_lock(&dl_b->lock);
1368 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1369 raw_spin_unlock(&dl_b->lock);
1370 __dl_clear_params(p);
1374 if (dl_se->dl_non_contending == 0)
1377 sub_running_bw(dl_se, &rq->dl);
1378 dl_se->dl_non_contending = 0;
1380 task_rq_unlock(rq, p, &rf);
1383 return HRTIMER_NORESTART;
1386 void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1388 struct hrtimer *timer = &dl_se->inactive_timer;
1390 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1391 timer->function = inactive_task_timer;
1396 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1398 struct rq *rq = rq_of_dl_rq(dl_rq);
1400 if (dl_rq->earliest_dl.curr == 0 ||
1401 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1402 if (dl_rq->earliest_dl.curr == 0)
1403 cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_HIGHER);
1404 dl_rq->earliest_dl.curr = deadline;
1405 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1409 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1411 struct rq *rq = rq_of_dl_rq(dl_rq);
1414 * Since we may have removed our earliest (and/or next earliest)
1415 * task we must recompute them.
1417 if (!dl_rq->dl_nr_running) {
1418 dl_rq->earliest_dl.curr = 0;
1419 dl_rq->earliest_dl.next = 0;
1420 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1421 cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio.curr);
1423 struct rb_node *leftmost = dl_rq->root.rb_leftmost;
1424 struct sched_dl_entity *entry;
1426 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
1427 dl_rq->earliest_dl.curr = entry->deadline;
1428 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1434 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1435 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1437 #endif /* CONFIG_SMP */
1440 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1442 int prio = dl_task_of(dl_se)->prio;
1443 u64 deadline = dl_se->deadline;
1445 WARN_ON(!dl_prio(prio));
1446 dl_rq->dl_nr_running++;
1447 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1449 inc_dl_deadline(dl_rq, deadline);
1450 inc_dl_migration(dl_se, dl_rq);
1454 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1456 int prio = dl_task_of(dl_se)->prio;
1458 WARN_ON(!dl_prio(prio));
1459 WARN_ON(!dl_rq->dl_nr_running);
1460 dl_rq->dl_nr_running--;
1461 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1463 dec_dl_deadline(dl_rq, dl_se->deadline);
1464 dec_dl_migration(dl_se, dl_rq);
1467 #define __node_2_dle(node) \
1468 rb_entry((node), struct sched_dl_entity, rb_node)
1470 static inline bool __dl_less(struct rb_node *a, const struct rb_node *b)
1472 return dl_time_before(__node_2_dle(a)->deadline, __node_2_dle(b)->deadline);
1475 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1477 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1479 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
1481 rb_add_cached(&dl_se->rb_node, &dl_rq->root, __dl_less);
1483 inc_dl_tasks(dl_se, dl_rq);
1486 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1488 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1490 if (RB_EMPTY_NODE(&dl_se->rb_node))
1493 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1495 RB_CLEAR_NODE(&dl_se->rb_node);
1497 dec_dl_tasks(dl_se, dl_rq);
1501 enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags)
1503 BUG_ON(on_dl_rq(dl_se));
1506 * If this is a wakeup or a new instance, the scheduling
1507 * parameters of the task might need updating. Otherwise,
1508 * we want a replenishment of its runtime.
1510 if (flags & ENQUEUE_WAKEUP) {
1511 task_contending(dl_se, flags);
1512 update_dl_entity(dl_se);
1513 } else if (flags & ENQUEUE_REPLENISH) {
1514 replenish_dl_entity(dl_se);
1515 } else if ((flags & ENQUEUE_RESTORE) &&
1516 dl_time_before(dl_se->deadline,
1517 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) {
1518 setup_new_dl_entity(dl_se);
1521 __enqueue_dl_entity(dl_se);
1524 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
1526 __dequeue_dl_entity(dl_se);
1529 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1531 if (is_dl_boosted(&p->dl)) {
1533 * Because of delays in the detection of the overrun of a
1534 * thread's runtime, it might be the case that a thread
1535 * goes to sleep in a rt mutex with negative runtime. As
1536 * a consequence, the thread will be throttled.
1538 * While waiting for the mutex, this thread can also be
1539 * boosted via PI, resulting in a thread that is throttled
1540 * and boosted at the same time.
1542 * In this case, the boost overrides the throttle.
1544 if (p->dl.dl_throttled) {
1546 * The replenish timer needs to be canceled. No
1547 * problem if it fires concurrently: boosted threads
1548 * are ignored in dl_task_timer().
