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);
1093 update_rq_clock(rq);
1095 if (p->flags & PF_KTHREAD) {
1097 * Kernel threads are allowed on online && !active CPUs
1099 cpu_valid_mask = cpu_online_mask;
1103 * Must re-check here, to close a race against __kthread_bind(),
1104 * sched_setaffinity() is not guaranteed to observe the flag.
1106 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1111 if (cpumask_equal(&p->cpus_allowed, new_mask))
1114 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1119 do_set_cpus_allowed(p, new_mask);
1121 if (p->flags & PF_KTHREAD) {
1123 * For kernel threads that do indeed end up on online &&
1124 * !active we want to ensure they are strict per-CPU threads.
1126 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1127 !cpumask_intersects(new_mask, cpu_active_mask) &&
1128 p->nr_cpus_allowed != 1);
1131 /* Can the task run on the task's current CPU? If so, we're done */
1132 if (cpumask_test_cpu(task_cpu(p), new_mask))
1135 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1136 if (task_running(rq, p) || p->state == TASK_WAKING) {
1137 struct migration_arg arg = { p, dest_cpu };
1138 /* Need help from migration thread: drop lock and wait. */
1139 task_rq_unlock(rq, p, &rf);
1140 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1141 tlb_migrate_finish(p->mm);
1143 } else if (task_on_rq_queued(p)) {
1145 * OK, since we're going to drop the lock immediately
1146 * afterwards anyway.
1148 rq_unpin_lock(rq, &rf);
1149 rq = move_queued_task(rq, p, dest_cpu);
1150 rq_repin_lock(rq, &rf);
1153 task_rq_unlock(rq, p, &rf);
1158 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1160 return __set_cpus_allowed_ptr(p, new_mask, false);
1162 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1164 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1166 #ifdef CONFIG_SCHED_DEBUG
1168 * We should never call set_task_cpu() on a blocked task,
1169 * ttwu() will sort out the placement.
1171 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1175 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1176 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1177 * time relying on p->on_rq.
1179 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1180 p->sched_class == &fair_sched_class &&
1181 (p->on_rq && !task_on_rq_migrating(p)));
1183 #ifdef CONFIG_LOCKDEP
1185 * The caller should hold either p->pi_lock or rq->lock, when changing
1186 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1188 * sched_move_task() holds both and thus holding either pins the cgroup,
1191 * Furthermore, all task_rq users should acquire both locks, see
1194 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1195 lockdep_is_held(&task_rq(p)->lock)));
1199 trace_sched_migrate_task(p, new_cpu);
1201 if (task_cpu(p) != new_cpu) {
1202 if (p->sched_class->migrate_task_rq)
1203 p->sched_class->migrate_task_rq(p);
1204 p->se.nr_migrations++;
1205 perf_event_task_migrate(p);
1208 __set_task_cpu(p, new_cpu);
1211 static void __migrate_swap_task(struct task_struct *p, int cpu)
1213 if (task_on_rq_queued(p)) {
1214 struct rq *src_rq, *dst_rq;
1216 src_rq = task_rq(p);
1217 dst_rq = cpu_rq(cpu);
1219 p->on_rq = TASK_ON_RQ_MIGRATING;
1220 deactivate_task(src_rq, p, 0);
1221 set_task_cpu(p, cpu);
1222 activate_task(dst_rq, p, 0);
1223 p->on_rq = TASK_ON_RQ_QUEUED;
1224 check_preempt_curr(dst_rq, p, 0);
1227 * Task isn't running anymore; make it appear like we migrated
1228 * it before it went to sleep. This means on wakeup we make the
1229 * previous CPU our target instead of where it really is.
1235 struct migration_swap_arg {
1236 struct task_struct *src_task, *dst_task;
1237 int src_cpu, dst_cpu;
1240 static int migrate_swap_stop(void *data)
1242 struct migration_swap_arg *arg = data;
1243 struct rq *src_rq, *dst_rq;
1246 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1249 src_rq = cpu_rq(arg->src_cpu);
1250 dst_rq = cpu_rq(arg->dst_cpu);
1252 double_raw_lock(&arg->src_task->pi_lock,
1253 &arg->dst_task->pi_lock);
1254 double_rq_lock(src_rq, dst_rq);
1256 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1259 if (task_cpu(arg->src_task) != arg->src_cpu)
1262 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1265 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1268 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1269 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1274 double_rq_unlock(src_rq, dst_rq);
1275 raw_spin_unlock(&arg->dst_task->pi_lock);
1276 raw_spin_unlock(&arg->src_task->pi_lock);
1282 * Cross migrate two tasks
1284 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1286 struct migration_swap_arg arg;
1289 arg = (struct migration_swap_arg){
1291 .src_cpu = task_cpu(cur),
1293 .dst_cpu = task_cpu(p),
1296 if (arg.src_cpu == arg.dst_cpu)
1300 * These three tests are all lockless; this is OK since all of them
1301 * will be re-checked with proper locks held further down the line.
1303 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1306 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1309 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1312 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1313 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1320 * wait_task_inactive - wait for a thread to unschedule.
1322 * If @match_state is nonzero, it's the @p->state value just checked and
1323 * not expected to change. If it changes, i.e. @p might have woken up,
1324 * then return zero. When we succeed in waiting for @p to be off its CPU,
1325 * we return a positive number (its total switch count). If a second call
1326 * a short while later returns the same number, the caller can be sure that
1327 * @p has remained unscheduled the whole time.
1329 * The caller must ensure that the task *will* unschedule sometime soon,
1330 * else this function might spin for a *long* time. This function can't
1331 * be called with interrupts off, or it may introduce deadlock with
1332 * smp_call_function() if an IPI is sent by the same process we are
1333 * waiting to become inactive.
1335 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1337 int running, queued;
1344 * We do the initial early heuristics without holding
1345 * any task-queue locks at all. We'll only try to get
1346 * the runqueue lock when things look like they will
1352 * If the task is actively running on another CPU
1353 * still, just relax and busy-wait without holding
1356 * NOTE! Since we don't hold any locks, it's not
1357 * even sure that "rq" stays as the right runqueue!
1358 * But we don't care, since "task_running()" will
1359 * return false if the runqueue has changed and p
1360 * is actually now running somewhere else!
1362 while (task_running(rq, p)) {
1363 if (match_state && unlikely(p->state != match_state))
1369 * Ok, time to look more closely! We need the rq
1370 * lock now, to be *sure*. If we're wrong, we'll
1371 * just go back and repeat.
1373 rq = task_rq_lock(p, &rf);
1374 trace_sched_wait_task(p);
1375 running = task_running(rq, p);
1376 queued = task_on_rq_queued(p);
1378 if (!match_state || p->state == match_state)
1379 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1380 task_rq_unlock(rq, p, &rf);
1383 * If it changed from the expected state, bail out now.
1385 if (unlikely(!ncsw))
1389 * Was it really running after all now that we
1390 * checked with the proper locks actually held?
1392 * Oops. Go back and try again..
1394 if (unlikely(running)) {
1400 * It's not enough that it's not actively running,
1401 * it must be off the runqueue _entirely_, and not
1404 * So if it was still runnable (but just not actively
1405 * running right now), it's preempted, and we should
1406 * yield - it could be a while.
1408 if (unlikely(queued)) {
1409 ktime_t to = NSEC_PER_SEC / HZ;
1411 set_current_state(TASK_UNINTERRUPTIBLE);
1412 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1417 * Ahh, all good. It wasn't running, and it wasn't
1418 * runnable, which means that it will never become
1419 * running in the future either. We're all done!
1428 * kick_process - kick a running thread to enter/exit the kernel
1429 * @p: the to-be-kicked thread
1431 * Cause a process which is running on another CPU to enter
1432 * kernel-mode, without any delay. (to get signals handled.)
1434 * NOTE: this function doesn't have to take the runqueue lock,
1435 * because all it wants to ensure is that the remote task enters
1436 * the kernel. If the IPI races and the task has been migrated
1437 * to another CPU then no harm is done and the purpose has been
1440 void kick_process(struct task_struct *p)
1446 if ((cpu != smp_processor_id()) && task_curr(p))
1447 smp_send_reschedule(cpu);
1450 EXPORT_SYMBOL_GPL(kick_process);
1453 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1455 * A few notes on cpu_active vs cpu_online:
1457 * - cpu_active must be a subset of cpu_online
1459 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1460 * see __set_cpus_allowed_ptr(). At this point the newly online
1461 * CPU isn't yet part of the sched domains, and balancing will not
1464 * - on CPU-down we clear cpu_active() to mask the sched domains and
1465 * avoid the load balancer to place new tasks on the to be removed
1466 * CPU. Existing tasks will remain running there and will be taken
1469 * This means that fallback selection must not select !active CPUs.
1470 * And can assume that any active CPU must be online. Conversely
1471 * select_task_rq() below may allow selection of !active CPUs in order
1472 * to satisfy the above rules.
1474 static int select_fallback_rq(int cpu, struct task_struct *p)
1476 int nid = cpu_to_node(cpu);
1477 const struct cpumask *nodemask = NULL;
1478 enum { cpuset, possible, fail } state = cpuset;
1482 * If the node that the CPU is on has been offlined, cpu_to_node()
1483 * will return -1. There is no CPU on the node, and we should
1484 * select the CPU on the other node.
1487 nodemask = cpumask_of_node(nid);
1489 /* Look for allowed, online CPU in same node. */
1490 for_each_cpu(dest_cpu, nodemask) {
1491 if (!cpu_active(dest_cpu))
1493 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1499 /* Any allowed, online CPU? */
1500 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1501 if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1503 if (!cpu_online(dest_cpu))
1508 /* No more Mr. Nice Guy. */
1511 if (IS_ENABLED(CONFIG_CPUSETS)) {
1512 cpuset_cpus_allowed_fallback(p);
1518 do_set_cpus_allowed(p, cpu_possible_mask);
1529 if (state != cpuset) {
1531 * Don't tell them about moving exiting tasks or
1532 * kernel threads (both mm NULL), since they never
1535 if (p->mm && printk_ratelimit()) {
1536 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1537 task_pid_nr(p), p->comm, cpu);
1545 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1548 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1550 lockdep_assert_held(&p->pi_lock);
1552 if (tsk_nr_cpus_allowed(p) > 1)
1553 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1555 cpu = cpumask_any(tsk_cpus_allowed(p));
1558 * In order not to call set_task_cpu() on a blocking task we need
1559 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1562 * Since this is common to all placement strategies, this lives here.
1564 * [ this allows ->select_task() to simply return task_cpu(p) and
1565 * not worry about this generic constraint ]
1567 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1569 cpu = select_fallback_rq(task_cpu(p), p);
1574 static void update_avg(u64 *avg, u64 sample)
1576 s64 diff = sample - *avg;
1582 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1583 const struct cpumask *new_mask, bool check)
1585 return set_cpus_allowed_ptr(p, new_mask);
1588 #endif /* CONFIG_SMP */
1591 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1595 if (!schedstat_enabled())
1601 if (cpu == rq->cpu) {
1602 schedstat_inc(rq->ttwu_local);
1603 schedstat_inc(p->se.statistics.nr_wakeups_local);
1605 struct sched_domain *sd;
1607 schedstat_inc(p->se.statistics.nr_wakeups_remote);
1609 for_each_domain(rq->cpu, sd) {
1610 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1611 schedstat_inc(sd->ttwu_wake_remote);
1618 if (wake_flags & WF_MIGRATED)
1619 schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1620 #endif /* CONFIG_SMP */
1622 schedstat_inc(rq->ttwu_count);
1623 schedstat_inc(p->se.statistics.nr_wakeups);
1625 if (wake_flags & WF_SYNC)
1626 schedstat_inc(p->se.statistics.nr_wakeups_sync);
1629 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1631 activate_task(rq, p, en_flags);
1632 p->on_rq = TASK_ON_RQ_QUEUED;
1634 /* If a worker is waking up, notify the workqueue: */
1635 if (p->flags & PF_WQ_WORKER)
1636 wq_worker_waking_up(p, cpu_of(rq));
1640 * Mark the task runnable and perform wakeup-preemption.
1642 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1643 struct rq_flags *rf)
1645 check_preempt_curr(rq, p, wake_flags);
1646 p->state = TASK_RUNNING;
1647 trace_sched_wakeup(p);
1650 if (p->sched_class->task_woken) {
1652 * Our task @p is fully woken up and running; so its safe to
1653 * drop the rq->lock, hereafter rq is only used for statistics.
1655 rq_unpin_lock(rq, rf);
1656 p->sched_class->task_woken(rq, p);
1657 rq_repin_lock(rq, rf);
1660 if (rq->idle_stamp) {
1661 u64 delta = rq_clock(rq) - rq->idle_stamp;
1662 u64 max = 2*rq->max_idle_balance_cost;
1664 update_avg(&rq->avg_idle, delta);
1666 if (rq->avg_idle > max)
1675 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1676 struct rq_flags *rf)
1678 int en_flags = ENQUEUE_WAKEUP;
1680 lockdep_assert_held(&rq->lock);
1683 if (p->sched_contributes_to_load)
1684 rq->nr_uninterruptible--;
1686 if (wake_flags & WF_MIGRATED)
1687 en_flags |= ENQUEUE_MIGRATED;
1690 ttwu_activate(rq, p, en_flags);
1691 ttwu_do_wakeup(rq, p, wake_flags, rf);
1695 * Called in case the task @p isn't fully descheduled from its runqueue,
1696 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1697 * since all we need to do is flip p->state to TASK_RUNNING, since
1698 * the task is still ->on_rq.
1700 static int ttwu_remote(struct task_struct *p, int wake_flags)
1706 rq = __task_rq_lock(p, &rf);
1707 if (task_on_rq_queued(p)) {
1708 /* check_preempt_curr() may use rq clock */
1709 update_rq_clock(rq);
1710 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1713 __task_rq_unlock(rq, &rf);
1719 void sched_ttwu_pending(void)
1721 struct rq *rq = this_rq();
1722 struct llist_node *llist = llist_del_all(&rq->wake_list);
1723 struct task_struct *p;
1724 unsigned long flags;
1730 raw_spin_lock_irqsave(&rq->lock, flags);
1731 rq_pin_lock(rq, &rf);
1736 p = llist_entry(llist, struct task_struct, wake_entry);
1737 llist = llist_next(llist);
1739 if (p->sched_remote_wakeup)
1740 wake_flags = WF_MIGRATED;
1742 ttwu_do_activate(rq, p, wake_flags, &rf);
1745 rq_unpin_lock(rq, &rf);
1746 raw_spin_unlock_irqrestore(&rq->lock, flags);
1749 void scheduler_ipi(void)
1752 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1753 * TIF_NEED_RESCHED remotely (for the first time) will also send
1756 preempt_fold_need_resched();
1758 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1762 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1763 * traditionally all their work was done from the interrupt return
1764 * path. Now that we actually do some work, we need to make sure
1767 * Some archs already do call them, luckily irq_enter/exit nest
1770 * Arguably we should visit all archs and update all handlers,
1771 * however a fair share of IPIs are still resched only so this would
1772 * somewhat pessimize the simple resched case.
1775 sched_ttwu_pending();
1778 * Check if someone kicked us for doing the nohz idle load balance.
1780 if (unlikely(got_nohz_idle_kick())) {
1781 this_rq()->idle_balance = 1;
1782 raise_softirq_irqoff(SCHED_SOFTIRQ);
1787 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1789 struct rq *rq = cpu_rq(cpu);
1791 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1793 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1794 if (!set_nr_if_polling(rq->idle))
1795 smp_send_reschedule(cpu);
1797 trace_sched_wake_idle_without_ipi(cpu);
1801 void wake_up_if_idle(int cpu)
1803 struct rq *rq = cpu_rq(cpu);
1804 unsigned long flags;
1808 if (!is_idle_task(rcu_dereference(rq->curr)))
1811 if (set_nr_if_polling(rq->idle)) {
1812 trace_sched_wake_idle_without_ipi(cpu);
1814 raw_spin_lock_irqsave(&rq->lock, flags);
1815 if (is_idle_task(rq->curr))
1816 smp_send_reschedule(cpu);
1817 /* Else CPU is not idle, do nothing here: */
1818 raw_spin_unlock_irqrestore(&rq->lock, flags);
1825 bool cpus_share_cache(int this_cpu, int that_cpu)
1827 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1829 #endif /* CONFIG_SMP */
1831 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1833 struct rq *rq = cpu_rq(cpu);
1836 #if defined(CONFIG_SMP)
1837 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1838 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1839 ttwu_queue_remote(p, cpu, wake_flags);
1844 raw_spin_lock(&rq->lock);
1845 rq_pin_lock(rq, &rf);
1846 ttwu_do_activate(rq, p, wake_flags, &rf);
1847 rq_unpin_lock(rq, &rf);
1848 raw_spin_unlock(&rq->lock);
1852 * Notes on Program-Order guarantees on SMP systems.
1856 * The basic program-order guarantee on SMP systems is that when a task [t]
1857 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1858 * execution on its new CPU [c1].
1860 * For migration (of runnable tasks) this is provided by the following means:
1862 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1863 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1864 * rq(c1)->lock (if not at the same time, then in that order).
1865 * C) LOCK of the rq(c1)->lock scheduling in task
1867 * Transitivity guarantees that B happens after A and C after B.
1868 * Note: we only require RCpc transitivity.
