4 * Core kernel scheduler code and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
10 #include <linux/nospec.h>
12 #include <linux/kcov.h>
14 #include <asm/switch_to.h>
17 #include "../workqueue_internal.h"
18 #include "../smpboot.h"
22 #define CREATE_TRACE_POINTS
23 #include <trace/events/sched.h>
25 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
27 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
29 * Debugging: various feature bits
31 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
32 * sysctl_sched_features, defined in sched.h, to allow constants propagation
33 * at compile time and compiler optimization based on features default.
35 #define SCHED_FEAT(name, enabled) \
36 (1UL << __SCHED_FEAT_##name) * enabled |
37 const_debug unsigned int sysctl_sched_features =
44 * Number of tasks to iterate in a single balance run.
45 * Limited because this is done with IRQs disabled.
47 const_debug unsigned int sysctl_sched_nr_migrate = 32;
50 * period over which we measure -rt task CPU usage in us.
53 unsigned int sysctl_sched_rt_period = 1000000;
55 __read_mostly int scheduler_running;
58 * part of the period that we allow rt tasks to run in us.
61 int sysctl_sched_rt_runtime = 950000;
64 * __task_rq_lock - lock the rq @p resides on.
66 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
71 lockdep_assert_held(&p->pi_lock);
75 raw_spin_lock(&rq->lock);
76 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
80 raw_spin_unlock(&rq->lock);
82 while (unlikely(task_on_rq_migrating(p)))
88 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
90 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
91 __acquires(p->pi_lock)
97 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
99 raw_spin_lock(&rq->lock);
101 * move_queued_task() task_rq_lock()
104 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
105 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
106 * [S] ->cpu = new_cpu [L] task_rq()
110 * If we observe the old CPU in task_rq_lock, the acquire of
111 * the old rq->lock will fully serialize against the stores.
113 * If we observe the new CPU in task_rq_lock, the acquire will
114 * pair with the WMB to ensure we must then also see migrating.
116 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
120 raw_spin_unlock(&rq->lock);
121 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
123 while (unlikely(task_on_rq_migrating(p)))
129 * RQ-clock updating methods:
132 static void update_rq_clock_task(struct rq *rq, s64 delta)
135 * In theory, the compile should just see 0 here, and optimize out the call
136 * to sched_rt_avg_update. But I don't trust it...
138 s64 __maybe_unused steal = 0, irq_delta = 0;
140 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
141 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
144 * Since irq_time is only updated on {soft,}irq_exit, we might run into
145 * this case when a previous update_rq_clock() happened inside a
148 * When this happens, we stop ->clock_task and only update the
149 * prev_irq_time stamp to account for the part that fit, so that a next
150 * update will consume the rest. This ensures ->clock_task is
153 * It does however cause some slight miss-attribution of {soft,}irq
154 * time, a more accurate solution would be to update the irq_time using
155 * the current rq->clock timestamp, except that would require using
158 if (irq_delta > delta)
161 rq->prev_irq_time += irq_delta;
164 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
165 if (static_key_false((¶virt_steal_rq_enabled))) {
166 steal = paravirt_steal_clock(cpu_of(rq));
167 steal -= rq->prev_steal_time_rq;
169 if (unlikely(steal > delta))
172 rq->prev_steal_time_rq += steal;
177 rq->clock_task += delta;
179 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
180 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
181 update_irq_load_avg(rq, irq_delta + steal);
185 void update_rq_clock(struct rq *rq)
189 lockdep_assert_held(&rq->lock);
191 if (rq->clock_update_flags & RQCF_ACT_SKIP)
194 #ifdef CONFIG_SCHED_DEBUG
195 if (sched_feat(WARN_DOUBLE_CLOCK))
196 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
197 rq->clock_update_flags |= RQCF_UPDATED;
200 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
204 update_rq_clock_task(rq, delta);
208 #ifdef CONFIG_SCHED_HRTICK
210 * Use HR-timers to deliver accurate preemption points.
213 static void hrtick_clear(struct rq *rq)
215 if (hrtimer_active(&rq->hrtick_timer))
216 hrtimer_cancel(&rq->hrtick_timer);
220 * High-resolution timer tick.
221 * Runs from hardirq context with interrupts disabled.
223 static enum hrtimer_restart hrtick(struct hrtimer *timer)
225 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
228 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
232 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
235 return HRTIMER_NORESTART;
240 static void __hrtick_restart(struct rq *rq)
242 struct hrtimer *timer = &rq->hrtick_timer;
244 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
248 * called from hardirq (IPI) context
250 static void __hrtick_start(void *arg)
256 __hrtick_restart(rq);
257 rq->hrtick_csd_pending = 0;
262 * Called to set the hrtick timer state.
264 * called with rq->lock held and irqs disabled
266 void hrtick_start(struct rq *rq, u64 delay)
268 struct hrtimer *timer = &rq->hrtick_timer;
273 * Don't schedule slices shorter than 10000ns, that just
274 * doesn't make sense and can cause timer DoS.
276 delta = max_t(s64, delay, 10000LL);
277 time = ktime_add_ns(timer->base->get_time(), delta);
279 hrtimer_set_expires(timer, time);
281 if (rq == this_rq()) {
282 __hrtick_restart(rq);
283 } else if (!rq->hrtick_csd_pending) {
284 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
285 rq->hrtick_csd_pending = 1;
291 * Called to set the hrtick timer state.
293 * called with rq->lock held and irqs disabled
295 void hrtick_start(struct rq *rq, u64 delay)
298 * Don't schedule slices shorter than 10000ns, that just
299 * doesn't make sense. Rely on vruntime for fairness.
301 delay = max_t(u64, delay, 10000LL);
302 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
303 HRTIMER_MODE_REL_PINNED);
305 #endif /* CONFIG_SMP */
307 static void hrtick_rq_init(struct rq *rq)
310 rq->hrtick_csd_pending = 0;
312 rq->hrtick_csd.flags = 0;
313 rq->hrtick_csd.func = __hrtick_start;
314 rq->hrtick_csd.info = rq;
317 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
318 rq->hrtick_timer.function = hrtick;
320 #else /* CONFIG_SCHED_HRTICK */
321 static inline void hrtick_clear(struct rq *rq)
325 static inline void hrtick_rq_init(struct rq *rq)
328 #endif /* CONFIG_SCHED_HRTICK */
331 * cmpxchg based fetch_or, macro so it works for different integer types
333 #define fetch_or(ptr, mask) \
335 typeof(ptr) _ptr = (ptr); \
336 typeof(mask) _mask = (mask); \
337 typeof(*_ptr) _old, _val = *_ptr; \
340 _old = cmpxchg(_ptr, _val, _val | _mask); \
348 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
350 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
351 * this avoids any races wrt polling state changes and thereby avoids
354 static bool set_nr_and_not_polling(struct task_struct *p)
356 struct thread_info *ti = task_thread_info(p);
357 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
361 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
363 * If this returns true, then the idle task promises to call
364 * sched_ttwu_pending() and reschedule soon.
366 static bool set_nr_if_polling(struct task_struct *p)
368 struct thread_info *ti = task_thread_info(p);
369 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
372 if (!(val & _TIF_POLLING_NRFLAG))
374 if (val & _TIF_NEED_RESCHED)
376 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
385 static bool set_nr_and_not_polling(struct task_struct *p)
387 set_tsk_need_resched(p);
392 static bool set_nr_if_polling(struct task_struct *p)
399 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
401 struct wake_q_node *node = &task->wake_q;
404 * Atomically grab the task, if ->wake_q is !nil already it means
405 * its already queued (either by us or someone else) and will get the
406 * wakeup due to that.
408 * This cmpxchg() executes a full barrier, which pairs with the full
409 * barrier executed by the wakeup in wake_up_q().
411 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
414 get_task_struct(task);
417 * The head is context local, there can be no concurrency.
420 head->lastp = &node->next;
423 void wake_up_q(struct wake_q_head *head)
425 struct wake_q_node *node = head->first;
427 while (node != WAKE_Q_TAIL) {
428 struct task_struct *task;
430 task = container_of(node, struct task_struct, wake_q);
432 /* Task can safely be re-inserted now: */
434 task->wake_q.next = NULL;
437 * wake_up_process() executes a full barrier, which pairs with
438 * the queueing in wake_q_add() so as not to miss wakeups.
440 wake_up_process(task);
441 put_task_struct(task);
446 * resched_curr - mark rq's current task 'to be rescheduled now'.
448 * On UP this means the setting of the need_resched flag, on SMP it
449 * might also involve a cross-CPU call to trigger the scheduler on
452 void resched_curr(struct rq *rq)
454 struct task_struct *curr = rq->curr;
457 lockdep_assert_held(&rq->lock);
459 if (test_tsk_need_resched(curr))
464 if (cpu == smp_processor_id()) {
465 set_tsk_need_resched(curr);
466 set_preempt_need_resched();
470 if (set_nr_and_not_polling(curr))
471 smp_send_reschedule(cpu);
473 trace_sched_wake_idle_without_ipi(cpu);
476 void resched_cpu(int cpu)
478 struct rq *rq = cpu_rq(cpu);
481 raw_spin_lock_irqsave(&rq->lock, flags);
482 if (cpu_online(cpu) || cpu == smp_processor_id())
484 raw_spin_unlock_irqrestore(&rq->lock, flags);
488 #ifdef CONFIG_NO_HZ_COMMON
490 * In the semi idle case, use the nearest busy CPU for migrating timers
491 * from an idle CPU. This is good for power-savings.
493 * We don't do similar optimization for completely idle system, as
494 * selecting an idle CPU will add more delays to the timers than intended
495 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
497 int get_nohz_timer_target(void)
499 int i, cpu = smp_processor_id();
500 struct sched_domain *sd;
502 if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
506 for_each_domain(cpu, sd) {
507 for_each_cpu(i, sched_domain_span(sd)) {
511 if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
518 if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
519 cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
526 * When add_timer_on() enqueues a timer into the timer wheel of an
527 * idle CPU then this timer might expire before the next timer event
528 * which is scheduled to wake up that CPU. In case of a completely
529 * idle system the next event might even be infinite time into the
530 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
531 * leaves the inner idle loop so the newly added timer is taken into
532 * account when the CPU goes back to idle and evaluates the timer
533 * wheel for the next timer event.
535 static void wake_up_idle_cpu(int cpu)
537 struct rq *rq = cpu_rq(cpu);
539 if (cpu == smp_processor_id())
542 if (set_nr_and_not_polling(rq->idle))
543 smp_send_reschedule(cpu);
545 trace_sched_wake_idle_without_ipi(cpu);
548 static bool wake_up_full_nohz_cpu(int cpu)
551 * We just need the target to call irq_exit() and re-evaluate
552 * the next tick. The nohz full kick at least implies that.
553 * If needed we can still optimize that later with an
556 if (cpu_is_offline(cpu))
557 return true; /* Don't try to wake offline CPUs. */
558 if (tick_nohz_full_cpu(cpu)) {
559 if (cpu != smp_processor_id() ||
560 tick_nohz_tick_stopped())
561 tick_nohz_full_kick_cpu(cpu);
569 * Wake up the specified CPU. If the CPU is going offline, it is the
570 * caller's responsibility to deal with the lost wakeup, for example,
571 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
573 void wake_up_nohz_cpu(int cpu)
575 if (!wake_up_full_nohz_cpu(cpu))
576 wake_up_idle_cpu(cpu);
579 static inline bool got_nohz_idle_kick(void)
581 int cpu = smp_processor_id();
583 if (!(atomic_read(nohz_flags(cpu)) & NOHZ_KICK_MASK))
586 if (idle_cpu(cpu) && !need_resched())
590 * We can't run Idle Load Balance on this CPU for this time so we
591 * cancel it and clear NOHZ_BALANCE_KICK
593 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
597 #else /* CONFIG_NO_HZ_COMMON */
599 static inline bool got_nohz_idle_kick(void)
604 #endif /* CONFIG_NO_HZ_COMMON */
606 #ifdef CONFIG_NO_HZ_FULL
607 bool sched_can_stop_tick(struct rq *rq)
611 /* Deadline tasks, even if single, need the tick */
612 if (rq->dl.dl_nr_running)
616 * If there are more than one RR tasks, we need the tick to effect the
617 * actual RR behaviour.
619 if (rq->rt.rr_nr_running) {
620 if (rq->rt.rr_nr_running == 1)
627 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
628 * forced preemption between FIFO tasks.
630 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
635 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
636 * if there's more than one we need the tick for involuntary
639 if (rq->nr_running > 1)
644 #endif /* CONFIG_NO_HZ_FULL */
645 #endif /* CONFIG_SMP */
647 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
648 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
650 * Iterate task_group tree rooted at *from, calling @down when first entering a
651 * node and @up when leaving it for the final time.
653 * Caller must hold rcu_lock or sufficient equivalent.
655 int walk_tg_tree_from(struct task_group *from,
656 tg_visitor down, tg_visitor up, void *data)
658 struct task_group *parent, *child;
664 ret = (*down)(parent, data);
667 list_for_each_entry_rcu(child, &parent->children, siblings) {
674 ret = (*up)(parent, data);
675 if (ret || parent == from)
679 parent = parent->parent;
686 int tg_nop(struct task_group *tg, void *data)
692 static void set_load_weight(struct task_struct *p, bool update_load)
694 int prio = p->static_prio - MAX_RT_PRIO;
695 struct load_weight *load = &p->se.load;
698 * SCHED_IDLE tasks get minimal weight:
700 if (idle_policy(p->policy)) {
701 load->weight = scale_load(WEIGHT_IDLEPRIO);
702 load->inv_weight = WMULT_IDLEPRIO;
703 p->se.runnable_weight = load->weight;
708 * SCHED_OTHER tasks have to update their load when changing their
711 if (update_load && p->sched_class == &fair_sched_class) {
712 reweight_task(p, prio);
714 load->weight = scale_load(sched_prio_to_weight[prio]);
715 load->inv_weight = sched_prio_to_wmult[prio];
716 p->se.runnable_weight = load->weight;
720 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
722 if (!(flags & ENQUEUE_NOCLOCK))
725 if (!(flags & ENQUEUE_RESTORE))
726 sched_info_queued(rq, p);
728 p->sched_class->enqueue_task(rq, p, flags);
731 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
733 if (!(flags & DEQUEUE_NOCLOCK))
736 if (!(flags & DEQUEUE_SAVE))
737 sched_info_dequeued(rq, p);
739 p->sched_class->dequeue_task(rq, p, flags);
742 void activate_task(struct rq *rq, struct task_struct *p, int flags)
744 if (task_contributes_to_load(p))
745 rq->nr_uninterruptible--;
747 enqueue_task(rq, p, flags);
750 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
752 if (task_contributes_to_load(p))
753 rq->nr_uninterruptible++;
755 dequeue_task(rq, p, flags);
759 * __normal_prio - return the priority that is based on the static prio
761 static inline int __normal_prio(struct task_struct *p)
763 return p->static_prio;
767 * Calculate the expected normal priority: i.e. priority
768 * without taking RT-inheritance into account. Might be
769 * boosted by interactivity modifiers. Changes upon fork,
770 * setprio syscalls, and whenever the interactivity
771 * estimator recalculates.
773 static inline int normal_prio(struct task_struct *p)
777 if (task_has_dl_policy(p))
778 prio = MAX_DL_PRIO-1;
779 else if (task_has_rt_policy(p))
780 prio = MAX_RT_PRIO-1 - p->rt_priority;
782 prio = __normal_prio(p);
787 * Calculate the current priority, i.e. the priority
788 * taken into account by the scheduler. This value might
789 * be boosted by RT tasks, or might be boosted by
790 * interactivity modifiers. Will be RT if the task got
791 * RT-boosted. If not then it returns p->normal_prio.
793 static int effective_prio(struct task_struct *p)
795 p->normal_prio = normal_prio(p);
797 * If we are RT tasks or we were boosted to RT priority,
798 * keep the priority unchanged. Otherwise, update priority
799 * to the normal priority:
801 if (!rt_prio(p->prio))
802 return p->normal_prio;
807 * task_curr - is this task currently executing on a CPU?
808 * @p: the task in question.
810 * Return: 1 if the task is currently executing. 0 otherwise.
812 inline int task_curr(const struct task_struct *p)
814 return cpu_curr(task_cpu(p)) == p;
818 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
819 * use the balance_callback list if you want balancing.
821 * this means any call to check_class_changed() must be followed by a call to
822 * balance_callback().
824 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
825 const struct sched_class *prev_class,
828 if (prev_class != p->sched_class) {
829 if (prev_class->switched_from)
830 prev_class->switched_from(rq, p);
832 p->sched_class->switched_to(rq, p);
833 } else if (oldprio != p->prio || dl_task(p))
834 p->sched_class->prio_changed(rq, p, oldprio);
837 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
839 const struct sched_class *class;
841 if (p->sched_class == rq->curr->sched_class) {
842 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
844 for_each_class(class) {
845 if (class == rq->curr->sched_class)
847 if (class == p->sched_class) {
855 * A queue event has occurred, and we're going to schedule. In
856 * this case, we can save a useless back to back clock update.
858 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
859 rq_clock_skip_update(rq);
864 static inline bool is_per_cpu_kthread(struct task_struct *p)
866 if (!(p->flags & PF_KTHREAD))
869 if (p->nr_cpus_allowed != 1)
876 * Per-CPU kthreads are allowed to run on !actie && online CPUs, see
877 * __set_cpus_allowed_ptr() and select_fallback_rq().
879 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
881 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
884 if (is_per_cpu_kthread(p))
885 return cpu_online(cpu);
887 return cpu_active(cpu);
891 * This is how migration works:
893 * 1) we invoke migration_cpu_stop() on the target CPU using
895 * 2) stopper starts to run (implicitly forcing the migrated thread
897 * 3) it checks whether the migrated task is still in the wrong runqueue.
898 * 4) if it's in the wrong runqueue then the migration thread removes
899 * it and puts it into the right queue.
900 * 5) stopper completes and stop_one_cpu() returns and the migration
905 * move_queued_task - move a queued task to new rq.
907 * Returns (locked) new rq. Old rq's lock is released.
909 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
910 struct task_struct *p, int new_cpu)
912 lockdep_assert_held(&rq->lock);
914 p->on_rq = TASK_ON_RQ_MIGRATING;
915 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
916 set_task_cpu(p, new_cpu);
919 rq = cpu_rq(new_cpu);
922 BUG_ON(task_cpu(p) != new_cpu);
923 enqueue_task(rq, p, 0);
924 p->on_rq = TASK_ON_RQ_QUEUED;
925 check_preempt_curr(rq, p, 0);
930 struct migration_arg {
931 struct task_struct *task;
936 * Move (not current) task off this CPU, onto the destination CPU. We're doing
937 * this because either it can't run here any more (set_cpus_allowed()
938 * away from this CPU, or CPU going down), or because we're
939 * attempting to rebalance this task on exec (sched_exec).
941 * So we race with normal scheduler movements, but that's OK, as long
942 * as the task is no longer on this CPU.
944 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
945 struct task_struct *p, int dest_cpu)
947 /* Affinity changed (again). */
948 if (!is_cpu_allowed(p, dest_cpu))
952 rq = move_queued_task(rq, rf, p, dest_cpu);
958 * migration_cpu_stop - this will be executed by a highprio stopper thread
959 * and performs thread migration by bumping thread off CPU then
960 * 'pushing' onto another runqueue.
