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(CONFIG_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)
400 * wake_q_add() - queue a wakeup for 'later' waking.
401 * @head: the wake_q_head to add @task to
402 * @task: the task to queue for 'later' wakeup
404 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
405 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
408 * This function must be used as-if it were wake_up_process(); IOW the task
409 * must be ready to be woken at this location.
411 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
413 struct wake_q_node *node = &task->wake_q;
416 * Atomically grab the task, if ->wake_q is !nil already it means
417 * its already queued (either by us or someone else) and will get the
418 * wakeup due to that.
420 * In order to ensure that a pending wakeup will observe our pending
421 * state, even in the failed case, an explicit smp_mb() must be used.
423 smp_mb__before_atomic();
424 if (cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL))
427 get_task_struct(task);
430 * The head is context local, there can be no concurrency.
433 head->lastp = &node->next;
436 void wake_up_q(struct wake_q_head *head)
438 struct wake_q_node *node = head->first;
440 while (node != WAKE_Q_TAIL) {
441 struct task_struct *task;
443 task = container_of(node, struct task_struct, wake_q);
445 /* Task can safely be re-inserted now: */
447 task->wake_q.next = NULL;
450 * wake_up_process() executes a full barrier, which pairs with
451 * the queueing in wake_q_add() so as not to miss wakeups.
453 wake_up_process(task);
454 put_task_struct(task);
459 * resched_curr - mark rq's current task 'to be rescheduled now'.
461 * On UP this means the setting of the need_resched flag, on SMP it
462 * might also involve a cross-CPU call to trigger the scheduler on
465 void resched_curr(struct rq *rq)
467 struct task_struct *curr = rq->curr;
470 lockdep_assert_held(&rq->lock);
472 if (test_tsk_need_resched(curr))
477 if (cpu == smp_processor_id()) {
478 set_tsk_need_resched(curr);
479 set_preempt_need_resched();
483 if (set_nr_and_not_polling(curr))
484 smp_send_reschedule(cpu);
486 trace_sched_wake_idle_without_ipi(cpu);
489 void resched_cpu(int cpu)
491 struct rq *rq = cpu_rq(cpu);
494 raw_spin_lock_irqsave(&rq->lock, flags);
495 if (cpu_online(cpu) || cpu == smp_processor_id())
497 raw_spin_unlock_irqrestore(&rq->lock, flags);
501 #ifdef CONFIG_NO_HZ_COMMON
503 * In the semi idle case, use the nearest busy CPU for migrating timers
504 * from an idle CPU. This is good for power-savings.
506 * We don't do similar optimization for completely idle system, as
507 * selecting an idle CPU will add more delays to the timers than intended
508 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
510 int get_nohz_timer_target(void)
512 int i, cpu = smp_processor_id();
513 struct sched_domain *sd;
515 if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
519 for_each_domain(cpu, sd) {
520 for_each_cpu(i, sched_domain_span(sd)) {
524 if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
531 if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
532 cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
539 * When add_timer_on() enqueues a timer into the timer wheel of an
540 * idle CPU then this timer might expire before the next timer event
541 * which is scheduled to wake up that CPU. In case of a completely
542 * idle system the next event might even be infinite time into the
543 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
544 * leaves the inner idle loop so the newly added timer is taken into
545 * account when the CPU goes back to idle and evaluates the timer
546 * wheel for the next timer event.
548 static void wake_up_idle_cpu(int cpu)
550 struct rq *rq = cpu_rq(cpu);
552 if (cpu == smp_processor_id())
555 if (set_nr_and_not_polling(rq->idle))
556 smp_send_reschedule(cpu);
558 trace_sched_wake_idle_without_ipi(cpu);
561 static bool wake_up_full_nohz_cpu(int cpu)
564 * We just need the target to call irq_exit() and re-evaluate
565 * the next tick. The nohz full kick at least implies that.
566 * If needed we can still optimize that later with an
569 if (cpu_is_offline(cpu))
570 return true; /* Don't try to wake offline CPUs. */
571 if (tick_nohz_full_cpu(cpu)) {
572 if (cpu != smp_processor_id() ||
573 tick_nohz_tick_stopped())
574 tick_nohz_full_kick_cpu(cpu);
582 * Wake up the specified CPU. If the CPU is going offline, it is the
583 * caller's responsibility to deal with the lost wakeup, for example,
584 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
586 void wake_up_nohz_cpu(int cpu)
588 if (!wake_up_full_nohz_cpu(cpu))
589 wake_up_idle_cpu(cpu);
592 static inline bool got_nohz_idle_kick(void)
594 int cpu = smp_processor_id();
596 if (!(atomic_read(nohz_flags(cpu)) & NOHZ_KICK_MASK))
599 if (idle_cpu(cpu) && !need_resched())
603 * We can't run Idle Load Balance on this CPU for this time so we
604 * cancel it and clear NOHZ_BALANCE_KICK
606 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
610 #else /* CONFIG_NO_HZ_COMMON */
612 static inline bool got_nohz_idle_kick(void)
617 #endif /* CONFIG_NO_HZ_COMMON */
619 #ifdef CONFIG_NO_HZ_FULL
620 bool sched_can_stop_tick(struct rq *rq)
624 /* Deadline tasks, even if single, need the tick */
625 if (rq->dl.dl_nr_running)
629 * If there are more than one RR tasks, we need the tick to effect the
630 * actual RR behaviour.
632 if (rq->rt.rr_nr_running) {
633 if (rq->rt.rr_nr_running == 1)
640 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
641 * forced preemption between FIFO tasks.
643 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
648 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
649 * if there's more than one we need the tick for involuntary
652 if (rq->nr_running > 1)
657 #endif /* CONFIG_NO_HZ_FULL */
658 #endif /* CONFIG_SMP */
660 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
661 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
663 * Iterate task_group tree rooted at *from, calling @down when first entering a
664 * node and @up when leaving it for the final time.
666 * Caller must hold rcu_lock or sufficient equivalent.
668 int walk_tg_tree_from(struct task_group *from,
669 tg_visitor down, tg_visitor up, void *data)
671 struct task_group *parent, *child;
677 ret = (*down)(parent, data);
680 list_for_each_entry_rcu(child, &parent->children, siblings) {
687 ret = (*up)(parent, data);
688 if (ret || parent == from)
692 parent = parent->parent;
699 int tg_nop(struct task_group *tg, void *data)
705 static void set_load_weight(struct task_struct *p, bool update_load)
707 int prio = p->static_prio - MAX_RT_PRIO;
708 struct load_weight *load = &p->se.load;
711 * SCHED_IDLE tasks get minimal weight:
713 if (task_has_idle_policy(p)) {
714 load->weight = scale_load(WEIGHT_IDLEPRIO);
715 load->inv_weight = WMULT_IDLEPRIO;
716 p->se.runnable_weight = load->weight;
721 * SCHED_OTHER tasks have to update their load when changing their
724 if (update_load && p->sched_class == &fair_sched_class) {
725 reweight_task(p, prio);
727 load->weight = scale_load(sched_prio_to_weight[prio]);
728 load->inv_weight = sched_prio_to_wmult[prio];
729 p->se.runnable_weight = load->weight;
733 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
735 if (!(flags & ENQUEUE_NOCLOCK))
738 if (!(flags & ENQUEUE_RESTORE)) {
739 sched_info_queued(rq, p);
740 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
743 p->sched_class->enqueue_task(rq, p, flags);
746 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
748 if (!(flags & DEQUEUE_NOCLOCK))
751 if (!(flags & DEQUEUE_SAVE)) {
752 sched_info_dequeued(rq, p);
753 psi_dequeue(p, flags & DEQUEUE_SLEEP);
756 p->sched_class->dequeue_task(rq, p, flags);
759 void activate_task(struct rq *rq, struct task_struct *p, int flags)
761 if (task_contributes_to_load(p))
762 rq->nr_uninterruptible--;
764 enqueue_task(rq, p, flags);
767 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
769 if (task_contributes_to_load(p))
770 rq->nr_uninterruptible++;
772 dequeue_task(rq, p, flags);
776 * __normal_prio - return the priority that is based on the static prio
778 static inline int __normal_prio(struct task_struct *p)
780 return p->static_prio;
784 * Calculate the expected normal priority: i.e. priority
785 * without taking RT-inheritance into account. Might be
786 * boosted by interactivity modifiers. Changes upon fork,
787 * setprio syscalls, and whenever the interactivity
788 * estimator recalculates.
