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;
707 * SCHED_OTHER tasks have to update their load when changing their
710 if (update_load && p->sched_class == &fair_sched_class) {
711 reweight_task(p, prio);
713 load->weight = scale_load(sched_prio_to_weight[prio]);
714 load->inv_weight = sched_prio_to_wmult[prio];
718 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
720 if (!(flags & ENQUEUE_NOCLOCK))
723 if (!(flags & ENQUEUE_RESTORE))
724 sched_info_queued(rq, p);
726 p->sched_class->enqueue_task(rq, p, flags);
729 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
731 if (!(flags & DEQUEUE_NOCLOCK))
734 if (!(flags & DEQUEUE_SAVE))
735 sched_info_dequeued(rq, p);
737 p->sched_class->dequeue_task(rq, p, flags);
740 void activate_task(struct rq *rq, struct task_struct *p, int flags)
742 if (task_contributes_to_load(p))
743 rq->nr_uninterruptible--;
745 enqueue_task(rq, p, flags);
748 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
750 if (task_contributes_to_load(p))
751 rq->nr_uninterruptible++;
753 dequeue_task(rq, p, flags);
757 * __normal_prio - return the priority that is based on the static prio
759 static inline int __normal_prio(struct task_struct *p)
761 return p->static_prio;
765 * Calculate the expected normal priority: i.e. priority
766 * without taking RT-inheritance into account. Might be
767 * boosted by interactivity modifiers. Changes upon fork,
768 * setprio syscalls, and whenever the interactivity
769 * estimator recalculates.
771 static inline int normal_prio(struct task_struct *p)
775 if (task_has_dl_policy(p))
776 prio = MAX_DL_PRIO-1;
777 else if (task_has_rt_policy(p))
778 prio = MAX_RT_PRIO-1 - p->rt_priority;
780 prio = __normal_prio(p);
785 * Calculate the current priority, i.e. the priority
786 * taken into account by the scheduler. This value might
787 * be boosted by RT tasks, or might be boosted by
788 * interactivity modifiers. Will be RT if the task got
789 * RT-boosted. If not then it returns p->normal_prio.
791 static int effective_prio(struct task_struct *p)
793 p->normal_prio = normal_prio(p);
795 * If we are RT tasks or we were boosted to RT priority,
796 * keep the priority unchanged. Otherwise, update priority
797 * to the normal priority:
799 if (!rt_prio(p->prio))
800 return p->normal_prio;
805 * task_curr - is this task currently executing on a CPU?
806 * @p: the task in question.
808 * Return: 1 if the task is currently executing. 0 otherwise.
810 inline int task_curr(const struct task_struct *p)
812 return cpu_curr(task_cpu(p)) == p;
816 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
817 * use the balance_callback list if you want balancing.
819 * this means any call to check_class_changed() must be followed by a call to
820 * balance_callback().
822 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
823 const struct sched_class *prev_class,
826 if (prev_class != p->sched_class) {
827 if (prev_class->switched_from)
828 prev_class->switched_from(rq, p);
830 p->sched_class->switched_to(rq, p);
831 } else if (oldprio != p->prio || dl_task(p))
832 p->sched_class->prio_changed(rq, p, oldprio);
835 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
837 const struct sched_class *class;
839 if (p->sched_class == rq->curr->sched_class) {
840 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
842 for_each_class(class) {
843 if (class == rq->curr->sched_class)
845 if (class == p->sched_class) {
853 * A queue event has occurred, and we're going to schedule. In
854 * this case, we can save a useless back to back clock update.
856 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
857 rq_clock_skip_update(rq);
862 static inline bool is_per_cpu_kthread(struct task_struct *p)
864 if (!(p->flags & PF_KTHREAD))
867 if (p->nr_cpus_allowed != 1)
874 * Per-CPU kthreads are allowed to run on !actie && online CPUs, see
875 * __set_cpus_allowed_ptr() and select_fallback_rq().
877 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
879 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
882 if (is_per_cpu_kthread(p))
883 return cpu_online(cpu);
885 return cpu_active(cpu);
889 * This is how migration works:
891 * 1) we invoke migration_cpu_stop() on the target CPU using
893 * 2) stopper starts to run (implicitly forcing the migrated thread
895 * 3) it checks whether the migrated task is still in the wrong runqueue.
896 * 4) if it's in the wrong runqueue then the migration thread removes
897 * it and puts it into the right queue.
898 * 5) stopper completes and stop_one_cpu() returns and the migration
903 * move_queued_task - move a queued task to new rq.
905 * Returns (locked) new rq. Old rq's lock is released.
907 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
908 struct task_struct *p, int new_cpu)
910 lockdep_assert_held(&rq->lock);
912 p->on_rq = TASK_ON_RQ_MIGRATING;
913 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
914 set_task_cpu(p, new_cpu);
917 rq = cpu_rq(new_cpu);
920 BUG_ON(task_cpu(p) != new_cpu);
921 enqueue_task(rq, p, 0);
922 p->on_rq = TASK_ON_RQ_QUEUED;
923 check_preempt_curr(rq, p, 0);
928 struct migration_arg {
929 struct task_struct *task;
934 * Move (not current) task off this CPU, onto the destination CPU. We're doing
935 * this because either it can't run here any more (set_cpus_allowed()
936 * away from this CPU, or CPU going down), or because we're
937 * attempting to rebalance this task on exec (sched_exec).
939 * So we race with normal scheduler movements, but that's OK, as long
940 * as the task is no longer on this CPU.
942 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
943 struct task_struct *p, int dest_cpu)
945 /* Affinity changed (again). */
946 if (!is_cpu_allowed(p, dest_cpu))
950 rq = move_queued_task(rq, rf, p, dest_cpu);
956 * migration_cpu_stop - this will be executed by a highprio stopper thread
957 * and performs thread migration by bumping thread off CPU then
958 * 'pushing' onto another runqueue.
960 static int migration_cpu_stop(void *data)
962 struct migration_arg *arg = data;
963 struct task_struct *p = arg->task;
964 struct rq *rq = this_rq();
968 * The original target CPU might have gone down and we might
969 * be on another CPU but it doesn't matter.
973 * We need to explicitly wake pending tasks before running
974 * __migrate_task() such that we will not miss enforcing cpus_allowed
975 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
977 sched_ttwu_pending();
979 raw_spin_lock(&p->pi_lock);
982 * If task_rq(p) != rq, it cannot be migrated here, because we're
983 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
984 * we're holding p->pi_lock.
986 if (task_rq(p) == rq) {
987 if (task_on_rq_queued(p))
988 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
990 p->wake_cpu = arg->dest_cpu;
993 raw_spin_unlock(&p->pi_lock);
1000 * sched_class::set_cpus_allowed must do the below, but is not required to
1001 * actually call this function.
1003 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1005 cpumask_copy(&p->cpus_allowed, new_mask);
1006 p->nr_cpus_allowed = cpumask_weight(new_mask);
1009 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1011 struct rq *rq = task_rq(p);
1012 bool queued, running;
1014 lockdep_assert_held(&p->pi_lock);
1016 queued = task_on_rq_queued(p);
1017 running = task_current(rq, p);
1021 * Because __kthread_bind() calls this on blocked tasks without
1024 lockdep_assert_held(&rq->lock);
1025 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1028 put_prev_task(rq, p);
1030 p->sched_class->set_cpus_allowed(p, new_mask);
1033 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1035 set_curr_task(rq, p);
1039 * Change a given task's CPU affinity. Migrate the thread to a
1040 * proper CPU and schedule it away if the CPU it's executing on
1041 * is removed from the allowed bitmask.
1043 * NOTE: the caller must have a valid reference to the task, the
1044 * task must not exit() & deallocate itself prematurely. The
1045 * call is not atomic; no spinlocks may be held.
1047 static int __set_cpus_allowed_ptr(struct task_struct *p,
1048 const struct cpumask *new_mask, bool check)
1050 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1051 unsigned int dest_cpu;
1056 rq = task_rq_lock(p, &rf);
1057 update_rq_clock(rq);
1059 if (p->flags & PF_KTHREAD) {
1061 * Kernel threads are allowed on online && !active CPUs
1063 cpu_valid_mask = cpu_online_mask;
1067 * Must re-check here, to close a race against __kthread_bind(),
1068 * sched_setaffinity() is not guaranteed to observe the flag.
1070 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1075 if (cpumask_equal(&p->cpus_allowed, new_mask))
1078 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1083 do_set_cpus_allowed(p, new_mask);
1085 if (p->flags & PF_KTHREAD) {
1087 * For kernel threads that do indeed end up on online &&
1088 * !active we want to ensure they are strict per-CPU threads.
1090 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1091 !cpumask_intersects(new_mask, cpu_active_mask) &&
1092 p->nr_cpus_allowed != 1);
1095 /* Can the task run on the task's current CPU? If so, we're done */
1096 if (cpumask_test_cpu(task_cpu(p), new_mask))
1099 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1100 if (task_running(rq, p) || p->state == TASK_WAKING) {
1101 struct migration_arg arg = { p, dest_cpu };
1102 /* Need help from migration thread: drop lock and wait. */
1103 task_rq_unlock(rq, p, &rf);
1104 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1105 tlb_migrate_finish(p->mm);
1107 } else if (task_on_rq_queued(p)) {
1109 * OK, since we're going to drop the lock immediately
1110 * afterwards anyway.
1112 rq = move_queued_task(rq, &rf, p, dest_cpu);
1115 task_rq_unlock(rq, p, &rf);
1120 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1122 return __set_cpus_allowed_ptr(p, new_mask, false);
1124 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1126 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1128 #ifdef CONFIG_SCHED_DEBUG
1130 * We should never call set_task_cpu() on a blocked task,
1131 * ttwu() will sort out the placement.
1133 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1137 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1138 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1139 * time relying on p->on_rq.
1141 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1142 p->sched_class == &fair_sched_class &&
1143 (p->on_rq && !task_on_rq_migrating(p)));
1145 #ifdef CONFIG_LOCKDEP
1147 * The caller should hold either p->pi_lock or rq->lock, when changing
1148 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1150 * sched_move_task() holds both and thus holding either pins the cgroup,
1153 * Furthermore, all task_rq users should acquire both locks, see
1156 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1157 lockdep_is_held(&task_rq(p)->lock)));
1160 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1162 WARN_ON_ONCE(!cpu_online(new_cpu));
1165 trace_sched_migrate_task(p, new_cpu);
1167 if (task_cpu(p) != new_cpu) {
1168 if (p->sched_class->migrate_task_rq)
1169 p->sched_class->migrate_task_rq(p, new_cpu);
1170 p->se.nr_migrations++;
1172 perf_event_task_migrate(p);
1175 __set_task_cpu(p, new_cpu);
1178 #ifdef CONFIG_NUMA_BALANCING
1179 static void __migrate_swap_task(struct task_struct *p, int cpu)
1181 if (task_on_rq_queued(p)) {
1182 struct rq *src_rq, *dst_rq;
1183 struct rq_flags srf, drf;
1185 src_rq = task_rq(p);
1186 dst_rq = cpu_rq(cpu);
1188 rq_pin_lock(src_rq, &srf);
1189 rq_pin_lock(dst_rq, &drf);
1191 p->on_rq = TASK_ON_RQ_MIGRATING;
1192 deactivate_task(src_rq, p, 0);
1193 set_task_cpu(p, cpu);
1194 activate_task(dst_rq, p, 0);
1195 p->on_rq = TASK_ON_RQ_QUEUED;
1196 check_preempt_curr(dst_rq, p, 0);
1198 rq_unpin_lock(dst_rq, &drf);
1199 rq_unpin_lock(src_rq, &srf);
1203 * Task isn't running anymore; make it appear like we migrated
1204 * it before it went to sleep. This means on wakeup we make the
1205 * previous CPU our target instead of where it really is.
1211 struct migration_swap_arg {
1212 struct task_struct *src_task, *dst_task;
1213 int src_cpu, dst_cpu;
1216 static int migrate_swap_stop(void *data)
1218 struct migration_swap_arg *arg = data;
1219 struct rq *src_rq, *dst_rq;
1222 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1225 src_rq = cpu_rq(arg->src_cpu);
1226 dst_rq = cpu_rq(arg->dst_cpu);
1228 double_raw_lock(&arg->src_task->pi_lock,
1229 &arg->dst_task->pi_lock);
1230 double_rq_lock(src_rq, dst_rq);
1232 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1235 if (task_cpu(arg->src_task) != arg->src_cpu)
1238 if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1241 if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1244 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1245 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1250 double_rq_unlock(src_rq, dst_rq);
1251 raw_spin_unlock(&arg->dst_task->pi_lock);
1252 raw_spin_unlock(&arg->src_task->pi_lock);
1258 * Cross migrate two tasks
1260 int migrate_swap(struct task_struct *cur, struct task_struct *p,
1261 int target_cpu, int curr_cpu)
1263 struct migration_swap_arg arg;
1266 arg = (struct migration_swap_arg){
1268 .src_cpu = curr_cpu,
1270 .dst_cpu = target_cpu,
1273 if (arg.src_cpu == arg.dst_cpu)
1277 * These three tests are all lockless; this is OK since all of them
1278 * will be re-checked with proper locks held further down the line.