962 static int migration_cpu_stop(void *data)
964 struct migration_arg *arg = data;
965 struct task_struct *p = arg->task;
966 struct rq *rq = this_rq();
970 * The original target CPU might have gone down and we might
971 * be on another CPU but it doesn't matter.
975 * We need to explicitly wake pending tasks before running
976 * __migrate_task() such that we will not miss enforcing cpus_allowed
977 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
979 sched_ttwu_pending();
981 raw_spin_lock(&p->pi_lock);
984 * If task_rq(p) != rq, it cannot be migrated here, because we're
985 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
986 * we're holding p->pi_lock.
988 if (task_rq(p) == rq) {
989 if (task_on_rq_queued(p))
990 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
992 p->wake_cpu = arg->dest_cpu;
995 raw_spin_unlock(&p->pi_lock);
1002 * sched_class::set_cpus_allowed must do the below, but is not required to
1003 * actually call this function.
1005 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1007 cpumask_copy(&p->cpus_allowed, new_mask);
1008 p->nr_cpus_allowed = cpumask_weight(new_mask);
1011 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1013 struct rq *rq = task_rq(p);
1014 bool queued, running;
1016 lockdep_assert_held(&p->pi_lock);
1018 queued = task_on_rq_queued(p);
1019 running = task_current(rq, p);
1023 * Because __kthread_bind() calls this on blocked tasks without
1026 lockdep_assert_held(&rq->lock);
1027 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1030 put_prev_task(rq, p);
1032 p->sched_class->set_cpus_allowed(p, new_mask);
1035 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1037 set_curr_task(rq, p);
1041 * Change a given task's CPU affinity. Migrate the thread to a
1042 * proper CPU and schedule it away if the CPU it's executing on
1043 * is removed from the allowed bitmask.
1045 * NOTE: the caller must have a valid reference to the task, the
1046 * task must not exit() & deallocate itself prematurely. The
1047 * call is not atomic; no spinlocks may be held.
1049 static int __set_cpus_allowed_ptr(struct task_struct *p,
1050 const struct cpumask *new_mask, bool check)
1052 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1053 unsigned int dest_cpu;
1058 rq = task_rq_lock(p, &rf);
1059 update_rq_clock(rq);
1061 if (p->flags & PF_KTHREAD) {
1063 * Kernel threads are allowed on online && !active CPUs
1065 cpu_valid_mask = cpu_online_mask;
1069 * Must re-check here, to close a race against __kthread_bind(),
1070 * sched_setaffinity() is not guaranteed to observe the flag.
1072 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1077 if (cpumask_equal(&p->cpus_allowed, new_mask))
1080 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1085 do_set_cpus_allowed(p, new_mask);
1087 if (p->flags & PF_KTHREAD) {
1089 * For kernel threads that do indeed end up on online &&
1090 * !active we want to ensure they are strict per-CPU threads.
1092 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1093 !cpumask_intersects(new_mask, cpu_active_mask) &&
1094 p->nr_cpus_allowed != 1);
1097 /* Can the task run on the task's current CPU? If so, we're done */
1098 if (cpumask_test_cpu(task_cpu(p), new_mask))
1101 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1102 if (task_running(rq, p) || p->state == TASK_WAKING) {
1103 struct migration_arg arg = { p, dest_cpu };
1104 /* Need help from migration thread: drop lock and wait. */
1105 task_rq_unlock(rq, p, &rf);
1106 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1107 tlb_migrate_finish(p->mm);
1109 } else if (task_on_rq_queued(p)) {
1111 * OK, since we're going to drop the lock immediately
1112 * afterwards anyway.
1114 rq = move_queued_task(rq, &rf, p, dest_cpu);
1117 task_rq_unlock(rq, p, &rf);
1122 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1124 return __set_cpus_allowed_ptr(p, new_mask, false);
1126 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1128 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1130 #ifdef CONFIG_SCHED_DEBUG
1132 * We should never call set_task_cpu() on a blocked task,
1133 * ttwu() will sort out the placement.
1135 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1139 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1140 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1141 * time relying on p->on_rq.
1143 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1144 p->sched_class == &fair_sched_class &&
1145 (p->on_rq && !task_on_rq_migrating(p)));
1147 #ifdef CONFIG_LOCKDEP
1149 * The caller should hold either p->pi_lock or rq->lock, when changing
1150 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1152 * sched_move_task() holds both and thus holding either pins the cgroup,
1155 * Furthermore, all task_rq users should acquire both locks, see
1158 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1159 lockdep_is_held(&task_rq(p)->lock)));
1162 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1164 WARN_ON_ONCE(!cpu_online(new_cpu));
1167 trace_sched_migrate_task(p, new_cpu);
1169 if (task_cpu(p) != new_cpu) {
1170 if (p->sched_class->migrate_task_rq)
1171 p->sched_class->migrate_task_rq(p, new_cpu);
1172 p->se.nr_migrations++;
1174 perf_event_task_migrate(p);
1177 __set_task_cpu(p, new_cpu);
1180 #ifdef CONFIG_NUMA_BALANCING
1181 static void __migrate_swap_task(struct task_struct *p, int cpu)
1183 if (task_on_rq_queued(p)) {
1184 struct rq *src_rq, *dst_rq;
1185 struct rq_flags srf, drf;
1187 src_rq = task_rq(p);
1188 dst_rq = cpu_rq(cpu);
1190 rq_pin_lock(src_rq, &srf);
1191 rq_pin_lock(dst_rq, &drf);
1193 p->on_rq = TASK_ON_RQ_MIGRATING;
1194 deactivate_task(src_rq, p, 0);
1195 set_task_cpu(p, cpu);
1196 activate_task(dst_rq, p, 0);
1197 p->on_rq = TASK_ON_RQ_QUEUED;
1198 check_preempt_curr(dst_rq, p, 0);
1200 rq_unpin_lock(dst_rq, &drf);
1201 rq_unpin_lock(src_rq, &srf);
1205 * Task isn't running anymore; make it appear like we migrated
1206 * it before it went to sleep. This means on wakeup we make the
1207 * previous CPU our target instead of where it really is.
1213 struct migration_swap_arg {
1214 struct task_struct *src_task, *dst_task;
1215 int src_cpu, dst_cpu;
1218 static int migrate_swap_stop(void *data)
1220 struct migration_swap_arg *arg = data;
1221 struct rq *src_rq, *dst_rq;
1224 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1227 src_rq = cpu_rq(arg->src_cpu);
1228 dst_rq = cpu_rq(arg->dst_cpu);
1230 double_raw_lock(&arg->src_task->pi_lock,
1231 &arg->dst_task->pi_lock);
1232 double_rq_lock(src_rq, dst_rq);
1234 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1237 if (task_cpu(arg->src_task) != arg->src_cpu)
1240 if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1243 if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1246 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1247 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1252 double_rq_unlock(src_rq, dst_rq);
1253 raw_spin_unlock(&arg->dst_task->pi_lock);
1254 raw_spin_unlock(&arg->src_task->pi_lock);
1260 * Cross migrate two tasks
1262 int migrate_swap(struct task_struct *cur, struct task_struct *p,
1263 int target_cpu, int curr_cpu)
1265 struct migration_swap_arg arg;
1268 arg = (struct migration_swap_arg){
1270 .src_cpu = curr_cpu,
1272 .dst_cpu = target_cpu,
1275 if (arg.src_cpu == arg.dst_cpu)
1279 * These three tests are all lockless; this is OK since all of them
1280 * will be re-checked with proper locks held further down the line.
1282 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1285 if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1288 if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1291 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1292 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1297 #endif /* CONFIG_NUMA_BALANCING */
1300 * wait_task_inactive - wait for a thread to unschedule.
1302 * If @match_state is nonzero, it's the @p->state value just checked and
1303 * not expected to change. If it changes, i.e. @p might have woken up,
1304 * then return zero. When we succeed in waiting for @p to be off its CPU,
1305 * we return a positive number (its total switch count). If a second call
1306 * a short while later returns the same number, the caller can be sure that
1307 * @p has remained unscheduled the whole time.
1309 * The caller must ensure that the task *will* unschedule sometime soon,
1310 * else this function might spin for a *long* time. This function can't
1311 * be called with interrupts off, or it may introduce deadlock with
1312 * smp_call_function() if an IPI is sent by the same process we are
1313 * waiting to become inactive.
1315 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1317 int running, queued;
1324 * We do the initial early heuristics without holding
1325 * any task-queue locks at all. We'll only try to get
1326 * the runqueue lock when things look like they will
1332 * If the task is actively running on another CPU
1333 * still, just relax and busy-wait without holding
1336 * NOTE! Since we don't hold any locks, it's not
1337 * even sure that "rq" stays as the right runqueue!
1338 * But we don't care, since "task_running()" will
1339 * return false if the runqueue has changed and p
1340 * is actually now running somewhere else!
1342 while (task_running(rq, p)) {
1343 if (match_state && unlikely(p->state != match_state))
1349 * Ok, time to look more closely! We need the rq
1350 * lock now, to be *sure*. If we're wrong, we'll
1351 * just go back and repeat.
1353 rq = task_rq_lock(p, &rf);
1354 trace_sched_wait_task(p);
1355 running = task_running(rq, p);
1356 queued = task_on_rq_queued(p);
1358 if (!match_state || p->state == match_state)
1359 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1360 task_rq_unlock(rq, p, &rf);
1363 * If it changed from the expected state, bail out now.
1365 if (unlikely(!ncsw))
1369 * Was it really running after all now that we
1370 * checked with the proper locks actually held?
1372 * Oops. Go back and try again..
1374 if (unlikely(running)) {
1380 * It's not enough that it's not actively running,
1381 * it must be off the runqueue _entirely_, and not
1384 * So if it was still runnable (but just not actively
1385 * running right now), it's preempted, and we should
1386 * yield - it could be a while.
1388 if (unlikely(queued)) {
1389 ktime_t to = NSEC_PER_SEC / HZ;
1391 set_current_state(TASK_UNINTERRUPTIBLE);
1392 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1397 * Ahh, all good. It wasn't running, and it wasn't
1398 * runnable, which means that it will never become
1399 * running in the future either. We're all done!
1408 * kick_process - kick a running thread to enter/exit the kernel
1409 * @p: the to-be-kicked thread
1411 * Cause a process which is running on another CPU to enter
1412 * kernel-mode, without any delay. (to get signals handled.)
1414 * NOTE: this function doesn't have to take the runqueue lock,
1415 * because all it wants to ensure is that the remote task enters
1416 * the kernel. If the IPI races and the task has been migrated
1417 * to another CPU then no harm is done and the purpose has been
1420 void kick_process(struct task_struct *p)
1426 if ((cpu != smp_processor_id()) && task_curr(p))
1427 smp_send_reschedule(cpu);
1430 EXPORT_SYMBOL_GPL(kick_process);
1433 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1435 * A few notes on cpu_active vs cpu_online:
1437 * - cpu_active must be a subset of cpu_online
1439 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
1440 * see __set_cpus_allowed_ptr(). At this point the newly online
1441 * CPU isn't yet part of the sched domains, and balancing will not
1444 * - on CPU-down we clear cpu_active() to mask the sched domains and
1445 * avoid the load balancer to place new tasks on the to be removed
1446 * CPU. Existing tasks will remain running there and will be taken
1449 * This means that fallback selection must not select !active CPUs.
1450 * And can assume that any active CPU must be online. Conversely
1451 * select_task_rq() below may allow selection of !active CPUs in order
1452 * to satisfy the above rules.
1454 static int select_fallback_rq(int cpu, struct task_struct *p)
1456 int nid = cpu_to_node(cpu);
1457 const struct cpumask *nodemask = NULL;
1458 enum { cpuset, possible, fail } state = cpuset;
1462 * If the node that the CPU is on has been offlined, cpu_to_node()
1463 * will return -1. There is no CPU on the node, and we should
1464 * select the CPU on the other node.
1467 nodemask = cpumask_of_node(nid);
1469 /* Look for allowed, online CPU in same node. */
1470 for_each_cpu(dest_cpu, nodemask) {
1471 if (!cpu_active(dest_cpu))
1473 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1479 /* Any allowed, online CPU? */
1480 for_each_cpu(dest_cpu, &p->cpus_allowed) {
1481 if (!is_cpu_allowed(p, dest_cpu))
1487 /* No more Mr. Nice Guy. */
1490 if (IS_ENABLED(CONFIG_CPUSETS)) {
1491 cpuset_cpus_allowed_fallback(p);
1497 do_set_cpus_allowed(p, cpu_possible_mask);
1508 if (state != cpuset) {
1510 * Don't tell them about moving exiting tasks or
1511 * kernel threads (both mm NULL), since they never
1514 if (p->mm && printk_ratelimit()) {
1515 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1516 task_pid_nr(p), p->comm, cpu);
1524 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1527 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1529 lockdep_assert_held(&p->pi_lock);
1531 if (p->nr_cpus_allowed > 1)
1532 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1534 cpu = cpumask_any(&p->cpus_allowed);
1537 * In order not to call set_task_cpu() on a blocking task we need
1538 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1541 * Since this is common to all placement strategies, this lives here.
1543 * [ this allows ->select_task() to simply return task_cpu(p) and
1544 * not worry about this generic constraint ]
1546 if (unlikely(!is_cpu_allowed(p, cpu)))
1547 cpu = select_fallback_rq(task_cpu(p), p);
1552 static void update_avg(u64 *avg, u64 sample)
1554 s64 diff = sample - *avg;
1558 void sched_set_stop_task(int cpu, struct task_struct *stop)
1560 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1561 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1565 * Make it appear like a SCHED_FIFO task, its something
1566 * userspace knows about and won't get confused about.
1568 * Also, it will make PI more or less work without too
1569 * much confusion -- but then, stop work should not
1570 * rely on PI working anyway.
1572 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1574 stop->sched_class = &stop_sched_class;
1577 cpu_rq(cpu)->stop = stop;
1581 * Reset it back to a normal scheduling class so that
1582 * it can die in pieces.
1584 old_stop->sched_class = &rt_sched_class;
1590 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1591 const struct cpumask *new_mask, bool check)
1593 return set_cpus_allowed_ptr(p, new_mask);
1596 #endif /* CONFIG_SMP */
1599 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1603 if (!schedstat_enabled())
1609 if (cpu == rq->cpu) {
1610 __schedstat_inc(rq->ttwu_local);
1611 __schedstat_inc(p->se.statistics.nr_wakeups_local);
1613 struct sched_domain *sd;
1615 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
1617 for_each_domain(rq->cpu, sd) {
1618 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1619 __schedstat_inc(sd->ttwu_wake_remote);
1626 if (wake_flags & WF_MIGRATED)
1627 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1628 #endif /* CONFIG_SMP */
1630 __schedstat_inc(rq->ttwu_count);
1631 __schedstat_inc(p->se.statistics.nr_wakeups);
1633 if (wake_flags & WF_SYNC)
1634 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
1637 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1639 activate_task(rq, p, en_flags);
1640 p->on_rq = TASK_ON_RQ_QUEUED;
1642 /* If a worker is waking up, notify the workqueue: */
1643 if (p->flags & PF_WQ_WORKER)
1644 wq_worker_waking_up(p, cpu_of(rq));
1648 * Mark the task runnable and perform wakeup-preemption.
1650 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1651 struct rq_flags *rf)
1653 check_preempt_curr(rq, p, wake_flags);
1654 p->state = TASK_RUNNING;
1655 trace_sched_wakeup(p);
1658 if (p->sched_class->task_woken) {
1660 * Our task @p is fully woken up and running; so its safe to
1661 * drop the rq->lock, hereafter rq is only used for statistics.
1663 rq_unpin_lock(rq, rf);
1664 p->sched_class->task_woken(rq, p);
1665 rq_repin_lock(rq, rf);
1668 if (rq->idle_stamp) {
1669 u64 delta = rq_clock(rq) - rq->idle_stamp;
1670 u64 max = 2*rq->max_idle_balance_cost;
1672 update_avg(&rq->avg_idle, delta);
1674 if (rq->avg_idle > max)
1683 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1684 struct rq_flags *rf)
1686 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
1688 lockdep_assert_held(&rq->lock);
1691 if (p->sched_contributes_to_load)
1692 rq->nr_uninterruptible--;
1694 if (wake_flags & WF_MIGRATED)
1695 en_flags |= ENQUEUE_MIGRATED;
1698 ttwu_activate(rq, p, en_flags);
1699 ttwu_do_wakeup(rq, p, wake_flags, rf);
1703 * Called in case the task @p isn't fully descheduled from its runqueue,
1704 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1705 * since all we need to do is flip p->state to TASK_RUNNING, since
1706 * the task is still ->on_rq.
1708 static int ttwu_remote(struct task_struct *p, int wake_flags)
1714 rq = __task_rq_lock(p, &rf);
1715 if (task_on_rq_queued(p)) {
1716 /* check_preempt_curr() may use rq clock */
1717 update_rq_clock(rq);
1718 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1721 __task_rq_unlock(rq, &rf);
1727 void sched_ttwu_pending(void)
1729 struct rq *rq = this_rq();
1730 struct llist_node *llist = llist_del_all(&rq->wake_list);
1731 struct task_struct *p, *t;
1737 rq_lock_irqsave(rq, &rf);
1738 update_rq_clock(rq);
1740 llist_for_each_entry_safe(p, t, llist, wake_entry)
1741 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
1743 rq_unlock_irqrestore(rq, &rf);
1746 void scheduler_ipi(void)
1749 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1750 * TIF_NEED_RESCHED remotely (for the first time) will also send
1753 preempt_fold_need_resched();
1755 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1759 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1760 * traditionally all their work was done from the interrupt return
1761 * path. Now that we actually do some work, we need to make sure
1764 * Some archs already do call them, luckily irq_enter/exit nest
1767 * Arguably we should visit all archs and update all handlers,
1768 * however a fair share of IPIs are still resched only so this would
1769 * somewhat pessimize the simple resched case.
1772 sched_ttwu_pending();
1775 * Check if someone kicked us for doing the nohz idle load balance.
1777 if (unlikely(got_nohz_idle_kick())) {
1778 this_rq()->idle_balance = 1;
1779 raise_softirq_irqoff(SCHED_SOFTIRQ);
1784 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1786 struct rq *rq = cpu_rq(cpu);
1788 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1790 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1791 if (!set_nr_if_polling(rq->idle))
1792 smp_send_reschedule(cpu);
1794 trace_sched_wake_idle_without_ipi(cpu);
1798 void wake_up_if_idle(int cpu)
1800 struct rq *rq = cpu_rq(cpu);
1805 if (!is_idle_task(rcu_dereference(rq->curr)))
1808 if (set_nr_if_polling(rq->idle)) {
1809 trace_sched_wake_idle_without_ipi(cpu);
1811 rq_lock_irqsave(rq, &rf);
1812 if (is_idle_task(rq->curr))
1813 smp_send_reschedule(cpu);
1814 /* Else CPU is not idle, do nothing here: */
1815 rq_unlock_irqrestore(rq, &rf);
1822 bool cpus_share_cache(int this_cpu, int that_cpu)
1824 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1826 #endif /* CONFIG_SMP */
1828 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1830 struct rq *rq = cpu_rq(cpu);
1833 #if defined(CONFIG_SMP)
1834 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1835 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1836 ttwu_queue_remote(p, cpu, wake_flags);
1842 update_rq_clock(rq);
1843 ttwu_do_activate(rq, p, wake_flags, &rf);
1848 * Notes on Program-Order guarantees on SMP systems.