1550 hrtimer_try_to_cancel(&p->dl.dl_timer);
1551 p->dl.dl_throttled = 0;
1553 } else if (!dl_prio(p->normal_prio)) {
1555 * Special case in which we have a !SCHED_DEADLINE task that is going
1556 * to be deboosted, but exceeds its runtime while doing so. No point in
1557 * replenishing it, as it's going to return back to its original
1558 * scheduling class after this. If it has been throttled, we need to
1559 * clear the flag, otherwise the task may wake up as throttled after
1560 * being boosted again with no means to replenish the runtime and clear
1563 p->dl.dl_throttled = 0;
1564 BUG_ON(!is_dl_boosted(&p->dl) || flags != ENQUEUE_REPLENISH);
1569 * Check if a constrained deadline task was activated
1570 * after the deadline but before the next period.
1571 * If that is the case, the task will be throttled and
1572 * the replenishment timer will be set to the next period.
1574 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
1575 dl_check_constrained_dl(&p->dl);
1577 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
1578 add_rq_bw(&p->dl, &rq->dl);
1579 add_running_bw(&p->dl, &rq->dl);
1583 * If p is throttled, we do not enqueue it. In fact, if it exhausted
1584 * its budget it needs a replenishment and, since it now is on
1585 * its rq, the bandwidth timer callback (which clearly has not
1586 * run yet) will take care of this.
1587 * However, the active utilization does not depend on the fact
1588 * that the task is on the runqueue or not (but depends on the
1589 * task's state - in GRUB parlance, "inactive" vs "active contending").
1590 * In other words, even if a task is throttled its utilization must
1591 * be counted in the active utilization; hence, we need to call
1594 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
1595 if (flags & ENQUEUE_WAKEUP)
1596 task_contending(&p->dl, flags);
1601 enqueue_dl_entity(&p->dl, flags);
1603 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
1604 enqueue_pushable_dl_task(rq, p);
1607 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1609 dequeue_dl_entity(&p->dl);
1610 dequeue_pushable_dl_task(rq, p);
1613 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1616 __dequeue_task_dl(rq, p, flags);
1618 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
1619 sub_running_bw(&p->dl, &rq->dl);
1620 sub_rq_bw(&p->dl, &rq->dl);
1624 * This check allows to start the inactive timer (or to immediately
1625 * decrease the active utilization, if needed) in two cases:
1626 * when the task blocks and when it is terminating
1627 * (p->state == TASK_DEAD). We can handle the two cases in the same
1628 * way, because from GRUB's point of view the same thing is happening
1629 * (the task moves from "active contending" to "active non contending"
1632 if (flags & DEQUEUE_SLEEP)
1633 task_non_contending(p);
1637 * Yield task semantic for -deadline tasks is:
1639 * get off from the CPU until our next instance, with
1640 * a new runtime. This is of little use now, since we
1641 * don't have a bandwidth reclaiming mechanism. Anyway,
1642 * bandwidth reclaiming is planned for the future, and
1643 * yield_task_dl will indicate that some spare budget
1644 * is available for other task instances to use it.
1646 static void yield_task_dl(struct rq *rq)
1649 * We make the task go to sleep until its current deadline by
1650 * forcing its runtime to zero. This way, update_curr_dl() stops
1651 * it and the bandwidth timer will wake it up and will give it
1652 * new scheduling parameters (thanks to dl_yielded=1).
1654 rq->curr->dl.dl_yielded = 1;
1656 update_rq_clock(rq);
1659 * Tell update_rq_clock() that we've just updated,
1660 * so we don't do microscopic update in schedule()
1661 * and double the fastpath cost.
1663 rq_clock_skip_update(rq);
1668 static int find_later_rq(struct task_struct *task);
1671 select_task_rq_dl(struct task_struct *p, int cpu, int flags)
1673 struct task_struct *curr;
1677 if (!(flags & WF_TTWU))
1683 curr = READ_ONCE(rq->curr); /* unlocked access */
1686 * If we are dealing with a -deadline task, we must
1687 * decide where to wake it up.
1688 * If it has a later deadline and the current task
1689 * on this rq can't move (provided the waking task
1690 * can!) we prefer to send it somewhere else. On the
1691 * other hand, if it has a shorter deadline, we
1692 * try to make it stay here, it might be important.
1694 select_rq = unlikely(dl_task(curr)) &&
1695 (curr->nr_cpus_allowed < 2 ||
1696 !dl_entity_preempt(&p->dl, &curr->dl)) &&
1697 p->nr_cpus_allowed > 1;
1700 * Take the capacity of the CPU into account to
1701 * ensure it fits the requirement of the task.
1703 if (static_branch_unlikely(&sched_asym_cpucapacity))
1704 select_rq |= !dl_task_fits_capacity(p, cpu);
1707 int target = find_later_rq(p);
1710 (dl_time_before(p->dl.deadline,
1711 cpu_rq(target)->dl.earliest_dl.curr) ||
1712 (cpu_rq(target)->dl.dl_nr_running == 0)))
1721 static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
1725 if (p->state != TASK_WAKING)
1730 * Since p->state == TASK_WAKING, set_task_cpu() has been called
1731 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
1732 * rq->lock is not... So, lock it
1734 raw_spin_lock(&rq->lock);
1735 if (p->dl.dl_non_contending) {
1736 sub_running_bw(&p->dl, &rq->dl);
1737 p->dl.dl_non_contending = 0;
1739 * If the timer handler is currently running and the
1740 * timer cannot be cancelled, inactive_task_timer()
1741 * will see that dl_not_contending is not set, and
1742 * will not touch the rq's active utilization,
1743 * so we are still safe.
