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 #ifdef CONFIG_SCHED_DEBUG
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 address
114 * dependency headed by '[L] rq = task_rq()' and the acquire
115 * will pair with the WMB to ensure we then also see migrating.
117 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
121 raw_spin_unlock(&rq->lock);
122 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
124 while (unlikely(task_on_rq_migrating(p)))
130 * RQ-clock updating methods:
133 static void update_rq_clock_task(struct rq *rq, s64 delta)
136 * In theory, the compile should just see 0 here, and optimize out the call
137 * to sched_rt_avg_update. But I don't trust it...
139 s64 __maybe_unused steal = 0, irq_delta = 0;
141 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
142 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
145 * Since irq_time is only updated on {soft,}irq_exit, we might run into
146 * this case when a previous update_rq_clock() happened inside a
149 * When this happens, we stop ->clock_task and only update the
150 * prev_irq_time stamp to account for the part that fit, so that a next
151 * update will consume the rest. This ensures ->clock_task is
154 * It does however cause some slight miss-attribution of {soft,}irq
155 * time, a more accurate solution would be to update the irq_time using
156 * the current rq->clock timestamp, except that would require using
159 if (irq_delta > delta)
162 rq->prev_irq_time += irq_delta;
165 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
166 if (static_key_false((¶virt_steal_rq_enabled))) {
167 steal = paravirt_steal_clock(cpu_of(rq));
168 steal -= rq->prev_steal_time_rq;
170 if (unlikely(steal > delta))
173 rq->prev_steal_time_rq += steal;
178 rq->clock_task += delta;
180 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
181 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
182 update_irq_load_avg(rq, irq_delta + steal);
186 void update_rq_clock(struct rq *rq)
190 lockdep_assert_held(&rq->lock);
192 if (rq->clock_update_flags & RQCF_ACT_SKIP)
195 #ifdef CONFIG_SCHED_DEBUG
196 if (sched_feat(WARN_DOUBLE_CLOCK))
197 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
198 rq->clock_update_flags |= RQCF_UPDATED;
201 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
205 update_rq_clock_task(rq, delta);
209 #ifdef CONFIG_SCHED_HRTICK
211 * Use HR-timers to deliver accurate preemption points.
214 static void hrtick_clear(struct rq *rq)
216 if (hrtimer_active(&rq->hrtick_timer))
217 hrtimer_cancel(&rq->hrtick_timer);
221 * High-resolution timer tick.
222 * Runs from hardirq context with interrupts disabled.
224 static enum hrtimer_restart hrtick(struct hrtimer *timer)
226 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
229 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
233 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
236 return HRTIMER_NORESTART;
241 static void __hrtick_restart(struct rq *rq)
243 struct hrtimer *timer = &rq->hrtick_timer;
245 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
249 * called from hardirq (IPI) context
251 static void __hrtick_start(void *arg)
257 __hrtick_restart(rq);
258 rq->hrtick_csd_pending = 0;
263 * Called to set the hrtick timer state.
265 * called with rq->lock held and irqs disabled
267 void hrtick_start(struct rq *rq, u64 delay)
269 struct hrtimer *timer = &rq->hrtick_timer;
274 * Don't schedule slices shorter than 10000ns, that just
275 * doesn't make sense and can cause timer DoS.
277 delta = max_t(s64, delay, 10000LL);
278 time = ktime_add_ns(timer->base->get_time(), delta);
280 hrtimer_set_expires(timer, time);
282 if (rq == this_rq()) {
283 __hrtick_restart(rq);
284 } else if (!rq->hrtick_csd_pending) {
285 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
286 rq->hrtick_csd_pending = 1;
292 * Called to set the hrtick timer state.
294 * called with rq->lock held and irqs disabled
296 void hrtick_start(struct rq *rq, u64 delay)
299 * Don't schedule slices shorter than 10000ns, that just
300 * doesn't make sense. Rely on vruntime for fairness.
302 delay = max_t(u64, delay, 10000LL);
303 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
304 HRTIMER_MODE_REL_PINNED);
306 #endif /* CONFIG_SMP */
308 static void hrtick_rq_init(struct rq *rq)
311 rq->hrtick_csd_pending = 0;
313 rq->hrtick_csd.flags = 0;
314 rq->hrtick_csd.func = __hrtick_start;
315 rq->hrtick_csd.info = rq;
318 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
319 rq->hrtick_timer.function = hrtick;
321 #else /* CONFIG_SCHED_HRTICK */
322 static inline void hrtick_clear(struct rq *rq)
326 static inline void hrtick_rq_init(struct rq *rq)
329 #endif /* CONFIG_SCHED_HRTICK */
332 * cmpxchg based fetch_or, macro so it works for different integer types
334 #define fetch_or(ptr, mask) \
336 typeof(ptr) _ptr = (ptr); \
337 typeof(mask) _mask = (mask); \
338 typeof(*_ptr) _old, _val = *_ptr; \
341 _old = cmpxchg(_ptr, _val, _val | _mask); \
349 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
351 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
352 * this avoids any races wrt polling state changes and thereby avoids
355 static bool set_nr_and_not_polling(struct task_struct *p)
357 struct thread_info *ti = task_thread_info(p);
358 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
362 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
364 * If this returns true, then the idle task promises to call
365 * sched_ttwu_pending() and reschedule soon.
367 static bool set_nr_if_polling(struct task_struct *p)
369 struct thread_info *ti = task_thread_info(p);
370 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
373 if (!(val & _TIF_POLLING_NRFLAG))
375 if (val & _TIF_NEED_RESCHED)
377 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
386 static bool set_nr_and_not_polling(struct task_struct *p)
388 set_tsk_need_resched(p);
393 static bool set_nr_if_polling(struct task_struct *p)
400 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
402 struct wake_q_node *node = &task->wake_q;
405 * Atomically grab the task, if ->wake_q is !nil already it means
406 * its already queued (either by us or someone else) and will get the
407 * wakeup due to that.
409 * In order to ensure that a pending wakeup will observe our pending
410 * state, even in the failed case, an explicit smp_mb() must be used.
412 smp_mb__before_atomic();
413 if (cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL))
416 get_task_struct(task);
419 * The head is context local, there can be no concurrency.
422 head->lastp = &node->next;
425 void wake_up_q(struct wake_q_head *head)
427 struct wake_q_node *node = head->first;
429 while (node != WAKE_Q_TAIL) {
430 struct task_struct *task;
432 task = container_of(node, struct task_struct, wake_q);
434 /* Task can safely be re-inserted now: */
436 task->wake_q.next = NULL;
439 * wake_up_process() executes a full barrier, which pairs with
440 * the queueing in wake_q_add() so as not to miss wakeups.
442 wake_up_process(task);
443 put_task_struct(task);
448 * resched_curr - mark rq's current task 'to be rescheduled now'.
450 * On UP this means the setting of the need_resched flag, on SMP it
451 * might also involve a cross-CPU call to trigger the scheduler on
454 void resched_curr(struct rq *rq)
456 struct task_struct *curr = rq->curr;
459 lockdep_assert_held(&rq->lock);
461 if (test_tsk_need_resched(curr))
466 if (cpu == smp_processor_id()) {
467 set_tsk_need_resched(curr);
468 set_preempt_need_resched();
472 if (set_nr_and_not_polling(curr))
473 smp_send_reschedule(cpu);
475 trace_sched_wake_idle_without_ipi(cpu);
478 void resched_cpu(int cpu)
480 struct rq *rq = cpu_rq(cpu);
483 raw_spin_lock_irqsave(&rq->lock, flags);
484 if (cpu_online(cpu) || cpu == smp_processor_id())
486 raw_spin_unlock_irqrestore(&rq->lock, flags);
490 #ifdef CONFIG_NO_HZ_COMMON
492 * In the semi idle case, use the nearest busy CPU for migrating timers
493 * from an idle CPU. This is good for power-savings.
495 * We don't do similar optimization for completely idle system, as
496 * selecting an idle CPU will add more delays to the timers than intended
497 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
499 int get_nohz_timer_target(void)
501 int i, cpu = smp_processor_id();
502 struct sched_domain *sd;
504 if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
508 for_each_domain(cpu, sd) {
509 for_each_cpu(i, sched_domain_span(sd)) {
513 if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
520 if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
521 cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
528 * When add_timer_on() enqueues a timer into the timer wheel of an
529 * idle CPU then this timer might expire before the next timer event
530 * which is scheduled to wake up that CPU. In case of a completely
531 * idle system the next event might even be infinite time into the
532 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
533 * leaves the inner idle loop so the newly added timer is taken into
534 * account when the CPU goes back to idle and evaluates the timer
535 * wheel for the next timer event.
537 static void wake_up_idle_cpu(int cpu)
539 struct rq *rq = cpu_rq(cpu);
541 if (cpu == smp_processor_id())
544 if (set_nr_and_not_polling(rq->idle))
545 smp_send_reschedule(cpu);
547 trace_sched_wake_idle_without_ipi(cpu);
550 static bool wake_up_full_nohz_cpu(int cpu)
553 * We just need the target to call irq_exit() and re-evaluate
554 * the next tick. The nohz full kick at least implies that.
555 * If needed we can still optimize that later with an
558 if (cpu_is_offline(cpu))
559 return true; /* Don't try to wake offline CPUs. */
560 if (tick_nohz_full_cpu(cpu)) {
561 if (cpu != smp_processor_id() ||
562 tick_nohz_tick_stopped())
563 tick_nohz_full_kick_cpu(cpu);
571 * Wake up the specified CPU. If the CPU is going offline, it is the
572 * caller's responsibility to deal with the lost wakeup, for example,
573 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
575 void wake_up_nohz_cpu(int cpu)
577 if (!wake_up_full_nohz_cpu(cpu))
578 wake_up_idle_cpu(cpu);
581 static inline bool got_nohz_idle_kick(void)
583 int cpu = smp_processor_id();
585 if (!(atomic_read(nohz_flags(cpu)) & NOHZ_KICK_MASK))
588 if (idle_cpu(cpu) && !need_resched())
592 * We can't run Idle Load Balance on this CPU for this time so we
593 * cancel it and clear NOHZ_BALANCE_KICK
595 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
599 #else /* CONFIG_NO_HZ_COMMON */
601 static inline bool got_nohz_idle_kick(void)
606 #endif /* CONFIG_NO_HZ_COMMON */
608 #ifdef CONFIG_NO_HZ_FULL
609 bool sched_can_stop_tick(struct rq *rq)
613 /* Deadline tasks, even if single, need the tick */
614 if (rq->dl.dl_nr_running)
618 * If there are more than one RR tasks, we need the tick to effect the
619 * actual RR behaviour.
621 if (rq->rt.rr_nr_running) {
622 if (rq->rt.rr_nr_running == 1)
629 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
630 * forced preemption between FIFO tasks.
632 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
637 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
638 * if there's more than one we need the tick for involuntary
641 if (rq->nr_running > 1)
646 #endif /* CONFIG_NO_HZ_FULL */
647 #endif /* CONFIG_SMP */
649 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
650 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
652 * Iterate task_group tree rooted at *from, calling @down when first entering a
653 * node and @up when leaving it for the final time.
655 * Caller must hold rcu_lock or sufficient equivalent.
657 int walk_tg_tree_from(struct task_group *from,
658 tg_visitor down, tg_visitor up, void *data)
660 struct task_group *parent, *child;
666 ret = (*down)(parent, data);
669 list_for_each_entry_rcu(child, &parent->children, siblings) {
676 ret = (*up)(parent, data);
677 if (ret || parent == from)
681 parent = parent->parent;
688 int tg_nop(struct task_group *tg, void *data)
694 static void set_load_weight(struct task_struct *p, bool update_load)
696 int prio = p->static_prio - MAX_RT_PRIO;
697 struct load_weight *load = &p->se.load;
700 * SCHED_IDLE tasks get minimal weight:
702 if (idle_policy(p->policy)) {
703 load->weight = scale_load(WEIGHT_IDLEPRIO);
704 load->inv_weight = WMULT_IDLEPRIO;
709 * SCHED_OTHER tasks have to update their load when changing their
712 if (update_load && p->sched_class == &fair_sched_class) {
713 reweight_task(p, prio);
715 load->weight = scale_load(sched_prio_to_weight[prio]);
716 load->inv_weight = sched_prio_to_wmult[prio];
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 WRITE_ONCE(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 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1081 if (dest_cpu >= nr_cpu_ids) {
1086 do_set_cpus_allowed(p, new_mask);
1088 if (p->flags & PF_KTHREAD) {
1090 * For kernel threads that do indeed end up on online &&
1091 * !active we want to ensure they are strict per-CPU threads.
1093 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1094 !cpumask_intersects(new_mask, cpu_active_mask) &&
1095 p->nr_cpus_allowed != 1);
1098 /* Can the task run on the task's current CPU? If so, we're done */
1099 if (cpumask_test_cpu(task_cpu(p), 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 if (this_cpu == that_cpu)
1827 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1829 #endif /* CONFIG_SMP */
1831 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1833 struct rq *rq = cpu_rq(cpu);
1836 #if defined(CONFIG_SMP)
1837 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1838 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1839 ttwu_queue_remote(p, cpu, wake_flags);
1845 update_rq_clock(rq);
1846 ttwu_do_activate(rq, p, wake_flags, &rf);
1851 * Notes on Program-Order guarantees on SMP systems.
1855 * The basic program-order guarantee on SMP systems is that when a task [t]
1856 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1857 * execution on its new CPU [c1].
1859 * For migration (of runnable tasks) this is provided by the following means:
1861 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1862 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1863 * rq(c1)->lock (if not at the same time, then in that order).
1864 * C) LOCK of the rq(c1)->lock scheduling in task
1866 * Release/acquire chaining guarantees that B happens after A and C after B.
1867 * Note: the CPU doing B need not be c0 or c1
1876 * UNLOCK rq(0)->lock
1878 * LOCK rq(0)->lock // orders against CPU0
1880 * UNLOCK rq(0)->lock
1884 * UNLOCK rq(1)->lock
1886 * LOCK rq(1)->lock // orders against CPU2
1889 * UNLOCK rq(1)->lock
1892 * BLOCKING -- aka. SLEEP + WAKEUP
1894 * For blocking we (obviously) need to provide the same guarantee as for
1895 * migration. However the means are completely different as there is no lock
1896 * chain to provide order. Instead we do:
1898 * 1) smp_store_release(X->on_cpu, 0)
1899 * 2) smp_cond_load_acquire(!X->on_cpu)
1903 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1905 * LOCK rq(0)->lock LOCK X->pi_lock
1908 * smp_store_release(X->on_cpu, 0);
1910 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1916 * X->state = RUNNING
1917 * UNLOCK rq(2)->lock
1919 * LOCK rq(2)->lock // orders against CPU1
1922 * UNLOCK rq(2)->lock
1925 * UNLOCK rq(0)->lock
1928 * However, for wakeups there is a second guarantee we must provide, namely we
1929 * must ensure that CONDITION=1 done by the caller can not be reordered with
1930 * accesses to the task state; see try_to_wake_up() and set_current_state().
1934 * try_to_wake_up - wake up a thread
1935 * @p: the thread to be awakened
1936 * @state: the mask of task states that can be woken
1937 * @wake_flags: wake modifier flags (WF_*)
1939 * If (@state & @p->state) @p->state = TASK_RUNNING.
1941 * If the task was not queued/runnable, also place it back on a runqueue.
1943 * Atomic against schedule() which would dequeue a task, also see
1944 * set_current_state().
1946 * This function executes a full memory barrier before accessing the task
1947 * state; see set_current_state().
1949 * Return: %true if @p->state changes (an actual wakeup was done),
1953 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1955 unsigned long flags;
1956 int cpu, success = 0;
1959 * If we are going to wake up a thread waiting for CONDITION we
1960 * need to ensure that CONDITION=1 done by the caller can not be
1961 * reordered with p->state check below. This pairs with mb() in
1962 * set_current_state() the waiting thread does.
1964 raw_spin_lock_irqsave(&p->pi_lock, flags);
1965 smp_mb__after_spinlock();
1966 if (!(p->state & state))
1969 trace_sched_waking(p);
1971 /* We're going to change ->state: */
1976 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1977 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1978 * in smp_cond_load_acquire() below.