1280 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1283 if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1286 if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1289 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1290 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1295 #endif /* CONFIG_NUMA_BALANCING */
1298 * wait_task_inactive - wait for a thread to unschedule.
1300 * If @match_state is nonzero, it's the @p->state value just checked and
1301 * not expected to change. If it changes, i.e. @p might have woken up,
1302 * then return zero. When we succeed in waiting for @p to be off its CPU,
1303 * we return a positive number (its total switch count). If a second call
1304 * a short while later returns the same number, the caller can be sure that
1305 * @p has remained unscheduled the whole time.
1307 * The caller must ensure that the task *will* unschedule sometime soon,
1308 * else this function might spin for a *long* time. This function can't
1309 * be called with interrupts off, or it may introduce deadlock with
1310 * smp_call_function() if an IPI is sent by the same process we are
1311 * waiting to become inactive.
1313 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1315 int running, queued;
1322 * We do the initial early heuristics without holding
1323 * any task-queue locks at all. We'll only try to get
1324 * the runqueue lock when things look like they will
1330 * If the task is actively running on another CPU
1331 * still, just relax and busy-wait without holding
1334 * NOTE! Since we don't hold any locks, it's not
1335 * even sure that "rq" stays as the right runqueue!
1336 * But we don't care, since "task_running()" will
1337 * return false if the runqueue has changed and p
1338 * is actually now running somewhere else!
1340 while (task_running(rq, p)) {
1341 if (match_state && unlikely(p->state != match_state))
1347 * Ok, time to look more closely! We need the rq
1348 * lock now, to be *sure*. If we're wrong, we'll
1349 * just go back and repeat.
1351 rq = task_rq_lock(p, &rf);
1352 trace_sched_wait_task(p);
1353 running = task_running(rq, p);
1354 queued = task_on_rq_queued(p);
1356 if (!match_state || p->state == match_state)
1357 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1358 task_rq_unlock(rq, p, &rf);
1361 * If it changed from the expected state, bail out now.
1363 if (unlikely(!ncsw))
1367 * Was it really running after all now that we
1368 * checked with the proper locks actually held?
1370 * Oops. Go back and try again..
1372 if (unlikely(running)) {
1378 * It's not enough that it's not actively running,
1379 * it must be off the runqueue _entirely_, and not
1382 * So if it was still runnable (but just not actively
1383 * running right now), it's preempted, and we should
1384 * yield - it could be a while.
1386 if (unlikely(queued)) {
1387 ktime_t to = NSEC_PER_SEC / HZ;
1389 set_current_state(TASK_UNINTERRUPTIBLE);
1390 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1395 * Ahh, all good. It wasn't running, and it wasn't
1396 * runnable, which means that it will never become
1397 * running in the future either. We're all done!
1406 * kick_process - kick a running thread to enter/exit the kernel
1407 * @p: the to-be-kicked thread
1409 * Cause a process which is running on another CPU to enter
1410 * kernel-mode, without any delay. (to get signals handled.)
1412 * NOTE: this function doesn't have to take the runqueue lock,
1413 * because all it wants to ensure is that the remote task enters
1414 * the kernel. If the IPI races and the task has been migrated
1415 * to another CPU then no harm is done and the purpose has been
1418 void kick_process(struct task_struct *p)
1424 if ((cpu != smp_processor_id()) && task_curr(p))
1425 smp_send_reschedule(cpu);
1428 EXPORT_SYMBOL_GPL(kick_process);
1431 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1433 * A few notes on cpu_active vs cpu_online:
1435 * - cpu_active must be a subset of cpu_online
1437 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
1438 * see __set_cpus_allowed_ptr(). At this point the newly online
1439 * CPU isn't yet part of the sched domains, and balancing will not
1442 * - on CPU-down we clear cpu_active() to mask the sched domains and
1443 * avoid the load balancer to place new tasks on the to be removed
1444 * CPU. Existing tasks will remain running there and will be taken
1447 * This means that fallback selection must not select !active CPUs.
1448 * And can assume that any active CPU must be online. Conversely
1449 * select_task_rq() below may allow selection of !active CPUs in order
1450 * to satisfy the above rules.
1452 static int select_fallback_rq(int cpu, struct task_struct *p)
1454 int nid = cpu_to_node(cpu);
1455 const struct cpumask *nodemask = NULL;
1456 enum { cpuset, possible, fail } state = cpuset;
1460 * If the node that the CPU is on has been offlined, cpu_to_node()
1461 * will return -1. There is no CPU on the node, and we should
1462 * select the CPU on the other node.
1465 nodemask = cpumask_of_node(nid);
1467 /* Look for allowed, online CPU in same node. */
1468 for_each_cpu(dest_cpu, nodemask) {
1469 if (!cpu_active(dest_cpu))
1471 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1477 /* Any allowed, online CPU? */
1478 for_each_cpu(dest_cpu, &p->cpus_allowed) {
1479 if (!is_cpu_allowed(p, dest_cpu))
1485 /* No more Mr. Nice Guy. */
1488 if (IS_ENABLED(CONFIG_CPUSETS)) {
1489 cpuset_cpus_allowed_fallback(p);
1495 do_set_cpus_allowed(p, cpu_possible_mask);
1506 if (state != cpuset) {
1508 * Don't tell them about moving exiting tasks or
1509 * kernel threads (both mm NULL), since they never
1512 if (p->mm && printk_ratelimit()) {
1513 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1514 task_pid_nr(p), p->comm, cpu);
1522 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1525 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1527 lockdep_assert_held(&p->pi_lock);
1529 if (p->nr_cpus_allowed > 1)
1530 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1532 cpu = cpumask_any(&p->cpus_allowed);
1535 * In order not to call set_task_cpu() on a blocking task we need
1536 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1539 * Since this is common to all placement strategies, this lives here.
1541 * [ this allows ->select_task() to simply return task_cpu(p) and
1542 * not worry about this generic constraint ]
1544 if (unlikely(!is_cpu_allowed(p, cpu)))
1545 cpu = select_fallback_rq(task_cpu(p), p);
1550 static void update_avg(u64 *avg, u64 sample)
1552 s64 diff = sample - *avg;
1556 void sched_set_stop_task(int cpu, struct task_struct *stop)
1558 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1559 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1563 * Make it appear like a SCHED_FIFO task, its something
1564 * userspace knows about and won't get confused about.
1566 * Also, it will make PI more or less work without too
1567 * much confusion -- but then, stop work should not
1568 * rely on PI working anyway.
1570 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1572 stop->sched_class = &stop_sched_class;
1575 cpu_rq(cpu)->stop = stop;
1579 * Reset it back to a normal scheduling class so that
1580 * it can die in pieces.
1582 old_stop->sched_class = &rt_sched_class;
1588 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1589 const struct cpumask *new_mask, bool check)
1591 return set_cpus_allowed_ptr(p, new_mask);
1594 #endif /* CONFIG_SMP */
1597 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1601 if (!schedstat_enabled())
1607 if (cpu == rq->cpu) {
1608 __schedstat_inc(rq->ttwu_local);
1609 __schedstat_inc(p->se.statistics.nr_wakeups_local);
1611 struct sched_domain *sd;
1613 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
1615 for_each_domain(rq->cpu, sd) {
1616 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1617 __schedstat_inc(sd->ttwu_wake_remote);
1624 if (wake_flags & WF_MIGRATED)
1625 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1626 #endif /* CONFIG_SMP */
1628 __schedstat_inc(rq->ttwu_count);
1629 __schedstat_inc(p->se.statistics.nr_wakeups);
1631 if (wake_flags & WF_SYNC)
1632 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
1635 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1637 activate_task(rq, p, en_flags);
1638 p->on_rq = TASK_ON_RQ_QUEUED;
1640 /* If a worker is waking up, notify the workqueue: */
1641 if (p->flags & PF_WQ_WORKER)
1642 wq_worker_waking_up(p, cpu_of(rq));
1646 * Mark the task runnable and perform wakeup-preemption.
1648 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1649 struct rq_flags *rf)
1651 check_preempt_curr(rq, p, wake_flags);
1652 p->state = TASK_RUNNING;
1653 trace_sched_wakeup(p);
1656 if (p->sched_class->task_woken) {
1658 * Our task @p is fully woken up and running; so its safe to
1659 * drop the rq->lock, hereafter rq is only used for statistics.
1661 rq_unpin_lock(rq, rf);
1662 p->sched_class->task_woken(rq, p);
1663 rq_repin_lock(rq, rf);
1666 if (rq->idle_stamp) {
1667 u64 delta = rq_clock(rq) - rq->idle_stamp;
1668 u64 max = 2*rq->max_idle_balance_cost;
1670 update_avg(&rq->avg_idle, delta);
1672 if (rq->avg_idle > max)
1681 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1682 struct rq_flags *rf)
1684 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
1686 lockdep_assert_held(&rq->lock);
1689 if (p->sched_contributes_to_load)
1690 rq->nr_uninterruptible--;
1692 if (wake_flags & WF_MIGRATED)
1693 en_flags |= ENQUEUE_MIGRATED;
1696 ttwu_activate(rq, p, en_flags);
1697 ttwu_do_wakeup(rq, p, wake_flags, rf);
1701 * Called in case the task @p isn't fully descheduled from its runqueue,
1702 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1703 * since all we need to do is flip p->state to TASK_RUNNING, since
1704 * the task is still ->on_rq.
1706 static int ttwu_remote(struct task_struct *p, int wake_flags)
1712 rq = __task_rq_lock(p, &rf);
1713 if (task_on_rq_queued(p)) {
1714 /* check_preempt_curr() may use rq clock */
1715 update_rq_clock(rq);
1716 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1719 __task_rq_unlock(rq, &rf);
1725 void sched_ttwu_pending(void)
1727 struct rq *rq = this_rq();
1728 struct llist_node *llist = llist_del_all(&rq->wake_list);
1729 struct task_struct *p, *t;
1735 rq_lock_irqsave(rq, &rf);
1736 update_rq_clock(rq);
1738 llist_for_each_entry_safe(p, t, llist, wake_entry)
1739 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
1741 rq_unlock_irqrestore(rq, &rf);
1744 void scheduler_ipi(void)
1747 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1748 * TIF_NEED_RESCHED remotely (for the first time) will also send
1751 preempt_fold_need_resched();
1753 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1757 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1758 * traditionally all their work was done from the interrupt return
1759 * path. Now that we actually do some work, we need to make sure
1762 * Some archs already do call them, luckily irq_enter/exit nest
1765 * Arguably we should visit all archs and update all handlers,
1766 * however a fair share of IPIs are still resched only so this would
1767 * somewhat pessimize the simple resched case.
1770 sched_ttwu_pending();
1773 * Check if someone kicked us for doing the nohz idle load balance.
1775 if (unlikely(got_nohz_idle_kick())) {
1776 this_rq()->idle_balance = 1;
1777 raise_softirq_irqoff(SCHED_SOFTIRQ);
1782 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1784 struct rq *rq = cpu_rq(cpu);
1786 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1788 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1789 if (!set_nr_if_polling(rq->idle))
1790 smp_send_reschedule(cpu);
1792 trace_sched_wake_idle_without_ipi(cpu);
1796 void wake_up_if_idle(int cpu)
1798 struct rq *rq = cpu_rq(cpu);
1803 if (!is_idle_task(rcu_dereference(rq->curr)))
1806 if (set_nr_if_polling(rq->idle)) {
1807 trace_sched_wake_idle_without_ipi(cpu);
1809 rq_lock_irqsave(rq, &rf);
1810 if (is_idle_task(rq->curr))
1811 smp_send_reschedule(cpu);
1812 /* Else CPU is not idle, do nothing here: */
1813 rq_unlock_irqrestore(rq, &rf);
1820 bool cpus_share_cache(int this_cpu, int that_cpu)
1822 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1824 #endif /* CONFIG_SMP */
1826 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1828 struct rq *rq = cpu_rq(cpu);
1831 #if defined(CONFIG_SMP)
1832 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1833 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1834 ttwu_queue_remote(p, cpu, wake_flags);
1840 update_rq_clock(rq);
1841 ttwu_do_activate(rq, p, wake_flags, &rf);
1846 * Notes on Program-Order guarantees on SMP systems.
