4 * Core kernel scheduler code and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 #include <linux/sched.h>
9 #include <linux/sched/clock.h>
10 #include <uapi/linux/sched/types.h>
11 #include <linux/sched/loadavg.h>
12 #include <linux/cpuset.h>
13 #include <linux/delayacct.h>
14 #include <linux/init_task.h>
15 #include <linux/context_tracking.h>
16 #include <linux/rcupdate_wait.h>
18 #include <linux/blkdev.h>
19 #include <linux/kprobes.h>
20 #include <linux/mmu_context.h>
21 #include <linux/module.h>
22 #include <linux/nmi.h>
23 #include <linux/prefetch.h>
24 #include <linux/profile.h>
25 #include <linux/security.h>
26 #include <linux/syscalls.h>
28 #include <asm/switch_to.h>
30 #ifdef CONFIG_PARAVIRT
31 #include <asm/paravirt.h>
35 #include "../workqueue_internal.h"
36 #include "../smpboot.h"
38 #define CREATE_TRACE_POINTS
39 #include <trace/events/sched.h>
41 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
44 * Debugging: various feature bits
47 #define SCHED_FEAT(name, enabled) \
48 (1UL << __SCHED_FEAT_##name) * enabled |
50 const_debug unsigned int sysctl_sched_features =
57 * Number of tasks to iterate in a single balance run.
58 * Limited because this is done with IRQs disabled.
60 const_debug unsigned int sysctl_sched_nr_migrate = 32;
63 * period over which we average the RT time consumption, measured
68 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
71 * period over which we measure -rt task CPU usage in us.
74 unsigned int sysctl_sched_rt_period = 1000000;
76 __read_mostly int scheduler_running;
79 * part of the period that we allow rt tasks to run in us.
82 int sysctl_sched_rt_runtime = 950000;
84 /* CPUs with isolated domains */
85 cpumask_var_t cpu_isolated_map;
88 * this_rq_lock - lock this runqueue and disable interrupts.
90 static struct rq *this_rq_lock(void)
97 raw_spin_lock(&rq->lock);
103 * __task_rq_lock - lock the rq @p resides on.
105 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
110 lockdep_assert_held(&p->pi_lock);
114 raw_spin_lock(&rq->lock);
115 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
119 raw_spin_unlock(&rq->lock);
121 while (unlikely(task_on_rq_migrating(p)))
127 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
129 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
130 __acquires(p->pi_lock)
136 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
138 raw_spin_lock(&rq->lock);
140 * move_queued_task() task_rq_lock()
143 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
144 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
145 * [S] ->cpu = new_cpu [L] task_rq()
149 * If we observe the old cpu in task_rq_lock, the acquire of
150 * the old rq->lock will fully serialize against the stores.
152 * If we observe the new CPU in task_rq_lock, the acquire will
153 * pair with the WMB to ensure we must then also see migrating.
155 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
159 raw_spin_unlock(&rq->lock);
160 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
162 while (unlikely(task_on_rq_migrating(p)))
168 * RQ-clock updating methods:
171 static void update_rq_clock_task(struct rq *rq, s64 delta)
174 * In theory, the compile should just see 0 here, and optimize out the call
175 * to sched_rt_avg_update. But I don't trust it...
177 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
178 s64 steal = 0, irq_delta = 0;
180 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
181 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
184 * Since irq_time is only updated on {soft,}irq_exit, we might run into
185 * this case when a previous update_rq_clock() happened inside a
188 * When this happens, we stop ->clock_task and only update the
189 * prev_irq_time stamp to account for the part that fit, so that a next
190 * update will consume the rest. This ensures ->clock_task is
193 * It does however cause some slight miss-attribution of {soft,}irq
194 * time, a more accurate solution would be to update the irq_time using
195 * the current rq->clock timestamp, except that would require using
198 if (irq_delta > delta)
201 rq->prev_irq_time += irq_delta;
204 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
205 if (static_key_false((¶virt_steal_rq_enabled))) {
206 steal = paravirt_steal_clock(cpu_of(rq));
207 steal -= rq->prev_steal_time_rq;
209 if (unlikely(steal > delta))
212 rq->prev_steal_time_rq += steal;
217 rq->clock_task += delta;
219 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
220 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
221 sched_rt_avg_update(rq, irq_delta + steal);
225 void update_rq_clock(struct rq *rq)
229 lockdep_assert_held(&rq->lock);
231 if (rq->clock_update_flags & RQCF_ACT_SKIP)
234 #ifdef CONFIG_SCHED_DEBUG
235 rq->clock_update_flags |= RQCF_UPDATED;
237 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
241 update_rq_clock_task(rq, delta);
245 #ifdef CONFIG_SCHED_HRTICK
247 * Use HR-timers to deliver accurate preemption points.
250 static void hrtick_clear(struct rq *rq)
252 if (hrtimer_active(&rq->hrtick_timer))
253 hrtimer_cancel(&rq->hrtick_timer);
257 * High-resolution timer tick.
258 * Runs from hardirq context with interrupts disabled.
260 static enum hrtimer_restart hrtick(struct hrtimer *timer)
262 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
264 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
266 raw_spin_lock(&rq->lock);
268 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
269 raw_spin_unlock(&rq->lock);
271 return HRTIMER_NORESTART;
276 static void __hrtick_restart(struct rq *rq)
278 struct hrtimer *timer = &rq->hrtick_timer;
280 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
284 * called from hardirq (IPI) context
286 static void __hrtick_start(void *arg)
290 raw_spin_lock(&rq->lock);
291 __hrtick_restart(rq);
292 rq->hrtick_csd_pending = 0;
293 raw_spin_unlock(&rq->lock);
297 * Called to set the hrtick timer state.
299 * called with rq->lock held and irqs disabled
301 void hrtick_start(struct rq *rq, u64 delay)
303 struct hrtimer *timer = &rq->hrtick_timer;
308 * Don't schedule slices shorter than 10000ns, that just
309 * doesn't make sense and can cause timer DoS.
311 delta = max_t(s64, delay, 10000LL);
312 time = ktime_add_ns(timer->base->get_time(), delta);
314 hrtimer_set_expires(timer, time);
316 if (rq == this_rq()) {
317 __hrtick_restart(rq);
318 } else if (!rq->hrtick_csd_pending) {
319 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
320 rq->hrtick_csd_pending = 1;
326 * Called to set the hrtick timer state.
328 * called with rq->lock held and irqs disabled
330 void hrtick_start(struct rq *rq, u64 delay)
333 * Don't schedule slices shorter than 10000ns, that just
334 * doesn't make sense. Rely on vruntime for fairness.
336 delay = max_t(u64, delay, 10000LL);
337 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
338 HRTIMER_MODE_REL_PINNED);
340 #endif /* CONFIG_SMP */
342 static void init_rq_hrtick(struct rq *rq)
345 rq->hrtick_csd_pending = 0;
347 rq->hrtick_csd.flags = 0;
348 rq->hrtick_csd.func = __hrtick_start;
349 rq->hrtick_csd.info = rq;
352 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
353 rq->hrtick_timer.function = hrtick;
355 #else /* CONFIG_SCHED_HRTICK */
356 static inline void hrtick_clear(struct rq *rq)
360 static inline void init_rq_hrtick(struct rq *rq)
363 #endif /* CONFIG_SCHED_HRTICK */
366 * cmpxchg based fetch_or, macro so it works for different integer types
368 #define fetch_or(ptr, mask) \
370 typeof(ptr) _ptr = (ptr); \
371 typeof(mask) _mask = (mask); \
372 typeof(*_ptr) _old, _val = *_ptr; \
375 _old = cmpxchg(_ptr, _val, _val | _mask); \
383 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
385 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
386 * this avoids any races wrt polling state changes and thereby avoids
389 static bool set_nr_and_not_polling(struct task_struct *p)
391 struct thread_info *ti = task_thread_info(p);
392 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
396 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
398 * If this returns true, then the idle task promises to call
399 * sched_ttwu_pending() and reschedule soon.
401 static bool set_nr_if_polling(struct task_struct *p)
403 struct thread_info *ti = task_thread_info(p);
404 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
407 if (!(val & _TIF_POLLING_NRFLAG))
409 if (val & _TIF_NEED_RESCHED)
411 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
420 static bool set_nr_and_not_polling(struct task_struct *p)
422 set_tsk_need_resched(p);
427 static bool set_nr_if_polling(struct task_struct *p)
434 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
436 struct wake_q_node *node = &task->wake_q;
439 * Atomically grab the task, if ->wake_q is !nil already it means
440 * its already queued (either by us or someone else) and will get the
441 * wakeup due to that.
443 * This cmpxchg() implies a full barrier, which pairs with the write
444 * barrier implied by the wakeup in wake_up_q().
446 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
449 get_task_struct(task);
452 * The head is context local, there can be no concurrency.
455 head->lastp = &node->next;
458 void wake_up_q(struct wake_q_head *head)
460 struct wake_q_node *node = head->first;
462 while (node != WAKE_Q_TAIL) {
463 struct task_struct *task;
465 task = container_of(node, struct task_struct, wake_q);
467 /* Task can safely be re-inserted now: */
469 task->wake_q.next = NULL;
472 * wake_up_process() implies a wmb() to pair with the queueing
473 * in wake_q_add() so as not to miss wakeups.
475 wake_up_process(task);
476 put_task_struct(task);
481 * resched_curr - mark rq's current task 'to be rescheduled now'.
483 * On UP this means the setting of the need_resched flag, on SMP it
484 * might also involve a cross-CPU call to trigger the scheduler on
487 void resched_curr(struct rq *rq)
489 struct task_struct *curr = rq->curr;
492 lockdep_assert_held(&rq->lock);
494 if (test_tsk_need_resched(curr))
499 if (cpu == smp_processor_id()) {
500 set_tsk_need_resched(curr);
501 set_preempt_need_resched();
505 if (set_nr_and_not_polling(curr))
506 smp_send_reschedule(cpu);
508 trace_sched_wake_idle_without_ipi(cpu);
511 void resched_cpu(int cpu)
513 struct rq *rq = cpu_rq(cpu);
516 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
519 raw_spin_unlock_irqrestore(&rq->lock, flags);
523 #ifdef CONFIG_NO_HZ_COMMON
525 * In the semi idle case, use the nearest busy CPU for migrating timers
526 * from an idle CPU. This is good for power-savings.
528 * We don't do similar optimization for completely idle system, as
529 * selecting an idle CPU will add more delays to the timers than intended
530 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
532 int get_nohz_timer_target(void)
534 int i, cpu = smp_processor_id();
535 struct sched_domain *sd;
537 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
541 for_each_domain(cpu, sd) {
542 for_each_cpu(i, sched_domain_span(sd)) {
546 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
553 if (!is_housekeeping_cpu(cpu))
554 cpu = housekeeping_any_cpu();
561 * When add_timer_on() enqueues a timer into the timer wheel of an
562 * idle CPU then this timer might expire before the next timer event
563 * which is scheduled to wake up that CPU. In case of a completely
564 * idle system the next event might even be infinite time into the
565 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
566 * leaves the inner idle loop so the newly added timer is taken into
567 * account when the CPU goes back to idle and evaluates the timer
568 * wheel for the next timer event.
570 static void wake_up_idle_cpu(int cpu)
572 struct rq *rq = cpu_rq(cpu);
574 if (cpu == smp_processor_id())
577 if (set_nr_and_not_polling(rq->idle))
578 smp_send_reschedule(cpu);
580 trace_sched_wake_idle_without_ipi(cpu);
583 static bool wake_up_full_nohz_cpu(int cpu)
586 * We just need the target to call irq_exit() and re-evaluate
587 * the next tick. The nohz full kick at least implies that.
588 * If needed we can still optimize that later with an
591 if (cpu_is_offline(cpu))
592 return true; /* Don't try to wake offline CPUs. */
593 if (tick_nohz_full_cpu(cpu)) {
594 if (cpu != smp_processor_id() ||
595 tick_nohz_tick_stopped())
596 tick_nohz_full_kick_cpu(cpu);
604 * Wake up the specified CPU. If the CPU is going offline, it is the
605 * caller's responsibility to deal with the lost wakeup, for example,
606 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
608 void wake_up_nohz_cpu(int cpu)
610 if (!wake_up_full_nohz_cpu(cpu))
611 wake_up_idle_cpu(cpu);
614 static inline bool got_nohz_idle_kick(void)
616 int cpu = smp_processor_id();
618 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
621 if (idle_cpu(cpu) && !need_resched())
625 * We can't run Idle Load Balance on this CPU for this time so we
626 * cancel it and clear NOHZ_BALANCE_KICK
628 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
632 #else /* CONFIG_NO_HZ_COMMON */
634 static inline bool got_nohz_idle_kick(void)
639 #endif /* CONFIG_NO_HZ_COMMON */
641 #ifdef CONFIG_NO_HZ_FULL
642 bool sched_can_stop_tick(struct rq *rq)
646 /* Deadline tasks, even if single, need the tick */
647 if (rq->dl.dl_nr_running)
651 * If there are more than one RR tasks, we need the tick to effect the
652 * actual RR behaviour.
654 if (rq->rt.rr_nr_running) {
655 if (rq->rt.rr_nr_running == 1)
662 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
663 * forced preemption between FIFO tasks.
665 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
670 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
671 * if there's more than one we need the tick for involuntary
674 if (rq->nr_running > 1)
679 #endif /* CONFIG_NO_HZ_FULL */
681 void sched_avg_update(struct rq *rq)
683 s64 period = sched_avg_period();
685 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
687 * Inline assembly required to prevent the compiler
688 * optimising this loop into a divmod call.
689 * See __iter_div_u64_rem() for another example of this.
691 asm("" : "+rm" (rq->age_stamp));
692 rq->age_stamp += period;
697 #endif /* CONFIG_SMP */
699 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
700 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
702 * Iterate task_group tree rooted at *from, calling @down when first entering a
703 * node and @up when leaving it for the final time.
705 * Caller must hold rcu_lock or sufficient equivalent.
707 int walk_tg_tree_from(struct task_group *from,
708 tg_visitor down, tg_visitor up, void *data)
710 struct task_group *parent, *child;
716 ret = (*down)(parent, data);
719 list_for_each_entry_rcu(child, &parent->children, siblings) {
726 ret = (*up)(parent, data);
727 if (ret || parent == from)
731 parent = parent->parent;
738 int tg_nop(struct task_group *tg, void *data)
744 static void set_load_weight(struct task_struct *p)
746 int prio = p->static_prio - MAX_RT_PRIO;
747 struct load_weight *load = &p->se.load;
750 * SCHED_IDLE tasks get minimal weight:
752 if (idle_policy(p->policy)) {
753 load->weight = scale_load(WEIGHT_IDLEPRIO);
754 load->inv_weight = WMULT_IDLEPRIO;
758 load->weight = scale_load(sched_prio_to_weight[prio]);
759 load->inv_weight = sched_prio_to_wmult[prio];
762 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
765 if (!(flags & ENQUEUE_RESTORE))
766 sched_info_queued(rq, p);
767 p->sched_class->enqueue_task(rq, p, flags);
770 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
773 if (!(flags & DEQUEUE_SAVE))
774 sched_info_dequeued(rq, p);
775 p->sched_class->dequeue_task(rq, p, flags);
778 void activate_task(struct rq *rq, struct task_struct *p, int flags)
780 if (task_contributes_to_load(p))
781 rq->nr_uninterruptible--;
783 enqueue_task(rq, p, flags);
786 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
788 if (task_contributes_to_load(p))
789 rq->nr_uninterruptible++;
791 dequeue_task(rq, p, flags);
794 void sched_set_stop_task(int cpu, struct task_struct *stop)
796 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
797 struct task_struct *old_stop = cpu_rq(cpu)->stop;
801 * Make it appear like a SCHED_FIFO task, its something
802 * userspace knows about and won't get confused about.
804 * Also, it will make PI more or less work without too
805 * much confusion -- but then, stop work should not
806 * rely on PI working anyway.
808 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
810 stop->sched_class = &stop_sched_class;
813 cpu_rq(cpu)->stop = stop;
817 * Reset it back to a normal scheduling class so that
818 * it can die in pieces.
820 old_stop->sched_class = &rt_sched_class;
825 * __normal_prio - return the priority that is based on the static prio
827 static inline int __normal_prio(struct task_struct *p)
829 return p->static_prio;
833 * Calculate the expected normal priority: i.e. priority
834 * without taking RT-inheritance into account. Might be
835 * boosted by interactivity modifiers. Changes upon fork,
836 * setprio syscalls, and whenever the interactivity
837 * estimator recalculates.
839 static inline int normal_prio(struct task_struct *p)
843 if (task_has_dl_policy(p))
844 prio = MAX_DL_PRIO-1;
845 else if (task_has_rt_policy(p))
846 prio = MAX_RT_PRIO-1 - p->rt_priority;
848 prio = __normal_prio(p);
853 * Calculate the current priority, i.e. the priority
854 * taken into account by the scheduler. This value might
855 * be boosted by RT tasks, or might be boosted by
856 * interactivity modifiers. Will be RT if the task got
857 * RT-boosted. If not then it returns p->normal_prio.
859 static int effective_prio(struct task_struct *p)
861 p->normal_prio = normal_prio(p);
863 * If we are RT tasks or we were boosted to RT priority,
864 * keep the priority unchanged. Otherwise, update priority
865 * to the normal priority:
867 if (!rt_prio(p->prio))
868 return p->normal_prio;
873 * task_curr - is this task currently executing on a CPU?
874 * @p: the task in question.
876 * Return: 1 if the task is currently executing. 0 otherwise.
878 inline int task_curr(const struct task_struct *p)
880 return cpu_curr(task_cpu(p)) == p;
884 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
885 * use the balance_callback list if you want balancing.
887 * this means any call to check_class_changed() must be followed by a call to
888 * balance_callback().
890 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
891 const struct sched_class *prev_class,
894 if (prev_class != p->sched_class) {
895 if (prev_class->switched_from)
896 prev_class->switched_from(rq, p);
898 p->sched_class->switched_to(rq, p);
899 } else if (oldprio != p->prio || dl_task(p))
900 p->sched_class->prio_changed(rq, p, oldprio);
903 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
905 const struct sched_class *class;
907 if (p->sched_class == rq->curr->sched_class) {
908 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
910 for_each_class(class) {
911 if (class == rq->curr->sched_class)
913 if (class == p->sched_class) {
921 * A queue event has occurred, and we're going to schedule. In
922 * this case, we can save a useless back to back clock update.
924 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
925 rq_clock_skip_update(rq, true);
930 * This is how migration works:
932 * 1) we invoke migration_cpu_stop() on the target CPU using
934 * 2) stopper starts to run (implicitly forcing the migrated thread
936 * 3) it checks whether the migrated task is still in the wrong runqueue.
937 * 4) if it's in the wrong runqueue then the migration thread removes
938 * it and puts it into the right queue.
939 * 5) stopper completes and stop_one_cpu() returns and the migration
944 * move_queued_task - move a queued task to new rq.
946 * Returns (locked) new rq. Old rq's lock is released.
948 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
950 lockdep_assert_held(&rq->lock);
952 p->on_rq = TASK_ON_RQ_MIGRATING;
953 dequeue_task(rq, p, 0);
954 set_task_cpu(p, new_cpu);
955 raw_spin_unlock(&rq->lock);
957 rq = cpu_rq(new_cpu);
959 raw_spin_lock(&rq->lock);
960 BUG_ON(task_cpu(p) != new_cpu);
961 enqueue_task(rq, p, 0);
962 p->on_rq = TASK_ON_RQ_QUEUED;
963 check_preempt_curr(rq, p, 0);
968 struct migration_arg {
969 struct task_struct *task;
974 * Move (not current) task off this CPU, onto the destination CPU. We're doing
975 * this because either it can't run here any more (set_cpus_allowed()
976 * away from this CPU, or CPU going down), or because we're
977 * attempting to rebalance this task on exec (sched_exec).
979 * So we race with normal scheduler movements, but that's OK, as long
980 * as the task is no longer on this CPU.
982 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
984 if (unlikely(!cpu_active(dest_cpu)))
987 /* Affinity changed (again). */
988 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
991 rq = move_queued_task(rq, p, dest_cpu);
997 * migration_cpu_stop - this will be executed by a highprio stopper thread
998 * and performs thread migration by bumping thread off CPU then
999 * 'pushing' onto another runqueue.