1869 * Note: the CPU doing B need not be c0 or c1
1878 * UNLOCK rq(0)->lock
1880 * LOCK rq(0)->lock // orders against CPU0
1882 * UNLOCK rq(0)->lock
1886 * UNLOCK rq(1)->lock
1888 * LOCK rq(1)->lock // orders against CPU2
1891 * UNLOCK rq(1)->lock
1894 * BLOCKING -- aka. SLEEP + WAKEUP
1896 * For blocking we (obviously) need to provide the same guarantee as for
1897 * migration. However the means are completely different as there is no lock
1898 * chain to provide order. Instead we do:
1900 * 1) smp_store_release(X->on_cpu, 0)
1901 * 2) smp_cond_load_acquire(!X->on_cpu)
1905 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1907 * LOCK rq(0)->lock LOCK X->pi_lock
1910 * smp_store_release(X->on_cpu, 0);
1912 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1918 * X->state = RUNNING
1919 * UNLOCK rq(2)->lock
1921 * LOCK rq(2)->lock // orders against CPU1
1924 * UNLOCK rq(2)->lock
1927 * UNLOCK rq(0)->lock
1930 * However; for wakeups there is a second guarantee we must provide, namely we
1931 * must observe the state that lead to our wakeup. That is, not only must our
1932 * task observe its own prior state, it must also observe the stores prior to
1935 * This means that any means of doing remote wakeups must order the CPU doing
1936 * the wakeup against the CPU the task is going to end up running on. This,
1937 * however, is already required for the regular Program-Order guarantee above,
1938 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1943 * try_to_wake_up - wake up a thread
1944 * @p: the thread to be awakened
1945 * @state: the mask of task states that can be woken
1946 * @wake_flags: wake modifier flags (WF_*)
1948 * If (@state & @p->state) @p->state = TASK_RUNNING.
1950 * If the task was not queued/runnable, also place it back on a runqueue.
1952 * Atomic against schedule() which would dequeue a task, also see
1953 * set_current_state().
1955 * Return: %true if @p->state changes (an actual wakeup was done),
1959 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1961 unsigned long flags;
1962 int cpu, success = 0;
1965 * If we are going to wake up a thread waiting for CONDITION we
1966 * need to ensure that CONDITION=1 done by the caller can not be
1967 * reordered with p->state check below. This pairs with mb() in
1968 * set_current_state() the waiting thread does.
1970 smp_mb__before_spinlock();
1971 raw_spin_lock_irqsave(&p->pi_lock, flags);
1972 if (!(p->state & state))
1975 trace_sched_waking(p);
1977 /* We're going to change ->state: */
1982 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1983 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1984 * in smp_cond_load_acquire() below.
1986 * sched_ttwu_pending() try_to_wake_up()
1987 * [S] p->on_rq = 1; [L] P->state
1988 * UNLOCK rq->lock -----.
1992 * LOCK rq->lock -----'
1996 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
1998 * Pairs with the UNLOCK+LOCK on rq->lock from the
1999 * last wakeup of our task and the schedule that got our task
2003 if (p->on_rq && ttwu_remote(p, wake_flags))
2008 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2009 * possible to, falsely, observe p->on_cpu == 0.
2011 * One must be running (->on_cpu == 1) in order to remove oneself
2012 * from the runqueue.
2014 * [S] ->on_cpu = 1; [L] ->on_rq
2018 * [S] ->on_rq = 0; [L] ->on_cpu
2020 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2021 * from the consecutive calls to schedule(); the first switching to our
2022 * task, the second putting it to sleep.
2027 * If the owning (remote) CPU is still in the middle of schedule() with
2028 * this task as prev, wait until its done referencing the task.
2030 * Pairs with the smp_store_release() in finish_lock_switch().
2032 * This ensures that tasks getting woken will be fully ordered against
2033 * their previous state and preserve Program Order.
2035 smp_cond_load_acquire(&p->on_cpu, !VAL);
2037 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2038 p->state = TASK_WAKING;
2041 delayacct_blkio_end();
2042 atomic_dec(&task_rq(p)->nr_iowait);
2045 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2046 if (task_cpu(p) != cpu) {
2047 wake_flags |= WF_MIGRATED;
2048 set_task_cpu(p, cpu);
2051 #else /* CONFIG_SMP */
2054 delayacct_blkio_end();
2055 atomic_dec(&task_rq(p)->nr_iowait);
2058 #endif /* CONFIG_SMP */
2060 ttwu_queue(p, cpu, wake_flags);
2062 ttwu_stat(p, cpu, wake_flags);
2064 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2070 * try_to_wake_up_local - try to wake up a local task with rq lock held
2071 * @p: the thread to be awakened
2072 * @cookie: context's cookie for pinning
2074 * Put @p on the run-queue if it's not already there. The caller must
2075 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2078 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2080 struct rq *rq = task_rq(p);
2082 if (WARN_ON_ONCE(rq != this_rq()) ||
2083 WARN_ON_ONCE(p == current))
2086 lockdep_assert_held(&rq->lock);
2088 if (!raw_spin_trylock(&p->pi_lock)) {
2090 * This is OK, because current is on_cpu, which avoids it being
2091 * picked for load-balance and preemption/IRQs are still
2092 * disabled avoiding further scheduler activity on it and we've
2093 * not yet picked a replacement task.
2095 rq_unpin_lock(rq, rf);
2096 raw_spin_unlock(&rq->lock);
2097 raw_spin_lock(&p->pi_lock);
2098 raw_spin_lock(&rq->lock);
2099 rq_repin_lock(rq, rf);
2102 if (!(p->state & TASK_NORMAL))
2105 trace_sched_waking(p);
2107 if (!task_on_rq_queued(p)) {
2109 delayacct_blkio_end();
2110 atomic_dec(&rq->nr_iowait);
2112 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2115 ttwu_do_wakeup(rq, p, 0, rf);
2116 ttwu_stat(p, smp_processor_id(), 0);
2118 raw_spin_unlock(&p->pi_lock);
2122 * wake_up_process - Wake up a specific process
2123 * @p: The process to be woken up.
2125 * Attempt to wake up the nominated process and move it to the set of runnable
2128 * Return: 1 if the process was woken up, 0 if it was already running.
2130 * It may be assumed that this function implies a write memory barrier before
2131 * changing the task state if and only if any tasks are woken up.
2133 int wake_up_process(struct task_struct *p)
2135 return try_to_wake_up(p, TASK_NORMAL, 0);
2137 EXPORT_SYMBOL(wake_up_process);
2139 int wake_up_state(struct task_struct *p, unsigned int state)
2141 return try_to_wake_up(p, state, 0);
2145 * This function clears the sched_dl_entity static params.
2147 void __dl_clear_params(struct task_struct *p)
2149 struct sched_dl_entity *dl_se = &p->dl;
2151 dl_se->dl_runtime = 0;
2152 dl_se->dl_deadline = 0;
2153 dl_se->dl_period = 0;
2157 dl_se->dl_throttled = 0;
2158 dl_se->dl_yielded = 0;
2162 * Perform scheduler related setup for a newly forked process p.
2163 * p is forked by current.
2165 * __sched_fork() is basic setup used by init_idle() too:
2167 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2172 p->se.exec_start = 0;
2173 p->se.sum_exec_runtime = 0;
2174 p->se.prev_sum_exec_runtime = 0;
2175 p->se.nr_migrations = 0;
2177 INIT_LIST_HEAD(&p->se.group_node);
2179 #ifdef CONFIG_FAIR_GROUP_SCHED
2180 p->se.cfs_rq = NULL;
2183 #ifdef CONFIG_SCHEDSTATS
2184 /* Even if schedstat is disabled, there should not be garbage */
2185 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2188 RB_CLEAR_NODE(&p->dl.rb_node);
2189 init_dl_task_timer(&p->dl);
2190 __dl_clear_params(p);
2192 INIT_LIST_HEAD(&p->rt.run_list);
2194 p->rt.time_slice = sched_rr_timeslice;
2198 #ifdef CONFIG_PREEMPT_NOTIFIERS
2199 INIT_HLIST_HEAD(&p->preempt_notifiers);
2202 #ifdef CONFIG_NUMA_BALANCING
2203 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2204 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2205 p->mm->numa_scan_seq = 0;
2208 if (clone_flags & CLONE_VM)
2209 p->numa_preferred_nid = current->numa_preferred_nid;
2211 p->numa_preferred_nid = -1;
2213 p->node_stamp = 0ULL;
2214 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2215 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2216 p->numa_work.next = &p->numa_work;
2217 p->numa_faults = NULL;
2218 p->last_task_numa_placement = 0;
2219 p->last_sum_exec_runtime = 0;
2221 p->numa_group = NULL;
2222 #endif /* CONFIG_NUMA_BALANCING */
2225 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2227 #ifdef CONFIG_NUMA_BALANCING
2229 void set_numabalancing_state(bool enabled)
2232 static_branch_enable(&sched_numa_balancing);
2234 static_branch_disable(&sched_numa_balancing);
2237 #ifdef CONFIG_PROC_SYSCTL
2238 int sysctl_numa_balancing(struct ctl_table *table, int write,
2239 void __user *buffer, size_t *lenp, loff_t *ppos)
2243 int state = static_branch_likely(&sched_numa_balancing);
2245 if (write && !capable(CAP_SYS_ADMIN))
2250 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2254 set_numabalancing_state(state);
2260 #ifdef CONFIG_SCHEDSTATS
2262 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2263 static bool __initdata __sched_schedstats = false;
2265 static void set_schedstats(bool enabled)
2268 static_branch_enable(&sched_schedstats);
2270 static_branch_disable(&sched_schedstats);
2273 void force_schedstat_enabled(void)
2275 if (!schedstat_enabled()) {
2276 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2277 static_branch_enable(&sched_schedstats);
2281 static int __init setup_schedstats(char *str)
2288 * This code is called before jump labels have been set up, so we can't
2289 * change the static branch directly just yet. Instead set a temporary
2290 * variable so init_schedstats() can do it later.
2292 if (!strcmp(str, "enable")) {
2293 __sched_schedstats = true;
2295 } else if (!strcmp(str, "disable")) {
2296 __sched_schedstats = false;
2301 pr_warn("Unable to parse schedstats=\n");
2305 __setup("schedstats=", setup_schedstats);
2307 static void __init init_schedstats(void)
2309 set_schedstats(__sched_schedstats);
2312 #ifdef CONFIG_PROC_SYSCTL
2313 int sysctl_schedstats(struct ctl_table *table, int write,
2314 void __user *buffer, size_t *lenp, loff_t *ppos)
2318 int state = static_branch_likely(&sched_schedstats);
2320 if (write && !capable(CAP_SYS_ADMIN))
2325 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2329 set_schedstats(state);
2332 #endif /* CONFIG_PROC_SYSCTL */
2333 #else /* !CONFIG_SCHEDSTATS */
2334 static inline void init_schedstats(void) {}
2335 #endif /* CONFIG_SCHEDSTATS */
2338 * fork()/clone()-time setup:
2340 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2342 unsigned long flags;
2343 int cpu = get_cpu();
2345 __sched_fork(clone_flags, p);
2347 * We mark the process as NEW here. This guarantees that
2348 * nobody will actually run it, and a signal or other external
2349 * event cannot wake it up and insert it on the runqueue either.
2351 p->state = TASK_NEW;
2354 * Make sure we do not leak PI boosting priority to the child.
2356 p->prio = current->normal_prio;
2359 * Revert to default priority/policy on fork if requested.
2361 if (unlikely(p->sched_reset_on_fork)) {
2362 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2363 p->policy = SCHED_NORMAL;
2364 p->static_prio = NICE_TO_PRIO(0);
2366 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2367 p->static_prio = NICE_TO_PRIO(0);
2369 p->prio = p->normal_prio = __normal_prio(p);
2373 * We don't need the reset flag anymore after the fork. It has
2374 * fulfilled its duty:
2376 p->sched_reset_on_fork = 0;
2379 if (dl_prio(p->prio)) {
2382 } else if (rt_prio(p->prio)) {
2383 p->sched_class = &rt_sched_class;
2385 p->sched_class = &fair_sched_class;
2388 init_entity_runnable_average(&p->se);
2391 * The child is not yet in the pid-hash so no cgroup attach races,
2392 * and the cgroup is pinned to this child due to cgroup_fork()
2393 * is ran before sched_fork().
2395 * Silence PROVE_RCU.
2397 raw_spin_lock_irqsave(&p->pi_lock, flags);
2399 * We're setting the CPU for the first time, we don't migrate,
2400 * so use __set_task_cpu().
2402 __set_task_cpu(p, cpu);
2403 if (p->sched_class->task_fork)
2404 p->sched_class->task_fork(p);
2405 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2407 #ifdef CONFIG_SCHED_INFO
2408 if (likely(sched_info_on()))
2409 memset(&p->sched_info, 0, sizeof(p->sched_info));
2411 #if defined(CONFIG_SMP)
2414 init_task_preempt_count(p);
2416 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2417 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2424 unsigned long to_ratio(u64 period, u64 runtime)
2426 if (runtime == RUNTIME_INF)
2430 * Doing this here saves a lot of checks in all
2431 * the calling paths, and returning zero seems
2432 * safe for them anyway.
2437 return div64_u64(runtime << 20, period);
2441 inline struct dl_bw *dl_bw_of(int i)
2443 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2444 "sched RCU must be held");
2445 return &cpu_rq(i)->rd->dl_bw;
2448 static inline int dl_bw_cpus(int i)
2450 struct root_domain *rd = cpu_rq(i)->rd;
2453 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2454 "sched RCU must be held");
2455 for_each_cpu_and(i, rd->span, cpu_active_mask)
2461 inline struct dl_bw *dl_bw_of(int i)
2463 return &cpu_rq(i)->dl.dl_bw;
2466 static inline int dl_bw_cpus(int i)
2473 * We must be sure that accepting a new task (or allowing changing the
2474 * parameters of an existing one) is consistent with the bandwidth
2475 * constraints. If yes, this function also accordingly updates the currently
2476 * allocated bandwidth to reflect the new situation.
2478 * This function is called while holding p's rq->lock.
2480 * XXX we should delay bw change until the task's 0-lag point, see
2483 static int dl_overflow(struct task_struct *p, int policy,
2484 const struct sched_attr *attr)
2487 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2488 u64 period = attr->sched_period ?: attr->sched_deadline;
2489 u64 runtime = attr->sched_runtime;
2490 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2493 /* !deadline task may carry old deadline bandwidth */
2494 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2498 * Either if a task, enters, leave, or stays -deadline but changes
2499 * its parameters, we may need to update accordingly the total
2500 * allocated bandwidth of the container.
2502 raw_spin_lock(&dl_b->lock);
2503 cpus = dl_bw_cpus(task_cpu(p));
2504 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2505 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2506 __dl_add(dl_b, new_bw);
2508 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2509 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2510 __dl_clear(dl_b, p->dl.dl_bw);
2511 __dl_add(dl_b, new_bw);
2513 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2514 __dl_clear(dl_b, p->dl.dl_bw);
2517 raw_spin_unlock(&dl_b->lock);
2522 extern void init_dl_bw(struct dl_bw *dl_b);
2525 * wake_up_new_task - wake up a newly created task for the first time.
2527 * This function will do some initial scheduler statistics housekeeping
2528 * that must be done for every newly created context, then puts the task
2529 * on the runqueue and wakes it.
2531 void wake_up_new_task(struct task_struct *p)
2536 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2537 p->state = TASK_RUNNING;
2540 * Fork balancing, do it here and not earlier because:
2541 * - cpus_allowed can change in the fork path
2542 * - any previously selected CPU might disappear through hotplug
2544 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2545 * as we're not fully set-up yet.
2547 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2549 rq = __task_rq_lock(p, &rf);
2550 update_rq_clock(rq);
2551 post_init_entity_util_avg(&p->se);
2553 activate_task(rq, p, 0);
2554 p->on_rq = TASK_ON_RQ_QUEUED;
2555 trace_sched_wakeup_new(p);
2556 check_preempt_curr(rq, p, WF_FORK);
2558 if (p->sched_class->task_woken) {
2560 * Nothing relies on rq->lock after this, so its fine to
2563 rq_unpin_lock(rq, &rf);
2564 p->sched_class->task_woken(rq, p);
2565 rq_repin_lock(rq, &rf);
2568 task_rq_unlock(rq, p, &rf);
2571 #ifdef CONFIG_PREEMPT_NOTIFIERS
2573 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2575 void preempt_notifier_inc(void)
2577 static_key_slow_inc(&preempt_notifier_key);
2579 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2581 void preempt_notifier_dec(void)
2583 static_key_slow_dec(&preempt_notifier_key);
2585 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2588 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2589 * @notifier: notifier struct to register
2591 void preempt_notifier_register(struct preempt_notifier *notifier)
2593 if (!static_key_false(&preempt_notifier_key))
2594 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2596 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2598 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2601 * preempt_notifier_unregister - no longer interested in preemption notifications
2602 * @notifier: notifier struct to unregister
2604 * This is *not* safe to call from within a preemption notifier.