1850 * The basic program-order guarantee on SMP systems is that when a task [t]
1851 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1852 * execution on its new CPU [c1].
1854 * For migration (of runnable tasks) this is provided by the following means:
1856 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1857 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1858 * rq(c1)->lock (if not at the same time, then in that order).
1859 * C) LOCK of the rq(c1)->lock scheduling in task
1861 * Release/acquire chaining guarantees that B happens after A and C after B.
1862 * Note: the CPU doing B need not be c0 or c1
1871 * UNLOCK rq(0)->lock
1873 * LOCK rq(0)->lock // orders against CPU0
1875 * UNLOCK rq(0)->lock
1879 * UNLOCK rq(1)->lock
1881 * LOCK rq(1)->lock // orders against CPU2
1884 * UNLOCK rq(1)->lock
1887 * BLOCKING -- aka. SLEEP + WAKEUP
1889 * For blocking we (obviously) need to provide the same guarantee as for
1890 * migration. However the means are completely different as there is no lock
1891 * chain to provide order. Instead we do:
1893 * 1) smp_store_release(X->on_cpu, 0)
1894 * 2) smp_cond_load_acquire(!X->on_cpu)
1898 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1900 * LOCK rq(0)->lock LOCK X->pi_lock
1903 * smp_store_release(X->on_cpu, 0);
1905 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1911 * X->state = RUNNING
1912 * UNLOCK rq(2)->lock
1914 * LOCK rq(2)->lock // orders against CPU1
1917 * UNLOCK rq(2)->lock
1920 * UNLOCK rq(0)->lock
1923 * However, for wakeups there is a second guarantee we must provide, namely we
1924 * must ensure that CONDITION=1 done by the caller can not be reordered with
1925 * accesses to the task state; see try_to_wake_up() and set_current_state().
1929 * try_to_wake_up - wake up a thread
1930 * @p: the thread to be awakened
1931 * @state: the mask of task states that can be woken
1932 * @wake_flags: wake modifier flags (WF_*)
1934 * If (@state & @p->state) @p->state = TASK_RUNNING.
1936 * If the task was not queued/runnable, also place it back on a runqueue.
1938 * Atomic against schedule() which would dequeue a task, also see
1939 * set_current_state().
1941 * This function executes a full memory barrier before accessing the task
1942 * state; see set_current_state().
1944 * Return: %true if @p->state changes (an actual wakeup was done),
1948 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1950 unsigned long flags;
1951 int cpu, success = 0;
1954 * If we are going to wake up a thread waiting for CONDITION we
1955 * need to ensure that CONDITION=1 done by the caller can not be
1956 * reordered with p->state check below. This pairs with mb() in
1957 * set_current_state() the waiting thread does.
1959 raw_spin_lock_irqsave(&p->pi_lock, flags);
1960 smp_mb__after_spinlock();
1961 if (!(p->state & state))
1964 trace_sched_waking(p);
1966 /* We're going to change ->state: */
1971 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1972 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1973 * in smp_cond_load_acquire() below.
1975 * sched_ttwu_pending() try_to_wake_up()
1976 * STORE p->on_rq = 1 LOAD p->state
1979 * __schedule() (switch to task 'p')
1980 * LOCK rq->lock smp_rmb();
1981 * smp_mb__after_spinlock();
1985 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
1987 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
1988 * __schedule(). See the comment for smp_mb__after_spinlock().
1991 if (p->on_rq && ttwu_remote(p, wake_flags))
1996 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1997 * possible to, falsely, observe p->on_cpu == 0.
1999 * One must be running (->on_cpu == 1) in order to remove oneself
2000 * from the runqueue.
2002 * __schedule() (switch to task 'p') try_to_wake_up()
2003 * STORE p->on_cpu = 1 LOAD p->on_rq
2006 * __schedule() (put 'p' to sleep)
2007 * LOCK rq->lock smp_rmb();
2008 * smp_mb__after_spinlock();
2009 * STORE p->on_rq = 0 LOAD p->on_cpu
2011 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2012 * __schedule(). See the comment for smp_mb__after_spinlock().
2017 * If the owning (remote) CPU is still in the middle of schedule() with
2018 * this task as prev, wait until its done referencing the task.
2020 * Pairs with the smp_store_release() in finish_task().
2022 * This ensures that tasks getting woken will be fully ordered against
2023 * their previous state and preserve Program Order.
2025 smp_cond_load_acquire(&p->on_cpu, !VAL);
2027 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2028 p->state = TASK_WAKING;
2031 delayacct_blkio_end(p);
2032 atomic_dec(&task_rq(p)->nr_iowait);
2035 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2036 if (task_cpu(p) != cpu) {
2037 wake_flags |= WF_MIGRATED;
2038 set_task_cpu(p, cpu);
2041 #else /* CONFIG_SMP */
2044 delayacct_blkio_end(p);
2045 atomic_dec(&task_rq(p)->nr_iowait);
2048 #endif /* CONFIG_SMP */
2050 ttwu_queue(p, cpu, wake_flags);
2052 ttwu_stat(p, cpu, wake_flags);
2054 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2060 * try_to_wake_up_local - try to wake up a local task with rq lock held
2061 * @p: the thread to be awakened
2062 * @rf: request-queue flags for pinning
2064 * Put @p on the run-queue if it's not already there. The caller must
2065 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2068 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2070 struct rq *rq = task_rq(p);
2072 if (WARN_ON_ONCE(rq != this_rq()) ||
2073 WARN_ON_ONCE(p == current))
2076 lockdep_assert_held(&rq->lock);
2078 if (!raw_spin_trylock(&p->pi_lock)) {
2080 * This is OK, because current is on_cpu, which avoids it being
2081 * picked for load-balance and preemption/IRQs are still
2082 * disabled avoiding further scheduler activity on it and we've
2083 * not yet picked a replacement task.
2086 raw_spin_lock(&p->pi_lock);
2090 if (!(p->state & TASK_NORMAL))
2093 trace_sched_waking(p);
2095 if (!task_on_rq_queued(p)) {
2097 delayacct_blkio_end(p);
2098 atomic_dec(&rq->nr_iowait);
2100 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK);
2103 ttwu_do_wakeup(rq, p, 0, rf);
2104 ttwu_stat(p, smp_processor_id(), 0);
2106 raw_spin_unlock(&p->pi_lock);
2110 * wake_up_process - Wake up a specific process
2111 * @p: The process to be woken up.
2113 * Attempt to wake up the nominated process and move it to the set of runnable
2116 * Return: 1 if the process was woken up, 0 if it was already running.
2118 * This function executes a full memory barrier before accessing the task state.
2120 int wake_up_process(struct task_struct *p)
2122 return try_to_wake_up(p, TASK_NORMAL, 0);
2124 EXPORT_SYMBOL(wake_up_process);
2126 int wake_up_state(struct task_struct *p, unsigned int state)
2128 return try_to_wake_up(p, state, 0);
2132 * Perform scheduler related setup for a newly forked process p.
2133 * p is forked by current.
2135 * __sched_fork() is basic setup used by init_idle() too:
2137 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2142 p->se.exec_start = 0;
2143 p->se.sum_exec_runtime = 0;
2144 p->se.prev_sum_exec_runtime = 0;
2145 p->se.nr_migrations = 0;
2147 INIT_LIST_HEAD(&p->se.group_node);
2149 #ifdef CONFIG_FAIR_GROUP_SCHED
2150 p->se.cfs_rq = NULL;
2153 #ifdef CONFIG_SCHEDSTATS
2154 /* Even if schedstat is disabled, there should not be garbage */
2155 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2158 RB_CLEAR_NODE(&p->dl.rb_node);
2159 init_dl_task_timer(&p->dl);
2160 init_dl_inactive_task_timer(&p->dl);
2161 __dl_clear_params(p);
2163 INIT_LIST_HEAD(&p->rt.run_list);
2165 p->rt.time_slice = sched_rr_timeslice;
2169 #ifdef CONFIG_PREEMPT_NOTIFIERS
2170 INIT_HLIST_HEAD(&p->preempt_notifiers);
2173 init_numa_balancing(clone_flags, p);
2176 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2178 #ifdef CONFIG_NUMA_BALANCING
2180 void set_numabalancing_state(bool enabled)
2183 static_branch_enable(&sched_numa_balancing);
2185 static_branch_disable(&sched_numa_balancing);
2188 #ifdef CONFIG_PROC_SYSCTL
2189 int sysctl_numa_balancing(struct ctl_table *table, int write,
2190 void __user *buffer, size_t *lenp, loff_t *ppos)
2194 int state = static_branch_likely(&sched_numa_balancing);
2196 if (write && !capable(CAP_SYS_ADMIN))
2201 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2205 set_numabalancing_state(state);
2211 #ifdef CONFIG_SCHEDSTATS
2213 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2214 static bool __initdata __sched_schedstats = false;
2216 static void set_schedstats(bool enabled)
2219 static_branch_enable(&sched_schedstats);
2221 static_branch_disable(&sched_schedstats);
2224 void force_schedstat_enabled(void)
2226 if (!schedstat_enabled()) {
2227 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2228 static_branch_enable(&sched_schedstats);
2232 static int __init setup_schedstats(char *str)
2239 * This code is called before jump labels have been set up, so we can't
2240 * change the static branch directly just yet. Instead set a temporary
2241 * variable so init_schedstats() can do it later.
2243 if (!strcmp(str, "enable")) {
2244 __sched_schedstats = true;
2246 } else if (!strcmp(str, "disable")) {
2247 __sched_schedstats = false;
2252 pr_warn("Unable to parse schedstats=\n");
2256 __setup("schedstats=", setup_schedstats);
2258 static void __init init_schedstats(void)
2260 set_schedstats(__sched_schedstats);
2263 #ifdef CONFIG_PROC_SYSCTL
2264 int sysctl_schedstats(struct ctl_table *table, int write,
2265 void __user *buffer, size_t *lenp, loff_t *ppos)
2269 int state = static_branch_likely(&sched_schedstats);
2271 if (write && !capable(CAP_SYS_ADMIN))
2276 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2280 set_schedstats(state);
2283 #endif /* CONFIG_PROC_SYSCTL */
2284 #else /* !CONFIG_SCHEDSTATS */
2285 static inline void init_schedstats(void) {}
2286 #endif /* CONFIG_SCHEDSTATS */
2289 * fork()/clone()-time setup:
2291 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2293 unsigned long flags;
2295 __sched_fork(clone_flags, p);
2297 * We mark the process as NEW here. This guarantees that
2298 * nobody will actually run it, and a signal or other external
2299 * event cannot wake it up and insert it on the runqueue either.
2301 p->state = TASK_NEW;
2304 * Make sure we do not leak PI boosting priority to the child.
2306 p->prio = current->normal_prio;
2309 * Revert to default priority/policy on fork if requested.
2311 if (unlikely(p->sched_reset_on_fork)) {
2312 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2313 p->policy = SCHED_NORMAL;
2314 p->static_prio = NICE_TO_PRIO(0);
2316 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2317 p->static_prio = NICE_TO_PRIO(0);
2319 p->prio = p->normal_prio = __normal_prio(p);
2320 set_load_weight(p, false);
2323 * We don't need the reset flag anymore after the fork. It has
2324 * fulfilled its duty:
2326 p->sched_reset_on_fork = 0;
2329 if (dl_prio(p->prio))
2331 else if (rt_prio(p->prio))
2332 p->sched_class = &rt_sched_class;
2334 p->sched_class = &fair_sched_class;
2336 init_entity_runnable_average(&p->se);
2339 * The child is not yet in the pid-hash so no cgroup attach races,
2340 * and the cgroup is pinned to this child due to cgroup_fork()
2341 * is ran before sched_fork().
2343 * Silence PROVE_RCU.
2345 raw_spin_lock_irqsave(&p->pi_lock, flags);
2347 * We're setting the CPU for the first time, we don't migrate,
2348 * so use __set_task_cpu().
2350 __set_task_cpu(p, smp_processor_id());
2351 if (p->sched_class->task_fork)
2352 p->sched_class->task_fork(p);
2353 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2355 #ifdef CONFIG_SCHED_INFO
2356 if (likely(sched_info_on()))
2357 memset(&p->sched_info, 0, sizeof(p->sched_info));
2359 #if defined(CONFIG_SMP)
2362 init_task_preempt_count(p);
2364 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2365 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2370 unsigned long to_ratio(u64 period, u64 runtime)
2372 if (runtime == RUNTIME_INF)
2376 * Doing this here saves a lot of checks in all
2377 * the calling paths, and returning zero seems
2378 * safe for them anyway.
2383 return div64_u64(runtime << BW_SHIFT, period);
2387 * wake_up_new_task - wake up a newly created task for the first time.