1745 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
1748 sub_rq_bw(&p->dl, &rq->dl);
1749 raw_spin_unlock(&rq->lock);
1752 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
1755 * Current can't be migrated, useless to reschedule,
1756 * let's hope p can move out.
1758 if (rq->curr->nr_cpus_allowed == 1 ||
1759 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
1763 * p is migratable, so let's not schedule it and
1764 * see if it is pushed or pulled somewhere else.
1766 if (p->nr_cpus_allowed != 1 &&
1767 cpudl_find(&rq->rd->cpudl, p, NULL))
1773 static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1775 if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) {
1777 * This is OK, because current is on_cpu, which avoids it being
1778 * picked for load-balance and preemption/IRQs are still
1779 * disabled avoiding further scheduler activity on it and we've
1780 * not yet started the picking loop.
1782 rq_unpin_lock(rq, rf);
1784 rq_repin_lock(rq, rf);
1787 return sched_stop_runnable(rq) || sched_dl_runnable(rq);
1789 #endif /* CONFIG_SMP */
1792 * Only called when both the current and waking task are -deadline
1795 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
1798 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
1805 * In the unlikely case current and p have the same deadline
1806 * let us try to decide what's the best thing to do...
1808 if ((p->dl.deadline == rq->curr->dl.deadline) &&
1809 !test_tsk_need_resched(rq->curr))
1810 check_preempt_equal_dl(rq, p);
1811 #endif /* CONFIG_SMP */
1814 #ifdef CONFIG_SCHED_HRTICK
1815 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1817 hrtick_start(rq, p->dl.runtime);
1819 #else /* !CONFIG_SCHED_HRTICK */
1820 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
1825 static void set_next_task_dl(struct rq *rq, struct task_struct *p, bool first)
1827 p->se.exec_start = rq_clock_task(rq);
1829 /* You can't push away the running task */
1830 dequeue_pushable_dl_task(rq, p);
1835 if (hrtick_enabled_dl(rq))
1836 start_hrtick_dl(rq, p);
1838 if (rq->curr->sched_class != &dl_sched_class)
1839 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
1841 deadline_queue_push_tasks(rq);
1844 static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
1845 struct dl_rq *dl_rq)
1847 struct rb_node *left = rb_first_cached(&dl_rq->root);
1852 return rb_entry(left, struct sched_dl_entity, rb_node);
1855 static struct task_struct *pick_next_task_dl(struct rq *rq)
1857 struct sched_dl_entity *dl_se;
1858 struct dl_rq *dl_rq = &rq->dl;
1859 struct task_struct *p;
1861 if (!sched_dl_runnable(rq))
1864 dl_se = pick_next_dl_entity(rq, dl_rq);
1866 p = dl_task_of(dl_se);
1867 set_next_task_dl(rq, p, true);
1871 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
1875 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1876 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
1877 enqueue_pushable_dl_task(rq, p);
1881 * scheduler tick hitting a task of our scheduling class.
1883 * NOTE: This function can be called remotely by the tick offload that
1884 * goes along full dynticks. Therefore no local assumption can be made
1885 * and everything must be accessed through the @rq and @curr passed in
1888 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1892 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
1894 * Even when we have runtime, update_curr_dl() might have resulted in us
1895 * not being the leftmost task anymore. In that case NEED_RESCHED will
1896 * be set and schedule() will start a new hrtick for the next task.
1898 if (hrtick_enabled_dl(rq) && queued && p->dl.runtime > 0 &&
1899 is_leftmost(p, &rq->dl))
1900 start_hrtick_dl(rq, p);
1903 static void task_fork_dl(struct task_struct *p)
1906 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1913 /* Only try algorithms three times */
1914 #define DL_MAX_TRIES 3
1916 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
1918 if (!task_running(rq, p) &&
1919 cpumask_test_cpu(cpu, &p->cpus_mask))
1925 * Return the earliest pushable rq's task, which is suitable to be executed
1926 * on the CPU, NULL otherwise:
1928 static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
1930 struct rb_node *next_node = rq->dl.pushable_dl_tasks_root.rb_leftmost;
1931 struct task_struct *p = NULL;
1933 if (!has_pushable_dl_tasks(rq))
1938 p = rb_entry(next_node, struct task_struct, pushable_dl_tasks);
1940 if (pick_dl_task(rq, p, cpu))
1943 next_node = rb_next(next_node);
1950 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
1952 static int find_later_rq(struct task_struct *task)
1954 struct sched_domain *sd;
1955 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
1956 int this_cpu = smp_processor_id();
1957 int cpu = task_cpu(task);
1959 /* Make sure the mask is initialized first */
1960 if (unlikely(!later_mask))
1963 if (task->nr_cpus_allowed == 1)
1967 * We have to consider system topology and task affinity
1968 * first, then we can look for a suitable CPU.