790 static inline int normal_prio(struct task_struct *p)
794 if (task_has_dl_policy(p))
795 prio = MAX_DL_PRIO-1;
796 else if (task_has_rt_policy(p))
797 prio = MAX_RT_PRIO-1 - p->rt_priority;
799 prio = __normal_prio(p);
804 * Calculate the current priority, i.e. the priority
805 * taken into account by the scheduler. This value might
806 * be boosted by RT tasks, or might be boosted by
807 * interactivity modifiers. Will be RT if the task got
808 * RT-boosted. If not then it returns p->normal_prio.
810 static int effective_prio(struct task_struct *p)
812 p->normal_prio = normal_prio(p);
814 * If we are RT tasks or we were boosted to RT priority,
815 * keep the priority unchanged. Otherwise, update priority
816 * to the normal priority:
818 if (!rt_prio(p->prio))
819 return p->normal_prio;
824 * task_curr - is this task currently executing on a CPU?
825 * @p: the task in question.
827 * Return: 1 if the task is currently executing. 0 otherwise.
829 inline int task_curr(const struct task_struct *p)
831 return cpu_curr(task_cpu(p)) == p;
835 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
836 * use the balance_callback list if you want balancing.
838 * this means any call to check_class_changed() must be followed by a call to
839 * balance_callback().
841 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
842 const struct sched_class *prev_class,
845 if (prev_class != p->sched_class) {
846 if (prev_class->switched_from)
847 prev_class->switched_from(rq, p);
849 p->sched_class->switched_to(rq, p);
850 } else if (oldprio != p->prio || dl_task(p))
851 p->sched_class->prio_changed(rq, p, oldprio);
854 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
856 const struct sched_class *class;
858 if (p->sched_class == rq->curr->sched_class) {
859 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
861 for_each_class(class) {
862 if (class == rq->curr->sched_class)
864 if (class == p->sched_class) {
872 * A queue event has occurred, and we're going to schedule. In
873 * this case, we can save a useless back to back clock update.
875 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
876 rq_clock_skip_update(rq);
881 static inline bool is_per_cpu_kthread(struct task_struct *p)
883 if (!(p->flags & PF_KTHREAD))
886 if (p->nr_cpus_allowed != 1)
893 * Per-CPU kthreads are allowed to run on !actie && online CPUs, see
894 * __set_cpus_allowed_ptr() and select_fallback_rq().
896 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
898 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
901 if (is_per_cpu_kthread(p))
902 return cpu_online(cpu);
904 return cpu_active(cpu);
908 * This is how migration works:
910 * 1) we invoke migration_cpu_stop() on the target CPU using
912 * 2) stopper starts to run (implicitly forcing the migrated thread
914 * 3) it checks whether the migrated task is still in the wrong runqueue.
915 * 4) if it's in the wrong runqueue then the migration thread removes
916 * it and puts it into the right queue.
917 * 5) stopper completes and stop_one_cpu() returns and the migration
922 * move_queued_task - move a queued task to new rq.
924 * Returns (locked) new rq. Old rq's lock is released.
926 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
927 struct task_struct *p, int new_cpu)
929 lockdep_assert_held(&rq->lock);
931 p->on_rq = TASK_ON_RQ_MIGRATING;
932 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
933 set_task_cpu(p, new_cpu);
936 rq = cpu_rq(new_cpu);
939 BUG_ON(task_cpu(p) != new_cpu);
940 enqueue_task(rq, p, 0);
941 p->on_rq = TASK_ON_RQ_QUEUED;
942 check_preempt_curr(rq, p, 0);
947 struct migration_arg {
948 struct task_struct *task;
953 * Move (not current) task off this CPU, onto the destination CPU. We're doing
954 * this because either it can't run here any more (set_cpus_allowed()
955 * away from this CPU, or CPU going down), or because we're
956 * attempting to rebalance this task on exec (sched_exec).
958 * So we race with normal scheduler movements, but that's OK, as long
959 * as the task is no longer on this CPU.
961 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
962 struct task_struct *p, int dest_cpu)
964 /* Affinity changed (again). */
965 if (!is_cpu_allowed(p, dest_cpu))
969 rq = move_queued_task(rq, rf, p, dest_cpu);
975 * migration_cpu_stop - this will be executed by a highprio stopper thread
976 * and performs thread migration by bumping thread off CPU then
977 * 'pushing' onto another runqueue.
979 static int migration_cpu_stop(void *data)
981 struct migration_arg *arg = data;
982 struct task_struct *p = arg->task;
983 struct rq *rq = this_rq();
987 * The original target CPU might have gone down and we might
988 * be on another CPU but it doesn't matter.
992 * We need to explicitly wake pending tasks before running
993 * __migrate_task() such that we will not miss enforcing cpus_allowed
994 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
996 sched_ttwu_pending();
998 raw_spin_lock(&p->pi_lock);
1001 * If task_rq(p) != rq, it cannot be migrated here, because we're
1002 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1003 * we're holding p->pi_lock.
1005 if (task_rq(p) == rq) {
1006 if (task_on_rq_queued(p))
1007 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1009 p->wake_cpu = arg->dest_cpu;
1012 raw_spin_unlock(&p->pi_lock);
1019 * sched_class::set_cpus_allowed must do the below, but is not required to
1020 * actually call this function.
1022 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1024 cpumask_copy(&p->cpus_allowed, new_mask);
1025 p->nr_cpus_allowed = cpumask_weight(new_mask);
1028 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1030 struct rq *rq = task_rq(p);
1031 bool queued, running;
1033 lockdep_assert_held(&p->pi_lock);
1035 queued = task_on_rq_queued(p);
1036 running = task_current(rq, p);
1040 * Because __kthread_bind() calls this on blocked tasks without
1043 lockdep_assert_held(&rq->lock);
1044 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1047 put_prev_task(rq, p);
1049 p->sched_class->set_cpus_allowed(p, new_mask);
1052 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1054 set_curr_task(rq, p);
1058 * Change a given task's CPU affinity. Migrate the thread to a
1059 * proper CPU and schedule it away if the CPU it's executing on
1060 * is removed from the allowed bitmask.
1062 * NOTE: the caller must have a valid reference to the task, the
1063 * task must not exit() & deallocate itself prematurely. The
1064 * call is not atomic; no spinlocks may be held.
1066 static int __set_cpus_allowed_ptr(struct task_struct *p,
1067 const struct cpumask *new_mask, bool check)
1069 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1070 unsigned int dest_cpu;
1075 rq = task_rq_lock(p, &rf);
1076 update_rq_clock(rq);
1078 if (p->flags & PF_KTHREAD) {
1080 * Kernel threads are allowed on online && !active CPUs
1082 cpu_valid_mask = cpu_online_mask;
1086 * Must re-check here, to close a race against __kthread_bind(),
1087 * sched_setaffinity() is not guaranteed to observe the flag.
1089 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1094 if (cpumask_equal(&p->cpus_allowed, new_mask))
1097 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1102 do_set_cpus_allowed(p, new_mask);
1104 if (p->flags & PF_KTHREAD) {
1106 * For kernel threads that do indeed end up on online &&
1107 * !active we want to ensure they are strict per-CPU threads.
1109 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1110 !cpumask_intersects(new_mask, cpu_active_mask) &&
1111 p->nr_cpus_allowed != 1);
1114 /* Can the task run on the task's current CPU? If so, we're done */
1115 if (cpumask_test_cpu(task_cpu(p), new_mask))
1118 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1119 if (task_running(rq, p) || p->state == TASK_WAKING) {
1120 struct migration_arg arg = { p, dest_cpu };
1121 /* Need help from migration thread: drop lock and wait. */
1122 task_rq_unlock(rq, p, &rf);
1123 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1124 tlb_migrate_finish(p->mm);
1126 } else if (task_on_rq_queued(p)) {
1128 * OK, since we're going to drop the lock immediately
1129 * afterwards anyway.