2389 * This function will do some initial scheduler statistics housekeeping
2390 * that must be done for every newly created context, then puts the task
2391 * on the runqueue and wakes it.
2393 void wake_up_new_task(struct task_struct *p)
2398 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2399 p->state = TASK_RUNNING;
2402 * Fork balancing, do it here and not earlier because:
2403 * - cpus_allowed can change in the fork path
2404 * - any previously selected CPU might disappear through hotplug
2406 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2407 * as we're not fully set-up yet.
2409 p->recent_used_cpu = task_cpu(p);
2410 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2412 rq = __task_rq_lock(p, &rf);
2413 update_rq_clock(rq);
2414 post_init_entity_util_avg(&p->se);
2416 activate_task(rq, p, ENQUEUE_NOCLOCK);
2417 p->on_rq = TASK_ON_RQ_QUEUED;
2418 trace_sched_wakeup_new(p);
2419 check_preempt_curr(rq, p, WF_FORK);
2421 if (p->sched_class->task_woken) {
2423 * Nothing relies on rq->lock after this, so its fine to
2426 rq_unpin_lock(rq, &rf);
2427 p->sched_class->task_woken(rq, p);
2428 rq_repin_lock(rq, &rf);
2431 task_rq_unlock(rq, p, &rf);
2434 #ifdef CONFIG_PREEMPT_NOTIFIERS
2436 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2438 void preempt_notifier_inc(void)
2440 static_branch_inc(&preempt_notifier_key);
2442 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2444 void preempt_notifier_dec(void)
2446 static_branch_dec(&preempt_notifier_key);
2448 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2451 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2452 * @notifier: notifier struct to register
2454 void preempt_notifier_register(struct preempt_notifier *notifier)
2456 if (!static_branch_unlikely(&preempt_notifier_key))
2457 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2459 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2461 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2464 * preempt_notifier_unregister - no longer interested in preemption notifications
2465 * @notifier: notifier struct to unregister
2467 * This is *not* safe to call from within a preemption notifier.
2469 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2471 hlist_del(¬ifier->link);
2473 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2475 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2477 struct preempt_notifier *notifier;
2479 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2480 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2483 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2485 if (static_branch_unlikely(&preempt_notifier_key))
2486 __fire_sched_in_preempt_notifiers(curr);
2490 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2491 struct task_struct *next)
2493 struct preempt_notifier *notifier;
2495 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2496 notifier->ops->sched_out(notifier, next);
2499 static __always_inline void
2500 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2501 struct task_struct *next)
2503 if (static_branch_unlikely(&preempt_notifier_key))
2504 __fire_sched_out_preempt_notifiers(curr, next);
2507 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2509 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2514 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2515 struct task_struct *next)
2519 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2521 static inline void prepare_task(struct task_struct *next)
2525 * Claim the task as running, we do this before switching to it
2526 * such that any running task will have this set.
2532 static inline void finish_task(struct task_struct *prev)
2536 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2537 * We must ensure this doesn't happen until the switch is completely
2540 * In particular, the load of prev->state in finish_task_switch() must
2541 * happen before this.
2543 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2545 smp_store_release(&prev->on_cpu, 0);
2550 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
2553 * Since the runqueue lock will be released by the next
2554 * task (which is an invalid locking op but in the case
2555 * of the scheduler it's an obvious special-case), so we
2556 * do an early lockdep release here:
2558 rq_unpin_lock(rq, rf);
2559 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2560 #ifdef CONFIG_DEBUG_SPINLOCK
2561 /* this is a valid case when another task releases the spinlock */
2562 rq->lock.owner = next;
2566 static inline void finish_lock_switch(struct rq *rq)
2569 * If we are tracking spinlock dependencies then we have to
2570 * fix up the runqueue lock - which gets 'carried over' from
2571 * prev into current:
2573 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
2574 raw_spin_unlock_irq(&rq->lock);
2578 * NOP if the arch has not defined these:
2581 #ifndef prepare_arch_switch
2582 # define prepare_arch_switch(next) do { } while (0)
2585 #ifndef finish_arch_post_lock_switch
2586 # define finish_arch_post_lock_switch() do { } while (0)
2590 * prepare_task_switch - prepare to switch tasks
2591 * @rq: the runqueue preparing to switch
2592 * @prev: the current task that is being switched out
2593 * @next: the task we are going to switch to.
2595 * This is called with the rq lock held and interrupts off. It must
2596 * be paired with a subsequent finish_task_switch after the context
2599 * prepare_task_switch sets up locking and calls architecture specific
2603 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2604 struct task_struct *next)
2606 kcov_prepare_switch(prev);
2607 sched_info_switch(rq, prev, next);
2608 perf_event_task_sched_out(prev, next);
2610 fire_sched_out_preempt_notifiers(prev, next);
2612 prepare_arch_switch(next);
2616 * finish_task_switch - clean up after a task-switch
2617 * @prev: the thread we just switched away from.
2619 * finish_task_switch must be called after the context switch, paired
2620 * with a prepare_task_switch call before the context switch.
2621 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2622 * and do any other architecture-specific cleanup actions.
2624 * Note that we may have delayed dropping an mm in context_switch(). If
2625 * so, we finish that here outside of the runqueue lock. (Doing it
2626 * with the lock held can cause deadlocks; see schedule() for
2629 * The context switch have flipped the stack from under us and restored the
2630 * local variables which were saved when this task called schedule() in the
2631 * past. prev == current is still correct but we need to recalculate this_rq
2632 * because prev may have moved to another CPU.
2634 static struct rq *finish_task_switch(struct task_struct *prev)
2635 __releases(rq->lock)
2637 struct rq *rq = this_rq();
2638 struct mm_struct *mm = rq->prev_mm;
2642 * The previous task will have left us with a preempt_count of 2
2643 * because it left us after:
2646 * preempt_disable(); // 1
2648 * raw_spin_lock_irq(&rq->lock) // 2
2650 * Also, see FORK_PREEMPT_COUNT.
2652 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2653 "corrupted preempt_count: %s/%d/0x%x\n",
2654 current->comm, current->pid, preempt_count()))
2655 preempt_count_set(FORK_PREEMPT_COUNT);
2660 * A task struct has one reference for the use as "current".
2661 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2662 * schedule one last time. The schedule call will never return, and
2663 * the scheduled task must drop that reference.
2665 * We must observe prev->state before clearing prev->on_cpu (in
2666 * finish_task), otherwise a concurrent wakeup can get prev
2667 * running on another CPU and we could rave with its RUNNING -> DEAD
2668 * transition, resulting in a double drop.
2670 prev_state = prev->state;
2671 vtime_task_switch(prev);
2672 perf_event_task_sched_in(prev, current);
2674 finish_lock_switch(rq);
2675 finish_arch_post_lock_switch();
2676 kcov_finish_switch(current);
2678 fire_sched_in_preempt_notifiers(current);
2680 * When switching through a kernel thread, the loop in
2681 * membarrier_{private,global}_expedited() may have observed that
2682 * kernel thread and not issued an IPI. It is therefore possible to
2683 * schedule between user->kernel->user threads without passing though
2684 * switch_mm(). Membarrier requires a barrier after storing to
2685 * rq->curr, before returning to userspace, so provide them here:
2687 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
2688 * provided by mmdrop(),
2689 * - a sync_core for SYNC_CORE.
2692 membarrier_mm_sync_core_before_usermode(mm);
2695 if (unlikely(prev_state == TASK_DEAD)) {
2696 if (prev->sched_class->task_dead)
2697 prev->sched_class->task_dead(prev);
2700 * Remove function-return probe instances associated with this
2701 * task and put them back on the free list.
2703 kprobe_flush_task(prev);
2705 /* Task is done with its stack. */
2706 put_task_stack(prev);
2708 put_task_struct(prev);
2711 tick_nohz_task_switch();
2717 /* rq->lock is NOT held, but preemption is disabled */
2718 static void __balance_callback(struct rq *rq)
2720 struct callback_head *head, *next;
2721 void (*func)(struct rq *rq);
2722 unsigned long flags;
2724 raw_spin_lock_irqsave(&rq->lock, flags);
2725 head = rq->balance_callback;
2726 rq->balance_callback = NULL;
2728 func = (void (*)(struct rq *))head->func;
2735 raw_spin_unlock_irqrestore(&rq->lock, flags);
2738 static inline void balance_callback(struct rq *rq)
2740 if (unlikely(rq->balance_callback))
2741 __balance_callback(rq);
2746 static inline void balance_callback(struct rq *rq)
2753 * schedule_tail - first thing a freshly forked thread must call.
2754 * @prev: the thread we just switched away from.
2756 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2757 __releases(rq->lock)
2762 * New tasks start with FORK_PREEMPT_COUNT, see there and
2763 * finish_task_switch() for details.
2765 * finish_task_switch() will drop rq->lock() and lower preempt_count
2766 * and the preempt_enable() will end up enabling preemption (on
2767 * PREEMPT_COUNT kernels).
2770 rq = finish_task_switch(prev);
2771 balance_callback(rq);
2774 if (current->set_child_tid)
2775 put_user(task_pid_vnr(current), current->set_child_tid);
2777 calculate_sigpending();
2781 * context_switch - switch to the new MM and the new thread's register state.
2783 static __always_inline struct rq *
2784 context_switch(struct rq *rq, struct task_struct *prev,
2785 struct task_struct *next, struct rq_flags *rf)
2787 struct mm_struct *mm, *oldmm;
2789 prepare_task_switch(rq, prev, next);
2792 oldmm = prev->active_mm;
2794 * For paravirt, this is coupled with an exit in switch_to to
2795 * combine the page table reload and the switch backend into
2798 arch_start_context_switch(prev);
2801 * If mm is non-NULL, we pass through switch_mm(). If mm is
2802 * NULL, we will pass through mmdrop() in finish_task_switch().
2803 * Both of these contain the full memory barrier required by
2804 * membarrier after storing to rq->curr, before returning to
2808 next->active_mm = oldmm;
2810 enter_lazy_tlb(oldmm, next);
2812 switch_mm_irqs_off(oldmm, mm, next);
2815 prev->active_mm = NULL;
2816 rq->prev_mm = oldmm;
2819 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2821 prepare_lock_switch(rq, next, rf);
2823 /* Here we just switch the register state and the stack. */
2824 switch_to(prev, next, prev);
2827 return finish_task_switch(prev);
2831 * nr_running and nr_context_switches:
2833 * externally visible scheduler statistics: current number of runnable
2834 * threads, total number of context switches performed since bootup.
2836 unsigned long nr_running(void)
2838 unsigned long i, sum = 0;
2840 for_each_online_cpu(i)
2841 sum += cpu_rq(i)->nr_running;
2847 * Check if only the current task is running on the CPU.
2849 * Caution: this function does not check that the caller has disabled
2850 * preemption, thus the result might have a time-of-check-to-time-of-use
2851 * race. The caller is responsible to use it correctly, for example:
2853 * - from a non-preemptable section (of course)
2855 * - from a thread that is bound to a single CPU
2857 * - in a loop with very short iterations (e.g. a polling loop)
2859 bool single_task_running(void)
2861 return raw_rq()->nr_running == 1;
2863 EXPORT_SYMBOL(single_task_running);
2865 unsigned long long nr_context_switches(void)
2868 unsigned long long sum = 0;
2870 for_each_possible_cpu(i)
2871 sum += cpu_rq(i)->nr_switches;
2877 * IO-wait accounting, and how its mostly bollocks (on SMP).
2879 * The idea behind IO-wait account is to account the idle time that we could
2880 * have spend running if it were not for IO. That is, if we were to improve the
2881 * storage performance, we'd have a proportional reduction in IO-wait time.
2883 * This all works nicely on UP, where, when a task blocks on IO, we account
2884 * idle time as IO-wait, because if the storage were faster, it could've been
2885 * running and we'd not be idle.
2887 * This has been extended to SMP, by doing the same for each CPU. This however
2890 * Imagine for instance the case where two tasks block on one CPU, only the one
2891 * CPU will have IO-wait accounted, while the other has regular idle. Even
2892 * though, if the storage were faster, both could've ran at the same time,
2893 * utilising both CPUs.
2895 * This means, that when looking globally, the current IO-wait accounting on
2896 * SMP is a lower bound, by reason of under accounting.
2898 * Worse, since the numbers are provided per CPU, they are sometimes
2899 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2900 * associated with any one particular CPU, it can wake to another CPU than it
2901 * blocked on. This means the per CPU IO-wait number is meaningless.
2903 * Task CPU affinities can make all that even more 'interesting'.