1852 * The basic program-order guarantee on SMP systems is that when a task [t]
1853 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1854 * execution on its new CPU [c1].
1856 * For migration (of runnable tasks) this is provided by the following means:
1858 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1859 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1860 * rq(c1)->lock (if not at the same time, then in that order).
1861 * C) LOCK of the rq(c1)->lock scheduling in task
1863 * Release/acquire chaining guarantees that B happens after A and C after B.
1864 * Note: the CPU doing B need not be c0 or c1
1873 * UNLOCK rq(0)->lock
1875 * LOCK rq(0)->lock // orders against CPU0
1877 * UNLOCK rq(0)->lock
1881 * UNLOCK rq(1)->lock
1883 * LOCK rq(1)->lock // orders against CPU2
1886 * UNLOCK rq(1)->lock
1889 * BLOCKING -- aka. SLEEP + WAKEUP
1891 * For blocking we (obviously) need to provide the same guarantee as for
1892 * migration. However the means are completely different as there is no lock
1893 * chain to provide order. Instead we do:
1895 * 1) smp_store_release(X->on_cpu, 0)
1896 * 2) smp_cond_load_acquire(!X->on_cpu)
1900 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1902 * LOCK rq(0)->lock LOCK X->pi_lock
1905 * smp_store_release(X->on_cpu, 0);
1907 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1913 * X->state = RUNNING
1914 * UNLOCK rq(2)->lock
1916 * LOCK rq(2)->lock // orders against CPU1
1919 * UNLOCK rq(2)->lock
1922 * UNLOCK rq(0)->lock
1925 * However, for wakeups there is a second guarantee we must provide, namely we
1926 * must ensure that CONDITION=1 done by the caller can not be reordered with
1927 * accesses to the task state; see try_to_wake_up() and set_current_state().
1931 * try_to_wake_up - wake up a thread
1932 * @p: the thread to be awakened
1933 * @state: the mask of task states that can be woken
1934 * @wake_flags: wake modifier flags (WF_*)
1936 * If (@state & @p->state) @p->state = TASK_RUNNING.
1938 * If the task was not queued/runnable, also place it back on a runqueue.
1940 * Atomic against schedule() which would dequeue a task, also see
1941 * set_current_state().
1943 * This function executes a full memory barrier before accessing the task
1944 * state; see set_current_state().
1946 * Return: %true if @p->state changes (an actual wakeup was done),
1950 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1952 unsigned long flags;
1953 int cpu, success = 0;
1956 * If we are going to wake up a thread waiting for CONDITION we
1957 * need to ensure that CONDITION=1 done by the caller can not be
1958 * reordered with p->state check below. This pairs with mb() in
1959 * set_current_state() the waiting thread does.
1961 raw_spin_lock_irqsave(&p->pi_lock, flags);
1962 smp_mb__after_spinlock();
1963 if (!(p->state & state))
1966 trace_sched_waking(p);
1968 /* We're going to change ->state: */
1973 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1974 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1975 * in smp_cond_load_acquire() below.
1977 * sched_ttwu_pending() try_to_wake_up()
1978 * STORE p->on_rq = 1 LOAD p->state
1981 * __schedule() (switch to task 'p')
1982 * LOCK rq->lock smp_rmb();
1983 * smp_mb__after_spinlock();
1987 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
1989 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
1990 * __schedule(). See the comment for smp_mb__after_spinlock().
1993 if (p->on_rq && ttwu_remote(p, wake_flags))
1998 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1999 * possible to, falsely, observe p->on_cpu == 0.
2001 * One must be running (->on_cpu == 1) in order to remove oneself
2002 * from the runqueue.
2004 * __schedule() (switch to task 'p') try_to_wake_up()
2005 * STORE p->on_cpu = 1 LOAD p->on_rq
2008 * __schedule() (put 'p' to sleep)
2009 * LOCK rq->lock smp_rmb();
2010 * smp_mb__after_spinlock();
2011 * STORE p->on_rq = 0 LOAD p->on_cpu
2013 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2014 * __schedule(). See the comment for smp_mb__after_spinlock().
2019 * If the owning (remote) CPU is still in the middle of schedule() with
2020 * this task as prev, wait until its done referencing the task.
2022 * Pairs with the smp_store_release() in finish_task().
2024 * This ensures that tasks getting woken will be fully ordered against
2025 * their previous state and preserve Program Order.
2027 smp_cond_load_acquire(&p->on_cpu, !VAL);
2029 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2030 p->state = TASK_WAKING;
2033 delayacct_blkio_end(p);
2034 atomic_dec(&task_rq(p)->nr_iowait);
2037 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2038 if (task_cpu(p) != cpu) {
2039 wake_flags |= WF_MIGRATED;
2040 set_task_cpu(p, cpu);
2043 #else /* CONFIG_SMP */
2046 delayacct_blkio_end(p);
2047 atomic_dec(&task_rq(p)->nr_iowait);
2050 #endif /* CONFIG_SMP */
2052 ttwu_queue(p, cpu, wake_flags);
2054 ttwu_stat(p, cpu, wake_flags);
2056 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2062 * try_to_wake_up_local - try to wake up a local task with rq lock held
2063 * @p: the thread to be awakened
2064 * @rf: request-queue flags for pinning
2066 * Put @p on the run-queue if it's not already there. The caller must
2067 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2070 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2072 struct rq *rq = task_rq(p);
2074 if (WARN_ON_ONCE(rq != this_rq()) ||
2075 WARN_ON_ONCE(p == current))
2078 lockdep_assert_held(&rq->lock);
2080 if (!raw_spin_trylock(&p->pi_lock)) {
2082 * This is OK, because current is on_cpu, which avoids it being
2083 * picked for load-balance and preemption/IRQs are still
2084 * disabled avoiding further scheduler activity on it and we've
2085 * not yet picked a replacement task.
2088 raw_spin_lock(&p->pi_lock);
2092 if (!(p->state & TASK_NORMAL))
2095 trace_sched_waking(p);
2097 if (!task_on_rq_queued(p)) {
2099 delayacct_blkio_end(p);
2100 atomic_dec(&rq->nr_iowait);
2102 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK);
2105 ttwu_do_wakeup(rq, p, 0, rf);
2106 ttwu_stat(p, smp_processor_id(), 0);
2108 raw_spin_unlock(&p->pi_lock);
2112 * wake_up_process - Wake up a specific process
2113 * @p: The process to be woken up.
2115 * Attempt to wake up the nominated process and move it to the set of runnable
2118 * Return: 1 if the process was woken up, 0 if it was already running.
2120 * This function executes a full memory barrier before accessing the task state.
2122 int wake_up_process(struct task_struct *p)
2124 return try_to_wake_up(p, TASK_NORMAL, 0);
2126 EXPORT_SYMBOL(wake_up_process);
2128 int wake_up_state(struct task_struct *p, unsigned int state)
2130 return try_to_wake_up(p, state, 0);
2134 * Perform scheduler related setup for a newly forked process p.
2135 * p is forked by current.
2137 * __sched_fork() is basic setup used by init_idle() too:
2139 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2144 p->se.exec_start = 0;
2145 p->se.sum_exec_runtime = 0;
2146 p->se.prev_sum_exec_runtime = 0;
2147 p->se.nr_migrations = 0;
2149 INIT_LIST_HEAD(&p->se.group_node);
2151 #ifdef CONFIG_FAIR_GROUP_SCHED
2152 p->se.cfs_rq = NULL;
2155 #ifdef CONFIG_SCHEDSTATS
2156 /* Even if schedstat is disabled, there should not be garbage */
2157 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2160 RB_CLEAR_NODE(&p->dl.rb_node);
2161 init_dl_task_timer(&p->dl);
2162 init_dl_inactive_task_timer(&p->dl);
2163 __dl_clear_params(p);
2165 INIT_LIST_HEAD(&p->rt.run_list);
2167 p->rt.time_slice = sched_rr_timeslice;
2171 #ifdef CONFIG_PREEMPT_NOTIFIERS
2172 INIT_HLIST_HEAD(&p->preempt_notifiers);
2175 init_numa_balancing(clone_flags, p);
2178 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2180 #ifdef CONFIG_NUMA_BALANCING
2182 void set_numabalancing_state(bool enabled)
2185 static_branch_enable(&sched_numa_balancing);
2187 static_branch_disable(&sched_numa_balancing);
2190 #ifdef CONFIG_PROC_SYSCTL
2191 int sysctl_numa_balancing(struct ctl_table *table, int write,
2192 void __user *buffer, size_t *lenp, loff_t *ppos)
2196 int state = static_branch_likely(&sched_numa_balancing);
2198 if (write && !capable(CAP_SYS_ADMIN))
2203 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2207 set_numabalancing_state(state);
2213 #ifdef CONFIG_SCHEDSTATS
2215 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2216 static bool __initdata __sched_schedstats = false;
2218 static void set_schedstats(bool enabled)
2221 static_branch_enable(&sched_schedstats);
2223 static_branch_disable(&sched_schedstats);
2226 void force_schedstat_enabled(void)
2228 if (!schedstat_enabled()) {
2229 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2230 static_branch_enable(&sched_schedstats);
2234 static int __init setup_schedstats(char *str)
2241 * This code is called before jump labels have been set up, so we can't
2242 * change the static branch directly just yet. Instead set a temporary
2243 * variable so init_schedstats() can do it later.
2245 if (!strcmp(str, "enable")) {
2246 __sched_schedstats = true;
2248 } else if (!strcmp(str, "disable")) {
2249 __sched_schedstats = false;
2254 pr_warn("Unable to parse schedstats=\n");
2258 __setup("schedstats=", setup_schedstats);
2260 static void __init init_schedstats(void)
2262 set_schedstats(__sched_schedstats);
2265 #ifdef CONFIG_PROC_SYSCTL
2266 int sysctl_schedstats(struct ctl_table *table, int write,
2267 void __user *buffer, size_t *lenp, loff_t *ppos)
2271 int state = static_branch_likely(&sched_schedstats);
2273 if (write && !capable(CAP_SYS_ADMIN))
2278 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2282 set_schedstats(state);
2285 #endif /* CONFIG_PROC_SYSCTL */
2286 #else /* !CONFIG_SCHEDSTATS */
2287 static inline void init_schedstats(void) {}
2288 #endif /* CONFIG_SCHEDSTATS */
2291 * fork()/clone()-time setup:
2293 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2295 unsigned long flags;
2297 __sched_fork(clone_flags, p);
2299 * We mark the process as NEW here. This guarantees that
2300 * nobody will actually run it, and a signal or other external
2301 * event cannot wake it up and insert it on the runqueue either.
2303 p->state = TASK_NEW;
2306 * Make sure we do not leak PI boosting priority to the child.
2308 p->prio = current->normal_prio;
2311 * Revert to default priority/policy on fork if requested.
2313 if (unlikely(p->sched_reset_on_fork)) {
2314 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2315 p->policy = SCHED_NORMAL;
2316 p->static_prio = NICE_TO_PRIO(0);
2318 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2319 p->static_prio = NICE_TO_PRIO(0);
2321 p->prio = p->normal_prio = __normal_prio(p);
2322 set_load_weight(p, false);
2325 * We don't need the reset flag anymore after the fork. It has
2326 * fulfilled its duty:
2328 p->sched_reset_on_fork = 0;
2331 if (dl_prio(p->prio))
2333 else if (rt_prio(p->prio))
2334 p->sched_class = &rt_sched_class;
2336 p->sched_class = &fair_sched_class;
2338 init_entity_runnable_average(&p->se);
2341 * The child is not yet in the pid-hash so no cgroup attach races,
2342 * and the cgroup is pinned to this child due to cgroup_fork()
2343 * is ran before sched_fork().
2345 * Silence PROVE_RCU.
2347 raw_spin_lock_irqsave(&p->pi_lock, flags);
2349 * We're setting the CPU for the first time, we don't migrate,
2350 * so use __set_task_cpu().
2352 __set_task_cpu(p, smp_processor_id());
2353 if (p->sched_class->task_fork)
2354 p->sched_class->task_fork(p);
2355 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2357 #ifdef CONFIG_SCHED_INFO
2358 if (likely(sched_info_on()))
2359 memset(&p->sched_info, 0, sizeof(p->sched_info));
2361 #if defined(CONFIG_SMP)
2364 init_task_preempt_count(p);
2366 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2367 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2372 unsigned long to_ratio(u64 period, u64 runtime)
2374 if (runtime == RUNTIME_INF)
2378 * Doing this here saves a lot of checks in all
2379 * the calling paths, and returning zero seems
2380 * safe for them anyway.
2385 return div64_u64(runtime << BW_SHIFT, period);
2389 * wake_up_new_task - wake up a newly created task for the first time.
2391 * This function will do some initial scheduler statistics housekeeping
2392 * that must be done for every newly created context, then puts the task
2393 * on the runqueue and wakes it.
2395 void wake_up_new_task(struct task_struct *p)
2400 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2401 p->state = TASK_RUNNING;
2404 * Fork balancing, do it here and not earlier because:
2405 * - cpus_allowed can change in the fork path
2406 * - any previously selected CPU might disappear through hotplug
2408 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2409 * as we're not fully set-up yet.
2411 p->recent_used_cpu = task_cpu(p);
2412 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2414 rq = __task_rq_lock(p, &rf);
2415 update_rq_clock(rq);
2416 post_init_entity_util_avg(&p->se);
2418 activate_task(rq, p, ENQUEUE_NOCLOCK);
2419 p->on_rq = TASK_ON_RQ_QUEUED;
2420 trace_sched_wakeup_new(p);
2421 check_preempt_curr(rq, p, WF_FORK);
2423 if (p->sched_class->task_woken) {
2425 * Nothing relies on rq->lock after this, so its fine to
2428 rq_unpin_lock(rq, &rf);
2429 p->sched_class->task_woken(rq, p);
2430 rq_repin_lock(rq, &rf);
2433 task_rq_unlock(rq, p, &rf);
2436 #ifdef CONFIG_PREEMPT_NOTIFIERS
2438 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2440 void preempt_notifier_inc(void)
2442 static_branch_inc(&preempt_notifier_key);
2444 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2446 void preempt_notifier_dec(void)
2448 static_branch_dec(&preempt_notifier_key);
2450 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2453 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2454 * @notifier: notifier struct to register
2456 void preempt_notifier_register(struct preempt_notifier *notifier)
2458 if (!static_branch_unlikely(&preempt_notifier_key))
2459 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2461 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2463 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2466 * preempt_notifier_unregister - no longer interested in preemption notifications
2467 * @notifier: notifier struct to unregister
2469 * This is *not* safe to call from within a preemption notifier.
2471 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2473 hlist_del(¬ifier->link);
2475 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2477 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2479 struct preempt_notifier *notifier;
2481 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2482 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2485 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2487 if (static_branch_unlikely(&preempt_notifier_key))
2488 __fire_sched_in_preempt_notifiers(curr);
2492 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2493 struct task_struct *next)
2495 struct preempt_notifier *notifier;
2497 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2498 notifier->ops->sched_out(notifier, next);
2501 static __always_inline void
2502 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2503 struct task_struct *next)
2505 if (static_branch_unlikely(&preempt_notifier_key))
2506 __fire_sched_out_preempt_notifiers(curr, next);
2509 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2511 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2516 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2517 struct task_struct *next)
2521 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2523 static inline void prepare_task(struct task_struct *next)
2527 * Claim the task as running, we do this before switching to it
2528 * such that any running task will have this set.
2534 static inline void finish_task(struct task_struct *prev)
2538 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2539 * We must ensure this doesn't happen until the switch is completely
2542 * In particular, the load of prev->state in finish_task_switch() must
2543 * happen before this.
2545 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2547 smp_store_release(&prev->on_cpu, 0);
2552 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
2555 * Since the runqueue lock will be released by the next
2556 * task (which is an invalid locking op but in the case
2557 * of the scheduler it's an obvious special-case), so we
2558 * do an early lockdep release here:
2560 rq_unpin_lock(rq, rf);
2561 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2562 #ifdef CONFIG_DEBUG_SPINLOCK
2563 /* this is a valid case when another task releases the spinlock */
2564 rq->lock.owner = next;
2568 static inline void finish_lock_switch(struct rq *rq)
2571 * If we are tracking spinlock dependencies then we have to
2572 * fix up the runqueue lock - which gets 'carried over' from
2573 * prev into current:
2575 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
2576 raw_spin_unlock_irq(&rq->lock);
2580 * NOP if the arch has not defined these:
2583 #ifndef prepare_arch_switch
2584 # define prepare_arch_switch(next) do { } while (0)
2587 #ifndef finish_arch_post_lock_switch
2588 # define finish_arch_post_lock_switch() do { } while (0)
2592 * prepare_task_switch - prepare to switch tasks
2593 * @rq: the runqueue preparing to switch
2594 * @prev: the current task that is being switched out
2595 * @next: the task we are going to switch to.
2597 * This is called with the rq lock held and interrupts off. It must
2598 * be paired with a subsequent finish_task_switch after the context
2601 * prepare_task_switch sets up locking and calls architecture specific
2605 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2606 struct task_struct *next)
2608 kcov_prepare_switch(prev);
2609 sched_info_switch(rq, prev, next);
2610 perf_event_task_sched_out(prev, next);
2612 fire_sched_out_preempt_notifiers(prev, next);
2614 prepare_arch_switch(next);
2618 * finish_task_switch - clean up after a task-switch
2619 * @prev: the thread we just switched away from.
2621 * finish_task_switch must be called after the context switch, paired
2622 * with a prepare_task_switch call before the context switch.
2623 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2624 * and do any other architecture-specific cleanup actions.
2626 * Note that we may have delayed dropping an mm in context_switch(). If
2627 * so, we finish that here outside of the runqueue lock. (Doing it
2628 * with the lock held can cause deadlocks; see schedule() for
2631 * The context switch have flipped the stack from under us and restored the
2632 * local variables which were saved when this task called schedule() in the
2633 * past. prev == current is still correct but we need to recalculate this_rq
2634 * because prev may have moved to another CPU.
2636 static struct rq *finish_task_switch(struct task_struct *prev)
2637 __releases(rq->lock)
2639 struct rq *rq = this_rq();
2640 struct mm_struct *mm = rq->prev_mm;
2644 * The previous task will have left us with a preempt_count of 2
2645 * because it left us after:
2648 * preempt_disable(); // 1
2650 * raw_spin_lock_irq(&rq->lock) // 2
2652 * Also, see FORK_PREEMPT_COUNT.
2654 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2655 "corrupted preempt_count: %s/%d/0x%x\n",
2656 current->comm, current->pid, preempt_count()))
2657 preempt_count_set(FORK_PREEMPT_COUNT);
2662 * A task struct has one reference for the use as "current".
2663 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2664 * schedule one last time. The schedule call will never return, and
2665 * the scheduled task must drop that reference.
2667 * We must observe prev->state before clearing prev->on_cpu (in
2668 * finish_task), otherwise a concurrent wakeup can get prev
2669 * running on another CPU and we could rave with its RUNNING -> DEAD
2670 * transition, resulting in a double drop.
