4 * Core kernel scheduler code and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
10 #include <linux/nospec.h>
12 #include <linux/kcov.h>
14 #include <asm/switch_to.h>
17 #include "../workqueue_internal.h"
18 #include "../smpboot.h"
22 #define CREATE_TRACE_POINTS
23 #include <trace/events/sched.h>
25 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
27 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
29 * Debugging: various feature bits
31 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
32 * sysctl_sched_features, defined in sched.h, to allow constants propagation
33 * at compile time and compiler optimization based on features default.
35 #define SCHED_FEAT(name, enabled) \
36 (1UL << __SCHED_FEAT_##name) * enabled |
37 const_debug unsigned int sysctl_sched_features =
44 * Number of tasks to iterate in a single balance run.
45 * Limited because this is done with IRQs disabled.
47 const_debug unsigned int sysctl_sched_nr_migrate = 32;
50 * period over which we measure -rt task CPU usage in us.
53 unsigned int sysctl_sched_rt_period = 1000000;
55 __read_mostly int scheduler_running;
58 * part of the period that we allow rt tasks to run in us.
61 int sysctl_sched_rt_runtime = 950000;
64 * __task_rq_lock - lock the rq @p resides on.
66 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
71 lockdep_assert_held(&p->pi_lock);
75 raw_spin_lock(&rq->lock);
76 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
80 raw_spin_unlock(&rq->lock);
82 while (unlikely(task_on_rq_migrating(p)))
88 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
90 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
91 __acquires(p->pi_lock)
97 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
99 raw_spin_lock(&rq->lock);
101 * move_queued_task() task_rq_lock()
104 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
105 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
106 * [S] ->cpu = new_cpu [L] task_rq()
110 * If we observe the old CPU in task_rq_lock, the acquire of
111 * the old rq->lock will fully serialize against the stores.
113 * If we observe the new CPU in task_rq_lock, the acquire will
114 * pair with the WMB to ensure we must then also see migrating.
116 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
120 raw_spin_unlock(&rq->lock);
121 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
123 while (unlikely(task_on_rq_migrating(p)))
129 * RQ-clock updating methods:
132 static void update_rq_clock_task(struct rq *rq, s64 delta)
135 * In theory, the compile should just see 0 here, and optimize out the call
136 * to sched_rt_avg_update. But I don't trust it...
138 s64 __maybe_unused steal = 0, irq_delta = 0;
140 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
141 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
144 * Since irq_time is only updated on {soft,}irq_exit, we might run into
145 * this case when a previous update_rq_clock() happened inside a
148 * When this happens, we stop ->clock_task and only update the
149 * prev_irq_time stamp to account for the part that fit, so that a next
150 * update will consume the rest. This ensures ->clock_task is
153 * It does however cause some slight miss-attribution of {soft,}irq
154 * time, a more accurate solution would be to update the irq_time using
155 * the current rq->clock timestamp, except that would require using
158 if (irq_delta > delta)
161 rq->prev_irq_time += irq_delta;
164 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
165 if (static_key_false((¶virt_steal_rq_enabled))) {
166 steal = paravirt_steal_clock(cpu_of(rq));
167 steal -= rq->prev_steal_time_rq;
169 if (unlikely(steal > delta))
172 rq->prev_steal_time_rq += steal;
177 rq->clock_task += delta;
179 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
180 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
181 update_irq_load_avg(rq, irq_delta + steal);
185 void update_rq_clock(struct rq *rq)
189 lockdep_assert_held(&rq->lock);
191 if (rq->clock_update_flags & RQCF_ACT_SKIP)
194 #ifdef CONFIG_SCHED_DEBUG
195 if (sched_feat(WARN_DOUBLE_CLOCK))
196 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
197 rq->clock_update_flags |= RQCF_UPDATED;
200 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
204 update_rq_clock_task(rq, delta);
208 #ifdef CONFIG_SCHED_HRTICK
210 * Use HR-timers to deliver accurate preemption points.
213 static void hrtick_clear(struct rq *rq)
215 if (hrtimer_active(&rq->hrtick_timer))
216 hrtimer_cancel(&rq->hrtick_timer);
220 * High-resolution timer tick.
221 * Runs from hardirq context with interrupts disabled.
223 static enum hrtimer_restart hrtick(struct hrtimer *timer)
225 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
228 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
232 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
235 return HRTIMER_NORESTART;
240 static void __hrtick_restart(struct rq *rq)
242 struct hrtimer *timer = &rq->hrtick_timer;
244 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
248 * called from hardirq (IPI) context
250 static void __hrtick_start(void *arg)
256 __hrtick_restart(rq);
257 rq->hrtick_csd_pending = 0;
262 * Called to set the hrtick timer state.
264 * called with rq->lock held and irqs disabled
266 void hrtick_start(struct rq *rq, u64 delay)
268 struct hrtimer *timer = &rq->hrtick_timer;
273 * Don't schedule slices shorter than 10000ns, that just
274 * doesn't make sense and can cause timer DoS.
276 delta = max_t(s64, delay, 10000LL);
277 time = ktime_add_ns(timer->base->get_time(), delta);
279 hrtimer_set_expires(timer, time);
281 if (rq == this_rq()) {
282 __hrtick_restart(rq);
283 } else if (!rq->hrtick_csd_pending) {
284 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
285 rq->hrtick_csd_pending = 1;
291 * Called to set the hrtick timer state.
293 * called with rq->lock held and irqs disabled
295 void hrtick_start(struct rq *rq, u64 delay)
298 * Don't schedule slices shorter than 10000ns, that just
299 * doesn't make sense. Rely on vruntime for fairness.
301 delay = max_t(u64, delay, 10000LL);
302 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
303 HRTIMER_MODE_REL_PINNED);
305 #endif /* CONFIG_SMP */
307 static void hrtick_rq_init(struct rq *rq)
310 rq->hrtick_csd_pending = 0;
312 rq->hrtick_csd.flags = 0;
313 rq->hrtick_csd.func = __hrtick_start;
314 rq->hrtick_csd.info = rq;
317 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
318 rq->hrtick_timer.function = hrtick;
320 #else /* CONFIG_SCHED_HRTICK */
321 static inline void hrtick_clear(struct rq *rq)
325 static inline void hrtick_rq_init(struct rq *rq)
328 #endif /* CONFIG_SCHED_HRTICK */
331 * cmpxchg based fetch_or, macro so it works for different integer types
333 #define fetch_or(ptr, mask) \
335 typeof(ptr) _ptr = (ptr); \
336 typeof(mask) _mask = (mask); \
337 typeof(*_ptr) _old, _val = *_ptr; \
340 _old = cmpxchg(_ptr, _val, _val | _mask); \
348 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
350 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
351 * this avoids any races wrt polling state changes and thereby avoids
354 static bool set_nr_and_not_polling(struct task_struct *p)
356 struct thread_info *ti = task_thread_info(p);
357 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
361 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
363 * If this returns true, then the idle task promises to call
364 * sched_ttwu_pending() and reschedule soon.
366 static bool set_nr_if_polling(struct task_struct *p)
368 struct thread_info *ti = task_thread_info(p);
369 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
372 if (!(val & _TIF_POLLING_NRFLAG))
374 if (val & _TIF_NEED_RESCHED)
376 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
385 static bool set_nr_and_not_polling(struct task_struct *p)
387 set_tsk_need_resched(p);
392 static bool set_nr_if_polling(struct task_struct *p)
399 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
401 struct wake_q_node *node = &task->wake_q;
404 * Atomically grab the task, if ->wake_q is !nil already it means
405 * its already queued (either by us or someone else) and will get the
406 * wakeup due to that.
408 * This cmpxchg() executes a full barrier, which pairs with the full
409 * barrier executed by the wakeup in wake_up_q().
411 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
414 get_task_struct(task);
417 * The head is context local, there can be no concurrency.
420 head->lastp = &node->next;
423 void wake_up_q(struct wake_q_head *head)
425 struct wake_q_node *node = head->first;
427 while (node != WAKE_Q_TAIL) {
428 struct task_struct *task;
430 task = container_of(node, struct task_struct, wake_q);
432 /* Task can safely be re-inserted now: */
434 task->wake_q.next = NULL;
437 * wake_up_process() executes a full barrier, which pairs with
438 * the queueing in wake_q_add() so as not to miss wakeups.
440 wake_up_process(task);
441 put_task_struct(task);
446 * resched_curr - mark rq's current task 'to be rescheduled now'.
448 * On UP this means the setting of the need_resched flag, on SMP it
449 * might also involve a cross-CPU call to trigger the scheduler on
452 void resched_curr(struct rq *rq)
454 struct task_struct *curr = rq->curr;
457 lockdep_assert_held(&rq->lock);
459 if (test_tsk_need_resched(curr))
464 if (cpu == smp_processor_id()) {
465 set_tsk_need_resched(curr);
466 set_preempt_need_resched();
470 if (set_nr_and_not_polling(curr))
471 smp_send_reschedule(cpu);
473 trace_sched_wake_idle_without_ipi(cpu);
476 void resched_cpu(int cpu)
478 struct rq *rq = cpu_rq(cpu);
481 raw_spin_lock_irqsave(&rq->lock, flags);
482 if (cpu_online(cpu) || cpu == smp_processor_id())
484 raw_spin_unlock_irqrestore(&rq->lock, flags);
488 #ifdef CONFIG_NO_HZ_COMMON
490 * In the semi idle case, use the nearest busy CPU for migrating timers
491 * from an idle CPU. This is good for power-savings.
493 * We don't do similar optimization for completely idle system, as
494 * selecting an idle CPU will add more delays to the timers than intended
495 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
497 int get_nohz_timer_target(void)
499 int i, cpu = smp_processor_id();
500 struct sched_domain *sd;
502 if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
506 for_each_domain(cpu, sd) {
507 for_each_cpu(i, sched_domain_span(sd)) {
511 if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
518 if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
519 cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
526 * When add_timer_on() enqueues a timer into the timer wheel of an
527 * idle CPU then this timer might expire before the next timer event
528 * which is scheduled to wake up that CPU. In case of a completely
529 * idle system the next event might even be infinite time into the
530 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
531 * leaves the inner idle loop so the newly added timer is taken into
532 * account when the CPU goes back to idle and evaluates the timer
533 * wheel for the next timer event.
535 static void wake_up_idle_cpu(int cpu)
537 struct rq *rq = cpu_rq(cpu);
539 if (cpu == smp_processor_id())
542 if (set_nr_and_not_polling(rq->idle))
543 smp_send_reschedule(cpu);
545 trace_sched_wake_idle_without_ipi(cpu);
548 static bool wake_up_full_nohz_cpu(int cpu)
551 * We just need the target to call irq_exit() and re-evaluate
552 * the next tick. The nohz full kick at least implies that.
553 * If needed we can still optimize that later with an
556 if (cpu_is_offline(cpu))
557 return true; /* Don't try to wake offline CPUs. */
558 if (tick_nohz_full_cpu(cpu)) {
559 if (cpu != smp_processor_id() ||
560 tick_nohz_tick_stopped())
561 tick_nohz_full_kick_cpu(cpu);
569 * Wake up the specified CPU. If the CPU is going offline, it is the
570 * caller's responsibility to deal with the lost wakeup, for example,
571 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
573 void wake_up_nohz_cpu(int cpu)
575 if (!wake_up_full_nohz_cpu(cpu))
576 wake_up_idle_cpu(cpu);
579 static inline bool got_nohz_idle_kick(void)
581 int cpu = smp_processor_id();
583 if (!(atomic_read(nohz_flags(cpu)) & NOHZ_KICK_MASK))
586 if (idle_cpu(cpu) && !need_resched())
590 * We can't run Idle Load Balance on this CPU for this time so we
591 * cancel it and clear NOHZ_BALANCE_KICK
593 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
597 #else /* CONFIG_NO_HZ_COMMON */
599 static inline bool got_nohz_idle_kick(void)
604 #endif /* CONFIG_NO_HZ_COMMON */
606 #ifdef CONFIG_NO_HZ_FULL
607 bool sched_can_stop_tick(struct rq *rq)
611 /* Deadline tasks, even if single, need the tick */
612 if (rq->dl.dl_nr_running)
616 * If there are more than one RR tasks, we need the tick to effect the
617 * actual RR behaviour.
619 if (rq->rt.rr_nr_running) {
620 if (rq->rt.rr_nr_running == 1)
627 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
628 * forced preemption between FIFO tasks.
630 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
635 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
636 * if there's more than one we need the tick for involuntary
639 if (rq->nr_running > 1)
644 #endif /* CONFIG_NO_HZ_FULL */
645 #endif /* CONFIG_SMP */
647 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
648 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
650 * Iterate task_group tree rooted at *from, calling @down when first entering a
651 * node and @up when leaving it for the final time.
653 * Caller must hold rcu_lock or sufficient equivalent.
655 int walk_tg_tree_from(struct task_group *from,
656 tg_visitor down, tg_visitor up, void *data)
658 struct task_group *parent, *child;
664 ret = (*down)(parent, data);
667 list_for_each_entry_rcu(child, &parent->children, siblings) {
674 ret = (*up)(parent, data);
675 if (ret || parent == from)
679 parent = parent->parent;
686 int tg_nop(struct task_group *tg, void *data)
692 static void set_load_weight(struct task_struct *p, bool update_load)
694 int prio = p->static_prio - MAX_RT_PRIO;
695 struct load_weight *load = &p->se.load;
698 * SCHED_IDLE tasks get minimal weight:
700 if (idle_policy(p->policy)) {
701 load->weight = scale_load(WEIGHT_IDLEPRIO);
702 load->inv_weight = WMULT_IDLEPRIO;
703 p->se.runnable_weight = load->weight;
708 * SCHED_OTHER tasks have to update their load when changing their
711 if (update_load && p->sched_class == &fair_sched_class) {
712 reweight_task(p, prio);
714 load->weight = scale_load(sched_prio_to_weight[prio]);
715 load->inv_weight = sched_prio_to_wmult[prio];
716 p->se.runnable_weight = load->weight;
720 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
722 if (!(flags & ENQUEUE_NOCLOCK))
725 if (!(flags & ENQUEUE_RESTORE)) {
726 sched_info_queued(rq, p);
727 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
730 p->sched_class->enqueue_task(rq, p, flags);
733 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
735 if (!(flags & DEQUEUE_NOCLOCK))
738 if (!(flags & DEQUEUE_SAVE)) {
739 sched_info_dequeued(rq, p);
740 psi_dequeue(p, flags & DEQUEUE_SLEEP);
743 p->sched_class->dequeue_task(rq, p, flags);
746 void activate_task(struct rq *rq, struct task_struct *p, int flags)
748 if (task_contributes_to_load(p))
749 rq->nr_uninterruptible--;
751 enqueue_task(rq, p, flags);
754 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
756 if (task_contributes_to_load(p))
757 rq->nr_uninterruptible++;
759 dequeue_task(rq, p, flags);
763 * __normal_prio - return the priority that is based on the static prio
765 static inline int __normal_prio(struct task_struct *p)
767 return p->static_prio;
771 * Calculate the expected normal priority: i.e. priority
772 * without taking RT-inheritance into account. Might be
773 * boosted by interactivity modifiers. Changes upon fork,
774 * setprio syscalls, and whenever the interactivity
775 * estimator recalculates.
777 static inline int normal_prio(struct task_struct *p)
781 if (task_has_dl_policy(p))
782 prio = MAX_DL_PRIO-1;
783 else if (task_has_rt_policy(p))
784 prio = MAX_RT_PRIO-1 - p->rt_priority;
786 prio = __normal_prio(p);
791 * Calculate the current priority, i.e. the priority
792 * taken into account by the scheduler. This value might
793 * be boosted by RT tasks, or might be boosted by
794 * interactivity modifiers. Will be RT if the task got
795 * RT-boosted. If not then it returns p->normal_prio.
797 static int effective_prio(struct task_struct *p)
799 p->normal_prio = normal_prio(p);
801 * If we are RT tasks or we were boosted to RT priority,
802 * keep the priority unchanged. Otherwise, update priority
803 * to the normal priority:
805 if (!rt_prio(p->prio))
806 return p->normal_prio;
811 * task_curr - is this task currently executing on a CPU?
812 * @p: the task in question.
814 * Return: 1 if the task is currently executing. 0 otherwise.
816 inline int task_curr(const struct task_struct *p)
818 return cpu_curr(task_cpu(p)) == p;
822 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
823 * use the balance_callback list if you want balancing.
825 * this means any call to check_class_changed() must be followed by a call to
826 * balance_callback().
828 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
829 const struct sched_class *prev_class,
832 if (prev_class != p->sched_class) {
833 if (prev_class->switched_from)
834 prev_class->switched_from(rq, p);
836 p->sched_class->switched_to(rq, p);
837 } else if (oldprio != p->prio || dl_task(p))
838 p->sched_class->prio_changed(rq, p, oldprio);
841 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
843 const struct sched_class *class;
845 if (p->sched_class == rq->curr->sched_class) {
846 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
848 for_each_class(class) {
849 if (class == rq->curr->sched_class)
851 if (class == p->sched_class) {
859 * A queue event has occurred, and we're going to schedule. In
860 * this case, we can save a useless back to back clock update.
862 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
863 rq_clock_skip_update(rq);
868 static inline bool is_per_cpu_kthread(struct task_struct *p)
870 if (!(p->flags & PF_KTHREAD))
873 if (p->nr_cpus_allowed != 1)
880 * Per-CPU kthreads are allowed to run on !actie && online CPUs, see
881 * __set_cpus_allowed_ptr() and select_fallback_rq().
883 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
885 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
888 if (is_per_cpu_kthread(p))
889 return cpu_online(cpu);
891 return cpu_active(cpu);
895 * This is how migration works:
897 * 1) we invoke migration_cpu_stop() on the target CPU using
899 * 2) stopper starts to run (implicitly forcing the migrated thread
901 * 3) it checks whether the migrated task is still in the wrong runqueue.
902 * 4) if it's in the wrong runqueue then the migration thread removes
903 * it and puts it into the right queue.
904 * 5) stopper completes and stop_one_cpu() returns and the migration
909 * move_queued_task - move a queued task to new rq.
911 * Returns (locked) new rq. Old rq's lock is released.
913 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
914 struct task_struct *p, int new_cpu)
916 lockdep_assert_held(&rq->lock);
918 p->on_rq = TASK_ON_RQ_MIGRATING;
919 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
920 set_task_cpu(p, new_cpu);
923 rq = cpu_rq(new_cpu);
926 BUG_ON(task_cpu(p) != new_cpu);
927 enqueue_task(rq, p, 0);
928 p->on_rq = TASK_ON_RQ_QUEUED;
929 check_preempt_curr(rq, p, 0);
934 struct migration_arg {
935 struct task_struct *task;
940 * Move (not current) task off this CPU, onto the destination CPU. We're doing
941 * this because either it can't run here any more (set_cpus_allowed()
942 * away from this CPU, or CPU going down), or because we're
943 * attempting to rebalance this task on exec (sched_exec).
945 * So we race with normal scheduler movements, but that's OK, as long
946 * as the task is no longer on this CPU.