2906 unsigned long nr_iowait(void)
2908 unsigned long i, sum = 0;
2910 for_each_possible_cpu(i)
2911 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2917 * Consumers of these two interfaces, like for example the cpufreq menu
2918 * governor are using nonsensical data. Boosting frequency for a CPU that has
2919 * IO-wait which might not even end up running the task when it does become
2923 unsigned long nr_iowait_cpu(int cpu)
2925 struct rq *this = cpu_rq(cpu);
2926 return atomic_read(&this->nr_iowait);
2929 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2931 struct rq *rq = this_rq();
2932 *nr_waiters = atomic_read(&rq->nr_iowait);
2933 *load = rq->load.weight;
2939 * sched_exec - execve() is a valuable balancing opportunity, because at
2940 * this point the task has the smallest effective memory and cache footprint.
2942 void sched_exec(void)
2944 struct task_struct *p = current;
2945 unsigned long flags;
2948 raw_spin_lock_irqsave(&p->pi_lock, flags);
2949 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2950 if (dest_cpu == smp_processor_id())
2953 if (likely(cpu_active(dest_cpu))) {
2954 struct migration_arg arg = { p, dest_cpu };
2956 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2957 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2961 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2966 DEFINE_PER_CPU(struct kernel_stat, kstat);
2967 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2969 EXPORT_PER_CPU_SYMBOL(kstat);
2970 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2973 * The function fair_sched_class.update_curr accesses the struct curr
2974 * and its field curr->exec_start; when called from task_sched_runtime(),
2975 * we observe a high rate of cache misses in practice.
2976 * Prefetching this data results in improved performance.
2978 static inline void prefetch_curr_exec_start(struct task_struct *p)
2980 #ifdef CONFIG_FAIR_GROUP_SCHED
2981 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
2983 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
2986 prefetch(&curr->exec_start);
2990 * Return accounted runtime for the task.
2991 * In case the task is currently running, return the runtime plus current's
2992 * pending runtime that have not been accounted yet.
2994 unsigned long long task_sched_runtime(struct task_struct *p)
3000 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3002 * 64-bit doesn't need locks to atomically read a 64-bit value.
3003 * So we have a optimization chance when the task's delta_exec is 0.
3004 * Reading ->on_cpu is racy, but this is ok.
3006 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3007 * If we race with it entering CPU, unaccounted time is 0. This is
3008 * indistinguishable from the read occurring a few cycles earlier.
3009 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3010 * been accounted, so we're correct here as well.
3012 if (!p->on_cpu || !task_on_rq_queued(p))
3013 return p->se.sum_exec_runtime;
3016 rq = task_rq_lock(p, &rf);
3018 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3019 * project cycles that may never be accounted to this
3020 * thread, breaking clock_gettime().
3022 if (task_current(rq, p) && task_on_rq_queued(p)) {
3023 prefetch_curr_exec_start(p);
3024 update_rq_clock(rq);
3025 p->sched_class->update_curr(rq);
3027 ns = p->se.sum_exec_runtime;
3028 task_rq_unlock(rq, p, &rf);
3034 * This function gets called by the timer code, with HZ frequency.
3035 * We call it with interrupts disabled.
3037 void scheduler_tick(void)
3039 int cpu = smp_processor_id();
3040 struct rq *rq = cpu_rq(cpu);
3041 struct task_struct *curr = rq->curr;
3048 update_rq_clock(rq);
3049 curr->sched_class->task_tick(rq, curr, 0);
3050 cpu_load_update_active(rq);
3051 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 local_irq_disable();
4938 schedstat_inc(rq->yld_count);
4939 current->sched_class->yield_task(rq);
4942 * Since we are going to call schedule() anyway, there's
4943 * no need to preempt or enable interrupts:
4947 sched_preempt_enable_no_resched();
4952 SYSCALL_DEFINE0(sched_yield)
4958 #ifndef CONFIG_PREEMPT
4959 int __sched _cond_resched(void)
4961 if (should_resched(0)) {
4962 preempt_schedule_common();
4968 EXPORT_SYMBOL(_cond_resched);
4972 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4973 * call schedule, and on return reacquire the lock.
4975 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4976 * operations here to prevent schedule() from being called twice (once via
4977 * spin_unlock(), once by hand).
4979 int __cond_resched_lock(spinlock_t *lock)
4981 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4984 lockdep_assert_held(lock);
4986 if (spin_needbreak(lock) || resched) {
4989 preempt_schedule_common();
4997 EXPORT_SYMBOL(__cond_resched_lock);
5000 * yield - yield the current processor to other threads.
5002 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5004 * The scheduler is at all times free to pick the calling task as the most
5005 * eligible task to run, if removing the yield() call from your code breaks
5006 * it, its already broken.
5008 * Typical broken usage is:
5013 * where one assumes that yield() will let 'the other' process run that will
5014 * make event true. If the current task is a SCHED_FIFO task that will never
5015 * happen. Never use yield() as a progress guarantee!!
5017 * If you want to use yield() to wait for something, use wait_event().
5018 * If you want to use yield() to be 'nice' for others, use cond_resched().
5019 * If you still want to use yield(), do not!
5021 void __sched yield(void)
5023 set_current_state(TASK_RUNNING);
5026 EXPORT_SYMBOL(yield);
5029 * yield_to - yield the current processor to another thread in
5030 * your thread group, or accelerate that thread toward the
5031 * processor it's on.
5033 * @preempt: whether task preemption is allowed or not
5035 * It's the caller's job to ensure that the target task struct
5036 * can't go away on us before we can do any checks.
5039 * true (>0) if we indeed boosted the target task.
5040 * false (0) if we failed to boost the target.
5041 * -ESRCH if there's no task to yield to.
5043 int __sched yield_to(struct task_struct *p, bool preempt)
5045 struct task_struct *curr = current;
5046 struct rq *rq, *p_rq;
5047 unsigned long flags;
5050 local_irq_save(flags);
5056 * If we're the only runnable task on the rq and target rq also
5057 * has only one task, there's absolutely no point in yielding.
5059 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5064 double_rq_lock(rq, p_rq);
5065 if (task_rq(p) != p_rq) {
5066 double_rq_unlock(rq, p_rq);
5070 if (!curr->sched_class->yield_to_task)
5073 if (curr->sched_class != p->sched_class)
5076 if (task_running(p_rq, p) || p->state)
5079 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5081 schedstat_inc(rq->yld_count);
5083 * Make p's CPU reschedule; pick_next_entity takes care of
5086 if (preempt && rq != p_rq)
5091 double_rq_unlock(rq, p_rq);
5093 local_irq_restore(flags);
5100 EXPORT_SYMBOL_GPL(yield_to);
5102 int io_schedule_prepare(void)
5104 int old_iowait = current->in_iowait;
5106 current->in_iowait = 1;
5107 blk_schedule_flush_plug(current);
5112 void io_schedule_finish(int token)
5114 current->in_iowait = token;
5118 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5119 * that process accounting knows that this is a task in IO wait state.
5121 long __sched io_schedule_timeout(long timeout)
5126 token = io_schedule_prepare();
5127 ret = schedule_timeout(timeout);
5128 io_schedule_finish(token);
5132 EXPORT_SYMBOL(io_schedule_timeout);
5134 void io_schedule(void)
5138 token = io_schedule_prepare();
5140 io_schedule_finish(token);
5142 EXPORT_SYMBOL(io_schedule);
5145 * sys_sched_get_priority_max - return maximum RT priority.
5146 * @policy: scheduling class.
5148 * Return: On success, this syscall returns the maximum
5149 * rt_priority that can be used by a given scheduling class.
5150 * On failure, a negative error code is returned.
5152 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5159 ret = MAX_USER_RT_PRIO-1;
5161 case SCHED_DEADLINE:
5172 * sys_sched_get_priority_min - return minimum RT priority.
5173 * @policy: scheduling class.
5175 * Return: On success, this syscall returns the minimum
5176 * rt_priority that can be used by a given scheduling class.
5177 * On failure, a negative error code is returned.
5179 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5188 case SCHED_DEADLINE:
5197 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5199 struct task_struct *p;
5200 unsigned int time_slice;
5210 p = find_process_by_pid(pid);
5214 retval = security_task_getscheduler(p);
5218 rq = task_rq_lock(p, &rf);
5220 if (p->sched_class->get_rr_interval)
5221 time_slice = p->sched_class->get_rr_interval(rq, p);
5222 task_rq_unlock(rq, p, &rf);
5225 jiffies_to_timespec64(time_slice, t);
5234 * sys_sched_rr_get_interval - return the default timeslice of a process.
5235 * @pid: pid of the process.
5236 * @interval: userspace pointer to the timeslice value.
5238 * this syscall writes the default timeslice value of a given process
5239 * into the user-space timespec buffer. A value of '0' means infinity.
5241 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5244 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5245 struct timespec __user *, interval)
5247 struct timespec64 t;
5248 int retval = sched_rr_get_interval(pid, &t);
5251 retval = put_timespec64(&t, interval);
5256 #ifdef CONFIG_COMPAT
5257 COMPAT_SYSCALL_DEFINE2(sched_rr_get_interval,
5259 struct compat_timespec __user *, interval)
5261 struct timespec64 t;
5262 int retval = sched_rr_get_interval(pid, &t);
5265 retval = compat_put_timespec64(&t, interval);
5270 void sched_show_task(struct task_struct *p)
5272 unsigned long free = 0;
5275 if (!try_get_task_stack(p))
5278 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5280 if (p->state == TASK_RUNNING)
5281 printk(KERN_CONT " running task ");
5282 #ifdef CONFIG_DEBUG_STACK_USAGE
5283 free = stack_not_used(p);
5288 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5290 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5291 task_pid_nr(p), ppid,
5292 (unsigned long)task_thread_info(p)->flags);
5294 print_worker_info(KERN_INFO, p);
5295 show_stack(p, NULL);
5298 EXPORT_SYMBOL_GPL(sched_show_task);
5301 state_filter_match(unsigned long state_filter, struct task_struct *p)
5303 /* no filter, everything matches */
5307 /* filter, but doesn't match */
5308 if (!(p->state & state_filter))
5312 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5315 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5322 void show_state_filter(unsigned long state_filter)
5324 struct task_struct *g, *p;
5326 #if BITS_PER_LONG == 32
5328 " task PC stack pid father\n");
5331 " task PC stack pid father\n");
5334 for_each_process_thread(g, p) {
5336 * reset the NMI-timeout, listing all files on a slow
5337 * console might take a lot of time:
5338 * Also, reset softlockup watchdogs on all CPUs, because
5339 * another CPU might be blocked waiting for us to process
5342 touch_nmi_watchdog();
5343 touch_all_softlockup_watchdogs();
5344 if (state_filter_match(state_filter, p))
5348 #ifdef CONFIG_SCHED_DEBUG
5350 sysrq_sched_debug_show();
5354 * Only show locks if all tasks are dumped:
5357 debug_show_all_locks();
5361 * init_idle - set up an idle thread for a given CPU
5362 * @idle: task in question
5363 * @cpu: CPU the idle task belongs to
5365 * NOTE: this function does not set the idle thread's NEED_RESCHED
5366 * flag, to make booting more robust.
5368 void init_idle(struct task_struct *idle, int cpu)
5370 struct rq *rq = cpu_rq(cpu);
5371 unsigned long flags;
5373 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5374 raw_spin_lock(&rq->lock);
5376 __sched_fork(0, idle);
5377 idle->state = TASK_RUNNING;
5378 idle->se.exec_start = sched_clock();
5379 idle->flags |= PF_IDLE;
5381 kasan_unpoison_task_stack(idle);
5385 * Its possible that init_idle() gets called multiple times on a task,
5386 * in that case do_set_cpus_allowed() will not do the right thing.
5388 * And since this is boot we can forgo the serialization.
5390 set_cpus_allowed_common(idle, cpumask_of(cpu));
5393 * We're having a chicken and egg problem, even though we are
5394 * holding rq->lock, the CPU isn't yet set to this CPU so the
5395 * lockdep check in task_group() will fail.
5397 * Similar case to sched_fork(). / Alternatively we could
5398 * use task_rq_lock() here and obtain the other rq->lock.
5403 __set_task_cpu(idle, cpu);
5406 rq->curr = rq->idle = idle;
5407 idle->on_rq = TASK_ON_RQ_QUEUED;
5411 raw_spin_unlock(&rq->lock);
5412 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5414 /* Set the preempt count _outside_ the spinlocks! */
5415 init_idle_preempt_count(idle, cpu);
5418 * The idle tasks have their own, simple scheduling class:
5420 idle->sched_class = &idle_sched_class;
5421 ftrace_graph_init_idle_task(idle, cpu);
5422 vtime_init_idle(idle, cpu);
5424 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5430 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5431 const struct cpumask *trial)
5435 if (!cpumask_weight(cur))
5438 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5443 int task_can_attach(struct task_struct *p,
5444 const struct cpumask *cs_cpus_allowed)
5449 * Kthreads which disallow setaffinity shouldn't be moved
5450 * to a new cpuset; we don't want to change their CPU
5451 * affinity and isolating such threads by their set of
5452 * allowed nodes is unnecessary. Thus, cpusets are not
5453 * applicable for such threads. This prevents checking for
5454 * success of set_cpus_allowed_ptr() on all attached tasks
5455 * before cpus_allowed may be changed.
