4 * Core kernel scheduler code and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 #include <linux/sched.h>
9 #include <linux/cpuset.h>
10 #include <linux/delayacct.h>
11 #include <linux/init_task.h>
12 #include <linux/context_tracking.h>
14 #include <linux/blkdev.h>
15 #include <linux/kprobes.h>
16 #include <linux/mmu_context.h>
17 #include <linux/module.h>
18 #include <linux/nmi.h>
19 #include <linux/prefetch.h>
20 #include <linux/profile.h>
21 #include <linux/security.h>
22 #include <linux/syscalls.h>
24 #include <asm/switch_to.h>
26 #ifdef CONFIG_PARAVIRT
27 #include <asm/paravirt.h>
31 #include "../workqueue_internal.h"
32 #include "../smpboot.h"
34 #define CREATE_TRACE_POINTS
35 #include <trace/events/sched.h>
37 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
40 * Debugging: various feature bits
43 #define SCHED_FEAT(name, enabled) \
44 (1UL << __SCHED_FEAT_##name) * enabled |
46 const_debug unsigned int sysctl_sched_features =
53 * Number of tasks to iterate in a single balance run.
54 * Limited because this is done with IRQs disabled.
56 const_debug unsigned int sysctl_sched_nr_migrate = 32;
59 * period over which we average the RT time consumption, measured
64 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
67 * period over which we measure -rt task CPU usage in us.
70 unsigned int sysctl_sched_rt_period = 1000000;
72 __read_mostly int scheduler_running;
75 * part of the period that we allow rt tasks to run in us.
78 int sysctl_sched_rt_runtime = 950000;
80 /* CPUs with isolated domains */
81 cpumask_var_t cpu_isolated_map;
84 * this_rq_lock - lock this runqueue and disable interrupts.
86 static struct rq *this_rq_lock(void)
93 raw_spin_lock(&rq->lock);
99 * __task_rq_lock - lock the rq @p resides on.
101 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
106 lockdep_assert_held(&p->pi_lock);
110 raw_spin_lock(&rq->lock);
111 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
115 raw_spin_unlock(&rq->lock);
117 while (unlikely(task_on_rq_migrating(p)))
123 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
125 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
126 __acquires(p->pi_lock)
132 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
134 raw_spin_lock(&rq->lock);
136 * move_queued_task() task_rq_lock()
139 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
140 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
141 * [S] ->cpu = new_cpu [L] task_rq()
145 * If we observe the old cpu in task_rq_lock, the acquire of
146 * the old rq->lock will fully serialize against the stores.
148 * If we observe the new CPU in task_rq_lock, the acquire will
149 * pair with the WMB to ensure we must then also see migrating.
151 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
155 raw_spin_unlock(&rq->lock);
156 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
158 while (unlikely(task_on_rq_migrating(p)))
164 * RQ-clock updating methods:
167 static void update_rq_clock_task(struct rq *rq, s64 delta)
170 * In theory, the compile should just see 0 here, and optimize out the call
171 * to sched_rt_avg_update. But I don't trust it...
173 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
174 s64 steal = 0, irq_delta = 0;
176 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
177 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
180 * Since irq_time is only updated on {soft,}irq_exit, we might run into
181 * this case when a previous update_rq_clock() happened inside a
184 * When this happens, we stop ->clock_task and only update the
185 * prev_irq_time stamp to account for the part that fit, so that a next
186 * update will consume the rest. This ensures ->clock_task is
189 * It does however cause some slight miss-attribution of {soft,}irq
190 * time, a more accurate solution would be to update the irq_time using
191 * the current rq->clock timestamp, except that would require using
194 if (irq_delta > delta)
197 rq->prev_irq_time += irq_delta;
200 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
201 if (static_key_false((¶virt_steal_rq_enabled))) {
202 steal = paravirt_steal_clock(cpu_of(rq));
203 steal -= rq->prev_steal_time_rq;
205 if (unlikely(steal > delta))
208 rq->prev_steal_time_rq += steal;
213 rq->clock_task += delta;
215 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
216 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
217 sched_rt_avg_update(rq, irq_delta + steal);
221 void update_rq_clock(struct rq *rq)
225 lockdep_assert_held(&rq->lock);
227 if (rq->clock_update_flags & RQCF_ACT_SKIP)
230 #ifdef CONFIG_SCHED_DEBUG
231 rq->clock_update_flags |= RQCF_UPDATED;
233 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
237 update_rq_clock_task(rq, delta);
241 #ifdef CONFIG_SCHED_HRTICK
243 * Use HR-timers to deliver accurate preemption points.
246 static void hrtick_clear(struct rq *rq)
248 if (hrtimer_active(&rq->hrtick_timer))
249 hrtimer_cancel(&rq->hrtick_timer);
253 * High-resolution timer tick.
254 * Runs from hardirq context with interrupts disabled.
256 static enum hrtimer_restart hrtick(struct hrtimer *timer)
258 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
260 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
262 raw_spin_lock(&rq->lock);
264 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
265 raw_spin_unlock(&rq->lock);
267 return HRTIMER_NORESTART;
272 static void __hrtick_restart(struct rq *rq)
274 struct hrtimer *timer = &rq->hrtick_timer;
276 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
280 * called from hardirq (IPI) context
282 static void __hrtick_start(void *arg)
286 raw_spin_lock(&rq->lock);
287 __hrtick_restart(rq);
288 rq->hrtick_csd_pending = 0;
289 raw_spin_unlock(&rq->lock);
293 * Called to set the hrtick timer state.
295 * called with rq->lock held and irqs disabled
297 void hrtick_start(struct rq *rq, u64 delay)
299 struct hrtimer *timer = &rq->hrtick_timer;
304 * Don't schedule slices shorter than 10000ns, that just
305 * doesn't make sense and can cause timer DoS.
307 delta = max_t(s64, delay, 10000LL);
308 time = ktime_add_ns(timer->base->get_time(), delta);
310 hrtimer_set_expires(timer, time);
312 if (rq == this_rq()) {
313 __hrtick_restart(rq);
314 } else if (!rq->hrtick_csd_pending) {
315 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
316 rq->hrtick_csd_pending = 1;
322 * Called to set the hrtick timer state.
324 * called with rq->lock held and irqs disabled
326 void hrtick_start(struct rq *rq, u64 delay)
329 * Don't schedule slices shorter than 10000ns, that just
330 * doesn't make sense. Rely on vruntime for fairness.
332 delay = max_t(u64, delay, 10000LL);
333 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
334 HRTIMER_MODE_REL_PINNED);
336 #endif /* CONFIG_SMP */
338 static void init_rq_hrtick(struct rq *rq)
341 rq->hrtick_csd_pending = 0;
343 rq->hrtick_csd.flags = 0;
344 rq->hrtick_csd.func = __hrtick_start;
345 rq->hrtick_csd.info = rq;
348 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
349 rq->hrtick_timer.function = hrtick;
351 #else /* CONFIG_SCHED_HRTICK */
352 static inline void hrtick_clear(struct rq *rq)
356 static inline void init_rq_hrtick(struct rq *rq)
359 #endif /* CONFIG_SCHED_HRTICK */
362 * cmpxchg based fetch_or, macro so it works for different integer types
364 #define fetch_or(ptr, mask) \
366 typeof(ptr) _ptr = (ptr); \
367 typeof(mask) _mask = (mask); \
368 typeof(*_ptr) _old, _val = *_ptr; \
371 _old = cmpxchg(_ptr, _val, _val | _mask); \
379 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
381 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
382 * this avoids any races wrt polling state changes and thereby avoids
385 static bool set_nr_and_not_polling(struct task_struct *p)
387 struct thread_info *ti = task_thread_info(p);
388 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
392 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
394 * If this returns true, then the idle task promises to call
395 * sched_ttwu_pending() and reschedule soon.
397 static bool set_nr_if_polling(struct task_struct *p)
399 struct thread_info *ti = task_thread_info(p);
400 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
403 if (!(val & _TIF_POLLING_NRFLAG))
405 if (val & _TIF_NEED_RESCHED)
407 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
416 static bool set_nr_and_not_polling(struct task_struct *p)
418 set_tsk_need_resched(p);
423 static bool set_nr_if_polling(struct task_struct *p)
430 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
432 struct wake_q_node *node = &task->wake_q;
435 * Atomically grab the task, if ->wake_q is !nil already it means
436 * its already queued (either by us or someone else) and will get the
437 * wakeup due to that.
439 * This cmpxchg() implies a full barrier, which pairs with the write
440 * barrier implied by the wakeup in wake_up_q().
442 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
445 get_task_struct(task);
448 * The head is context local, there can be no concurrency.
451 head->lastp = &node->next;
454 void wake_up_q(struct wake_q_head *head)
456 struct wake_q_node *node = head->first;
458 while (node != WAKE_Q_TAIL) {
459 struct task_struct *task;
461 task = container_of(node, struct task_struct, wake_q);
463 /* Task can safely be re-inserted now: */
465 task->wake_q.next = NULL;
468 * wake_up_process() implies a wmb() to pair with the queueing
469 * in wake_q_add() so as not to miss wakeups.
471 wake_up_process(task);
472 put_task_struct(task);
477 * resched_curr - mark rq's current task 'to be rescheduled now'.
479 * On UP this means the setting of the need_resched flag, on SMP it
480 * might also involve a cross-CPU call to trigger the scheduler on
483 void resched_curr(struct rq *rq)
485 struct task_struct *curr = rq->curr;
488 lockdep_assert_held(&rq->lock);
490 if (test_tsk_need_resched(curr))
495 if (cpu == smp_processor_id()) {
496 set_tsk_need_resched(curr);
497 set_preempt_need_resched();
501 if (set_nr_and_not_polling(curr))
502 smp_send_reschedule(cpu);
504 trace_sched_wake_idle_without_ipi(cpu);
507 void resched_cpu(int cpu)
509 struct rq *rq = cpu_rq(cpu);
512 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
515 raw_spin_unlock_irqrestore(&rq->lock, flags);
519 #ifdef CONFIG_NO_HZ_COMMON
521 * In the semi idle case, use the nearest busy CPU for migrating timers
522 * from an idle CPU. This is good for power-savings.
524 * We don't do similar optimization for completely idle system, as
525 * selecting an idle CPU will add more delays to the timers than intended
526 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
528 int get_nohz_timer_target(void)
530 int i, cpu = smp_processor_id();
531 struct sched_domain *sd;
533 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
537 for_each_domain(cpu, sd) {
538 for_each_cpu(i, sched_domain_span(sd)) {
542 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
549 if (!is_housekeeping_cpu(cpu))
550 cpu = housekeeping_any_cpu();
557 * When add_timer_on() enqueues a timer into the timer wheel of an
558 * idle CPU then this timer might expire before the next timer event
559 * which is scheduled to wake up that CPU. In case of a completely
560 * idle system the next event might even be infinite time into the
561 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
562 * leaves the inner idle loop so the newly added timer is taken into
563 * account when the CPU goes back to idle and evaluates the timer
564 * wheel for the next timer event.
566 static void wake_up_idle_cpu(int cpu)
568 struct rq *rq = cpu_rq(cpu);
570 if (cpu == smp_processor_id())
573 if (set_nr_and_not_polling(rq->idle))
574 smp_send_reschedule(cpu);
576 trace_sched_wake_idle_without_ipi(cpu);
579 static bool wake_up_full_nohz_cpu(int cpu)
582 * We just need the target to call irq_exit() and re-evaluate
583 * the next tick. The nohz full kick at least implies that.
584 * If needed we can still optimize that later with an
587 if (cpu_is_offline(cpu))
588 return true; /* Don't try to wake offline CPUs. */
589 if (tick_nohz_full_cpu(cpu)) {
590 if (cpu != smp_processor_id() ||
591 tick_nohz_tick_stopped())
592 tick_nohz_full_kick_cpu(cpu);
600 * Wake up the specified CPU. If the CPU is going offline, it is the
601 * caller's responsibility to deal with the lost wakeup, for example,
602 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
604 void wake_up_nohz_cpu(int cpu)
606 if (!wake_up_full_nohz_cpu(cpu))
607 wake_up_idle_cpu(cpu);
610 static inline bool got_nohz_idle_kick(void)
612 int cpu = smp_processor_id();
614 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
617 if (idle_cpu(cpu) && !need_resched())
621 * We can't run Idle Load Balance on this CPU for this time so we
622 * cancel it and clear NOHZ_BALANCE_KICK
624 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
628 #else /* CONFIG_NO_HZ_COMMON */
630 static inline bool got_nohz_idle_kick(void)
635 #endif /* CONFIG_NO_HZ_COMMON */
637 #ifdef CONFIG_NO_HZ_FULL
638 bool sched_can_stop_tick(struct rq *rq)
642 /* Deadline tasks, even if single, need the tick */
643 if (rq->dl.dl_nr_running)
647 * If there are more than one RR tasks, we need the tick to effect the
648 * actual RR behaviour.
650 if (rq->rt.rr_nr_running) {
651 if (rq->rt.rr_nr_running == 1)
658 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
659 * forced preemption between FIFO tasks.
661 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
666 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
667 * if there's more than one we need the tick for involuntary
670 if (rq->nr_running > 1)
675 #endif /* CONFIG_NO_HZ_FULL */
677 void sched_avg_update(struct rq *rq)
679 s64 period = sched_avg_period();
681 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
683 * Inline assembly required to prevent the compiler
684 * optimising this loop into a divmod call.
685 * See __iter_div_u64_rem() for another example of this.
687 asm("" : "+rm" (rq->age_stamp));
688 rq->age_stamp += period;
693 #endif /* CONFIG_SMP */
695 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
696 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
698 * Iterate task_group tree rooted at *from, calling @down when first entering a
699 * node and @up when leaving it for the final time.
701 * Caller must hold rcu_lock or sufficient equivalent.
703 int walk_tg_tree_from(struct task_group *from,
704 tg_visitor down, tg_visitor up, void *data)
706 struct task_group *parent, *child;
712 ret = (*down)(parent, data);
715 list_for_each_entry_rcu(child, &parent->children, siblings) {
722 ret = (*up)(parent, data);
723 if (ret || parent == from)
727 parent = parent->parent;
734 int tg_nop(struct task_group *tg, void *data)
740 static void set_load_weight(struct task_struct *p)
742 int prio = p->static_prio - MAX_RT_PRIO;
743 struct load_weight *load = &p->se.load;
746 * SCHED_IDLE tasks get minimal weight:
748 if (idle_policy(p->policy)) {
749 load->weight = scale_load(WEIGHT_IDLEPRIO);
750 load->inv_weight = WMULT_IDLEPRIO;
754 load->weight = scale_load(sched_prio_to_weight[prio]);
755 load->inv_weight = sched_prio_to_wmult[prio];
758 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
761 if (!(flags & ENQUEUE_RESTORE))
762 sched_info_queued(rq, p);
763 p->sched_class->enqueue_task(rq, p, flags);
766 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
769 if (!(flags & DEQUEUE_SAVE))
770 sched_info_dequeued(rq, p);
771 p->sched_class->dequeue_task(rq, p, flags);
774 void activate_task(struct rq *rq, struct task_struct *p, int flags)
776 if (task_contributes_to_load(p))
777 rq->nr_uninterruptible--;
779 enqueue_task(rq, p, flags);
782 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
784 if (task_contributes_to_load(p))
785 rq->nr_uninterruptible++;
787 dequeue_task(rq, p, flags);
790 void sched_set_stop_task(int cpu, struct task_struct *stop)
792 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
793 struct task_struct *old_stop = cpu_rq(cpu)->stop;
797 * Make it appear like a SCHED_FIFO task, its something
798 * userspace knows about and won't get confused about.
800 * Also, it will make PI more or less work without too
801 * much confusion -- but then, stop work should not
802 * rely on PI working anyway.
804 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
806 stop->sched_class = &stop_sched_class;
809 cpu_rq(cpu)->stop = stop;
813 * Reset it back to a normal scheduling class so that
814 * it can die in pieces.
816 old_stop->sched_class = &rt_sched_class;
821 * __normal_prio - return the priority that is based on the static prio
823 static inline int __normal_prio(struct task_struct *p)
825 return p->static_prio;
829 * Calculate the expected normal priority: i.e. priority
830 * without taking RT-inheritance into account. Might be
831 * boosted by interactivity modifiers. Changes upon fork,
832 * setprio syscalls, and whenever the interactivity
833 * estimator recalculates.
835 static inline int normal_prio(struct task_struct *p)
839 if (task_has_dl_policy(p))
840 prio = MAX_DL_PRIO-1;
841 else if (task_has_rt_policy(p))
842 prio = MAX_RT_PRIO-1 - p->rt_priority;
844 prio = __normal_prio(p);
849 * Calculate the current priority, i.e. the priority
850 * taken into account by the scheduler. This value might
851 * be boosted by RT tasks, or might be boosted by
852 * interactivity modifiers. Will be RT if the task got
853 * RT-boosted. If not then it returns p->normal_prio.
855 static int effective_prio(struct task_struct *p)
857 p->normal_prio = normal_prio(p);
859 * If we are RT tasks or we were boosted to RT priority,
860 * keep the priority unchanged. Otherwise, update priority
861 * to the normal priority:
863 if (!rt_prio(p->prio))
864 return p->normal_prio;
869 * task_curr - is this task currently executing on a CPU?
870 * @p: the task in question.
872 * Return: 1 if the task is currently executing. 0 otherwise.
874 inline int task_curr(const struct task_struct *p)
876 return cpu_curr(task_cpu(p)) == p;
880 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
881 * use the balance_callback list if you want balancing.
883 * this means any call to check_class_changed() must be followed by a call to
884 * balance_callback().
886 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
887 const struct sched_class *prev_class,
890 if (prev_class != p->sched_class) {
891 if (prev_class->switched_from)
892 prev_class->switched_from(rq, p);
894 p->sched_class->switched_to(rq, p);
895 } else if (oldprio != p->prio || dl_task(p))
896 p->sched_class->prio_changed(rq, p, oldprio);
899 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
901 const struct sched_class *class;
903 if (p->sched_class == rq->curr->sched_class) {
904 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
906 for_each_class(class) {
907 if (class == rq->curr->sched_class)
909 if (class == p->sched_class) {
917 * A queue event has occurred, and we're going to schedule. In
918 * this case, we can save a useless back to back clock update.
920 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
921 rq_clock_skip_update(rq, true);
926 * This is how migration works:
928 * 1) we invoke migration_cpu_stop() on the target CPU using
930 * 2) stopper starts to run (implicitly forcing the migrated thread
932 * 3) it checks whether the migrated task is still in the wrong runqueue.
933 * 4) if it's in the wrong runqueue then the migration thread removes
934 * it and puts it into the right queue.
935 * 5) stopper completes and stop_one_cpu() returns and the migration
940 * move_queued_task - move a queued task to new rq.
942 * Returns (locked) new rq. Old rq's lock is released.
944 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
946 lockdep_assert_held(&rq->lock);
948 p->on_rq = TASK_ON_RQ_MIGRATING;
949 dequeue_task(rq, p, 0);
950 set_task_cpu(p, new_cpu);
951 raw_spin_unlock(&rq->lock);
953 rq = cpu_rq(new_cpu);
955 raw_spin_lock(&rq->lock);
956 BUG_ON(task_cpu(p) != new_cpu);
957 enqueue_task(rq, p, 0);
958 p->on_rq = TASK_ON_RQ_QUEUED;
959 check_preempt_curr(rq, p, 0);
964 struct migration_arg {
965 struct task_struct *task;
970 * Move (not current) task off this CPU, onto the destination CPU. We're doing
971 * this because either it can't run here any more (set_cpus_allowed()
972 * away from this CPU, or CPU going down), or because we're
973 * attempting to rebalance this task on exec (sched_exec).
975 * So we race with normal scheduler movements, but that's OK, as long
976 * as the task is no longer on this CPU.
978 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
980 if (unlikely(!cpu_active(dest_cpu)))
983 /* Affinity changed (again). */
984 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
987 rq = move_queued_task(rq, p, dest_cpu);
993 * migration_cpu_stop - this will be executed by a highprio stopper thread
994 * and performs thread migration by bumping thread off CPU then
995 * 'pushing' onto another runqueue.
