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>
20 #include <linux/blkdev.h>
21 #include <linux/kprobes.h>
22 #include <linux/mmu_context.h>
23 #include <linux/module.h>
24 #include <linux/nmi.h>
25 #include <linux/prefetch.h>
26 #include <linux/profile.h>
27 #include <linux/security.h>
28 #include <linux/syscalls.h>
30 #include <asm/switch_to.h>
32 #ifdef CONFIG_PARAVIRT
33 #include <asm/paravirt.h>
37 #include "../workqueue_internal.h"
38 #include "../smpboot.h"
40 #define CREATE_TRACE_POINTS
41 #include <trace/events/sched.h>
43 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
46 * Debugging: various feature bits
49 #define SCHED_FEAT(name, enabled) \
50 (1UL << __SCHED_FEAT_##name) * enabled |
52 const_debug unsigned int sysctl_sched_features =
59 * Number of tasks to iterate in a single balance run.
60 * Limited because this is done with IRQs disabled.
62 const_debug unsigned int sysctl_sched_nr_migrate = 32;
65 * period over which we average the RT time consumption, measured
70 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
73 * period over which we measure -rt task CPU usage in us.
76 unsigned int sysctl_sched_rt_period = 1000000;
78 __read_mostly int scheduler_running;
81 * part of the period that we allow rt tasks to run in us.
84 int sysctl_sched_rt_runtime = 950000;
86 /* CPUs with isolated domains */
87 cpumask_var_t cpu_isolated_map;
90 * __task_rq_lock - lock the rq @p resides on.
92 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
97 lockdep_assert_held(&p->pi_lock);
101 raw_spin_lock(&rq->lock);
102 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
106 raw_spin_unlock(&rq->lock);
108 while (unlikely(task_on_rq_migrating(p)))
114 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
116 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
117 __acquires(p->pi_lock)
123 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
125 raw_spin_lock(&rq->lock);
127 * move_queued_task() task_rq_lock()
130 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
131 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
132 * [S] ->cpu = new_cpu [L] task_rq()
136 * If we observe the old cpu in task_rq_lock, the acquire of
137 * the old rq->lock will fully serialize against the stores.
139 * If we observe the new CPU in task_rq_lock, the acquire will
140 * pair with the WMB to ensure we must then also see migrating.
142 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
146 raw_spin_unlock(&rq->lock);
147 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
149 while (unlikely(task_on_rq_migrating(p)))
155 * RQ-clock updating methods:
158 static void update_rq_clock_task(struct rq *rq, s64 delta)
161 * In theory, the compile should just see 0 here, and optimize out the call
162 * to sched_rt_avg_update. But I don't trust it...
164 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
165 s64 steal = 0, irq_delta = 0;
167 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
168 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
171 * Since irq_time is only updated on {soft,}irq_exit, we might run into
172 * this case when a previous update_rq_clock() happened inside a
175 * When this happens, we stop ->clock_task and only update the
176 * prev_irq_time stamp to account for the part that fit, so that a next
177 * update will consume the rest. This ensures ->clock_task is
180 * It does however cause some slight miss-attribution of {soft,}irq
181 * time, a more accurate solution would be to update the irq_time using
182 * the current rq->clock timestamp, except that would require using
185 if (irq_delta > delta)
188 rq->prev_irq_time += irq_delta;
191 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
192 if (static_key_false((¶virt_steal_rq_enabled))) {
193 steal = paravirt_steal_clock(cpu_of(rq));
194 steal -= rq->prev_steal_time_rq;
196 if (unlikely(steal > delta))
199 rq->prev_steal_time_rq += steal;
204 rq->clock_task += delta;
206 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
207 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
208 sched_rt_avg_update(rq, irq_delta + steal);
212 void update_rq_clock(struct rq *rq)
216 lockdep_assert_held(&rq->lock);
218 if (rq->clock_update_flags & RQCF_ACT_SKIP)
221 #ifdef CONFIG_SCHED_DEBUG
222 if (sched_feat(WARN_DOUBLE_CLOCK))
223 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
224 rq->clock_update_flags |= RQCF_UPDATED;
227 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
231 update_rq_clock_task(rq, delta);
235 #ifdef CONFIG_SCHED_HRTICK
237 * Use HR-timers to deliver accurate preemption points.
240 static void hrtick_clear(struct rq *rq)
242 if (hrtimer_active(&rq->hrtick_timer))
243 hrtimer_cancel(&rq->hrtick_timer);
247 * High-resolution timer tick.
248 * Runs from hardirq context with interrupts disabled.
250 static enum hrtimer_restart hrtick(struct hrtimer *timer)
252 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
255 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
259 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
262 return HRTIMER_NORESTART;
267 static void __hrtick_restart(struct rq *rq)
269 struct hrtimer *timer = &rq->hrtick_timer;
271 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
275 * called from hardirq (IPI) context
277 static void __hrtick_start(void *arg)
283 __hrtick_restart(rq);
284 rq->hrtick_csd_pending = 0;
289 * Called to set the hrtick timer state.
291 * called with rq->lock held and irqs disabled
293 void hrtick_start(struct rq *rq, u64 delay)
295 struct hrtimer *timer = &rq->hrtick_timer;
300 * Don't schedule slices shorter than 10000ns, that just
301 * doesn't make sense and can cause timer DoS.
303 delta = max_t(s64, delay, 10000LL);
304 time = ktime_add_ns(timer->base->get_time(), delta);
306 hrtimer_set_expires(timer, time);
308 if (rq == this_rq()) {
309 __hrtick_restart(rq);
310 } else if (!rq->hrtick_csd_pending) {
311 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
312 rq->hrtick_csd_pending = 1;
318 * Called to set the hrtick timer state.
320 * called with rq->lock held and irqs disabled
322 void hrtick_start(struct rq *rq, u64 delay)
325 * Don't schedule slices shorter than 10000ns, that just
326 * doesn't make sense. Rely on vruntime for fairness.
328 delay = max_t(u64, delay, 10000LL);
329 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
330 HRTIMER_MODE_REL_PINNED);
332 #endif /* CONFIG_SMP */
334 static void init_rq_hrtick(struct rq *rq)
337 rq->hrtick_csd_pending = 0;
339 rq->hrtick_csd.flags = 0;
340 rq->hrtick_csd.func = __hrtick_start;
341 rq->hrtick_csd.info = rq;
344 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
345 rq->hrtick_timer.function = hrtick;
347 #else /* CONFIG_SCHED_HRTICK */
348 static inline void hrtick_clear(struct rq *rq)
352 static inline void init_rq_hrtick(struct rq *rq)
355 #endif /* CONFIG_SCHED_HRTICK */
358 * cmpxchg based fetch_or, macro so it works for different integer types
360 #define fetch_or(ptr, mask) \
362 typeof(ptr) _ptr = (ptr); \
363 typeof(mask) _mask = (mask); \
364 typeof(*_ptr) _old, _val = *_ptr; \
367 _old = cmpxchg(_ptr, _val, _val | _mask); \
375 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
377 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
378 * this avoids any races wrt polling state changes and thereby avoids
381 static bool set_nr_and_not_polling(struct task_struct *p)
383 struct thread_info *ti = task_thread_info(p);
384 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
388 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
390 * If this returns true, then the idle task promises to call
391 * sched_ttwu_pending() and reschedule soon.
393 static bool set_nr_if_polling(struct task_struct *p)
395 struct thread_info *ti = task_thread_info(p);
396 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
399 if (!(val & _TIF_POLLING_NRFLAG))
401 if (val & _TIF_NEED_RESCHED)
403 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
412 static bool set_nr_and_not_polling(struct task_struct *p)
414 set_tsk_need_resched(p);
419 static bool set_nr_if_polling(struct task_struct *p)
426 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
428 struct wake_q_node *node = &task->wake_q;
431 * Atomically grab the task, if ->wake_q is !nil already it means
432 * its already queued (either by us or someone else) and will get the
433 * wakeup due to that.
435 * This cmpxchg() implies a full barrier, which pairs with the write
436 * barrier implied by the wakeup in wake_up_q().
438 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
441 get_task_struct(task);
444 * The head is context local, there can be no concurrency.
447 head->lastp = &node->next;
450 void wake_up_q(struct wake_q_head *head)
452 struct wake_q_node *node = head->first;
454 while (node != WAKE_Q_TAIL) {
455 struct task_struct *task;
457 task = container_of(node, struct task_struct, wake_q);
459 /* Task can safely be re-inserted now: */
461 task->wake_q.next = NULL;
464 * wake_up_process() implies a wmb() to pair with the queueing
465 * in wake_q_add() so as not to miss wakeups.
467 wake_up_process(task);
468 put_task_struct(task);
473 * resched_curr - mark rq's current task 'to be rescheduled now'.
475 * On UP this means the setting of the need_resched flag, on SMP it
476 * might also involve a cross-CPU call to trigger the scheduler on
479 void resched_curr(struct rq *rq)
481 struct task_struct *curr = rq->curr;
484 lockdep_assert_held(&rq->lock);
486 if (test_tsk_need_resched(curr))
491 if (cpu == smp_processor_id()) {
492 set_tsk_need_resched(curr);
493 set_preempt_need_resched();
497 if (set_nr_and_not_polling(curr))
498 smp_send_reschedule(cpu);
500 trace_sched_wake_idle_without_ipi(cpu);
503 void resched_cpu(int cpu)
505 struct rq *rq = cpu_rq(cpu);
508 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
511 raw_spin_unlock_irqrestore(&rq->lock, flags);
515 #ifdef CONFIG_NO_HZ_COMMON
517 * In the semi idle case, use the nearest busy CPU for migrating timers
518 * from an idle CPU. This is good for power-savings.
520 * We don't do similar optimization for completely idle system, as
521 * selecting an idle CPU will add more delays to the timers than intended
522 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
524 int get_nohz_timer_target(void)
526 int i, cpu = smp_processor_id();
527 struct sched_domain *sd;
529 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
533 for_each_domain(cpu, sd) {
534 for_each_cpu(i, sched_domain_span(sd)) {
538 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
545 if (!is_housekeeping_cpu(cpu))
546 cpu = housekeeping_any_cpu();
553 * When add_timer_on() enqueues a timer into the timer wheel of an
554 * idle CPU then this timer might expire before the next timer event
555 * which is scheduled to wake up that CPU. In case of a completely
556 * idle system the next event might even be infinite time into the
557 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
558 * leaves the inner idle loop so the newly added timer is taken into
559 * account when the CPU goes back to idle and evaluates the timer
560 * wheel for the next timer event.
562 static void wake_up_idle_cpu(int cpu)
564 struct rq *rq = cpu_rq(cpu);
566 if (cpu == smp_processor_id())
569 if (set_nr_and_not_polling(rq->idle))
570 smp_send_reschedule(cpu);
572 trace_sched_wake_idle_without_ipi(cpu);
575 static bool wake_up_full_nohz_cpu(int cpu)
578 * We just need the target to call irq_exit() and re-evaluate
579 * the next tick. The nohz full kick at least implies that.
580 * If needed we can still optimize that later with an
583 if (cpu_is_offline(cpu))
584 return true; /* Don't try to wake offline CPUs. */
585 if (tick_nohz_full_cpu(cpu)) {
586 if (cpu != smp_processor_id() ||
587 tick_nohz_tick_stopped())
588 tick_nohz_full_kick_cpu(cpu);
596 * Wake up the specified CPU. If the CPU is going offline, it is the
597 * caller's responsibility to deal with the lost wakeup, for example,
598 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
600 void wake_up_nohz_cpu(int cpu)
602 if (!wake_up_full_nohz_cpu(cpu))
603 wake_up_idle_cpu(cpu);
606 static inline bool got_nohz_idle_kick(void)
608 int cpu = smp_processor_id();
610 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
613 if (idle_cpu(cpu) && !need_resched())
617 * We can't run Idle Load Balance on this CPU for this time so we
618 * cancel it and clear NOHZ_BALANCE_KICK
620 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
624 #else /* CONFIG_NO_HZ_COMMON */
626 static inline bool got_nohz_idle_kick(void)
631 #endif /* CONFIG_NO_HZ_COMMON */
633 #ifdef CONFIG_NO_HZ_FULL
634 bool sched_can_stop_tick(struct rq *rq)
638 /* Deadline tasks, even if single, need the tick */
639 if (rq->dl.dl_nr_running)
643 * If there are more than one RR tasks, we need the tick to effect the
644 * actual RR behaviour.
646 if (rq->rt.rr_nr_running) {
647 if (rq->rt.rr_nr_running == 1)
654 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
655 * forced preemption between FIFO tasks.
657 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
662 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
663 * if there's more than one we need the tick for involuntary
666 if (rq->nr_running > 1)
671 #endif /* CONFIG_NO_HZ_FULL */
673 void sched_avg_update(struct rq *rq)
675 s64 period = sched_avg_period();
677 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
679 * Inline assembly required to prevent the compiler
680 * optimising this loop into a divmod call.
681 * See __iter_div_u64_rem() for another example of this.
683 asm("" : "+rm" (rq->age_stamp));
684 rq->age_stamp += period;
689 #endif /* CONFIG_SMP */
691 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
692 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
694 * Iterate task_group tree rooted at *from, calling @down when first entering a
695 * node and @up when leaving it for the final time.
697 * Caller must hold rcu_lock or sufficient equivalent.
699 int walk_tg_tree_from(struct task_group *from,
700 tg_visitor down, tg_visitor up, void *data)
702 struct task_group *parent, *child;
708 ret = (*down)(parent, data);
711 list_for_each_entry_rcu(child, &parent->children, siblings) {
718 ret = (*up)(parent, data);
719 if (ret || parent == from)
723 parent = parent->parent;
730 int tg_nop(struct task_group *tg, void *data)
736 static void set_load_weight(struct task_struct *p)
738 int prio = p->static_prio - MAX_RT_PRIO;
739 struct load_weight *load = &p->se.load;
742 * SCHED_IDLE tasks get minimal weight:
744 if (idle_policy(p->policy)) {
745 load->weight = scale_load(WEIGHT_IDLEPRIO);
746 load->inv_weight = WMULT_IDLEPRIO;
750 load->weight = scale_load(sched_prio_to_weight[prio]);
751 load->inv_weight = sched_prio_to_wmult[prio];
754 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
756 if (!(flags & ENQUEUE_NOCLOCK))
759 if (!(flags & ENQUEUE_RESTORE))
760 sched_info_queued(rq, p);
762 p->sched_class->enqueue_task(rq, p, flags);
765 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
767 if (!(flags & DEQUEUE_NOCLOCK))
770 if (!(flags & DEQUEUE_SAVE))
771 sched_info_dequeued(rq, p);
773 p->sched_class->dequeue_task(rq, p, flags);
776 void activate_task(struct rq *rq, struct task_struct *p, int flags)
778 if (task_contributes_to_load(p))
779 rq->nr_uninterruptible--;
781 enqueue_task(rq, p, flags);
784 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
786 if (task_contributes_to_load(p))
787 rq->nr_uninterruptible++;
789 dequeue_task(rq, p, flags);
793 * __normal_prio - return the priority that is based on the static prio
795 static inline int __normal_prio(struct task_struct *p)
797 return p->static_prio;
801 * Calculate the expected normal priority: i.e. priority
802 * without taking RT-inheritance into account. Might be
803 * boosted by interactivity modifiers. Changes upon fork,
804 * setprio syscalls, and whenever the interactivity
805 * estimator recalculates.
807 static inline int normal_prio(struct task_struct *p)
811 if (task_has_dl_policy(p))
812 prio = MAX_DL_PRIO-1;
813 else if (task_has_rt_policy(p))
814 prio = MAX_RT_PRIO-1 - p->rt_priority;
816 prio = __normal_prio(p);
821 * Calculate the current priority, i.e. the priority
822 * taken into account by the scheduler. This value might
823 * be boosted by RT tasks, or might be boosted by
824 * interactivity modifiers. Will be RT if the task got
825 * RT-boosted. If not then it returns p->normal_prio.
827 static int effective_prio(struct task_struct *p)
829 p->normal_prio = normal_prio(p);
831 * If we are RT tasks or we were boosted to RT priority,
832 * keep the priority unchanged. Otherwise, update priority
833 * to the normal priority:
835 if (!rt_prio(p->prio))
836 return p->normal_prio;
841 * task_curr - is this task currently executing on a CPU?
842 * @p: the task in question.
844 * Return: 1 if the task is currently executing. 0 otherwise.
846 inline int task_curr(const struct task_struct *p)
848 return cpu_curr(task_cpu(p)) == p;
852 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
853 * use the balance_callback list if you want balancing.
855 * this means any call to check_class_changed() must be followed by a call to
856 * balance_callback().
858 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
859 const struct sched_class *prev_class,
862 if (prev_class != p->sched_class) {
863 if (prev_class->switched_from)
864 prev_class->switched_from(rq, p);
866 p->sched_class->switched_to(rq, p);
867 } else if (oldprio != p->prio || dl_task(p))
868 p->sched_class->prio_changed(rq, p, oldprio);
871 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
873 const struct sched_class *class;
875 if (p->sched_class == rq->curr->sched_class) {
876 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
878 for_each_class(class) {
879 if (class == rq->curr->sched_class)
881 if (class == p->sched_class) {
889 * A queue event has occurred, and we're going to schedule. In
890 * this case, we can save a useless back to back clock update.
892 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
893 rq_clock_skip_update(rq, true);
898 * This is how migration works:
900 * 1) we invoke migration_cpu_stop() on the target CPU using
902 * 2) stopper starts to run (implicitly forcing the migrated thread
904 * 3) it checks whether the migrated task is still in the wrong runqueue.
905 * 4) if it's in the wrong runqueue then the migration thread removes
906 * it and puts it into the right queue.
907 * 5) stopper completes and stop_one_cpu() returns and the migration
912 * move_queued_task - move a queued task to new rq.
914 * Returns (locked) new rq. Old rq's lock is released.
916 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
917 struct task_struct *p, int new_cpu)
919 lockdep_assert_held(&rq->lock);
921 p->on_rq = TASK_ON_RQ_MIGRATING;
922 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
923 set_task_cpu(p, new_cpu);
926 rq = cpu_rq(new_cpu);
929 BUG_ON(task_cpu(p) != new_cpu);
930 enqueue_task(rq, p, 0);
931 p->on_rq = TASK_ON_RQ_QUEUED;
932 check_preempt_curr(rq, p, 0);
937 struct migration_arg {
938 struct task_struct *task;
943 * Move (not current) task off this CPU, onto the destination CPU. We're doing
944 * this because either it can't run here any more (set_cpus_allowed()
945 * away from this CPU, or CPU going down), or because we're
946 * attempting to rebalance this task on exec (sched_exec).