1001 static int migration_cpu_stop(void *data)
1003 struct migration_arg *arg = data;
1004 struct task_struct *p = arg->task;
1005 struct rq *rq = this_rq();
1008 * The original target CPU might have gone down and we might
1009 * be on another CPU but it doesn't matter.
1011 local_irq_disable();
1013 * We need to explicitly wake pending tasks before running
1014 * __migrate_task() such that we will not miss enforcing cpus_allowed
1015 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1017 sched_ttwu_pending();
1019 raw_spin_lock(&p->pi_lock);
1020 raw_spin_lock(&rq->lock);
1022 * If task_rq(p) != rq, it cannot be migrated here, because we're
1023 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1024 * we're holding p->pi_lock.
1026 if (task_rq(p) == rq) {
1027 if (task_on_rq_queued(p))
1028 rq = __migrate_task(rq, p, arg->dest_cpu);
1030 p->wake_cpu = arg->dest_cpu;
1032 raw_spin_unlock(&rq->lock);
1033 raw_spin_unlock(&p->pi_lock);
1040 * sched_class::set_cpus_allowed must do the below, but is not required to
1041 * actually call this function.
1043 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1045 cpumask_copy(&p->cpus_allowed, new_mask);
1046 p->nr_cpus_allowed = cpumask_weight(new_mask);
1049 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1051 struct rq *rq = task_rq(p);
1052 bool queued, running;
1054 lockdep_assert_held(&p->pi_lock);
1056 queued = task_on_rq_queued(p);
1057 running = task_current(rq, p);
1061 * Because __kthread_bind() calls this on blocked tasks without
1064 lockdep_assert_held(&rq->lock);
1065 dequeue_task(rq, p, DEQUEUE_SAVE);
1068 put_prev_task(rq, p);
1070 p->sched_class->set_cpus_allowed(p, new_mask);
1073 enqueue_task(rq, p, ENQUEUE_RESTORE);
1075 set_curr_task(rq, p);
1079 * Change a given task's CPU affinity. Migrate the thread to a
1080 * proper CPU and schedule it away if the CPU it's executing on
1081 * is removed from the allowed bitmask.
1083 * NOTE: the caller must have a valid reference to the task, the
1084 * task must not exit() & deallocate itself prematurely. The
1085 * call is not atomic; no spinlocks may be held.
1087 static int __set_cpus_allowed_ptr(struct task_struct *p,
1088 const struct cpumask *new_mask, bool check)
1090 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1091 unsigned int dest_cpu;
1096 rq = task_rq_lock(p, &rf);
1097 update_rq_clock(rq);
1099 if (p->flags & PF_KTHREAD) {
1101 * Kernel threads are allowed on online && !active CPUs
1103 cpu_valid_mask = cpu_online_mask;
1107 * Must re-check here, to close a race against __kthread_bind(),
1108 * sched_setaffinity() is not guaranteed to observe the flag.
1110 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1115 if (cpumask_equal(&p->cpus_allowed, new_mask))
1118 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1123 do_set_cpus_allowed(p, new_mask);
1125 if (p->flags & PF_KTHREAD) {
1127 * For kernel threads that do indeed end up on online &&
1128 * !active we want to ensure they are strict per-CPU threads.
1130 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1131 !cpumask_intersects(new_mask, cpu_active_mask) &&
1132 p->nr_cpus_allowed != 1);
1135 /* Can the task run on the task's current CPU? If so, we're done */
1136 if (cpumask_test_cpu(task_cpu(p), new_mask))
1139 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1140 if (task_running(rq, p) || p->state == TASK_WAKING) {
1141 struct migration_arg arg = { p, dest_cpu };
1142 /* Need help from migration thread: drop lock and wait. */
1143 task_rq_unlock(rq, p, &rf);
1144 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1145 tlb_migrate_finish(p->mm);
1147 } else if (task_on_rq_queued(p)) {
1149 * OK, since we're going to drop the lock immediately
1150 * afterwards anyway.
1152 rq_unpin_lock(rq, &rf);
1153 rq = move_queued_task(rq, p, dest_cpu);
1154 rq_repin_lock(rq, &rf);
1157 task_rq_unlock(rq, p, &rf);
1162 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1164 return __set_cpus_allowed_ptr(p, new_mask, false);
1166 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1168 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1170 #ifdef CONFIG_SCHED_DEBUG
1172 * We should never call set_task_cpu() on a blocked task,
1173 * ttwu() will sort out the placement.
1175 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1179 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1180 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1181 * time relying on p->on_rq.
1183 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1184 p->sched_class == &fair_sched_class &&
1185 (p->on_rq && !task_on_rq_migrating(p)));
1187 #ifdef CONFIG_LOCKDEP
1189 * The caller should hold either p->pi_lock or rq->lock, when changing
1190 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1192 * sched_move_task() holds both and thus holding either pins the cgroup,
1195 * Furthermore, all task_rq users should acquire both locks, see
1198 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1199 lockdep_is_held(&task_rq(p)->lock)));
1203 trace_sched_migrate_task(p, new_cpu);
1205 if (task_cpu(p) != new_cpu) {
1206 if (p->sched_class->migrate_task_rq)
1207 p->sched_class->migrate_task_rq(p);
1208 p->se.nr_migrations++;
1209 perf_event_task_migrate(p);
1212 __set_task_cpu(p, new_cpu);
1215 static void __migrate_swap_task(struct task_struct *p, int cpu)
1217 if (task_on_rq_queued(p)) {
1218 struct rq *src_rq, *dst_rq;
1220 src_rq = task_rq(p);
1221 dst_rq = cpu_rq(cpu);
1223 p->on_rq = TASK_ON_RQ_MIGRATING;
1224 deactivate_task(src_rq, p, 0);
1225 set_task_cpu(p, cpu);
1226 activate_task(dst_rq, p, 0);
1227 p->on_rq = TASK_ON_RQ_QUEUED;
1228 check_preempt_curr(dst_rq, p, 0);
1231 * Task isn't running anymore; make it appear like we migrated
1232 * it before it went to sleep. This means on wakeup we make the
1233 * previous CPU our target instead of where it really is.
1239 struct migration_swap_arg {
1240 struct task_struct *src_task, *dst_task;
1241 int src_cpu, dst_cpu;
1244 static int migrate_swap_stop(void *data)
1246 struct migration_swap_arg *arg = data;
1247 struct rq *src_rq, *dst_rq;
1250 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1253 src_rq = cpu_rq(arg->src_cpu);
1254 dst_rq = cpu_rq(arg->dst_cpu);
1256 double_raw_lock(&arg->src_task->pi_lock,
1257 &arg->dst_task->pi_lock);
1258 double_rq_lock(src_rq, dst_rq);
1260 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1263 if (task_cpu(arg->src_task) != arg->src_cpu)
1266 if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1269 if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1272 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1273 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1278 double_rq_unlock(src_rq, dst_rq);
1279 raw_spin_unlock(&arg->dst_task->pi_lock);
1280 raw_spin_unlock(&arg->src_task->pi_lock);
1286 * Cross migrate two tasks
1288 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1290 struct migration_swap_arg arg;
1293 arg = (struct migration_swap_arg){
1295 .src_cpu = task_cpu(cur),
1297 .dst_cpu = task_cpu(p),
1300 if (arg.src_cpu == arg.dst_cpu)
1304 * These three tests are all lockless; this is OK since all of them
1305 * will be re-checked with proper locks held further down the line.
1307 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1310 if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1313 if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1316 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1317 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1324 * wait_task_inactive - wait for a thread to unschedule.
1326 * If @match_state is nonzero, it's the @p->state value just checked and
1327 * not expected to change. If it changes, i.e. @p might have woken up,
1328 * then return zero. When we succeed in waiting for @p to be off its CPU,
1329 * we return a positive number (its total switch count). If a second call
1330 * a short while later returns the same number, the caller can be sure that
1331 * @p has remained unscheduled the whole time.
1333 * The caller must ensure that the task *will* unschedule sometime soon,
1334 * else this function might spin for a *long* time. This function can't
1335 * be called with interrupts off, or it may introduce deadlock with
1336 * smp_call_function() if an IPI is sent by the same process we are
1337 * waiting to become inactive.
1339 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1341 int running, queued;
1348 * We do the initial early heuristics without holding
1349 * any task-queue locks at all. We'll only try to get
1350 * the runqueue lock when things look like they will
1356 * If the task is actively running on another CPU
1357 * still, just relax and busy-wait without holding
1360 * NOTE! Since we don't hold any locks, it's not
1361 * even sure that "rq" stays as the right runqueue!
1362 * But we don't care, since "task_running()" will
1363 * return false if the runqueue has changed and p
1364 * is actually now running somewhere else!
1366 while (task_running(rq, p)) {
1367 if (match_state && unlikely(p->state != match_state))
1373 * Ok, time to look more closely! We need the rq
1374 * lock now, to be *sure*. If we're wrong, we'll
1375 * just go back and repeat.
1377 rq = task_rq_lock(p, &rf);
1378 trace_sched_wait_task(p);
1379 running = task_running(rq, p);
1380 queued = task_on_rq_queued(p);
1382 if (!match_state || p->state == match_state)
1383 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1384 task_rq_unlock(rq, p, &rf);
1387 * If it changed from the expected state, bail out now.
1389 if (unlikely(!ncsw))
1393 * Was it really running after all now that we
1394 * checked with the proper locks actually held?
1396 * Oops. Go back and try again..
1398 if (unlikely(running)) {
1404 * It's not enough that it's not actively running,
1405 * it must be off the runqueue _entirely_, and not
1408 * So if it was still runnable (but just not actively
1409 * running right now), it's preempted, and we should
1410 * yield - it could be a while.
1412 if (unlikely(queued)) {
1413 ktime_t to = NSEC_PER_SEC / HZ;
1415 set_current_state(TASK_UNINTERRUPTIBLE);
1416 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1421 * Ahh, all good. It wasn't running, and it wasn't
1422 * runnable, which means that it will never become
1423 * running in the future either. We're all done!
1432 * kick_process - kick a running thread to enter/exit the kernel
1433 * @p: the to-be-kicked thread
1435 * Cause a process which is running on another CPU to enter
1436 * kernel-mode, without any delay. (to get signals handled.)
1438 * NOTE: this function doesn't have to take the runqueue lock,
1439 * because all it wants to ensure is that the remote task enters
1440 * the kernel. If the IPI races and the task has been migrated
1441 * to another CPU then no harm is done and the purpose has been
1444 void kick_process(struct task_struct *p)
1450 if ((cpu != smp_processor_id()) && task_curr(p))
1451 smp_send_reschedule(cpu);
1454 EXPORT_SYMBOL_GPL(kick_process);
1457 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1459 * A few notes on cpu_active vs cpu_online:
1461 * - cpu_active must be a subset of cpu_online
1463 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1464 * see __set_cpus_allowed_ptr(). At this point the newly online
1465 * CPU isn't yet part of the sched domains, and balancing will not
1468 * - on CPU-down we clear cpu_active() to mask the sched domains and
1469 * avoid the load balancer to place new tasks on the to be removed
1470 * CPU. Existing tasks will remain running there and will be taken
1473 * This means that fallback selection must not select !active CPUs.
1474 * And can assume that any active CPU must be online. Conversely
1475 * select_task_rq() below may allow selection of !active CPUs in order
1476 * to satisfy the above rules.
1478 static int select_fallback_rq(int cpu, struct task_struct *p)
1480 int nid = cpu_to_node(cpu);
1481 const struct cpumask *nodemask = NULL;
1482 enum { cpuset, possible, fail } state = cpuset;
1486 * If the node that the CPU is on has been offlined, cpu_to_node()
1487 * will return -1. There is no CPU on the node, and we should
1488 * select the CPU on the other node.
1491 nodemask = cpumask_of_node(nid);
1493 /* Look for allowed, online CPU in same node. */
1494 for_each_cpu(dest_cpu, nodemask) {
1495 if (!cpu_active(dest_cpu))
1497 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1503 /* Any allowed, online CPU? */
1504 for_each_cpu(dest_cpu, &p->cpus_allowed) {
1505 if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1507 if (!cpu_online(dest_cpu))
1512 /* No more Mr. Nice Guy. */
1515 if (IS_ENABLED(CONFIG_CPUSETS)) {
1516 cpuset_cpus_allowed_fallback(p);
1522 do_set_cpus_allowed(p, cpu_possible_mask);
1533 if (state != cpuset) {
1535 * Don't tell them about moving exiting tasks or
1536 * kernel threads (both mm NULL), since they never
1539 if (p->mm && printk_ratelimit()) {
1540 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1541 task_pid_nr(p), p->comm, cpu);
1549 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1552 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1554 lockdep_assert_held(&p->pi_lock);
1556 if (p->nr_cpus_allowed > 1)
1557 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1559 cpu = cpumask_any(&p->cpus_allowed);
1562 * In order not to call set_task_cpu() on a blocking task we need
1563 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1566 * Since this is common to all placement strategies, this lives here.
1568 * [ this allows ->select_task() to simply return task_cpu(p) and
1569 * not worry about this generic constraint ]
1571 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
1573 cpu = select_fallback_rq(task_cpu(p), p);
1578 static void update_avg(u64 *avg, u64 sample)
1580 s64 diff = sample - *avg;
1586 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1587 const struct cpumask *new_mask, bool check)
1589 return set_cpus_allowed_ptr(p, new_mask);
1592 #endif /* CONFIG_SMP */
1595 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1599 if (!schedstat_enabled())
1605 if (cpu == rq->cpu) {
1606 schedstat_inc(rq->ttwu_local);
1607 schedstat_inc(p->se.statistics.nr_wakeups_local);
1609 struct sched_domain *sd;
1611 schedstat_inc(p->se.statistics.nr_wakeups_remote);
1613 for_each_domain(rq->cpu, sd) {
1614 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1615 schedstat_inc(sd->ttwu_wake_remote);
1622 if (wake_flags & WF_MIGRATED)
1623 schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1624 #endif /* CONFIG_SMP */
1626 schedstat_inc(rq->ttwu_count);
1627 schedstat_inc(p->se.statistics.nr_wakeups);
1629 if (wake_flags & WF_SYNC)
1630 schedstat_inc(p->se.statistics.nr_wakeups_sync);
1633 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1635 activate_task(rq, p, en_flags);
1636 p->on_rq = TASK_ON_RQ_QUEUED;
1638 /* If a worker is waking up, notify the workqueue: */
1639 if (p->flags & PF_WQ_WORKER)
1640 wq_worker_waking_up(p, cpu_of(rq));
1644 * Mark the task runnable and perform wakeup-preemption.
1646 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1647 struct rq_flags *rf)
1649 check_preempt_curr(rq, p, wake_flags);
1650 p->state = TASK_RUNNING;
1651 trace_sched_wakeup(p);
1654 if (p->sched_class->task_woken) {
1656 * Our task @p is fully woken up and running; so its safe to
1657 * drop the rq->lock, hereafter rq is only used for statistics.
1659 rq_unpin_lock(rq, rf);
1660 p->sched_class->task_woken(rq, p);
1661 rq_repin_lock(rq, rf);
1664 if (rq->idle_stamp) {
1665 u64 delta = rq_clock(rq) - rq->idle_stamp;
1666 u64 max = 2*rq->max_idle_balance_cost;
1668 update_avg(&rq->avg_idle, delta);
1670 if (rq->avg_idle > max)
1679 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1680 struct rq_flags *rf)
1682 int en_flags = ENQUEUE_WAKEUP;
1684 lockdep_assert_held(&rq->lock);
1687 if (p->sched_contributes_to_load)
1688 rq->nr_uninterruptible--;
1690 if (wake_flags & WF_MIGRATED)
1691 en_flags |= ENQUEUE_MIGRATED;
1694 ttwu_activate(rq, p, en_flags);
1695 ttwu_do_wakeup(rq, p, wake_flags, rf);
1699 * Called in case the task @p isn't fully descheduled from its runqueue,
1700 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1701 * since all we need to do is flip p->state to TASK_RUNNING, since
1702 * the task is still ->on_rq.
1704 static int ttwu_remote(struct task_struct *p, int wake_flags)
1710 rq = __task_rq_lock(p, &rf);
1711 if (task_on_rq_queued(p)) {
1712 /* check_preempt_curr() may use rq clock */
1713 update_rq_clock(rq);
1714 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1717 __task_rq_unlock(rq, &rf);
1723 void sched_ttwu_pending(void)
1725 struct rq *rq = this_rq();
1726 struct llist_node *llist = llist_del_all(&rq->wake_list);
1727 struct task_struct *p;
1728 unsigned long flags;
1734 raw_spin_lock_irqsave(&rq->lock, flags);
1735 rq_pin_lock(rq, &rf);
1740 p = llist_entry(llist, struct task_struct, wake_entry);
1741 llist = llist_next(llist);
1743 if (p->sched_remote_wakeup)
1744 wake_flags = WF_MIGRATED;
1746 ttwu_do_activate(rq, p, wake_flags, &rf);
1749 rq_unpin_lock(rq, &rf);
1750 raw_spin_unlock_irqrestore(&rq->lock, flags);
1753 void scheduler_ipi(void)
1756 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1757 * TIF_NEED_RESCHED remotely (for the first time) will also send
1760 preempt_fold_need_resched();
1762 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1766 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1767 * traditionally all their work was done from the interrupt return
1768 * path. Now that we actually do some work, we need to make sure
1771 * Some archs already do call them, luckily irq_enter/exit nest
1774 * Arguably we should visit all archs and update all handlers,
1775 * however a fair share of IPIs are still resched only so this would
1776 * somewhat pessimize the simple resched case.
1779 sched_ttwu_pending();
1782 * Check if someone kicked us for doing the nohz idle load balance.
1784 if (unlikely(got_nohz_idle_kick())) {
1785 this_rq()->idle_balance = 1;
1786 raise_softirq_irqoff(SCHED_SOFTIRQ);
1791 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1793 struct rq *rq = cpu_rq(cpu);
1795 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1797 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1798 if (!set_nr_if_polling(rq->idle))
1799 smp_send_reschedule(cpu);
1801 trace_sched_wake_idle_without_ipi(cpu);
1805 void wake_up_if_idle(int cpu)
1807 struct rq *rq = cpu_rq(cpu);
1808 unsigned long flags;
1812 if (!is_idle_task(rcu_dereference(rq->curr)))
1815 if (set_nr_if_polling(rq->idle)) {
1816 trace_sched_wake_idle_without_ipi(cpu);
1818 raw_spin_lock_irqsave(&rq->lock, flags);
1819 if (is_idle_task(rq->curr))
1820 smp_send_reschedule(cpu);
1821 /* Else CPU is not idle, do nothing here: */
1822 raw_spin_unlock_irqrestore(&rq->lock, flags);
1829 bool cpus_share_cache(int this_cpu, int that_cpu)
1831 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1833 #endif /* CONFIG_SMP */
1835 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1837 struct rq *rq = cpu_rq(cpu);
1840 #if defined(CONFIG_SMP)
1841 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1842 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1843 ttwu_queue_remote(p, cpu, wake_flags);
1848 raw_spin_lock(&rq->lock);
1849 rq_pin_lock(rq, &rf);
1850 ttwu_do_activate(rq, p, wake_flags, &rf);
1851 rq_unpin_lock(rq, &rf);
1852 raw_spin_unlock(&rq->lock);
1856 * Notes on Program-Order guarantees on SMP systems.
1860 * The basic program-order guarantee on SMP systems is that when a task [t]
1861 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1862 * execution on its new CPU [c1].
1864 * For migration (of runnable tasks) this is provided by the following means:
1866 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1867 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1868 * rq(c1)->lock (if not at the same time, then in that order).
1869 * C) LOCK of the rq(c1)->lock scheduling in task
1871 * Transitivity guarantees that B happens after A and C after B.
1872 * Note: we only require RCpc transitivity.