997 static int migration_cpu_stop(void *data)
999 struct migration_arg *arg = data;
1000 struct task_struct *p = arg->task;
1001 struct rq *rq = this_rq();
1004 * The original target CPU might have gone down and we might
1005 * be on another CPU but it doesn't matter.
1007 local_irq_disable();
1009 * We need to explicitly wake pending tasks before running
1010 * __migrate_task() such that we will not miss enforcing cpus_allowed
1011 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1013 sched_ttwu_pending();
1015 raw_spin_lock(&p->pi_lock);
1016 raw_spin_lock(&rq->lock);
1018 * If task_rq(p) != rq, it cannot be migrated here, because we're
1019 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1020 * we're holding p->pi_lock.
1022 if (task_rq(p) == rq) {
1023 if (task_on_rq_queued(p))
1024 rq = __migrate_task(rq, p, arg->dest_cpu);
1026 p->wake_cpu = arg->dest_cpu;
1028 raw_spin_unlock(&rq->lock);
1029 raw_spin_unlock(&p->pi_lock);
1036 * sched_class::set_cpus_allowed must do the below, but is not required to
1037 * actually call this function.
1039 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1041 cpumask_copy(&p->cpus_allowed, new_mask);
1042 p->nr_cpus_allowed = cpumask_weight(new_mask);
1045 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1047 struct rq *rq = task_rq(p);
1048 bool queued, running;
1050 lockdep_assert_held(&p->pi_lock);
1052 queued = task_on_rq_queued(p);
1053 running = task_current(rq, p);
1057 * Because __kthread_bind() calls this on blocked tasks without
1060 lockdep_assert_held(&rq->lock);
1061 dequeue_task(rq, p, DEQUEUE_SAVE);
1064 put_prev_task(rq, p);
1066 p->sched_class->set_cpus_allowed(p, new_mask);
1069 enqueue_task(rq, p, ENQUEUE_RESTORE);
1071 set_curr_task(rq, p);
1075 * Change a given task's CPU affinity. Migrate the thread to a
1076 * proper CPU and schedule it away if the CPU it's executing on
1077 * is removed from the allowed bitmask.
1079 * NOTE: the caller must have a valid reference to the task, the
1080 * task must not exit() & deallocate itself prematurely. The
1081 * call is not atomic; no spinlocks may be held.
1083 static int __set_cpus_allowed_ptr(struct task_struct *p,
1084 const struct cpumask *new_mask, bool check)
1086 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1087 unsigned int dest_cpu;
1092 rq = task_rq_lock(p, &rf);
1094 if (p->flags & PF_KTHREAD) {
1096 * Kernel threads are allowed on online && !active CPUs
1098 cpu_valid_mask = cpu_online_mask;
1102 * Must re-check here, to close a race against __kthread_bind(),
1103 * sched_setaffinity() is not guaranteed to observe the flag.
1105 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1110 if (cpumask_equal(&p->cpus_allowed, new_mask))
1113 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1118 do_set_cpus_allowed(p, new_mask);
1120 if (p->flags & PF_KTHREAD) {
1122 * For kernel threads that do indeed end up on online &&
1123 * !active we want to ensure they are strict per-CPU threads.
1125 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1126 !cpumask_intersects(new_mask, cpu_active_mask) &&
1127 p->nr_cpus_allowed != 1);
1130 /* Can the task run on the task's current CPU? If so, we're done */
1131 if (cpumask_test_cpu(task_cpu(p), new_mask))
1134 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1135 if (task_running(rq, p) || p->state == TASK_WAKING) {
1136 struct migration_arg arg = { p, dest_cpu };
1137 /* Need help from migration thread: drop lock and wait. */
1138 task_rq_unlock(rq, p, &rf);
1139 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1140 tlb_migrate_finish(p->mm);
1142 } else if (task_on_rq_queued(p)) {
1144 * OK, since we're going to drop the lock immediately
1145 * afterwards anyway.
1147 rq_unpin_lock(rq, &rf);
1148 rq = move_queued_task(rq, p, dest_cpu);
1149 rq_repin_lock(rq, &rf);
1152 task_rq_unlock(rq, p, &rf);
1157 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1159 return __set_cpus_allowed_ptr(p, new_mask, false);
1161 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1163 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1165 #ifdef CONFIG_SCHED_DEBUG
1167 * We should never call set_task_cpu() on a blocked task,
1168 * ttwu() will sort out the placement.
1170 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1174 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1175 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1176 * time relying on p->on_rq.
1178 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1179 p->sched_class == &fair_sched_class &&
1180 (p->on_rq && !task_on_rq_migrating(p)));
1182 #ifdef CONFIG_LOCKDEP
1184 * The caller should hold either p->pi_lock or rq->lock, when changing
1185 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1187 * sched_move_task() holds both and thus holding either pins the cgroup,
1190 * Furthermore, all task_rq users should acquire both locks, see
1193 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1194 lockdep_is_held(&task_rq(p)->lock)));
1198 trace_sched_migrate_task(p, new_cpu);
1200 if (task_cpu(p) != new_cpu) {
1201 if (p->sched_class->migrate_task_rq)
1202 p->sched_class->migrate_task_rq(p);
1203 p->se.nr_migrations++;
1204 perf_event_task_migrate(p);
1207 __set_task_cpu(p, new_cpu);
1210 static void __migrate_swap_task(struct task_struct *p, int cpu)
1212 if (task_on_rq_queued(p)) {
1213 struct rq *src_rq, *dst_rq;
1215 src_rq = task_rq(p);
1216 dst_rq = cpu_rq(cpu);
1218 p->on_rq = TASK_ON_RQ_MIGRATING;
1219 deactivate_task(src_rq, p, 0);
1220 set_task_cpu(p, cpu);
1221 activate_task(dst_rq, p, 0);
1222 p->on_rq = TASK_ON_RQ_QUEUED;
1223 check_preempt_curr(dst_rq, p, 0);
1226 * Task isn't running anymore; make it appear like we migrated
1227 * it before it went to sleep. This means on wakeup we make the
1228 * previous CPU our target instead of where it really is.
1234 struct migration_swap_arg {
1235 struct task_struct *src_task, *dst_task;
1236 int src_cpu, dst_cpu;
1239 static int migrate_swap_stop(void *data)
1241 struct migration_swap_arg *arg = data;
1242 struct rq *src_rq, *dst_rq;
1245 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1248 src_rq = cpu_rq(arg->src_cpu);
1249 dst_rq = cpu_rq(arg->dst_cpu);
1251 double_raw_lock(&arg->src_task->pi_lock,
1252 &arg->dst_task->pi_lock);
1253 double_rq_lock(src_rq, dst_rq);
1255 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1258 if (task_cpu(arg->src_task) != arg->src_cpu)
1261 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1264 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1267 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1268 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1273 double_rq_unlock(src_rq, dst_rq);
1274 raw_spin_unlock(&arg->dst_task->pi_lock);
1275 raw_spin_unlock(&arg->src_task->pi_lock);
1281 * Cross migrate two tasks
1283 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1285 struct migration_swap_arg arg;
1288 arg = (struct migration_swap_arg){
1290 .src_cpu = task_cpu(cur),
1292 .dst_cpu = task_cpu(p),
1295 if (arg.src_cpu == arg.dst_cpu)
1299 * These three tests are all lockless; this is OK since all of them
1300 * will be re-checked with proper locks held further down the line.
1302 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1305 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1308 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1311 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1312 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1319 * wait_task_inactive - wait for a thread to unschedule.
1321 * If @match_state is nonzero, it's the @p->state value just checked and
1322 * not expected to change. If it changes, i.e. @p might have woken up,
1323 * then return zero. When we succeed in waiting for @p to be off its CPU,
1324 * we return a positive number (its total switch count). If a second call
1325 * a short while later returns the same number, the caller can be sure that
1326 * @p has remained unscheduled the whole time.
1328 * The caller must ensure that the task *will* unschedule sometime soon,
1329 * else this function might spin for a *long* time. This function can't
1330 * be called with interrupts off, or it may introduce deadlock with
1331 * smp_call_function() if an IPI is sent by the same process we are
1332 * waiting to become inactive.
1334 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1336 int running, queued;
1343 * We do the initial early heuristics without holding
1344 * any task-queue locks at all. We'll only try to get
1345 * the runqueue lock when things look like they will
1351 * If the task is actively running on another CPU
1352 * still, just relax and busy-wait without holding
1355 * NOTE! Since we don't hold any locks, it's not
1356 * even sure that "rq" stays as the right runqueue!
1357 * But we don't care, since "task_running()" will
1358 * return false if the runqueue has changed and p
1359 * is actually now running somewhere else!
1361 while (task_running(rq, p)) {
1362 if (match_state && unlikely(p->state != match_state))
1368 * Ok, time to look more closely! We need the rq
1369 * lock now, to be *sure*. If we're wrong, we'll
1370 * just go back and repeat.
1372 rq = task_rq_lock(p, &rf);
1373 trace_sched_wait_task(p);
1374 running = task_running(rq, p);
1375 queued = task_on_rq_queued(p);
1377 if (!match_state || p->state == match_state)
1378 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1379 task_rq_unlock(rq, p, &rf);
1382 * If it changed from the expected state, bail out now.
1384 if (unlikely(!ncsw))
1388 * Was it really running after all now that we
1389 * checked with the proper locks actually held?
1391 * Oops. Go back and try again..
1393 if (unlikely(running)) {
1399 * It's not enough that it's not actively running,
1400 * it must be off the runqueue _entirely_, and not
1403 * So if it was still runnable (but just not actively
1404 * running right now), it's preempted, and we should
1405 * yield - it could be a while.
1407 if (unlikely(queued)) {
1408 ktime_t to = NSEC_PER_SEC / HZ;
1410 set_current_state(TASK_UNINTERRUPTIBLE);
1411 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1416 * Ahh, all good. It wasn't running, and it wasn't
1417 * runnable, which means that it will never become
1418 * running in the future either. We're all done!
1427 * kick_process - kick a running thread to enter/exit the kernel
1428 * @p: the to-be-kicked thread
1430 * Cause a process which is running on another CPU to enter
1431 * kernel-mode, without any delay. (to get signals handled.)
1433 * NOTE: this function doesn't have to take the runqueue lock,
1434 * because all it wants to ensure is that the remote task enters
1435 * the kernel. If the IPI races and the task has been migrated
1436 * to another CPU then no harm is done and the purpose has been
1439 void kick_process(struct task_struct *p)
1445 if ((cpu != smp_processor_id()) && task_curr(p))
1446 smp_send_reschedule(cpu);
1449 EXPORT_SYMBOL_GPL(kick_process);
1452 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1454 * A few notes on cpu_active vs cpu_online:
1456 * - cpu_active must be a subset of cpu_online
1458 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1459 * see __set_cpus_allowed_ptr(). At this point the newly online
1460 * CPU isn't yet part of the sched domains, and balancing will not
1463 * - on CPU-down we clear cpu_active() to mask the sched domains and
1464 * avoid the load balancer to place new tasks on the to be removed
1465 * CPU. Existing tasks will remain running there and will be taken
1468 * This means that fallback selection must not select !active CPUs.
1469 * And can assume that any active CPU must be online. Conversely
1470 * select_task_rq() below may allow selection of !active CPUs in order
1471 * to satisfy the above rules.
1473 static int select_fallback_rq(int cpu, struct task_struct *p)
1475 int nid = cpu_to_node(cpu);
1476 const struct cpumask *nodemask = NULL;
1477 enum { cpuset, possible, fail } state = cpuset;
1481 * If the node that the CPU is on has been offlined, cpu_to_node()
1482 * will return -1. There is no CPU on the node, and we should
1483 * select the CPU on the other node.
1486 nodemask = cpumask_of_node(nid);
1488 /* Look for allowed, online CPU in same node. */
1489 for_each_cpu(dest_cpu, nodemask) {
1490 if (!cpu_active(dest_cpu))
1492 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1498 /* Any allowed, online CPU? */
1499 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1500 if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1502 if (!cpu_online(dest_cpu))
1507 /* No more Mr. Nice Guy. */
1510 if (IS_ENABLED(CONFIG_CPUSETS)) {
1511 cpuset_cpus_allowed_fallback(p);
1517 do_set_cpus_allowed(p, cpu_possible_mask);
1528 if (state != cpuset) {
1530 * Don't tell them about moving exiting tasks or
1531 * kernel threads (both mm NULL), since they never
1534 if (p->mm && printk_ratelimit()) {
1535 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1536 task_pid_nr(p), p->comm, cpu);
1544 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1547 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1549 lockdep_assert_held(&p->pi_lock);
1551 if (tsk_nr_cpus_allowed(p) > 1)
1552 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1554 cpu = cpumask_any(tsk_cpus_allowed(p));
1557 * In order not to call set_task_cpu() on a blocking task we need
1558 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1561 * Since this is common to all placement strategies, this lives here.
1563 * [ this allows ->select_task() to simply return task_cpu(p) and
1564 * not worry about this generic constraint ]
1566 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1568 cpu = select_fallback_rq(task_cpu(p), p);
1573 static void update_avg(u64 *avg, u64 sample)
1575 s64 diff = sample - *avg;
1581 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1582 const struct cpumask *new_mask, bool check)
1584 return set_cpus_allowed_ptr(p, new_mask);
1587 #endif /* CONFIG_SMP */
1590 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1594 if (!schedstat_enabled())
1600 if (cpu == rq->cpu) {
1601 schedstat_inc(rq->ttwu_local);
1602 schedstat_inc(p->se.statistics.nr_wakeups_local);
1604 struct sched_domain *sd;
1606 schedstat_inc(p->se.statistics.nr_wakeups_remote);
1608 for_each_domain(rq->cpu, sd) {
1609 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1610 schedstat_inc(sd->ttwu_wake_remote);
1617 if (wake_flags & WF_MIGRATED)
1618 schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1619 #endif /* CONFIG_SMP */
1621 schedstat_inc(rq->ttwu_count);
1622 schedstat_inc(p->se.statistics.nr_wakeups);
1624 if (wake_flags & WF_SYNC)
1625 schedstat_inc(p->se.statistics.nr_wakeups_sync);
1628 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1630 activate_task(rq, p, en_flags);
1631 p->on_rq = TASK_ON_RQ_QUEUED;
1633 /* If a worker is waking up, notify the workqueue: */
1634 if (p->flags & PF_WQ_WORKER)
1635 wq_worker_waking_up(p, cpu_of(rq));
1639 * Mark the task runnable and perform wakeup-preemption.
1641 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1642 struct rq_flags *rf)
1644 check_preempt_curr(rq, p, wake_flags);
1645 p->state = TASK_RUNNING;
1646 trace_sched_wakeup(p);
1649 if (p->sched_class->task_woken) {
1651 * Our task @p is fully woken up and running; so its safe to
1652 * drop the rq->lock, hereafter rq is only used for statistics.
1654 rq_unpin_lock(rq, rf);
1655 p->sched_class->task_woken(rq, p);
1656 rq_repin_lock(rq, rf);
1659 if (rq->idle_stamp) {
1660 u64 delta = rq_clock(rq) - rq->idle_stamp;
1661 u64 max = 2*rq->max_idle_balance_cost;
1663 update_avg(&rq->avg_idle, delta);
1665 if (rq->avg_idle > max)
1674 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1675 struct rq_flags *rf)
1677 int en_flags = ENQUEUE_WAKEUP;
1679 lockdep_assert_held(&rq->lock);
1682 if (p->sched_contributes_to_load)
1683 rq->nr_uninterruptible--;
1685 if (wake_flags & WF_MIGRATED)
1686 en_flags |= ENQUEUE_MIGRATED;
1689 ttwu_activate(rq, p, en_flags);
1690 ttwu_do_wakeup(rq, p, wake_flags, rf);
1694 * Called in case the task @p isn't fully descheduled from its runqueue,
1695 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1696 * since all we need to do is flip p->state to TASK_RUNNING, since
1697 * the task is still ->on_rq.
1699 static int ttwu_remote(struct task_struct *p, int wake_flags)
1705 rq = __task_rq_lock(p, &rf);
1706 if (task_on_rq_queued(p)) {
1707 /* check_preempt_curr() may use rq clock */
1708 update_rq_clock(rq);
1709 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1712 __task_rq_unlock(rq, &rf);
1718 void sched_ttwu_pending(void)
1720 struct rq *rq = this_rq();
1721 struct llist_node *llist = llist_del_all(&rq->wake_list);
1722 struct task_struct *p;
1723 unsigned long flags;
1729 raw_spin_lock_irqsave(&rq->lock, flags);
1730 rq_pin_lock(rq, &rf);
1735 p = llist_entry(llist, struct task_struct, wake_entry);
1736 llist = llist_next(llist);
1738 if (p->sched_remote_wakeup)
1739 wake_flags = WF_MIGRATED;
1741 ttwu_do_activate(rq, p, wake_flags, &rf);
1744 rq_unpin_lock(rq, &rf);
1745 raw_spin_unlock_irqrestore(&rq->lock, flags);
1748 void scheduler_ipi(void)
1751 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1752 * TIF_NEED_RESCHED remotely (for the first time) will also send
1755 preempt_fold_need_resched();
1757 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1761 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1762 * traditionally all their work was done from the interrupt return
1763 * path. Now that we actually do some work, we need to make sure
1766 * Some archs already do call them, luckily irq_enter/exit nest
1769 * Arguably we should visit all archs and update all handlers,
1770 * however a fair share of IPIs are still resched only so this would
1771 * somewhat pessimize the simple resched case.
1774 sched_ttwu_pending();
1777 * Check if someone kicked us for doing the nohz idle load balance.
1779 if (unlikely(got_nohz_idle_kick())) {
1780 this_rq()->idle_balance = 1;
1781 raise_softirq_irqoff(SCHED_SOFTIRQ);
1786 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1788 struct rq *rq = cpu_rq(cpu);
1790 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1792 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1793 if (!set_nr_if_polling(rq->idle))
1794 smp_send_reschedule(cpu);
1796 trace_sched_wake_idle_without_ipi(cpu);
1800 void wake_up_if_idle(int cpu)
1802 struct rq *rq = cpu_rq(cpu);
1803 unsigned long flags;
1807 if (!is_idle_task(rcu_dereference(rq->curr)))
1810 if (set_nr_if_polling(rq->idle)) {
1811 trace_sched_wake_idle_without_ipi(cpu);
1813 raw_spin_lock_irqsave(&rq->lock, flags);
1814 if (is_idle_task(rq->curr))
1815 smp_send_reschedule(cpu);
1816 /* Else CPU is not idle, do nothing here: */
1817 raw_spin_unlock_irqrestore(&rq->lock, flags);
1824 bool cpus_share_cache(int this_cpu, int that_cpu)
1826 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1828 #endif /* CONFIG_SMP */
1830 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1832 struct rq *rq = cpu_rq(cpu);
1835 #if defined(CONFIG_SMP)
1836 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1837 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1838 ttwu_queue_remote(p, cpu, wake_flags);
1843 raw_spin_lock(&rq->lock);
1844 rq_pin_lock(rq, &rf);
1845 ttwu_do_activate(rq, p, wake_flags, &rf);
1846 rq_unpin_lock(rq, &rf);
1847 raw_spin_unlock(&rq->lock);
1851 * Notes on Program-Order guarantees on SMP systems.
1855 * The basic program-order guarantee on SMP systems is that when a task [t]
1856 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1857 * execution on its new CPU [c1].
1859 * For migration (of runnable tasks) this is provided by the following means:
1861 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1862 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1863 * rq(c1)->lock (if not at the same time, then in that order).
1864 * C) LOCK of the rq(c1)->lock scheduling in task
1866 * Transitivity guarantees that B happens after A and C after B.
1867 * Note: we only require RCpc transitivity.