5457 if (p->flags & PF_NO_SETAFFINITY) {
5462 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5464 ret = dl_task_can_attach(p, cs_cpus_allowed);
5470 bool sched_smp_initialized __read_mostly;
5472 #ifdef CONFIG_NUMA_BALANCING
5473 /* Migrate current task p to target_cpu */
5474 int migrate_task_to(struct task_struct *p, int target_cpu)
5476 struct migration_arg arg = { p, target_cpu };
5477 int curr_cpu = task_cpu(p);
5479 if (curr_cpu == target_cpu)
5482 if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5485 /* TODO: This is not properly updating schedstats */
5487 trace_sched_move_numa(p, curr_cpu, target_cpu);
5488 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5492 * Requeue a task on a given node and accurately track the number of NUMA
5493 * tasks on the runqueues
5495 void sched_setnuma(struct task_struct *p, int nid)
5497 bool queued, running;
5501 rq = task_rq_lock(p, &rf);
5502 queued = task_on_rq_queued(p);
5503 running = task_current(rq, p);
5506 dequeue_task(rq, p, DEQUEUE_SAVE);
5508 put_prev_task(rq, p);
5510 p->numa_preferred_nid = nid;
5513 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5515 set_curr_task(rq, p);
5516 task_rq_unlock(rq, p, &rf);
5518 #endif /* CONFIG_NUMA_BALANCING */
5520 #ifdef CONFIG_HOTPLUG_CPU
5522 * Ensure that the idle task is using init_mm right before its CPU goes
5525 void idle_task_exit(void)
5527 struct mm_struct *mm = current->active_mm;
5529 BUG_ON(cpu_online(smp_processor_id()));
5531 if (mm != &init_mm) {
5532 switch_mm(mm, &init_mm, current);
5533 current->active_mm = &init_mm;
5534 finish_arch_post_lock_switch();
5540 * Since this CPU is going 'away' for a while, fold any nr_active delta
5541 * we might have. Assumes we're called after migrate_tasks() so that the
5542 * nr_active count is stable. We need to take the teardown thread which
5543 * is calling this into account, so we hand in adjust = 1 to the load
5546 * Also see the comment "Global load-average calculations".
5548 static void calc_load_migrate(struct rq *rq)
5550 long delta = calc_load_fold_active(rq, 1);
5552 atomic_long_add(delta, &calc_load_tasks);
5555 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5559 static const struct sched_class fake_sched_class = {
5560 .put_prev_task = put_prev_task_fake,
5563 static struct task_struct fake_task = {
5565 * Avoid pull_{rt,dl}_task()
5567 .prio = MAX_PRIO + 1,
5568 .sched_class = &fake_sched_class,
5572 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5573 * try_to_wake_up()->select_task_rq().
5575 * Called with rq->lock held even though we'er in stop_machine() and
5576 * there's no concurrency possible, we hold the required locks anyway
5577 * because of lock validation efforts.
5579 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5581 struct rq *rq = dead_rq;
5582 struct task_struct *next, *stop = rq->stop;
5583 struct rq_flags orf = *rf;
5587 * Fudge the rq selection such that the below task selection loop
5588 * doesn't get stuck on the currently eligible stop task.
5590 * We're currently inside stop_machine() and the rq is either stuck
5591 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5592 * either way we should never end up calling schedule() until we're
5598 * put_prev_task() and pick_next_task() sched
5599 * class method both need to have an up-to-date
5600 * value of rq->clock[_task]
5602 update_rq_clock(rq);
5606 * There's this thread running, bail when that's the only
5609 if (rq->nr_running == 1)
5613 * pick_next_task() assumes pinned rq->lock:
5615 next = pick_next_task(rq, &fake_task, rf);
5617 put_prev_task(rq, next);
5620 * Rules for changing task_struct::cpus_allowed are holding
5621 * both pi_lock and rq->lock, such that holding either
5622 * stabilizes the mask.
5624 * Drop rq->lock is not quite as disastrous as it usually is
5625 * because !cpu_active at this point, which means load-balance
5626 * will not interfere. Also, stop-machine.
5629 raw_spin_lock(&next->pi_lock);
5633 * Since we're inside stop-machine, _nothing_ should have
5634 * changed the task, WARN if weird stuff happened, because in
5635 * that case the above rq->lock drop is a fail too.
5637 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5638 raw_spin_unlock(&next->pi_lock);
5642 /* Find suitable destination for @next, with force if needed. */
5643 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5644 rq = __migrate_task(rq, rf, next, dest_cpu);
5645 if (rq != dead_rq) {
5651 raw_spin_unlock(&next->pi_lock);
5656 #endif /* CONFIG_HOTPLUG_CPU */
5658 void set_rq_online(struct rq *rq)
5661 const struct sched_class *class;
5663 cpumask_set_cpu(rq->cpu, rq->rd->online);
5666 for_each_class(class) {
5667 if (class->rq_online)
5668 class->rq_online(rq);
5673 void set_rq_offline(struct rq *rq)
5676 const struct sched_class *class;
5678 for_each_class(class) {
5679 if (class->rq_offline)
5680 class->rq_offline(rq);
5683 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5689 * used to mark begin/end of suspend/resume:
5691 static int num_cpus_frozen;
5694 * Update cpusets according to cpu_active mask. If cpusets are
5695 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5696 * around partition_sched_domains().
5698 * If we come here as part of a suspend/resume, don't touch cpusets because we
5699 * want to restore it back to its original state upon resume anyway.
5701 static void cpuset_cpu_active(void)
5703 if (cpuhp_tasks_frozen) {
5705 * num_cpus_frozen tracks how many CPUs are involved in suspend
5706 * resume sequence. As long as this is not the last online
5707 * operation in the resume sequence, just build a single sched
5708 * domain, ignoring cpusets.
5710 partition_sched_domains(1, NULL, NULL);
5711 if (--num_cpus_frozen)
5714 * This is the last CPU online operation. So fall through and
5715 * restore the original sched domains by considering the
5716 * cpuset configurations.
5718 cpuset_force_rebuild();
5720 cpuset_update_active_cpus();
5723 static int cpuset_cpu_inactive(unsigned int cpu)
5725 if (!cpuhp_tasks_frozen) {
5726 if (dl_cpu_busy(cpu))
5728 cpuset_update_active_cpus();
5731 partition_sched_domains(1, NULL, NULL);
5736 int sched_cpu_activate(unsigned int cpu)
5738 struct rq *rq = cpu_rq(cpu);
5741 #ifdef CONFIG_SCHED_SMT
5743 * The sched_smt_present static key needs to be evaluated on every
5744 * hotplug event because at boot time SMT might be disabled when
5745 * the number of booted CPUs is limited.
5747 * If then later a sibling gets hotplugged, then the key would stay
5748 * off and SMT scheduling would never be functional.
5750 if (cpumask_weight(cpu_smt_mask(cpu)) > 1)
5751 static_branch_enable_cpuslocked(&sched_smt_present);
5753 set_cpu_active(cpu, true);
5755 if (sched_smp_initialized) {
5756 sched_domains_numa_masks_set(cpu);
5757 cpuset_cpu_active();
5761 * Put the rq online, if not already. This happens:
5763 * 1) In the early boot process, because we build the real domains
5764 * after all CPUs have been brought up.
5766 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5769 rq_lock_irqsave(rq, &rf);
5771 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5774 rq_unlock_irqrestore(rq, &rf);
5776 update_max_interval();
5781 int sched_cpu_deactivate(unsigned int cpu)
5785 set_cpu_active(cpu, false);
5787 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5788 * users of this state to go away such that all new such users will
5791 * Do sync before park smpboot threads to take care the rcu boost case.
5793 synchronize_rcu_mult(call_rcu, call_rcu_sched);
5795 if (!sched_smp_initialized)
5798 ret = cpuset_cpu_inactive(cpu);
5800 set_cpu_active(cpu, true);
5803 sched_domains_numa_masks_clear(cpu);
5807 static void sched_rq_cpu_starting(unsigned int cpu)
5809 struct rq *rq = cpu_rq(cpu);
5811 rq->calc_load_update = calc_load_update;
5812 update_max_interval();
5815 int sched_cpu_starting(unsigned int cpu)
5817 sched_rq_cpu_starting(cpu);
5818 sched_tick_start(cpu);
5822 #ifdef CONFIG_HOTPLUG_CPU
5823 int sched_cpu_dying(unsigned int cpu)
5825 struct rq *rq = cpu_rq(cpu);
5828 /* Handle pending wakeups and then migrate everything off */
5829 sched_ttwu_pending();
5830 sched_tick_stop(cpu);
5832 rq_lock_irqsave(rq, &rf);
5834 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5837 migrate_tasks(rq, &rf);
5838 BUG_ON(rq->nr_running != 1);
5839 rq_unlock_irqrestore(rq, &rf);
5841 calc_load_migrate(rq);
5842 update_max_interval();
5843 nohz_balance_exit_idle(rq);
5849 void __init sched_init_smp(void)
5854 * There's no userspace yet to cause hotplug operations; hence all the
5855 * CPU masks are stable and all blatant races in the below code cannot
5858 mutex_lock(&sched_domains_mutex);
5859 sched_init_domains(cpu_active_mask);
5860 mutex_unlock(&sched_domains_mutex);
5862 /* Move init over to a non-isolated CPU */
5863 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
5865 sched_init_granularity();
5867 init_sched_rt_class();
5868 init_sched_dl_class();
5870 sched_smp_initialized = true;
5873 static int __init migration_init(void)
5875 sched_rq_cpu_starting(smp_processor_id());
5878 early_initcall(migration_init);
5881 void __init sched_init_smp(void)
5883 sched_init_granularity();
5885 #endif /* CONFIG_SMP */
5887 int in_sched_functions(unsigned long addr)
5889 return in_lock_functions(addr) ||
5890 (addr >= (unsigned long)__sched_text_start
5891 && addr < (unsigned long)__sched_text_end);
5894 #ifdef CONFIG_CGROUP_SCHED
5896 * Default task group.
5897 * Every task in system belongs to this group at bootup.
5899 struct task_group root_task_group;
5900 LIST_HEAD(task_groups);
5902 /* Cacheline aligned slab cache for task_group */
5903 static struct kmem_cache *task_group_cache __read_mostly;
5906 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5907 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5909 void __init sched_init(void)
5912 unsigned long alloc_size = 0, ptr;
5916 #ifdef CONFIG_FAIR_GROUP_SCHED
5917 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5919 #ifdef CONFIG_RT_GROUP_SCHED
5920 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5923 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5925 #ifdef CONFIG_FAIR_GROUP_SCHED
5926 root_task_group.se = (struct sched_entity **)ptr;
5927 ptr += nr_cpu_ids * sizeof(void **);
5929 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5930 ptr += nr_cpu_ids * sizeof(void **);
5932 #endif /* CONFIG_FAIR_GROUP_SCHED */
5933 #ifdef CONFIG_RT_GROUP_SCHED
5934 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5935 ptr += nr_cpu_ids * sizeof(void **);
5937 root_task_group.rt_rq = (struct rt_rq **)ptr;
5938 ptr += nr_cpu_ids * sizeof(void **);
5940 #endif /* CONFIG_RT_GROUP_SCHED */
5942 #ifdef CONFIG_CPUMASK_OFFSTACK
5943 for_each_possible_cpu(i) {
5944 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5945 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5946 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
5947 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5949 #endif /* CONFIG_CPUMASK_OFFSTACK */
5951 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
5952 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
5955 init_defrootdomain();
5958 #ifdef CONFIG_RT_GROUP_SCHED
5959 init_rt_bandwidth(&root_task_group.rt_bandwidth,
5960 global_rt_period(), global_rt_runtime());
5961 #endif /* CONFIG_RT_GROUP_SCHED */
5963 #ifdef CONFIG_CGROUP_SCHED
5964 task_group_cache = KMEM_CACHE(task_group, 0);
5966 list_add(&root_task_group.list, &task_groups);
5967 INIT_LIST_HEAD(&root_task_group.children);
5968 INIT_LIST_HEAD(&root_task_group.siblings);
5969 autogroup_init(&init_task);
5970 #endif /* CONFIG_CGROUP_SCHED */
5972 for_each_possible_cpu(i) {
5976 raw_spin_lock_init(&rq->lock);
5978 rq->calc_load_active = 0;
5979 rq->calc_load_update = jiffies + LOAD_FREQ;
5980 init_cfs_rq(&rq->cfs);
5981 init_rt_rq(&rq->rt);
5982 init_dl_rq(&rq->dl);
5983 #ifdef CONFIG_FAIR_GROUP_SCHED
5984 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
5985 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
5986 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
5988 * How much CPU bandwidth does root_task_group get?