2606 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2608 hlist_del(¬ifier->link);
2610 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2612 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2614 struct preempt_notifier *notifier;
2616 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2617 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2620 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2622 if (static_key_false(&preempt_notifier_key))
2623 __fire_sched_in_preempt_notifiers(curr);
2627 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2628 struct task_struct *next)
2630 struct preempt_notifier *notifier;
2632 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2633 notifier->ops->sched_out(notifier, next);
2636 static __always_inline void
2637 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2638 struct task_struct *next)
2640 if (static_key_false(&preempt_notifier_key))
2641 __fire_sched_out_preempt_notifiers(curr, next);
2644 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2646 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2651 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2652 struct task_struct *next)
2656 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2659 * prepare_task_switch - prepare to switch tasks
2660 * @rq: the runqueue preparing to switch
2661 * @prev: the current task that is being switched out
2662 * @next: the task we are going to switch to.
2664 * This is called with the rq lock held and interrupts off. It must
2665 * be paired with a subsequent finish_task_switch after the context
2668 * prepare_task_switch sets up locking and calls architecture specific
2672 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2673 struct task_struct *next)
2675 sched_info_switch(rq, prev, next);
2676 perf_event_task_sched_out(prev, next);
2677 fire_sched_out_preempt_notifiers(prev, next);
2678 prepare_lock_switch(rq, next);
2679 prepare_arch_switch(next);
2683 * finish_task_switch - clean up after a task-switch
2684 * @prev: the thread we just switched away from.
2686 * finish_task_switch must be called after the context switch, paired
2687 * with a prepare_task_switch call before the context switch.
2688 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2689 * and do any other architecture-specific cleanup actions.
2691 * Note that we may have delayed dropping an mm in context_switch(). If
2692 * so, we finish that here outside of the runqueue lock. (Doing it
2693 * with the lock held can cause deadlocks; see schedule() for
2696 * The context switch have flipped the stack from under us and restored the
2697 * local variables which were saved when this task called schedule() in the
2698 * past. prev == current is still correct but we need to recalculate this_rq
2699 * because prev may have moved to another CPU.
2701 static struct rq *finish_task_switch(struct task_struct *prev)
2702 __releases(rq->lock)
2704 struct rq *rq = this_rq();
2705 struct mm_struct *mm = rq->prev_mm;
2709 * The previous task will have left us with a preempt_count of 2
2710 * because it left us after:
2713 * preempt_disable(); // 1
2715 * raw_spin_lock_irq(&rq->lock) // 2
2717 * Also, see FORK_PREEMPT_COUNT.
2719 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2720 "corrupted preempt_count: %s/%d/0x%x\n",
2721 current->comm, current->pid, preempt_count()))
2722 preempt_count_set(FORK_PREEMPT_COUNT);
2727 * A task struct has one reference for the use as "current".
2728 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2729 * schedule one last time. The schedule call will never return, and
2730 * the scheduled task must drop that reference.
2732 * We must observe prev->state before clearing prev->on_cpu (in
2733 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2734 * running on another CPU and we could rave with its RUNNING -> DEAD
2735 * transition, resulting in a double drop.
2737 prev_state = prev->state;
2738 vtime_task_switch(prev);
2739 perf_event_task_sched_in(prev, current);
2740 finish_lock_switch(rq, prev);
2741 finish_arch_post_lock_switch();
2743 fire_sched_in_preempt_notifiers(current);
2746 if (unlikely(prev_state == TASK_DEAD)) {
2747 if (prev->sched_class->task_dead)
2748 prev->sched_class->task_dead(prev);
2751 * Remove function-return probe instances associated with this
2752 * task and put them back on the free list.
2754 kprobe_flush_task(prev);
2756 /* Task is done with its stack. */
2757 put_task_stack(prev);
2759 put_task_struct(prev);
2762 tick_nohz_task_switch();
2768 /* rq->lock is NOT held, but preemption is disabled */
2769 static void __balance_callback(struct rq *rq)
2771 struct callback_head *head, *next;
2772 void (*func)(struct rq *rq);
2773 unsigned long flags;
2775 raw_spin_lock_irqsave(&rq->lock, flags);
2776 head = rq->balance_callback;
2777 rq->balance_callback = NULL;
2779 func = (void (*)(struct rq *))head->func;
2786 raw_spin_unlock_irqrestore(&rq->lock, flags);
2789 static inline void balance_callback(struct rq *rq)
2791 if (unlikely(rq->balance_callback))
2792 __balance_callback(rq);
2797 static inline void balance_callback(struct rq *rq)
2804 * schedule_tail - first thing a freshly forked thread must call.
2805 * @prev: the thread we just switched away from.
2807 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2808 __releases(rq->lock)
2813 * New tasks start with FORK_PREEMPT_COUNT, see there and
2814 * finish_task_switch() for details.
2816 * finish_task_switch() will drop rq->lock() and lower preempt_count
2817 * and the preempt_enable() will end up enabling preemption (on
2818 * PREEMPT_COUNT kernels).
2821 rq = finish_task_switch(prev);
2822 balance_callback(rq);
2825 if (current->set_child_tid)
2826 put_user(task_pid_vnr(current), current->set_child_tid);
2830 * context_switch - switch to the new MM and the new thread's register state.
2832 static __always_inline struct rq *
2833 context_switch(struct rq *rq, struct task_struct *prev,
2834 struct task_struct *next, struct rq_flags *rf)
2836 struct mm_struct *mm, *oldmm;
2838 prepare_task_switch(rq, prev, next);
2841 oldmm = prev->active_mm;
2843 * For paravirt, this is coupled with an exit in switch_to to
2844 * combine the page table reload and the switch backend into
2847 arch_start_context_switch(prev);
2850 next->active_mm = oldmm;
2852 enter_lazy_tlb(oldmm, next);
2854 switch_mm_irqs_off(oldmm, mm, next);
2857 prev->active_mm = NULL;
2858 rq->prev_mm = oldmm;
2861 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2864 * Since the runqueue lock will be released by the next
2865 * task (which is an invalid locking op but in the case
2866 * of the scheduler it's an obvious special-case), so we
2867 * do an early lockdep release here:
2869 rq_unpin_lock(rq, rf);
2870 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2872 /* Here we just switch the register state and the stack. */
2873 switch_to(prev, next, prev);
2876 return finish_task_switch(prev);
2880 * nr_running and nr_context_switches:
2882 * externally visible scheduler statistics: current number of runnable
2883 * threads, total number of context switches performed since bootup.
2885 unsigned long nr_running(void)
2887 unsigned long i, sum = 0;
2889 for_each_online_cpu(i)
2890 sum += cpu_rq(i)->nr_running;
2896 * Check if only the current task is running on the CPU.
2898 * Caution: this function does not check that the caller has disabled
2899 * preemption, thus the result might have a time-of-check-to-time-of-use
2900 * race. The caller is responsible to use it correctly, for example:
2902 * - from a non-preemptable section (of course)
2904 * - from a thread that is bound to a single CPU
2906 * - in a loop with very short iterations (e.g. a polling loop)
2908 bool single_task_running(void)
2910 return raw_rq()->nr_running == 1;
2912 EXPORT_SYMBOL(single_task_running);
2914 unsigned long long nr_context_switches(void)
2917 unsigned long long sum = 0;
2919 for_each_possible_cpu(i)
2920 sum += cpu_rq(i)->nr_switches;
2926 * IO-wait accounting, and how its mostly bollocks (on SMP).
2928 * The idea behind IO-wait account is to account the idle time that we could
2929 * have spend running if it were not for IO. That is, if we were to improve the
2930 * storage performance, we'd have a proportional reduction in IO-wait time.
2932 * This all works nicely on UP, where, when a task blocks on IO, we account
2933 * idle time as IO-wait, because if the storage were faster, it could've been
2934 * running and we'd not be idle.
2936 * This has been extended to SMP, by doing the same for each CPU. This however
2939 * Imagine for instance the case where two tasks block on one CPU, only the one
2940 * CPU will have IO-wait accounted, while the other has regular idle. Even
2941 * though, if the storage were faster, both could've ran at the same time,
2942 * utilising both CPUs.
2944 * This means, that when looking globally, the current IO-wait accounting on
2945 * SMP is a lower bound, by reason of under accounting.
2947 * Worse, since the numbers are provided per CPU, they are sometimes
2948 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2949 * associated with any one particular CPU, it can wake to another CPU than it
2950 * blocked on. This means the per CPU IO-wait number is meaningless.
2952 * Task CPU affinities can make all that even more 'interesting'.
2955 unsigned long nr_iowait(void)
2957 unsigned long i, sum = 0;
2959 for_each_possible_cpu(i)
2960 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2966 * Consumers of these two interfaces, like for example the cpufreq menu
2967 * governor are using nonsensical data. Boosting frequency for a CPU that has
2968 * IO-wait which might not even end up running the task when it does become
2972 unsigned long nr_iowait_cpu(int cpu)
2974 struct rq *this = cpu_rq(cpu);
2975 return atomic_read(&this->nr_iowait);
2978 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2980 struct rq *rq = this_rq();
2981 *nr_waiters = atomic_read(&rq->nr_iowait);
2982 *load = rq->load.weight;
2988 * sched_exec - execve() is a valuable balancing opportunity, because at
2989 * this point the task has the smallest effective memory and cache footprint.
2991 void sched_exec(void)
2993 struct task_struct *p = current;
2994 unsigned long flags;
2997 raw_spin_lock_irqsave(&p->pi_lock, flags);
2998 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2999 if (dest_cpu == smp_processor_id())
3002 if (likely(cpu_active(dest_cpu))) {
3003 struct migration_arg arg = { p, dest_cpu };
3005 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3006 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3010 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3015 DEFINE_PER_CPU(struct kernel_stat, kstat);
3016 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3018 EXPORT_PER_CPU_SYMBOL(kstat);
3019 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3022 * The function fair_sched_class.update_curr accesses the struct curr
3023 * and its field curr->exec_start; when called from task_sched_runtime(),
3024 * we observe a high rate of cache misses in practice.
3025 * Prefetching this data results in improved performance.
3027 static inline void prefetch_curr_exec_start(struct task_struct *p)
3029 #ifdef CONFIG_FAIR_GROUP_SCHED
3030 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3032 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3035 prefetch(&curr->exec_start);
3039 * Return accounted runtime for the task.
3040 * In case the task is currently running, return the runtime plus current's
3041 * pending runtime that have not been accounted yet.
3043 unsigned long long task_sched_runtime(struct task_struct *p)
3049 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3051 * 64-bit doesn't need locks to atomically read a 64bit value.
3052 * So we have a optimization chance when the task's delta_exec is 0.
3053 * Reading ->on_cpu is racy, but this is ok.
3055 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3056 * If we race with it entering CPU, unaccounted time is 0. This is
3057 * indistinguishable from the read occurring a few cycles earlier.
3058 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3059 * been accounted, so we're correct here as well.
3061 if (!p->on_cpu || !task_on_rq_queued(p))
3062 return p->se.sum_exec_runtime;
3065 rq = task_rq_lock(p, &rf);
3067 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3068 * project cycles that may never be accounted to this
3069 * thread, breaking clock_gettime().
3071 if (task_current(rq, p) && task_on_rq_queued(p)) {
3072 prefetch_curr_exec_start(p);
3073 update_rq_clock(rq);
3074 p->sched_class->update_curr(rq);
3076 ns = p->se.sum_exec_runtime;
3077 task_rq_unlock(rq, p, &rf);
3083 * This function gets called by the timer code, with HZ frequency.
3084 * We call it with interrupts disabled.
3086 void scheduler_tick(void)
3088 int cpu = smp_processor_id();
3089 struct rq *rq = cpu_rq(cpu);
3090 struct task_struct *curr = rq->curr;
3094 raw_spin_lock(&rq->lock);
3095 update_rq_clock(rq);
3096 curr->sched_class->task_tick(rq, curr, 0);
3097 cpu_load_update_active(rq);
3098 calc_global_load_tick(rq);
3099 raw_spin_unlock(&rq->lock);
3101 perf_event_task_tick();
3104 rq->idle_balance = idle_cpu(cpu);
3105 trigger_load_balance(rq);
3107 rq_last_tick_reset(rq);
3110 #ifdef CONFIG_NO_HZ_FULL
3112 * scheduler_tick_max_deferment
3114 * Keep at least one tick per second when a single
3115 * active task is running because the scheduler doesn't
3116 * yet completely support full dynticks environment.
3118 * This makes sure that uptime, CFS vruntime, load
3119 * balancing, etc... continue to move forward, even
3120 * with a very low granularity.
3122 * Return: Maximum deferment in nanoseconds.
3124 u64 scheduler_tick_max_deferment(void)
3126 struct rq *rq = this_rq();
3127 unsigned long next, now = READ_ONCE(jiffies);
3129 next = rq->last_sched_tick + HZ;
3131 if (time_before_eq(next, now))
3134 return jiffies_to_nsecs(next - now);
3138 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3139 defined(CONFIG_PREEMPT_TRACER))
3141 * If the value passed in is equal to the current preempt count
3142 * then we just disabled preemption. Start timing the latency.
3144 static inline void preempt_latency_start(int val)
3146 if (preempt_count() == val) {
3147 unsigned long ip = get_lock_parent_ip();
3148 #ifdef CONFIG_DEBUG_PREEMPT
3149 current->preempt_disable_ip = ip;
3151 trace_preempt_off(CALLER_ADDR0, ip);
3155 void preempt_count_add(int val)
3157 #ifdef CONFIG_DEBUG_PREEMPT
3161 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3164 __preempt_count_add(val);
3165 #ifdef CONFIG_DEBUG_PREEMPT
3167 * Spinlock count overflowing soon?
3169 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3172 preempt_latency_start(val);
3174 EXPORT_SYMBOL(preempt_count_add);
3175 NOKPROBE_SYMBOL(preempt_count_add);
3178 * If the value passed in equals to the current preempt count
3179 * then we just enabled preemption. Stop timing the latency.
3181 static inline void preempt_latency_stop(int val)
3183 if (preempt_count() == val)
3184 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3187 void preempt_count_sub(int val)
3189 #ifdef CONFIG_DEBUG_PREEMPT
3193 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3196 * Is the spinlock portion underflowing?
3198 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3199 !(preempt_count() & PREEMPT_MASK)))
3203 preempt_latency_stop(val);
3204 __preempt_count_sub(val);
3206 EXPORT_SYMBOL(preempt_count_sub);
3207 NOKPROBE_SYMBOL(preempt_count_sub);
3210 static inline void preempt_latency_start(int val) { }
3211 static inline void preempt_latency_stop(int val) { }
3215 * Print scheduling while atomic bug:
3217 static noinline void __schedule_bug(struct task_struct *prev)
3219 /* Save this before calling printk(), since that will clobber it */
3220 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3222 if (oops_in_progress)
3225 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3226 prev->comm, prev->pid, preempt_count());
3228 debug_show_held_locks(prev);
3230 if (irqs_disabled())
3231 print_irqtrace_events(prev);
3232 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3233 && in_atomic_preempt_off()) {
3234 pr_err("Preemption disabled at:");
3235 print_ip_sym(preempt_disable_ip);
3239 panic("scheduling while atomic\n");
3242 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3246 * Various schedule()-time debugging checks and statistics:
3248 static inline void schedule_debug(struct task_struct *prev)
3250 #ifdef CONFIG_SCHED_STACK_END_CHECK
3251 if (task_stack_end_corrupted(prev))
3252 panic("corrupted stack end detected inside scheduler\n");
3255 if (unlikely(in_atomic_preempt_off())) {
3256 __schedule_bug(prev);
3257 preempt_count_set(PREEMPT_DISABLED);
3261 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3263 schedstat_inc(this_rq()->sched_count);
3267 * Pick up the highest-prio task:
3269 static inline struct task_struct *
3270 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3272 const struct sched_class *class;
3273 struct task_struct *p;
3276 * Optimization: we know that if all tasks are in
3277 * the fair class we can call that function directly:
3279 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3280 p = fair_sched_class.pick_next_task(rq, prev, rf);
3281 if (unlikely(p == RETRY_TASK))
3284 /* Assumes fair_sched_class->next == idle_sched_class */
3286 p = idle_sched_class.pick_next_task(rq, prev, rf);
3292 for_each_class(class) {
3293 p = class->pick_next_task(rq, prev, rf);
3295 if (unlikely(p == RETRY_TASK))
3301 /* The idle class should always have a runnable task: */
3306 * __schedule() is the main scheduler function.
3308 * The main means of driving the scheduler and thus entering this function are:
3310 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3312 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3313 * paths. For example, see arch/x86/entry_64.S.
3315 * To drive preemption between tasks, the scheduler sets the flag in timer
3316 * interrupt handler scheduler_tick().
3318 * 3. Wakeups don't really cause entry into schedule(). They add a
3319 * task to the run-queue and that's it.
3321 * Now, if the new task added to the run-queue preempts the current
3322 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3323 * called on the nearest possible occasion:
3325 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3327 * - in syscall or exception context, at the next outmost
3328 * preempt_enable(). (this might be as soon as the wake_up()'s
3331 * - in IRQ context, return from interrupt-handler to
3332 * preemptible context
3334 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3337 * - cond_resched() call
3338 * - explicit schedule() call
3339 * - return from syscall or exception to user-space
3340 * - return from interrupt-handler to user-space
3342 * WARNING: must be called with preemption disabled!
3344 static void __sched notrace __schedule(bool preempt)
3346 struct task_struct *prev, *next;
3347 unsigned long *switch_count;
3352 cpu = smp_processor_id();
3356 schedule_debug(prev);
3358 if (sched_feat(HRTICK))
3361 local_irq_disable();
3362 rcu_note_context_switch();
3365 * Make sure that signal_pending_state()->signal_pending() below
3366 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3367 * done by the caller to avoid the race with signal_wake_up().