1980 * sched_ttwu_pending() try_to_wake_up()
1981 * STORE p->on_rq = 1 LOAD p->state
1984 * __schedule() (switch to task 'p')
1985 * LOCK rq->lock smp_rmb();
1986 * smp_mb__after_spinlock();
1990 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
1992 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
1993 * __schedule(). See the comment for smp_mb__after_spinlock().
1996 if (p->on_rq && ttwu_remote(p, wake_flags))
2001 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2002 * possible to, falsely, observe p->on_cpu == 0.
2004 * One must be running (->on_cpu == 1) in order to remove oneself
2005 * from the runqueue.
2007 * __schedule() (switch to task 'p') try_to_wake_up()
2008 * STORE p->on_cpu = 1 LOAD p->on_rq
2011 * __schedule() (put 'p' to sleep)
2012 * LOCK rq->lock smp_rmb();
2013 * smp_mb__after_spinlock();
2014 * STORE p->on_rq = 0 LOAD p->on_cpu
2016 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2017 * __schedule(). See the comment for smp_mb__after_spinlock().
2022 * If the owning (remote) CPU is still in the middle of schedule() with
2023 * this task as prev, wait until its done referencing the task.
2025 * Pairs with the smp_store_release() in finish_task().
2027 * This ensures that tasks getting woken will be fully ordered against
2028 * their previous state and preserve Program Order.
2030 smp_cond_load_acquire(&p->on_cpu, !VAL);
2032 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2033 p->state = TASK_WAKING;
2036 delayacct_blkio_end(p);
2037 atomic_dec(&task_rq(p)->nr_iowait);
2040 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2041 if (task_cpu(p) != cpu) {
2042 wake_flags |= WF_MIGRATED;
2043 set_task_cpu(p, cpu);
2046 #else /* CONFIG_SMP */
2049 delayacct_blkio_end(p);
2050 atomic_dec(&task_rq(p)->nr_iowait);
2053 #endif /* CONFIG_SMP */
2055 ttwu_queue(p, cpu, wake_flags);
2057 ttwu_stat(p, cpu, wake_flags);
2059 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2065 * try_to_wake_up_local - try to wake up a local task with rq lock held
2066 * @p: the thread to be awakened
2067 * @rf: request-queue flags for pinning
2069 * Put @p on the run-queue if it's not already there. The caller must
2070 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2073 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2075 struct rq *rq = task_rq(p);
2077 if (WARN_ON_ONCE(rq != this_rq()) ||
2078 WARN_ON_ONCE(p == current))
2081 lockdep_assert_held(&rq->lock);
2083 if (!raw_spin_trylock(&p->pi_lock)) {
2085 * This is OK, because current is on_cpu, which avoids it being
2086 * picked for load-balance and preemption/IRQs are still
2087 * disabled avoiding further scheduler activity on it and we've
2088 * not yet picked a replacement task.
2091 raw_spin_lock(&p->pi_lock);
2095 if (!(p->state & TASK_NORMAL))
2098 trace_sched_waking(p);
2100 if (!task_on_rq_queued(p)) {
2102 delayacct_blkio_end(p);
2103 atomic_dec(&rq->nr_iowait);
2105 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK);
2108 ttwu_do_wakeup(rq, p, 0, rf);
2109 ttwu_stat(p, smp_processor_id(), 0);
2111 raw_spin_unlock(&p->pi_lock);
2115 * wake_up_process - Wake up a specific process
2116 * @p: The process to be woken up.
2118 * Attempt to wake up the nominated process and move it to the set of runnable
2121 * Return: 1 if the process was woken up, 0 if it was already running.
2123 * This function executes a full memory barrier before accessing the task state.
2125 int wake_up_process(struct task_struct *p)
2127 return try_to_wake_up(p, TASK_NORMAL, 0);
2129 EXPORT_SYMBOL(wake_up_process);
2131 int wake_up_state(struct task_struct *p, unsigned int state)
2133 return try_to_wake_up(p, state, 0);
2137 * Perform scheduler related setup for a newly forked process p.
2138 * p is forked by current.
2140 * __sched_fork() is basic setup used by init_idle() too:
2142 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2147 p->se.exec_start = 0;
2148 p->se.sum_exec_runtime = 0;
2149 p->se.prev_sum_exec_runtime = 0;
2150 p->se.nr_migrations = 0;
2152 INIT_LIST_HEAD(&p->se.group_node);
2154 #ifdef CONFIG_FAIR_GROUP_SCHED
2155 p->se.cfs_rq = NULL;
2158 #ifdef CONFIG_SCHEDSTATS
2159 /* Even if schedstat is disabled, there should not be garbage */
2160 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2163 RB_CLEAR_NODE(&p->dl.rb_node);
2164 init_dl_task_timer(&p->dl);
2165 init_dl_inactive_task_timer(&p->dl);
2166 __dl_clear_params(p);
2168 INIT_LIST_HEAD(&p->rt.run_list);
2170 p->rt.time_slice = sched_rr_timeslice;
2174 #ifdef CONFIG_PREEMPT_NOTIFIERS
2175 INIT_HLIST_HEAD(&p->preempt_notifiers);
2178 init_numa_balancing(clone_flags, p);
2181 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2183 #ifdef CONFIG_NUMA_BALANCING
2185 void set_numabalancing_state(bool enabled)
2188 static_branch_enable(&sched_numa_balancing);
2190 static_branch_disable(&sched_numa_balancing);
2193 #ifdef CONFIG_PROC_SYSCTL
2194 int sysctl_numa_balancing(struct ctl_table *table, int write,
2195 void __user *buffer, size_t *lenp, loff_t *ppos)
2199 int state = static_branch_likely(&sched_numa_balancing);
2201 if (write && !capable(CAP_SYS_ADMIN))
2206 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2210 set_numabalancing_state(state);
2216 #ifdef CONFIG_SCHEDSTATS
2218 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2219 static bool __initdata __sched_schedstats = false;
2221 static void set_schedstats(bool enabled)
2224 static_branch_enable(&sched_schedstats);
2226 static_branch_disable(&sched_schedstats);
2229 void force_schedstat_enabled(void)
2231 if (!schedstat_enabled()) {
2232 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2233 static_branch_enable(&sched_schedstats);
2237 static int __init setup_schedstats(char *str)
2244 * This code is called before jump labels have been set up, so we can't
2245 * change the static branch directly just yet. Instead set a temporary
2246 * variable so init_schedstats() can do it later.
2248 if (!strcmp(str, "enable")) {
2249 __sched_schedstats = true;
2251 } else if (!strcmp(str, "disable")) {
2252 __sched_schedstats = false;
2257 pr_warn("Unable to parse schedstats=\n");
2261 __setup("schedstats=", setup_schedstats);
2263 static void __init init_schedstats(void)
2265 set_schedstats(__sched_schedstats);
2268 #ifdef CONFIG_PROC_SYSCTL
2269 int sysctl_schedstats(struct ctl_table *table, int write,
2270 void __user *buffer, size_t *lenp, loff_t *ppos)
2274 int state = static_branch_likely(&sched_schedstats);
2276 if (write && !capable(CAP_SYS_ADMIN))
2281 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2285 set_schedstats(state);
2288 #endif /* CONFIG_PROC_SYSCTL */
2289 #else /* !CONFIG_SCHEDSTATS */
2290 static inline void init_schedstats(void) {}
2291 #endif /* CONFIG_SCHEDSTATS */
2294 * fork()/clone()-time setup:
2296 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2298 unsigned long flags;
2300 __sched_fork(clone_flags, p);
2302 * We mark the process as NEW here. This guarantees that
2303 * nobody will actually run it, and a signal or other external
2304 * event cannot wake it up and insert it on the runqueue either.
2306 p->state = TASK_NEW;
2309 * Make sure we do not leak PI boosting priority to the child.
2311 p->prio = current->normal_prio;
2314 * Revert to default priority/policy on fork if requested.
2316 if (unlikely(p->sched_reset_on_fork)) {
2317 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2318 p->policy = SCHED_NORMAL;
2319 p->static_prio = NICE_TO_PRIO(0);
2321 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2322 p->static_prio = NICE_TO_PRIO(0);
2324 p->prio = p->normal_prio = __normal_prio(p);
2325 set_load_weight(p, false);
2328 * We don't need the reset flag anymore after the fork. It has
2329 * fulfilled its duty:
2331 p->sched_reset_on_fork = 0;
2334 if (dl_prio(p->prio))
2336 else if (rt_prio(p->prio))
2337 p->sched_class = &rt_sched_class;
2339 p->sched_class = &fair_sched_class;
2341 init_entity_runnable_average(&p->se);
2344 * The child is not yet in the pid-hash so no cgroup attach races,
2345 * and the cgroup is pinned to this child due to cgroup_fork()
2346 * is ran before sched_fork().
2348 * Silence PROVE_RCU.
2350 raw_spin_lock_irqsave(&p->pi_lock, flags);
2353 * We're setting the CPU for the first time, we don't migrate,
2354 * so use __set_task_cpu().
2356 __set_task_cpu(p, smp_processor_id());
2357 if (p->sched_class->task_fork)
2358 p->sched_class->task_fork(p);
2359 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2361 #ifdef CONFIG_SCHED_INFO
2362 if (likely(sched_info_on()))
2363 memset(&p->sched_info, 0, sizeof(p->sched_info));
2365 #if defined(CONFIG_SMP)
2368 init_task_preempt_count(p);
2370 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2371 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2376 unsigned long to_ratio(u64 period, u64 runtime)
2378 if (runtime == RUNTIME_INF)
2382 * Doing this here saves a lot of checks in all
2383 * the calling paths, and returning zero seems
2384 * safe for them anyway.
2389 return div64_u64(runtime << BW_SHIFT, period);
2393 * wake_up_new_task - wake up a newly created task for the first time.
2395 * This function will do some initial scheduler statistics housekeeping
2396 * that must be done for every newly created context, then puts the task
2397 * on the runqueue and wakes it.
2399 void wake_up_new_task(struct task_struct *p)
2404 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2405 p->state = TASK_RUNNING;
2408 * Fork balancing, do it here and not earlier because:
2409 * - cpus_allowed can change in the fork path
2410 * - any previously selected CPU might disappear through hotplug
2412 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2413 * as we're not fully set-up yet.
2415 p->recent_used_cpu = task_cpu(p);
2417 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2419 rq = __task_rq_lock(p, &rf);
2420 update_rq_clock(rq);
2421 post_init_entity_util_avg(&p->se);
2423 activate_task(rq, p, ENQUEUE_NOCLOCK);
2424 p->on_rq = TASK_ON_RQ_QUEUED;
2425 trace_sched_wakeup_new(p);
2426 check_preempt_curr(rq, p, WF_FORK);
2428 if (p->sched_class->task_woken) {
2430 * Nothing relies on rq->lock after this, so its fine to
2433 rq_unpin_lock(rq, &rf);
2434 p->sched_class->task_woken(rq, p);
2435 rq_repin_lock(rq, &rf);
2438 task_rq_unlock(rq, p, &rf);
2441 #ifdef CONFIG_PREEMPT_NOTIFIERS
2443 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2445 void preempt_notifier_inc(void)
2447 static_branch_inc(&preempt_notifier_key);
2449 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2451 void preempt_notifier_dec(void)
2453 static_branch_dec(&preempt_notifier_key);
2455 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2458 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2459 * @notifier: notifier struct to register
2461 void preempt_notifier_register(struct preempt_notifier *notifier)
2463 if (!static_branch_unlikely(&preempt_notifier_key))
2464 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2466 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2468 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2471 * preempt_notifier_unregister - no longer interested in preemption notifications
2472 * @notifier: notifier struct to unregister
2474 * This is *not* safe to call from within a preemption notifier.
2476 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2478 hlist_del(¬ifier->link);
2480 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2482 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2484 struct preempt_notifier *notifier;
2486 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2487 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2490 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2492 if (static_branch_unlikely(&preempt_notifier_key))
2493 __fire_sched_in_preempt_notifiers(curr);
2497 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2498 struct task_struct *next)
2500 struct preempt_notifier *notifier;
2502 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2503 notifier->ops->sched_out(notifier, next);
2506 static __always_inline void
2507 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2508 struct task_struct *next)
2510 if (static_branch_unlikely(&preempt_notifier_key))
2511 __fire_sched_out_preempt_notifiers(curr, next);
2514 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2516 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2521 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2522 struct task_struct *next)
2526 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2528 static inline void prepare_task(struct task_struct *next)
2532 * Claim the task as running, we do this before switching to it
2533 * such that any running task will have this set.
2539 static inline void finish_task(struct task_struct *prev)
2543 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2544 * We must ensure this doesn't happen until the switch is completely
2547 * In particular, the load of prev->state in finish_task_switch() must
2548 * happen before this.
2550 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2552 smp_store_release(&prev->on_cpu, 0);
2557 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
2560 * Since the runqueue lock will be released by the next
2561 * task (which is an invalid locking op but in the case
2562 * of the scheduler it's an obvious special-case), so we
2563 * do an early lockdep release here:
2565 rq_unpin_lock(rq, rf);
2566 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2567 #ifdef CONFIG_DEBUG_SPINLOCK
2568 /* this is a valid case when another task releases the spinlock */
2569 rq->lock.owner = next;
2573 static inline void finish_lock_switch(struct rq *rq)
2576 * If we are tracking spinlock dependencies then we have to
2577 * fix up the runqueue lock - which gets 'carried over' from
2578 * prev into current:
2580 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
2581 raw_spin_unlock_irq(&rq->lock);
2585 * NOP if the arch has not defined these:
2588 #ifndef prepare_arch_switch
2589 # define prepare_arch_switch(next) do { } while (0)
2592 #ifndef finish_arch_post_lock_switch
2593 # define finish_arch_post_lock_switch() do { } while (0)
2597 * prepare_task_switch - prepare to switch tasks
2598 * @rq: the runqueue preparing to switch
2599 * @prev: the current task that is being switched out
2600 * @next: the task we are going to switch to.
2602 * This is called with the rq lock held and interrupts off. It must
2603 * be paired with a subsequent finish_task_switch after the context
2606 * prepare_task_switch sets up locking and calls architecture specific
2610 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2611 struct task_struct *next)
2613 kcov_prepare_switch(prev);
2614 sched_info_switch(rq, prev, next);
2615 perf_event_task_sched_out(prev, next);
2617 fire_sched_out_preempt_notifiers(prev, next);
2619 prepare_arch_switch(next);
2623 * finish_task_switch - clean up after a task-switch
2624 * @prev: the thread we just switched away from.
2626 * finish_task_switch must be called after the context switch, paired
2627 * with a prepare_task_switch call before the context switch.
2628 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2629 * and do any other architecture-specific cleanup actions.
2631 * Note that we may have delayed dropping an mm in context_switch(). If
2632 * so, we finish that here outside of the runqueue lock. (Doing it
2633 * with the lock held can cause deadlocks; see schedule() for
2636 * The context switch have flipped the stack from under us and restored the
2637 * local variables which were saved when this task called schedule() in the
2638 * past. prev == current is still correct but we need to recalculate this_rq
2639 * because prev may have moved to another CPU.
2641 static struct rq *finish_task_switch(struct task_struct *prev)
2642 __releases(rq->lock)
2644 struct rq *rq = this_rq();
2645 struct mm_struct *mm = rq->prev_mm;
2649 * The previous task will have left us with a preempt_count of 2
2650 * because it left us after:
2653 * preempt_disable(); // 1
2655 * raw_spin_lock_irq(&rq->lock) // 2
2657 * Also, see FORK_PREEMPT_COUNT.
2659 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2660 "corrupted preempt_count: %s/%d/0x%x\n",
2661 current->comm, current->pid, preempt_count()))
2662 preempt_count_set(FORK_PREEMPT_COUNT);
2667 * A task struct has one reference for the use as "current".
2668 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2669 * schedule one last time. The schedule call will never return, and
2670 * the scheduled task must drop that reference.
2672 * We must observe prev->state before clearing prev->on_cpu (in
2673 * finish_task), otherwise a concurrent wakeup can get prev
2674 * running on another CPU and we could rave with its RUNNING -> DEAD
2675 * transition, resulting in a double drop.