5990 * In case of task-groups formed thr' the cgroup filesystem, it
5991 * gets 100% of the CPU resources in the system. This overall
5992 * system CPU resource is divided among the tasks of
5993 * root_task_group and its child task-groups in a fair manner,
5994 * based on each entity's (task or task-group's) weight
5995 * (se->load.weight).
5997 * In other words, if root_task_group has 10 tasks of weight
5998 * 1024) and two child groups A0 and A1 (of weight 1024 each),
5999 * then A0's share of the CPU resource is:
6001 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6003 * We achieve this by letting root_task_group's tasks sit
6004 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6006 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6007 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6008 #endif /* CONFIG_FAIR_GROUP_SCHED */
6010 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6011 #ifdef CONFIG_RT_GROUP_SCHED
6012 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6015 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6016 rq->cpu_load[j] = 0;
6021 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6022 rq->balance_callback = NULL;
6023 rq->active_balance = 0;
6024 rq->next_balance = jiffies;
6029 rq->avg_idle = 2*sysctl_sched_migration_cost;
6030 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6032 INIT_LIST_HEAD(&rq->cfs_tasks);
6034 rq_attach_root(rq, &def_root_domain);
6035 #ifdef CONFIG_NO_HZ_COMMON
6036 rq->last_load_update_tick = jiffies;
6037 rq->last_blocked_load_update_tick = jiffies;
6038 atomic_set(&rq->nohz_flags, 0);
6040 #endif /* CONFIG_SMP */
6042 atomic_set(&rq->nr_iowait, 0);
6045 set_load_weight(&init_task, false);
6048 * The boot idle thread does lazy MMU switching as well:
6051 enter_lazy_tlb(&init_mm, current);
6054 * Make us the idle thread. Technically, schedule() should not be
6055 * called from this thread, however somewhere below it might be,
6056 * but because we are the idle thread, we just pick up running again
6057 * when this runqueue becomes "idle".
6059 init_idle(current, smp_processor_id());
6061 calc_load_update = jiffies + LOAD_FREQ;
6064 idle_thread_set_boot_cpu();
6066 init_sched_fair_class();
6070 scheduler_running = 1;
6073 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6074 static inline int preempt_count_equals(int preempt_offset)
6076 int nested = preempt_count() + rcu_preempt_depth();
6078 return (nested == preempt_offset);
6081 void __might_sleep(const char *file, int line, int preempt_offset)
6084 * Blocking primitives will set (and therefore destroy) current->state,
6085 * since we will exit with TASK_RUNNING make sure we enter with it,
6086 * otherwise we will destroy state.
6088 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6089 "do not call blocking ops when !TASK_RUNNING; "
6090 "state=%lx set at [<%p>] %pS\n",
6092 (void *)current->task_state_change,
6093 (void *)current->task_state_change);
6095 ___might_sleep(file, line, preempt_offset);
6097 EXPORT_SYMBOL(__might_sleep);
6099 void ___might_sleep(const char *file, int line, int preempt_offset)
6101 /* Ratelimiting timestamp: */
6102 static unsigned long prev_jiffy;
6104 unsigned long preempt_disable_ip;
6106 /* WARN_ON_ONCE() by default, no rate limit required: */
6109 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6110 !is_idle_task(current)) ||
6111 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6115 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6117 prev_jiffy = jiffies;
6119 /* Save this before calling printk(), since that will clobber it: */
6120 preempt_disable_ip = get_preempt_disable_ip(current);
6123 "BUG: sleeping function called from invalid context at %s:%d\n",
6126 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6127 in_atomic(), irqs_disabled(),
6128 current->pid, current->comm);
6130 if (task_stack_end_corrupted(current))
6131 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6133 debug_show_held_locks(current);
6134 if (irqs_disabled())
6135 print_irqtrace_events(current);
6136 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6137 && !preempt_count_equals(preempt_offset)) {
6138 pr_err("Preemption disabled at:");
6139 print_ip_sym(preempt_disable_ip);
6143 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6145 EXPORT_SYMBOL(___might_sleep);
6148 #ifdef CONFIG_MAGIC_SYSRQ
6149 void normalize_rt_tasks(void)
6151 struct task_struct *g, *p;
6152 struct sched_attr attr = {
6153 .sched_policy = SCHED_NORMAL,
6156 read_lock(&tasklist_lock);
6157 for_each_process_thread(g, p) {
6159 * Only normalize user tasks:
6161 if (p->flags & PF_KTHREAD)
6164 p->se.exec_start = 0;
6165 schedstat_set(p->se.statistics.wait_start, 0);
6166 schedstat_set(p->se.statistics.sleep_start, 0);
6167 schedstat_set(p->se.statistics.block_start, 0);
6169 if (!dl_task(p) && !rt_task(p)) {
6171 * Renice negative nice level userspace
6174 if (task_nice(p) < 0)
6175 set_user_nice(p, 0);
6179 __sched_setscheduler(p, &attr, false, false);
6181 read_unlock(&tasklist_lock);
6184 #endif /* CONFIG_MAGIC_SYSRQ */
6186 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6188 * These functions are only useful for the IA64 MCA handling, or kdb.
6190 * They can only be called when the whole system has been
6191 * stopped - every CPU needs to be quiescent, and no scheduling
6192 * activity can take place. Using them for anything else would
6193 * be a serious bug, and as a result, they aren't even visible
6194 * under any other configuration.
6198 * curr_task - return the current task for a given CPU.
6199 * @cpu: the processor in question.
6201 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6203 * Return: The current task for @cpu.
6205 struct task_struct *curr_task(int cpu)
6207 return cpu_curr(cpu);
6210 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6214 * set_curr_task - set the current task for a given CPU.
6215 * @cpu: the processor in question.
6216 * @p: the task pointer to set.
6218 * Description: This function must only be used when non-maskable interrupts
6219 * are serviced on a separate stack. It allows the architecture to switch the
6220 * notion of the current task on a CPU in a non-blocking manner. This function
6221 * must be called with all CPU's synchronized, and interrupts disabled, the
6222 * and caller must save the original value of the current task (see
6223 * curr_task() above) and restore that value before reenabling interrupts and
6224 * re-starting the system.
6226 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6228 void ia64_set_curr_task(int cpu, struct task_struct *p)
6235 #ifdef CONFIG_CGROUP_SCHED
6236 /* task_group_lock serializes the addition/removal of task groups */
6237 static DEFINE_SPINLOCK(task_group_lock);
6239 static void sched_free_group(struct task_group *tg)
6241 free_fair_sched_group(tg);
6242 free_rt_sched_group(tg);
6244 kmem_cache_free(task_group_cache, tg);
6247 /* allocate runqueue etc for a new task group */
6248 struct task_group *sched_create_group(struct task_group *parent)
6250 struct task_group *tg;
6252 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6254 return ERR_PTR(-ENOMEM);
6256 if (!alloc_fair_sched_group(tg, parent))
6259 if (!alloc_rt_sched_group(tg, parent))
6265 sched_free_group(tg);
6266 return ERR_PTR(-ENOMEM);
6269 void sched_online_group(struct task_group *tg, struct task_group *parent)
6271 unsigned long flags;
6273 spin_lock_irqsave(&task_group_lock, flags);
6274 list_add_rcu(&tg->list, &task_groups);
6276 /* Root should already exist: */
6279 tg->parent = parent;
6280 INIT_LIST_HEAD(&tg->children);
6281 list_add_rcu(&tg->siblings, &parent->children);
6282 spin_unlock_irqrestore(&task_group_lock, flags);
6284 online_fair_sched_group(tg);
6287 /* rcu callback to free various structures associated with a task group */
6288 static void sched_free_group_rcu(struct rcu_head *rhp)
6290 /* Now it should be safe to free those cfs_rqs: */
6291 sched_free_group(container_of(rhp, struct task_group, rcu));
6294 void sched_destroy_group(struct task_group *tg)
6296 /* Wait for possible concurrent references to cfs_rqs complete: */
6297 call_rcu(&tg->rcu, sched_free_group_rcu);
6300 void sched_offline_group(struct task_group *tg)
6302 unsigned long flags;
6304 /* End participation in shares distribution: */
6305 unregister_fair_sched_group(tg);
6307 spin_lock_irqsave(&task_group_lock, flags);
6308 list_del_rcu(&tg->list);
6309 list_del_rcu(&tg->siblings);
6310 spin_unlock_irqrestore(&task_group_lock, flags);
6313 static void sched_change_group(struct task_struct *tsk, int type)
6315 struct task_group *tg;
6318 * All callers are synchronized by task_rq_lock(); we do not use RCU
6319 * which is pointless here. Thus, we pass "true" to task_css_check()
6320 * to prevent lockdep warnings.
6322 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6323 struct task_group, css);
6324 tg = autogroup_task_group(tsk, tg);
6325 tsk->sched_task_group = tg;
6327 #ifdef CONFIG_FAIR_GROUP_SCHED
6328 if (tsk->sched_class->task_change_group)
6329 tsk->sched_class->task_change_group(tsk, type);
6332 set_task_rq(tsk, task_cpu(tsk));
6336 * Change task's runqueue when it moves between groups.
6338 * The caller of this function should have put the task in its new group by
6339 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6342 void sched_move_task(struct task_struct *tsk)
6344 int queued, running, queue_flags =
6345 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6349 rq = task_rq_lock(tsk, &rf);
6350 update_rq_clock(rq);
6352 running = task_current(rq, tsk);
6353 queued = task_on_rq_queued(tsk);
6356 dequeue_task(rq, tsk, queue_flags);
6358 put_prev_task(rq, tsk);
6360 sched_change_group(tsk, TASK_MOVE_GROUP);
6363 enqueue_task(rq, tsk, queue_flags);
6365 set_curr_task(rq, tsk);
6367 task_rq_unlock(rq, tsk, &rf);
6370 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6372 return css ? container_of(css, struct task_group, css) : NULL;
6375 static struct cgroup_subsys_state *
6376 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6378 struct task_group *parent = css_tg(parent_css);
6379 struct task_group *tg;
6382 /* This is early initialization for the top cgroup */
6383 return &root_task_group.css;
6386 tg = sched_create_group(parent);
6388 return ERR_PTR(-ENOMEM);
6393 /* Expose task group only after completing cgroup initialization */
6394 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6396 struct task_group *tg = css_tg(css);
6397 struct task_group *parent = css_tg(css->parent);
6400 sched_online_group(tg, parent);
6404 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6406 struct task_group *tg = css_tg(css);
6408 sched_offline_group(tg);
6411 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6413 struct task_group *tg = css_tg(css);
6416 * Relies on the RCU grace period between css_released() and this.
6418 sched_free_group(tg);
6422 * This is called before wake_up_new_task(), therefore we really only
6423 * have to set its group bits, all the other stuff does not apply.
6425 static void cpu_cgroup_fork(struct task_struct *task)
6430 rq = task_rq_lock(task, &rf);
6432 update_rq_clock(rq);
6433 sched_change_group(task, TASK_SET_GROUP);
6435 task_rq_unlock(rq, task, &rf);
6438 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6440 struct task_struct *task;
6441 struct cgroup_subsys_state *css;
6444 cgroup_taskset_for_each(task, css, tset) {
6445 #ifdef CONFIG_RT_GROUP_SCHED
6446 if (!sched_rt_can_attach(css_tg(css), task))
6449 /* We don't support RT-tasks being in separate groups */
6450 if (task->sched_class != &fair_sched_class)
6454 * Serialize against wake_up_new_task() such that if its
6455 * running, we're sure to observe its full state.
6457 raw_spin_lock_irq(&task->pi_lock);
6459 * Avoid calling sched_move_task() before wake_up_new_task()
6460 * has happened. This would lead to problems with PELT, due to
6461 * move wanting to detach+attach while we're not attached yet.
6463 if (task->state == TASK_NEW)
6465 raw_spin_unlock_irq(&task->pi_lock);
6473 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6475 struct task_struct *task;
6476 struct cgroup_subsys_state *css;
6478 cgroup_taskset_for_each(task, css, tset)
6479 sched_move_task(task);
6482 #ifdef CONFIG_FAIR_GROUP_SCHED
6483 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6484 struct cftype *cftype, u64 shareval)
6486 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6489 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6492 struct task_group *tg = css_tg(css);
6494 return (u64) scale_load_down(tg->shares);
6497 #ifdef CONFIG_CFS_BANDWIDTH
6498 static DEFINE_MUTEX(cfs_constraints_mutex);
6500 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6501 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6503 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6505 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6507 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6508 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6510 if (tg == &root_task_group)
6514 * Ensure we have at some amount of bandwidth every period. This is
6515 * to prevent reaching a state of large arrears when throttled via
6516 * entity_tick() resulting in prolonged exit starvation.