2672 prev_state = prev->state;
2673 vtime_task_switch(prev);
2674 perf_event_task_sched_in(prev, current);
2676 finish_lock_switch(rq);
2677 finish_arch_post_lock_switch();
2678 kcov_finish_switch(current);
2680 fire_sched_in_preempt_notifiers(current);
2682 * When switching through a kernel thread, the loop in
2683 * membarrier_{private,global}_expedited() may have observed that
2684 * kernel thread and not issued an IPI. It is therefore possible to
2685 * schedule between user->kernel->user threads without passing though
2686 * switch_mm(). Membarrier requires a barrier after storing to
2687 * rq->curr, before returning to userspace, so provide them here:
2689 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
2690 * provided by mmdrop(),
2691 * - a sync_core for SYNC_CORE.
2694 membarrier_mm_sync_core_before_usermode(mm);
2697 if (unlikely(prev_state == TASK_DEAD)) {
2698 if (prev->sched_class->task_dead)
2699 prev->sched_class->task_dead(prev);
2702 * Remove function-return probe instances associated with this
2703 * task and put them back on the free list.
2705 kprobe_flush_task(prev);
2707 /* Task is done with its stack. */
2708 put_task_stack(prev);
2710 put_task_struct(prev);
2713 tick_nohz_task_switch();
2719 /* rq->lock is NOT held, but preemption is disabled */
2720 static void __balance_callback(struct rq *rq)
2722 struct callback_head *head, *next;
2723 void (*func)(struct rq *rq);
2724 unsigned long flags;
2726 raw_spin_lock_irqsave(&rq->lock, flags);
2727 head = rq->balance_callback;
2728 rq->balance_callback = NULL;
2730 func = (void (*)(struct rq *))head->func;
2737 raw_spin_unlock_irqrestore(&rq->lock, flags);
2740 static inline void balance_callback(struct rq *rq)
2742 if (unlikely(rq->balance_callback))
2743 __balance_callback(rq);
2748 static inline void balance_callback(struct rq *rq)
2755 * schedule_tail - first thing a freshly forked thread must call.
2756 * @prev: the thread we just switched away from.
2758 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2759 __releases(rq->lock)
2764 * New tasks start with FORK_PREEMPT_COUNT, see there and
2765 * finish_task_switch() for details.
2767 * finish_task_switch() will drop rq->lock() and lower preempt_count
2768 * and the preempt_enable() will end up enabling preemption (on
2769 * PREEMPT_COUNT kernels).
2772 rq = finish_task_switch(prev);
2773 balance_callback(rq);
2776 if (current->set_child_tid)
2777 put_user(task_pid_vnr(current), current->set_child_tid);
2779 calculate_sigpending();
2783 * context_switch - switch to the new MM and the new thread's register state.
2785 static __always_inline struct rq *
2786 context_switch(struct rq *rq, struct task_struct *prev,
2787 struct task_struct *next, struct rq_flags *rf)
2789 struct mm_struct *mm, *oldmm;
2791 prepare_task_switch(rq, prev, next);
2794 oldmm = prev->active_mm;
2796 * For paravirt, this is coupled with an exit in switch_to to
2797 * combine the page table reload and the switch backend into
2800 arch_start_context_switch(prev);
2803 * If mm is non-NULL, we pass through switch_mm(). If mm is
2804 * NULL, we will pass through mmdrop() in finish_task_switch().
2805 * Both of these contain the full memory barrier required by
2806 * membarrier after storing to rq->curr, before returning to
2810 next->active_mm = oldmm;
2812 enter_lazy_tlb(oldmm, next);
2814 switch_mm_irqs_off(oldmm, mm, next);
2817 prev->active_mm = NULL;
2818 rq->prev_mm = oldmm;
2821 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2823 prepare_lock_switch(rq, next, rf);
2825 /* Here we just switch the register state and the stack. */
2826 switch_to(prev, next, prev);
2829 return finish_task_switch(prev);
2833 * nr_running and nr_context_switches:
2835 * externally visible scheduler statistics: current number of runnable
2836 * threads, total number of context switches performed since bootup.
2838 unsigned long nr_running(void)
2840 unsigned long i, sum = 0;
2842 for_each_online_cpu(i)
2843 sum += cpu_rq(i)->nr_running;
2849 * Check if only the current task is running on the CPU.
2851 * Caution: this function does not check that the caller has disabled
2852 * preemption, thus the result might have a time-of-check-to-time-of-use
2853 * race. The caller is responsible to use it correctly, for example:
2855 * - from a non-preemptable section (of course)
2857 * - from a thread that is bound to a single CPU
2859 * - in a loop with very short iterations (e.g. a polling loop)
2861 bool single_task_running(void)
2863 return raw_rq()->nr_running == 1;
2865 EXPORT_SYMBOL(single_task_running);
2867 unsigned long long nr_context_switches(void)
2870 unsigned long long sum = 0;
2872 for_each_possible_cpu(i)
2873 sum += cpu_rq(i)->nr_switches;
2879 * IO-wait accounting, and how its mostly bollocks (on SMP).
2881 * The idea behind IO-wait account is to account the idle time that we could
2882 * have spend running if it were not for IO. That is, if we were to improve the
2883 * storage performance, we'd have a proportional reduction in IO-wait time.
2885 * This all works nicely on UP, where, when a task blocks on IO, we account
2886 * idle time as IO-wait, because if the storage were faster, it could've been
2887 * running and we'd not be idle.
2889 * This has been extended to SMP, by doing the same for each CPU. This however
2892 * Imagine for instance the case where two tasks block on one CPU, only the one
2893 * CPU will have IO-wait accounted, while the other has regular idle. Even
2894 * though, if the storage were faster, both could've ran at the same time,
2895 * utilising both CPUs.
2897 * This means, that when looking globally, the current IO-wait accounting on
2898 * SMP is a lower bound, by reason of under accounting.
2900 * Worse, since the numbers are provided per CPU, they are sometimes
2901 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2902 * associated with any one particular CPU, it can wake to another CPU than it
2903 * blocked on. This means the per CPU IO-wait number is meaningless.
2905 * Task CPU affinities can make all that even more 'interesting'.
2908 unsigned long nr_iowait(void)
2910 unsigned long i, sum = 0;
2912 for_each_possible_cpu(i)
2913 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2919 * Consumers of these two interfaces, like for example the cpuidle menu
2920 * governor, are using nonsensical data. Preferring shallow idle state selection
2921 * for a CPU that has IO-wait which might not even end up running the task when
2922 * it does become runnable.
2925 unsigned long nr_iowait_cpu(int cpu)
2927 struct rq *this = cpu_rq(cpu);
2928 return atomic_read(&this->nr_iowait);
2931 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2933 struct rq *rq = this_rq();
2934 *nr_waiters = atomic_read(&rq->nr_iowait);
2935 *load = rq->load.weight;
2941 * sched_exec - execve() is a valuable balancing opportunity, because at
2942 * this point the task has the smallest effective memory and cache footprint.
2944 void sched_exec(void)
2946 struct task_struct *p = current;
2947 unsigned long flags;
2950 raw_spin_lock_irqsave(&p->pi_lock, flags);
2951 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2952 if (dest_cpu == smp_processor_id())
2955 if (likely(cpu_active(dest_cpu))) {
2956 struct migration_arg arg = { p, dest_cpu };
2958 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2959 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2963 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2968 DEFINE_PER_CPU(struct kernel_stat, kstat);
2969 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2971 EXPORT_PER_CPU_SYMBOL(kstat);
2972 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2975 * The function fair_sched_class.update_curr accesses the struct curr
2976 * and its field curr->exec_start; when called from task_sched_runtime(),
2977 * we observe a high rate of cache misses in practice.
2978 * Prefetching this data results in improved performance.
2980 static inline void prefetch_curr_exec_start(struct task_struct *p)
2982 #ifdef CONFIG_FAIR_GROUP_SCHED
2983 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
2985 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
2988 prefetch(&curr->exec_start);
2992 * Return accounted runtime for the task.
2993 * In case the task is currently running, return the runtime plus current's
2994 * pending runtime that have not been accounted yet.
2996 unsigned long long task_sched_runtime(struct task_struct *p)
3002 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3004 * 64-bit doesn't need locks to atomically read a 64-bit value.
3005 * So we have a optimization chance when the task's delta_exec is 0.
3006 * Reading ->on_cpu is racy, but this is ok.
3008 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3009 * If we race with it entering CPU, unaccounted time is 0. This is
3010 * indistinguishable from the read occurring a few cycles earlier.
3011 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3012 * been accounted, so we're correct here as well.
3014 if (!p->on_cpu || !task_on_rq_queued(p))
3015 return p->se.sum_exec_runtime;
3018 rq = task_rq_lock(p, &rf);
3020 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3021 * project cycles that may never be accounted to this
3022 * thread, breaking clock_gettime().
3024 if (task_current(rq, p) && task_on_rq_queued(p)) {
3025 prefetch_curr_exec_start(p);
3026 update_rq_clock(rq);
3027 p->sched_class->update_curr(rq);
3029 ns = p->se.sum_exec_runtime;
3030 task_rq_unlock(rq, p, &rf);
3036 * This function gets called by the timer code, with HZ frequency.
3037 * We call it with interrupts disabled.
3039 void scheduler_tick(void)
3041 int cpu = smp_processor_id();
3042 struct rq *rq = cpu_rq(cpu);
3043 struct task_struct *curr = rq->curr;
3050 update_rq_clock(rq);
3051 curr->sched_class->task_tick(rq, curr, 0);
3052 cpu_load_update_active(rq);
3053 calc_global_load_tick(rq);
3057 perf_event_task_tick();
3060 rq->idle_balance = idle_cpu(cpu);
3061 trigger_load_balance(rq);
3065 #ifdef CONFIG_NO_HZ_FULL
3069 struct delayed_work work;
3072 static struct tick_work __percpu *tick_work_cpu;
3074 static void sched_tick_remote(struct work_struct *work)
3076 struct delayed_work *dwork = to_delayed_work(work);
3077 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3078 int cpu = twork->cpu;
3079 struct rq *rq = cpu_rq(cpu);
3080 struct task_struct *curr;
3085 * Handle the tick only if it appears the remote CPU is running in full
3086 * dynticks mode. The check is racy by nature, but missing a tick or
3087 * having one too much is no big deal because the scheduler tick updates
3088 * statistics and checks timeslices in a time-independent way, regardless
3089 * of when exactly it is running.
3091 if (idle_cpu(cpu) || !tick_nohz_tick_stopped_cpu(cpu))
3094 rq_lock_irq(rq, &rf);
3096 if (is_idle_task(curr))
3099 update_rq_clock(rq);
3100 delta = rq_clock_task(rq) - curr->se.exec_start;
3103 * Make sure the next tick runs within a reasonable
3106 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3107 curr->sched_class->task_tick(rq, curr, 0);
3110 rq_unlock_irq(rq, &rf);
3114 * Run the remote tick once per second (1Hz). This arbitrary
3115 * frequency is large enough to avoid overload but short enough
3116 * to keep scheduler internal stats reasonably up to date.
3118 queue_delayed_work(system_unbound_wq, dwork, HZ);
3121 static void sched_tick_start(int cpu)
3123 struct tick_work *twork;
3125 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3128 WARN_ON_ONCE(!tick_work_cpu);
3130 twork = per_cpu_ptr(tick_work_cpu, cpu);
3132 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3133 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3136 #ifdef CONFIG_HOTPLUG_CPU
3137 static void sched_tick_stop(int cpu)
3139 struct tick_work *twork;
3141 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3144 WARN_ON_ONCE(!tick_work_cpu);
3146 twork = per_cpu_ptr(tick_work_cpu, cpu);
3147 cancel_delayed_work_sync(&twork->work);
3149 #endif /* CONFIG_HOTPLUG_CPU */
3151 int __init sched_tick_offload_init(void)
3153 tick_work_cpu = alloc_percpu(struct tick_work);
3154 BUG_ON(!tick_work_cpu);
3159 #else /* !CONFIG_NO_HZ_FULL */
3160 static inline void sched_tick_start(int cpu) { }
3161 static inline void sched_tick_stop(int cpu) { }
3164 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3165 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3167 * If the value passed in is equal to the current preempt count
3168 * then we just disabled preemption. Start timing the latency.
3170 static inline void preempt_latency_start(int val)
3172 if (preempt_count() == val) {
3173 unsigned long ip = get_lock_parent_ip();
3174 #ifdef CONFIG_DEBUG_PREEMPT
3175 current->preempt_disable_ip = ip;
3177 trace_preempt_off(CALLER_ADDR0, ip);
3181 void preempt_count_add(int val)
3183 #ifdef CONFIG_DEBUG_PREEMPT
3187 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3190 __preempt_count_add(val);
3191 #ifdef CONFIG_DEBUG_PREEMPT
3193 * Spinlock count overflowing soon?
3195 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3198 preempt_latency_start(val);
3200 EXPORT_SYMBOL(preempt_count_add);
3201 NOKPROBE_SYMBOL(preempt_count_add);
3204 * If the value passed in equals to the current preempt count
3205 * then we just enabled preemption. Stop timing the latency.
3207 static inline void preempt_latency_stop(int val)
3209 if (preempt_count() == val)
3210 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3213 void preempt_count_sub(int val)
3215 #ifdef CONFIG_DEBUG_PREEMPT
3219 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3222 * Is the spinlock portion underflowing?
3224 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3225 !(preempt_count() & PREEMPT_MASK)))
3229 preempt_latency_stop(val);
3230 __preempt_count_sub(val);
3232 EXPORT_SYMBOL(preempt_count_sub);
3233 NOKPROBE_SYMBOL(preempt_count_sub);
3236 static inline void preempt_latency_start(int val) { }
3237 static inline void preempt_latency_stop(int val) { }
3240 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3242 #ifdef CONFIG_DEBUG_PREEMPT
3243 return p->preempt_disable_ip;
3250 * Print scheduling while atomic bug:
3252 static noinline void __schedule_bug(struct task_struct *prev)
3254 /* Save this before calling printk(), since that will clobber it */
3255 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3257 if (oops_in_progress)
3260 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3261 prev->comm, prev->pid, preempt_count());
3263 debug_show_held_locks(prev);
3265 if (irqs_disabled())
3266 print_irqtrace_events(prev);
3267 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3268 && in_atomic_preempt_off()) {
3269 pr_err("Preemption disabled at:");
3270 print_ip_sym(preempt_disable_ip);
3274 panic("scheduling while atomic\n");
3277 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3281 * Various schedule()-time debugging checks and statistics:
3283 static inline void schedule_debug(struct task_struct *prev)
3285 #ifdef CONFIG_SCHED_STACK_END_CHECK
3286 if (task_stack_end_corrupted(prev))
3287 panic("corrupted stack end detected inside scheduler\n");
3290 if (unlikely(in_atomic_preempt_off())) {
3291 __schedule_bug(prev);
3292 preempt_count_set(PREEMPT_DISABLED);
3296 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3298 schedstat_inc(this_rq()->sched_count);
3302 * Pick up the highest-prio task:
3304 static inline struct task_struct *
3305 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3307 const struct sched_class *class;
3308 struct task_struct *p;
3311 * Optimization: we know that if all tasks are in the fair class we can
3312 * call that function directly, but only if the @prev task wasn't of a
3313 * higher scheduling class, because otherwise those loose the
3314 * opportunity to pull in more work from other CPUs.
3316 if (likely((prev->sched_class == &idle_sched_class ||
3317 prev->sched_class == &fair_sched_class) &&
3318 rq->nr_running == rq->cfs.h_nr_running)) {
3320 p = fair_sched_class.pick_next_task(rq, prev, rf);
3321 if (unlikely(p == RETRY_TASK))
3324 /* Assumes fair_sched_class->next == idle_sched_class */
3326 p = idle_sched_class.pick_next_task(rq, prev, rf);
3332 for_each_class(class) {
3333 p = class->pick_next_task(rq, prev, rf);
3335 if (unlikely(p == RETRY_TASK))
3341 /* The idle class should always have a runnable task: */
3346 * __schedule() is the main scheduler function.
3348 * The main means of driving the scheduler and thus entering this function are:
3350 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3352 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3353 * paths. For example, see arch/x86/entry_64.S.
3355 * To drive preemption between tasks, the scheduler sets the flag in timer
3356 * interrupt handler scheduler_tick().
3358 * 3. Wakeups don't really cause entry into schedule(). They add a
3359 * task to the run-queue and that's it.
3361 * Now, if the new task added to the run-queue preempts the current
3362 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3363 * called on the nearest possible occasion:
3365 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3367 * - in syscall or exception context, at the next outmost
3368 * preempt_enable(). (this might be as soon as the wake_up()'s
3371 * - in IRQ context, return from interrupt-handler to
3372 * preemptible context
3374 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3377 * - cond_resched() call
3378 * - explicit schedule() call
3379 * - return from syscall or exception to user-space
3380 * - return from interrupt-handler to user-space
3382 * WARNING: must be called with preemption disabled!
3384 static void __sched notrace __schedule(bool preempt)
3386 struct task_struct *prev, *next;
3387 unsigned long *switch_count;
3392 cpu = smp_processor_id();
3396 schedule_debug(prev);
3398 if (sched_feat(HRTICK))
3401 local_irq_disable();
3402 rcu_note_context_switch(preempt);
3405 * Make sure that signal_pending_state()->signal_pending() below
3406 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3407 * done by the caller to avoid the race with signal_wake_up().
3409 * The membarrier system call requires a full memory barrier
3410 * after coming from user-space, before storing to rq->curr.
3413 smp_mb__after_spinlock();
3415 /* Promote REQ to ACT */
3416 rq->clock_update_flags <<= 1;
3417 update_rq_clock(rq);
3419 switch_count = &prev->nivcsw;
3420 if (!preempt && prev->state) {
3421 if (unlikely(signal_pending_state(prev->state, prev))) {
3422 prev->state = TASK_RUNNING;
3424 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3427 if (prev->in_iowait) {
3428 atomic_inc(&rq->nr_iowait);
3429 delayacct_blkio_start();
3433 * If a worker went to sleep, notify and ask workqueue
3434 * whether it wants to wake up a task to maintain
3437 if (prev->flags & PF_WQ_WORKER) {
3438 struct task_struct *to_wakeup;
3440 to_wakeup = wq_worker_sleeping(prev);
3442 try_to_wake_up_local(to_wakeup, &rf);
3445 switch_count = &prev->nvcsw;
3448 next = pick_next_task(rq, prev, &rf);
3449 clear_tsk_need_resched(prev);
3450 clear_preempt_need_resched();
3452 if (likely(prev != next)) {
3456 * The membarrier system call requires each architecture
3457 * to have a full memory barrier after updating
3458 * rq->curr, before returning to user-space.
3460 * Here are the schemes providing that barrier on the
3461 * various architectures:
3462 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3463 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3464 * - finish_lock_switch() for weakly-ordered
3465 * architectures where spin_unlock is a full barrier,
3466 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3467 * is a RELEASE barrier),
3471 trace_sched_switch(preempt, prev, next);
3473 /* Also unlocks the rq: */
3474 rq = context_switch(rq, prev, next, &rf);
3476 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3477 rq_unlock_irq(rq, &rf);
3480 balance_callback(rq);
3483 void __noreturn do_task_dead(void)
3485 /* Causes final put_task_struct in finish_task_switch(): */
3486 set_special_state(TASK_DEAD);
3488 /* Tell freezer to ignore us: */
3489 current->flags |= PF_NOFREEZE;
3494 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3499 static inline void sched_submit_work(struct task_struct *tsk)
3501 if (!tsk->state || tsk_is_pi_blocked(tsk))
3504 * If we are going to sleep and we have plugged IO queued,
3505 * make sure to submit it to avoid deadlocks.