948 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
949 struct task_struct *p, int dest_cpu)
951 /* Affinity changed (again). */
952 if (!is_cpu_allowed(p, dest_cpu))
956 rq = move_queued_task(rq, rf, p, dest_cpu);
962 * migration_cpu_stop - this will be executed by a highprio stopper thread
963 * and performs thread migration by bumping thread off CPU then
964 * 'pushing' onto another runqueue.
966 static int migration_cpu_stop(void *data)
968 struct migration_arg *arg = data;
969 struct task_struct *p = arg->task;
970 struct rq *rq = this_rq();
974 * The original target CPU might have gone down and we might
975 * be on another CPU but it doesn't matter.
979 * We need to explicitly wake pending tasks before running
980 * __migrate_task() such that we will not miss enforcing cpus_allowed
981 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
983 sched_ttwu_pending();
985 raw_spin_lock(&p->pi_lock);
988 * If task_rq(p) != rq, it cannot be migrated here, because we're
989 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
990 * we're holding p->pi_lock.
992 if (task_rq(p) == rq) {
993 if (task_on_rq_queued(p))
994 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
996 p->wake_cpu = arg->dest_cpu;
999 raw_spin_unlock(&p->pi_lock);
1006 * sched_class::set_cpus_allowed must do the below, but is not required to
1007 * actually call this function.
1009 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1011 cpumask_copy(&p->cpus_allowed, new_mask);
1012 p->nr_cpus_allowed = cpumask_weight(new_mask);
1015 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1017 struct rq *rq = task_rq(p);
1018 bool queued, running;
1020 lockdep_assert_held(&p->pi_lock);
1022 queued = task_on_rq_queued(p);
1023 running = task_current(rq, p);
1027 * Because __kthread_bind() calls this on blocked tasks without
1030 lockdep_assert_held(&rq->lock);
1031 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1034 put_prev_task(rq, p);
1036 p->sched_class->set_cpus_allowed(p, new_mask);
1039 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1041 set_curr_task(rq, p);
1045 * Change a given task's CPU affinity. Migrate the thread to a
1046 * proper CPU and schedule it away if the CPU it's executing on
1047 * is removed from the allowed bitmask.
1049 * NOTE: the caller must have a valid reference to the task, the
1050 * task must not exit() & deallocate itself prematurely. The
1051 * call is not atomic; no spinlocks may be held.
1053 static int __set_cpus_allowed_ptr(struct task_struct *p,
1054 const struct cpumask *new_mask, bool check)
1056 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1057 unsigned int dest_cpu;
1062 rq = task_rq_lock(p, &rf);
1063 update_rq_clock(rq);
1065 if (p->flags & PF_KTHREAD) {
1067 * Kernel threads are allowed on online && !active CPUs
1069 cpu_valid_mask = cpu_online_mask;
1073 * Must re-check here, to close a race against __kthread_bind(),
1074 * sched_setaffinity() is not guaranteed to observe the flag.
1076 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1081 if (cpumask_equal(&p->cpus_allowed, new_mask))
1084 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1089 do_set_cpus_allowed(p, new_mask);
1091 if (p->flags & PF_KTHREAD) {
1093 * For kernel threads that do indeed end up on online &&
1094 * !active we want to ensure they are strict per-CPU threads.
1096 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1097 !cpumask_intersects(new_mask, cpu_active_mask) &&
1098 p->nr_cpus_allowed != 1);
1101 /* Can the task run on the task's current CPU? If so, we're done */
1102 if (cpumask_test_cpu(task_cpu(p), new_mask))
1105 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1106 if (task_running(rq, p) || p->state == TASK_WAKING) {
1107 struct migration_arg arg = { p, dest_cpu };
1108 /* Need help from migration thread: drop lock and wait. */
1109 task_rq_unlock(rq, p, &rf);
1110 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1111 tlb_migrate_finish(p->mm);
1113 } else if (task_on_rq_queued(p)) {
1115 * OK, since we're going to drop the lock immediately
1116 * afterwards anyway.
1118 rq = move_queued_task(rq, &rf, p, dest_cpu);
1121 task_rq_unlock(rq, p, &rf);
1126 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1128 return __set_cpus_allowed_ptr(p, new_mask, false);
1130 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1132 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1134 #ifdef CONFIG_SCHED_DEBUG
1136 * We should never call set_task_cpu() on a blocked task,
1137 * ttwu() will sort out the placement.
1139 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1143 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1144 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1145 * time relying on p->on_rq.
1147 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1148 p->sched_class == &fair_sched_class &&
1149 (p->on_rq && !task_on_rq_migrating(p)));
1151 #ifdef CONFIG_LOCKDEP
1153 * The caller should hold either p->pi_lock or rq->lock, when changing
1154 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1156 * sched_move_task() holds both and thus holding either pins the cgroup,
1159 * Furthermore, all task_rq users should acquire both locks, see
1162 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1163 lockdep_is_held(&task_rq(p)->lock)));
1166 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1168 WARN_ON_ONCE(!cpu_online(new_cpu));
1171 trace_sched_migrate_task(p, new_cpu);
1173 if (task_cpu(p) != new_cpu) {
1174 if (p->sched_class->migrate_task_rq)
1175 p->sched_class->migrate_task_rq(p, new_cpu);
1176 p->se.nr_migrations++;
1178 perf_event_task_migrate(p);
1181 __set_task_cpu(p, new_cpu);
1184 #ifdef CONFIG_NUMA_BALANCING
1185 static void __migrate_swap_task(struct task_struct *p, int cpu)
1187 if (task_on_rq_queued(p)) {
1188 struct rq *src_rq, *dst_rq;
1189 struct rq_flags srf, drf;
1191 src_rq = task_rq(p);
1192 dst_rq = cpu_rq(cpu);
1194 rq_pin_lock(src_rq, &srf);
1195 rq_pin_lock(dst_rq, &drf);
1197 p->on_rq = TASK_ON_RQ_MIGRATING;
1198 deactivate_task(src_rq, p, 0);
1199 set_task_cpu(p, cpu);
1200 activate_task(dst_rq, p, 0);
1201 p->on_rq = TASK_ON_RQ_QUEUED;
1202 check_preempt_curr(dst_rq, p, 0);
1204 rq_unpin_lock(dst_rq, &drf);
1205 rq_unpin_lock(src_rq, &srf);
1209 * Task isn't running anymore; make it appear like we migrated
1210 * it before it went to sleep. This means on wakeup we make the
1211 * previous CPU our target instead of where it really is.
1217 struct migration_swap_arg {
1218 struct task_struct *src_task, *dst_task;
1219 int src_cpu, dst_cpu;
1222 static int migrate_swap_stop(void *data)
1224 struct migration_swap_arg *arg = data;
1225 struct rq *src_rq, *dst_rq;
1228 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1231 src_rq = cpu_rq(arg->src_cpu);
1232 dst_rq = cpu_rq(arg->dst_cpu);
1234 double_raw_lock(&arg->src_task->pi_lock,
1235 &arg->dst_task->pi_lock);
1236 double_rq_lock(src_rq, dst_rq);
1238 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1241 if (task_cpu(arg->src_task) != arg->src_cpu)
1244 if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1247 if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1250 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1251 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1256 double_rq_unlock(src_rq, dst_rq);
1257 raw_spin_unlock(&arg->dst_task->pi_lock);
1258 raw_spin_unlock(&arg->src_task->pi_lock);
1264 * Cross migrate two tasks
1266 int migrate_swap(struct task_struct *cur, struct task_struct *p,
1267 int target_cpu, int curr_cpu)
1269 struct migration_swap_arg arg;
1272 arg = (struct migration_swap_arg){
1274 .src_cpu = curr_cpu,
1276 .dst_cpu = target_cpu,
1279 if (arg.src_cpu == arg.dst_cpu)
1283 * These three tests are all lockless; this is OK since all of them
1284 * will be re-checked with proper locks held further down the line.
1286 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1289 if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1292 if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1295 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1296 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1301 #endif /* CONFIG_NUMA_BALANCING */
1304 * wait_task_inactive - wait for a thread to unschedule.
1306 * If @match_state is nonzero, it's the @p->state value just checked and
1307 * not expected to change. If it changes, i.e. @p might have woken up,
1308 * then return zero. When we succeed in waiting for @p to be off its CPU,
1309 * we return a positive number (its total switch count). If a second call
1310 * a short while later returns the same number, the caller can be sure that
1311 * @p has remained unscheduled the whole time.
1313 * The caller must ensure that the task *will* unschedule sometime soon,
1314 * else this function might spin for a *long* time. This function can't
1315 * be called with interrupts off, or it may introduce deadlock with
1316 * smp_call_function() if an IPI is sent by the same process we are
1317 * waiting to become inactive.
1319 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1321 int running, queued;
1328 * We do the initial early heuristics without holding
1329 * any task-queue locks at all. We'll only try to get
1330 * the runqueue lock when things look like they will
1336 * If the task is actively running on another CPU
1337 * still, just relax and busy-wait without holding
1340 * NOTE! Since we don't hold any locks, it's not
1341 * even sure that "rq" stays as the right runqueue!
1342 * But we don't care, since "task_running()" will
1343 * return false if the runqueue has changed and p
1344 * is actually now running somewhere else!
1346 while (task_running(rq, p)) {
1347 if (match_state && unlikely(p->state != match_state))
1353 * Ok, time to look more closely! We need the rq
1354 * lock now, to be *sure*. If we're wrong, we'll
1355 * just go back and repeat.
1357 rq = task_rq_lock(p, &rf);
1358 trace_sched_wait_task(p);
1359 running = task_running(rq, p);
1360 queued = task_on_rq_queued(p);
1362 if (!match_state || p->state == match_state)
1363 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1364 task_rq_unlock(rq, p, &rf);
1367 * If it changed from the expected state, bail out now.
1369 if (unlikely(!ncsw))
1373 * Was it really running after all now that we
1374 * checked with the proper locks actually held?
1376 * Oops. Go back and try again..
1378 if (unlikely(running)) {
1384 * It's not enough that it's not actively running,
1385 * it must be off the runqueue _entirely_, and not
1388 * So if it was still runnable (but just not actively
1389 * running right now), it's preempted, and we should
1390 * yield - it could be a while.
1392 if (unlikely(queued)) {
1393 ktime_t to = NSEC_PER_SEC / HZ;
1395 set_current_state(TASK_UNINTERRUPTIBLE);
1396 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1401 * Ahh, all good. It wasn't running, and it wasn't
1402 * runnable, which means that it will never become
1403 * running in the future either. We're all done!
1412 * kick_process - kick a running thread to enter/exit the kernel
1413 * @p: the to-be-kicked thread
1415 * Cause a process which is running on another CPU to enter
1416 * kernel-mode, without any delay. (to get signals handled.)
1418 * NOTE: this function doesn't have to take the runqueue lock,
1419 * because all it wants to ensure is that the remote task enters
1420 * the kernel. If the IPI races and the task has been migrated
1421 * to another CPU then no harm is done and the purpose has been
1424 void kick_process(struct task_struct *p)
1430 if ((cpu != smp_processor_id()) && task_curr(p))
1431 smp_send_reschedule(cpu);
1434 EXPORT_SYMBOL_GPL(kick_process);
1437 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1439 * A few notes on cpu_active vs cpu_online:
1441 * - cpu_active must be a subset of cpu_online
1443 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
1444 * see __set_cpus_allowed_ptr(). At this point the newly online
1445 * CPU isn't yet part of the sched domains, and balancing will not
1448 * - on CPU-down we clear cpu_active() to mask the sched domains and
1449 * avoid the load balancer to place new tasks on the to be removed
1450 * CPU. Existing tasks will remain running there and will be taken
1453 * This means that fallback selection must not select !active CPUs.
1454 * And can assume that any active CPU must be online. Conversely
1455 * select_task_rq() below may allow selection of !active CPUs in order
1456 * to satisfy the above rules.
1458 static int select_fallback_rq(int cpu, struct task_struct *p)
1460 int nid = cpu_to_node(cpu);
1461 const struct cpumask *nodemask = NULL;
1462 enum { cpuset, possible, fail } state = cpuset;
1466 * If the node that the CPU is on has been offlined, cpu_to_node()
1467 * will return -1. There is no CPU on the node, and we should
1468 * select the CPU on the other node.
1471 nodemask = cpumask_of_node(nid);
1473 /* Look for allowed, online CPU in same node. */
1474 for_each_cpu(dest_cpu, nodemask) {
1475 if (!cpu_active(dest_cpu))
1477 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1483 /* Any allowed, online CPU? */
1484 for_each_cpu(dest_cpu, &p->cpus_allowed) {
1485 if (!is_cpu_allowed(p, dest_cpu))
1491 /* No more Mr. Nice Guy. */
1494 if (IS_ENABLED(CONFIG_CPUSETS)) {
1495 cpuset_cpus_allowed_fallback(p);
1501 do_set_cpus_allowed(p, cpu_possible_mask);
1512 if (state != cpuset) {
1514 * Don't tell them about moving exiting tasks or
1515 * kernel threads (both mm NULL), since they never
1518 if (p->mm && printk_ratelimit()) {
1519 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1520 task_pid_nr(p), p->comm, cpu);
1528 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1531 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1533 lockdep_assert_held(&p->pi_lock);
1535 if (p->nr_cpus_allowed > 1)
1536 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1538 cpu = cpumask_any(&p->cpus_allowed);
1541 * In order not to call set_task_cpu() on a blocking task we need
1542 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1545 * Since this is common to all placement strategies, this lives here.
1547 * [ this allows ->select_task() to simply return task_cpu(p) and
1548 * not worry about this generic constraint ]
1550 if (unlikely(!is_cpu_allowed(p, cpu)))
1551 cpu = select_fallback_rq(task_cpu(p), p);
1556 static void update_avg(u64 *avg, u64 sample)
1558 s64 diff = sample - *avg;
1562 void sched_set_stop_task(int cpu, struct task_struct *stop)
1564 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1565 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1569 * Make it appear like a SCHED_FIFO task, its something
1570 * userspace knows about and won't get confused about.
1572 * Also, it will make PI more or less work without too
1573 * much confusion -- but then, stop work should not
1574 * rely on PI working anyway.
1576 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1578 stop->sched_class = &stop_sched_class;
1581 cpu_rq(cpu)->stop = stop;
1585 * Reset it back to a normal scheduling class so that
1586 * it can die in pieces.
1588 old_stop->sched_class = &rt_sched_class;
1594 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1595 const struct cpumask *new_mask, bool check)
1597 return set_cpus_allowed_ptr(p, new_mask);
1600 #endif /* CONFIG_SMP */
1603 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1607 if (!schedstat_enabled())
1613 if (cpu == rq->cpu) {
1614 __schedstat_inc(rq->ttwu_local);
1615 __schedstat_inc(p->se.statistics.nr_wakeups_local);
1617 struct sched_domain *sd;
1619 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
1621 for_each_domain(rq->cpu, sd) {
1622 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1623 __schedstat_inc(sd->ttwu_wake_remote);
1630 if (wake_flags & WF_MIGRATED)
1631 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1632 #endif /* CONFIG_SMP */
1634 __schedstat_inc(rq->ttwu_count);
1635 __schedstat_inc(p->se.statistics.nr_wakeups);
1637 if (wake_flags & WF_SYNC)
1638 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
1641 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1643 activate_task(rq, p, en_flags);
1644 p->on_rq = TASK_ON_RQ_QUEUED;
1646 /* If a worker is waking up, notify the workqueue: */
1647 if (p->flags & PF_WQ_WORKER)
1648 wq_worker_waking_up(p, cpu_of(rq));
1652 * Mark the task runnable and perform wakeup-preemption.
1654 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1655 struct rq_flags *rf)
1657 check_preempt_curr(rq, p, wake_flags);
1658 p->state = TASK_RUNNING;
1659 trace_sched_wakeup(p);
1662 if (p->sched_class->task_woken) {
1664 * Our task @p is fully woken up and running; so its safe to
1665 * drop the rq->lock, hereafter rq is only used for statistics.
1667 rq_unpin_lock(rq, rf);
1668 p->sched_class->task_woken(rq, p);
1669 rq_repin_lock(rq, rf);
1672 if (rq->idle_stamp) {
1673 u64 delta = rq_clock(rq) - rq->idle_stamp;
1674 u64 max = 2*rq->max_idle_balance_cost;
1676 update_avg(&rq->avg_idle, delta);
1678 if (rq->avg_idle > max)
1687 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1688 struct rq_flags *rf)
1690 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
1692 lockdep_assert_held(&rq->lock);
1695 if (p->sched_contributes_to_load)
1696 rq->nr_uninterruptible--;
1698 if (wake_flags & WF_MIGRATED)
1699 en_flags |= ENQUEUE_MIGRATED;
1702 ttwu_activate(rq, p, en_flags);
1703 ttwu_do_wakeup(rq, p, wake_flags, rf);
1707 * Called in case the task @p isn't fully descheduled from its runqueue,
1708 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1709 * since all we need to do is flip p->state to TASK_RUNNING, since
1710 * the task is still ->on_rq.
1712 static int ttwu_remote(struct task_struct *p, int wake_flags)
1718 rq = __task_rq_lock(p, &rf);
1719 if (task_on_rq_queued(p)) {
1720 /* check_preempt_curr() may use rq clock */
1721 update_rq_clock(rq);
1722 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1725 __task_rq_unlock(rq, &rf);
1731 void sched_ttwu_pending(void)
1733 struct rq *rq = this_rq();
1734 struct llist_node *llist = llist_del_all(&rq->wake_list);
1735 struct task_struct *p, *t;
1741 rq_lock_irqsave(rq, &rf);
1742 update_rq_clock(rq);
1744 llist_for_each_entry_safe(p, t, llist, wake_entry)
1745 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
1747 rq_unlock_irqrestore(rq, &rf);
1750 void scheduler_ipi(void)
1753 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1754 * TIF_NEED_RESCHED remotely (for the first time) will also send
1757 preempt_fold_need_resched();
1759 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1763 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1764 * traditionally all their work was done from the interrupt return
1765 * path. Now that we actually do some work, we need to make sure
1768 * Some archs already do call them, luckily irq_enter/exit nest
1771 * Arguably we should visit all archs and update all handlers,
1772 * however a fair share of IPIs are still resched only so this would
1773 * somewhat pessimize the simple resched case.
1776 sched_ttwu_pending();
1779 * Check if someone kicked us for doing the nohz idle load balance.
1781 if (unlikely(got_nohz_idle_kick())) {
1782 this_rq()->idle_balance = 1;
1783 raise_softirq_irqoff(SCHED_SOFTIRQ);
1788 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1790 struct rq *rq = cpu_rq(cpu);
1792 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1794 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1795 if (!set_nr_if_polling(rq->idle))
1796 smp_send_reschedule(cpu);
1798 trace_sched_wake_idle_without_ipi(cpu);
1802 void wake_up_if_idle(int cpu)
1804 struct rq *rq = cpu_rq(cpu);
1809 if (!is_idle_task(rcu_dereference(rq->curr)))
1812 if (set_nr_if_polling(rq->idle)) {
1813 trace_sched_wake_idle_without_ipi(cpu);
1815 rq_lock_irqsave(rq, &rf);
1816 if (is_idle_task(rq->curr))
1817 smp_send_reschedule(cpu);
1818 /* Else CPU is not idle, do nothing here: */
1819 rq_unlock_irqrestore(rq, &rf);
1826 bool cpus_share_cache(int this_cpu, int that_cpu)
1828 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1830 #endif /* CONFIG_SMP */
1832 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1834 struct rq *rq = cpu_rq(cpu);
1837 #if defined(CONFIG_SMP)
1838 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1839 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1840 ttwu_queue_remote(p, cpu, wake_flags);
1846 update_rq_clock(rq);
1847 ttwu_do_activate(rq, p, wake_flags, &rf);
1852 * Notes on Program-Order guarantees on SMP systems.