1131 rq = move_queued_task(rq, &rf, p, dest_cpu);
1134 task_rq_unlock(rq, p, &rf);
1139 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1141 return __set_cpus_allowed_ptr(p, new_mask, false);
1143 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1145 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1147 #ifdef CONFIG_SCHED_DEBUG
1149 * We should never call set_task_cpu() on a blocked task,
1150 * ttwu() will sort out the placement.
1152 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1156 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1157 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1158 * time relying on p->on_rq.
1160 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1161 p->sched_class == &fair_sched_class &&
1162 (p->on_rq && !task_on_rq_migrating(p)));
1164 #ifdef CONFIG_LOCKDEP
1166 * The caller should hold either p->pi_lock or rq->lock, when changing
1167 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1169 * sched_move_task() holds both and thus holding either pins the cgroup,
1172 * Furthermore, all task_rq users should acquire both locks, see
1175 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1176 lockdep_is_held(&task_rq(p)->lock)));
1179 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1181 WARN_ON_ONCE(!cpu_online(new_cpu));
1184 trace_sched_migrate_task(p, new_cpu);
1186 if (task_cpu(p) != new_cpu) {
1187 if (p->sched_class->migrate_task_rq)
1188 p->sched_class->migrate_task_rq(p, new_cpu);
1189 p->se.nr_migrations++;
1191 perf_event_task_migrate(p);
1194 __set_task_cpu(p, new_cpu);
1197 #ifdef CONFIG_NUMA_BALANCING
1198 static void __migrate_swap_task(struct task_struct *p, int cpu)
1200 if (task_on_rq_queued(p)) {
1201 struct rq *src_rq, *dst_rq;
1202 struct rq_flags srf, drf;
1204 src_rq = task_rq(p);
1205 dst_rq = cpu_rq(cpu);
1207 rq_pin_lock(src_rq, &srf);
1208 rq_pin_lock(dst_rq, &drf);
1210 p->on_rq = TASK_ON_RQ_MIGRATING;
1211 deactivate_task(src_rq, p, 0);
1212 set_task_cpu(p, cpu);
1213 activate_task(dst_rq, p, 0);
1214 p->on_rq = TASK_ON_RQ_QUEUED;
1215 check_preempt_curr(dst_rq, p, 0);
1217 rq_unpin_lock(dst_rq, &drf);
1218 rq_unpin_lock(src_rq, &srf);
1222 * Task isn't running anymore; make it appear like we migrated
1223 * it before it went to sleep. This means on wakeup we make the
1224 * previous CPU our target instead of where it really is.
1230 struct migration_swap_arg {
1231 struct task_struct *src_task, *dst_task;
1232 int src_cpu, dst_cpu;
1235 static int migrate_swap_stop(void *data)
1237 struct migration_swap_arg *arg = data;
1238 struct rq *src_rq, *dst_rq;
1241 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1244 src_rq = cpu_rq(arg->src_cpu);
1245 dst_rq = cpu_rq(arg->dst_cpu);
1247 double_raw_lock(&arg->src_task->pi_lock,
1248 &arg->dst_task->pi_lock);
1249 double_rq_lock(src_rq, dst_rq);
1251 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1254 if (task_cpu(arg->src_task) != arg->src_cpu)
1257 if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1260 if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1263 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1264 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1269 double_rq_unlock(src_rq, dst_rq);
1270 raw_spin_unlock(&arg->dst_task->pi_lock);
1271 raw_spin_unlock(&arg->src_task->pi_lock);
1277 * Cross migrate two tasks
1279 int migrate_swap(struct task_struct *cur, struct task_struct *p,
1280 int target_cpu, int curr_cpu)
1282 struct migration_swap_arg arg;
1285 arg = (struct migration_swap_arg){
1287 .src_cpu = curr_cpu,
1289 .dst_cpu = target_cpu,
1292 if (arg.src_cpu == arg.dst_cpu)
1296 * These three tests are all lockless; this is OK since all of them
1297 * will be re-checked with proper locks held further down the line.
1299 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1302 if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1305 if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1308 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1309 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1314 #endif /* CONFIG_NUMA_BALANCING */
1317 * wait_task_inactive - wait for a thread to unschedule.
1319 * If @match_state is nonzero, it's the @p->state value just checked and
1320 * not expected to change. If it changes, i.e. @p might have woken up,
1321 * then return zero. When we succeed in waiting for @p to be off its CPU,
1322 * we return a positive number (its total switch count). If a second call
1323 * a short while later returns the same number, the caller can be sure that
1324 * @p has remained unscheduled the whole time.
1326 * The caller must ensure that the task *will* unschedule sometime soon,
1327 * else this function might spin for a *long* time. This function can't
1328 * be called with interrupts off, or it may introduce deadlock with
1329 * smp_call_function() if an IPI is sent by the same process we are
1330 * waiting to become inactive.
1332 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1334 int running, queued;
1341 * We do the initial early heuristics without holding
1342 * any task-queue locks at all. We'll only try to get
1343 * the runqueue lock when things look like they will
1349 * If the task is actively running on another CPU
1350 * still, just relax and busy-wait without holding
1353 * NOTE! Since we don't hold any locks, it's not
1354 * even sure that "rq" stays as the right runqueue!
1355 * But we don't care, since "task_running()" will
1356 * return false if the runqueue has changed and p
1357 * is actually now running somewhere else!
1359 while (task_running(rq, p)) {
1360 if (match_state && unlikely(p->state != match_state))
1366 * Ok, time to look more closely! We need the rq
1367 * lock now, to be *sure*. If we're wrong, we'll
1368 * just go back and repeat.
1370 rq = task_rq_lock(p, &rf);
1371 trace_sched_wait_task(p);
1372 running = task_running(rq, p);
1373 queued = task_on_rq_queued(p);
1375 if (!match_state || p->state == match_state)
1376 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1377 task_rq_unlock(rq, p, &rf);
1380 * If it changed from the expected state, bail out now.
1382 if (unlikely(!ncsw))
1386 * Was it really running after all now that we
1387 * checked with the proper locks actually held?
1389 * Oops. Go back and try again..
1391 if (unlikely(running)) {
1397 * It's not enough that it's not actively running,
1398 * it must be off the runqueue _entirely_, and not
1401 * So if it was still runnable (but just not actively
1402 * running right now), it's preempted, and we should
1403 * yield - it could be a while.
1405 if (unlikely(queued)) {
1406 ktime_t to = NSEC_PER_SEC / HZ;
1408 set_current_state(TASK_UNINTERRUPTIBLE);
1409 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1414 * Ahh, all good. It wasn't running, and it wasn't
1415 * runnable, which means that it will never become
1416 * running in the future either. We're all done!
1425 * kick_process - kick a running thread to enter/exit the kernel
1426 * @p: the to-be-kicked thread
1428 * Cause a process which is running on another CPU to enter
1429 * kernel-mode, without any delay. (to get signals handled.)
1431 * NOTE: this function doesn't have to take the runqueue lock,
1432 * because all it wants to ensure is that the remote task enters
1433 * the kernel. If the IPI races and the task has been migrated
1434 * to another CPU then no harm is done and the purpose has been
1437 void kick_process(struct task_struct *p)
1443 if ((cpu != smp_processor_id()) && task_curr(p))
1444 smp_send_reschedule(cpu);
1447 EXPORT_SYMBOL_GPL(kick_process);
1450 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1452 * A few notes on cpu_active vs cpu_online:
1454 * - cpu_active must be a subset of cpu_online
1456 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
1457 * see __set_cpus_allowed_ptr(). At this point the newly online
1458 * CPU isn't yet part of the sched domains, and balancing will not
1461 * - on CPU-down we clear cpu_active() to mask the sched domains and
1462 * avoid the load balancer to place new tasks on the to be removed
1463 * CPU. Existing tasks will remain running there and will be taken
1466 * This means that fallback selection must not select !active CPUs.