948 * So we race with normal scheduler movements, but that's OK, as long
949 * as the task is no longer on this CPU.
951 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
952 struct task_struct *p, int dest_cpu)
954 if (p->flags & PF_KTHREAD) {
955 if (unlikely(!cpu_online(dest_cpu)))
958 if (unlikely(!cpu_active(dest_cpu)))
962 /* Affinity changed (again). */
963 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
967 rq = move_queued_task(rq, rf, p, dest_cpu);
973 * migration_cpu_stop - this will be executed by a highprio stopper thread
974 * and performs thread migration by bumping thread off CPU then
975 * 'pushing' onto another runqueue.
977 static int migration_cpu_stop(void *data)
979 struct migration_arg *arg = data;
980 struct task_struct *p = arg->task;
981 struct rq *rq = this_rq();
985 * The original target CPU might have gone down and we might
986 * be on another CPU but it doesn't matter.
990 * We need to explicitly wake pending tasks before running
991 * __migrate_task() such that we will not miss enforcing cpus_allowed
992 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
994 sched_ttwu_pending();
996 raw_spin_lock(&p->pi_lock);
999 * If task_rq(p) != rq, it cannot be migrated here, because we're
1000 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1001 * we're holding p->pi_lock.
1003 if (task_rq(p) == rq) {
1004 if (task_on_rq_queued(p))
1005 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1007 p->wake_cpu = arg->dest_cpu;
1010 raw_spin_unlock(&p->pi_lock);
1017 * sched_class::set_cpus_allowed must do the below, but is not required to
1018 * actually call this function.
1020 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1022 cpumask_copy(&p->cpus_allowed, new_mask);
1023 p->nr_cpus_allowed = cpumask_weight(new_mask);
1026 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1028 struct rq *rq = task_rq(p);
1029 bool queued, running;
1031 lockdep_assert_held(&p->pi_lock);
1033 queued = task_on_rq_queued(p);
1034 running = task_current(rq, p);
1038 * Because __kthread_bind() calls this on blocked tasks without
1041 lockdep_assert_held(&rq->lock);
1042 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1045 put_prev_task(rq, p);
1047 p->sched_class->set_cpus_allowed(p, new_mask);
1050 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1052 set_curr_task(rq, p);
1056 * Change a given task's CPU affinity. Migrate the thread to a
1057 * proper CPU and schedule it away if the CPU it's executing on
1058 * is removed from the allowed bitmask.
1060 * NOTE: the caller must have a valid reference to the task, the
1061 * task must not exit() & deallocate itself prematurely. The
1062 * call is not atomic; no spinlocks may be held.
1064 static int __set_cpus_allowed_ptr(struct task_struct *p,
1065 const struct cpumask *new_mask, bool check)
1067 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1068 unsigned int dest_cpu;
1073 rq = task_rq_lock(p, &rf);
1074 update_rq_clock(rq);
1076 if (p->flags & PF_KTHREAD) {
1078 * Kernel threads are allowed on online && !active CPUs
1080 cpu_valid_mask = cpu_online_mask;
1084 * Must re-check here, to close a race against __kthread_bind(),
1085 * sched_setaffinity() is not guaranteed to observe the flag.
1087 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1092 if (cpumask_equal(&p->cpus_allowed, new_mask))
1095 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1100 do_set_cpus_allowed(p, new_mask);
1102 if (p->flags & PF_KTHREAD) {
1104 * For kernel threads that do indeed end up on online &&
1105 * !active we want to ensure they are strict per-CPU threads.
1107 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1108 !cpumask_intersects(new_mask, cpu_active_mask) &&
1109 p->nr_cpus_allowed != 1);
1112 /* Can the task run on the task's current CPU? If so, we're done */
1113 if (cpumask_test_cpu(task_cpu(p), new_mask))
1116 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1117 if (task_running(rq, p) || p->state == TASK_WAKING) {
1118 struct migration_arg arg = { p, dest_cpu };
1119 /* Need help from migration thread: drop lock and wait. */
1120 task_rq_unlock(rq, p, &rf);
1121 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1122 tlb_migrate_finish(p->mm);
1124 } else if (task_on_rq_queued(p)) {
1126 * OK, since we're going to drop the lock immediately
1127 * afterwards anyway.
1129 rq = move_queued_task(rq, &rf, p, dest_cpu);
1132 task_rq_unlock(rq, p, &rf);
1137 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1139 return __set_cpus_allowed_ptr(p, new_mask, false);
1141 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1143 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1145 #ifdef CONFIG_SCHED_DEBUG
1147 * We should never call set_task_cpu() on a blocked task,
1148 * ttwu() will sort out the placement.
1150 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1154 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1155 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1156 * time relying on p->on_rq.
1158 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1159 p->sched_class == &fair_sched_class &&
1160 (p->on_rq && !task_on_rq_migrating(p)));
1162 #ifdef CONFIG_LOCKDEP
1164 * The caller should hold either p->pi_lock or rq->lock, when changing
1165 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1167 * sched_move_task() holds both and thus holding either pins the cgroup,
1170 * Furthermore, all task_rq users should acquire both locks, see
1173 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1174 lockdep_is_held(&task_rq(p)->lock)));
1178 trace_sched_migrate_task(p, new_cpu);
1180 if (task_cpu(p) != new_cpu) {
1181 if (p->sched_class->migrate_task_rq)
1182 p->sched_class->migrate_task_rq(p);
1183 p->se.nr_migrations++;
1184 perf_event_task_migrate(p);
1187 __set_task_cpu(p, new_cpu);
1190 static void __migrate_swap_task(struct task_struct *p, int cpu)
1192 if (task_on_rq_queued(p)) {
1193 struct rq *src_rq, *dst_rq;
1194 struct rq_flags srf, drf;
1196 src_rq = task_rq(p);
1197 dst_rq = cpu_rq(cpu);
1199 rq_pin_lock(src_rq, &srf);
1200 rq_pin_lock(dst_rq, &drf);
1202 p->on_rq = TASK_ON_RQ_MIGRATING;
1203 deactivate_task(src_rq, p, 0);
1204 set_task_cpu(p, cpu);
1205 activate_task(dst_rq, p, 0);
1206 p->on_rq = TASK_ON_RQ_QUEUED;
1207 check_preempt_curr(dst_rq, p, 0);
1209 rq_unpin_lock(dst_rq, &drf);
1210 rq_unpin_lock(src_rq, &srf);
1214 * Task isn't running anymore; make it appear like we migrated
1215 * it before it went to sleep. This means on wakeup we make the
1216 * previous CPU our target instead of where it really is.
1222 struct migration_swap_arg {
1223 struct task_struct *src_task, *dst_task;
1224 int src_cpu, dst_cpu;
1227 static int migrate_swap_stop(void *data)
1229 struct migration_swap_arg *arg = data;
1230 struct rq *src_rq, *dst_rq;
1233 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1236 src_rq = cpu_rq(arg->src_cpu);
1237 dst_rq = cpu_rq(arg->dst_cpu);
1239 double_raw_lock(&arg->src_task->pi_lock,
1240 &arg->dst_task->pi_lock);
1241 double_rq_lock(src_rq, dst_rq);
1243 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1246 if (task_cpu(arg->src_task) != arg->src_cpu)
1249 if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1252 if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1255 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1256 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1261 double_rq_unlock(src_rq, dst_rq);
1262 raw_spin_unlock(&arg->dst_task->pi_lock);
1263 raw_spin_unlock(&arg->src_task->pi_lock);
1269 * Cross migrate two tasks
1271 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1273 struct migration_swap_arg arg;
1276 arg = (struct migration_swap_arg){
1278 .src_cpu = task_cpu(cur),
1280 .dst_cpu = task_cpu(p),
1283 if (arg.src_cpu == arg.dst_cpu)
1287 * These three tests are all lockless; this is OK since all of them
1288 * will be re-checked with proper locks held further down the line.
1290 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1293 if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1296 if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1299 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1300 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1307 * wait_task_inactive - wait for a thread to unschedule.
1309 * If @match_state is nonzero, it's the @p->state value just checked and
1310 * not expected to change. If it changes, i.e. @p might have woken up,
1311 * then return zero. When we succeed in waiting for @p to be off its CPU,
1312 * we return a positive number (its total switch count). If a second call
1313 * a short while later returns the same number, the caller can be sure that
1314 * @p has remained unscheduled the whole time.
1316 * The caller must ensure that the task *will* unschedule sometime soon,
1317 * else this function might spin for a *long* time. This function can't
1318 * be called with interrupts off, or it may introduce deadlock with
1319 * smp_call_function() if an IPI is sent by the same process we are
1320 * waiting to become inactive.
1322 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1324 int running, queued;
1331 * We do the initial early heuristics without holding
1332 * any task-queue locks at all. We'll only try to get
1333 * the runqueue lock when things look like they will
1339 * If the task is actively running on another CPU
1340 * still, just relax and busy-wait without holding
1343 * NOTE! Since we don't hold any locks, it's not
1344 * even sure that "rq" stays as the right runqueue!
1345 * But we don't care, since "task_running()" will
1346 * return false if the runqueue has changed and p
1347 * is actually now running somewhere else!
1349 while (task_running(rq, p)) {
1350 if (match_state && unlikely(p->state != match_state))
1356 * Ok, time to look more closely! We need the rq
1357 * lock now, to be *sure*. If we're wrong, we'll
1358 * just go back and repeat.
1360 rq = task_rq_lock(p, &rf);
1361 trace_sched_wait_task(p);
1362 running = task_running(rq, p);
1363 queued = task_on_rq_queued(p);
1365 if (!match_state || p->state == match_state)
1366 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1367 task_rq_unlock(rq, p, &rf);
1370 * If it changed from the expected state, bail out now.
1372 if (unlikely(!ncsw))
1376 * Was it really running after all now that we
1377 * checked with the proper locks actually held?
1379 * Oops. Go back and try again..
1381 if (unlikely(running)) {
1387 * It's not enough that it's not actively running,
1388 * it must be off the runqueue _entirely_, and not
1391 * So if it was still runnable (but just not actively
1392 * running right now), it's preempted, and we should
1393 * yield - it could be a while.
1395 if (unlikely(queued)) {
1396 ktime_t to = NSEC_PER_SEC / HZ;
1398 set_current_state(TASK_UNINTERRUPTIBLE);
1399 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1404 * Ahh, all good. It wasn't running, and it wasn't
1405 * runnable, which means that it will never become
1406 * running in the future either. We're all done!
1415 * kick_process - kick a running thread to enter/exit the kernel
1416 * @p: the to-be-kicked thread
1418 * Cause a process which is running on another CPU to enter
1419 * kernel-mode, without any delay. (to get signals handled.)
1421 * NOTE: this function doesn't have to take the runqueue lock,
1422 * because all it wants to ensure is that the remote task enters
1423 * the kernel. If the IPI races and the task has been migrated
1424 * to another CPU then no harm is done and the purpose has been
1427 void kick_process(struct task_struct *p)
1433 if ((cpu != smp_processor_id()) && task_curr(p))
1434 smp_send_reschedule(cpu);
1437 EXPORT_SYMBOL_GPL(kick_process);
1440 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1442 * A few notes on cpu_active vs cpu_online:
1444 * - cpu_active must be a subset of cpu_online
1446 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1447 * see __set_cpus_allowed_ptr(). At this point the newly online
1448 * CPU isn't yet part of the sched domains, and balancing will not
1451 * - on CPU-down we clear cpu_active() to mask the sched domains and
1452 * avoid the load balancer to place new tasks on the to be removed
1453 * CPU. Existing tasks will remain running there and will be taken
1456 * This means that fallback selection must not select !active CPUs.
1457 * And can assume that any active CPU must be online. Conversely
1458 * select_task_rq() below may allow selection of !active CPUs in order
1459 * to satisfy the above rules.
1461 static int select_fallback_rq(int cpu, struct task_struct *p)
1463 int nid = cpu_to_node(cpu);
1464 const struct cpumask *nodemask = NULL;
1465 enum { cpuset, possible, fail } state = cpuset;
1469 * If the node that the CPU is on has been offlined, cpu_to_node()
1470 * will return -1. There is no CPU on the node, and we should
1471 * select the CPU on the other node.
1474 nodemask = cpumask_of_node(nid);
1476 /* Look for allowed, online CPU in same node. */
1477 for_each_cpu(dest_cpu, nodemask) {
1478 if (!cpu_active(dest_cpu))
1480 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1486 /* Any allowed, online CPU? */
1487 for_each_cpu(dest_cpu, &p->cpus_allowed) {
1488 if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1490 if (!cpu_online(dest_cpu))
1495 /* No more Mr. Nice Guy. */
1498 if (IS_ENABLED(CONFIG_CPUSETS)) {
1499 cpuset_cpus_allowed_fallback(p);
1505 do_set_cpus_allowed(p, cpu_possible_mask);
1516 if (state != cpuset) {
1518 * Don't tell them about moving exiting tasks or
1519 * kernel threads (both mm NULL), since they never
1522 if (p->mm && printk_ratelimit()) {
1523 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1524 task_pid_nr(p), p->comm, cpu);
1532 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1535 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1537 lockdep_assert_held(&p->pi_lock);
1539 if (p->nr_cpus_allowed > 1)
1540 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1542 cpu = cpumask_any(&p->cpus_allowed);
1545 * In order not to call set_task_cpu() on a blocking task we need
1546 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1549 * Since this is common to all placement strategies, this lives here.
1551 * [ this allows ->select_task() to simply return task_cpu(p) and
1552 * not worry about this generic constraint ]
1554 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
1556 cpu = select_fallback_rq(task_cpu(p), p);
1561 static void update_avg(u64 *avg, u64 sample)
1563 s64 diff = sample - *avg;
1567 void sched_set_stop_task(int cpu, struct task_struct *stop)
1569 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1570 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1574 * Make it appear like a SCHED_FIFO task, its something
1575 * userspace knows about and won't get confused about.
1577 * Also, it will make PI more or less work without too
1578 * much confusion -- but then, stop work should not
1579 * rely on PI working anyway.
1581 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1583 stop->sched_class = &stop_sched_class;
1586 cpu_rq(cpu)->stop = stop;
1590 * Reset it back to a normal scheduling class so that
1591 * it can die in pieces.
1593 old_stop->sched_class = &rt_sched_class;
1599 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1600 const struct cpumask *new_mask, bool check)
1602 return set_cpus_allowed_ptr(p, new_mask);
1605 #endif /* CONFIG_SMP */
1608 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1612 if (!schedstat_enabled())
1618 if (cpu == rq->cpu) {
1619 schedstat_inc(rq->ttwu_local);
1620 schedstat_inc(p->se.statistics.nr_wakeups_local);
1622 struct sched_domain *sd;
1624 schedstat_inc(p->se.statistics.nr_wakeups_remote);
1626 for_each_domain(rq->cpu, sd) {
1627 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1628 schedstat_inc(sd->ttwu_wake_remote);
1635 if (wake_flags & WF_MIGRATED)
1636 schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1637 #endif /* CONFIG_SMP */
1639 schedstat_inc(rq->ttwu_count);
1640 schedstat_inc(p->se.statistics.nr_wakeups);
1642 if (wake_flags & WF_SYNC)
1643 schedstat_inc(p->se.statistics.nr_wakeups_sync);
1646 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1648 activate_task(rq, p, en_flags);
1649 p->on_rq = TASK_ON_RQ_QUEUED;
1651 /* If a worker is waking up, notify the workqueue: */
1652 if (p->flags & PF_WQ_WORKER)
1653 wq_worker_waking_up(p, cpu_of(rq));
1657 * Mark the task runnable and perform wakeup-preemption.
1659 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1660 struct rq_flags *rf)
1662 check_preempt_curr(rq, p, wake_flags);
1663 p->state = TASK_RUNNING;
1664 trace_sched_wakeup(p);
1667 if (p->sched_class->task_woken) {
1669 * Our task @p is fully woken up and running; so its safe to
1670 * drop the rq->lock, hereafter rq is only used for statistics.
1672 rq_unpin_lock(rq, rf);
1673 p->sched_class->task_woken(rq, p);
1674 rq_repin_lock(rq, rf);
1677 if (rq->idle_stamp) {
1678 u64 delta = rq_clock(rq) - rq->idle_stamp;
1679 u64 max = 2*rq->max_idle_balance_cost;
1681 update_avg(&rq->avg_idle, delta);
1683 if (rq->avg_idle > max)
1692 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1693 struct rq_flags *rf)
1695 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
1697 lockdep_assert_held(&rq->lock);
1700 if (p->sched_contributes_to_load)
1701 rq->nr_uninterruptible--;
1703 if (wake_flags & WF_MIGRATED)
1704 en_flags |= ENQUEUE_MIGRATED;
1707 ttwu_activate(rq, p, en_flags);
1708 ttwu_do_wakeup(rq, p, wake_flags, rf);
1712 * Called in case the task @p isn't fully descheduled from its runqueue,
1713 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1714 * since all we need to do is flip p->state to TASK_RUNNING, since
1715 * the task is still ->on_rq.
1717 static int ttwu_remote(struct task_struct *p, int wake_flags)
1723 rq = __task_rq_lock(p, &rf);
1724 if (task_on_rq_queued(p)) {
1725 /* check_preempt_curr() may use rq clock */
1726 update_rq_clock(rq);
1727 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1730 __task_rq_unlock(rq, &rf);
1736 void sched_ttwu_pending(void)
1738 struct rq *rq = this_rq();
1739 struct llist_node *llist = llist_del_all(&rq->wake_list);
1740 struct task_struct *p, *t;
1746 rq_lock_irqsave(rq, &rf);
1747 update_rq_clock(rq);
1749 llist_for_each_entry_safe(p, t, llist, wake_entry)
1750 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
1752 rq_unlock_irqrestore(rq, &rf);
1755 void scheduler_ipi(void)
1758 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1759 * TIF_NEED_RESCHED remotely (for the first time) will also send
1762 preempt_fold_need_resched();
1764 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1768 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1769 * traditionally all their work was done from the interrupt return
1770 * path. Now that we actually do some work, we need to make sure
1773 * Some archs already do call them, luckily irq_enter/exit nest
1776 * Arguably we should visit all archs and update all handlers,
1777 * however a fair share of IPIs are still resched only so this would
1778 * somewhat pessimize the simple resched case.
1781 sched_ttwu_pending();
1784 * Check if someone kicked us for doing the nohz idle load balance.