3507 if (blk_needs_flush_plug(tsk))
3508 blk_schedule_flush_plug(tsk);
3511 asmlinkage __visible void __sched schedule(void)
3513 struct task_struct *tsk = current;
3515 sched_submit_work(tsk);
3519 sched_preempt_enable_no_resched();
3520 } while (need_resched());
3522 EXPORT_SYMBOL(schedule);
3525 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3526 * state (have scheduled out non-voluntarily) by making sure that all
3527 * tasks have either left the run queue or have gone into user space.
3528 * As idle tasks do not do either, they must not ever be preempted
3529 * (schedule out non-voluntarily).
3531 * schedule_idle() is similar to schedule_preempt_disable() except that it
3532 * never enables preemption because it does not call sched_submit_work().
3534 void __sched schedule_idle(void)
3537 * As this skips calling sched_submit_work(), which the idle task does
3538 * regardless because that function is a nop when the task is in a
3539 * TASK_RUNNING state, make sure this isn't used someplace that the
3540 * current task can be in any other state. Note, idle is always in the
3541 * TASK_RUNNING state.
3543 WARN_ON_ONCE(current->state);
3546 } while (need_resched());
3549 #ifdef CONFIG_CONTEXT_TRACKING
3550 asmlinkage __visible void __sched schedule_user(void)
3553 * If we come here after a random call to set_need_resched(),
3554 * or we have been woken up remotely but the IPI has not yet arrived,
3555 * we haven't yet exited the RCU idle mode. Do it here manually until
3556 * we find a better solution.
3558 * NB: There are buggy callers of this function. Ideally we
3559 * should warn if prev_state != CONTEXT_USER, but that will trigger
3560 * too frequently to make sense yet.
3562 enum ctx_state prev_state = exception_enter();
3564 exception_exit(prev_state);
3569 * schedule_preempt_disabled - called with preemption disabled
3571 * Returns with preemption disabled. Note: preempt_count must be 1
3573 void __sched schedule_preempt_disabled(void)
3575 sched_preempt_enable_no_resched();
3580 static void __sched notrace preempt_schedule_common(void)
3584 * Because the function tracer can trace preempt_count_sub()
3585 * and it also uses preempt_enable/disable_notrace(), if
3586 * NEED_RESCHED is set, the preempt_enable_notrace() called
3587 * by the function tracer will call this function again and
3588 * cause infinite recursion.
3590 * Preemption must be disabled here before the function
3591 * tracer can trace. Break up preempt_disable() into two
3592 * calls. One to disable preemption without fear of being
3593 * traced. The other to still record the preemption latency,
3594 * which can also be traced by the function tracer.
3596 preempt_disable_notrace();
3597 preempt_latency_start(1);
3599 preempt_latency_stop(1);
3600 preempt_enable_no_resched_notrace();
3603 * Check again in case we missed a preemption opportunity
3604 * between schedule and now.
3606 } while (need_resched());
3609 #ifdef CONFIG_PREEMPT
3611 * this is the entry point to schedule() from in-kernel preemption
3612 * off of preempt_enable. Kernel preemptions off return from interrupt
3613 * occur there and call schedule directly.
3615 asmlinkage __visible void __sched notrace preempt_schedule(void)
3618 * If there is a non-zero preempt_count or interrupts are disabled,
3619 * we do not want to preempt the current task. Just return..
3621 if (likely(!preemptible()))
3624 preempt_schedule_common();
3626 NOKPROBE_SYMBOL(preempt_schedule);
3627 EXPORT_SYMBOL(preempt_schedule);
3630 * preempt_schedule_notrace - preempt_schedule called by tracing
3632 * The tracing infrastructure uses preempt_enable_notrace to prevent
3633 * recursion and tracing preempt enabling caused by the tracing
3634 * infrastructure itself. But as tracing can happen in areas coming
3635 * from userspace or just about to enter userspace, a preempt enable
3636 * can occur before user_exit() is called. This will cause the scheduler
3637 * to be called when the system is still in usermode.
3639 * To prevent this, the preempt_enable_notrace will use this function
3640 * instead of preempt_schedule() to exit user context if needed before
3641 * calling the scheduler.
3643 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3645 enum ctx_state prev_ctx;
3647 if (likely(!preemptible()))
3652 * Because the function tracer can trace preempt_count_sub()
3653 * and it also uses preempt_enable/disable_notrace(), if
3654 * NEED_RESCHED is set, the preempt_enable_notrace() called
3655 * by the function tracer will call this function again and
3656 * cause infinite recursion.
3658 * Preemption must be disabled here before the function
3659 * tracer can trace. Break up preempt_disable() into two
3660 * calls. One to disable preemption without fear of being
3661 * traced. The other to still record the preemption latency,
3662 * which can also be traced by the function tracer.
3664 preempt_disable_notrace();
3665 preempt_latency_start(1);
3667 * Needs preempt disabled in case user_exit() is traced
3668 * and the tracer calls preempt_enable_notrace() causing
3669 * an infinite recursion.
3671 prev_ctx = exception_enter();
3673 exception_exit(prev_ctx);
3675 preempt_latency_stop(1);
3676 preempt_enable_no_resched_notrace();
3677 } while (need_resched());
3679 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3681 #endif /* CONFIG_PREEMPT */
3684 * this is the entry point to schedule() from kernel preemption
3685 * off of irq context.
3686 * Note, that this is called and return with irqs disabled. This will
3687 * protect us against recursive calling from irq.
3689 asmlinkage __visible void __sched preempt_schedule_irq(void)
3691 enum ctx_state prev_state;
3693 /* Catch callers which need to be fixed */
3694 BUG_ON(preempt_count() || !irqs_disabled());
3696 prev_state = exception_enter();
3702 local_irq_disable();
3703 sched_preempt_enable_no_resched();
3704 } while (need_resched());
3706 exception_exit(prev_state);
3709 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
3712 return try_to_wake_up(curr->private, mode, wake_flags);
3714 EXPORT_SYMBOL(default_wake_function);
3716 #ifdef CONFIG_RT_MUTEXES
3718 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
3721 prio = min(prio, pi_task->prio);
3726 static inline int rt_effective_prio(struct task_struct *p, int prio)
3728 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3730 return __rt_effective_prio(pi_task, prio);
3734 * rt_mutex_setprio - set the current priority of a task
3736 * @pi_task: donor task
3738 * This function changes the 'effective' priority of a task. It does
3739 * not touch ->normal_prio like __setscheduler().
3741 * Used by the rt_mutex code to implement priority inheritance
3742 * logic. Call site only calls if the priority of the task changed.
3744 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
3746 int prio, oldprio, queued, running, queue_flag =
3747 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
3748 const struct sched_class *prev_class;
3752 /* XXX used to be waiter->prio, not waiter->task->prio */
3753 prio = __rt_effective_prio(pi_task, p->normal_prio);
3756 * If nothing changed; bail early.
3758 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
3761 rq = __task_rq_lock(p, &rf);
3762 update_rq_clock(rq);
3764 * Set under pi_lock && rq->lock, such that the value can be used under
3767 * Note that there is loads of tricky to make this pointer cache work
3768 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3769 * ensure a task is de-boosted (pi_task is set to NULL) before the
3770 * task is allowed to run again (and can exit). This ensures the pointer
3771 * points to a blocked task -- which guaratees the task is present.
3773 p->pi_top_task = pi_task;
3776 * For FIFO/RR we only need to set prio, if that matches we're done.
3778 if (prio == p->prio && !dl_prio(prio))
3782 * Idle task boosting is a nono in general. There is one
3783 * exception, when PREEMPT_RT and NOHZ is active:
3785 * The idle task calls get_next_timer_interrupt() and holds
3786 * the timer wheel base->lock on the CPU and another CPU wants
3787 * to access the timer (probably to cancel it). We can safely
3788 * ignore the boosting request, as the idle CPU runs this code
3789 * with interrupts disabled and will complete the lock
3790 * protected section without being interrupted. So there is no
3791 * real need to boost.
3793 if (unlikely(p == rq->idle)) {
3794 WARN_ON(p != rq->curr);
3795 WARN_ON(p->pi_blocked_on);
3799 trace_sched_pi_setprio(p, pi_task);
3802 if (oldprio == prio)
3803 queue_flag &= ~DEQUEUE_MOVE;
3805 prev_class = p->sched_class;
3806 queued = task_on_rq_queued(p);
3807 running = task_current(rq, p);
3809 dequeue_task(rq, p, queue_flag);
3811 put_prev_task(rq, p);
3814 * Boosting condition are:
3815 * 1. -rt task is running and holds mutex A
3816 * --> -dl task blocks on mutex A
3818 * 2. -dl task is running and holds mutex A
3819 * --> -dl task blocks on mutex A and could preempt the
3822 if (dl_prio(prio)) {
3823 if (!dl_prio(p->normal_prio) ||
3824 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3825 p->dl.dl_boosted = 1;
3826 queue_flag |= ENQUEUE_REPLENISH;
3828 p->dl.dl_boosted = 0;
3829 p->sched_class = &dl_sched_class;
3830 } else if (rt_prio(prio)) {
3831 if (dl_prio(oldprio))
3832 p->dl.dl_boosted = 0;
3834 queue_flag |= ENQUEUE_HEAD;
3835 p->sched_class = &rt_sched_class;
3837 if (dl_prio(oldprio))
3838 p->dl.dl_boosted = 0;
3839 if (rt_prio(oldprio))
3841 p->sched_class = &fair_sched_class;
3847 enqueue_task(rq, p, queue_flag);
3849 set_curr_task(rq, p);
3851 check_class_changed(rq, p, prev_class, oldprio);
3853 /* Avoid rq from going away on us: */
3855 __task_rq_unlock(rq, &rf);
3857 balance_callback(rq);
3861 static inline int rt_effective_prio(struct task_struct *p, int prio)
3867 void set_user_nice(struct task_struct *p, long nice)
3869 bool queued, running;
3870 int old_prio, delta;
3874 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3877 * We have to be careful, if called from sys_setpriority(),
3878 * the task might be in the middle of scheduling on another CPU.
3880 rq = task_rq_lock(p, &rf);
3881 update_rq_clock(rq);
3884 * The RT priorities are set via sched_setscheduler(), but we still
3885 * allow the 'normal' nice value to be set - but as expected
3886 * it wont have any effect on scheduling until the task is
3887 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3889 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3890 p->static_prio = NICE_TO_PRIO(nice);
3893 queued = task_on_rq_queued(p);
3894 running = task_current(rq, p);
3896 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
3898 put_prev_task(rq, p);
3900 p->static_prio = NICE_TO_PRIO(nice);
3901 set_load_weight(p, true);
3903 p->prio = effective_prio(p);
3904 delta = p->prio - old_prio;
3907 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
3909 * If the task increased its priority or is running and
3910 * lowered its priority, then reschedule its CPU:
3912 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3916 set_curr_task(rq, p);
3918 task_rq_unlock(rq, p, &rf);
3920 EXPORT_SYMBOL(set_user_nice);
3923 * can_nice - check if a task can reduce its nice value
3927 int can_nice(const struct task_struct *p, const int nice)
3929 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3930 int nice_rlim = nice_to_rlimit(nice);
3932 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3933 capable(CAP_SYS_NICE));
3936 #ifdef __ARCH_WANT_SYS_NICE
3939 * sys_nice - change the priority of the current process.
3940 * @increment: priority increment
3942 * sys_setpriority is a more generic, but much slower function that
3943 * does similar things.
3945 SYSCALL_DEFINE1(nice, int, increment)
3950 * Setpriority might change our priority at the same moment.
3951 * We don't have to worry. Conceptually one call occurs first
3952 * and we have a single winner.
3954 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3955 nice = task_nice(current) + increment;
3957 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3958 if (increment < 0 && !can_nice(current, nice))
3961 retval = security_task_setnice(current, nice);
3965 set_user_nice(current, nice);
3972 * task_prio - return the priority value of a given task.
3973 * @p: the task in question.
3975 * Return: The priority value as seen by users in /proc.
3976 * RT tasks are offset by -200. Normal tasks are centered
3977 * around 0, value goes from -16 to +15.
3979 int task_prio(const struct task_struct *p)
3981 return p->prio - MAX_RT_PRIO;
3985 * idle_cpu - is a given CPU idle currently?
3986 * @cpu: the processor in question.
3988 * Return: 1 if the CPU is currently idle. 0 otherwise.
3990 int idle_cpu(int cpu)
3992 struct rq *rq = cpu_rq(cpu);
3994 if (rq->curr != rq->idle)
4001 if (!llist_empty(&rq->wake_list))
4009 * available_idle_cpu - is a given CPU idle for enqueuing work.
4010 * @cpu: the CPU in question.
4012 * Return: 1 if the CPU is currently idle. 0 otherwise.
4014 int available_idle_cpu(int cpu)
4019 if (vcpu_is_preempted(cpu))
4026 * idle_task - return the idle task for a given CPU.
4027 * @cpu: the processor in question.
4029 * Return: The idle task for the CPU @cpu.
4031 struct task_struct *idle_task(int cpu)
4033 return cpu_rq(cpu)->idle;
4037 * find_process_by_pid - find a process with a matching PID value.
4038 * @pid: the pid in question.
4040 * The task of @pid, if found. %NULL otherwise.
4042 static struct task_struct *find_process_by_pid(pid_t pid)
4044 return pid ? find_task_by_vpid(pid) : current;
4048 * sched_setparam() passes in -1 for its policy, to let the functions
4049 * it calls know not to change it.
4051 #define SETPARAM_POLICY -1
4053 static void __setscheduler_params(struct task_struct *p,
4054 const struct sched_attr *attr)
4056 int policy = attr->sched_policy;
4058 if (policy == SETPARAM_POLICY)
4063 if (dl_policy(policy))
4064 __setparam_dl(p, attr);
4065 else if (fair_policy(policy))
4066 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4069 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4070 * !rt_policy. Always setting this ensures that things like
4071 * getparam()/getattr() don't report silly values for !rt tasks.
4073 p->rt_priority = attr->sched_priority;
4074 p->normal_prio = normal_prio(p);
4075 set_load_weight(p, true);
4078 /* Actually do priority change: must hold pi & rq lock. */
4079 static void __setscheduler(struct rq *rq, struct task_struct *p,
4080 const struct sched_attr *attr, bool keep_boost)
4082 __setscheduler_params(p, attr);
4085 * Keep a potential priority boosting if called from
4086 * sched_setscheduler().
4088 p->prio = normal_prio(p);
4090 p->prio = rt_effective_prio(p, p->prio);
4092 if (dl_prio(p->prio))
4093 p->sched_class = &dl_sched_class;
4094 else if (rt_prio(p->prio))
4095 p->sched_class = &rt_sched_class;
4097 p->sched_class = &fair_sched_class;
4101 * Check the target process has a UID that matches the current process's:
4103 static bool check_same_owner(struct task_struct *p)
4105 const struct cred *cred = current_cred(), *pcred;
4109 pcred = __task_cred(p);
4110 match = (uid_eq(cred->euid, pcred->euid) ||
4111 uid_eq(cred->euid, pcred->uid));
4116 static int __sched_setscheduler(struct task_struct *p,
4117 const struct sched_attr *attr,
4120 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4121 MAX_RT_PRIO - 1 - attr->sched_priority;
4122 int retval, oldprio, oldpolicy = -1, queued, running;
4123 int new_effective_prio, policy = attr->sched_policy;
4124 const struct sched_class *prev_class;
4127 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4130 /* The pi code expects interrupts enabled */
4131 BUG_ON(pi && in_interrupt());
4133 /* Double check policy once rq lock held: */
4135 reset_on_fork = p->sched_reset_on_fork;
4136 policy = oldpolicy = p->policy;
4138 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4140 if (!valid_policy(policy))
4144 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4148 * Valid priorities for SCHED_FIFO and SCHED_RR are
4149 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4150 * SCHED_BATCH and SCHED_IDLE is 0.
4152 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4153 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4155 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4156 (rt_policy(policy) != (attr->sched_priority != 0)))
4160 * Allow unprivileged RT tasks to decrease priority:
4162 if (user && !capable(CAP_SYS_NICE)) {
4163 if (fair_policy(policy)) {
4164 if (attr->sched_nice < task_nice(p) &&
4165 !can_nice(p, attr->sched_nice))
4169 if (rt_policy(policy)) {
4170 unsigned long rlim_rtprio =
4171 task_rlimit(p, RLIMIT_RTPRIO);
4173 /* Can't set/change the rt policy: */
4174 if (policy != p->policy && !rlim_rtprio)
4177 /* Can't increase priority: */
4178 if (attr->sched_priority > p->rt_priority &&
4179 attr->sched_priority > rlim_rtprio)
4184 * Can't set/change SCHED_DEADLINE policy at all for now
4185 * (safest behavior); in the future we would like to allow
4186 * unprivileged DL tasks to increase their relative deadline
4187 * or reduce their runtime (both ways reducing utilization)
4189 if (dl_policy(policy))
4193 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4194 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4196 if (idle_policy(p->policy) && !idle_policy(policy)) {
4197 if (!can_nice(p, task_nice(p)))
4201 /* Can't change other user's priorities: */
4202 if (!check_same_owner(p))
4205 /* Normal users shall not reset the sched_reset_on_fork flag: */
4206 if (p->sched_reset_on_fork && !reset_on_fork)
4211 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4214 retval = security_task_setscheduler(p);
4220 * Make sure no PI-waiters arrive (or leave) while we are
4221 * changing the priority of the task:
4223 * To be able to change p->policy safely, the appropriate
4224 * runqueue lock must be held.
4226 rq = task_rq_lock(p, &rf);
4227 update_rq_clock(rq);
4230 * Changing the policy of the stop threads its a very bad idea:
4232 if (p == rq->stop) {
4233 task_rq_unlock(rq, p, &rf);
4238 * If not changing anything there's no need to proceed further,
4239 * but store a possible modification of reset_on_fork.
4241 if (unlikely(policy == p->policy)) {
4242 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4244 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4246 if (dl_policy(policy) && dl_param_changed(p, attr))
4249 p->sched_reset_on_fork = reset_on_fork;
4250 task_rq_unlock(rq, p, &rf);
4256 #ifdef CONFIG_RT_GROUP_SCHED
4258 * Do not allow realtime tasks into groups that have no runtime
4261 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4262 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4263 !task_group_is_autogroup(task_group(p))) {
4264 task_rq_unlock(rq, p, &rf);
4269 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4270 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4271 cpumask_t *span = rq->rd->span;
4274 * Don't allow tasks with an affinity mask smaller than
4275 * the entire root_domain to become SCHED_DEADLINE. We
4276 * will also fail if there's no bandwidth available.