3369 smp_mb__before_spinlock();
3370 raw_spin_lock(&rq->lock);
3371 rq_pin_lock(rq, &rf);
3373 /* Promote REQ to ACT */
3374 rq->clock_update_flags <<= 1;
3376 switch_count = &prev->nivcsw;
3377 if (!preempt && prev->state) {
3378 if (unlikely(signal_pending_state(prev->state, prev))) {
3379 prev->state = TASK_RUNNING;
3381 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3384 if (prev->in_iowait) {
3385 atomic_inc(&rq->nr_iowait);
3386 delayacct_blkio_start();
3390 * If a worker went to sleep, notify and ask workqueue
3391 * whether it wants to wake up a task to maintain
3394 if (prev->flags & PF_WQ_WORKER) {
3395 struct task_struct *to_wakeup;
3397 to_wakeup = wq_worker_sleeping(prev);
3399 try_to_wake_up_local(to_wakeup, &rf);
3402 switch_count = &prev->nvcsw;
3405 if (task_on_rq_queued(prev))
3406 update_rq_clock(rq);
3408 next = pick_next_task(rq, prev, &rf);
3409 clear_tsk_need_resched(prev);
3410 clear_preempt_need_resched();
3412 if (likely(prev != next)) {
3417 trace_sched_switch(preempt, prev, next);
3419 /* Also unlocks the rq: */
3420 rq = context_switch(rq, prev, next, &rf);
3422 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3423 rq_unpin_lock(rq, &rf);
3424 raw_spin_unlock_irq(&rq->lock);
3427 balance_callback(rq);
3430 void __noreturn do_task_dead(void)
3433 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3434 * when the following two conditions become true.
3435 * - There is race condition of mmap_sem (It is acquired by
3437 * - SMI occurs before setting TASK_RUNINNG.
3438 * (or hypervisor of virtual machine switches to other guest)
3439 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3441 * To avoid it, we have to wait for releasing tsk->pi_lock which
3442 * is held by try_to_wake_up()
3445 raw_spin_unlock_wait(¤t->pi_lock);
3447 /* Causes final put_task_struct in finish_task_switch(): */
3448 __set_current_state(TASK_DEAD);
3450 /* Tell freezer to ignore us: */
3451 current->flags |= PF_NOFREEZE;
3456 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3461 static inline void sched_submit_work(struct task_struct *tsk)
3463 if (!tsk->state || tsk_is_pi_blocked(tsk))
3466 * If we are going to sleep and we have plugged IO queued,
3467 * make sure to submit it to avoid deadlocks.
3469 if (blk_needs_flush_plug(tsk))
3470 blk_schedule_flush_plug(tsk);
3473 asmlinkage __visible void __sched schedule(void)
3475 struct task_struct *tsk = current;
3477 sched_submit_work(tsk);
3481 sched_preempt_enable_no_resched();
3482 } while (need_resched());
3484 EXPORT_SYMBOL(schedule);
3486 #ifdef CONFIG_CONTEXT_TRACKING
3487 asmlinkage __visible void __sched schedule_user(void)
3490 * If we come here after a random call to set_need_resched(),
3491 * or we have been woken up remotely but the IPI has not yet arrived,
3492 * we haven't yet exited the RCU idle mode. Do it here manually until
3493 * we find a better solution.
3495 * NB: There are buggy callers of this function. Ideally we
3496 * should warn if prev_state != CONTEXT_USER, but that will trigger
3497 * too frequently to make sense yet.
3499 enum ctx_state prev_state = exception_enter();
3501 exception_exit(prev_state);
3506 * schedule_preempt_disabled - called with preemption disabled
3508 * Returns with preemption disabled. Note: preempt_count must be 1
3510 void __sched schedule_preempt_disabled(void)
3512 sched_preempt_enable_no_resched();
3517 static void __sched notrace preempt_schedule_common(void)
3521 * Because the function tracer can trace preempt_count_sub()
3522 * and it also uses preempt_enable/disable_notrace(), if
3523 * NEED_RESCHED is set, the preempt_enable_notrace() called
3524 * by the function tracer will call this function again and
3525 * cause infinite recursion.
3527 * Preemption must be disabled here before the function
3528 * tracer can trace. Break up preempt_disable() into two
3529 * calls. One to disable preemption without fear of being
3530 * traced. The other to still record the preemption latency,
3531 * which can also be traced by the function tracer.
3533 preempt_disable_notrace();
3534 preempt_latency_start(1);
3536 preempt_latency_stop(1);
3537 preempt_enable_no_resched_notrace();
3540 * Check again in case we missed a preemption opportunity
3541 * between schedule and now.
3543 } while (need_resched());
3546 #ifdef CONFIG_PREEMPT
3548 * this is the entry point to schedule() from in-kernel preemption
3549 * off of preempt_enable. Kernel preemptions off return from interrupt
3550 * occur there and call schedule directly.
3552 asmlinkage __visible void __sched notrace preempt_schedule(void)
3555 * If there is a non-zero preempt_count or interrupts are disabled,
3556 * we do not want to preempt the current task. Just return..
3558 if (likely(!preemptible()))
3561 preempt_schedule_common();
3563 NOKPROBE_SYMBOL(preempt_schedule);
3564 EXPORT_SYMBOL(preempt_schedule);
3567 * preempt_schedule_notrace - preempt_schedule called by tracing
3569 * The tracing infrastructure uses preempt_enable_notrace to prevent
3570 * recursion and tracing preempt enabling caused by the tracing
3571 * infrastructure itself. But as tracing can happen in areas coming
3572 * from userspace or just about to enter userspace, a preempt enable
3573 * can occur before user_exit() is called. This will cause the scheduler
3574 * to be called when the system is still in usermode.
3576 * To prevent this, the preempt_enable_notrace will use this function
3577 * instead of preempt_schedule() to exit user context if needed before
3578 * calling the scheduler.
3580 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3582 enum ctx_state prev_ctx;
3584 if (likely(!preemptible()))
3589 * Because the function tracer can trace preempt_count_sub()
3590 * and it also uses preempt_enable/disable_notrace(), if
3591 * NEED_RESCHED is set, the preempt_enable_notrace() called
3592 * by the function tracer will call this function again and
3593 * cause infinite recursion.
3595 * Preemption must be disabled here before the function
3596 * tracer can trace. Break up preempt_disable() into two
3597 * calls. One to disable preemption without fear of being
3598 * traced. The other to still record the preemption latency,
3599 * which can also be traced by the function tracer.
3601 preempt_disable_notrace();
3602 preempt_latency_start(1);
3604 * Needs preempt disabled in case user_exit() is traced
3605 * and the tracer calls preempt_enable_notrace() causing
3606 * an infinite recursion.
3608 prev_ctx = exception_enter();
3610 exception_exit(prev_ctx);
3612 preempt_latency_stop(1);
3613 preempt_enable_no_resched_notrace();
3614 } while (need_resched());
3616 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3618 #endif /* CONFIG_PREEMPT */
3621 * this is the entry point to schedule() from kernel preemption
3622 * off of irq context.
3623 * Note, that this is called and return with irqs disabled. This will
3624 * protect us against recursive calling from irq.
3626 asmlinkage __visible void __sched preempt_schedule_irq(void)
3628 enum ctx_state prev_state;
3630 /* Catch callers which need to be fixed */
3631 BUG_ON(preempt_count() || !irqs_disabled());
3633 prev_state = exception_enter();
3639 local_irq_disable();
3640 sched_preempt_enable_no_resched();
3641 } while (need_resched());
3643 exception_exit(prev_state);
3646 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3649 return try_to_wake_up(curr->private, mode, wake_flags);
3651 EXPORT_SYMBOL(default_wake_function);
3653 #ifdef CONFIG_RT_MUTEXES
3656 * rt_mutex_setprio - set the current priority of a task
3658 * @prio: prio value (kernel-internal form)
3660 * This function changes the 'effective' priority of a task. It does
3661 * not touch ->normal_prio like __setscheduler().
3663 * Used by the rt_mutex code to implement priority inheritance
3664 * logic. Call site only calls if the priority of the task changed.
3666 void rt_mutex_setprio(struct task_struct *p, int prio)
3668 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3669 const struct sched_class *prev_class;
3673 BUG_ON(prio > MAX_PRIO);
3675 rq = __task_rq_lock(p, &rf);
3676 update_rq_clock(rq);
3679 * Idle task boosting is a nono in general. There is one
3680 * exception, when PREEMPT_RT and NOHZ is active:
3682 * The idle task calls get_next_timer_interrupt() and holds
3683 * the timer wheel base->lock on the CPU and another CPU wants
3684 * to access the timer (probably to cancel it). We can safely
3685 * ignore the boosting request, as the idle CPU runs this code
3686 * with interrupts disabled and will complete the lock
3687 * protected section without being interrupted. So there is no
3688 * real need to boost.
3690 if (unlikely(p == rq->idle)) {
3691 WARN_ON(p != rq->curr);
3692 WARN_ON(p->pi_blocked_on);
3696 trace_sched_pi_setprio(p, prio);
3699 if (oldprio == prio)
3700 queue_flag &= ~DEQUEUE_MOVE;
3702 prev_class = p->sched_class;
3703 queued = task_on_rq_queued(p);
3704 running = task_current(rq, p);
3706 dequeue_task(rq, p, queue_flag);
3708 put_prev_task(rq, p);
3711 * Boosting condition are:
3712 * 1. -rt task is running and holds mutex A
3713 * --> -dl task blocks on mutex A
3715 * 2. -dl task is running and holds mutex A
3716 * --> -dl task blocks on mutex A and could preempt the
3719 if (dl_prio(prio)) {
3720 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3721 if (!dl_prio(p->normal_prio) ||
3722 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3723 p->dl.dl_boosted = 1;
3724 queue_flag |= ENQUEUE_REPLENISH;
3726 p->dl.dl_boosted = 0;
3727 p->sched_class = &dl_sched_class;
3728 } else if (rt_prio(prio)) {
3729 if (dl_prio(oldprio))
3730 p->dl.dl_boosted = 0;
3732 queue_flag |= ENQUEUE_HEAD;
3733 p->sched_class = &rt_sched_class;
3735 if (dl_prio(oldprio))
3736 p->dl.dl_boosted = 0;
3737 if (rt_prio(oldprio))
3739 p->sched_class = &fair_sched_class;
3745 enqueue_task(rq, p, queue_flag);
3747 set_curr_task(rq, p);
3749 check_class_changed(rq, p, prev_class, oldprio);
3751 /* Avoid rq from going away on us: */
3753 __task_rq_unlock(rq, &rf);
3755 balance_callback(rq);
3760 void set_user_nice(struct task_struct *p, long nice)
3762 bool queued, running;
3763 int old_prio, delta;
3767 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3770 * We have to be careful, if called from sys_setpriority(),
3771 * the task might be in the middle of scheduling on another CPU.
3773 rq = task_rq_lock(p, &rf);
3774 update_rq_clock(rq);
3777 * The RT priorities are set via sched_setscheduler(), but we still
3778 * allow the 'normal' nice value to be set - but as expected
3779 * it wont have any effect on scheduling until the task is
3780 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3782 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3783 p->static_prio = NICE_TO_PRIO(nice);
3786 queued = task_on_rq_queued(p);
3787 running = task_current(rq, p);
3789 dequeue_task(rq, p, DEQUEUE_SAVE);
3791 put_prev_task(rq, p);
3793 p->static_prio = NICE_TO_PRIO(nice);
3796 p->prio = effective_prio(p);
3797 delta = p->prio - old_prio;
3800 enqueue_task(rq, p, ENQUEUE_RESTORE);
3802 * If the task increased its priority or is running and
3803 * lowered its priority, then reschedule its CPU:
3805 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3809 set_curr_task(rq, p);
3811 task_rq_unlock(rq, p, &rf);
3813 EXPORT_SYMBOL(set_user_nice);
3816 * can_nice - check if a task can reduce its nice value
3820 int can_nice(const struct task_struct *p, const int nice)
3822 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3823 int nice_rlim = nice_to_rlimit(nice);
3825 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3826 capable(CAP_SYS_NICE));
3829 #ifdef __ARCH_WANT_SYS_NICE
3832 * sys_nice - change the priority of the current process.
3833 * @increment: priority increment
3835 * sys_setpriority is a more generic, but much slower function that
3836 * does similar things.
3838 SYSCALL_DEFINE1(nice, int, increment)
3843 * Setpriority might change our priority at the same moment.
3844 * We don't have to worry. Conceptually one call occurs first
3845 * and we have a single winner.
3847 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3848 nice = task_nice(current) + increment;
3850 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3851 if (increment < 0 && !can_nice(current, nice))
3854 retval = security_task_setnice(current, nice);
3858 set_user_nice(current, nice);
3865 * task_prio - return the priority value of a given task.
3866 * @p: the task in question.
3868 * Return: The priority value as seen by users in /proc.
3869 * RT tasks are offset by -200. Normal tasks are centered
3870 * around 0, value goes from -16 to +15.
3872 int task_prio(const struct task_struct *p)
3874 return p->prio - MAX_RT_PRIO;
3878 * idle_cpu - is a given CPU idle currently?
3879 * @cpu: the processor in question.
3881 * Return: 1 if the CPU is currently idle. 0 otherwise.
3883 int idle_cpu(int cpu)
3885 struct rq *rq = cpu_rq(cpu);
3887 if (rq->curr != rq->idle)
3894 if (!llist_empty(&rq->wake_list))
3902 * idle_task - return the idle task for a given CPU.
3903 * @cpu: the processor in question.
3905 * Return: The idle task for the CPU @cpu.
3907 struct task_struct *idle_task(int cpu)
3909 return cpu_rq(cpu)->idle;
3913 * find_process_by_pid - find a process with a matching PID value.
3914 * @pid: the pid in question.
3916 * The task of @pid, if found. %NULL otherwise.
3918 static struct task_struct *find_process_by_pid(pid_t pid)
3920 return pid ? find_task_by_vpid(pid) : current;
3924 * This function initializes the sched_dl_entity of a newly becoming
3925 * SCHED_DEADLINE task.
3927 * Only the static values are considered here, the actual runtime and the
3928 * absolute deadline will be properly calculated when the task is enqueued
3929 * for the first time with its new policy.
3932 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3934 struct sched_dl_entity *dl_se = &p->dl;
3936 dl_se->dl_runtime = attr->sched_runtime;
3937 dl_se->dl_deadline = attr->sched_deadline;
3938 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3939 dl_se->flags = attr->sched_flags;
3940 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3943 * Changing the parameters of a task is 'tricky' and we're not doing
3944 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3946 * What we SHOULD do is delay the bandwidth release until the 0-lag
3947 * point. This would include retaining the task_struct until that time
3948 * and change dl_overflow() to not immediately decrement the current
3951 * Instead we retain the current runtime/deadline and let the new
3952 * parameters take effect after the current reservation period lapses.
3953 * This is safe (albeit pessimistic) because the 0-lag point is always
3954 * before the current scheduling deadline.
3956 * We can still have temporary overloads because we do not delay the
3957 * change in bandwidth until that time; so admission control is
3958 * not on the safe side. It does however guarantee tasks will never
3959 * consume more than promised.
3964 * sched_setparam() passes in -1 for its policy, to let the functions
3965 * it calls know not to change it.
3967 #define SETPARAM_POLICY -1
3969 static void __setscheduler_params(struct task_struct *p,
3970 const struct sched_attr *attr)
3972 int policy = attr->sched_policy;
3974 if (policy == SETPARAM_POLICY)
3979 if (dl_policy(policy))
3980 __setparam_dl(p, attr);
3981 else if (fair_policy(policy))
3982 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3985 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3986 * !rt_policy. Always setting this ensures that things like
3987 * getparam()/getattr() don't report silly values for !rt tasks.
3989 p->rt_priority = attr->sched_priority;
3990 p->normal_prio = normal_prio(p);
3994 /* Actually do priority change: must hold pi & rq lock. */
3995 static void __setscheduler(struct rq *rq, struct task_struct *p,
3996 const struct sched_attr *attr, bool keep_boost)
3998 __setscheduler_params(p, attr);
4001 * Keep a potential priority boosting if called from
4002 * sched_setscheduler().
4005 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
4007 p->prio = normal_prio(p);
4009 if (dl_prio(p->prio))
4010 p->sched_class = &dl_sched_class;
4011 else if (rt_prio(p->prio))
4012 p->sched_class = &rt_sched_class;
4014 p->sched_class = &fair_sched_class;
4018 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
4020 struct sched_dl_entity *dl_se = &p->dl;
4022 attr->sched_priority = p->rt_priority;
4023 attr->sched_runtime = dl_se->dl_runtime;
4024 attr->sched_deadline = dl_se->dl_deadline;
4025 attr->sched_period = dl_se->dl_period;
4026 attr->sched_flags = dl_se->flags;
4030 * This function validates the new parameters of a -deadline task.
4031 * We ask for the deadline not being zero, and greater or equal
4032 * than the runtime, as well as the period of being zero or
4033 * greater than deadline. Furthermore, we have to be sure that
4034 * user parameters are above the internal resolution of 1us (we
4035 * check sched_runtime only since it is always the smaller one) and
4036 * below 2^63 ns (we have to check both sched_deadline and
4037 * sched_period, as the latter can be zero).