1786 if (unlikely(got_nohz_idle_kick())) {
1787 this_rq()->idle_balance = 1;
1788 raise_softirq_irqoff(SCHED_SOFTIRQ);
1793 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1795 struct rq *rq = cpu_rq(cpu);
1797 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1799 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1800 if (!set_nr_if_polling(rq->idle))
1801 smp_send_reschedule(cpu);
1803 trace_sched_wake_idle_without_ipi(cpu);
1807 void wake_up_if_idle(int cpu)
1809 struct rq *rq = cpu_rq(cpu);
1814 if (!is_idle_task(rcu_dereference(rq->curr)))
1817 if (set_nr_if_polling(rq->idle)) {
1818 trace_sched_wake_idle_without_ipi(cpu);
1820 rq_lock_irqsave(rq, &rf);
1821 if (is_idle_task(rq->curr))
1822 smp_send_reschedule(cpu);
1823 /* Else CPU is not idle, do nothing here: */
1824 rq_unlock_irqrestore(rq, &rf);
1831 bool cpus_share_cache(int this_cpu, int that_cpu)
1833 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1835 #endif /* CONFIG_SMP */
1837 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1839 struct rq *rq = cpu_rq(cpu);
1842 #if defined(CONFIG_SMP)
1843 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1844 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1845 ttwu_queue_remote(p, cpu, wake_flags);
1851 update_rq_clock(rq);
1852 ttwu_do_activate(rq, p, wake_flags, &rf);
1857 * Notes on Program-Order guarantees on SMP systems.
1861 * The basic program-order guarantee on SMP systems is that when a task [t]
1862 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1863 * execution on its new CPU [c1].
1865 * For migration (of runnable tasks) this is provided by the following means:
1867 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1868 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1869 * rq(c1)->lock (if not at the same time, then in that order).
1870 * C) LOCK of the rq(c1)->lock scheduling in task
1872 * Transitivity guarantees that B happens after A and C after B.
1873 * Note: we only require RCpc transitivity.
1874 * Note: the CPU doing B need not be c0 or c1
1883 * UNLOCK rq(0)->lock
1885 * LOCK rq(0)->lock // orders against CPU0
1887 * UNLOCK rq(0)->lock
1891 * UNLOCK rq(1)->lock
1893 * LOCK rq(1)->lock // orders against CPU2
1896 * UNLOCK rq(1)->lock
1899 * BLOCKING -- aka. SLEEP + WAKEUP
1901 * For blocking we (obviously) need to provide the same guarantee as for
1902 * migration. However the means are completely different as there is no lock
1903 * chain to provide order. Instead we do:
1905 * 1) smp_store_release(X->on_cpu, 0)
1906 * 2) smp_cond_load_acquire(!X->on_cpu)
1910 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1912 * LOCK rq(0)->lock LOCK X->pi_lock
1915 * smp_store_release(X->on_cpu, 0);
1917 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1923 * X->state = RUNNING
1924 * UNLOCK rq(2)->lock
1926 * LOCK rq(2)->lock // orders against CPU1
1929 * UNLOCK rq(2)->lock
1932 * UNLOCK rq(0)->lock
1935 * However; for wakeups there is a second guarantee we must provide, namely we
1936 * must observe the state that lead to our wakeup. That is, not only must our
1937 * task observe its own prior state, it must also observe the stores prior to
1940 * This means that any means of doing remote wakeups must order the CPU doing
1941 * the wakeup against the CPU the task is going to end up running on. This,
1942 * however, is already required for the regular Program-Order guarantee above,
1943 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1948 * try_to_wake_up - wake up a thread
1949 * @p: the thread to be awakened
1950 * @state: the mask of task states that can be woken
1951 * @wake_flags: wake modifier flags (WF_*)
1953 * If (@state & @p->state) @p->state = TASK_RUNNING.
1955 * If the task was not queued/runnable, also place it back on a runqueue.
1957 * Atomic against schedule() which would dequeue a task, also see
1958 * set_current_state().
1960 * Return: %true if @p->state changes (an actual wakeup was done),
1964 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1966 unsigned long flags;
1967 int cpu, success = 0;
1970 * If we are going to wake up a thread waiting for CONDITION we
1971 * need to ensure that CONDITION=1 done by the caller can not be
1972 * reordered with p->state check below. This pairs with mb() in
1973 * set_current_state() the waiting thread does.
1975 smp_mb__before_spinlock();
1976 raw_spin_lock_irqsave(&p->pi_lock, flags);
1977 if (!(p->state & state))
1980 trace_sched_waking(p);
1982 /* We're going to change ->state: */
1987 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1988 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1989 * in smp_cond_load_acquire() below.
1991 * sched_ttwu_pending() try_to_wake_up()
1992 * [S] p->on_rq = 1; [L] P->state
1993 * UNLOCK rq->lock -----.
1997 * LOCK rq->lock -----'
2001 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2003 * Pairs with the UNLOCK+LOCK on rq->lock from the
2004 * last wakeup of our task and the schedule that got our task
2008 if (p->on_rq && ttwu_remote(p, wake_flags))
2013 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2014 * possible to, falsely, observe p->on_cpu == 0.
2016 * One must be running (->on_cpu == 1) in order to remove oneself
2017 * from the runqueue.
2019 * [S] ->on_cpu = 1; [L] ->on_rq
2023 * [S] ->on_rq = 0; [L] ->on_cpu
2025 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2026 * from the consecutive calls to schedule(); the first switching to our
2027 * task, the second putting it to sleep.
2032 * If the owning (remote) CPU is still in the middle of schedule() with
2033 * this task as prev, wait until its done referencing the task.
2035 * Pairs with the smp_store_release() in finish_lock_switch().
2037 * This ensures that tasks getting woken will be fully ordered against
2038 * their previous state and preserve Program Order.
2040 smp_cond_load_acquire(&p->on_cpu, !VAL);
2042 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2043 p->state = TASK_WAKING;
2046 delayacct_blkio_end();
2047 atomic_dec(&task_rq(p)->nr_iowait);
2050 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2051 if (task_cpu(p) != cpu) {
2052 wake_flags |= WF_MIGRATED;
2053 set_task_cpu(p, cpu);
2056 #else /* CONFIG_SMP */
2059 delayacct_blkio_end();
2060 atomic_dec(&task_rq(p)->nr_iowait);
2063 #endif /* CONFIG_SMP */
2065 ttwu_queue(p, cpu, wake_flags);
2067 ttwu_stat(p, cpu, wake_flags);
2069 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2075 * try_to_wake_up_local - try to wake up a local task with rq lock held
2076 * @p: the thread to be awakened
2077 * @rf: request-queue flags for pinning
2079 * Put @p on the run-queue if it's not already there. The caller must
2080 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2083 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2085 struct rq *rq = task_rq(p);
2087 if (WARN_ON_ONCE(rq != this_rq()) ||
2088 WARN_ON_ONCE(p == current))
2091 lockdep_assert_held(&rq->lock);
2093 if (!raw_spin_trylock(&p->pi_lock)) {
2095 * This is OK, because current is on_cpu, which avoids it being
2096 * picked for load-balance and preemption/IRQs are still
2097 * disabled avoiding further scheduler activity on it and we've
2098 * not yet picked a replacement task.
2101 raw_spin_lock(&p->pi_lock);
2105 if (!(p->state & TASK_NORMAL))
2108 trace_sched_waking(p);
2110 if (!task_on_rq_queued(p)) {
2112 delayacct_blkio_end();
2113 atomic_dec(&rq->nr_iowait);
2115 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK);
2118 ttwu_do_wakeup(rq, p, 0, rf);
2119 ttwu_stat(p, smp_processor_id(), 0);
2121 raw_spin_unlock(&p->pi_lock);
2125 * wake_up_process - Wake up a specific process
2126 * @p: The process to be woken up.
2128 * Attempt to wake up the nominated process and move it to the set of runnable
2131 * Return: 1 if the process was woken up, 0 if it was already running.
2133 * It may be assumed that this function implies a write memory barrier before
2134 * changing the task state if and only if any tasks are woken up.
2136 int wake_up_process(struct task_struct *p)
2138 return try_to_wake_up(p, TASK_NORMAL, 0);
2140 EXPORT_SYMBOL(wake_up_process);
2142 int wake_up_state(struct task_struct *p, unsigned int state)
2144 return try_to_wake_up(p, state, 0);
2148 * Perform scheduler related setup for a newly forked process p.
2149 * p is forked by current.
2151 * __sched_fork() is basic setup used by init_idle() too:
2153 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2158 p->se.exec_start = 0;
2159 p->se.sum_exec_runtime = 0;
2160 p->se.prev_sum_exec_runtime = 0;
2161 p->se.nr_migrations = 0;
2163 INIT_LIST_HEAD(&p->se.group_node);
2165 #ifdef CONFIG_FAIR_GROUP_SCHED
2166 p->se.cfs_rq = NULL;
2169 #ifdef CONFIG_SCHEDSTATS
2170 /* Even if schedstat is disabled, there should not be garbage */
2171 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2174 RB_CLEAR_NODE(&p->dl.rb_node);
2175 init_dl_task_timer(&p->dl);
2176 init_dl_inactive_task_timer(&p->dl);
2177 __dl_clear_params(p);
2179 INIT_LIST_HEAD(&p->rt.run_list);
2181 p->rt.time_slice = sched_rr_timeslice;
2185 #ifdef CONFIG_PREEMPT_NOTIFIERS
2186 INIT_HLIST_HEAD(&p->preempt_notifiers);
2189 #ifdef CONFIG_NUMA_BALANCING
2190 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2191 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2192 p->mm->numa_scan_seq = 0;
2195 if (clone_flags & CLONE_VM)
2196 p->numa_preferred_nid = current->numa_preferred_nid;
2198 p->numa_preferred_nid = -1;
2200 p->node_stamp = 0ULL;
2201 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2202 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2203 p->numa_work.next = &p->numa_work;
2204 p->numa_faults = NULL;
2205 p->last_task_numa_placement = 0;
2206 p->last_sum_exec_runtime = 0;
2208 p->numa_group = NULL;
2209 #endif /* CONFIG_NUMA_BALANCING */
2212 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2214 #ifdef CONFIG_NUMA_BALANCING
2216 void set_numabalancing_state(bool enabled)
2219 static_branch_enable(&sched_numa_balancing);
2221 static_branch_disable(&sched_numa_balancing);
2224 #ifdef CONFIG_PROC_SYSCTL
2225 int sysctl_numa_balancing(struct ctl_table *table, int write,
2226 void __user *buffer, size_t *lenp, loff_t *ppos)
2230 int state = static_branch_likely(&sched_numa_balancing);
2232 if (write && !capable(CAP_SYS_ADMIN))
2237 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2241 set_numabalancing_state(state);
2247 #ifdef CONFIG_SCHEDSTATS
2249 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2250 static bool __initdata __sched_schedstats = false;
2252 static void set_schedstats(bool enabled)
2255 static_branch_enable(&sched_schedstats);
2257 static_branch_disable(&sched_schedstats);
2260 void force_schedstat_enabled(void)
2262 if (!schedstat_enabled()) {
2263 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2264 static_branch_enable(&sched_schedstats);
2268 static int __init setup_schedstats(char *str)
2275 * This code is called before jump labels have been set up, so we can't
2276 * change the static branch directly just yet. Instead set a temporary
2277 * variable so init_schedstats() can do it later.
2279 if (!strcmp(str, "enable")) {
2280 __sched_schedstats = true;
2282 } else if (!strcmp(str, "disable")) {
2283 __sched_schedstats = false;
2288 pr_warn("Unable to parse schedstats=\n");
2292 __setup("schedstats=", setup_schedstats);
2294 static void __init init_schedstats(void)
2296 set_schedstats(__sched_schedstats);
2299 #ifdef CONFIG_PROC_SYSCTL
2300 int sysctl_schedstats(struct ctl_table *table, int write,
2301 void __user *buffer, size_t *lenp, loff_t *ppos)
2305 int state = static_branch_likely(&sched_schedstats);
2307 if (write && !capable(CAP_SYS_ADMIN))
2312 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2316 set_schedstats(state);
2319 #endif /* CONFIG_PROC_SYSCTL */
2320 #else /* !CONFIG_SCHEDSTATS */
2321 static inline void init_schedstats(void) {}
2322 #endif /* CONFIG_SCHEDSTATS */
2325 * fork()/clone()-time setup:
2327 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2329 unsigned long flags;
2330 int cpu = get_cpu();
2332 __sched_fork(clone_flags, p);
2334 * We mark the process as NEW here. This guarantees that
2335 * nobody will actually run it, and a signal or other external
2336 * event cannot wake it up and insert it on the runqueue either.
2338 p->state = TASK_NEW;
2341 * Make sure we do not leak PI boosting priority to the child.
2343 p->prio = current->normal_prio;
2346 * Revert to default priority/policy on fork if requested.
2348 if (unlikely(p->sched_reset_on_fork)) {
2349 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2350 p->policy = SCHED_NORMAL;
2351 p->static_prio = NICE_TO_PRIO(0);
2353 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2354 p->static_prio = NICE_TO_PRIO(0);
2356 p->prio = p->normal_prio = __normal_prio(p);
2360 * We don't need the reset flag anymore after the fork. It has
2361 * fulfilled its duty:
2363 p->sched_reset_on_fork = 0;
2366 if (dl_prio(p->prio)) {
2369 } else if (rt_prio(p->prio)) {
2370 p->sched_class = &rt_sched_class;
2372 p->sched_class = &fair_sched_class;
2375 init_entity_runnable_average(&p->se);
2378 * The child is not yet in the pid-hash so no cgroup attach races,
2379 * and the cgroup is pinned to this child due to cgroup_fork()
2380 * is ran before sched_fork().
2382 * Silence PROVE_RCU.
2384 raw_spin_lock_irqsave(&p->pi_lock, flags);
2386 * We're setting the CPU for the first time, we don't migrate,
2387 * so use __set_task_cpu().
2389 __set_task_cpu(p, cpu);
2390 if (p->sched_class->task_fork)
2391 p->sched_class->task_fork(p);
2392 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2394 #ifdef CONFIG_SCHED_INFO
2395 if (likely(sched_info_on()))
2396 memset(&p->sched_info, 0, sizeof(p->sched_info));
2398 #if defined(CONFIG_SMP)
2401 init_task_preempt_count(p);
2403 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2404 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2411 unsigned long to_ratio(u64 period, u64 runtime)
2413 if (runtime == RUNTIME_INF)
2417 * Doing this here saves a lot of checks in all
2418 * the calling paths, and returning zero seems
2419 * safe for them anyway.
2424 return div64_u64(runtime << BW_SHIFT, period);
2428 * wake_up_new_task - wake up a newly created task for the first time.
2430 * This function will do some initial scheduler statistics housekeeping
2431 * that must be done for every newly created context, then puts the task
2432 * on the runqueue and wakes it.
2434 void wake_up_new_task(struct task_struct *p)
2439 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2440 p->state = TASK_RUNNING;
2443 * Fork balancing, do it here and not earlier because:
2444 * - cpus_allowed can change in the fork path
2445 * - any previously selected CPU might disappear through hotplug
2447 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2448 * as we're not fully set-up yet.
2450 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2452 rq = __task_rq_lock(p, &rf);
2453 update_rq_clock(rq);
2454 post_init_entity_util_avg(&p->se);
2456 activate_task(rq, p, ENQUEUE_NOCLOCK);
2457 p->on_rq = TASK_ON_RQ_QUEUED;
2458 trace_sched_wakeup_new(p);
2459 check_preempt_curr(rq, p, WF_FORK);
2461 if (p->sched_class->task_woken) {
2463 * Nothing relies on rq->lock after this, so its fine to
2466 rq_unpin_lock(rq, &rf);
2467 p->sched_class->task_woken(rq, p);
2468 rq_repin_lock(rq, &rf);
2471 task_rq_unlock(rq, p, &rf);
2474 #ifdef CONFIG_PREEMPT_NOTIFIERS
2476 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2478 void preempt_notifier_inc(void)
2480 static_key_slow_inc(&preempt_notifier_key);
2482 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2484 void preempt_notifier_dec(void)
2486 static_key_slow_dec(&preempt_notifier_key);
2488 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2491 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2492 * @notifier: notifier struct to register
2494 void preempt_notifier_register(struct preempt_notifier *notifier)
2496 if (!static_key_false(&preempt_notifier_key))
2497 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2499 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2501 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2504 * preempt_notifier_unregister - no longer interested in preemption notifications
2505 * @notifier: notifier struct to unregister
2507 * This is *not* safe to call from within a preemption notifier.
2509 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2511 hlist_del(¬ifier->link);
2513 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2515 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2517 struct preempt_notifier *notifier;
2519 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2520 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2523 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2525 if (static_key_false(&preempt_notifier_key))
2526 __fire_sched_in_preempt_notifiers(curr);
2530 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2531 struct task_struct *next)
2533 struct preempt_notifier *notifier;
2535 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2536 notifier->ops->sched_out(notifier, next);
2539 static __always_inline void
2540 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2541 struct task_struct *next)
2543 if (static_key_false(&preempt_notifier_key))
2544 __fire_sched_out_preempt_notifiers(curr, next);
2547 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2549 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2554 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2555 struct task_struct *next)
2559 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2562 * prepare_task_switch - prepare to switch tasks
2563 * @rq: the runqueue preparing to switch
2564 * @prev: the current task that is being switched out
2565 * @next: the task we are going to switch to.
2567 * This is called with the rq lock held and interrupts off. It must
2568 * be paired with a subsequent finish_task_switch after the context
2571 * prepare_task_switch sets up locking and calls architecture specific
2575 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2576 struct task_struct *next)
2578 sched_info_switch(rq, prev, next);
2579 perf_event_task_sched_out(prev, next);
2580 fire_sched_out_preempt_notifiers(prev, next);
2581 prepare_lock_switch(rq, next);
2582 prepare_arch_switch(next);
2586 * finish_task_switch - clean up after a task-switch
2587 * @prev: the thread we just switched away from.
2589 * finish_task_switch must be called after the context switch, paired
2590 * with a prepare_task_switch call before the context switch.
2591 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2592 * and do any other architecture-specific cleanup actions.
2594 * Note that we may have delayed dropping an mm in context_switch(). If
2595 * so, we finish that here outside of the runqueue lock. (Doing it
2596 * with the lock held can cause deadlocks; see schedule() for
2599 * The context switch have flipped the stack from under us and restored the
2600 * local variables which were saved when this task called schedule() in the
2601 * past. prev == current is still correct but we need to recalculate this_rq
2602 * because prev may have moved to another CPU.