2677 prev_state = prev->state;
2678 vtime_task_switch(prev);
2679 perf_event_task_sched_in(prev, current);
2681 finish_lock_switch(rq);
2682 finish_arch_post_lock_switch();
2683 kcov_finish_switch(current);
2685 fire_sched_in_preempt_notifiers(current);
2687 * When switching through a kernel thread, the loop in
2688 * membarrier_{private,global}_expedited() may have observed that
2689 * kernel thread and not issued an IPI. It is therefore possible to
2690 * schedule between user->kernel->user threads without passing though
2691 * switch_mm(). Membarrier requires a barrier after storing to
2692 * rq->curr, before returning to userspace, so provide them here:
2694 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
2695 * provided by mmdrop(),
2696 * - a sync_core for SYNC_CORE.
2699 membarrier_mm_sync_core_before_usermode(mm);
2702 if (unlikely(prev_state == TASK_DEAD)) {
2703 if (prev->sched_class->task_dead)
2704 prev->sched_class->task_dead(prev);
2707 * Remove function-return probe instances associated with this
2708 * task and put them back on the free list.
2710 kprobe_flush_task(prev);
2712 /* Task is done with its stack. */
2713 put_task_stack(prev);
2715 put_task_struct(prev);
2718 tick_nohz_task_switch();
2724 /* rq->lock is NOT held, but preemption is disabled */
2725 static void __balance_callback(struct rq *rq)
2727 struct callback_head *head, *next;
2728 void (*func)(struct rq *rq);
2729 unsigned long flags;
2731 raw_spin_lock_irqsave(&rq->lock, flags);
2732 head = rq->balance_callback;
2733 rq->balance_callback = NULL;
2735 func = (void (*)(struct rq *))head->func;
2742 raw_spin_unlock_irqrestore(&rq->lock, flags);
2745 static inline void balance_callback(struct rq *rq)
2747 if (unlikely(rq->balance_callback))
2748 __balance_callback(rq);
2753 static inline void balance_callback(struct rq *rq)
2760 * schedule_tail - first thing a freshly forked thread must call.
2761 * @prev: the thread we just switched away from.
2763 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2764 __releases(rq->lock)
2769 * New tasks start with FORK_PREEMPT_COUNT, see there and
2770 * finish_task_switch() for details.
2772 * finish_task_switch() will drop rq->lock() and lower preempt_count
2773 * and the preempt_enable() will end up enabling preemption (on
2774 * PREEMPT_COUNT kernels).
2777 rq = finish_task_switch(prev);
2778 balance_callback(rq);
2781 if (current->set_child_tid)
2782 put_user(task_pid_vnr(current), current->set_child_tid);
2784 calculate_sigpending();
2788 * context_switch - switch to the new MM and the new thread's register state.
2790 static __always_inline struct rq *
2791 context_switch(struct rq *rq, struct task_struct *prev,
2792 struct task_struct *next, struct rq_flags *rf)
2794 struct mm_struct *mm, *oldmm;
2796 prepare_task_switch(rq, prev, next);
2799 oldmm = prev->active_mm;
2801 * For paravirt, this is coupled with an exit in switch_to to
2802 * combine the page table reload and the switch backend into
2805 arch_start_context_switch(prev);
2808 * If mm is non-NULL, we pass through switch_mm(). If mm is
2809 * NULL, we will pass through mmdrop() in finish_task_switch().
2810 * Both of these contain the full memory barrier required by
2811 * membarrier after storing to rq->curr, before returning to
2815 next->active_mm = oldmm;
2817 enter_lazy_tlb(oldmm, next);
2819 switch_mm_irqs_off(oldmm, mm, next);
2822 prev->active_mm = NULL;
2823 rq->prev_mm = oldmm;
2826 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2828 prepare_lock_switch(rq, next, rf);
2830 /* Here we just switch the register state and the stack. */
2831 switch_to(prev, next, prev);
2834 return finish_task_switch(prev);
2838 * nr_running and nr_context_switches:
2840 * externally visible scheduler statistics: current number of runnable
2841 * threads, total number of context switches performed since bootup.
2843 unsigned long nr_running(void)
2845 unsigned long i, sum = 0;
2847 for_each_online_cpu(i)
2848 sum += cpu_rq(i)->nr_running;
2854 * Check if only the current task is running on the CPU.
2856 * Caution: this function does not check that the caller has disabled
2857 * preemption, thus the result might have a time-of-check-to-time-of-use
2858 * race. The caller is responsible to use it correctly, for example:
2860 * - from a non-preemptable section (of course)
2862 * - from a thread that is bound to a single CPU
2864 * - in a loop with very short iterations (e.g. a polling loop)
2866 bool single_task_running(void)
2868 return raw_rq()->nr_running == 1;
2870 EXPORT_SYMBOL(single_task_running);
2872 unsigned long long nr_context_switches(void)
2875 unsigned long long sum = 0;
2877 for_each_possible_cpu(i)
2878 sum += cpu_rq(i)->nr_switches;
2884 * IO-wait accounting, and how its mostly bollocks (on SMP).
2886 * The idea behind IO-wait account is to account the idle time that we could
2887 * have spend running if it were not for IO. That is, if we were to improve the
2888 * storage performance, we'd have a proportional reduction in IO-wait time.
2890 * This all works nicely on UP, where, when a task blocks on IO, we account
2891 * idle time as IO-wait, because if the storage were faster, it could've been
2892 * running and we'd not be idle.
2894 * This has been extended to SMP, by doing the same for each CPU. This however
2897 * Imagine for instance the case where two tasks block on one CPU, only the one
2898 * CPU will have IO-wait accounted, while the other has regular idle. Even
2899 * though, if the storage were faster, both could've ran at the same time,
2900 * utilising both CPUs.
2902 * This means, that when looking globally, the current IO-wait accounting on
2903 * SMP is a lower bound, by reason of under accounting.
2905 * Worse, since the numbers are provided per CPU, they are sometimes
2906 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2907 * associated with any one particular CPU, it can wake to another CPU than it
2908 * blocked on. This means the per CPU IO-wait number is meaningless.
2910 * Task CPU affinities can make all that even more 'interesting'.
2913 unsigned long nr_iowait(void)
2915 unsigned long i, sum = 0;
2917 for_each_possible_cpu(i)
2918 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2924 * Consumers of these two interfaces, like for example the cpufreq menu
2925 * governor are using nonsensical data. Boosting frequency for a CPU that has
2926 * IO-wait which might not even end up running the task when it does become
2930 unsigned long nr_iowait_cpu(int cpu)
2932 struct rq *this = cpu_rq(cpu);
2933 return atomic_read(&this->nr_iowait);
2936 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2938 struct rq *rq = this_rq();
2939 *nr_waiters = atomic_read(&rq->nr_iowait);
2940 *load = rq->load.weight;
2946 * sched_exec - execve() is a valuable balancing opportunity, because at
2947 * this point the task has the smallest effective memory and cache footprint.
2949 void sched_exec(void)
2951 struct task_struct *p = current;
2952 unsigned long flags;
2955 raw_spin_lock_irqsave(&p->pi_lock, flags);
2956 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2957 if (dest_cpu == smp_processor_id())
2960 if (likely(cpu_active(dest_cpu))) {
2961 struct migration_arg arg = { p, dest_cpu };
2963 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2964 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2968 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2973 DEFINE_PER_CPU(struct kernel_stat, kstat);
2974 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2976 EXPORT_PER_CPU_SYMBOL(kstat);
2977 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2980 * The function fair_sched_class.update_curr accesses the struct curr
2981 * and its field curr->exec_start; when called from task_sched_runtime(),
2982 * we observe a high rate of cache misses in practice.
2983 * Prefetching this data results in improved performance.
2985 static inline void prefetch_curr_exec_start(struct task_struct *p)
2987 #ifdef CONFIG_FAIR_GROUP_SCHED
2988 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
2990 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
2993 prefetch(&curr->exec_start);
2997 * Return accounted runtime for the task.
2998 * In case the task is currently running, return the runtime plus current's
2999 * pending runtime that have not been accounted yet.
3001 unsigned long long task_sched_runtime(struct task_struct *p)
3007 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3009 * 64-bit doesn't need locks to atomically read a 64-bit value.
3010 * So we have a optimization chance when the task's delta_exec is 0.
3011 * Reading ->on_cpu is racy, but this is ok.
3013 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3014 * If we race with it entering CPU, unaccounted time is 0. This is
3015 * indistinguishable from the read occurring a few cycles earlier.
3016 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3017 * been accounted, so we're correct here as well.
3019 if (!p->on_cpu || !task_on_rq_queued(p))
3020 return p->se.sum_exec_runtime;
3023 rq = task_rq_lock(p, &rf);
3025 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3026 * project cycles that may never be accounted to this
3027 * thread, breaking clock_gettime().
3029 if (task_current(rq, p) && task_on_rq_queued(p)) {
3030 prefetch_curr_exec_start(p);
3031 update_rq_clock(rq);
3032 p->sched_class->update_curr(rq);
3034 ns = p->se.sum_exec_runtime;
3035 task_rq_unlock(rq, p, &rf);
3041 * This function gets called by the timer code, with HZ frequency.
3042 * We call it with interrupts disabled.
3044 void scheduler_tick(void)
3046 int cpu = smp_processor_id();
3047 struct rq *rq = cpu_rq(cpu);
3048 struct task_struct *curr = rq->curr;
3055 update_rq_clock(rq);
3056 curr->sched_class->task_tick(rq, curr, 0);
3057 cpu_load_update_active(rq);
3058 calc_global_load_tick(rq);
3062 perf_event_task_tick();
3065 rq->idle_balance = idle_cpu(cpu);
3066 trigger_load_balance(rq);
3070 #ifdef CONFIG_NO_HZ_FULL
3075 struct delayed_work work;
3077 /* Values for ->state, see diagram below. */
3078 #define TICK_SCHED_REMOTE_OFFLINE 0
3079 #define TICK_SCHED_REMOTE_OFFLINING 1
3080 #define TICK_SCHED_REMOTE_RUNNING 2
3083 * State diagram for ->state:
3086 * TICK_SCHED_REMOTE_OFFLINE
3089 * | | sched_tick_remote()
3092 * +--TICK_SCHED_REMOTE_OFFLINING
3095 * sched_tick_start() | | sched_tick_stop()
3098 * TICK_SCHED_REMOTE_RUNNING
3101 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
3102 * and sched_tick_start() are happy to leave the state in RUNNING.
3105 static struct tick_work __percpu *tick_work_cpu;
3107 static void sched_tick_remote(struct work_struct *work)
3109 struct delayed_work *dwork = to_delayed_work(work);
3110 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3111 int cpu = twork->cpu;
3112 struct rq *rq = cpu_rq(cpu);
3113 struct task_struct *curr;
3119 * Handle the tick only if it appears the remote CPU is running in full
3120 * dynticks mode. The check is racy by nature, but missing a tick or
3121 * having one too much is no big deal because the scheduler tick updates
3122 * statistics and checks timeslices in a time-independent way, regardless
3123 * of when exactly it is running.
3125 if (idle_cpu(cpu) || !tick_nohz_tick_stopped_cpu(cpu))
3128 rq_lock_irq(rq, &rf);
3130 if (is_idle_task(curr) || cpu_is_offline(cpu))
3133 update_rq_clock(rq);
3134 delta = rq_clock_task(rq) - curr->se.exec_start;
3137 * Make sure the next tick runs within a reasonable
3140 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3141 curr->sched_class->task_tick(rq, curr, 0);
3144 rq_unlock_irq(rq, &rf);
3148 * Run the remote tick once per second (1Hz). This arbitrary
3149 * frequency is large enough to avoid overload but short enough
3150 * to keep scheduler internal stats reasonably up to date. But
3151 * first update state to reflect hotplug activity if required.
3153 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
3154 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
3155 if (os == TICK_SCHED_REMOTE_RUNNING)
3156 queue_delayed_work(system_unbound_wq, dwork, HZ);
3159 static void sched_tick_start(int cpu)
3162 struct tick_work *twork;
3164 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3167 WARN_ON_ONCE(!tick_work_cpu);
3169 twork = per_cpu_ptr(tick_work_cpu, cpu);
3170 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
3171 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
3172 if (os == TICK_SCHED_REMOTE_OFFLINE) {
3174 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3175 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3179 #ifdef CONFIG_HOTPLUG_CPU
3180 static void sched_tick_stop(int cpu)
3182 struct tick_work *twork;
3185 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3188 WARN_ON_ONCE(!tick_work_cpu);
3190 twork = per_cpu_ptr(tick_work_cpu, cpu);
3191 /* There cannot be competing actions, but don't rely on stop-machine. */
3192 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
3193 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
3194 /* Don't cancel, as this would mess up the state machine. */
3196 #endif /* CONFIG_HOTPLUG_CPU */
3198 int __init sched_tick_offload_init(void)
3200 tick_work_cpu = alloc_percpu(struct tick_work);
3201 BUG_ON(!tick_work_cpu);
3205 #else /* !CONFIG_NO_HZ_FULL */
3206 static inline void sched_tick_start(int cpu) { }
3207 static inline void sched_tick_stop(int cpu) { }
3210 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3211 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3213 * If the value passed in is equal to the current preempt count
3214 * then we just disabled preemption. Start timing the latency.
3216 static inline void preempt_latency_start(int val)
3218 if (preempt_count() == val) {
3219 unsigned long ip = get_lock_parent_ip();
3220 #ifdef CONFIG_DEBUG_PREEMPT
3221 current->preempt_disable_ip = ip;
3223 trace_preempt_off(CALLER_ADDR0, ip);
3227 void preempt_count_add(int val)
3229 #ifdef CONFIG_DEBUG_PREEMPT
3233 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3236 __preempt_count_add(val);
3237 #ifdef CONFIG_DEBUG_PREEMPT
3239 * Spinlock count overflowing soon?
3241 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3244 preempt_latency_start(val);
3246 EXPORT_SYMBOL(preempt_count_add);
3247 NOKPROBE_SYMBOL(preempt_count_add);
3250 * If the value passed in equals to the current preempt count
3251 * then we just enabled preemption. Stop timing the latency.
3253 static inline void preempt_latency_stop(int val)
3255 if (preempt_count() == val)
3256 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3259 void preempt_count_sub(int val)
3261 #ifdef CONFIG_DEBUG_PREEMPT
3265 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3268 * Is the spinlock portion underflowing?
3270 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3271 !(preempt_count() & PREEMPT_MASK)))
3275 preempt_latency_stop(val);
3276 __preempt_count_sub(val);
3278 EXPORT_SYMBOL(preempt_count_sub);
3279 NOKPROBE_SYMBOL(preempt_count_sub);
3282 static inline void preempt_latency_start(int val) { }
3283 static inline void preempt_latency_stop(int val) { }
3286 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3288 #ifdef CONFIG_DEBUG_PREEMPT
3289 return p->preempt_disable_ip;
3296 * Print scheduling while atomic bug:
3298 static noinline void __schedule_bug(struct task_struct *prev)
3300 /* Save this before calling printk(), since that will clobber it */
3301 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3303 if (oops_in_progress)
3306 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3307 prev->comm, prev->pid, preempt_count());
3309 debug_show_held_locks(prev);
3311 if (irqs_disabled())
3312 print_irqtrace_events(prev);
3313 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3314 && in_atomic_preempt_off()) {
3315 pr_err("Preemption disabled at:");
3316 print_ip_sym(preempt_disable_ip);
3320 panic("scheduling while atomic\n");
3323 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3327 * Various schedule()-time debugging checks and statistics:
3329 static inline void schedule_debug(struct task_struct *prev)
3331 #ifdef CONFIG_SCHED_STACK_END_CHECK
3332 if (task_stack_end_corrupted(prev))
3333 panic("corrupted stack end detected inside scheduler\n");
3336 if (unlikely(in_atomic_preempt_off())) {
3337 __schedule_bug(prev);
3338 preempt_count_set(PREEMPT_DISABLED);
3342 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3344 schedstat_inc(this_rq()->sched_count);
3348 * Pick up the highest-prio task:
3350 static inline struct task_struct *
3351 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3353 const struct sched_class *class;
3354 struct task_struct *p;
3357 * Optimization: we know that if all tasks are in the fair class we can
3358 * call that function directly, but only if the @prev task wasn't of a
3359 * higher scheduling class, because otherwise those loose the
3360 * opportunity to pull in more work from other CPUs.
