4 * Core kernel scheduler code and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 #include <linux/sched.h>
9 #include <linux/sched/clock.h>
10 #include <uapi/linux/sched/types.h>
11 #include <linux/sched/loadavg.h>
12 #include <linux/sched/hotplug.h>
13 #include <linux/wait_bit.h>
14 #include <linux/cpuset.h>
15 #include <linux/delayacct.h>
16 #include <linux/init_task.h>
17 #include <linux/context_tracking.h>
18 #include <linux/rcupdate_wait.h>
19 #include <linux/compat.h>
21 #include <linux/blkdev.h>
22 #include <linux/kprobes.h>
23 #include <linux/mmu_context.h>
24 #include <linux/module.h>
25 #include <linux/nmi.h>
26 #include <linux/prefetch.h>
27 #include <linux/profile.h>
28 #include <linux/security.h>
29 #include <linux/syscalls.h>
30 #include <linux/sched/isolation.h>
32 #include <asm/switch_to.h>
34 #ifdef CONFIG_PARAVIRT
35 #include <asm/paravirt.h>
39 #include "../workqueue_internal.h"
40 #include "../smpboot.h"
42 #define CREATE_TRACE_POINTS
43 #include <trace/events/sched.h>
45 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
47 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
49 * Debugging: various feature bits
51 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
52 * sysctl_sched_features, defined in sched.h, to allow constants propagation
53 * at compile time and compiler optimization based on features default.
55 #define SCHED_FEAT(name, enabled) \
56 (1UL << __SCHED_FEAT_##name) * enabled |
57 const_debug unsigned int sysctl_sched_features =
64 * Number of tasks to iterate in a single balance run.
65 * Limited because this is done with IRQs disabled.
67 const_debug unsigned int sysctl_sched_nr_migrate = 32;
70 * period over which we average the RT time consumption, measured
75 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
78 * period over which we measure -rt task CPU usage in us.
81 unsigned int sysctl_sched_rt_period = 1000000;
83 __read_mostly int scheduler_running;
86 * part of the period that we allow rt tasks to run in us.
89 int sysctl_sched_rt_runtime = 950000;
92 * __task_rq_lock - lock the rq @p resides on.
94 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
99 lockdep_assert_held(&p->pi_lock);
103 raw_spin_lock(&rq->lock);
104 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
108 raw_spin_unlock(&rq->lock);
110 while (unlikely(task_on_rq_migrating(p)))
116 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
118 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
119 __acquires(p->pi_lock)
125 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
127 raw_spin_lock(&rq->lock);
129 * move_queued_task() task_rq_lock()
132 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
133 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
134 * [S] ->cpu = new_cpu [L] task_rq()
138 * If we observe the old cpu in task_rq_lock, the acquire of
139 * the old rq->lock will fully serialize against the stores.
141 * If we observe the new CPU in task_rq_lock, the acquire will
142 * pair with the WMB to ensure we must then also see migrating.
144 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
148 raw_spin_unlock(&rq->lock);
149 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
151 while (unlikely(task_on_rq_migrating(p)))
157 * RQ-clock updating methods:
160 static void update_rq_clock_task(struct rq *rq, s64 delta)
163 * In theory, the compile should just see 0 here, and optimize out the call
164 * to sched_rt_avg_update. But I don't trust it...
166 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
167 s64 steal = 0, irq_delta = 0;
169 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
170 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
173 * Since irq_time is only updated on {soft,}irq_exit, we might run into
174 * this case when a previous update_rq_clock() happened inside a
177 * When this happens, we stop ->clock_task and only update the
178 * prev_irq_time stamp to account for the part that fit, so that a next
179 * update will consume the rest. This ensures ->clock_task is
182 * It does however cause some slight miss-attribution of {soft,}irq
183 * time, a more accurate solution would be to update the irq_time using
184 * the current rq->clock timestamp, except that would require using
187 if (irq_delta > delta)
190 rq->prev_irq_time += irq_delta;
193 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
194 if (static_key_false((¶virt_steal_rq_enabled))) {
195 steal = paravirt_steal_clock(cpu_of(rq));
196 steal -= rq->prev_steal_time_rq;
198 if (unlikely(steal > delta))
201 rq->prev_steal_time_rq += steal;
206 rq->clock_task += delta;
208 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
209 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
210 sched_rt_avg_update(rq, irq_delta + steal);
214 void update_rq_clock(struct rq *rq)
218 lockdep_assert_held(&rq->lock);
220 if (rq->clock_update_flags & RQCF_ACT_SKIP)
223 #ifdef CONFIG_SCHED_DEBUG
224 if (sched_feat(WARN_DOUBLE_CLOCK))
225 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
226 rq->clock_update_flags |= RQCF_UPDATED;
229 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
233 update_rq_clock_task(rq, delta);
237 #ifdef CONFIG_SCHED_HRTICK
239 * Use HR-timers to deliver accurate preemption points.
242 static void hrtick_clear(struct rq *rq)
244 if (hrtimer_active(&rq->hrtick_timer))
245 hrtimer_cancel(&rq->hrtick_timer);
249 * High-resolution timer tick.
250 * Runs from hardirq context with interrupts disabled.
252 static enum hrtimer_restart hrtick(struct hrtimer *timer)
254 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
257 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
261 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
264 return HRTIMER_NORESTART;
269 static void __hrtick_restart(struct rq *rq)
271 struct hrtimer *timer = &rq->hrtick_timer;
273 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
277 * called from hardirq (IPI) context
279 static void __hrtick_start(void *arg)
285 __hrtick_restart(rq);
286 rq->hrtick_csd_pending = 0;
291 * Called to set the hrtick timer state.
293 * called with rq->lock held and irqs disabled
295 void hrtick_start(struct rq *rq, u64 delay)
297 struct hrtimer *timer = &rq->hrtick_timer;
302 * Don't schedule slices shorter than 10000ns, that just
303 * doesn't make sense and can cause timer DoS.
305 delta = max_t(s64, delay, 10000LL);
306 time = ktime_add_ns(timer->base->get_time(), delta);
308 hrtimer_set_expires(timer, time);
310 if (rq == this_rq()) {
311 __hrtick_restart(rq);
312 } else if (!rq->hrtick_csd_pending) {
313 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
314 rq->hrtick_csd_pending = 1;
320 * Called to set the hrtick timer state.
322 * called with rq->lock held and irqs disabled
324 void hrtick_start(struct rq *rq, u64 delay)
327 * Don't schedule slices shorter than 10000ns, that just
328 * doesn't make sense. Rely on vruntime for fairness.
330 delay = max_t(u64, delay, 10000LL);
331 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
332 HRTIMER_MODE_REL_PINNED);
334 #endif /* CONFIG_SMP */
336 static void init_rq_hrtick(struct rq *rq)
339 rq->hrtick_csd_pending = 0;
341 rq->hrtick_csd.flags = 0;
342 rq->hrtick_csd.func = __hrtick_start;
343 rq->hrtick_csd.info = rq;
346 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
347 rq->hrtick_timer.function = hrtick;
349 #else /* CONFIG_SCHED_HRTICK */
350 static inline void hrtick_clear(struct rq *rq)
354 static inline void init_rq_hrtick(struct rq *rq)
357 #endif /* CONFIG_SCHED_HRTICK */
360 * cmpxchg based fetch_or, macro so it works for different integer types
362 #define fetch_or(ptr, mask) \
364 typeof(ptr) _ptr = (ptr); \
365 typeof(mask) _mask = (mask); \
366 typeof(*_ptr) _old, _val = *_ptr; \
369 _old = cmpxchg(_ptr, _val, _val | _mask); \
377 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
379 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
380 * this avoids any races wrt polling state changes and thereby avoids
383 static bool set_nr_and_not_polling(struct task_struct *p)
385 struct thread_info *ti = task_thread_info(p);
386 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
390 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
392 * If this returns true, then the idle task promises to call
393 * sched_ttwu_pending() and reschedule soon.
395 static bool set_nr_if_polling(struct task_struct *p)
397 struct thread_info *ti = task_thread_info(p);
398 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
401 if (!(val & _TIF_POLLING_NRFLAG))
403 if (val & _TIF_NEED_RESCHED)
405 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
414 static bool set_nr_and_not_polling(struct task_struct *p)
416 set_tsk_need_resched(p);
421 static bool set_nr_if_polling(struct task_struct *p)
428 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
430 struct wake_q_node *node = &task->wake_q;
433 * Atomically grab the task, if ->wake_q is !nil already it means
434 * its already queued (either by us or someone else) and will get the
435 * wakeup due to that.
437 * This cmpxchg() implies a full barrier, which pairs with the write
438 * barrier implied by the wakeup in wake_up_q().
440 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
443 get_task_struct(task);
446 * The head is context local, there can be no concurrency.
449 head->lastp = &node->next;
452 void wake_up_q(struct wake_q_head *head)
454 struct wake_q_node *node = head->first;
456 while (node != WAKE_Q_TAIL) {
457 struct task_struct *task;
459 task = container_of(node, struct task_struct, wake_q);
461 /* Task can safely be re-inserted now: */
463 task->wake_q.next = NULL;
466 * wake_up_process() implies a wmb() to pair with the queueing
467 * in wake_q_add() so as not to miss wakeups.
469 wake_up_process(task);
470 put_task_struct(task);
475 * resched_curr - mark rq's current task 'to be rescheduled now'.
477 * On UP this means the setting of the need_resched flag, on SMP it
478 * might also involve a cross-CPU call to trigger the scheduler on
481 void resched_curr(struct rq *rq)
483 struct task_struct *curr = rq->curr;
486 lockdep_assert_held(&rq->lock);
488 if (test_tsk_need_resched(curr))
493 if (cpu == smp_processor_id()) {
494 set_tsk_need_resched(curr);
495 set_preempt_need_resched();
499 if (set_nr_and_not_polling(curr))
500 smp_send_reschedule(cpu);
502 trace_sched_wake_idle_without_ipi(cpu);
505 void resched_cpu(int cpu)
507 struct rq *rq = cpu_rq(cpu);
510 raw_spin_lock_irqsave(&rq->lock, flags);
511 if (cpu_online(cpu) || cpu == smp_processor_id())
513 raw_spin_unlock_irqrestore(&rq->lock, flags);
517 #ifdef CONFIG_NO_HZ_COMMON
519 * In the semi idle case, use the nearest busy CPU for migrating timers
520 * from an idle CPU. This is good for power-savings.
522 * We don't do similar optimization for completely idle system, as
523 * selecting an idle CPU will add more delays to the timers than intended
524 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
526 int get_nohz_timer_target(void)
528 int i, cpu = smp_processor_id();
529 struct sched_domain *sd;
531 if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
535 for_each_domain(cpu, sd) {
536 for_each_cpu(i, sched_domain_span(sd)) {
540 if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
547 if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
548 cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
555 * When add_timer_on() enqueues a timer into the timer wheel of an
556 * idle CPU then this timer might expire before the next timer event
557 * which is scheduled to wake up that CPU. In case of a completely
558 * idle system the next event might even be infinite time into the
559 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
560 * leaves the inner idle loop so the newly added timer is taken into
561 * account when the CPU goes back to idle and evaluates the timer
562 * wheel for the next timer event.
564 static void wake_up_idle_cpu(int cpu)
566 struct rq *rq = cpu_rq(cpu);
568 if (cpu == smp_processor_id())
571 if (set_nr_and_not_polling(rq->idle))
572 smp_send_reschedule(cpu);
574 trace_sched_wake_idle_without_ipi(cpu);
577 static bool wake_up_full_nohz_cpu(int cpu)
580 * We just need the target to call irq_exit() and re-evaluate
581 * the next tick. The nohz full kick at least implies that.
582 * If needed we can still optimize that later with an
585 if (cpu_is_offline(cpu))
586 return true; /* Don't try to wake offline CPUs. */
587 if (tick_nohz_full_cpu(cpu)) {
588 if (cpu != smp_processor_id() ||
589 tick_nohz_tick_stopped())
590 tick_nohz_full_kick_cpu(cpu);
598 * Wake up the specified CPU. If the CPU is going offline, it is the
599 * caller's responsibility to deal with the lost wakeup, for example,
600 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
602 void wake_up_nohz_cpu(int cpu)
604 if (!wake_up_full_nohz_cpu(cpu))
605 wake_up_idle_cpu(cpu);
608 static inline bool got_nohz_idle_kick(void)
610 int cpu = smp_processor_id();
612 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
615 if (idle_cpu(cpu) && !need_resched())
619 * We can't run Idle Load Balance on this CPU for this time so we
620 * cancel it and clear NOHZ_BALANCE_KICK
622 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
626 #else /* CONFIG_NO_HZ_COMMON */
628 static inline bool got_nohz_idle_kick(void)
633 #endif /* CONFIG_NO_HZ_COMMON */
635 #ifdef CONFIG_NO_HZ_FULL
636 bool sched_can_stop_tick(struct rq *rq)
640 /* Deadline tasks, even if single, need the tick */
641 if (rq->dl.dl_nr_running)
645 * If there are more than one RR tasks, we need the tick to effect the
646 * actual RR behaviour.
648 if (rq->rt.rr_nr_running) {
649 if (rq->rt.rr_nr_running == 1)
656 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
657 * forced preemption between FIFO tasks.
659 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
664 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
665 * if there's more than one we need the tick for involuntary
668 if (rq->nr_running > 1)
673 #endif /* CONFIG_NO_HZ_FULL */
675 void sched_avg_update(struct rq *rq)
677 s64 period = sched_avg_period();
679 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
681 * Inline assembly required to prevent the compiler
682 * optimising this loop into a divmod call.
683 * See __iter_div_u64_rem() for another example of this.
685 asm("" : "+rm" (rq->age_stamp));
686 rq->age_stamp += period;
691 #endif /* CONFIG_SMP */
693 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
694 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
696 * Iterate task_group tree rooted at *from, calling @down when first entering a
697 * node and @up when leaving it for the final time.
699 * Caller must hold rcu_lock or sufficient equivalent.
701 int walk_tg_tree_from(struct task_group *from,
702 tg_visitor down, tg_visitor up, void *data)
704 struct task_group *parent, *child;
710 ret = (*down)(parent, data);
713 list_for_each_entry_rcu(child, &parent->children, siblings) {
720 ret = (*up)(parent, data);
721 if (ret || parent == from)
725 parent = parent->parent;
732 int tg_nop(struct task_group *tg, void *data)
738 static void set_load_weight(struct task_struct *p, bool update_load)
740 int prio = p->static_prio - MAX_RT_PRIO;
741 struct load_weight *load = &p->se.load;
744 * SCHED_IDLE tasks get minimal weight:
746 if (idle_policy(p->policy)) {
747 load->weight = scale_load(WEIGHT_IDLEPRIO);
748 load->inv_weight = WMULT_IDLEPRIO;
753 * SCHED_OTHER tasks have to update their load when changing their
756 if (update_load && p->sched_class == &fair_sched_class) {
757 reweight_task(p, prio);
759 load->weight = scale_load(sched_prio_to_weight[prio]);
760 load->inv_weight = sched_prio_to_wmult[prio];
764 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
766 if (!(flags & ENQUEUE_NOCLOCK))
769 if (!(flags & ENQUEUE_RESTORE))
770 sched_info_queued(rq, p);
772 p->sched_class->enqueue_task(rq, p, flags);
775 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
777 if (!(flags & DEQUEUE_NOCLOCK))
780 if (!(flags & DEQUEUE_SAVE))
781 sched_info_dequeued(rq, p);
783 p->sched_class->dequeue_task(rq, p, flags);
786 void activate_task(struct rq *rq, struct task_struct *p, int flags)
788 if (task_contributes_to_load(p))
789 rq->nr_uninterruptible--;
791 enqueue_task(rq, p, flags);
794 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
796 if (task_contributes_to_load(p))
797 rq->nr_uninterruptible++;
799 dequeue_task(rq, p, flags);
803 * __normal_prio - return the priority that is based on the static prio
805 static inline int __normal_prio(struct task_struct *p)
807 return p->static_prio;
811 * Calculate the expected normal priority: i.e. priority
812 * without taking RT-inheritance into account. Might be
813 * boosted by interactivity modifiers. Changes upon fork,
814 * setprio syscalls, and whenever the interactivity
815 * estimator recalculates.
817 static inline int normal_prio(struct task_struct *p)
821 if (task_has_dl_policy(p))
822 prio = MAX_DL_PRIO-1;
823 else if (task_has_rt_policy(p))
824 prio = MAX_RT_PRIO-1 - p->rt_priority;
826 prio = __normal_prio(p);
831 * Calculate the current priority, i.e. the priority
832 * taken into account by the scheduler. This value might
833 * be boosted by RT tasks, or might be boosted by
834 * interactivity modifiers. Will be RT if the task got
835 * RT-boosted. If not then it returns p->normal_prio.
837 static int effective_prio(struct task_struct *p)
839 p->normal_prio = normal_prio(p);
841 * If we are RT tasks or we were boosted to RT priority,
842 * keep the priority unchanged. Otherwise, update priority
843 * to the normal priority:
845 if (!rt_prio(p->prio))
846 return p->normal_prio;
851 * task_curr - is this task currently executing on a CPU?
852 * @p: the task in question.
854 * Return: 1 if the task is currently executing. 0 otherwise.
856 inline int task_curr(const struct task_struct *p)
858 return cpu_curr(task_cpu(p)) == p;
862 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
863 * use the balance_callback list if you want balancing.
865 * this means any call to check_class_changed() must be followed by a call to
866 * balance_callback().
868 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
869 const struct sched_class *prev_class,
872 if (prev_class != p->sched_class) {
873 if (prev_class->switched_from)
874 prev_class->switched_from(rq, p);
876 p->sched_class->switched_to(rq, p);
877 } else if (oldprio != p->prio || dl_task(p))
878 p->sched_class->prio_changed(rq, p, oldprio);
881 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
883 const struct sched_class *class;
885 if (p->sched_class == rq->curr->sched_class) {
886 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
888 for_each_class(class) {
889 if (class == rq->curr->sched_class)
891 if (class == p->sched_class) {
899 * A queue event has occurred, and we're going to schedule. In
900 * this case, we can save a useless back to back clock update.
902 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
903 rq_clock_skip_update(rq, true);
908 * This is how migration works:
910 * 1) we invoke migration_cpu_stop() on the target CPU using
912 * 2) stopper starts to run (implicitly forcing the migrated thread
914 * 3) it checks whether the migrated task is still in the wrong runqueue.
915 * 4) if it's in the wrong runqueue then the migration thread removes
916 * it and puts it into the right queue.
917 * 5) stopper completes and stop_one_cpu() returns and the migration
922 * move_queued_task - move a queued task to new rq.
924 * Returns (locked) new rq. Old rq's lock is released.
926 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
927 struct task_struct *p, int new_cpu)
929 lockdep_assert_held(&rq->lock);
931 p->on_rq = TASK_ON_RQ_MIGRATING;
932 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
933 set_task_cpu(p, new_cpu);
936 rq = cpu_rq(new_cpu);
939 BUG_ON(task_cpu(p) != new_cpu);
940 enqueue_task(rq, p, 0);
941 p->on_rq = TASK_ON_RQ_QUEUED;
942 check_preempt_curr(rq, p, 0);
947 struct migration_arg {
948 struct task_struct *task;
953 * Move (not current) task off this CPU, onto the destination CPU. We're doing
954 * this because either it can't run here any more (set_cpus_allowed()
955 * away from this CPU, or CPU going down), or because we're
956 * attempting to rebalance this task on exec (sched_exec).