1868 * Note: the CPU doing B need not be c0 or c1
1877 * UNLOCK rq(0)->lock
1879 * LOCK rq(0)->lock // orders against CPU0
1881 * UNLOCK rq(0)->lock
1885 * UNLOCK rq(1)->lock
1887 * LOCK rq(1)->lock // orders against CPU2
1890 * UNLOCK rq(1)->lock
1893 * BLOCKING -- aka. SLEEP + WAKEUP
1895 * For blocking we (obviously) need to provide the same guarantee as for
1896 * migration. However the means are completely different as there is no lock
1897 * chain to provide order. Instead we do:
1899 * 1) smp_store_release(X->on_cpu, 0)
1900 * 2) smp_cond_load_acquire(!X->on_cpu)
1904 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1906 * LOCK rq(0)->lock LOCK X->pi_lock
1909 * smp_store_release(X->on_cpu, 0);
1911 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1917 * X->state = RUNNING
1918 * UNLOCK rq(2)->lock
1920 * LOCK rq(2)->lock // orders against CPU1
1923 * UNLOCK rq(2)->lock
1926 * UNLOCK rq(0)->lock
1929 * However; for wakeups there is a second guarantee we must provide, namely we
1930 * must observe the state that lead to our wakeup. That is, not only must our
1931 * task observe its own prior state, it must also observe the stores prior to
1934 * This means that any means of doing remote wakeups must order the CPU doing
1935 * the wakeup against the CPU the task is going to end up running on. This,
1936 * however, is already required for the regular Program-Order guarantee above,
1937 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1942 * try_to_wake_up - wake up a thread
1943 * @p: the thread to be awakened
1944 * @state: the mask of task states that can be woken
1945 * @wake_flags: wake modifier flags (WF_*)
1947 * If (@state & @p->state) @p->state = TASK_RUNNING.
1949 * If the task was not queued/runnable, also place it back on a runqueue.
1951 * Atomic against schedule() which would dequeue a task, also see
1952 * set_current_state().
1954 * Return: %true if @p->state changes (an actual wakeup was done),
1958 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1960 unsigned long flags;
1961 int cpu, success = 0;
1964 * If we are going to wake up a thread waiting for CONDITION we
1965 * need to ensure that CONDITION=1 done by the caller can not be
1966 * reordered with p->state check below. This pairs with mb() in
1967 * set_current_state() the waiting thread does.
1969 smp_mb__before_spinlock();
1970 raw_spin_lock_irqsave(&p->pi_lock, flags);
1971 if (!(p->state & state))
1974 trace_sched_waking(p);
1976 /* We're going to change ->state: */
1981 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1982 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1983 * in smp_cond_load_acquire() below.
1985 * sched_ttwu_pending() try_to_wake_up()
1986 * [S] p->on_rq = 1; [L] P->state
1987 * UNLOCK rq->lock -----.
1991 * LOCK rq->lock -----'
1995 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
1997 * Pairs with the UNLOCK+LOCK on rq->lock from the
1998 * last wakeup of our task and the schedule that got our task
2002 if (p->on_rq && ttwu_remote(p, wake_flags))
2007 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2008 * possible to, falsely, observe p->on_cpu == 0.
2010 * One must be running (->on_cpu == 1) in order to remove oneself
2011 * from the runqueue.
2013 * [S] ->on_cpu = 1; [L] ->on_rq
2017 * [S] ->on_rq = 0; [L] ->on_cpu
2019 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2020 * from the consecutive calls to schedule(); the first switching to our
2021 * task, the second putting it to sleep.
2026 * If the owning (remote) CPU is still in the middle of schedule() with
2027 * this task as prev, wait until its done referencing the task.
2029 * Pairs with the smp_store_release() in finish_lock_switch().
2031 * This ensures that tasks getting woken will be fully ordered against
2032 * their previous state and preserve Program Order.
2034 smp_cond_load_acquire(&p->on_cpu, !VAL);
2036 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2037 p->state = TASK_WAKING;
2040 delayacct_blkio_end();
2041 atomic_dec(&task_rq(p)->nr_iowait);
2044 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2045 if (task_cpu(p) != cpu) {
2046 wake_flags |= WF_MIGRATED;
2047 set_task_cpu(p, cpu);
2050 #else /* CONFIG_SMP */
2053 delayacct_blkio_end();
2054 atomic_dec(&task_rq(p)->nr_iowait);
2057 #endif /* CONFIG_SMP */
2059 ttwu_queue(p, cpu, wake_flags);
2061 ttwu_stat(p, cpu, wake_flags);
2063 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2069 * try_to_wake_up_local - try to wake up a local task with rq lock held
2070 * @p: the thread to be awakened
2071 * @cookie: context's cookie for pinning
2073 * Put @p on the run-queue if it's not already there. The caller must
2074 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2077 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2079 struct rq *rq = task_rq(p);
2081 if (WARN_ON_ONCE(rq != this_rq()) ||
2082 WARN_ON_ONCE(p == current))
2085 lockdep_assert_held(&rq->lock);
2087 if (!raw_spin_trylock(&p->pi_lock)) {
2089 * This is OK, because current is on_cpu, which avoids it being
2090 * picked for load-balance and preemption/IRQs are still
2091 * disabled avoiding further scheduler activity on it and we've
2092 * not yet picked a replacement task.
2094 rq_unpin_lock(rq, rf);
2095 raw_spin_unlock(&rq->lock);
2096 raw_spin_lock(&p->pi_lock);
2097 raw_spin_lock(&rq->lock);
2098 rq_repin_lock(rq, rf);
2101 if (!(p->state & TASK_NORMAL))
2104 trace_sched_waking(p);
2106 if (!task_on_rq_queued(p)) {
2108 delayacct_blkio_end();
2109 atomic_dec(&rq->nr_iowait);
2111 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2114 ttwu_do_wakeup(rq, p, 0, rf);
2115 ttwu_stat(p, smp_processor_id(), 0);
2117 raw_spin_unlock(&p->pi_lock);
2121 * wake_up_process - Wake up a specific process
2122 * @p: The process to be woken up.
2124 * Attempt to wake up the nominated process and move it to the set of runnable
2127 * Return: 1 if the process was woken up, 0 if it was already running.
2129 * It may be assumed that this function implies a write memory barrier before
2130 * changing the task state if and only if any tasks are woken up.
2132 int wake_up_process(struct task_struct *p)
2134 return try_to_wake_up(p, TASK_NORMAL, 0);
2136 EXPORT_SYMBOL(wake_up_process);
2138 int wake_up_state(struct task_struct *p, unsigned int state)
2140 return try_to_wake_up(p, state, 0);
2144 * This function clears the sched_dl_entity static params.
2146 void __dl_clear_params(struct task_struct *p)
2148 struct sched_dl_entity *dl_se = &p->dl;
2150 dl_se->dl_runtime = 0;
2151 dl_se->dl_deadline = 0;
2152 dl_se->dl_period = 0;
2156 dl_se->dl_throttled = 0;
2157 dl_se->dl_yielded = 0;
2161 * Perform scheduler related setup for a newly forked process p.
2162 * p is forked by current.
2164 * __sched_fork() is basic setup used by init_idle() too:
2166 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2171 p->se.exec_start = 0;
2172 p->se.sum_exec_runtime = 0;
2173 p->se.prev_sum_exec_runtime = 0;
2174 p->se.nr_migrations = 0;
2176 INIT_LIST_HEAD(&p->se.group_node);
2178 #ifdef CONFIG_FAIR_GROUP_SCHED
2179 p->se.cfs_rq = NULL;
2182 #ifdef CONFIG_SCHEDSTATS
2183 /* Even if schedstat is disabled, there should not be garbage */
2184 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2187 RB_CLEAR_NODE(&p->dl.rb_node);
2188 init_dl_task_timer(&p->dl);
2189 __dl_clear_params(p);
2191 INIT_LIST_HEAD(&p->rt.run_list);
2193 p->rt.time_slice = sched_rr_timeslice;
2197 #ifdef CONFIG_PREEMPT_NOTIFIERS
2198 INIT_HLIST_HEAD(&p->preempt_notifiers);
2201 #ifdef CONFIG_NUMA_BALANCING
2202 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2203 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2204 p->mm->numa_scan_seq = 0;
2207 if (clone_flags & CLONE_VM)
2208 p->numa_preferred_nid = current->numa_preferred_nid;
2210 p->numa_preferred_nid = -1;
2212 p->node_stamp = 0ULL;
2213 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2214 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2215 p->numa_work.next = &p->numa_work;
2216 p->numa_faults = NULL;
2217 p->last_task_numa_placement = 0;
2218 p->last_sum_exec_runtime = 0;
2220 p->numa_group = NULL;
2221 #endif /* CONFIG_NUMA_BALANCING */
2224 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2226 #ifdef CONFIG_NUMA_BALANCING
2228 void set_numabalancing_state(bool enabled)
2231 static_branch_enable(&sched_numa_balancing);
2233 static_branch_disable(&sched_numa_balancing);
2236 #ifdef CONFIG_PROC_SYSCTL
2237 int sysctl_numa_balancing(struct ctl_table *table, int write,
2238 void __user *buffer, size_t *lenp, loff_t *ppos)
2242 int state = static_branch_likely(&sched_numa_balancing);
2244 if (write && !capable(CAP_SYS_ADMIN))
2249 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2253 set_numabalancing_state(state);
2259 #ifdef CONFIG_SCHEDSTATS
2261 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2262 static bool __initdata __sched_schedstats = false;
2264 static void set_schedstats(bool enabled)
2267 static_branch_enable(&sched_schedstats);
2269 static_branch_disable(&sched_schedstats);
2272 void force_schedstat_enabled(void)
2274 if (!schedstat_enabled()) {
2275 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2276 static_branch_enable(&sched_schedstats);
2280 static int __init setup_schedstats(char *str)
2287 * This code is called before jump labels have been set up, so we can't
2288 * change the static branch directly just yet. Instead set a temporary
2289 * variable so init_schedstats() can do it later.
2291 if (!strcmp(str, "enable")) {
2292 __sched_schedstats = true;
2294 } else if (!strcmp(str, "disable")) {
2295 __sched_schedstats = false;
2300 pr_warn("Unable to parse schedstats=\n");
2304 __setup("schedstats=", setup_schedstats);
2306 static void __init init_schedstats(void)
2308 set_schedstats(__sched_schedstats);
2311 #ifdef CONFIG_PROC_SYSCTL
2312 int sysctl_schedstats(struct ctl_table *table, int write,
2313 void __user *buffer, size_t *lenp, loff_t *ppos)
2317 int state = static_branch_likely(&sched_schedstats);
2319 if (write && !capable(CAP_SYS_ADMIN))
2324 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2328 set_schedstats(state);
2331 #endif /* CONFIG_PROC_SYSCTL */
2332 #else /* !CONFIG_SCHEDSTATS */
2333 static inline void init_schedstats(void) {}
2334 #endif /* CONFIG_SCHEDSTATS */
2337 * fork()/clone()-time setup:
2339 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2341 unsigned long flags;
2342 int cpu = get_cpu();
2344 __sched_fork(clone_flags, p);
2346 * We mark the process as NEW here. This guarantees that
2347 * nobody will actually run it, and a signal or other external
2348 * event cannot wake it up and insert it on the runqueue either.
2350 p->state = TASK_NEW;
2353 * Make sure we do not leak PI boosting priority to the child.
2355 p->prio = current->normal_prio;
2358 * Revert to default priority/policy on fork if requested.
2360 if (unlikely(p->sched_reset_on_fork)) {
2361 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2362 p->policy = SCHED_NORMAL;
2363 p->static_prio = NICE_TO_PRIO(0);
2365 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2366 p->static_prio = NICE_TO_PRIO(0);
2368 p->prio = p->normal_prio = __normal_prio(p);
2372 * We don't need the reset flag anymore after the fork. It has
2373 * fulfilled its duty:
2375 p->sched_reset_on_fork = 0;
2378 if (dl_prio(p->prio)) {
2381 } else if (rt_prio(p->prio)) {
2382 p->sched_class = &rt_sched_class;
2384 p->sched_class = &fair_sched_class;
2387 init_entity_runnable_average(&p->se);
2390 * The child is not yet in the pid-hash so no cgroup attach races,
2391 * and the cgroup is pinned to this child due to cgroup_fork()
2392 * is ran before sched_fork().
2394 * Silence PROVE_RCU.
2396 raw_spin_lock_irqsave(&p->pi_lock, flags);
2398 * We're setting the CPU for the first time, we don't migrate,
2399 * so use __set_task_cpu().
2401 __set_task_cpu(p, cpu);
2402 if (p->sched_class->task_fork)
2403 p->sched_class->task_fork(p);
2404 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2406 #ifdef CONFIG_SCHED_INFO
2407 if (likely(sched_info_on()))
2408 memset(&p->sched_info, 0, sizeof(p->sched_info));
2410 #if defined(CONFIG_SMP)
2413 init_task_preempt_count(p);
2415 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2416 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2423 unsigned long to_ratio(u64 period, u64 runtime)
2425 if (runtime == RUNTIME_INF)
2429 * Doing this here saves a lot of checks in all
2430 * the calling paths, and returning zero seems
2431 * safe for them anyway.
2436 return div64_u64(runtime << 20, period);
2440 inline struct dl_bw *dl_bw_of(int i)
2442 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2443 "sched RCU must be held");
2444 return &cpu_rq(i)->rd->dl_bw;
2447 static inline int dl_bw_cpus(int i)
2449 struct root_domain *rd = cpu_rq(i)->rd;
2452 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2453 "sched RCU must be held");
2454 for_each_cpu_and(i, rd->span, cpu_active_mask)
2460 inline struct dl_bw *dl_bw_of(int i)
2462 return &cpu_rq(i)->dl.dl_bw;
2465 static inline int dl_bw_cpus(int i)
2472 * We must be sure that accepting a new task (or allowing changing the
2473 * parameters of an existing one) is consistent with the bandwidth
2474 * constraints. If yes, this function also accordingly updates the currently
2475 * allocated bandwidth to reflect the new situation.
2477 * This function is called while holding p's rq->lock.
2479 * XXX we should delay bw change until the task's 0-lag point, see
2482 static int dl_overflow(struct task_struct *p, int policy,
2483 const struct sched_attr *attr)
2486 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2487 u64 period = attr->sched_period ?: attr->sched_deadline;
2488 u64 runtime = attr->sched_runtime;
2489 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2492 /* !deadline task may carry old deadline bandwidth */
2493 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2497 * Either if a task, enters, leave, or stays -deadline but changes
2498 * its parameters, we may need to update accordingly the total
2499 * allocated bandwidth of the container.
2501 raw_spin_lock(&dl_b->lock);
2502 cpus = dl_bw_cpus(task_cpu(p));
2503 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2504 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2505 __dl_add(dl_b, new_bw);
2507 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2508 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2509 __dl_clear(dl_b, p->dl.dl_bw);
2510 __dl_add(dl_b, new_bw);
2512 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2513 __dl_clear(dl_b, p->dl.dl_bw);
2516 raw_spin_unlock(&dl_b->lock);
2521 extern void init_dl_bw(struct dl_bw *dl_b);
2524 * wake_up_new_task - wake up a newly created task for the first time.
2526 * This function will do some initial scheduler statistics housekeeping
2527 * that must be done for every newly created context, then puts the task
2528 * on the runqueue and wakes it.
2530 void wake_up_new_task(struct task_struct *p)
2535 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2536 p->state = TASK_RUNNING;
2539 * Fork balancing, do it here and not earlier because:
2540 * - cpus_allowed can change in the fork path
2541 * - any previously selected CPU might disappear through hotplug
2543 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2544 * as we're not fully set-up yet.
2546 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2548 rq = __task_rq_lock(p, &rf);
2549 update_rq_clock(rq);
2550 post_init_entity_util_avg(&p->se);
2552 activate_task(rq, p, 0);
2553 p->on_rq = TASK_ON_RQ_QUEUED;
2554 trace_sched_wakeup_new(p);
2555 check_preempt_curr(rq, p, WF_FORK);
2557 if (p->sched_class->task_woken) {
2559 * Nothing relies on rq->lock after this, so its fine to
2562 rq_unpin_lock(rq, &rf);
2563 p->sched_class->task_woken(rq, p);
2564 rq_repin_lock(rq, &rf);
2567 task_rq_unlock(rq, p, &rf);
2570 #ifdef CONFIG_PREEMPT_NOTIFIERS
2572 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2574 void preempt_notifier_inc(void)
2576 static_key_slow_inc(&preempt_notifier_key);
2578 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2580 void preempt_notifier_dec(void)
2582 static_key_slow_dec(&preempt_notifier_key);
2584 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2587 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2588 * @notifier: notifier struct to register
2590 void preempt_notifier_register(struct preempt_notifier *notifier)
2592 if (!static_key_false(&preempt_notifier_key))
2593 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2595 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2597 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2600 * preempt_notifier_unregister - no longer interested in preemption notifications
2601 * @notifier: notifier struct to unregister
2603 * This is *not* safe to call from within a preemption notifier.
2605 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2607 hlist_del(¬ifier->link);
2609 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2611 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2613 struct preempt_notifier *notifier;
2615 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2616 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2619 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2621 if (static_key_false(&preempt_notifier_key))
2622 __fire_sched_in_preempt_notifiers(curr);
2626 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2627 struct task_struct *next)
2629 struct preempt_notifier *notifier;
2631 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2632 notifier->ops->sched_out(notifier, next);
2635 static __always_inline void
2636 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2637 struct task_struct *next)
2639 if (static_key_false(&preempt_notifier_key))
2640 __fire_sched_out_preempt_notifiers(curr, next);
2643 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2645 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2650 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2651 struct task_struct *next)
2655 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2658 * prepare_task_switch - prepare to switch tasks
2659 * @rq: the runqueue preparing to switch
2660 * @prev: the current task that is being switched out
2661 * @next: the task we are going to switch to.
2663 * This is called with the rq lock held and interrupts off. It must
2664 * be paired with a subsequent finish_task_switch after the context
2667 * prepare_task_switch sets up locking and calls architecture specific
2671 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2672 struct task_struct *next)
2674 sched_info_switch(rq, prev, next);
2675 perf_event_task_sched_out(prev, next);
2676 fire_sched_out_preempt_notifiers(prev, next);
2677 prepare_lock_switch(rq, next);
2678 prepare_arch_switch(next);
2682 * finish_task_switch - clean up after a task-switch
2683 * @prev: the thread we just switched away from.
2685 * finish_task_switch must be called after the context switch, paired
2686 * with a prepare_task_switch call before the context switch.
2687 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2688 * and do any other architecture-specific cleanup actions.
2690 * Note that we may have delayed dropping an mm in context_switch(). If
2691 * so, we finish that here outside of the runqueue lock. (Doing it
2692 * with the lock held can cause deadlocks; see schedule() for
2695 * The context switch have flipped the stack from under us and restored the
2696 * local variables which were saved when this task called schedule() in the
2697 * past. prev == current is still correct but we need to recalculate this_rq
2698 * because prev may have moved to another CPU.
2700 static struct rq *finish_task_switch(struct task_struct *prev)
2701 __releases(rq->lock)
2703 struct rq *rq = this_rq();
2704 struct mm_struct *mm = rq->prev_mm;
2708 * The previous task will have left us with a preempt_count of 2
2709 * because it left us after:
2712 * preempt_disable(); // 1
2714 * raw_spin_lock_irq(&rq->lock) // 2
2716 * Also, see FORK_PREEMPT_COUNT.
2718 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2719 "corrupted preempt_count: %s/%d/0x%x\n",
2720 current->comm, current->pid, preempt_count()))
2721 preempt_count_set(FORK_PREEMPT_COUNT);
2726 * A task struct has one reference for the use as "current".
2727 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2728 * schedule one last time. The schedule call will never return, and
2729 * the scheduled task must drop that reference.
2731 * We must observe prev->state before clearing prev->on_cpu (in
2732 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2733 * running on another CPU and we could rave with its RUNNING -> DEAD
2734 * transition, resulting in a double drop.
2736 prev_state = prev->state;
2737 vtime_task_switch(prev);
2738 perf_event_task_sched_in(prev, current);
2739 finish_lock_switch(rq, prev);
2740 finish_arch_post_lock_switch();
2742 fire_sched_in_preempt_notifiers(current);
2745 if (unlikely(prev_state == TASK_DEAD)) {
2746 if (prev->sched_class->task_dead)
2747 prev->sched_class->task_dead(prev);
2750 * Remove function-return probe instances associated with this
2751 * task and put them back on the free list.