4278 if (!cpumask_subset(span, &p->cpus_allowed) ||
4279 rq->rd->dl_bw.bw == 0) {
4280 task_rq_unlock(rq, p, &rf);
4287 /* Re-check policy now with rq lock held: */
4288 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4289 policy = oldpolicy = -1;
4290 task_rq_unlock(rq, p, &rf);
4295 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4296 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4299 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4300 task_rq_unlock(rq, p, &rf);
4304 p->sched_reset_on_fork = reset_on_fork;
4309 * Take priority boosted tasks into account. If the new
4310 * effective priority is unchanged, we just store the new
4311 * normal parameters and do not touch the scheduler class and
4312 * the runqueue. This will be done when the task deboost
4315 new_effective_prio = rt_effective_prio(p, newprio);
4316 if (new_effective_prio == oldprio)
4317 queue_flags &= ~DEQUEUE_MOVE;
4320 queued = task_on_rq_queued(p);
4321 running = task_current(rq, p);
4323 dequeue_task(rq, p, queue_flags);
4325 put_prev_task(rq, p);
4327 prev_class = p->sched_class;
4328 __setscheduler(rq, p, attr, pi);
4332 * We enqueue to tail when the priority of a task is
4333 * increased (user space view).
4335 if (oldprio < p->prio)
4336 queue_flags |= ENQUEUE_HEAD;
4338 enqueue_task(rq, p, queue_flags);
4341 set_curr_task(rq, p);
4343 check_class_changed(rq, p, prev_class, oldprio);
4345 /* Avoid rq from going away on us: */
4347 task_rq_unlock(rq, p, &rf);
4350 rt_mutex_adjust_pi(p);
4352 /* Run balance callbacks after we've adjusted the PI chain: */
4353 balance_callback(rq);
4359 static int _sched_setscheduler(struct task_struct *p, int policy,
4360 const struct sched_param *param, bool check)
4362 struct sched_attr attr = {
4363 .sched_policy = policy,
4364 .sched_priority = param->sched_priority,
4365 .sched_nice = PRIO_TO_NICE(p->static_prio),
4368 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4369 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4370 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4371 policy &= ~SCHED_RESET_ON_FORK;
4372 attr.sched_policy = policy;
4375 return __sched_setscheduler(p, &attr, check, true);
4378 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4379 * @p: the task in question.
4380 * @policy: new policy.
4381 * @param: structure containing the new RT priority.
4383 * Return: 0 on success. An error code otherwise.
4385 * NOTE that the task may be already dead.
4387 int sched_setscheduler(struct task_struct *p, int policy,
4388 const struct sched_param *param)
4390 return _sched_setscheduler(p, policy, param, true);
4392 EXPORT_SYMBOL_GPL(sched_setscheduler);
4394 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4396 return __sched_setscheduler(p, attr, true, true);
4398 EXPORT_SYMBOL_GPL(sched_setattr);
4400 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
4402 return __sched_setscheduler(p, attr, false, true);
4406 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4407 * @p: the task in question.
4408 * @policy: new policy.
4409 * @param: structure containing the new RT priority.
4411 * Just like sched_setscheduler, only don't bother checking if the
4412 * current context has permission. For example, this is needed in
4413 * stop_machine(): we create temporary high priority worker threads,
4414 * but our caller might not have that capability.
4416 * Return: 0 on success. An error code otherwise.
4418 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4419 const struct sched_param *param)
4421 return _sched_setscheduler(p, policy, param, false);
4423 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4426 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4428 struct sched_param lparam;
4429 struct task_struct *p;
4432 if (!param || pid < 0)
4434 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4439 p = find_process_by_pid(pid);
4441 retval = sched_setscheduler(p, policy, &lparam);
4448 * Mimics kernel/events/core.c perf_copy_attr().
4450 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4455 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4458 /* Zero the full structure, so that a short copy will be nice: */
4459 memset(attr, 0, sizeof(*attr));
4461 ret = get_user(size, &uattr->size);
4465 /* Bail out on silly large: */
4466 if (size > PAGE_SIZE)
4469 /* ABI compatibility quirk: */
4471 size = SCHED_ATTR_SIZE_VER0;
4473 if (size < SCHED_ATTR_SIZE_VER0)
4477 * If we're handed a bigger struct than we know of,
4478 * ensure all the unknown bits are 0 - i.e. new
4479 * user-space does not rely on any kernel feature
4480 * extensions we dont know about yet.
4482 if (size > sizeof(*attr)) {
4483 unsigned char __user *addr;
4484 unsigned char __user *end;
4487 addr = (void __user *)uattr + sizeof(*attr);
4488 end = (void __user *)uattr + size;
4490 for (; addr < end; addr++) {
4491 ret = get_user(val, addr);
4497 size = sizeof(*attr);
4500 ret = copy_from_user(attr, uattr, size);
4505 * XXX: Do we want to be lenient like existing syscalls; or do we want
4506 * to be strict and return an error on out-of-bounds values?
4508 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4513 put_user(sizeof(*attr), &uattr->size);
4518 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4519 * @pid: the pid in question.
4520 * @policy: new policy.
4521 * @param: structure containing the new RT priority.
4523 * Return: 0 on success. An error code otherwise.
4525 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4530 return do_sched_setscheduler(pid, policy, param);
4534 * sys_sched_setparam - set/change the RT priority of a thread
4535 * @pid: the pid in question.
4536 * @param: structure containing the new RT priority.
4538 * Return: 0 on success. An error code otherwise.
4540 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4542 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4546 * sys_sched_setattr - same as above, but with extended sched_attr
4547 * @pid: the pid in question.
4548 * @uattr: structure containing the extended parameters.
4549 * @flags: for future extension.
4551 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4552 unsigned int, flags)
4554 struct sched_attr attr;
4555 struct task_struct *p;
4558 if (!uattr || pid < 0 || flags)
4561 retval = sched_copy_attr(uattr, &attr);
4565 if ((int)attr.sched_policy < 0)
4570 p = find_process_by_pid(pid);
4572 retval = sched_setattr(p, &attr);
4579 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4580 * @pid: the pid in question.
4582 * Return: On success, the policy of the thread. Otherwise, a negative error
4585 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4587 struct task_struct *p;
4595 p = find_process_by_pid(pid);
4597 retval = security_task_getscheduler(p);
4600 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4607 * sys_sched_getparam - get the RT priority of a thread
4608 * @pid: the pid in question.
4609 * @param: structure containing the RT priority.
4611 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4614 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4616 struct sched_param lp = { .sched_priority = 0 };
4617 struct task_struct *p;
4620 if (!param || pid < 0)
4624 p = find_process_by_pid(pid);
4629 retval = security_task_getscheduler(p);
4633 if (task_has_rt_policy(p))
4634 lp.sched_priority = p->rt_priority;
4638 * This one might sleep, we cannot do it with a spinlock held ...
4640 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4649 static int sched_read_attr(struct sched_attr __user *uattr,
4650 struct sched_attr *attr,
4655 if (!access_ok(VERIFY_WRITE, uattr, usize))
4659 * If we're handed a smaller struct than we know of,
4660 * ensure all the unknown bits are 0 - i.e. old
4661 * user-space does not get uncomplete information.
4663 if (usize < sizeof(*attr)) {
4664 unsigned char *addr;
4667 addr = (void *)attr + usize;
4668 end = (void *)attr + sizeof(*attr);
4670 for (; addr < end; addr++) {
4678 ret = copy_to_user(uattr, attr, attr->size);
4686 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4687 * @pid: the pid in question.
4688 * @uattr: structure containing the extended parameters.
4689 * @size: sizeof(attr) for fwd/bwd comp.
4690 * @flags: for future extension.
4692 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4693 unsigned int, size, unsigned int, flags)
4695 struct sched_attr attr = {
4696 .size = sizeof(struct sched_attr),
4698 struct task_struct *p;
4701 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4702 size < SCHED_ATTR_SIZE_VER0 || flags)
4706 p = find_process_by_pid(pid);
4711 retval = security_task_getscheduler(p);
4715 attr.sched_policy = p->policy;
4716 if (p->sched_reset_on_fork)
4717 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4718 if (task_has_dl_policy(p))
4719 __getparam_dl(p, &attr);
4720 else if (task_has_rt_policy(p))
4721 attr.sched_priority = p->rt_priority;
4723 attr.sched_nice = task_nice(p);
4727 retval = sched_read_attr(uattr, &attr, size);
4735 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4737 cpumask_var_t cpus_allowed, new_mask;
4738 struct task_struct *p;
4743 p = find_process_by_pid(pid);
4749 /* Prevent p going away */
4753 if (p->flags & PF_NO_SETAFFINITY) {
4757 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4761 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4763 goto out_free_cpus_allowed;
4766 if (!check_same_owner(p)) {
4768 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4770 goto out_free_new_mask;
4775 retval = security_task_setscheduler(p);
4777 goto out_free_new_mask;
4780 cpuset_cpus_allowed(p, cpus_allowed);
4781 cpumask_and(new_mask, in_mask, cpus_allowed);
4784 * Since bandwidth control happens on root_domain basis,
4785 * if admission test is enabled, we only admit -deadline
4786 * tasks allowed to run on all the CPUs in the task's
4790 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4792 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4795 goto out_free_new_mask;
4801 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4804 cpuset_cpus_allowed(p, cpus_allowed);
4805 if (!cpumask_subset(new_mask, cpus_allowed)) {
4807 * We must have raced with a concurrent cpuset
4808 * update. Just reset the cpus_allowed to the
4809 * cpuset's cpus_allowed
4811 cpumask_copy(new_mask, cpus_allowed);
4816 free_cpumask_var(new_mask);
4817 out_free_cpus_allowed:
4818 free_cpumask_var(cpus_allowed);
4824 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4825 struct cpumask *new_mask)
4827 if (len < cpumask_size())
4828 cpumask_clear(new_mask);
4829 else if (len > cpumask_size())
4830 len = cpumask_size();
4832 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4836 * sys_sched_setaffinity - set the CPU affinity of a process
4837 * @pid: pid of the process
4838 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4839 * @user_mask_ptr: user-space pointer to the new CPU mask
4841 * Return: 0 on success. An error code otherwise.
4843 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4844 unsigned long __user *, user_mask_ptr)
4846 cpumask_var_t new_mask;
4849 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4852 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4854 retval = sched_setaffinity(pid, new_mask);
4855 free_cpumask_var(new_mask);
4859 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4861 struct task_struct *p;
4862 unsigned long flags;
4868 p = find_process_by_pid(pid);
4872 retval = security_task_getscheduler(p);
4876 raw_spin_lock_irqsave(&p->pi_lock, flags);
4877 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4878 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4887 * sys_sched_getaffinity - get the CPU affinity of a process
4888 * @pid: pid of the process
4889 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4890 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4892 * Return: size of CPU mask copied to user_mask_ptr on success. An
4893 * error code otherwise.
4895 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4896 unsigned long __user *, user_mask_ptr)
4901 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4903 if (len & (sizeof(unsigned long)-1))
4906 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4909 ret = sched_getaffinity(pid, mask);
4911 unsigned int retlen = min(len, cpumask_size());
4913 if (copy_to_user(user_mask_ptr, mask, retlen))
4918 free_cpumask_var(mask);
4924 * sys_sched_yield - yield the current processor to other threads.
4926 * This function yields the current CPU to other tasks. If there are no
4927 * other threads running on this CPU then this function will return.
4931 static void do_sched_yield(void)
4936 local_irq_disable();
4940 schedstat_inc(rq->yld_count);
4941 current->sched_class->yield_task(rq);
4944 * Since we are going to call schedule() anyway, there's
4945 * no need to preempt or enable interrupts:
4949 sched_preempt_enable_no_resched();
4954 SYSCALL_DEFINE0(sched_yield)
4960 #ifndef CONFIG_PREEMPT
4961 int __sched _cond_resched(void)
4963 if (should_resched(0)) {
4964 preempt_schedule_common();
4970 EXPORT_SYMBOL(_cond_resched);
4974 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4975 * call schedule, and on return reacquire the lock.
4977 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4978 * operations here to prevent schedule() from being called twice (once via
4979 * spin_unlock(), once by hand).
4981 int __cond_resched_lock(spinlock_t *lock)
4983 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4986 lockdep_assert_held(lock);
4988 if (spin_needbreak(lock) || resched) {
4991 preempt_schedule_common();
4999 EXPORT_SYMBOL(__cond_resched_lock);
5002 * yield - yield the current processor to other threads.
5004 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5006 * The scheduler is at all times free to pick the calling task as the most
5007 * eligible task to run, if removing the yield() call from your code breaks
5008 * it, its already broken.
5010 * Typical broken usage is:
5015 * where one assumes that yield() will let 'the other' process run that will
5016 * make event true. If the current task is a SCHED_FIFO task that will never
5017 * happen. Never use yield() as a progress guarantee!!
5019 * If you want to use yield() to wait for something, use wait_event().
5020 * If you want to use yield() to be 'nice' for others, use cond_resched().
5021 * If you still want to use yield(), do not!
5023 void __sched yield(void)
5025 set_current_state(TASK_RUNNING);
5028 EXPORT_SYMBOL(yield);
5031 * yield_to - yield the current processor to another thread in
5032 * your thread group, or accelerate that thread toward the
5033 * processor it's on.
5035 * @preempt: whether task preemption is allowed or not
5037 * It's the caller's job to ensure that the target task struct
5038 * can't go away on us before we can do any checks.
5041 * true (>0) if we indeed boosted the target task.
5042 * false (0) if we failed to boost the target.
5043 * -ESRCH if there's no task to yield to.
5045 int __sched yield_to(struct task_struct *p, bool preempt)
5047 struct task_struct *curr = current;
5048 struct rq *rq, *p_rq;
5049 unsigned long flags;
5052 local_irq_save(flags);
5058 * If we're the only runnable task on the rq and target rq also
5059 * has only one task, there's absolutely no point in yielding.
5061 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5066 double_rq_lock(rq, p_rq);
5067 if (task_rq(p) != p_rq) {
5068 double_rq_unlock(rq, p_rq);
5072 if (!curr->sched_class->yield_to_task)
5075 if (curr->sched_class != p->sched_class)
5078 if (task_running(p_rq, p) || p->state)
5081 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5083 schedstat_inc(rq->yld_count);
5085 * Make p's CPU reschedule; pick_next_entity takes care of
5088 if (preempt && rq != p_rq)
5093 double_rq_unlock(rq, p_rq);
5095 local_irq_restore(flags);
5102 EXPORT_SYMBOL_GPL(yield_to);
5104 int io_schedule_prepare(void)
5106 int old_iowait = current->in_iowait;
5108 current->in_iowait = 1;
5109 blk_schedule_flush_plug(current);
5114 void io_schedule_finish(int token)
5116 current->in_iowait = token;
5120 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5121 * that process accounting knows that this is a task in IO wait state.
5123 long __sched io_schedule_timeout(long timeout)
5128 token = io_schedule_prepare();
5129 ret = schedule_timeout(timeout);
5130 io_schedule_finish(token);
5134 EXPORT_SYMBOL(io_schedule_timeout);
5136 void io_schedule(void)
5140 token = io_schedule_prepare();
5142 io_schedule_finish(token);
5144 EXPORT_SYMBOL(io_schedule);
5147 * sys_sched_get_priority_max - return maximum RT priority.
5148 * @policy: scheduling class.
5150 * Return: On success, this syscall returns the maximum
5151 * rt_priority that can be used by a given scheduling class.
5152 * On failure, a negative error code is returned.
5154 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5161 ret = MAX_USER_RT_PRIO-1;
5163 case SCHED_DEADLINE:
5174 * sys_sched_get_priority_min - return minimum RT priority.
5175 * @policy: scheduling class.
5177 * Return: On success, this syscall returns the minimum
5178 * rt_priority that can be used by a given scheduling class.
5179 * On failure, a negative error code is returned.
5181 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5190 case SCHED_DEADLINE:
5199 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5201 struct task_struct *p;
5202 unsigned int time_slice;
5212 p = find_process_by_pid(pid);
5216 retval = security_task_getscheduler(p);
5220 rq = task_rq_lock(p, &rf);
5222 if (p->sched_class->get_rr_interval)
5223 time_slice = p->sched_class->get_rr_interval(rq, p);
5224 task_rq_unlock(rq, p, &rf);
5227 jiffies_to_timespec64(time_slice, t);
5236 * sys_sched_rr_get_interval - return the default timeslice of a process.
5237 * @pid: pid of the process.
5238 * @interval: userspace pointer to the timeslice value.
5240 * this syscall writes the default timeslice value of a given process
5241 * into the user-space timespec buffer. A value of '0' means infinity.
5243 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5246 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5247 struct __kernel_timespec __user *, interval)
5249 struct timespec64 t;
5250 int retval = sched_rr_get_interval(pid, &t);
5253 retval = put_timespec64(&t, interval);
5258 #ifdef CONFIG_COMPAT_32BIT_TIME
5259 COMPAT_SYSCALL_DEFINE2(sched_rr_get_interval,
5261 struct old_timespec32 __user *, interval)
5263 struct timespec64 t;
5264 int retval = sched_rr_get_interval(pid, &t);
5267 retval = put_old_timespec32(&t, interval);
5272 void sched_show_task(struct task_struct *p)
5274 unsigned long free = 0;
5277 if (!try_get_task_stack(p))
5280 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5282 if (p->state == TASK_RUNNING)
5283 printk(KERN_CONT " running task ");
5284 #ifdef CONFIG_DEBUG_STACK_USAGE
5285 free = stack_not_used(p);
5290 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5292 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5293 task_pid_nr(p), ppid,
5294 (unsigned long)task_thread_info(p)->flags);
5296 print_worker_info(KERN_INFO, p);
5297 show_stack(p, NULL);
5300 EXPORT_SYMBOL_GPL(sched_show_task);
5303 state_filter_match(unsigned long state_filter, struct task_struct *p)
5305 /* no filter, everything matches */
5309 /* filter, but doesn't match */
5310 if (!(p->state & state_filter))
5314 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5317 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5324 void show_state_filter(unsigned long state_filter)
5326 struct task_struct *g, *p;
5328 #if BITS_PER_LONG == 32
5330 " task PC stack pid father\n");
5333 " task PC stack pid father\n");
5336 for_each_process_thread(g, p) {
5338 * reset the NMI-timeout, listing all files on a slow
5339 * console might take a lot of time:
5340 * Also, reset softlockup watchdogs on all CPUs, because
5341 * another CPU might be blocked waiting for us to process
5344 touch_nmi_watchdog();
5345 touch_all_softlockup_watchdogs();
5346 if (state_filter_match(state_filter, p))
5350 #ifdef CONFIG_SCHED_DEBUG
5352 sysrq_sched_debug_show();
5356 * Only show locks if all tasks are dumped:
5359 debug_show_all_locks();
5363 * init_idle - set up an idle thread for a given CPU
5364 * @idle: task in question
5365 * @cpu: CPU the idle task belongs to
5367 * NOTE: this function does not set the idle thread's NEED_RESCHED
5368 * flag, to make booting more robust.
5370 void init_idle(struct task_struct *idle, int cpu)
5372 struct rq *rq = cpu_rq(cpu);
5373 unsigned long flags;
5375 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5376 raw_spin_lock(&rq->lock);
5378 __sched_fork(0, idle);
5379 idle->state = TASK_RUNNING;
5380 idle->se.exec_start = sched_clock();
5381 idle->flags |= PF_IDLE;
5383 kasan_unpoison_task_stack(idle);
5387 * Its possible that init_idle() gets called multiple times on a task,
5388 * in that case do_set_cpus_allowed() will not do the right thing.
5390 * And since this is boot we can forgo the serialization.
5392 set_cpus_allowed_common(idle, cpumask_of(cpu));
5395 * We're having a chicken and egg problem, even though we are
5396 * holding rq->lock, the CPU isn't yet set to this CPU so the
5397 * lockdep check in task_group() will fail.