1467 * And can assume that any active CPU must be online. Conversely
1468 * select_task_rq() below may allow selection of !active CPUs in order
1469 * to satisfy the above rules.
1471 static int select_fallback_rq(int cpu, struct task_struct *p)
1473 int nid = cpu_to_node(cpu);
1474 const struct cpumask *nodemask = NULL;
1475 enum { cpuset, possible, fail } state = cpuset;
1479 * If the node that the CPU is on has been offlined, cpu_to_node()
1480 * will return -1. There is no CPU on the node, and we should
1481 * select the CPU on the other node.
1484 nodemask = cpumask_of_node(nid);
1486 /* Look for allowed, online CPU in same node. */
1487 for_each_cpu(dest_cpu, nodemask) {
1488 if (!cpu_active(dest_cpu))
1490 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1496 /* Any allowed, online CPU? */
1497 for_each_cpu(dest_cpu, &p->cpus_allowed) {
1498 if (!is_cpu_allowed(p, dest_cpu))
1504 /* No more Mr. Nice Guy. */
1507 if (IS_ENABLED(CONFIG_CPUSETS)) {
1508 cpuset_cpus_allowed_fallback(p);
1514 do_set_cpus_allowed(p, cpu_possible_mask);
1525 if (state != cpuset) {
1527 * Don't tell them about moving exiting tasks or
1528 * kernel threads (both mm NULL), since they never
1531 if (p->mm && printk_ratelimit()) {
1532 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1533 task_pid_nr(p), p->comm, cpu);
1541 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1544 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1546 lockdep_assert_held(&p->pi_lock);
1548 if (p->nr_cpus_allowed > 1)
1549 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1551 cpu = cpumask_any(&p->cpus_allowed);
1554 * In order not to call set_task_cpu() on a blocking task we need
1555 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1558 * Since this is common to all placement strategies, this lives here.
1560 * [ this allows ->select_task() to simply return task_cpu(p) and
1561 * not worry about this generic constraint ]
1563 if (unlikely(!is_cpu_allowed(p, cpu)))
1564 cpu = select_fallback_rq(task_cpu(p), p);
1569 static void update_avg(u64 *avg, u64 sample)
1571 s64 diff = sample - *avg;
1575 void sched_set_stop_task(int cpu, struct task_struct *stop)
1577 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1578 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1582 * Make it appear like a SCHED_FIFO task, its something
1583 * userspace knows about and won't get confused about.
1585 * Also, it will make PI more or less work without too
1586 * much confusion -- but then, stop work should not
1587 * rely on PI working anyway.
1589 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1591 stop->sched_class = &stop_sched_class;
1594 cpu_rq(cpu)->stop = stop;
1598 * Reset it back to a normal scheduling class so that
1599 * it can die in pieces.
1601 old_stop->sched_class = &rt_sched_class;
1607 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1608 const struct cpumask *new_mask, bool check)
1610 return set_cpus_allowed_ptr(p, new_mask);
1613 #endif /* CONFIG_SMP */
1616 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1620 if (!schedstat_enabled())
1626 if (cpu == rq->cpu) {
1627 __schedstat_inc(rq->ttwu_local);
1628 __schedstat_inc(p->se.statistics.nr_wakeups_local);
1630 struct sched_domain *sd;
1632 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
1634 for_each_domain(rq->cpu, sd) {
1635 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1636 __schedstat_inc(sd->ttwu_wake_remote);
1643 if (wake_flags & WF_MIGRATED)
1644 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1645 #endif /* CONFIG_SMP */
1647 __schedstat_inc(rq->ttwu_count);
1648 __schedstat_inc(p->se.statistics.nr_wakeups);
1650 if (wake_flags & WF_SYNC)
1651 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
1654 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1656 activate_task(rq, p, en_flags);
1657 p->on_rq = TASK_ON_RQ_QUEUED;
1659 /* If a worker is waking up, notify the workqueue: */
1660 if (p->flags & PF_WQ_WORKER)
1661 wq_worker_waking_up(p, cpu_of(rq));
1665 * Mark the task runnable and perform wakeup-preemption.
1667 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1668 struct rq_flags *rf)
1670 check_preempt_curr(rq, p, wake_flags);
1671 p->state = TASK_RUNNING;
1672 trace_sched_wakeup(p);
1675 if (p->sched_class->task_woken) {
1677 * Our task @p is fully woken up and running; so its safe to
1678 * drop the rq->lock, hereafter rq is only used for statistics.
1680 rq_unpin_lock(rq, rf);
1681 p->sched_class->task_woken(rq, p);
1682 rq_repin_lock(rq, rf);
1685 if (rq->idle_stamp) {
1686 u64 delta = rq_clock(rq) - rq->idle_stamp;
1687 u64 max = 2*rq->max_idle_balance_cost;
1689 update_avg(&rq->avg_idle, delta);
1691 if (rq->avg_idle > max)
1700 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1701 struct rq_flags *rf)
1703 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
1705 lockdep_assert_held(&rq->lock);
1708 if (p->sched_contributes_to_load)
1709 rq->nr_uninterruptible--;
1711 if (wake_flags & WF_MIGRATED)
1712 en_flags |= ENQUEUE_MIGRATED;
1715 ttwu_activate(rq, p, en_flags);
1716 ttwu_do_wakeup(rq, p, wake_flags, rf);
1720 * Called in case the task @p isn't fully descheduled from its runqueue,
1721 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1722 * since all we need to do is flip p->state to TASK_RUNNING, since
1723 * the task is still ->on_rq.
1725 static int ttwu_remote(struct task_struct *p, int wake_flags)
1731 rq = __task_rq_lock(p, &rf);
1732 if (task_on_rq_queued(p)) {
1733 /* check_preempt_curr() may use rq clock */
1734 update_rq_clock(rq);
1735 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1738 __task_rq_unlock(rq, &rf);
1744 void sched_ttwu_pending(void)
1746 struct rq *rq = this_rq();
1747 struct llist_node *llist = llist_del_all(&rq->wake_list);
1748 struct task_struct *p, *t;
1754 rq_lock_irqsave(rq, &rf);
1755 update_rq_clock(rq);
1757 llist_for_each_entry_safe(p, t, llist, wake_entry)
1758 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
1760 rq_unlock_irqrestore(rq, &rf);
1763 void scheduler_ipi(void)
1766 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1767 * TIF_NEED_RESCHED remotely (for the first time) will also send
1770 preempt_fold_need_resched();
1772 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1776 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1777 * traditionally all their work was done from the interrupt return
1778 * path. Now that we actually do some work, we need to make sure
1781 * Some archs already do call them, luckily irq_enter/exit nest
1784 * Arguably we should visit all archs and update all handlers,
1785 * however a fair share of IPIs are still resched only so this would
1786 * somewhat pessimize the simple resched case.
1789 sched_ttwu_pending();
1792 * Check if someone kicked us for doing the nohz idle load balance.
1794 if (unlikely(got_nohz_idle_kick())) {
1795 this_rq()->idle_balance = 1;
1796 raise_softirq_irqoff(SCHED_SOFTIRQ);
1801 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1803 struct rq *rq = cpu_rq(cpu);
1805 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1807 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1808 if (!set_nr_if_polling(rq->idle))
1809 smp_send_reschedule(cpu);
1811 trace_sched_wake_idle_without_ipi(cpu);
1815 void wake_up_if_idle(int cpu)
1817 struct rq *rq = cpu_rq(cpu);
1822 if (!is_idle_task(rcu_dereference(rq->curr)))
1825 if (set_nr_if_polling(rq->idle)) {
1826 trace_sched_wake_idle_without_ipi(cpu);
1828 rq_lock_irqsave(rq, &rf);
1829 if (is_idle_task(rq->curr))
1830 smp_send_reschedule(cpu);
1831 /* Else CPU is not idle, do nothing here: */
1832 rq_unlock_irqrestore(rq, &rf);
1839 bool cpus_share_cache(int this_cpu, int that_cpu)
1841 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1843 #endif /* CONFIG_SMP */
1845 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1847 struct rq *rq = cpu_rq(cpu);
1850 #if defined(CONFIG_SMP)
1851 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1852 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1853 ttwu_queue_remote(p, cpu, wake_flags);
1859 update_rq_clock(rq);
1860 ttwu_do_activate(rq, p, wake_flags, &rf);
1865 * Notes on Program-Order guarantees on SMP systems.