2753 kprobe_flush_task(prev);
2755 /* Task is done with its stack. */
2756 put_task_stack(prev);
2758 put_task_struct(prev);
2761 tick_nohz_task_switch();
2767 /* rq->lock is NOT held, but preemption is disabled */
2768 static void __balance_callback(struct rq *rq)
2770 struct callback_head *head, *next;
2771 void (*func)(struct rq *rq);
2772 unsigned long flags;
2774 raw_spin_lock_irqsave(&rq->lock, flags);
2775 head = rq->balance_callback;
2776 rq->balance_callback = NULL;
2778 func = (void (*)(struct rq *))head->func;
2785 raw_spin_unlock_irqrestore(&rq->lock, flags);
2788 static inline void balance_callback(struct rq *rq)
2790 if (unlikely(rq->balance_callback))
2791 __balance_callback(rq);
2796 static inline void balance_callback(struct rq *rq)
2803 * schedule_tail - first thing a freshly forked thread must call.
2804 * @prev: the thread we just switched away from.
2806 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2807 __releases(rq->lock)
2812 * New tasks start with FORK_PREEMPT_COUNT, see there and
2813 * finish_task_switch() for details.
2815 * finish_task_switch() will drop rq->lock() and lower preempt_count
2816 * and the preempt_enable() will end up enabling preemption (on
2817 * PREEMPT_COUNT kernels).
2820 rq = finish_task_switch(prev);
2821 balance_callback(rq);
2824 if (current->set_child_tid)
2825 put_user(task_pid_vnr(current), current->set_child_tid);
2829 * context_switch - switch to the new MM and the new thread's register state.
2831 static __always_inline struct rq *
2832 context_switch(struct rq *rq, struct task_struct *prev,
2833 struct task_struct *next, struct rq_flags *rf)
2835 struct mm_struct *mm, *oldmm;
2837 prepare_task_switch(rq, prev, next);
2840 oldmm = prev->active_mm;
2842 * For paravirt, this is coupled with an exit in switch_to to
2843 * combine the page table reload and the switch backend into
2846 arch_start_context_switch(prev);
2849 next->active_mm = oldmm;
2850 atomic_inc(&oldmm->mm_count);
2851 enter_lazy_tlb(oldmm, next);
2853 switch_mm_irqs_off(oldmm, mm, next);
2856 prev->active_mm = NULL;
2857 rq->prev_mm = oldmm;
2860 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2863 * Since the runqueue lock will be released by the next
2864 * task (which is an invalid locking op but in the case
2865 * of the scheduler it's an obvious special-case), so we
2866 * do an early lockdep release here:
2868 rq_unpin_lock(rq, rf);
2869 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2871 /* Here we just switch the register state and the stack. */
2872 switch_to(prev, next, prev);
2875 return finish_task_switch(prev);
2879 * nr_running and nr_context_switches:
2881 * externally visible scheduler statistics: current number of runnable
2882 * threads, total number of context switches performed since bootup.
2884 unsigned long nr_running(void)
2886 unsigned long i, sum = 0;
2888 for_each_online_cpu(i)
2889 sum += cpu_rq(i)->nr_running;
2895 * Check if only the current task is running on the CPU.
2897 * Caution: this function does not check that the caller has disabled
2898 * preemption, thus the result might have a time-of-check-to-time-of-use
2899 * race. The caller is responsible to use it correctly, for example:
2901 * - from a non-preemptable section (of course)
2903 * - from a thread that is bound to a single CPU
2905 * - in a loop with very short iterations (e.g. a polling loop)
2907 bool single_task_running(void)
2909 return raw_rq()->nr_running == 1;
2911 EXPORT_SYMBOL(single_task_running);
2913 unsigned long long nr_context_switches(void)
2916 unsigned long long sum = 0;
2918 for_each_possible_cpu(i)
2919 sum += cpu_rq(i)->nr_switches;
2925 * IO-wait accounting, and how its mostly bollocks (on SMP).
2927 * The idea behind IO-wait account is to account the idle time that we could
2928 * have spend running if it were not for IO. That is, if we were to improve the
2929 * storage performance, we'd have a proportional reduction in IO-wait time.
2931 * This all works nicely on UP, where, when a task blocks on IO, we account
2932 * idle time as IO-wait, because if the storage were faster, it could've been
2933 * running and we'd not be idle.
2935 * This has been extended to SMP, by doing the same for each CPU. This however
2938 * Imagine for instance the case where two tasks block on one CPU, only the one
2939 * CPU will have IO-wait accounted, while the other has regular idle. Even
2940 * though, if the storage were faster, both could've ran at the same time,
2941 * utilising both CPUs.
2943 * This means, that when looking globally, the current IO-wait accounting on
2944 * SMP is a lower bound, by reason of under accounting.
2946 * Worse, since the numbers are provided per CPU, they are sometimes
2947 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2948 * associated with any one particular CPU, it can wake to another CPU than it
2949 * blocked on. This means the per CPU IO-wait number is meaningless.
2951 * Task CPU affinities can make all that even more 'interesting'.
2954 unsigned long nr_iowait(void)
2956 unsigned long i, sum = 0;
2958 for_each_possible_cpu(i)
2959 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2965 * Consumers of these two interfaces, like for example the cpufreq menu
2966 * governor are using nonsensical data. Boosting frequency for a CPU that has
2967 * IO-wait which might not even end up running the task when it does become
2971 unsigned long nr_iowait_cpu(int cpu)
2973 struct rq *this = cpu_rq(cpu);
2974 return atomic_read(&this->nr_iowait);
2977 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2979 struct rq *rq = this_rq();
2980 *nr_waiters = atomic_read(&rq->nr_iowait);
2981 *load = rq->load.weight;
2987 * sched_exec - execve() is a valuable balancing opportunity, because at
2988 * this point the task has the smallest effective memory and cache footprint.
2990 void sched_exec(void)
2992 struct task_struct *p = current;
2993 unsigned long flags;
2996 raw_spin_lock_irqsave(&p->pi_lock, flags);
2997 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2998 if (dest_cpu == smp_processor_id())
3001 if (likely(cpu_active(dest_cpu))) {
3002 struct migration_arg arg = { p, dest_cpu };
3004 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3005 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3009 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3014 DEFINE_PER_CPU(struct kernel_stat, kstat);
3015 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3017 EXPORT_PER_CPU_SYMBOL(kstat);
3018 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3021 * The function fair_sched_class.update_curr accesses the struct curr
3022 * and its field curr->exec_start; when called from task_sched_runtime(),
3023 * we observe a high rate of cache misses in practice.
3024 * Prefetching this data results in improved performance.
3026 static inline void prefetch_curr_exec_start(struct task_struct *p)
3028 #ifdef CONFIG_FAIR_GROUP_SCHED
3029 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3031 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3034 prefetch(&curr->exec_start);
3038 * Return accounted runtime for the task.
3039 * In case the task is currently running, return the runtime plus current's
3040 * pending runtime that have not been accounted yet.
3042 unsigned long long task_sched_runtime(struct task_struct *p)
3048 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3050 * 64-bit doesn't need locks to atomically read a 64bit value.
3051 * So we have a optimization chance when the task's delta_exec is 0.
3052 * Reading ->on_cpu is racy, but this is ok.
3054 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3055 * If we race with it entering CPU, unaccounted time is 0. This is
3056 * indistinguishable from the read occurring a few cycles earlier.
3057 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3058 * been accounted, so we're correct here as well.
3060 if (!p->on_cpu || !task_on_rq_queued(p))
3061 return p->se.sum_exec_runtime;
3064 rq = task_rq_lock(p, &rf);
3066 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3067 * project cycles that may never be accounted to this
3068 * thread, breaking clock_gettime().
3070 if (task_current(rq, p) && task_on_rq_queued(p)) {
3071 prefetch_curr_exec_start(p);
3072 update_rq_clock(rq);
3073 p->sched_class->update_curr(rq);
3075 ns = p->se.sum_exec_runtime;
3076 task_rq_unlock(rq, p, &rf);
3082 * This function gets called by the timer code, with HZ frequency.
3083 * We call it with interrupts disabled.
3085 void scheduler_tick(void)
3087 int cpu = smp_processor_id();
3088 struct rq *rq = cpu_rq(cpu);
3089 struct task_struct *curr = rq->curr;
3093 raw_spin_lock(&rq->lock);
3094 update_rq_clock(rq);
3095 curr->sched_class->task_tick(rq, curr, 0);
3096 cpu_load_update_active(rq);
3097 calc_global_load_tick(rq);
3098 raw_spin_unlock(&rq->lock);
3100 perf_event_task_tick();
3103 rq->idle_balance = idle_cpu(cpu);
3104 trigger_load_balance(rq);
3106 rq_last_tick_reset(rq);
3109 #ifdef CONFIG_NO_HZ_FULL
3111 * scheduler_tick_max_deferment
3113 * Keep at least one tick per second when a single
3114 * active task is running because the scheduler doesn't
3115 * yet completely support full dynticks environment.
3117 * This makes sure that uptime, CFS vruntime, load
3118 * balancing, etc... continue to move forward, even
3119 * with a very low granularity.
3121 * Return: Maximum deferment in nanoseconds.
3123 u64 scheduler_tick_max_deferment(void)
3125 struct rq *rq = this_rq();
3126 unsigned long next, now = READ_ONCE(jiffies);
3128 next = rq->last_sched_tick + HZ;
3130 if (time_before_eq(next, now))
3133 return jiffies_to_nsecs(next - now);
3137 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3138 defined(CONFIG_PREEMPT_TRACER))
3140 * If the value passed in is equal to the current preempt count
3141 * then we just disabled preemption. Start timing the latency.
3143 static inline void preempt_latency_start(int val)
3145 if (preempt_count() == val) {
3146 unsigned long ip = get_lock_parent_ip();
3147 #ifdef CONFIG_DEBUG_PREEMPT
3148 current->preempt_disable_ip = ip;
3150 trace_preempt_off(CALLER_ADDR0, ip);
3154 void preempt_count_add(int val)
3156 #ifdef CONFIG_DEBUG_PREEMPT
3160 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3163 __preempt_count_add(val);
3164 #ifdef CONFIG_DEBUG_PREEMPT
3166 * Spinlock count overflowing soon?
3168 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3171 preempt_latency_start(val);
3173 EXPORT_SYMBOL(preempt_count_add);
3174 NOKPROBE_SYMBOL(preempt_count_add);
3177 * If the value passed in equals to the current preempt count
3178 * then we just enabled preemption. Stop timing the latency.
3180 static inline void preempt_latency_stop(int val)
3182 if (preempt_count() == val)
3183 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3186 void preempt_count_sub(int val)
3188 #ifdef CONFIG_DEBUG_PREEMPT
3192 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3195 * Is the spinlock portion underflowing?
3197 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3198 !(preempt_count() & PREEMPT_MASK)))
3202 preempt_latency_stop(val);
3203 __preempt_count_sub(val);
3205 EXPORT_SYMBOL(preempt_count_sub);
3206 NOKPROBE_SYMBOL(preempt_count_sub);
3209 static inline void preempt_latency_start(int val) { }
3210 static inline void preempt_latency_stop(int val) { }
3214 * Print scheduling while atomic bug:
3216 static noinline void __schedule_bug(struct task_struct *prev)
3218 /* Save this before calling printk(), since that will clobber it */
3219 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3221 if (oops_in_progress)
3224 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3225 prev->comm, prev->pid, preempt_count());
3227 debug_show_held_locks(prev);
3229 if (irqs_disabled())
3230 print_irqtrace_events(prev);
3231 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3232 && in_atomic_preempt_off()) {
3233 pr_err("Preemption disabled at:");
3234 print_ip_sym(preempt_disable_ip);
3238 panic("scheduling while atomic\n");
3241 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3245 * Various schedule()-time debugging checks and statistics:
3247 static inline void schedule_debug(struct task_struct *prev)
3249 #ifdef CONFIG_SCHED_STACK_END_CHECK
3250 if (task_stack_end_corrupted(prev))
3251 panic("corrupted stack end detected inside scheduler\n");
3254 if (unlikely(in_atomic_preempt_off())) {
3255 __schedule_bug(prev);
3256 preempt_count_set(PREEMPT_DISABLED);
3260 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3262 schedstat_inc(this_rq()->sched_count);
3266 * Pick up the highest-prio task:
3268 static inline struct task_struct *
3269 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3271 const struct sched_class *class;
3272 struct task_struct *p;
3275 * Optimization: we know that if all tasks are in
3276 * the fair class we can call that function directly:
3278 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3279 p = fair_sched_class.pick_next_task(rq, prev, rf);
3280 if (unlikely(p == RETRY_TASK))
3283 /* Assumes fair_sched_class->next == idle_sched_class */
3285 p = idle_sched_class.pick_next_task(rq, prev, rf);
3291 for_each_class(class) {
3292 p = class->pick_next_task(rq, prev, rf);
3294 if (unlikely(p == RETRY_TASK))
3300 /* The idle class should always have a runnable task: */
3305 * __schedule() is the main scheduler function.
3307 * The main means of driving the scheduler and thus entering this function are:
3309 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3311 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3312 * paths. For example, see arch/x86/entry_64.S.
3314 * To drive preemption between tasks, the scheduler sets the flag in timer
3315 * interrupt handler scheduler_tick().
3317 * 3. Wakeups don't really cause entry into schedule(). They add a
3318 * task to the run-queue and that's it.
3320 * Now, if the new task added to the run-queue preempts the current
3321 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3322 * called on the nearest possible occasion:
3324 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3326 * - in syscall or exception context, at the next outmost
3327 * preempt_enable(). (this might be as soon as the wake_up()'s
3330 * - in IRQ context, return from interrupt-handler to
3331 * preemptible context
3333 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3336 * - cond_resched() call
3337 * - explicit schedule() call
3338 * - return from syscall or exception to user-space
3339 * - return from interrupt-handler to user-space
3341 * WARNING: must be called with preemption disabled!
3343 static void __sched notrace __schedule(bool preempt)
3345 struct task_struct *prev, *next;
3346 unsigned long *switch_count;
3351 cpu = smp_processor_id();
3355 schedule_debug(prev);
3357 if (sched_feat(HRTICK))
3360 local_irq_disable();
3361 rcu_note_context_switch();
3364 * Make sure that signal_pending_state()->signal_pending() below
3365 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3366 * done by the caller to avoid the race with signal_wake_up().
3368 smp_mb__before_spinlock();
3369 raw_spin_lock(&rq->lock);
3370 rq_pin_lock(rq, &rf);
3372 /* Promote REQ to ACT */
3373 rq->clock_update_flags <<= 1;
3375 switch_count = &prev->nivcsw;
3376 if (!preempt && prev->state) {
3377 if (unlikely(signal_pending_state(prev->state, prev))) {
3378 prev->state = TASK_RUNNING;
3380 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3383 if (prev->in_iowait) {
3384 atomic_inc(&rq->nr_iowait);
3385 delayacct_blkio_start();
3389 * If a worker went to sleep, notify and ask workqueue
3390 * whether it wants to wake up a task to maintain
3393 if (prev->flags & PF_WQ_WORKER) {
3394 struct task_struct *to_wakeup;
3396 to_wakeup = wq_worker_sleeping(prev);
3398 try_to_wake_up_local(to_wakeup, &rf);
3401 switch_count = &prev->nvcsw;
3404 if (task_on_rq_queued(prev))
3405 update_rq_clock(rq);
3407 next = pick_next_task(rq, prev, &rf);
3408 clear_tsk_need_resched(prev);
3409 clear_preempt_need_resched();
3411 if (likely(prev != next)) {
3416 trace_sched_switch(preempt, prev, next);
3418 /* Also unlocks the rq: */
3419 rq = context_switch(rq, prev, next, &rf);
3421 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3422 rq_unpin_lock(rq, &rf);
3423 raw_spin_unlock_irq(&rq->lock);
3426 balance_callback(rq);
3429 void __noreturn do_task_dead(void)
3432 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3433 * when the following two conditions become true.
3434 * - There is race condition of mmap_sem (It is acquired by
3436 * - SMI occurs before setting TASK_RUNINNG.
3437 * (or hypervisor of virtual machine switches to other guest)
3438 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3440 * To avoid it, we have to wait for releasing tsk->pi_lock which
3441 * is held by try_to_wake_up()
3444 raw_spin_unlock_wait(¤t->pi_lock);
3446 /* Causes final put_task_struct in finish_task_switch(): */
3447 __set_current_state(TASK_DEAD);
3449 /* Tell freezer to ignore us: */
3450 current->flags |= PF_NOFREEZE;
3455 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3460 static inline void sched_submit_work(struct task_struct *tsk)
3462 if (!tsk->state || tsk_is_pi_blocked(tsk))
3465 * If we are going to sleep and we have plugged IO queued,
3466 * make sure to submit it to avoid deadlocks.
3468 if (blk_needs_flush_plug(tsk))
3469 blk_schedule_flush_plug(tsk);
3472 asmlinkage __visible void __sched schedule(void)
3474 struct task_struct *tsk = current;
3476 sched_submit_work(tsk);
3480 sched_preempt_enable_no_resched();
3481 } while (need_resched());
3483 EXPORT_SYMBOL(schedule);
3485 #ifdef CONFIG_CONTEXT_TRACKING
3486 asmlinkage __visible void __sched schedule_user(void)
3489 * If we come here after a random call to set_need_resched(),
3490 * or we have been woken up remotely but the IPI has not yet arrived,
3491 * we haven't yet exited the RCU idle mode. Do it here manually until
3492 * we find a better solution.
3494 * NB: There are buggy callers of this function. Ideally we
3495 * should warn if prev_state != CONTEXT_USER, but that will trigger
3496 * too frequently to make sense yet.
3498 enum ctx_state prev_state = exception_enter();
3500 exception_exit(prev_state);
3505 * schedule_preempt_disabled - called with preemption disabled
3507 * Returns with preemption disabled. Note: preempt_count must be 1
3509 void __sched schedule_preempt_disabled(void)
3511 sched_preempt_enable_no_resched();
3516 static void __sched notrace preempt_schedule_common(void)
3520 * Because the function tracer can trace preempt_count_sub()
3521 * and it also uses preempt_enable/disable_notrace(), if
3522 * NEED_RESCHED is set, the preempt_enable_notrace() called
3523 * by the function tracer will call this function again and
3524 * cause infinite recursion.
3526 * Preemption must be disabled here before the function
3527 * tracer can trace. Break up preempt_disable() into two
3528 * calls. One to disable preemption without fear of being
3529 * traced. The other to still record the preemption latency,
3530 * which can also be traced by the function tracer.
3532 preempt_disable_notrace();
3533 preempt_latency_start(1);
3535 preempt_latency_stop(1);
3536 preempt_enable_no_resched_notrace();
3539 * Check again in case we missed a preemption opportunity
3540 * between schedule and now.
3542 } while (need_resched());
3545 #ifdef CONFIG_PREEMPT
3547 * this is the entry point to schedule() from in-kernel preemption
3548 * off of preempt_enable. Kernel preemptions off return from interrupt
3549 * occur there and call schedule directly.
3551 asmlinkage __visible void __sched notrace preempt_schedule(void)
3554 * If there is a non-zero preempt_count or interrupts are disabled,
3555 * we do not want to preempt the current task. Just return..
3557 if (likely(!preemptible()))
3560 preempt_schedule_common();
3562 NOKPROBE_SYMBOL(preempt_schedule);
3563 EXPORT_SYMBOL(preempt_schedule);
3566 * preempt_schedule_notrace - preempt_schedule called by tracing
3568 * The tracing infrastructure uses preempt_enable_notrace to prevent
3569 * recursion and tracing preempt enabling caused by the tracing
3570 * infrastructure itself. But as tracing can happen in areas coming
3571 * from userspace or just about to enter userspace, a preempt enable
3572 * can occur before user_exit() is called. This will cause the scheduler
3573 * to be called when the system is still in usermode.
3575 * To prevent this, the preempt_enable_notrace will use this function
3576 * instead of preempt_schedule() to exit user context if needed before
3577 * calling the scheduler.