958 * So we race with normal scheduler movements, but that's OK, as long
959 * as the task is no longer on this CPU.
961 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
962 struct task_struct *p, int dest_cpu)
964 if (p->flags & PF_KTHREAD) {
965 if (unlikely(!cpu_online(dest_cpu)))
968 if (unlikely(!cpu_active(dest_cpu)))
972 /* Affinity changed (again). */
973 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
977 rq = move_queued_task(rq, rf, p, dest_cpu);
983 * migration_cpu_stop - this will be executed by a highprio stopper thread
984 * and performs thread migration by bumping thread off CPU then
985 * 'pushing' onto another runqueue.
987 static int migration_cpu_stop(void *data)
989 struct migration_arg *arg = data;
990 struct task_struct *p = arg->task;
991 struct rq *rq = this_rq();
995 * The original target CPU might have gone down and we might
996 * be on another CPU but it doesn't matter.
1000 * We need to explicitly wake pending tasks before running
1001 * __migrate_task() such that we will not miss enforcing cpus_allowed
1002 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1004 sched_ttwu_pending();
1006 raw_spin_lock(&p->pi_lock);
1009 * If task_rq(p) != rq, it cannot be migrated here, because we're
1010 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1011 * we're holding p->pi_lock.
1013 if (task_rq(p) == rq) {
1014 if (task_on_rq_queued(p))
1015 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1017 p->wake_cpu = arg->dest_cpu;
1020 raw_spin_unlock(&p->pi_lock);
1027 * sched_class::set_cpus_allowed must do the below, but is not required to
1028 * actually call this function.
1030 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1032 cpumask_copy(&p->cpus_allowed, new_mask);
1033 p->nr_cpus_allowed = cpumask_weight(new_mask);
1036 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1038 struct rq *rq = task_rq(p);
1039 bool queued, running;
1041 lockdep_assert_held(&p->pi_lock);
1043 queued = task_on_rq_queued(p);
1044 running = task_current(rq, p);
1048 * Because __kthread_bind() calls this on blocked tasks without
1051 lockdep_assert_held(&rq->lock);
1052 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1055 put_prev_task(rq, p);
1057 p->sched_class->set_cpus_allowed(p, new_mask);
1060 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1062 set_curr_task(rq, p);
1066 * Change a given task's CPU affinity. Migrate the thread to a
1067 * proper CPU and schedule it away if the CPU it's executing on
1068 * is removed from the allowed bitmask.
1070 * NOTE: the caller must have a valid reference to the task, the
1071 * task must not exit() & deallocate itself prematurely. The
1072 * call is not atomic; no spinlocks may be held.
1074 static int __set_cpus_allowed_ptr(struct task_struct *p,
1075 const struct cpumask *new_mask, bool check)
1077 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1078 unsigned int dest_cpu;
1083 rq = task_rq_lock(p, &rf);
1084 update_rq_clock(rq);
1086 if (p->flags & PF_KTHREAD) {
1088 * Kernel threads are allowed on online && !active CPUs
1090 cpu_valid_mask = cpu_online_mask;
1094 * Must re-check here, to close a race against __kthread_bind(),
1095 * sched_setaffinity() is not guaranteed to observe the flag.
1097 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1102 if (cpumask_equal(&p->cpus_allowed, new_mask))
1105 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1110 do_set_cpus_allowed(p, new_mask);
1112 if (p->flags & PF_KTHREAD) {
1114 * For kernel threads that do indeed end up on online &&
1115 * !active we want to ensure they are strict per-CPU threads.
1117 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1118 !cpumask_intersects(new_mask, cpu_active_mask) &&
1119 p->nr_cpus_allowed != 1);
1122 /* Can the task run on the task's current CPU? If so, we're done */
1123 if (cpumask_test_cpu(task_cpu(p), new_mask))
1126 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1127 if (task_running(rq, p) || p->state == TASK_WAKING) {
1128 struct migration_arg arg = { p, dest_cpu };
1129 /* Need help from migration thread: drop lock and wait. */
1130 task_rq_unlock(rq, p, &rf);
1131 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1132 tlb_migrate_finish(p->mm);
1134 } else if (task_on_rq_queued(p)) {
1136 * OK, since we're going to drop the lock immediately
1137 * afterwards anyway.
1139 rq = move_queued_task(rq, &rf, p, dest_cpu);
1142 task_rq_unlock(rq, p, &rf);
1147 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1149 return __set_cpus_allowed_ptr(p, new_mask, false);
1151 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1153 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1155 #ifdef CONFIG_SCHED_DEBUG
1157 * We should never call set_task_cpu() on a blocked task,
1158 * ttwu() will sort out the placement.
1160 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1164 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1165 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1166 * time relying on p->on_rq.
1168 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1169 p->sched_class == &fair_sched_class &&
1170 (p->on_rq && !task_on_rq_migrating(p)));
1172 #ifdef CONFIG_LOCKDEP
1174 * The caller should hold either p->pi_lock or rq->lock, when changing
1175 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1177 * sched_move_task() holds both and thus holding either pins the cgroup,
1180 * Furthermore, all task_rq users should acquire both locks, see
1183 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1184 lockdep_is_held(&task_rq(p)->lock)));
1187 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1189 WARN_ON_ONCE(!cpu_online(new_cpu));
1192 trace_sched_migrate_task(p, new_cpu);
1194 if (task_cpu(p) != new_cpu) {
1195 if (p->sched_class->migrate_task_rq)
1196 p->sched_class->migrate_task_rq(p);
1197 p->se.nr_migrations++;
1198 perf_event_task_migrate(p);
1201 __set_task_cpu(p, new_cpu);
1204 static void __migrate_swap_task(struct task_struct *p, int cpu)
1206 if (task_on_rq_queued(p)) {
1207 struct rq *src_rq, *dst_rq;
1208 struct rq_flags srf, drf;
1210 src_rq = task_rq(p);
1211 dst_rq = cpu_rq(cpu);
1213 rq_pin_lock(src_rq, &srf);
1214 rq_pin_lock(dst_rq, &drf);
1216 p->on_rq = TASK_ON_RQ_MIGRATING;
1217 deactivate_task(src_rq, p, 0);
1218 set_task_cpu(p, cpu);
1219 activate_task(dst_rq, p, 0);
1220 p->on_rq = TASK_ON_RQ_QUEUED;
1221 check_preempt_curr(dst_rq, p, 0);
1223 rq_unpin_lock(dst_rq, &drf);
1224 rq_unpin_lock(src_rq, &srf);
1228 * Task isn't running anymore; make it appear like we migrated
1229 * it before it went to sleep. This means on wakeup we make the
1230 * previous CPU our target instead of where it really is.
1236 struct migration_swap_arg {
1237 struct task_struct *src_task, *dst_task;
1238 int src_cpu, dst_cpu;
1241 static int migrate_swap_stop(void *data)
1243 struct migration_swap_arg *arg = data;
1244 struct rq *src_rq, *dst_rq;
1247 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1250 src_rq = cpu_rq(arg->src_cpu);
1251 dst_rq = cpu_rq(arg->dst_cpu);
1253 double_raw_lock(&arg->src_task->pi_lock,
1254 &arg->dst_task->pi_lock);
1255 double_rq_lock(src_rq, dst_rq);
1257 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1260 if (task_cpu(arg->src_task) != arg->src_cpu)
1263 if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1266 if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1269 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1270 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1275 double_rq_unlock(src_rq, dst_rq);
1276 raw_spin_unlock(&arg->dst_task->pi_lock);
1277 raw_spin_unlock(&arg->src_task->pi_lock);
1283 * Cross migrate two tasks
1285 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1287 struct migration_swap_arg arg;
1290 arg = (struct migration_swap_arg){
1292 .src_cpu = task_cpu(cur),
1294 .dst_cpu = task_cpu(p),
1297 if (arg.src_cpu == arg.dst_cpu)
1301 * These three tests are all lockless; this is OK since all of them
1302 * will be re-checked with proper locks held further down the line.
1304 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1307 if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1310 if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1313 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1314 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1321 * wait_task_inactive - wait for a thread to unschedule.
1323 * If @match_state is nonzero, it's the @p->state value just checked and
1324 * not expected to change. If it changes, i.e. @p might have woken up,
1325 * then return zero. When we succeed in waiting for @p to be off its CPU,
1326 * we return a positive number (its total switch count). If a second call
1327 * a short while later returns the same number, the caller can be sure that
1328 * @p has remained unscheduled the whole time.
1330 * The caller must ensure that the task *will* unschedule sometime soon,
1331 * else this function might spin for a *long* time. This function can't
1332 * be called with interrupts off, or it may introduce deadlock with
1333 * smp_call_function() if an IPI is sent by the same process we are
1334 * waiting to become inactive.
1336 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1338 int running, queued;
1345 * We do the initial early heuristics without holding
1346 * any task-queue locks at all. We'll only try to get
1347 * the runqueue lock when things look like they will
1353 * If the task is actively running on another CPU
1354 * still, just relax and busy-wait without holding
1357 * NOTE! Since we don't hold any locks, it's not
1358 * even sure that "rq" stays as the right runqueue!
1359 * But we don't care, since "task_running()" will
1360 * return false if the runqueue has changed and p
1361 * is actually now running somewhere else!
1363 while (task_running(rq, p)) {
1364 if (match_state && unlikely(p->state != match_state))
1370 * Ok, time to look more closely! We need the rq
1371 * lock now, to be *sure*. If we're wrong, we'll
1372 * just go back and repeat.
1374 rq = task_rq_lock(p, &rf);
1375 trace_sched_wait_task(p);
1376 running = task_running(rq, p);
1377 queued = task_on_rq_queued(p);
1379 if (!match_state || p->state == match_state)
1380 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1381 task_rq_unlock(rq, p, &rf);
1384 * If it changed from the expected state, bail out now.
1386 if (unlikely(!ncsw))
1390 * Was it really running after all now that we
1391 * checked with the proper locks actually held?
1393 * Oops. Go back and try again..
1395 if (unlikely(running)) {
1401 * It's not enough that it's not actively running,
1402 * it must be off the runqueue _entirely_, and not
1405 * So if it was still runnable (but just not actively
1406 * running right now), it's preempted, and we should
1407 * yield - it could be a while.
1409 if (unlikely(queued)) {
1410 ktime_t to = NSEC_PER_SEC / HZ;
1412 set_current_state(TASK_UNINTERRUPTIBLE);
1413 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1418 * Ahh, all good. It wasn't running, and it wasn't
1419 * runnable, which means that it will never become
1420 * running in the future either. We're all done!
1429 * kick_process - kick a running thread to enter/exit the kernel
1430 * @p: the to-be-kicked thread
1432 * Cause a process which is running on another CPU to enter
1433 * kernel-mode, without any delay. (to get signals handled.)
1435 * NOTE: this function doesn't have to take the runqueue lock,
1436 * because all it wants to ensure is that the remote task enters
1437 * the kernel. If the IPI races and the task has been migrated
1438 * to another CPU then no harm is done and the purpose has been
1441 void kick_process(struct task_struct *p)
1447 if ((cpu != smp_processor_id()) && task_curr(p))
1448 smp_send_reschedule(cpu);
1451 EXPORT_SYMBOL_GPL(kick_process);
1454 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1456 * A few notes on cpu_active vs cpu_online:
1458 * - cpu_active must be a subset of cpu_online
1460 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1461 * see __set_cpus_allowed_ptr(). At this point the newly online
1462 * CPU isn't yet part of the sched domains, and balancing will not
1465 * - on CPU-down we clear cpu_active() to mask the sched domains and
1466 * avoid the load balancer to place new tasks on the to be removed
1467 * CPU. Existing tasks will remain running there and will be taken
1470 * This means that fallback selection must not select !active CPUs.
1471 * And can assume that any active CPU must be online. Conversely
1472 * select_task_rq() below may allow selection of !active CPUs in order
1473 * to satisfy the above rules.
1475 static int select_fallback_rq(int cpu, struct task_struct *p)
1477 int nid = cpu_to_node(cpu);
1478 const struct cpumask *nodemask = NULL;
1479 enum { cpuset, possible, fail } state = cpuset;
1483 * If the node that the CPU is on has been offlined, cpu_to_node()
1484 * will return -1. There is no CPU on the node, and we should
1485 * select the CPU on the other node.
1488 nodemask = cpumask_of_node(nid);
1490 /* Look for allowed, online CPU in same node. */
1491 for_each_cpu(dest_cpu, nodemask) {
1492 if (!cpu_active(dest_cpu))
1494 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1500 /* Any allowed, online CPU? */
1501 for_each_cpu(dest_cpu, &p->cpus_allowed) {
1502 if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1504 if (!cpu_online(dest_cpu))
1509 /* No more Mr. Nice Guy. */
1512 if (IS_ENABLED(CONFIG_CPUSETS)) {
1513 cpuset_cpus_allowed_fallback(p);
1519 do_set_cpus_allowed(p, cpu_possible_mask);
1530 if (state != cpuset) {
1532 * Don't tell them about moving exiting tasks or
1533 * kernel threads (both mm NULL), since they never
1536 if (p->mm && printk_ratelimit()) {
1537 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1538 task_pid_nr(p), p->comm, cpu);
1546 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1549 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1551 lockdep_assert_held(&p->pi_lock);
1553 if (p->nr_cpus_allowed > 1)
1554 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1556 cpu = cpumask_any(&p->cpus_allowed);
1559 * In order not to call set_task_cpu() on a blocking task we need
1560 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1563 * Since this is common to all placement strategies, this lives here.
1565 * [ this allows ->select_task() to simply return task_cpu(p) and
1566 * not worry about this generic constraint ]
1568 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
1570 cpu = select_fallback_rq(task_cpu(p), p);
1575 static void update_avg(u64 *avg, u64 sample)
1577 s64 diff = sample - *avg;
1581 void sched_set_stop_task(int cpu, struct task_struct *stop)
1583 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1584 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1588 * Make it appear like a SCHED_FIFO task, its something
1589 * userspace knows about and won't get confused about.
1591 * Also, it will make PI more or less work without too
1592 * much confusion -- but then, stop work should not
1593 * rely on PI working anyway.
1595 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1597 stop->sched_class = &stop_sched_class;
1600 cpu_rq(cpu)->stop = stop;
1604 * Reset it back to a normal scheduling class so that
1605 * it can die in pieces.
1607 old_stop->sched_class = &rt_sched_class;
1613 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1614 const struct cpumask *new_mask, bool check)
1616 return set_cpus_allowed_ptr(p, new_mask);
1619 #endif /* CONFIG_SMP */
1622 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1626 if (!schedstat_enabled())
1632 if (cpu == rq->cpu) {
1633 schedstat_inc(rq->ttwu_local);
1634 schedstat_inc(p->se.statistics.nr_wakeups_local);
1636 struct sched_domain *sd;
1638 schedstat_inc(p->se.statistics.nr_wakeups_remote);
1640 for_each_domain(rq->cpu, sd) {
1641 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1642 schedstat_inc(sd->ttwu_wake_remote);
1649 if (wake_flags & WF_MIGRATED)
1650 schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1651 #endif /* CONFIG_SMP */
1653 schedstat_inc(rq->ttwu_count);
1654 schedstat_inc(p->se.statistics.nr_wakeups);
1656 if (wake_flags & WF_SYNC)
1657 schedstat_inc(p->se.statistics.nr_wakeups_sync);
1660 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1662 activate_task(rq, p, en_flags);
1663 p->on_rq = TASK_ON_RQ_QUEUED;
1665 /* If a worker is waking up, notify the workqueue: */
1666 if (p->flags & PF_WQ_WORKER)
1667 wq_worker_waking_up(p, cpu_of(rq));
1671 * Mark the task runnable and perform wakeup-preemption.
1673 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1674 struct rq_flags *rf)
1676 check_preempt_curr(rq, p, wake_flags);
1677 p->state = TASK_RUNNING;
1678 trace_sched_wakeup(p);
1681 if (p->sched_class->task_woken) {
1683 * Our task @p is fully woken up and running; so its safe to
1684 * drop the rq->lock, hereafter rq is only used for statistics.
1686 rq_unpin_lock(rq, rf);
1687 p->sched_class->task_woken(rq, p);
1688 rq_repin_lock(rq, rf);
1691 if (rq->idle_stamp) {
1692 u64 delta = rq_clock(rq) - rq->idle_stamp;
1693 u64 max = 2*rq->max_idle_balance_cost;
1695 update_avg(&rq->avg_idle, delta);
1697 if (rq->avg_idle > max)
1706 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1707 struct rq_flags *rf)
1709 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
1711 lockdep_assert_held(&rq->lock);
1714 if (p->sched_contributes_to_load)
1715 rq->nr_uninterruptible--;
1717 if (wake_flags & WF_MIGRATED)
1718 en_flags |= ENQUEUE_MIGRATED;
1721 ttwu_activate(rq, p, en_flags);
1722 ttwu_do_wakeup(rq, p, wake_flags, rf);
1726 * Called in case the task @p isn't fully descheduled from its runqueue,
1727 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1728 * since all we need to do is flip p->state to TASK_RUNNING, since
1729 * the task is still ->on_rq.
1731 static int ttwu_remote(struct task_struct *p, int wake_flags)
1737 rq = __task_rq_lock(p, &rf);
1738 if (task_on_rq_queued(p)) {
1739 /* check_preempt_curr() may use rq clock */
1740 update_rq_clock(rq);
1741 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1744 __task_rq_unlock(rq, &rf);
1750 void sched_ttwu_pending(void)
1752 struct rq *rq = this_rq();
1753 struct llist_node *llist = llist_del_all(&rq->wake_list);
1754 struct task_struct *p, *t;
1760 rq_lock_irqsave(rq, &rf);
1761 update_rq_clock(rq);
1763 llist_for_each_entry_safe(p, t, llist, wake_entry)
1764 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
1766 rq_unlock_irqrestore(rq, &rf);
1769 void scheduler_ipi(void)
1772 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1773 * TIF_NEED_RESCHED remotely (for the first time) will also send
1776 preempt_fold_need_resched();
1778 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1782 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1783 * traditionally all their work was done from the interrupt return
1784 * path. Now that we actually do some work, we need to make sure
1787 * Some archs already do call them, luckily irq_enter/exit nest
1790 * Arguably we should visit all archs and update all handlers,
1791 * however a fair share of IPIs are still resched only so this would
1792 * somewhat pessimize the simple resched case.
1795 sched_ttwu_pending();
1798 * Check if someone kicked us for doing the nohz idle load balance.