1970 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
1974 * If we are here, some targets have been found, including
1975 * the most suitable which is, among the runqueues where the
1976 * current tasks have later deadlines than the task's one, the
1977 * rq with the latest possible one.
1979 * Now we check how well this matches with task's
1980 * affinity and system topology.
1982 * The last CPU where the task run is our first
1983 * guess, since it is most likely cache-hot there.
1985 if (cpumask_test_cpu(cpu, later_mask))
1988 * Check if this_cpu is to be skipped (i.e., it is
1989 * not in the mask) or not.
1991 if (!cpumask_test_cpu(this_cpu, later_mask))
1995 for_each_domain(cpu, sd) {
1996 if (sd->flags & SD_WAKE_AFFINE) {
2000 * If possible, preempting this_cpu is
2001 * cheaper than migrating.
2003 if (this_cpu != -1 &&
2004 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
2009 best_cpu = cpumask_any_and_distribute(later_mask,
2010 sched_domain_span(sd));
2012 * Last chance: if a CPU being in both later_mask
2013 * and current sd span is valid, that becomes our
2014 * choice. Of course, the latest possible CPU is
2015 * already under consideration through later_mask.
2017 if (best_cpu < nr_cpu_ids) {
2026 * At this point, all our guesses failed, we just return
2027 * 'something', and let the caller sort the things out.
2032 cpu = cpumask_any_distribute(later_mask);
2033 if (cpu < nr_cpu_ids)
2039 /* Locks the rq it finds */
2040 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
2042 struct rq *later_rq = NULL;
2046 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
2047 cpu = find_later_rq(task);
2049 if ((cpu == -1) || (cpu == rq->cpu))
2052 later_rq = cpu_rq(cpu);
2054 if (later_rq->dl.dl_nr_running &&
2055 !dl_time_before(task->dl.deadline,
2056 later_rq->dl.earliest_dl.curr)) {
2058 * Target rq has tasks of equal or earlier deadline,
2059 * retrying does not release any lock and is unlikely
2060 * to yield a different result.
2066 /* Retry if something changed. */
2067 if (double_lock_balance(rq, later_rq)) {
2068 if (unlikely(task_rq(task) != rq ||
2069 !cpumask_test_cpu(later_rq->cpu, &task->cpus_mask) ||
2070 task_running(rq, task) ||
2072 !task_on_rq_queued(task))) {
2073 double_unlock_balance(rq, later_rq);
2080 * If the rq we found has no -deadline task, or
2081 * its earliest one has a later deadline than our
2082 * task, the rq is a good one.
2084 if (!later_rq->dl.dl_nr_running ||
2085 dl_time_before(task->dl.deadline,
2086 later_rq->dl.earliest_dl.curr))
2089 /* Otherwise we try again. */
2090 double_unlock_balance(rq, later_rq);
2097 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2099 struct task_struct *p;
2101 if (!has_pushable_dl_tasks(rq))
2104 p = rb_entry(rq->dl.pushable_dl_tasks_root.rb_leftmost,
2105 struct task_struct, pushable_dl_tasks);
2107 BUG_ON(rq->cpu != task_cpu(p));
2108 BUG_ON(task_current(rq, p));
2109 BUG_ON(p->nr_cpus_allowed <= 1);
2111 BUG_ON(!task_on_rq_queued(p));
2112 BUG_ON(!dl_task(p));
2118 * See if the non running -deadline tasks on this rq
2119 * can be sent to some other CPU where they can preempt
2120 * and start executing.
2122 static int push_dl_task(struct rq *rq)
2124 struct task_struct *next_task;
2125 struct rq *later_rq;
2128 if (!rq->dl.overloaded)
2131 next_task = pick_next_pushable_dl_task(rq);
2136 if (is_migration_disabled(next_task))
2139 if (WARN_ON(next_task == rq->curr))
2143 * If next_task preempts rq->curr, and rq->curr
2144 * can move away, it makes sense to just reschedule
2145 * without going further in pushing next_task.
2147 if (dl_task(rq->curr) &&
2148 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2149 rq->curr->nr_cpus_allowed > 1) {
2154 /* We might release rq lock */
2155 get_task_struct(next_task);
2157 /* Will lock the rq it'll find */
2158 later_rq = find_lock_later_rq(next_task, rq);
2160 struct task_struct *task;
2163 * We must check all this again, since
2164 * find_lock_later_rq releases rq->lock and it is
2165 * then possible that next_task has migrated.
2167 task = pick_next_pushable_dl_task(rq);
2168 if (task == next_task) {
2170 * The task is still there. We don't try
2171 * again, some other CPU will pull it when ready.