5399 * Similar case to sched_fork(). / Alternatively we could
5400 * use task_rq_lock() here and obtain the other rq->lock.
5405 __set_task_cpu(idle, cpu);
5408 rq->curr = rq->idle = idle;
5409 idle->on_rq = TASK_ON_RQ_QUEUED;
5413 raw_spin_unlock(&rq->lock);
5414 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5416 /* Set the preempt count _outside_ the spinlocks! */
5417 init_idle_preempt_count(idle, cpu);
5420 * The idle tasks have their own, simple scheduling class:
5422 idle->sched_class = &idle_sched_class;
5423 ftrace_graph_init_idle_task(idle, cpu);
5424 vtime_init_idle(idle, cpu);
5426 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5432 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5433 const struct cpumask *trial)
5437 if (!cpumask_weight(cur))
5440 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5445 int task_can_attach(struct task_struct *p,
5446 const struct cpumask *cs_cpus_allowed)
5451 * Kthreads which disallow setaffinity shouldn't be moved
5452 * to a new cpuset; we don't want to change their CPU
5453 * affinity and isolating such threads by their set of
5454 * allowed nodes is unnecessary. Thus, cpusets are not
5455 * applicable for such threads. This prevents checking for
5456 * success of set_cpus_allowed_ptr() on all attached tasks
5457 * before cpus_allowed may be changed.
5459 if (p->flags & PF_NO_SETAFFINITY) {
5464 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5466 ret = dl_task_can_attach(p, cs_cpus_allowed);
5472 bool sched_smp_initialized __read_mostly;
5474 #ifdef CONFIG_NUMA_BALANCING
5475 /* Migrate current task p to target_cpu */
5476 int migrate_task_to(struct task_struct *p, int target_cpu)
5478 struct migration_arg arg = { p, target_cpu };
5479 int curr_cpu = task_cpu(p);
5481 if (curr_cpu == target_cpu)
5484 if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5487 /* TODO: This is not properly updating schedstats */
5489 trace_sched_move_numa(p, curr_cpu, target_cpu);
5490 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5494 * Requeue a task on a given node and accurately track the number of NUMA
5495 * tasks on the runqueues
5497 void sched_setnuma(struct task_struct *p, int nid)
5499 bool queued, running;
5503 rq = task_rq_lock(p, &rf);
5504 queued = task_on_rq_queued(p);
5505 running = task_current(rq, p);
5508 dequeue_task(rq, p, DEQUEUE_SAVE);
5510 put_prev_task(rq, p);
5512 p->numa_preferred_nid = nid;
5515 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5517 set_curr_task(rq, p);
5518 task_rq_unlock(rq, p, &rf);
5520 #endif /* CONFIG_NUMA_BALANCING */
5522 #ifdef CONFIG_HOTPLUG_CPU
5524 * Ensure that the idle task is using init_mm right before its CPU goes
5527 void idle_task_exit(void)
5529 struct mm_struct *mm = current->active_mm;
5531 BUG_ON(cpu_online(smp_processor_id()));
5533 if (mm != &init_mm) {
5534 switch_mm(mm, &init_mm, current);
5535 current->active_mm = &init_mm;
5536 finish_arch_post_lock_switch();
5542 * Since this CPU is going 'away' for a while, fold any nr_active delta
5543 * we might have. Assumes we're called after migrate_tasks() so that the
5544 * nr_active count is stable. We need to take the teardown thread which
5545 * is calling this into account, so we hand in adjust = 1 to the load
5548 * Also see the comment "Global load-average calculations".
5550 static void calc_load_migrate(struct rq *rq)
5552 long delta = calc_load_fold_active(rq, 1);
5554 atomic_long_add(delta, &calc_load_tasks);
5557 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5561 static const struct sched_class fake_sched_class = {
5562 .put_prev_task = put_prev_task_fake,
5565 static struct task_struct fake_task = {
5567 * Avoid pull_{rt,dl}_task()
5569 .prio = MAX_PRIO + 1,
5570 .sched_class = &fake_sched_class,
5574 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5575 * try_to_wake_up()->select_task_rq().
5577 * Called with rq->lock held even though we'er in stop_machine() and
5578 * there's no concurrency possible, we hold the required locks anyway
5579 * because of lock validation efforts.
5581 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5583 struct rq *rq = dead_rq;
5584 struct task_struct *next, *stop = rq->stop;
5585 struct rq_flags orf = *rf;
5589 * Fudge the rq selection such that the below task selection loop
5590 * doesn't get stuck on the currently eligible stop task.
5592 * We're currently inside stop_machine() and the rq is either stuck
5593 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5594 * either way we should never end up calling schedule() until we're
5600 * put_prev_task() and pick_next_task() sched
5601 * class method both need to have an up-to-date
5602 * value of rq->clock[_task]
5604 update_rq_clock(rq);
5608 * There's this thread running, bail when that's the only
5611 if (rq->nr_running == 1)
5615 * pick_next_task() assumes pinned rq->lock:
5617 next = pick_next_task(rq, &fake_task, rf);
5619 put_prev_task(rq, next);
5622 * Rules for changing task_struct::cpus_allowed are holding
5623 * both pi_lock and rq->lock, such that holding either
5624 * stabilizes the mask.
5626 * Drop rq->lock is not quite as disastrous as it usually is
5627 * because !cpu_active at this point, which means load-balance
5628 * will not interfere. Also, stop-machine.
5631 raw_spin_lock(&next->pi_lock);
5635 * Since we're inside stop-machine, _nothing_ should have
5636 * changed the task, WARN if weird stuff happened, because in
5637 * that case the above rq->lock drop is a fail too.
5639 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5640 raw_spin_unlock(&next->pi_lock);
5644 /* Find suitable destination for @next, with force if needed. */
5645 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5646 rq = __migrate_task(rq, rf, next, dest_cpu);
5647 if (rq != dead_rq) {
5653 raw_spin_unlock(&next->pi_lock);
5658 #endif /* CONFIG_HOTPLUG_CPU */
5660 void set_rq_online(struct rq *rq)
5663 const struct sched_class *class;
5665 cpumask_set_cpu(rq->cpu, rq->rd->online);
5668 for_each_class(class) {
5669 if (class->rq_online)
5670 class->rq_online(rq);
5675 void set_rq_offline(struct rq *rq)
5678 const struct sched_class *class;
5680 for_each_class(class) {
5681 if (class->rq_offline)
5682 class->rq_offline(rq);
5685 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5691 * used to mark begin/end of suspend/resume:
5693 static int num_cpus_frozen;
5696 * Update cpusets according to cpu_active mask. If cpusets are
5697 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5698 * around partition_sched_domains().
5700 * If we come here as part of a suspend/resume, don't touch cpusets because we
5701 * want to restore it back to its original state upon resume anyway.
5703 static void cpuset_cpu_active(void)
5705 if (cpuhp_tasks_frozen) {
5707 * num_cpus_frozen tracks how many CPUs are involved in suspend
5708 * resume sequence. As long as this is not the last online
5709 * operation in the resume sequence, just build a single sched
5710 * domain, ignoring cpusets.
5712 partition_sched_domains(1, NULL, NULL);
5713 if (--num_cpus_frozen)
5716 * This is the last CPU online operation. So fall through and
5717 * restore the original sched domains by considering the
5718 * cpuset configurations.
5720 cpuset_force_rebuild();
5722 cpuset_update_active_cpus();
5725 static int cpuset_cpu_inactive(unsigned int cpu)
5727 if (!cpuhp_tasks_frozen) {
5728 if (dl_cpu_busy(cpu))
5730 cpuset_update_active_cpus();
5733 partition_sched_domains(1, NULL, NULL);
5738 int sched_cpu_activate(unsigned int cpu)
5740 struct rq *rq = cpu_rq(cpu);
5743 #ifdef CONFIG_SCHED_SMT
5745 * The sched_smt_present static key needs to be evaluated on every
5746 * hotplug event because at boot time SMT might be disabled when
5747 * the number of booted CPUs is limited.
5749 * If then later a sibling gets hotplugged, then the key would stay
5750 * off and SMT scheduling would never be functional.
5752 if (cpumask_weight(cpu_smt_mask(cpu)) > 1)
5753 static_branch_enable_cpuslocked(&sched_smt_present);
5755 set_cpu_active(cpu, true);
5757 if (sched_smp_initialized) {
5758 sched_domains_numa_masks_set(cpu);
5759 cpuset_cpu_active();
5763 * Put the rq online, if not already. This happens:
5765 * 1) In the early boot process, because we build the real domains
5766 * after all CPUs have been brought up.
5768 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5771 rq_lock_irqsave(rq, &rf);
5773 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5776 rq_unlock_irqrestore(rq, &rf);
5778 update_max_interval();
5783 int sched_cpu_deactivate(unsigned int cpu)
5787 set_cpu_active(cpu, false);
5789 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5790 * users of this state to go away such that all new such users will
5793 * Do sync before park smpboot threads to take care the rcu boost case.
5795 synchronize_rcu_mult(call_rcu, call_rcu_sched);
5797 if (!sched_smp_initialized)
5800 ret = cpuset_cpu_inactive(cpu);
5802 set_cpu_active(cpu, true);
5805 sched_domains_numa_masks_clear(cpu);
5809 static void sched_rq_cpu_starting(unsigned int cpu)
5811 struct rq *rq = cpu_rq(cpu);
5813 rq->calc_load_update = calc_load_update;
5814 update_max_interval();
5817 int sched_cpu_starting(unsigned int cpu)
5819 sched_rq_cpu_starting(cpu);
5820 sched_tick_start(cpu);
5824 #ifdef CONFIG_HOTPLUG_CPU
5825 int sched_cpu_dying(unsigned int cpu)
5827 struct rq *rq = cpu_rq(cpu);
5830 /* Handle pending wakeups and then migrate everything off */
5831 sched_ttwu_pending();
5832 sched_tick_stop(cpu);
5834 rq_lock_irqsave(rq, &rf);
5836 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5839 migrate_tasks(rq, &rf);
5840 BUG_ON(rq->nr_running != 1);
5841 rq_unlock_irqrestore(rq, &rf);
5843 calc_load_migrate(rq);
5844 update_max_interval();
5845 nohz_balance_exit_idle(rq);
5851 void __init sched_init_smp(void)
5856 * There's no userspace yet to cause hotplug operations; hence all the
5857 * CPU masks are stable and all blatant races in the below code cannot
5860 mutex_lock(&sched_domains_mutex);
5861 sched_init_domains(cpu_active_mask);
5862 mutex_unlock(&sched_domains_mutex);
5864 /* Move init over to a non-isolated CPU */
5865 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
5867 sched_init_granularity();
5869 init_sched_rt_class();
5870 init_sched_dl_class();
5872 sched_smp_initialized = true;
5875 static int __init migration_init(void)
5877 sched_rq_cpu_starting(smp_processor_id());
5880 early_initcall(migration_init);
5883 void __init sched_init_smp(void)
5885 sched_init_granularity();
5887 #endif /* CONFIG_SMP */
5889 int in_sched_functions(unsigned long addr)
5891 return in_lock_functions(addr) ||
5892 (addr >= (unsigned long)__sched_text_start
5893 && addr < (unsigned long)__sched_text_end);
5896 #ifdef CONFIG_CGROUP_SCHED
5898 * Default task group.
5899 * Every task in system belongs to this group at bootup.
5901 struct task_group root_task_group;
5902 LIST_HEAD(task_groups);
5904 /* Cacheline aligned slab cache for task_group */
5905 static struct kmem_cache *task_group_cache __read_mostly;
5908 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5909 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5911 void __init sched_init(void)
5914 unsigned long alloc_size = 0, ptr;
5918 #ifdef CONFIG_FAIR_GROUP_SCHED
5919 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5921 #ifdef CONFIG_RT_GROUP_SCHED
5922 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5925 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5927 #ifdef CONFIG_FAIR_GROUP_SCHED
5928 root_task_group.se = (struct sched_entity **)ptr;
5929 ptr += nr_cpu_ids * sizeof(void **);
5931 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5932 ptr += nr_cpu_ids * sizeof(void **);
5934 #endif /* CONFIG_FAIR_GROUP_SCHED */
5935 #ifdef CONFIG_RT_GROUP_SCHED
5936 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5937 ptr += nr_cpu_ids * sizeof(void **);
5939 root_task_group.rt_rq = (struct rt_rq **)ptr;
5940 ptr += nr_cpu_ids * sizeof(void **);
5942 #endif /* CONFIG_RT_GROUP_SCHED */
5944 #ifdef CONFIG_CPUMASK_OFFSTACK
5945 for_each_possible_cpu(i) {
5946 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5947 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5948 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
5949 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5951 #endif /* CONFIG_CPUMASK_OFFSTACK */
5953 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
5954 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
5957 init_defrootdomain();
5960 #ifdef CONFIG_RT_GROUP_SCHED
5961 init_rt_bandwidth(&root_task_group.rt_bandwidth,
5962 global_rt_period(), global_rt_runtime());
5963 #endif /* CONFIG_RT_GROUP_SCHED */
5965 #ifdef CONFIG_CGROUP_SCHED
5966 task_group_cache = KMEM_CACHE(task_group, 0);
5968 list_add(&root_task_group.list, &task_groups);
5969 INIT_LIST_HEAD(&root_task_group.children);
5970 INIT_LIST_HEAD(&root_task_group.siblings);
5971 autogroup_init(&init_task);
5972 #endif /* CONFIG_CGROUP_SCHED */
5974 for_each_possible_cpu(i) {
5978 raw_spin_lock_init(&rq->lock);
5980 rq->calc_load_active = 0;
5981 rq->calc_load_update = jiffies + LOAD_FREQ;
5982 init_cfs_rq(&rq->cfs);
5983 init_rt_rq(&rq->rt);
5984 init_dl_rq(&rq->dl);
5985 #ifdef CONFIG_FAIR_GROUP_SCHED
5986 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
5987 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
5988 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
5990 * How much CPU bandwidth does root_task_group get?
5992 * In case of task-groups formed thr' the cgroup filesystem, it
5993 * gets 100% of the CPU resources in the system. This overall
5994 * system CPU resource is divided among the tasks of
5995 * root_task_group and its child task-groups in a fair manner,
5996 * based on each entity's (task or task-group's) weight
5997 * (se->load.weight).
5999 * In other words, if root_task_group has 10 tasks of weight
6000 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6001 * then A0's share of the CPU resource is:
6003 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6005 * We achieve this by letting root_task_group's tasks sit
6006 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6008 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6009 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6010 #endif /* CONFIG_FAIR_GROUP_SCHED */
6012 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6013 #ifdef CONFIG_RT_GROUP_SCHED
6014 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6017 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6018 rq->cpu_load[j] = 0;
6023 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6024 rq->balance_callback = NULL;
6025 rq->active_balance = 0;
6026 rq->next_balance = jiffies;
6031 rq->avg_idle = 2*sysctl_sched_migration_cost;
6032 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6034 INIT_LIST_HEAD(&rq->cfs_tasks);
6036 rq_attach_root(rq, &def_root_domain);
6037 #ifdef CONFIG_NO_HZ_COMMON
6038 rq->last_load_update_tick = jiffies;
6039 rq->last_blocked_load_update_tick = jiffies;
6040 atomic_set(&rq->nohz_flags, 0);
6042 #endif /* CONFIG_SMP */
6044 atomic_set(&rq->nr_iowait, 0);
6047 set_load_weight(&init_task, false);
6050 * The boot idle thread does lazy MMU switching as well:
6053 enter_lazy_tlb(&init_mm, current);
6056 * Make us the idle thread. Technically, schedule() should not be
6057 * called from this thread, however somewhere below it might be,
6058 * but because we are the idle thread, we just pick up running again
6059 * when this runqueue becomes "idle".
6061 init_idle(current, smp_processor_id());
6063 calc_load_update = jiffies + LOAD_FREQ;
6066 idle_thread_set_boot_cpu();
6068 init_sched_fair_class();
6072 scheduler_running = 1;
6075 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6076 static inline int preempt_count_equals(int preempt_offset)
6078 int nested = preempt_count() + rcu_preempt_depth();
6080 return (nested == preempt_offset);
6083 void __might_sleep(const char *file, int line, int preempt_offset)
6086 * Blocking primitives will set (and therefore destroy) current->state,
6087 * since we will exit with TASK_RUNNING make sure we enter with it,
6088 * otherwise we will destroy state.
6090 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6091 "do not call blocking ops when !TASK_RUNNING; "
6092 "state=%lx set at [<%p>] %pS\n",
6094 (void *)current->task_state_change,
6095 (void *)current->task_state_change);
6097 ___might_sleep(file, line, preempt_offset);
6099 EXPORT_SYMBOL(__might_sleep);
6101 void ___might_sleep(const char *file, int line, int preempt_offset)
6103 /* Ratelimiting timestamp: */
6104 static unsigned long prev_jiffy;
6106 unsigned long preempt_disable_ip;
6108 /* WARN_ON_ONCE() by default, no rate limit required: */
6111 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6112 !is_idle_task(current)) ||
6113 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6117 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6119 prev_jiffy = jiffies;
6121 /* Save this before calling printk(), since that will clobber it: */
6122 preempt_disable_ip = get_preempt_disable_ip(current);
6125 "BUG: sleeping function called from invalid context at %s:%d\n",
6128 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6129 in_atomic(), irqs_disabled(),
6130 current->pid, current->comm);
6132 if (task_stack_end_corrupted(current))
6133 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6135 debug_show_held_locks(current);
6136 if (irqs_disabled())
6137 print_irqtrace_events(current);
6138 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6139 && !preempt_count_equals(preempt_offset)) {
6140 pr_err("Preemption disabled at:");
6141 print_ip_sym(preempt_disable_ip);
6145 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6147 EXPORT_SYMBOL(___might_sleep);
6150 #ifdef CONFIG_MAGIC_SYSRQ
6151 void normalize_rt_tasks(void)
6153 struct task_struct *g, *p;
6154 struct sched_attr attr = {
6155 .sched_policy = SCHED_NORMAL,
6158 read_lock(&tasklist_lock);
6159 for_each_process_thread(g, p) {
6161 * Only normalize user tasks:
6163 if (p->flags & PF_KTHREAD)
6166 p->se.exec_start = 0;
6167 schedstat_set(p->se.statistics.wait_start, 0);
6168 schedstat_set(p->se.statistics.sleep_start, 0);
6169 schedstat_set(p->se.statistics.block_start, 0);
6171 if (!dl_task(p) && !rt_task(p)) {
6173 * Renice negative nice level userspace
6176 if (task_nice(p) < 0)
6177 set_user_nice(p, 0);
6181 __sched_setscheduler(p, &attr, false, false);
6183 read_unlock(&tasklist_lock);
6186 #endif /* CONFIG_MAGIC_SYSRQ */
6188 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6190 * These functions are only useful for the IA64 MCA handling, or kdb.
6192 * They can only be called when the whole system has been
6193 * stopped - every CPU needs to be quiescent, and no scheduling
6194 * activity can take place. Using them for anything else would
6195 * be a serious bug, and as a result, they aren't even visible
6196 * under any other configuration.
6200 * curr_task - return the current task for a given CPU.
6201 * @cpu: the processor in question.
6203 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6205 * Return: The current task for @cpu.
6207 struct task_struct *curr_task(int cpu)
6209 return cpu_curr(cpu);
6212 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6216 * set_curr_task - set the current task for a given CPU.