1800 if (unlikely(got_nohz_idle_kick())) {
1801 this_rq()->idle_balance = 1;
1802 raise_softirq_irqoff(SCHED_SOFTIRQ);
1807 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1809 struct rq *rq = cpu_rq(cpu);
1811 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1813 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1814 if (!set_nr_if_polling(rq->idle))
1815 smp_send_reschedule(cpu);
1817 trace_sched_wake_idle_without_ipi(cpu);
1821 void wake_up_if_idle(int cpu)
1823 struct rq *rq = cpu_rq(cpu);
1828 if (!is_idle_task(rcu_dereference(rq->curr)))
1831 if (set_nr_if_polling(rq->idle)) {
1832 trace_sched_wake_idle_without_ipi(cpu);
1834 rq_lock_irqsave(rq, &rf);
1835 if (is_idle_task(rq->curr))
1836 smp_send_reschedule(cpu);
1837 /* Else CPU is not idle, do nothing here: */
1838 rq_unlock_irqrestore(rq, &rf);
1845 bool cpus_share_cache(int this_cpu, int that_cpu)
1847 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1849 #endif /* CONFIG_SMP */
1851 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1853 struct rq *rq = cpu_rq(cpu);
1856 #if defined(CONFIG_SMP)
1857 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1858 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1859 ttwu_queue_remote(p, cpu, wake_flags);
1865 update_rq_clock(rq);
1866 ttwu_do_activate(rq, p, wake_flags, &rf);
1871 * Notes on Program-Order guarantees on SMP systems.
1875 * The basic program-order guarantee on SMP systems is that when a task [t]
1876 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1877 * execution on its new CPU [c1].
1879 * For migration (of runnable tasks) this is provided by the following means:
1881 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1882 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1883 * rq(c1)->lock (if not at the same time, then in that order).
1884 * C) LOCK of the rq(c1)->lock scheduling in task
1886 * Transitivity guarantees that B happens after A and C after B.
1887 * Note: we only require RCpc transitivity.
1888 * Note: the CPU doing B need not be c0 or c1
1897 * UNLOCK rq(0)->lock
1899 * LOCK rq(0)->lock // orders against CPU0
1901 * UNLOCK rq(0)->lock
1905 * UNLOCK rq(1)->lock
1907 * LOCK rq(1)->lock // orders against CPU2
1910 * UNLOCK rq(1)->lock
1913 * BLOCKING -- aka. SLEEP + WAKEUP
1915 * For blocking we (obviously) need to provide the same guarantee as for
1916 * migration. However the means are completely different as there is no lock
1917 * chain to provide order. Instead we do:
1919 * 1) smp_store_release(X->on_cpu, 0)
1920 * 2) smp_cond_load_acquire(!X->on_cpu)
1924 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1926 * LOCK rq(0)->lock LOCK X->pi_lock
1929 * smp_store_release(X->on_cpu, 0);
1931 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1937 * X->state = RUNNING
1938 * UNLOCK rq(2)->lock
1940 * LOCK rq(2)->lock // orders against CPU1
1943 * UNLOCK rq(2)->lock
1946 * UNLOCK rq(0)->lock
1949 * However; for wakeups there is a second guarantee we must provide, namely we
1950 * must observe the state that lead to our wakeup. That is, not only must our
1951 * task observe its own prior state, it must also observe the stores prior to
1954 * This means that any means of doing remote wakeups must order the CPU doing
1955 * the wakeup against the CPU the task is going to end up running on. This,
1956 * however, is already required for the regular Program-Order guarantee above,
1957 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1962 * try_to_wake_up - wake up a thread
1963 * @p: the thread to be awakened
1964 * @state: the mask of task states that can be woken
1965 * @wake_flags: wake modifier flags (WF_*)
1967 * If (@state & @p->state) @p->state = TASK_RUNNING.
1969 * If the task was not queued/runnable, also place it back on a runqueue.
1971 * Atomic against schedule() which would dequeue a task, also see
1972 * set_current_state().
1974 * Return: %true if @p->state changes (an actual wakeup was done),
1978 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1980 unsigned long flags;
1981 int cpu, success = 0;
1984 * If we are going to wake up a thread waiting for CONDITION we
1985 * need to ensure that CONDITION=1 done by the caller can not be
1986 * reordered with p->state check below. This pairs with mb() in
1987 * set_current_state() the waiting thread does.
1989 raw_spin_lock_irqsave(&p->pi_lock, flags);
1990 smp_mb__after_spinlock();
1991 if (!(p->state & state))
1994 trace_sched_waking(p);
1996 /* We're going to change ->state: */
2001 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2002 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2003 * in smp_cond_load_acquire() below.
2005 * sched_ttwu_pending() try_to_wake_up()
2006 * [S] p->on_rq = 1; [L] P->state
2007 * UNLOCK rq->lock -----.
2011 * LOCK rq->lock -----'
2015 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2017 * Pairs with the UNLOCK+LOCK on rq->lock from the
2018 * last wakeup of our task and the schedule that got our task
2022 if (p->on_rq && ttwu_remote(p, wake_flags))
2027 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2028 * possible to, falsely, observe p->on_cpu == 0.
2030 * One must be running (->on_cpu == 1) in order to remove oneself
2031 * from the runqueue.
2033 * [S] ->on_cpu = 1; [L] ->on_rq
2037 * [S] ->on_rq = 0; [L] ->on_cpu
2039 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2040 * from the consecutive calls to schedule(); the first switching to our
2041 * task, the second putting it to sleep.
2046 * If the owning (remote) CPU is still in the middle of schedule() with
2047 * this task as prev, wait until its done referencing the task.
2049 * Pairs with the smp_store_release() in finish_task().
2051 * This ensures that tasks getting woken will be fully ordered against
2052 * their previous state and preserve Program Order.
2054 smp_cond_load_acquire(&p->on_cpu, !VAL);
2056 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2057 p->state = TASK_WAKING;
2060 delayacct_blkio_end(p);
2061 atomic_dec(&task_rq(p)->nr_iowait);
2064 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2065 if (task_cpu(p) != cpu) {
2066 wake_flags |= WF_MIGRATED;
2067 set_task_cpu(p, cpu);
2070 #else /* CONFIG_SMP */
2073 delayacct_blkio_end(p);
2074 atomic_dec(&task_rq(p)->nr_iowait);
2077 #endif /* CONFIG_SMP */
2079 ttwu_queue(p, cpu, wake_flags);
2081 ttwu_stat(p, cpu, wake_flags);
2083 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2089 * try_to_wake_up_local - try to wake up a local task with rq lock held
2090 * @p: the thread to be awakened
2091 * @rf: request-queue flags for pinning
2093 * Put @p on the run-queue if it's not already there. The caller must
2094 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2097 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2099 struct rq *rq = task_rq(p);
2101 if (WARN_ON_ONCE(rq != this_rq()) ||
2102 WARN_ON_ONCE(p == current))
2105 lockdep_assert_held(&rq->lock);
2107 if (!raw_spin_trylock(&p->pi_lock)) {
2109 * This is OK, because current is on_cpu, which avoids it being
2110 * picked for load-balance and preemption/IRQs are still
2111 * disabled avoiding further scheduler activity on it and we've
2112 * not yet picked a replacement task.
2115 raw_spin_lock(&p->pi_lock);
2119 if (!(p->state & TASK_NORMAL))
2122 trace_sched_waking(p);
2124 if (!task_on_rq_queued(p)) {
2126 delayacct_blkio_end(p);
2127 atomic_dec(&rq->nr_iowait);
2129 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK);
2132 ttwu_do_wakeup(rq, p, 0, rf);
2133 ttwu_stat(p, smp_processor_id(), 0);
2135 raw_spin_unlock(&p->pi_lock);
2139 * wake_up_process - Wake up a specific process
2140 * @p: The process to be woken up.
2142 * Attempt to wake up the nominated process and move it to the set of runnable
2145 * Return: 1 if the process was woken up, 0 if it was already running.
2147 * It may be assumed that this function implies a write memory barrier before
2148 * changing the task state if and only if any tasks are woken up.
2150 int wake_up_process(struct task_struct *p)
2152 return try_to_wake_up(p, TASK_NORMAL, 0);
2154 EXPORT_SYMBOL(wake_up_process);
2156 int wake_up_state(struct task_struct *p, unsigned int state)
2158 return try_to_wake_up(p, state, 0);
2162 * Perform scheduler related setup for a newly forked process p.
2163 * p is forked by current.
2165 * __sched_fork() is basic setup used by init_idle() too:
2167 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2172 p->se.exec_start = 0;
2173 p->se.sum_exec_runtime = 0;
2174 p->se.prev_sum_exec_runtime = 0;
2175 p->se.nr_migrations = 0;
2177 INIT_LIST_HEAD(&p->se.group_node);
2179 #ifdef CONFIG_FAIR_GROUP_SCHED
2180 p->se.cfs_rq = NULL;
2183 #ifdef CONFIG_SCHEDSTATS
2184 /* Even if schedstat is disabled, there should not be garbage */
2185 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2188 RB_CLEAR_NODE(&p->dl.rb_node);
2189 init_dl_task_timer(&p->dl);
2190 init_dl_inactive_task_timer(&p->dl);
2191 __dl_clear_params(p);
2193 INIT_LIST_HEAD(&p->rt.run_list);
2195 p->rt.time_slice = sched_rr_timeslice;
2199 #ifdef CONFIG_PREEMPT_NOTIFIERS
2200 INIT_HLIST_HEAD(&p->preempt_notifiers);
2203 #ifdef CONFIG_NUMA_BALANCING
2204 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2205 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2206 p->mm->numa_scan_seq = 0;
2209 if (clone_flags & CLONE_VM)
2210 p->numa_preferred_nid = current->numa_preferred_nid;
2212 p->numa_preferred_nid = -1;
2214 p->node_stamp = 0ULL;
2215 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2216 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2217 p->numa_work.next = &p->numa_work;
2218 p->numa_faults = NULL;
2219 p->last_task_numa_placement = 0;
2220 p->last_sum_exec_runtime = 0;
2222 p->numa_group = NULL;
2223 #endif /* CONFIG_NUMA_BALANCING */
2226 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2228 #ifdef CONFIG_NUMA_BALANCING
2230 void set_numabalancing_state(bool enabled)
2233 static_branch_enable(&sched_numa_balancing);
2235 static_branch_disable(&sched_numa_balancing);
2238 #ifdef CONFIG_PROC_SYSCTL
2239 int sysctl_numa_balancing(struct ctl_table *table, int write,
2240 void __user *buffer, size_t *lenp, loff_t *ppos)
2244 int state = static_branch_likely(&sched_numa_balancing);
2246 if (write && !capable(CAP_SYS_ADMIN))
2251 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2255 set_numabalancing_state(state);
2261 #ifdef CONFIG_SCHEDSTATS
2263 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2264 static bool __initdata __sched_schedstats = false;
2266 static void set_schedstats(bool enabled)
2269 static_branch_enable(&sched_schedstats);
2271 static_branch_disable(&sched_schedstats);
2274 void force_schedstat_enabled(void)
2276 if (!schedstat_enabled()) {
2277 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2278 static_branch_enable(&sched_schedstats);
2282 static int __init setup_schedstats(char *str)
2289 * This code is called before jump labels have been set up, so we can't
2290 * change the static branch directly just yet. Instead set a temporary
2291 * variable so init_schedstats() can do it later.
2293 if (!strcmp(str, "enable")) {
2294 __sched_schedstats = true;
2296 } else if (!strcmp(str, "disable")) {
2297 __sched_schedstats = false;
2302 pr_warn("Unable to parse schedstats=\n");
2306 __setup("schedstats=", setup_schedstats);
2308 static void __init init_schedstats(void)
2310 set_schedstats(__sched_schedstats);
2313 #ifdef CONFIG_PROC_SYSCTL
2314 int sysctl_schedstats(struct ctl_table *table, int write,
2315 void __user *buffer, size_t *lenp, loff_t *ppos)
2319 int state = static_branch_likely(&sched_schedstats);
2321 if (write && !capable(CAP_SYS_ADMIN))
2326 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2330 set_schedstats(state);
2333 #endif /* CONFIG_PROC_SYSCTL */
2334 #else /* !CONFIG_SCHEDSTATS */
2335 static inline void init_schedstats(void) {}
2336 #endif /* CONFIG_SCHEDSTATS */
2339 * fork()/clone()-time setup:
2341 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2343 unsigned long flags;
2344 int cpu = get_cpu();
2346 __sched_fork(clone_flags, p);
2348 * We mark the process as NEW here. This guarantees that
2349 * nobody will actually run it, and a signal or other external
2350 * event cannot wake it up and insert it on the runqueue either.
2352 p->state = TASK_NEW;
2355 * Make sure we do not leak PI boosting priority to the child.
2357 p->prio = current->normal_prio;
2360 * Revert to default priority/policy on fork if requested.
2362 if (unlikely(p->sched_reset_on_fork)) {
2363 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2364 p->policy = SCHED_NORMAL;
2365 p->static_prio = NICE_TO_PRIO(0);
2367 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2368 p->static_prio = NICE_TO_PRIO(0);
2370 p->prio = p->normal_prio = __normal_prio(p);
2371 set_load_weight(p, false);
2374 * We don't need the reset flag anymore after the fork. It has
2375 * fulfilled its duty:
2377 p->sched_reset_on_fork = 0;
2380 if (dl_prio(p->prio)) {
2383 } else if (rt_prio(p->prio)) {
2384 p->sched_class = &rt_sched_class;
2386 p->sched_class = &fair_sched_class;
2389 init_entity_runnable_average(&p->se);
2392 * The child is not yet in the pid-hash so no cgroup attach races,
2393 * and the cgroup is pinned to this child due to cgroup_fork()
2394 * is ran before sched_fork().
2396 * Silence PROVE_RCU.
2398 raw_spin_lock_irqsave(&p->pi_lock, flags);
2400 * We're setting the CPU for the first time, we don't migrate,
2401 * so use __set_task_cpu().
2403 __set_task_cpu(p, cpu);
2404 if (p->sched_class->task_fork)
2405 p->sched_class->task_fork(p);
2406 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2408 #ifdef CONFIG_SCHED_INFO
2409 if (likely(sched_info_on()))
2410 memset(&p->sched_info, 0, sizeof(p->sched_info));
2412 #if defined(CONFIG_SMP)
2415 init_task_preempt_count(p);
2417 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2418 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2425 unsigned long to_ratio(u64 period, u64 runtime)
2427 if (runtime == RUNTIME_INF)
2431 * Doing this here saves a lot of checks in all
2432 * the calling paths, and returning zero seems
2433 * safe for them anyway.
2438 return div64_u64(runtime << BW_SHIFT, period);
2442 * wake_up_new_task - wake up a newly created task for the first time.
2444 * This function will do some initial scheduler statistics housekeeping
2445 * that must be done for every newly created context, then puts the task
2446 * on the runqueue and wakes it.
2448 void wake_up_new_task(struct task_struct *p)
2453 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2454 p->state = TASK_RUNNING;
2457 * Fork balancing, do it here and not earlier because:
2458 * - cpus_allowed can change in the fork path
2459 * - any previously selected CPU might disappear through hotplug
2461 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2462 * as we're not fully set-up yet.
2464 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2466 rq = __task_rq_lock(p, &rf);
2467 update_rq_clock(rq);
2468 post_init_entity_util_avg(&p->se);
2470 activate_task(rq, p, ENQUEUE_NOCLOCK);
2471 p->on_rq = TASK_ON_RQ_QUEUED;
2472 trace_sched_wakeup_new(p);
2473 check_preempt_curr(rq, p, WF_FORK);
2475 if (p->sched_class->task_woken) {
2477 * Nothing relies on rq->lock after this, so its fine to
2480 rq_unpin_lock(rq, &rf);
2481 p->sched_class->task_woken(rq, p);
2482 rq_repin_lock(rq, &rf);
2485 task_rq_unlock(rq, p, &rf);
2488 #ifdef CONFIG_PREEMPT_NOTIFIERS
2490 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2492 void preempt_notifier_inc(void)
2494 static_key_slow_inc(&preempt_notifier_key);
2496 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2498 void preempt_notifier_dec(void)
2500 static_key_slow_dec(&preempt_notifier_key);
2502 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2505 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2506 * @notifier: notifier struct to register
2508 void preempt_notifier_register(struct preempt_notifier *notifier)
2510 if (!static_key_false(&preempt_notifier_key))
2511 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2513 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2515 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2518 * preempt_notifier_unregister - no longer interested in preemption notifications
2519 * @notifier: notifier struct to unregister
2521 * This is *not* safe to call from within a preemption notifier.
2523 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2525 hlist_del(¬ifier->link);
2527 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2529 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2531 struct preempt_notifier *notifier;
2533 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2534 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2537 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2539 if (static_key_false(&preempt_notifier_key))
2540 __fire_sched_in_preempt_notifiers(curr);
2544 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2545 struct task_struct *next)
2547 struct preempt_notifier *notifier;
2549 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2550 notifier->ops->sched_out(notifier, next);
2553 static __always_inline void
2554 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2555 struct task_struct *next)
2557 if (static_key_false(&preempt_notifier_key))
2558 __fire_sched_out_preempt_notifiers(curr, next);
2561 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2563 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2568 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2569 struct task_struct *next)
2573 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2575 static inline void prepare_task(struct task_struct *next)
2579 * Claim the task as running, we do this before switching to it
2580 * such that any running task will have this set.
2586 static inline void finish_task(struct task_struct *prev)
2590 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2591 * We must ensure this doesn't happen until the switch is completely
2594 * In particular, the load of prev->state in finish_task_switch() must
2595 * happen before this.
2597 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2599 smp_store_release(&prev->on_cpu, 0);
2603 static inline void finish_lock_switch(struct rq *rq)
2605 #ifdef CONFIG_DEBUG_SPINLOCK
2606 /* this is a valid case when another task releases the spinlock */
2607 rq->lock.owner = current;
2610 * If we are tracking spinlock dependencies then we have to
2611 * fix up the runqueue lock - which gets 'carried over' from
2612 * prev into current:
2614 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
2616 raw_spin_unlock_irq(&rq->lock);
2620 * prepare_task_switch - prepare to switch tasks
2621 * @rq: the runqueue preparing to switch
2622 * @prev: the current task that is being switched out
2623 * @next: the task we are going to switch to.
2625 * This is called with the rq lock held and interrupts off. It must
2626 * be paired with a subsequent finish_task_switch after the context
2629 * prepare_task_switch sets up locking and calls architecture specific
2633 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2634 struct task_struct *next)
2636 sched_info_switch(rq, prev, next);
2637 perf_event_task_sched_out(prev, next);
2638 fire_sched_out_preempt_notifiers(prev, next);
2640 prepare_arch_switch(next);
2644 * finish_task_switch - clean up after a task-switch
2645 * @prev: the thread we just switched away from.
2647 * finish_task_switch must be called after the context switch, paired
2648 * with a prepare_task_switch call before the context switch.
2649 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2650 * and do any other architecture-specific cleanup actions.
2652 * Note that we may have delayed dropping an mm in context_switch(). If
2653 * so, we finish that here outside of the runqueue lock. (Doing it
2654 * with the lock held can cause deadlocks; see schedule() for
2657 * The context switch have flipped the stack from under us and restored the
2658 * local variables which were saved when this task called schedule() in the
2659 * past. prev == current is still correct but we need to recalculate this_rq
2660 * because prev may have moved to another CPU.
2662 static struct rq *finish_task_switch(struct task_struct *prev)
2663 __releases(rq->lock)
2665 struct rq *rq = this_rq();
2666 struct mm_struct *mm = rq->prev_mm;
2670 * The previous task will have left us with a preempt_count of 2
2671 * because it left us after:
2674 * preempt_disable(); // 1
2676 * raw_spin_lock_irq(&rq->lock) // 2
2678 * Also, see FORK_PREEMPT_COUNT.