3362 if (likely((prev->sched_class == &idle_sched_class ||
3363 prev->sched_class == &fair_sched_class) &&
3364 rq->nr_running == rq->cfs.h_nr_running)) {
3366 p = fair_sched_class.pick_next_task(rq, prev, rf);
3367 if (unlikely(p == RETRY_TASK))
3370 /* Assumes fair_sched_class->next == idle_sched_class */
3372 p = idle_sched_class.pick_next_task(rq, prev, rf);
3378 for_each_class(class) {
3379 p = class->pick_next_task(rq, prev, rf);
3381 if (unlikely(p == RETRY_TASK))
3387 /* The idle class should always have a runnable task: */
3392 * __schedule() is the main scheduler function.
3394 * The main means of driving the scheduler and thus entering this function are:
3396 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3398 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3399 * paths. For example, see arch/x86/entry_64.S.
3401 * To drive preemption between tasks, the scheduler sets the flag in timer
3402 * interrupt handler scheduler_tick().
3404 * 3. Wakeups don't really cause entry into schedule(). They add a
3405 * task to the run-queue and that's it.
3407 * Now, if the new task added to the run-queue preempts the current
3408 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3409 * called on the nearest possible occasion:
3411 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3413 * - in syscall or exception context, at the next outmost
3414 * preempt_enable(). (this might be as soon as the wake_up()'s
3417 * - in IRQ context, return from interrupt-handler to
3418 * preemptible context
3420 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3423 * - cond_resched() call
3424 * - explicit schedule() call
3425 * - return from syscall or exception to user-space
3426 * - return from interrupt-handler to user-space
3428 * WARNING: must be called with preemption disabled!
3430 static void __sched notrace __schedule(bool preempt)
3432 struct task_struct *prev, *next;
3433 unsigned long *switch_count;
3438 cpu = smp_processor_id();
3442 schedule_debug(prev);
3444 if (sched_feat(HRTICK))
3447 local_irq_disable();
3448 rcu_note_context_switch(preempt);
3451 * Make sure that signal_pending_state()->signal_pending() below
3452 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3453 * done by the caller to avoid the race with signal_wake_up().
3455 * The membarrier system call requires a full memory barrier
3456 * after coming from user-space, before storing to rq->curr.
3459 smp_mb__after_spinlock();
3461 /* Promote REQ to ACT */
3462 rq->clock_update_flags <<= 1;
3463 update_rq_clock(rq);
3465 switch_count = &prev->nivcsw;
3466 if (!preempt && prev->state) {
3467 if (unlikely(signal_pending_state(prev->state, prev))) {
3468 prev->state = TASK_RUNNING;
3470 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3473 if (prev->in_iowait) {
3474 atomic_inc(&rq->nr_iowait);
3475 delayacct_blkio_start();
3479 * If a worker went to sleep, notify and ask workqueue
3480 * whether it wants to wake up a task to maintain
3483 if (prev->flags & PF_WQ_WORKER) {
3484 struct task_struct *to_wakeup;
3486 to_wakeup = wq_worker_sleeping(prev);
3488 try_to_wake_up_local(to_wakeup, &rf);
3491 switch_count = &prev->nvcsw;
3494 next = pick_next_task(rq, prev, &rf);
3495 clear_tsk_need_resched(prev);
3496 clear_preempt_need_resched();
3498 if (likely(prev != next)) {
3502 * The membarrier system call requires each architecture
3503 * to have a full memory barrier after updating
3504 * rq->curr, before returning to user-space.
3506 * Here are the schemes providing that barrier on the
3507 * various architectures:
3508 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3509 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3510 * - finish_lock_switch() for weakly-ordered
3511 * architectures where spin_unlock is a full barrier,
3512 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3513 * is a RELEASE barrier),
3517 trace_sched_switch(preempt, prev, next);
3519 /* Also unlocks the rq: */
3520 rq = context_switch(rq, prev, next, &rf);
3522 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3523 rq_unlock_irq(rq, &rf);
3526 balance_callback(rq);
3529 void __noreturn do_task_dead(void)
3531 /* Causes final put_task_struct in finish_task_switch(): */
3532 set_special_state(TASK_DEAD);
3534 /* Tell freezer to ignore us: */
3535 current->flags |= PF_NOFREEZE;
3540 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3545 static inline void sched_submit_work(struct task_struct *tsk)
3547 if (!tsk->state || tsk_is_pi_blocked(tsk))
3550 * If we are going to sleep and we have plugged IO queued,
3551 * make sure to submit it to avoid deadlocks.
3553 if (blk_needs_flush_plug(tsk))
3554 blk_schedule_flush_plug(tsk);
3557 asmlinkage __visible void __sched schedule(void)
3559 struct task_struct *tsk = current;
3561 sched_submit_work(tsk);
3565 sched_preempt_enable_no_resched();
3566 } while (need_resched());
3568 EXPORT_SYMBOL(schedule);
3571 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3572 * state (have scheduled out non-voluntarily) by making sure that all
3573 * tasks have either left the run queue or have gone into user space.
3574 * As idle tasks do not do either, they must not ever be preempted
3575 * (schedule out non-voluntarily).
3577 * schedule_idle() is similar to schedule_preempt_disable() except that it
3578 * never enables preemption because it does not call sched_submit_work().
3580 void __sched schedule_idle(void)
3583 * As this skips calling sched_submit_work(), which the idle task does
3584 * regardless because that function is a nop when the task is in a
3585 * TASK_RUNNING state, make sure this isn't used someplace that the
3586 * current task can be in any other state. Note, idle is always in the
3587 * TASK_RUNNING state.
3589 WARN_ON_ONCE(current->state);
3592 } while (need_resched());
3595 #ifdef CONFIG_CONTEXT_TRACKING
3596 asmlinkage __visible void __sched schedule_user(void)
3599 * If we come here after a random call to set_need_resched(),
3600 * or we have been woken up remotely but the IPI has not yet arrived,
3601 * we haven't yet exited the RCU idle mode. Do it here manually until
3602 * we find a better solution.
3604 * NB: There are buggy callers of this function. Ideally we
3605 * should warn if prev_state != CONTEXT_USER, but that will trigger
3606 * too frequently to make sense yet.
3608 enum ctx_state prev_state = exception_enter();
3610 exception_exit(prev_state);
3615 * schedule_preempt_disabled - called with preemption disabled
3617 * Returns with preemption disabled. Note: preempt_count must be 1
3619 void __sched schedule_preempt_disabled(void)
3621 sched_preempt_enable_no_resched();
3626 static void __sched notrace preempt_schedule_common(void)
3630 * Because the function tracer can trace preempt_count_sub()
3631 * and it also uses preempt_enable/disable_notrace(), if
3632 * NEED_RESCHED is set, the preempt_enable_notrace() called
3633 * by the function tracer will call this function again and
3634 * cause infinite recursion.
3636 * Preemption must be disabled here before the function
3637 * tracer can trace. Break up preempt_disable() into two
3638 * calls. One to disable preemption without fear of being
3639 * traced. The other to still record the preemption latency,
3640 * which can also be traced by the function tracer.
3642 preempt_disable_notrace();
3643 preempt_latency_start(1);
3645 preempt_latency_stop(1);
3646 preempt_enable_no_resched_notrace();
3649 * Check again in case we missed a preemption opportunity
3650 * between schedule and now.
3652 } while (need_resched());
3655 #ifdef CONFIG_PREEMPT
3657 * this is the entry point to schedule() from in-kernel preemption
3658 * off of preempt_enable. Kernel preemptions off return from interrupt
3659 * occur there and call schedule directly.
3661 asmlinkage __visible void __sched notrace preempt_schedule(void)
3664 * If there is a non-zero preempt_count or interrupts are disabled,
3665 * we do not want to preempt the current task. Just return..
3667 if (likely(!preemptible()))
3670 preempt_schedule_common();
3672 NOKPROBE_SYMBOL(preempt_schedule);
3673 EXPORT_SYMBOL(preempt_schedule);
3676 * preempt_schedule_notrace - preempt_schedule called by tracing
3678 * The tracing infrastructure uses preempt_enable_notrace to prevent
3679 * recursion and tracing preempt enabling caused by the tracing
3680 * infrastructure itself. But as tracing can happen in areas coming
3681 * from userspace or just about to enter userspace, a preempt enable
3682 * can occur before user_exit() is called. This will cause the scheduler
3683 * to be called when the system is still in usermode.
3685 * To prevent this, the preempt_enable_notrace will use this function
3686 * instead of preempt_schedule() to exit user context if needed before
3687 * calling the scheduler.
3689 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3691 enum ctx_state prev_ctx;
3693 if (likely(!preemptible()))
3698 * Because the function tracer can trace preempt_count_sub()
3699 * and it also uses preempt_enable/disable_notrace(), if
3700 * NEED_RESCHED is set, the preempt_enable_notrace() called
3701 * by the function tracer will call this function again and
3702 * cause infinite recursion.
3704 * Preemption must be disabled here before the function
3705 * tracer can trace. Break up preempt_disable() into two
3706 * calls. One to disable preemption without fear of being
3707 * traced. The other to still record the preemption latency,
3708 * which can also be traced by the function tracer.
3710 preempt_disable_notrace();
3711 preempt_latency_start(1);
3713 * Needs preempt disabled in case user_exit() is traced
3714 * and the tracer calls preempt_enable_notrace() causing
3715 * an infinite recursion.
3717 prev_ctx = exception_enter();
3719 exception_exit(prev_ctx);
3721 preempt_latency_stop(1);
3722 preempt_enable_no_resched_notrace();
3723 } while (need_resched());
3725 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3727 #endif /* CONFIG_PREEMPT */
3730 * this is the entry point to schedule() from kernel preemption
3731 * off of irq context.
3732 * Note, that this is called and return with irqs disabled. This will
3733 * protect us against recursive calling from irq.
3735 asmlinkage __visible void __sched preempt_schedule_irq(void)
3737 enum ctx_state prev_state;
3739 /* Catch callers which need to be fixed */
3740 BUG_ON(preempt_count() || !irqs_disabled());
3742 prev_state = exception_enter();
3748 local_irq_disable();
3749 sched_preempt_enable_no_resched();
3750 } while (need_resched());
3752 exception_exit(prev_state);
3755 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
3758 return try_to_wake_up(curr->private, mode, wake_flags);
3760 EXPORT_SYMBOL(default_wake_function);
3762 #ifdef CONFIG_RT_MUTEXES
3764 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
3767 prio = min(prio, pi_task->prio);
3772 static inline int rt_effective_prio(struct task_struct *p, int prio)
3774 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3776 return __rt_effective_prio(pi_task, prio);
3780 * rt_mutex_setprio - set the current priority of a task
3782 * @pi_task: donor task
3784 * This function changes the 'effective' priority of a task. It does
3785 * not touch ->normal_prio like __setscheduler().
3787 * Used by the rt_mutex code to implement priority inheritance
3788 * logic. Call site only calls if the priority of the task changed.
3790 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
3792 int prio, oldprio, queued, running, queue_flag =
3793 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
3794 const struct sched_class *prev_class;
3798 /* XXX used to be waiter->prio, not waiter->task->prio */
3799 prio = __rt_effective_prio(pi_task, p->normal_prio);
3802 * If nothing changed; bail early.
3804 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
3807 rq = __task_rq_lock(p, &rf);
3808 update_rq_clock(rq);
3810 * Set under pi_lock && rq->lock, such that the value can be used under
3813 * Note that there is loads of tricky to make this pointer cache work
3814 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3815 * ensure a task is de-boosted (pi_task is set to NULL) before the
3816 * task is allowed to run again (and can exit). This ensures the pointer
3817 * points to a blocked task -- which guaratees the task is present.
3819 p->pi_top_task = pi_task;
3822 * For FIFO/RR we only need to set prio, if that matches we're done.
3824 if (prio == p->prio && !dl_prio(prio))
3828 * Idle task boosting is a nono in general. There is one
3829 * exception, when PREEMPT_RT and NOHZ is active:
3831 * The idle task calls get_next_timer_interrupt() and holds
3832 * the timer wheel base->lock on the CPU and another CPU wants
3833 * to access the timer (probably to cancel it). We can safely
3834 * ignore the boosting request, as the idle CPU runs this code
3835 * with interrupts disabled and will complete the lock
3836 * protected section without being interrupted. So there is no
3837 * real need to boost.
3839 if (unlikely(p == rq->idle)) {
3840 WARN_ON(p != rq->curr);
3841 WARN_ON(p->pi_blocked_on);
3845 trace_sched_pi_setprio(p, pi_task);
3848 if (oldprio == prio)
3849 queue_flag &= ~DEQUEUE_MOVE;
3851 prev_class = p->sched_class;
3852 queued = task_on_rq_queued(p);
3853 running = task_current(rq, p);
3855 dequeue_task(rq, p, queue_flag);
3857 put_prev_task(rq, p);
3860 * Boosting condition are:
3861 * 1. -rt task is running and holds mutex A
3862 * --> -dl task blocks on mutex A
3864 * 2. -dl task is running and holds mutex A
3865 * --> -dl task blocks on mutex A and could preempt the
3868 if (dl_prio(prio)) {
3869 if (!dl_prio(p->normal_prio) ||
3870 (pi_task && dl_prio(pi_task->prio) &&
3871 dl_entity_preempt(&pi_task->dl, &p->dl))) {
3872 p->dl.dl_boosted = 1;
3873 queue_flag |= ENQUEUE_REPLENISH;
3875 p->dl.dl_boosted = 0;
3876 p->sched_class = &dl_sched_class;
3877 } else if (rt_prio(prio)) {
3878 if (dl_prio(oldprio))
3879 p->dl.dl_boosted = 0;
3881 queue_flag |= ENQUEUE_HEAD;
3882 p->sched_class = &rt_sched_class;
3884 if (dl_prio(oldprio))
3885 p->dl.dl_boosted = 0;
3886 if (rt_prio(oldprio))
3888 p->sched_class = &fair_sched_class;
3894 enqueue_task(rq, p, queue_flag);
3896 set_curr_task(rq, p);
3898 check_class_changed(rq, p, prev_class, oldprio);
3900 /* Avoid rq from going away on us: */
3902 __task_rq_unlock(rq, &rf);
3904 balance_callback(rq);
3908 static inline int rt_effective_prio(struct task_struct *p, int prio)
3914 void set_user_nice(struct task_struct *p, long nice)
3916 bool queued, running;
3917 int old_prio, delta;
3921 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3924 * We have to be careful, if called from sys_setpriority(),
3925 * the task might be in the middle of scheduling on another CPU.
3927 rq = task_rq_lock(p, &rf);
3928 update_rq_clock(rq);
3931 * The RT priorities are set via sched_setscheduler(), but we still
3932 * allow the 'normal' nice value to be set - but as expected
3933 * it wont have any effect on scheduling until the task is
3934 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3936 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3937 p->static_prio = NICE_TO_PRIO(nice);
3940 queued = task_on_rq_queued(p);
3941 running = task_current(rq, p);
3943 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
3945 put_prev_task(rq, p);
3947 p->static_prio = NICE_TO_PRIO(nice);
3948 set_load_weight(p, true);
3950 p->prio = effective_prio(p);
3951 delta = p->prio - old_prio;
3954 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
3956 * If the task increased its priority or is running and
3957 * lowered its priority, then reschedule its CPU:
3959 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3963 set_curr_task(rq, p);
3965 task_rq_unlock(rq, p, &rf);
3967 EXPORT_SYMBOL(set_user_nice);
3970 * can_nice - check if a task can reduce its nice value
3974 int can_nice(const struct task_struct *p, const int nice)
3976 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3977 int nice_rlim = nice_to_rlimit(nice);
3979 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3980 capable(CAP_SYS_NICE));
3983 #ifdef __ARCH_WANT_SYS_NICE
3986 * sys_nice - change the priority of the current process.
3987 * @increment: priority increment
3989 * sys_setpriority is a more generic, but much slower function that
3990 * does similar things.
3992 SYSCALL_DEFINE1(nice, int, increment)
3997 * Setpriority might change our priority at the same moment.
3998 * We don't have to worry. Conceptually one call occurs first
3999 * and we have a single winner.
4001 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
4002 nice = task_nice(current) + increment;
4004 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
4005 if (increment < 0 && !can_nice(current, nice))
4008 retval = security_task_setnice(current, nice);
4012 set_user_nice(current, nice);
4019 * task_prio - return the priority value of a given task.
4020 * @p: the task in question.
4022 * Return: The priority value as seen by users in /proc.
4023 * RT tasks are offset by -200. Normal tasks are centered
4024 * around 0, value goes from -16 to +15.