4040 __checkparam_dl(const struct sched_attr *attr)
4043 if (attr->sched_deadline == 0)
4047 * Since we truncate DL_SCALE bits, make sure we're at least
4050 if (attr->sched_runtime < (1ULL << DL_SCALE))
4054 * Since we use the MSB for wrap-around and sign issues, make
4055 * sure it's not set (mind that period can be equal to zero).
4057 if (attr->sched_deadline & (1ULL << 63) ||
4058 attr->sched_period & (1ULL << 63))
4061 /* runtime <= deadline <= period (if period != 0) */
4062 if ((attr->sched_period != 0 &&
4063 attr->sched_period < attr->sched_deadline) ||
4064 attr->sched_deadline < attr->sched_runtime)
4071 * Check the target process has a UID that matches the current process's:
4073 static bool check_same_owner(struct task_struct *p)
4075 const struct cred *cred = current_cred(), *pcred;
4079 pcred = __task_cred(p);
4080 match = (uid_eq(cred->euid, pcred->euid) ||
4081 uid_eq(cred->euid, pcred->uid));
4086 static bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
4088 struct sched_dl_entity *dl_se = &p->dl;
4090 if (dl_se->dl_runtime != attr->sched_runtime ||
4091 dl_se->dl_deadline != attr->sched_deadline ||
4092 dl_se->dl_period != attr->sched_period ||
4093 dl_se->flags != attr->sched_flags)
4099 static int __sched_setscheduler(struct task_struct *p,
4100 const struct sched_attr *attr,
4103 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4104 MAX_RT_PRIO - 1 - attr->sched_priority;
4105 int retval, oldprio, oldpolicy = -1, queued, running;
4106 int new_effective_prio, policy = attr->sched_policy;
4107 const struct sched_class *prev_class;
4110 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
4113 /* May grab non-irq protected spin_locks: */
4114 BUG_ON(in_interrupt());
4116 /* Double check policy once rq lock held: */
4118 reset_on_fork = p->sched_reset_on_fork;
4119 policy = oldpolicy = p->policy;
4121 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4123 if (!valid_policy(policy))
4127 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4131 * Valid priorities for SCHED_FIFO and SCHED_RR are
4132 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4133 * SCHED_BATCH and SCHED_IDLE is 0.
4135 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4136 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4138 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4139 (rt_policy(policy) != (attr->sched_priority != 0)))
4143 * Allow unprivileged RT tasks to decrease priority:
4145 if (user && !capable(CAP_SYS_NICE)) {
4146 if (fair_policy(policy)) {
4147 if (attr->sched_nice < task_nice(p) &&
4148 !can_nice(p, attr->sched_nice))
4152 if (rt_policy(policy)) {
4153 unsigned long rlim_rtprio =
4154 task_rlimit(p, RLIMIT_RTPRIO);
4156 /* Can't set/change the rt policy: */
4157 if (policy != p->policy && !rlim_rtprio)
4160 /* Can't increase priority: */
4161 if (attr->sched_priority > p->rt_priority &&
4162 attr->sched_priority > rlim_rtprio)
4167 * Can't set/change SCHED_DEADLINE policy at all for now
4168 * (safest behavior); in the future we would like to allow
4169 * unprivileged DL tasks to increase their relative deadline
4170 * or reduce their runtime (both ways reducing utilization)
4172 if (dl_policy(policy))
4176 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4177 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4179 if (idle_policy(p->policy) && !idle_policy(policy)) {
4180 if (!can_nice(p, task_nice(p)))
4184 /* Can't change other user's priorities: */
4185 if (!check_same_owner(p))
4188 /* Normal users shall not reset the sched_reset_on_fork flag: */
4189 if (p->sched_reset_on_fork && !reset_on_fork)
4194 retval = security_task_setscheduler(p);
4200 * Make sure no PI-waiters arrive (or leave) while we are
4201 * changing the priority of the task:
4203 * To be able to change p->policy safely, the appropriate
4204 * runqueue lock must be held.
4206 rq = task_rq_lock(p, &rf);
4207 update_rq_clock(rq);
4210 * Changing the policy of the stop threads its a very bad idea:
4212 if (p == rq->stop) {
4213 task_rq_unlock(rq, p, &rf);
4218 * If not changing anything there's no need to proceed further,
4219 * but store a possible modification of reset_on_fork.
4221 if (unlikely(policy == p->policy)) {
4222 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4224 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4226 if (dl_policy(policy) && dl_param_changed(p, attr))
4229 p->sched_reset_on_fork = reset_on_fork;
4230 task_rq_unlock(rq, p, &rf);
4236 #ifdef CONFIG_RT_GROUP_SCHED
4238 * Do not allow realtime tasks into groups that have no runtime
4241 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4242 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4243 !task_group_is_autogroup(task_group(p))) {
4244 task_rq_unlock(rq, p, &rf);
4249 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4250 cpumask_t *span = rq->rd->span;
4253 * Don't allow tasks with an affinity mask smaller than
4254 * the entire root_domain to become SCHED_DEADLINE. We
4255 * will also fail if there's no bandwidth available.
4257 if (!cpumask_subset(span, &p->cpus_allowed) ||
4258 rq->rd->dl_bw.bw == 0) {
4259 task_rq_unlock(rq, p, &rf);
4266 /* Re-check policy now with rq lock held: */
4267 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4268 policy = oldpolicy = -1;
4269 task_rq_unlock(rq, p, &rf);
4274 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4275 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4278 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4279 task_rq_unlock(rq, p, &rf);
4283 p->sched_reset_on_fork = reset_on_fork;
4288 * Take priority boosted tasks into account. If the new
4289 * effective priority is unchanged, we just store the new
4290 * normal parameters and do not touch the scheduler class and
4291 * the runqueue. This will be done when the task deboost
4294 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4295 if (new_effective_prio == oldprio)
4296 queue_flags &= ~DEQUEUE_MOVE;
4299 queued = task_on_rq_queued(p);
4300 running = task_current(rq, p);
4302 dequeue_task(rq, p, queue_flags);
4304 put_prev_task(rq, p);
4306 prev_class = p->sched_class;
4307 __setscheduler(rq, p, attr, pi);
4311 * We enqueue to tail when the priority of a task is
4312 * increased (user space view).
4314 if (oldprio < p->prio)
4315 queue_flags |= ENQUEUE_HEAD;
4317 enqueue_task(rq, p, queue_flags);
4320 set_curr_task(rq, p);
4322 check_class_changed(rq, p, prev_class, oldprio);
4324 /* Avoid rq from going away on us: */
4326 task_rq_unlock(rq, p, &rf);
4329 rt_mutex_adjust_pi(p);
4331 /* Run balance callbacks after we've adjusted the PI chain: */
4332 balance_callback(rq);
4338 static int _sched_setscheduler(struct task_struct *p, int policy,
4339 const struct sched_param *param, bool check)
4341 struct sched_attr attr = {
4342 .sched_policy = policy,
4343 .sched_priority = param->sched_priority,
4344 .sched_nice = PRIO_TO_NICE(p->static_prio),
4347 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4348 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4349 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4350 policy &= ~SCHED_RESET_ON_FORK;
4351 attr.sched_policy = policy;
4354 return __sched_setscheduler(p, &attr, check, true);
4357 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4358 * @p: the task in question.
4359 * @policy: new policy.
4360 * @param: structure containing the new RT priority.
4362 * Return: 0 on success. An error code otherwise.
4364 * NOTE that the task may be already dead.
4366 int sched_setscheduler(struct task_struct *p, int policy,
4367 const struct sched_param *param)
4369 return _sched_setscheduler(p, policy, param, true);
4371 EXPORT_SYMBOL_GPL(sched_setscheduler);
4373 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4375 return __sched_setscheduler(p, attr, true, true);
4377 EXPORT_SYMBOL_GPL(sched_setattr);
4380 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4381 * @p: the task in question.
4382 * @policy: new policy.
4383 * @param: structure containing the new RT priority.
4385 * Just like sched_setscheduler, only don't bother checking if the
4386 * current context has permission. For example, this is needed in
4387 * stop_machine(): we create temporary high priority worker threads,
4388 * but our caller might not have that capability.
4390 * Return: 0 on success. An error code otherwise.
4392 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4393 const struct sched_param *param)
4395 return _sched_setscheduler(p, policy, param, false);
4397 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4400 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4402 struct sched_param lparam;
4403 struct task_struct *p;
4406 if (!param || pid < 0)
4408 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4413 p = find_process_by_pid(pid);
4415 retval = sched_setscheduler(p, policy, &lparam);
4422 * Mimics kernel/events/core.c perf_copy_attr().
4424 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4429 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4432 /* Zero the full structure, so that a short copy will be nice: */
4433 memset(attr, 0, sizeof(*attr));
4435 ret = get_user(size, &uattr->size);
4439 /* Bail out on silly large: */
4440 if (size > PAGE_SIZE)
4443 /* ABI compatibility quirk: */
4445 size = SCHED_ATTR_SIZE_VER0;
4447 if (size < SCHED_ATTR_SIZE_VER0)
4451 * If we're handed a bigger struct than we know of,
4452 * ensure all the unknown bits are 0 - i.e. new
4453 * user-space does not rely on any kernel feature
4454 * extensions we dont know about yet.
4456 if (size > sizeof(*attr)) {
4457 unsigned char __user *addr;
4458 unsigned char __user *end;
4461 addr = (void __user *)uattr + sizeof(*attr);
4462 end = (void __user *)uattr + size;
4464 for (; addr < end; addr++) {
4465 ret = get_user(val, addr);
4471 size = sizeof(*attr);
4474 ret = copy_from_user(attr, uattr, size);
4479 * XXX: Do we want to be lenient like existing syscalls; or do we want
4480 * to be strict and return an error on out-of-bounds values?
4482 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4487 put_user(sizeof(*attr), &uattr->size);
4492 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4493 * @pid: the pid in question.
4494 * @policy: new policy.
4495 * @param: structure containing the new RT priority.
4497 * Return: 0 on success. An error code otherwise.
4499 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4504 return do_sched_setscheduler(pid, policy, param);
4508 * sys_sched_setparam - set/change the RT priority of a thread
4509 * @pid: the pid in question.
4510 * @param: structure containing the new RT priority.
4512 * Return: 0 on success. An error code otherwise.
4514 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4516 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4520 * sys_sched_setattr - same as above, but with extended sched_attr
4521 * @pid: the pid in question.
4522 * @uattr: structure containing the extended parameters.
4523 * @flags: for future extension.
4525 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4526 unsigned int, flags)
4528 struct sched_attr attr;
4529 struct task_struct *p;
4532 if (!uattr || pid < 0 || flags)
4535 retval = sched_copy_attr(uattr, &attr);
4539 if ((int)attr.sched_policy < 0)
4544 p = find_process_by_pid(pid);
4546 retval = sched_setattr(p, &attr);
4553 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4554 * @pid: the pid in question.
4556 * Return: On success, the policy of the thread. Otherwise, a negative error
4559 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4561 struct task_struct *p;
4569 p = find_process_by_pid(pid);
4571 retval = security_task_getscheduler(p);
4574 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4581 * sys_sched_getparam - get the RT priority of a thread
4582 * @pid: the pid in question.
4583 * @param: structure containing the RT priority.
4585 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4588 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4590 struct sched_param lp = { .sched_priority = 0 };
4591 struct task_struct *p;
4594 if (!param || pid < 0)
4598 p = find_process_by_pid(pid);
4603 retval = security_task_getscheduler(p);
4607 if (task_has_rt_policy(p))
4608 lp.sched_priority = p->rt_priority;
4612 * This one might sleep, we cannot do it with a spinlock held ...
4614 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4623 static int sched_read_attr(struct sched_attr __user *uattr,
4624 struct sched_attr *attr,
4629 if (!access_ok(VERIFY_WRITE, uattr, usize))
4633 * If we're handed a smaller struct than we know of,
4634 * ensure all the unknown bits are 0 - i.e. old
4635 * user-space does not get uncomplete information.
4637 if (usize < sizeof(*attr)) {
4638 unsigned char *addr;
4641 addr = (void *)attr + usize;
4642 end = (void *)attr + sizeof(*attr);
4644 for (; addr < end; addr++) {
4652 ret = copy_to_user(uattr, attr, attr->size);
4660 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4661 * @pid: the pid in question.
4662 * @uattr: structure containing the extended parameters.
4663 * @size: sizeof(attr) for fwd/bwd comp.
4664 * @flags: for future extension.
4666 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4667 unsigned int, size, unsigned int, flags)
4669 struct sched_attr attr = {
4670 .size = sizeof(struct sched_attr),
4672 struct task_struct *p;
4675 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4676 size < SCHED_ATTR_SIZE_VER0 || flags)
4680 p = find_process_by_pid(pid);
4685 retval = security_task_getscheduler(p);
4689 attr.sched_policy = p->policy;
4690 if (p->sched_reset_on_fork)
4691 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4692 if (task_has_dl_policy(p))
4693 __getparam_dl(p, &attr);
4694 else if (task_has_rt_policy(p))
4695 attr.sched_priority = p->rt_priority;
4697 attr.sched_nice = task_nice(p);
4701 retval = sched_read_attr(uattr, &attr, size);
4709 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4711 cpumask_var_t cpus_allowed, new_mask;
4712 struct task_struct *p;
4717 p = find_process_by_pid(pid);
4723 /* Prevent p going away */
4727 if (p->flags & PF_NO_SETAFFINITY) {
4731 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4735 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4737 goto out_free_cpus_allowed;
4740 if (!check_same_owner(p)) {
4742 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4744 goto out_free_new_mask;
4749 retval = security_task_setscheduler(p);
4751 goto out_free_new_mask;
4754 cpuset_cpus_allowed(p, cpus_allowed);
4755 cpumask_and(new_mask, in_mask, cpus_allowed);
4758 * Since bandwidth control happens on root_domain basis,
4759 * if admission test is enabled, we only admit -deadline
4760 * tasks allowed to run on all the CPUs in the task's
4764 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4766 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4769 goto out_free_new_mask;
4775 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4778 cpuset_cpus_allowed(p, cpus_allowed);
4779 if (!cpumask_subset(new_mask, cpus_allowed)) {
4781 * We must have raced with a concurrent cpuset
4782 * update. Just reset the cpus_allowed to the
4783 * cpuset's cpus_allowed
4785 cpumask_copy(new_mask, cpus_allowed);
4790 free_cpumask_var(new_mask);
4791 out_free_cpus_allowed:
4792 free_cpumask_var(cpus_allowed);
4798 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4799 struct cpumask *new_mask)
4801 if (len < cpumask_size())
4802 cpumask_clear(new_mask);
4803 else if (len > cpumask_size())
4804 len = cpumask_size();
4806 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4810 * sys_sched_setaffinity - set the CPU affinity of a process
4811 * @pid: pid of the process
4812 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4813 * @user_mask_ptr: user-space pointer to the new CPU mask
4815 * Return: 0 on success. An error code otherwise.
4817 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4818 unsigned long __user *, user_mask_ptr)
4820 cpumask_var_t new_mask;
4823 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4826 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4828 retval = sched_setaffinity(pid, new_mask);
4829 free_cpumask_var(new_mask);
4833 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4835 struct task_struct *p;
4836 unsigned long flags;
4842 p = find_process_by_pid(pid);
4846 retval = security_task_getscheduler(p);
4850 raw_spin_lock_irqsave(&p->pi_lock, flags);
4851 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4852 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4861 * sys_sched_getaffinity - get the CPU affinity of a process
4862 * @pid: pid of the process
4863 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4864 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4866 * Return: size of CPU mask copied to user_mask_ptr on success. An
4867 * error code otherwise.
4869 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4870 unsigned long __user *, user_mask_ptr)
4875 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4877 if (len & (sizeof(unsigned long)-1))
4880 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4883 ret = sched_getaffinity(pid, mask);
4885 size_t retlen = min_t(size_t, len, cpumask_size());
4887 if (copy_to_user(user_mask_ptr, mask, retlen))
4892 free_cpumask_var(mask);
4898 * sys_sched_yield - yield the current processor to other threads.
4900 * This function yields the current CPU to other tasks. If there are no
4901 * other threads running on this CPU then this function will return.
4905 SYSCALL_DEFINE0(sched_yield)
4907 struct rq *rq = this_rq_lock();
4909 schedstat_inc(rq->yld_count);
4910 current->sched_class->yield_task(rq);
4913 * Since we are going to call schedule() anyway, there's
4914 * no need to preempt or enable interrupts:
4916 __release(rq->lock);
4917 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4918 do_raw_spin_unlock(&rq->lock);
4919 sched_preempt_enable_no_resched();
4926 #ifndef CONFIG_PREEMPT
4927 int __sched _cond_resched(void)
4929 if (should_resched(0)) {
4930 preempt_schedule_common();
4935 EXPORT_SYMBOL(_cond_resched);
4939 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4940 * call schedule, and on return reacquire the lock.
4942 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4943 * operations here to prevent schedule() from being called twice (once via
4944 * spin_unlock(), once by hand).
4946 int __cond_resched_lock(spinlock_t *lock)
4948 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4951 lockdep_assert_held(lock);
4953 if (spin_needbreak(lock) || resched) {
4956 preempt_schedule_common();
4964 EXPORT_SYMBOL(__cond_resched_lock);
4966 int __sched __cond_resched_softirq(void)
4968 BUG_ON(!in_softirq());
4970 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4972 preempt_schedule_common();
4978 EXPORT_SYMBOL(__cond_resched_softirq);
4981 * yield - yield the current processor to other threads.