2680 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2681 "corrupted preempt_count: %s/%d/0x%x\n",
2682 current->comm, current->pid, preempt_count()))
2683 preempt_count_set(FORK_PREEMPT_COUNT);
2688 * A task struct has one reference for the use as "current".
2689 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2690 * schedule one last time. The schedule call will never return, and
2691 * the scheduled task must drop that reference.
2693 * We must observe prev->state before clearing prev->on_cpu (in
2694 * finish_task), otherwise a concurrent wakeup can get prev
2695 * running on another CPU and we could rave with its RUNNING -> DEAD
2696 * transition, resulting in a double drop.
2698 prev_state = prev->state;
2699 vtime_task_switch(prev);
2700 perf_event_task_sched_in(prev, current);
2702 finish_lock_switch(rq);
2703 finish_arch_post_lock_switch();
2705 fire_sched_in_preempt_notifiers(current);
2707 * When transitioning from a kernel thread to a userspace
2708 * thread, mmdrop()'s implicit full barrier is required by the
2709 * membarrier system call, because the current ->active_mm can
2710 * become the current mm without going through switch_mm().
2714 if (unlikely(prev_state == TASK_DEAD)) {
2715 if (prev->sched_class->task_dead)
2716 prev->sched_class->task_dead(prev);
2719 * Remove function-return probe instances associated with this
2720 * task and put them back on the free list.
2722 kprobe_flush_task(prev);
2724 /* Task is done with its stack. */
2725 put_task_stack(prev);
2727 put_task_struct(prev);
2730 tick_nohz_task_switch();
2736 /* rq->lock is NOT held, but preemption is disabled */
2737 static void __balance_callback(struct rq *rq)
2739 struct callback_head *head, *next;
2740 void (*func)(struct rq *rq);
2741 unsigned long flags;
2743 raw_spin_lock_irqsave(&rq->lock, flags);
2744 head = rq->balance_callback;
2745 rq->balance_callback = NULL;
2747 func = (void (*)(struct rq *))head->func;
2754 raw_spin_unlock_irqrestore(&rq->lock, flags);
2757 static inline void balance_callback(struct rq *rq)
2759 if (unlikely(rq->balance_callback))
2760 __balance_callback(rq);
2765 static inline void balance_callback(struct rq *rq)
2772 * schedule_tail - first thing a freshly forked thread must call.
2773 * @prev: the thread we just switched away from.
2775 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2776 __releases(rq->lock)
2781 * New tasks start with FORK_PREEMPT_COUNT, see there and
2782 * finish_task_switch() for details.
2784 * finish_task_switch() will drop rq->lock() and lower preempt_count
2785 * and the preempt_enable() will end up enabling preemption (on
2786 * PREEMPT_COUNT kernels).
2789 rq = finish_task_switch(prev);
2790 balance_callback(rq);
2793 if (current->set_child_tid)
2794 put_user(task_pid_vnr(current), current->set_child_tid);
2798 * context_switch - switch to the new MM and the new thread's register state.
2800 static __always_inline struct rq *
2801 context_switch(struct rq *rq, struct task_struct *prev,
2802 struct task_struct *next, struct rq_flags *rf)
2804 struct mm_struct *mm, *oldmm;
2806 prepare_task_switch(rq, prev, next);
2809 oldmm = prev->active_mm;
2811 * For paravirt, this is coupled with an exit in switch_to to
2812 * combine the page table reload and the switch backend into
2815 arch_start_context_switch(prev);
2818 * If mm is non-NULL, we pass through switch_mm(). If mm is
2819 * NULL, we will pass through mmdrop() in finish_task_switch().
2820 * Both of these contain the full memory barrier required by
2821 * membarrier after storing to rq->curr, before returning to
2825 next->active_mm = oldmm;
2827 enter_lazy_tlb(oldmm, next);
2829 switch_mm_irqs_off(oldmm, mm, next);
2832 prev->active_mm = NULL;
2833 rq->prev_mm = oldmm;
2836 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2839 * Since the runqueue lock will be released by the next
2840 * task (which is an invalid locking op but in the case
2841 * of the scheduler it's an obvious special-case), so we
2842 * do an early lockdep release here:
2844 rq_unpin_lock(rq, rf);
2845 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2847 /* Here we just switch the register state and the stack. */
2848 switch_to(prev, next, prev);
2851 return finish_task_switch(prev);
2855 * nr_running and nr_context_switches:
2857 * externally visible scheduler statistics: current number of runnable
2858 * threads, total number of context switches performed since bootup.
2860 unsigned long nr_running(void)
2862 unsigned long i, sum = 0;
2864 for_each_online_cpu(i)
2865 sum += cpu_rq(i)->nr_running;
2871 * Check if only the current task is running on the CPU.
2873 * Caution: this function does not check that the caller has disabled
2874 * preemption, thus the result might have a time-of-check-to-time-of-use
2875 * race. The caller is responsible to use it correctly, for example:
2877 * - from a non-preemptable section (of course)
2879 * - from a thread that is bound to a single CPU
2881 * - in a loop with very short iterations (e.g. a polling loop)
2883 bool single_task_running(void)
2885 return raw_rq()->nr_running == 1;
2887 EXPORT_SYMBOL(single_task_running);
2889 unsigned long long nr_context_switches(void)
2892 unsigned long long sum = 0;
2894 for_each_possible_cpu(i)
2895 sum += cpu_rq(i)->nr_switches;
2901 * IO-wait accounting, and how its mostly bollocks (on SMP).
2903 * The idea behind IO-wait account is to account the idle time that we could
2904 * have spend running if it were not for IO. That is, if we were to improve the
2905 * storage performance, we'd have a proportional reduction in IO-wait time.
2907 * This all works nicely on UP, where, when a task blocks on IO, we account
2908 * idle time as IO-wait, because if the storage were faster, it could've been
2909 * running and we'd not be idle.
2911 * This has been extended to SMP, by doing the same for each CPU. This however
2914 * Imagine for instance the case where two tasks block on one CPU, only the one
2915 * CPU will have IO-wait accounted, while the other has regular idle. Even
2916 * though, if the storage were faster, both could've ran at the same time,
2917 * utilising both CPUs.
2919 * This means, that when looking globally, the current IO-wait accounting on
2920 * SMP is a lower bound, by reason of under accounting.
2922 * Worse, since the numbers are provided per CPU, they are sometimes
2923 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2924 * associated with any one particular CPU, it can wake to another CPU than it
2925 * blocked on. This means the per CPU IO-wait number is meaningless.
2927 * Task CPU affinities can make all that even more 'interesting'.
2930 unsigned long nr_iowait(void)
2932 unsigned long i, sum = 0;
2934 for_each_possible_cpu(i)
2935 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2941 * Consumers of these two interfaces, like for example the cpufreq menu
2942 * governor are using nonsensical data. Boosting frequency for a CPU that has
2943 * IO-wait which might not even end up running the task when it does become
2947 unsigned long nr_iowait_cpu(int cpu)
2949 struct rq *this = cpu_rq(cpu);
2950 return atomic_read(&this->nr_iowait);
2953 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2955 struct rq *rq = this_rq();
2956 *nr_waiters = atomic_read(&rq->nr_iowait);
2957 *load = rq->load.weight;
2963 * sched_exec - execve() is a valuable balancing opportunity, because at
2964 * this point the task has the smallest effective memory and cache footprint.
2966 void sched_exec(void)
2968 struct task_struct *p = current;
2969 unsigned long flags;
2972 raw_spin_lock_irqsave(&p->pi_lock, flags);
2973 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2974 if (dest_cpu == smp_processor_id())
2977 if (likely(cpu_active(dest_cpu))) {
2978 struct migration_arg arg = { p, dest_cpu };
2980 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2981 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2985 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2990 DEFINE_PER_CPU(struct kernel_stat, kstat);
2991 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2993 EXPORT_PER_CPU_SYMBOL(kstat);
2994 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2997 * The function fair_sched_class.update_curr accesses the struct curr
2998 * and its field curr->exec_start; when called from task_sched_runtime(),
2999 * we observe a high rate of cache misses in practice.
3000 * Prefetching this data results in improved performance.
3002 static inline void prefetch_curr_exec_start(struct task_struct *p)
3004 #ifdef CONFIG_FAIR_GROUP_SCHED
3005 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3007 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3010 prefetch(&curr->exec_start);
3014 * Return accounted runtime for the task.
3015 * In case the task is currently running, return the runtime plus current's
3016 * pending runtime that have not been accounted yet.
3018 unsigned long long task_sched_runtime(struct task_struct *p)
3024 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3026 * 64-bit doesn't need locks to atomically read a 64bit value.
3027 * So we have a optimization chance when the task's delta_exec is 0.
3028 * Reading ->on_cpu is racy, but this is ok.
3030 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3031 * If we race with it entering CPU, unaccounted time is 0. This is
3032 * indistinguishable from the read occurring a few cycles earlier.
3033 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3034 * been accounted, so we're correct here as well.
3036 if (!p->on_cpu || !task_on_rq_queued(p))
3037 return p->se.sum_exec_runtime;
3040 rq = task_rq_lock(p, &rf);
3042 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3043 * project cycles that may never be accounted to this
3044 * thread, breaking clock_gettime().
3046 if (task_current(rq, p) && task_on_rq_queued(p)) {
3047 prefetch_curr_exec_start(p);
3048 update_rq_clock(rq);
3049 p->sched_class->update_curr(rq);
3051 ns = p->se.sum_exec_runtime;
3052 task_rq_unlock(rq, p, &rf);
3058 * This function gets called by the timer code, with HZ frequency.
3059 * We call it with interrupts disabled.
3061 void scheduler_tick(void)
3063 int cpu = smp_processor_id();
3064 struct rq *rq = cpu_rq(cpu);
3065 struct task_struct *curr = rq->curr;
3072 update_rq_clock(rq);
3073 curr->sched_class->task_tick(rq, curr, 0);
3074 cpu_load_update_active(rq);
3075 calc_global_load_tick(rq);
3079 perf_event_task_tick();
3082 rq->idle_balance = idle_cpu(cpu);
3083 trigger_load_balance(rq);
3085 rq_last_tick_reset(rq);
3088 #ifdef CONFIG_NO_HZ_FULL
3090 * scheduler_tick_max_deferment
3092 * Keep at least one tick per second when a single
3093 * active task is running because the scheduler doesn't
3094 * yet completely support full dynticks environment.
3096 * This makes sure that uptime, CFS vruntime, load
3097 * balancing, etc... continue to move forward, even
3098 * with a very low granularity.
3100 * Return: Maximum deferment in nanoseconds.
3102 u64 scheduler_tick_max_deferment(void)
3104 struct rq *rq = this_rq();
3105 unsigned long next, now = READ_ONCE(jiffies);
3107 next = rq->last_sched_tick + HZ;
3109 if (time_before_eq(next, now))
3112 return jiffies_to_nsecs(next - now);
3116 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3117 defined(CONFIG_PREEMPT_TRACER))
3119 * If the value passed in is equal to the current preempt count
3120 * then we just disabled preemption. Start timing the latency.
3122 static inline void preempt_latency_start(int val)
3124 if (preempt_count() == val) {
3125 unsigned long ip = get_lock_parent_ip();
3126 #ifdef CONFIG_DEBUG_PREEMPT
3127 current->preempt_disable_ip = ip;
3129 trace_preempt_off(CALLER_ADDR0, ip);
3133 void preempt_count_add(int val)
3135 #ifdef CONFIG_DEBUG_PREEMPT
3139 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3142 __preempt_count_add(val);
3143 #ifdef CONFIG_DEBUG_PREEMPT
3145 * Spinlock count overflowing soon?
3147 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3150 preempt_latency_start(val);
3152 EXPORT_SYMBOL(preempt_count_add);
3153 NOKPROBE_SYMBOL(preempt_count_add);
3156 * If the value passed in equals to the current preempt count
3157 * then we just enabled preemption. Stop timing the latency.
3159 static inline void preempt_latency_stop(int val)
3161 if (preempt_count() == val)
3162 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3165 void preempt_count_sub(int val)
3167 #ifdef CONFIG_DEBUG_PREEMPT
3171 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3174 * Is the spinlock portion underflowing?
3176 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3177 !(preempt_count() & PREEMPT_MASK)))
3181 preempt_latency_stop(val);
3182 __preempt_count_sub(val);
3184 EXPORT_SYMBOL(preempt_count_sub);
3185 NOKPROBE_SYMBOL(preempt_count_sub);
3188 static inline void preempt_latency_start(int val) { }
3189 static inline void preempt_latency_stop(int val) { }
3192 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3194 #ifdef CONFIG_DEBUG_PREEMPT
3195 return p->preempt_disable_ip;
3202 * Print scheduling while atomic bug:
3204 static noinline void __schedule_bug(struct task_struct *prev)
3206 /* Save this before calling printk(), since that will clobber it */
3207 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3209 if (oops_in_progress)
3212 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3213 prev->comm, prev->pid, preempt_count());
3215 debug_show_held_locks(prev);
3217 if (irqs_disabled())
3218 print_irqtrace_events(prev);
3219 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3220 && in_atomic_preempt_off()) {
3221 pr_err("Preemption disabled at:");
3222 print_ip_sym(preempt_disable_ip);
3226 panic("scheduling while atomic\n");
3229 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3233 * Various schedule()-time debugging checks and statistics:
3235 static inline void schedule_debug(struct task_struct *prev)
3237 #ifdef CONFIG_SCHED_STACK_END_CHECK
3238 if (task_stack_end_corrupted(prev))
3239 panic("corrupted stack end detected inside scheduler\n");
3242 if (unlikely(in_atomic_preempt_off())) {
3243 __schedule_bug(prev);
3244 preempt_count_set(PREEMPT_DISABLED);
3248 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3250 schedstat_inc(this_rq()->sched_count);
3254 * Pick up the highest-prio task:
3256 static inline struct task_struct *
3257 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3259 const struct sched_class *class;
3260 struct task_struct *p;
3263 * Optimization: we know that if all tasks are in the fair class we can
3264 * call that function directly, but only if the @prev task wasn't of a
3265 * higher scheduling class, because otherwise those loose the
3266 * opportunity to pull in more work from other CPUs.
3268 if (likely((prev->sched_class == &idle_sched_class ||
3269 prev->sched_class == &fair_sched_class) &&
3270 rq->nr_running == rq->cfs.h_nr_running)) {
3272 p = fair_sched_class.pick_next_task(rq, prev, rf);
3273 if (unlikely(p == RETRY_TASK))
3276 /* Assumes fair_sched_class->next == idle_sched_class */
3278 p = idle_sched_class.pick_next_task(rq, prev, rf);
3284 for_each_class(class) {
3285 p = class->pick_next_task(rq, prev, rf);
3287 if (unlikely(p == RETRY_TASK))
3293 /* The idle class should always have a runnable task: */
3298 * __schedule() is the main scheduler function.
3300 * The main means of driving the scheduler and thus entering this function are:
3302 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3304 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3305 * paths. For example, see arch/x86/entry_64.S.
3307 * To drive preemption between tasks, the scheduler sets the flag in timer
3308 * interrupt handler scheduler_tick().
3310 * 3. Wakeups don't really cause entry into schedule(). They add a
3311 * task to the run-queue and that's it.
3313 * Now, if the new task added to the run-queue preempts the current
3314 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3315 * called on the nearest possible occasion:
3317 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3319 * - in syscall or exception context, at the next outmost
3320 * preempt_enable(). (this might be as soon as the wake_up()'s
3323 * - in IRQ context, return from interrupt-handler to
3324 * preemptible context
3326 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3329 * - cond_resched() call
3330 * - explicit schedule() call
3331 * - return from syscall or exception to user-space
3332 * - return from interrupt-handler to user-space
3334 * WARNING: must be called with preemption disabled!
3336 static void __sched notrace __schedule(bool preempt)
3338 struct task_struct *prev, *next;
3339 unsigned long *switch_count;
3344 cpu = smp_processor_id();
3348 schedule_debug(prev);
3350 if (sched_feat(HRTICK))
3353 local_irq_disable();
3354 rcu_note_context_switch(preempt);
3357 * Make sure that signal_pending_state()->signal_pending() below
3358 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3359 * done by the caller to avoid the race with signal_wake_up().
3361 * The membarrier system call requires a full memory barrier
3362 * after coming from user-space, before storing to rq->curr.
3365 smp_mb__after_spinlock();
3367 /* Promote REQ to ACT */
3368 rq->clock_update_flags <<= 1;
3369 update_rq_clock(rq);
3371 switch_count = &prev->nivcsw;
3372 if (!preempt && prev->state) {
3373 if (unlikely(signal_pending_state(prev->state, prev))) {
3374 prev->state = TASK_RUNNING;
3376 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3379 if (prev->in_iowait) {
3380 atomic_inc(&rq->nr_iowait);
3381 delayacct_blkio_start();
3385 * If a worker went to sleep, notify and ask workqueue
3386 * whether it wants to wake up a task to maintain
3389 if (prev->flags & PF_WQ_WORKER) {
3390 struct task_struct *to_wakeup;
3392 to_wakeup = wq_worker_sleeping(prev);
3394 try_to_wake_up_local(to_wakeup, &rf);
3397 switch_count = &prev->nvcsw;
3400 next = pick_next_task(rq, prev, &rf);
3401 clear_tsk_need_resched(prev);
3402 clear_preempt_need_resched();
3404 if (likely(prev != next)) {
3408 * The membarrier system call requires each architecture
3409 * to have a full memory barrier after updating
3410 * rq->curr, before returning to user-space.
3412 * Here are the schemes providing that barrier on the
3413 * various architectures:
3414 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3415 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3416 * - finish_lock_switch() for weakly-ordered
3417 * architectures where spin_unlock is a full barrier,
3418 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3419 * is a RELEASE barrier),
3423 trace_sched_switch(preempt, prev, next);
3425 /* Also unlocks the rq: */
3426 rq = context_switch(rq, prev, next, &rf);
3428 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3429 rq_unlock_irq(rq, &rf);
3432 balance_callback(rq);
3435 void __noreturn do_task_dead(void)
3438 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3439 * when the following two conditions become true.
3440 * - There is race condition of mmap_sem (It is acquired by
3442 * - SMI occurs before setting TASK_RUNINNG.
3443 * (or hypervisor of virtual machine switches to other guest)
3444 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3446 * To avoid it, we have to wait for releasing tsk->pi_lock which
3447 * is held by try_to_wake_up()
3449 raw_spin_lock_irq(¤t->pi_lock);
3450 raw_spin_unlock_irq(¤t->pi_lock);
3452 /* Causes final put_task_struct in finish_task_switch(): */
3453 __set_current_state(TASK_DEAD);
3455 /* Tell freezer to ignore us: */
3456 current->flags |= PF_NOFREEZE;
3461 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3466 static inline void sched_submit_work(struct task_struct *tsk)
3468 if (!tsk->state || tsk_is_pi_blocked(tsk))
3471 * If we are going to sleep and we have plugged IO queued,
3472 * make sure to submit it to avoid deadlocks.