4026 int task_prio(const struct task_struct *p)
4028 return p->prio - MAX_RT_PRIO;
4032 * idle_cpu - is a given CPU idle currently?
4033 * @cpu: the processor in question.
4035 * Return: 1 if the CPU is currently idle. 0 otherwise.
4037 int idle_cpu(int cpu)
4039 struct rq *rq = cpu_rq(cpu);
4041 if (rq->curr != rq->idle)
4048 if (!llist_empty(&rq->wake_list))
4056 * available_idle_cpu - is a given CPU idle for enqueuing work.
4057 * @cpu: the CPU in question.
4059 * Return: 1 if the CPU is currently idle. 0 otherwise.
4061 int available_idle_cpu(int cpu)
4066 if (vcpu_is_preempted(cpu))
4073 * idle_task - return the idle task for a given CPU.
4074 * @cpu: the processor in question.
4076 * Return: The idle task for the CPU @cpu.
4078 struct task_struct *idle_task(int cpu)
4080 return cpu_rq(cpu)->idle;
4084 * find_process_by_pid - find a process with a matching PID value.
4085 * @pid: the pid in question.
4087 * The task of @pid, if found. %NULL otherwise.
4089 static struct task_struct *find_process_by_pid(pid_t pid)
4091 return pid ? find_task_by_vpid(pid) : current;
4095 * sched_setparam() passes in -1 for its policy, to let the functions
4096 * it calls know not to change it.
4098 #define SETPARAM_POLICY -1
4100 static void __setscheduler_params(struct task_struct *p,
4101 const struct sched_attr *attr)
4103 int policy = attr->sched_policy;
4105 if (policy == SETPARAM_POLICY)
4110 if (dl_policy(policy))
4111 __setparam_dl(p, attr);
4112 else if (fair_policy(policy))
4113 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4116 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4117 * !rt_policy. Always setting this ensures that things like
4118 * getparam()/getattr() don't report silly values for !rt tasks.
4120 p->rt_priority = attr->sched_priority;
4121 p->normal_prio = normal_prio(p);
4122 set_load_weight(p, true);
4125 /* Actually do priority change: must hold pi & rq lock. */
4126 static void __setscheduler(struct rq *rq, struct task_struct *p,
4127 const struct sched_attr *attr, bool keep_boost)
4129 __setscheduler_params(p, attr);
4132 * Keep a potential priority boosting if called from
4133 * sched_setscheduler().
4135 p->prio = normal_prio(p);
4137 p->prio = rt_effective_prio(p, p->prio);
4139 if (dl_prio(p->prio))
4140 p->sched_class = &dl_sched_class;
4141 else if (rt_prio(p->prio))
4142 p->sched_class = &rt_sched_class;
4144 p->sched_class = &fair_sched_class;
4148 * Check the target process has a UID that matches the current process's:
4150 static bool check_same_owner(struct task_struct *p)
4152 const struct cred *cred = current_cred(), *pcred;
4156 pcred = __task_cred(p);
4157 match = (uid_eq(cred->euid, pcred->euid) ||
4158 uid_eq(cred->euid, pcred->uid));
4163 static int __sched_setscheduler(struct task_struct *p,
4164 const struct sched_attr *attr,
4167 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4168 MAX_RT_PRIO - 1 - attr->sched_priority;
4169 int retval, oldprio, oldpolicy = -1, queued, running;
4170 int new_effective_prio, policy = attr->sched_policy;
4171 const struct sched_class *prev_class;
4174 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4177 /* The pi code expects interrupts enabled */
4178 BUG_ON(pi && in_interrupt());
4180 /* Double check policy once rq lock held: */
4182 reset_on_fork = p->sched_reset_on_fork;
4183 policy = oldpolicy = p->policy;
4185 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4187 if (!valid_policy(policy))
4191 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4195 * Valid priorities for SCHED_FIFO and SCHED_RR are
4196 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4197 * SCHED_BATCH and SCHED_IDLE is 0.
4199 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4200 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4202 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4203 (rt_policy(policy) != (attr->sched_priority != 0)))
4207 * Allow unprivileged RT tasks to decrease priority:
4209 if (user && !capable(CAP_SYS_NICE)) {
4210 if (fair_policy(policy)) {
4211 if (attr->sched_nice < task_nice(p) &&
4212 !can_nice(p, attr->sched_nice))
4216 if (rt_policy(policy)) {
4217 unsigned long rlim_rtprio =
4218 task_rlimit(p, RLIMIT_RTPRIO);
4220 /* Can't set/change the rt policy: */
4221 if (policy != p->policy && !rlim_rtprio)
4224 /* Can't increase priority: */
4225 if (attr->sched_priority > p->rt_priority &&
4226 attr->sched_priority > rlim_rtprio)
4231 * Can't set/change SCHED_DEADLINE policy at all for now
4232 * (safest behavior); in the future we would like to allow
4233 * unprivileged DL tasks to increase their relative deadline
4234 * or reduce their runtime (both ways reducing utilization)
4236 if (dl_policy(policy))
4240 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4241 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4243 if (idle_policy(p->policy) && !idle_policy(policy)) {
4244 if (!can_nice(p, task_nice(p)))
4248 /* Can't change other user's priorities: */
4249 if (!check_same_owner(p))
4252 /* Normal users shall not reset the sched_reset_on_fork flag: */
4253 if (p->sched_reset_on_fork && !reset_on_fork)
4258 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4261 retval = security_task_setscheduler(p);
4267 * Make sure no PI-waiters arrive (or leave) while we are
4268 * changing the priority of the task:
4270 * To be able to change p->policy safely, the appropriate
4271 * runqueue lock must be held.
4273 rq = task_rq_lock(p, &rf);
4274 update_rq_clock(rq);
4277 * Changing the policy of the stop threads its a very bad idea:
4279 if (p == rq->stop) {
4280 task_rq_unlock(rq, p, &rf);
4285 * If not changing anything there's no need to proceed further,
4286 * but store a possible modification of reset_on_fork.
4288 if (unlikely(policy == p->policy)) {
4289 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4291 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4293 if (dl_policy(policy) && dl_param_changed(p, attr))
4296 p->sched_reset_on_fork = reset_on_fork;
4297 task_rq_unlock(rq, p, &rf);
4303 #ifdef CONFIG_RT_GROUP_SCHED
4305 * Do not allow realtime tasks into groups that have no runtime
4308 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4309 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4310 !task_group_is_autogroup(task_group(p))) {
4311 task_rq_unlock(rq, p, &rf);
4316 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4317 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4318 cpumask_t *span = rq->rd->span;
4321 * Don't allow tasks with an affinity mask smaller than
4322 * the entire root_domain to become SCHED_DEADLINE. We
4323 * will also fail if there's no bandwidth available.
4325 if (!cpumask_subset(span, &p->cpus_allowed) ||
4326 rq->rd->dl_bw.bw == 0) {
4327 task_rq_unlock(rq, p, &rf);
4334 /* Re-check policy now with rq lock held: */
4335 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4336 policy = oldpolicy = -1;
4337 task_rq_unlock(rq, p, &rf);
4342 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4343 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4346 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4347 task_rq_unlock(rq, p, &rf);
4351 p->sched_reset_on_fork = reset_on_fork;
4356 * Take priority boosted tasks into account. If the new
4357 * effective priority is unchanged, we just store the new
4358 * normal parameters and do not touch the scheduler class and
4359 * the runqueue. This will be done when the task deboost
4362 new_effective_prio = rt_effective_prio(p, newprio);
4363 if (new_effective_prio == oldprio)
4364 queue_flags &= ~DEQUEUE_MOVE;
4367 queued = task_on_rq_queued(p);
4368 running = task_current(rq, p);
4370 dequeue_task(rq, p, queue_flags);
4372 put_prev_task(rq, p);
4374 prev_class = p->sched_class;
4375 __setscheduler(rq, p, attr, pi);
4379 * We enqueue to tail when the priority of a task is
4380 * increased (user space view).
4382 if (oldprio < p->prio)
4383 queue_flags |= ENQUEUE_HEAD;
4385 enqueue_task(rq, p, queue_flags);
4388 set_curr_task(rq, p);
4390 check_class_changed(rq, p, prev_class, oldprio);
4392 /* Avoid rq from going away on us: */
4394 task_rq_unlock(rq, p, &rf);
4397 rt_mutex_adjust_pi(p);
4399 /* Run balance callbacks after we've adjusted the PI chain: */
4400 balance_callback(rq);
4406 static int _sched_setscheduler(struct task_struct *p, int policy,
4407 const struct sched_param *param, bool check)
4409 struct sched_attr attr = {
4410 .sched_policy = policy,
4411 .sched_priority = param->sched_priority,
4412 .sched_nice = PRIO_TO_NICE(p->static_prio),
4415 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4416 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4417 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4418 policy &= ~SCHED_RESET_ON_FORK;
4419 attr.sched_policy = policy;
4422 return __sched_setscheduler(p, &attr, check, true);
4425 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4426 * @p: the task in question.
4427 * @policy: new policy.
4428 * @param: structure containing the new RT priority.
4430 * Return: 0 on success. An error code otherwise.
4432 * NOTE that the task may be already dead.
4434 int sched_setscheduler(struct task_struct *p, int policy,
4435 const struct sched_param *param)
4437 return _sched_setscheduler(p, policy, param, true);
4439 EXPORT_SYMBOL_GPL(sched_setscheduler);
4441 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4443 return __sched_setscheduler(p, attr, true, true);
4445 EXPORT_SYMBOL_GPL(sched_setattr);
4447 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
4449 return __sched_setscheduler(p, attr, false, true);
4453 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4454 * @p: the task in question.
4455 * @policy: new policy.
4456 * @param: structure containing the new RT priority.
4458 * Just like sched_setscheduler, only don't bother checking if the
4459 * current context has permission. For example, this is needed in
4460 * stop_machine(): we create temporary high priority worker threads,
4461 * but our caller might not have that capability.
4463 * Return: 0 on success. An error code otherwise.
4465 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4466 const struct sched_param *param)
4468 return _sched_setscheduler(p, policy, param, false);
4470 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4473 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4475 struct sched_param lparam;
4476 struct task_struct *p;
4479 if (!param || pid < 0)
4481 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4486 p = find_process_by_pid(pid);
4488 retval = sched_setscheduler(p, policy, &lparam);
4495 * Mimics kernel/events/core.c perf_copy_attr().
4497 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4502 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4505 /* Zero the full structure, so that a short copy will be nice: */
4506 memset(attr, 0, sizeof(*attr));
4508 ret = get_user(size, &uattr->size);
4512 /* Bail out on silly large: */
4513 if (size > PAGE_SIZE)
4516 /* ABI compatibility quirk: */
4518 size = SCHED_ATTR_SIZE_VER0;
4520 if (size < SCHED_ATTR_SIZE_VER0)
4524 * If we're handed a bigger struct than we know of,
4525 * ensure all the unknown bits are 0 - i.e. new
4526 * user-space does not rely on any kernel feature
4527 * extensions we dont know about yet.
4529 if (size > sizeof(*attr)) {
4530 unsigned char __user *addr;
4531 unsigned char __user *end;
4534 addr = (void __user *)uattr + sizeof(*attr);
4535 end = (void __user *)uattr + size;
4537 for (; addr < end; addr++) {
4538 ret = get_user(val, addr);
4544 size = sizeof(*attr);
4547 ret = copy_from_user(attr, uattr, size);
4552 * XXX: Do we want to be lenient like existing syscalls; or do we want
4553 * to be strict and return an error on out-of-bounds values?
4555 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4560 put_user(sizeof(*attr), &uattr->size);
4565 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4566 * @pid: the pid in question.
4567 * @policy: new policy.
4568 * @param: structure containing the new RT priority.
4570 * Return: 0 on success. An error code otherwise.
4572 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4577 return do_sched_setscheduler(pid, policy, param);
4581 * sys_sched_setparam - set/change the RT priority of a thread
4582 * @pid: the pid in question.
4583 * @param: structure containing the new RT priority.
4585 * Return: 0 on success. An error code otherwise.
4587 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4589 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4593 * sys_sched_setattr - same as above, but with extended sched_attr
4594 * @pid: the pid in question.
4595 * @uattr: structure containing the extended parameters.
4596 * @flags: for future extension.
4598 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4599 unsigned int, flags)
4601 struct sched_attr attr;
4602 struct task_struct *p;
4605 if (!uattr || pid < 0 || flags)
4608 retval = sched_copy_attr(uattr, &attr);
4612 if ((int)attr.sched_policy < 0)
4617 p = find_process_by_pid(pid);
4619 retval = sched_setattr(p, &attr);
4626 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4627 * @pid: the pid in question.
4629 * Return: On success, the policy of the thread. Otherwise, a negative error
4632 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4634 struct task_struct *p;
4642 p = find_process_by_pid(pid);
4644 retval = security_task_getscheduler(p);
4647 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4654 * sys_sched_getparam - get the RT priority of a thread
4655 * @pid: the pid in question.
4656 * @param: structure containing the RT priority.
4658 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4661 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4663 struct sched_param lp = { .sched_priority = 0 };
4664 struct task_struct *p;
4667 if (!param || pid < 0)
4671 p = find_process_by_pid(pid);
4676 retval = security_task_getscheduler(p);
4680 if (task_has_rt_policy(p))
4681 lp.sched_priority = p->rt_priority;
4685 * This one might sleep, we cannot do it with a spinlock held ...
4687 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4696 static int sched_read_attr(struct sched_attr __user *uattr,
4697 struct sched_attr *attr,
4702 if (!access_ok(VERIFY_WRITE, uattr, usize))
4706 * If we're handed a smaller struct than we know of,
4707 * ensure all the unknown bits are 0 - i.e. old
4708 * user-space does not get uncomplete information.
4710 if (usize < sizeof(*attr)) {
4711 unsigned char *addr;
4714 addr = (void *)attr + usize;
4715 end = (void *)attr + sizeof(*attr);
4717 for (; addr < end; addr++) {
4725 ret = copy_to_user(uattr, attr, attr->size);
4733 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4734 * @pid: the pid in question.
4735 * @uattr: structure containing the extended parameters.
4736 * @size: sizeof(attr) for fwd/bwd comp.
4737 * @flags: for future extension.
4739 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4740 unsigned int, size, unsigned int, flags)
4742 struct sched_attr attr = {
4743 .size = sizeof(struct sched_attr),
4745 struct task_struct *p;
4748 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4749 size < SCHED_ATTR_SIZE_VER0 || flags)
4753 p = find_process_by_pid(pid);
4758 retval = security_task_getscheduler(p);
4762 attr.sched_policy = p->policy;
4763 if (p->sched_reset_on_fork)
4764 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4765 if (task_has_dl_policy(p))
4766 __getparam_dl(p, &attr);
4767 else if (task_has_rt_policy(p))
4768 attr.sched_priority = p->rt_priority;
4770 attr.sched_nice = task_nice(p);
4774 retval = sched_read_attr(uattr, &attr, size);
4782 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4784 cpumask_var_t cpus_allowed, new_mask;
4785 struct task_struct *p;
4790 p = find_process_by_pid(pid);
4796 /* Prevent p going away */
4800 if (p->flags & PF_NO_SETAFFINITY) {
4804 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4808 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4810 goto out_free_cpus_allowed;
4813 if (!check_same_owner(p)) {
4815 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4817 goto out_free_new_mask;
4822 retval = security_task_setscheduler(p);
4824 goto out_free_new_mask;
4827 cpuset_cpus_allowed(p, cpus_allowed);
4828 cpumask_and(new_mask, in_mask, cpus_allowed);
4831 * Since bandwidth control happens on root_domain basis,
4832 * if admission test is enabled, we only admit -deadline
4833 * tasks allowed to run on all the CPUs in the task's
4837 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4839 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4842 goto out_free_new_mask;
4848 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4851 cpuset_cpus_allowed(p, cpus_allowed);
4852 if (!cpumask_subset(new_mask, cpus_allowed)) {
4854 * We must have raced with a concurrent cpuset
4855 * update. Just reset the cpus_allowed to the
4856 * cpuset's cpus_allowed
4858 cpumask_copy(new_mask, cpus_allowed);
4863 free_cpumask_var(new_mask);
4864 out_free_cpus_allowed:
4865 free_cpumask_var(cpus_allowed);
4871 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4872 struct cpumask *new_mask)
4874 if (len < cpumask_size())
4875 cpumask_clear(new_mask);
4876 else if (len > cpumask_size())
4877 len = cpumask_size();
4879 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4883 * sys_sched_setaffinity - set the CPU affinity of a process
4884 * @pid: pid of the process
4885 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4886 * @user_mask_ptr: user-space pointer to the new CPU mask
4888 * Return: 0 on success. An error code otherwise.