1856 * The basic program-order guarantee on SMP systems is that when a task [t]
1857 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1858 * execution on its new CPU [c1].
1860 * For migration (of runnable tasks) this is provided by the following means:
1862 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1863 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1864 * rq(c1)->lock (if not at the same time, then in that order).
1865 * C) LOCK of the rq(c1)->lock scheduling in task
1867 * Release/acquire chaining guarantees that B happens after A and C after B.
1868 * Note: the CPU doing B need not be c0 or c1
1877 * UNLOCK rq(0)->lock
1879 * LOCK rq(0)->lock // orders against CPU0
1881 * UNLOCK rq(0)->lock
1885 * UNLOCK rq(1)->lock
1887 * LOCK rq(1)->lock // orders against CPU2
1890 * UNLOCK rq(1)->lock
1893 * BLOCKING -- aka. SLEEP + WAKEUP
1895 * For blocking we (obviously) need to provide the same guarantee as for
1896 * migration. However the means are completely different as there is no lock
1897 * chain to provide order. Instead we do:
1899 * 1) smp_store_release(X->on_cpu, 0)
1900 * 2) smp_cond_load_acquire(!X->on_cpu)
1904 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1906 * LOCK rq(0)->lock LOCK X->pi_lock
1909 * smp_store_release(X->on_cpu, 0);
1911 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1917 * X->state = RUNNING
1918 * UNLOCK rq(2)->lock
1920 * LOCK rq(2)->lock // orders against CPU1
1923 * UNLOCK rq(2)->lock
1926 * UNLOCK rq(0)->lock
1929 * However, for wakeups there is a second guarantee we must provide, namely we
1930 * must ensure that CONDITION=1 done by the caller can not be reordered with
1931 * accesses to the task state; see try_to_wake_up() and set_current_state().
1935 * try_to_wake_up - wake up a thread
1936 * @p: the thread to be awakened
1937 * @state: the mask of task states that can be woken
1938 * @wake_flags: wake modifier flags (WF_*)
1940 * If (@state & @p->state) @p->state = TASK_RUNNING.
1942 * If the task was not queued/runnable, also place it back on a runqueue.
1944 * Atomic against schedule() which would dequeue a task, also see
1945 * set_current_state().
1947 * This function executes a full memory barrier before accessing the task
1948 * state; see set_current_state().
1950 * Return: %true if @p->state changes (an actual wakeup was done),
1954 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1956 unsigned long flags;
1957 int cpu, success = 0;
1960 * If we are going to wake up a thread waiting for CONDITION we
1961 * need to ensure that CONDITION=1 done by the caller can not be
1962 * reordered with p->state check below. This pairs with mb() in
1963 * set_current_state() the waiting thread does.
1965 raw_spin_lock_irqsave(&p->pi_lock, flags);
1966 smp_mb__after_spinlock();
1967 if (!(p->state & state))
1970 trace_sched_waking(p);
1972 /* We're going to change ->state: */
1977 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1978 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1979 * in smp_cond_load_acquire() below.
1981 * sched_ttwu_pending() try_to_wake_up()
1982 * STORE p->on_rq = 1 LOAD p->state
1985 * __schedule() (switch to task 'p')
1986 * LOCK rq->lock smp_rmb();
1987 * smp_mb__after_spinlock();
1991 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
1993 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
1994 * __schedule(). See the comment for smp_mb__after_spinlock().
1997 if (p->on_rq && ttwu_remote(p, wake_flags))
2002 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2003 * possible to, falsely, observe p->on_cpu == 0.
2005 * One must be running (->on_cpu == 1) in order to remove oneself
2006 * from the runqueue.
2008 * __schedule() (switch to task 'p') try_to_wake_up()
2009 * STORE p->on_cpu = 1 LOAD p->on_rq
2012 * __schedule() (put 'p' to sleep)
2013 * LOCK rq->lock smp_rmb();
2014 * smp_mb__after_spinlock();
2015 * STORE p->on_rq = 0 LOAD p->on_cpu
2017 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2018 * __schedule(). See the comment for smp_mb__after_spinlock().
2023 * If the owning (remote) CPU is still in the middle of schedule() with
2024 * this task as prev, wait until its done referencing the task.
2026 * Pairs with the smp_store_release() in finish_task().
2028 * This ensures that tasks getting woken will be fully ordered against
2029 * their previous state and preserve Program Order.
2031 smp_cond_load_acquire(&p->on_cpu, !VAL);
2033 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2034 p->state = TASK_WAKING;
2037 delayacct_blkio_end(p);
2038 atomic_dec(&task_rq(p)->nr_iowait);
2041 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2042 if (task_cpu(p) != cpu) {
2043 wake_flags |= WF_MIGRATED;
2044 psi_ttwu_dequeue(p);
2045 set_task_cpu(p, cpu);
2048 #else /* CONFIG_SMP */
2051 delayacct_blkio_end(p);
2052 atomic_dec(&task_rq(p)->nr_iowait);
2055 #endif /* CONFIG_SMP */
2057 ttwu_queue(p, cpu, wake_flags);
2059 ttwu_stat(p, cpu, wake_flags);
2061 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2067 * try_to_wake_up_local - try to wake up a local task with rq lock held
2068 * @p: the thread to be awakened
2069 * @rf: request-queue flags for pinning
2071 * Put @p on the run-queue if it's not already there. The caller must
2072 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2075 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2077 struct rq *rq = task_rq(p);
2079 if (WARN_ON_ONCE(rq != this_rq()) ||
2080 WARN_ON_ONCE(p == current))
2083 lockdep_assert_held(&rq->lock);
2085 if (!raw_spin_trylock(&p->pi_lock)) {
2087 * This is OK, because current is on_cpu, which avoids it being
2088 * picked for load-balance and preemption/IRQs are still
2089 * disabled avoiding further scheduler activity on it and we've
2090 * not yet picked a replacement task.
2093 raw_spin_lock(&p->pi_lock);
2097 if (!(p->state & TASK_NORMAL))
2100 trace_sched_waking(p);
2102 if (!task_on_rq_queued(p)) {
2104 delayacct_blkio_end(p);
2105 atomic_dec(&rq->nr_iowait);
2107 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK);
2110 ttwu_do_wakeup(rq, p, 0, rf);
2111 ttwu_stat(p, smp_processor_id(), 0);
2113 raw_spin_unlock(&p->pi_lock);
2117 * wake_up_process - Wake up a specific process
2118 * @p: The process to be woken up.
2120 * Attempt to wake up the nominated process and move it to the set of runnable
2123 * Return: 1 if the process was woken up, 0 if it was already running.
2125 * This function executes a full memory barrier before accessing the task state.
2127 int wake_up_process(struct task_struct *p)
2129 return try_to_wake_up(p, TASK_NORMAL, 0);
2131 EXPORT_SYMBOL(wake_up_process);
2133 int wake_up_state(struct task_struct *p, unsigned int state)
2135 return try_to_wake_up(p, state, 0);
2139 * Perform scheduler related setup for a newly forked process p.
2140 * p is forked by current.
2142 * __sched_fork() is basic setup used by init_idle() too:
2144 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2149 p->se.exec_start = 0;
2150 p->se.sum_exec_runtime = 0;
2151 p->se.prev_sum_exec_runtime = 0;
2152 p->se.nr_migrations = 0;
2154 INIT_LIST_HEAD(&p->se.group_node);
2156 #ifdef CONFIG_FAIR_GROUP_SCHED
2157 p->se.cfs_rq = NULL;
2160 #ifdef CONFIG_SCHEDSTATS
2161 /* Even if schedstat is disabled, there should not be garbage */
2162 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2165 RB_CLEAR_NODE(&p->dl.rb_node);
2166 init_dl_task_timer(&p->dl);
2167 init_dl_inactive_task_timer(&p->dl);
2168 __dl_clear_params(p);
2170 INIT_LIST_HEAD(&p->rt.run_list);
2172 p->rt.time_slice = sched_rr_timeslice;
2176 #ifdef CONFIG_PREEMPT_NOTIFIERS
2177 INIT_HLIST_HEAD(&p->preempt_notifiers);
2180 init_numa_balancing(clone_flags, p);
2183 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2185 #ifdef CONFIG_NUMA_BALANCING
2187 void set_numabalancing_state(bool enabled)
2190 static_branch_enable(&sched_numa_balancing);
2192 static_branch_disable(&sched_numa_balancing);
2195 #ifdef CONFIG_PROC_SYSCTL
2196 int sysctl_numa_balancing(struct ctl_table *table, int write,
2197 void __user *buffer, size_t *lenp, loff_t *ppos)
2201 int state = static_branch_likely(&sched_numa_balancing);
2203 if (write && !capable(CAP_SYS_ADMIN))
2208 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2212 set_numabalancing_state(state);
2218 #ifdef CONFIG_SCHEDSTATS
2220 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2221 static bool __initdata __sched_schedstats = false;
2223 static void set_schedstats(bool enabled)
2226 static_branch_enable(&sched_schedstats);
2228 static_branch_disable(&sched_schedstats);
2231 void force_schedstat_enabled(void)
2233 if (!schedstat_enabled()) {
2234 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2235 static_branch_enable(&sched_schedstats);
2239 static int __init setup_schedstats(char *str)
2246 * This code is called before jump labels have been set up, so we can't
2247 * change the static branch directly just yet. Instead set a temporary
2248 * variable so init_schedstats() can do it later.
2250 if (!strcmp(str, "enable")) {
2251 __sched_schedstats = true;
2253 } else if (!strcmp(str, "disable")) {
2254 __sched_schedstats = false;
2259 pr_warn("Unable to parse schedstats=\n");
2263 __setup("schedstats=", setup_schedstats);
2265 static void __init init_schedstats(void)
2267 set_schedstats(__sched_schedstats);
2270 #ifdef CONFIG_PROC_SYSCTL
2271 int sysctl_schedstats(struct ctl_table *table, int write,
2272 void __user *buffer, size_t *lenp, loff_t *ppos)
2276 int state = static_branch_likely(&sched_schedstats);
2278 if (write && !capable(CAP_SYS_ADMIN))
2283 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2287 set_schedstats(state);
2290 #endif /* CONFIG_PROC_SYSCTL */
2291 #else /* !CONFIG_SCHEDSTATS */
2292 static inline void init_schedstats(void) {}
2293 #endif /* CONFIG_SCHEDSTATS */
2296 * fork()/clone()-time setup:
2298 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2300 unsigned long flags;
2302 __sched_fork(clone_flags, p);
2304 * We mark the process as NEW here. This guarantees that
2305 * nobody will actually run it, and a signal or other external
2306 * event cannot wake it up and insert it on the runqueue either.
2308 p->state = TASK_NEW;
2311 * Make sure we do not leak PI boosting priority to the child.
2313 p->prio = current->normal_prio;
2316 * Revert to default priority/policy on fork if requested.
2318 if (unlikely(p->sched_reset_on_fork)) {
2319 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2320 p->policy = SCHED_NORMAL;
2321 p->static_prio = NICE_TO_PRIO(0);
2323 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2324 p->static_prio = NICE_TO_PRIO(0);
2326 p->prio = p->normal_prio = __normal_prio(p);
2327 set_load_weight(p, false);
2330 * We don't need the reset flag anymore after the fork. It has
2331 * fulfilled its duty:
2333 p->sched_reset_on_fork = 0;
2336 if (dl_prio(p->prio))
2338 else if (rt_prio(p->prio))
2339 p->sched_class = &rt_sched_class;
2341 p->sched_class = &fair_sched_class;
2343 init_entity_runnable_average(&p->se);
2346 * The child is not yet in the pid-hash so no cgroup attach races,
2347 * and the cgroup is pinned to this child due to cgroup_fork()
2348 * is ran before sched_fork().
2350 * Silence PROVE_RCU.
2352 raw_spin_lock_irqsave(&p->pi_lock, flags);
2354 * We're setting the CPU for the first time, we don't migrate,
2355 * so use __set_task_cpu().
2357 __set_task_cpu(p, smp_processor_id());
2358 if (p->sched_class->task_fork)
2359 p->sched_class->task_fork(p);
2360 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2362 #ifdef CONFIG_SCHED_INFO
2363 if (likely(sched_info_on()))
2364 memset(&p->sched_info, 0, sizeof(p->sched_info));
2366 #if defined(CONFIG_SMP)
2369 init_task_preempt_count(p);
2371 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2372 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2377 unsigned long to_ratio(u64 period, u64 runtime)
2379 if (runtime == RUNTIME_INF)
2383 * Doing this here saves a lot of checks in all
2384 * the calling paths, and returning zero seems
2385 * safe for them anyway.
2390 return div64_u64(runtime << BW_SHIFT, period);
2394 * wake_up_new_task - wake up a newly created task for the first time.
2396 * This function will do some initial scheduler statistics housekeeping
2397 * that must be done for every newly created context, then puts the task
2398 * on the runqueue and wakes it.
2400 void wake_up_new_task(struct task_struct *p)
2405 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2406 p->state = TASK_RUNNING;
2409 * Fork balancing, do it here and not earlier because:
2410 * - cpus_allowed can change in the fork path
2411 * - any previously selected CPU might disappear through hotplug
2413 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2414 * as we're not fully set-up yet.
2416 p->recent_used_cpu = task_cpu(p);
2417 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2419 rq = __task_rq_lock(p, &rf);
2420 update_rq_clock(rq);
2421 post_init_entity_util_avg(&p->se);
2423 activate_task(rq, p, ENQUEUE_NOCLOCK);
2424 p->on_rq = TASK_ON_RQ_QUEUED;
2425 trace_sched_wakeup_new(p);
2426 check_preempt_curr(rq, p, WF_FORK);
2428 if (p->sched_class->task_woken) {
2430 * Nothing relies on rq->lock after this, so its fine to
2433 rq_unpin_lock(rq, &rf);
2434 p->sched_class->task_woken(rq, p);
2435 rq_repin_lock(rq, &rf);
2438 task_rq_unlock(rq, p, &rf);
2441 #ifdef CONFIG_PREEMPT_NOTIFIERS
2443 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2445 void preempt_notifier_inc(void)
2447 static_branch_inc(&preempt_notifier_key);
2449 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2451 void preempt_notifier_dec(void)
2453 static_branch_dec(&preempt_notifier_key);
2455 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2458 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2459 * @notifier: notifier struct to register
2461 void preempt_notifier_register(struct preempt_notifier *notifier)
2463 if (!static_branch_unlikely(&preempt_notifier_key))
2464 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2466 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2468 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2471 * preempt_notifier_unregister - no longer interested in preemption notifications
2472 * @notifier: notifier struct to unregister
2474 * This is *not* safe to call from within a preemption notifier.
2476 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2478 hlist_del(¬ifier->link);
2480 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2482 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2484 struct preempt_notifier *notifier;
2486 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2487 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2490 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2492 if (static_branch_unlikely(&preempt_notifier_key))
2493 __fire_sched_in_preempt_notifiers(curr);
2497 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2498 struct task_struct *next)
2500 struct preempt_notifier *notifier;
2502 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2503 notifier->ops->sched_out(notifier, next);
2506 static __always_inline void
2507 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2508 struct task_struct *next)
2510 if (static_branch_unlikely(&preempt_notifier_key))
2511 __fire_sched_out_preempt_notifiers(curr, next);
2514 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2516 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2521 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2522 struct task_struct *next)
2526 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2528 static inline void prepare_task(struct task_struct *next)
2532 * Claim the task as running, we do this before switching to it
2533 * such that any running task will have this set.
2539 static inline void finish_task(struct task_struct *prev)
2543 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2544 * We must ensure this doesn't happen until the switch is completely
2547 * In particular, the load of prev->state in finish_task_switch() must
2548 * happen before this.
2550 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2552 smp_store_release(&prev->on_cpu, 0);
2557 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
2560 * Since the runqueue lock will be released by the next
2561 * task (which is an invalid locking op but in the case
2562 * of the scheduler it's an obvious special-case), so we
2563 * do an early lockdep release here:
2565 rq_unpin_lock(rq, rf);
2566 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2567 #ifdef CONFIG_DEBUG_SPINLOCK
2568 /* this is a valid case when another task releases the spinlock */
2569 rq->lock.owner = next;
2573 static inline void finish_lock_switch(struct rq *rq)
2576 * If we are tracking spinlock dependencies then we have to
2577 * fix up the runqueue lock - which gets 'carried over' from
2578 * prev into current:
2580 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
2581 raw_spin_unlock_irq(&rq->lock);
2585 * NOP if the arch has not defined these:
2588 #ifndef prepare_arch_switch
2589 # define prepare_arch_switch(next) do { } while (0)
2592 #ifndef finish_arch_post_lock_switch
2593 # define finish_arch_post_lock_switch() do { } while (0)
2597 * prepare_task_switch - prepare to switch tasks
2598 * @rq: the runqueue preparing to switch
2599 * @prev: the current task that is being switched out
2600 * @next: the task we are going to switch to.
2602 * This is called with the rq lock held and interrupts off. It must
2603 * be paired with a subsequent finish_task_switch after the context
2606 * prepare_task_switch sets up locking and calls architecture specific
2610 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2611 struct task_struct *next)
2613 kcov_prepare_switch(prev);
2614 sched_info_switch(rq, prev, next);
2615 perf_event_task_sched_out(prev, next);
2617 fire_sched_out_preempt_notifiers(prev, next);
2619 prepare_arch_switch(next);
2623 * finish_task_switch - clean up after a task-switch
2624 * @prev: the thread we just switched away from.
2626 * finish_task_switch must be called after the context switch, paired
2627 * with a prepare_task_switch call before the context switch.
2628 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2629 * and do any other architecture-specific cleanup actions.
2631 * Note that we may have delayed dropping an mm in context_switch(). If
2632 * so, we finish that here outside of the runqueue lock. (Doing it
2633 * with the lock held can cause deadlocks; see schedule() for
2636 * The context switch have flipped the stack from under us and restored the
2637 * local variables which were saved when this task called schedule() in the
2638 * past. prev == current is still correct but we need to recalculate this_rq
2639 * because prev may have moved to another CPU.
2641 static struct rq *finish_task_switch(struct task_struct *prev)
2642 __releases(rq->lock)
2644 struct rq *rq = this_rq();
2645 struct mm_struct *mm = rq->prev_mm;
2649 * The previous task will have left us with a preempt_count of 2
2650 * because it left us after:
2653 * preempt_disable(); // 1
2655 * raw_spin_lock_irq(&rq->lock) // 2
2657 * Also, see FORK_PREEMPT_COUNT.
2659 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2660 "corrupted preempt_count: %s/%d/0x%x\n",
2661 current->comm, current->pid, preempt_count()))
2662 preempt_count_set(FORK_PREEMPT_COUNT);
2667 * A task struct has one reference for the use as "current".