3474 if (blk_needs_flush_plug(tsk))
3475 blk_schedule_flush_plug(tsk);
3478 asmlinkage __visible void __sched schedule(void)
3480 struct task_struct *tsk = current;
3482 sched_submit_work(tsk);
3486 sched_preempt_enable_no_resched();
3487 } while (need_resched());
3489 EXPORT_SYMBOL(schedule);
3492 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3493 * state (have scheduled out non-voluntarily) by making sure that all
3494 * tasks have either left the run queue or have gone into user space.
3495 * As idle tasks do not do either, they must not ever be preempted
3496 * (schedule out non-voluntarily).
3498 * schedule_idle() is similar to schedule_preempt_disable() except that it
3499 * never enables preemption because it does not call sched_submit_work().
3501 void __sched schedule_idle(void)
3504 * As this skips calling sched_submit_work(), which the idle task does
3505 * regardless because that function is a nop when the task is in a
3506 * TASK_RUNNING state, make sure this isn't used someplace that the
3507 * current task can be in any other state. Note, idle is always in the
3508 * TASK_RUNNING state.
3510 WARN_ON_ONCE(current->state);
3513 } while (need_resched());
3516 #ifdef CONFIG_CONTEXT_TRACKING
3517 asmlinkage __visible void __sched schedule_user(void)
3520 * If we come here after a random call to set_need_resched(),
3521 * or we have been woken up remotely but the IPI has not yet arrived,
3522 * we haven't yet exited the RCU idle mode. Do it here manually until
3523 * we find a better solution.
3525 * NB: There are buggy callers of this function. Ideally we
3526 * should warn if prev_state != CONTEXT_USER, but that will trigger
3527 * too frequently to make sense yet.
3529 enum ctx_state prev_state = exception_enter();
3531 exception_exit(prev_state);
3536 * schedule_preempt_disabled - called with preemption disabled
3538 * Returns with preemption disabled. Note: preempt_count must be 1
3540 void __sched schedule_preempt_disabled(void)
3542 sched_preempt_enable_no_resched();
3547 static void __sched notrace preempt_schedule_common(void)
3551 * Because the function tracer can trace preempt_count_sub()
3552 * and it also uses preempt_enable/disable_notrace(), if
3553 * NEED_RESCHED is set, the preempt_enable_notrace() called
3554 * by the function tracer will call this function again and
3555 * cause infinite recursion.
3557 * Preemption must be disabled here before the function
3558 * tracer can trace. Break up preempt_disable() into two
3559 * calls. One to disable preemption without fear of being
3560 * traced. The other to still record the preemption latency,
3561 * which can also be traced by the function tracer.
3563 preempt_disable_notrace();
3564 preempt_latency_start(1);
3566 preempt_latency_stop(1);
3567 preempt_enable_no_resched_notrace();
3570 * Check again in case we missed a preemption opportunity
3571 * between schedule and now.
3573 } while (need_resched());
3576 #ifdef CONFIG_PREEMPT
3578 * this is the entry point to schedule() from in-kernel preemption
3579 * off of preempt_enable. Kernel preemptions off return from interrupt
3580 * occur there and call schedule directly.
3582 asmlinkage __visible void __sched notrace preempt_schedule(void)
3585 * If there is a non-zero preempt_count or interrupts are disabled,
3586 * we do not want to preempt the current task. Just return..
3588 if (likely(!preemptible()))
3591 preempt_schedule_common();
3593 NOKPROBE_SYMBOL(preempt_schedule);
3594 EXPORT_SYMBOL(preempt_schedule);
3597 * preempt_schedule_notrace - preempt_schedule called by tracing
3599 * The tracing infrastructure uses preempt_enable_notrace to prevent
3600 * recursion and tracing preempt enabling caused by the tracing
3601 * infrastructure itself. But as tracing can happen in areas coming
3602 * from userspace or just about to enter userspace, a preempt enable
3603 * can occur before user_exit() is called. This will cause the scheduler
3604 * to be called when the system is still in usermode.
3606 * To prevent this, the preempt_enable_notrace will use this function
3607 * instead of preempt_schedule() to exit user context if needed before
3608 * calling the scheduler.
3610 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3612 enum ctx_state prev_ctx;
3614 if (likely(!preemptible()))
3619 * Because the function tracer can trace preempt_count_sub()
3620 * and it also uses preempt_enable/disable_notrace(), if
3621 * NEED_RESCHED is set, the preempt_enable_notrace() called
3622 * by the function tracer will call this function again and
3623 * cause infinite recursion.
3625 * Preemption must be disabled here before the function
3626 * tracer can trace. Break up preempt_disable() into two
3627 * calls. One to disable preemption without fear of being
3628 * traced. The other to still record the preemption latency,
3629 * which can also be traced by the function tracer.
3631 preempt_disable_notrace();
3632 preempt_latency_start(1);
3634 * Needs preempt disabled in case user_exit() is traced
3635 * and the tracer calls preempt_enable_notrace() causing
3636 * an infinite recursion.
3638 prev_ctx = exception_enter();
3640 exception_exit(prev_ctx);
3642 preempt_latency_stop(1);
3643 preempt_enable_no_resched_notrace();
3644 } while (need_resched());
3646 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3648 #endif /* CONFIG_PREEMPT */
3651 * this is the entry point to schedule() from kernel preemption
3652 * off of irq context.
3653 * Note, that this is called and return with irqs disabled. This will
3654 * protect us against recursive calling from irq.
3656 asmlinkage __visible void __sched preempt_schedule_irq(void)
3658 enum ctx_state prev_state;
3660 /* Catch callers which need to be fixed */
3661 BUG_ON(preempt_count() || !irqs_disabled());
3663 prev_state = exception_enter();
3669 local_irq_disable();
3670 sched_preempt_enable_no_resched();
3671 } while (need_resched());
3673 exception_exit(prev_state);
3676 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
3679 return try_to_wake_up(curr->private, mode, wake_flags);
3681 EXPORT_SYMBOL(default_wake_function);
3683 #ifdef CONFIG_RT_MUTEXES
3685 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
3688 prio = min(prio, pi_task->prio);
3693 static inline int rt_effective_prio(struct task_struct *p, int prio)
3695 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3697 return __rt_effective_prio(pi_task, prio);
3701 * rt_mutex_setprio - set the current priority of a task
3703 * @pi_task: donor task
3705 * This function changes the 'effective' priority of a task. It does
3706 * not touch ->normal_prio like __setscheduler().
3708 * Used by the rt_mutex code to implement priority inheritance
3709 * logic. Call site only calls if the priority of the task changed.
3711 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
3713 int prio, oldprio, queued, running, queue_flag =
3714 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
3715 const struct sched_class *prev_class;
3719 /* XXX used to be waiter->prio, not waiter->task->prio */
3720 prio = __rt_effective_prio(pi_task, p->normal_prio);
3723 * If nothing changed; bail early.
3725 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
3728 rq = __task_rq_lock(p, &rf);
3729 update_rq_clock(rq);
3731 * Set under pi_lock && rq->lock, such that the value can be used under
3734 * Note that there is loads of tricky to make this pointer cache work
3735 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3736 * ensure a task is de-boosted (pi_task is set to NULL) before the
3737 * task is allowed to run again (and can exit). This ensures the pointer
3738 * points to a blocked task -- which guaratees the task is present.
3740 p->pi_top_task = pi_task;
3743 * For FIFO/RR we only need to set prio, if that matches we're done.
3745 if (prio == p->prio && !dl_prio(prio))
3749 * Idle task boosting is a nono in general. There is one
3750 * exception, when PREEMPT_RT and NOHZ is active:
3752 * The idle task calls get_next_timer_interrupt() and holds
3753 * the timer wheel base->lock on the CPU and another CPU wants
3754 * to access the timer (probably to cancel it). We can safely
3755 * ignore the boosting request, as the idle CPU runs this code
3756 * with interrupts disabled and will complete the lock
3757 * protected section without being interrupted. So there is no
3758 * real need to boost.
3760 if (unlikely(p == rq->idle)) {
3761 WARN_ON(p != rq->curr);
3762 WARN_ON(p->pi_blocked_on);
3766 trace_sched_pi_setprio(p, pi_task);
3769 if (oldprio == prio)
3770 queue_flag &= ~DEQUEUE_MOVE;
3772 prev_class = p->sched_class;
3773 queued = task_on_rq_queued(p);
3774 running = task_current(rq, p);
3776 dequeue_task(rq, p, queue_flag);
3778 put_prev_task(rq, p);
3781 * Boosting condition are:
3782 * 1. -rt task is running and holds mutex A
3783 * --> -dl task blocks on mutex A
3785 * 2. -dl task is running and holds mutex A
3786 * --> -dl task blocks on mutex A and could preempt the
3789 if (dl_prio(prio)) {
3790 if (!dl_prio(p->normal_prio) ||
3791 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3792 p->dl.dl_boosted = 1;
3793 queue_flag |= ENQUEUE_REPLENISH;
3795 p->dl.dl_boosted = 0;
3796 p->sched_class = &dl_sched_class;
3797 } else if (rt_prio(prio)) {
3798 if (dl_prio(oldprio))
3799 p->dl.dl_boosted = 0;
3801 queue_flag |= ENQUEUE_HEAD;
3802 p->sched_class = &rt_sched_class;
3804 if (dl_prio(oldprio))
3805 p->dl.dl_boosted = 0;
3806 if (rt_prio(oldprio))
3808 p->sched_class = &fair_sched_class;
3814 enqueue_task(rq, p, queue_flag);
3816 set_curr_task(rq, p);
3818 check_class_changed(rq, p, prev_class, oldprio);
3820 /* Avoid rq from going away on us: */
3822 __task_rq_unlock(rq, &rf);
3824 balance_callback(rq);
3828 static inline int rt_effective_prio(struct task_struct *p, int prio)
3834 void set_user_nice(struct task_struct *p, long nice)
3836 bool queued, running;
3837 int old_prio, delta;
3841 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3844 * We have to be careful, if called from sys_setpriority(),
3845 * the task might be in the middle of scheduling on another CPU.
3847 rq = task_rq_lock(p, &rf);
3848 update_rq_clock(rq);
3851 * The RT priorities are set via sched_setscheduler(), but we still
3852 * allow the 'normal' nice value to be set - but as expected
3853 * it wont have any effect on scheduling until the task is
3854 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3856 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3857 p->static_prio = NICE_TO_PRIO(nice);
3860 queued = task_on_rq_queued(p);
3861 running = task_current(rq, p);
3863 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
3865 put_prev_task(rq, p);
3867 p->static_prio = NICE_TO_PRIO(nice);
3868 set_load_weight(p, true);
3870 p->prio = effective_prio(p);
3871 delta = p->prio - old_prio;
3874 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
3876 * If the task increased its priority or is running and
3877 * lowered its priority, then reschedule its CPU:
3879 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3883 set_curr_task(rq, p);
3885 task_rq_unlock(rq, p, &rf);
3887 EXPORT_SYMBOL(set_user_nice);
3890 * can_nice - check if a task can reduce its nice value
3894 int can_nice(const struct task_struct *p, const int nice)
3896 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3897 int nice_rlim = nice_to_rlimit(nice);
3899 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3900 capable(CAP_SYS_NICE));
3903 #ifdef __ARCH_WANT_SYS_NICE
3906 * sys_nice - change the priority of the current process.
3907 * @increment: priority increment
3909 * sys_setpriority is a more generic, but much slower function that
3910 * does similar things.
3912 SYSCALL_DEFINE1(nice, int, increment)
3917 * Setpriority might change our priority at the same moment.
3918 * We don't have to worry. Conceptually one call occurs first
3919 * and we have a single winner.
3921 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3922 nice = task_nice(current) + increment;
3924 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3925 if (increment < 0 && !can_nice(current, nice))
3928 retval = security_task_setnice(current, nice);
3932 set_user_nice(current, nice);
3939 * task_prio - return the priority value of a given task.
3940 * @p: the task in question.
3942 * Return: The priority value as seen by users in /proc.
3943 * RT tasks are offset by -200. Normal tasks are centered
3944 * around 0, value goes from -16 to +15.
3946 int task_prio(const struct task_struct *p)
3948 return p->prio - MAX_RT_PRIO;
3952 * idle_cpu - is a given CPU idle currently?
3953 * @cpu: the processor in question.
3955 * Return: 1 if the CPU is currently idle. 0 otherwise.
3957 int idle_cpu(int cpu)
3959 struct rq *rq = cpu_rq(cpu);
3961 if (rq->curr != rq->idle)
3968 if (!llist_empty(&rq->wake_list))
3976 * idle_task - return the idle task for a given CPU.
3977 * @cpu: the processor in question.
3979 * Return: The idle task for the CPU @cpu.
3981 struct task_struct *idle_task(int cpu)
3983 return cpu_rq(cpu)->idle;
3987 * find_process_by_pid - find a process with a matching PID value.
3988 * @pid: the pid in question.
3990 * The task of @pid, if found. %NULL otherwise.
3992 static struct task_struct *find_process_by_pid(pid_t pid)
3994 return pid ? find_task_by_vpid(pid) : current;
3998 * sched_setparam() passes in -1 for its policy, to let the functions
3999 * it calls know not to change it.
4001 #define SETPARAM_POLICY -1
4003 static void __setscheduler_params(struct task_struct *p,
4004 const struct sched_attr *attr)
4006 int policy = attr->sched_policy;
4008 if (policy == SETPARAM_POLICY)
4013 if (dl_policy(policy))
4014 __setparam_dl(p, attr);
4015 else if (fair_policy(policy))
4016 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4019 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4020 * !rt_policy. Always setting this ensures that things like
4021 * getparam()/getattr() don't report silly values for !rt tasks.
4023 p->rt_priority = attr->sched_priority;
4024 p->normal_prio = normal_prio(p);
4025 set_load_weight(p, true);
4028 /* Actually do priority change: must hold pi & rq lock. */
4029 static void __setscheduler(struct rq *rq, struct task_struct *p,
4030 const struct sched_attr *attr, bool keep_boost)
4032 __setscheduler_params(p, attr);
4035 * Keep a potential priority boosting if called from
4036 * sched_setscheduler().
4038 p->prio = normal_prio(p);
4040 p->prio = rt_effective_prio(p, p->prio);
4042 if (dl_prio(p->prio))
4043 p->sched_class = &dl_sched_class;
4044 else if (rt_prio(p->prio))
4045 p->sched_class = &rt_sched_class;
4047 p->sched_class = &fair_sched_class;
4051 * Check the target process has a UID that matches the current process's:
4053 static bool check_same_owner(struct task_struct *p)
4055 const struct cred *cred = current_cred(), *pcred;
4059 pcred = __task_cred(p);
4060 match = (uid_eq(cred->euid, pcred->euid) ||
4061 uid_eq(cred->euid, pcred->uid));
4066 static int __sched_setscheduler(struct task_struct *p,
4067 const struct sched_attr *attr,
4070 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4071 MAX_RT_PRIO - 1 - attr->sched_priority;
4072 int retval, oldprio, oldpolicy = -1, queued, running;
4073 int new_effective_prio, policy = attr->sched_policy;
4074 const struct sched_class *prev_class;
4077 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4080 /* The pi code expects interrupts enabled */
4081 BUG_ON(pi && in_interrupt());
4083 /* Double check policy once rq lock held: */
4085 reset_on_fork = p->sched_reset_on_fork;
4086 policy = oldpolicy = p->policy;
4088 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4090 if (!valid_policy(policy))
4094 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4098 * Valid priorities for SCHED_FIFO and SCHED_RR are
4099 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4100 * SCHED_BATCH and SCHED_IDLE is 0.
4102 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4103 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4105 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4106 (rt_policy(policy) != (attr->sched_priority != 0)))
4110 * Allow unprivileged RT tasks to decrease priority:
4112 if (user && !capable(CAP_SYS_NICE)) {
4113 if (fair_policy(policy)) {
4114 if (attr->sched_nice < task_nice(p) &&
4115 !can_nice(p, attr->sched_nice))
4119 if (rt_policy(policy)) {
4120 unsigned long rlim_rtprio =
4121 task_rlimit(p, RLIMIT_RTPRIO);
4123 /* Can't set/change the rt policy: */
4124 if (policy != p->policy && !rlim_rtprio)
4127 /* Can't increase priority: */
4128 if (attr->sched_priority > p->rt_priority &&
4129 attr->sched_priority > rlim_rtprio)
4134 * Can't set/change SCHED_DEADLINE policy at all for now
4135 * (safest behavior); in the future we would like to allow
4136 * unprivileged DL tasks to increase their relative deadline
4137 * or reduce their runtime (both ways reducing utilization)
4139 if (dl_policy(policy))
4143 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4144 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4146 if (idle_policy(p->policy) && !idle_policy(policy)) {
4147 if (!can_nice(p, task_nice(p)))
4151 /* Can't change other user's priorities: */
4152 if (!check_same_owner(p))
4155 /* Normal users shall not reset the sched_reset_on_fork flag: */
4156 if (p->sched_reset_on_fork && !reset_on_fork)
4161 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4164 retval = security_task_setscheduler(p);
4170 * Make sure no PI-waiters arrive (or leave) while we are
4171 * changing the priority of the task:
4173 * To be able to change p->policy safely, the appropriate
4174 * runqueue lock must be held.
4176 rq = task_rq_lock(p, &rf);
4177 update_rq_clock(rq);
4180 * Changing the policy of the stop threads its a very bad idea:
4182 if (p == rq->stop) {
4183 task_rq_unlock(rq, p, &rf);
4188 * If not changing anything there's no need to proceed further,
4189 * but store a possible modification of reset_on_fork.
4191 if (unlikely(policy == p->policy)) {
4192 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4194 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4196 if (dl_policy(policy) && dl_param_changed(p, attr))
4199 p->sched_reset_on_fork = reset_on_fork;
4200 task_rq_unlock(rq, p, &rf);
4206 #ifdef CONFIG_RT_GROUP_SCHED
4208 * Do not allow realtime tasks into groups that have no runtime
4211 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4212 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4213 !task_group_is_autogroup(task_group(p))) {
4214 task_rq_unlock(rq, p, &rf);
4219 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4220 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4221 cpumask_t *span = rq->rd->span;
4224 * Don't allow tasks with an affinity mask smaller than
4225 * the entire root_domain to become SCHED_DEADLINE. We
4226 * will also fail if there's no bandwidth available.
4228 if (!cpumask_subset(span, &p->cpus_allowed) ||
4229 rq->rd->dl_bw.bw == 0) {
4230 task_rq_unlock(rq, p, &rf);
4237 /* Re-check policy now with rq lock held: */
4238 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4239 policy = oldpolicy = -1;
4240 task_rq_unlock(rq, p, &rf);
4245 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4246 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4249 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4250 task_rq_unlock(rq, p, &rf);
4254 p->sched_reset_on_fork = reset_on_fork;
4259 * Take priority boosted tasks into account. If the new
4260 * effective priority is unchanged, we just store the new
4261 * normal parameters and do not touch the scheduler class and
4262 * the runqueue. This will be done when the task deboost
4265 new_effective_prio = rt_effective_prio(p, newprio);
4266 if (new_effective_prio == oldprio)
4267 queue_flags &= ~DEQUEUE_MOVE;
4270 queued = task_on_rq_queued(p);
4271 running = task_current(rq, p);
4273 dequeue_task(rq, p, queue_flags);
4275 put_prev_task(rq, p);
4277 prev_class = p->sched_class;
4278 __setscheduler(rq, p, attr, pi);
4282 * We enqueue to tail when the priority of a task is
4283 * increased (user space view).