2180 put_task_struct(next_task);
2185 deactivate_task(rq, next_task, 0);
2186 set_task_cpu(next_task, later_rq->cpu);
2189 * Update the later_rq clock here, because the clock is used
2190 * by the cpufreq_update_util() inside __add_running_bw().
2192 update_rq_clock(later_rq);
2193 activate_task(later_rq, next_task, ENQUEUE_NOCLOCK);
2196 resched_curr(later_rq);
2198 double_unlock_balance(rq, later_rq);
2201 put_task_struct(next_task);
2206 static void push_dl_tasks(struct rq *rq)
2208 /* push_dl_task() will return true if it moved a -deadline task */
2209 while (push_dl_task(rq))
2213 static void pull_dl_task(struct rq *this_rq)
2215 int this_cpu = this_rq->cpu, cpu;
2216 struct task_struct *p, *push_task;
2217 bool resched = false;
2219 u64 dmin = LONG_MAX;
2221 if (likely(!dl_overloaded(this_rq)))
2225 * Match the barrier from dl_set_overloaded; this guarantees that if we
2226 * see overloaded we must also see the dlo_mask bit.
2230 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2231 if (this_cpu == cpu)
2234 src_rq = cpu_rq(cpu);
2237 * It looks racy, abd it is! However, as in sched_rt.c,
2238 * we are fine with this.
2240 if (this_rq->dl.dl_nr_running &&
2241 dl_time_before(this_rq->dl.earliest_dl.curr,
2242 src_rq->dl.earliest_dl.next))
2245 /* Might drop this_rq->lock */
2247 double_lock_balance(this_rq, src_rq);
2250 * If there are no more pullable tasks on the
2251 * rq, we're done with it.
2253 if (src_rq->dl.dl_nr_running <= 1)
2256 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2259 * We found a task to be pulled if:
2260 * - it preempts our current (if there's one),
2261 * - it will preempt the last one we pulled (if any).
2263 if (p && dl_time_before(p->dl.deadline, dmin) &&
2264 (!this_rq->dl.dl_nr_running ||
2265 dl_time_before(p->dl.deadline,
2266 this_rq->dl.earliest_dl.curr))) {
2267 WARN_ON(p == src_rq->curr);
2268 WARN_ON(!task_on_rq_queued(p));
2271 * Then we pull iff p has actually an earlier
2272 * deadline than the current task of its runqueue.
2274 if (dl_time_before(p->dl.deadline,
2275 src_rq->curr->dl.deadline))
2278 if (is_migration_disabled(p)) {
2279 push_task = get_push_task(src_rq);
2281 deactivate_task(src_rq, p, 0);
2282 set_task_cpu(p, this_cpu);
2283 activate_task(this_rq, p, 0);
2284 dmin = p->dl.deadline;
2288 /* Is there any other task even earlier? */
2291 double_unlock_balance(this_rq, src_rq);
2294 raw_spin_unlock(&this_rq->lock);
2295 stop_one_cpu_nowait(src_rq->cpu, push_cpu_stop,
2296 push_task, &src_rq->push_work);
2297 raw_spin_lock(&this_rq->lock);
2302 resched_curr(this_rq);
2306 * Since the task is not running and a reschedule is not going to happen
2307 * anytime soon on its runqueue, we try pushing it away now.
2309 static void task_woken_dl(struct rq *rq, struct task_struct *p)
2311 if (!task_running(rq, p) &&
2312 !test_tsk_need_resched(rq->curr) &&
2313 p->nr_cpus_allowed > 1 &&
2314 dl_task(rq->curr) &&
2315 (rq->curr->nr_cpus_allowed < 2 ||
2316 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2321 static void set_cpus_allowed_dl(struct task_struct *p,
2322 const struct cpumask *new_mask,
2325 struct root_domain *src_rd;
2328 BUG_ON(!dl_task(p));
2333 * Migrating a SCHED_DEADLINE task between exclusive
2334 * cpusets (different root_domains) entails a bandwidth
2335 * update. We already made space for us in the destination
2336 * domain (see cpuset_can_attach()).
2338 if (!cpumask_intersects(src_rd->span, new_mask)) {
2339 struct dl_bw *src_dl_b;
2341 src_dl_b = dl_bw_of(cpu_of(rq));
2343 * We now free resources of the root_domain we are migrating
2344 * off. In the worst case, sched_setattr() may temporary fail
2345 * until we complete the update.