2668 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2669 * schedule one last time. The schedule call will never return, and
2670 * the scheduled task must drop that reference.
2672 * We must observe prev->state before clearing prev->on_cpu (in
2673 * finish_task), otherwise a concurrent wakeup can get prev
2674 * running on another CPU and we could rave with its RUNNING -> DEAD
2675 * transition, resulting in a double drop.
2677 prev_state = prev->state;
2678 vtime_task_switch(prev);
2679 perf_event_task_sched_in(prev, current);
2681 finish_lock_switch(rq);
2682 finish_arch_post_lock_switch();
2683 kcov_finish_switch(current);
2685 fire_sched_in_preempt_notifiers(current);
2687 * When switching through a kernel thread, the loop in
2688 * membarrier_{private,global}_expedited() may have observed that
2689 * kernel thread and not issued an IPI. It is therefore possible to
2690 * schedule between user->kernel->user threads without passing though
2691 * switch_mm(). Membarrier requires a barrier after storing to
2692 * rq->curr, before returning to userspace, so provide them here:
2694 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
2695 * provided by mmdrop(),
2696 * - a sync_core for SYNC_CORE.
2699 membarrier_mm_sync_core_before_usermode(mm);
2702 if (unlikely(prev_state == TASK_DEAD)) {
2703 if (prev->sched_class->task_dead)
2704 prev->sched_class->task_dead(prev);
2707 * Remove function-return probe instances associated with this
2708 * task and put them back on the free list.
2710 kprobe_flush_task(prev);
2712 /* Task is done with its stack. */
2713 put_task_stack(prev);
2715 put_task_struct(prev);
2718 tick_nohz_task_switch();
2724 /* rq->lock is NOT held, but preemption is disabled */
2725 static void __balance_callback(struct rq *rq)
2727 struct callback_head *head, *next;
2728 void (*func)(struct rq *rq);
2729 unsigned long flags;
2731 raw_spin_lock_irqsave(&rq->lock, flags);
2732 head = rq->balance_callback;
2733 rq->balance_callback = NULL;
2735 func = (void (*)(struct rq *))head->func;
2742 raw_spin_unlock_irqrestore(&rq->lock, flags);
2745 static inline void balance_callback(struct rq *rq)
2747 if (unlikely(rq->balance_callback))
2748 __balance_callback(rq);
2753 static inline void balance_callback(struct rq *rq)
2760 * schedule_tail - first thing a freshly forked thread must call.
2761 * @prev: the thread we just switched away from.
2763 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2764 __releases(rq->lock)
2769 * New tasks start with FORK_PREEMPT_COUNT, see there and
2770 * finish_task_switch() for details.
2772 * finish_task_switch() will drop rq->lock() and lower preempt_count
2773 * and the preempt_enable() will end up enabling preemption (on
2774 * PREEMPT_COUNT kernels).
2777 rq = finish_task_switch(prev);
2778 balance_callback(rq);
2781 if (current->set_child_tid)
2782 put_user(task_pid_vnr(current), current->set_child_tid);
2784 calculate_sigpending();
2788 * context_switch - switch to the new MM and the new thread's register state.
2790 static __always_inline struct rq *
2791 context_switch(struct rq *rq, struct task_struct *prev,
2792 struct task_struct *next, struct rq_flags *rf)
2794 struct mm_struct *mm, *oldmm;
2796 prepare_task_switch(rq, prev, next);
2799 oldmm = prev->active_mm;
2801 * For paravirt, this is coupled with an exit in switch_to to
2802 * combine the page table reload and the switch backend into
2805 arch_start_context_switch(prev);
2808 * If mm is non-NULL, we pass through switch_mm(). If mm is
2809 * NULL, we will pass through mmdrop() in finish_task_switch().
2810 * Both of these contain the full memory barrier required by
2811 * membarrier after storing to rq->curr, before returning to
2815 next->active_mm = oldmm;
2817 enter_lazy_tlb(oldmm, next);
2819 switch_mm_irqs_off(oldmm, mm, next);
2822 prev->active_mm = NULL;
2823 rq->prev_mm = oldmm;
2826 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2828 prepare_lock_switch(rq, next, rf);
2830 /* Here we just switch the register state and the stack. */
2831 switch_to(prev, next, prev);
2834 return finish_task_switch(prev);
2838 * nr_running and nr_context_switches:
2840 * externally visible scheduler statistics: current number of runnable
2841 * threads, total number of context switches performed since bootup.
2843 unsigned long nr_running(void)
2845 unsigned long i, sum = 0;
2847 for_each_online_cpu(i)
2848 sum += cpu_rq(i)->nr_running;
2854 * Check if only the current task is running on the CPU.
2856 * Caution: this function does not check that the caller has disabled
2857 * preemption, thus the result might have a time-of-check-to-time-of-use
2858 * race. The caller is responsible to use it correctly, for example:
2860 * - from a non-preemptable section (of course)
2862 * - from a thread that is bound to a single CPU
2864 * - in a loop with very short iterations (e.g. a polling loop)
2866 bool single_task_running(void)
2868 return raw_rq()->nr_running == 1;
2870 EXPORT_SYMBOL(single_task_running);
2872 unsigned long long nr_context_switches(void)
2875 unsigned long long sum = 0;
2877 for_each_possible_cpu(i)
2878 sum += cpu_rq(i)->nr_switches;
2884 * Consumers of these two interfaces, like for example the cpuidle menu
2885 * governor, are using nonsensical data. Preferring shallow idle state selection
2886 * for a CPU that has IO-wait which might not even end up running the task when
2887 * it does become runnable.
2890 unsigned long nr_iowait_cpu(int cpu)
2892 return atomic_read(&cpu_rq(cpu)->nr_iowait);
2896 * IO-wait accounting, and how its mostly bollocks (on SMP).
2898 * The idea behind IO-wait account is to account the idle time that we could
2899 * have spend running if it were not for IO. That is, if we were to improve the
2900 * storage performance, we'd have a proportional reduction in IO-wait time.
2902 * This all works nicely on UP, where, when a task blocks on IO, we account
2903 * idle time as IO-wait, because if the storage were faster, it could've been
2904 * running and we'd not be idle.
2906 * This has been extended to SMP, by doing the same for each CPU. This however
2909 * Imagine for instance the case where two tasks block on one CPU, only the one
2910 * CPU will have IO-wait accounted, while the other has regular idle. Even
2911 * though, if the storage were faster, both could've ran at the same time,
2912 * utilising both CPUs.
2914 * This means, that when looking globally, the current IO-wait accounting on
2915 * SMP is a lower bound, by reason of under accounting.
2917 * Worse, since the numbers are provided per CPU, they are sometimes
2918 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2919 * associated with any one particular CPU, it can wake to another CPU than it
2920 * blocked on. This means the per CPU IO-wait number is meaningless.
2922 * Task CPU affinities can make all that even more 'interesting'.
2925 unsigned long nr_iowait(void)
2927 unsigned long i, sum = 0;
2929 for_each_possible_cpu(i)
2930 sum += nr_iowait_cpu(i);
2938 * sched_exec - execve() is a valuable balancing opportunity, because at
2939 * this point the task has the smallest effective memory and cache footprint.
2941 void sched_exec(void)
2943 struct task_struct *p = current;
2944 unsigned long flags;
2947 raw_spin_lock_irqsave(&p->pi_lock, flags);
2948 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2949 if (dest_cpu == smp_processor_id())
2952 if (likely(cpu_active(dest_cpu))) {
2953 struct migration_arg arg = { p, dest_cpu };
2955 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2956 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2960 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2965 DEFINE_PER_CPU(struct kernel_stat, kstat);
2966 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2968 EXPORT_PER_CPU_SYMBOL(kstat);
2969 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2972 * The function fair_sched_class.update_curr accesses the struct curr
2973 * and its field curr->exec_start; when called from task_sched_runtime(),
2974 * we observe a high rate of cache misses in practice.
2975 * Prefetching this data results in improved performance.
2977 static inline void prefetch_curr_exec_start(struct task_struct *p)
2979 #ifdef CONFIG_FAIR_GROUP_SCHED
2980 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
2982 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
2985 prefetch(&curr->exec_start);
2989 * Return accounted runtime for the task.
2990 * In case the task is currently running, return the runtime plus current's
2991 * pending runtime that have not been accounted yet.
2993 unsigned long long task_sched_runtime(struct task_struct *p)
2999 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3001 * 64-bit doesn't need locks to atomically read a 64-bit value.
3002 * So we have a optimization chance when the task's delta_exec is 0.
3003 * Reading ->on_cpu is racy, but this is ok.
3005 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3006 * If we race with it entering CPU, unaccounted time is 0. This is
3007 * indistinguishable from the read occurring a few cycles earlier.
3008 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3009 * been accounted, so we're correct here as well.
3011 if (!p->on_cpu || !task_on_rq_queued(p))
3012 return p->se.sum_exec_runtime;
3015 rq = task_rq_lock(p, &rf);
3017 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3018 * project cycles that may never be accounted to this
3019 * thread, breaking clock_gettime().
3021 if (task_current(rq, p) && task_on_rq_queued(p)) {
3022 prefetch_curr_exec_start(p);
3023 update_rq_clock(rq);
3024 p->sched_class->update_curr(rq);
3026 ns = p->se.sum_exec_runtime;
3027 task_rq_unlock(rq, p, &rf);
3033 * This function gets called by the timer code, with HZ frequency.
3034 * We call it with interrupts disabled.
3036 void scheduler_tick(void)
3038 int cpu = smp_processor_id();
3039 struct rq *rq = cpu_rq(cpu);
3040 struct task_struct *curr = rq->curr;
3047 update_rq_clock(rq);
3048 curr->sched_class->task_tick(rq, curr, 0);
3049 cpu_load_update_active(rq);
3050 calc_global_load_tick(rq);
3055 perf_event_task_tick();
3058 rq->idle_balance = idle_cpu(cpu);
3059 trigger_load_balance(rq);
3063 #ifdef CONFIG_NO_HZ_FULL
3067 struct delayed_work work;
3070 static struct tick_work __percpu *tick_work_cpu;
3072 static void sched_tick_remote(struct work_struct *work)
3074 struct delayed_work *dwork = to_delayed_work(work);
3075 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3076 int cpu = twork->cpu;
3077 struct rq *rq = cpu_rq(cpu);
3078 struct task_struct *curr;
3083 * Handle the tick only if it appears the remote CPU is running in full
3084 * dynticks mode. The check is racy by nature, but missing a tick or
3085 * having one too much is no big deal because the scheduler tick updates
3086 * statistics and checks timeslices in a time-independent way, regardless
3087 * of when exactly it is running.
3089 if (idle_cpu(cpu) || !tick_nohz_tick_stopped_cpu(cpu))
3092 rq_lock_irq(rq, &rf);
3094 if (is_idle_task(curr))
3097 update_rq_clock(rq);
3098 delta = rq_clock_task(rq) - curr->se.exec_start;
3101 * Make sure the next tick runs within a reasonable
3104 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3105 curr->sched_class->task_tick(rq, curr, 0);
3108 rq_unlock_irq(rq, &rf);
3112 * Run the remote tick once per second (1Hz). This arbitrary
3113 * frequency is large enough to avoid overload but short enough
3114 * to keep scheduler internal stats reasonably up to date.
3116 queue_delayed_work(system_unbound_wq, dwork, HZ);
3119 static void sched_tick_start(int cpu)
3121 struct tick_work *twork;
3123 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3126 WARN_ON_ONCE(!tick_work_cpu);
3128 twork = per_cpu_ptr(tick_work_cpu, cpu);
3130 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3131 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3134 #ifdef CONFIG_HOTPLUG_CPU
3135 static void sched_tick_stop(int cpu)
3137 struct tick_work *twork;
3139 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3142 WARN_ON_ONCE(!tick_work_cpu);
3144 twork = per_cpu_ptr(tick_work_cpu, cpu);
3145 cancel_delayed_work_sync(&twork->work);
3147 #endif /* CONFIG_HOTPLUG_CPU */
3149 int __init sched_tick_offload_init(void)
3151 tick_work_cpu = alloc_percpu(struct tick_work);
3152 BUG_ON(!tick_work_cpu);
3157 #else /* !CONFIG_NO_HZ_FULL */
3158 static inline void sched_tick_start(int cpu) { }
3159 static inline void sched_tick_stop(int cpu) { }
3162 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3163 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3165 * If the value passed in is equal to the current preempt count
3166 * then we just disabled preemption. Start timing the latency.
3168 static inline void preempt_latency_start(int val)
3170 if (preempt_count() == val) {
3171 unsigned long ip = get_lock_parent_ip();
3172 #ifdef CONFIG_DEBUG_PREEMPT
3173 current->preempt_disable_ip = ip;
3175 trace_preempt_off(CALLER_ADDR0, ip);
3179 void preempt_count_add(int val)
3181 #ifdef CONFIG_DEBUG_PREEMPT
3185 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3188 __preempt_count_add(val);
3189 #ifdef CONFIG_DEBUG_PREEMPT
3191 * Spinlock count overflowing soon?
3193 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3196 preempt_latency_start(val);
3198 EXPORT_SYMBOL(preempt_count_add);
3199 NOKPROBE_SYMBOL(preempt_count_add);
3202 * If the value passed in equals to the current preempt count
3203 * then we just enabled preemption. Stop timing the latency.
3205 static inline void preempt_latency_stop(int val)
3207 if (preempt_count() == val)
3208 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3211 void preempt_count_sub(int val)
3213 #ifdef CONFIG_DEBUG_PREEMPT
3217 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3220 * Is the spinlock portion underflowing?
3222 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3223 !(preempt_count() & PREEMPT_MASK)))
3227 preempt_latency_stop(val);
3228 __preempt_count_sub(val);
3230 EXPORT_SYMBOL(preempt_count_sub);
3231 NOKPROBE_SYMBOL(preempt_count_sub);
3234 static inline void preempt_latency_start(int val) { }
3235 static inline void preempt_latency_stop(int val) { }
3238 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3240 #ifdef CONFIG_DEBUG_PREEMPT
3241 return p->preempt_disable_ip;
3248 * Print scheduling while atomic bug:
3250 static noinline void __schedule_bug(struct task_struct *prev)
3252 /* Save this before calling printk(), since that will clobber it */
3253 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3255 if (oops_in_progress)
3258 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3259 prev->comm, prev->pid, preempt_count());
3261 debug_show_held_locks(prev);
3263 if (irqs_disabled())
3264 print_irqtrace_events(prev);
3265 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3266 && in_atomic_preempt_off()) {
3267 pr_err("Preemption disabled at:");
3268 print_ip_sym(preempt_disable_ip);
3272 panic("scheduling while atomic\n");
3275 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3279 * Various schedule()-time debugging checks and statistics:
3281 static inline void schedule_debug(struct task_struct *prev)
3283 #ifdef CONFIG_SCHED_STACK_END_CHECK
3284 if (task_stack_end_corrupted(prev))
3285 panic("corrupted stack end detected inside scheduler\n");
3288 if (unlikely(in_atomic_preempt_off())) {
3289 __schedule_bug(prev);
3290 preempt_count_set(PREEMPT_DISABLED);
3294 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3296 schedstat_inc(this_rq()->sched_count);
3300 * Pick up the highest-prio task:
3302 static inline struct task_struct *
3303 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3305 const struct sched_class *class;
3306 struct task_struct *p;
3309 * Optimization: we know that if all tasks are in the fair class we can
3310 * call that function directly, but only if the @prev task wasn't of a
3311 * higher scheduling class, because otherwise those loose the
3312 * opportunity to pull in more work from other CPUs.
3314 if (likely((prev->sched_class == &idle_sched_class ||
3315 prev->sched_class == &fair_sched_class) &&
3316 rq->nr_running == rq->cfs.h_nr_running)) {
3318 p = fair_sched_class.pick_next_task(rq, prev, rf);
3319 if (unlikely(p == RETRY_TASK))
3322 /* Assumes fair_sched_class->next == idle_sched_class */
3324 p = idle_sched_class.pick_next_task(rq, prev, rf);
3330 for_each_class(class) {
3331 p = class->pick_next_task(rq, prev, rf);
3333 if (unlikely(p == RETRY_TASK))
3339 /* The idle class should always have a runnable task: */
3344 * __schedule() is the main scheduler function.
3346 * The main means of driving the scheduler and thus entering this function are:
3348 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3350 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3351 * paths. For example, see arch/x86/entry_64.S.
3353 * To drive preemption between tasks, the scheduler sets the flag in timer
3354 * interrupt handler scheduler_tick().
3356 * 3. Wakeups don't really cause entry into schedule(). They add a
3357 * task to the run-queue and that's it.
3359 * Now, if the new task added to the run-queue preempts the current
3360 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3361 * called on the nearest possible occasion:
3363 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3365 * - in syscall or exception context, at the next outmost
3366 * preempt_enable(). (this might be as soon as the wake_up()'s
3369 * - in IRQ context, return from interrupt-handler to
3370 * preemptible context
3372 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3375 * - cond_resched() call
3376 * - explicit schedule() call
3377 * - return from syscall or exception to user-space
3378 * - return from interrupt-handler to user-space
3380 * WARNING: must be called with preemption disabled!
3382 static void __sched notrace __schedule(bool preempt)
3384 struct task_struct *prev, *next;
3385 unsigned long *switch_count;
3390 cpu = smp_processor_id();
3394 schedule_debug(prev);
3396 if (sched_feat(HRTICK))
3399 local_irq_disable();
3400 rcu_note_context_switch(preempt);
3403 * Make sure that signal_pending_state()->signal_pending() below
3404 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3405 * done by the caller to avoid the race with signal_wake_up().
3407 * The membarrier system call requires a full memory barrier
3408 * after coming from user-space, before storing to rq->curr.
3411 smp_mb__after_spinlock();
3413 /* Promote REQ to ACT */
3414 rq->clock_update_flags <<= 1;
3415 update_rq_clock(rq);
3417 switch_count = &prev->nivcsw;
3418 if (!preempt && prev->state) {
3419 if (unlikely(signal_pending_state(prev->state, prev))) {
3420 prev->state = TASK_RUNNING;
3422 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3425 if (prev->in_iowait) {
3426 atomic_inc(&rq->nr_iowait);
3427 delayacct_blkio_start();
3431 * If a worker went to sleep, notify and ask workqueue
3432 * whether it wants to wake up a task to maintain
3435 if (prev->flags & PF_WQ_WORKER) {
3436 struct task_struct *to_wakeup;
3438 to_wakeup = wq_worker_sleeping(prev);
3440 try_to_wake_up_local(to_wakeup, &rf);
3443 switch_count = &prev->nvcsw;
3446 next = pick_next_task(rq, prev, &rf);
3447 clear_tsk_need_resched(prev);
3448 clear_preempt_need_resched();
3450 if (likely(prev != next)) {
3454 * The membarrier system call requires each architecture
3455 * to have a full memory barrier after updating
3456 * rq->curr, before returning to user-space.