4285 if (oldprio < p->prio)
4286 queue_flags |= ENQUEUE_HEAD;
4288 enqueue_task(rq, p, queue_flags);
4291 set_curr_task(rq, p);
4293 check_class_changed(rq, p, prev_class, oldprio);
4295 /* Avoid rq from going away on us: */
4297 task_rq_unlock(rq, p, &rf);
4300 rt_mutex_adjust_pi(p);
4302 /* Run balance callbacks after we've adjusted the PI chain: */
4303 balance_callback(rq);
4309 static int _sched_setscheduler(struct task_struct *p, int policy,
4310 const struct sched_param *param, bool check)
4312 struct sched_attr attr = {
4313 .sched_policy = policy,
4314 .sched_priority = param->sched_priority,
4315 .sched_nice = PRIO_TO_NICE(p->static_prio),
4318 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4319 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4320 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4321 policy &= ~SCHED_RESET_ON_FORK;
4322 attr.sched_policy = policy;
4325 return __sched_setscheduler(p, &attr, check, true);
4328 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4329 * @p: the task in question.
4330 * @policy: new policy.
4331 * @param: structure containing the new RT priority.
4333 * Return: 0 on success. An error code otherwise.
4335 * NOTE that the task may be already dead.
4337 int sched_setscheduler(struct task_struct *p, int policy,
4338 const struct sched_param *param)
4340 return _sched_setscheduler(p, policy, param, true);
4342 EXPORT_SYMBOL_GPL(sched_setscheduler);
4344 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4346 return __sched_setscheduler(p, attr, true, true);
4348 EXPORT_SYMBOL_GPL(sched_setattr);
4350 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
4352 return __sched_setscheduler(p, attr, false, true);
4356 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4357 * @p: the task in question.
4358 * @policy: new policy.
4359 * @param: structure containing the new RT priority.
4361 * Just like sched_setscheduler, only don't bother checking if the
4362 * current context has permission. For example, this is needed in
4363 * stop_machine(): we create temporary high priority worker threads,
4364 * but our caller might not have that capability.
4366 * Return: 0 on success. An error code otherwise.
4368 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4369 const struct sched_param *param)
4371 return _sched_setscheduler(p, policy, param, false);
4373 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4376 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4378 struct sched_param lparam;
4379 struct task_struct *p;
4382 if (!param || pid < 0)
4384 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4389 p = find_process_by_pid(pid);
4391 retval = sched_setscheduler(p, policy, &lparam);
4398 * Mimics kernel/events/core.c perf_copy_attr().
4400 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4405 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4408 /* Zero the full structure, so that a short copy will be nice: */
4409 memset(attr, 0, sizeof(*attr));
4411 ret = get_user(size, &uattr->size);
4415 /* Bail out on silly large: */
4416 if (size > PAGE_SIZE)
4419 /* ABI compatibility quirk: */
4421 size = SCHED_ATTR_SIZE_VER0;
4423 if (size < SCHED_ATTR_SIZE_VER0)
4427 * If we're handed a bigger struct than we know of,
4428 * ensure all the unknown bits are 0 - i.e. new
4429 * user-space does not rely on any kernel feature
4430 * extensions we dont know about yet.
4432 if (size > sizeof(*attr)) {
4433 unsigned char __user *addr;
4434 unsigned char __user *end;
4437 addr = (void __user *)uattr + sizeof(*attr);
4438 end = (void __user *)uattr + size;
4440 for (; addr < end; addr++) {
4441 ret = get_user(val, addr);
4447 size = sizeof(*attr);
4450 ret = copy_from_user(attr, uattr, size);
4455 * XXX: Do we want to be lenient like existing syscalls; or do we want
4456 * to be strict and return an error on out-of-bounds values?
4458 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4463 put_user(sizeof(*attr), &uattr->size);
4468 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4469 * @pid: the pid in question.
4470 * @policy: new policy.
4471 * @param: structure containing the new RT priority.
4473 * Return: 0 on success. An error code otherwise.
4475 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4480 return do_sched_setscheduler(pid, policy, param);
4484 * sys_sched_setparam - set/change the RT priority of a thread
4485 * @pid: the pid in question.
4486 * @param: structure containing the new RT priority.
4488 * Return: 0 on success. An error code otherwise.
4490 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4492 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4496 * sys_sched_setattr - same as above, but with extended sched_attr
4497 * @pid: the pid in question.
4498 * @uattr: structure containing the extended parameters.
4499 * @flags: for future extension.
4501 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4502 unsigned int, flags)
4504 struct sched_attr attr;
4505 struct task_struct *p;
4508 if (!uattr || pid < 0 || flags)
4511 retval = sched_copy_attr(uattr, &attr);
4515 if ((int)attr.sched_policy < 0)
4520 p = find_process_by_pid(pid);
4522 retval = sched_setattr(p, &attr);
4529 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4530 * @pid: the pid in question.
4532 * Return: On success, the policy of the thread. Otherwise, a negative error
4535 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4537 struct task_struct *p;
4545 p = find_process_by_pid(pid);
4547 retval = security_task_getscheduler(p);
4550 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4557 * sys_sched_getparam - get the RT priority of a thread
4558 * @pid: the pid in question.
4559 * @param: structure containing the RT priority.
4561 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4564 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4566 struct sched_param lp = { .sched_priority = 0 };
4567 struct task_struct *p;
4570 if (!param || pid < 0)
4574 p = find_process_by_pid(pid);
4579 retval = security_task_getscheduler(p);
4583 if (task_has_rt_policy(p))
4584 lp.sched_priority = p->rt_priority;
4588 * This one might sleep, we cannot do it with a spinlock held ...
4590 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4599 static int sched_read_attr(struct sched_attr __user *uattr,
4600 struct sched_attr *attr,
4605 if (!access_ok(VERIFY_WRITE, uattr, usize))
4609 * If we're handed a smaller struct than we know of,
4610 * ensure all the unknown bits are 0 - i.e. old
4611 * user-space does not get uncomplete information.
4613 if (usize < sizeof(*attr)) {
4614 unsigned char *addr;
4617 addr = (void *)attr + usize;
4618 end = (void *)attr + sizeof(*attr);
4620 for (; addr < end; addr++) {
4628 ret = copy_to_user(uattr, attr, attr->size);
4636 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4637 * @pid: the pid in question.
4638 * @uattr: structure containing the extended parameters.
4639 * @size: sizeof(attr) for fwd/bwd comp.
4640 * @flags: for future extension.
4642 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4643 unsigned int, size, unsigned int, flags)
4645 struct sched_attr attr = {
4646 .size = sizeof(struct sched_attr),
4648 struct task_struct *p;
4651 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4652 size < SCHED_ATTR_SIZE_VER0 || flags)
4656 p = find_process_by_pid(pid);
4661 retval = security_task_getscheduler(p);
4665 attr.sched_policy = p->policy;
4666 if (p->sched_reset_on_fork)
4667 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4668 if (task_has_dl_policy(p))
4669 __getparam_dl(p, &attr);
4670 else if (task_has_rt_policy(p))
4671 attr.sched_priority = p->rt_priority;
4673 attr.sched_nice = task_nice(p);
4677 retval = sched_read_attr(uattr, &attr, size);
4685 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4687 cpumask_var_t cpus_allowed, new_mask;
4688 struct task_struct *p;
4693 p = find_process_by_pid(pid);
4699 /* Prevent p going away */
4703 if (p->flags & PF_NO_SETAFFINITY) {
4707 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4711 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4713 goto out_free_cpus_allowed;
4716 if (!check_same_owner(p)) {
4718 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4720 goto out_free_new_mask;
4725 retval = security_task_setscheduler(p);
4727 goto out_free_new_mask;
4730 cpuset_cpus_allowed(p, cpus_allowed);
4731 cpumask_and(new_mask, in_mask, cpus_allowed);
4734 * Since bandwidth control happens on root_domain basis,
4735 * if admission test is enabled, we only admit -deadline
4736 * tasks allowed to run on all the CPUs in the task's
4740 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4742 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4745 goto out_free_new_mask;
4751 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4754 cpuset_cpus_allowed(p, cpus_allowed);
4755 if (!cpumask_subset(new_mask, cpus_allowed)) {
4757 * We must have raced with a concurrent cpuset
4758 * update. Just reset the cpus_allowed to the
4759 * cpuset's cpus_allowed
4761 cpumask_copy(new_mask, cpus_allowed);
4766 free_cpumask_var(new_mask);
4767 out_free_cpus_allowed:
4768 free_cpumask_var(cpus_allowed);
4774 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4775 struct cpumask *new_mask)
4777 if (len < cpumask_size())
4778 cpumask_clear(new_mask);
4779 else if (len > cpumask_size())
4780 len = cpumask_size();
4782 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4786 * sys_sched_setaffinity - set the CPU affinity of a process
4787 * @pid: pid of the process
4788 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4789 * @user_mask_ptr: user-space pointer to the new CPU mask
4791 * Return: 0 on success. An error code otherwise.
4793 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4794 unsigned long __user *, user_mask_ptr)
4796 cpumask_var_t new_mask;
4799 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4802 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4804 retval = sched_setaffinity(pid, new_mask);
4805 free_cpumask_var(new_mask);
4809 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4811 struct task_struct *p;
4812 unsigned long flags;
4818 p = find_process_by_pid(pid);
4822 retval = security_task_getscheduler(p);
4826 raw_spin_lock_irqsave(&p->pi_lock, flags);
4827 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4828 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4837 * sys_sched_getaffinity - get the CPU affinity of a process
4838 * @pid: pid of the process
4839 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4840 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4842 * Return: size of CPU mask copied to user_mask_ptr on success. An
4843 * error code otherwise.
4845 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4846 unsigned long __user *, user_mask_ptr)
4851 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4853 if (len & (sizeof(unsigned long)-1))
4856 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4859 ret = sched_getaffinity(pid, mask);
4861 size_t retlen = min_t(size_t, len, cpumask_size());
4863 if (copy_to_user(user_mask_ptr, mask, retlen))
4868 free_cpumask_var(mask);
4874 * sys_sched_yield - yield the current processor to other threads.
4876 * This function yields the current CPU to other tasks. If there are no
4877 * other threads running on this CPU then this function will return.
4881 SYSCALL_DEFINE0(sched_yield)
4886 local_irq_disable();
4890 schedstat_inc(rq->yld_count);
4891 current->sched_class->yield_task(rq);
4894 * Since we are going to call schedule() anyway, there's
4895 * no need to preempt or enable interrupts:
4899 sched_preempt_enable_no_resched();
4906 #ifndef CONFIG_PREEMPT
4907 int __sched _cond_resched(void)
4909 if (should_resched(0)) {
4910 preempt_schedule_common();
4916 EXPORT_SYMBOL(_cond_resched);
4920 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4921 * call schedule, and on return reacquire the lock.
4923 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4924 * operations here to prevent schedule() from being called twice (once via
4925 * spin_unlock(), once by hand).
4927 int __cond_resched_lock(spinlock_t *lock)
4929 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4932 lockdep_assert_held(lock);
4934 if (spin_needbreak(lock) || resched) {
4937 preempt_schedule_common();
4945 EXPORT_SYMBOL(__cond_resched_lock);
4947 int __sched __cond_resched_softirq(void)
4949 BUG_ON(!in_softirq());
4951 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4953 preempt_schedule_common();
4959 EXPORT_SYMBOL(__cond_resched_softirq);
4962 * yield - yield the current processor to other threads.
4964 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4966 * The scheduler is at all times free to pick the calling task as the most
4967 * eligible task to run, if removing the yield() call from your code breaks
4968 * it, its already broken.
4970 * Typical broken usage is:
4975 * where one assumes that yield() will let 'the other' process run that will
4976 * make event true. If the current task is a SCHED_FIFO task that will never
4977 * happen. Never use yield() as a progress guarantee!!
4979 * If you want to use yield() to wait for something, use wait_event().
4980 * If you want to use yield() to be 'nice' for others, use cond_resched().
4981 * If you still want to use yield(), do not!
4983 void __sched yield(void)
4985 set_current_state(TASK_RUNNING);
4988 EXPORT_SYMBOL(yield);
4991 * yield_to - yield the current processor to another thread in
4992 * your thread group, or accelerate that thread toward the
4993 * processor it's on.
4995 * @preempt: whether task preemption is allowed or not
4997 * It's the caller's job to ensure that the target task struct
4998 * can't go away on us before we can do any checks.
5001 * true (>0) if we indeed boosted the target task.
5002 * false (0) if we failed to boost the target.
5003 * -ESRCH if there's no task to yield to.
5005 int __sched yield_to(struct task_struct *p, bool preempt)
5007 struct task_struct *curr = current;
5008 struct rq *rq, *p_rq;
5009 unsigned long flags;
5012 local_irq_save(flags);
5018 * If we're the only runnable task on the rq and target rq also
5019 * has only one task, there's absolutely no point in yielding.
5021 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5026 double_rq_lock(rq, p_rq);
5027 if (task_rq(p) != p_rq) {
5028 double_rq_unlock(rq, p_rq);
5032 if (!curr->sched_class->yield_to_task)
5035 if (curr->sched_class != p->sched_class)
5038 if (task_running(p_rq, p) || p->state)
5041 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5043 schedstat_inc(rq->yld_count);
5045 * Make p's CPU reschedule; pick_next_entity takes care of
5048 if (preempt && rq != p_rq)
5053 double_rq_unlock(rq, p_rq);
5055 local_irq_restore(flags);
5062 EXPORT_SYMBOL_GPL(yield_to);
5064 int io_schedule_prepare(void)
5066 int old_iowait = current->in_iowait;
5068 current->in_iowait = 1;
5069 blk_schedule_flush_plug(current);
5074 void io_schedule_finish(int token)
5076 current->in_iowait = token;
5080 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5081 * that process accounting knows that this is a task in IO wait state.
5083 long __sched io_schedule_timeout(long timeout)
5088 token = io_schedule_prepare();
5089 ret = schedule_timeout(timeout);
5090 io_schedule_finish(token);
5094 EXPORT_SYMBOL(io_schedule_timeout);
5096 void io_schedule(void)
5100 token = io_schedule_prepare();
5102 io_schedule_finish(token);
5104 EXPORT_SYMBOL(io_schedule);
5107 * sys_sched_get_priority_max - return maximum RT priority.
5108 * @policy: scheduling class.
5110 * Return: On success, this syscall returns the maximum
5111 * rt_priority that can be used by a given scheduling class.
5112 * On failure, a negative error code is returned.
5114 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5121 ret = MAX_USER_RT_PRIO-1;
5123 case SCHED_DEADLINE:
5134 * sys_sched_get_priority_min - return minimum RT priority.
5135 * @policy: scheduling class.
5137 * Return: On success, this syscall returns the minimum
5138 * rt_priority that can be used by a given scheduling class.
5139 * On failure, a negative error code is returned.
5141 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5150 case SCHED_DEADLINE:
5159 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5161 struct task_struct *p;
5162 unsigned int time_slice;
5172 p = find_process_by_pid(pid);
5176 retval = security_task_getscheduler(p);
5180 rq = task_rq_lock(p, &rf);
5182 if (p->sched_class->get_rr_interval)
5183 time_slice = p->sched_class->get_rr_interval(rq, p);
5184 task_rq_unlock(rq, p, &rf);
5187 jiffies_to_timespec64(time_slice, t);
5196 * sys_sched_rr_get_interval - return the default timeslice of a process.
5197 * @pid: pid of the process.
5198 * @interval: userspace pointer to the timeslice value.
5200 * this syscall writes the default timeslice value of a given process
5201 * into the user-space timespec buffer. A value of '0' means infinity.
5203 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5206 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5207 struct timespec __user *, interval)
5209 struct timespec64 t;
5210 int retval = sched_rr_get_interval(pid, &t);
5213 retval = put_timespec64(&t, interval);
5218 #ifdef CONFIG_COMPAT
5219 COMPAT_SYSCALL_DEFINE2(sched_rr_get_interval,
5221 struct compat_timespec __user *, interval)
5223 struct timespec64 t;
5224 int retval = sched_rr_get_interval(pid, &t);
5227 retval = compat_put_timespec64(&t, interval);
5232 void sched_show_task(struct task_struct *p)
5234 unsigned long free = 0;
5237 if (!try_get_task_stack(p))
5240 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5242 if (p->state == TASK_RUNNING)
5243 printk(KERN_CONT " running task ");
5244 #ifdef CONFIG_DEBUG_STACK_USAGE
5245 free = stack_not_used(p);
5250 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5252 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5253 task_pid_nr(p), ppid,
5254 (unsigned long)task_thread_info(p)->flags);
5256 print_worker_info(KERN_INFO, p);
5257 show_stack(p, NULL);
5260 EXPORT_SYMBOL_GPL(sched_show_task);
5263 state_filter_match(unsigned long state_filter, struct task_struct *p)
5265 /* no filter, everything matches */
5269 /* filter, but doesn't match */
5270 if (!(p->state & state_filter))
5274 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5277 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5284 void show_state_filter(unsigned long state_filter)
5286 struct task_struct *g, *p;
5288 #if BITS_PER_LONG == 32
5290 " task PC stack pid father\n");
5293 " task PC stack pid father\n");
5296 for_each_process_thread(g, p) {
5298 * reset the NMI-timeout, listing all files on a slow
5299 * console might take a lot of time:
5300 * Also, reset softlockup watchdogs on all CPUs, because
5301 * another CPU might be blocked waiting for us to process
5304 touch_nmi_watchdog();
5305 touch_all_softlockup_watchdogs();
5306 if (state_filter_match(state_filter, p))
5310 #ifdef CONFIG_SCHED_DEBUG
5312 sysrq_sched_debug_show();
5316 * Only show locks if all tasks are dumped:
5319 debug_show_all_locks();
5323 * init_idle - set up an idle thread for a given CPU
5324 * @idle: task in question
5325 * @cpu: CPU the idle task belongs to
5327 * NOTE: this function does not set the idle thread's NEED_RESCHED
5328 * flag, to make booting more robust.
5330 void init_idle(struct task_struct *idle, int cpu)
5332 struct rq *rq = cpu_rq(cpu);
5333 unsigned long flags;
5335 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5336 raw_spin_lock(&rq->lock);
5338 __sched_fork(0, idle);
5339 idle->state = TASK_RUNNING;
5340 idle->se.exec_start = sched_clock();
5341 idle->flags |= PF_IDLE;
5343 kasan_unpoison_task_stack(idle);
5347 * Its possible that init_idle() gets called multiple times on a task,
5348 * in that case do_set_cpus_allowed() will not do the right thing.
5350 * And since this is boot we can forgo the serialization.
5352 set_cpus_allowed_common(idle, cpumask_of(cpu));
5355 * We're having a chicken and egg problem, even though we are
5356 * holding rq->lock, the CPU isn't yet set to this CPU so the
5357 * lockdep check in task_group() will fail.
5359 * Similar case to sched_fork(). / Alternatively we could
5360 * use task_rq_lock() here and obtain the other rq->lock.