3579 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3581 enum ctx_state prev_ctx;
3583 if (likely(!preemptible()))
3588 * Because the function tracer can trace preempt_count_sub()
3589 * and it also uses preempt_enable/disable_notrace(), if
3590 * NEED_RESCHED is set, the preempt_enable_notrace() called
3591 * by the function tracer will call this function again and
3592 * cause infinite recursion.
3594 * Preemption must be disabled here before the function
3595 * tracer can trace. Break up preempt_disable() into two
3596 * calls. One to disable preemption without fear of being
3597 * traced. The other to still record the preemption latency,
3598 * which can also be traced by the function tracer.
3600 preempt_disable_notrace();
3601 preempt_latency_start(1);
3603 * Needs preempt disabled in case user_exit() is traced
3604 * and the tracer calls preempt_enable_notrace() causing
3605 * an infinite recursion.
3607 prev_ctx = exception_enter();
3609 exception_exit(prev_ctx);
3611 preempt_latency_stop(1);
3612 preempt_enable_no_resched_notrace();
3613 } while (need_resched());
3615 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3617 #endif /* CONFIG_PREEMPT */
3620 * this is the entry point to schedule() from kernel preemption
3621 * off of irq context.
3622 * Note, that this is called and return with irqs disabled. This will
3623 * protect us against recursive calling from irq.
3625 asmlinkage __visible void __sched preempt_schedule_irq(void)
3627 enum ctx_state prev_state;
3629 /* Catch callers which need to be fixed */
3630 BUG_ON(preempt_count() || !irqs_disabled());
3632 prev_state = exception_enter();
3638 local_irq_disable();
3639 sched_preempt_enable_no_resched();
3640 } while (need_resched());
3642 exception_exit(prev_state);
3645 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3648 return try_to_wake_up(curr->private, mode, wake_flags);
3650 EXPORT_SYMBOL(default_wake_function);
3652 #ifdef CONFIG_RT_MUTEXES
3655 * rt_mutex_setprio - set the current priority of a task
3657 * @prio: prio value (kernel-internal form)
3659 * This function changes the 'effective' priority of a task. It does
3660 * not touch ->normal_prio like __setscheduler().
3662 * Used by the rt_mutex code to implement priority inheritance
3663 * logic. Call site only calls if the priority of the task changed.
3665 void rt_mutex_setprio(struct task_struct *p, int prio)
3667 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3668 const struct sched_class *prev_class;
3672 BUG_ON(prio > MAX_PRIO);
3674 rq = __task_rq_lock(p, &rf);
3675 update_rq_clock(rq);
3678 * Idle task boosting is a nono in general. There is one
3679 * exception, when PREEMPT_RT and NOHZ is active:
3681 * The idle task calls get_next_timer_interrupt() and holds
3682 * the timer wheel base->lock on the CPU and another CPU wants
3683 * to access the timer (probably to cancel it). We can safely
3684 * ignore the boosting request, as the idle CPU runs this code
3685 * with interrupts disabled and will complete the lock
3686 * protected section without being interrupted. So there is no
3687 * real need to boost.
3689 if (unlikely(p == rq->idle)) {
3690 WARN_ON(p != rq->curr);
3691 WARN_ON(p->pi_blocked_on);
3695 trace_sched_pi_setprio(p, prio);
3698 if (oldprio == prio)
3699 queue_flag &= ~DEQUEUE_MOVE;
3701 prev_class = p->sched_class;
3702 queued = task_on_rq_queued(p);
3703 running = task_current(rq, p);
3705 dequeue_task(rq, p, queue_flag);
3707 put_prev_task(rq, p);
3710 * Boosting condition are:
3711 * 1. -rt task is running and holds mutex A
3712 * --> -dl task blocks on mutex A
3714 * 2. -dl task is running and holds mutex A
3715 * --> -dl task blocks on mutex A and could preempt the
3718 if (dl_prio(prio)) {
3719 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3720 if (!dl_prio(p->normal_prio) ||
3721 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3722 p->dl.dl_boosted = 1;
3723 queue_flag |= ENQUEUE_REPLENISH;
3725 p->dl.dl_boosted = 0;
3726 p->sched_class = &dl_sched_class;
3727 } else if (rt_prio(prio)) {
3728 if (dl_prio(oldprio))
3729 p->dl.dl_boosted = 0;
3731 queue_flag |= ENQUEUE_HEAD;
3732 p->sched_class = &rt_sched_class;
3734 if (dl_prio(oldprio))
3735 p->dl.dl_boosted = 0;
3736 if (rt_prio(oldprio))
3738 p->sched_class = &fair_sched_class;
3744 enqueue_task(rq, p, queue_flag);
3746 set_curr_task(rq, p);
3748 check_class_changed(rq, p, prev_class, oldprio);
3750 /* Avoid rq from going away on us: */
3752 __task_rq_unlock(rq, &rf);
3754 balance_callback(rq);
3759 void set_user_nice(struct task_struct *p, long nice)
3761 bool queued, running;
3762 int old_prio, delta;
3766 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3769 * We have to be careful, if called from sys_setpriority(),
3770 * the task might be in the middle of scheduling on another CPU.
3772 rq = task_rq_lock(p, &rf);
3773 update_rq_clock(rq);
3776 * The RT priorities are set via sched_setscheduler(), but we still
3777 * allow the 'normal' nice value to be set - but as expected
3778 * it wont have any effect on scheduling until the task is
3779 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3781 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3782 p->static_prio = NICE_TO_PRIO(nice);
3785 queued = task_on_rq_queued(p);
3786 running = task_current(rq, p);
3788 dequeue_task(rq, p, DEQUEUE_SAVE);
3790 put_prev_task(rq, p);
3792 p->static_prio = NICE_TO_PRIO(nice);
3795 p->prio = effective_prio(p);
3796 delta = p->prio - old_prio;
3799 enqueue_task(rq, p, ENQUEUE_RESTORE);
3801 * If the task increased its priority or is running and
3802 * lowered its priority, then reschedule its CPU:
3804 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3808 set_curr_task(rq, p);
3810 task_rq_unlock(rq, p, &rf);
3812 EXPORT_SYMBOL(set_user_nice);
3815 * can_nice - check if a task can reduce its nice value
3819 int can_nice(const struct task_struct *p, const int nice)
3821 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3822 int nice_rlim = nice_to_rlimit(nice);
3824 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3825 capable(CAP_SYS_NICE));
3828 #ifdef __ARCH_WANT_SYS_NICE
3831 * sys_nice - change the priority of the current process.
3832 * @increment: priority increment
3834 * sys_setpriority is a more generic, but much slower function that
3835 * does similar things.
3837 SYSCALL_DEFINE1(nice, int, increment)
3842 * Setpriority might change our priority at the same moment.
3843 * We don't have to worry. Conceptually one call occurs first
3844 * and we have a single winner.
3846 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3847 nice = task_nice(current) + increment;
3849 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3850 if (increment < 0 && !can_nice(current, nice))
3853 retval = security_task_setnice(current, nice);
3857 set_user_nice(current, nice);
3864 * task_prio - return the priority value of a given task.
3865 * @p: the task in question.
3867 * Return: The priority value as seen by users in /proc.
3868 * RT tasks are offset by -200. Normal tasks are centered
3869 * around 0, value goes from -16 to +15.
3871 int task_prio(const struct task_struct *p)
3873 return p->prio - MAX_RT_PRIO;
3877 * idle_cpu - is a given CPU idle currently?
3878 * @cpu: the processor in question.
3880 * Return: 1 if the CPU is currently idle. 0 otherwise.
3882 int idle_cpu(int cpu)
3884 struct rq *rq = cpu_rq(cpu);
3886 if (rq->curr != rq->idle)
3893 if (!llist_empty(&rq->wake_list))
3901 * idle_task - return the idle task for a given CPU.
3902 * @cpu: the processor in question.
3904 * Return: The idle task for the CPU @cpu.
3906 struct task_struct *idle_task(int cpu)
3908 return cpu_rq(cpu)->idle;
3912 * find_process_by_pid - find a process with a matching PID value.
3913 * @pid: the pid in question.
3915 * The task of @pid, if found. %NULL otherwise.
3917 static struct task_struct *find_process_by_pid(pid_t pid)
3919 return pid ? find_task_by_vpid(pid) : current;
3923 * This function initializes the sched_dl_entity of a newly becoming
3924 * SCHED_DEADLINE task.
3926 * Only the static values are considered here, the actual runtime and the
3927 * absolute deadline will be properly calculated when the task is enqueued
3928 * for the first time with its new policy.
3931 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3933 struct sched_dl_entity *dl_se = &p->dl;
3935 dl_se->dl_runtime = attr->sched_runtime;
3936 dl_se->dl_deadline = attr->sched_deadline;
3937 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3938 dl_se->flags = attr->sched_flags;
3939 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3942 * Changing the parameters of a task is 'tricky' and we're not doing
3943 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3945 * What we SHOULD do is delay the bandwidth release until the 0-lag
3946 * point. This would include retaining the task_struct until that time
3947 * and change dl_overflow() to not immediately decrement the current
3950 * Instead we retain the current runtime/deadline and let the new
3951 * parameters take effect after the current reservation period lapses.
3952 * This is safe (albeit pessimistic) because the 0-lag point is always
3953 * before the current scheduling deadline.
3955 * We can still have temporary overloads because we do not delay the
3956 * change in bandwidth until that time; so admission control is
3957 * not on the safe side. It does however guarantee tasks will never
3958 * consume more than promised.
3963 * sched_setparam() passes in -1 for its policy, to let the functions
3964 * it calls know not to change it.
3966 #define SETPARAM_POLICY -1
3968 static void __setscheduler_params(struct task_struct *p,
3969 const struct sched_attr *attr)
3971 int policy = attr->sched_policy;
3973 if (policy == SETPARAM_POLICY)
3978 if (dl_policy(policy))
3979 __setparam_dl(p, attr);
3980 else if (fair_policy(policy))
3981 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3984 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3985 * !rt_policy. Always setting this ensures that things like
3986 * getparam()/getattr() don't report silly values for !rt tasks.
3988 p->rt_priority = attr->sched_priority;
3989 p->normal_prio = normal_prio(p);
3993 /* Actually do priority change: must hold pi & rq lock. */
3994 static void __setscheduler(struct rq *rq, struct task_struct *p,
3995 const struct sched_attr *attr, bool keep_boost)
3997 __setscheduler_params(p, attr);
4000 * Keep a potential priority boosting if called from
4001 * sched_setscheduler().
4004 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
4006 p->prio = normal_prio(p);
4008 if (dl_prio(p->prio))
4009 p->sched_class = &dl_sched_class;
4010 else if (rt_prio(p->prio))
4011 p->sched_class = &rt_sched_class;
4013 p->sched_class = &fair_sched_class;
4017 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
4019 struct sched_dl_entity *dl_se = &p->dl;
4021 attr->sched_priority = p->rt_priority;
4022 attr->sched_runtime = dl_se->dl_runtime;
4023 attr->sched_deadline = dl_se->dl_deadline;
4024 attr->sched_period = dl_se->dl_period;
4025 attr->sched_flags = dl_se->flags;
4029 * This function validates the new parameters of a -deadline task.
4030 * We ask for the deadline not being zero, and greater or equal
4031 * than the runtime, as well as the period of being zero or
4032 * greater than deadline. Furthermore, we have to be sure that
4033 * user parameters are above the internal resolution of 1us (we
4034 * check sched_runtime only since it is always the smaller one) and
4035 * below 2^63 ns (we have to check both sched_deadline and
4036 * sched_period, as the latter can be zero).
4039 __checkparam_dl(const struct sched_attr *attr)
4042 if (attr->sched_deadline == 0)
4046 * Since we truncate DL_SCALE bits, make sure we're at least
4049 if (attr->sched_runtime < (1ULL << DL_SCALE))
4053 * Since we use the MSB for wrap-around and sign issues, make
4054 * sure it's not set (mind that period can be equal to zero).
4056 if (attr->sched_deadline & (1ULL << 63) ||
4057 attr->sched_period & (1ULL << 63))
4060 /* runtime <= deadline <= period (if period != 0) */
4061 if ((attr->sched_period != 0 &&
4062 attr->sched_period < attr->sched_deadline) ||
4063 attr->sched_deadline < attr->sched_runtime)
4070 * Check the target process has a UID that matches the current process's:
4072 static bool check_same_owner(struct task_struct *p)
4074 const struct cred *cred = current_cred(), *pcred;
4078 pcred = __task_cred(p);
4079 match = (uid_eq(cred->euid, pcred->euid) ||
4080 uid_eq(cred->euid, pcred->uid));
4085 static bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
4087 struct sched_dl_entity *dl_se = &p->dl;
4089 if (dl_se->dl_runtime != attr->sched_runtime ||
4090 dl_se->dl_deadline != attr->sched_deadline ||
4091 dl_se->dl_period != attr->sched_period ||
4092 dl_se->flags != attr->sched_flags)
4098 static int __sched_setscheduler(struct task_struct *p,
4099 const struct sched_attr *attr,
4102 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4103 MAX_RT_PRIO - 1 - attr->sched_priority;
4104 int retval, oldprio, oldpolicy = -1, queued, running;
4105 int new_effective_prio, policy = attr->sched_policy;
4106 const struct sched_class *prev_class;
4109 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
4112 /* May grab non-irq protected spin_locks: */
4113 BUG_ON(in_interrupt());
4115 /* Double check policy once rq lock held: */
4117 reset_on_fork = p->sched_reset_on_fork;
4118 policy = oldpolicy = p->policy;
4120 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4122 if (!valid_policy(policy))
4126 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4130 * Valid priorities for SCHED_FIFO and SCHED_RR are
4131 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4132 * SCHED_BATCH and SCHED_IDLE is 0.
4134 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4135 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4137 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4138 (rt_policy(policy) != (attr->sched_priority != 0)))
4142 * Allow unprivileged RT tasks to decrease priority:
4144 if (user && !capable(CAP_SYS_NICE)) {
4145 if (fair_policy(policy)) {
4146 if (attr->sched_nice < task_nice(p) &&
4147 !can_nice(p, attr->sched_nice))
4151 if (rt_policy(policy)) {
4152 unsigned long rlim_rtprio =
4153 task_rlimit(p, RLIMIT_RTPRIO);
4155 /* Can't set/change the rt policy: */
4156 if (policy != p->policy && !rlim_rtprio)
4159 /* Can't increase priority: */
4160 if (attr->sched_priority > p->rt_priority &&
4161 attr->sched_priority > rlim_rtprio)
4166 * Can't set/change SCHED_DEADLINE policy at all for now
4167 * (safest behavior); in the future we would like to allow
4168 * unprivileged DL tasks to increase their relative deadline
4169 * or reduce their runtime (both ways reducing utilization)
4171 if (dl_policy(policy))
4175 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4176 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4178 if (idle_policy(p->policy) && !idle_policy(policy)) {
4179 if (!can_nice(p, task_nice(p)))
4183 /* Can't change other user's priorities: */
4184 if (!check_same_owner(p))
4187 /* Normal users shall not reset the sched_reset_on_fork flag: */
4188 if (p->sched_reset_on_fork && !reset_on_fork)
4193 retval = security_task_setscheduler(p);
4199 * Make sure no PI-waiters arrive (or leave) while we are
4200 * changing the priority of the task:
4202 * To be able to change p->policy safely, the appropriate
4203 * runqueue lock must be held.
4205 rq = task_rq_lock(p, &rf);
4206 update_rq_clock(rq);
4209 * Changing the policy of the stop threads its a very bad idea:
4211 if (p == rq->stop) {
4212 task_rq_unlock(rq, p, &rf);
4217 * If not changing anything there's no need to proceed further,
4218 * but store a possible modification of reset_on_fork.
4220 if (unlikely(policy == p->policy)) {
4221 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4223 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4225 if (dl_policy(policy) && dl_param_changed(p, attr))
4228 p->sched_reset_on_fork = reset_on_fork;
4229 task_rq_unlock(rq, p, &rf);
4235 #ifdef CONFIG_RT_GROUP_SCHED
4237 * Do not allow realtime tasks into groups that have no runtime
4240 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4241 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4242 !task_group_is_autogroup(task_group(p))) {
4243 task_rq_unlock(rq, p, &rf);
4248 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4249 cpumask_t *span = rq->rd->span;
4252 * Don't allow tasks with an affinity mask smaller than
4253 * the entire root_domain to become SCHED_DEADLINE. We
4254 * will also fail if there's no bandwidth available.
4256 if (!cpumask_subset(span, &p->cpus_allowed) ||
4257 rq->rd->dl_bw.bw == 0) {
4258 task_rq_unlock(rq, p, &rf);
4265 /* Re-check policy now with rq lock held: */
4266 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4267 policy = oldpolicy = -1;
4268 task_rq_unlock(rq, p, &rf);
4273 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4274 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4277 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4278 task_rq_unlock(rq, p, &rf);
4282 p->sched_reset_on_fork = reset_on_fork;
4287 * Take priority boosted tasks into account. If the new
4288 * effective priority is unchanged, we just store the new
4289 * normal parameters and do not touch the scheduler class and
4290 * the runqueue. This will be done when the task deboost
4293 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4294 if (new_effective_prio == oldprio)
4295 queue_flags &= ~DEQUEUE_MOVE;
4298 queued = task_on_rq_queued(p);
4299 running = task_current(rq, p);
4301 dequeue_task(rq, p, queue_flags);
4303 put_prev_task(rq, p);
4305 prev_class = p->sched_class;
4306 __setscheduler(rq, p, attr, pi);
4310 * We enqueue to tail when the priority of a task is
4311 * increased (user space view).
4313 if (oldprio < p->prio)
4314 queue_flags |= ENQUEUE_HEAD;
4316 enqueue_task(rq, p, queue_flags);
4319 set_curr_task(rq, p);
4321 check_class_changed(rq, p, prev_class, oldprio);
4323 /* Avoid rq from going away on us: */
4325 task_rq_unlock(rq, p, &rf);
4328 rt_mutex_adjust_pi(p);
4330 /* Run balance callbacks after we've adjusted the PI chain: */
4331 balance_callback(rq);
4337 static int _sched_setscheduler(struct task_struct *p, int policy,
4338 const struct sched_param *param, bool check)
4340 struct sched_attr attr = {
4341 .sched_policy = policy,
4342 .sched_priority = param->sched_priority,
4343 .sched_nice = PRIO_TO_NICE(p->static_prio),
4346 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4347 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4348 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4349 policy &= ~SCHED_RESET_ON_FORK;
4350 attr.sched_policy = policy;
4353 return __sched_setscheduler(p, &attr, check, true);
4356 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4357 * @p: the task in question.
4358 * @policy: new policy.
4359 * @param: structure containing the new RT priority.
4361 * Return: 0 on success. An error code otherwise.
4363 * NOTE that the task may be already dead.
4365 int sched_setscheduler(struct task_struct *p, int policy,
4366 const struct sched_param *param)
4368 return _sched_setscheduler(p, policy, param, true);
4370 EXPORT_SYMBOL_GPL(sched_setscheduler);
4372 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4374 return __sched_setscheduler(p, attr, true, true);
4376 EXPORT_SYMBOL_GPL(sched_setattr);
4379 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4380 * @p: the task in question.
4381 * @policy: new policy.
4382 * @param: structure containing the new RT priority.
4384 * Just like sched_setscheduler, only don't bother checking if the
4385 * current context has permission. For example, this is needed in
4386 * stop_machine(): we create temporary high priority worker threads,
4387 * but our caller might not have that capability.
4389 * Return: 0 on success. An error code otherwise.
4391 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4392 const struct sched_param *param)
4394 return _sched_setscheduler(p, policy, param, false);
4396 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4399 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4401 struct sched_param lparam;
4402 struct task_struct *p;
4405 if (!param || pid < 0)
4407 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4412 p = find_process_by_pid(pid);
4414 retval = sched_setscheduler(p, policy, &lparam);
4421 * Mimics kernel/events/core.c perf_copy_attr().