3458 * Here are the schemes providing that barrier on the
3459 * various architectures:
3460 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3461 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3462 * - finish_lock_switch() for weakly-ordered
3463 * architectures where spin_unlock is a full barrier,
3464 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3465 * is a RELEASE barrier),
3469 trace_sched_switch(preempt, prev, next);
3471 /* Also unlocks the rq: */
3472 rq = context_switch(rq, prev, next, &rf);
3474 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3475 rq_unlock_irq(rq, &rf);
3478 balance_callback(rq);
3481 void __noreturn do_task_dead(void)
3483 /* Causes final put_task_struct in finish_task_switch(): */
3484 set_special_state(TASK_DEAD);
3486 /* Tell freezer to ignore us: */
3487 current->flags |= PF_NOFREEZE;
3492 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3497 static inline void sched_submit_work(struct task_struct *tsk)
3499 if (!tsk->state || tsk_is_pi_blocked(tsk))
3502 * If we are going to sleep and we have plugged IO queued,
3503 * make sure to submit it to avoid deadlocks.
3505 if (blk_needs_flush_plug(tsk))
3506 blk_schedule_flush_plug(tsk);
3509 asmlinkage __visible void __sched schedule(void)
3511 struct task_struct *tsk = current;
3513 sched_submit_work(tsk);
3517 sched_preempt_enable_no_resched();
3518 } while (need_resched());
3520 EXPORT_SYMBOL(schedule);
3523 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3524 * state (have scheduled out non-voluntarily) by making sure that all
3525 * tasks have either left the run queue or have gone into user space.
3526 * As idle tasks do not do either, they must not ever be preempted
3527 * (schedule out non-voluntarily).
3529 * schedule_idle() is similar to schedule_preempt_disable() except that it
3530 * never enables preemption because it does not call sched_submit_work().
3532 void __sched schedule_idle(void)
3535 * As this skips calling sched_submit_work(), which the idle task does
3536 * regardless because that function is a nop when the task is in a
3537 * TASK_RUNNING state, make sure this isn't used someplace that the
3538 * current task can be in any other state. Note, idle is always in the
3539 * TASK_RUNNING state.
3541 WARN_ON_ONCE(current->state);
3544 } while (need_resched());
3547 #ifdef CONFIG_CONTEXT_TRACKING
3548 asmlinkage __visible void __sched schedule_user(void)
3551 * If we come here after a random call to set_need_resched(),
3552 * or we have been woken up remotely but the IPI has not yet arrived,
3553 * we haven't yet exited the RCU idle mode. Do it here manually until
3554 * we find a better solution.
3556 * NB: There are buggy callers of this function. Ideally we
3557 * should warn if prev_state != CONTEXT_USER, but that will trigger
3558 * too frequently to make sense yet.
3560 enum ctx_state prev_state = exception_enter();
3562 exception_exit(prev_state);
3567 * schedule_preempt_disabled - called with preemption disabled
3569 * Returns with preemption disabled. Note: preempt_count must be 1
3571 void __sched schedule_preempt_disabled(void)
3573 sched_preempt_enable_no_resched();
3578 static void __sched notrace preempt_schedule_common(void)
3582 * Because the function tracer can trace preempt_count_sub()
3583 * and it also uses preempt_enable/disable_notrace(), if
3584 * NEED_RESCHED is set, the preempt_enable_notrace() called
3585 * by the function tracer will call this function again and
3586 * cause infinite recursion.
3588 * Preemption must be disabled here before the function
3589 * tracer can trace. Break up preempt_disable() into two
3590 * calls. One to disable preemption without fear of being
3591 * traced. The other to still record the preemption latency,
3592 * which can also be traced by the function tracer.
3594 preempt_disable_notrace();
3595 preempt_latency_start(1);
3597 preempt_latency_stop(1);
3598 preempt_enable_no_resched_notrace();
3601 * Check again in case we missed a preemption opportunity
3602 * between schedule and now.
3604 } while (need_resched());
3607 #ifdef CONFIG_PREEMPT
3609 * this is the entry point to schedule() from in-kernel preemption
3610 * off of preempt_enable. Kernel preemptions off return from interrupt
3611 * occur there and call schedule directly.
3613 asmlinkage __visible void __sched notrace preempt_schedule(void)
3616 * If there is a non-zero preempt_count or interrupts are disabled,
3617 * we do not want to preempt the current task. Just return..
3619 if (likely(!preemptible()))
3622 preempt_schedule_common();
3624 NOKPROBE_SYMBOL(preempt_schedule);
3625 EXPORT_SYMBOL(preempt_schedule);
3628 * preempt_schedule_notrace - preempt_schedule called by tracing
3630 * The tracing infrastructure uses preempt_enable_notrace to prevent
3631 * recursion and tracing preempt enabling caused by the tracing
3632 * infrastructure itself. But as tracing can happen in areas coming
3633 * from userspace or just about to enter userspace, a preempt enable
3634 * can occur before user_exit() is called. This will cause the scheduler
3635 * to be called when the system is still in usermode.
3637 * To prevent this, the preempt_enable_notrace will use this function
3638 * instead of preempt_schedule() to exit user context if needed before
3639 * calling the scheduler.
3641 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3643 enum ctx_state prev_ctx;
3645 if (likely(!preemptible()))
3650 * Because the function tracer can trace preempt_count_sub()
3651 * and it also uses preempt_enable/disable_notrace(), if
3652 * NEED_RESCHED is set, the preempt_enable_notrace() called
3653 * by the function tracer will call this function again and
3654 * cause infinite recursion.
3656 * Preemption must be disabled here before the function
3657 * tracer can trace. Break up preempt_disable() into two
3658 * calls. One to disable preemption without fear of being
3659 * traced. The other to still record the preemption latency,
3660 * which can also be traced by the function tracer.
3662 preempt_disable_notrace();
3663 preempt_latency_start(1);
3665 * Needs preempt disabled in case user_exit() is traced
3666 * and the tracer calls preempt_enable_notrace() causing
3667 * an infinite recursion.
3669 prev_ctx = exception_enter();
3671 exception_exit(prev_ctx);
3673 preempt_latency_stop(1);
3674 preempt_enable_no_resched_notrace();
3675 } while (need_resched());
3677 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3679 #endif /* CONFIG_PREEMPT */
3682 * this is the entry point to schedule() from kernel preemption
3683 * off of irq context.
3684 * Note, that this is called and return with irqs disabled. This will
3685 * protect us against recursive calling from irq.
3687 asmlinkage __visible void __sched preempt_schedule_irq(void)
3689 enum ctx_state prev_state;
3691 /* Catch callers which need to be fixed */
3692 BUG_ON(preempt_count() || !irqs_disabled());
3694 prev_state = exception_enter();
3700 local_irq_disable();
3701 sched_preempt_enable_no_resched();
3702 } while (need_resched());
3704 exception_exit(prev_state);
3707 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
3710 return try_to_wake_up(curr->private, mode, wake_flags);
3712 EXPORT_SYMBOL(default_wake_function);
3714 #ifdef CONFIG_RT_MUTEXES
3716 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
3719 prio = min(prio, pi_task->prio);
3724 static inline int rt_effective_prio(struct task_struct *p, int prio)
3726 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3728 return __rt_effective_prio(pi_task, prio);
3732 * rt_mutex_setprio - set the current priority of a task
3734 * @pi_task: donor task
3736 * This function changes the 'effective' priority of a task. It does
3737 * not touch ->normal_prio like __setscheduler().
3739 * Used by the rt_mutex code to implement priority inheritance
3740 * logic. Call site only calls if the priority of the task changed.
3742 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
3744 int prio, oldprio, queued, running, queue_flag =
3745 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
3746 const struct sched_class *prev_class;
3750 /* XXX used to be waiter->prio, not waiter->task->prio */
3751 prio = __rt_effective_prio(pi_task, p->normal_prio);
3754 * If nothing changed; bail early.
3756 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
3759 rq = __task_rq_lock(p, &rf);
3760 update_rq_clock(rq);
3762 * Set under pi_lock && rq->lock, such that the value can be used under
3765 * Note that there is loads of tricky to make this pointer cache work
3766 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3767 * ensure a task is de-boosted (pi_task is set to NULL) before the
3768 * task is allowed to run again (and can exit). This ensures the pointer
3769 * points to a blocked task -- which guaratees the task is present.
3771 p->pi_top_task = pi_task;
3774 * For FIFO/RR we only need to set prio, if that matches we're done.
3776 if (prio == p->prio && !dl_prio(prio))
3780 * Idle task boosting is a nono in general. There is one
3781 * exception, when PREEMPT_RT and NOHZ is active:
3783 * The idle task calls get_next_timer_interrupt() and holds
3784 * the timer wheel base->lock on the CPU and another CPU wants
3785 * to access the timer (probably to cancel it). We can safely
3786 * ignore the boosting request, as the idle CPU runs this code
3787 * with interrupts disabled and will complete the lock
3788 * protected section without being interrupted. So there is no
3789 * real need to boost.
3791 if (unlikely(p == rq->idle)) {
3792 WARN_ON(p != rq->curr);
3793 WARN_ON(p->pi_blocked_on);
3797 trace_sched_pi_setprio(p, pi_task);
3800 if (oldprio == prio)
3801 queue_flag &= ~DEQUEUE_MOVE;
3803 prev_class = p->sched_class;
3804 queued = task_on_rq_queued(p);
3805 running = task_current(rq, p);
3807 dequeue_task(rq, p, queue_flag);
3809 put_prev_task(rq, p);
3812 * Boosting condition are:
3813 * 1. -rt task is running and holds mutex A
3814 * --> -dl task blocks on mutex A
3816 * 2. -dl task is running and holds mutex A
3817 * --> -dl task blocks on mutex A and could preempt the
3820 if (dl_prio(prio)) {
3821 if (!dl_prio(p->normal_prio) ||
3822 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3823 p->dl.dl_boosted = 1;
3824 queue_flag |= ENQUEUE_REPLENISH;
3826 p->dl.dl_boosted = 0;
3827 p->sched_class = &dl_sched_class;
3828 } else if (rt_prio(prio)) {
3829 if (dl_prio(oldprio))
3830 p->dl.dl_boosted = 0;
3832 queue_flag |= ENQUEUE_HEAD;
3833 p->sched_class = &rt_sched_class;
3835 if (dl_prio(oldprio))
3836 p->dl.dl_boosted = 0;
3837 if (rt_prio(oldprio))
3839 p->sched_class = &fair_sched_class;
3845 enqueue_task(rq, p, queue_flag);
3847 set_curr_task(rq, p);
3849 check_class_changed(rq, p, prev_class, oldprio);
3851 /* Avoid rq from going away on us: */
3853 __task_rq_unlock(rq, &rf);
3855 balance_callback(rq);
3859 static inline int rt_effective_prio(struct task_struct *p, int prio)
3865 void set_user_nice(struct task_struct *p, long nice)
3867 bool queued, running;
3868 int old_prio, delta;
3872 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3875 * We have to be careful, if called from sys_setpriority(),
3876 * the task might be in the middle of scheduling on another CPU.
3878 rq = task_rq_lock(p, &rf);
3879 update_rq_clock(rq);
3882 * The RT priorities are set via sched_setscheduler(), but we still
3883 * allow the 'normal' nice value to be set - but as expected
3884 * it wont have any effect on scheduling until the task is
3885 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3887 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3888 p->static_prio = NICE_TO_PRIO(nice);
3891 queued = task_on_rq_queued(p);
3892 running = task_current(rq, p);
3894 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
3896 put_prev_task(rq, p);
3898 p->static_prio = NICE_TO_PRIO(nice);
3899 set_load_weight(p, true);
3901 p->prio = effective_prio(p);
3902 delta = p->prio - old_prio;
3905 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
3907 * If the task increased its priority or is running and
3908 * lowered its priority, then reschedule its CPU:
3910 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3914 set_curr_task(rq, p);
3916 task_rq_unlock(rq, p, &rf);
3918 EXPORT_SYMBOL(set_user_nice);
3921 * can_nice - check if a task can reduce its nice value
3925 int can_nice(const struct task_struct *p, const int nice)
3927 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3928 int nice_rlim = nice_to_rlimit(nice);
3930 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3931 capable(CAP_SYS_NICE));
3934 #ifdef __ARCH_WANT_SYS_NICE
3937 * sys_nice - change the priority of the current process.
3938 * @increment: priority increment
3940 * sys_setpriority is a more generic, but much slower function that
3941 * does similar things.
3943 SYSCALL_DEFINE1(nice, int, increment)
3948 * Setpriority might change our priority at the same moment.
3949 * We don't have to worry. Conceptually one call occurs first
3950 * and we have a single winner.
3952 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3953 nice = task_nice(current) + increment;
3955 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3956 if (increment < 0 && !can_nice(current, nice))
3959 retval = security_task_setnice(current, nice);
3963 set_user_nice(current, nice);
3970 * task_prio - return the priority value of a given task.
3971 * @p: the task in question.
3973 * Return: The priority value as seen by users in /proc.
3974 * RT tasks are offset by -200. Normal tasks are centered
3975 * around 0, value goes from -16 to +15.
3977 int task_prio(const struct task_struct *p)
3979 return p->prio - MAX_RT_PRIO;
3983 * idle_cpu - is a given CPU idle currently?
3984 * @cpu: the processor in question.
3986 * Return: 1 if the CPU is currently idle. 0 otherwise.
3988 int idle_cpu(int cpu)
3990 struct rq *rq = cpu_rq(cpu);
3992 if (rq->curr != rq->idle)
3999 if (!llist_empty(&rq->wake_list))
4007 * available_idle_cpu - is a given CPU idle for enqueuing work.
4008 * @cpu: the CPU in question.
4010 * Return: 1 if the CPU is currently idle. 0 otherwise.
4012 int available_idle_cpu(int cpu)
4017 if (vcpu_is_preempted(cpu))
4024 * idle_task - return the idle task for a given CPU.
4025 * @cpu: the processor in question.
4027 * Return: The idle task for the CPU @cpu.
4029 struct task_struct *idle_task(int cpu)
4031 return cpu_rq(cpu)->idle;
4035 * find_process_by_pid - find a process with a matching PID value.
4036 * @pid: the pid in question.
4038 * The task of @pid, if found. %NULL otherwise.
4040 static struct task_struct *find_process_by_pid(pid_t pid)
4042 return pid ? find_task_by_vpid(pid) : current;
4046 * sched_setparam() passes in -1 for its policy, to let the functions
4047 * it calls know not to change it.
4049 #define SETPARAM_POLICY -1
4051 static void __setscheduler_params(struct task_struct *p,
4052 const struct sched_attr *attr)
4054 int policy = attr->sched_policy;
4056 if (policy == SETPARAM_POLICY)
4061 if (dl_policy(policy))
4062 __setparam_dl(p, attr);
4063 else if (fair_policy(policy))
4064 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4067 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4068 * !rt_policy. Always setting this ensures that things like
4069 * getparam()/getattr() don't report silly values for !rt tasks.
4071 p->rt_priority = attr->sched_priority;
4072 p->normal_prio = normal_prio(p);
4073 set_load_weight(p, true);
4076 /* Actually do priority change: must hold pi & rq lock. */
4077 static void __setscheduler(struct rq *rq, struct task_struct *p,
4078 const struct sched_attr *attr, bool keep_boost)
4080 __setscheduler_params(p, attr);
4083 * Keep a potential priority boosting if called from
4084 * sched_setscheduler().
4086 p->prio = normal_prio(p);
4088 p->prio = rt_effective_prio(p, p->prio);
4090 if (dl_prio(p->prio))
4091 p->sched_class = &dl_sched_class;
4092 else if (rt_prio(p->prio))
4093 p->sched_class = &rt_sched_class;
4095 p->sched_class = &fair_sched_class;
4099 * Check the target process has a UID that matches the current process's:
4101 static bool check_same_owner(struct task_struct *p)
4103 const struct cred *cred = current_cred(), *pcred;
4107 pcred = __task_cred(p);
4108 match = (uid_eq(cred->euid, pcred->euid) ||
4109 uid_eq(cred->euid, pcred->uid));
4114 static int __sched_setscheduler(struct task_struct *p,
4115 const struct sched_attr *attr,
4118 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4119 MAX_RT_PRIO - 1 - attr->sched_priority;
4120 int retval, oldprio, oldpolicy = -1, queued, running;
4121 int new_effective_prio, policy = attr->sched_policy;
4122 const struct sched_class *prev_class;
4125 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4128 /* The pi code expects interrupts enabled */
4129 BUG_ON(pi && in_interrupt());
4131 /* Double check policy once rq lock held: */
4133 reset_on_fork = p->sched_reset_on_fork;
4134 policy = oldpolicy = p->policy;
4136 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4138 if (!valid_policy(policy))
4142 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4146 * Valid priorities for SCHED_FIFO and SCHED_RR are
4147 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4148 * SCHED_BATCH and SCHED_IDLE is 0.
4150 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4151 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4153 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4154 (rt_policy(policy) != (attr->sched_priority != 0)))
4158 * Allow unprivileged RT tasks to decrease priority:
4160 if (user && !capable(CAP_SYS_NICE)) {
4161 if (fair_policy(policy)) {
4162 if (attr->sched_nice < task_nice(p) &&
4163 !can_nice(p, attr->sched_nice))
4167 if (rt_policy(policy)) {
4168 unsigned long rlim_rtprio =
4169 task_rlimit(p, RLIMIT_RTPRIO);
4171 /* Can't set/change the rt policy: */
4172 if (policy != p->policy && !rlim_rtprio)
4175 /* Can't increase priority: */
4176 if (attr->sched_priority > p->rt_priority &&
4177 attr->sched_priority > rlim_rtprio)
4182 * Can't set/change SCHED_DEADLINE policy at all for now
4183 * (safest behavior); in the future we would like to allow
4184 * unprivileged DL tasks to increase their relative deadline
4185 * or reduce their runtime (both ways reducing utilization)
4187 if (dl_policy(policy))
4191 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4192 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4194 if (idle_policy(p->policy) && !idle_policy(policy)) {
4195 if (!can_nice(p, task_nice(p)))
4199 /* Can't change other user's priorities: */
4200 if (!check_same_owner(p))
4203 /* Normal users shall not reset the sched_reset_on_fork flag: */
4204 if (p->sched_reset_on_fork && !reset_on_fork)
4209 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4212 retval = security_task_setscheduler(p);
4218 * Make sure no PI-waiters arrive (or leave) while we are
4219 * changing the priority of the task:
4221 * To be able to change p->policy safely, the appropriate
4222 * runqueue lock must be held.
4224 rq = task_rq_lock(p, &rf);
4225 update_rq_clock(rq);
4228 * Changing the policy of the stop threads its a very bad idea:
4230 if (p == rq->stop) {
4231 task_rq_unlock(rq, p, &rf);
4236 * If not changing anything there's no need to proceed further,
4237 * but store a possible modification of reset_on_fork.
4239 if (unlikely(policy == p->policy)) {
4240 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4242 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4244 if (dl_policy(policy) && dl_param_changed(p, attr))
4247 p->sched_reset_on_fork = reset_on_fork;
4248 task_rq_unlock(rq, p, &rf);
4254 #ifdef CONFIG_RT_GROUP_SCHED
4256 * Do not allow realtime tasks into groups that have no runtime
4259 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4260 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4261 !task_group_is_autogroup(task_group(p))) {
4262 task_rq_unlock(rq, p, &rf);
4267 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4268 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4269 cpumask_t *span = rq->rd->span;
4272 * Don't allow tasks with an affinity mask smaller than
4273 * the entire root_domain to become SCHED_DEADLINE. We
4274 * will also fail if there's no bandwidth available.