1873 * Note: the CPU doing B need not be c0 or c1
1882 * UNLOCK rq(0)->lock
1884 * LOCK rq(0)->lock // orders against CPU0
1886 * UNLOCK rq(0)->lock
1890 * UNLOCK rq(1)->lock
1892 * LOCK rq(1)->lock // orders against CPU2
1895 * UNLOCK rq(1)->lock
1898 * BLOCKING -- aka. SLEEP + WAKEUP
1900 * For blocking we (obviously) need to provide the same guarantee as for
1901 * migration. However the means are completely different as there is no lock
1902 * chain to provide order. Instead we do:
1904 * 1) smp_store_release(X->on_cpu, 0)
1905 * 2) smp_cond_load_acquire(!X->on_cpu)
1909 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1911 * LOCK rq(0)->lock LOCK X->pi_lock
1914 * smp_store_release(X->on_cpu, 0);
1916 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1922 * X->state = RUNNING
1923 * UNLOCK rq(2)->lock
1925 * LOCK rq(2)->lock // orders against CPU1
1928 * UNLOCK rq(2)->lock
1931 * UNLOCK rq(0)->lock
1934 * However; for wakeups there is a second guarantee we must provide, namely we
1935 * must observe the state that lead to our wakeup. That is, not only must our
1936 * task observe its own prior state, it must also observe the stores prior to
1939 * This means that any means of doing remote wakeups must order the CPU doing
1940 * the wakeup against the CPU the task is going to end up running on. This,
1941 * however, is already required for the regular Program-Order guarantee above,
1942 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1947 * try_to_wake_up - wake up a thread
1948 * @p: the thread to be awakened
1949 * @state: the mask of task states that can be woken
1950 * @wake_flags: wake modifier flags (WF_*)
1952 * If (@state & @p->state) @p->state = TASK_RUNNING.
1954 * If the task was not queued/runnable, also place it back on a runqueue.
1956 * Atomic against schedule() which would dequeue a task, also see
1957 * set_current_state().
1959 * Return: %true if @p->state changes (an actual wakeup was done),
1963 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1965 unsigned long flags;
1966 int cpu, success = 0;
1969 * If we are going to wake up a thread waiting for CONDITION we
1970 * need to ensure that CONDITION=1 done by the caller can not be
1971 * reordered with p->state check below. This pairs with mb() in
1972 * set_current_state() the waiting thread does.
1974 smp_mb__before_spinlock();
1975 raw_spin_lock_irqsave(&p->pi_lock, flags);
1976 if (!(p->state & state))
1979 trace_sched_waking(p);
1981 /* We're going to change ->state: */
1986 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1987 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1988 * in smp_cond_load_acquire() below.
1990 * sched_ttwu_pending() try_to_wake_up()
1991 * [S] p->on_rq = 1; [L] P->state
1992 * UNLOCK rq->lock -----.
1996 * LOCK rq->lock -----'
2000 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2002 * Pairs with the UNLOCK+LOCK on rq->lock from the
2003 * last wakeup of our task and the schedule that got our task
2007 if (p->on_rq && ttwu_remote(p, wake_flags))
2012 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2013 * possible to, falsely, observe p->on_cpu == 0.
2015 * One must be running (->on_cpu == 1) in order to remove oneself
2016 * from the runqueue.
2018 * [S] ->on_cpu = 1; [L] ->on_rq
2022 * [S] ->on_rq = 0; [L] ->on_cpu
2024 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2025 * from the consecutive calls to schedule(); the first switching to our
2026 * task, the second putting it to sleep.
2031 * If the owning (remote) CPU is still in the middle of schedule() with
2032 * this task as prev, wait until its done referencing the task.
2034 * Pairs with the smp_store_release() in finish_lock_switch().
2036 * This ensures that tasks getting woken will be fully ordered against
2037 * their previous state and preserve Program Order.
2039 smp_cond_load_acquire(&p->on_cpu, !VAL);
2041 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2042 p->state = TASK_WAKING;
2045 delayacct_blkio_end();
2046 atomic_dec(&task_rq(p)->nr_iowait);
2049 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2050 if (task_cpu(p) != cpu) {
2051 wake_flags |= WF_MIGRATED;
2052 set_task_cpu(p, cpu);
2055 #else /* CONFIG_SMP */
2058 delayacct_blkio_end();
2059 atomic_dec(&task_rq(p)->nr_iowait);
2062 #endif /* CONFIG_SMP */
2064 ttwu_queue(p, cpu, wake_flags);
2066 ttwu_stat(p, cpu, wake_flags);
2068 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2074 * try_to_wake_up_local - try to wake up a local task with rq lock held
2075 * @p: the thread to be awakened
2076 * @cookie: context's cookie for pinning
2078 * Put @p on the run-queue if it's not already there. The caller must
2079 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2082 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2084 struct rq *rq = task_rq(p);
2086 if (WARN_ON_ONCE(rq != this_rq()) ||
2087 WARN_ON_ONCE(p == current))
2090 lockdep_assert_held(&rq->lock);
2092 if (!raw_spin_trylock(&p->pi_lock)) {
2094 * This is OK, because current is on_cpu, which avoids it being
2095 * picked for load-balance and preemption/IRQs are still
2096 * disabled avoiding further scheduler activity on it and we've
2097 * not yet picked a replacement task.
2099 rq_unpin_lock(rq, rf);
2100 raw_spin_unlock(&rq->lock);
2101 raw_spin_lock(&p->pi_lock);
2102 raw_spin_lock(&rq->lock);
2103 rq_repin_lock(rq, rf);
2106 if (!(p->state & TASK_NORMAL))
2109 trace_sched_waking(p);
2111 if (!task_on_rq_queued(p)) {
2113 delayacct_blkio_end();
2114 atomic_dec(&rq->nr_iowait);
2116 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2119 ttwu_do_wakeup(rq, p, 0, rf);
2120 ttwu_stat(p, smp_processor_id(), 0);
2122 raw_spin_unlock(&p->pi_lock);
2126 * wake_up_process - Wake up a specific process
2127 * @p: The process to be woken up.
2129 * Attempt to wake up the nominated process and move it to the set of runnable
2132 * Return: 1 if the process was woken up, 0 if it was already running.
2134 * It may be assumed that this function implies a write memory barrier before
2135 * changing the task state if and only if any tasks are woken up.
2137 int wake_up_process(struct task_struct *p)
2139 return try_to_wake_up(p, TASK_NORMAL, 0);
2141 EXPORT_SYMBOL(wake_up_process);
2143 int wake_up_state(struct task_struct *p, unsigned int state)
2145 return try_to_wake_up(p, state, 0);
2149 * This function clears the sched_dl_entity static params.
2151 void __dl_clear_params(struct task_struct *p)
2153 struct sched_dl_entity *dl_se = &p->dl;
2155 dl_se->dl_runtime = 0;
2156 dl_se->dl_deadline = 0;
2157 dl_se->dl_period = 0;
2161 dl_se->dl_throttled = 0;
2162 dl_se->dl_yielded = 0;
2166 * Perform scheduler related setup for a newly forked process p.
2167 * p is forked by current.
2169 * __sched_fork() is basic setup used by init_idle() too:
2171 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2176 p->se.exec_start = 0;
2177 p->se.sum_exec_runtime = 0;
2178 p->se.prev_sum_exec_runtime = 0;
2179 p->se.nr_migrations = 0;
2181 INIT_LIST_HEAD(&p->se.group_node);
2183 #ifdef CONFIG_FAIR_GROUP_SCHED
2184 p->se.cfs_rq = NULL;
2187 #ifdef CONFIG_SCHEDSTATS
2188 /* Even if schedstat is disabled, there should not be garbage */
2189 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2192 RB_CLEAR_NODE(&p->dl.rb_node);
2193 init_dl_task_timer(&p->dl);
2194 __dl_clear_params(p);
2196 INIT_LIST_HEAD(&p->rt.run_list);
2198 p->rt.time_slice = sched_rr_timeslice;
2202 #ifdef CONFIG_PREEMPT_NOTIFIERS
2203 INIT_HLIST_HEAD(&p->preempt_notifiers);
2206 #ifdef CONFIG_NUMA_BALANCING
2207 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2208 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2209 p->mm->numa_scan_seq = 0;
2212 if (clone_flags & CLONE_VM)
2213 p->numa_preferred_nid = current->numa_preferred_nid;
2215 p->numa_preferred_nid = -1;
2217 p->node_stamp = 0ULL;
2218 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2219 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2220 p->numa_work.next = &p->numa_work;
2221 p->numa_faults = NULL;
2222 p->last_task_numa_placement = 0;
2223 p->last_sum_exec_runtime = 0;
2225 p->numa_group = NULL;
2226 #endif /* CONFIG_NUMA_BALANCING */
2229 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2231 #ifdef CONFIG_NUMA_BALANCING
2233 void set_numabalancing_state(bool enabled)
2236 static_branch_enable(&sched_numa_balancing);
2238 static_branch_disable(&sched_numa_balancing);
2241 #ifdef CONFIG_PROC_SYSCTL
2242 int sysctl_numa_balancing(struct ctl_table *table, int write,
2243 void __user *buffer, size_t *lenp, loff_t *ppos)
2247 int state = static_branch_likely(&sched_numa_balancing);
2249 if (write && !capable(CAP_SYS_ADMIN))
2254 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2258 set_numabalancing_state(state);
2264 #ifdef CONFIG_SCHEDSTATS
2266 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2267 static bool __initdata __sched_schedstats = false;
2269 static void set_schedstats(bool enabled)
2272 static_branch_enable(&sched_schedstats);
2274 static_branch_disable(&sched_schedstats);
2277 void force_schedstat_enabled(void)
2279 if (!schedstat_enabled()) {
2280 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2281 static_branch_enable(&sched_schedstats);
2285 static int __init setup_schedstats(char *str)
2292 * This code is called before jump labels have been set up, so we can't
2293 * change the static branch directly just yet. Instead set a temporary
2294 * variable so init_schedstats() can do it later.
2296 if (!strcmp(str, "enable")) {
2297 __sched_schedstats = true;
2299 } else if (!strcmp(str, "disable")) {
2300 __sched_schedstats = false;
2305 pr_warn("Unable to parse schedstats=\n");
2309 __setup("schedstats=", setup_schedstats);
2311 static void __init init_schedstats(void)
2313 set_schedstats(__sched_schedstats);
2316 #ifdef CONFIG_PROC_SYSCTL
2317 int sysctl_schedstats(struct ctl_table *table, int write,
2318 void __user *buffer, size_t *lenp, loff_t *ppos)
2322 int state = static_branch_likely(&sched_schedstats);
2324 if (write && !capable(CAP_SYS_ADMIN))
2329 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2333 set_schedstats(state);
2336 #endif /* CONFIG_PROC_SYSCTL */
2337 #else /* !CONFIG_SCHEDSTATS */
2338 static inline void init_schedstats(void) {}
2339 #endif /* CONFIG_SCHEDSTATS */
2342 * fork()/clone()-time setup:
2344 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2346 unsigned long flags;
2347 int cpu = get_cpu();
2349 __sched_fork(clone_flags, p);
2351 * We mark the process as NEW here. This guarantees that
2352 * nobody will actually run it, and a signal or other external
2353 * event cannot wake it up and insert it on the runqueue either.
2355 p->state = TASK_NEW;
2358 * Make sure we do not leak PI boosting priority to the child.
2360 p->prio = current->normal_prio;
2363 * Revert to default priority/policy on fork if requested.
2365 if (unlikely(p->sched_reset_on_fork)) {
2366 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2367 p->policy = SCHED_NORMAL;
2368 p->static_prio = NICE_TO_PRIO(0);
2370 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2371 p->static_prio = NICE_TO_PRIO(0);
2373 p->prio = p->normal_prio = __normal_prio(p);
2377 * We don't need the reset flag anymore after the fork. It has
2378 * fulfilled its duty:
2380 p->sched_reset_on_fork = 0;
2383 if (dl_prio(p->prio)) {
2386 } else if (rt_prio(p->prio)) {
2387 p->sched_class = &rt_sched_class;
2389 p->sched_class = &fair_sched_class;
2392 init_entity_runnable_average(&p->se);
2395 * The child is not yet in the pid-hash so no cgroup attach races,
2396 * and the cgroup is pinned to this child due to cgroup_fork()
2397 * is ran before sched_fork().
2399 * Silence PROVE_RCU.
2401 raw_spin_lock_irqsave(&p->pi_lock, flags);
2403 * We're setting the CPU for the first time, we don't migrate,
2404 * so use __set_task_cpu().
2406 __set_task_cpu(p, cpu);
2407 if (p->sched_class->task_fork)
2408 p->sched_class->task_fork(p);
2409 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2411 #ifdef CONFIG_SCHED_INFO
2412 if (likely(sched_info_on()))
2413 memset(&p->sched_info, 0, sizeof(p->sched_info));
2415 #if defined(CONFIG_SMP)
2418 init_task_preempt_count(p);
2420 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2421 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2428 unsigned long to_ratio(u64 period, u64 runtime)
2430 if (runtime == RUNTIME_INF)
2434 * Doing this here saves a lot of checks in all
2435 * the calling paths, and returning zero seems
2436 * safe for them anyway.
2441 return div64_u64(runtime << 20, period);
2445 inline struct dl_bw *dl_bw_of(int i)
2447 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2448 "sched RCU must be held");
2449 return &cpu_rq(i)->rd->dl_bw;
2452 static inline int dl_bw_cpus(int i)
2454 struct root_domain *rd = cpu_rq(i)->rd;
2457 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2458 "sched RCU must be held");
2459 for_each_cpu_and(i, rd->span, cpu_active_mask)
2465 inline struct dl_bw *dl_bw_of(int i)
2467 return &cpu_rq(i)->dl.dl_bw;
2470 static inline int dl_bw_cpus(int i)
2477 * We must be sure that accepting a new task (or allowing changing the
2478 * parameters of an existing one) is consistent with the bandwidth
2479 * constraints. If yes, this function also accordingly updates the currently
2480 * allocated bandwidth to reflect the new situation.
2482 * This function is called while holding p's rq->lock.
2484 * XXX we should delay bw change until the task's 0-lag point, see
2487 static int dl_overflow(struct task_struct *p, int policy,
2488 const struct sched_attr *attr)
2491 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2492 u64 period = attr->sched_period ?: attr->sched_deadline;
2493 u64 runtime = attr->sched_runtime;
2494 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2497 /* !deadline task may carry old deadline bandwidth */
2498 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2502 * Either if a task, enters, leave, or stays -deadline but changes
2503 * its parameters, we may need to update accordingly the total
2504 * allocated bandwidth of the container.
2506 raw_spin_lock(&dl_b->lock);
2507 cpus = dl_bw_cpus(task_cpu(p));
2508 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2509 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2510 __dl_add(dl_b, new_bw);
2512 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2513 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2514 __dl_clear(dl_b, p->dl.dl_bw);
2515 __dl_add(dl_b, new_bw);
2517 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2518 __dl_clear(dl_b, p->dl.dl_bw);
2521 raw_spin_unlock(&dl_b->lock);
2526 extern void init_dl_bw(struct dl_bw *dl_b);
2529 * wake_up_new_task - wake up a newly created task for the first time.
2531 * This function will do some initial scheduler statistics housekeeping
2532 * that must be done for every newly created context, then puts the task
2533 * on the runqueue and wakes it.
2535 void wake_up_new_task(struct task_struct *p)
2540 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2541 p->state = TASK_RUNNING;
2544 * Fork balancing, do it here and not earlier because:
2545 * - cpus_allowed can change in the fork path
2546 * - any previously selected CPU might disappear through hotplug
2548 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2549 * as we're not fully set-up yet.
2551 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2553 rq = __task_rq_lock(p, &rf);
2554 update_rq_clock(rq);
2555 post_init_entity_util_avg(&p->se);
2557 activate_task(rq, p, 0);
2558 p->on_rq = TASK_ON_RQ_QUEUED;
2559 trace_sched_wakeup_new(p);
2560 check_preempt_curr(rq, p, WF_FORK);
2562 if (p->sched_class->task_woken) {
2564 * Nothing relies on rq->lock after this, so its fine to
2567 rq_unpin_lock(rq, &rf);
2568 p->sched_class->task_woken(rq, p);
2569 rq_repin_lock(rq, &rf);
2572 task_rq_unlock(rq, p, &rf);
2575 #ifdef CONFIG_PREEMPT_NOTIFIERS
2577 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2579 void preempt_notifier_inc(void)
2581 static_key_slow_inc(&preempt_notifier_key);
2583 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2585 void preempt_notifier_dec(void)
2587 static_key_slow_dec(&preempt_notifier_key);
2589 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2592 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2593 * @notifier: notifier struct to register
2595 void preempt_notifier_register(struct preempt_notifier *notifier)
2597 if (!static_key_false(&preempt_notifier_key))
2598 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2600 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2602 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2605 * preempt_notifier_unregister - no longer interested in preemption notifications
2606 * @notifier: notifier struct to unregister
2608 * This is *not* safe to call from within a preemption notifier.
2610 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2612 hlist_del(¬ifier->link);
2614 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2616 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2618 struct preempt_notifier *notifier;
2620 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2621 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2624 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2626 if (static_key_false(&preempt_notifier_key))
2627 __fire_sched_in_preempt_notifiers(curr);
2631 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2632 struct task_struct *next)
2634 struct preempt_notifier *notifier;
2636 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2637 notifier->ops->sched_out(notifier, next);
2640 static __always_inline void
2641 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2642 struct task_struct *next)
2644 if (static_key_false(&preempt_notifier_key))
2645 __fire_sched_out_preempt_notifiers(curr, next);
2648 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2650 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2655 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2656 struct task_struct *next)
2660 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2663 * prepare_task_switch - prepare to switch tasks
2664 * @rq: the runqueue preparing to switch
2665 * @prev: the current task that is being switched out
2666 * @next: the task we are going to switch to.
2668 * This is called with the rq lock held and interrupts off. It must
2669 * be paired with a subsequent finish_task_switch after the context
2672 * prepare_task_switch sets up locking and calls architecture specific
2676 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2677 struct task_struct *next)
2679 sched_info_switch(rq, prev, next);
2680 perf_event_task_sched_out(prev, next);
2681 fire_sched_out_preempt_notifiers(prev, next);
2682 prepare_lock_switch(rq, next);
2683 prepare_arch_switch(next);
2687 * finish_task_switch - clean up after a task-switch
2688 * @prev: the thread we just switched away from.
2690 * finish_task_switch must be called after the context switch, paired
2691 * with a prepare_task_switch call before the context switch.
2692 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2693 * and do any other architecture-specific cleanup actions.
2695 * Note that we may have delayed dropping an mm in context_switch(). If
2696 * so, we finish that here outside of the runqueue lock. (Doing it
2697 * with the lock held can cause deadlocks; see schedule() for
2700 * The context switch have flipped the stack from under us and restored the
2701 * local variables which were saved when this task called schedule() in the
2702 * past. prev == current is still correct but we need to recalculate this_rq
2703 * because prev may have moved to another CPU.
2705 static struct rq *finish_task_switch(struct task_struct *prev)
2706 __releases(rq->lock)
2708 struct rq *rq = this_rq();
2709 struct mm_struct *mm = rq->prev_mm;
2713 * The previous task will have left us with a preempt_count of 2
2714 * because it left us after:
2717 * preempt_disable(); // 1
2719 * raw_spin_lock_irq(&rq->lock) // 2
2721 * Also, see FORK_PREEMPT_COUNT.
2723 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2724 "corrupted preempt_count: %s/%d/0x%x\n",
2725 current->comm, current->pid, preempt_count()))
2726 preempt_count_set(FORK_PREEMPT_COUNT);
2731 * A task struct has one reference for the use as "current".
2732 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2733 * schedule one last time. The schedule call will never return, and
2734 * the scheduled task must drop that reference.
2736 * We must observe prev->state before clearing prev->on_cpu (in
2737 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2738 * running on another CPU and we could rave with its RUNNING -> DEAD
2739 * transition, resulting in a double drop.
2741 prev_state = prev->state;
2742 vtime_task_switch(prev);
2743 perf_event_task_sched_in(prev, current);
2744 finish_lock_switch(rq, prev);
2745 finish_arch_post_lock_switch();
2747 fire_sched_in_preempt_notifiers(current);
2750 if (unlikely(prev_state == TASK_DEAD)) {
2751 if (prev->sched_class->task_dead)
2752 prev->sched_class->task_dead(prev);
2755 * Remove function-return probe instances associated with this
2756 * task and put them back on the free list.