2604 static struct rq *finish_task_switch(struct task_struct *prev)
2605 __releases(rq->lock)
2607 struct rq *rq = this_rq();
2608 struct mm_struct *mm = rq->prev_mm;
2612 * The previous task will have left us with a preempt_count of 2
2613 * because it left us after:
2616 * preempt_disable(); // 1
2618 * raw_spin_lock_irq(&rq->lock) // 2
2620 * Also, see FORK_PREEMPT_COUNT.
2622 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2623 "corrupted preempt_count: %s/%d/0x%x\n",
2624 current->comm, current->pid, preempt_count()))
2625 preempt_count_set(FORK_PREEMPT_COUNT);
2630 * A task struct has one reference for the use as "current".
2631 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2632 * schedule one last time. The schedule call will never return, and
2633 * the scheduled task must drop that reference.
2635 * We must observe prev->state before clearing prev->on_cpu (in
2636 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2637 * running on another CPU and we could rave with its RUNNING -> DEAD
2638 * transition, resulting in a double drop.
2640 prev_state = prev->state;
2641 vtime_task_switch(prev);
2642 perf_event_task_sched_in(prev, current);
2644 * The membarrier system call requires a full memory barrier
2645 * after storing to rq->curr, before going back to user-space.
2647 * TODO: This smp_mb__after_unlock_lock can go away if PPC end
2648 * up adding a full barrier to switch_mm(), or we should figure
2649 * out if a smp_mb__after_unlock_lock is really the proper API
2652 smp_mb__after_unlock_lock();
2653 finish_lock_switch(rq, prev);
2654 finish_arch_post_lock_switch();
2656 fire_sched_in_preempt_notifiers(current);
2659 if (unlikely(prev_state == TASK_DEAD)) {
2660 if (prev->sched_class->task_dead)
2661 prev->sched_class->task_dead(prev);
2664 * Remove function-return probe instances associated with this
2665 * task and put them back on the free list.
2667 kprobe_flush_task(prev);
2669 /* Task is done with its stack. */
2670 put_task_stack(prev);
2672 put_task_struct(prev);
2675 tick_nohz_task_switch();
2681 /* rq->lock is NOT held, but preemption is disabled */
2682 static void __balance_callback(struct rq *rq)
2684 struct callback_head *head, *next;
2685 void (*func)(struct rq *rq);
2686 unsigned long flags;
2688 raw_spin_lock_irqsave(&rq->lock, flags);
2689 head = rq->balance_callback;
2690 rq->balance_callback = NULL;
2692 func = (void (*)(struct rq *))head->func;
2699 raw_spin_unlock_irqrestore(&rq->lock, flags);
2702 static inline void balance_callback(struct rq *rq)
2704 if (unlikely(rq->balance_callback))
2705 __balance_callback(rq);
2710 static inline void balance_callback(struct rq *rq)
2717 * schedule_tail - first thing a freshly forked thread must call.
2718 * @prev: the thread we just switched away from.
2720 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2721 __releases(rq->lock)
2726 * New tasks start with FORK_PREEMPT_COUNT, see there and
2727 * finish_task_switch() for details.
2729 * finish_task_switch() will drop rq->lock() and lower preempt_count
2730 * and the preempt_enable() will end up enabling preemption (on
2731 * PREEMPT_COUNT kernels).
2734 rq = finish_task_switch(prev);
2735 balance_callback(rq);
2738 if (current->set_child_tid)
2739 put_user(task_pid_vnr(current), current->set_child_tid);
2743 * context_switch - switch to the new MM and the new thread's register state.
2745 static __always_inline struct rq *
2746 context_switch(struct rq *rq, struct task_struct *prev,
2747 struct task_struct *next, struct rq_flags *rf)
2749 struct mm_struct *mm, *oldmm;
2751 prepare_task_switch(rq, prev, next);
2754 oldmm = prev->active_mm;
2756 * For paravirt, this is coupled with an exit in switch_to to
2757 * combine the page table reload and the switch backend into
2760 arch_start_context_switch(prev);
2763 next->active_mm = oldmm;
2765 enter_lazy_tlb(oldmm, next);
2767 switch_mm_irqs_off(oldmm, mm, next);
2770 prev->active_mm = NULL;
2771 rq->prev_mm = oldmm;
2774 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2777 * Since the runqueue lock will be released by the next
2778 * task (which is an invalid locking op but in the case
2779 * of the scheduler it's an obvious special-case), so we
2780 * do an early lockdep release here:
2782 rq_unpin_lock(rq, rf);
2783 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2785 /* Here we just switch the register state and the stack. */
2786 switch_to(prev, next, prev);
2789 return finish_task_switch(prev);
2793 * nr_running and nr_context_switches:
2795 * externally visible scheduler statistics: current number of runnable
2796 * threads, total number of context switches performed since bootup.
2798 unsigned long nr_running(void)
2800 unsigned long i, sum = 0;
2802 for_each_online_cpu(i)
2803 sum += cpu_rq(i)->nr_running;
2809 * Check if only the current task is running on the CPU.
2811 * Caution: this function does not check that the caller has disabled
2812 * preemption, thus the result might have a time-of-check-to-time-of-use
2813 * race. The caller is responsible to use it correctly, for example:
2815 * - from a non-preemptable section (of course)
2817 * - from a thread that is bound to a single CPU
2819 * - in a loop with very short iterations (e.g. a polling loop)
2821 bool single_task_running(void)
2823 return raw_rq()->nr_running == 1;
2825 EXPORT_SYMBOL(single_task_running);
2827 unsigned long long nr_context_switches(void)
2830 unsigned long long sum = 0;
2832 for_each_possible_cpu(i)
2833 sum += cpu_rq(i)->nr_switches;
2839 * IO-wait accounting, and how its mostly bollocks (on SMP).
2841 * The idea behind IO-wait account is to account the idle time that we could
2842 * have spend running if it were not for IO. That is, if we were to improve the
2843 * storage performance, we'd have a proportional reduction in IO-wait time.
2845 * This all works nicely on UP, where, when a task blocks on IO, we account
2846 * idle time as IO-wait, because if the storage were faster, it could've been
2847 * running and we'd not be idle.
2849 * This has been extended to SMP, by doing the same for each CPU. This however
2852 * Imagine for instance the case where two tasks block on one CPU, only the one
2853 * CPU will have IO-wait accounted, while the other has regular idle. Even
2854 * though, if the storage were faster, both could've ran at the same time,
2855 * utilising both CPUs.
2857 * This means, that when looking globally, the current IO-wait accounting on
2858 * SMP is a lower bound, by reason of under accounting.
2860 * Worse, since the numbers are provided per CPU, they are sometimes
2861 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2862 * associated with any one particular CPU, it can wake to another CPU than it
2863 * blocked on. This means the per CPU IO-wait number is meaningless.
2865 * Task CPU affinities can make all that even more 'interesting'.
2868 unsigned long nr_iowait(void)
2870 unsigned long i, sum = 0;
2872 for_each_possible_cpu(i)
2873 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2879 * Consumers of these two interfaces, like for example the cpufreq menu
2880 * governor are using nonsensical data. Boosting frequency for a CPU that has
2881 * IO-wait which might not even end up running the task when it does become
2885 unsigned long nr_iowait_cpu(int cpu)
2887 struct rq *this = cpu_rq(cpu);
2888 return atomic_read(&this->nr_iowait);
2891 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2893 struct rq *rq = this_rq();
2894 *nr_waiters = atomic_read(&rq->nr_iowait);
2895 *load = rq->load.weight;
2901 * sched_exec - execve() is a valuable balancing opportunity, because at
2902 * this point the task has the smallest effective memory and cache footprint.
2904 void sched_exec(void)
2906 struct task_struct *p = current;
2907 unsigned long flags;
2910 raw_spin_lock_irqsave(&p->pi_lock, flags);
2911 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2912 if (dest_cpu == smp_processor_id())
2915 if (likely(cpu_active(dest_cpu))) {
2916 struct migration_arg arg = { p, dest_cpu };
2918 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2919 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2923 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2928 DEFINE_PER_CPU(struct kernel_stat, kstat);
2929 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2931 EXPORT_PER_CPU_SYMBOL(kstat);
2932 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2935 * The function fair_sched_class.update_curr accesses the struct curr
2936 * and its field curr->exec_start; when called from task_sched_runtime(),
2937 * we observe a high rate of cache misses in practice.
2938 * Prefetching this data results in improved performance.
2940 static inline void prefetch_curr_exec_start(struct task_struct *p)
2942 #ifdef CONFIG_FAIR_GROUP_SCHED
2943 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
2945 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
2948 prefetch(&curr->exec_start);
2952 * Return accounted runtime for the task.
2953 * In case the task is currently running, return the runtime plus current's
2954 * pending runtime that have not been accounted yet.
2956 unsigned long long task_sched_runtime(struct task_struct *p)
2962 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2964 * 64-bit doesn't need locks to atomically read a 64bit value.
2965 * So we have a optimization chance when the task's delta_exec is 0.
2966 * Reading ->on_cpu is racy, but this is ok.
2968 * If we race with it leaving CPU, we'll take a lock. So we're correct.
2969 * If we race with it entering CPU, unaccounted time is 0. This is
2970 * indistinguishable from the read occurring a few cycles earlier.
2971 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2972 * been accounted, so we're correct here as well.
2974 if (!p->on_cpu || !task_on_rq_queued(p))
2975 return p->se.sum_exec_runtime;
2978 rq = task_rq_lock(p, &rf);
2980 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2981 * project cycles that may never be accounted to this
2982 * thread, breaking clock_gettime().
2984 if (task_current(rq, p) && task_on_rq_queued(p)) {
2985 prefetch_curr_exec_start(p);
2986 update_rq_clock(rq);
2987 p->sched_class->update_curr(rq);
2989 ns = p->se.sum_exec_runtime;
2990 task_rq_unlock(rq, p, &rf);
2996 * This function gets called by the timer code, with HZ frequency.
2997 * We call it with interrupts disabled.
2999 void scheduler_tick(void)
3001 int cpu = smp_processor_id();
3002 struct rq *rq = cpu_rq(cpu);
3003 struct task_struct *curr = rq->curr;
3010 update_rq_clock(rq);
3011 curr->sched_class->task_tick(rq, curr, 0);
3012 cpu_load_update_active(rq);
3013 calc_global_load_tick(rq);
3017 perf_event_task_tick();
3020 rq->idle_balance = idle_cpu(cpu);
3021 trigger_load_balance(rq);
3023 rq_last_tick_reset(rq);
3026 #ifdef CONFIG_NO_HZ_FULL
3028 * scheduler_tick_max_deferment
3030 * Keep at least one tick per second when a single
3031 * active task is running because the scheduler doesn't
3032 * yet completely support full dynticks environment.
3034 * This makes sure that uptime, CFS vruntime, load
3035 * balancing, etc... continue to move forward, even
3036 * with a very low granularity.
3038 * Return: Maximum deferment in nanoseconds.
3040 u64 scheduler_tick_max_deferment(void)
3042 struct rq *rq = this_rq();
3043 unsigned long next, now = READ_ONCE(jiffies);
3045 next = rq->last_sched_tick + HZ;
3047 if (time_before_eq(next, now))
3050 return jiffies_to_nsecs(next - now);
3054 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3055 defined(CONFIG_PREEMPT_TRACER))
3057 * If the value passed in is equal to the current preempt count
3058 * then we just disabled preemption. Start timing the latency.
3060 static inline void preempt_latency_start(int val)
3062 if (preempt_count() == val) {
3063 unsigned long ip = get_lock_parent_ip();
3064 #ifdef CONFIG_DEBUG_PREEMPT
3065 current->preempt_disable_ip = ip;
3067 trace_preempt_off(CALLER_ADDR0, ip);
3071 void preempt_count_add(int val)
3073 #ifdef CONFIG_DEBUG_PREEMPT
3077 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3080 __preempt_count_add(val);
3081 #ifdef CONFIG_DEBUG_PREEMPT
3083 * Spinlock count overflowing soon?
3085 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3088 preempt_latency_start(val);
3090 EXPORT_SYMBOL(preempt_count_add);
3091 NOKPROBE_SYMBOL(preempt_count_add);
3094 * If the value passed in equals to the current preempt count
3095 * then we just enabled preemption. Stop timing the latency.
3097 static inline void preempt_latency_stop(int val)
3099 if (preempt_count() == val)
3100 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3103 void preempt_count_sub(int val)
3105 #ifdef CONFIG_DEBUG_PREEMPT
3109 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3112 * Is the spinlock portion underflowing?
3114 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3115 !(preempt_count() & PREEMPT_MASK)))
3119 preempt_latency_stop(val);
3120 __preempt_count_sub(val);
3122 EXPORT_SYMBOL(preempt_count_sub);
3123 NOKPROBE_SYMBOL(preempt_count_sub);
3126 static inline void preempt_latency_start(int val) { }
3127 static inline void preempt_latency_stop(int val) { }
3130 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3132 #ifdef CONFIG_DEBUG_PREEMPT
3133 return p->preempt_disable_ip;
3140 * Print scheduling while atomic bug:
3142 static noinline void __schedule_bug(struct task_struct *prev)
3144 /* Save this before calling printk(), since that will clobber it */
3145 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3147 if (oops_in_progress)
3150 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3151 prev->comm, prev->pid, preempt_count());
3153 debug_show_held_locks(prev);
3155 if (irqs_disabled())
3156 print_irqtrace_events(prev);
3157 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3158 && in_atomic_preempt_off()) {
3159 pr_err("Preemption disabled at:");
3160 print_ip_sym(preempt_disable_ip);
3164 panic("scheduling while atomic\n");
3167 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3171 * Various schedule()-time debugging checks and statistics:
3173 static inline void schedule_debug(struct task_struct *prev)
3175 #ifdef CONFIG_SCHED_STACK_END_CHECK
3176 if (task_stack_end_corrupted(prev))
3177 panic("corrupted stack end detected inside scheduler\n");
3180 if (unlikely(in_atomic_preempt_off())) {
3181 __schedule_bug(prev);
3182 preempt_count_set(PREEMPT_DISABLED);
3186 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3188 schedstat_inc(this_rq()->sched_count);
3192 * Pick up the highest-prio task:
3194 static inline struct task_struct *
3195 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3197 const struct sched_class *class;
3198 struct task_struct *p;
3201 * Optimization: we know that if all tasks are in the fair class we can
3202 * call that function directly, but only if the @prev task wasn't of a
3203 * higher scheduling class, because otherwise those loose the
3204 * opportunity to pull in more work from other CPUs.
3206 if (likely((prev->sched_class == &idle_sched_class ||
3207 prev->sched_class == &fair_sched_class) &&
3208 rq->nr_running == rq->cfs.h_nr_running)) {
3210 p = fair_sched_class.pick_next_task(rq, prev, rf);
3211 if (unlikely(p == RETRY_TASK))
3214 /* Assumes fair_sched_class->next == idle_sched_class */
3216 p = idle_sched_class.pick_next_task(rq, prev, rf);
3222 for_each_class(class) {
3223 p = class->pick_next_task(rq, prev, rf);
3225 if (unlikely(p == RETRY_TASK))
3231 /* The idle class should always have a runnable task: */
3236 * __schedule() is the main scheduler function.
3238 * The main means of driving the scheduler and thus entering this function are:
3240 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3242 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3243 * paths. For example, see arch/x86/entry_64.S.
3245 * To drive preemption between tasks, the scheduler sets the flag in timer
3246 * interrupt handler scheduler_tick().
3248 * 3. Wakeups don't really cause entry into schedule(). They add a
3249 * task to the run-queue and that's it.
3251 * Now, if the new task added to the run-queue preempts the current
3252 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3253 * called on the nearest possible occasion:
3255 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3257 * - in syscall or exception context, at the next outmost
3258 * preempt_enable(). (this might be as soon as the wake_up()'s
3261 * - in IRQ context, return from interrupt-handler to
3262 * preemptible context
3264 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3267 * - cond_resched() call
3268 * - explicit schedule() call
3269 * - return from syscall or exception to user-space
3270 * - return from interrupt-handler to user-space
3272 * WARNING: must be called with preemption disabled!
3274 static void __sched notrace __schedule(bool preempt)
3276 struct task_struct *prev, *next;
3277 unsigned long *switch_count;
3282 cpu = smp_processor_id();
3286 schedule_debug(prev);
3288 if (sched_feat(HRTICK))
3291 local_irq_disable();
3292 rcu_note_context_switch(preempt);
3295 * Make sure that signal_pending_state()->signal_pending() below
3296 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3297 * done by the caller to avoid the race with signal_wake_up().
3299 smp_mb__before_spinlock();
3302 /* Promote REQ to ACT */
3303 rq->clock_update_flags <<= 1;
3304 update_rq_clock(rq);
3306 switch_count = &prev->nivcsw;
3307 if (!preempt && prev->state) {
3308 if (unlikely(signal_pending_state(prev->state, prev))) {
3309 prev->state = TASK_RUNNING;
3311 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3314 if (prev->in_iowait) {
3315 atomic_inc(&rq->nr_iowait);
3316 delayacct_blkio_start();
3320 * If a worker went to sleep, notify and ask workqueue
3321 * whether it wants to wake up a task to maintain
3324 if (prev->flags & PF_WQ_WORKER) {
3325 struct task_struct *to_wakeup;
3327 to_wakeup = wq_worker_sleeping(prev);
3329 try_to_wake_up_local(to_wakeup, &rf);
3332 switch_count = &prev->nvcsw;
3335 next = pick_next_task(rq, prev, &rf);
3336 clear_tsk_need_resched(prev);
3337 clear_preempt_need_resched();
3339 if (likely(prev != next)) {
3343 * The membarrier system call requires each architecture
3344 * to have a full memory barrier after updating
3345 * rq->curr, before returning to user-space. For TSO
3346 * (e.g. x86), the architecture must provide its own
3347 * barrier in switch_mm(). For weakly ordered machines
3348 * for which spin_unlock() acts as a full memory
3349 * barrier, finish_lock_switch() in common code takes
3350 * care of this barrier. For weakly ordered machines for
3351 * which spin_unlock() acts as a RELEASE barrier (only
3352 * arm64 and PowerPC), arm64 has a full barrier in
3353 * switch_to(), and PowerPC has
3354 * smp_mb__after_unlock_lock() before
3355 * finish_lock_switch().
3359 trace_sched_switch(preempt, prev, next);
3361 /* Also unlocks the rq: */
3362 rq = context_switch(rq, prev, next, &rf);
3364 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3365 rq_unlock_irq(rq, &rf);
3368 balance_callback(rq);
3371 void __noreturn do_task_dead(void)
3374 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3375 * when the following two conditions become true.