4276 if (!cpumask_subset(span, &p->cpus_allowed) ||
4277 rq->rd->dl_bw.bw == 0) {
4278 task_rq_unlock(rq, p, &rf);
4285 /* Re-check policy now with rq lock held: */
4286 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4287 policy = oldpolicy = -1;
4288 task_rq_unlock(rq, p, &rf);
4293 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4294 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4297 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4298 task_rq_unlock(rq, p, &rf);
4302 p->sched_reset_on_fork = reset_on_fork;
4307 * Take priority boosted tasks into account. If the new
4308 * effective priority is unchanged, we just store the new
4309 * normal parameters and do not touch the scheduler class and
4310 * the runqueue. This will be done when the task deboost
4313 new_effective_prio = rt_effective_prio(p, newprio);
4314 if (new_effective_prio == oldprio)
4315 queue_flags &= ~DEQUEUE_MOVE;
4318 queued = task_on_rq_queued(p);
4319 running = task_current(rq, p);
4321 dequeue_task(rq, p, queue_flags);
4323 put_prev_task(rq, p);
4325 prev_class = p->sched_class;
4326 __setscheduler(rq, p, attr, pi);
4330 * We enqueue to tail when the priority of a task is
4331 * increased (user space view).
4333 if (oldprio < p->prio)
4334 queue_flags |= ENQUEUE_HEAD;
4336 enqueue_task(rq, p, queue_flags);
4339 set_curr_task(rq, p);
4341 check_class_changed(rq, p, prev_class, oldprio);
4343 /* Avoid rq from going away on us: */
4345 task_rq_unlock(rq, p, &rf);
4348 rt_mutex_adjust_pi(p);
4350 /* Run balance callbacks after we've adjusted the PI chain: */
4351 balance_callback(rq);
4357 static int _sched_setscheduler(struct task_struct *p, int policy,
4358 const struct sched_param *param, bool check)
4360 struct sched_attr attr = {
4361 .sched_policy = policy,
4362 .sched_priority = param->sched_priority,
4363 .sched_nice = PRIO_TO_NICE(p->static_prio),
4366 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4367 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4368 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4369 policy &= ~SCHED_RESET_ON_FORK;
4370 attr.sched_policy = policy;
4373 return __sched_setscheduler(p, &attr, check, true);
4376 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4377 * @p: the task in question.
4378 * @policy: new policy.
4379 * @param: structure containing the new RT priority.
4381 * Return: 0 on success. An error code otherwise.
4383 * NOTE that the task may be already dead.
4385 int sched_setscheduler(struct task_struct *p, int policy,
4386 const struct sched_param *param)
4388 return _sched_setscheduler(p, policy, param, true);
4390 EXPORT_SYMBOL_GPL(sched_setscheduler);
4392 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4394 return __sched_setscheduler(p, attr, true, true);
4396 EXPORT_SYMBOL_GPL(sched_setattr);
4398 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
4400 return __sched_setscheduler(p, attr, false, true);
4404 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4405 * @p: the task in question.
4406 * @policy: new policy.
4407 * @param: structure containing the new RT priority.
4409 * Just like sched_setscheduler, only don't bother checking if the
4410 * current context has permission. For example, this is needed in
4411 * stop_machine(): we create temporary high priority worker threads,
4412 * but our caller might not have that capability.
4414 * Return: 0 on success. An error code otherwise.
4416 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4417 const struct sched_param *param)
4419 return _sched_setscheduler(p, policy, param, false);
4421 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4424 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4426 struct sched_param lparam;
4427 struct task_struct *p;
4430 if (!param || pid < 0)
4432 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4437 p = find_process_by_pid(pid);
4439 retval = sched_setscheduler(p, policy, &lparam);
4446 * Mimics kernel/events/core.c perf_copy_attr().
4448 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4453 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4456 /* Zero the full structure, so that a short copy will be nice: */
4457 memset(attr, 0, sizeof(*attr));
4459 ret = get_user(size, &uattr->size);
4463 /* Bail out on silly large: */
4464 if (size > PAGE_SIZE)
4467 /* ABI compatibility quirk: */
4469 size = SCHED_ATTR_SIZE_VER0;
4471 if (size < SCHED_ATTR_SIZE_VER0)
4475 * If we're handed a bigger struct than we know of,
4476 * ensure all the unknown bits are 0 - i.e. new
4477 * user-space does not rely on any kernel feature
4478 * extensions we dont know about yet.
4480 if (size > sizeof(*attr)) {
4481 unsigned char __user *addr;
4482 unsigned char __user *end;
4485 addr = (void __user *)uattr + sizeof(*attr);
4486 end = (void __user *)uattr + size;
4488 for (; addr < end; addr++) {
4489 ret = get_user(val, addr);
4495 size = sizeof(*attr);
4498 ret = copy_from_user(attr, uattr, size);
4503 * XXX: Do we want to be lenient like existing syscalls; or do we want
4504 * to be strict and return an error on out-of-bounds values?
4506 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4511 put_user(sizeof(*attr), &uattr->size);
4516 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4517 * @pid: the pid in question.
4518 * @policy: new policy.
4519 * @param: structure containing the new RT priority.
4521 * Return: 0 on success. An error code otherwise.
4523 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4528 return do_sched_setscheduler(pid, policy, param);
4532 * sys_sched_setparam - set/change the RT priority of a thread
4533 * @pid: the pid in question.
4534 * @param: structure containing the new RT priority.
4536 * Return: 0 on success. An error code otherwise.
4538 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4540 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4544 * sys_sched_setattr - same as above, but with extended sched_attr
4545 * @pid: the pid in question.
4546 * @uattr: structure containing the extended parameters.
4547 * @flags: for future extension.
4549 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4550 unsigned int, flags)
4552 struct sched_attr attr;
4553 struct task_struct *p;
4556 if (!uattr || pid < 0 || flags)
4559 retval = sched_copy_attr(uattr, &attr);
4563 if ((int)attr.sched_policy < 0)
4568 p = find_process_by_pid(pid);
4570 retval = sched_setattr(p, &attr);
4577 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4578 * @pid: the pid in question.
4580 * Return: On success, the policy of the thread. Otherwise, a negative error
4583 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4585 struct task_struct *p;
4593 p = find_process_by_pid(pid);
4595 retval = security_task_getscheduler(p);
4598 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4605 * sys_sched_getparam - get the RT priority of a thread
4606 * @pid: the pid in question.
4607 * @param: structure containing the RT priority.
4609 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4612 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4614 struct sched_param lp = { .sched_priority = 0 };
4615 struct task_struct *p;
4618 if (!param || pid < 0)
4622 p = find_process_by_pid(pid);
4627 retval = security_task_getscheduler(p);
4631 if (task_has_rt_policy(p))
4632 lp.sched_priority = p->rt_priority;
4636 * This one might sleep, we cannot do it with a spinlock held ...
4638 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4647 static int sched_read_attr(struct sched_attr __user *uattr,
4648 struct sched_attr *attr,
4653 if (!access_ok(VERIFY_WRITE, uattr, usize))
4657 * If we're handed a smaller struct than we know of,
4658 * ensure all the unknown bits are 0 - i.e. old
4659 * user-space does not get uncomplete information.
4661 if (usize < sizeof(*attr)) {
4662 unsigned char *addr;
4665 addr = (void *)attr + usize;
4666 end = (void *)attr + sizeof(*attr);
4668 for (; addr < end; addr++) {
4676 ret = copy_to_user(uattr, attr, attr->size);
4684 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4685 * @pid: the pid in question.
4686 * @uattr: structure containing the extended parameters.
4687 * @size: sizeof(attr) for fwd/bwd comp.
4688 * @flags: for future extension.
4690 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4691 unsigned int, size, unsigned int, flags)
4693 struct sched_attr attr = {
4694 .size = sizeof(struct sched_attr),
4696 struct task_struct *p;
4699 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4700 size < SCHED_ATTR_SIZE_VER0 || flags)
4704 p = find_process_by_pid(pid);
4709 retval = security_task_getscheduler(p);
4713 attr.sched_policy = p->policy;
4714 if (p->sched_reset_on_fork)
4715 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4716 if (task_has_dl_policy(p))
4717 __getparam_dl(p, &attr);
4718 else if (task_has_rt_policy(p))
4719 attr.sched_priority = p->rt_priority;
4721 attr.sched_nice = task_nice(p);
4725 retval = sched_read_attr(uattr, &attr, size);
4733 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4735 cpumask_var_t cpus_allowed, new_mask;
4736 struct task_struct *p;
4741 p = find_process_by_pid(pid);
4747 /* Prevent p going away */
4751 if (p->flags & PF_NO_SETAFFINITY) {
4755 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4759 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4761 goto out_free_cpus_allowed;
4764 if (!check_same_owner(p)) {
4766 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4768 goto out_free_new_mask;
4773 retval = security_task_setscheduler(p);
4775 goto out_free_new_mask;
4778 cpuset_cpus_allowed(p, cpus_allowed);
4779 cpumask_and(new_mask, in_mask, cpus_allowed);
4782 * Since bandwidth control happens on root_domain basis,
4783 * if admission test is enabled, we only admit -deadline
4784 * tasks allowed to run on all the CPUs in the task's
4788 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4790 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4793 goto out_free_new_mask;
4799 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4802 cpuset_cpus_allowed(p, cpus_allowed);
4803 if (!cpumask_subset(new_mask, cpus_allowed)) {
4805 * We must have raced with a concurrent cpuset
4806 * update. Just reset the cpus_allowed to the
4807 * cpuset's cpus_allowed
4809 cpumask_copy(new_mask, cpus_allowed);
4814 free_cpumask_var(new_mask);
4815 out_free_cpus_allowed:
4816 free_cpumask_var(cpus_allowed);
4822 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4823 struct cpumask *new_mask)
4825 if (len < cpumask_size())
4826 cpumask_clear(new_mask);
4827 else if (len > cpumask_size())
4828 len = cpumask_size();
4830 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4834 * sys_sched_setaffinity - set the CPU affinity of a process
4835 * @pid: pid of the process
4836 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4837 * @user_mask_ptr: user-space pointer to the new CPU mask
4839 * Return: 0 on success. An error code otherwise.
4841 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4842 unsigned long __user *, user_mask_ptr)
4844 cpumask_var_t new_mask;
4847 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4850 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4852 retval = sched_setaffinity(pid, new_mask);
4853 free_cpumask_var(new_mask);
4857 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4859 struct task_struct *p;
4860 unsigned long flags;
4866 p = find_process_by_pid(pid);
4870 retval = security_task_getscheduler(p);
4874 raw_spin_lock_irqsave(&p->pi_lock, flags);
4875 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4876 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4885 * sys_sched_getaffinity - get the CPU affinity of a process
4886 * @pid: pid of the process
4887 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4888 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4890 * Return: size of CPU mask copied to user_mask_ptr on success. An
4891 * error code otherwise.
4893 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4894 unsigned long __user *, user_mask_ptr)
4899 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4901 if (len & (sizeof(unsigned long)-1))
4904 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4907 ret = sched_getaffinity(pid, mask);
4909 unsigned int retlen = min(len, cpumask_size());
4911 if (copy_to_user(user_mask_ptr, mask, retlen))
4916 free_cpumask_var(mask);
4922 * sys_sched_yield - yield the current processor to other threads.
4924 * This function yields the current CPU to other tasks. If there are no
4925 * other threads running on this CPU then this function will return.
4929 static void do_sched_yield(void)
4934 rq = this_rq_lock_irq(&rf);
4936 schedstat_inc(rq->yld_count);
4937 current->sched_class->yield_task(rq);
4940 * Since we are going to call schedule() anyway, there's
4941 * no need to preempt or enable interrupts:
4945 sched_preempt_enable_no_resched();
4950 SYSCALL_DEFINE0(sched_yield)
4956 #ifndef CONFIG_PREEMPT
4957 int __sched _cond_resched(void)
4959 if (should_resched(0)) {
4960 preempt_schedule_common();
4966 EXPORT_SYMBOL(_cond_resched);
4970 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4971 * call schedule, and on return reacquire the lock.
4973 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4974 * operations here to prevent schedule() from being called twice (once via
4975 * spin_unlock(), once by hand).
4977 int __cond_resched_lock(spinlock_t *lock)
4979 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4982 lockdep_assert_held(lock);
4984 if (spin_needbreak(lock) || resched) {
4987 preempt_schedule_common();
4995 EXPORT_SYMBOL(__cond_resched_lock);
4998 * yield - yield the current processor to other threads.
5000 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5002 * The scheduler is at all times free to pick the calling task as the most
5003 * eligible task to run, if removing the yield() call from your code breaks
5004 * it, its already broken.
5006 * Typical broken usage is:
5011 * where one assumes that yield() will let 'the other' process run that will
5012 * make event true. If the current task is a SCHED_FIFO task that will never
5013 * happen. Never use yield() as a progress guarantee!!
5015 * If you want to use yield() to wait for something, use wait_event().
5016 * If you want to use yield() to be 'nice' for others, use cond_resched().
5017 * If you still want to use yield(), do not!
5019 void __sched yield(void)
5021 set_current_state(TASK_RUNNING);
5024 EXPORT_SYMBOL(yield);
5027 * yield_to - yield the current processor to another thread in
5028 * your thread group, or accelerate that thread toward the
5029 * processor it's on.
5031 * @preempt: whether task preemption is allowed or not
5033 * It's the caller's job to ensure that the target task struct
5034 * can't go away on us before we can do any checks.
5037 * true (>0) if we indeed boosted the target task.
5038 * false (0) if we failed to boost the target.
5039 * -ESRCH if there's no task to yield to.
5041 int __sched yield_to(struct task_struct *p, bool preempt)
5043 struct task_struct *curr = current;
5044 struct rq *rq, *p_rq;
5045 unsigned long flags;
5048 local_irq_save(flags);
5054 * If we're the only runnable task on the rq and target rq also
5055 * has only one task, there's absolutely no point in yielding.
5057 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5062 double_rq_lock(rq, p_rq);
5063 if (task_rq(p) != p_rq) {
5064 double_rq_unlock(rq, p_rq);
5068 if (!curr->sched_class->yield_to_task)
5071 if (curr->sched_class != p->sched_class)
5074 if (task_running(p_rq, p) || p->state)
5077 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5079 schedstat_inc(rq->yld_count);
5081 * Make p's CPU reschedule; pick_next_entity takes care of
5084 if (preempt && rq != p_rq)
5089 double_rq_unlock(rq, p_rq);
5091 local_irq_restore(flags);
5098 EXPORT_SYMBOL_GPL(yield_to);
5100 int io_schedule_prepare(void)
5102 int old_iowait = current->in_iowait;
5104 current->in_iowait = 1;
5105 blk_schedule_flush_plug(current);
5110 void io_schedule_finish(int token)
5112 current->in_iowait = token;
5116 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5117 * that process accounting knows that this is a task in IO wait state.
5119 long __sched io_schedule_timeout(long timeout)
5124 token = io_schedule_prepare();
5125 ret = schedule_timeout(timeout);
5126 io_schedule_finish(token);
5130 EXPORT_SYMBOL(io_schedule_timeout);
5132 void io_schedule(void)
5136 token = io_schedule_prepare();
5138 io_schedule_finish(token);
5140 EXPORT_SYMBOL(io_schedule);
5143 * sys_sched_get_priority_max - return maximum RT priority.
5144 * @policy: scheduling class.
5146 * Return: On success, this syscall returns the maximum
5147 * rt_priority that can be used by a given scheduling class.
5148 * On failure, a negative error code is returned.
5150 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5157 ret = MAX_USER_RT_PRIO-1;
5159 case SCHED_DEADLINE:
5170 * sys_sched_get_priority_min - return minimum RT priority.
5171 * @policy: scheduling class.
5173 * Return: On success, this syscall returns the minimum
5174 * rt_priority that can be used by a given scheduling class.
5175 * On failure, a negative error code is returned.
5177 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5186 case SCHED_DEADLINE:
5195 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5197 struct task_struct *p;
5198 unsigned int time_slice;
5208 p = find_process_by_pid(pid);
5212 retval = security_task_getscheduler(p);
5216 rq = task_rq_lock(p, &rf);
5218 if (p->sched_class->get_rr_interval)
5219 time_slice = p->sched_class->get_rr_interval(rq, p);
5220 task_rq_unlock(rq, p, &rf);
5223 jiffies_to_timespec64(time_slice, t);
5232 * sys_sched_rr_get_interval - return the default timeslice of a process.
5233 * @pid: pid of the process.
5234 * @interval: userspace pointer to the timeslice value.
5236 * this syscall writes the default timeslice value of a given process
5237 * into the user-space timespec buffer. A value of '0' means infinity.
5239 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5242 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5243 struct __kernel_timespec __user *, interval)
5245 struct timespec64 t;
5246 int retval = sched_rr_get_interval(pid, &t);
5249 retval = put_timespec64(&t, interval);
5254 #ifdef CONFIG_COMPAT_32BIT_TIME
5255 COMPAT_SYSCALL_DEFINE2(sched_rr_get_interval,
5257 struct old_timespec32 __user *, interval)
5259 struct timespec64 t;
5260 int retval = sched_rr_get_interval(pid, &t);
5263 retval = put_old_timespec32(&t, interval);
5268 void sched_show_task(struct task_struct *p)
5270 unsigned long free = 0;
5273 if (!try_get_task_stack(p))
5276 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5278 if (p->state == TASK_RUNNING)
5279 printk(KERN_CONT " running task ");
5280 #ifdef CONFIG_DEBUG_STACK_USAGE
5281 free = stack_not_used(p);
5286 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5288 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5289 task_pid_nr(p), ppid,
5290 (unsigned long)task_thread_info(p)->flags);
5292 print_worker_info(KERN_INFO, p);
5293 show_stack(p, NULL);
5296 EXPORT_SYMBOL_GPL(sched_show_task);
5299 state_filter_match(unsigned long state_filter, struct task_struct *p)
5301 /* no filter, everything matches */
5305 /* filter, but doesn't match */
5306 if (!(p->state & state_filter))
5310 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5313 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5320 void show_state_filter(unsigned long state_filter)
5322 struct task_struct *g, *p;
5324 #if BITS_PER_LONG == 32
5326 " task PC stack pid father\n");
5329 " task PC stack pid father\n");
5332 for_each_process_thread(g, p) {
5334 * reset the NMI-timeout, listing all files on a slow
5335 * console might take a lot of time:
5336 * Also, reset softlockup watchdogs on all CPUs, because
5337 * another CPU might be blocked waiting for us to process
5340 touch_nmi_watchdog();
5341 touch_all_softlockup_watchdogs();
5342 if (state_filter_match(state_filter, p))
5346 #ifdef CONFIG_SCHED_DEBUG
5348 sysrq_sched_debug_show();
5352 * Only show locks if all tasks are dumped:
5355 debug_show_all_locks();
5359 * init_idle - set up an idle thread for a given CPU
5360 * @idle: task in question
5361 * @cpu: CPU the idle task belongs to
5363 * NOTE: this function does not set the idle thread's NEED_RESCHED
5364 * flag, to make booting more robust.