6518 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6522 * Likewise, bound things on the otherside by preventing insane quota
6523 * periods. This also allows us to normalize in computing quota
6526 if (period > max_cfs_quota_period)
6530 * Prevent race between setting of cfs_rq->runtime_enabled and
6531 * unthrottle_offline_cfs_rqs().
6534 mutex_lock(&cfs_constraints_mutex);
6535 ret = __cfs_schedulable(tg, period, quota);
6539 runtime_enabled = quota != RUNTIME_INF;
6540 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6542 * If we need to toggle cfs_bandwidth_used, off->on must occur
6543 * before making related changes, and on->off must occur afterwards
6545 if (runtime_enabled && !runtime_was_enabled)
6546 cfs_bandwidth_usage_inc();
6547 raw_spin_lock_irq(&cfs_b->lock);
6548 cfs_b->period = ns_to_ktime(period);
6549 cfs_b->quota = quota;
6551 __refill_cfs_bandwidth_runtime(cfs_b);
6553 /* Restart the period timer (if active) to handle new period expiry: */
6554 if (runtime_enabled)
6555 start_cfs_bandwidth(cfs_b);
6557 raw_spin_unlock_irq(&cfs_b->lock);
6559 for_each_online_cpu(i) {
6560 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6561 struct rq *rq = cfs_rq->rq;
6564 rq_lock_irq(rq, &rf);
6565 cfs_rq->runtime_enabled = runtime_enabled;
6566 cfs_rq->runtime_remaining = 0;
6568 if (cfs_rq->throttled)
6569 unthrottle_cfs_rq(cfs_rq);
6570 rq_unlock_irq(rq, &rf);
6572 if (runtime_was_enabled && !runtime_enabled)
6573 cfs_bandwidth_usage_dec();
6575 mutex_unlock(&cfs_constraints_mutex);
6581 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
6585 period = ktime_to_ns(tg->cfs_bandwidth.period);
6586 if (cfs_quota_us < 0)
6587 quota = RUNTIME_INF;
6589 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
6591 return tg_set_cfs_bandwidth(tg, period, quota);
6594 long tg_get_cfs_quota(struct task_group *tg)
6598 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
6601 quota_us = tg->cfs_bandwidth.quota;
6602 do_div(quota_us, NSEC_PER_USEC);
6607 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
6611 period = (u64)cfs_period_us * NSEC_PER_USEC;
6612 quota = tg->cfs_bandwidth.quota;
6614 return tg_set_cfs_bandwidth(tg, period, quota);
6617 long tg_get_cfs_period(struct task_group *tg)
6621 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
6622 do_div(cfs_period_us, NSEC_PER_USEC);
6624 return cfs_period_us;
6627 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
6630 return tg_get_cfs_quota(css_tg(css));
6633 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
6634 struct cftype *cftype, s64 cfs_quota_us)
6636 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
6639 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
6642 return tg_get_cfs_period(css_tg(css));
6645 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
6646 struct cftype *cftype, u64 cfs_period_us)
6648 return tg_set_cfs_period(css_tg(css), cfs_period_us);
6651 struct cfs_schedulable_data {
6652 struct task_group *tg;
6657 * normalize group quota/period to be quota/max_period
6658 * note: units are usecs
6660 static u64 normalize_cfs_quota(struct task_group *tg,
6661 struct cfs_schedulable_data *d)
6669 period = tg_get_cfs_period(tg);
6670 quota = tg_get_cfs_quota(tg);
6673 /* note: these should typically be equivalent */
6674 if (quota == RUNTIME_INF || quota == -1)
6677 return to_ratio(period, quota);
6680 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
6682 struct cfs_schedulable_data *d = data;
6683 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6684 s64 quota = 0, parent_quota = -1;
6687 quota = RUNTIME_INF;
6689 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
6691 quota = normalize_cfs_quota(tg, d);
6692 parent_quota = parent_b->hierarchical_quota;
6695 * Ensure max(child_quota) <= parent_quota. On cgroup2,
6696 * always take the min. On cgroup1, only inherit when no
6699 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
6700 quota = min(quota, parent_quota);
6702 if (quota == RUNTIME_INF)
6703 quota = parent_quota;
6704 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
6708 cfs_b->hierarchical_quota = quota;
6713 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
6716 struct cfs_schedulable_data data = {
6722 if (quota != RUNTIME_INF) {
6723 do_div(data.period, NSEC_PER_USEC);
6724 do_div(data.quota, NSEC_PER_USEC);
6728 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
6734 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
6736 struct task_group *tg = css_tg(seq_css(sf));
6737 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6739 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
6740 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
6741 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
6743 if (schedstat_enabled() && tg != &root_task_group) {
6747 for_each_possible_cpu(i)
6748 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
6750 seq_printf(sf, "wait_sum %llu\n", ws);
6755 #endif /* CONFIG_CFS_BANDWIDTH */
6756 #endif /* CONFIG_FAIR_GROUP_SCHED */
6758 #ifdef CONFIG_RT_GROUP_SCHED
6759 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
6760 struct cftype *cft, s64 val)
6762 return sched_group_set_rt_runtime(css_tg(css), val);
6765 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
6768 return sched_group_rt_runtime(css_tg(css));
6771 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
6772 struct cftype *cftype, u64 rt_period_us)
6774 return sched_group_set_rt_period(css_tg(css), rt_period_us);
6777 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
6780 return sched_group_rt_period(css_tg(css));
6782 #endif /* CONFIG_RT_GROUP_SCHED */
6784 static struct cftype cpu_legacy_files[] = {
6785 #ifdef CONFIG_FAIR_GROUP_SCHED
6788 .read_u64 = cpu_shares_read_u64,
6789 .write_u64 = cpu_shares_write_u64,
6792 #ifdef CONFIG_CFS_BANDWIDTH
6794 .name = "cfs_quota_us",
6795 .read_s64 = cpu_cfs_quota_read_s64,
6796 .write_s64 = cpu_cfs_quota_write_s64,
6799 .name = "cfs_period_us",
6800 .read_u64 = cpu_cfs_period_read_u64,
6801 .write_u64 = cpu_cfs_period_write_u64,
6805 .seq_show = cpu_cfs_stat_show,
6808 #ifdef CONFIG_RT_GROUP_SCHED
6810 .name = "rt_runtime_us",
6811 .read_s64 = cpu_rt_runtime_read,
6812 .write_s64 = cpu_rt_runtime_write,
6815 .name = "rt_period_us",
6816 .read_u64 = cpu_rt_period_read_uint,
6817 .write_u64 = cpu_rt_period_write_uint,
6823 static int cpu_extra_stat_show(struct seq_file *sf,
6824 struct cgroup_subsys_state *css)
6826 #ifdef CONFIG_CFS_BANDWIDTH
6828 struct task_group *tg = css_tg(css);
6829 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6832 throttled_usec = cfs_b->throttled_time;
6833 do_div(throttled_usec, NSEC_PER_USEC);
6835 seq_printf(sf, "nr_periods %d\n"
6837 "throttled_usec %llu\n",
6838 cfs_b->nr_periods, cfs_b->nr_throttled,
6845 #ifdef CONFIG_FAIR_GROUP_SCHED
6846 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
6849 struct task_group *tg = css_tg(css);
6850 u64 weight = scale_load_down(tg->shares);
6852 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
6855 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
6856 struct cftype *cft, u64 weight)
6859 * cgroup weight knobs should use the common MIN, DFL and MAX
6860 * values which are 1, 100 and 10000 respectively. While it loses
6861 * a bit of range on both ends, it maps pretty well onto the shares
6862 * value used by scheduler and the round-trip conversions preserve
6863 * the original value over the entire range.
6865 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
6868 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
6870 return sched_group_set_shares(css_tg(css), scale_load(weight));
6873 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
6876 unsigned long weight = scale_load_down(css_tg(css)->shares);
6877 int last_delta = INT_MAX;
6880 /* find the closest nice value to the current weight */
6881 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
6882 delta = abs(sched_prio_to_weight[prio] - weight);
6883 if (delta >= last_delta)
6888 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
6891 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
6892 struct cftype *cft, s64 nice)
6894 unsigned long weight;
6897 if (nice < MIN_NICE || nice > MAX_NICE)
6900 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
6901 idx = array_index_nospec(idx, 40);
6902 weight = sched_prio_to_weight[idx];
6904 return sched_group_set_shares(css_tg(css), scale_load(weight));
6908 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
6909 long period, long quota)
6912 seq_puts(sf, "max");
6914 seq_printf(sf, "%ld", quota);
6916 seq_printf(sf, " %ld\n", period);
6919 /* caller should put the current value in *@periodp before calling */
6920 static int __maybe_unused cpu_period_quota_parse(char *buf,
6921 u64 *periodp, u64 *quotap)
6923 char tok[21]; /* U64_MAX */
6925 if (!sscanf(buf, "%s %llu", tok, periodp))
6928 *periodp *= NSEC_PER_USEC;
6930 if (sscanf(tok, "%llu", quotap))
6931 *quotap *= NSEC_PER_USEC;
6932 else if (!strcmp(tok, "max"))
6933 *quotap = RUNTIME_INF;
6940 #ifdef CONFIG_CFS_BANDWIDTH
6941 static int cpu_max_show(struct seq_file *sf, void *v)
6943 struct task_group *tg = css_tg(seq_css(sf));
6945 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
6949 static ssize_t cpu_max_write(struct kernfs_open_file *of,
6950 char *buf, size_t nbytes, loff_t off)
6952 struct task_group *tg = css_tg(of_css(of));
6953 u64 period = tg_get_cfs_period(tg);
6957 ret = cpu_period_quota_parse(buf, &period, "a);
6959 ret = tg_set_cfs_bandwidth(tg, period, quota);
6960 return ret ?: nbytes;
6964 static struct cftype cpu_files[] = {
6965 #ifdef CONFIG_FAIR_GROUP_SCHED
6968 .flags = CFTYPE_NOT_ON_ROOT,
6969 .read_u64 = cpu_weight_read_u64,
6970 .write_u64 = cpu_weight_write_u64,
6973 .name = "weight.nice",
6974 .flags = CFTYPE_NOT_ON_ROOT,
6975 .read_s64 = cpu_weight_nice_read_s64,
6976 .write_s64 = cpu_weight_nice_write_s64,
6979 #ifdef CONFIG_CFS_BANDWIDTH
6982 .flags = CFTYPE_NOT_ON_ROOT,
6983 .seq_show = cpu_max_show,
6984 .write = cpu_max_write,
6990 struct cgroup_subsys cpu_cgrp_subsys = {
6991 .css_alloc = cpu_cgroup_css_alloc,
6992 .css_online = cpu_cgroup_css_online,
6993 .css_released = cpu_cgroup_css_released,
6994 .css_free = cpu_cgroup_css_free,
6995 .css_extra_stat_show = cpu_extra_stat_show,
6996 .fork = cpu_cgroup_fork,
6997 .can_attach = cpu_cgroup_can_attach,
6998 .attach = cpu_cgroup_attach,
6999 .legacy_cftypes = cpu_legacy_files,
7000 .dfl_cftypes = cpu_files,
7005 #endif /* CONFIG_CGROUP_SCHED */
7007 void dump_cpu_task(int cpu)
7009 pr_info("Task dump for CPU %d:\n", cpu);
7010 sched_show_task(cpu_curr(cpu));
7014 * Nice levels are multiplicative, with a gentle 10% change for every
7015 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7016 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7017 * that remained on nice 0.
7019 * The "10% effect" is relative and cumulative: from _any_ nice level,
7020 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7021 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7022 * If a task goes up by ~10% and another task goes down by ~10% then
7023 * the relative distance between them is ~25%.)
7025 const int sched_prio_to_weight[40] = {
7026 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7027 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7028 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7029 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7030 /* 0 */ 1024, 820, 655, 526, 423,
7031 /* 5 */ 335, 272, 215, 172, 137,
7032 /* 10 */ 110, 87, 70, 56, 45,
7033 /* 15 */ 36, 29, 23, 18, 15,
7037 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7039 * In cases where the weight does not change often, we can use the
7040 * precalculated inverse to speed up arithmetics by turning divisions
7041 * into multiplications:
7043 const u32 sched_prio_to_wmult[40] = {
7044 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7045 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7046 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7047 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7048 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7049 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7050 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7051 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7054 #undef CREATE_TRACE_POINTS