2347 raw_spin_lock(&src_dl_b->lock);
2348 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2349 raw_spin_unlock(&src_dl_b->lock);
2352 set_cpus_allowed_common(p, new_mask, flags);
2355 /* Assumes rq->lock is held */
2356 static void rq_online_dl(struct rq *rq)
2358 if (rq->dl.overloaded)
2359 dl_set_overload(rq);
2361 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2362 if (rq->dl.dl_nr_running > 0)
2363 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2366 /* Assumes rq->lock is held */
2367 static void rq_offline_dl(struct rq *rq)
2369 if (rq->dl.overloaded)
2370 dl_clear_overload(rq);
2372 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2373 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2376 void __init init_sched_dl_class(void)
2380 for_each_possible_cpu(i)
2381 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2382 GFP_KERNEL, cpu_to_node(i));
2385 void dl_add_task_root_domain(struct task_struct *p)
2391 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2393 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
2397 rq = __task_rq_lock(p, &rf);
2399 dl_b = &rq->rd->dl_bw;
2400 raw_spin_lock(&dl_b->lock);
2402 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2404 raw_spin_unlock(&dl_b->lock);
2406 task_rq_unlock(rq, p, &rf);
2409 void dl_clear_root_domain(struct root_domain *rd)
2411 unsigned long flags;
2413 raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
2414 rd->dl_bw.total_bw = 0;
2415 raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
2418 #endif /* CONFIG_SMP */
2420 static void switched_from_dl(struct rq *rq, struct task_struct *p)
2423 * task_non_contending() can start the "inactive timer" (if the 0-lag
2424 * time is in the future). If the task switches back to dl before
2425 * the "inactive timer" fires, it can continue to consume its current
2426 * runtime using its current deadline. If it stays outside of
2427 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2428 * will reset the task parameters.
2430 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2431 task_non_contending(p);
2433 if (!task_on_rq_queued(p)) {
2435 * Inactive timer is armed. However, p is leaving DEADLINE and
2436 * might migrate away from this rq while continuing to run on
2437 * some other class. We need to remove its contribution from
2438 * this rq running_bw now, or sub_rq_bw (below) will complain.
2440 if (p->dl.dl_non_contending)
2441 sub_running_bw(&p->dl, &rq->dl);
2442 sub_rq_bw(&p->dl, &rq->dl);
2446 * We cannot use inactive_task_timer() to invoke sub_running_bw()
2447 * at the 0-lag time, because the task could have been migrated
2448 * while SCHED_OTHER in the meanwhile.
2450 if (p->dl.dl_non_contending)
2451 p->dl.dl_non_contending = 0;
2454 * Since this might be the only -deadline task on the rq,
2455 * this is the right place to try to pull some other one
2456 * from an overloaded CPU, if any.
2458 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
2461 deadline_queue_pull_task(rq);
2465 * When switching to -deadline, we may overload the rq, then
2466 * we try to push someone off, if possible.
2468 static void switched_to_dl(struct rq *rq, struct task_struct *p)
2470 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2473 /* If p is not queued we will update its parameters at next wakeup. */
2474 if (!task_on_rq_queued(p)) {
2475 add_rq_bw(&p->dl, &rq->dl);
2480 if (rq->curr != p) {
2482 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
2483 deadline_queue_push_tasks(rq);
2485 if (dl_task(rq->curr))
2486 check_preempt_curr_dl(rq, p, 0);
2493 * If the scheduling parameters of a -deadline task changed,
2494 * a push or pull operation might be needed.
2496 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
2499 if (task_on_rq_queued(p) || task_current(rq, p)) {
2502 * This might be too much, but unfortunately
2503 * we don't have the old deadline value, and
2504 * we can't argue if the task is increasing
2505 * or lowering its prio, so...
2507 if (!rq->dl.overloaded)
2508 deadline_queue_pull_task(rq);
2511 * If we now have a earlier deadline task than p,
2512 * then reschedule, provided p is still on this
2515 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
2519 * Again, we don't know if p has a earlier
2520 * or later deadline, so let's blindly set a
2521 * (maybe not needed) rescheduling point.
2524 #endif /* CONFIG_SMP */
2528 DEFINE_SCHED_CLASS(dl) = {
2530 .enqueue_task = enqueue_task_dl,
2531 .dequeue_task = dequeue_task_dl,
2532 .yield_task = yield_task_dl,
2534 .check_preempt_curr = check_preempt_curr_dl,
2536 .pick_next_task = pick_next_task_dl,
2537 .put_prev_task = put_prev_task_dl,
2538 .set_next_task = set_next_task_dl,
2541 .balance = balance_dl,
2542 .select_task_rq = select_task_rq_dl,
2543 .migrate_task_rq = migrate_task_rq_dl,
2544 .set_cpus_allowed = set_cpus_allowed_dl,
2545 .rq_online = rq_online_dl,
2546 .rq_offline = rq_offline_dl,
2547 .task_woken = task_woken_dl,
2548 .find_lock_rq = find_lock_later_rq,
2551 .task_tick = task_tick_dl,
2552 .task_fork = task_fork_dl,
2554 .prio_changed = prio_changed_dl,
2555 .switched_from = switched_from_dl,
2556 .switched_to = switched_to_dl,
2558 .update_curr = update_curr_dl,
2561 /* Used for dl_bw check and update, used under sched_rt_handler()::mutex */
2562 static u64 dl_generation;
2564 int sched_dl_global_validate(void)
2566 u64 runtime = global_rt_runtime();
2567 u64 period = global_rt_period();
2568 u64 new_bw = to_ratio(period, runtime);
2569 u64 gen = ++dl_generation;
2571 int cpu, cpus, ret = 0;
2572 unsigned long flags;
2575 * Here we want to check the bandwidth not being set to some
2576 * value smaller than the currently allocated bandwidth in
2577 * any of the root_domains.