4423 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4428 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4431 /* Zero the full structure, so that a short copy will be nice: */
4432 memset(attr, 0, sizeof(*attr));
4434 ret = get_user(size, &uattr->size);
4438 /* Bail out on silly large: */
4439 if (size > PAGE_SIZE)
4442 /* ABI compatibility quirk: */
4444 size = SCHED_ATTR_SIZE_VER0;
4446 if (size < SCHED_ATTR_SIZE_VER0)
4450 * If we're handed a bigger struct than we know of,
4451 * ensure all the unknown bits are 0 - i.e. new
4452 * user-space does not rely on any kernel feature
4453 * extensions we dont know about yet.
4455 if (size > sizeof(*attr)) {
4456 unsigned char __user *addr;
4457 unsigned char __user *end;
4460 addr = (void __user *)uattr + sizeof(*attr);
4461 end = (void __user *)uattr + size;
4463 for (; addr < end; addr++) {
4464 ret = get_user(val, addr);
4470 size = sizeof(*attr);
4473 ret = copy_from_user(attr, uattr, size);
4478 * XXX: Do we want to be lenient like existing syscalls; or do we want
4479 * to be strict and return an error on out-of-bounds values?
4481 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4486 put_user(sizeof(*attr), &uattr->size);
4491 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4492 * @pid: the pid in question.
4493 * @policy: new policy.
4494 * @param: structure containing the new RT priority.
4496 * Return: 0 on success. An error code otherwise.
4498 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4503 return do_sched_setscheduler(pid, policy, param);
4507 * sys_sched_setparam - set/change the RT priority of a thread
4508 * @pid: the pid in question.
4509 * @param: structure containing the new RT priority.
4511 * Return: 0 on success. An error code otherwise.
4513 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4515 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4519 * sys_sched_setattr - same as above, but with extended sched_attr
4520 * @pid: the pid in question.
4521 * @uattr: structure containing the extended parameters.
4522 * @flags: for future extension.
4524 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4525 unsigned int, flags)
4527 struct sched_attr attr;
4528 struct task_struct *p;
4531 if (!uattr || pid < 0 || flags)
4534 retval = sched_copy_attr(uattr, &attr);
4538 if ((int)attr.sched_policy < 0)
4543 p = find_process_by_pid(pid);
4545 retval = sched_setattr(p, &attr);
4552 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4553 * @pid: the pid in question.
4555 * Return: On success, the policy of the thread. Otherwise, a negative error
4558 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4560 struct task_struct *p;
4568 p = find_process_by_pid(pid);
4570 retval = security_task_getscheduler(p);
4573 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4580 * sys_sched_getparam - get the RT priority of a thread
4581 * @pid: the pid in question.
4582 * @param: structure containing the RT priority.
4584 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4587 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4589 struct sched_param lp = { .sched_priority = 0 };
4590 struct task_struct *p;
4593 if (!param || pid < 0)
4597 p = find_process_by_pid(pid);
4602 retval = security_task_getscheduler(p);
4606 if (task_has_rt_policy(p))
4607 lp.sched_priority = p->rt_priority;
4611 * This one might sleep, we cannot do it with a spinlock held ...
4613 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4622 static int sched_read_attr(struct sched_attr __user *uattr,
4623 struct sched_attr *attr,
4628 if (!access_ok(VERIFY_WRITE, uattr, usize))
4632 * If we're handed a smaller struct than we know of,
4633 * ensure all the unknown bits are 0 - i.e. old
4634 * user-space does not get uncomplete information.
4636 if (usize < sizeof(*attr)) {
4637 unsigned char *addr;
4640 addr = (void *)attr + usize;
4641 end = (void *)attr + sizeof(*attr);
4643 for (; addr < end; addr++) {
4651 ret = copy_to_user(uattr, attr, attr->size);
4659 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4660 * @pid: the pid in question.
4661 * @uattr: structure containing the extended parameters.
4662 * @size: sizeof(attr) for fwd/bwd comp.
4663 * @flags: for future extension.
4665 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4666 unsigned int, size, unsigned int, flags)
4668 struct sched_attr attr = {
4669 .size = sizeof(struct sched_attr),
4671 struct task_struct *p;
4674 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4675 size < SCHED_ATTR_SIZE_VER0 || flags)
4679 p = find_process_by_pid(pid);
4684 retval = security_task_getscheduler(p);
4688 attr.sched_policy = p->policy;
4689 if (p->sched_reset_on_fork)
4690 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4691 if (task_has_dl_policy(p))
4692 __getparam_dl(p, &attr);
4693 else if (task_has_rt_policy(p))
4694 attr.sched_priority = p->rt_priority;
4696 attr.sched_nice = task_nice(p);
4700 retval = sched_read_attr(uattr, &attr, size);
4708 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4710 cpumask_var_t cpus_allowed, new_mask;
4711 struct task_struct *p;
4716 p = find_process_by_pid(pid);
4722 /* Prevent p going away */
4726 if (p->flags & PF_NO_SETAFFINITY) {
4730 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4734 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4736 goto out_free_cpus_allowed;
4739 if (!check_same_owner(p)) {
4741 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4743 goto out_free_new_mask;
4748 retval = security_task_setscheduler(p);
4750 goto out_free_new_mask;
4753 cpuset_cpus_allowed(p, cpus_allowed);
4754 cpumask_and(new_mask, in_mask, cpus_allowed);
4757 * Since bandwidth control happens on root_domain basis,
4758 * if admission test is enabled, we only admit -deadline
4759 * tasks allowed to run on all the CPUs in the task's
4763 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4765 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4768 goto out_free_new_mask;
4774 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4777 cpuset_cpus_allowed(p, cpus_allowed);
4778 if (!cpumask_subset(new_mask, cpus_allowed)) {
4780 * We must have raced with a concurrent cpuset
4781 * update. Just reset the cpus_allowed to the
4782 * cpuset's cpus_allowed
4784 cpumask_copy(new_mask, cpus_allowed);
4789 free_cpumask_var(new_mask);
4790 out_free_cpus_allowed:
4791 free_cpumask_var(cpus_allowed);
4797 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4798 struct cpumask *new_mask)
4800 if (len < cpumask_size())
4801 cpumask_clear(new_mask);
4802 else if (len > cpumask_size())
4803 len = cpumask_size();
4805 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4809 * sys_sched_setaffinity - set the CPU affinity of a process
4810 * @pid: pid of the process
4811 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4812 * @user_mask_ptr: user-space pointer to the new CPU mask
4814 * Return: 0 on success. An error code otherwise.
4816 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4817 unsigned long __user *, user_mask_ptr)
4819 cpumask_var_t new_mask;
4822 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4825 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4827 retval = sched_setaffinity(pid, new_mask);
4828 free_cpumask_var(new_mask);
4832 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4834 struct task_struct *p;
4835 unsigned long flags;
4841 p = find_process_by_pid(pid);
4845 retval = security_task_getscheduler(p);
4849 raw_spin_lock_irqsave(&p->pi_lock, flags);
4850 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4851 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4860 * sys_sched_getaffinity - get the CPU affinity of a process
4861 * @pid: pid of the process
4862 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4863 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4865 * Return: size of CPU mask copied to user_mask_ptr on success. An
4866 * error code otherwise.
4868 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4869 unsigned long __user *, user_mask_ptr)
4874 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4876 if (len & (sizeof(unsigned long)-1))
4879 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4882 ret = sched_getaffinity(pid, mask);
4884 size_t retlen = min_t(size_t, len, cpumask_size());
4886 if (copy_to_user(user_mask_ptr, mask, retlen))
4891 free_cpumask_var(mask);
4897 * sys_sched_yield - yield the current processor to other threads.
4899 * This function yields the current CPU to other tasks. If there are no
4900 * other threads running on this CPU then this function will return.
4904 SYSCALL_DEFINE0(sched_yield)
4906 struct rq *rq = this_rq_lock();
4908 schedstat_inc(rq->yld_count);
4909 current->sched_class->yield_task(rq);
4912 * Since we are going to call schedule() anyway, there's
4913 * no need to preempt or enable interrupts:
4915 __release(rq->lock);
4916 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4917 do_raw_spin_unlock(&rq->lock);
4918 sched_preempt_enable_no_resched();
4925 #ifndef CONFIG_PREEMPT
4926 int __sched _cond_resched(void)
4928 if (should_resched(0)) {
4929 preempt_schedule_common();
4934 EXPORT_SYMBOL(_cond_resched);
4938 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4939 * call schedule, and on return reacquire the lock.
4941 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4942 * operations here to prevent schedule() from being called twice (once via
4943 * spin_unlock(), once by hand).
4945 int __cond_resched_lock(spinlock_t *lock)
4947 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4950 lockdep_assert_held(lock);
4952 if (spin_needbreak(lock) || resched) {
4955 preempt_schedule_common();
4963 EXPORT_SYMBOL(__cond_resched_lock);
4965 int __sched __cond_resched_softirq(void)
4967 BUG_ON(!in_softirq());
4969 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4971 preempt_schedule_common();
4977 EXPORT_SYMBOL(__cond_resched_softirq);
4980 * yield - yield the current processor to other threads.
4982 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4984 * The scheduler is at all times free to pick the calling task as the most
4985 * eligible task to run, if removing the yield() call from your code breaks
4986 * it, its already broken.
4988 * Typical broken usage is:
4993 * where one assumes that yield() will let 'the other' process run that will
4994 * make event true. If the current task is a SCHED_FIFO task that will never
4995 * happen. Never use yield() as a progress guarantee!!
4997 * If you want to use yield() to wait for something, use wait_event().
4998 * If you want to use yield() to be 'nice' for others, use cond_resched().
4999 * If you still want to use yield(), do not!
5001 void __sched yield(void)
5003 set_current_state(TASK_RUNNING);
5006 EXPORT_SYMBOL(yield);
5009 * yield_to - yield the current processor to another thread in
5010 * your thread group, or accelerate that thread toward the
5011 * processor it's on.
5013 * @preempt: whether task preemption is allowed or not
5015 * It's the caller's job to ensure that the target task struct
5016 * can't go away on us before we can do any checks.
5019 * true (>0) if we indeed boosted the target task.
5020 * false (0) if we failed to boost the target.
5021 * -ESRCH if there's no task to yield to.
5023 int __sched yield_to(struct task_struct *p, bool preempt)
5025 struct task_struct *curr = current;
5026 struct rq *rq, *p_rq;
5027 unsigned long flags;
5030 local_irq_save(flags);
5036 * If we're the only runnable task on the rq and target rq also
5037 * has only one task, there's absolutely no point in yielding.
5039 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5044 double_rq_lock(rq, p_rq);
5045 if (task_rq(p) != p_rq) {
5046 double_rq_unlock(rq, p_rq);
5050 if (!curr->sched_class->yield_to_task)
5053 if (curr->sched_class != p->sched_class)
5056 if (task_running(p_rq, p) || p->state)
5059 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5061 schedstat_inc(rq->yld_count);
5063 * Make p's CPU reschedule; pick_next_entity takes care of
5066 if (preempt && rq != p_rq)
5071 double_rq_unlock(rq, p_rq);
5073 local_irq_restore(flags);
5080 EXPORT_SYMBOL_GPL(yield_to);
5082 int io_schedule_prepare(void)
5084 int old_iowait = current->in_iowait;
5086 current->in_iowait = 1;
5087 blk_schedule_flush_plug(current);
5092 void io_schedule_finish(int token)
5094 current->in_iowait = token;
5098 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5099 * that process accounting knows that this is a task in IO wait state.
5101 long __sched io_schedule_timeout(long timeout)
5106 token = io_schedule_prepare();
5107 ret = schedule_timeout(timeout);
5108 io_schedule_finish(token);
5112 EXPORT_SYMBOL(io_schedule_timeout);
5114 void io_schedule(void)
5118 token = io_schedule_prepare();
5120 io_schedule_finish(token);
5122 EXPORT_SYMBOL(io_schedule);
5125 * sys_sched_get_priority_max - return maximum RT priority.
5126 * @policy: scheduling class.
5128 * Return: On success, this syscall returns the maximum
5129 * rt_priority that can be used by a given scheduling class.
5130 * On failure, a negative error code is returned.
5132 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5139 ret = MAX_USER_RT_PRIO-1;
5141 case SCHED_DEADLINE:
5152 * sys_sched_get_priority_min - return minimum RT priority.
5153 * @policy: scheduling class.
5155 * Return: On success, this syscall returns the minimum
5156 * rt_priority that can be used by a given scheduling class.
5157 * On failure, a negative error code is returned.
5159 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5168 case SCHED_DEADLINE:
5178 * sys_sched_rr_get_interval - return the default timeslice of a process.
5179 * @pid: pid of the process.
5180 * @interval: userspace pointer to the timeslice value.
5182 * this syscall writes the default timeslice value of a given process
5183 * into the user-space timespec buffer. A value of '0' means infinity.
5185 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5188 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5189 struct timespec __user *, interval)
5191 struct task_struct *p;
5192 unsigned int time_slice;
5203 p = find_process_by_pid(pid);
5207 retval = security_task_getscheduler(p);
5211 rq = task_rq_lock(p, &rf);
5213 if (p->sched_class->get_rr_interval)
5214 time_slice = p->sched_class->get_rr_interval(rq, p);
5215 task_rq_unlock(rq, p, &rf);
5218 jiffies_to_timespec(time_slice, &t);
5219 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5227 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5229 void sched_show_task(struct task_struct *p)
5231 unsigned long free = 0;
5233 unsigned long state = p->state;
5235 if (!try_get_task_stack(p))
5238 state = __ffs(state) + 1;
5239 printk(KERN_INFO "%-15.15s %c", p->comm,
5240 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5241 if (state == TASK_RUNNING)
5242 printk(KERN_CONT " running task ");
5243 #ifdef CONFIG_DEBUG_STACK_USAGE
5244 free = stack_not_used(p);
5249 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5251 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5252 task_pid_nr(p), ppid,
5253 (unsigned long)task_thread_info(p)->flags);
5255 print_worker_info(KERN_INFO, p);
5256 show_stack(p, NULL);
5260 void show_state_filter(unsigned long state_filter)
5262 struct task_struct *g, *p;
5264 #if BITS_PER_LONG == 32
5266 " task PC stack pid father\n");
5269 " task PC stack pid father\n");
5272 for_each_process_thread(g, p) {
5274 * reset the NMI-timeout, listing all files on a slow
5275 * console might take a lot of time:
5276 * Also, reset softlockup watchdogs on all CPUs, because
5277 * another CPU might be blocked waiting for us to process
5280 touch_nmi_watchdog();
5281 touch_all_softlockup_watchdogs();
5282 if (!state_filter || (p->state & state_filter))
5286 #ifdef CONFIG_SCHED_DEBUG
5288 sysrq_sched_debug_show();
5292 * Only show locks if all tasks are dumped:
5295 debug_show_all_locks();
5298 void init_idle_bootup_task(struct task_struct *idle)
5300 idle->sched_class = &idle_sched_class;
5304 * init_idle - set up an idle thread for a given CPU
5305 * @idle: task in question
5306 * @cpu: CPU the idle task belongs to
5308 * NOTE: this function does not set the idle thread's NEED_RESCHED
5309 * flag, to make booting more robust.
5311 void init_idle(struct task_struct *idle, int cpu)
5313 struct rq *rq = cpu_rq(cpu);
5314 unsigned long flags;
5316 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5317 raw_spin_lock(&rq->lock);
5319 __sched_fork(0, idle);
5320 idle->state = TASK_RUNNING;
5321 idle->se.exec_start = sched_clock();
5322 idle->flags |= PF_IDLE;
5324 kasan_unpoison_task_stack(idle);
5328 * Its possible that init_idle() gets called multiple times on a task,
5329 * in that case do_set_cpus_allowed() will not do the right thing.
5331 * And since this is boot we can forgo the serialization.
5333 set_cpus_allowed_common(idle, cpumask_of(cpu));
5336 * We're having a chicken and egg problem, even though we are
5337 * holding rq->lock, the CPU isn't yet set to this CPU so the
5338 * lockdep check in task_group() will fail.
5340 * Similar case to sched_fork(). / Alternatively we could
5341 * use task_rq_lock() here and obtain the other rq->lock.
5346 __set_task_cpu(idle, cpu);
5349 rq->curr = rq->idle = idle;
5350 idle->on_rq = TASK_ON_RQ_QUEUED;
5354 raw_spin_unlock(&rq->lock);
5355 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5357 /* Set the preempt count _outside_ the spinlocks! */
5358 init_idle_preempt_count(idle, cpu);
5361 * The idle tasks have their own, simple scheduling class:
5363 idle->sched_class = &idle_sched_class;
5364 ftrace_graph_init_idle_task(idle, cpu);
5365 vtime_init_idle(idle, cpu);
5367 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5371 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5372 const struct cpumask *trial)
5374 int ret = 1, trial_cpus;
5375 struct dl_bw *cur_dl_b;
5376 unsigned long flags;
5378 if (!cpumask_weight(cur))
5381 rcu_read_lock_sched();
5382 cur_dl_b = dl_bw_of(cpumask_any(cur));
5383 trial_cpus = cpumask_weight(trial);
5385 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5386 if (cur_dl_b->bw != -1 &&
5387 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5389 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5390 rcu_read_unlock_sched();
5395 int task_can_attach(struct task_struct *p,
5396 const struct cpumask *cs_cpus_allowed)
5401 * Kthreads which disallow setaffinity shouldn't be moved
5402 * to a new cpuset; we don't want to change their CPU
5403 * affinity and isolating such threads by their set of
5404 * allowed nodes is unnecessary. Thus, cpusets are not
5405 * applicable for such threads. This prevents checking for
5406 * success of set_cpus_allowed_ptr() on all attached tasks
5407 * before cpus_allowed may be changed.
5409 if (p->flags & PF_NO_SETAFFINITY) {
5415 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5417 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5422 unsigned long flags;
5424 rcu_read_lock_sched();
5425 dl_b = dl_bw_of(dest_cpu);
5426 raw_spin_lock_irqsave(&dl_b->lock, flags);
5427 cpus = dl_bw_cpus(dest_cpu);
5428 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5433 * We reserve space for this task in the destination
5434 * root_domain, as we can't fail after this point.
5435 * We will free resources in the source root_domain
5436 * later on (see set_cpus_allowed_dl()).
5438 __dl_add(dl_b, p->dl.dl_bw);
5440 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5441 rcu_read_unlock_sched();
5451 bool sched_smp_initialized __read_mostly;
5453 #ifdef CONFIG_NUMA_BALANCING
5454 /* Migrate current task p to target_cpu */
5455 int migrate_task_to(struct task_struct *p, int target_cpu)
5457 struct migration_arg arg = { p, target_cpu };
5458 int curr_cpu = task_cpu(p);
5460 if (curr_cpu == target_cpu)
5463 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5466 /* TODO: This is not properly updating schedstats */
5468 trace_sched_move_numa(p, curr_cpu, target_cpu);
5469 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5473 * Requeue a task on a given node and accurately track the number of NUMA
5474 * tasks on the runqueues
5476 void sched_setnuma(struct task_struct *p, int nid)
5478 bool queued, running;
5482 rq = task_rq_lock(p, &rf);
5483 queued = task_on_rq_queued(p);
5484 running = task_current(rq, p);
5487 dequeue_task(rq, p, DEQUEUE_SAVE);
5489 put_prev_task(rq, p);
5491 p->numa_preferred_nid = nid;
5494 enqueue_task(rq, p, ENQUEUE_RESTORE);
5496 set_curr_task(rq, p);
5497 task_rq_unlock(rq, p, &rf);
5499 #endif /* CONFIG_NUMA_BALANCING */
5501 #ifdef CONFIG_HOTPLUG_CPU
5503 * Ensure that the idle task is using init_mm right before its CPU goes
5506 void idle_task_exit(void)
5508 struct mm_struct *mm = current->active_mm;
5510 BUG_ON(cpu_online(smp_processor_id()));
5512 if (mm != &init_mm) {
5513 switch_mm_irqs_off(mm, &init_mm, current);
5514 finish_arch_post_lock_switch();
5520 * Since this CPU is going 'away' for a while, fold any nr_active delta
5521 * we might have. Assumes we're called after migrate_tasks() so that the
5522 * nr_active count is stable. We need to take the teardown thread which
5523 * is calling this into account, so we hand in adjust = 1 to the load
5526 * Also see the comment "Global load-average calculations".