1869 * The basic program-order guarantee on SMP systems is that when a task [t]
1870 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1871 * execution on its new CPU [c1].
1873 * For migration (of runnable tasks) this is provided by the following means:
1875 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1876 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1877 * rq(c1)->lock (if not at the same time, then in that order).
1878 * C) LOCK of the rq(c1)->lock scheduling in task
1880 * Release/acquire chaining guarantees that B happens after A and C after B.
1881 * Note: the CPU doing B need not be c0 or c1
1890 * UNLOCK rq(0)->lock
1892 * LOCK rq(0)->lock // orders against CPU0
1894 * UNLOCK rq(0)->lock
1898 * UNLOCK rq(1)->lock
1900 * LOCK rq(1)->lock // orders against CPU2
1903 * UNLOCK rq(1)->lock
1906 * BLOCKING -- aka. SLEEP + WAKEUP
1908 * For blocking we (obviously) need to provide the same guarantee as for
1909 * migration. However the means are completely different as there is no lock
1910 * chain to provide order. Instead we do:
1912 * 1) smp_store_release(X->on_cpu, 0)
1913 * 2) smp_cond_load_acquire(!X->on_cpu)
1917 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1919 * LOCK rq(0)->lock LOCK X->pi_lock
1922 * smp_store_release(X->on_cpu, 0);
1924 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1930 * X->state = RUNNING
1931 * UNLOCK rq(2)->lock
1933 * LOCK rq(2)->lock // orders against CPU1
1936 * UNLOCK rq(2)->lock
1939 * UNLOCK rq(0)->lock
1942 * However, for wakeups there is a second guarantee we must provide, namely we
1943 * must ensure that CONDITION=1 done by the caller can not be reordered with
1944 * accesses to the task state; see try_to_wake_up() and set_current_state().
1948 * try_to_wake_up - wake up a thread
1949 * @p: the thread to be awakened
1950 * @state: the mask of task states that can be woken
1951 * @wake_flags: wake modifier flags (WF_*)
1953 * If (@state & @p->state) @p->state = TASK_RUNNING.
1955 * If the task was not queued/runnable, also place it back on a runqueue.
1957 * Atomic against schedule() which would dequeue a task, also see
1958 * set_current_state().
1960 * This function executes a full memory barrier before accessing the task
1961 * state; see set_current_state().
1963 * Return: %true if @p->state changes (an actual wakeup was done),
1967 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1969 unsigned long flags;
1970 int cpu, success = 0;
1973 * If we are going to wake up a thread waiting for CONDITION we
1974 * need to ensure that CONDITION=1 done by the caller can not be
1975 * reordered with p->state check below. This pairs with mb() in
1976 * set_current_state() the waiting thread does.
1978 raw_spin_lock_irqsave(&p->pi_lock, flags);
1979 smp_mb__after_spinlock();
1980 if (!(p->state & state))
1983 trace_sched_waking(p);
1985 /* We're going to change ->state: */
1990 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1991 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1992 * in smp_cond_load_acquire() below.
1994 * sched_ttwu_pending() try_to_wake_up()
1995 * STORE p->on_rq = 1 LOAD p->state
1998 * __schedule() (switch to task 'p')
1999 * LOCK rq->lock smp_rmb();
2000 * smp_mb__after_spinlock();
2004 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2006 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2007 * __schedule(). See the comment for smp_mb__after_spinlock().
2010 if (p->on_rq && ttwu_remote(p, wake_flags))
2015 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2016 * possible to, falsely, observe p->on_cpu == 0.
2018 * One must be running (->on_cpu == 1) in order to remove oneself
2019 * from the runqueue.
2021 * __schedule() (switch to task 'p') try_to_wake_up()
2022 * STORE p->on_cpu = 1 LOAD p->on_rq
2025 * __schedule() (put 'p' to sleep)
2026 * LOCK rq->lock smp_rmb();
2027 * smp_mb__after_spinlock();
2028 * STORE p->on_rq = 0 LOAD p->on_cpu
2030 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2031 * __schedule(). See the comment for smp_mb__after_spinlock().
2036 * If the owning (remote) CPU is still in the middle of schedule() with
2037 * this task as prev, wait until its done referencing the task.
2039 * Pairs with the smp_store_release() in finish_task().
2041 * This ensures that tasks getting woken will be fully ordered against
2042 * their previous state and preserve Program Order.
2044 smp_cond_load_acquire(&p->on_cpu, !VAL);
2046 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2047 p->state = TASK_WAKING;
2050 delayacct_blkio_end(p);
2051 atomic_dec(&task_rq(p)->nr_iowait);
2054 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2055 if (task_cpu(p) != cpu) {
2056 wake_flags |= WF_MIGRATED;
2057 psi_ttwu_dequeue(p);
2058 set_task_cpu(p, cpu);
2061 #else /* CONFIG_SMP */
2064 delayacct_blkio_end(p);
2065 atomic_dec(&task_rq(p)->nr_iowait);
2068 #endif /* CONFIG_SMP */
2070 ttwu_queue(p, cpu, wake_flags);
2072 ttwu_stat(p, cpu, wake_flags);
2074 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2080 * try_to_wake_up_local - try to wake up a local task with rq lock held
2081 * @p: the thread to be awakened
2082 * @rf: request-queue flags for pinning
2084 * Put @p on the run-queue if it's not already there. The caller must
2085 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2088 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2090 struct rq *rq = task_rq(p);
2092 if (WARN_ON_ONCE(rq != this_rq()) ||
2093 WARN_ON_ONCE(p == current))
2096 lockdep_assert_held(&rq->lock);
2098 if (!raw_spin_trylock(&p->pi_lock)) {
2100 * This is OK, because current is on_cpu, which avoids it being
2101 * picked for load-balance and preemption/IRQs are still
2102 * disabled avoiding further scheduler activity on it and we've
2103 * not yet picked a replacement task.
2106 raw_spin_lock(&p->pi_lock);
2110 if (!(p->state & TASK_NORMAL))
2113 trace_sched_waking(p);
2115 if (!task_on_rq_queued(p)) {
2117 delayacct_blkio_end(p);
2118 atomic_dec(&rq->nr_iowait);
2120 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK);
2123 ttwu_do_wakeup(rq, p, 0, rf);
2124 ttwu_stat(p, smp_processor_id(), 0);
2126 raw_spin_unlock(&p->pi_lock);
2130 * wake_up_process - Wake up a specific process
2131 * @p: The process to be woken up.
2133 * Attempt to wake up the nominated process and move it to the set of runnable
2136 * Return: 1 if the process was woken up, 0 if it was already running.
2138 * This function executes a full memory barrier before accessing the task state.
2140 int wake_up_process(struct task_struct *p)
2142 return try_to_wake_up(p, TASK_NORMAL, 0);
2144 EXPORT_SYMBOL(wake_up_process);
2146 int wake_up_state(struct task_struct *p, unsigned int state)
2148 return try_to_wake_up(p, state, 0);
2152 * Perform scheduler related setup for a newly forked process p.
2153 * p is forked by current.