5365 __set_task_cpu(idle, cpu);
5368 rq->curr = rq->idle = idle;
5369 idle->on_rq = TASK_ON_RQ_QUEUED;
5373 raw_spin_unlock(&rq->lock);
5374 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5376 /* Set the preempt count _outside_ the spinlocks! */
5377 init_idle_preempt_count(idle, cpu);
5380 * The idle tasks have their own, simple scheduling class:
5382 idle->sched_class = &idle_sched_class;
5383 ftrace_graph_init_idle_task(idle, cpu);
5384 vtime_init_idle(idle, cpu);
5386 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5392 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5393 const struct cpumask *trial)
5397 if (!cpumask_weight(cur))
5400 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5405 int task_can_attach(struct task_struct *p,
5406 const struct cpumask *cs_cpus_allowed)
5411 * Kthreads which disallow setaffinity shouldn't be moved
5412 * to a new cpuset; we don't want to change their CPU
5413 * affinity and isolating such threads by their set of
5414 * allowed nodes is unnecessary. Thus, cpusets are not
5415 * applicable for such threads. This prevents checking for
5416 * success of set_cpus_allowed_ptr() on all attached tasks
5417 * before cpus_allowed may be changed.
5419 if (p->flags & PF_NO_SETAFFINITY) {
5424 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5426 ret = dl_task_can_attach(p, cs_cpus_allowed);
5432 bool sched_smp_initialized __read_mostly;
5434 #ifdef CONFIG_NUMA_BALANCING
5435 /* Migrate current task p to target_cpu */
5436 int migrate_task_to(struct task_struct *p, int target_cpu)
5438 struct migration_arg arg = { p, target_cpu };
5439 int curr_cpu = task_cpu(p);
5441 if (curr_cpu == target_cpu)
5444 if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5447 /* TODO: This is not properly updating schedstats */
5449 trace_sched_move_numa(p, curr_cpu, target_cpu);
5450 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5454 * Requeue a task on a given node and accurately track the number of NUMA
5455 * tasks on the runqueues
5457 void sched_setnuma(struct task_struct *p, int nid)
5459 bool queued, running;
5463 rq = task_rq_lock(p, &rf);
5464 queued = task_on_rq_queued(p);
5465 running = task_current(rq, p);
5468 dequeue_task(rq, p, DEQUEUE_SAVE);
5470 put_prev_task(rq, p);
5472 p->numa_preferred_nid = nid;
5475 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5477 set_curr_task(rq, p);
5478 task_rq_unlock(rq, p, &rf);
5480 #endif /* CONFIG_NUMA_BALANCING */
5482 #ifdef CONFIG_HOTPLUG_CPU
5484 * Ensure that the idle task is using init_mm right before its CPU goes
5487 void idle_task_exit(void)
5489 struct mm_struct *mm = current->active_mm;
5491 BUG_ON(cpu_online(smp_processor_id()));
5493 if (mm != &init_mm) {
5494 switch_mm(mm, &init_mm, current);
5495 finish_arch_post_lock_switch();
5501 * Since this CPU is going 'away' for a while, fold any nr_active delta
5502 * we might have. Assumes we're called after migrate_tasks() so that the
5503 * nr_active count is stable. We need to take the teardown thread which
5504 * is calling this into account, so we hand in adjust = 1 to the load
5507 * Also see the comment "Global load-average calculations".
5509 static void calc_load_migrate(struct rq *rq)
5511 long delta = calc_load_fold_active(rq, 1);
5513 atomic_long_add(delta, &calc_load_tasks);
5516 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5520 static const struct sched_class fake_sched_class = {
5521 .put_prev_task = put_prev_task_fake,
5524 static struct task_struct fake_task = {
5526 * Avoid pull_{rt,dl}_task()
5528 .prio = MAX_PRIO + 1,
5529 .sched_class = &fake_sched_class,
5533 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5534 * try_to_wake_up()->select_task_rq().
5536 * Called with rq->lock held even though we'er in stop_machine() and
5537 * there's no concurrency possible, we hold the required locks anyway
5538 * because of lock validation efforts.
5540 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5542 struct rq *rq = dead_rq;
5543 struct task_struct *next, *stop = rq->stop;
5544 struct rq_flags orf = *rf;
5548 * Fudge the rq selection such that the below task selection loop
5549 * doesn't get stuck on the currently eligible stop task.
5551 * We're currently inside stop_machine() and the rq is either stuck
5552 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5553 * either way we should never end up calling schedule() until we're
5559 * put_prev_task() and pick_next_task() sched
5560 * class method both need to have an up-to-date
5561 * value of rq->clock[_task]
5563 update_rq_clock(rq);
5567 * There's this thread running, bail when that's the only
5570 if (rq->nr_running == 1)
5574 * pick_next_task() assumes pinned rq->lock:
5576 next = pick_next_task(rq, &fake_task, rf);
5578 put_prev_task(rq, next);
5581 * Rules for changing task_struct::cpus_allowed are holding
5582 * both pi_lock and rq->lock, such that holding either
5583 * stabilizes the mask.
5585 * Drop rq->lock is not quite as disastrous as it usually is
5586 * because !cpu_active at this point, which means load-balance
5587 * will not interfere. Also, stop-machine.
5590 raw_spin_lock(&next->pi_lock);
5594 * Since we're inside stop-machine, _nothing_ should have
5595 * changed the task, WARN if weird stuff happened, because in
5596 * that case the above rq->lock drop is a fail too.
5598 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5599 raw_spin_unlock(&next->pi_lock);
5603 /* Find suitable destination for @next, with force if needed. */
5604 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5605 rq = __migrate_task(rq, rf, next, dest_cpu);
5606 if (rq != dead_rq) {
5612 raw_spin_unlock(&next->pi_lock);
5617 #endif /* CONFIG_HOTPLUG_CPU */
5619 void set_rq_online(struct rq *rq)
5622 const struct sched_class *class;
5624 cpumask_set_cpu(rq->cpu, rq->rd->online);
5627 for_each_class(class) {
5628 if (class->rq_online)
5629 class->rq_online(rq);
5634 void set_rq_offline(struct rq *rq)
5637 const struct sched_class *class;
5639 for_each_class(class) {
5640 if (class->rq_offline)
5641 class->rq_offline(rq);
5644 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5649 static void set_cpu_rq_start_time(unsigned int cpu)
5651 struct rq *rq = cpu_rq(cpu);
5653 rq->age_stamp = sched_clock_cpu(cpu);
5657 * used to mark begin/end of suspend/resume:
5659 static int num_cpus_frozen;
5662 * Update cpusets according to cpu_active mask. If cpusets are
5663 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5664 * around partition_sched_domains().
5666 * If we come here as part of a suspend/resume, don't touch cpusets because we
5667 * want to restore it back to its original state upon resume anyway.
5669 static void cpuset_cpu_active(void)
5671 if (cpuhp_tasks_frozen) {
5673 * num_cpus_frozen tracks how many CPUs are involved in suspend
5674 * resume sequence. As long as this is not the last online
5675 * operation in the resume sequence, just build a single sched
5676 * domain, ignoring cpusets.
5678 partition_sched_domains(1, NULL, NULL);
5679 if (--num_cpus_frozen)
5682 * This is the last CPU online operation. So fall through and
5683 * restore the original sched domains by considering the
5684 * cpuset configurations.
5686 cpuset_force_rebuild();
5688 cpuset_update_active_cpus();
5691 static int cpuset_cpu_inactive(unsigned int cpu)
5693 if (!cpuhp_tasks_frozen) {
5694 if (dl_cpu_busy(cpu))
5696 cpuset_update_active_cpus();
5699 partition_sched_domains(1, NULL, NULL);
5704 int sched_cpu_activate(unsigned int cpu)
5706 struct rq *rq = cpu_rq(cpu);
5709 set_cpu_active(cpu, true);
5711 if (sched_smp_initialized) {
5712 sched_domains_numa_masks_set(cpu);
5713 cpuset_cpu_active();
5717 * Put the rq online, if not already. This happens:
5719 * 1) In the early boot process, because we build the real domains
5720 * after all CPUs have been brought up.
5722 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5725 rq_lock_irqsave(rq, &rf);
5727 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5730 rq_unlock_irqrestore(rq, &rf);
5732 update_max_interval();
5737 int sched_cpu_deactivate(unsigned int cpu)
5741 set_cpu_active(cpu, false);
5743 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5744 * users of this state to go away such that all new such users will
5747 * Do sync before park smpboot threads to take care the rcu boost case.
5749 synchronize_rcu_mult(call_rcu, call_rcu_sched);
5751 if (!sched_smp_initialized)
5754 ret = cpuset_cpu_inactive(cpu);
5756 set_cpu_active(cpu, true);
5759 sched_domains_numa_masks_clear(cpu);
5763 static void sched_rq_cpu_starting(unsigned int cpu)
5765 struct rq *rq = cpu_rq(cpu);
5767 rq->calc_load_update = calc_load_update;
5768 update_max_interval();
5771 int sched_cpu_starting(unsigned int cpu)
5773 set_cpu_rq_start_time(cpu);
5774 sched_rq_cpu_starting(cpu);
5778 #ifdef CONFIG_HOTPLUG_CPU
5779 int sched_cpu_dying(unsigned int cpu)
5781 struct rq *rq = cpu_rq(cpu);
5784 /* Handle pending wakeups and then migrate everything off */
5785 sched_ttwu_pending();
5787 rq_lock_irqsave(rq, &rf);
5789 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5792 migrate_tasks(rq, &rf);
5793 BUG_ON(rq->nr_running != 1);
5794 rq_unlock_irqrestore(rq, &rf);
5796 calc_load_migrate(rq);
5797 update_max_interval();
5798 nohz_balance_exit_idle(cpu);
5804 #ifdef CONFIG_SCHED_SMT
5805 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5807 static void sched_init_smt(void)
5810 * We've enumerated all CPUs and will assume that if any CPU
5811 * has SMT siblings, CPU0 will too.
5813 if (cpumask_weight(cpu_smt_mask(0)) > 1)
5814 static_branch_enable(&sched_smt_present);
5817 static inline void sched_init_smt(void) { }
5820 void __init sched_init_smp(void)
5825 * There's no userspace yet to cause hotplug operations; hence all the
5826 * CPU masks are stable and all blatant races in the below code cannot
5829 mutex_lock(&sched_domains_mutex);
5830 sched_init_domains(cpu_active_mask);
5831 mutex_unlock(&sched_domains_mutex);
5833 /* Move init over to a non-isolated CPU */
5834 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
5836 sched_init_granularity();
5838 init_sched_rt_class();
5839 init_sched_dl_class();
5843 sched_smp_initialized = true;
5846 static int __init migration_init(void)
5848 sched_rq_cpu_starting(smp_processor_id());
5851 early_initcall(migration_init);
5854 void __init sched_init_smp(void)
5856 sched_init_granularity();
5858 #endif /* CONFIG_SMP */
5860 int in_sched_functions(unsigned long addr)
5862 return in_lock_functions(addr) ||
5863 (addr >= (unsigned long)__sched_text_start
5864 && addr < (unsigned long)__sched_text_end);
5867 #ifdef CONFIG_CGROUP_SCHED
5869 * Default task group.
5870 * Every task in system belongs to this group at bootup.
5872 struct task_group root_task_group;
5873 LIST_HEAD(task_groups);
5875 /* Cacheline aligned slab cache for task_group */
5876 static struct kmem_cache *task_group_cache __read_mostly;
5879 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5880 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5882 void __init sched_init(void)
5885 unsigned long alloc_size = 0, ptr;
5890 #ifdef CONFIG_FAIR_GROUP_SCHED
5891 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5893 #ifdef CONFIG_RT_GROUP_SCHED
5894 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5897 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5899 #ifdef CONFIG_FAIR_GROUP_SCHED
5900 root_task_group.se = (struct sched_entity **)ptr;
5901 ptr += nr_cpu_ids * sizeof(void **);
5903 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5904 ptr += nr_cpu_ids * sizeof(void **);
5906 #endif /* CONFIG_FAIR_GROUP_SCHED */
5907 #ifdef CONFIG_RT_GROUP_SCHED
5908 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5909 ptr += nr_cpu_ids * sizeof(void **);
5911 root_task_group.rt_rq = (struct rt_rq **)ptr;
5912 ptr += nr_cpu_ids * sizeof(void **);
5914 #endif /* CONFIG_RT_GROUP_SCHED */
5916 #ifdef CONFIG_CPUMASK_OFFSTACK
5917 for_each_possible_cpu(i) {
5918 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5919 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5920 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
5921 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5923 #endif /* CONFIG_CPUMASK_OFFSTACK */
5925 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
5926 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
5929 init_defrootdomain();
5932 #ifdef CONFIG_RT_GROUP_SCHED
5933 init_rt_bandwidth(&root_task_group.rt_bandwidth,
5934 global_rt_period(), global_rt_runtime());
5935 #endif /* CONFIG_RT_GROUP_SCHED */
5937 #ifdef CONFIG_CGROUP_SCHED
5938 task_group_cache = KMEM_CACHE(task_group, 0);
5940 list_add(&root_task_group.list, &task_groups);
5941 INIT_LIST_HEAD(&root_task_group.children);
5942 INIT_LIST_HEAD(&root_task_group.siblings);
5943 autogroup_init(&init_task);
5944 #endif /* CONFIG_CGROUP_SCHED */
5946 for_each_possible_cpu(i) {
5950 raw_spin_lock_init(&rq->lock);
5952 rq->calc_load_active = 0;
5953 rq->calc_load_update = jiffies + LOAD_FREQ;
5954 init_cfs_rq(&rq->cfs);
5955 init_rt_rq(&rq->rt);
5956 init_dl_rq(&rq->dl);
5957 #ifdef CONFIG_FAIR_GROUP_SCHED
5958 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
5959 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
5960 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
5962 * How much CPU bandwidth does root_task_group get?
5964 * In case of task-groups formed thr' the cgroup filesystem, it
5965 * gets 100% of the CPU resources in the system. This overall
5966 * system CPU resource is divided among the tasks of
5967 * root_task_group and its child task-groups in a fair manner,
5968 * based on each entity's (task or task-group's) weight
5969 * (se->load.weight).
5971 * In other words, if root_task_group has 10 tasks of weight
5972 * 1024) and two child groups A0 and A1 (of weight 1024 each),
5973 * then A0's share of the CPU resource is:
5975 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
5977 * We achieve this by letting root_task_group's tasks sit
5978 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
5980 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
5981 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
5982 #endif /* CONFIG_FAIR_GROUP_SCHED */
5984 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
5985 #ifdef CONFIG_RT_GROUP_SCHED
5986 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
5989 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
5990 rq->cpu_load[j] = 0;
5995 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
5996 rq->balance_callback = NULL;
5997 rq->active_balance = 0;
5998 rq->next_balance = jiffies;
6003 rq->avg_idle = 2*sysctl_sched_migration_cost;
6004 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6006 INIT_LIST_HEAD(&rq->cfs_tasks);
6008 rq_attach_root(rq, &def_root_domain);
6009 #ifdef CONFIG_NO_HZ_COMMON
6010 rq->last_load_update_tick = jiffies;
6013 #ifdef CONFIG_NO_HZ_FULL
6014 rq->last_sched_tick = 0;
6016 #endif /* CONFIG_SMP */
6018 atomic_set(&rq->nr_iowait, 0);
6021 set_load_weight(&init_task, false);
6024 * The boot idle thread does lazy MMU switching as well:
6027 enter_lazy_tlb(&init_mm, current);
6030 * Make us the idle thread. Technically, schedule() should not be
6031 * called from this thread, however somewhere below it might be,
6032 * but because we are the idle thread, we just pick up running again
6033 * when this runqueue becomes "idle".
6035 init_idle(current, smp_processor_id());
6037 calc_load_update = jiffies + LOAD_FREQ;
6040 idle_thread_set_boot_cpu();
6041 set_cpu_rq_start_time(smp_processor_id());
6043 init_sched_fair_class();
6047 scheduler_running = 1;
6050 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6051 static inline int preempt_count_equals(int preempt_offset)
6053 int nested = preempt_count() + rcu_preempt_depth();
6055 return (nested == preempt_offset);
6058 void __might_sleep(const char *file, int line, int preempt_offset)
6061 * Blocking primitives will set (and therefore destroy) current->state,
6062 * since we will exit with TASK_RUNNING make sure we enter with it,
6063 * otherwise we will destroy state.
6065 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6066 "do not call blocking ops when !TASK_RUNNING; "
6067 "state=%lx set at [<%p>] %pS\n",
6069 (void *)current->task_state_change,
6070 (void *)current->task_state_change);
6072 ___might_sleep(file, line, preempt_offset);
6074 EXPORT_SYMBOL(__might_sleep);
6076 void ___might_sleep(const char *file, int line, int preempt_offset)
6078 /* Ratelimiting timestamp: */
6079 static unsigned long prev_jiffy;
6081 unsigned long preempt_disable_ip;
6083 /* WARN_ON_ONCE() by default, no rate limit required: */
6086 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6087 !is_idle_task(current)) ||
6088 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6092 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6094 prev_jiffy = jiffies;
6096 /* Save this before calling printk(), since that will clobber it: */
6097 preempt_disable_ip = get_preempt_disable_ip(current);
6100 "BUG: sleeping function called from invalid context at %s:%d\n",
6103 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6104 in_atomic(), irqs_disabled(),
6105 current->pid, current->comm);
6107 if (task_stack_end_corrupted(current))
6108 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6110 debug_show_held_locks(current);
6111 if (irqs_disabled())
6112 print_irqtrace_events(current);
6113 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6114 && !preempt_count_equals(preempt_offset)) {
6115 pr_err("Preemption disabled at:");
6116 print_ip_sym(preempt_disable_ip);
6120 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6122 EXPORT_SYMBOL(___might_sleep);
6125 #ifdef CONFIG_MAGIC_SYSRQ
6126 void normalize_rt_tasks(void)
6128 struct task_struct *g, *p;
6129 struct sched_attr attr = {
6130 .sched_policy = SCHED_NORMAL,
6133 read_lock(&tasklist_lock);
6134 for_each_process_thread(g, p) {
6136 * Only normalize user tasks:
6138 if (p->flags & PF_KTHREAD)
6141 p->se.exec_start = 0;
6142 schedstat_set(p->se.statistics.wait_start, 0);
6143 schedstat_set(p->se.statistics.sleep_start, 0);
6144 schedstat_set(p->se.statistics.block_start, 0);
6146 if (!dl_task(p) && !rt_task(p)) {
6148 * Renice negative nice level userspace
6151 if (task_nice(p) < 0)
6152 set_user_nice(p, 0);
6156 __sched_setscheduler(p, &attr, false, false);
6158 read_unlock(&tasklist_lock);
6161 #endif /* CONFIG_MAGIC_SYSRQ */
6163 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6165 * These functions are only useful for the IA64 MCA handling, or kdb.
6167 * They can only be called when the whole system has been
6168 * stopped - every CPU needs to be quiescent, and no scheduling
6169 * activity can take place. Using them for anything else would
6170 * be a serious bug, and as a result, they aren't even visible
6171 * under any other configuration.
6175 * curr_task - return the current task for a given CPU.
6176 * @cpu: the processor in question.
6178 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6180 * Return: The current task for @cpu.
6182 struct task_struct *curr_task(int cpu)
6184 return cpu_curr(cpu);
6187 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6191 * set_curr_task - set the current task for a given CPU.
6192 * @cpu: the processor in question.