3376 * - There is race condition of mmap_sem (It is acquired by
3378 * - SMI occurs before setting TASK_RUNINNG.
3379 * (or hypervisor of virtual machine switches to other guest)
3380 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3382 * To avoid it, we have to wait for releasing tsk->pi_lock which
3383 * is held by try_to_wake_up()
3385 raw_spin_lock_irq(¤t->pi_lock);
3386 raw_spin_unlock_irq(¤t->pi_lock);
3388 /* Causes final put_task_struct in finish_task_switch(): */
3389 __set_current_state(TASK_DEAD);
3391 /* Tell freezer to ignore us: */
3392 current->flags |= PF_NOFREEZE;
3397 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3402 static inline void sched_submit_work(struct task_struct *tsk)
3404 if (!tsk->state || tsk_is_pi_blocked(tsk))
3407 * If we are going to sleep and we have plugged IO queued,
3408 * make sure to submit it to avoid deadlocks.
3410 if (blk_needs_flush_plug(tsk))
3411 blk_schedule_flush_plug(tsk);
3414 asmlinkage __visible void __sched schedule(void)
3416 struct task_struct *tsk = current;
3418 sched_submit_work(tsk);
3422 sched_preempt_enable_no_resched();
3423 } while (need_resched());
3425 EXPORT_SYMBOL(schedule);
3428 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3429 * state (have scheduled out non-voluntarily) by making sure that all
3430 * tasks have either left the run queue or have gone into user space.
3431 * As idle tasks do not do either, they must not ever be preempted
3432 * (schedule out non-voluntarily).
3434 * schedule_idle() is similar to schedule_preempt_disable() except that it
3435 * never enables preemption because it does not call sched_submit_work().
3437 void __sched schedule_idle(void)
3440 * As this skips calling sched_submit_work(), which the idle task does
3441 * regardless because that function is a nop when the task is in a
3442 * TASK_RUNNING state, make sure this isn't used someplace that the
3443 * current task can be in any other state. Note, idle is always in the
3444 * TASK_RUNNING state.
3446 WARN_ON_ONCE(current->state);
3449 } while (need_resched());
3452 #ifdef CONFIG_CONTEXT_TRACKING
3453 asmlinkage __visible void __sched schedule_user(void)
3456 * If we come here after a random call to set_need_resched(),
3457 * or we have been woken up remotely but the IPI has not yet arrived,
3458 * we haven't yet exited the RCU idle mode. Do it here manually until
3459 * we find a better solution.
3461 * NB: There are buggy callers of this function. Ideally we
3462 * should warn if prev_state != CONTEXT_USER, but that will trigger
3463 * too frequently to make sense yet.
3465 enum ctx_state prev_state = exception_enter();
3467 exception_exit(prev_state);
3472 * schedule_preempt_disabled - called with preemption disabled
3474 * Returns with preemption disabled. Note: preempt_count must be 1
3476 void __sched schedule_preempt_disabled(void)
3478 sched_preempt_enable_no_resched();
3483 static void __sched notrace preempt_schedule_common(void)
3487 * Because the function tracer can trace preempt_count_sub()
3488 * and it also uses preempt_enable/disable_notrace(), if
3489 * NEED_RESCHED is set, the preempt_enable_notrace() called
3490 * by the function tracer will call this function again and
3491 * cause infinite recursion.
3493 * Preemption must be disabled here before the function
3494 * tracer can trace. Break up preempt_disable() into two
3495 * calls. One to disable preemption without fear of being
3496 * traced. The other to still record the preemption latency,
3497 * which can also be traced by the function tracer.
3499 preempt_disable_notrace();
3500 preempt_latency_start(1);
3502 preempt_latency_stop(1);
3503 preempt_enable_no_resched_notrace();
3506 * Check again in case we missed a preemption opportunity
3507 * between schedule and now.
3509 } while (need_resched());
3512 #ifdef CONFIG_PREEMPT
3514 * this is the entry point to schedule() from in-kernel preemption
3515 * off of preempt_enable. Kernel preemptions off return from interrupt
3516 * occur there and call schedule directly.
3518 asmlinkage __visible void __sched notrace preempt_schedule(void)
3521 * If there is a non-zero preempt_count or interrupts are disabled,
3522 * we do not want to preempt the current task. Just return..
3524 if (likely(!preemptible()))
3527 preempt_schedule_common();
3529 NOKPROBE_SYMBOL(preempt_schedule);
3530 EXPORT_SYMBOL(preempt_schedule);
3533 * preempt_schedule_notrace - preempt_schedule called by tracing
3535 * The tracing infrastructure uses preempt_enable_notrace to prevent
3536 * recursion and tracing preempt enabling caused by the tracing
3537 * infrastructure itself. But as tracing can happen in areas coming
3538 * from userspace or just about to enter userspace, a preempt enable
3539 * can occur before user_exit() is called. This will cause the scheduler
3540 * to be called when the system is still in usermode.
3542 * To prevent this, the preempt_enable_notrace will use this function
3543 * instead of preempt_schedule() to exit user context if needed before
3544 * calling the scheduler.
3546 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3548 enum ctx_state prev_ctx;
3550 if (likely(!preemptible()))
3555 * Because the function tracer can trace preempt_count_sub()
3556 * and it also uses preempt_enable/disable_notrace(), if
3557 * NEED_RESCHED is set, the preempt_enable_notrace() called
3558 * by the function tracer will call this function again and
3559 * cause infinite recursion.
3561 * Preemption must be disabled here before the function
3562 * tracer can trace. Break up preempt_disable() into two
3563 * calls. One to disable preemption without fear of being
3564 * traced. The other to still record the preemption latency,
3565 * which can also be traced by the function tracer.
3567 preempt_disable_notrace();
3568 preempt_latency_start(1);
3570 * Needs preempt disabled in case user_exit() is traced
3571 * and the tracer calls preempt_enable_notrace() causing
3572 * an infinite recursion.
3574 prev_ctx = exception_enter();
3576 exception_exit(prev_ctx);
3578 preempt_latency_stop(1);
3579 preempt_enable_no_resched_notrace();
3580 } while (need_resched());
3582 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3584 #endif /* CONFIG_PREEMPT */
3587 * this is the entry point to schedule() from kernel preemption
3588 * off of irq context.
3589 * Note, that this is called and return with irqs disabled. This will
3590 * protect us against recursive calling from irq.
3592 asmlinkage __visible void __sched preempt_schedule_irq(void)
3594 enum ctx_state prev_state;
3596 /* Catch callers which need to be fixed */
3597 BUG_ON(preempt_count() || !irqs_disabled());
3599 prev_state = exception_enter();
3605 local_irq_disable();
3606 sched_preempt_enable_no_resched();
3607 } while (need_resched());
3609 exception_exit(prev_state);
3612 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
3615 return try_to_wake_up(curr->private, mode, wake_flags);
3617 EXPORT_SYMBOL(default_wake_function);
3619 #ifdef CONFIG_RT_MUTEXES
3621 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
3624 prio = min(prio, pi_task->prio);
3629 static inline int rt_effective_prio(struct task_struct *p, int prio)
3631 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3633 return __rt_effective_prio(pi_task, prio);
3637 * rt_mutex_setprio - set the current priority of a task
3639 * @pi_task: donor task
3641 * This function changes the 'effective' priority of a task. It does
3642 * not touch ->normal_prio like __setscheduler().
3644 * Used by the rt_mutex code to implement priority inheritance
3645 * logic. Call site only calls if the priority of the task changed.
3647 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
3649 int prio, oldprio, queued, running, queue_flag =
3650 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
3651 const struct sched_class *prev_class;
3655 /* XXX used to be waiter->prio, not waiter->task->prio */
3656 prio = __rt_effective_prio(pi_task, p->normal_prio);
3659 * If nothing changed; bail early.
3661 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
3664 rq = __task_rq_lock(p, &rf);
3665 update_rq_clock(rq);
3667 * Set under pi_lock && rq->lock, such that the value can be used under
3670 * Note that there is loads of tricky to make this pointer cache work
3671 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3672 * ensure a task is de-boosted (pi_task is set to NULL) before the
3673 * task is allowed to run again (and can exit). This ensures the pointer
3674 * points to a blocked task -- which guaratees the task is present.
3676 p->pi_top_task = pi_task;
3679 * For FIFO/RR we only need to set prio, if that matches we're done.
3681 if (prio == p->prio && !dl_prio(prio))
3685 * Idle task boosting is a nono in general. There is one
3686 * exception, when PREEMPT_RT and NOHZ is active:
3688 * The idle task calls get_next_timer_interrupt() and holds
3689 * the timer wheel base->lock on the CPU and another CPU wants
3690 * to access the timer (probably to cancel it). We can safely
3691 * ignore the boosting request, as the idle CPU runs this code
3692 * with interrupts disabled and will complete the lock
3693 * protected section without being interrupted. So there is no
3694 * real need to boost.
3696 if (unlikely(p == rq->idle)) {
3697 WARN_ON(p != rq->curr);
3698 WARN_ON(p->pi_blocked_on);
3702 trace_sched_pi_setprio(p, pi_task);
3705 if (oldprio == prio)
3706 queue_flag &= ~DEQUEUE_MOVE;
3708 prev_class = p->sched_class;
3709 queued = task_on_rq_queued(p);
3710 running = task_current(rq, p);
3712 dequeue_task(rq, p, queue_flag);
3714 put_prev_task(rq, p);
3717 * Boosting condition are:
3718 * 1. -rt task is running and holds mutex A
3719 * --> -dl task blocks on mutex A
3721 * 2. -dl task is running and holds mutex A
3722 * --> -dl task blocks on mutex A and could preempt the
3725 if (dl_prio(prio)) {
3726 if (!dl_prio(p->normal_prio) ||
3727 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3728 p->dl.dl_boosted = 1;
3729 queue_flag |= ENQUEUE_REPLENISH;
3731 p->dl.dl_boosted = 0;
3732 p->sched_class = &dl_sched_class;
3733 } else if (rt_prio(prio)) {
3734 if (dl_prio(oldprio))
3735 p->dl.dl_boosted = 0;
3737 queue_flag |= ENQUEUE_HEAD;
3738 p->sched_class = &rt_sched_class;
3740 if (dl_prio(oldprio))
3741 p->dl.dl_boosted = 0;
3742 if (rt_prio(oldprio))
3744 p->sched_class = &fair_sched_class;
3750 enqueue_task(rq, p, queue_flag);
3752 set_curr_task(rq, p);
3754 check_class_changed(rq, p, prev_class, oldprio);
3756 /* Avoid rq from going away on us: */
3758 __task_rq_unlock(rq, &rf);
3760 balance_callback(rq);
3764 static inline int rt_effective_prio(struct task_struct *p, int prio)
3770 void set_user_nice(struct task_struct *p, long nice)
3772 bool queued, running;
3773 int old_prio, delta;
3777 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3780 * We have to be careful, if called from sys_setpriority(),
3781 * the task might be in the middle of scheduling on another CPU.
3783 rq = task_rq_lock(p, &rf);
3784 update_rq_clock(rq);
3787 * The RT priorities are set via sched_setscheduler(), but we still
3788 * allow the 'normal' nice value to be set - but as expected
3789 * it wont have any effect on scheduling until the task is
3790 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3792 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3793 p->static_prio = NICE_TO_PRIO(nice);
3796 queued = task_on_rq_queued(p);
3797 running = task_current(rq, p);
3799 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
3801 put_prev_task(rq, p);
3803 p->static_prio = NICE_TO_PRIO(nice);
3806 p->prio = effective_prio(p);
3807 delta = p->prio - old_prio;
3810 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
3812 * If the task increased its priority or is running and
3813 * lowered its priority, then reschedule its CPU:
3815 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3819 set_curr_task(rq, p);
3821 task_rq_unlock(rq, p, &rf);
3823 EXPORT_SYMBOL(set_user_nice);
3826 * can_nice - check if a task can reduce its nice value
3830 int can_nice(const struct task_struct *p, const int nice)
3832 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3833 int nice_rlim = nice_to_rlimit(nice);
3835 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3836 capable(CAP_SYS_NICE));
3839 #ifdef __ARCH_WANT_SYS_NICE
3842 * sys_nice - change the priority of the current process.
3843 * @increment: priority increment
3845 * sys_setpriority is a more generic, but much slower function that
3846 * does similar things.
3848 SYSCALL_DEFINE1(nice, int, increment)
3853 * Setpriority might change our priority at the same moment.
3854 * We don't have to worry. Conceptually one call occurs first
3855 * and we have a single winner.
3857 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3858 nice = task_nice(current) + increment;
3860 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3861 if (increment < 0 && !can_nice(current, nice))
3864 retval = security_task_setnice(current, nice);
3868 set_user_nice(current, nice);
3875 * task_prio - return the priority value of a given task.
3876 * @p: the task in question.
3878 * Return: The priority value as seen by users in /proc.
3879 * RT tasks are offset by -200. Normal tasks are centered
3880 * around 0, value goes from -16 to +15.
3882 int task_prio(const struct task_struct *p)
3884 return p->prio - MAX_RT_PRIO;
3888 * idle_cpu - is a given CPU idle currently?
3889 * @cpu: the processor in question.
3891 * Return: 1 if the CPU is currently idle. 0 otherwise.
3893 int idle_cpu(int cpu)
3895 struct rq *rq = cpu_rq(cpu);
3897 if (rq->curr != rq->idle)
3904 if (!llist_empty(&rq->wake_list))
3912 * idle_task - return the idle task for a given CPU.
3913 * @cpu: the processor in question.
3915 * Return: The idle task for the CPU @cpu.
3917 struct task_struct *idle_task(int cpu)
3919 return cpu_rq(cpu)->idle;
3923 * find_process_by_pid - find a process with a matching PID value.
3924 * @pid: the pid in question.
3926 * The task of @pid, if found. %NULL otherwise.
3928 static struct task_struct *find_process_by_pid(pid_t pid)
3930 return pid ? find_task_by_vpid(pid) : current;
3934 * sched_setparam() passes in -1 for its policy, to let the functions
3935 * it calls know not to change it.
3937 #define SETPARAM_POLICY -1
3939 static void __setscheduler_params(struct task_struct *p,
3940 const struct sched_attr *attr)
3942 int policy = attr->sched_policy;
3944 if (policy == SETPARAM_POLICY)
3949 if (dl_policy(policy))
3950 __setparam_dl(p, attr);
3951 else if (fair_policy(policy))
3952 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3955 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3956 * !rt_policy. Always setting this ensures that things like
3957 * getparam()/getattr() don't report silly values for !rt tasks.
3959 p->rt_priority = attr->sched_priority;
3960 p->normal_prio = normal_prio(p);
3964 /* Actually do priority change: must hold pi & rq lock. */
3965 static void __setscheduler(struct rq *rq, struct task_struct *p,
3966 const struct sched_attr *attr, bool keep_boost)
3968 __setscheduler_params(p, attr);
3971 * Keep a potential priority boosting if called from
3972 * sched_setscheduler().
3974 p->prio = normal_prio(p);
3976 p->prio = rt_effective_prio(p, p->prio);
3978 if (dl_prio(p->prio))
3979 p->sched_class = &dl_sched_class;
3980 else if (rt_prio(p->prio))
3981 p->sched_class = &rt_sched_class;
3983 p->sched_class = &fair_sched_class;
3987 * Check the target process has a UID that matches the current process's:
3989 static bool check_same_owner(struct task_struct *p)
3991 const struct cred *cred = current_cred(), *pcred;
3995 pcred = __task_cred(p);
3996 match = (uid_eq(cred->euid, pcred->euid) ||
3997 uid_eq(cred->euid, pcred->uid));
4002 static int __sched_setscheduler(struct task_struct *p,
4003 const struct sched_attr *attr,
4006 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4007 MAX_RT_PRIO - 1 - attr->sched_priority;
4008 int retval, oldprio, oldpolicy = -1, queued, running;
4009 int new_effective_prio, policy = attr->sched_policy;
4010 const struct sched_class *prev_class;
4013 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4016 /* The pi code expects interrupts enabled */
4017 BUG_ON(pi && in_interrupt());
4019 /* Double check policy once rq lock held: */
4021 reset_on_fork = p->sched_reset_on_fork;
4022 policy = oldpolicy = p->policy;
4024 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4026 if (!valid_policy(policy))
4030 if (attr->sched_flags &
4031 ~(SCHED_FLAG_RESET_ON_FORK | SCHED_FLAG_RECLAIM))
4035 * Valid priorities for SCHED_FIFO and SCHED_RR are
4036 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4037 * SCHED_BATCH and SCHED_IDLE is 0.
4039 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4040 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4042 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4043 (rt_policy(policy) != (attr->sched_priority != 0)))
4047 * Allow unprivileged RT tasks to decrease priority:
4049 if (user && !capable(CAP_SYS_NICE)) {
4050 if (fair_policy(policy)) {
4051 if (attr->sched_nice < task_nice(p) &&
4052 !can_nice(p, attr->sched_nice))
4056 if (rt_policy(policy)) {
4057 unsigned long rlim_rtprio =
4058 task_rlimit(p, RLIMIT_RTPRIO);
4060 /* Can't set/change the rt policy: */
4061 if (policy != p->policy && !rlim_rtprio)
4064 /* Can't increase priority: */
4065 if (attr->sched_priority > p->rt_priority &&
4066 attr->sched_priority > rlim_rtprio)
4071 * Can't set/change SCHED_DEADLINE policy at all for now
4072 * (safest behavior); in the future we would like to allow
4073 * unprivileged DL tasks to increase their relative deadline
4074 * or reduce their runtime (both ways reducing utilization)
4076 if (dl_policy(policy))
4080 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4081 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4083 if (idle_policy(p->policy) && !idle_policy(policy)) {
4084 if (!can_nice(p, task_nice(p)))
4088 /* Can't change other user's priorities: */
4089 if (!check_same_owner(p))
4092 /* Normal users shall not reset the sched_reset_on_fork flag: */
4093 if (p->sched_reset_on_fork && !reset_on_fork)
4098 retval = security_task_setscheduler(p);
4104 * Make sure no PI-waiters arrive (or leave) while we are
4105 * changing the priority of the task:
4107 * To be able to change p->policy safely, the appropriate
4108 * runqueue lock must be held.
4110 rq = task_rq_lock(p, &rf);
4111 update_rq_clock(rq);
4114 * Changing the policy of the stop threads its a very bad idea:
4116 if (p == rq->stop) {
4117 task_rq_unlock(rq, p, &rf);
4122 * If not changing anything there's no need to proceed further,
4123 * but store a possible modification of reset_on_fork.