5528 static void calc_load_migrate(struct rq *rq)
5530 long delta = calc_load_fold_active(rq, 1);
5532 atomic_long_add(delta, &calc_load_tasks);
5535 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5539 static const struct sched_class fake_sched_class = {
5540 .put_prev_task = put_prev_task_fake,
5543 static struct task_struct fake_task = {
5545 * Avoid pull_{rt,dl}_task()
5547 .prio = MAX_PRIO + 1,
5548 .sched_class = &fake_sched_class,
5552 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5553 * try_to_wake_up()->select_task_rq().
5555 * Called with rq->lock held even though we'er in stop_machine() and
5556 * there's no concurrency possible, we hold the required locks anyway
5557 * because of lock validation efforts.
5559 static void migrate_tasks(struct rq *dead_rq)
5561 struct rq *rq = dead_rq;
5562 struct task_struct *next, *stop = rq->stop;
5563 struct rq_flags rf, old_rf;
5567 * Fudge the rq selection such that the below task selection loop
5568 * doesn't get stuck on the currently eligible stop task.
5570 * We're currently inside stop_machine() and the rq is either stuck
5571 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5572 * either way we should never end up calling schedule() until we're
5578 * put_prev_task() and pick_next_task() sched
5579 * class method both need to have an up-to-date
5580 * value of rq->clock[_task]
5582 update_rq_clock(rq);
5586 * There's this thread running, bail when that's the only
5589 if (rq->nr_running == 1)
5593 * pick_next_task() assumes pinned rq->lock:
5595 rq_pin_lock(rq, &rf);
5596 next = pick_next_task(rq, &fake_task, &rf);
5598 next->sched_class->put_prev_task(rq, next);
5601 * Rules for changing task_struct::cpus_allowed are holding
5602 * both pi_lock and rq->lock, such that holding either
5603 * stabilizes the mask.
5605 * Drop rq->lock is not quite as disastrous as it usually is
5606 * because !cpu_active at this point, which means load-balance
5607 * will not interfere. Also, stop-machine.
5609 rq_unpin_lock(rq, &rf);
5610 raw_spin_unlock(&rq->lock);
5611 raw_spin_lock(&next->pi_lock);
5612 raw_spin_lock(&rq->lock);
5615 * Since we're inside stop-machine, _nothing_ should have
5616 * changed the task, WARN if weird stuff happened, because in
5617 * that case the above rq->lock drop is a fail too.
5619 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5620 raw_spin_unlock(&next->pi_lock);
5625 * __migrate_task() may return with a different
5626 * rq->lock held and a new cookie in 'rf', but we need
5627 * to preserve rf::clock_update_flags for 'dead_rq'.
5631 /* Find suitable destination for @next, with force if needed. */
5632 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5634 rq = __migrate_task(rq, next, dest_cpu);
5635 if (rq != dead_rq) {
5636 raw_spin_unlock(&rq->lock);
5638 raw_spin_lock(&rq->lock);
5641 raw_spin_unlock(&next->pi_lock);
5646 #endif /* CONFIG_HOTPLUG_CPU */
5648 void set_rq_online(struct rq *rq)
5651 const struct sched_class *class;
5653 cpumask_set_cpu(rq->cpu, rq->rd->online);
5656 for_each_class(class) {
5657 if (class->rq_online)
5658 class->rq_online(rq);
5663 void set_rq_offline(struct rq *rq)
5666 const struct sched_class *class;
5668 for_each_class(class) {
5669 if (class->rq_offline)
5670 class->rq_offline(rq);
5673 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5678 static void set_cpu_rq_start_time(unsigned int cpu)
5680 struct rq *rq = cpu_rq(cpu);
5682 rq->age_stamp = sched_clock_cpu(cpu);
5686 * used to mark begin/end of suspend/resume:
5688 static int num_cpus_frozen;
5691 * Update cpusets according to cpu_active mask. If cpusets are
5692 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5693 * around partition_sched_domains().
5695 * If we come here as part of a suspend/resume, don't touch cpusets because we
5696 * want to restore it back to its original state upon resume anyway.
5698 static void cpuset_cpu_active(void)
5700 if (cpuhp_tasks_frozen) {
5702 * num_cpus_frozen tracks how many CPUs are involved in suspend
5703 * resume sequence. As long as this is not the last online
5704 * operation in the resume sequence, just build a single sched
5705 * domain, ignoring cpusets.
5708 if (likely(num_cpus_frozen)) {
5709 partition_sched_domains(1, NULL, NULL);
5713 * This is the last CPU online operation. So fall through and
5714 * restore the original sched domains by considering the
5715 * cpuset configurations.
5718 cpuset_update_active_cpus(true);
5721 static int cpuset_cpu_inactive(unsigned int cpu)
5723 unsigned long flags;
5728 if (!cpuhp_tasks_frozen) {
5729 rcu_read_lock_sched();
5730 dl_b = dl_bw_of(cpu);
5732 raw_spin_lock_irqsave(&dl_b->lock, flags);
5733 cpus = dl_bw_cpus(cpu);
5734 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5735 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5737 rcu_read_unlock_sched();
5741 cpuset_update_active_cpus(false);
5744 partition_sched_domains(1, NULL, NULL);
5749 int sched_cpu_activate(unsigned int cpu)
5751 struct rq *rq = cpu_rq(cpu);
5752 unsigned long flags;
5754 set_cpu_active(cpu, true);
5756 if (sched_smp_initialized) {
5757 sched_domains_numa_masks_set(cpu);
5758 cpuset_cpu_active();
5762 * Put the rq online, if not already. This happens:
5764 * 1) In the early boot process, because we build the real domains
5765 * after all CPUs have been brought up.
5767 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5770 raw_spin_lock_irqsave(&rq->lock, flags);
5772 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5775 raw_spin_unlock_irqrestore(&rq->lock, flags);
5777 update_max_interval();
5782 int sched_cpu_deactivate(unsigned int cpu)
5786 set_cpu_active(cpu, false);
5788 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5789 * users of this state to go away such that all new such users will
5792 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
5793 * not imply sync_sched(), so wait for both.
5795 * Do sync before park smpboot threads to take care the rcu boost case.
5797 if (IS_ENABLED(CONFIG_PREEMPT))
5798 synchronize_rcu_mult(call_rcu, call_rcu_sched);
5802 if (!sched_smp_initialized)
5805 ret = cpuset_cpu_inactive(cpu);
5807 set_cpu_active(cpu, true);
5810 sched_domains_numa_masks_clear(cpu);
5814 static void sched_rq_cpu_starting(unsigned int cpu)
5816 struct rq *rq = cpu_rq(cpu);
5818 rq->calc_load_update = calc_load_update;
5819 update_max_interval();
5822 int sched_cpu_starting(unsigned int cpu)
5824 set_cpu_rq_start_time(cpu);
5825 sched_rq_cpu_starting(cpu);
5829 #ifdef CONFIG_HOTPLUG_CPU
5830 int sched_cpu_dying(unsigned int cpu)
5832 struct rq *rq = cpu_rq(cpu);
5833 unsigned long flags;
5835 /* Handle pending wakeups and then migrate everything off */
5836 sched_ttwu_pending();
5837 raw_spin_lock_irqsave(&rq->lock, flags);
5839 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5843 BUG_ON(rq->nr_running != 1);
5844 raw_spin_unlock_irqrestore(&rq->lock, flags);
5845 calc_load_migrate(rq);
5846 update_max_interval();
5847 nohz_balance_exit_idle(cpu);
5853 #ifdef CONFIG_SCHED_SMT
5854 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5856 static void sched_init_smt(void)
5859 * We've enumerated all CPUs and will assume that if any CPU
5860 * has SMT siblings, CPU0 will too.
5862 if (cpumask_weight(cpu_smt_mask(0)) > 1)
5863 static_branch_enable(&sched_smt_present);
5866 static inline void sched_init_smt(void) { }
5869 void __init sched_init_smp(void)
5871 cpumask_var_t non_isolated_cpus;
5873 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
5874 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
5879 * There's no userspace yet to cause hotplug operations; hence all the
5880 * CPU masks are stable and all blatant races in the below code cannot
5883 mutex_lock(&sched_domains_mutex);
5884 init_sched_domains(cpu_active_mask);
5885 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
5886 if (cpumask_empty(non_isolated_cpus))
5887 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
5888 mutex_unlock(&sched_domains_mutex);
5890 /* Move init over to a non-isolated CPU */
5891 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
5893 sched_init_granularity();
5894 free_cpumask_var(non_isolated_cpus);
5896 init_sched_rt_class();
5897 init_sched_dl_class();
5900 sched_clock_init_late();
5902 sched_smp_initialized = true;
5905 static int __init migration_init(void)
5907 sched_rq_cpu_starting(smp_processor_id());
5910 early_initcall(migration_init);
5913 void __init sched_init_smp(void)
5915 sched_init_granularity();
5916 sched_clock_init_late();
5918 #endif /* CONFIG_SMP */
5920 int in_sched_functions(unsigned long addr)
5922 return in_lock_functions(addr) ||
5923 (addr >= (unsigned long)__sched_text_start
5924 && addr < (unsigned long)__sched_text_end);
5927 #ifdef CONFIG_CGROUP_SCHED
5929 * Default task group.
5930 * Every task in system belongs to this group at bootup.
5932 struct task_group root_task_group;
5933 LIST_HEAD(task_groups);
5935 /* Cacheline aligned slab cache for task_group */
5936 static struct kmem_cache *task_group_cache __read_mostly;
5939 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5940 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5942 #define WAIT_TABLE_BITS 8
5943 #define WAIT_TABLE_SIZE (1 << WAIT_TABLE_BITS)
5944 static wait_queue_head_t bit_wait_table[WAIT_TABLE_SIZE] __cacheline_aligned;
5946 wait_queue_head_t *bit_waitqueue(void *word, int bit)
5948 const int shift = BITS_PER_LONG == 32 ? 5 : 6;
5949 unsigned long val = (unsigned long)word << shift | bit;
5951 return bit_wait_table + hash_long(val, WAIT_TABLE_BITS);
5953 EXPORT_SYMBOL(bit_waitqueue);
5955 void __init sched_init(void)
5958 unsigned long alloc_size = 0, ptr;
5962 for (i = 0; i < WAIT_TABLE_SIZE; i++)
5963 init_waitqueue_head(bit_wait_table + i);
5965 #ifdef CONFIG_FAIR_GROUP_SCHED
5966 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5968 #ifdef CONFIG_RT_GROUP_SCHED
5969 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5972 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5974 #ifdef CONFIG_FAIR_GROUP_SCHED
5975 root_task_group.se = (struct sched_entity **)ptr;
5976 ptr += nr_cpu_ids * sizeof(void **);
5978 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5979 ptr += nr_cpu_ids * sizeof(void **);
5981 #endif /* CONFIG_FAIR_GROUP_SCHED */
5982 #ifdef CONFIG_RT_GROUP_SCHED
5983 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5984 ptr += nr_cpu_ids * sizeof(void **);
5986 root_task_group.rt_rq = (struct rt_rq **)ptr;
5987 ptr += nr_cpu_ids * sizeof(void **);
5989 #endif /* CONFIG_RT_GROUP_SCHED */
5991 #ifdef CONFIG_CPUMASK_OFFSTACK
5992 for_each_possible_cpu(i) {
5993 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5994 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5995 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
5996 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5998 #endif /* CONFIG_CPUMASK_OFFSTACK */
6000 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
6001 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6004 init_defrootdomain();
6007 #ifdef CONFIG_RT_GROUP_SCHED
6008 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6009 global_rt_period(), global_rt_runtime());
6010 #endif /* CONFIG_RT_GROUP_SCHED */
6012 #ifdef CONFIG_CGROUP_SCHED
6013 task_group_cache = KMEM_CACHE(task_group, 0);
6015 list_add(&root_task_group.list, &task_groups);
6016 INIT_LIST_HEAD(&root_task_group.children);
6017 INIT_LIST_HEAD(&root_task_group.siblings);
6018 autogroup_init(&init_task);
6019 #endif /* CONFIG_CGROUP_SCHED */
6021 for_each_possible_cpu(i) {
6025 raw_spin_lock_init(&rq->lock);
6027 rq->calc_load_active = 0;
6028 rq->calc_load_update = jiffies + LOAD_FREQ;
6029 init_cfs_rq(&rq->cfs);
6030 init_rt_rq(&rq->rt);
6031 init_dl_rq(&rq->dl);
6032 #ifdef CONFIG_FAIR_GROUP_SCHED
6033 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6034 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6035 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6037 * How much CPU bandwidth does root_task_group get?
6039 * In case of task-groups formed thr' the cgroup filesystem, it
6040 * gets 100% of the CPU resources in the system. This overall
6041 * system CPU resource is divided among the tasks of
6042 * root_task_group and its child task-groups in a fair manner,
6043 * based on each entity's (task or task-group's) weight
6044 * (se->load.weight).
6046 * In other words, if root_task_group has 10 tasks of weight
6047 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6048 * then A0's share of the CPU resource is:
6050 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6052 * We achieve this by letting root_task_group's tasks sit
6053 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6055 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6056 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6057 #endif /* CONFIG_FAIR_GROUP_SCHED */
6059 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6060 #ifdef CONFIG_RT_GROUP_SCHED
6061 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6064 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6065 rq->cpu_load[j] = 0;
6070 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6071 rq->balance_callback = NULL;
6072 rq->active_balance = 0;
6073 rq->next_balance = jiffies;
6078 rq->avg_idle = 2*sysctl_sched_migration_cost;
6079 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6081 INIT_LIST_HEAD(&rq->cfs_tasks);
6083 rq_attach_root(rq, &def_root_domain);
6084 #ifdef CONFIG_NO_HZ_COMMON
6085 rq->last_load_update_tick = jiffies;
6088 #ifdef CONFIG_NO_HZ_FULL
6089 rq->last_sched_tick = 0;
6091 #endif /* CONFIG_SMP */
6093 atomic_set(&rq->nr_iowait, 0);
6096 set_load_weight(&init_task);
6099 * The boot idle thread does lazy MMU switching as well:
6101 atomic_inc(&init_mm.mm_count);
6102 enter_lazy_tlb(&init_mm, current);
6105 * Make us the idle thread. Technically, schedule() should not be
6106 * called from this thread, however somewhere below it might be,
6107 * but because we are the idle thread, we just pick up running again
6108 * when this runqueue becomes "idle".
6110 init_idle(current, smp_processor_id());
6112 calc_load_update = jiffies + LOAD_FREQ;
6115 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6116 /* May be allocated at isolcpus cmdline parse time */
6117 if (cpu_isolated_map == NULL)
6118 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6119 idle_thread_set_boot_cpu();
6120 set_cpu_rq_start_time(smp_processor_id());
6122 init_sched_fair_class();
6126 scheduler_running = 1;
6129 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6130 static inline int preempt_count_equals(int preempt_offset)
6132 int nested = preempt_count() + rcu_preempt_depth();
6134 return (nested == preempt_offset);
6137 void __might_sleep(const char *file, int line, int preempt_offset)
6140 * Blocking primitives will set (and therefore destroy) current->state,
6141 * since we will exit with TASK_RUNNING make sure we enter with it,
6142 * otherwise we will destroy state.
6144 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6145 "do not call blocking ops when !TASK_RUNNING; "
6146 "state=%lx set at [<%p>] %pS\n",
6148 (void *)current->task_state_change,
6149 (void *)current->task_state_change);
6151 ___might_sleep(file, line, preempt_offset);
6153 EXPORT_SYMBOL(__might_sleep);
6155 void ___might_sleep(const char *file, int line, int preempt_offset)
6157 /* Ratelimiting timestamp: */
6158 static unsigned long prev_jiffy;
6160 unsigned long preempt_disable_ip;
6162 /* WARN_ON_ONCE() by default, no rate limit required: */
6165 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6166 !is_idle_task(current)) ||
6167 system_state != SYSTEM_RUNNING || oops_in_progress)
6169 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6171 prev_jiffy = jiffies;
6173 /* Save this before calling printk(), since that will clobber it: */
6174 preempt_disable_ip = get_preempt_disable_ip(current);
6177 "BUG: sleeping function called from invalid context at %s:%d\n",
6180 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6181 in_atomic(), irqs_disabled(),
6182 current->pid, current->comm);
6184 if (task_stack_end_corrupted(current))
6185 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6187 debug_show_held_locks(current);
6188 if (irqs_disabled())
6189 print_irqtrace_events(current);
6190 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6191 && !preempt_count_equals(preempt_offset)) {
6192 pr_err("Preemption disabled at:");
6193 print_ip_sym(preempt_disable_ip);
6197 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6199 EXPORT_SYMBOL(___might_sleep);
6202 #ifdef CONFIG_MAGIC_SYSRQ
6203 void normalize_rt_tasks(void)
6205 struct task_struct *g, *p;
6206 struct sched_attr attr = {
6207 .sched_policy = SCHED_NORMAL,
6210 read_lock(&tasklist_lock);
6211 for_each_process_thread(g, p) {
6213 * Only normalize user tasks:
6215 if (p->flags & PF_KTHREAD)
6218 p->se.exec_start = 0;
6219 schedstat_set(p->se.statistics.wait_start, 0);
6220 schedstat_set(p->se.statistics.sleep_start, 0);
6221 schedstat_set(p->se.statistics.block_start, 0);
6223 if (!dl_task(p) && !rt_task(p)) {
6225 * Renice negative nice level userspace
6228 if (task_nice(p) < 0)
6229 set_user_nice(p, 0);
6233 __sched_setscheduler(p, &attr, false, false);
6235 read_unlock(&tasklist_lock);
6238 #endif /* CONFIG_MAGIC_SYSRQ */
6240 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6242 * These functions are only useful for the IA64 MCA handling, or kdb.
6244 * They can only be called when the whole system has been
6245 * stopped - every CPU needs to be quiescent, and no scheduling
6246 * activity can take place. Using them for anything else would
6247 * be a serious bug, and as a result, they aren't even visible
6248 * under any other configuration.
6252 * curr_task - return the current task for a given CPU.
6253 * @cpu: the processor in question.
6255 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6257 * Return: The current task for @cpu.
6259 struct task_struct *curr_task(int cpu)
6261 return cpu_curr(cpu);
6264 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6268 * set_curr_task - set the current task for a given CPU.
6269 * @cpu: the processor in question.
6270 * @p: the task pointer to set.
6272 * Description: This function must only be used when non-maskable interrupts
6273 * are serviced on a separate stack. It allows the architecture to switch the
6274 * notion of the current task on a CPU in a non-blocking manner. This function
6275 * must be called with all CPU's synchronized, and interrupts disabled, the
6276 * and caller must save the original value of the current task (see
6277 * curr_task() above) and restore that value before reenabling interrupts and
6278 * re-starting the system.
6280 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6282 void ia64_set_curr_task(int cpu, struct task_struct *p)
6289 #ifdef CONFIG_CGROUP_SCHED
6290 /* task_group_lock serializes the addition/removal of task groups */
6291 static DEFINE_SPINLOCK(task_group_lock);
6293 static void sched_free_group(struct task_group *tg)
6295 free_fair_sched_group(tg);
6296 free_rt_sched_group(tg);
6298 kmem_cache_free(task_group_cache, tg);
6301 /* allocate runqueue etc for a new task group */
6302 struct task_group *sched_create_group(struct task_group *parent)
6304 struct task_group *tg;
6306 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6308 return ERR_PTR(-ENOMEM);
6310 if (!alloc_fair_sched_group(tg, parent))
6313 if (!alloc_rt_sched_group(tg, parent))
6319 sched_free_group(tg);
6320 return ERR_PTR(-ENOMEM);
6323 void sched_online_group(struct task_group *tg, struct task_group *parent)
6325 unsigned long flags;
6327 spin_lock_irqsave(&task_group_lock, flags);
6328 list_add_rcu(&tg->list, &task_groups);
6330 /* Root should already exist: */
6333 tg->parent = parent;
6334 INIT_LIST_HEAD(&tg->children);
6335 list_add_rcu(&tg->siblings, &parent->children);
6336 spin_unlock_irqrestore(&task_group_lock, flags);
6338 online_fair_sched_group(tg);
6341 /* rcu callback to free various structures associated with a task group */
6342 static void sched_free_group_rcu(struct rcu_head *rhp)
6344 /* Now it should be safe to free those cfs_rqs: */
6345 sched_free_group(container_of(rhp, struct task_group, rcu));
6348 void sched_destroy_group(struct task_group *tg)
6350 /* Wait for possible concurrent references to cfs_rqs complete: */
6351 call_rcu(&tg->rcu, sched_free_group_rcu);
6354 void sched_offline_group(struct task_group *tg)
6356 unsigned long flags;
6358 /* End participation in shares distribution: */
6359 unregister_fair_sched_group(tg);
6361 spin_lock_irqsave(&task_group_lock, flags);
6362 list_del_rcu(&tg->list);
6363 list_del_rcu(&tg->siblings);
6364 spin_unlock_irqrestore(&task_group_lock, flags);
6367 static void sched_change_group(struct task_struct *tsk, int type)
6369 struct task_group *tg;
6372 * All callers are synchronized by task_rq_lock(); we do not use RCU
6373 * which is pointless here. Thus, we pass "true" to task_css_check()
6374 * to prevent lockdep warnings.