2155 * __sched_fork() is basic setup used by init_idle() too:
2157 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2162 p->se.exec_start = 0;
2163 p->se.sum_exec_runtime = 0;
2164 p->se.prev_sum_exec_runtime = 0;
2165 p->se.nr_migrations = 0;
2167 INIT_LIST_HEAD(&p->se.group_node);
2169 #ifdef CONFIG_FAIR_GROUP_SCHED
2170 p->se.cfs_rq = NULL;
2173 #ifdef CONFIG_SCHEDSTATS
2174 /* Even if schedstat is disabled, there should not be garbage */
2175 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2178 RB_CLEAR_NODE(&p->dl.rb_node);
2179 init_dl_task_timer(&p->dl);
2180 init_dl_inactive_task_timer(&p->dl);
2181 __dl_clear_params(p);
2183 INIT_LIST_HEAD(&p->rt.run_list);
2185 p->rt.time_slice = sched_rr_timeslice;
2189 #ifdef CONFIG_PREEMPT_NOTIFIERS
2190 INIT_HLIST_HEAD(&p->preempt_notifiers);
2193 init_numa_balancing(clone_flags, p);
2196 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2198 #ifdef CONFIG_NUMA_BALANCING
2200 void set_numabalancing_state(bool enabled)
2203 static_branch_enable(&sched_numa_balancing);
2205 static_branch_disable(&sched_numa_balancing);
2208 #ifdef CONFIG_PROC_SYSCTL
2209 int sysctl_numa_balancing(struct ctl_table *table, int write,
2210 void __user *buffer, size_t *lenp, loff_t *ppos)
2214 int state = static_branch_likely(&sched_numa_balancing);
2216 if (write && !capable(CAP_SYS_ADMIN))
2221 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2225 set_numabalancing_state(state);
2231 #ifdef CONFIG_SCHEDSTATS
2233 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2234 static bool __initdata __sched_schedstats = false;
2236 static void set_schedstats(bool enabled)
2239 static_branch_enable(&sched_schedstats);
2241 static_branch_disable(&sched_schedstats);
2244 void force_schedstat_enabled(void)
2246 if (!schedstat_enabled()) {
2247 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2248 static_branch_enable(&sched_schedstats);
2252 static int __init setup_schedstats(char *str)
2259 * This code is called before jump labels have been set up, so we can't
2260 * change the static branch directly just yet. Instead set a temporary
2261 * variable so init_schedstats() can do it later.
2263 if (!strcmp(str, "enable")) {
2264 __sched_schedstats = true;
2266 } else if (!strcmp(str, "disable")) {
2267 __sched_schedstats = false;
2272 pr_warn("Unable to parse schedstats=\n");
2276 __setup("schedstats=", setup_schedstats);
2278 static void __init init_schedstats(void)
2280 set_schedstats(__sched_schedstats);
2283 #ifdef CONFIG_PROC_SYSCTL
2284 int sysctl_schedstats(struct ctl_table *table, int write,
2285 void __user *buffer, size_t *lenp, loff_t *ppos)
2289 int state = static_branch_likely(&sched_schedstats);
2291 if (write && !capable(CAP_SYS_ADMIN))
2296 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2300 set_schedstats(state);
2303 #endif /* CONFIG_PROC_SYSCTL */
2304 #else /* !CONFIG_SCHEDSTATS */
2305 static inline void init_schedstats(void) {}
2306 #endif /* CONFIG_SCHEDSTATS */
2309 * fork()/clone()-time setup:
2311 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2313 unsigned long flags;
2315 __sched_fork(clone_flags, p);
2317 * We mark the process as NEW here. This guarantees that
2318 * nobody will actually run it, and a signal or other external
2319 * event cannot wake it up and insert it on the runqueue either.
2321 p->state = TASK_NEW;
2324 * Make sure we do not leak PI boosting priority to the child.
2326 p->prio = current->normal_prio;
2329 * Revert to default priority/policy on fork if requested.
2331 if (unlikely(p->sched_reset_on_fork)) {
2332 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2333 p->policy = SCHED_NORMAL;
2334 p->static_prio = NICE_TO_PRIO(0);
2336 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2337 p->static_prio = NICE_TO_PRIO(0);
2339 p->prio = p->normal_prio = __normal_prio(p);
2340 set_load_weight(p, false);
2343 * We don't need the reset flag anymore after the fork. It has
2344 * fulfilled its duty:
2346 p->sched_reset_on_fork = 0;
2349 if (dl_prio(p->prio))
2351 else if (rt_prio(p->prio))
2352 p->sched_class = &rt_sched_class;
2354 p->sched_class = &fair_sched_class;
2356 init_entity_runnable_average(&p->se);
2359 * The child is not yet in the pid-hash so no cgroup attach races,
2360 * and the cgroup is pinned to this child due to cgroup_fork()
2361 * is ran before sched_fork().
2363 * Silence PROVE_RCU.
2365 raw_spin_lock_irqsave(&p->pi_lock, flags);
2367 * We're setting the CPU for the first time, we don't migrate,
2368 * so use __set_task_cpu().
2370 __set_task_cpu(p, smp_processor_id());
2371 if (p->sched_class->task_fork)
2372 p->sched_class->task_fork(p);
2373 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2375 #ifdef CONFIG_SCHED_INFO
2376 if (likely(sched_info_on()))
2377 memset(&p->sched_info, 0, sizeof(p->sched_info));
2379 #if defined(CONFIG_SMP)
2382 init_task_preempt_count(p);
2384 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2385 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2390 unsigned long to_ratio(u64 period, u64 runtime)
2392 if (runtime == RUNTIME_INF)
2396 * Doing this here saves a lot of checks in all
2397 * the calling paths, and returning zero seems
2398 * safe for them anyway.
2403 return div64_u64(runtime << BW_SHIFT, period);
2407 * wake_up_new_task - wake up a newly created task for the first time.
2409 * This function will do some initial scheduler statistics housekeeping
2410 * that must be done for every newly created context, then puts the task
2411 * on the runqueue and wakes it.
2413 void wake_up_new_task(struct task_struct *p)
2418 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2419 p->state = TASK_RUNNING;
2422 * Fork balancing, do it here and not earlier because:
2423 * - cpus_allowed can change in the fork path
2424 * - any previously selected CPU might disappear through hotplug
2426 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2427 * as we're not fully set-up yet.
2429 p->recent_used_cpu = task_cpu(p);
2430 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2432 rq = __task_rq_lock(p, &rf);
2433 update_rq_clock(rq);
2434 post_init_entity_util_avg(&p->se);
2436 activate_task(rq, p, ENQUEUE_NOCLOCK);
2437 p->on_rq = TASK_ON_RQ_QUEUED;
2438 trace_sched_wakeup_new(p);
2439 check_preempt_curr(rq, p, WF_FORK);
2441 if (p->sched_class->task_woken) {
2443 * Nothing relies on rq->lock after this, so its fine to
2446 rq_unpin_lock(rq, &rf);
2447 p->sched_class->task_woken(rq, p);
2448 rq_repin_lock(rq, &rf);
2451 task_rq_unlock(rq, p, &rf);
2454 #ifdef CONFIG_PREEMPT_NOTIFIERS
2456 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2458 void preempt_notifier_inc(void)
2460 static_branch_inc(&preempt_notifier_key);
2462 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2464 void preempt_notifier_dec(void)
2466 static_branch_dec(&preempt_notifier_key);
2468 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2471 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2472 * @notifier: notifier struct to register
2474 void preempt_notifier_register(struct preempt_notifier *notifier)
2476 if (!static_branch_unlikely(&preempt_notifier_key))
2477 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2479 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2481 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2484 * preempt_notifier_unregister - no longer interested in preemption notifications
2485 * @notifier: notifier struct to unregister
2487 * This is *not* safe to call from within a preemption notifier.
2489 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2491 hlist_del(¬ifier->link);
2493 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2495 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2497 struct preempt_notifier *notifier;
2499 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2500 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2503 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2505 if (static_branch_unlikely(&preempt_notifier_key))
2506 __fire_sched_in_preempt_notifiers(curr);
2510 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2511 struct task_struct *next)
2513 struct preempt_notifier *notifier;
2515 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2516 notifier->ops->sched_out(notifier, next);
2519 static __always_inline void
2520 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2521 struct task_struct *next)
2523 if (static_branch_unlikely(&preempt_notifier_key))
2524 __fire_sched_out_preempt_notifiers(curr, next);
2527 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2529 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2534 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2535 struct task_struct *next)
2539 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2541 static inline void prepare_task(struct task_struct *next)
2545 * Claim the task as running, we do this before switching to it
2546 * such that any running task will have this set.
2552 static inline void finish_task(struct task_struct *prev)
2556 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2557 * We must ensure this doesn't happen until the switch is completely
2560 * In particular, the load of prev->state in finish_task_switch() must
2561 * happen before this.