4890 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4891 unsigned long __user *, user_mask_ptr)
4893 cpumask_var_t new_mask;
4896 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4899 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4901 retval = sched_setaffinity(pid, new_mask);
4902 free_cpumask_var(new_mask);
4906 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4908 struct task_struct *p;
4909 unsigned long flags;
4915 p = find_process_by_pid(pid);
4919 retval = security_task_getscheduler(p);
4923 raw_spin_lock_irqsave(&p->pi_lock, flags);
4924 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4925 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4934 * sys_sched_getaffinity - get the CPU affinity of a process
4935 * @pid: pid of the process
4936 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4937 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4939 * Return: size of CPU mask copied to user_mask_ptr on success. An
4940 * error code otherwise.
4942 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4943 unsigned long __user *, user_mask_ptr)
4948 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4950 if (len & (sizeof(unsigned long)-1))
4953 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4956 ret = sched_getaffinity(pid, mask);
4958 unsigned int retlen = min(len, cpumask_size());
4960 if (copy_to_user(user_mask_ptr, mask, retlen))
4965 free_cpumask_var(mask);
4971 * sys_sched_yield - yield the current processor to other threads.
4973 * This function yields the current CPU to other tasks. If there are no
4974 * other threads running on this CPU then this function will return.
4978 static void do_sched_yield(void)
4983 local_irq_disable();
4987 schedstat_inc(rq->yld_count);
4988 current->sched_class->yield_task(rq);
4991 rq_unlock_irq(rq, &rf);
4992 sched_preempt_enable_no_resched();
4997 SYSCALL_DEFINE0(sched_yield)
5003 #ifndef CONFIG_PREEMPT
5004 int __sched _cond_resched(void)
5006 if (should_resched(0)) {
5007 preempt_schedule_common();
5013 EXPORT_SYMBOL(_cond_resched);
5017 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5018 * call schedule, and on return reacquire the lock.
5020 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5021 * operations here to prevent schedule() from being called twice (once via
5022 * spin_unlock(), once by hand).
5024 int __cond_resched_lock(spinlock_t *lock)
5026 int resched = should_resched(PREEMPT_LOCK_OFFSET);
5029 lockdep_assert_held(lock);
5031 if (spin_needbreak(lock) || resched) {
5034 preempt_schedule_common();
5042 EXPORT_SYMBOL(__cond_resched_lock);
5045 * yield - yield the current processor to other threads.
5047 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5049 * The scheduler is at all times free to pick the calling task as the most
5050 * eligible task to run, if removing the yield() call from your code breaks
5051 * it, its already broken.
5053 * Typical broken usage is:
5058 * where one assumes that yield() will let 'the other' process run that will
5059 * make event true. If the current task is a SCHED_FIFO task that will never
5060 * happen. Never use yield() as a progress guarantee!!
5062 * If you want to use yield() to wait for something, use wait_event().
5063 * If you want to use yield() to be 'nice' for others, use cond_resched().
5064 * If you still want to use yield(), do not!
5066 void __sched yield(void)
5068 set_current_state(TASK_RUNNING);
5071 EXPORT_SYMBOL(yield);
5074 * yield_to - yield the current processor to another thread in
5075 * your thread group, or accelerate that thread toward the
5076 * processor it's on.
5078 * @preempt: whether task preemption is allowed or not
5080 * It's the caller's job to ensure that the target task struct
5081 * can't go away on us before we can do any checks.
5084 * true (>0) if we indeed boosted the target task.
5085 * false (0) if we failed to boost the target.
5086 * -ESRCH if there's no task to yield to.
5088 int __sched yield_to(struct task_struct *p, bool preempt)
5090 struct task_struct *curr = current;
5091 struct rq *rq, *p_rq;
5092 unsigned long flags;
5095 local_irq_save(flags);
5101 * If we're the only runnable task on the rq and target rq also
5102 * has only one task, there's absolutely no point in yielding.
5104 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5109 double_rq_lock(rq, p_rq);
5110 if (task_rq(p) != p_rq) {
5111 double_rq_unlock(rq, p_rq);
5115 if (!curr->sched_class->yield_to_task)
5118 if (curr->sched_class != p->sched_class)
5121 if (task_running(p_rq, p) || p->state)
5124 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5126 schedstat_inc(rq->yld_count);
5128 * Make p's CPU reschedule; pick_next_entity takes care of
5131 if (preempt && rq != p_rq)
5136 double_rq_unlock(rq, p_rq);
5138 local_irq_restore(flags);
5145 EXPORT_SYMBOL_GPL(yield_to);
5147 int io_schedule_prepare(void)
5149 int old_iowait = current->in_iowait;
5151 current->in_iowait = 1;
5152 blk_schedule_flush_plug(current);
5157 void io_schedule_finish(int token)
5159 current->in_iowait = token;
5163 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5164 * that process accounting knows that this is a task in IO wait state.
5166 long __sched io_schedule_timeout(long timeout)
5171 token = io_schedule_prepare();
5172 ret = schedule_timeout(timeout);
5173 io_schedule_finish(token);
5177 EXPORT_SYMBOL(io_schedule_timeout);
5179 void __sched io_schedule(void)
5183 token = io_schedule_prepare();
5185 io_schedule_finish(token);
5187 EXPORT_SYMBOL(io_schedule);
5190 * sys_sched_get_priority_max - return maximum RT priority.
5191 * @policy: scheduling class.
5193 * Return: On success, this syscall returns the maximum
5194 * rt_priority that can be used by a given scheduling class.
5195 * On failure, a negative error code is returned.
5197 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5204 ret = MAX_USER_RT_PRIO-1;
5206 case SCHED_DEADLINE:
5217 * sys_sched_get_priority_min - return minimum RT priority.
5218 * @policy: scheduling class.
5220 * Return: On success, this syscall returns the minimum
5221 * rt_priority that can be used by a given scheduling class.
5222 * On failure, a negative error code is returned.
5224 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5233 case SCHED_DEADLINE:
5242 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5244 struct task_struct *p;
5245 unsigned int time_slice;
5255 p = find_process_by_pid(pid);
5259 retval = security_task_getscheduler(p);
5263 rq = task_rq_lock(p, &rf);
5265 if (p->sched_class->get_rr_interval)
5266 time_slice = p->sched_class->get_rr_interval(rq, p);
5267 task_rq_unlock(rq, p, &rf);
5270 jiffies_to_timespec64(time_slice, t);
5279 * sys_sched_rr_get_interval - return the default timeslice of a process.
5280 * @pid: pid of the process.
5281 * @interval: userspace pointer to the timeslice value.
5283 * this syscall writes the default timeslice value of a given process
5284 * into the user-space timespec buffer. A value of '0' means infinity.
5286 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5289 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5290 struct timespec __user *, interval)
5292 struct timespec64 t;
5293 int retval = sched_rr_get_interval(pid, &t);
5296 retval = put_timespec64(&t, interval);
5301 #ifdef CONFIG_COMPAT
5302 COMPAT_SYSCALL_DEFINE2(sched_rr_get_interval,
5304 struct compat_timespec __user *, interval)
5306 struct timespec64 t;
5307 int retval = sched_rr_get_interval(pid, &t);
5310 retval = compat_put_timespec64(&t, interval);
5315 void sched_show_task(struct task_struct *p)
5317 unsigned long free = 0;
5320 if (!try_get_task_stack(p))
5323 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5325 if (p->state == TASK_RUNNING)
5326 printk(KERN_CONT " running task ");
5327 #ifdef CONFIG_DEBUG_STACK_USAGE
5328 free = stack_not_used(p);
5333 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5335 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5336 task_pid_nr(p), ppid,
5337 (unsigned long)task_thread_info(p)->flags);
5339 print_worker_info(KERN_INFO, p);
5340 show_stack(p, NULL);
5343 EXPORT_SYMBOL_GPL(sched_show_task);
5346 state_filter_match(unsigned long state_filter, struct task_struct *p)
5348 /* no filter, everything matches */
5352 /* filter, but doesn't match */
5353 if (!(p->state & state_filter))
5357 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5360 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5367 void show_state_filter(unsigned long state_filter)
5369 struct task_struct *g, *p;
5371 #if BITS_PER_LONG == 32
5373 " task PC stack pid father\n");
5376 " task PC stack pid father\n");
5379 for_each_process_thread(g, p) {
5381 * reset the NMI-timeout, listing all files on a slow
5382 * console might take a lot of time:
5383 * Also, reset softlockup watchdogs on all CPUs, because
5384 * another CPU might be blocked waiting for us to process
5387 touch_nmi_watchdog();
5388 touch_all_softlockup_watchdogs();
5389 if (state_filter_match(state_filter, p))
5393 #ifdef CONFIG_SCHED_DEBUG
5395 sysrq_sched_debug_show();
5399 * Only show locks if all tasks are dumped:
5402 debug_show_all_locks();
5406 * init_idle - set up an idle thread for a given CPU
5407 * @idle: task in question
5408 * @cpu: CPU the idle task belongs to
5410 * NOTE: this function does not set the idle thread's NEED_RESCHED
5411 * flag, to make booting more robust.
5413 void init_idle(struct task_struct *idle, int cpu)
5415 struct rq *rq = cpu_rq(cpu);
5416 unsigned long flags;
5418 __sched_fork(0, idle);
5420 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5421 raw_spin_lock(&rq->lock);
5423 idle->state = TASK_RUNNING;
5424 idle->se.exec_start = sched_clock();
5425 idle->flags |= PF_IDLE;
5427 kasan_unpoison_task_stack(idle);
5431 * Its possible that init_idle() gets called multiple times on a task,
5432 * in that case do_set_cpus_allowed() will not do the right thing.
5434 * And since this is boot we can forgo the serialization.
5436 set_cpus_allowed_common(idle, cpumask_of(cpu));
5439 * We're having a chicken and egg problem, even though we are
5440 * holding rq->lock, the CPU isn't yet set to this CPU so the
5441 * lockdep check in task_group() will fail.
5443 * Similar case to sched_fork(). / Alternatively we could
5444 * use task_rq_lock() here and obtain the other rq->lock.
5449 __set_task_cpu(idle, cpu);
5452 rq->curr = rq->idle = idle;
5453 idle->on_rq = TASK_ON_RQ_QUEUED;
5457 raw_spin_unlock(&rq->lock);
5458 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5460 /* Set the preempt count _outside_ the spinlocks! */
5461 init_idle_preempt_count(idle, cpu);
5464 * The idle tasks have their own, simple scheduling class:
5466 idle->sched_class = &idle_sched_class;
5467 ftrace_graph_init_idle_task(idle, cpu);
5468 vtime_init_idle(idle, cpu);
5470 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5476 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5477 const struct cpumask *trial)
5481 if (!cpumask_weight(cur))
5484 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5489 int task_can_attach(struct task_struct *p,
5490 const struct cpumask *cs_cpus_allowed)
5495 * Kthreads which disallow setaffinity shouldn't be moved
5496 * to a new cpuset; we don't want to change their CPU
5497 * affinity and isolating such threads by their set of
5498 * allowed nodes is unnecessary. Thus, cpusets are not
5499 * applicable for such threads. This prevents checking for
5500 * success of set_cpus_allowed_ptr() on all attached tasks
5501 * before cpus_allowed may be changed.
5503 if (p->flags & PF_NO_SETAFFINITY) {
5508 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5510 ret = dl_task_can_attach(p, cs_cpus_allowed);
5516 bool sched_smp_initialized __read_mostly;
5518 #ifdef CONFIG_NUMA_BALANCING
5519 /* Migrate current task p to target_cpu */
5520 int migrate_task_to(struct task_struct *p, int target_cpu)
5522 struct migration_arg arg = { p, target_cpu };
5523 int curr_cpu = task_cpu(p);
5525 if (curr_cpu == target_cpu)
5528 if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5531 /* TODO: This is not properly updating schedstats */
5533 trace_sched_move_numa(p, curr_cpu, target_cpu);
5534 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5538 * Requeue a task on a given node and accurately track the number of NUMA
5539 * tasks on the runqueues
5541 void sched_setnuma(struct task_struct *p, int nid)
5543 bool queued, running;
5547 rq = task_rq_lock(p, &rf);
5548 queued = task_on_rq_queued(p);
5549 running = task_current(rq, p);
5552 dequeue_task(rq, p, DEQUEUE_SAVE);
5554 put_prev_task(rq, p);
5556 p->numa_preferred_nid = nid;
5559 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5561 set_curr_task(rq, p);
5562 task_rq_unlock(rq, p, &rf);
5564 #endif /* CONFIG_NUMA_BALANCING */
5566 #ifdef CONFIG_HOTPLUG_CPU
5568 * Ensure that the idle task is using init_mm right before its CPU goes
5571 void idle_task_exit(void)
5573 struct mm_struct *mm = current->active_mm;
5575 BUG_ON(cpu_online(smp_processor_id()));
5576 BUG_ON(current != this_rq()->idle);
5578 if (mm != &init_mm) {
5579 switch_mm(mm, &init_mm, current);
5580 finish_arch_post_lock_switch();
5583 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
5587 * Since this CPU is going 'away' for a while, fold any nr_active delta
5588 * we might have. Assumes we're called after migrate_tasks() so that the
5589 * nr_active count is stable. We need to take the teardown thread which
5590 * is calling this into account, so we hand in adjust = 1 to the load
5593 * Also see the comment "Global load-average calculations".
5595 static void calc_load_migrate(struct rq *rq)
5597 long delta = calc_load_fold_active(rq, 1);
5599 atomic_long_add(delta, &calc_load_tasks);
5602 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5606 static const struct sched_class fake_sched_class = {
5607 .put_prev_task = put_prev_task_fake,
5610 static struct task_struct fake_task = {
5612 * Avoid pull_{rt,dl}_task()
5614 .prio = MAX_PRIO + 1,
5615 .sched_class = &fake_sched_class,
5619 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5620 * try_to_wake_up()->select_task_rq().
5622 * Called with rq->lock held even though we'er in stop_machine() and
5623 * there's no concurrency possible, we hold the required locks anyway
5624 * because of lock validation efforts.
5626 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5628 struct rq *rq = dead_rq;
5629 struct task_struct *next, *stop = rq->stop;
5630 struct rq_flags orf = *rf;
5634 * Fudge the rq selection such that the below task selection loop
5635 * doesn't get stuck on the currently eligible stop task.
5637 * We're currently inside stop_machine() and the rq is either stuck
5638 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5639 * either way we should never end up calling schedule() until we're
5645 * put_prev_task() and pick_next_task() sched
5646 * class method both need to have an up-to-date
5647 * value of rq->clock[_task]
5649 update_rq_clock(rq);
5653 * There's this thread running, bail when that's the only
5656 if (rq->nr_running == 1)
5660 * pick_next_task() assumes pinned rq->lock:
5662 next = pick_next_task(rq, &fake_task, rf);
5664 put_prev_task(rq, next);
5667 * Rules for changing task_struct::cpus_allowed are holding
5668 * both pi_lock and rq->lock, such that holding either
5669 * stabilizes the mask.
5671 * Drop rq->lock is not quite as disastrous as it usually is
5672 * because !cpu_active at this point, which means load-balance
5673 * will not interfere. Also, stop-machine.
5676 raw_spin_lock(&next->pi_lock);
5680 * Since we're inside stop-machine, _nothing_ should have
5681 * changed the task, WARN if weird stuff happened, because in
5682 * that case the above rq->lock drop is a fail too.
5684 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5685 raw_spin_unlock(&next->pi_lock);
5689 /* Find suitable destination for @next, with force if needed. */
5690 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5691 rq = __migrate_task(rq, rf, next, dest_cpu);
5692 if (rq != dead_rq) {
5698 raw_spin_unlock(&next->pi_lock);
5703 #endif /* CONFIG_HOTPLUG_CPU */
5705 void set_rq_online(struct rq *rq)
5708 const struct sched_class *class;
5710 cpumask_set_cpu(rq->cpu, rq->rd->online);
5713 for_each_class(class) {
5714 if (class->rq_online)
5715 class->rq_online(rq);
5720 void set_rq_offline(struct rq *rq)
5723 const struct sched_class *class;
5725 for_each_class(class) {
5726 if (class->rq_offline)
5727 class->rq_offline(rq);
5730 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5736 * used to mark begin/end of suspend/resume:
5738 static int num_cpus_frozen;
5741 * Update cpusets according to cpu_active mask. If cpusets are
5742 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5743 * around partition_sched_domains().