4125 if (unlikely(policy == p->policy)) {
4126 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4128 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4130 if (dl_policy(policy) && dl_param_changed(p, attr))
4133 p->sched_reset_on_fork = reset_on_fork;
4134 task_rq_unlock(rq, p, &rf);
4140 #ifdef CONFIG_RT_GROUP_SCHED
4142 * Do not allow realtime tasks into groups that have no runtime
4145 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4146 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4147 !task_group_is_autogroup(task_group(p))) {
4148 task_rq_unlock(rq, p, &rf);
4153 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4154 cpumask_t *span = rq->rd->span;
4157 * Don't allow tasks with an affinity mask smaller than
4158 * the entire root_domain to become SCHED_DEADLINE. We
4159 * will also fail if there's no bandwidth available.
4161 if (!cpumask_subset(span, &p->cpus_allowed) ||
4162 rq->rd->dl_bw.bw == 0) {
4163 task_rq_unlock(rq, p, &rf);
4170 /* Re-check policy now with rq lock held: */
4171 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4172 policy = oldpolicy = -1;
4173 task_rq_unlock(rq, p, &rf);
4178 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4179 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4182 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4183 task_rq_unlock(rq, p, &rf);
4187 p->sched_reset_on_fork = reset_on_fork;
4192 * Take priority boosted tasks into account. If the new
4193 * effective priority is unchanged, we just store the new
4194 * normal parameters and do not touch the scheduler class and
4195 * the runqueue. This will be done when the task deboost
4198 new_effective_prio = rt_effective_prio(p, newprio);
4199 if (new_effective_prio == oldprio)
4200 queue_flags &= ~DEQUEUE_MOVE;
4203 queued = task_on_rq_queued(p);
4204 running = task_current(rq, p);
4206 dequeue_task(rq, p, queue_flags);
4208 put_prev_task(rq, p);
4210 prev_class = p->sched_class;
4211 __setscheduler(rq, p, attr, pi);
4215 * We enqueue to tail when the priority of a task is
4216 * increased (user space view).
4218 if (oldprio < p->prio)
4219 queue_flags |= ENQUEUE_HEAD;
4221 enqueue_task(rq, p, queue_flags);
4224 set_curr_task(rq, p);
4226 check_class_changed(rq, p, prev_class, oldprio);
4228 /* Avoid rq from going away on us: */
4230 task_rq_unlock(rq, p, &rf);
4233 rt_mutex_adjust_pi(p);
4235 /* Run balance callbacks after we've adjusted the PI chain: */
4236 balance_callback(rq);
4242 static int _sched_setscheduler(struct task_struct *p, int policy,
4243 const struct sched_param *param, bool check)
4245 struct sched_attr attr = {
4246 .sched_policy = policy,
4247 .sched_priority = param->sched_priority,
4248 .sched_nice = PRIO_TO_NICE(p->static_prio),
4251 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4252 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4253 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4254 policy &= ~SCHED_RESET_ON_FORK;
4255 attr.sched_policy = policy;
4258 return __sched_setscheduler(p, &attr, check, true);
4261 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4262 * @p: the task in question.
4263 * @policy: new policy.
4264 * @param: structure containing the new RT priority.
4266 * Return: 0 on success. An error code otherwise.
4268 * NOTE that the task may be already dead.
4270 int sched_setscheduler(struct task_struct *p, int policy,
4271 const struct sched_param *param)
4273 return _sched_setscheduler(p, policy, param, true);
4275 EXPORT_SYMBOL_GPL(sched_setscheduler);
4277 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4279 return __sched_setscheduler(p, attr, true, true);
4281 EXPORT_SYMBOL_GPL(sched_setattr);
4284 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4285 * @p: the task in question.
4286 * @policy: new policy.
4287 * @param: structure containing the new RT priority.
4289 * Just like sched_setscheduler, only don't bother checking if the
4290 * current context has permission. For example, this is needed in
4291 * stop_machine(): we create temporary high priority worker threads,
4292 * but our caller might not have that capability.
4294 * Return: 0 on success. An error code otherwise.
4296 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4297 const struct sched_param *param)
4299 return _sched_setscheduler(p, policy, param, false);
4301 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4304 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4306 struct sched_param lparam;
4307 struct task_struct *p;
4310 if (!param || pid < 0)
4312 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4317 p = find_process_by_pid(pid);
4319 retval = sched_setscheduler(p, policy, &lparam);
4326 * Mimics kernel/events/core.c perf_copy_attr().
4328 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4333 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4336 /* Zero the full structure, so that a short copy will be nice: */
4337 memset(attr, 0, sizeof(*attr));
4339 ret = get_user(size, &uattr->size);
4343 /* Bail out on silly large: */
4344 if (size > PAGE_SIZE)
4347 /* ABI compatibility quirk: */
4349 size = SCHED_ATTR_SIZE_VER0;
4351 if (size < SCHED_ATTR_SIZE_VER0)
4355 * If we're handed a bigger struct than we know of,
4356 * ensure all the unknown bits are 0 - i.e. new
4357 * user-space does not rely on any kernel feature
4358 * extensions we dont know about yet.
4360 if (size > sizeof(*attr)) {
4361 unsigned char __user *addr;
4362 unsigned char __user *end;
4365 addr = (void __user *)uattr + sizeof(*attr);
4366 end = (void __user *)uattr + size;
4368 for (; addr < end; addr++) {
4369 ret = get_user(val, addr);
4375 size = sizeof(*attr);
4378 ret = copy_from_user(attr, uattr, size);
4383 * XXX: Do we want to be lenient like existing syscalls; or do we want
4384 * to be strict and return an error on out-of-bounds values?
4386 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4391 put_user(sizeof(*attr), &uattr->size);
4396 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4397 * @pid: the pid in question.
4398 * @policy: new policy.
4399 * @param: structure containing the new RT priority.
4401 * Return: 0 on success. An error code otherwise.
4403 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4408 return do_sched_setscheduler(pid, policy, param);
4412 * sys_sched_setparam - set/change the RT priority of a thread
4413 * @pid: the pid in question.
4414 * @param: structure containing the new RT priority.
4416 * Return: 0 on success. An error code otherwise.
4418 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4420 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4424 * sys_sched_setattr - same as above, but with extended sched_attr
4425 * @pid: the pid in question.
4426 * @uattr: structure containing the extended parameters.
4427 * @flags: for future extension.
4429 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4430 unsigned int, flags)
4432 struct sched_attr attr;
4433 struct task_struct *p;
4436 if (!uattr || pid < 0 || flags)
4439 retval = sched_copy_attr(uattr, &attr);
4443 if ((int)attr.sched_policy < 0)
4448 p = find_process_by_pid(pid);
4450 retval = sched_setattr(p, &attr);
4457 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4458 * @pid: the pid in question.
4460 * Return: On success, the policy of the thread. Otherwise, a negative error
4463 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4465 struct task_struct *p;
4473 p = find_process_by_pid(pid);
4475 retval = security_task_getscheduler(p);
4478 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4485 * sys_sched_getparam - get the RT priority of a thread
4486 * @pid: the pid in question.
4487 * @param: structure containing the RT priority.
4489 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4492 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4494 struct sched_param lp = { .sched_priority = 0 };
4495 struct task_struct *p;
4498 if (!param || pid < 0)
4502 p = find_process_by_pid(pid);
4507 retval = security_task_getscheduler(p);
4511 if (task_has_rt_policy(p))
4512 lp.sched_priority = p->rt_priority;
4516 * This one might sleep, we cannot do it with a spinlock held ...
4518 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4527 static int sched_read_attr(struct sched_attr __user *uattr,
4528 struct sched_attr *attr,
4533 if (!access_ok(VERIFY_WRITE, uattr, usize))
4537 * If we're handed a smaller struct than we know of,
4538 * ensure all the unknown bits are 0 - i.e. old
4539 * user-space does not get uncomplete information.
4541 if (usize < sizeof(*attr)) {
4542 unsigned char *addr;
4545 addr = (void *)attr + usize;
4546 end = (void *)attr + sizeof(*attr);
4548 for (; addr < end; addr++) {
4556 ret = copy_to_user(uattr, attr, attr->size);
4564 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4565 * @pid: the pid in question.
4566 * @uattr: structure containing the extended parameters.
4567 * @size: sizeof(attr) for fwd/bwd comp.
4568 * @flags: for future extension.
4570 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4571 unsigned int, size, unsigned int, flags)
4573 struct sched_attr attr = {
4574 .size = sizeof(struct sched_attr),
4576 struct task_struct *p;
4579 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4580 size < SCHED_ATTR_SIZE_VER0 || flags)
4584 p = find_process_by_pid(pid);
4589 retval = security_task_getscheduler(p);
4593 attr.sched_policy = p->policy;
4594 if (p->sched_reset_on_fork)
4595 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4596 if (task_has_dl_policy(p))
4597 __getparam_dl(p, &attr);
4598 else if (task_has_rt_policy(p))
4599 attr.sched_priority = p->rt_priority;
4601 attr.sched_nice = task_nice(p);
4605 retval = sched_read_attr(uattr, &attr, size);
4613 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4615 cpumask_var_t cpus_allowed, new_mask;
4616 struct task_struct *p;
4621 p = find_process_by_pid(pid);
4627 /* Prevent p going away */
4631 if (p->flags & PF_NO_SETAFFINITY) {
4635 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4639 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4641 goto out_free_cpus_allowed;
4644 if (!check_same_owner(p)) {
4646 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4648 goto out_free_new_mask;
4653 retval = security_task_setscheduler(p);
4655 goto out_free_new_mask;
4658 cpuset_cpus_allowed(p, cpus_allowed);
4659 cpumask_and(new_mask, in_mask, cpus_allowed);
4662 * Since bandwidth control happens on root_domain basis,
4663 * if admission test is enabled, we only admit -deadline
4664 * tasks allowed to run on all the CPUs in the task's
4668 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4670 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4673 goto out_free_new_mask;
4679 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4682 cpuset_cpus_allowed(p, cpus_allowed);
4683 if (!cpumask_subset(new_mask, cpus_allowed)) {
4685 * We must have raced with a concurrent cpuset
4686 * update. Just reset the cpus_allowed to the
4687 * cpuset's cpus_allowed
4689 cpumask_copy(new_mask, cpus_allowed);
4694 free_cpumask_var(new_mask);
4695 out_free_cpus_allowed:
4696 free_cpumask_var(cpus_allowed);
4702 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4703 struct cpumask *new_mask)
4705 if (len < cpumask_size())
4706 cpumask_clear(new_mask);
4707 else if (len > cpumask_size())
4708 len = cpumask_size();
4710 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4714 * sys_sched_setaffinity - set the CPU affinity of a process
4715 * @pid: pid of the process
4716 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4717 * @user_mask_ptr: user-space pointer to the new CPU mask
4719 * Return: 0 on success. An error code otherwise.
4721 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4722 unsigned long __user *, user_mask_ptr)
4724 cpumask_var_t new_mask;
4727 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4730 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4732 retval = sched_setaffinity(pid, new_mask);
4733 free_cpumask_var(new_mask);
4737 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4739 struct task_struct *p;
4740 unsigned long flags;
4746 p = find_process_by_pid(pid);
4750 retval = security_task_getscheduler(p);
4754 raw_spin_lock_irqsave(&p->pi_lock, flags);
4755 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4756 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4765 * sys_sched_getaffinity - get the CPU affinity of a process
4766 * @pid: pid of the process
4767 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4768 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4770 * Return: size of CPU mask copied to user_mask_ptr on success. An
4771 * error code otherwise.
4773 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4774 unsigned long __user *, user_mask_ptr)
4779 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4781 if (len & (sizeof(unsigned long)-1))
4784 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4787 ret = sched_getaffinity(pid, mask);
4789 size_t retlen = min_t(size_t, len, cpumask_size());
4791 if (copy_to_user(user_mask_ptr, mask, retlen))
4796 free_cpumask_var(mask);
4802 * sys_sched_yield - yield the current processor to other threads.
4804 * This function yields the current CPU to other tasks. If there are no
4805 * other threads running on this CPU then this function will return.
4809 SYSCALL_DEFINE0(sched_yield)
4814 local_irq_disable();
4818 schedstat_inc(rq->yld_count);
4819 current->sched_class->yield_task(rq);
4822 * Since we are going to call schedule() anyway, there's
4823 * no need to preempt or enable interrupts:
4827 sched_preempt_enable_no_resched();
4834 #ifndef CONFIG_PREEMPT
4835 int __sched _cond_resched(void)
4837 if (should_resched(0)) {
4838 preempt_schedule_common();
4843 EXPORT_SYMBOL(_cond_resched);
4847 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4848 * call schedule, and on return reacquire the lock.
4850 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4851 * operations here to prevent schedule() from being called twice (once via
4852 * spin_unlock(), once by hand).
4854 int __cond_resched_lock(spinlock_t *lock)
4856 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4859 lockdep_assert_held(lock);
4861 if (spin_needbreak(lock) || resched) {
4864 preempt_schedule_common();
4872 EXPORT_SYMBOL(__cond_resched_lock);
4874 int __sched __cond_resched_softirq(void)
4876 BUG_ON(!in_softirq());
4878 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4880 preempt_schedule_common();
4886 EXPORT_SYMBOL(__cond_resched_softirq);
4889 * yield - yield the current processor to other threads.
4891 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4893 * The scheduler is at all times free to pick the calling task as the most
4894 * eligible task to run, if removing the yield() call from your code breaks
4895 * it, its already broken.
4897 * Typical broken usage is:
4902 * where one assumes that yield() will let 'the other' process run that will
4903 * make event true. If the current task is a SCHED_FIFO task that will never
4904 * happen. Never use yield() as a progress guarantee!!
4906 * If you want to use yield() to wait for something, use wait_event().
4907 * If you want to use yield() to be 'nice' for others, use cond_resched().
4908 * If you still want to use yield(), do not!
4910 void __sched yield(void)
4912 set_current_state(TASK_RUNNING);
4915 EXPORT_SYMBOL(yield);
4918 * yield_to - yield the current processor to another thread in
4919 * your thread group, or accelerate that thread toward the
4920 * processor it's on.
4922 * @preempt: whether task preemption is allowed or not
4924 * It's the caller's job to ensure that the target task struct
4925 * can't go away on us before we can do any checks.
4928 * true (>0) if we indeed boosted the target task.
4929 * false (0) if we failed to boost the target.
4930 * -ESRCH if there's no task to yield to.
4932 int __sched yield_to(struct task_struct *p, bool preempt)
4934 struct task_struct *curr = current;
4935 struct rq *rq, *p_rq;
4936 unsigned long flags;
4939 local_irq_save(flags);
4945 * If we're the only runnable task on the rq and target rq also
4946 * has only one task, there's absolutely no point in yielding.
4948 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4953 double_rq_lock(rq, p_rq);
4954 if (task_rq(p) != p_rq) {
4955 double_rq_unlock(rq, p_rq);
4959 if (!curr->sched_class->yield_to_task)
4962 if (curr->sched_class != p->sched_class)
4965 if (task_running(p_rq, p) || p->state)
4968 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4970 schedstat_inc(rq->yld_count);
4972 * Make p's CPU reschedule; pick_next_entity takes care of
4975 if (preempt && rq != p_rq)
4980 double_rq_unlock(rq, p_rq);
4982 local_irq_restore(flags);
4989 EXPORT_SYMBOL_GPL(yield_to);
4991 int io_schedule_prepare(void)
4993 int old_iowait = current->in_iowait;
4995 current->in_iowait = 1;
4996 blk_schedule_flush_plug(current);
5001 void io_schedule_finish(int token)
5003 current->in_iowait = token;
5007 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5008 * that process accounting knows that this is a task in IO wait state.
5010 long __sched io_schedule_timeout(long timeout)
5015 token = io_schedule_prepare();
5016 ret = schedule_timeout(timeout);
5017 io_schedule_finish(token);
5021 EXPORT_SYMBOL(io_schedule_timeout);
5023 void io_schedule(void)
5027 token = io_schedule_prepare();
5029 io_schedule_finish(token);
5031 EXPORT_SYMBOL(io_schedule);
5034 * sys_sched_get_priority_max - return maximum RT priority.
5035 * @policy: scheduling class.
5037 * Return: On success, this syscall returns the maximum
5038 * rt_priority that can be used by a given scheduling class.
5039 * On failure, a negative error code is returned.
5041 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5048 ret = MAX_USER_RT_PRIO-1;
5050 case SCHED_DEADLINE:
5061 * sys_sched_get_priority_min - return minimum RT priority.
5062 * @policy: scheduling class.
5064 * Return: On success, this syscall returns the minimum
5065 * rt_priority that can be used by a given scheduling class.
5066 * On failure, a negative error code is returned.
5068 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5077 case SCHED_DEADLINE:
5087 * sys_sched_rr_get_interval - return the default timeslice of a process.
5088 * @pid: pid of the process.
5089 * @interval: userspace pointer to the timeslice value.
5091 * this syscall writes the default timeslice value of a given process
5092 * into the user-space timespec buffer. A value of '0' means infinity.
5094 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5097 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5098 struct timespec __user *, interval)
5100 struct task_struct *p;
5101 unsigned int time_slice;
5112 p = find_process_by_pid(pid);
5116 retval = security_task_getscheduler(p);
5120 rq = task_rq_lock(p, &rf);
5122 if (p->sched_class->get_rr_interval)
5123 time_slice = p->sched_class->get_rr_interval(rq, p);
5124 task_rq_unlock(rq, p, &rf);
5127 jiffies_to_timespec(time_slice, &t);
5128 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5136 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5138 void sched_show_task(struct task_struct *p)
5140 unsigned long free = 0;
5142 unsigned long state = p->state;
5144 /* Make sure the string lines up properly with the number of task states: */
5145 BUILD_BUG_ON(sizeof(TASK_STATE_TO_CHAR_STR)-1 != ilog2(TASK_STATE_MAX)+1);
5147 if (!try_get_task_stack(p))
5150 state = __ffs(state) + 1;
5151 printk(KERN_INFO "%-15.15s %c", p->comm,
5152 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5153 if (state == TASK_RUNNING)
5154 printk(KERN_CONT " running task ");
5155 #ifdef CONFIG_DEBUG_STACK_USAGE
5156 free = stack_not_used(p);
5161 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5163 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5164 task_pid_nr(p), ppid,
5165 (unsigned long)task_thread_info(p)->flags);
5167 print_worker_info(KERN_INFO, p);
5168 show_stack(p, NULL);
5172 void show_state_filter(unsigned long state_filter)
5174 struct task_struct *g, *p;
5176 #if BITS_PER_LONG == 32
5178 " task PC stack pid father\n");
5181 " task PC stack pid father\n");
5184 for_each_process_thread(g, p) {
5186 * reset the NMI-timeout, listing all files on a slow
5187 * console might take a lot of time:
5188 * Also, reset softlockup watchdogs on all CPUs, because
5189 * another CPU might be blocked waiting for us to process
5192 touch_nmi_watchdog();
5193 touch_all_softlockup_watchdogs();
5194 if (!state_filter || (p->state & state_filter))
5198 #ifdef CONFIG_SCHED_DEBUG
5200 sysrq_sched_debug_show();
5204 * Only show locks if all tasks are dumped:
5207 debug_show_all_locks();
5210 void init_idle_bootup_task(struct task_struct *idle)
5212 idle->sched_class = &idle_sched_class;
5216 * init_idle - set up an idle thread for a given CPU
5217 * @idle: task in question
5218 * @cpu: CPU the idle task belongs to
5220 * NOTE: this function does not set the idle thread's NEED_RESCHED
5221 * flag, to make booting more robust.