2758 kprobe_flush_task(prev);
2760 /* Task is done with its stack. */
2761 put_task_stack(prev);
2763 put_task_struct(prev);
2766 tick_nohz_task_switch();
2772 /* rq->lock is NOT held, but preemption is disabled */
2773 static void __balance_callback(struct rq *rq)
2775 struct callback_head *head, *next;
2776 void (*func)(struct rq *rq);
2777 unsigned long flags;
2779 raw_spin_lock_irqsave(&rq->lock, flags);
2780 head = rq->balance_callback;
2781 rq->balance_callback = NULL;
2783 func = (void (*)(struct rq *))head->func;
2790 raw_spin_unlock_irqrestore(&rq->lock, flags);
2793 static inline void balance_callback(struct rq *rq)
2795 if (unlikely(rq->balance_callback))
2796 __balance_callback(rq);
2801 static inline void balance_callback(struct rq *rq)
2808 * schedule_tail - first thing a freshly forked thread must call.
2809 * @prev: the thread we just switched away from.
2811 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2812 __releases(rq->lock)
2817 * New tasks start with FORK_PREEMPT_COUNT, see there and
2818 * finish_task_switch() for details.
2820 * finish_task_switch() will drop rq->lock() and lower preempt_count
2821 * and the preempt_enable() will end up enabling preemption (on
2822 * PREEMPT_COUNT kernels).
2825 rq = finish_task_switch(prev);
2826 balance_callback(rq);
2829 if (current->set_child_tid)
2830 put_user(task_pid_vnr(current), current->set_child_tid);
2834 * context_switch - switch to the new MM and the new thread's register state.
2836 static __always_inline struct rq *
2837 context_switch(struct rq *rq, struct task_struct *prev,
2838 struct task_struct *next, struct rq_flags *rf)
2840 struct mm_struct *mm, *oldmm;
2842 prepare_task_switch(rq, prev, next);
2845 oldmm = prev->active_mm;
2847 * For paravirt, this is coupled with an exit in switch_to to
2848 * combine the page table reload and the switch backend into
2851 arch_start_context_switch(prev);
2854 next->active_mm = oldmm;
2856 enter_lazy_tlb(oldmm, next);
2858 switch_mm_irqs_off(oldmm, mm, next);
2861 prev->active_mm = NULL;
2862 rq->prev_mm = oldmm;
2865 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2868 * Since the runqueue lock will be released by the next
2869 * task (which is an invalid locking op but in the case
2870 * of the scheduler it's an obvious special-case), so we
2871 * do an early lockdep release here:
2873 rq_unpin_lock(rq, rf);
2874 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2876 /* Here we just switch the register state and the stack. */
2877 switch_to(prev, next, prev);
2880 return finish_task_switch(prev);
2884 * nr_running and nr_context_switches:
2886 * externally visible scheduler statistics: current number of runnable
2887 * threads, total number of context switches performed since bootup.
2889 unsigned long nr_running(void)
2891 unsigned long i, sum = 0;
2893 for_each_online_cpu(i)
2894 sum += cpu_rq(i)->nr_running;
2900 * Check if only the current task is running on the CPU.
2902 * Caution: this function does not check that the caller has disabled
2903 * preemption, thus the result might have a time-of-check-to-time-of-use
2904 * race. The caller is responsible to use it correctly, for example:
2906 * - from a non-preemptable section (of course)
2908 * - from a thread that is bound to a single CPU
2910 * - in a loop with very short iterations (e.g. a polling loop)
2912 bool single_task_running(void)
2914 return raw_rq()->nr_running == 1;
2916 EXPORT_SYMBOL(single_task_running);
2918 unsigned long long nr_context_switches(void)
2921 unsigned long long sum = 0;
2923 for_each_possible_cpu(i)
2924 sum += cpu_rq(i)->nr_switches;
2930 * IO-wait accounting, and how its mostly bollocks (on SMP).
2932 * The idea behind IO-wait account is to account the idle time that we could
2933 * have spend running if it were not for IO. That is, if we were to improve the
2934 * storage performance, we'd have a proportional reduction in IO-wait time.
2936 * This all works nicely on UP, where, when a task blocks on IO, we account
2937 * idle time as IO-wait, because if the storage were faster, it could've been
2938 * running and we'd not be idle.
2940 * This has been extended to SMP, by doing the same for each CPU. This however
2943 * Imagine for instance the case where two tasks block on one CPU, only the one
2944 * CPU will have IO-wait accounted, while the other has regular idle. Even
2945 * though, if the storage were faster, both could've ran at the same time,
2946 * utilising both CPUs.
2948 * This means, that when looking globally, the current IO-wait accounting on
2949 * SMP is a lower bound, by reason of under accounting.
2951 * Worse, since the numbers are provided per CPU, they are sometimes
2952 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2953 * associated with any one particular CPU, it can wake to another CPU than it
2954 * blocked on. This means the per CPU IO-wait number is meaningless.
2956 * Task CPU affinities can make all that even more 'interesting'.
2959 unsigned long nr_iowait(void)
2961 unsigned long i, sum = 0;
2963 for_each_possible_cpu(i)
2964 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2970 * Consumers of these two interfaces, like for example the cpufreq menu
2971 * governor are using nonsensical data. Boosting frequency for a CPU that has
2972 * IO-wait which might not even end up running the task when it does become
2976 unsigned long nr_iowait_cpu(int cpu)
2978 struct rq *this = cpu_rq(cpu);
2979 return atomic_read(&this->nr_iowait);
2982 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2984 struct rq *rq = this_rq();
2985 *nr_waiters = atomic_read(&rq->nr_iowait);
2986 *load = rq->load.weight;
2992 * sched_exec - execve() is a valuable balancing opportunity, because at
2993 * this point the task has the smallest effective memory and cache footprint.
2995 void sched_exec(void)
2997 struct task_struct *p = current;
2998 unsigned long flags;
3001 raw_spin_lock_irqsave(&p->pi_lock, flags);
3002 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3003 if (dest_cpu == smp_processor_id())
3006 if (likely(cpu_active(dest_cpu))) {
3007 struct migration_arg arg = { p, dest_cpu };
3009 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3010 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3014 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3019 DEFINE_PER_CPU(struct kernel_stat, kstat);
3020 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3022 EXPORT_PER_CPU_SYMBOL(kstat);
3023 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3026 * The function fair_sched_class.update_curr accesses the struct curr
3027 * and its field curr->exec_start; when called from task_sched_runtime(),
3028 * we observe a high rate of cache misses in practice.
3029 * Prefetching this data results in improved performance.
3031 static inline void prefetch_curr_exec_start(struct task_struct *p)
3033 #ifdef CONFIG_FAIR_GROUP_SCHED
3034 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3036 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3039 prefetch(&curr->exec_start);
3043 * Return accounted runtime for the task.
3044 * In case the task is currently running, return the runtime plus current's
3045 * pending runtime that have not been accounted yet.
3047 unsigned long long task_sched_runtime(struct task_struct *p)
3053 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3055 * 64-bit doesn't need locks to atomically read a 64bit value.
3056 * So we have a optimization chance when the task's delta_exec is 0.
3057 * Reading ->on_cpu is racy, but this is ok.
3059 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3060 * If we race with it entering CPU, unaccounted time is 0. This is
3061 * indistinguishable from the read occurring a few cycles earlier.
3062 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3063 * been accounted, so we're correct here as well.
3065 if (!p->on_cpu || !task_on_rq_queued(p))
3066 return p->se.sum_exec_runtime;
3069 rq = task_rq_lock(p, &rf);
3071 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3072 * project cycles that may never be accounted to this
3073 * thread, breaking clock_gettime().
3075 if (task_current(rq, p) && task_on_rq_queued(p)) {
3076 prefetch_curr_exec_start(p);
3077 update_rq_clock(rq);
3078 p->sched_class->update_curr(rq);
3080 ns = p->se.sum_exec_runtime;
3081 task_rq_unlock(rq, p, &rf);
3087 * This function gets called by the timer code, with HZ frequency.
3088 * We call it with interrupts disabled.
3090 void scheduler_tick(void)
3092 int cpu = smp_processor_id();
3093 struct rq *rq = cpu_rq(cpu);
3094 struct task_struct *curr = rq->curr;
3098 raw_spin_lock(&rq->lock);
3099 update_rq_clock(rq);
3100 curr->sched_class->task_tick(rq, curr, 0);
3101 cpu_load_update_active(rq);
3102 calc_global_load_tick(rq);
3103 raw_spin_unlock(&rq->lock);
3105 perf_event_task_tick();
3108 rq->idle_balance = idle_cpu(cpu);
3109 trigger_load_balance(rq);
3111 rq_last_tick_reset(rq);
3114 #ifdef CONFIG_NO_HZ_FULL
3116 * scheduler_tick_max_deferment
3118 * Keep at least one tick per second when a single
3119 * active task is running because the scheduler doesn't
3120 * yet completely support full dynticks environment.
3122 * This makes sure that uptime, CFS vruntime, load
3123 * balancing, etc... continue to move forward, even
3124 * with a very low granularity.
3126 * Return: Maximum deferment in nanoseconds.
3128 u64 scheduler_tick_max_deferment(void)
3130 struct rq *rq = this_rq();
3131 unsigned long next, now = READ_ONCE(jiffies);
3133 next = rq->last_sched_tick + HZ;
3135 if (time_before_eq(next, now))
3138 return jiffies_to_nsecs(next - now);
3142 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3143 defined(CONFIG_PREEMPT_TRACER))
3145 * If the value passed in is equal to the current preempt count
3146 * then we just disabled preemption. Start timing the latency.
3148 static inline void preempt_latency_start(int val)
3150 if (preempt_count() == val) {
3151 unsigned long ip = get_lock_parent_ip();
3152 #ifdef CONFIG_DEBUG_PREEMPT
3153 current->preempt_disable_ip = ip;
3155 trace_preempt_off(CALLER_ADDR0, ip);
3159 void preempt_count_add(int val)
3161 #ifdef CONFIG_DEBUG_PREEMPT
3165 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3168 __preempt_count_add(val);
3169 #ifdef CONFIG_DEBUG_PREEMPT
3171 * Spinlock count overflowing soon?
3173 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3176 preempt_latency_start(val);
3178 EXPORT_SYMBOL(preempt_count_add);
3179 NOKPROBE_SYMBOL(preempt_count_add);
3182 * If the value passed in equals to the current preempt count
3183 * then we just enabled preemption. Stop timing the latency.
3185 static inline void preempt_latency_stop(int val)
3187 if (preempt_count() == val)
3188 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3191 void preempt_count_sub(int val)
3193 #ifdef CONFIG_DEBUG_PREEMPT
3197 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3200 * Is the spinlock portion underflowing?
3202 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3203 !(preempt_count() & PREEMPT_MASK)))
3207 preempt_latency_stop(val);
3208 __preempt_count_sub(val);
3210 EXPORT_SYMBOL(preempt_count_sub);
3211 NOKPROBE_SYMBOL(preempt_count_sub);
3214 static inline void preempt_latency_start(int val) { }
3215 static inline void preempt_latency_stop(int val) { }
3218 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3220 #ifdef CONFIG_DEBUG_PREEMPT
3221 return p->preempt_disable_ip;
3228 * Print scheduling while atomic bug:
3230 static noinline void __schedule_bug(struct task_struct *prev)
3232 /* Save this before calling printk(), since that will clobber it */
3233 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3235 if (oops_in_progress)
3238 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3239 prev->comm, prev->pid, preempt_count());
3241 debug_show_held_locks(prev);
3243 if (irqs_disabled())
3244 print_irqtrace_events(prev);
3245 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3246 && in_atomic_preempt_off()) {
3247 pr_err("Preemption disabled at:");
3248 print_ip_sym(preempt_disable_ip);
3252 panic("scheduling while atomic\n");
3255 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3259 * Various schedule()-time debugging checks and statistics:
3261 static inline void schedule_debug(struct task_struct *prev)
3263 #ifdef CONFIG_SCHED_STACK_END_CHECK
3264 if (task_stack_end_corrupted(prev))
3265 panic("corrupted stack end detected inside scheduler\n");
3268 if (unlikely(in_atomic_preempt_off())) {
3269 __schedule_bug(prev);
3270 preempt_count_set(PREEMPT_DISABLED);
3274 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3276 schedstat_inc(this_rq()->sched_count);
3280 * Pick up the highest-prio task:
3282 static inline struct task_struct *
3283 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3285 const struct sched_class *class;
3286 struct task_struct *p;
3289 * Optimization: we know that if all tasks are in
3290 * the fair class we can call that function directly:
3292 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3293 p = fair_sched_class.pick_next_task(rq, prev, rf);
3294 if (unlikely(p == RETRY_TASK))
3297 /* Assumes fair_sched_class->next == idle_sched_class */
3299 p = idle_sched_class.pick_next_task(rq, prev, rf);
3305 for_each_class(class) {
3306 p = class->pick_next_task(rq, prev, rf);
3308 if (unlikely(p == RETRY_TASK))
3314 /* The idle class should always have a runnable task: */
3319 * __schedule() is the main scheduler function.
3321 * The main means of driving the scheduler and thus entering this function are:
3323 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3325 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3326 * paths. For example, see arch/x86/entry_64.S.
3328 * To drive preemption between tasks, the scheduler sets the flag in timer
3329 * interrupt handler scheduler_tick().
3331 * 3. Wakeups don't really cause entry into schedule(). They add a
3332 * task to the run-queue and that's it.
3334 * Now, if the new task added to the run-queue preempts the current
3335 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3336 * called on the nearest possible occasion:
3338 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3340 * - in syscall or exception context, at the next outmost
3341 * preempt_enable(). (this might be as soon as the wake_up()'s
3344 * - in IRQ context, return from interrupt-handler to
3345 * preemptible context
3347 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3350 * - cond_resched() call
3351 * - explicit schedule() call
3352 * - return from syscall or exception to user-space
3353 * - return from interrupt-handler to user-space
3355 * WARNING: must be called with preemption disabled!
3357 static void __sched notrace __schedule(bool preempt)
3359 struct task_struct *prev, *next;
3360 unsigned long *switch_count;
3365 cpu = smp_processor_id();
3369 schedule_debug(prev);
3371 if (sched_feat(HRTICK))
3374 local_irq_disable();
3375 rcu_note_context_switch();
3378 * Make sure that signal_pending_state()->signal_pending() below
3379 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3380 * done by the caller to avoid the race with signal_wake_up().
3382 smp_mb__before_spinlock();
3383 raw_spin_lock(&rq->lock);
3384 rq_pin_lock(rq, &rf);
3386 /* Promote REQ to ACT */
3387 rq->clock_update_flags <<= 1;
3389 switch_count = &prev->nivcsw;
3390 if (!preempt && prev->state) {
3391 if (unlikely(signal_pending_state(prev->state, prev))) {
3392 prev->state = TASK_RUNNING;
3394 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3397 if (prev->in_iowait) {
3398 atomic_inc(&rq->nr_iowait);
3399 delayacct_blkio_start();
3403 * If a worker went to sleep, notify and ask workqueue
3404 * whether it wants to wake up a task to maintain
3407 if (prev->flags & PF_WQ_WORKER) {
3408 struct task_struct *to_wakeup;
3410 to_wakeup = wq_worker_sleeping(prev);
3412 try_to_wake_up_local(to_wakeup, &rf);
3415 switch_count = &prev->nvcsw;
3418 if (task_on_rq_queued(prev))
3419 update_rq_clock(rq);
3421 next = pick_next_task(rq, prev, &rf);
3422 clear_tsk_need_resched(prev);
3423 clear_preempt_need_resched();
3425 if (likely(prev != next)) {
3430 trace_sched_switch(preempt, prev, next);
3432 /* Also unlocks the rq: */
3433 rq = context_switch(rq, prev, next, &rf);
3435 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3436 rq_unpin_lock(rq, &rf);
3437 raw_spin_unlock_irq(&rq->lock);
3440 balance_callback(rq);
3443 void __noreturn do_task_dead(void)
3446 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3447 * when the following two conditions become true.
3448 * - There is race condition of mmap_sem (It is acquired by
3450 * - SMI occurs before setting TASK_RUNINNG.
3451 * (or hypervisor of virtual machine switches to other guest)
3452 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3454 * To avoid it, we have to wait for releasing tsk->pi_lock which
3455 * is held by try_to_wake_up()
3458 raw_spin_unlock_wait(¤t->pi_lock);
3460 /* Causes final put_task_struct in finish_task_switch(): */
3461 __set_current_state(TASK_DEAD);
3463 /* Tell freezer to ignore us: */
3464 current->flags |= PF_NOFREEZE;
3469 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3474 static inline void sched_submit_work(struct task_struct *tsk)
3476 if (!tsk->state || tsk_is_pi_blocked(tsk))
3479 * If we are going to sleep and we have plugged IO queued,
3480 * make sure to submit it to avoid deadlocks.
3482 if (blk_needs_flush_plug(tsk))
3483 blk_schedule_flush_plug(tsk);
3486 asmlinkage __visible void __sched schedule(void)
3488 struct task_struct *tsk = current;
3490 sched_submit_work(tsk);
3494 sched_preempt_enable_no_resched();
3495 } while (need_resched());
3497 EXPORT_SYMBOL(schedule);
3499 #ifdef CONFIG_CONTEXT_TRACKING
3500 asmlinkage __visible void __sched schedule_user(void)
3503 * If we come here after a random call to set_need_resched(),
3504 * or we have been woken up remotely but the IPI has not yet arrived,
3505 * we haven't yet exited the RCU idle mode. Do it here manually until
3506 * we find a better solution.
3508 * NB: There are buggy callers of this function. Ideally we
3509 * should warn if prev_state != CONTEXT_USER, but that will trigger
3510 * too frequently to make sense yet.
3512 enum ctx_state prev_state = exception_enter();
3514 exception_exit(prev_state);
3519 * schedule_preempt_disabled - called with preemption disabled
3521 * Returns with preemption disabled. Note: preempt_count must be 1
3523 void __sched schedule_preempt_disabled(void)
3525 sched_preempt_enable_no_resched();
3530 static void __sched notrace preempt_schedule_common(void)
3534 * Because the function tracer can trace preempt_count_sub()
3535 * and it also uses preempt_enable/disable_notrace(), if
3536 * NEED_RESCHED is set, the preempt_enable_notrace() called
3537 * by the function tracer will call this function again and
3538 * cause infinite recursion.
3540 * Preemption must be disabled here before the function
3541 * tracer can trace. Break up preempt_disable() into two
3542 * calls. One to disable preemption without fear of being
3543 * traced. The other to still record the preemption latency,
3544 * which can also be traced by the function tracer.
3546 preempt_disable_notrace();
3547 preempt_latency_start(1);
3549 preempt_latency_stop(1);
3550 preempt_enable_no_resched_notrace();
3553 * Check again in case we missed a preemption opportunity
3554 * between schedule and now.
3556 } while (need_resched());
3559 #ifdef CONFIG_PREEMPT
3561 * this is the entry point to schedule() from in-kernel preemption
3562 * off of preempt_enable. Kernel preemptions off return from interrupt
3563 * occur there and call schedule directly.
3565 asmlinkage __visible void __sched notrace preempt_schedule(void)
3568 * If there is a non-zero preempt_count or interrupts are disabled,
3569 * we do not want to preempt the current task. Just return..
3571 if (likely(!preemptible()))
3574 preempt_schedule_common();
3576 NOKPROBE_SYMBOL(preempt_schedule);
3577 EXPORT_SYMBOL(preempt_schedule);
3580 * preempt_schedule_notrace - preempt_schedule called by tracing
3582 * The tracing infrastructure uses preempt_enable_notrace to prevent
3583 * recursion and tracing preempt enabling caused by the tracing
3584 * infrastructure itself. But as tracing can happen in areas coming
3585 * from userspace or just about to enter userspace, a preempt enable
3586 * can occur before user_exit() is called. This will cause the scheduler
3587 * to be called when the system is still in usermode.
3589 * To prevent this, the preempt_enable_notrace will use this function
3590 * instead of preempt_schedule() to exit user context if needed before
3591 * calling the scheduler.