5745 * If we come here as part of a suspend/resume, don't touch cpusets because we
5746 * want to restore it back to its original state upon resume anyway.
5748 static void cpuset_cpu_active(void)
5750 if (cpuhp_tasks_frozen) {
5752 * num_cpus_frozen tracks how many CPUs are involved in suspend
5753 * resume sequence. As long as this is not the last online
5754 * operation in the resume sequence, just build a single sched
5755 * domain, ignoring cpusets.
5757 partition_sched_domains(1, NULL, NULL);
5758 if (--num_cpus_frozen)
5761 * This is the last CPU online operation. So fall through and
5762 * restore the original sched domains by considering the
5763 * cpuset configurations.
5765 cpuset_force_rebuild();
5767 cpuset_update_active_cpus();
5770 static int cpuset_cpu_inactive(unsigned int cpu)
5772 if (!cpuhp_tasks_frozen) {
5773 if (dl_cpu_busy(cpu))
5775 cpuset_update_active_cpus();
5778 partition_sched_domains(1, NULL, NULL);
5783 int sched_cpu_activate(unsigned int cpu)
5785 struct rq *rq = cpu_rq(cpu);
5788 #ifdef CONFIG_SCHED_SMT
5790 * When going up, increment the number of cores with SMT present.
5792 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
5793 static_branch_inc_cpuslocked(&sched_smt_present);
5795 set_cpu_active(cpu, true);
5797 if (sched_smp_initialized) {
5798 sched_domains_numa_masks_set(cpu);
5799 cpuset_cpu_active();
5803 * Put the rq online, if not already. This happens:
5805 * 1) In the early boot process, because we build the real domains
5806 * after all CPUs have been brought up.
5808 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5811 rq_lock_irqsave(rq, &rf);
5813 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5816 rq_unlock_irqrestore(rq, &rf);
5818 update_max_interval();
5823 int sched_cpu_deactivate(unsigned int cpu)
5827 set_cpu_active(cpu, false);
5829 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5830 * users of this state to go away such that all new such users will
5833 * Do sync before park smpboot threads to take care the rcu boost case.
5835 synchronize_rcu_mult(call_rcu, call_rcu_sched);
5837 #ifdef CONFIG_SCHED_SMT
5839 * When going down, decrement the number of cores with SMT present.
5841 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
5842 static_branch_dec_cpuslocked(&sched_smt_present);
5845 if (!sched_smp_initialized)
5848 ret = cpuset_cpu_inactive(cpu);
5850 set_cpu_active(cpu, true);
5853 sched_domains_numa_masks_clear(cpu);
5857 static void sched_rq_cpu_starting(unsigned int cpu)
5859 struct rq *rq = cpu_rq(cpu);
5861 rq->calc_load_update = calc_load_update;
5862 update_max_interval();
5865 int sched_cpu_starting(unsigned int cpu)
5867 sched_rq_cpu_starting(cpu);
5868 sched_tick_start(cpu);
5872 #ifdef CONFIG_HOTPLUG_CPU
5873 int sched_cpu_dying(unsigned int cpu)
5875 struct rq *rq = cpu_rq(cpu);
5878 /* Handle pending wakeups and then migrate everything off */
5879 sched_ttwu_pending();
5880 sched_tick_stop(cpu);
5882 rq_lock_irqsave(rq, &rf);
5884 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5887 migrate_tasks(rq, &rf);
5888 BUG_ON(rq->nr_running != 1);
5889 rq_unlock_irqrestore(rq, &rf);
5891 calc_load_migrate(rq);
5892 update_max_interval();
5893 nohz_balance_exit_idle(rq);
5899 void __init sched_init_smp(void)
5904 * There's no userspace yet to cause hotplug operations; hence all the
5905 * CPU masks are stable and all blatant races in the below code cannot
5906 * happen. The hotplug lock is nevertheless taken to satisfy lockdep,
5907 * but there won't be any contention on it.
5910 mutex_lock(&sched_domains_mutex);
5911 sched_init_domains(cpu_active_mask);
5912 mutex_unlock(&sched_domains_mutex);
5915 /* Move init over to a non-isolated CPU */
5916 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
5918 sched_init_granularity();
5920 init_sched_rt_class();
5921 init_sched_dl_class();
5923 sched_smp_initialized = true;
5926 static int __init migration_init(void)
5928 sched_rq_cpu_starting(smp_processor_id());
5931 early_initcall(migration_init);
5934 void __init sched_init_smp(void)
5936 sched_init_granularity();
5938 #endif /* CONFIG_SMP */
5940 int in_sched_functions(unsigned long addr)
5942 return in_lock_functions(addr) ||
5943 (addr >= (unsigned long)__sched_text_start
5944 && addr < (unsigned long)__sched_text_end);
5947 #ifdef CONFIG_CGROUP_SCHED
5949 * Default task group.
5950 * Every task in system belongs to this group at bootup.
5952 struct task_group root_task_group;
5953 LIST_HEAD(task_groups);
5955 /* Cacheline aligned slab cache for task_group */
5956 static struct kmem_cache *task_group_cache __read_mostly;
5959 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5960 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5962 void __init sched_init(void)
5965 unsigned long alloc_size = 0, ptr;
5969 #ifdef CONFIG_FAIR_GROUP_SCHED
5970 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5972 #ifdef CONFIG_RT_GROUP_SCHED
5973 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5976 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5978 #ifdef CONFIG_FAIR_GROUP_SCHED
5979 root_task_group.se = (struct sched_entity **)ptr;
5980 ptr += nr_cpu_ids * sizeof(void **);
5982 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5983 ptr += nr_cpu_ids * sizeof(void **);
5985 #endif /* CONFIG_FAIR_GROUP_SCHED */
5986 #ifdef CONFIG_RT_GROUP_SCHED
5987 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5988 ptr += nr_cpu_ids * sizeof(void **);
5990 root_task_group.rt_rq = (struct rt_rq **)ptr;
5991 ptr += nr_cpu_ids * sizeof(void **);
5993 #endif /* CONFIG_RT_GROUP_SCHED */
5995 #ifdef CONFIG_CPUMASK_OFFSTACK
5996 for_each_possible_cpu(i) {
5997 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5998 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5999 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
6000 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6002 #endif /* CONFIG_CPUMASK_OFFSTACK */
6004 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
6005 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6008 init_defrootdomain();
6011 #ifdef CONFIG_RT_GROUP_SCHED
6012 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6013 global_rt_period(), global_rt_runtime());
6014 #endif /* CONFIG_RT_GROUP_SCHED */
6016 #ifdef CONFIG_CGROUP_SCHED
6017 task_group_cache = KMEM_CACHE(task_group, 0);
6019 list_add(&root_task_group.list, &task_groups);
6020 INIT_LIST_HEAD(&root_task_group.children);
6021 INIT_LIST_HEAD(&root_task_group.siblings);
6022 autogroup_init(&init_task);
6023 #endif /* CONFIG_CGROUP_SCHED */
6025 for_each_possible_cpu(i) {
6029 raw_spin_lock_init(&rq->lock);
6031 rq->calc_load_active = 0;
6032 rq->calc_load_update = jiffies + LOAD_FREQ;
6033 init_cfs_rq(&rq->cfs);
6034 init_rt_rq(&rq->rt);
6035 init_dl_rq(&rq->dl);
6036 #ifdef CONFIG_FAIR_GROUP_SCHED
6037 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6038 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6039 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6041 * How much CPU bandwidth does root_task_group get?
6043 * In case of task-groups formed thr' the cgroup filesystem, it
6044 * gets 100% of the CPU resources in the system. This overall
6045 * system CPU resource is divided among the tasks of
6046 * root_task_group and its child task-groups in a fair manner,
6047 * based on each entity's (task or task-group's) weight
6048 * (se->load.weight).
6050 * In other words, if root_task_group has 10 tasks of weight
6051 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6052 * then A0's share of the CPU resource is:
6054 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6056 * We achieve this by letting root_task_group's tasks sit
6057 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6059 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6060 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6061 #endif /* CONFIG_FAIR_GROUP_SCHED */
6063 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6064 #ifdef CONFIG_RT_GROUP_SCHED
6065 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6068 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6069 rq->cpu_load[j] = 0;
6074 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6075 rq->balance_callback = NULL;
6076 rq->active_balance = 0;
6077 rq->next_balance = jiffies;
6082 rq->avg_idle = 2*sysctl_sched_migration_cost;
6083 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6085 INIT_LIST_HEAD(&rq->cfs_tasks);
6087 rq_attach_root(rq, &def_root_domain);
6088 #ifdef CONFIG_NO_HZ_COMMON
6089 rq->last_load_update_tick = jiffies;
6090 rq->last_blocked_load_update_tick = jiffies;
6091 atomic_set(&rq->nohz_flags, 0);
6093 #endif /* CONFIG_SMP */
6095 atomic_set(&rq->nr_iowait, 0);
6098 set_load_weight(&init_task, false);
6101 * The boot idle thread does lazy MMU switching as well:
6104 enter_lazy_tlb(&init_mm, current);
6107 * Make us the idle thread. Technically, schedule() should not be
6108 * called from this thread, however somewhere below it might be,
6109 * but because we are the idle thread, we just pick up running again
6110 * when this runqueue becomes "idle".
6112 init_idle(current, smp_processor_id());
6114 calc_load_update = jiffies + LOAD_FREQ;
6117 idle_thread_set_boot_cpu();
6119 init_sched_fair_class();
6123 scheduler_running = 1;
6126 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6127 static inline int preempt_count_equals(int preempt_offset)
6129 int nested = preempt_count() + rcu_preempt_depth();
6131 return (nested == preempt_offset);
6134 void __might_sleep(const char *file, int line, int preempt_offset)
6137 * Blocking primitives will set (and therefore destroy) current->state,
6138 * since we will exit with TASK_RUNNING make sure we enter with it,
6139 * otherwise we will destroy state.
6141 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6142 "do not call blocking ops when !TASK_RUNNING; "
6143 "state=%lx set at [<%p>] %pS\n",
6145 (void *)current->task_state_change,
6146 (void *)current->task_state_change);
6148 ___might_sleep(file, line, preempt_offset);
6150 EXPORT_SYMBOL(__might_sleep);
6152 void ___might_sleep(const char *file, int line, int preempt_offset)
6154 /* Ratelimiting timestamp: */
6155 static unsigned long prev_jiffy;
6157 unsigned long preempt_disable_ip;
6159 /* WARN_ON_ONCE() by default, no rate limit required: */
6162 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6163 !is_idle_task(current)) ||
6164 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6168 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6170 prev_jiffy = jiffies;
6172 /* Save this before calling printk(), since that will clobber it: */
6173 preempt_disable_ip = get_preempt_disable_ip(current);
6176 "BUG: sleeping function called from invalid context at %s:%d\n",
6179 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6180 in_atomic(), irqs_disabled(),
6181 current->pid, current->comm);
6183 if (task_stack_end_corrupted(current))
6184 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6186 debug_show_held_locks(current);
6187 if (irqs_disabled())
6188 print_irqtrace_events(current);
6189 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6190 && !preempt_count_equals(preempt_offset)) {
6191 pr_err("Preemption disabled at:");
6192 print_ip_sym(preempt_disable_ip);
6196 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6198 EXPORT_SYMBOL(___might_sleep);
6201 #ifdef CONFIG_MAGIC_SYSRQ
6202 void normalize_rt_tasks(void)
6204 struct task_struct *g, *p;
6205 struct sched_attr attr = {
6206 .sched_policy = SCHED_NORMAL,
6209 read_lock(&tasklist_lock);
6210 for_each_process_thread(g, p) {
6212 * Only normalize user tasks:
6214 if (p->flags & PF_KTHREAD)
6217 p->se.exec_start = 0;
6218 schedstat_set(p->se.statistics.wait_start, 0);
6219 schedstat_set(p->se.statistics.sleep_start, 0);
6220 schedstat_set(p->se.statistics.block_start, 0);
6222 if (!dl_task(p) && !rt_task(p)) {
6224 * Renice negative nice level userspace
6227 if (task_nice(p) < 0)
6228 set_user_nice(p, 0);
6232 __sched_setscheduler(p, &attr, false, false);
6234 read_unlock(&tasklist_lock);
6237 #endif /* CONFIG_MAGIC_SYSRQ */
6239 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6241 * These functions are only useful for the IA64 MCA handling, or kdb.
6243 * They can only be called when the whole system has been
6244 * stopped - every CPU needs to be quiescent, and no scheduling
6245 * activity can take place. Using them for anything else would
6246 * be a serious bug, and as a result, they aren't even visible
6247 * under any other configuration.
6251 * curr_task - return the current task for a given CPU.
6252 * @cpu: the processor in question.
6254 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6256 * Return: The current task for @cpu.
6258 struct task_struct *curr_task(int cpu)
6260 return cpu_curr(cpu);
6263 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6267 * set_curr_task - set the current task for a given CPU.
6268 * @cpu: the processor in question.
6269 * @p: the task pointer to set.
6271 * Description: This function must only be used when non-maskable interrupts
6272 * are serviced on a separate stack. It allows the architecture to switch the
6273 * notion of the current task on a CPU in a non-blocking manner. This function
6274 * must be called with all CPU's synchronized, and interrupts disabled, the
6275 * and caller must save the original value of the current task (see
6276 * curr_task() above) and restore that value before reenabling interrupts and
6277 * re-starting the system.
6279 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6281 void ia64_set_curr_task(int cpu, struct task_struct *p)
6288 #ifdef CONFIG_CGROUP_SCHED
6289 /* task_group_lock serializes the addition/removal of task groups */
6290 static DEFINE_SPINLOCK(task_group_lock);
6292 static void sched_free_group(struct task_group *tg)
6294 free_fair_sched_group(tg);
6295 free_rt_sched_group(tg);
6297 kmem_cache_free(task_group_cache, tg);
6300 /* allocate runqueue etc for a new task group */
6301 struct task_group *sched_create_group(struct task_group *parent)
6303 struct task_group *tg;
6305 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6307 return ERR_PTR(-ENOMEM);
6309 if (!alloc_fair_sched_group(tg, parent))
6312 if (!alloc_rt_sched_group(tg, parent))
6318 sched_free_group(tg);
6319 return ERR_PTR(-ENOMEM);
6322 void sched_online_group(struct task_group *tg, struct task_group *parent)
6324 unsigned long flags;
6326 spin_lock_irqsave(&task_group_lock, flags);
6327 list_add_rcu(&tg->list, &task_groups);
6329 /* Root should already exist: */
6332 tg->parent = parent;
6333 INIT_LIST_HEAD(&tg->children);
6334 list_add_rcu(&tg->siblings, &parent->children);
6335 spin_unlock_irqrestore(&task_group_lock, flags);
6337 online_fair_sched_group(tg);
6340 /* rcu callback to free various structures associated with a task group */
6341 static void sched_free_group_rcu(struct rcu_head *rhp)
6343 /* Now it should be safe to free those cfs_rqs: */
6344 sched_free_group(container_of(rhp, struct task_group, rcu));
6347 void sched_destroy_group(struct task_group *tg)
6349 /* Wait for possible concurrent references to cfs_rqs complete: */
6350 call_rcu(&tg->rcu, sched_free_group_rcu);
6353 void sched_offline_group(struct task_group *tg)
6355 unsigned long flags;
6357 /* End participation in shares distribution: */
6358 unregister_fair_sched_group(tg);
6360 spin_lock_irqsave(&task_group_lock, flags);
6361 list_del_rcu(&tg->list);
6362 list_del_rcu(&tg->siblings);
6363 spin_unlock_irqrestore(&task_group_lock, flags);
6366 static void sched_change_group(struct task_struct *tsk, int type)
6368 struct task_group *tg;
6371 * All callers are synchronized by task_rq_lock(); we do not use RCU
6372 * which is pointless here. Thus, we pass "true" to task_css_check()
6373 * to prevent lockdep warnings.