6217 * @cpu: the processor in question.
6218 * @p: the task pointer to set.
6220 * Description: This function must only be used when non-maskable interrupts
6221 * are serviced on a separate stack. It allows the architecture to switch the
6222 * notion of the current task on a CPU in a non-blocking manner. This function
6223 * must be called with all CPU's synchronized, and interrupts disabled, the
6224 * and caller must save the original value of the current task (see
6225 * curr_task() above) and restore that value before reenabling interrupts and
6226 * re-starting the system.
6228 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6230 void ia64_set_curr_task(int cpu, struct task_struct *p)
6237 #ifdef CONFIG_CGROUP_SCHED
6238 /* task_group_lock serializes the addition/removal of task groups */
6239 static DEFINE_SPINLOCK(task_group_lock);
6241 static void sched_free_group(struct task_group *tg)
6243 free_fair_sched_group(tg);
6244 free_rt_sched_group(tg);
6246 kmem_cache_free(task_group_cache, tg);
6249 /* allocate runqueue etc for a new task group */
6250 struct task_group *sched_create_group(struct task_group *parent)
6252 struct task_group *tg;
6254 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6256 return ERR_PTR(-ENOMEM);
6258 if (!alloc_fair_sched_group(tg, parent))
6261 if (!alloc_rt_sched_group(tg, parent))
6267 sched_free_group(tg);
6268 return ERR_PTR(-ENOMEM);
6271 void sched_online_group(struct task_group *tg, struct task_group *parent)
6273 unsigned long flags;
6275 spin_lock_irqsave(&task_group_lock, flags);
6276 list_add_rcu(&tg->list, &task_groups);
6278 /* Root should already exist: */
6281 tg->parent = parent;
6282 INIT_LIST_HEAD(&tg->children);
6283 list_add_rcu(&tg->siblings, &parent->children);
6284 spin_unlock_irqrestore(&task_group_lock, flags);
6286 online_fair_sched_group(tg);
6289 /* rcu callback to free various structures associated with a task group */
6290 static void sched_free_group_rcu(struct rcu_head *rhp)
6292 /* Now it should be safe to free those cfs_rqs: */
6293 sched_free_group(container_of(rhp, struct task_group, rcu));
6296 void sched_destroy_group(struct task_group *tg)
6298 /* Wait for possible concurrent references to cfs_rqs complete: */
6299 call_rcu(&tg->rcu, sched_free_group_rcu);
6302 void sched_offline_group(struct task_group *tg)
6304 unsigned long flags;
6306 /* End participation in shares distribution: */
6307 unregister_fair_sched_group(tg);
6309 spin_lock_irqsave(&task_group_lock, flags);
6310 list_del_rcu(&tg->list);
6311 list_del_rcu(&tg->siblings);
6312 spin_unlock_irqrestore(&task_group_lock, flags);
6315 static void sched_change_group(struct task_struct *tsk, int type)
6317 struct task_group *tg;
6320 * All callers are synchronized by task_rq_lock(); we do not use RCU
6321 * which is pointless here. Thus, we pass "true" to task_css_check()
6322 * to prevent lockdep warnings.
6324 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6325 struct task_group, css);
6326 tg = autogroup_task_group(tsk, tg);
6327 tsk->sched_task_group = tg;
6329 #ifdef CONFIG_FAIR_GROUP_SCHED
6330 if (tsk->sched_class->task_change_group)
6331 tsk->sched_class->task_change_group(tsk, type);
6334 set_task_rq(tsk, task_cpu(tsk));
6338 * Change task's runqueue when it moves between groups.
6340 * The caller of this function should have put the task in its new group by
6341 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6344 void sched_move_task(struct task_struct *tsk)
6346 int queued, running, queue_flags =
6347 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6351 rq = task_rq_lock(tsk, &rf);
6352 update_rq_clock(rq);
6354 running = task_current(rq, tsk);
6355 queued = task_on_rq_queued(tsk);
6358 dequeue_task(rq, tsk, queue_flags);
6360 put_prev_task(rq, tsk);
6362 sched_change_group(tsk, TASK_MOVE_GROUP);
6365 enqueue_task(rq, tsk, queue_flags);
6367 set_curr_task(rq, tsk);
6369 task_rq_unlock(rq, tsk, &rf);
6372 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6374 return css ? container_of(css, struct task_group, css) : NULL;
6377 static struct cgroup_subsys_state *
6378 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6380 struct task_group *parent = css_tg(parent_css);
6381 struct task_group *tg;
6384 /* This is early initialization for the top cgroup */
6385 return &root_task_group.css;
6388 tg = sched_create_group(parent);
6390 return ERR_PTR(-ENOMEM);
6395 /* Expose task group only after completing cgroup initialization */
6396 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6398 struct task_group *tg = css_tg(css);
6399 struct task_group *parent = css_tg(css->parent);
6402 sched_online_group(tg, parent);
6406 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6408 struct task_group *tg = css_tg(css);
6410 sched_offline_group(tg);
6413 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6415 struct task_group *tg = css_tg(css);
6418 * Relies on the RCU grace period between css_released() and this.
6420 sched_free_group(tg);
6424 * This is called before wake_up_new_task(), therefore we really only
6425 * have to set its group bits, all the other stuff does not apply.
6427 static void cpu_cgroup_fork(struct task_struct *task)
6432 rq = task_rq_lock(task, &rf);
6434 update_rq_clock(rq);
6435 sched_change_group(task, TASK_SET_GROUP);
6437 task_rq_unlock(rq, task, &rf);
6440 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6442 struct task_struct *task;
6443 struct cgroup_subsys_state *css;
6446 cgroup_taskset_for_each(task, css, tset) {
6447 #ifdef CONFIG_RT_GROUP_SCHED
6448 if (!sched_rt_can_attach(css_tg(css), task))
6451 /* We don't support RT-tasks being in separate groups */
6452 if (task->sched_class != &fair_sched_class)
6456 * Serialize against wake_up_new_task() such that if its
6457 * running, we're sure to observe its full state.
6459 raw_spin_lock_irq(&task->pi_lock);
6461 * Avoid calling sched_move_task() before wake_up_new_task()
6462 * has happened. This would lead to problems with PELT, due to
6463 * move wanting to detach+attach while we're not attached yet.
6465 if (task->state == TASK_NEW)
6467 raw_spin_unlock_irq(&task->pi_lock);
6475 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6477 struct task_struct *task;
6478 struct cgroup_subsys_state *css;
6480 cgroup_taskset_for_each(task, css, tset)
6481 sched_move_task(task);
6484 #ifdef CONFIG_FAIR_GROUP_SCHED
6485 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6486 struct cftype *cftype, u64 shareval)
6488 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6491 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6494 struct task_group *tg = css_tg(css);
6496 return (u64) scale_load_down(tg->shares);
6499 #ifdef CONFIG_CFS_BANDWIDTH
6500 static DEFINE_MUTEX(cfs_constraints_mutex);
6502 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6503 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6505 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6507 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6509 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6510 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6512 if (tg == &root_task_group)
6516 * Ensure we have at some amount of bandwidth every period. This is
6517 * to prevent reaching a state of large arrears when throttled via
6518 * entity_tick() resulting in prolonged exit starvation.
6520 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6524 * Likewise, bound things on the otherside by preventing insane quota
6525 * periods. This also allows us to normalize in computing quota
6528 if (period > max_cfs_quota_period)
6532 * Prevent race between setting of cfs_rq->runtime_enabled and
6533 * unthrottle_offline_cfs_rqs().
6536 mutex_lock(&cfs_constraints_mutex);
6537 ret = __cfs_schedulable(tg, period, quota);
6541 runtime_enabled = quota != RUNTIME_INF;
6542 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6544 * If we need to toggle cfs_bandwidth_used, off->on must occur
6545 * before making related changes, and on->off must occur afterwards
6547 if (runtime_enabled && !runtime_was_enabled)
6548 cfs_bandwidth_usage_inc();
6549 raw_spin_lock_irq(&cfs_b->lock);
6550 cfs_b->period = ns_to_ktime(period);
6551 cfs_b->quota = quota;
6553 __refill_cfs_bandwidth_runtime(cfs_b);
6555 /* Restart the period timer (if active) to handle new period expiry: */
6556 if (runtime_enabled)
6557 start_cfs_bandwidth(cfs_b);
6559 raw_spin_unlock_irq(&cfs_b->lock);
6561 for_each_online_cpu(i) {
6562 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6563 struct rq *rq = cfs_rq->rq;
6566 rq_lock_irq(rq, &rf);
6567 cfs_rq->runtime_enabled = runtime_enabled;
6568 cfs_rq->runtime_remaining = 0;
6570 if (cfs_rq->throttled)
6571 unthrottle_cfs_rq(cfs_rq);
6572 rq_unlock_irq(rq, &rf);
6574 if (runtime_was_enabled && !runtime_enabled)
6575 cfs_bandwidth_usage_dec();
6577 mutex_unlock(&cfs_constraints_mutex);
6583 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
6587 period = ktime_to_ns(tg->cfs_bandwidth.period);
6588 if (cfs_quota_us < 0)
6589 quota = RUNTIME_INF;
6591 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
6593 return tg_set_cfs_bandwidth(tg, period, quota);
6596 long tg_get_cfs_quota(struct task_group *tg)
6600 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
6603 quota_us = tg->cfs_bandwidth.quota;
6604 do_div(quota_us, NSEC_PER_USEC);
6609 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
6613 period = (u64)cfs_period_us * NSEC_PER_USEC;
6614 quota = tg->cfs_bandwidth.quota;
6616 return tg_set_cfs_bandwidth(tg, period, quota);
6619 long tg_get_cfs_period(struct task_group *tg)
6623 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
6624 do_div(cfs_period_us, NSEC_PER_USEC);
6626 return cfs_period_us;
6629 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
6632 return tg_get_cfs_quota(css_tg(css));
6635 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
6636 struct cftype *cftype, s64 cfs_quota_us)
6638 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
6641 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
6644 return tg_get_cfs_period(css_tg(css));
6647 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
6648 struct cftype *cftype, u64 cfs_period_us)
6650 return tg_set_cfs_period(css_tg(css), cfs_period_us);
6653 struct cfs_schedulable_data {
6654 struct task_group *tg;
6659 * normalize group quota/period to be quota/max_period
6660 * note: units are usecs
6662 static u64 normalize_cfs_quota(struct task_group *tg,
6663 struct cfs_schedulable_data *d)
6671 period = tg_get_cfs_period(tg);
6672 quota = tg_get_cfs_quota(tg);
6675 /* note: these should typically be equivalent */
6676 if (quota == RUNTIME_INF || quota == -1)
6679 return to_ratio(period, quota);
6682 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
6684 struct cfs_schedulable_data *d = data;
6685 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6686 s64 quota = 0, parent_quota = -1;
6689 quota = RUNTIME_INF;
6691 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
6693 quota = normalize_cfs_quota(tg, d);
6694 parent_quota = parent_b->hierarchical_quota;
6697 * Ensure max(child_quota) <= parent_quota. On cgroup2,
6698 * always take the min. On cgroup1, only inherit when no
6701 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
6702 quota = min(quota, parent_quota);
6704 if (quota == RUNTIME_INF)
6705 quota = parent_quota;
6706 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
6710 cfs_b->hierarchical_quota = quota;
6715 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
6718 struct cfs_schedulable_data data = {
6724 if (quota != RUNTIME_INF) {
6725 do_div(data.period, NSEC_PER_USEC);
6726 do_div(data.quota, NSEC_PER_USEC);
6730 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
6736 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
6738 struct task_group *tg = css_tg(seq_css(sf));
6739 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6741 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
6742 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
6743 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
6745 if (schedstat_enabled() && tg != &root_task_group) {
6749 for_each_possible_cpu(i)
6750 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
6752 seq_printf(sf, "wait_sum %llu\n", ws);
6757 #endif /* CONFIG_CFS_BANDWIDTH */
6758 #endif /* CONFIG_FAIR_GROUP_SCHED */
6760 #ifdef CONFIG_RT_GROUP_SCHED
6761 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
6762 struct cftype *cft, s64 val)
6764 return sched_group_set_rt_runtime(css_tg(css), val);
6767 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
6770 return sched_group_rt_runtime(css_tg(css));
6773 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
6774 struct cftype *cftype, u64 rt_period_us)
6776 return sched_group_set_rt_period(css_tg(css), rt_period_us);
6779 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
6782 return sched_group_rt_period(css_tg(css));
6784 #endif /* CONFIG_RT_GROUP_SCHED */
6786 static struct cftype cpu_legacy_files[] = {
6787 #ifdef CONFIG_FAIR_GROUP_SCHED
6790 .read_u64 = cpu_shares_read_u64,
6791 .write_u64 = cpu_shares_write_u64,
6794 #ifdef CONFIG_CFS_BANDWIDTH
6796 .name = "cfs_quota_us",
6797 .read_s64 = cpu_cfs_quota_read_s64,
6798 .write_s64 = cpu_cfs_quota_write_s64,
6801 .name = "cfs_period_us",
6802 .read_u64 = cpu_cfs_period_read_u64,
6803 .write_u64 = cpu_cfs_period_write_u64,
6807 .seq_show = cpu_cfs_stat_show,
6810 #ifdef CONFIG_RT_GROUP_SCHED
6812 .name = "rt_runtime_us",
6813 .read_s64 = cpu_rt_runtime_read,
6814 .write_s64 = cpu_rt_runtime_write,
6817 .name = "rt_period_us",
6818 .read_u64 = cpu_rt_period_read_uint,
6819 .write_u64 = cpu_rt_period_write_uint,
6825 static int cpu_extra_stat_show(struct seq_file *sf,
6826 struct cgroup_subsys_state *css)
6828 #ifdef CONFIG_CFS_BANDWIDTH
6830 struct task_group *tg = css_tg(css);
6831 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6834 throttled_usec = cfs_b->throttled_time;
6835 do_div(throttled_usec, NSEC_PER_USEC);
6837 seq_printf(sf, "nr_periods %d\n"
6839 "throttled_usec %llu\n",
6840 cfs_b->nr_periods, cfs_b->nr_throttled,
6847 #ifdef CONFIG_FAIR_GROUP_SCHED
6848 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
6851 struct task_group *tg = css_tg(css);
6852 u64 weight = scale_load_down(tg->shares);
6854 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
6857 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
6858 struct cftype *cft, u64 weight)
6861 * cgroup weight knobs should use the common MIN, DFL and MAX
6862 * values which are 1, 100 and 10000 respectively. While it loses
6863 * a bit of range on both ends, it maps pretty well onto the shares
6864 * value used by scheduler and the round-trip conversions preserve
6865 * the original value over the entire range.
6867 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
6870 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
6872 return sched_group_set_shares(css_tg(css), scale_load(weight));
6875 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
6878 unsigned long weight = scale_load_down(css_tg(css)->shares);
6879 int last_delta = INT_MAX;
6882 /* find the closest nice value to the current weight */
6883 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
6884 delta = abs(sched_prio_to_weight[prio] - weight);
6885 if (delta >= last_delta)
6890 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
6893 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
6894 struct cftype *cft, s64 nice)
6896 unsigned long weight;
6899 if (nice < MIN_NICE || nice > MAX_NICE)
6902 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
6903 idx = array_index_nospec(idx, 40);
6904 weight = sched_prio_to_weight[idx];
6906 return sched_group_set_shares(css_tg(css), scale_load(weight));
6910 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
6911 long period, long quota)
6914 seq_puts(sf, "max");
6916 seq_printf(sf, "%ld", quota);
6918 seq_printf(sf, " %ld\n", period);
6921 /* caller should put the current value in *@periodp before calling */
6922 static int __maybe_unused cpu_period_quota_parse(char *buf,
6923 u64 *periodp, u64 *quotap)
6925 char tok[21]; /* U64_MAX */
6927 if (!sscanf(buf, "%s %llu", tok, periodp))
6930 *periodp *= NSEC_PER_USEC;
6932 if (sscanf(tok, "%llu", quotap))
6933 *quotap *= NSEC_PER_USEC;
6934 else if (!strcmp(tok, "max"))
6935 *quotap = RUNTIME_INF;
6942 #ifdef CONFIG_CFS_BANDWIDTH
6943 static int cpu_max_show(struct seq_file *sf, void *v)
6945 struct task_group *tg = css_tg(seq_css(sf));
6947 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
6951 static ssize_t cpu_max_write(struct kernfs_open_file *of,
6952 char *buf, size_t nbytes, loff_t off)
6954 struct task_group *tg = css_tg(of_css(of));
6955 u64 period = tg_get_cfs_period(tg);
6959 ret = cpu_period_quota_parse(buf, &period, "a);
6961 ret = tg_set_cfs_bandwidth(tg, period, quota);
6962 return ret ?: nbytes;
6966 static struct cftype cpu_files[] = {
6967 #ifdef CONFIG_FAIR_GROUP_SCHED
6970 .flags = CFTYPE_NOT_ON_ROOT,
6971 .read_u64 = cpu_weight_read_u64,
6972 .write_u64 = cpu_weight_write_u64,
6975 .name = "weight.nice",
6976 .flags = CFTYPE_NOT_ON_ROOT,
6977 .read_s64 = cpu_weight_nice_read_s64,
6978 .write_s64 = cpu_weight_nice_write_s64,
6981 #ifdef CONFIG_CFS_BANDWIDTH
6984 .flags = CFTYPE_NOT_ON_ROOT,
6985 .seq_show = cpu_max_show,
6986 .write = cpu_max_write,
6992 struct cgroup_subsys cpu_cgrp_subsys = {
6993 .css_alloc = cpu_cgroup_css_alloc,
6994 .css_online = cpu_cgroup_css_online,
6995 .css_released = cpu_cgroup_css_released,
6996 .css_free = cpu_cgroup_css_free,
6997 .css_extra_stat_show = cpu_extra_stat_show,
6998 .fork = cpu_cgroup_fork,
6999 .can_attach = cpu_cgroup_can_attach,
7000 .attach = cpu_cgroup_attach,
7001 .legacy_cftypes = cpu_legacy_files,
7002 .dfl_cftypes = cpu_files,
7007 #endif /* CONFIG_CGROUP_SCHED */
7009 void dump_cpu_task(int cpu)
7011 pr_info("Task dump for CPU %d:\n", cpu);
7012 sched_show_task(cpu_curr(cpu));
7016 * Nice levels are multiplicative, with a gentle 10% change for every
7017 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7018 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7019 * that remained on nice 0.
7021 * The "10% effect" is relative and cumulative: from _any_ nice level,
7022 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7023 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7024 * If a task goes up by ~10% and another task goes down by ~10% then
7025 * the relative distance between them is ~25%.)
7027 const int sched_prio_to_weight[40] = {
7028 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7029 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7030 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7031 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7032 /* 0 */ 1024, 820, 655, 526, 423,
7033 /* 5 */ 335, 272, 215, 172, 137,
7034 /* 10 */ 110, 87, 70, 56, 45,
7035 /* 15 */ 36, 29, 23, 18, 15,
7039 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7041 * In cases where the weight does not change often, we can use the
7042 * precalculated inverse to speed up arithmetics by turning divisions
7043 * into multiplications:
7045 const u32 sched_prio_to_wmult[40] = {
7046 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7047 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7048 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7049 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7050 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7051 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7052 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7053 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7056 #undef CREATE_TRACE_POINTS