3593 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3595 enum ctx_state prev_ctx;
3597 if (likely(!preemptible()))
3602 * Because the function tracer can trace preempt_count_sub()
3603 * and it also uses preempt_enable/disable_notrace(), if
3604 * NEED_RESCHED is set, the preempt_enable_notrace() called
3605 * by the function tracer will call this function again and
3606 * cause infinite recursion.
3608 * Preemption must be disabled here before the function
3609 * tracer can trace. Break up preempt_disable() into two
3610 * calls. One to disable preemption without fear of being
3611 * traced. The other to still record the preemption latency,
3612 * which can also be traced by the function tracer.
3614 preempt_disable_notrace();
3615 preempt_latency_start(1);
3617 * Needs preempt disabled in case user_exit() is traced
3618 * and the tracer calls preempt_enable_notrace() causing
3619 * an infinite recursion.
3621 prev_ctx = exception_enter();
3623 exception_exit(prev_ctx);
3625 preempt_latency_stop(1);
3626 preempt_enable_no_resched_notrace();
3627 } while (need_resched());
3629 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3631 #endif /* CONFIG_PREEMPT */
3634 * this is the entry point to schedule() from kernel preemption
3635 * off of irq context.
3636 * Note, that this is called and return with irqs disabled. This will
3637 * protect us against recursive calling from irq.
3639 asmlinkage __visible void __sched preempt_schedule_irq(void)
3641 enum ctx_state prev_state;
3643 /* Catch callers which need to be fixed */
3644 BUG_ON(preempt_count() || !irqs_disabled());
3646 prev_state = exception_enter();
3652 local_irq_disable();
3653 sched_preempt_enable_no_resched();
3654 } while (need_resched());
3656 exception_exit(prev_state);
3659 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3662 return try_to_wake_up(curr->private, mode, wake_flags);
3664 EXPORT_SYMBOL(default_wake_function);
3666 #ifdef CONFIG_RT_MUTEXES
3669 * rt_mutex_setprio - set the current priority of a task
3671 * @prio: prio value (kernel-internal form)
3673 * This function changes the 'effective' priority of a task. It does
3674 * not touch ->normal_prio like __setscheduler().
3676 * Used by the rt_mutex code to implement priority inheritance
3677 * logic. Call site only calls if the priority of the task changed.
3679 void rt_mutex_setprio(struct task_struct *p, int prio)
3681 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3682 const struct sched_class *prev_class;
3686 BUG_ON(prio > MAX_PRIO);
3688 rq = __task_rq_lock(p, &rf);
3689 update_rq_clock(rq);
3692 * Idle task boosting is a nono in general. There is one
3693 * exception, when PREEMPT_RT and NOHZ is active:
3695 * The idle task calls get_next_timer_interrupt() and holds
3696 * the timer wheel base->lock on the CPU and another CPU wants
3697 * to access the timer (probably to cancel it). We can safely
3698 * ignore the boosting request, as the idle CPU runs this code
3699 * with interrupts disabled and will complete the lock
3700 * protected section without being interrupted. So there is no
3701 * real need to boost.
3703 if (unlikely(p == rq->idle)) {
3704 WARN_ON(p != rq->curr);
3705 WARN_ON(p->pi_blocked_on);
3709 trace_sched_pi_setprio(p, prio);
3712 if (oldprio == prio)
3713 queue_flag &= ~DEQUEUE_MOVE;
3715 prev_class = p->sched_class;
3716 queued = task_on_rq_queued(p);
3717 running = task_current(rq, p);
3719 dequeue_task(rq, p, queue_flag);
3721 put_prev_task(rq, p);
3724 * Boosting condition are:
3725 * 1. -rt task is running and holds mutex A
3726 * --> -dl task blocks on mutex A
3728 * 2. -dl task is running and holds mutex A
3729 * --> -dl task blocks on mutex A and could preempt the
3732 if (dl_prio(prio)) {
3733 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3734 if (!dl_prio(p->normal_prio) ||
3735 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3736 p->dl.dl_boosted = 1;
3737 queue_flag |= ENQUEUE_REPLENISH;
3739 p->dl.dl_boosted = 0;
3740 p->sched_class = &dl_sched_class;
3741 } else if (rt_prio(prio)) {
3742 if (dl_prio(oldprio))
3743 p->dl.dl_boosted = 0;
3745 queue_flag |= ENQUEUE_HEAD;
3746 p->sched_class = &rt_sched_class;
3748 if (dl_prio(oldprio))
3749 p->dl.dl_boosted = 0;
3750 if (rt_prio(oldprio))
3752 p->sched_class = &fair_sched_class;
3758 enqueue_task(rq, p, queue_flag);
3760 set_curr_task(rq, p);
3762 check_class_changed(rq, p, prev_class, oldprio);
3764 /* Avoid rq from going away on us: */
3766 __task_rq_unlock(rq, &rf);
3768 balance_callback(rq);
3773 void set_user_nice(struct task_struct *p, long nice)
3775 bool queued, running;
3776 int old_prio, delta;
3780 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3783 * We have to be careful, if called from sys_setpriority(),
3784 * the task might be in the middle of scheduling on another CPU.
3786 rq = task_rq_lock(p, &rf);
3787 update_rq_clock(rq);
3790 * The RT priorities are set via sched_setscheduler(), but we still
3791 * allow the 'normal' nice value to be set - but as expected
3792 * it wont have any effect on scheduling until the task is
3793 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3795 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3796 p->static_prio = NICE_TO_PRIO(nice);
3799 queued = task_on_rq_queued(p);
3800 running = task_current(rq, p);
3802 dequeue_task(rq, p, DEQUEUE_SAVE);
3804 put_prev_task(rq, p);
3806 p->static_prio = NICE_TO_PRIO(nice);
3809 p->prio = effective_prio(p);
3810 delta = p->prio - old_prio;
3813 enqueue_task(rq, p, ENQUEUE_RESTORE);
3815 * If the task increased its priority or is running and
3816 * lowered its priority, then reschedule its CPU:
3818 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3822 set_curr_task(rq, p);
3824 task_rq_unlock(rq, p, &rf);
3826 EXPORT_SYMBOL(set_user_nice);
3829 * can_nice - check if a task can reduce its nice value
3833 int can_nice(const struct task_struct *p, const int nice)
3835 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3836 int nice_rlim = nice_to_rlimit(nice);
3838 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3839 capable(CAP_SYS_NICE));
3842 #ifdef __ARCH_WANT_SYS_NICE
3845 * sys_nice - change the priority of the current process.
3846 * @increment: priority increment
3848 * sys_setpriority is a more generic, but much slower function that
3849 * does similar things.
3851 SYSCALL_DEFINE1(nice, int, increment)
3856 * Setpriority might change our priority at the same moment.
3857 * We don't have to worry. Conceptually one call occurs first
3858 * and we have a single winner.
3860 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3861 nice = task_nice(current) + increment;
3863 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3864 if (increment < 0 && !can_nice(current, nice))
3867 retval = security_task_setnice(current, nice);
3871 set_user_nice(current, nice);
3878 * task_prio - return the priority value of a given task.
3879 * @p: the task in question.
3881 * Return: The priority value as seen by users in /proc.
3882 * RT tasks are offset by -200. Normal tasks are centered
3883 * around 0, value goes from -16 to +15.
3885 int task_prio(const struct task_struct *p)
3887 return p->prio - MAX_RT_PRIO;
3891 * idle_cpu - is a given CPU idle currently?
3892 * @cpu: the processor in question.
3894 * Return: 1 if the CPU is currently idle. 0 otherwise.
3896 int idle_cpu(int cpu)
3898 struct rq *rq = cpu_rq(cpu);
3900 if (rq->curr != rq->idle)
3907 if (!llist_empty(&rq->wake_list))
3915 * idle_task - return the idle task for a given CPU.
3916 * @cpu: the processor in question.
3918 * Return: The idle task for the CPU @cpu.
3920 struct task_struct *idle_task(int cpu)
3922 return cpu_rq(cpu)->idle;
3926 * find_process_by_pid - find a process with a matching PID value.
3927 * @pid: the pid in question.
3929 * The task of @pid, if found. %NULL otherwise.
3931 static struct task_struct *find_process_by_pid(pid_t pid)
3933 return pid ? find_task_by_vpid(pid) : current;
3937 * This function initializes the sched_dl_entity of a newly becoming
3938 * SCHED_DEADLINE task.
3940 * Only the static values are considered here, the actual runtime and the
3941 * absolute deadline will be properly calculated when the task is enqueued
3942 * for the first time with its new policy.
3945 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3947 struct sched_dl_entity *dl_se = &p->dl;
3949 dl_se->dl_runtime = attr->sched_runtime;
3950 dl_se->dl_deadline = attr->sched_deadline;
3951 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3952 dl_se->flags = attr->sched_flags;
3953 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3956 * Changing the parameters of a task is 'tricky' and we're not doing
3957 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3959 * What we SHOULD do is delay the bandwidth release until the 0-lag
3960 * point. This would include retaining the task_struct until that time
3961 * and change dl_overflow() to not immediately decrement the current
3964 * Instead we retain the current runtime/deadline and let the new
3965 * parameters take effect after the current reservation period lapses.
3966 * This is safe (albeit pessimistic) because the 0-lag point is always
3967 * before the current scheduling deadline.
3969 * We can still have temporary overloads because we do not delay the
3970 * change in bandwidth until that time; so admission control is
3971 * not on the safe side. It does however guarantee tasks will never
3972 * consume more than promised.
3977 * sched_setparam() passes in -1 for its policy, to let the functions
3978 * it calls know not to change it.
3980 #define SETPARAM_POLICY -1
3982 static void __setscheduler_params(struct task_struct *p,
3983 const struct sched_attr *attr)
3985 int policy = attr->sched_policy;
3987 if (policy == SETPARAM_POLICY)
3992 if (dl_policy(policy))
3993 __setparam_dl(p, attr);
3994 else if (fair_policy(policy))
3995 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3998 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3999 * !rt_policy. Always setting this ensures that things like
4000 * getparam()/getattr() don't report silly values for !rt tasks.
4002 p->rt_priority = attr->sched_priority;
4003 p->normal_prio = normal_prio(p);
4007 /* Actually do priority change: must hold pi & rq lock. */
4008 static void __setscheduler(struct rq *rq, struct task_struct *p,
4009 const struct sched_attr *attr, bool keep_boost)
4011 __setscheduler_params(p, attr);
4014 * Keep a potential priority boosting if called from
4015 * sched_setscheduler().
4018 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
4020 p->prio = normal_prio(p);
4022 if (dl_prio(p->prio))
4023 p->sched_class = &dl_sched_class;
4024 else if (rt_prio(p->prio))
4025 p->sched_class = &rt_sched_class;
4027 p->sched_class = &fair_sched_class;
4031 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
4033 struct sched_dl_entity *dl_se = &p->dl;
4035 attr->sched_priority = p->rt_priority;
4036 attr->sched_runtime = dl_se->dl_runtime;
4037 attr->sched_deadline = dl_se->dl_deadline;
4038 attr->sched_period = dl_se->dl_period;
4039 attr->sched_flags = dl_se->flags;
4043 * This function validates the new parameters of a -deadline task.
4044 * We ask for the deadline not being zero, and greater or equal
4045 * than the runtime, as well as the period of being zero or
4046 * greater than deadline. Furthermore, we have to be sure that
4047 * user parameters are above the internal resolution of 1us (we
4048 * check sched_runtime only since it is always the smaller one) and
4049 * below 2^63 ns (we have to check both sched_deadline and
4050 * sched_period, as the latter can be zero).
4053 __checkparam_dl(const struct sched_attr *attr)
4056 if (attr->sched_deadline == 0)
4060 * Since we truncate DL_SCALE bits, make sure we're at least
4063 if (attr->sched_runtime < (1ULL << DL_SCALE))
4067 * Since we use the MSB for wrap-around and sign issues, make
4068 * sure it's not set (mind that period can be equal to zero).
4070 if (attr->sched_deadline & (1ULL << 63) ||
4071 attr->sched_period & (1ULL << 63))
4074 /* runtime <= deadline <= period (if period != 0) */
4075 if ((attr->sched_period != 0 &&
4076 attr->sched_period < attr->sched_deadline) ||
4077 attr->sched_deadline < attr->sched_runtime)
4084 * Check the target process has a UID that matches the current process's:
4086 static bool check_same_owner(struct task_struct *p)
4088 const struct cred *cred = current_cred(), *pcred;
4092 pcred = __task_cred(p);
4093 match = (uid_eq(cred->euid, pcred->euid) ||
4094 uid_eq(cred->euid, pcred->uid));
4099 static bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
4101 struct sched_dl_entity *dl_se = &p->dl;
4103 if (dl_se->dl_runtime != attr->sched_runtime ||
4104 dl_se->dl_deadline != attr->sched_deadline ||
4105 dl_se->dl_period != attr->sched_period ||
4106 dl_se->flags != attr->sched_flags)
4112 static int __sched_setscheduler(struct task_struct *p,
4113 const struct sched_attr *attr,
4116 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4117 MAX_RT_PRIO - 1 - attr->sched_priority;
4118 int retval, oldprio, oldpolicy = -1, queued, running;
4119 int new_effective_prio, policy = attr->sched_policy;
4120 const struct sched_class *prev_class;
4123 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
4126 /* May grab non-irq protected spin_locks: */
4127 BUG_ON(in_interrupt());
4129 /* Double check policy once rq lock held: */
4131 reset_on_fork = p->sched_reset_on_fork;
4132 policy = oldpolicy = p->policy;
4134 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4136 if (!valid_policy(policy))
4140 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4144 * Valid priorities for SCHED_FIFO and SCHED_RR are
4145 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4146 * SCHED_BATCH and SCHED_IDLE is 0.
4148 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4149 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4151 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4152 (rt_policy(policy) != (attr->sched_priority != 0)))
4156 * Allow unprivileged RT tasks to decrease priority:
4158 if (user && !capable(CAP_SYS_NICE)) {
4159 if (fair_policy(policy)) {
4160 if (attr->sched_nice < task_nice(p) &&
4161 !can_nice(p, attr->sched_nice))
4165 if (rt_policy(policy)) {
4166 unsigned long rlim_rtprio =
4167 task_rlimit(p, RLIMIT_RTPRIO);
4169 /* Can't set/change the rt policy: */
4170 if (policy != p->policy && !rlim_rtprio)
4173 /* Can't increase priority: */
4174 if (attr->sched_priority > p->rt_priority &&
4175 attr->sched_priority > rlim_rtprio)
4180 * Can't set/change SCHED_DEADLINE policy at all for now
4181 * (safest behavior); in the future we would like to allow
4182 * unprivileged DL tasks to increase their relative deadline
4183 * or reduce their runtime (both ways reducing utilization)
4185 if (dl_policy(policy))
4189 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4190 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4192 if (idle_policy(p->policy) && !idle_policy(policy)) {
4193 if (!can_nice(p, task_nice(p)))
4197 /* Can't change other user's priorities: */
4198 if (!check_same_owner(p))
4201 /* Normal users shall not reset the sched_reset_on_fork flag: */
4202 if (p->sched_reset_on_fork && !reset_on_fork)
4207 retval = security_task_setscheduler(p);
4213 * Make sure no PI-waiters arrive (or leave) while we are
4214 * changing the priority of the task:
4216 * To be able to change p->policy safely, the appropriate
4217 * runqueue lock must be held.
4219 rq = task_rq_lock(p, &rf);
4220 update_rq_clock(rq);
4223 * Changing the policy of the stop threads its a very bad idea:
4225 if (p == rq->stop) {
4226 task_rq_unlock(rq, p, &rf);
4231 * If not changing anything there's no need to proceed further,
4232 * but store a possible modification of reset_on_fork.
4234 if (unlikely(policy == p->policy)) {
4235 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4237 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4239 if (dl_policy(policy) && dl_param_changed(p, attr))
4242 p->sched_reset_on_fork = reset_on_fork;
4243 task_rq_unlock(rq, p, &rf);
4249 #ifdef CONFIG_RT_GROUP_SCHED
4251 * Do not allow realtime tasks into groups that have no runtime
4254 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4255 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4256 !task_group_is_autogroup(task_group(p))) {
4257 task_rq_unlock(rq, p, &rf);
4262 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4263 cpumask_t *span = rq->rd->span;
4266 * Don't allow tasks with an affinity mask smaller than
4267 * the entire root_domain to become SCHED_DEADLINE. We
4268 * will also fail if there's no bandwidth available.
4270 if (!cpumask_subset(span, &p->cpus_allowed) ||
4271 rq->rd->dl_bw.bw == 0) {
4272 task_rq_unlock(rq, p, &rf);
4279 /* Re-check policy now with rq lock held: */
4280 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4281 policy = oldpolicy = -1;
4282 task_rq_unlock(rq, p, &rf);
4287 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4288 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4291 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4292 task_rq_unlock(rq, p, &rf);
4296 p->sched_reset_on_fork = reset_on_fork;
4301 * Take priority boosted tasks into account. If the new
4302 * effective priority is unchanged, we just store the new
4303 * normal parameters and do not touch the scheduler class and
4304 * the runqueue. This will be done when the task deboost
4307 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4308 if (new_effective_prio == oldprio)
4309 queue_flags &= ~DEQUEUE_MOVE;
4312 queued = task_on_rq_queued(p);
4313 running = task_current(rq, p);
4315 dequeue_task(rq, p, queue_flags);
4317 put_prev_task(rq, p);
4319 prev_class = p->sched_class;
4320 __setscheduler(rq, p, attr, pi);
4324 * We enqueue to tail when the priority of a task is
4325 * increased (user space view).
4327 if (oldprio < p->prio)
4328 queue_flags |= ENQUEUE_HEAD;
4330 enqueue_task(rq, p, queue_flags);
4333 set_curr_task(rq, p);
4335 check_class_changed(rq, p, prev_class, oldprio);
4337 /* Avoid rq from going away on us: */
4339 task_rq_unlock(rq, p, &rf);
4342 rt_mutex_adjust_pi(p);
4344 /* Run balance callbacks after we've adjusted the PI chain: */
4345 balance_callback(rq);
4351 static int _sched_setscheduler(struct task_struct *p, int policy,
4352 const struct sched_param *param, bool check)
4354 struct sched_attr attr = {
4355 .sched_policy = policy,
4356 .sched_priority = param->sched_priority,
4357 .sched_nice = PRIO_TO_NICE(p->static_prio),
4360 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4361 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4362 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4363 policy &= ~SCHED_RESET_ON_FORK;
4364 attr.sched_policy = policy;
4367 return __sched_setscheduler(p, &attr, check, true);
4370 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4371 * @p: the task in question.
4372 * @policy: new policy.
4373 * @param: structure containing the new RT priority.
4375 * Return: 0 on success. An error code otherwise.
4377 * NOTE that the task may be already dead.
4379 int sched_setscheduler(struct task_struct *p, int policy,
4380 const struct sched_param *param)
4382 return _sched_setscheduler(p, policy, param, true);
4384 EXPORT_SYMBOL_GPL(sched_setscheduler);
4386 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4388 return __sched_setscheduler(p, attr, true, true);
4390 EXPORT_SYMBOL_GPL(sched_setattr);
4393 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4394 * @p: the task in question.
4395 * @policy: new policy.
4396 * @param: structure containing the new RT priority.
4398 * Just like sched_setscheduler, only don't bother checking if the
4399 * current context has permission. For example, this is needed in
4400 * stop_machine(): we create temporary high priority worker threads,
4401 * but our caller might not have that capability.
4403 * Return: 0 on success. An error code otherwise.
4405 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4406 const struct sched_param *param)
4408 return _sched_setscheduler(p, policy, param, false);
4410 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4413 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4415 struct sched_param lparam;
4416 struct task_struct *p;
4419 if (!param || pid < 0)
4421 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4426 p = find_process_by_pid(pid);
4428 retval = sched_setscheduler(p, policy, &lparam);
4435 * Mimics kernel/events/core.c perf_copy_attr().
4437 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4442 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4445 /* Zero the full structure, so that a short copy will be nice: */
4446 memset(attr, 0, sizeof(*attr));
4448 ret = get_user(size, &uattr->size);
4452 /* Bail out on silly large: */
4453 if (size > PAGE_SIZE)
4456 /* ABI compatibility quirk: */
4458 size = SCHED_ATTR_SIZE_VER0;
4460 if (size < SCHED_ATTR_SIZE_VER0)
4464 * If we're handed a bigger struct than we know of,
4465 * ensure all the unknown bits are 0 - i.e. new
4466 * user-space does not rely on any kernel feature
4467 * extensions we dont know about yet.