6375 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6376 struct task_group, css);
6377 tg = autogroup_task_group(tsk, tg);
6378 tsk->sched_task_group = tg;
6380 #ifdef CONFIG_FAIR_GROUP_SCHED
6381 if (tsk->sched_class->task_change_group)
6382 tsk->sched_class->task_change_group(tsk, type);
6385 set_task_rq(tsk, task_cpu(tsk));
6389 * Change task's runqueue when it moves between groups.
6391 * The caller of this function should have put the task in its new group by
6392 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6395 void sched_move_task(struct task_struct *tsk)
6397 int queued, running, queue_flags =
6398 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6402 rq = task_rq_lock(tsk, &rf);
6403 update_rq_clock(rq);
6405 running = task_current(rq, tsk);
6406 queued = task_on_rq_queued(tsk);
6409 dequeue_task(rq, tsk, queue_flags);
6411 put_prev_task(rq, tsk);
6413 sched_change_group(tsk, TASK_MOVE_GROUP);
6416 enqueue_task(rq, tsk, queue_flags);
6418 set_curr_task(rq, tsk);
6420 task_rq_unlock(rq, tsk, &rf);
6423 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6425 return css ? container_of(css, struct task_group, css) : NULL;
6428 static struct cgroup_subsys_state *
6429 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6431 struct task_group *parent = css_tg(parent_css);
6432 struct task_group *tg;
6435 /* This is early initialization for the top cgroup */
6436 return &root_task_group.css;
6439 tg = sched_create_group(parent);
6441 return ERR_PTR(-ENOMEM);
6446 /* Expose task group only after completing cgroup initialization */
6447 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6449 struct task_group *tg = css_tg(css);
6450 struct task_group *parent = css_tg(css->parent);
6453 sched_online_group(tg, parent);
6457 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6459 struct task_group *tg = css_tg(css);
6461 sched_offline_group(tg);
6464 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6466 struct task_group *tg = css_tg(css);
6469 * Relies on the RCU grace period between css_released() and this.
6471 sched_free_group(tg);
6475 * This is called before wake_up_new_task(), therefore we really only
6476 * have to set its group bits, all the other stuff does not apply.
6478 static void cpu_cgroup_fork(struct task_struct *task)
6483 rq = task_rq_lock(task, &rf);
6485 update_rq_clock(rq);
6486 sched_change_group(task, TASK_SET_GROUP);
6488 task_rq_unlock(rq, task, &rf);
6491 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6493 struct task_struct *task;
6494 struct cgroup_subsys_state *css;
6497 cgroup_taskset_for_each(task, css, tset) {
6498 #ifdef CONFIG_RT_GROUP_SCHED
6499 if (!sched_rt_can_attach(css_tg(css), task))
6503 * Serialize against wake_up_new_task() such that if its
6504 * running, we're sure to observe its full state.
6506 raw_spin_lock_irq(&task->pi_lock);
6508 * Avoid calling sched_move_task() before wake_up_new_task()
6509 * has happened. This would lead to problems with PELT, due to
6510 * move wanting to detach+attach while we're not attached yet.
6512 if (task->state == TASK_NEW)
6514 raw_spin_unlock_irq(&task->pi_lock);
6522 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6524 struct task_struct *task;
6525 struct cgroup_subsys_state *css;
6527 cgroup_taskset_for_each(task, css, tset)
6528 sched_move_task(task);
6531 #ifdef CONFIG_FAIR_GROUP_SCHED
6532 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6533 struct cftype *cftype, u64 shareval)
6535 if (shareval > scale_load_down(ULONG_MAX))
6536 shareval = MAX_SHARES;
6537 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6540 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6543 struct task_group *tg = css_tg(css);
6545 return (u64) scale_load_down(tg->shares);
6548 #ifdef CONFIG_CFS_BANDWIDTH
6549 static DEFINE_MUTEX(cfs_constraints_mutex);
6551 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6552 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6554 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6556 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6558 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6559 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6561 if (tg == &root_task_group)
6565 * Ensure we have at some amount of bandwidth every period. This is
6566 * to prevent reaching a state of large arrears when throttled via
6567 * entity_tick() resulting in prolonged exit starvation.
6569 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6573 * Likewise, bound things on the otherside by preventing insane quota
6574 * periods. This also allows us to normalize in computing quota
6577 if (period > max_cfs_quota_period)
6581 * Prevent race between setting of cfs_rq->runtime_enabled and
6582 * unthrottle_offline_cfs_rqs().
6585 mutex_lock(&cfs_constraints_mutex);
6586 ret = __cfs_schedulable(tg, period, quota);
6590 runtime_enabled = quota != RUNTIME_INF;
6591 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6593 * If we need to toggle cfs_bandwidth_used, off->on must occur
6594 * before making related changes, and on->off must occur afterwards
6596 if (runtime_enabled && !runtime_was_enabled)
6597 cfs_bandwidth_usage_inc();
6598 raw_spin_lock_irq(&cfs_b->lock);
6599 cfs_b->period = ns_to_ktime(period);
6600 cfs_b->quota = quota;
6602 __refill_cfs_bandwidth_runtime(cfs_b);
6604 /* Restart the period timer (if active) to handle new period expiry: */
6605 if (runtime_enabled)
6606 start_cfs_bandwidth(cfs_b);
6608 raw_spin_unlock_irq(&cfs_b->lock);
6610 for_each_online_cpu(i) {
6611 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6612 struct rq *rq = cfs_rq->rq;
6615 rq_lock_irq(rq, &rf);
6616 cfs_rq->runtime_enabled = runtime_enabled;
6617 cfs_rq->runtime_remaining = 0;
6619 if (cfs_rq->throttled)
6620 unthrottle_cfs_rq(cfs_rq);
6621 rq_unlock_irq(rq, &rf);
6623 if (runtime_was_enabled && !runtime_enabled)
6624 cfs_bandwidth_usage_dec();
6626 mutex_unlock(&cfs_constraints_mutex);
6632 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
6636 period = ktime_to_ns(tg->cfs_bandwidth.period);
6637 if (cfs_quota_us < 0)
6638 quota = RUNTIME_INF;
6639 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
6640 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
6644 return tg_set_cfs_bandwidth(tg, period, quota);
6647 long tg_get_cfs_quota(struct task_group *tg)
6651 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
6654 quota_us = tg->cfs_bandwidth.quota;
6655 do_div(quota_us, NSEC_PER_USEC);
6660 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
6664 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
6667 period = (u64)cfs_period_us * NSEC_PER_USEC;
6668 quota = tg->cfs_bandwidth.quota;
6670 return tg_set_cfs_bandwidth(tg, period, quota);
6673 long tg_get_cfs_period(struct task_group *tg)
6677 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
6678 do_div(cfs_period_us, NSEC_PER_USEC);
6680 return cfs_period_us;
6683 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
6686 return tg_get_cfs_quota(css_tg(css));
6689 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
6690 struct cftype *cftype, s64 cfs_quota_us)
6692 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
6695 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
6698 return tg_get_cfs_period(css_tg(css));
6701 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
6702 struct cftype *cftype, u64 cfs_period_us)
6704 return tg_set_cfs_period(css_tg(css), cfs_period_us);
6707 struct cfs_schedulable_data {
6708 struct task_group *tg;
6713 * normalize group quota/period to be quota/max_period
6714 * note: units are usecs
6716 static u64 normalize_cfs_quota(struct task_group *tg,
6717 struct cfs_schedulable_data *d)
6725 period = tg_get_cfs_period(tg);
6726 quota = tg_get_cfs_quota(tg);
6729 /* note: these should typically be equivalent */
6730 if (quota == RUNTIME_INF || quota == -1)
6733 return to_ratio(period, quota);
6736 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
6738 struct cfs_schedulable_data *d = data;
6739 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6740 s64 quota = 0, parent_quota = -1;
6743 quota = RUNTIME_INF;
6745 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
6747 quota = normalize_cfs_quota(tg, d);
6748 parent_quota = parent_b->hierarchical_quota;
6751 * Ensure max(child_quota) <= parent_quota. On cgroup2,
6752 * always take the min. On cgroup1, only inherit when no
6755 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
6756 quota = min(quota, parent_quota);
6758 if (quota == RUNTIME_INF)
6759 quota = parent_quota;
6760 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
6764 cfs_b->hierarchical_quota = quota;
6769 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
6772 struct cfs_schedulable_data data = {
6778 if (quota != RUNTIME_INF) {
6779 do_div(data.period, NSEC_PER_USEC);
6780 do_div(data.quota, NSEC_PER_USEC);
6784 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
6790 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
6792 struct task_group *tg = css_tg(seq_css(sf));
6793 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6795 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
6796 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
6797 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
6799 if (schedstat_enabled() && tg != &root_task_group) {
6803 for_each_possible_cpu(i)
6804 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
6806 seq_printf(sf, "wait_sum %llu\n", ws);
6811 #endif /* CONFIG_CFS_BANDWIDTH */
6812 #endif /* CONFIG_FAIR_GROUP_SCHED */
6814 #ifdef CONFIG_RT_GROUP_SCHED
6815 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
6816 struct cftype *cft, s64 val)
6818 return sched_group_set_rt_runtime(css_tg(css), val);
6821 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
6824 return sched_group_rt_runtime(css_tg(css));
6827 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
6828 struct cftype *cftype, u64 rt_period_us)
6830 return sched_group_set_rt_period(css_tg(css), rt_period_us);
6833 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
6836 return sched_group_rt_period(css_tg(css));
6838 #endif /* CONFIG_RT_GROUP_SCHED */
6840 static struct cftype cpu_legacy_files[] = {
6841 #ifdef CONFIG_FAIR_GROUP_SCHED
6844 .read_u64 = cpu_shares_read_u64,
6845 .write_u64 = cpu_shares_write_u64,
6848 #ifdef CONFIG_CFS_BANDWIDTH
6850 .name = "cfs_quota_us",
6851 .read_s64 = cpu_cfs_quota_read_s64,
6852 .write_s64 = cpu_cfs_quota_write_s64,
6855 .name = "cfs_period_us",
6856 .read_u64 = cpu_cfs_period_read_u64,
6857 .write_u64 = cpu_cfs_period_write_u64,
6861 .seq_show = cpu_cfs_stat_show,
6864 #ifdef CONFIG_RT_GROUP_SCHED
6866 .name = "rt_runtime_us",
6867 .read_s64 = cpu_rt_runtime_read,
6868 .write_s64 = cpu_rt_runtime_write,
6871 .name = "rt_period_us",
6872 .read_u64 = cpu_rt_period_read_uint,
6873 .write_u64 = cpu_rt_period_write_uint,
6879 static int cpu_extra_stat_show(struct seq_file *sf,
6880 struct cgroup_subsys_state *css)
6882 #ifdef CONFIG_CFS_BANDWIDTH
6884 struct task_group *tg = css_tg(css);
6885 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6888 throttled_usec = cfs_b->throttled_time;
6889 do_div(throttled_usec, NSEC_PER_USEC);
6891 seq_printf(sf, "nr_periods %d\n"
6893 "throttled_usec %llu\n",
6894 cfs_b->nr_periods, cfs_b->nr_throttled,
6901 #ifdef CONFIG_FAIR_GROUP_SCHED
6902 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
6905 struct task_group *tg = css_tg(css);
6906 u64 weight = scale_load_down(tg->shares);
6908 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
6911 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
6912 struct cftype *cft, u64 weight)
6915 * cgroup weight knobs should use the common MIN, DFL and MAX
6916 * values which are 1, 100 and 10000 respectively. While it loses
6917 * a bit of range on both ends, it maps pretty well onto the shares
6918 * value used by scheduler and the round-trip conversions preserve
6919 * the original value over the entire range.
6921 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
6924 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
6926 return sched_group_set_shares(css_tg(css), scale_load(weight));
6929 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
6932 unsigned long weight = scale_load_down(css_tg(css)->shares);
6933 int last_delta = INT_MAX;
6936 /* find the closest nice value to the current weight */
6937 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
6938 delta = abs(sched_prio_to_weight[prio] - weight);
6939 if (delta >= last_delta)
6944 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
6947 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
6948 struct cftype *cft, s64 nice)
6950 unsigned long weight;
6953 if (nice < MIN_NICE || nice > MAX_NICE)
6956 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
6957 idx = array_index_nospec(idx, 40);
6958 weight = sched_prio_to_weight[idx];
6960 return sched_group_set_shares(css_tg(css), scale_load(weight));
6964 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
6965 long period, long quota)
6968 seq_puts(sf, "max");
6970 seq_printf(sf, "%ld", quota);
6972 seq_printf(sf, " %ld\n", period);
6975 /* caller should put the current value in *@periodp before calling */
6976 static int __maybe_unused cpu_period_quota_parse(char *buf,
6977 u64 *periodp, u64 *quotap)
6979 char tok[21]; /* U64_MAX */
6981 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
6984 *periodp *= NSEC_PER_USEC;
6986 if (sscanf(tok, "%llu", quotap))
6987 *quotap *= NSEC_PER_USEC;
6988 else if (!strcmp(tok, "max"))
6989 *quotap = RUNTIME_INF;
6996 #ifdef CONFIG_CFS_BANDWIDTH
6997 static int cpu_max_show(struct seq_file *sf, void *v)
6999 struct task_group *tg = css_tg(seq_css(sf));
7001 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
7005 static ssize_t cpu_max_write(struct kernfs_open_file *of,
7006 char *buf, size_t nbytes, loff_t off)
7008 struct task_group *tg = css_tg(of_css(of));
7009 u64 period = tg_get_cfs_period(tg);
7013 ret = cpu_period_quota_parse(buf, &period, "a);
7015 ret = tg_set_cfs_bandwidth(tg, period, quota);
7016 return ret ?: nbytes;
7020 static struct cftype cpu_files[] = {
7021 #ifdef CONFIG_FAIR_GROUP_SCHED
7024 .flags = CFTYPE_NOT_ON_ROOT,
7025 .read_u64 = cpu_weight_read_u64,
7026 .write_u64 = cpu_weight_write_u64,
7029 .name = "weight.nice",
7030 .flags = CFTYPE_NOT_ON_ROOT,
7031 .read_s64 = cpu_weight_nice_read_s64,
7032 .write_s64 = cpu_weight_nice_write_s64,
7035 #ifdef CONFIG_CFS_BANDWIDTH
7038 .flags = CFTYPE_NOT_ON_ROOT,
7039 .seq_show = cpu_max_show,
7040 .write = cpu_max_write,
7046 struct cgroup_subsys cpu_cgrp_subsys = {
7047 .css_alloc = cpu_cgroup_css_alloc,
7048 .css_online = cpu_cgroup_css_online,
7049 .css_released = cpu_cgroup_css_released,
7050 .css_free = cpu_cgroup_css_free,
7051 .css_extra_stat_show = cpu_extra_stat_show,
7052 .fork = cpu_cgroup_fork,
7053 .can_attach = cpu_cgroup_can_attach,
7054 .attach = cpu_cgroup_attach,
7055 .legacy_cftypes = cpu_legacy_files,
7056 .dfl_cftypes = cpu_files,
7061 #endif /* CONFIG_CGROUP_SCHED */
7063 void dump_cpu_task(int cpu)
7065 pr_info("Task dump for CPU %d:\n", cpu);
7066 sched_show_task(cpu_curr(cpu));
7070 * Nice levels are multiplicative, with a gentle 10% change for every
7071 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7072 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7073 * that remained on nice 0.
7075 * The "10% effect" is relative and cumulative: from _any_ nice level,
7076 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7077 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7078 * If a task goes up by ~10% and another task goes down by ~10% then
7079 * the relative distance between them is ~25%.)
7081 const int sched_prio_to_weight[40] = {
7082 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7083 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7084 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7085 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7086 /* 0 */ 1024, 820, 655, 526, 423,
7087 /* 5 */ 335, 272, 215, 172, 137,
7088 /* 10 */ 110, 87, 70, 56, 45,
7089 /* 15 */ 36, 29, 23, 18, 15,
7093 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7095 * In cases where the weight does not change often, we can use the
7096 * precalculated inverse to speed up arithmetics by turning divisions
7097 * into multiplications:
7099 const u32 sched_prio_to_wmult[40] = {
7100 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7101 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7102 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7103 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7104 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7105 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7106 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7107 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7110 #undef CREATE_TRACE_POINTS