2563 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2565 smp_store_release(&prev->on_cpu, 0);
2570 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
2573 * Since the runqueue lock will be released by the next
2574 * task (which is an invalid locking op but in the case
2575 * of the scheduler it's an obvious special-case), so we
2576 * do an early lockdep release here:
2578 rq_unpin_lock(rq, rf);
2579 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2580 #ifdef CONFIG_DEBUG_SPINLOCK
2581 /* this is a valid case when another task releases the spinlock */
2582 rq->lock.owner = next;
2586 static inline void finish_lock_switch(struct rq *rq)
2589 * If we are tracking spinlock dependencies then we have to
2590 * fix up the runqueue lock - which gets 'carried over' from
2591 * prev into current:
2593 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
2594 raw_spin_unlock_irq(&rq->lock);
2598 * NOP if the arch has not defined these:
2601 #ifndef prepare_arch_switch
2602 # define prepare_arch_switch(next) do { } while (0)
2605 #ifndef finish_arch_post_lock_switch
2606 # define finish_arch_post_lock_switch() do { } while (0)
2610 * prepare_task_switch - prepare to switch tasks
2611 * @rq: the runqueue preparing to switch
2612 * @prev: the current task that is being switched out
2613 * @next: the task we are going to switch to.
2615 * This is called with the rq lock held and interrupts off. It must
2616 * be paired with a subsequent finish_task_switch after the context
2619 * prepare_task_switch sets up locking and calls architecture specific
2623 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2624 struct task_struct *next)
2626 kcov_prepare_switch(prev);
2627 sched_info_switch(rq, prev, next);
2628 perf_event_task_sched_out(prev, next);
2630 fire_sched_out_preempt_notifiers(prev, next);
2632 prepare_arch_switch(next);
2636 * finish_task_switch - clean up after a task-switch
2637 * @prev: the thread we just switched away from.
2639 * finish_task_switch must be called after the context switch, paired
2640 * with a prepare_task_switch call before the context switch.
2641 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2642 * and do any other architecture-specific cleanup actions.
2644 * Note that we may have delayed dropping an mm in context_switch(). If
2645 * so, we finish that here outside of the runqueue lock. (Doing it
2646 * with the lock held can cause deadlocks; see schedule() for
2649 * The context switch have flipped the stack from under us and restored the
2650 * local variables which were saved when this task called schedule() in the
2651 * past. prev == current is still correct but we need to recalculate this_rq
2652 * because prev may have moved to another CPU.
2654 static struct rq *finish_task_switch(struct task_struct *prev)
2655 __releases(rq->lock)
2657 struct rq *rq = this_rq();
2658 struct mm_struct *mm = rq->prev_mm;
2662 * The previous task will have left us with a preempt_count of 2
2663 * because it left us after:
2666 * preempt_disable(); // 1
2668 * raw_spin_lock_irq(&rq->lock) // 2
2670 * Also, see FORK_PREEMPT_COUNT.
2672 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2673 "corrupted preempt_count: %s/%d/0x%x\n",
2674 current->comm, current->pid, preempt_count()))
2675 preempt_count_set(FORK_PREEMPT_COUNT);
2680 * A task struct has one reference for the use as "current".
2681 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2682 * schedule one last time. The schedule call will never return, and
2683 * the scheduled task must drop that reference.
2685 * We must observe prev->state before clearing prev->on_cpu (in
2686 * finish_task), otherwise a concurrent wakeup can get prev
2687 * running on another CPU and we could rave with its RUNNING -> DEAD
2688 * transition, resulting in a double drop.
2690 prev_state = prev->state;
2691 vtime_task_switch(prev);
2692 perf_event_task_sched_in(prev, current);
2694 finish_lock_switch(rq);
2695 finish_arch_post_lock_switch();
2696 kcov_finish_switch(current);
2698 fire_sched_in_preempt_notifiers(current);
2700 * When switching through a kernel thread, the loop in
2701 * membarrier_{private,global}_expedited() may have observed that
2702 * kernel thread and not issued an IPI. It is therefore possible to
2703 * schedule between user->kernel->user threads without passing though
2704 * switch_mm(). Membarrier requires a barrier after storing to
2705 * rq->curr, before returning to userspace, so provide them here:
2707 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
2708 * provided by mmdrop(),
2709 * - a sync_core for SYNC_CORE.
2712 membarrier_mm_sync_core_before_usermode(mm);
2715 if (unlikely(prev_state == TASK_DEAD)) {
2716 if (prev->sched_class->task_dead)
2717 prev->sched_class->task_dead(prev);
2720 * Remove function-return probe instances associated with this
2721 * task and put them back on the free list.
2723 kprobe_flush_task(prev);
2725 /* Task is done with its stack. */
2726 put_task_stack(prev);
2728 put_task_struct(prev);
2731 tick_nohz_task_switch();
2737 /* rq->lock is NOT held, but preemption is disabled */
2738 static void __balance_callback(struct rq *rq)
2740 struct callback_head *head, *next;
2741 void (*func)(struct rq *rq);
2742 unsigned long flags;
2744 raw_spin_lock_irqsave(&rq->lock, flags);
2745 head = rq->balance_callback;
2746 rq->balance_callback = NULL;
2748 func = (void (*)(struct rq *))head->func;
2755 raw_spin_unlock_irqrestore(&rq->lock, flags);
2758 static inline void balance_callback(struct rq *rq)
2760 if (unlikely(rq->balance_callback))
2761 __balance_callback(rq);
2766 static inline void balance_callback(struct rq *rq)
2773 * schedule_tail - first thing a freshly forked thread must call.
2774 * @prev: the thread we just switched away from.
2776 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2777 __releases(rq->lock)
2782 * New tasks start with FORK_PREEMPT_COUNT, see there and
2783 * finish_task_switch() for details.
2785 * finish_task_switch() will drop rq->lock() and lower preempt_count
2786 * and the preempt_enable() will end up enabling preemption (on
2787 * PREEMPT_COUNT kernels).
2790 rq = finish_task_switch(prev);
2791 balance_callback(rq);
2794 if (current->set_child_tid)
2795 put_user(task_pid_vnr(current), current->set_child_tid);
2797 calculate_sigpending();
2801 * context_switch - switch to the new MM and the new thread's register state.
2803 static __always_inline struct rq *
2804 context_switch(struct rq *rq, struct task_struct *prev,
2805 struct task_struct *next, struct rq_flags *rf)
2807 struct mm_struct *mm, *oldmm;
2809 prepare_task_switch(rq, prev, next);
2812 oldmm = prev->active_mm;
2814 * For paravirt, this is coupled with an exit in switch_to to
2815 * combine the page table reload and the switch backend into
2818 arch_start_context_switch(prev);
2821 * If mm is non-NULL, we pass through switch_mm(). If mm is
2822 * NULL, we will pass through mmdrop() in finish_task_switch().
2823 * Both of these contain the full memory barrier required by
2824 * membarrier after storing to rq->curr, before returning to
2828 next->active_mm = oldmm;
2830 enter_lazy_tlb(oldmm, next);
2832 switch_mm_irqs_off(oldmm, mm, next);
2835 prev->active_mm = NULL;
2836 rq->prev_mm = oldmm;
2839 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2841 prepare_lock_switch(rq, next, rf);
2843 /* Here we just switch the register state and the stack. */
2844 switch_to(prev, next, prev);
2847 return finish_task_switch(prev);
2851 * nr_running and nr_context_switches:
2853 * externally visible scheduler statistics: current number of runnable
2854 * threads, total number of context switches performed since bootup.
2856 unsigned long nr_running(void)
2858 unsigned long i, sum = 0;
2860 for_each_online_cpu(i)
2861 sum += cpu_rq(i)->nr_running;
2867 * Check if only the current task is running on the CPU.