4469 if (size > sizeof(*attr)) {
4470 unsigned char __user *addr;
4471 unsigned char __user *end;
4474 addr = (void __user *)uattr + sizeof(*attr);
4475 end = (void __user *)uattr + size;
4477 for (; addr < end; addr++) {
4478 ret = get_user(val, addr);
4484 size = sizeof(*attr);
4487 ret = copy_from_user(attr, uattr, size);
4492 * XXX: Do we want to be lenient like existing syscalls; or do we want
4493 * to be strict and return an error on out-of-bounds values?
4495 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4500 put_user(sizeof(*attr), &uattr->size);
4505 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4506 * @pid: the pid in question.
4507 * @policy: new policy.
4508 * @param: structure containing the new RT priority.
4510 * Return: 0 on success. An error code otherwise.
4512 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4517 return do_sched_setscheduler(pid, policy, param);
4521 * sys_sched_setparam - set/change the RT priority of a thread
4522 * @pid: the pid in question.
4523 * @param: structure containing the new RT priority.
4525 * Return: 0 on success. An error code otherwise.
4527 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4529 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4533 * sys_sched_setattr - same as above, but with extended sched_attr
4534 * @pid: the pid in question.
4535 * @uattr: structure containing the extended parameters.
4536 * @flags: for future extension.
4538 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4539 unsigned int, flags)
4541 struct sched_attr attr;
4542 struct task_struct *p;
4545 if (!uattr || pid < 0 || flags)
4548 retval = sched_copy_attr(uattr, &attr);
4552 if ((int)attr.sched_policy < 0)
4557 p = find_process_by_pid(pid);
4559 retval = sched_setattr(p, &attr);
4566 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4567 * @pid: the pid in question.
4569 * Return: On success, the policy of the thread. Otherwise, a negative error
4572 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4574 struct task_struct *p;
4582 p = find_process_by_pid(pid);
4584 retval = security_task_getscheduler(p);
4587 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4594 * sys_sched_getparam - get the RT priority of a thread
4595 * @pid: the pid in question.
4596 * @param: structure containing the RT priority.
4598 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4601 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4603 struct sched_param lp = { .sched_priority = 0 };
4604 struct task_struct *p;
4607 if (!param || pid < 0)
4611 p = find_process_by_pid(pid);
4616 retval = security_task_getscheduler(p);
4620 if (task_has_rt_policy(p))
4621 lp.sched_priority = p->rt_priority;
4625 * This one might sleep, we cannot do it with a spinlock held ...
4627 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4636 static int sched_read_attr(struct sched_attr __user *uattr,
4637 struct sched_attr *attr,
4642 if (!access_ok(VERIFY_WRITE, uattr, usize))
4646 * If we're handed a smaller struct than we know of,
4647 * ensure all the unknown bits are 0 - i.e. old
4648 * user-space does not get uncomplete information.
4650 if (usize < sizeof(*attr)) {
4651 unsigned char *addr;
4654 addr = (void *)attr + usize;
4655 end = (void *)attr + sizeof(*attr);
4657 for (; addr < end; addr++) {
4665 ret = copy_to_user(uattr, attr, attr->size);
4673 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4674 * @pid: the pid in question.
4675 * @uattr: structure containing the extended parameters.
4676 * @size: sizeof(attr) for fwd/bwd comp.
4677 * @flags: for future extension.
4679 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4680 unsigned int, size, unsigned int, flags)
4682 struct sched_attr attr = {
4683 .size = sizeof(struct sched_attr),
4685 struct task_struct *p;
4688 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4689 size < SCHED_ATTR_SIZE_VER0 || flags)
4693 p = find_process_by_pid(pid);
4698 retval = security_task_getscheduler(p);
4702 attr.sched_policy = p->policy;
4703 if (p->sched_reset_on_fork)
4704 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4705 if (task_has_dl_policy(p))
4706 __getparam_dl(p, &attr);
4707 else if (task_has_rt_policy(p))
4708 attr.sched_priority = p->rt_priority;
4710 attr.sched_nice = task_nice(p);
4714 retval = sched_read_attr(uattr, &attr, size);
4722 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4724 cpumask_var_t cpus_allowed, new_mask;
4725 struct task_struct *p;
4730 p = find_process_by_pid(pid);
4736 /* Prevent p going away */
4740 if (p->flags & PF_NO_SETAFFINITY) {
4744 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4748 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4750 goto out_free_cpus_allowed;
4753 if (!check_same_owner(p)) {
4755 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4757 goto out_free_new_mask;
4762 retval = security_task_setscheduler(p);
4764 goto out_free_new_mask;
4767 cpuset_cpus_allowed(p, cpus_allowed);
4768 cpumask_and(new_mask, in_mask, cpus_allowed);
4771 * Since bandwidth control happens on root_domain basis,
4772 * if admission test is enabled, we only admit -deadline
4773 * tasks allowed to run on all the CPUs in the task's
4777 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4779 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4782 goto out_free_new_mask;
4788 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4791 cpuset_cpus_allowed(p, cpus_allowed);
4792 if (!cpumask_subset(new_mask, cpus_allowed)) {
4794 * We must have raced with a concurrent cpuset
4795 * update. Just reset the cpus_allowed to the
4796 * cpuset's cpus_allowed
4798 cpumask_copy(new_mask, cpus_allowed);
4803 free_cpumask_var(new_mask);
4804 out_free_cpus_allowed:
4805 free_cpumask_var(cpus_allowed);
4811 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4812 struct cpumask *new_mask)
4814 if (len < cpumask_size())
4815 cpumask_clear(new_mask);
4816 else if (len > cpumask_size())
4817 len = cpumask_size();
4819 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4823 * sys_sched_setaffinity - set the CPU affinity of a process
4824 * @pid: pid of the process
4825 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4826 * @user_mask_ptr: user-space pointer to the new CPU mask
4828 * Return: 0 on success. An error code otherwise.
4830 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4831 unsigned long __user *, user_mask_ptr)
4833 cpumask_var_t new_mask;
4836 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4839 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4841 retval = sched_setaffinity(pid, new_mask);
4842 free_cpumask_var(new_mask);
4846 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4848 struct task_struct *p;
4849 unsigned long flags;
4855 p = find_process_by_pid(pid);
4859 retval = security_task_getscheduler(p);
4863 raw_spin_lock_irqsave(&p->pi_lock, flags);
4864 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4865 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4874 * sys_sched_getaffinity - get the CPU affinity of a process
4875 * @pid: pid of the process
4876 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4877 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4879 * Return: size of CPU mask copied to user_mask_ptr on success. An
4880 * error code otherwise.
4882 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4883 unsigned long __user *, user_mask_ptr)
4888 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4890 if (len & (sizeof(unsigned long)-1))
4893 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4896 ret = sched_getaffinity(pid, mask);
4898 size_t retlen = min_t(size_t, len, cpumask_size());
4900 if (copy_to_user(user_mask_ptr, mask, retlen))
4905 free_cpumask_var(mask);
4911 * sys_sched_yield - yield the current processor to other threads.
4913 * This function yields the current CPU to other tasks. If there are no
4914 * other threads running on this CPU then this function will return.
4918 SYSCALL_DEFINE0(sched_yield)
4920 struct rq *rq = this_rq_lock();
4922 schedstat_inc(rq->yld_count);
4923 current->sched_class->yield_task(rq);
4926 * Since we are going to call schedule() anyway, there's
4927 * no need to preempt or enable interrupts:
4929 __release(rq->lock);
4930 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4931 do_raw_spin_unlock(&rq->lock);
4932 sched_preempt_enable_no_resched();
4939 #ifndef CONFIG_PREEMPT
4940 int __sched _cond_resched(void)
4942 if (should_resched(0)) {
4943 preempt_schedule_common();
4948 EXPORT_SYMBOL(_cond_resched);
4952 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4953 * call schedule, and on return reacquire the lock.
4955 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4956 * operations here to prevent schedule() from being called twice (once via
4957 * spin_unlock(), once by hand).
4959 int __cond_resched_lock(spinlock_t *lock)
4961 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4964 lockdep_assert_held(lock);
4966 if (spin_needbreak(lock) || resched) {
4969 preempt_schedule_common();
4977 EXPORT_SYMBOL(__cond_resched_lock);
4979 int __sched __cond_resched_softirq(void)
4981 BUG_ON(!in_softirq());
4983 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4985 preempt_schedule_common();
4991 EXPORT_SYMBOL(__cond_resched_softirq);
4994 * yield - yield the current processor to other threads.
4996 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4998 * The scheduler is at all times free to pick the calling task as the most
4999 * eligible task to run, if removing the yield() call from your code breaks
5000 * it, its already broken.
5002 * Typical broken usage is:
5007 * where one assumes that yield() will let 'the other' process run that will
5008 * make event true. If the current task is a SCHED_FIFO task that will never
5009 * happen. Never use yield() as a progress guarantee!!
5011 * If you want to use yield() to wait for something, use wait_event().
5012 * If you want to use yield() to be 'nice' for others, use cond_resched().
5013 * If you still want to use yield(), do not!
5015 void __sched yield(void)
5017 set_current_state(TASK_RUNNING);
5020 EXPORT_SYMBOL(yield);
5023 * yield_to - yield the current processor to another thread in
5024 * your thread group, or accelerate that thread toward the
5025 * processor it's on.
5027 * @preempt: whether task preemption is allowed or not
5029 * It's the caller's job to ensure that the target task struct
5030 * can't go away on us before we can do any checks.
5033 * true (>0) if we indeed boosted the target task.
5034 * false (0) if we failed to boost the target.
5035 * -ESRCH if there's no task to yield to.
5037 int __sched yield_to(struct task_struct *p, bool preempt)
5039 struct task_struct *curr = current;
5040 struct rq *rq, *p_rq;
5041 unsigned long flags;
5044 local_irq_save(flags);
5050 * If we're the only runnable task on the rq and target rq also
5051 * has only one task, there's absolutely no point in yielding.
5053 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5058 double_rq_lock(rq, p_rq);
5059 if (task_rq(p) != p_rq) {
5060 double_rq_unlock(rq, p_rq);
5064 if (!curr->sched_class->yield_to_task)
5067 if (curr->sched_class != p->sched_class)
5070 if (task_running(p_rq, p) || p->state)
5073 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5075 schedstat_inc(rq->yld_count);
5077 * Make p's CPU reschedule; pick_next_entity takes care of
5080 if (preempt && rq != p_rq)
5085 double_rq_unlock(rq, p_rq);
5087 local_irq_restore(flags);
5094 EXPORT_SYMBOL_GPL(yield_to);
5096 int io_schedule_prepare(void)
5098 int old_iowait = current->in_iowait;
5100 current->in_iowait = 1;
5101 blk_schedule_flush_plug(current);
5106 void io_schedule_finish(int token)
5108 current->in_iowait = token;
5112 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5113 * that process accounting knows that this is a task in IO wait state.
5115 long __sched io_schedule_timeout(long timeout)
5120 token = io_schedule_prepare();
5121 ret = schedule_timeout(timeout);
5122 io_schedule_finish(token);
5126 EXPORT_SYMBOL(io_schedule_timeout);
5128 void io_schedule(void)
5132 token = io_schedule_prepare();
5134 io_schedule_finish(token);
5136 EXPORT_SYMBOL(io_schedule);
5139 * sys_sched_get_priority_max - return maximum RT priority.
5140 * @policy: scheduling class.
5142 * Return: On success, this syscall returns the maximum
5143 * rt_priority that can be used by a given scheduling class.
5144 * On failure, a negative error code is returned.
5146 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5153 ret = MAX_USER_RT_PRIO-1;
5155 case SCHED_DEADLINE:
5166 * sys_sched_get_priority_min - return minimum RT priority.
5167 * @policy: scheduling class.
5169 * Return: On success, this syscall returns the minimum
5170 * rt_priority that can be used by a given scheduling class.
5171 * On failure, a negative error code is returned.
5173 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5182 case SCHED_DEADLINE:
5192 * sys_sched_rr_get_interval - return the default timeslice of a process.
5193 * @pid: pid of the process.
5194 * @interval: userspace pointer to the timeslice value.
5196 * this syscall writes the default timeslice value of a given process
5197 * into the user-space timespec buffer. A value of '0' means infinity.
5199 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5202 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5203 struct timespec __user *, interval)
5205 struct task_struct *p;
5206 unsigned int time_slice;
5217 p = find_process_by_pid(pid);
5221 retval = security_task_getscheduler(p);
5225 rq = task_rq_lock(p, &rf);
5227 if (p->sched_class->get_rr_interval)
5228 time_slice = p->sched_class->get_rr_interval(rq, p);
5229 task_rq_unlock(rq, p, &rf);
5232 jiffies_to_timespec(time_slice, &t);
5233 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5241 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5243 void sched_show_task(struct task_struct *p)
5245 unsigned long free = 0;
5247 unsigned long state = p->state;
5249 /* Make sure the string lines up properly with the number of task states: */
5250 BUILD_BUG_ON(sizeof(TASK_STATE_TO_CHAR_STR)-1 != ilog2(TASK_STATE_MAX)+1);
5252 if (!try_get_task_stack(p))
5255 state = __ffs(state) + 1;
5256 printk(KERN_INFO "%-15.15s %c", p->comm,
5257 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5258 if (state == TASK_RUNNING)
5259 printk(KERN_CONT " running task ");
5260 #ifdef CONFIG_DEBUG_STACK_USAGE
5261 free = stack_not_used(p);
5266 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5268 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5269 task_pid_nr(p), ppid,
5270 (unsigned long)task_thread_info(p)->flags);
5272 print_worker_info(KERN_INFO, p);
5273 show_stack(p, NULL);
5277 void show_state_filter(unsigned long state_filter)
5279 struct task_struct *g, *p;
5281 #if BITS_PER_LONG == 32
5283 " task PC stack pid father\n");
5286 " task PC stack pid father\n");
5289 for_each_process_thread(g, p) {
5291 * reset the NMI-timeout, listing all files on a slow
5292 * console might take a lot of time:
5293 * Also, reset softlockup watchdogs on all CPUs, because
5294 * another CPU might be blocked waiting for us to process
5297 touch_nmi_watchdog();
5298 touch_all_softlockup_watchdogs();
5299 if (!state_filter || (p->state & state_filter))
5303 #ifdef CONFIG_SCHED_DEBUG
5305 sysrq_sched_debug_show();
5309 * Only show locks if all tasks are dumped:
5312 debug_show_all_locks();
5315 void init_idle_bootup_task(struct task_struct *idle)
5317 idle->sched_class = &idle_sched_class;
5321 * init_idle - set up an idle thread for a given CPU
5322 * @idle: task in question
5323 * @cpu: CPU the idle task belongs to
5325 * NOTE: this function does not set the idle thread's NEED_RESCHED
5326 * flag, to make booting more robust.
5328 void init_idle(struct task_struct *idle, int cpu)
5330 struct rq *rq = cpu_rq(cpu);
5331 unsigned long flags;
5333 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5334 raw_spin_lock(&rq->lock);
5336 __sched_fork(0, idle);
5337 idle->state = TASK_RUNNING;
5338 idle->se.exec_start = sched_clock();
5339 idle->flags |= PF_IDLE;
5341 kasan_unpoison_task_stack(idle);
5345 * Its possible that init_idle() gets called multiple times on a task,
5346 * in that case do_set_cpus_allowed() will not do the right thing.
5348 * And since this is boot we can forgo the serialization.
5350 set_cpus_allowed_common(idle, cpumask_of(cpu));
5353 * We're having a chicken and egg problem, even though we are
5354 * holding rq->lock, the CPU isn't yet set to this CPU so the
5355 * lockdep check in task_group() will fail.
5357 * Similar case to sched_fork(). / Alternatively we could
5358 * use task_rq_lock() here and obtain the other rq->lock.
5363 __set_task_cpu(idle, cpu);
5366 rq->curr = rq->idle = idle;
5367 idle->on_rq = TASK_ON_RQ_QUEUED;
5371 raw_spin_unlock(&rq->lock);
5372 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5374 /* Set the preempt count _outside_ the spinlocks! */
5375 init_idle_preempt_count(idle, cpu);
5378 * The idle tasks have their own, simple scheduling class:
5380 idle->sched_class = &idle_sched_class;
5381 ftrace_graph_init_idle_task(idle, cpu);
5382 vtime_init_idle(idle, cpu);
5384 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5388 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5389 const struct cpumask *trial)
5391 int ret = 1, trial_cpus;
5392 struct dl_bw *cur_dl_b;
5393 unsigned long flags;
5395 if (!cpumask_weight(cur))
5398 rcu_read_lock_sched();
5399 cur_dl_b = dl_bw_of(cpumask_any(cur));
5400 trial_cpus = cpumask_weight(trial);
5402 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5403 if (cur_dl_b->bw != -1 &&
5404 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5406 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5407 rcu_read_unlock_sched();
5412 int task_can_attach(struct task_struct *p,
5413 const struct cpumask *cs_cpus_allowed)
5418 * Kthreads which disallow setaffinity shouldn't be moved
5419 * to a new cpuset; we don't want to change their CPU
5420 * affinity and isolating such threads by their set of
5421 * allowed nodes is unnecessary. Thus, cpusets are not
5422 * applicable for such threads. This prevents checking for
5423 * success of set_cpus_allowed_ptr() on all attached tasks
5424 * before cpus_allowed may be changed.
5426 if (p->flags & PF_NO_SETAFFINITY) {
5432 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5434 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5439 unsigned long flags;
5441 rcu_read_lock_sched();
5442 dl_b = dl_bw_of(dest_cpu);
5443 raw_spin_lock_irqsave(&dl_b->lock, flags);
5444 cpus = dl_bw_cpus(dest_cpu);
5445 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5450 * We reserve space for this task in the destination
5451 * root_domain, as we can't fail after this point.
5452 * We will free resources in the source root_domain
5453 * later on (see set_cpus_allowed_dl()).
5455 __dl_add(dl_b, p->dl.dl_bw);
5457 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5458 rcu_read_unlock_sched();
5468 bool sched_smp_initialized __read_mostly;
5470 #ifdef CONFIG_NUMA_BALANCING
5471 /* Migrate current task p to target_cpu */
5472 int migrate_task_to(struct task_struct *p, int target_cpu)
5474 struct migration_arg arg = { p, target_cpu };
5475 int curr_cpu = task_cpu(p);
5477 if (curr_cpu == target_cpu)
5480 if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5483 /* TODO: This is not properly updating schedstats */
5485 trace_sched_move_numa(p, curr_cpu, target_cpu);
5486 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5490 * Requeue a task on a given node and accurately track the number of NUMA
5491 * tasks on the runqueues
5493 void sched_setnuma(struct task_struct *p, int nid)
5495 bool queued, running;
5499 rq = task_rq_lock(p, &rf);
5500 queued = task_on_rq_queued(p);
5501 running = task_current(rq, p);
5504 dequeue_task(rq, p, DEQUEUE_SAVE);
5506 put_prev_task(rq, p);
5508 p->numa_preferred_nid = nid;
5511 enqueue_task(rq, p, ENQUEUE_RESTORE);
5513 set_curr_task(rq, p);
5514 task_rq_unlock(rq, p, &rf);
5516 #endif /* CONFIG_NUMA_BALANCING */
5518 #ifdef CONFIG_HOTPLUG_CPU
5520 * Ensure that the idle task is using init_mm right before its CPU goes
5523 void idle_task_exit(void)
5525 struct mm_struct *mm = current->active_mm;
5527 BUG_ON(cpu_online(smp_processor_id()));
5529 if (mm != &init_mm) {
5530 switch_mm_irqs_off(mm, &init_mm, current);
5531 finish_arch_post_lock_switch();
5537 * Since this CPU is going 'away' for a while, fold any nr_active delta
5538 * we might have. Assumes we're called after migrate_tasks() so that the
5539 * nr_active count is stable. We need to take the teardown thread which
5540 * is calling this into account, so we hand in adjust = 1 to the load
5543 * Also see the comment "Global load-average calculations".