5366 void init_idle(struct task_struct *idle, int cpu)
5368 struct rq *rq = cpu_rq(cpu);
5369 unsigned long flags;
5371 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5372 raw_spin_lock(&rq->lock);
5374 __sched_fork(0, idle);
5375 idle->state = TASK_RUNNING;
5376 idle->se.exec_start = sched_clock();
5377 idle->flags |= PF_IDLE;
5379 kasan_unpoison_task_stack(idle);
5383 * Its possible that init_idle() gets called multiple times on a task,
5384 * in that case do_set_cpus_allowed() will not do the right thing.
5386 * And since this is boot we can forgo the serialization.
5388 set_cpus_allowed_common(idle, cpumask_of(cpu));
5391 * We're having a chicken and egg problem, even though we are
5392 * holding rq->lock, the CPU isn't yet set to this CPU so the
5393 * lockdep check in task_group() will fail.
5395 * Similar case to sched_fork(). / Alternatively we could
5396 * use task_rq_lock() here and obtain the other rq->lock.
5401 __set_task_cpu(idle, cpu);
5404 rq->curr = rq->idle = idle;
5405 idle->on_rq = TASK_ON_RQ_QUEUED;
5409 raw_spin_unlock(&rq->lock);
5410 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5412 /* Set the preempt count _outside_ the spinlocks! */
5413 init_idle_preempt_count(idle, cpu);
5416 * The idle tasks have their own, simple scheduling class:
5418 idle->sched_class = &idle_sched_class;
5419 ftrace_graph_init_idle_task(idle, cpu);
5420 vtime_init_idle(idle, cpu);
5422 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5428 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5429 const struct cpumask *trial)
5433 if (!cpumask_weight(cur))
5436 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5441 int task_can_attach(struct task_struct *p,
5442 const struct cpumask *cs_cpus_allowed)
5447 * Kthreads which disallow setaffinity shouldn't be moved
5448 * to a new cpuset; we don't want to change their CPU
5449 * affinity and isolating such threads by their set of
5450 * allowed nodes is unnecessary. Thus, cpusets are not
5451 * applicable for such threads. This prevents checking for
5452 * success of set_cpus_allowed_ptr() on all attached tasks
5453 * before cpus_allowed may be changed.
5455 if (p->flags & PF_NO_SETAFFINITY) {
5460 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5462 ret = dl_task_can_attach(p, cs_cpus_allowed);
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 | ENQUEUE_NOCLOCK);
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(mm, &init_mm, current);
5531 current->active_mm = &init_mm;
5532 finish_arch_post_lock_switch();
5538 * Since this CPU is going 'away' for a while, fold any nr_active delta
5539 * we might have. Assumes we're called after migrate_tasks() so that the
5540 * nr_active count is stable. We need to take the teardown thread which
5541 * is calling this into account, so we hand in adjust = 1 to the load
5544 * Also see the comment "Global load-average calculations".
5546 static void calc_load_migrate(struct rq *rq)
5548 long delta = calc_load_fold_active(rq, 1);
5550 atomic_long_add(delta, &calc_load_tasks);
5553 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5557 static const struct sched_class fake_sched_class = {
5558 .put_prev_task = put_prev_task_fake,
5561 static struct task_struct fake_task = {
5563 * Avoid pull_{rt,dl}_task()
5565 .prio = MAX_PRIO + 1,
5566 .sched_class = &fake_sched_class,
5570 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5571 * try_to_wake_up()->select_task_rq().
5573 * Called with rq->lock held even though we'er in stop_machine() and
5574 * there's no concurrency possible, we hold the required locks anyway
5575 * because of lock validation efforts.
5577 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5579 struct rq *rq = dead_rq;
5580 struct task_struct *next, *stop = rq->stop;
5581 struct rq_flags orf = *rf;
5585 * Fudge the rq selection such that the below task selection loop
5586 * doesn't get stuck on the currently eligible stop task.
5588 * We're currently inside stop_machine() and the rq is either stuck
5589 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5590 * either way we should never end up calling schedule() until we're
5596 * put_prev_task() and pick_next_task() sched
5597 * class method both need to have an up-to-date
5598 * value of rq->clock[_task]
5600 update_rq_clock(rq);
5604 * There's this thread running, bail when that's the only
5607 if (rq->nr_running == 1)
5611 * pick_next_task() assumes pinned rq->lock:
5613 next = pick_next_task(rq, &fake_task, rf);
5615 put_prev_task(rq, next);
5618 * Rules for changing task_struct::cpus_allowed are holding
5619 * both pi_lock and rq->lock, such that holding either
5620 * stabilizes the mask.
5622 * Drop rq->lock is not quite as disastrous as it usually is
5623 * because !cpu_active at this point, which means load-balance
5624 * will not interfere. Also, stop-machine.
5627 raw_spin_lock(&next->pi_lock);
5631 * Since we're inside stop-machine, _nothing_ should have
5632 * changed the task, WARN if weird stuff happened, because in
5633 * that case the above rq->lock drop is a fail too.
5635 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5636 raw_spin_unlock(&next->pi_lock);
5640 /* Find suitable destination for @next, with force if needed. */
5641 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5642 rq = __migrate_task(rq, rf, next, dest_cpu);
5643 if (rq != dead_rq) {
5649 raw_spin_unlock(&next->pi_lock);
5654 #endif /* CONFIG_HOTPLUG_CPU */
5656 void set_rq_online(struct rq *rq)
5659 const struct sched_class *class;
5661 cpumask_set_cpu(rq->cpu, rq->rd->online);
5664 for_each_class(class) {
5665 if (class->rq_online)
5666 class->rq_online(rq);
5671 void set_rq_offline(struct rq *rq)
5674 const struct sched_class *class;
5676 for_each_class(class) {
5677 if (class->rq_offline)
5678 class->rq_offline(rq);
5681 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5687 * used to mark begin/end of suspend/resume:
5689 static int num_cpus_frozen;
5692 * Update cpusets according to cpu_active mask. If cpusets are
5693 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5694 * around partition_sched_domains().
5696 * If we come here as part of a suspend/resume, don't touch cpusets because we
5697 * want to restore it back to its original state upon resume anyway.
5699 static void cpuset_cpu_active(void)
5701 if (cpuhp_tasks_frozen) {
5703 * num_cpus_frozen tracks how many CPUs are involved in suspend
5704 * resume sequence. As long as this is not the last online
5705 * operation in the resume sequence, just build a single sched
5706 * domain, ignoring cpusets.
5708 partition_sched_domains(1, NULL, NULL);
5709 if (--num_cpus_frozen)
5712 * This is the last CPU online operation. So fall through and
5713 * restore the original sched domains by considering the
5714 * cpuset configurations.
5716 cpuset_force_rebuild();
5718 cpuset_update_active_cpus();
5721 static int cpuset_cpu_inactive(unsigned int cpu)
5723 if (!cpuhp_tasks_frozen) {
5724 if (dl_cpu_busy(cpu))
5726 cpuset_update_active_cpus();
5729 partition_sched_domains(1, NULL, NULL);
5734 int sched_cpu_activate(unsigned int cpu)
5736 struct rq *rq = cpu_rq(cpu);
5739 #ifdef CONFIG_SCHED_SMT
5741 * The sched_smt_present static key needs to be evaluated on every
5742 * hotplug event because at boot time SMT might be disabled when
5743 * the number of booted CPUs is limited.
5745 * If then later a sibling gets hotplugged, then the key would stay
5746 * off and SMT scheduling would never be functional.
5748 if (cpumask_weight(cpu_smt_mask(cpu)) > 1)
5749 static_branch_enable_cpuslocked(&sched_smt_present);
5751 set_cpu_active(cpu, true);
5753 if (sched_smp_initialized) {
5754 sched_domains_numa_masks_set(cpu);
5755 cpuset_cpu_active();
5759 * Put the rq online, if not already. This happens:
5761 * 1) In the early boot process, because we build the real domains
5762 * after all CPUs have been brought up.
5764 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5767 rq_lock_irqsave(rq, &rf);
5769 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5772 rq_unlock_irqrestore(rq, &rf);
5774 update_max_interval();
5779 int sched_cpu_deactivate(unsigned int cpu)
5783 set_cpu_active(cpu, false);
5785 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5786 * users of this state to go away such that all new such users will
5789 * Do sync before park smpboot threads to take care the rcu boost case.
5791 synchronize_rcu_mult(call_rcu, call_rcu_sched);
5793 if (!sched_smp_initialized)
5796 ret = cpuset_cpu_inactive(cpu);
5798 set_cpu_active(cpu, true);
5801 sched_domains_numa_masks_clear(cpu);
5805 static void sched_rq_cpu_starting(unsigned int cpu)
5807 struct rq *rq = cpu_rq(cpu);
5809 rq->calc_load_update = calc_load_update;
5810 update_max_interval();
5813 int sched_cpu_starting(unsigned int cpu)
5815 sched_rq_cpu_starting(cpu);
5816 sched_tick_start(cpu);
5820 #ifdef CONFIG_HOTPLUG_CPU
5821 int sched_cpu_dying(unsigned int cpu)
5823 struct rq *rq = cpu_rq(cpu);
5826 /* Handle pending wakeups and then migrate everything off */
5827 sched_ttwu_pending();
5828 sched_tick_stop(cpu);
5830 rq_lock_irqsave(rq, &rf);
5832 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5835 migrate_tasks(rq, &rf);
5836 BUG_ON(rq->nr_running != 1);
5837 rq_unlock_irqrestore(rq, &rf);
5839 calc_load_migrate(rq);
5840 update_max_interval();
5841 nohz_balance_exit_idle(rq);
5847 void __init sched_init_smp(void)
5852 * There's no userspace yet to cause hotplug operations; hence all the
5853 * CPU masks are stable and all blatant races in the below code cannot
5854 * happen. The hotplug lock is nevertheless taken to satisfy lockdep,
5855 * but there won't be any contention on it.
5858 mutex_lock(&sched_domains_mutex);
5859 sched_init_domains(cpu_active_mask);
5860 mutex_unlock(&sched_domains_mutex);
5863 /* Move init over to a non-isolated CPU */
5864 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
5866 sched_init_granularity();
5868 init_sched_rt_class();
5869 init_sched_dl_class();
5871 sched_smp_initialized = true;
5874 static int __init migration_init(void)
5876 sched_rq_cpu_starting(smp_processor_id());
5879 early_initcall(migration_init);
5882 void __init sched_init_smp(void)
5884 sched_init_granularity();
5886 #endif /* CONFIG_SMP */
5888 int in_sched_functions(unsigned long addr)
5890 return in_lock_functions(addr) ||
5891 (addr >= (unsigned long)__sched_text_start
5892 && addr < (unsigned long)__sched_text_end);
5895 #ifdef CONFIG_CGROUP_SCHED
5897 * Default task group.
5898 * Every task in system belongs to this group at bootup.
5900 struct task_group root_task_group;
5901 LIST_HEAD(task_groups);
5903 /* Cacheline aligned slab cache for task_group */
5904 static struct kmem_cache *task_group_cache __read_mostly;
5907 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5908 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5910 void __init sched_init(void)
5913 unsigned long alloc_size = 0, ptr;
5917 #ifdef CONFIG_FAIR_GROUP_SCHED
5918 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5920 #ifdef CONFIG_RT_GROUP_SCHED
5921 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5924 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5926 #ifdef CONFIG_FAIR_GROUP_SCHED
5927 root_task_group.se = (struct sched_entity **)ptr;
5928 ptr += nr_cpu_ids * sizeof(void **);
5930 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5931 ptr += nr_cpu_ids * sizeof(void **);
5933 #endif /* CONFIG_FAIR_GROUP_SCHED */
5934 #ifdef CONFIG_RT_GROUP_SCHED
5935 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5936 ptr += nr_cpu_ids * sizeof(void **);
5938 root_task_group.rt_rq = (struct rt_rq **)ptr;
5939 ptr += nr_cpu_ids * sizeof(void **);
5941 #endif /* CONFIG_RT_GROUP_SCHED */
5943 #ifdef CONFIG_CPUMASK_OFFSTACK
5944 for_each_possible_cpu(i) {
5945 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5946 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5947 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
5948 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5950 #endif /* CONFIG_CPUMASK_OFFSTACK */
5952 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
5953 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
5956 init_defrootdomain();
5959 #ifdef CONFIG_RT_GROUP_SCHED
5960 init_rt_bandwidth(&root_task_group.rt_bandwidth,
5961 global_rt_period(), global_rt_runtime());
5962 #endif /* CONFIG_RT_GROUP_SCHED */
5964 #ifdef CONFIG_CGROUP_SCHED
5965 task_group_cache = KMEM_CACHE(task_group, 0);
5967 list_add(&root_task_group.list, &task_groups);
5968 INIT_LIST_HEAD(&root_task_group.children);
5969 INIT_LIST_HEAD(&root_task_group.siblings);
5970 autogroup_init(&init_task);
5971 #endif /* CONFIG_CGROUP_SCHED */
5973 for_each_possible_cpu(i) {
5977 raw_spin_lock_init(&rq->lock);
5979 rq->calc_load_active = 0;
5980 rq->calc_load_update = jiffies + LOAD_FREQ;
5981 init_cfs_rq(&rq->cfs);
5982 init_rt_rq(&rq->rt);
5983 init_dl_rq(&rq->dl);
5984 #ifdef CONFIG_FAIR_GROUP_SCHED
5985 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
5986 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
5987 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
5989 * How much CPU bandwidth does root_task_group get?
5991 * In case of task-groups formed thr' the cgroup filesystem, it
5992 * gets 100% of the CPU resources in the system. This overall
5993 * system CPU resource is divided among the tasks of
5994 * root_task_group and its child task-groups in a fair manner,
5995 * based on each entity's (task or task-group's) weight
5996 * (se->load.weight).
5998 * In other words, if root_task_group has 10 tasks of weight
5999 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6000 * then A0's share of the CPU resource is:
6002 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6004 * We achieve this by letting root_task_group's tasks sit
6005 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6007 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6008 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6009 #endif /* CONFIG_FAIR_GROUP_SCHED */
6011 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6012 #ifdef CONFIG_RT_GROUP_SCHED
6013 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6016 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6017 rq->cpu_load[j] = 0;
6022 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6023 rq->balance_callback = NULL;
6024 rq->active_balance = 0;
6025 rq->next_balance = jiffies;
6030 rq->avg_idle = 2*sysctl_sched_migration_cost;
6031 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6033 INIT_LIST_HEAD(&rq->cfs_tasks);
6035 rq_attach_root(rq, &def_root_domain);
6036 #ifdef CONFIG_NO_HZ_COMMON
6037 rq->last_load_update_tick = jiffies;
6038 rq->last_blocked_load_update_tick = jiffies;
6039 atomic_set(&rq->nohz_flags, 0);
6041 #endif /* CONFIG_SMP */
6043 atomic_set(&rq->nr_iowait, 0);
6046 set_load_weight(&init_task, false);
6049 * The boot idle thread does lazy MMU switching as well:
6052 enter_lazy_tlb(&init_mm, current);
6055 * Make us the idle thread. Technically, schedule() should not be
6056 * called from this thread, however somewhere below it might be,
6057 * but because we are the idle thread, we just pick up running again
6058 * when this runqueue becomes "idle".
6060 init_idle(current, smp_processor_id());
6062 calc_load_update = jiffies + LOAD_FREQ;
6065 idle_thread_set_boot_cpu();
6067 init_sched_fair_class();
6073 scheduler_running = 1;
6076 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6077 static inline int preempt_count_equals(int preempt_offset)
6079 int nested = preempt_count() + rcu_preempt_depth();
6081 return (nested == preempt_offset);
6084 void __might_sleep(const char *file, int line, int preempt_offset)
6087 * Blocking primitives will set (and therefore destroy) current->state,
6088 * since we will exit with TASK_RUNNING make sure we enter with it,
6089 * otherwise we will destroy state.
6091 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6092 "do not call blocking ops when !TASK_RUNNING; "
6093 "state=%lx set at [<%p>] %pS\n",
6095 (void *)current->task_state_change,
6096 (void *)current->task_state_change);
6098 ___might_sleep(file, line, preempt_offset);
6100 EXPORT_SYMBOL(__might_sleep);
6102 void ___might_sleep(const char *file, int line, int preempt_offset)
6104 /* Ratelimiting timestamp: */
6105 static unsigned long prev_jiffy;
6107 unsigned long preempt_disable_ip;
6109 /* WARN_ON_ONCE() by default, no rate limit required: */
6112 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6113 !is_idle_task(current)) ||
6114 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6118 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6120 prev_jiffy = jiffies;
6122 /* Save this before calling printk(), since that will clobber it: */
6123 preempt_disable_ip = get_preempt_disable_ip(current);
6126 "BUG: sleeping function called from invalid context at %s:%d\n",
6129 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6130 in_atomic(), irqs_disabled(),
6131 current->pid, current->comm);
6133 if (task_stack_end_corrupted(current))
6134 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6136 debug_show_held_locks(current);
6137 if (irqs_disabled())
6138 print_irqtrace_events(current);
6139 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6140 && !preempt_count_equals(preempt_offset)) {
6141 pr_err("Preemption disabled at:");
6142 print_ip_sym(preempt_disable_ip);
6146 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6148 EXPORT_SYMBOL(___might_sleep);
6151 #ifdef CONFIG_MAGIC_SYSRQ
6152 void normalize_rt_tasks(void)
6154 struct task_struct *g, *p;
6155 struct sched_attr attr = {
6156 .sched_policy = SCHED_NORMAL,
6159 read_lock(&tasklist_lock);
6160 for_each_process_thread(g, p) {
6162 * Only normalize user tasks:
6164 if (p->flags & PF_KTHREAD)
6167 p->se.exec_start = 0;
6168 schedstat_set(p->se.statistics.wait_start, 0);
6169 schedstat_set(p->se.statistics.sleep_start, 0);
6170 schedstat_set(p->se.statistics.block_start, 0);
6172 if (!dl_task(p) && !rt_task(p)) {
6174 * Renice negative nice level userspace
6177 if (task_nice(p) < 0)
6178 set_user_nice(p, 0);
6182 __sched_setscheduler(p, &attr, false, false);
6184 read_unlock(&tasklist_lock);
6187 #endif /* CONFIG_MAGIC_SYSRQ */
6189 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6191 * These functions are only useful for the IA64 MCA handling, or kdb.
6193 * They can only be called when the whole system has been
6194 * stopped - every CPU needs to be quiescent, and no scheduling
6195 * activity can take place. Using them for anything else would
6196 * be a serious bug, and as a result, they aren't even visible
6197 * under any other configuration.
6201 * curr_task - return the current task for a given CPU.
6202 * @cpu: the processor in question.
6204 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6206 * Return: The current task for @cpu.
6208 struct task_struct *curr_task(int cpu)
6210 return cpu_curr(cpu);
6213 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6217 * set_curr_task - set the current task for a given CPU.
6218 * @cpu: the processor in question.
6219 * @p: the task pointer to set.