2579 for_each_possible_cpu(cpu) {
2580 rcu_read_lock_sched();
2582 if (dl_bw_visited(cpu, gen))
2585 dl_b = dl_bw_of(cpu);
2586 cpus = dl_bw_cpus(cpu);
2588 raw_spin_lock_irqsave(&dl_b->lock, flags);
2589 if (new_bw * cpus < dl_b->total_bw)
2591 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2594 rcu_read_unlock_sched();
2603 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
2605 if (global_rt_runtime() == RUNTIME_INF) {
2606 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
2607 dl_rq->extra_bw = 1 << BW_SHIFT;
2609 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
2610 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
2611 dl_rq->extra_bw = to_ratio(global_rt_period(),
2612 global_rt_runtime());
2616 void sched_dl_do_global(void)
2619 u64 gen = ++dl_generation;
2622 unsigned long flags;
2624 def_dl_bandwidth.dl_period = global_rt_period();
2625 def_dl_bandwidth.dl_runtime = global_rt_runtime();
2627 if (global_rt_runtime() != RUNTIME_INF)
2628 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
2630 for_each_possible_cpu(cpu) {
2631 rcu_read_lock_sched();
2633 if (dl_bw_visited(cpu, gen)) {
2634 rcu_read_unlock_sched();
2638 dl_b = dl_bw_of(cpu);
2640 raw_spin_lock_irqsave(&dl_b->lock, flags);
2642 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2644 rcu_read_unlock_sched();
2645 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
2650 * We must be sure that accepting a new task (or allowing changing the
2651 * parameters of an existing one) is consistent with the bandwidth
2652 * constraints. If yes, this function also accordingly updates the currently
2653 * allocated bandwidth to reflect the new situation.
2655 * This function is called while holding p's rq->lock.
2657 int sched_dl_overflow(struct task_struct *p, int policy,
2658 const struct sched_attr *attr)
2660 u64 period = attr->sched_period ?: attr->sched_deadline;
2661 u64 runtime = attr->sched_runtime;
2662 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2663 int cpus, err = -1, cpu = task_cpu(p);
2664 struct dl_bw *dl_b = dl_bw_of(cpu);
2667 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2670 /* !deadline task may carry old deadline bandwidth */
2671 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2675 * Either if a task, enters, leave, or stays -deadline but changes
2676 * its parameters, we may need to update accordingly the total
2677 * allocated bandwidth of the container.
2679 raw_spin_lock(&dl_b->lock);
2680 cpus = dl_bw_cpus(cpu);
2681 cap = dl_bw_capacity(cpu);
2683 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2684 !__dl_overflow(dl_b, cap, 0, new_bw)) {
2685 if (hrtimer_active(&p->dl.inactive_timer))
2686 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2687 __dl_add(dl_b, new_bw, cpus);
2689 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2690 !__dl_overflow(dl_b, cap, p->dl.dl_bw, new_bw)) {
2692 * XXX this is slightly incorrect: when the task
2693 * utilization decreases, we should delay the total
2694 * utilization change until the task's 0-lag point.
2695 * But this would require to set the task's "inactive
2696 * timer" when the task is not inactive.
2698 __dl_sub(dl_b, p->dl.dl_bw, cpus);
2699 __dl_add(dl_b, new_bw, cpus);
2700 dl_change_utilization(p, new_bw);
2702 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2704 * Do not decrease the total deadline utilization here,
2705 * switched_from_dl() will take care to do it at the correct
2710 raw_spin_unlock(&dl_b->lock);
2716 * This function initializes the sched_dl_entity of a newly becoming
2717 * SCHED_DEADLINE task.
2719 * Only the static values are considered here, the actual runtime and the
2720 * absolute deadline will be properly calculated when the task is enqueued
2721 * for the first time with its new policy.
2723 void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
2725 struct sched_dl_entity *dl_se = &p->dl;
2727 dl_se->dl_runtime = attr->sched_runtime;
2728 dl_se->dl_deadline = attr->sched_deadline;
2729 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
2730 dl_se->flags = attr->sched_flags;
2731 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
2732 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
2735 void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
2737 struct sched_dl_entity *dl_se = &p->dl;
2739 attr->sched_priority = p->rt_priority;
2740 attr->sched_runtime = dl_se->dl_runtime;
2741 attr->sched_deadline = dl_se->dl_deadline;
2742 attr->sched_period = dl_se->dl_period;
2743 attr->sched_flags = dl_se->flags;
2747 * Default limits for DL period; on the top end we guard against small util
2748 * tasks still getting rediculous long effective runtimes, on the bottom end we
2749 * guard against timer DoS.