6376 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6377 struct task_group, css);
6378 tg = autogroup_task_group(tsk, tg);
6379 tsk->sched_task_group = tg;
6381 #ifdef CONFIG_FAIR_GROUP_SCHED
6382 if (tsk->sched_class->task_change_group)
6383 tsk->sched_class->task_change_group(tsk, type);
6386 set_task_rq(tsk, task_cpu(tsk));
6390 * Change task's runqueue when it moves between groups.
6392 * The caller of this function should have put the task in its new group by
6393 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6396 void sched_move_task(struct task_struct *tsk)
6398 int queued, running;
6402 rq = task_rq_lock(tsk, &rf);
6403 update_rq_clock(rq);
6405 running = task_current(rq, tsk);
6406 queued = task_on_rq_queued(tsk);
6409 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
6411 put_prev_task(rq, tsk);
6413 sched_change_group(tsk, TASK_MOVE_GROUP);
6416 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
6418 set_curr_task(rq, tsk);
6420 task_rq_unlock(rq, tsk, &rf);
6422 #endif /* CONFIG_CGROUP_SCHED */
6424 #ifdef CONFIG_RT_GROUP_SCHED
6426 * Ensure that the real time constraints are schedulable.
6428 static DEFINE_MUTEX(rt_constraints_mutex);
6430 /* Must be called with tasklist_lock held */
6431 static inline int tg_has_rt_tasks(struct task_group *tg)
6433 struct task_struct *g, *p;
6436 * Autogroups do not have RT tasks; see autogroup_create().
6438 if (task_group_is_autogroup(tg))
6441 for_each_process_thread(g, p) {
6442 if (rt_task(p) && task_group(p) == tg)
6449 struct rt_schedulable_data {
6450 struct task_group *tg;
6455 static int tg_rt_schedulable(struct task_group *tg, void *data)
6457 struct rt_schedulable_data *d = data;
6458 struct task_group *child;
6459 unsigned long total, sum = 0;
6460 u64 period, runtime;
6462 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6463 runtime = tg->rt_bandwidth.rt_runtime;
6466 period = d->rt_period;
6467 runtime = d->rt_runtime;
6471 * Cannot have more runtime than the period.
6473 if (runtime > period && runtime != RUNTIME_INF)
6477 * Ensure we don't starve existing RT tasks.
6479 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
6482 total = to_ratio(period, runtime);
6485 * Nobody can have more than the global setting allows.
6487 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
6491 * The sum of our children's runtime should not exceed our own.
6493 list_for_each_entry_rcu(child, &tg->children, siblings) {
6494 period = ktime_to_ns(child->rt_bandwidth.rt_period);
6495 runtime = child->rt_bandwidth.rt_runtime;
6497 if (child == d->tg) {
6498 period = d->rt_period;
6499 runtime = d->rt_runtime;
6502 sum += to_ratio(period, runtime);
6511 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
6515 struct rt_schedulable_data data = {
6517 .rt_period = period,
6518 .rt_runtime = runtime,
6522 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
6528 static int tg_set_rt_bandwidth(struct task_group *tg,
6529 u64 rt_period, u64 rt_runtime)
6534 * Disallowing the root group RT runtime is BAD, it would disallow the
6535 * kernel creating (and or operating) RT threads.
6537 if (tg == &root_task_group && rt_runtime == 0)
6540 /* No period doesn't make any sense. */
6544 mutex_lock(&rt_constraints_mutex);
6545 read_lock(&tasklist_lock);
6546 err = __rt_schedulable(tg, rt_period, rt_runtime);
6550 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6551 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
6552 tg->rt_bandwidth.rt_runtime = rt_runtime;
6554 for_each_possible_cpu(i) {
6555 struct rt_rq *rt_rq = tg->rt_rq[i];
6557 raw_spin_lock(&rt_rq->rt_runtime_lock);
6558 rt_rq->rt_runtime = rt_runtime;
6559 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6561 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6563 read_unlock(&tasklist_lock);
6564 mutex_unlock(&rt_constraints_mutex);
6569 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
6571 u64 rt_runtime, rt_period;
6573 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6574 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
6575 if (rt_runtime_us < 0)
6576 rt_runtime = RUNTIME_INF;
6578 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6581 static long sched_group_rt_runtime(struct task_group *tg)
6585 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
6588 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
6589 do_div(rt_runtime_us, NSEC_PER_USEC);
6590 return rt_runtime_us;
6593 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
6595 u64 rt_runtime, rt_period;
6597 rt_period = rt_period_us * NSEC_PER_USEC;
6598 rt_runtime = tg->rt_bandwidth.rt_runtime;
6600 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6603 static long sched_group_rt_period(struct task_group *tg)
6607 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
6608 do_div(rt_period_us, NSEC_PER_USEC);
6609 return rt_period_us;
6611 #endif /* CONFIG_RT_GROUP_SCHED */
6613 #ifdef CONFIG_RT_GROUP_SCHED
6614 static int sched_rt_global_constraints(void)
6618 mutex_lock(&rt_constraints_mutex);
6619 read_lock(&tasklist_lock);
6620 ret = __rt_schedulable(NULL, 0, 0);
6621 read_unlock(&tasklist_lock);
6622 mutex_unlock(&rt_constraints_mutex);
6627 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
6629 /* Don't accept realtime tasks when there is no way for them to run */
6630 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
6636 #else /* !CONFIG_RT_GROUP_SCHED */
6637 static int sched_rt_global_constraints(void)
6639 unsigned long flags;
6642 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
6643 for_each_possible_cpu(i) {
6644 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
6646 raw_spin_lock(&rt_rq->rt_runtime_lock);
6647 rt_rq->rt_runtime = global_rt_runtime();
6648 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6650 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
6654 #endif /* CONFIG_RT_GROUP_SCHED */
6656 static int sched_dl_global_validate(void)
6658 u64 runtime = global_rt_runtime();
6659 u64 period = global_rt_period();
6660 u64 new_bw = to_ratio(period, runtime);
6663 unsigned long flags;
6666 * Here we want to check the bandwidth not being set to some
6667 * value smaller than the currently allocated bandwidth in
6668 * any of the root_domains.
6670 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
6671 * cycling on root_domains... Discussion on different/better
6672 * solutions is welcome!
6674 for_each_possible_cpu(cpu) {
6675 rcu_read_lock_sched();
6676 dl_b = dl_bw_of(cpu);
6678 raw_spin_lock_irqsave(&dl_b->lock, flags);
6679 if (new_bw < dl_b->total_bw)
6681 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
6683 rcu_read_unlock_sched();
6692 static void sched_dl_do_global(void)
6697 unsigned long flags;
6699 def_dl_bandwidth.dl_period = global_rt_period();
6700 def_dl_bandwidth.dl_runtime = global_rt_runtime();
6702 if (global_rt_runtime() != RUNTIME_INF)
6703 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
6706 * FIXME: As above...
6708 for_each_possible_cpu(cpu) {
6709 rcu_read_lock_sched();
6710 dl_b = dl_bw_of(cpu);
6712 raw_spin_lock_irqsave(&dl_b->lock, flags);
6714 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
6716 rcu_read_unlock_sched();
6720 static int sched_rt_global_validate(void)
6722 if (sysctl_sched_rt_period <= 0)
6725 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
6726 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
6732 static void sched_rt_do_global(void)
6734 def_rt_bandwidth.rt_runtime = global_rt_runtime();
6735 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
6738 int sched_rt_handler(struct ctl_table *table, int write,
6739 void __user *buffer, size_t *lenp,
6742 int old_period, old_runtime;
6743 static DEFINE_MUTEX(mutex);
6747 old_period = sysctl_sched_rt_period;
6748 old_runtime = sysctl_sched_rt_runtime;
6750 ret = proc_dointvec(table, write, buffer, lenp, ppos);
6752 if (!ret && write) {
6753 ret = sched_rt_global_validate();
6757 ret = sched_dl_global_validate();
6761 ret = sched_rt_global_constraints();
6765 sched_rt_do_global();
6766 sched_dl_do_global();
6770 sysctl_sched_rt_period = old_period;
6771 sysctl_sched_rt_runtime = old_runtime;
6773 mutex_unlock(&mutex);
6778 int sched_rr_handler(struct ctl_table *table, int write,
6779 void __user *buffer, size_t *lenp,
6783 static DEFINE_MUTEX(mutex);
6786 ret = proc_dointvec(table, write, buffer, lenp, ppos);
6788 * Make sure that internally we keep jiffies.
6789 * Also, writing zero resets the timeslice to default:
6791 if (!ret && write) {
6792 sched_rr_timeslice =
6793 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
6794 msecs_to_jiffies(sysctl_sched_rr_timeslice);
6796 mutex_unlock(&mutex);
6800 #ifdef CONFIG_CGROUP_SCHED
6802 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6804 return css ? container_of(css, struct task_group, css) : NULL;
6807 static struct cgroup_subsys_state *
6808 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6810 struct task_group *parent = css_tg(parent_css);
6811 struct task_group *tg;
6814 /* This is early initialization for the top cgroup */
6815 return &root_task_group.css;
6818 tg = sched_create_group(parent);
6820 return ERR_PTR(-ENOMEM);
6822 sched_online_group(tg, parent);
6827 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6829 struct task_group *tg = css_tg(css);
6831 sched_offline_group(tg);
6834 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6836 struct task_group *tg = css_tg(css);
6839 * Relies on the RCU grace period between css_released() and this.
6841 sched_free_group(tg);
6845 * This is called before wake_up_new_task(), therefore we really only
6846 * have to set its group bits, all the other stuff does not apply.
6848 static void cpu_cgroup_fork(struct task_struct *task)
6853 rq = task_rq_lock(task, &rf);
6855 update_rq_clock(rq);
6856 sched_change_group(task, TASK_SET_GROUP);
6858 task_rq_unlock(rq, task, &rf);
6861 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6863 struct task_struct *task;
6864 struct cgroup_subsys_state *css;
6867 cgroup_taskset_for_each(task, css, tset) {
6868 #ifdef CONFIG_RT_GROUP_SCHED
6869 if (!sched_rt_can_attach(css_tg(css), task))
6872 /* We don't support RT-tasks being in separate groups */
6873 if (task->sched_class != &fair_sched_class)
6877 * Serialize against wake_up_new_task() such that if its
6878 * running, we're sure to observe its full state.
6880 raw_spin_lock_irq(&task->pi_lock);
6882 * Avoid calling sched_move_task() before wake_up_new_task()
6883 * has happened. This would lead to problems with PELT, due to
6884 * move wanting to detach+attach while we're not attached yet.
6886 if (task->state == TASK_NEW)
6888 raw_spin_unlock_irq(&task->pi_lock);
6896 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6898 struct task_struct *task;
6899 struct cgroup_subsys_state *css;
6901 cgroup_taskset_for_each(task, css, tset)
6902 sched_move_task(task);
6905 #ifdef CONFIG_FAIR_GROUP_SCHED
6906 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6907 struct cftype *cftype, u64 shareval)
6909 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6912 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6915 struct task_group *tg = css_tg(css);
6917 return (u64) scale_load_down(tg->shares);
6920 #ifdef CONFIG_CFS_BANDWIDTH
6921 static DEFINE_MUTEX(cfs_constraints_mutex);
6923 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6924 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6926 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6928 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6930 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6931 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6933 if (tg == &root_task_group)
6937 * Ensure we have at some amount of bandwidth every period. This is
6938 * to prevent reaching a state of large arrears when throttled via
6939 * entity_tick() resulting in prolonged exit starvation.
6941 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6945 * Likewise, bound things on the otherside by preventing insane quota
6946 * periods. This also allows us to normalize in computing quota
6949 if (period > max_cfs_quota_period)
6953 * Prevent race between setting of cfs_rq->runtime_enabled and
6954 * unthrottle_offline_cfs_rqs().
6957 mutex_lock(&cfs_constraints_mutex);
6958 ret = __cfs_schedulable(tg, period, quota);
6962 runtime_enabled = quota != RUNTIME_INF;
6963 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6965 * If we need to toggle cfs_bandwidth_used, off->on must occur
6966 * before making related changes, and on->off must occur afterwards
6968 if (runtime_enabled && !runtime_was_enabled)
6969 cfs_bandwidth_usage_inc();
6970 raw_spin_lock_irq(&cfs_b->lock);
6971 cfs_b->period = ns_to_ktime(period);
6972 cfs_b->quota = quota;
6974 __refill_cfs_bandwidth_runtime(cfs_b);
6976 /* Restart the period timer (if active) to handle new period expiry: */
6977 if (runtime_enabled)
6978 start_cfs_bandwidth(cfs_b);
6980 raw_spin_unlock_irq(&cfs_b->lock);
6982 for_each_online_cpu(i) {
6983 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6984 struct rq *rq = cfs_rq->rq;
6986 raw_spin_lock_irq(&rq->lock);
6987 cfs_rq->runtime_enabled = runtime_enabled;
6988 cfs_rq->runtime_remaining = 0;
6990 if (cfs_rq->throttled)
6991 unthrottle_cfs_rq(cfs_rq);
6992 raw_spin_unlock_irq(&rq->lock);
6994 if (runtime_was_enabled && !runtime_enabled)
6995 cfs_bandwidth_usage_dec();
6997 mutex_unlock(&cfs_constraints_mutex);
7003 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7007 period = ktime_to_ns(tg->cfs_bandwidth.period);
7008 if (cfs_quota_us < 0)
7009 quota = RUNTIME_INF;
7011 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7013 return tg_set_cfs_bandwidth(tg, period, quota);
7016 long tg_get_cfs_quota(struct task_group *tg)
7020 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7023 quota_us = tg->cfs_bandwidth.quota;
7024 do_div(quota_us, NSEC_PER_USEC);
7029 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7033 period = (u64)cfs_period_us * NSEC_PER_USEC;
7034 quota = tg->cfs_bandwidth.quota;
7036 return tg_set_cfs_bandwidth(tg, period, quota);
7039 long tg_get_cfs_period(struct task_group *tg)
7043 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7044 do_div(cfs_period_us, NSEC_PER_USEC);
7046 return cfs_period_us;
7049 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7052 return tg_get_cfs_quota(css_tg(css));
7055 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7056 struct cftype *cftype, s64 cfs_quota_us)
7058 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7061 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7064 return tg_get_cfs_period(css_tg(css));
7067 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7068 struct cftype *cftype, u64 cfs_period_us)
7070 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7073 struct cfs_schedulable_data {
7074 struct task_group *tg;
7079 * normalize group quota/period to be quota/max_period
7080 * note: units are usecs
7082 static u64 normalize_cfs_quota(struct task_group *tg,
7083 struct cfs_schedulable_data *d)
7091 period = tg_get_cfs_period(tg);
7092 quota = tg_get_cfs_quota(tg);
7095 /* note: these should typically be equivalent */
7096 if (quota == RUNTIME_INF || quota == -1)
7099 return to_ratio(period, quota);
7102 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7104 struct cfs_schedulable_data *d = data;
7105 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7106 s64 quota = 0, parent_quota = -1;
7109 quota = RUNTIME_INF;
7111 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7113 quota = normalize_cfs_quota(tg, d);
7114 parent_quota = parent_b->hierarchical_quota;
7117 * Ensure max(child_quota) <= parent_quota, inherit when no
7120 if (quota == RUNTIME_INF)
7121 quota = parent_quota;
7122 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7125 cfs_b->hierarchical_quota = quota;
7130 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7133 struct cfs_schedulable_data data = {
7139 if (quota != RUNTIME_INF) {
7140 do_div(data.period, NSEC_PER_USEC);
7141 do_div(data.quota, NSEC_PER_USEC);
7145 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7151 static int cpu_stats_show(struct seq_file *sf, void *v)
7153 struct task_group *tg = css_tg(seq_css(sf));
7154 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7156 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7157 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7158 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7162 #endif /* CONFIG_CFS_BANDWIDTH */
7163 #endif /* CONFIG_FAIR_GROUP_SCHED */
7165 #ifdef CONFIG_RT_GROUP_SCHED
7166 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7167 struct cftype *cft, s64 val)
7169 return sched_group_set_rt_runtime(css_tg(css), val);
7172 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7175 return sched_group_rt_runtime(css_tg(css));
7178 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7179 struct cftype *cftype, u64 rt_period_us)
7181 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7184 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7187 return sched_group_rt_period(css_tg(css));
7189 #endif /* CONFIG_RT_GROUP_SCHED */
7191 static struct cftype cpu_files[] = {
7192 #ifdef CONFIG_FAIR_GROUP_SCHED
7195 .read_u64 = cpu_shares_read_u64,
7196 .write_u64 = cpu_shares_write_u64,
7199 #ifdef CONFIG_CFS_BANDWIDTH
7201 .name = "cfs_quota_us",
7202 .read_s64 = cpu_cfs_quota_read_s64,
7203 .write_s64 = cpu_cfs_quota_write_s64,
7206 .name = "cfs_period_us",
7207 .read_u64 = cpu_cfs_period_read_u64,
7208 .write_u64 = cpu_cfs_period_write_u64,
7212 .seq_show = cpu_stats_show,
7215 #ifdef CONFIG_RT_GROUP_SCHED
7217 .name = "rt_runtime_us",
7218 .read_s64 = cpu_rt_runtime_read,
7219 .write_s64 = cpu_rt_runtime_write,
7222 .name = "rt_period_us",
7223 .read_u64 = cpu_rt_period_read_uint,
7224 .write_u64 = cpu_rt_period_write_uint,
7230 struct cgroup_subsys cpu_cgrp_subsys = {
7231 .css_alloc = cpu_cgroup_css_alloc,
7232 .css_released = cpu_cgroup_css_released,
7233 .css_free = cpu_cgroup_css_free,
7234 .fork = cpu_cgroup_fork,
7235 .can_attach = cpu_cgroup_can_attach,
7236 .attach = cpu_cgroup_attach,
7237 .legacy_cftypes = cpu_files,
7241 #endif /* CONFIG_CGROUP_SCHED */
7243 void dump_cpu_task(int cpu)
7245 pr_info("Task dump for CPU %d:\n", cpu);
7246 sched_show_task(cpu_curr(cpu));
7250 * Nice levels are multiplicative, with a gentle 10% change for every
7251 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7252 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7253 * that remained on nice 0.
7255 * The "10% effect" is relative and cumulative: from _any_ nice level,
7256 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7257 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7258 * If a task goes up by ~10% and another task goes down by ~10% then
7259 * the relative distance between them is ~25%.)
7261 const int sched_prio_to_weight[40] = {
7262 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7263 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7264 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7265 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7266 /* 0 */ 1024, 820, 655, 526, 423,
7267 /* 5 */ 335, 272, 215, 172, 137,
7268 /* 10 */ 110, 87, 70, 56, 45,
7269 /* 15 */ 36, 29, 23, 18, 15,
7273 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7275 * In cases where the weight does not change often, we can use the
7276 * precalculated inverse to speed up arithmetics by turning divisions
7277 * into multiplications:
7279 const u32 sched_prio_to_wmult[40] = {
7280 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7281 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7282 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7283 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7284 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7285 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7286 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7287 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,