6221 * Description: This function must only be used when non-maskable interrupts
6222 * are serviced on a separate stack. It allows the architecture to switch the
6223 * notion of the current task on a CPU in a non-blocking manner. This function
6224 * must be called with all CPU's synchronized, and interrupts disabled, the
6225 * and caller must save the original value of the current task (see
6226 * curr_task() above) and restore that value before reenabling interrupts and
6227 * re-starting the system.
6229 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6231 void ia64_set_curr_task(int cpu, struct task_struct *p)
6238 #ifdef CONFIG_CGROUP_SCHED
6239 /* task_group_lock serializes the addition/removal of task groups */
6240 static DEFINE_SPINLOCK(task_group_lock);
6242 static void sched_free_group(struct task_group *tg)
6244 free_fair_sched_group(tg);
6245 free_rt_sched_group(tg);
6247 kmem_cache_free(task_group_cache, tg);
6250 /* allocate runqueue etc for a new task group */
6251 struct task_group *sched_create_group(struct task_group *parent)
6253 struct task_group *tg;
6255 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6257 return ERR_PTR(-ENOMEM);
6259 if (!alloc_fair_sched_group(tg, parent))
6262 if (!alloc_rt_sched_group(tg, parent))
6268 sched_free_group(tg);
6269 return ERR_PTR(-ENOMEM);
6272 void sched_online_group(struct task_group *tg, struct task_group *parent)
6274 unsigned long flags;
6276 spin_lock_irqsave(&task_group_lock, flags);
6277 list_add_rcu(&tg->list, &task_groups);
6279 /* Root should already exist: */
6282 tg->parent = parent;
6283 INIT_LIST_HEAD(&tg->children);
6284 list_add_rcu(&tg->siblings, &parent->children);
6285 spin_unlock_irqrestore(&task_group_lock, flags);
6287 online_fair_sched_group(tg);
6290 /* rcu callback to free various structures associated with a task group */
6291 static void sched_free_group_rcu(struct rcu_head *rhp)
6293 /* Now it should be safe to free those cfs_rqs: */
6294 sched_free_group(container_of(rhp, struct task_group, rcu));
6297 void sched_destroy_group(struct task_group *tg)
6299 /* Wait for possible concurrent references to cfs_rqs complete: */
6300 call_rcu(&tg->rcu, sched_free_group_rcu);
6303 void sched_offline_group(struct task_group *tg)
6305 unsigned long flags;
6307 /* End participation in shares distribution: */
6308 unregister_fair_sched_group(tg);
6310 spin_lock_irqsave(&task_group_lock, flags);
6311 list_del_rcu(&tg->list);
6312 list_del_rcu(&tg->siblings);
6313 spin_unlock_irqrestore(&task_group_lock, flags);
6316 static void sched_change_group(struct task_struct *tsk, int type)
6318 struct task_group *tg;
6321 * All callers are synchronized by task_rq_lock(); we do not use RCU
6322 * which is pointless here. Thus, we pass "true" to task_css_check()
6323 * to prevent lockdep warnings.
6325 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6326 struct task_group, css);
6327 tg = autogroup_task_group(tsk, tg);
6328 tsk->sched_task_group = tg;
6330 #ifdef CONFIG_FAIR_GROUP_SCHED
6331 if (tsk->sched_class->task_change_group)
6332 tsk->sched_class->task_change_group(tsk, type);
6335 set_task_rq(tsk, task_cpu(tsk));
6339 * Change task's runqueue when it moves between groups.
6341 * The caller of this function should have put the task in its new group by
6342 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6345 void sched_move_task(struct task_struct *tsk)
6347 int queued, running, queue_flags =
6348 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6352 rq = task_rq_lock(tsk, &rf);
6353 update_rq_clock(rq);
6355 running = task_current(rq, tsk);
6356 queued = task_on_rq_queued(tsk);
6359 dequeue_task(rq, tsk, queue_flags);
6361 put_prev_task(rq, tsk);
6363 sched_change_group(tsk, TASK_MOVE_GROUP);
6366 enqueue_task(rq, tsk, queue_flags);
6368 set_curr_task(rq, tsk);
6370 task_rq_unlock(rq, tsk, &rf);
6373 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6375 return css ? container_of(css, struct task_group, css) : NULL;
6378 static struct cgroup_subsys_state *
6379 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6381 struct task_group *parent = css_tg(parent_css);
6382 struct task_group *tg;
6385 /* This is early initialization for the top cgroup */
6386 return &root_task_group.css;
6389 tg = sched_create_group(parent);
6391 return ERR_PTR(-ENOMEM);
6396 /* Expose task group only after completing cgroup initialization */
6397 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6399 struct task_group *tg = css_tg(css);
6400 struct task_group *parent = css_tg(css->parent);
6403 sched_online_group(tg, parent);
6407 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6409 struct task_group *tg = css_tg(css);
6411 sched_offline_group(tg);
6414 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6416 struct task_group *tg = css_tg(css);
6419 * Relies on the RCU grace period between css_released() and this.
6421 sched_free_group(tg);
6425 * This is called before wake_up_new_task(), therefore we really only
6426 * have to set its group bits, all the other stuff does not apply.
6428 static void cpu_cgroup_fork(struct task_struct *task)
6433 rq = task_rq_lock(task, &rf);
6435 update_rq_clock(rq);
6436 sched_change_group(task, TASK_SET_GROUP);
6438 task_rq_unlock(rq, task, &rf);
6441 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6443 struct task_struct *task;
6444 struct cgroup_subsys_state *css;
6447 cgroup_taskset_for_each(task, css, tset) {
6448 #ifdef CONFIG_RT_GROUP_SCHED
6449 if (!sched_rt_can_attach(css_tg(css), task))
6452 /* We don't support RT-tasks being in separate groups */
6453 if (task->sched_class != &fair_sched_class)
6457 * Serialize against wake_up_new_task() such that if its
6458 * running, we're sure to observe its full state.
6460 raw_spin_lock_irq(&task->pi_lock);
6462 * Avoid calling sched_move_task() before wake_up_new_task()
6463 * has happened. This would lead to problems with PELT, due to
6464 * move wanting to detach+attach while we're not attached yet.
6466 if (task->state == TASK_NEW)
6468 raw_spin_unlock_irq(&task->pi_lock);
6476 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6478 struct task_struct *task;
6479 struct cgroup_subsys_state *css;
6481 cgroup_taskset_for_each(task, css, tset)
6482 sched_move_task(task);
6485 #ifdef CONFIG_FAIR_GROUP_SCHED
6486 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6487 struct cftype *cftype, u64 shareval)
6489 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6492 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6495 struct task_group *tg = css_tg(css);
6497 return (u64) scale_load_down(tg->shares);
6500 #ifdef CONFIG_CFS_BANDWIDTH
6501 static DEFINE_MUTEX(cfs_constraints_mutex);
6503 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6504 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6506 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6508 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6510 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6511 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6513 if (tg == &root_task_group)
6517 * Ensure we have at some amount of bandwidth every period. This is
6518 * to prevent reaching a state of large arrears when throttled via
6519 * entity_tick() resulting in prolonged exit starvation.
6521 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6525 * Likewise, bound things on the otherside by preventing insane quota
6526 * periods. This also allows us to normalize in computing quota
6529 if (period > max_cfs_quota_period)
6533 * Prevent race between setting of cfs_rq->runtime_enabled and
6534 * unthrottle_offline_cfs_rqs().
6537 mutex_lock(&cfs_constraints_mutex);
6538 ret = __cfs_schedulable(tg, period, quota);
6542 runtime_enabled = quota != RUNTIME_INF;
6543 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6545 * If we need to toggle cfs_bandwidth_used, off->on must occur
6546 * before making related changes, and on->off must occur afterwards
6548 if (runtime_enabled && !runtime_was_enabled)
6549 cfs_bandwidth_usage_inc();
6550 raw_spin_lock_irq(&cfs_b->lock);
6551 cfs_b->period = ns_to_ktime(period);
6552 cfs_b->quota = quota;
6554 __refill_cfs_bandwidth_runtime(cfs_b);
6556 /* Restart the period timer (if active) to handle new period expiry: */
6557 if (runtime_enabled)
6558 start_cfs_bandwidth(cfs_b);
6560 raw_spin_unlock_irq(&cfs_b->lock);
6562 for_each_online_cpu(i) {
6563 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6564 struct rq *rq = cfs_rq->rq;
6567 rq_lock_irq(rq, &rf);
6568 cfs_rq->runtime_enabled = runtime_enabled;
6569 cfs_rq->runtime_remaining = 0;
6571 if (cfs_rq->throttled)
6572 unthrottle_cfs_rq(cfs_rq);
6573 rq_unlock_irq(rq, &rf);
6575 if (runtime_was_enabled && !runtime_enabled)
6576 cfs_bandwidth_usage_dec();
6578 mutex_unlock(&cfs_constraints_mutex);
6584 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
6588 period = ktime_to_ns(tg->cfs_bandwidth.period);
6589 if (cfs_quota_us < 0)
6590 quota = RUNTIME_INF;
6592 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
6594 return tg_set_cfs_bandwidth(tg, period, quota);
6597 long tg_get_cfs_quota(struct task_group *tg)
6601 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
6604 quota_us = tg->cfs_bandwidth.quota;
6605 do_div(quota_us, NSEC_PER_USEC);
6610 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
6614 period = (u64)cfs_period_us * NSEC_PER_USEC;
6615 quota = tg->cfs_bandwidth.quota;
6617 return tg_set_cfs_bandwidth(tg, period, quota);
6620 long tg_get_cfs_period(struct task_group *tg)
6624 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
6625 do_div(cfs_period_us, NSEC_PER_USEC);
6627 return cfs_period_us;
6630 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
6633 return tg_get_cfs_quota(css_tg(css));
6636 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
6637 struct cftype *cftype, s64 cfs_quota_us)
6639 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
6642 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
6645 return tg_get_cfs_period(css_tg(css));
6648 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
6649 struct cftype *cftype, u64 cfs_period_us)
6651 return tg_set_cfs_period(css_tg(css), cfs_period_us);
6654 struct cfs_schedulable_data {
6655 struct task_group *tg;
6660 * normalize group quota/period to be quota/max_period
6661 * note: units are usecs
6663 static u64 normalize_cfs_quota(struct task_group *tg,
6664 struct cfs_schedulable_data *d)
6672 period = tg_get_cfs_period(tg);
6673 quota = tg_get_cfs_quota(tg);
6676 /* note: these should typically be equivalent */
6677 if (quota == RUNTIME_INF || quota == -1)
6680 return to_ratio(period, quota);
6683 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
6685 struct cfs_schedulable_data *d = data;
6686 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6687 s64 quota = 0, parent_quota = -1;
6690 quota = RUNTIME_INF;
6692 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
6694 quota = normalize_cfs_quota(tg, d);
6695 parent_quota = parent_b->hierarchical_quota;
6698 * Ensure max(child_quota) <= parent_quota. On cgroup2,
6699 * always take the min. On cgroup1, only inherit when no
6702 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
6703 quota = min(quota, parent_quota);
6705 if (quota == RUNTIME_INF)
6706 quota = parent_quota;
6707 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
6711 cfs_b->hierarchical_quota = quota;
6716 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
6719 struct cfs_schedulable_data data = {
6725 if (quota != RUNTIME_INF) {
6726 do_div(data.period, NSEC_PER_USEC);
6727 do_div(data.quota, NSEC_PER_USEC);
6731 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
6737 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
6739 struct task_group *tg = css_tg(seq_css(sf));
6740 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6742 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
6743 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
6744 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
6746 if (schedstat_enabled() && tg != &root_task_group) {
6750 for_each_possible_cpu(i)
6751 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
6753 seq_printf(sf, "wait_sum %llu\n", ws);
6758 #endif /* CONFIG_CFS_BANDWIDTH */
6759 #endif /* CONFIG_FAIR_GROUP_SCHED */
6761 #ifdef CONFIG_RT_GROUP_SCHED
6762 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
6763 struct cftype *cft, s64 val)
6765 return sched_group_set_rt_runtime(css_tg(css), val);
6768 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
6771 return sched_group_rt_runtime(css_tg(css));
6774 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
6775 struct cftype *cftype, u64 rt_period_us)
6777 return sched_group_set_rt_period(css_tg(css), rt_period_us);
6780 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
6783 return sched_group_rt_period(css_tg(css));
6785 #endif /* CONFIG_RT_GROUP_SCHED */
6787 static struct cftype cpu_legacy_files[] = {
6788 #ifdef CONFIG_FAIR_GROUP_SCHED
6791 .read_u64 = cpu_shares_read_u64,
6792 .write_u64 = cpu_shares_write_u64,
6795 #ifdef CONFIG_CFS_BANDWIDTH
6797 .name = "cfs_quota_us",
6798 .read_s64 = cpu_cfs_quota_read_s64,
6799 .write_s64 = cpu_cfs_quota_write_s64,
6802 .name = "cfs_period_us",
6803 .read_u64 = cpu_cfs_period_read_u64,
6804 .write_u64 = cpu_cfs_period_write_u64,
6808 .seq_show = cpu_cfs_stat_show,
6811 #ifdef CONFIG_RT_GROUP_SCHED
6813 .name = "rt_runtime_us",
6814 .read_s64 = cpu_rt_runtime_read,
6815 .write_s64 = cpu_rt_runtime_write,
6818 .name = "rt_period_us",
6819 .read_u64 = cpu_rt_period_read_uint,
6820 .write_u64 = cpu_rt_period_write_uint,
6826 static int cpu_extra_stat_show(struct seq_file *sf,
6827 struct cgroup_subsys_state *css)
6829 #ifdef CONFIG_CFS_BANDWIDTH
6831 struct task_group *tg = css_tg(css);
6832 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6835 throttled_usec = cfs_b->throttled_time;
6836 do_div(throttled_usec, NSEC_PER_USEC);
6838 seq_printf(sf, "nr_periods %d\n"
6840 "throttled_usec %llu\n",
6841 cfs_b->nr_periods, cfs_b->nr_throttled,
6848 #ifdef CONFIG_FAIR_GROUP_SCHED
6849 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
6852 struct task_group *tg = css_tg(css);
6853 u64 weight = scale_load_down(tg->shares);
6855 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
6858 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
6859 struct cftype *cft, u64 weight)
6862 * cgroup weight knobs should use the common MIN, DFL and MAX
6863 * values which are 1, 100 and 10000 respectively. While it loses
6864 * a bit of range on both ends, it maps pretty well onto the shares
6865 * value used by scheduler and the round-trip conversions preserve
6866 * the original value over the entire range.
6868 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
6871 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
6873 return sched_group_set_shares(css_tg(css), scale_load(weight));
6876 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
6879 unsigned long weight = scale_load_down(css_tg(css)->shares);
6880 int last_delta = INT_MAX;
6883 /* find the closest nice value to the current weight */
6884 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
6885 delta = abs(sched_prio_to_weight[prio] - weight);
6886 if (delta >= last_delta)
6891 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
6894 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
6895 struct cftype *cft, s64 nice)
6897 unsigned long weight;
6900 if (nice < MIN_NICE || nice > MAX_NICE)
6903 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
6904 idx = array_index_nospec(idx, 40);
6905 weight = sched_prio_to_weight[idx];
6907 return sched_group_set_shares(css_tg(css), scale_load(weight));
6911 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
6912 long period, long quota)
6915 seq_puts(sf, "max");
6917 seq_printf(sf, "%ld", quota);
6919 seq_printf(sf, " %ld\n", period);
6922 /* caller should put the current value in *@periodp before calling */
6923 static int __maybe_unused cpu_period_quota_parse(char *buf,
6924 u64 *periodp, u64 *quotap)
6926 char tok[21]; /* U64_MAX */
6928 if (!sscanf(buf, "%s %llu", tok, periodp))
6931 *periodp *= NSEC_PER_USEC;
6933 if (sscanf(tok, "%llu", quotap))
6934 *quotap *= NSEC_PER_USEC;
6935 else if (!strcmp(tok, "max"))
6936 *quotap = RUNTIME_INF;
6943 #ifdef CONFIG_CFS_BANDWIDTH
6944 static int cpu_max_show(struct seq_file *sf, void *v)
6946 struct task_group *tg = css_tg(seq_css(sf));
6948 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
6952 static ssize_t cpu_max_write(struct kernfs_open_file *of,
6953 char *buf, size_t nbytes, loff_t off)
6955 struct task_group *tg = css_tg(of_css(of));
6956 u64 period = tg_get_cfs_period(tg);
6960 ret = cpu_period_quota_parse(buf, &period, "a);
6962 ret = tg_set_cfs_bandwidth(tg, period, quota);
6963 return ret ?: nbytes;
6967 static struct cftype cpu_files[] = {
6968 #ifdef CONFIG_FAIR_GROUP_SCHED
6971 .flags = CFTYPE_NOT_ON_ROOT,
6972 .read_u64 = cpu_weight_read_u64,
6973 .write_u64 = cpu_weight_write_u64,
6976 .name = "weight.nice",
6977 .flags = CFTYPE_NOT_ON_ROOT,
6978 .read_s64 = cpu_weight_nice_read_s64,
6979 .write_s64 = cpu_weight_nice_write_s64,
6982 #ifdef CONFIG_CFS_BANDWIDTH
6985 .flags = CFTYPE_NOT_ON_ROOT,
6986 .seq_show = cpu_max_show,
6987 .write = cpu_max_write,
6993 struct cgroup_subsys cpu_cgrp_subsys = {
6994 .css_alloc = cpu_cgroup_css_alloc,
6995 .css_online = cpu_cgroup_css_online,
6996 .css_released = cpu_cgroup_css_released,
6997 .css_free = cpu_cgroup_css_free,
6998 .css_extra_stat_show = cpu_extra_stat_show,
6999 .fork = cpu_cgroup_fork,
7000 .can_attach = cpu_cgroup_can_attach,
7001 .attach = cpu_cgroup_attach,
7002 .legacy_cftypes = cpu_legacy_files,
7003 .dfl_cftypes = cpu_files,
7008 #endif /* CONFIG_CGROUP_SCHED */
7010 void dump_cpu_task(int cpu)
7012 pr_info("Task dump for CPU %d:\n", cpu);
7013 sched_show_task(cpu_curr(cpu));
7017 * Nice levels are multiplicative, with a gentle 10% change for every
7018 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7019 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7020 * that remained on nice 0.
7022 * The "10% effect" is relative and cumulative: from _any_ nice level,
7023 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7024 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7025 * If a task goes up by ~10% and another task goes down by ~10% then
7026 * the relative distance between them is ~25%.)
7028 const int sched_prio_to_weight[40] = {
7029 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7030 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7031 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7032 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7033 /* 0 */ 1024, 820, 655, 526, 423,
7034 /* 5 */ 335, 272, 215, 172, 137,
7035 /* 10 */ 110, 87, 70, 56, 45,
7036 /* 15 */ 36, 29, 23, 18, 15,
7040 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7042 * In cases where the weight does not change often, we can use the
7043 * precalculated inverse to speed up arithmetics by turning divisions
7044 * into multiplications:
7046 const u32 sched_prio_to_wmult[40] = {
7047 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7048 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7049 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7050 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7051 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7052 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7053 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7054 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7057 #undef CREATE_TRACE_POINTS