2751 unsigned int sysctl_sched_dl_period_max = 1 << 22; /* ~4 seconds */
2752 unsigned int sysctl_sched_dl_period_min = 100; /* 100 us */
2755 * This function validates the new parameters of a -deadline task.
2756 * We ask for the deadline not being zero, and greater or equal
2757 * than the runtime, as well as the period of being zero or
2758 * greater than deadline. Furthermore, we have to be sure that
2759 * user parameters are above the internal resolution of 1us (we
2760 * check sched_runtime only since it is always the smaller one) and
2761 * below 2^63 ns (we have to check both sched_deadline and
2762 * sched_period, as the latter can be zero).
2764 bool __checkparam_dl(const struct sched_attr *attr)
2766 u64 period, max, min;
2768 /* special dl tasks don't actually use any parameter */
2769 if (attr->sched_flags & SCHED_FLAG_SUGOV)
2773 if (attr->sched_deadline == 0)
2777 * Since we truncate DL_SCALE bits, make sure we're at least
2780 if (attr->sched_runtime < (1ULL << DL_SCALE))
2784 * Since we use the MSB for wrap-around and sign issues, make
2785 * sure it's not set (mind that period can be equal to zero).
2787 if (attr->sched_deadline & (1ULL << 63) ||
2788 attr->sched_period & (1ULL << 63))
2791 period = attr->sched_period;
2793 period = attr->sched_deadline;
2795 /* runtime <= deadline <= period (if period != 0) */
2796 if (period < attr->sched_deadline ||
2797 attr->sched_deadline < attr->sched_runtime)
2800 max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
2801 min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
2803 if (period < min || period > max)
2810 * This function clears the sched_dl_entity static params.
2812 void __dl_clear_params(struct task_struct *p)
2814 struct sched_dl_entity *dl_se = &p->dl;
2816 dl_se->dl_runtime = 0;
2817 dl_se->dl_deadline = 0;
2818 dl_se->dl_period = 0;
2821 dl_se->dl_density = 0;
2823 dl_se->dl_throttled = 0;
2824 dl_se->dl_yielded = 0;
2825 dl_se->dl_non_contending = 0;
2826 dl_se->dl_overrun = 0;
2828 #ifdef CONFIG_RT_MUTEXES
2829 dl_se->pi_se = dl_se;
2833 bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
2835 struct sched_dl_entity *dl_se = &p->dl;
2837 if (dl_se->dl_runtime != attr->sched_runtime ||
2838 dl_se->dl_deadline != attr->sched_deadline ||
2839 dl_se->dl_period != attr->sched_period ||
2840 dl_se->flags != attr->sched_flags)
2847 int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed)
2849 unsigned long flags, cap;
2850 unsigned int dest_cpu;
2855 dest_cpu = cpumask_any_and(cpu_active_mask, cs_cpus_allowed);
2857 rcu_read_lock_sched();
2858 dl_b = dl_bw_of(dest_cpu);
2859 raw_spin_lock_irqsave(&dl_b->lock, flags);
2860 cap = dl_bw_capacity(dest_cpu);
2861 overflow = __dl_overflow(dl_b, cap, 0, p->dl.dl_bw);
2866 * We reserve space for this task in the destination
2867 * root_domain, as we can't fail after this point.
2868 * We will free resources in the source root_domain
2869 * later on (see set_cpus_allowed_dl()).
2871 int cpus = dl_bw_cpus(dest_cpu);
2873 __dl_add(dl_b, p->dl.dl_bw, cpus);
2876 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2877 rcu_read_unlock_sched();
2882 int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
2883 const struct cpumask *trial)
2885 int ret = 1, trial_cpus;
2886 struct dl_bw *cur_dl_b;
2887 unsigned long flags;
2889 rcu_read_lock_sched();
2890 cur_dl_b = dl_bw_of(cpumask_any(cur));
2891 trial_cpus = cpumask_weight(trial);
2893 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
2894 if (cur_dl_b->bw != -1 &&
2895 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
2897 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
2898 rcu_read_unlock_sched();
2903 bool dl_cpu_busy(unsigned int cpu)
2905 unsigned long flags, cap;
2909 rcu_read_lock_sched();
2910 dl_b = dl_bw_of(cpu);
2911 raw_spin_lock_irqsave(&dl_b->lock, flags);
2912 cap = dl_bw_capacity(cpu);
2913 overflow = __dl_overflow(dl_b, cap, 0, 0);
2914 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
2915 rcu_read_unlock_sched();
2921 #ifdef CONFIG_SCHED_DEBUG
2922 void print_dl_stats(struct seq_file *m, int cpu)
2924 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
2926 #endif /* CONFIG_SCHED_DEBUG */