5223 void init_idle(struct task_struct *idle, int cpu)
5225 struct rq *rq = cpu_rq(cpu);
5226 unsigned long flags;
5228 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5229 raw_spin_lock(&rq->lock);
5231 __sched_fork(0, idle);
5232 idle->state = TASK_RUNNING;
5233 idle->se.exec_start = sched_clock();
5234 idle->flags |= PF_IDLE;
5236 kasan_unpoison_task_stack(idle);
5240 * Its possible that init_idle() gets called multiple times on a task,
5241 * in that case do_set_cpus_allowed() will not do the right thing.
5243 * And since this is boot we can forgo the serialization.
5245 set_cpus_allowed_common(idle, cpumask_of(cpu));
5248 * We're having a chicken and egg problem, even though we are
5249 * holding rq->lock, the CPU isn't yet set to this CPU so the
5250 * lockdep check in task_group() will fail.
5252 * Similar case to sched_fork(). / Alternatively we could
5253 * use task_rq_lock() here and obtain the other rq->lock.
5258 __set_task_cpu(idle, cpu);
5261 rq->curr = rq->idle = idle;
5262 idle->on_rq = TASK_ON_RQ_QUEUED;
5266 raw_spin_unlock(&rq->lock);
5267 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5269 /* Set the preempt count _outside_ the spinlocks! */
5270 init_idle_preempt_count(idle, cpu);
5273 * The idle tasks have their own, simple scheduling class:
5275 idle->sched_class = &idle_sched_class;
5276 ftrace_graph_init_idle_task(idle, cpu);
5277 vtime_init_idle(idle, cpu);
5279 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5285 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5286 const struct cpumask *trial)
5290 if (!cpumask_weight(cur))
5293 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5298 int task_can_attach(struct task_struct *p,
5299 const struct cpumask *cs_cpus_allowed)
5304 * Kthreads which disallow setaffinity shouldn't be moved
5305 * to a new cpuset; we don't want to change their CPU
5306 * affinity and isolating such threads by their set of
5307 * allowed nodes is unnecessary. Thus, cpusets are not
5308 * applicable for such threads. This prevents checking for
5309 * success of set_cpus_allowed_ptr() on all attached tasks
5310 * before cpus_allowed may be changed.
5312 if (p->flags & PF_NO_SETAFFINITY) {
5317 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5319 ret = dl_task_can_attach(p, cs_cpus_allowed);
5325 bool sched_smp_initialized __read_mostly;
5327 #ifdef CONFIG_NUMA_BALANCING
5328 /* Migrate current task p to target_cpu */
5329 int migrate_task_to(struct task_struct *p, int target_cpu)
5331 struct migration_arg arg = { p, target_cpu };
5332 int curr_cpu = task_cpu(p);
5334 if (curr_cpu == target_cpu)
5337 if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5340 /* TODO: This is not properly updating schedstats */
5342 trace_sched_move_numa(p, curr_cpu, target_cpu);
5343 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5347 * Requeue a task on a given node and accurately track the number of NUMA
5348 * tasks on the runqueues
5350 void sched_setnuma(struct task_struct *p, int nid)
5352 bool queued, running;
5356 rq = task_rq_lock(p, &rf);
5357 queued = task_on_rq_queued(p);
5358 running = task_current(rq, p);
5361 dequeue_task(rq, p, DEQUEUE_SAVE);
5363 put_prev_task(rq, p);
5365 p->numa_preferred_nid = nid;
5368 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5370 set_curr_task(rq, p);
5371 task_rq_unlock(rq, p, &rf);
5373 #endif /* CONFIG_NUMA_BALANCING */
5375 #ifdef CONFIG_HOTPLUG_CPU
5377 * Ensure that the idle task is using init_mm right before its CPU goes
5380 void idle_task_exit(void)
5382 struct mm_struct *mm = current->active_mm;
5384 BUG_ON(cpu_online(smp_processor_id()));
5386 if (mm != &init_mm) {
5387 switch_mm(mm, &init_mm, current);
5388 finish_arch_post_lock_switch();
5394 * Since this CPU is going 'away' for a while, fold any nr_active delta
5395 * we might have. Assumes we're called after migrate_tasks() so that the
5396 * nr_active count is stable. We need to take the teardown thread which
5397 * is calling this into account, so we hand in adjust = 1 to the load
5400 * Also see the comment "Global load-average calculations".
5402 static void calc_load_migrate(struct rq *rq)
5404 long delta = calc_load_fold_active(rq, 1);
5406 atomic_long_add(delta, &calc_load_tasks);
5409 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5413 static const struct sched_class fake_sched_class = {
5414 .put_prev_task = put_prev_task_fake,
5417 static struct task_struct fake_task = {
5419 * Avoid pull_{rt,dl}_task()
5421 .prio = MAX_PRIO + 1,
5422 .sched_class = &fake_sched_class,
5426 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5427 * try_to_wake_up()->select_task_rq().
5429 * Called with rq->lock held even though we'er in stop_machine() and
5430 * there's no concurrency possible, we hold the required locks anyway
5431 * because of lock validation efforts.
5433 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5435 struct rq *rq = dead_rq;
5436 struct task_struct *next, *stop = rq->stop;
5437 struct rq_flags orf = *rf;
5441 * Fudge the rq selection such that the below task selection loop
5442 * doesn't get stuck on the currently eligible stop task.
5444 * We're currently inside stop_machine() and the rq is either stuck
5445 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5446 * either way we should never end up calling schedule() until we're
5452 * put_prev_task() and pick_next_task() sched
5453 * class method both need to have an up-to-date
5454 * value of rq->clock[_task]
5456 update_rq_clock(rq);
5460 * There's this thread running, bail when that's the only
5463 if (rq->nr_running == 1)
5467 * pick_next_task() assumes pinned rq->lock:
5469 next = pick_next_task(rq, &fake_task, rf);
5471 next->sched_class->put_prev_task(rq, next);
5474 * Rules for changing task_struct::cpus_allowed are holding
5475 * both pi_lock and rq->lock, such that holding either
5476 * stabilizes the mask.
5478 * Drop rq->lock is not quite as disastrous as it usually is
5479 * because !cpu_active at this point, which means load-balance
5480 * will not interfere. Also, stop-machine.
5483 raw_spin_lock(&next->pi_lock);
5487 * Since we're inside stop-machine, _nothing_ should have
5488 * changed the task, WARN if weird stuff happened, because in
5489 * that case the above rq->lock drop is a fail too.
5491 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5492 raw_spin_unlock(&next->pi_lock);
5496 /* Find suitable destination for @next, with force if needed. */
5497 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5498 rq = __migrate_task(rq, rf, next, dest_cpu);
5499 if (rq != dead_rq) {
5505 raw_spin_unlock(&next->pi_lock);
5510 #endif /* CONFIG_HOTPLUG_CPU */
5512 void set_rq_online(struct rq *rq)
5515 const struct sched_class *class;
5517 cpumask_set_cpu(rq->cpu, rq->rd->online);
5520 for_each_class(class) {
5521 if (class->rq_online)
5522 class->rq_online(rq);
5527 void set_rq_offline(struct rq *rq)
5530 const struct sched_class *class;
5532 for_each_class(class) {
5533 if (class->rq_offline)
5534 class->rq_offline(rq);
5537 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5542 static void set_cpu_rq_start_time(unsigned int cpu)
5544 struct rq *rq = cpu_rq(cpu);
5546 rq->age_stamp = sched_clock_cpu(cpu);
5550 * used to mark begin/end of suspend/resume:
5552 static int num_cpus_frozen;
5555 * Update cpusets according to cpu_active mask. If cpusets are
5556 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5557 * around partition_sched_domains().
5559 * If we come here as part of a suspend/resume, don't touch cpusets because we
5560 * want to restore it back to its original state upon resume anyway.
5562 static void cpuset_cpu_active(void)
5564 if (cpuhp_tasks_frozen) {
5566 * num_cpus_frozen tracks how many CPUs are involved in suspend
5567 * resume sequence. As long as this is not the last online
5568 * operation in the resume sequence, just build a single sched
5569 * domain, ignoring cpusets.
5572 if (likely(num_cpus_frozen)) {
5573 partition_sched_domains(1, NULL, NULL);
5577 * This is the last CPU online operation. So fall through and
5578 * restore the original sched domains by considering the
5579 * cpuset configurations.
5582 cpuset_update_active_cpus();
5585 static int cpuset_cpu_inactive(unsigned int cpu)
5587 if (!cpuhp_tasks_frozen) {
5588 if (dl_cpu_busy(cpu))
5590 cpuset_update_active_cpus();
5593 partition_sched_domains(1, NULL, NULL);
5598 int sched_cpu_activate(unsigned int cpu)
5600 struct rq *rq = cpu_rq(cpu);
5603 set_cpu_active(cpu, true);
5605 if (sched_smp_initialized) {
5606 sched_domains_numa_masks_set(cpu);
5607 cpuset_cpu_active();
5611 * Put the rq online, if not already. This happens:
5613 * 1) In the early boot process, because we build the real domains
5614 * after all CPUs have been brought up.
5616 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5619 rq_lock_irqsave(rq, &rf);
5621 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5624 rq_unlock_irqrestore(rq, &rf);
5626 update_max_interval();
5631 int sched_cpu_deactivate(unsigned int cpu)
5635 set_cpu_active(cpu, false);
5637 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5638 * users of this state to go away such that all new such users will
5641 * Do sync before park smpboot threads to take care the rcu boost case.
5643 synchronize_rcu_mult(call_rcu, call_rcu_sched);
5645 if (!sched_smp_initialized)
5648 ret = cpuset_cpu_inactive(cpu);
5650 set_cpu_active(cpu, true);
5653 sched_domains_numa_masks_clear(cpu);
5657 static void sched_rq_cpu_starting(unsigned int cpu)
5659 struct rq *rq = cpu_rq(cpu);
5661 rq->calc_load_update = calc_load_update;
5662 update_max_interval();
5665 int sched_cpu_starting(unsigned int cpu)
5667 set_cpu_rq_start_time(cpu);
5668 sched_rq_cpu_starting(cpu);
5672 #ifdef CONFIG_HOTPLUG_CPU
5673 int sched_cpu_dying(unsigned int cpu)
5675 struct rq *rq = cpu_rq(cpu);
5678 /* Handle pending wakeups and then migrate everything off */
5679 sched_ttwu_pending();
5681 rq_lock_irqsave(rq, &rf);
5683 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5686 migrate_tasks(rq, &rf);
5687 BUG_ON(rq->nr_running != 1);
5688 rq_unlock_irqrestore(rq, &rf);
5690 calc_load_migrate(rq);
5691 update_max_interval();
5692 nohz_balance_exit_idle(cpu);
5698 #ifdef CONFIG_SCHED_SMT
5699 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5701 static void sched_init_smt(void)
5704 * We've enumerated all CPUs and will assume that if any CPU
5705 * has SMT siblings, CPU0 will too.
5707 if (cpumask_weight(cpu_smt_mask(0)) > 1)
5708 static_branch_enable(&sched_smt_present);
5711 static inline void sched_init_smt(void) { }
5714 void __init sched_init_smp(void)
5716 cpumask_var_t non_isolated_cpus;
5718 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
5723 * There's no userspace yet to cause hotplug operations; hence all the
5724 * CPU masks are stable and all blatant races in the below code cannot
5727 mutex_lock(&sched_domains_mutex);
5728 sched_init_domains(cpu_active_mask);
5729 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
5730 if (cpumask_empty(non_isolated_cpus))
5731 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
5732 mutex_unlock(&sched_domains_mutex);
5734 /* Move init over to a non-isolated CPU */
5735 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
5737 sched_init_granularity();
5738 free_cpumask_var(non_isolated_cpus);
5740 init_sched_rt_class();
5741 init_sched_dl_class();
5745 sched_smp_initialized = true;
5748 static int __init migration_init(void)
5750 sched_rq_cpu_starting(smp_processor_id());
5753 early_initcall(migration_init);
5756 void __init sched_init_smp(void)
5758 sched_init_granularity();
5760 #endif /* CONFIG_SMP */
5762 int in_sched_functions(unsigned long addr)
5764 return in_lock_functions(addr) ||
5765 (addr >= (unsigned long)__sched_text_start
5766 && addr < (unsigned long)__sched_text_end);
5769 #ifdef CONFIG_CGROUP_SCHED
5771 * Default task group.
5772 * Every task in system belongs to this group at bootup.
5774 struct task_group root_task_group;
5775 LIST_HEAD(task_groups);
5777 /* Cacheline aligned slab cache for task_group */
5778 static struct kmem_cache *task_group_cache __read_mostly;
5781 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5782 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5784 void __init sched_init(void)
5787 unsigned long alloc_size = 0, ptr;
5792 #ifdef CONFIG_FAIR_GROUP_SCHED
5793 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5795 #ifdef CONFIG_RT_GROUP_SCHED
5796 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5799 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5801 #ifdef CONFIG_FAIR_GROUP_SCHED
5802 root_task_group.se = (struct sched_entity **)ptr;
5803 ptr += nr_cpu_ids * sizeof(void **);
5805 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5806 ptr += nr_cpu_ids * sizeof(void **);
5808 #endif /* CONFIG_FAIR_GROUP_SCHED */
5809 #ifdef CONFIG_RT_GROUP_SCHED
5810 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5811 ptr += nr_cpu_ids * sizeof(void **);
5813 root_task_group.rt_rq = (struct rt_rq **)ptr;
5814 ptr += nr_cpu_ids * sizeof(void **);
5816 #endif /* CONFIG_RT_GROUP_SCHED */
5818 #ifdef CONFIG_CPUMASK_OFFSTACK
5819 for_each_possible_cpu(i) {
5820 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5821 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5822 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
5823 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5825 #endif /* CONFIG_CPUMASK_OFFSTACK */
5827 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
5828 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
5831 init_defrootdomain();
5834 #ifdef CONFIG_RT_GROUP_SCHED
5835 init_rt_bandwidth(&root_task_group.rt_bandwidth,
5836 global_rt_period(), global_rt_runtime());
5837 #endif /* CONFIG_RT_GROUP_SCHED */
5839 #ifdef CONFIG_CGROUP_SCHED
5840 task_group_cache = KMEM_CACHE(task_group, 0);
5842 list_add(&root_task_group.list, &task_groups);
5843 INIT_LIST_HEAD(&root_task_group.children);
5844 INIT_LIST_HEAD(&root_task_group.siblings);
5845 autogroup_init(&init_task);
5846 #endif /* CONFIG_CGROUP_SCHED */
5848 for_each_possible_cpu(i) {
5852 raw_spin_lock_init(&rq->lock);
5854 rq->calc_load_active = 0;
5855 rq->calc_load_update = jiffies + LOAD_FREQ;
5856 init_cfs_rq(&rq->cfs);
5857 init_rt_rq(&rq->rt);
5858 init_dl_rq(&rq->dl);
5859 #ifdef CONFIG_FAIR_GROUP_SCHED
5860 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
5861 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
5862 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
5864 * How much CPU bandwidth does root_task_group get?
5866 * In case of task-groups formed thr' the cgroup filesystem, it
5867 * gets 100% of the CPU resources in the system. This overall
5868 * system CPU resource is divided among the tasks of
5869 * root_task_group and its child task-groups in a fair manner,
5870 * based on each entity's (task or task-group's) weight
5871 * (se->load.weight).
5873 * In other words, if root_task_group has 10 tasks of weight
5874 * 1024) and two child groups A0 and A1 (of weight 1024 each),
5875 * then A0's share of the CPU resource is:
5877 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
5879 * We achieve this by letting root_task_group's tasks sit
5880 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
5882 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
5883 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
5884 #endif /* CONFIG_FAIR_GROUP_SCHED */
5886 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
5887 #ifdef CONFIG_RT_GROUP_SCHED
5888 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
5891 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
5892 rq->cpu_load[j] = 0;
5897 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
5898 rq->balance_callback = NULL;
5899 rq->active_balance = 0;
5900 rq->next_balance = jiffies;
5905 rq->avg_idle = 2*sysctl_sched_migration_cost;
5906 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
5908 INIT_LIST_HEAD(&rq->cfs_tasks);
5910 rq_attach_root(rq, &def_root_domain);
5911 #ifdef CONFIG_NO_HZ_COMMON
5912 rq->last_load_update_tick = jiffies;
5915 #ifdef CONFIG_NO_HZ_FULL
5916 rq->last_sched_tick = 0;
5918 #endif /* CONFIG_SMP */
5920 atomic_set(&rq->nr_iowait, 0);
5923 set_load_weight(&init_task);
5926 * The boot idle thread does lazy MMU switching as well:
5929 enter_lazy_tlb(&init_mm, current);
5932 * Make us the idle thread. Technically, schedule() should not be
5933 * called from this thread, however somewhere below it might be,
5934 * but because we are the idle thread, we just pick up running again
5935 * when this runqueue becomes "idle".
5937 init_idle(current, smp_processor_id());
5939 calc_load_update = jiffies + LOAD_FREQ;
5942 /* May be allocated at isolcpus cmdline parse time */
5943 if (cpu_isolated_map == NULL)
5944 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
5945 idle_thread_set_boot_cpu();
5946 set_cpu_rq_start_time(smp_processor_id());
5948 init_sched_fair_class();
5952 scheduler_running = 1;
5955 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
5956 static inline int preempt_count_equals(int preempt_offset)
5958 int nested = preempt_count() + rcu_preempt_depth();
5960 return (nested == preempt_offset);
5963 void __might_sleep(const char *file, int line, int preempt_offset)
5966 * Blocking primitives will set (and therefore destroy) current->state,
5967 * since we will exit with TASK_RUNNING make sure we enter with it,
5968 * otherwise we will destroy state.