5545 static void calc_load_migrate(struct rq *rq)
5547 long delta = calc_load_fold_active(rq, 1);
5549 atomic_long_add(delta, &calc_load_tasks);
5552 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5556 static const struct sched_class fake_sched_class = {
5557 .put_prev_task = put_prev_task_fake,
5560 static struct task_struct fake_task = {
5562 * Avoid pull_{rt,dl}_task()
5564 .prio = MAX_PRIO + 1,
5565 .sched_class = &fake_sched_class,
5569 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5570 * try_to_wake_up()->select_task_rq().
5572 * Called with rq->lock held even though we'er in stop_machine() and
5573 * there's no concurrency possible, we hold the required locks anyway
5574 * because of lock validation efforts.
5576 static void migrate_tasks(struct rq *dead_rq)
5578 struct rq *rq = dead_rq;
5579 struct task_struct *next, *stop = rq->stop;
5584 * Fudge the rq selection such that the below task selection loop
5585 * doesn't get stuck on the currently eligible stop task.
5587 * We're currently inside stop_machine() and the rq is either stuck
5588 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5589 * either way we should never end up calling schedule() until we're
5595 * put_prev_task() and pick_next_task() sched
5596 * class method both need to have an up-to-date
5597 * value of rq->clock[_task]
5599 rq_pin_lock(rq, &rf);
5600 update_rq_clock(rq);
5601 rq_unpin_lock(rq, &rf);
5605 * There's this thread running, bail when that's the only
5608 if (rq->nr_running == 1)
5612 * pick_next_task() assumes pinned rq->lock:
5614 rq_repin_lock(rq, &rf);
5615 next = pick_next_task(rq, &fake_task, &rf);
5617 next->sched_class->put_prev_task(rq, next);
5620 * Rules for changing task_struct::cpus_allowed are holding
5621 * both pi_lock and rq->lock, such that holding either
5622 * stabilizes the mask.
5624 * Drop rq->lock is not quite as disastrous as it usually is
5625 * because !cpu_active at this point, which means load-balance
5626 * will not interfere. Also, stop-machine.
5628 rq_unpin_lock(rq, &rf);
5629 raw_spin_unlock(&rq->lock);
5630 raw_spin_lock(&next->pi_lock);
5631 raw_spin_lock(&rq->lock);
5634 * Since we're inside stop-machine, _nothing_ should have
5635 * changed the task, WARN if weird stuff happened, because in
5636 * that case the above rq->lock drop is a fail too.
5638 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5639 raw_spin_unlock(&next->pi_lock);
5643 /* Find suitable destination for @next, with force if needed. */
5644 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5646 rq = __migrate_task(rq, next, dest_cpu);
5647 if (rq != dead_rq) {
5648 raw_spin_unlock(&rq->lock);
5650 raw_spin_lock(&rq->lock);
5652 raw_spin_unlock(&next->pi_lock);
5657 #endif /* CONFIG_HOTPLUG_CPU */
5659 void set_rq_online(struct rq *rq)
5662 const struct sched_class *class;
5664 cpumask_set_cpu(rq->cpu, rq->rd->online);
5667 for_each_class(class) {
5668 if (class->rq_online)
5669 class->rq_online(rq);
5674 void set_rq_offline(struct rq *rq)
5677 const struct sched_class *class;
5679 for_each_class(class) {
5680 if (class->rq_offline)
5681 class->rq_offline(rq);
5684 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5689 static void set_cpu_rq_start_time(unsigned int cpu)
5691 struct rq *rq = cpu_rq(cpu);
5693 rq->age_stamp = sched_clock_cpu(cpu);
5697 * used to mark begin/end of suspend/resume:
5699 static int num_cpus_frozen;
5702 * Update cpusets according to cpu_active mask. If cpusets are
5703 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5704 * around partition_sched_domains().
5706 * If we come here as part of a suspend/resume, don't touch cpusets because we
5707 * want to restore it back to its original state upon resume anyway.
5709 static void cpuset_cpu_active(void)
5711 if (cpuhp_tasks_frozen) {
5713 * num_cpus_frozen tracks how many CPUs are involved in suspend
5714 * resume sequence. As long as this is not the last online
5715 * operation in the resume sequence, just build a single sched
5716 * domain, ignoring cpusets.
5719 if (likely(num_cpus_frozen)) {
5720 partition_sched_domains(1, NULL, NULL);
5724 * This is the last CPU online operation. So fall through and
5725 * restore the original sched domains by considering the
5726 * cpuset configurations.
5729 cpuset_update_active_cpus(true);
5732 static int cpuset_cpu_inactive(unsigned int cpu)
5734 unsigned long flags;
5739 if (!cpuhp_tasks_frozen) {
5740 rcu_read_lock_sched();
5741 dl_b = dl_bw_of(cpu);
5743 raw_spin_lock_irqsave(&dl_b->lock, flags);
5744 cpus = dl_bw_cpus(cpu);
5745 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5746 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5748 rcu_read_unlock_sched();
5752 cpuset_update_active_cpus(false);
5755 partition_sched_domains(1, NULL, NULL);
5760 int sched_cpu_activate(unsigned int cpu)
5762 struct rq *rq = cpu_rq(cpu);
5763 unsigned long flags;
5765 set_cpu_active(cpu, true);
5767 if (sched_smp_initialized) {
5768 sched_domains_numa_masks_set(cpu);
5769 cpuset_cpu_active();
5773 * Put the rq online, if not already. This happens:
5775 * 1) In the early boot process, because we build the real domains
5776 * after all CPUs have been brought up.
5778 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5781 raw_spin_lock_irqsave(&rq->lock, flags);
5783 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5786 raw_spin_unlock_irqrestore(&rq->lock, flags);
5788 update_max_interval();
5793 int sched_cpu_deactivate(unsigned int cpu)
5797 set_cpu_active(cpu, false);
5799 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5800 * users of this state to go away such that all new such users will
5803 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
5804 * not imply sync_sched(), so wait for both.
5806 * Do sync before park smpboot threads to take care the rcu boost case.
5808 if (IS_ENABLED(CONFIG_PREEMPT))
5809 synchronize_rcu_mult(call_rcu, call_rcu_sched);
5813 if (!sched_smp_initialized)
5816 ret = cpuset_cpu_inactive(cpu);
5818 set_cpu_active(cpu, true);
5821 sched_domains_numa_masks_clear(cpu);
5825 static void sched_rq_cpu_starting(unsigned int cpu)
5827 struct rq *rq = cpu_rq(cpu);
5829 rq->calc_load_update = calc_load_update;
5830 update_max_interval();
5833 int sched_cpu_starting(unsigned int cpu)
5835 set_cpu_rq_start_time(cpu);
5836 sched_rq_cpu_starting(cpu);
5840 #ifdef CONFIG_HOTPLUG_CPU
5841 int sched_cpu_dying(unsigned int cpu)
5843 struct rq *rq = cpu_rq(cpu);
5844 unsigned long flags;
5846 /* Handle pending wakeups and then migrate everything off */
5847 sched_ttwu_pending();
5848 raw_spin_lock_irqsave(&rq->lock, flags);
5850 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5854 BUG_ON(rq->nr_running != 1);
5855 raw_spin_unlock_irqrestore(&rq->lock, flags);
5856 calc_load_migrate(rq);
5857 update_max_interval();
5858 nohz_balance_exit_idle(cpu);
5864 #ifdef CONFIG_SCHED_SMT
5865 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5867 static void sched_init_smt(void)
5870 * We've enumerated all CPUs and will assume that if any CPU
5871 * has SMT siblings, CPU0 will too.
5873 if (cpumask_weight(cpu_smt_mask(0)) > 1)
5874 static_branch_enable(&sched_smt_present);
5877 static inline void sched_init_smt(void) { }
5880 void __init sched_init_smp(void)
5882 cpumask_var_t non_isolated_cpus;
5884 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
5885 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
5890 * There's no userspace yet to cause hotplug operations; hence all the
5891 * CPU masks are stable and all blatant races in the below code cannot
5894 mutex_lock(&sched_domains_mutex);
5895 init_sched_domains(cpu_active_mask);
5896 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
5897 if (cpumask_empty(non_isolated_cpus))
5898 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
5899 mutex_unlock(&sched_domains_mutex);
5901 /* Move init over to a non-isolated CPU */
5902 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
5904 sched_init_granularity();
5905 free_cpumask_var(non_isolated_cpus);
5907 init_sched_rt_class();
5908 init_sched_dl_class();
5911 sched_clock_init_late();
5913 sched_smp_initialized = true;
5916 static int __init migration_init(void)
5918 sched_rq_cpu_starting(smp_processor_id());
5921 early_initcall(migration_init);
5924 void __init sched_init_smp(void)
5926 sched_init_granularity();
5927 sched_clock_init_late();
5929 #endif /* CONFIG_SMP */
5931 int in_sched_functions(unsigned long addr)
5933 return in_lock_functions(addr) ||
5934 (addr >= (unsigned long)__sched_text_start
5935 && addr < (unsigned long)__sched_text_end);
5938 #ifdef CONFIG_CGROUP_SCHED
5940 * Default task group.
5941 * Every task in system belongs to this group at bootup.
5943 struct task_group root_task_group;
5944 LIST_HEAD(task_groups);
5946 /* Cacheline aligned slab cache for task_group */
5947 static struct kmem_cache *task_group_cache __read_mostly;
5950 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5951 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5953 #define WAIT_TABLE_BITS 8
5954 #define WAIT_TABLE_SIZE (1 << WAIT_TABLE_BITS)
5955 static wait_queue_head_t bit_wait_table[WAIT_TABLE_SIZE] __cacheline_aligned;
5957 wait_queue_head_t *bit_waitqueue(void *word, int bit)
5959 const int shift = BITS_PER_LONG == 32 ? 5 : 6;
5960 unsigned long val = (unsigned long)word << shift | bit;
5962 return bit_wait_table + hash_long(val, WAIT_TABLE_BITS);
5964 EXPORT_SYMBOL(bit_waitqueue);
5966 void __init sched_init(void)
5969 unsigned long alloc_size = 0, ptr;
5973 for (i = 0; i < WAIT_TABLE_SIZE; i++)
5974 init_waitqueue_head(bit_wait_table + i);
5976 #ifdef CONFIG_FAIR_GROUP_SCHED
5977 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5979 #ifdef CONFIG_RT_GROUP_SCHED
5980 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5983 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5985 #ifdef CONFIG_FAIR_GROUP_SCHED
5986 root_task_group.se = (struct sched_entity **)ptr;
5987 ptr += nr_cpu_ids * sizeof(void **);
5989 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5990 ptr += nr_cpu_ids * sizeof(void **);
5992 #endif /* CONFIG_FAIR_GROUP_SCHED */
5993 #ifdef CONFIG_RT_GROUP_SCHED
5994 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5995 ptr += nr_cpu_ids * sizeof(void **);
5997 root_task_group.rt_rq = (struct rt_rq **)ptr;
5998 ptr += nr_cpu_ids * sizeof(void **);
6000 #endif /* CONFIG_RT_GROUP_SCHED */
6002 #ifdef CONFIG_CPUMASK_OFFSTACK
6003 for_each_possible_cpu(i) {
6004 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
6005 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6006 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
6007 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6009 #endif /* CONFIG_CPUMASK_OFFSTACK */
6011 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
6012 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6015 init_defrootdomain();
6018 #ifdef CONFIG_RT_GROUP_SCHED
6019 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6020 global_rt_period(), global_rt_runtime());
6021 #endif /* CONFIG_RT_GROUP_SCHED */
6023 #ifdef CONFIG_CGROUP_SCHED
6024 task_group_cache = KMEM_CACHE(task_group, 0);
6026 list_add(&root_task_group.list, &task_groups);
6027 INIT_LIST_HEAD(&root_task_group.children);
6028 INIT_LIST_HEAD(&root_task_group.siblings);
6029 autogroup_init(&init_task);
6030 #endif /* CONFIG_CGROUP_SCHED */
6032 for_each_possible_cpu(i) {
6036 raw_spin_lock_init(&rq->lock);
6038 rq->calc_load_active = 0;
6039 rq->calc_load_update = jiffies + LOAD_FREQ;
6040 init_cfs_rq(&rq->cfs);
6041 init_rt_rq(&rq->rt);
6042 init_dl_rq(&rq->dl);
6043 #ifdef CONFIG_FAIR_GROUP_SCHED
6044 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6045 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6046 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6048 * How much CPU bandwidth does root_task_group get?
6050 * In case of task-groups formed thr' the cgroup filesystem, it
6051 * gets 100% of the CPU resources in the system. This overall
6052 * system CPU resource is divided among the tasks of
6053 * root_task_group and its child task-groups in a fair manner,
6054 * based on each entity's (task or task-group's) weight
6055 * (se->load.weight).
6057 * In other words, if root_task_group has 10 tasks of weight
6058 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6059 * then A0's share of the CPU resource is:
6061 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6063 * We achieve this by letting root_task_group's tasks sit
6064 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6066 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6067 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6068 #endif /* CONFIG_FAIR_GROUP_SCHED */
6070 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6071 #ifdef CONFIG_RT_GROUP_SCHED
6072 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6075 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6076 rq->cpu_load[j] = 0;
6081 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6082 rq->balance_callback = NULL;
6083 rq->active_balance = 0;
6084 rq->next_balance = jiffies;
6089 rq->avg_idle = 2*sysctl_sched_migration_cost;
6090 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6092 INIT_LIST_HEAD(&rq->cfs_tasks);
6094 rq_attach_root(rq, &def_root_domain);
6095 #ifdef CONFIG_NO_HZ_COMMON
6096 rq->last_load_update_tick = jiffies;
6099 #ifdef CONFIG_NO_HZ_FULL
6100 rq->last_sched_tick = 0;
6102 #endif /* CONFIG_SMP */
6104 atomic_set(&rq->nr_iowait, 0);
6107 set_load_weight(&init_task);
6110 * The boot idle thread does lazy MMU switching as well:
6113 enter_lazy_tlb(&init_mm, current);
6116 * Make us the idle thread. Technically, schedule() should not be
6117 * called from this thread, however somewhere below it might be,
6118 * but because we are the idle thread, we just pick up running again
6119 * when this runqueue becomes "idle".
6121 init_idle(current, smp_processor_id());
6123 calc_load_update = jiffies + LOAD_FREQ;
6126 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6127 /* May be allocated at isolcpus cmdline parse time */
6128 if (cpu_isolated_map == NULL)
6129 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6130 idle_thread_set_boot_cpu();
6131 set_cpu_rq_start_time(smp_processor_id());
6133 init_sched_fair_class();
6137 scheduler_running = 1;
6140 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6141 static inline int preempt_count_equals(int preempt_offset)
6143 int nested = preempt_count() + rcu_preempt_depth();
6145 return (nested == preempt_offset);
6148 void __might_sleep(const char *file, int line, int preempt_offset)
6151 * Blocking primitives will set (and therefore destroy) current->state,
6152 * since we will exit with TASK_RUNNING make sure we enter with it,
6153 * otherwise we will destroy state.
6155 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6156 "do not call blocking ops when !TASK_RUNNING; "
6157 "state=%lx set at [<%p>] %pS\n",
6159 (void *)current->task_state_change,
6160 (void *)current->task_state_change);
6162 ___might_sleep(file, line, preempt_offset);
6164 EXPORT_SYMBOL(__might_sleep);
6166 void ___might_sleep(const char *file, int line, int preempt_offset)
6168 /* Ratelimiting timestamp: */
6169 static unsigned long prev_jiffy;
6171 unsigned long preempt_disable_ip;
6173 /* WARN_ON_ONCE() by default, no rate limit required: */
6176 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6177 !is_idle_task(current)) ||
6178 system_state != SYSTEM_RUNNING || oops_in_progress)
6180 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6182 prev_jiffy = jiffies;
6184 /* Save this before calling printk(), since that will clobber it: */
6185 preempt_disable_ip = get_preempt_disable_ip(current);
6188 "BUG: sleeping function called from invalid context at %s:%d\n",
6191 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6192 in_atomic(), irqs_disabled(),
6193 current->pid, current->comm);
6195 if (task_stack_end_corrupted(current))
6196 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6198 debug_show_held_locks(current);
6199 if (irqs_disabled())
6200 print_irqtrace_events(current);
6201 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6202 && !preempt_count_equals(preempt_offset)) {
6203 pr_err("Preemption disabled at:");
6204 print_ip_sym(preempt_disable_ip);
6208 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6210 EXPORT_SYMBOL(___might_sleep);
6213 #ifdef CONFIG_MAGIC_SYSRQ
6214 void normalize_rt_tasks(void)
6216 struct task_struct *g, *p;
6217 struct sched_attr attr = {
6218 .sched_policy = SCHED_NORMAL,
6221 read_lock(&tasklist_lock);
6222 for_each_process_thread(g, p) {
6224 * Only normalize user tasks:
6226 if (p->flags & PF_KTHREAD)
6229 p->se.exec_start = 0;
6230 schedstat_set(p->se.statistics.wait_start, 0);
6231 schedstat_set(p->se.statistics.sleep_start, 0);
6232 schedstat_set(p->se.statistics.block_start, 0);
6234 if (!dl_task(p) && !rt_task(p)) {
6236 * Renice negative nice level userspace
6239 if (task_nice(p) < 0)
6240 set_user_nice(p, 0);
6244 __sched_setscheduler(p, &attr, false, false);
6246 read_unlock(&tasklist_lock);
6249 #endif /* CONFIG_MAGIC_SYSRQ */
6251 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6253 * These functions are only useful for the IA64 MCA handling, or kdb.
6255 * They can only be called when the whole system has been
6256 * stopped - every CPU needs to be quiescent, and no scheduling
6257 * activity can take place. Using them for anything else would
6258 * be a serious bug, and as a result, they aren't even visible
6259 * under any other configuration.
6263 * curr_task - return the current task for a given CPU.
6264 * @cpu: the processor in question.
6266 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6268 * Return: The current task for @cpu.
6270 struct task_struct *curr_task(int cpu)
6272 return cpu_curr(cpu);
6275 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6279 * set_curr_task - set the current task for a given CPU.
6280 * @cpu: the processor in question.
6281 * @p: the task pointer to set.
6283 * Description: This function must only be used when non-maskable interrupts
6284 * are serviced on a separate stack. It allows the architecture to switch the
6285 * notion of the current task on a CPU in a non-blocking manner. This function
6286 * must be called with all CPU's synchronized, and interrupts disabled, the
6287 * and caller must save the original value of the current task (see
6288 * curr_task() above) and restore that value before reenabling interrupts and
6289 * re-starting the system.
6291 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6293 void ia64_set_curr_task(int cpu, struct task_struct *p)
6300 #ifdef CONFIG_CGROUP_SCHED
6301 /* task_group_lock serializes the addition/removal of task groups */
6302 static DEFINE_SPINLOCK(task_group_lock);
6304 static void sched_free_group(struct task_group *tg)
6306 free_fair_sched_group(tg);
6307 free_rt_sched_group(tg);
6309 kmem_cache_free(task_group_cache, tg);
6312 /* allocate runqueue etc for a new task group */
6313 struct task_group *sched_create_group(struct task_group *parent)
6315 struct task_group *tg;
6317 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6319 return ERR_PTR(-ENOMEM);
6321 if (!alloc_fair_sched_group(tg, parent))
6324 if (!alloc_rt_sched_group(tg, parent))
6330 sched_free_group(tg);
6331 return ERR_PTR(-ENOMEM);
6334 void sched_online_group(struct task_group *tg, struct task_group *parent)
6336 unsigned long flags;
6338 spin_lock_irqsave(&task_group_lock, flags);
6339 list_add_rcu(&tg->list, &task_groups);
6341 /* Root should already exist: */
6344 tg->parent = parent;
6345 INIT_LIST_HEAD(&tg->children);
6346 list_add_rcu(&tg->siblings, &parent->children);
6347 spin_unlock_irqrestore(&task_group_lock, flags);
6349 online_fair_sched_group(tg);
6352 /* rcu callback to free various structures associated with a task group */
6353 static void sched_free_group_rcu(struct rcu_head *rhp)
6355 /* Now it should be safe to free those cfs_rqs: */
6356 sched_free_group(container_of(rhp, struct task_group, rcu));
6359 void sched_destroy_group(struct task_group *tg)
6361 /* Wait for possible concurrent references to cfs_rqs complete: */
6362 call_rcu(&tg->rcu, sched_free_group_rcu);
6365 void sched_offline_group(struct task_group *tg)
6367 unsigned long flags;
6369 /* End participation in shares distribution: */
6370 unregister_fair_sched_group(tg);
6372 spin_lock_irqsave(&task_group_lock, flags);
6373 list_del_rcu(&tg->list);
6374 list_del_rcu(&tg->siblings);
6375 spin_unlock_irqrestore(&task_group_lock, flags);
6378 static void sched_change_group(struct task_struct *tsk, int type)
6380 struct task_group *tg;
6383 * All callers are synchronized by task_rq_lock(); we do not use RCU
6384 * which is pointless here. Thus, we pass "true" to task_css_check()
6385 * to prevent lockdep warnings.