4983 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4985 * The scheduler is at all times free to pick the calling task as the most
4986 * eligible task to run, if removing the yield() call from your code breaks
4987 * it, its already broken.
4989 * Typical broken usage is:
4994 * where one assumes that yield() will let 'the other' process run that will
4995 * make event true. If the current task is a SCHED_FIFO task that will never
4996 * happen. Never use yield() as a progress guarantee!!
4998 * If you want to use yield() to wait for something, use wait_event().
4999 * If you want to use yield() to be 'nice' for others, use cond_resched().
5000 * If you still want to use yield(), do not!
5002 void __sched yield(void)
5004 set_current_state(TASK_RUNNING);
5007 EXPORT_SYMBOL(yield);
5010 * yield_to - yield the current processor to another thread in
5011 * your thread group, or accelerate that thread toward the
5012 * processor it's on.
5014 * @preempt: whether task preemption is allowed or not
5016 * It's the caller's job to ensure that the target task struct
5017 * can't go away on us before we can do any checks.
5020 * true (>0) if we indeed boosted the target task.
5021 * false (0) if we failed to boost the target.
5022 * -ESRCH if there's no task to yield to.
5024 int __sched yield_to(struct task_struct *p, bool preempt)
5026 struct task_struct *curr = current;
5027 struct rq *rq, *p_rq;
5028 unsigned long flags;
5031 local_irq_save(flags);
5037 * If we're the only runnable task on the rq and target rq also
5038 * has only one task, there's absolutely no point in yielding.
5040 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5045 double_rq_lock(rq, p_rq);
5046 if (task_rq(p) != p_rq) {
5047 double_rq_unlock(rq, p_rq);
5051 if (!curr->sched_class->yield_to_task)
5054 if (curr->sched_class != p->sched_class)
5057 if (task_running(p_rq, p) || p->state)
5060 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5062 schedstat_inc(rq->yld_count);
5064 * Make p's CPU reschedule; pick_next_entity takes care of
5067 if (preempt && rq != p_rq)
5072 double_rq_unlock(rq, p_rq);
5074 local_irq_restore(flags);
5081 EXPORT_SYMBOL_GPL(yield_to);
5083 int io_schedule_prepare(void)
5085 int old_iowait = current->in_iowait;
5087 current->in_iowait = 1;
5088 blk_schedule_flush_plug(current);
5093 void io_schedule_finish(int token)
5095 current->in_iowait = token;
5099 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5100 * that process accounting knows that this is a task in IO wait state.
5102 long __sched io_schedule_timeout(long timeout)
5107 token = io_schedule_prepare();
5108 ret = schedule_timeout(timeout);
5109 io_schedule_finish(token);
5113 EXPORT_SYMBOL(io_schedule_timeout);
5115 void io_schedule(void)
5119 token = io_schedule_prepare();
5121 io_schedule_finish(token);
5123 EXPORT_SYMBOL(io_schedule);
5126 * sys_sched_get_priority_max - return maximum RT priority.
5127 * @policy: scheduling class.
5129 * Return: On success, this syscall returns the maximum
5130 * rt_priority that can be used by a given scheduling class.
5131 * On failure, a negative error code is returned.
5133 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5140 ret = MAX_USER_RT_PRIO-1;
5142 case SCHED_DEADLINE:
5153 * sys_sched_get_priority_min - return minimum RT priority.
5154 * @policy: scheduling class.
5156 * Return: On success, this syscall returns the minimum
5157 * rt_priority that can be used by a given scheduling class.
5158 * On failure, a negative error code is returned.
5160 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5169 case SCHED_DEADLINE:
5179 * sys_sched_rr_get_interval - return the default timeslice of a process.
5180 * @pid: pid of the process.
5181 * @interval: userspace pointer to the timeslice value.
5183 * this syscall writes the default timeslice value of a given process
5184 * into the user-space timespec buffer. A value of '0' means infinity.
5186 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5189 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5190 struct timespec __user *, interval)
5192 struct task_struct *p;
5193 unsigned int time_slice;
5204 p = find_process_by_pid(pid);
5208 retval = security_task_getscheduler(p);
5212 rq = task_rq_lock(p, &rf);
5214 if (p->sched_class->get_rr_interval)
5215 time_slice = p->sched_class->get_rr_interval(rq, p);
5216 task_rq_unlock(rq, p, &rf);
5219 jiffies_to_timespec(time_slice, &t);
5220 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5228 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5230 void sched_show_task(struct task_struct *p)
5232 unsigned long free = 0;
5234 unsigned long state = p->state;
5236 if (!try_get_task_stack(p))
5239 state = __ffs(state) + 1;
5240 printk(KERN_INFO "%-15.15s %c", p->comm,
5241 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5242 if (state == TASK_RUNNING)
5243 printk(KERN_CONT " running task ");
5244 #ifdef CONFIG_DEBUG_STACK_USAGE
5245 free = stack_not_used(p);
5250 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5252 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5253 task_pid_nr(p), ppid,
5254 (unsigned long)task_thread_info(p)->flags);
5256 print_worker_info(KERN_INFO, p);
5257 show_stack(p, NULL);
5261 void show_state_filter(unsigned long state_filter)
5263 struct task_struct *g, *p;
5265 #if BITS_PER_LONG == 32
5267 " task PC stack pid father\n");
5270 " task PC stack pid father\n");
5273 for_each_process_thread(g, p) {
5275 * reset the NMI-timeout, listing all files on a slow
5276 * console might take a lot of time:
5277 * Also, reset softlockup watchdogs on all CPUs, because
5278 * another CPU might be blocked waiting for us to process
5281 touch_nmi_watchdog();
5282 touch_all_softlockup_watchdogs();
5283 if (!state_filter || (p->state & state_filter))
5287 #ifdef CONFIG_SCHED_DEBUG
5289 sysrq_sched_debug_show();
5293 * Only show locks if all tasks are dumped:
5296 debug_show_all_locks();
5299 void init_idle_bootup_task(struct task_struct *idle)
5301 idle->sched_class = &idle_sched_class;
5305 * init_idle - set up an idle thread for a given CPU
5306 * @idle: task in question
5307 * @cpu: CPU the idle task belongs to
5309 * NOTE: this function does not set the idle thread's NEED_RESCHED
5310 * flag, to make booting more robust.
5312 void init_idle(struct task_struct *idle, int cpu)
5314 struct rq *rq = cpu_rq(cpu);
5315 unsigned long flags;
5317 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5318 raw_spin_lock(&rq->lock);
5320 __sched_fork(0, idle);
5321 idle->state = TASK_RUNNING;
5322 idle->se.exec_start = sched_clock();
5323 idle->flags |= PF_IDLE;
5325 kasan_unpoison_task_stack(idle);
5329 * Its possible that init_idle() gets called multiple times on a task,
5330 * in that case do_set_cpus_allowed() will not do the right thing.
5332 * And since this is boot we can forgo the serialization.
5334 set_cpus_allowed_common(idle, cpumask_of(cpu));
5337 * We're having a chicken and egg problem, even though we are
5338 * holding rq->lock, the CPU isn't yet set to this CPU so the
5339 * lockdep check in task_group() will fail.
5341 * Similar case to sched_fork(). / Alternatively we could
5342 * use task_rq_lock() here and obtain the other rq->lock.
5347 __set_task_cpu(idle, cpu);
5350 rq->curr = rq->idle = idle;
5351 idle->on_rq = TASK_ON_RQ_QUEUED;
5355 raw_spin_unlock(&rq->lock);
5356 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5358 /* Set the preempt count _outside_ the spinlocks! */
5359 init_idle_preempt_count(idle, cpu);
5362 * The idle tasks have their own, simple scheduling class:
5364 idle->sched_class = &idle_sched_class;
5365 ftrace_graph_init_idle_task(idle, cpu);
5366 vtime_init_idle(idle, cpu);
5368 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5372 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5373 const struct cpumask *trial)
5375 int ret = 1, trial_cpus;
5376 struct dl_bw *cur_dl_b;
5377 unsigned long flags;
5379 if (!cpumask_weight(cur))
5382 rcu_read_lock_sched();
5383 cur_dl_b = dl_bw_of(cpumask_any(cur));
5384 trial_cpus = cpumask_weight(trial);
5386 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5387 if (cur_dl_b->bw != -1 &&
5388 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5390 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5391 rcu_read_unlock_sched();
5396 int task_can_attach(struct task_struct *p,
5397 const struct cpumask *cs_cpus_allowed)
5402 * Kthreads which disallow setaffinity shouldn't be moved
5403 * to a new cpuset; we don't want to change their CPU
5404 * affinity and isolating such threads by their set of
5405 * allowed nodes is unnecessary. Thus, cpusets are not
5406 * applicable for such threads. This prevents checking for
5407 * success of set_cpus_allowed_ptr() on all attached tasks
5408 * before cpus_allowed may be changed.
5410 if (p->flags & PF_NO_SETAFFINITY) {
5416 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5418 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5423 unsigned long flags;
5425 rcu_read_lock_sched();
5426 dl_b = dl_bw_of(dest_cpu);
5427 raw_spin_lock_irqsave(&dl_b->lock, flags);
5428 cpus = dl_bw_cpus(dest_cpu);
5429 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5434 * We reserve space for this task in the destination
5435 * root_domain, as we can't fail after this point.
5436 * We will free resources in the source root_domain
5437 * later on (see set_cpus_allowed_dl()).
5439 __dl_add(dl_b, p->dl.dl_bw);
5441 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5442 rcu_read_unlock_sched();
5452 bool sched_smp_initialized __read_mostly;
5454 #ifdef CONFIG_NUMA_BALANCING
5455 /* Migrate current task p to target_cpu */
5456 int migrate_task_to(struct task_struct *p, int target_cpu)
5458 struct migration_arg arg = { p, target_cpu };
5459 int curr_cpu = task_cpu(p);
5461 if (curr_cpu == target_cpu)
5464 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5467 /* TODO: This is not properly updating schedstats */
5469 trace_sched_move_numa(p, curr_cpu, target_cpu);
5470 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5474 * Requeue a task on a given node and accurately track the number of NUMA
5475 * tasks on the runqueues
5477 void sched_setnuma(struct task_struct *p, int nid)
5479 bool queued, running;
5483 rq = task_rq_lock(p, &rf);
5484 queued = task_on_rq_queued(p);
5485 running = task_current(rq, p);
5488 dequeue_task(rq, p, DEQUEUE_SAVE);
5490 put_prev_task(rq, p);
5492 p->numa_preferred_nid = nid;
5495 enqueue_task(rq, p, ENQUEUE_RESTORE);
5497 set_curr_task(rq, p);
5498 task_rq_unlock(rq, p, &rf);
5500 #endif /* CONFIG_NUMA_BALANCING */
5502 #ifdef CONFIG_HOTPLUG_CPU
5504 * Ensure that the idle task is using init_mm right before its CPU goes
5507 void idle_task_exit(void)
5509 struct mm_struct *mm = current->active_mm;
5511 BUG_ON(cpu_online(smp_processor_id()));
5513 if (mm != &init_mm) {
5514 switch_mm_irqs_off(mm, &init_mm, current);
5515 finish_arch_post_lock_switch();
5521 * Since this CPU is going 'away' for a while, fold any nr_active delta
5522 * we might have. Assumes we're called after migrate_tasks() so that the
5523 * nr_active count is stable. We need to take the teardown thread which
5524 * is calling this into account, so we hand in adjust = 1 to the load
5527 * Also see the comment "Global load-average calculations".
5529 static void calc_load_migrate(struct rq *rq)
5531 long delta = calc_load_fold_active(rq, 1);
5533 atomic_long_add(delta, &calc_load_tasks);
5536 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5540 static const struct sched_class fake_sched_class = {
5541 .put_prev_task = put_prev_task_fake,
5544 static struct task_struct fake_task = {
5546 * Avoid pull_{rt,dl}_task()
5548 .prio = MAX_PRIO + 1,
5549 .sched_class = &fake_sched_class,
5553 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5554 * try_to_wake_up()->select_task_rq().
5556 * Called with rq->lock held even though we'er in stop_machine() and
5557 * there's no concurrency possible, we hold the required locks anyway
5558 * because of lock validation efforts.
5560 static void migrate_tasks(struct rq *dead_rq)
5562 struct rq *rq = dead_rq;
5563 struct task_struct *next, *stop = rq->stop;
5568 * Fudge the rq selection such that the below task selection loop
5569 * doesn't get stuck on the currently eligible stop task.
5571 * We're currently inside stop_machine() and the rq is either stuck
5572 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5573 * either way we should never end up calling schedule() until we're
5579 * put_prev_task() and pick_next_task() sched
5580 * class method both need to have an up-to-date
5581 * value of rq->clock[_task]
5583 rq_pin_lock(rq, &rf);
5584 update_rq_clock(rq);
5585 rq_unpin_lock(rq, &rf);
5589 * There's this thread running, bail when that's the only
5592 if (rq->nr_running == 1)
5596 * pick_next_task() assumes pinned rq->lock:
5598 rq_repin_lock(rq, &rf);
5599 next = pick_next_task(rq, &fake_task, &rf);
5601 next->sched_class->put_prev_task(rq, next);
5604 * Rules for changing task_struct::cpus_allowed are holding
5605 * both pi_lock and rq->lock, such that holding either
5606 * stabilizes the mask.
5608 * Drop rq->lock is not quite as disastrous as it usually is
5609 * because !cpu_active at this point, which means load-balance
5610 * will not interfere. Also, stop-machine.
5612 rq_unpin_lock(rq, &rf);
5613 raw_spin_unlock(&rq->lock);
5614 raw_spin_lock(&next->pi_lock);
5615 raw_spin_lock(&rq->lock);
5618 * Since we're inside stop-machine, _nothing_ should have
5619 * changed the task, WARN if weird stuff happened, because in
5620 * that case the above rq->lock drop is a fail too.
5622 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5623 raw_spin_unlock(&next->pi_lock);
5627 /* Find suitable destination for @next, with force if needed. */
5628 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5630 rq = __migrate_task(rq, next, dest_cpu);
5631 if (rq != dead_rq) {
5632 raw_spin_unlock(&rq->lock);
5634 raw_spin_lock(&rq->lock);
5636 raw_spin_unlock(&next->pi_lock);
5641 #endif /* CONFIG_HOTPLUG_CPU */
5643 void set_rq_online(struct rq *rq)
5646 const struct sched_class *class;
5648 cpumask_set_cpu(rq->cpu, rq->rd->online);
5651 for_each_class(class) {
5652 if (class->rq_online)
5653 class->rq_online(rq);
5658 void set_rq_offline(struct rq *rq)
5661 const struct sched_class *class;
5663 for_each_class(class) {
5664 if (class->rq_offline)
5665 class->rq_offline(rq);
5668 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5673 static void set_cpu_rq_start_time(unsigned int cpu)
5675 struct rq *rq = cpu_rq(cpu);
5677 rq->age_stamp = sched_clock_cpu(cpu);
5681 * used to mark begin/end of suspend/resume:
5683 static int num_cpus_frozen;
5686 * Update cpusets according to cpu_active mask. If cpusets are
5687 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5688 * around partition_sched_domains().
5690 * If we come here as part of a suspend/resume, don't touch cpusets because we
5691 * want to restore it back to its original state upon resume anyway.
5693 static void cpuset_cpu_active(void)
5695 if (cpuhp_tasks_frozen) {
5697 * num_cpus_frozen tracks how many CPUs are involved in suspend
5698 * resume sequence. As long as this is not the last online
5699 * operation in the resume sequence, just build a single sched
5700 * domain, ignoring cpusets.
5703 if (likely(num_cpus_frozen)) {
5704 partition_sched_domains(1, NULL, NULL);
5708 * This is the last CPU online operation. So fall through and
5709 * restore the original sched domains by considering the
5710 * cpuset configurations.
5713 cpuset_update_active_cpus(true);
5716 static int cpuset_cpu_inactive(unsigned int cpu)
5718 unsigned long flags;
5723 if (!cpuhp_tasks_frozen) {
5724 rcu_read_lock_sched();
5725 dl_b = dl_bw_of(cpu);
5727 raw_spin_lock_irqsave(&dl_b->lock, flags);
5728 cpus = dl_bw_cpus(cpu);
5729 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5730 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5732 rcu_read_unlock_sched();
5736 cpuset_update_active_cpus(false);
5739 partition_sched_domains(1, NULL, NULL);
5744 int sched_cpu_activate(unsigned int cpu)
5746 struct rq *rq = cpu_rq(cpu);
5747 unsigned long flags;
5749 set_cpu_active(cpu, true);
5751 if (sched_smp_initialized) {
5752 sched_domains_numa_masks_set(cpu);
5753 cpuset_cpu_active();
5757 * Put the rq online, if not already. This happens:
5759 * 1) In the early boot process, because we build the real domains
5760 * after all CPUs have been brought up.
5762 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5765 raw_spin_lock_irqsave(&rq->lock, flags);
5767 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5770 raw_spin_unlock_irqrestore(&rq->lock, flags);
5772 update_max_interval();
5777 int sched_cpu_deactivate(unsigned int cpu)
5781 set_cpu_active(cpu, false);
5783 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5784 * users of this state to go away such that all new such users will
5787 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
5788 * not imply sync_sched(), so wait for both.