5970 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
5971 "do not call blocking ops when !TASK_RUNNING; "
5972 "state=%lx set at [<%p>] %pS\n",
5974 (void *)current->task_state_change,
5975 (void *)current->task_state_change);
5977 ___might_sleep(file, line, preempt_offset);
5979 EXPORT_SYMBOL(__might_sleep);
5981 void ___might_sleep(const char *file, int line, int preempt_offset)
5983 /* Ratelimiting timestamp: */
5984 static unsigned long prev_jiffy;
5986 unsigned long preempt_disable_ip;
5988 /* WARN_ON_ONCE() by default, no rate limit required: */
5991 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
5992 !is_idle_task(current)) ||
5993 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
5997 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
5999 prev_jiffy = jiffies;
6001 /* Save this before calling printk(), since that will clobber it: */
6002 preempt_disable_ip = get_preempt_disable_ip(current);
6005 "BUG: sleeping function called from invalid context at %s:%d\n",
6008 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6009 in_atomic(), irqs_disabled(),
6010 current->pid, current->comm);
6012 if (task_stack_end_corrupted(current))
6013 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6015 debug_show_held_locks(current);
6016 if (irqs_disabled())
6017 print_irqtrace_events(current);
6018 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6019 && !preempt_count_equals(preempt_offset)) {
6020 pr_err("Preemption disabled at:");
6021 print_ip_sym(preempt_disable_ip);
6025 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6027 EXPORT_SYMBOL(___might_sleep);
6030 #ifdef CONFIG_MAGIC_SYSRQ
6031 void normalize_rt_tasks(void)
6033 struct task_struct *g, *p;
6034 struct sched_attr attr = {
6035 .sched_policy = SCHED_NORMAL,
6038 read_lock(&tasklist_lock);
6039 for_each_process_thread(g, p) {
6041 * Only normalize user tasks:
6043 if (p->flags & PF_KTHREAD)
6046 p->se.exec_start = 0;
6047 schedstat_set(p->se.statistics.wait_start, 0);
6048 schedstat_set(p->se.statistics.sleep_start, 0);
6049 schedstat_set(p->se.statistics.block_start, 0);
6051 if (!dl_task(p) && !rt_task(p)) {
6053 * Renice negative nice level userspace
6056 if (task_nice(p) < 0)
6057 set_user_nice(p, 0);
6061 __sched_setscheduler(p, &attr, false, false);
6063 read_unlock(&tasklist_lock);
6066 #endif /* CONFIG_MAGIC_SYSRQ */
6068 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6070 * These functions are only useful for the IA64 MCA handling, or kdb.
6072 * They can only be called when the whole system has been
6073 * stopped - every CPU needs to be quiescent, and no scheduling
6074 * activity can take place. Using them for anything else would
6075 * be a serious bug, and as a result, they aren't even visible
6076 * under any other configuration.
6080 * curr_task - return the current task for a given CPU.
6081 * @cpu: the processor in question.
6083 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6085 * Return: The current task for @cpu.
6087 struct task_struct *curr_task(int cpu)
6089 return cpu_curr(cpu);
6092 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6096 * set_curr_task - set the current task for a given CPU.
6097 * @cpu: the processor in question.
6098 * @p: the task pointer to set.
6100 * Description: This function must only be used when non-maskable interrupts
6101 * are serviced on a separate stack. It allows the architecture to switch the
6102 * notion of the current task on a CPU in a non-blocking manner. This function
6103 * must be called with all CPU's synchronized, and interrupts disabled, the
6104 * and caller must save the original value of the current task (see
6105 * curr_task() above) and restore that value before reenabling interrupts and
6106 * re-starting the system.
6108 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6110 void ia64_set_curr_task(int cpu, struct task_struct *p)
6117 #ifdef CONFIG_CGROUP_SCHED
6118 /* task_group_lock serializes the addition/removal of task groups */
6119 static DEFINE_SPINLOCK(task_group_lock);
6121 static void sched_free_group(struct task_group *tg)
6123 free_fair_sched_group(tg);
6124 free_rt_sched_group(tg);
6126 kmem_cache_free(task_group_cache, tg);
6129 /* allocate runqueue etc for a new task group */
6130 struct task_group *sched_create_group(struct task_group *parent)
6132 struct task_group *tg;
6134 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6136 return ERR_PTR(-ENOMEM);
6138 if (!alloc_fair_sched_group(tg, parent))
6141 if (!alloc_rt_sched_group(tg, parent))
6147 sched_free_group(tg);
6148 return ERR_PTR(-ENOMEM);
6151 void sched_online_group(struct task_group *tg, struct task_group *parent)
6153 unsigned long flags;
6155 spin_lock_irqsave(&task_group_lock, flags);
6156 list_add_rcu(&tg->list, &task_groups);
6158 /* Root should already exist: */
6161 tg->parent = parent;
6162 INIT_LIST_HEAD(&tg->children);
6163 list_add_rcu(&tg->siblings, &parent->children);
6164 spin_unlock_irqrestore(&task_group_lock, flags);
6166 online_fair_sched_group(tg);
6169 /* rcu callback to free various structures associated with a task group */
6170 static void sched_free_group_rcu(struct rcu_head *rhp)
6172 /* Now it should be safe to free those cfs_rqs: */
6173 sched_free_group(container_of(rhp, struct task_group, rcu));
6176 void sched_destroy_group(struct task_group *tg)
6178 /* Wait for possible concurrent references to cfs_rqs complete: */
6179 call_rcu(&tg->rcu, sched_free_group_rcu);
6182 void sched_offline_group(struct task_group *tg)
6184 unsigned long flags;
6186 /* End participation in shares distribution: */
6187 unregister_fair_sched_group(tg);
6189 spin_lock_irqsave(&task_group_lock, flags);
6190 list_del_rcu(&tg->list);
6191 list_del_rcu(&tg->siblings);
6192 spin_unlock_irqrestore(&task_group_lock, flags);
6195 static void sched_change_group(struct task_struct *tsk, int type)
6197 struct task_group *tg;
6200 * All callers are synchronized by task_rq_lock(); we do not use RCU
6201 * which is pointless here. Thus, we pass "true" to task_css_check()
6202 * to prevent lockdep warnings.
6204 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6205 struct task_group, css);
6206 tg = autogroup_task_group(tsk, tg);
6207 tsk->sched_task_group = tg;
6209 #ifdef CONFIG_FAIR_GROUP_SCHED
6210 if (tsk->sched_class->task_change_group)
6211 tsk->sched_class->task_change_group(tsk, type);
6214 set_task_rq(tsk, task_cpu(tsk));
6218 * Change task's runqueue when it moves between groups.
6220 * The caller of this function should have put the task in its new group by
6221 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6224 void sched_move_task(struct task_struct *tsk)
6226 int queued, running, queue_flags =
6227 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6231 rq = task_rq_lock(tsk, &rf);
6232 update_rq_clock(rq);
6234 running = task_current(rq, tsk);
6235 queued = task_on_rq_queued(tsk);
6238 dequeue_task(rq, tsk, queue_flags);
6240 put_prev_task(rq, tsk);
6242 sched_change_group(tsk, TASK_MOVE_GROUP);
6245 enqueue_task(rq, tsk, queue_flags);
6247 set_curr_task(rq, tsk);
6249 task_rq_unlock(rq, tsk, &rf);
6252 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6254 return css ? container_of(css, struct task_group, css) : NULL;
6257 static struct cgroup_subsys_state *
6258 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6260 struct task_group *parent = css_tg(parent_css);
6261 struct task_group *tg;
6264 /* This is early initialization for the top cgroup */
6265 return &root_task_group.css;
6268 tg = sched_create_group(parent);
6270 return ERR_PTR(-ENOMEM);
6275 /* Expose task group only after completing cgroup initialization */
6276 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6278 struct task_group *tg = css_tg(css);
6279 struct task_group *parent = css_tg(css->parent);
6282 sched_online_group(tg, parent);
6286 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6288 struct task_group *tg = css_tg(css);
6290 sched_offline_group(tg);
6293 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6295 struct task_group *tg = css_tg(css);
6298 * Relies on the RCU grace period between css_released() and this.
6300 sched_free_group(tg);
6304 * This is called before wake_up_new_task(), therefore we really only
6305 * have to set its group bits, all the other stuff does not apply.
6307 static void cpu_cgroup_fork(struct task_struct *task)
6312 rq = task_rq_lock(task, &rf);
6314 update_rq_clock(rq);
6315 sched_change_group(task, TASK_SET_GROUP);
6317 task_rq_unlock(rq, task, &rf);
6320 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6322 struct task_struct *task;
6323 struct cgroup_subsys_state *css;
6326 cgroup_taskset_for_each(task, css, tset) {
6327 #ifdef CONFIG_RT_GROUP_SCHED
6328 if (!sched_rt_can_attach(css_tg(css), task))
6331 /* We don't support RT-tasks being in separate groups */
6332 if (task->sched_class != &fair_sched_class)
6336 * Serialize against wake_up_new_task() such that if its
6337 * running, we're sure to observe its full state.
6339 raw_spin_lock_irq(&task->pi_lock);
6341 * Avoid calling sched_move_task() before wake_up_new_task()
6342 * has happened. This would lead to problems with PELT, due to
6343 * move wanting to detach+attach while we're not attached yet.
6345 if (task->state == TASK_NEW)
6347 raw_spin_unlock_irq(&task->pi_lock);
6355 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6357 struct task_struct *task;
6358 struct cgroup_subsys_state *css;
6360 cgroup_taskset_for_each(task, css, tset)
6361 sched_move_task(task);
6364 #ifdef CONFIG_FAIR_GROUP_SCHED
6365 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6366 struct cftype *cftype, u64 shareval)
6368 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6371 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6374 struct task_group *tg = css_tg(css);
6376 return (u64) scale_load_down(tg->shares);
6379 #ifdef CONFIG_CFS_BANDWIDTH
6380 static DEFINE_MUTEX(cfs_constraints_mutex);
6382 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6383 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6385 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6387 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6389 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6390 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6392 if (tg == &root_task_group)
6396 * Ensure we have at some amount of bandwidth every period. This is
6397 * to prevent reaching a state of large arrears when throttled via
6398 * entity_tick() resulting in prolonged exit starvation.
6400 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6404 * Likewise, bound things on the otherside by preventing insane quota
6405 * periods. This also allows us to normalize in computing quota
6408 if (period > max_cfs_quota_period)
6412 * Prevent race between setting of cfs_rq->runtime_enabled and
6413 * unthrottle_offline_cfs_rqs().
6416 mutex_lock(&cfs_constraints_mutex);
6417 ret = __cfs_schedulable(tg, period, quota);
6421 runtime_enabled = quota != RUNTIME_INF;
6422 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6424 * If we need to toggle cfs_bandwidth_used, off->on must occur
6425 * before making related changes, and on->off must occur afterwards
6427 if (runtime_enabled && !runtime_was_enabled)
6428 cfs_bandwidth_usage_inc();
6429 raw_spin_lock_irq(&cfs_b->lock);
6430 cfs_b->period = ns_to_ktime(period);
6431 cfs_b->quota = quota;
6433 __refill_cfs_bandwidth_runtime(cfs_b);
6435 /* Restart the period timer (if active) to handle new period expiry: */
6436 if (runtime_enabled)
6437 start_cfs_bandwidth(cfs_b);
6439 raw_spin_unlock_irq(&cfs_b->lock);
6441 for_each_online_cpu(i) {
6442 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6443 struct rq *rq = cfs_rq->rq;
6446 rq_lock_irq(rq, &rf);
6447 cfs_rq->runtime_enabled = runtime_enabled;
6448 cfs_rq->runtime_remaining = 0;
6450 if (cfs_rq->throttled)
6451 unthrottle_cfs_rq(cfs_rq);
6452 rq_unlock_irq(rq, &rf);
6454 if (runtime_was_enabled && !runtime_enabled)
6455 cfs_bandwidth_usage_dec();
6457 mutex_unlock(&cfs_constraints_mutex);
6463 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
6467 period = ktime_to_ns(tg->cfs_bandwidth.period);
6468 if (cfs_quota_us < 0)
6469 quota = RUNTIME_INF;
6471 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
6473 return tg_set_cfs_bandwidth(tg, period, quota);
6476 long tg_get_cfs_quota(struct task_group *tg)
6480 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
6483 quota_us = tg->cfs_bandwidth.quota;
6484 do_div(quota_us, NSEC_PER_USEC);
6489 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
6493 period = (u64)cfs_period_us * NSEC_PER_USEC;
6494 quota = tg->cfs_bandwidth.quota;
6496 return tg_set_cfs_bandwidth(tg, period, quota);
6499 long tg_get_cfs_period(struct task_group *tg)
6503 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
6504 do_div(cfs_period_us, NSEC_PER_USEC);
6506 return cfs_period_us;
6509 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
6512 return tg_get_cfs_quota(css_tg(css));
6515 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
6516 struct cftype *cftype, s64 cfs_quota_us)
6518 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
6521 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
6524 return tg_get_cfs_period(css_tg(css));
6527 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
6528 struct cftype *cftype, u64 cfs_period_us)
6530 return tg_set_cfs_period(css_tg(css), cfs_period_us);
6533 struct cfs_schedulable_data {
6534 struct task_group *tg;
6539 * normalize group quota/period to be quota/max_period
6540 * note: units are usecs
6542 static u64 normalize_cfs_quota(struct task_group *tg,
6543 struct cfs_schedulable_data *d)
6551 period = tg_get_cfs_period(tg);
6552 quota = tg_get_cfs_quota(tg);
6555 /* note: these should typically be equivalent */
6556 if (quota == RUNTIME_INF || quota == -1)
6559 return to_ratio(period, quota);
6562 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
6564 struct cfs_schedulable_data *d = data;
6565 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6566 s64 quota = 0, parent_quota = -1;
6569 quota = RUNTIME_INF;
6571 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
6573 quota = normalize_cfs_quota(tg, d);
6574 parent_quota = parent_b->hierarchical_quota;
6577 * Ensure max(child_quota) <= parent_quota, inherit when no
6580 if (quota == RUNTIME_INF)
6581 quota = parent_quota;
6582 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
6585 cfs_b->hierarchical_quota = quota;
6590 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
6593 struct cfs_schedulable_data data = {
6599 if (quota != RUNTIME_INF) {
6600 do_div(data.period, NSEC_PER_USEC);
6601 do_div(data.quota, NSEC_PER_USEC);
6605 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
6611 static int cpu_stats_show(struct seq_file *sf, void *v)
6613 struct task_group *tg = css_tg(seq_css(sf));
6614 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6616 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
6617 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
6618 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
6622 #endif /* CONFIG_CFS_BANDWIDTH */
6623 #endif /* CONFIG_FAIR_GROUP_SCHED */
6625 #ifdef CONFIG_RT_GROUP_SCHED
6626 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
6627 struct cftype *cft, s64 val)
6629 return sched_group_set_rt_runtime(css_tg(css), val);
6632 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
6635 return sched_group_rt_runtime(css_tg(css));
6638 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
6639 struct cftype *cftype, u64 rt_period_us)
6641 return sched_group_set_rt_period(css_tg(css), rt_period_us);
6644 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
6647 return sched_group_rt_period(css_tg(css));
6649 #endif /* CONFIG_RT_GROUP_SCHED */
6651 static struct cftype cpu_files[] = {
6652 #ifdef CONFIG_FAIR_GROUP_SCHED
6655 .read_u64 = cpu_shares_read_u64,
6656 .write_u64 = cpu_shares_write_u64,
6659 #ifdef CONFIG_CFS_BANDWIDTH
6661 .name = "cfs_quota_us",
6662 .read_s64 = cpu_cfs_quota_read_s64,
6663 .write_s64 = cpu_cfs_quota_write_s64,
6666 .name = "cfs_period_us",
6667 .read_u64 = cpu_cfs_period_read_u64,
6668 .write_u64 = cpu_cfs_period_write_u64,
6672 .seq_show = cpu_stats_show,
6675 #ifdef CONFIG_RT_GROUP_SCHED
6677 .name = "rt_runtime_us",
6678 .read_s64 = cpu_rt_runtime_read,
6679 .write_s64 = cpu_rt_runtime_write,
6682 .name = "rt_period_us",
6683 .read_u64 = cpu_rt_period_read_uint,
6684 .write_u64 = cpu_rt_period_write_uint,
6690 struct cgroup_subsys cpu_cgrp_subsys = {
6691 .css_alloc = cpu_cgroup_css_alloc,
6692 .css_online = cpu_cgroup_css_online,
6693 .css_released = cpu_cgroup_css_released,
6694 .css_free = cpu_cgroup_css_free,
6695 .fork = cpu_cgroup_fork,
6696 .can_attach = cpu_cgroup_can_attach,
6697 .attach = cpu_cgroup_attach,
6698 .legacy_cftypes = cpu_files,
6702 #endif /* CONFIG_CGROUP_SCHED */
6704 void dump_cpu_task(int cpu)
6706 pr_info("Task dump for CPU %d:\n", cpu);
6707 sched_show_task(cpu_curr(cpu));
6711 * Nice levels are multiplicative, with a gentle 10% change for every
6712 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
6713 * nice 1, it will get ~10% less CPU time than another CPU-bound task
6714 * that remained on nice 0.
6716 * The "10% effect" is relative and cumulative: from _any_ nice level,
6717 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
6718 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
6719 * If a task goes up by ~10% and another task goes down by ~10% then
6720 * the relative distance between them is ~25%.)
6722 const int sched_prio_to_weight[40] = {
6723 /* -20 */ 88761, 71755, 56483, 46273, 36291,
6724 /* -15 */ 29154, 23254, 18705, 14949, 11916,
6725 /* -10 */ 9548, 7620, 6100, 4904, 3906,
6726 /* -5 */ 3121, 2501, 1991, 1586, 1277,
6727 /* 0 */ 1024, 820, 655, 526, 423,
6728 /* 5 */ 335, 272, 215, 172, 137,
6729 /* 10 */ 110, 87, 70, 56, 45,
6730 /* 15 */ 36, 29, 23, 18, 15,
6734 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
6736 * In cases where the weight does not change often, we can use the
6737 * precalculated inverse to speed up arithmetics by turning divisions
6738 * into multiplications:
6740 const u32 sched_prio_to_wmult[40] = {
6741 /* -20 */ 48388, 59856, 76040, 92818, 118348,
6742 /* -15 */ 147320, 184698, 229616, 287308, 360437,
6743 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
6744 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
6745 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
6746 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
6747 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
6748 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,