6387 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6388 struct task_group, css);
6389 tg = autogroup_task_group(tsk, tg);
6390 tsk->sched_task_group = tg;
6392 #ifdef CONFIG_FAIR_GROUP_SCHED
6393 if (tsk->sched_class->task_change_group)
6394 tsk->sched_class->task_change_group(tsk, type);
6397 set_task_rq(tsk, task_cpu(tsk));
6401 * Change task's runqueue when it moves between groups.
6403 * The caller of this function should have put the task in its new group by
6404 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6407 void sched_move_task(struct task_struct *tsk)
6409 int queued, running;
6413 rq = task_rq_lock(tsk, &rf);
6414 update_rq_clock(rq);
6416 running = task_current(rq, tsk);
6417 queued = task_on_rq_queued(tsk);
6420 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
6422 put_prev_task(rq, tsk);
6424 sched_change_group(tsk, TASK_MOVE_GROUP);
6427 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
6429 set_curr_task(rq, tsk);
6431 task_rq_unlock(rq, tsk, &rf);
6433 #endif /* CONFIG_CGROUP_SCHED */
6435 #ifdef CONFIG_RT_GROUP_SCHED
6437 * Ensure that the real time constraints are schedulable.
6439 static DEFINE_MUTEX(rt_constraints_mutex);
6441 /* Must be called with tasklist_lock held */
6442 static inline int tg_has_rt_tasks(struct task_group *tg)
6444 struct task_struct *g, *p;
6447 * Autogroups do not have RT tasks; see autogroup_create().
6449 if (task_group_is_autogroup(tg))
6452 for_each_process_thread(g, p) {
6453 if (rt_task(p) && task_group(p) == tg)
6460 struct rt_schedulable_data {
6461 struct task_group *tg;
6466 static int tg_rt_schedulable(struct task_group *tg, void *data)
6468 struct rt_schedulable_data *d = data;
6469 struct task_group *child;
6470 unsigned long total, sum = 0;
6471 u64 period, runtime;
6473 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6474 runtime = tg->rt_bandwidth.rt_runtime;
6477 period = d->rt_period;
6478 runtime = d->rt_runtime;
6482 * Cannot have more runtime than the period.
6484 if (runtime > period && runtime != RUNTIME_INF)
6488 * Ensure we don't starve existing RT tasks.
6490 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
6493 total = to_ratio(period, runtime);
6496 * Nobody can have more than the global setting allows.
6498 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
6502 * The sum of our children's runtime should not exceed our own.
6504 list_for_each_entry_rcu(child, &tg->children, siblings) {
6505 period = ktime_to_ns(child->rt_bandwidth.rt_period);
6506 runtime = child->rt_bandwidth.rt_runtime;
6508 if (child == d->tg) {
6509 period = d->rt_period;
6510 runtime = d->rt_runtime;
6513 sum += to_ratio(period, runtime);
6522 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
6526 struct rt_schedulable_data data = {
6528 .rt_period = period,
6529 .rt_runtime = runtime,
6533 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
6539 static int tg_set_rt_bandwidth(struct task_group *tg,
6540 u64 rt_period, u64 rt_runtime)
6545 * Disallowing the root group RT runtime is BAD, it would disallow the
6546 * kernel creating (and or operating) RT threads.
6548 if (tg == &root_task_group && rt_runtime == 0)
6551 /* No period doesn't make any sense. */
6555 mutex_lock(&rt_constraints_mutex);
6556 read_lock(&tasklist_lock);
6557 err = __rt_schedulable(tg, rt_period, rt_runtime);
6561 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6562 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
6563 tg->rt_bandwidth.rt_runtime = rt_runtime;
6565 for_each_possible_cpu(i) {
6566 struct rt_rq *rt_rq = tg->rt_rq[i];
6568 raw_spin_lock(&rt_rq->rt_runtime_lock);
6569 rt_rq->rt_runtime = rt_runtime;
6570 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6572 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6574 read_unlock(&tasklist_lock);
6575 mutex_unlock(&rt_constraints_mutex);
6580 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
6582 u64 rt_runtime, rt_period;
6584 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6585 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
6586 if (rt_runtime_us < 0)
6587 rt_runtime = RUNTIME_INF;
6589 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6592 static long sched_group_rt_runtime(struct task_group *tg)
6596 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
6599 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
6600 do_div(rt_runtime_us, NSEC_PER_USEC);
6601 return rt_runtime_us;
6604 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
6606 u64 rt_runtime, rt_period;
6608 rt_period = rt_period_us * NSEC_PER_USEC;
6609 rt_runtime = tg->rt_bandwidth.rt_runtime;
6611 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6614 static long sched_group_rt_period(struct task_group *tg)
6618 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
6619 do_div(rt_period_us, NSEC_PER_USEC);
6620 return rt_period_us;
6622 #endif /* CONFIG_RT_GROUP_SCHED */
6624 #ifdef CONFIG_RT_GROUP_SCHED
6625 static int sched_rt_global_constraints(void)
6629 mutex_lock(&rt_constraints_mutex);
6630 read_lock(&tasklist_lock);
6631 ret = __rt_schedulable(NULL, 0, 0);
6632 read_unlock(&tasklist_lock);
6633 mutex_unlock(&rt_constraints_mutex);
6638 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
6640 /* Don't accept realtime tasks when there is no way for them to run */
6641 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
6647 #else /* !CONFIG_RT_GROUP_SCHED */
6648 static int sched_rt_global_constraints(void)
6650 unsigned long flags;
6653 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
6654 for_each_possible_cpu(i) {
6655 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
6657 raw_spin_lock(&rt_rq->rt_runtime_lock);
6658 rt_rq->rt_runtime = global_rt_runtime();
6659 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6661 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
6665 #endif /* CONFIG_RT_GROUP_SCHED */
6667 static int sched_dl_global_validate(void)
6669 u64 runtime = global_rt_runtime();
6670 u64 period = global_rt_period();
6671 u64 new_bw = to_ratio(period, runtime);
6674 unsigned long flags;
6677 * Here we want to check the bandwidth not being set to some
6678 * value smaller than the currently allocated bandwidth in
6679 * any of the root_domains.
6681 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
6682 * cycling on root_domains... Discussion on different/better
6683 * solutions is welcome!
6685 for_each_possible_cpu(cpu) {
6686 rcu_read_lock_sched();
6687 dl_b = dl_bw_of(cpu);
6689 raw_spin_lock_irqsave(&dl_b->lock, flags);
6690 if (new_bw < dl_b->total_bw)
6692 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
6694 rcu_read_unlock_sched();
6703 static void sched_dl_do_global(void)
6708 unsigned long flags;
6710 def_dl_bandwidth.dl_period = global_rt_period();
6711 def_dl_bandwidth.dl_runtime = global_rt_runtime();
6713 if (global_rt_runtime() != RUNTIME_INF)
6714 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
6717 * FIXME: As above...
6719 for_each_possible_cpu(cpu) {
6720 rcu_read_lock_sched();
6721 dl_b = dl_bw_of(cpu);
6723 raw_spin_lock_irqsave(&dl_b->lock, flags);
6725 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
6727 rcu_read_unlock_sched();
6731 static int sched_rt_global_validate(void)
6733 if (sysctl_sched_rt_period <= 0)
6736 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
6737 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
6743 static void sched_rt_do_global(void)
6745 def_rt_bandwidth.rt_runtime = global_rt_runtime();
6746 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
6749 int sched_rt_handler(struct ctl_table *table, int write,
6750 void __user *buffer, size_t *lenp,
6753 int old_period, old_runtime;
6754 static DEFINE_MUTEX(mutex);
6758 old_period = sysctl_sched_rt_period;
6759 old_runtime = sysctl_sched_rt_runtime;
6761 ret = proc_dointvec(table, write, buffer, lenp, ppos);
6763 if (!ret && write) {
6764 ret = sched_rt_global_validate();
6768 ret = sched_dl_global_validate();
6772 ret = sched_rt_global_constraints();
6776 sched_rt_do_global();
6777 sched_dl_do_global();
6781 sysctl_sched_rt_period = old_period;
6782 sysctl_sched_rt_runtime = old_runtime;
6784 mutex_unlock(&mutex);
6789 int sched_rr_handler(struct ctl_table *table, int write,
6790 void __user *buffer, size_t *lenp,
6794 static DEFINE_MUTEX(mutex);
6797 ret = proc_dointvec(table, write, buffer, lenp, ppos);
6799 * Make sure that internally we keep jiffies.
6800 * Also, writing zero resets the timeslice to default:
6802 if (!ret && write) {
6803 sched_rr_timeslice =
6804 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
6805 msecs_to_jiffies(sysctl_sched_rr_timeslice);
6807 mutex_unlock(&mutex);
6811 #ifdef CONFIG_CGROUP_SCHED
6813 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6815 return css ? container_of(css, struct task_group, css) : NULL;
6818 static struct cgroup_subsys_state *
6819 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6821 struct task_group *parent = css_tg(parent_css);
6822 struct task_group *tg;
6825 /* This is early initialization for the top cgroup */
6826 return &root_task_group.css;
6829 tg = sched_create_group(parent);
6831 return ERR_PTR(-ENOMEM);
6836 /* Expose task group only after completing cgroup initialization */
6837 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6839 struct task_group *tg = css_tg(css);
6840 struct task_group *parent = css_tg(css->parent);
6843 sched_online_group(tg, parent);
6847 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6849 struct task_group *tg = css_tg(css);
6851 sched_offline_group(tg);
6854 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6856 struct task_group *tg = css_tg(css);
6859 * Relies on the RCU grace period between css_released() and this.
6861 sched_free_group(tg);
6865 * This is called before wake_up_new_task(), therefore we really only
6866 * have to set its group bits, all the other stuff does not apply.
6868 static void cpu_cgroup_fork(struct task_struct *task)
6873 rq = task_rq_lock(task, &rf);
6875 update_rq_clock(rq);
6876 sched_change_group(task, TASK_SET_GROUP);
6878 task_rq_unlock(rq, task, &rf);
6881 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6883 struct task_struct *task;
6884 struct cgroup_subsys_state *css;
6887 cgroup_taskset_for_each(task, css, tset) {
6888 #ifdef CONFIG_RT_GROUP_SCHED
6889 if (!sched_rt_can_attach(css_tg(css), task))
6892 /* We don't support RT-tasks being in separate groups */
6893 if (task->sched_class != &fair_sched_class)
6897 * Serialize against wake_up_new_task() such that if its
6898 * running, we're sure to observe its full state.
6900 raw_spin_lock_irq(&task->pi_lock);
6902 * Avoid calling sched_move_task() before wake_up_new_task()
6903 * has happened. This would lead to problems with PELT, due to
6904 * move wanting to detach+attach while we're not attached yet.
6906 if (task->state == TASK_NEW)
6908 raw_spin_unlock_irq(&task->pi_lock);
6916 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6918 struct task_struct *task;
6919 struct cgroup_subsys_state *css;
6921 cgroup_taskset_for_each(task, css, tset)
6922 sched_move_task(task);
6925 #ifdef CONFIG_FAIR_GROUP_SCHED
6926 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6927 struct cftype *cftype, u64 shareval)
6929 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6932 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6935 struct task_group *tg = css_tg(css);
6937 return (u64) scale_load_down(tg->shares);
6940 #ifdef CONFIG_CFS_BANDWIDTH
6941 static DEFINE_MUTEX(cfs_constraints_mutex);
6943 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6944 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6946 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6948 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6950 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6951 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6953 if (tg == &root_task_group)
6957 * Ensure we have at some amount of bandwidth every period. This is
6958 * to prevent reaching a state of large arrears when throttled via
6959 * entity_tick() resulting in prolonged exit starvation.
6961 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6965 * Likewise, bound things on the otherside by preventing insane quota
6966 * periods. This also allows us to normalize in computing quota
6969 if (period > max_cfs_quota_period)
6973 * Prevent race between setting of cfs_rq->runtime_enabled and
6974 * unthrottle_offline_cfs_rqs().
6977 mutex_lock(&cfs_constraints_mutex);
6978 ret = __cfs_schedulable(tg, period, quota);
6982 runtime_enabled = quota != RUNTIME_INF;
6983 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6985 * If we need to toggle cfs_bandwidth_used, off->on must occur
6986 * before making related changes, and on->off must occur afterwards
6988 if (runtime_enabled && !runtime_was_enabled)
6989 cfs_bandwidth_usage_inc();
6990 raw_spin_lock_irq(&cfs_b->lock);
6991 cfs_b->period = ns_to_ktime(period);
6992 cfs_b->quota = quota;
6994 __refill_cfs_bandwidth_runtime(cfs_b);
6996 /* Restart the period timer (if active) to handle new period expiry: */
6997 if (runtime_enabled)
6998 start_cfs_bandwidth(cfs_b);
7000 raw_spin_unlock_irq(&cfs_b->lock);
7002 for_each_online_cpu(i) {
7003 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7004 struct rq *rq = cfs_rq->rq;
7006 raw_spin_lock_irq(&rq->lock);
7007 cfs_rq->runtime_enabled = runtime_enabled;
7008 cfs_rq->runtime_remaining = 0;
7010 if (cfs_rq->throttled)
7011 unthrottle_cfs_rq(cfs_rq);
7012 raw_spin_unlock_irq(&rq->lock);
7014 if (runtime_was_enabled && !runtime_enabled)
7015 cfs_bandwidth_usage_dec();
7017 mutex_unlock(&cfs_constraints_mutex);
7023 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7027 period = ktime_to_ns(tg->cfs_bandwidth.period);
7028 if (cfs_quota_us < 0)
7029 quota = RUNTIME_INF;
7031 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7033 return tg_set_cfs_bandwidth(tg, period, quota);
7036 long tg_get_cfs_quota(struct task_group *tg)
7040 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7043 quota_us = tg->cfs_bandwidth.quota;
7044 do_div(quota_us, NSEC_PER_USEC);
7049 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7053 period = (u64)cfs_period_us * NSEC_PER_USEC;
7054 quota = tg->cfs_bandwidth.quota;
7056 return tg_set_cfs_bandwidth(tg, period, quota);
7059 long tg_get_cfs_period(struct task_group *tg)
7063 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7064 do_div(cfs_period_us, NSEC_PER_USEC);
7066 return cfs_period_us;
7069 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7072 return tg_get_cfs_quota(css_tg(css));
7075 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7076 struct cftype *cftype, s64 cfs_quota_us)
7078 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7081 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7084 return tg_get_cfs_period(css_tg(css));
7087 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7088 struct cftype *cftype, u64 cfs_period_us)
7090 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7093 struct cfs_schedulable_data {
7094 struct task_group *tg;
7099 * normalize group quota/period to be quota/max_period
7100 * note: units are usecs
7102 static u64 normalize_cfs_quota(struct task_group *tg,
7103 struct cfs_schedulable_data *d)
7111 period = tg_get_cfs_period(tg);
7112 quota = tg_get_cfs_quota(tg);
7115 /* note: these should typically be equivalent */
7116 if (quota == RUNTIME_INF || quota == -1)
7119 return to_ratio(period, quota);
7122 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7124 struct cfs_schedulable_data *d = data;
7125 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7126 s64 quota = 0, parent_quota = -1;
7129 quota = RUNTIME_INF;
7131 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7133 quota = normalize_cfs_quota(tg, d);
7134 parent_quota = parent_b->hierarchical_quota;
7137 * Ensure max(child_quota) <= parent_quota, inherit when no
7140 if (quota == RUNTIME_INF)
7141 quota = parent_quota;
7142 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7145 cfs_b->hierarchical_quota = quota;
7150 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7153 struct cfs_schedulable_data data = {
7159 if (quota != RUNTIME_INF) {
7160 do_div(data.period, NSEC_PER_USEC);
7161 do_div(data.quota, NSEC_PER_USEC);
7165 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7171 static int cpu_stats_show(struct seq_file *sf, void *v)
7173 struct task_group *tg = css_tg(seq_css(sf));
7174 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7176 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7177 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7178 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7182 #endif /* CONFIG_CFS_BANDWIDTH */
7183 #endif /* CONFIG_FAIR_GROUP_SCHED */
7185 #ifdef CONFIG_RT_GROUP_SCHED
7186 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7187 struct cftype *cft, s64 val)
7189 return sched_group_set_rt_runtime(css_tg(css), val);
7192 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7195 return sched_group_rt_runtime(css_tg(css));
7198 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7199 struct cftype *cftype, u64 rt_period_us)
7201 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7204 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7207 return sched_group_rt_period(css_tg(css));
7209 #endif /* CONFIG_RT_GROUP_SCHED */
7211 static struct cftype cpu_files[] = {
7212 #ifdef CONFIG_FAIR_GROUP_SCHED
7215 .read_u64 = cpu_shares_read_u64,
7216 .write_u64 = cpu_shares_write_u64,
7219 #ifdef CONFIG_CFS_BANDWIDTH
7221 .name = "cfs_quota_us",
7222 .read_s64 = cpu_cfs_quota_read_s64,
7223 .write_s64 = cpu_cfs_quota_write_s64,
7226 .name = "cfs_period_us",
7227 .read_u64 = cpu_cfs_period_read_u64,
7228 .write_u64 = cpu_cfs_period_write_u64,
7232 .seq_show = cpu_stats_show,
7235 #ifdef CONFIG_RT_GROUP_SCHED
7237 .name = "rt_runtime_us",
7238 .read_s64 = cpu_rt_runtime_read,
7239 .write_s64 = cpu_rt_runtime_write,
7242 .name = "rt_period_us",
7243 .read_u64 = cpu_rt_period_read_uint,
7244 .write_u64 = cpu_rt_period_write_uint,
7250 struct cgroup_subsys cpu_cgrp_subsys = {
7251 .css_alloc = cpu_cgroup_css_alloc,
7252 .css_online = cpu_cgroup_css_online,
7253 .css_released = cpu_cgroup_css_released,
7254 .css_free = cpu_cgroup_css_free,
7255 .fork = cpu_cgroup_fork,
7256 .can_attach = cpu_cgroup_can_attach,
7257 .attach = cpu_cgroup_attach,
7258 .legacy_cftypes = cpu_files,
7262 #endif /* CONFIG_CGROUP_SCHED */
7264 void dump_cpu_task(int cpu)
7266 pr_info("Task dump for CPU %d:\n", cpu);
7267 sched_show_task(cpu_curr(cpu));
7271 * Nice levels are multiplicative, with a gentle 10% change for every
7272 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7273 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7274 * that remained on nice 0.
7276 * The "10% effect" is relative and cumulative: from _any_ nice level,
7277 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7278 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7279 * If a task goes up by ~10% and another task goes down by ~10% then
7280 * the relative distance between them is ~25%.)
7282 const int sched_prio_to_weight[40] = {
7283 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7284 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7285 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7286 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7287 /* 0 */ 1024, 820, 655, 526, 423,
7288 /* 5 */ 335, 272, 215, 172, 137,
7289 /* 10 */ 110, 87, 70, 56, 45,
7290 /* 15 */ 36, 29, 23, 18, 15,
7294 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7296 * In cases where the weight does not change often, we can use the
7297 * precalculated inverse to speed up arithmetics by turning divisions
7298 * into multiplications:
7300 const u32 sched_prio_to_wmult[40] = {
7301 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7302 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7303 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7304 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7305 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7306 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7307 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7308 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,