5790 * Do sync before park smpboot threads to take care the rcu boost case.
5792 if (IS_ENABLED(CONFIG_PREEMPT))
5793 synchronize_rcu_mult(call_rcu, call_rcu_sched);
5797 if (!sched_smp_initialized)
5800 ret = cpuset_cpu_inactive(cpu);
5802 set_cpu_active(cpu, true);
5805 sched_domains_numa_masks_clear(cpu);
5809 static void sched_rq_cpu_starting(unsigned int cpu)
5811 struct rq *rq = cpu_rq(cpu);
5813 rq->calc_load_update = calc_load_update;
5814 update_max_interval();
5817 int sched_cpu_starting(unsigned int cpu)
5819 set_cpu_rq_start_time(cpu);
5820 sched_rq_cpu_starting(cpu);
5824 #ifdef CONFIG_HOTPLUG_CPU
5825 int sched_cpu_dying(unsigned int cpu)
5827 struct rq *rq = cpu_rq(cpu);
5828 unsigned long flags;
5830 /* Handle pending wakeups and then migrate everything off */
5831 sched_ttwu_pending();
5832 raw_spin_lock_irqsave(&rq->lock, flags);
5834 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5838 BUG_ON(rq->nr_running != 1);
5839 raw_spin_unlock_irqrestore(&rq->lock, flags);
5840 calc_load_migrate(rq);
5841 update_max_interval();
5842 nohz_balance_exit_idle(cpu);
5848 #ifdef CONFIG_SCHED_SMT
5849 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5851 static void sched_init_smt(void)
5854 * We've enumerated all CPUs and will assume that if any CPU
5855 * has SMT siblings, CPU0 will too.
5857 if (cpumask_weight(cpu_smt_mask(0)) > 1)
5858 static_branch_enable(&sched_smt_present);
5861 static inline void sched_init_smt(void) { }
5864 void __init sched_init_smp(void)
5866 cpumask_var_t non_isolated_cpus;
5868 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
5869 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
5874 * There's no userspace yet to cause hotplug operations; hence all the
5875 * CPU masks are stable and all blatant races in the below code cannot
5878 mutex_lock(&sched_domains_mutex);
5879 init_sched_domains(cpu_active_mask);
5880 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
5881 if (cpumask_empty(non_isolated_cpus))
5882 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
5883 mutex_unlock(&sched_domains_mutex);
5885 /* Move init over to a non-isolated CPU */
5886 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
5888 sched_init_granularity();
5889 free_cpumask_var(non_isolated_cpus);
5891 init_sched_rt_class();
5892 init_sched_dl_class();
5895 sched_clock_init_late();
5897 sched_smp_initialized = true;
5900 static int __init migration_init(void)
5902 sched_rq_cpu_starting(smp_processor_id());
5905 early_initcall(migration_init);
5908 void __init sched_init_smp(void)
5910 sched_init_granularity();
5911 sched_clock_init_late();
5913 #endif /* CONFIG_SMP */
5915 int in_sched_functions(unsigned long addr)
5917 return in_lock_functions(addr) ||
5918 (addr >= (unsigned long)__sched_text_start
5919 && addr < (unsigned long)__sched_text_end);
5922 #ifdef CONFIG_CGROUP_SCHED
5924 * Default task group.
5925 * Every task in system belongs to this group at bootup.
5927 struct task_group root_task_group;
5928 LIST_HEAD(task_groups);
5930 /* Cacheline aligned slab cache for task_group */
5931 static struct kmem_cache *task_group_cache __read_mostly;
5934 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5935 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5937 #define WAIT_TABLE_BITS 8
5938 #define WAIT_TABLE_SIZE (1 << WAIT_TABLE_BITS)
5939 static wait_queue_head_t bit_wait_table[WAIT_TABLE_SIZE] __cacheline_aligned;
5941 wait_queue_head_t *bit_waitqueue(void *word, int bit)
5943 const int shift = BITS_PER_LONG == 32 ? 5 : 6;
5944 unsigned long val = (unsigned long)word << shift | bit;
5946 return bit_wait_table + hash_long(val, WAIT_TABLE_BITS);
5948 EXPORT_SYMBOL(bit_waitqueue);
5950 void __init sched_init(void)
5953 unsigned long alloc_size = 0, ptr;
5957 for (i = 0; i < WAIT_TABLE_SIZE; i++)
5958 init_waitqueue_head(bit_wait_table + i);
5960 #ifdef CONFIG_FAIR_GROUP_SCHED
5961 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5963 #ifdef CONFIG_RT_GROUP_SCHED
5964 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5967 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5969 #ifdef CONFIG_FAIR_GROUP_SCHED
5970 root_task_group.se = (struct sched_entity **)ptr;
5971 ptr += nr_cpu_ids * sizeof(void **);
5973 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5974 ptr += nr_cpu_ids * sizeof(void **);
5976 #endif /* CONFIG_FAIR_GROUP_SCHED */
5977 #ifdef CONFIG_RT_GROUP_SCHED
5978 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5979 ptr += nr_cpu_ids * sizeof(void **);
5981 root_task_group.rt_rq = (struct rt_rq **)ptr;
5982 ptr += nr_cpu_ids * sizeof(void **);
5984 #endif /* CONFIG_RT_GROUP_SCHED */
5986 #ifdef CONFIG_CPUMASK_OFFSTACK
5987 for_each_possible_cpu(i) {
5988 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5989 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5990 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
5991 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5993 #endif /* CONFIG_CPUMASK_OFFSTACK */
5995 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
5996 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
5999 init_defrootdomain();
6002 #ifdef CONFIG_RT_GROUP_SCHED
6003 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6004 global_rt_period(), global_rt_runtime());
6005 #endif /* CONFIG_RT_GROUP_SCHED */
6007 #ifdef CONFIG_CGROUP_SCHED
6008 task_group_cache = KMEM_CACHE(task_group, 0);
6010 list_add(&root_task_group.list, &task_groups);
6011 INIT_LIST_HEAD(&root_task_group.children);
6012 INIT_LIST_HEAD(&root_task_group.siblings);
6013 autogroup_init(&init_task);
6014 #endif /* CONFIG_CGROUP_SCHED */
6016 for_each_possible_cpu(i) {
6020 raw_spin_lock_init(&rq->lock);
6022 rq->calc_load_active = 0;
6023 rq->calc_load_update = jiffies + LOAD_FREQ;
6024 init_cfs_rq(&rq->cfs);
6025 init_rt_rq(&rq->rt);
6026 init_dl_rq(&rq->dl);
6027 #ifdef CONFIG_FAIR_GROUP_SCHED
6028 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6029 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6030 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6032 * How much CPU bandwidth does root_task_group get?
6034 * In case of task-groups formed thr' the cgroup filesystem, it
6035 * gets 100% of the CPU resources in the system. This overall
6036 * system CPU resource is divided among the tasks of
6037 * root_task_group and its child task-groups in a fair manner,
6038 * based on each entity's (task or task-group's) weight
6039 * (se->load.weight).
6041 * In other words, if root_task_group has 10 tasks of weight
6042 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6043 * then A0's share of the CPU resource is:
6045 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6047 * We achieve this by letting root_task_group's tasks sit
6048 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6050 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6051 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6052 #endif /* CONFIG_FAIR_GROUP_SCHED */
6054 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6055 #ifdef CONFIG_RT_GROUP_SCHED
6056 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6059 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6060 rq->cpu_load[j] = 0;
6065 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6066 rq->balance_callback = NULL;
6067 rq->active_balance = 0;
6068 rq->next_balance = jiffies;
6073 rq->avg_idle = 2*sysctl_sched_migration_cost;
6074 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6076 INIT_LIST_HEAD(&rq->cfs_tasks);
6078 rq_attach_root(rq, &def_root_domain);
6079 #ifdef CONFIG_NO_HZ_COMMON
6080 rq->last_load_update_tick = jiffies;
6083 #ifdef CONFIG_NO_HZ_FULL
6084 rq->last_sched_tick = 0;
6086 #endif /* CONFIG_SMP */
6088 atomic_set(&rq->nr_iowait, 0);
6091 set_load_weight(&init_task);
6094 * The boot idle thread does lazy MMU switching as well:
6097 enter_lazy_tlb(&init_mm, current);
6100 * Make us the idle thread. Technically, schedule() should not be
6101 * called from this thread, however somewhere below it might be,
6102 * but because we are the idle thread, we just pick up running again
6103 * when this runqueue becomes "idle".
6105 init_idle(current, smp_processor_id());
6107 calc_load_update = jiffies + LOAD_FREQ;
6110 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6111 /* May be allocated at isolcpus cmdline parse time */
6112 if (cpu_isolated_map == NULL)
6113 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6114 idle_thread_set_boot_cpu();
6115 set_cpu_rq_start_time(smp_processor_id());
6117 init_sched_fair_class();
6121 scheduler_running = 1;
6124 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6125 static inline int preempt_count_equals(int preempt_offset)
6127 int nested = preempt_count() + rcu_preempt_depth();
6129 return (nested == preempt_offset);
6132 void __might_sleep(const char *file, int line, int preempt_offset)
6135 * Blocking primitives will set (and therefore destroy) current->state,
6136 * since we will exit with TASK_RUNNING make sure we enter with it,
6137 * otherwise we will destroy state.
6139 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6140 "do not call blocking ops when !TASK_RUNNING; "
6141 "state=%lx set at [<%p>] %pS\n",
6143 (void *)current->task_state_change,
6144 (void *)current->task_state_change);
6146 ___might_sleep(file, line, preempt_offset);
6148 EXPORT_SYMBOL(__might_sleep);
6150 void ___might_sleep(const char *file, int line, int preempt_offset)
6152 /* Ratelimiting timestamp: */
6153 static unsigned long prev_jiffy;
6155 unsigned long preempt_disable_ip;
6157 /* WARN_ON_ONCE() by default, no rate limit required: */
6160 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6161 !is_idle_task(current)) ||
6162 system_state != SYSTEM_RUNNING || oops_in_progress)
6164 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6166 prev_jiffy = jiffies;
6168 /* Save this before calling printk(), since that will clobber it: */
6169 preempt_disable_ip = get_preempt_disable_ip(current);
6172 "BUG: sleeping function called from invalid context at %s:%d\n",
6175 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6176 in_atomic(), irqs_disabled(),
6177 current->pid, current->comm);
6179 if (task_stack_end_corrupted(current))
6180 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6182 debug_show_held_locks(current);
6183 if (irqs_disabled())
6184 print_irqtrace_events(current);
6185 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6186 && !preempt_count_equals(preempt_offset)) {
6187 pr_err("Preemption disabled at:");
6188 print_ip_sym(preempt_disable_ip);
6192 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6194 EXPORT_SYMBOL(___might_sleep);
6197 #ifdef CONFIG_MAGIC_SYSRQ
6198 void normalize_rt_tasks(void)
6200 struct task_struct *g, *p;
6201 struct sched_attr attr = {
6202 .sched_policy = SCHED_NORMAL,
6205 read_lock(&tasklist_lock);
6206 for_each_process_thread(g, p) {
6208 * Only normalize user tasks:
6210 if (p->flags & PF_KTHREAD)
6213 p->se.exec_start = 0;
6214 schedstat_set(p->se.statistics.wait_start, 0);
6215 schedstat_set(p->se.statistics.sleep_start, 0);
6216 schedstat_set(p->se.statistics.block_start, 0);
6218 if (!dl_task(p) && !rt_task(p)) {
6220 * Renice negative nice level userspace
6223 if (task_nice(p) < 0)
6224 set_user_nice(p, 0);
6228 __sched_setscheduler(p, &attr, false, false);
6230 read_unlock(&tasklist_lock);
6233 #endif /* CONFIG_MAGIC_SYSRQ */
6235 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6237 * These functions are only useful for the IA64 MCA handling, or kdb.
6239 * They can only be called when the whole system has been
6240 * stopped - every CPU needs to be quiescent, and no scheduling
6241 * activity can take place. Using them for anything else would
6242 * be a serious bug, and as a result, they aren't even visible
6243 * under any other configuration.
6247 * curr_task - return the current task for a given CPU.
6248 * @cpu: the processor in question.
6250 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6252 * Return: The current task for @cpu.
6254 struct task_struct *curr_task(int cpu)
6256 return cpu_curr(cpu);
6259 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6263 * set_curr_task - set the current task for a given CPU.
6264 * @cpu: the processor in question.
6265 * @p: the task pointer to set.
6267 * Description: This function must only be used when non-maskable interrupts
6268 * are serviced on a separate stack. It allows the architecture to switch the
6269 * notion of the current task on a CPU in a non-blocking manner. This function
6270 * must be called with all CPU's synchronized, and interrupts disabled, the
6271 * and caller must save the original value of the current task (see
6272 * curr_task() above) and restore that value before reenabling interrupts and
6273 * re-starting the system.
6275 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6277 void ia64_set_curr_task(int cpu, struct task_struct *p)
6284 #ifdef CONFIG_CGROUP_SCHED
6285 /* task_group_lock serializes the addition/removal of task groups */
6286 static DEFINE_SPINLOCK(task_group_lock);
6288 static void sched_free_group(struct task_group *tg)
6290 free_fair_sched_group(tg);
6291 free_rt_sched_group(tg);
6293 kmem_cache_free(task_group_cache, tg);
6296 /* allocate runqueue etc for a new task group */
6297 struct task_group *sched_create_group(struct task_group *parent)
6299 struct task_group *tg;
6301 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6303 return ERR_PTR(-ENOMEM);
6305 if (!alloc_fair_sched_group(tg, parent))
6308 if (!alloc_rt_sched_group(tg, parent))
6314 sched_free_group(tg);
6315 return ERR_PTR(-ENOMEM);
6318 void sched_online_group(struct task_group *tg, struct task_group *parent)
6320 unsigned long flags;
6322 spin_lock_irqsave(&task_group_lock, flags);
6323 list_add_rcu(&tg->list, &task_groups);
6325 /* Root should already exist: */
6328 tg->parent = parent;
6329 INIT_LIST_HEAD(&tg->children);
6330 list_add_rcu(&tg->siblings, &parent->children);
6331 spin_unlock_irqrestore(&task_group_lock, flags);
6333 online_fair_sched_group(tg);
6336 /* rcu callback to free various structures associated with a task group */
6337 static void sched_free_group_rcu(struct rcu_head *rhp)
6339 /* Now it should be safe to free those cfs_rqs: */
6340 sched_free_group(container_of(rhp, struct task_group, rcu));
6343 void sched_destroy_group(struct task_group *tg)
6345 /* Wait for possible concurrent references to cfs_rqs complete: */
6346 call_rcu(&tg->rcu, sched_free_group_rcu);
6349 void sched_offline_group(struct task_group *tg)
6351 unsigned long flags;
6353 /* End participation in shares distribution: */
6354 unregister_fair_sched_group(tg);
6356 spin_lock_irqsave(&task_group_lock, flags);
6357 list_del_rcu(&tg->list);
6358 list_del_rcu(&tg->siblings);
6359 spin_unlock_irqrestore(&task_group_lock, flags);
6362 static void sched_change_group(struct task_struct *tsk, int type)
6364 struct task_group *tg;
6367 * All callers are synchronized by task_rq_lock(); we do not use RCU
6368 * which is pointless here. Thus, we pass "true" to task_css_check()
6369 * to prevent lockdep warnings.
6371 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6372 struct task_group, css);
6373 tg = autogroup_task_group(tsk, tg);
6374 tsk->sched_task_group = tg;
6376 #ifdef CONFIG_FAIR_GROUP_SCHED
6377 if (tsk->sched_class->task_change_group)
6378 tsk->sched_class->task_change_group(tsk, type);
6381 set_task_rq(tsk, task_cpu(tsk));
6385 * Change task's runqueue when it moves between groups.
6387 * The caller of this function should have put the task in its new group by
6388 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6391 void sched_move_task(struct task_struct *tsk)
6393 int queued, running;
6397 rq = task_rq_lock(tsk, &rf);
6398 update_rq_clock(rq);
6400 running = task_current(rq, tsk);
6401 queued = task_on_rq_queued(tsk);
6404 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
6406 put_prev_task(rq, tsk);
6408 sched_change_group(tsk, TASK_MOVE_GROUP);
6411 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
6413 set_curr_task(rq, tsk);
6415 task_rq_unlock(rq, tsk, &rf);
6417 #endif /* CONFIG_CGROUP_SCHED */
6419 #ifdef CONFIG_RT_GROUP_SCHED
6421 * Ensure that the real time constraints are schedulable.
6423 static DEFINE_MUTEX(rt_constraints_mutex);
6425 /* Must be called with tasklist_lock held */
6426 static inline int tg_has_rt_tasks(struct task_group *tg)
6428 struct task_struct *g, *p;
6431 * Autogroups do not have RT tasks; see autogroup_create().
6433 if (task_group_is_autogroup(tg))
6436 for_each_process_thread(g, p) {
6437 if (rt_task(p) && task_group(p) == tg)
6444 struct rt_schedulable_data {
6445 struct task_group *tg;
6450 static int tg_rt_schedulable(struct task_group *tg, void *data)
6452 struct rt_schedulable_data *d = data;
6453 struct task_group *child;
6454 unsigned long total, sum = 0;
6455 u64 period, runtime;
6457 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6458 runtime = tg->rt_bandwidth.rt_runtime;
6461 period = d->rt_period;
6462 runtime = d->rt_runtime;
6466 * Cannot have more runtime than the period.