6193 * @p: the task pointer to set.
6195 * Description: This function must only be used when non-maskable interrupts
6196 * are serviced on a separate stack. It allows the architecture to switch the
6197 * notion of the current task on a CPU in a non-blocking manner. This function
6198 * must be called with all CPU's synchronized, and interrupts disabled, the
6199 * and caller must save the original value of the current task (see
6200 * curr_task() above) and restore that value before reenabling interrupts and
6201 * re-starting the system.
6203 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6205 void ia64_set_curr_task(int cpu, struct task_struct *p)
6212 #ifdef CONFIG_CGROUP_SCHED
6213 /* task_group_lock serializes the addition/removal of task groups */
6214 static DEFINE_SPINLOCK(task_group_lock);
6216 static void sched_free_group(struct task_group *tg)
6218 free_fair_sched_group(tg);
6219 free_rt_sched_group(tg);
6221 kmem_cache_free(task_group_cache, tg);
6224 /* allocate runqueue etc for a new task group */
6225 struct task_group *sched_create_group(struct task_group *parent)
6227 struct task_group *tg;
6229 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6231 return ERR_PTR(-ENOMEM);
6233 if (!alloc_fair_sched_group(tg, parent))
6236 if (!alloc_rt_sched_group(tg, parent))
6242 sched_free_group(tg);
6243 return ERR_PTR(-ENOMEM);
6246 void sched_online_group(struct task_group *tg, struct task_group *parent)
6248 unsigned long flags;
6250 spin_lock_irqsave(&task_group_lock, flags);
6251 list_add_rcu(&tg->list, &task_groups);
6253 /* Root should already exist: */
6256 tg->parent = parent;
6257 INIT_LIST_HEAD(&tg->children);
6258 list_add_rcu(&tg->siblings, &parent->children);
6259 spin_unlock_irqrestore(&task_group_lock, flags);
6261 online_fair_sched_group(tg);
6264 /* rcu callback to free various structures associated with a task group */
6265 static void sched_free_group_rcu(struct rcu_head *rhp)
6267 /* Now it should be safe to free those cfs_rqs: */
6268 sched_free_group(container_of(rhp, struct task_group, rcu));
6271 void sched_destroy_group(struct task_group *tg)
6273 /* Wait for possible concurrent references to cfs_rqs complete: */
6274 call_rcu(&tg->rcu, sched_free_group_rcu);
6277 void sched_offline_group(struct task_group *tg)
6279 unsigned long flags;
6281 /* End participation in shares distribution: */
6282 unregister_fair_sched_group(tg);
6284 spin_lock_irqsave(&task_group_lock, flags);
6285 list_del_rcu(&tg->list);
6286 list_del_rcu(&tg->siblings);
6287 spin_unlock_irqrestore(&task_group_lock, flags);
6290 static void sched_change_group(struct task_struct *tsk, int type)
6292 struct task_group *tg;
6295 * All callers are synchronized by task_rq_lock(); we do not use RCU
6296 * which is pointless here. Thus, we pass "true" to task_css_check()
6297 * to prevent lockdep warnings.
6299 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6300 struct task_group, css);
6301 tg = autogroup_task_group(tsk, tg);
6302 tsk->sched_task_group = tg;
6304 #ifdef CONFIG_FAIR_GROUP_SCHED
6305 if (tsk->sched_class->task_change_group)
6306 tsk->sched_class->task_change_group(tsk, type);
6309 set_task_rq(tsk, task_cpu(tsk));
6313 * Change task's runqueue when it moves between groups.
6315 * The caller of this function should have put the task in its new group by
6316 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6319 void sched_move_task(struct task_struct *tsk)
6321 int queued, running, queue_flags =
6322 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6326 rq = task_rq_lock(tsk, &rf);
6327 update_rq_clock(rq);
6329 running = task_current(rq, tsk);
6330 queued = task_on_rq_queued(tsk);
6333 dequeue_task(rq, tsk, queue_flags);
6335 put_prev_task(rq, tsk);
6337 sched_change_group(tsk, TASK_MOVE_GROUP);
6340 enqueue_task(rq, tsk, queue_flags);
6342 set_curr_task(rq, tsk);
6344 task_rq_unlock(rq, tsk, &rf);
6347 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6349 return css ? container_of(css, struct task_group, css) : NULL;
6352 static struct cgroup_subsys_state *
6353 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6355 struct task_group *parent = css_tg(parent_css);
6356 struct task_group *tg;
6359 /* This is early initialization for the top cgroup */
6360 return &root_task_group.css;
6363 tg = sched_create_group(parent);
6365 return ERR_PTR(-ENOMEM);
6370 /* Expose task group only after completing cgroup initialization */
6371 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6373 struct task_group *tg = css_tg(css);
6374 struct task_group *parent = css_tg(css->parent);
6377 sched_online_group(tg, parent);
6381 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6383 struct task_group *tg = css_tg(css);
6385 sched_offline_group(tg);
6388 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6390 struct task_group *tg = css_tg(css);
6393 * Relies on the RCU grace period between css_released() and this.
6395 sched_free_group(tg);
6399 * This is called before wake_up_new_task(), therefore we really only
6400 * have to set its group bits, all the other stuff does not apply.
6402 static void cpu_cgroup_fork(struct task_struct *task)
6407 rq = task_rq_lock(task, &rf);
6409 update_rq_clock(rq);
6410 sched_change_group(task, TASK_SET_GROUP);
6412 task_rq_unlock(rq, task, &rf);
6415 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6417 struct task_struct *task;
6418 struct cgroup_subsys_state *css;
6421 cgroup_taskset_for_each(task, css, tset) {
6422 #ifdef CONFIG_RT_GROUP_SCHED
6423 if (!sched_rt_can_attach(css_tg(css), task))
6426 /* We don't support RT-tasks being in separate groups */
6427 if (task->sched_class != &fair_sched_class)
6431 * Serialize against wake_up_new_task() such that if its
6432 * running, we're sure to observe its full state.
6434 raw_spin_lock_irq(&task->pi_lock);
6436 * Avoid calling sched_move_task() before wake_up_new_task()
6437 * has happened. This would lead to problems with PELT, due to
6438 * move wanting to detach+attach while we're not attached yet.
6440 if (task->state == TASK_NEW)
6442 raw_spin_unlock_irq(&task->pi_lock);
6450 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6452 struct task_struct *task;
6453 struct cgroup_subsys_state *css;
6455 cgroup_taskset_for_each(task, css, tset)
6456 sched_move_task(task);
6459 #ifdef CONFIG_FAIR_GROUP_SCHED
6460 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6461 struct cftype *cftype, u64 shareval)
6463 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6466 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6469 struct task_group *tg = css_tg(css);
6471 return (u64) scale_load_down(tg->shares);
6474 #ifdef CONFIG_CFS_BANDWIDTH
6475 static DEFINE_MUTEX(cfs_constraints_mutex);
6477 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6478 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6480 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6482 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6484 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6485 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6487 if (tg == &root_task_group)
6491 * Ensure we have at some amount of bandwidth every period. This is
6492 * to prevent reaching a state of large arrears when throttled via
6493 * entity_tick() resulting in prolonged exit starvation.
6495 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6499 * Likewise, bound things on the otherside by preventing insane quota
6500 * periods. This also allows us to normalize in computing quota
6503 if (period > max_cfs_quota_period)
6507 * Prevent race between setting of cfs_rq->runtime_enabled and
6508 * unthrottle_offline_cfs_rqs().
6511 mutex_lock(&cfs_constraints_mutex);
6512 ret = __cfs_schedulable(tg, period, quota);
6516 runtime_enabled = quota != RUNTIME_INF;
6517 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6519 * If we need to toggle cfs_bandwidth_used, off->on must occur
6520 * before making related changes, and on->off must occur afterwards
6522 if (runtime_enabled && !runtime_was_enabled)
6523 cfs_bandwidth_usage_inc();
6524 raw_spin_lock_irq(&cfs_b->lock);
6525 cfs_b->period = ns_to_ktime(period);
6526 cfs_b->quota = quota;
6528 __refill_cfs_bandwidth_runtime(cfs_b);
6530 /* Restart the period timer (if active) to handle new period expiry: */
6531 if (runtime_enabled)
6532 start_cfs_bandwidth(cfs_b);
6534 raw_spin_unlock_irq(&cfs_b->lock);
6536 for_each_online_cpu(i) {
6537 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6538 struct rq *rq = cfs_rq->rq;
6541 rq_lock_irq(rq, &rf);
6542 cfs_rq->runtime_enabled = runtime_enabled;
6543 cfs_rq->runtime_remaining = 0;
6545 if (cfs_rq->throttled)
6546 unthrottle_cfs_rq(cfs_rq);
6547 rq_unlock_irq(rq, &rf);
6549 if (runtime_was_enabled && !runtime_enabled)
6550 cfs_bandwidth_usage_dec();
6552 mutex_unlock(&cfs_constraints_mutex);
6558 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
6562 period = ktime_to_ns(tg->cfs_bandwidth.period);
6563 if (cfs_quota_us < 0)
6564 quota = RUNTIME_INF;
6566 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
6568 return tg_set_cfs_bandwidth(tg, period, quota);
6571 long tg_get_cfs_quota(struct task_group *tg)
6575 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
6578 quota_us = tg->cfs_bandwidth.quota;
6579 do_div(quota_us, NSEC_PER_USEC);
6584 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
6588 period = (u64)cfs_period_us * NSEC_PER_USEC;
6589 quota = tg->cfs_bandwidth.quota;
6591 return tg_set_cfs_bandwidth(tg, period, quota);
6594 long tg_get_cfs_period(struct task_group *tg)
6598 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
6599 do_div(cfs_period_us, NSEC_PER_USEC);
6601 return cfs_period_us;
6604 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
6607 return tg_get_cfs_quota(css_tg(css));
6610 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
6611 struct cftype *cftype, s64 cfs_quota_us)
6613 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
6616 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
6619 return tg_get_cfs_period(css_tg(css));
6622 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
6623 struct cftype *cftype, u64 cfs_period_us)
6625 return tg_set_cfs_period(css_tg(css), cfs_period_us);
6628 struct cfs_schedulable_data {
6629 struct task_group *tg;
6634 * normalize group quota/period to be quota/max_period
6635 * note: units are usecs
6637 static u64 normalize_cfs_quota(struct task_group *tg,
6638 struct cfs_schedulable_data *d)
6646 period = tg_get_cfs_period(tg);
6647 quota = tg_get_cfs_quota(tg);
6650 /* note: these should typically be equivalent */
6651 if (quota == RUNTIME_INF || quota == -1)
6654 return to_ratio(period, quota);
6657 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
6659 struct cfs_schedulable_data *d = data;
6660 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6661 s64 quota = 0, parent_quota = -1;
6664 quota = RUNTIME_INF;
6666 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
6668 quota = normalize_cfs_quota(tg, d);
6669 parent_quota = parent_b->hierarchical_quota;
6672 * Ensure max(child_quota) <= parent_quota, inherit when no
6675 if (quota == RUNTIME_INF)
6676 quota = parent_quota;
6677 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
6680 cfs_b->hierarchical_quota = quota;
6685 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
6688 struct cfs_schedulable_data data = {
6694 if (quota != RUNTIME_INF) {
6695 do_div(data.period, NSEC_PER_USEC);
6696 do_div(data.quota, NSEC_PER_USEC);
6700 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
6706 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
6708 struct task_group *tg = css_tg(seq_css(sf));
6709 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6711 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
6712 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
6713 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
6717 #endif /* CONFIG_CFS_BANDWIDTH */
6718 #endif /* CONFIG_FAIR_GROUP_SCHED */
6720 #ifdef CONFIG_RT_GROUP_SCHED
6721 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
6722 struct cftype *cft, s64 val)
6724 return sched_group_set_rt_runtime(css_tg(css), val);
6727 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
6730 return sched_group_rt_runtime(css_tg(css));
6733 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
6734 struct cftype *cftype, u64 rt_period_us)
6736 return sched_group_set_rt_period(css_tg(css), rt_period_us);
6739 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
6742 return sched_group_rt_period(css_tg(css));
6744 #endif /* CONFIG_RT_GROUP_SCHED */
6746 static struct cftype cpu_legacy_files[] = {
6747 #ifdef CONFIG_FAIR_GROUP_SCHED
6750 .read_u64 = cpu_shares_read_u64,
6751 .write_u64 = cpu_shares_write_u64,
6754 #ifdef CONFIG_CFS_BANDWIDTH
6756 .name = "cfs_quota_us",
6757 .read_s64 = cpu_cfs_quota_read_s64,
6758 .write_s64 = cpu_cfs_quota_write_s64,
6761 .name = "cfs_period_us",
6762 .read_u64 = cpu_cfs_period_read_u64,
6763 .write_u64 = cpu_cfs_period_write_u64,
6767 .seq_show = cpu_cfs_stat_show,
6770 #ifdef CONFIG_RT_GROUP_SCHED
6772 .name = "rt_runtime_us",
6773 .read_s64 = cpu_rt_runtime_read,
6774 .write_s64 = cpu_rt_runtime_write,
6777 .name = "rt_period_us",
6778 .read_u64 = cpu_rt_period_read_uint,
6779 .write_u64 = cpu_rt_period_write_uint,
6785 static int cpu_extra_stat_show(struct seq_file *sf,
6786 struct cgroup_subsys_state *css)
6788 #ifdef CONFIG_CFS_BANDWIDTH
6790 struct task_group *tg = css_tg(css);
6791 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6794 throttled_usec = cfs_b->throttled_time;
6795 do_div(throttled_usec, NSEC_PER_USEC);
6797 seq_printf(sf, "nr_periods %d\n"
6799 "throttled_usec %llu\n",
6800 cfs_b->nr_periods, cfs_b->nr_throttled,
6807 #ifdef CONFIG_FAIR_GROUP_SCHED
6808 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
6811 struct task_group *tg = css_tg(css);
6812 u64 weight = scale_load_down(tg->shares);
6814 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
6817 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
6818 struct cftype *cft, u64 weight)
6821 * cgroup weight knobs should use the common MIN, DFL and MAX
6822 * values which are 1, 100 and 10000 respectively. While it loses
6823 * a bit of range on both ends, it maps pretty well onto the shares
6824 * value used by scheduler and the round-trip conversions preserve
6825 * the original value over the entire range.
6827 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
6830 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
6832 return sched_group_set_shares(css_tg(css), scale_load(weight));
6835 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
6838 unsigned long weight = scale_load_down(css_tg(css)->shares);
6839 int last_delta = INT_MAX;
6842 /* find the closest nice value to the current weight */
6843 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
6844 delta = abs(sched_prio_to_weight[prio] - weight);
6845 if (delta >= last_delta)
6850 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
6853 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
6854 struct cftype *cft, s64 nice)
6856 unsigned long weight;
6858 if (nice < MIN_NICE || nice > MAX_NICE)
6861 weight = sched_prio_to_weight[NICE_TO_PRIO(nice) - MAX_RT_PRIO];
6862 return sched_group_set_shares(css_tg(css), scale_load(weight));
6866 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
6867 long period, long quota)
6870 seq_puts(sf, "max");
6872 seq_printf(sf, "%ld", quota);
6874 seq_printf(sf, " %ld\n", period);
6877 /* caller should put the current value in *@periodp before calling */
6878 static int __maybe_unused cpu_period_quota_parse(char *buf,
6879 u64 *periodp, u64 *quotap)
6881 char tok[21]; /* U64_MAX */
6883 if (!sscanf(buf, "%s %llu", tok, periodp))
6886 *periodp *= NSEC_PER_USEC;
6888 if (sscanf(tok, "%llu", quotap))
6889 *quotap *= NSEC_PER_USEC;
6890 else if (!strcmp(tok, "max"))
6891 *quotap = RUNTIME_INF;
6898 #ifdef CONFIG_CFS_BANDWIDTH
6899 static int cpu_max_show(struct seq_file *sf, void *v)
6901 struct task_group *tg = css_tg(seq_css(sf));
6903 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
6907 static ssize_t cpu_max_write(struct kernfs_open_file *of,
6908 char *buf, size_t nbytes, loff_t off)
6910 struct task_group *tg = css_tg(of_css(of));
6911 u64 period = tg_get_cfs_period(tg);
6915 ret = cpu_period_quota_parse(buf, &period, "a);
6917 ret = tg_set_cfs_bandwidth(tg, period, quota);
6918 return ret ?: nbytes;
6922 static struct cftype cpu_files[] = {
6923 #ifdef CONFIG_FAIR_GROUP_SCHED
6926 .flags = CFTYPE_NOT_ON_ROOT,
6927 .read_u64 = cpu_weight_read_u64,
6928 .write_u64 = cpu_weight_write_u64,
6931 .name = "weight.nice",
6932 .flags = CFTYPE_NOT_ON_ROOT,
6933 .read_s64 = cpu_weight_nice_read_s64,
6934 .write_s64 = cpu_weight_nice_write_s64,
6937 #ifdef CONFIG_CFS_BANDWIDTH
6940 .flags = CFTYPE_NOT_ON_ROOT,
6941 .seq_show = cpu_max_show,
6942 .write = cpu_max_write,
6948 struct cgroup_subsys cpu_cgrp_subsys = {
6949 .css_alloc = cpu_cgroup_css_alloc,
6950 .css_online = cpu_cgroup_css_online,
6951 .css_released = cpu_cgroup_css_released,
6952 .css_free = cpu_cgroup_css_free,
6953 .css_extra_stat_show = cpu_extra_stat_show,
6954 .fork = cpu_cgroup_fork,
6955 .can_attach = cpu_cgroup_can_attach,
6956 .attach = cpu_cgroup_attach,
6957 .legacy_cftypes = cpu_legacy_files,
6958 .dfl_cftypes = cpu_files,
6963 #endif /* CONFIG_CGROUP_SCHED */
6965 void dump_cpu_task(int cpu)
6967 pr_info("Task dump for CPU %d:\n", cpu);
6968 sched_show_task(cpu_curr(cpu));
6972 * Nice levels are multiplicative, with a gentle 10% change for every
6973 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
6974 * nice 1, it will get ~10% less CPU time than another CPU-bound task
6975 * that remained on nice 0.
6977 * The "10% effect" is relative and cumulative: from _any_ nice level,
6978 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
6979 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
6980 * If a task goes up by ~10% and another task goes down by ~10% then
6981 * the relative distance between them is ~25%.)
6983 const int sched_prio_to_weight[40] = {
6984 /* -20 */ 88761, 71755, 56483, 46273, 36291,
6985 /* -15 */ 29154, 23254, 18705, 14949, 11916,
6986 /* -10 */ 9548, 7620, 6100, 4904, 3906,
6987 /* -5 */ 3121, 2501, 1991, 1586, 1277,
6988 /* 0 */ 1024, 820, 655, 526, 423,
6989 /* 5 */ 335, 272, 215, 172, 137,
6990 /* 10 */ 110, 87, 70, 56, 45,
6991 /* 15 */ 36, 29, 23, 18, 15,
6995 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
6997 * In cases where the weight does not change often, we can use the
6998 * precalculated inverse to speed up arithmetics by turning divisions
6999 * into multiplications:
7001 const u32 sched_prio_to_wmult[40] = {
7002 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7003 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7004 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7005 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7006 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7007 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7008 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7009 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,