6468 if (runtime > period && runtime != RUNTIME_INF)
6472 * Ensure we don't starve existing RT tasks.
6474 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
6477 total = to_ratio(period, runtime);
6480 * Nobody can have more than the global setting allows.
6482 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
6486 * The sum of our children's runtime should not exceed our own.
6488 list_for_each_entry_rcu(child, &tg->children, siblings) {
6489 period = ktime_to_ns(child->rt_bandwidth.rt_period);
6490 runtime = child->rt_bandwidth.rt_runtime;
6492 if (child == d->tg) {
6493 period = d->rt_period;
6494 runtime = d->rt_runtime;
6497 sum += to_ratio(period, runtime);
6506 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
6510 struct rt_schedulable_data data = {
6512 .rt_period = period,
6513 .rt_runtime = runtime,
6517 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
6523 static int tg_set_rt_bandwidth(struct task_group *tg,
6524 u64 rt_period, u64 rt_runtime)
6529 * Disallowing the root group RT runtime is BAD, it would disallow the
6530 * kernel creating (and or operating) RT threads.
6532 if (tg == &root_task_group && rt_runtime == 0)
6535 /* No period doesn't make any sense. */
6539 mutex_lock(&rt_constraints_mutex);
6540 read_lock(&tasklist_lock);
6541 err = __rt_schedulable(tg, rt_period, rt_runtime);
6545 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6546 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
6547 tg->rt_bandwidth.rt_runtime = rt_runtime;
6549 for_each_possible_cpu(i) {
6550 struct rt_rq *rt_rq = tg->rt_rq[i];
6552 raw_spin_lock(&rt_rq->rt_runtime_lock);
6553 rt_rq->rt_runtime = rt_runtime;
6554 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6556 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6558 read_unlock(&tasklist_lock);
6559 mutex_unlock(&rt_constraints_mutex);
6564 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
6566 u64 rt_runtime, rt_period;
6568 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6569 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
6570 if (rt_runtime_us < 0)
6571 rt_runtime = RUNTIME_INF;
6573 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6576 static long sched_group_rt_runtime(struct task_group *tg)
6580 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
6583 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
6584 do_div(rt_runtime_us, NSEC_PER_USEC);
6585 return rt_runtime_us;
6588 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
6590 u64 rt_runtime, rt_period;
6592 rt_period = rt_period_us * NSEC_PER_USEC;
6593 rt_runtime = tg->rt_bandwidth.rt_runtime;
6595 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6598 static long sched_group_rt_period(struct task_group *tg)
6602 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
6603 do_div(rt_period_us, NSEC_PER_USEC);
6604 return rt_period_us;
6606 #endif /* CONFIG_RT_GROUP_SCHED */
6608 #ifdef CONFIG_RT_GROUP_SCHED
6609 static int sched_rt_global_constraints(void)
6613 mutex_lock(&rt_constraints_mutex);
6614 read_lock(&tasklist_lock);
6615 ret = __rt_schedulable(NULL, 0, 0);
6616 read_unlock(&tasklist_lock);
6617 mutex_unlock(&rt_constraints_mutex);
6622 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
6624 /* Don't accept realtime tasks when there is no way for them to run */
6625 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
6631 #else /* !CONFIG_RT_GROUP_SCHED */
6632 static int sched_rt_global_constraints(void)
6634 unsigned long flags;
6637 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
6638 for_each_possible_cpu(i) {
6639 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
6641 raw_spin_lock(&rt_rq->rt_runtime_lock);
6642 rt_rq->rt_runtime = global_rt_runtime();
6643 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6645 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
6649 #endif /* CONFIG_RT_GROUP_SCHED */
6651 static int sched_dl_global_validate(void)
6653 u64 runtime = global_rt_runtime();
6654 u64 period = global_rt_period();
6655 u64 new_bw = to_ratio(period, runtime);
6658 unsigned long flags;
6661 * Here we want to check the bandwidth not being set to some
6662 * value smaller than the currently allocated bandwidth in
6663 * any of the root_domains.
6665 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
6666 * cycling on root_domains... Discussion on different/better
6667 * solutions is welcome!
6669 for_each_possible_cpu(cpu) {
6670 rcu_read_lock_sched();
6671 dl_b = dl_bw_of(cpu);
6673 raw_spin_lock_irqsave(&dl_b->lock, flags);
6674 if (new_bw < dl_b->total_bw)
6676 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
6678 rcu_read_unlock_sched();
6687 static void sched_dl_do_global(void)
6692 unsigned long flags;
6694 def_dl_bandwidth.dl_period = global_rt_period();
6695 def_dl_bandwidth.dl_runtime = global_rt_runtime();
6697 if (global_rt_runtime() != RUNTIME_INF)
6698 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
6701 * FIXME: As above...
6703 for_each_possible_cpu(cpu) {
6704 rcu_read_lock_sched();
6705 dl_b = dl_bw_of(cpu);
6707 raw_spin_lock_irqsave(&dl_b->lock, flags);
6709 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
6711 rcu_read_unlock_sched();
6715 static int sched_rt_global_validate(void)
6717 if (sysctl_sched_rt_period <= 0)
6720 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
6721 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
6727 static void sched_rt_do_global(void)
6729 def_rt_bandwidth.rt_runtime = global_rt_runtime();
6730 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
6733 int sched_rt_handler(struct ctl_table *table, int write,
6734 void __user *buffer, size_t *lenp,
6737 int old_period, old_runtime;
6738 static DEFINE_MUTEX(mutex);
6742 old_period = sysctl_sched_rt_period;
6743 old_runtime = sysctl_sched_rt_runtime;
6745 ret = proc_dointvec(table, write, buffer, lenp, ppos);
6747 if (!ret && write) {
6748 ret = sched_rt_global_validate();
6752 ret = sched_dl_global_validate();
6756 ret = sched_rt_global_constraints();
6760 sched_rt_do_global();
6761 sched_dl_do_global();
6765 sysctl_sched_rt_period = old_period;
6766 sysctl_sched_rt_runtime = old_runtime;
6768 mutex_unlock(&mutex);
6773 int sched_rr_handler(struct ctl_table *table, int write,
6774 void __user *buffer, size_t *lenp,
6778 static DEFINE_MUTEX(mutex);
6781 ret = proc_dointvec(table, write, buffer, lenp, ppos);
6783 * Make sure that internally we keep jiffies.
6784 * Also, writing zero resets the timeslice to default:
6786 if (!ret && write) {
6787 sched_rr_timeslice =
6788 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
6789 msecs_to_jiffies(sysctl_sched_rr_timeslice);
6791 mutex_unlock(&mutex);
6795 #ifdef CONFIG_CGROUP_SCHED
6797 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6799 return css ? container_of(css, struct task_group, css) : NULL;
6802 static struct cgroup_subsys_state *
6803 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6805 struct task_group *parent = css_tg(parent_css);
6806 struct task_group *tg;
6809 /* This is early initialization for the top cgroup */
6810 return &root_task_group.css;
6813 tg = sched_create_group(parent);
6815 return ERR_PTR(-ENOMEM);
6820 /* Expose task group only after completing cgroup initialization */
6821 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6823 struct task_group *tg = css_tg(css);
6824 struct task_group *parent = css_tg(css->parent);
6827 sched_online_group(tg, parent);
6831 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6833 struct task_group *tg = css_tg(css);
6835 sched_offline_group(tg);
6838 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6840 struct task_group *tg = css_tg(css);
6843 * Relies on the RCU grace period between css_released() and this.
6845 sched_free_group(tg);
6849 * This is called before wake_up_new_task(), therefore we really only
6850 * have to set its group bits, all the other stuff does not apply.
6852 static void cpu_cgroup_fork(struct task_struct *task)
6857 rq = task_rq_lock(task, &rf);
6859 update_rq_clock(rq);
6860 sched_change_group(task, TASK_SET_GROUP);
6862 task_rq_unlock(rq, task, &rf);
6865 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6867 struct task_struct *task;
6868 struct cgroup_subsys_state *css;
6871 cgroup_taskset_for_each(task, css, tset) {
6872 #ifdef CONFIG_RT_GROUP_SCHED
6873 if (!sched_rt_can_attach(css_tg(css), task))
6876 /* We don't support RT-tasks being in separate groups */
6877 if (task->sched_class != &fair_sched_class)
6881 * Serialize against wake_up_new_task() such that if its
6882 * running, we're sure to observe its full state.
6884 raw_spin_lock_irq(&task->pi_lock);
6886 * Avoid calling sched_move_task() before wake_up_new_task()
6887 * has happened. This would lead to problems with PELT, due to
6888 * move wanting to detach+attach while we're not attached yet.
6890 if (task->state == TASK_NEW)
6892 raw_spin_unlock_irq(&task->pi_lock);
6900 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6902 struct task_struct *task;
6903 struct cgroup_subsys_state *css;
6905 cgroup_taskset_for_each(task, css, tset)
6906 sched_move_task(task);
6909 #ifdef CONFIG_FAIR_GROUP_SCHED
6910 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6911 struct cftype *cftype, u64 shareval)
6913 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6916 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6919 struct task_group *tg = css_tg(css);
6921 return (u64) scale_load_down(tg->shares);
6924 #ifdef CONFIG_CFS_BANDWIDTH
6925 static DEFINE_MUTEX(cfs_constraints_mutex);
6927 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6928 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6930 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6932 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6934 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6935 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6937 if (tg == &root_task_group)
6941 * Ensure we have at some amount of bandwidth every period. This is
6942 * to prevent reaching a state of large arrears when throttled via
6943 * entity_tick() resulting in prolonged exit starvation.
6945 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6949 * Likewise, bound things on the otherside by preventing insane quota
6950 * periods. This also allows us to normalize in computing quota
6953 if (period > max_cfs_quota_period)
6957 * Prevent race between setting of cfs_rq->runtime_enabled and
6958 * unthrottle_offline_cfs_rqs().
6961 mutex_lock(&cfs_constraints_mutex);
6962 ret = __cfs_schedulable(tg, period, quota);
6966 runtime_enabled = quota != RUNTIME_INF;
6967 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6969 * If we need to toggle cfs_bandwidth_used, off->on must occur
6970 * before making related changes, and on->off must occur afterwards
6972 if (runtime_enabled && !runtime_was_enabled)
6973 cfs_bandwidth_usage_inc();
6974 raw_spin_lock_irq(&cfs_b->lock);
6975 cfs_b->period = ns_to_ktime(period);
6976 cfs_b->quota = quota;
6978 __refill_cfs_bandwidth_runtime(cfs_b);
6980 /* Restart the period timer (if active) to handle new period expiry: */
6981 if (runtime_enabled)
6982 start_cfs_bandwidth(cfs_b);
6984 raw_spin_unlock_irq(&cfs_b->lock);
6986 for_each_online_cpu(i) {
6987 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6988 struct rq *rq = cfs_rq->rq;
6990 raw_spin_lock_irq(&rq->lock);
6991 cfs_rq->runtime_enabled = runtime_enabled;
6992 cfs_rq->runtime_remaining = 0;
6994 if (cfs_rq->throttled)
6995 unthrottle_cfs_rq(cfs_rq);
6996 raw_spin_unlock_irq(&rq->lock);
6998 if (runtime_was_enabled && !runtime_enabled)
6999 cfs_bandwidth_usage_dec();
7001 mutex_unlock(&cfs_constraints_mutex);
7007 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7011 period = ktime_to_ns(tg->cfs_bandwidth.period);
7012 if (cfs_quota_us < 0)
7013 quota = RUNTIME_INF;
7015 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7017 return tg_set_cfs_bandwidth(tg, period, quota);
7020 long tg_get_cfs_quota(struct task_group *tg)
7024 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7027 quota_us = tg->cfs_bandwidth.quota;
7028 do_div(quota_us, NSEC_PER_USEC);
7033 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7037 period = (u64)cfs_period_us * NSEC_PER_USEC;
7038 quota = tg->cfs_bandwidth.quota;
7040 return tg_set_cfs_bandwidth(tg, period, quota);
7043 long tg_get_cfs_period(struct task_group *tg)
7047 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7048 do_div(cfs_period_us, NSEC_PER_USEC);
7050 return cfs_period_us;
7053 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7056 return tg_get_cfs_quota(css_tg(css));
7059 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7060 struct cftype *cftype, s64 cfs_quota_us)
7062 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7065 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7068 return tg_get_cfs_period(css_tg(css));
7071 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7072 struct cftype *cftype, u64 cfs_period_us)
7074 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7077 struct cfs_schedulable_data {
7078 struct task_group *tg;
7083 * normalize group quota/period to be quota/max_period
7084 * note: units are usecs
7086 static u64 normalize_cfs_quota(struct task_group *tg,
7087 struct cfs_schedulable_data *d)
7095 period = tg_get_cfs_period(tg);
7096 quota = tg_get_cfs_quota(tg);
7099 /* note: these should typically be equivalent */
7100 if (quota == RUNTIME_INF || quota == -1)
7103 return to_ratio(period, quota);
7106 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7108 struct cfs_schedulable_data *d = data;
7109 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7110 s64 quota = 0, parent_quota = -1;
7113 quota = RUNTIME_INF;
7115 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7117 quota = normalize_cfs_quota(tg, d);
7118 parent_quota = parent_b->hierarchical_quota;
7121 * Ensure max(child_quota) <= parent_quota, inherit when no
7124 if (quota == RUNTIME_INF)
7125 quota = parent_quota;
7126 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7129 cfs_b->hierarchical_quota = quota;
7134 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7137 struct cfs_schedulable_data data = {
7143 if (quota != RUNTIME_INF) {
7144 do_div(data.period, NSEC_PER_USEC);
7145 do_div(data.quota, NSEC_PER_USEC);
7149 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7155 static int cpu_stats_show(struct seq_file *sf, void *v)
7157 struct task_group *tg = css_tg(seq_css(sf));
7158 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7160 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7161 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7162 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7166 #endif /* CONFIG_CFS_BANDWIDTH */
7167 #endif /* CONFIG_FAIR_GROUP_SCHED */
7169 #ifdef CONFIG_RT_GROUP_SCHED
7170 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7171 struct cftype *cft, s64 val)
7173 return sched_group_set_rt_runtime(css_tg(css), val);
7176 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7179 return sched_group_rt_runtime(css_tg(css));
7182 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7183 struct cftype *cftype, u64 rt_period_us)
7185 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7188 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7191 return sched_group_rt_period(css_tg(css));
7193 #endif /* CONFIG_RT_GROUP_SCHED */
7195 static struct cftype cpu_files[] = {
7196 #ifdef CONFIG_FAIR_GROUP_SCHED
7199 .read_u64 = cpu_shares_read_u64,
7200 .write_u64 = cpu_shares_write_u64,
7203 #ifdef CONFIG_CFS_BANDWIDTH
7205 .name = "cfs_quota_us",
7206 .read_s64 = cpu_cfs_quota_read_s64,
7207 .write_s64 = cpu_cfs_quota_write_s64,
7210 .name = "cfs_period_us",
7211 .read_u64 = cpu_cfs_period_read_u64,
7212 .write_u64 = cpu_cfs_period_write_u64,
7216 .seq_show = cpu_stats_show,
7219 #ifdef CONFIG_RT_GROUP_SCHED
7221 .name = "rt_runtime_us",
7222 .read_s64 = cpu_rt_runtime_read,
7223 .write_s64 = cpu_rt_runtime_write,
7226 .name = "rt_period_us",
7227 .read_u64 = cpu_rt_period_read_uint,
7228 .write_u64 = cpu_rt_period_write_uint,
7234 struct cgroup_subsys cpu_cgrp_subsys = {
7235 .css_alloc = cpu_cgroup_css_alloc,
7236 .css_online = cpu_cgroup_css_online,
7237 .css_released = cpu_cgroup_css_released,
7238 .css_free = cpu_cgroup_css_free,
7239 .fork = cpu_cgroup_fork,
7240 .can_attach = cpu_cgroup_can_attach,
7241 .attach = cpu_cgroup_attach,
7242 .legacy_cftypes = cpu_files,
7246 #endif /* CONFIG_CGROUP_SCHED */
7248 void dump_cpu_task(int cpu)
7250 pr_info("Task dump for CPU %d:\n", cpu);
7251 sched_show_task(cpu_curr(cpu));
7255 * Nice levels are multiplicative, with a gentle 10% change for every
7256 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7257 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7258 * that remained on nice 0.
7260 * The "10% effect" is relative and cumulative: from _any_ nice level,
7261 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7262 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7263 * If a task goes up by ~10% and another task goes down by ~10% then
7264 * the relative distance between them is ~25%.)
7266 const int sched_prio_to_weight[40] = {
7267 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7268 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7269 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7270 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7271 /* 0 */ 1024, 820, 655, 526, 423,
7272 /* 5 */ 335, 272, 215, 172, 137,
7273 /* 10 */ 110, 87, 70, 56, 45,
7274 /* 15 */ 36, 29, 23, 18, 15,
7278 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7280 * In cases where the weight does not change often, we can use the
7281 * precalculated inverse to speed up arithmetics by turning divisions
7282 * into multiplications:
7284 const u32 sched_prio_to_wmult[40] = {
7285 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7286 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7287 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7288 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7289 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7290 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7291 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7292 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,