2869 * Caution: this function does not check that the caller has disabled
2870 * preemption, thus the result might have a time-of-check-to-time-of-use
2871 * race. The caller is responsible to use it correctly, for example:
2873 * - from a non-preemptible section (of course)
2875 * - from a thread that is bound to a single CPU
2877 * - in a loop with very short iterations (e.g. a polling loop)
2879 bool single_task_running(void)
2881 return raw_rq()->nr_running == 1;
2883 EXPORT_SYMBOL(single_task_running);
2885 unsigned long long nr_context_switches(void)
2888 unsigned long long sum = 0;
2890 for_each_possible_cpu(i)
2891 sum += cpu_rq(i)->nr_switches;
2897 * Consumers of these two interfaces, like for example the cpuidle menu
2898 * governor, are using nonsensical data. Preferring shallow idle state selection
2899 * for a CPU that has IO-wait which might not even end up running the task when
2900 * it does become runnable.
2903 unsigned long nr_iowait_cpu(int cpu)
2905 return atomic_read(&cpu_rq(cpu)->nr_iowait);
2909 * IO-wait accounting, and how its mostly bollocks (on SMP).
2911 * The idea behind IO-wait account is to account the idle time that we could
2912 * have spend running if it were not for IO. That is, if we were to improve the
2913 * storage performance, we'd have a proportional reduction in IO-wait time.
2915 * This all works nicely on UP, where, when a task blocks on IO, we account
2916 * idle time as IO-wait, because if the storage were faster, it could've been
2917 * running and we'd not be idle.
2919 * This has been extended to SMP, by doing the same for each CPU. This however
2922 * Imagine for instance the case where two tasks block on one CPU, only the one
2923 * CPU will have IO-wait accounted, while the other has regular idle. Even
2924 * though, if the storage were faster, both could've ran at the same time,
2925 * utilising both CPUs.
2927 * This means, that when looking globally, the current IO-wait accounting on
2928 * SMP is a lower bound, by reason of under accounting.
2930 * Worse, since the numbers are provided per CPU, they are sometimes
2931 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2932 * associated with any one particular CPU, it can wake to another CPU than it
2933 * blocked on. This means the per CPU IO-wait number is meaningless.
2935 * Task CPU affinities can make all that even more 'interesting'.
2938 unsigned long nr_iowait(void)
2940 unsigned long i, sum = 0;
2942 for_each_possible_cpu(i)
2943 sum += nr_iowait_cpu(i);
2951 * sched_exec - execve() is a valuable balancing opportunity, because at
2952 * this point the task has the smallest effective memory and cache footprint.
2954 void sched_exec(void)
2956 struct task_struct *p = current;
2957 unsigned long flags;
2960 raw_spin_lock_irqsave(&p->pi_lock, flags);
2961 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2962 if (dest_cpu == smp_processor_id())
2965 if (likely(cpu_active(dest_cpu))) {
2966 struct migration_arg arg = { p, dest_cpu };
2968 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2969 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2973 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2978 DEFINE_PER_CPU(struct kernel_stat, kstat);
2979 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2981 EXPORT_PER_CPU_SYMBOL(kstat);
2982 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2985 * The function fair_sched_class.update_curr accesses the struct curr
2986 * and its field curr->exec_start; when called from task_sched_runtime(),
2987 * we observe a high rate of cache misses in practice.
2988 * Prefetching this data results in improved performance.
2990 static inline void prefetch_curr_exec_start(struct task_struct *p)
2992 #ifdef CONFIG_FAIR_GROUP_SCHED
2993 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
2995 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
2998 prefetch(&curr->exec_start);
3002 * Return accounted runtime for the task.
3003 * In case the task is currently running, return the runtime plus current's
3004 * pending runtime that have not been accounted yet.
3006 unsigned long long task_sched_runtime(struct task_struct *p)
3012 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3014 * 64-bit doesn't need locks to atomically read a 64-bit value.
3015 * So we have a optimization chance when the task's delta_exec is 0.
3016 * Reading ->on_cpu is racy, but this is ok.
3018 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3019 * If we race with it entering CPU, unaccounted time is 0. This is
3020 * indistinguishable from the read occurring a few cycles earlier.
3021 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3022 * been accounted, so we're correct here as well.
3024 if (!p->on_cpu || !task_on_rq_queued(p))
3025 return p->se.sum_exec_runtime;
3028 rq = task_rq_lock(p, &rf);
3030 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3031 * project cycles that may never be accounted to this
3032 * thread, breaking clock_gettime().
3034 if (task_current(rq, p) && task_on_rq_queued(p)) {
3035 prefetch_curr_exec_start(p);
3036 update_rq_clock(rq);
3037 p->sched_class->update_curr(rq);
3039 ns = p->se.sum_exec_runtime;
3040 task_rq_unlock(rq, p, &rf);
3046 * This function gets called by the timer code, with HZ frequency.
3047 * We call it with interrupts disabled.
3049 void scheduler_tick(void)
3051 int cpu = smp_processor_id();
3052 struct rq *rq = cpu_rq(cpu);
3053 struct task_struct *curr = rq->curr;
3060 update_rq_clock(rq);
3061 curr->sched_class->task_tick(rq, curr, 0);
3062 cpu_load_update_active(rq);
3063 calc_global_load_tick(rq);
3068 perf_event_task_tick();
3071 rq->idle_balance = idle_cpu(cpu);
3072 trigger_load_balance(rq);
3076 #ifdef CONFIG_NO_HZ_FULL
3080 struct delayed_work work;
3083 static struct tick_work __percpu *tick_work_cpu;
3085 static void sched_tick_remote(struct work_struct *work)
3087 struct delayed_work *dwork = to_delayed_work(work);
3088 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3089 int cpu = twork->cpu;
3090 struct rq *rq = cpu_rq(cpu);
3091 struct task_struct *curr;
3096 * Handle the tick only if it appears the remote CPU is running in full
3097 * dynticks mode. The check is racy by nature, but missing a tick or
3098 * having one too much is no big deal because the scheduler tick updates
3099 * statistics and checks timeslices in a time-independent way, regardless
3100 * of when exactly it is running.
3102 if (idle_cpu(cpu) || !tick_nohz_tick_stopped_cpu(cpu))
3105 rq_lock_irq(rq, &rf);
3107 if (is_idle_task(curr))
3110 update_rq_clock(rq);
3111 delta = rq_clock_task(rq) - curr->se.exec_start;
3114 * Make sure the next tick runs within a reasonable
3117 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3118 curr->sched_class->task_tick(rq, curr, 0);
3121 rq_unlock_irq(rq, &rf);
3125 * Run the remote tick once per second (1Hz). This arbitrary
3126 * frequency is large enough to avoid overload but short enough
3127 * to keep scheduler internal stats reasonably up to date.
3129 queue_delayed_work(system_unbound_wq, dwork, HZ);
3132 static void sched_tick_start(int cpu)
3134 struct tick_work *twork;
3136 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3139 WARN_ON_ONCE(!tick_work_cpu);
3141 twork = per_cpu_ptr(tick_work_cpu, cpu);
3143 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3144 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3147 #ifdef CONFIG_HOTPLUG_CPU
3148 static void sched_tick_stop(int cpu)
3150 struct tick_work *twork;
3152 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3155 WARN_ON_ONCE(!tick_work_cpu);
3157 twork = per_cpu_ptr(tick_work_cpu, cpu);
3158 cancel_delayed_work_sync(&twork->work);
3160 #endif /* CONFIG_HOTPLUG_CPU */
3162 int __init sched_tick_offload_init(void)
3164 tick_work_cpu = alloc_percpu(struct tick_work);
3165 BUG_ON(!tick_work_cpu);
3170 #else /* !CONFIG_NO_HZ_FULL */
3171 static inline void sched_tick_start(int cpu) { }
3172 static inline void sched_tick_stop(int cpu) { }
3175 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3176 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3178 * If the value passed in is equal to the current preempt count
3179 * then we just disabled preemption. Start timing the latency.
3181 static inline void preempt_latency_start(int val)
3183 if (preempt_count() == val) {
3184 unsigned long ip = get_lock_parent_ip();
3185 #ifdef CONFIG_DEBUG_PREEMPT
3186 current->preempt_disable_ip = ip;
3188 trace_preempt_off(CALLER_ADDR0, ip);
3192 void preempt_count_add(int val)
3194 #ifdef CONFIG_DEBUG_PREEMPT
3198 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3201 __preempt_count_add(val);
3202 #ifdef CONFIG_DEBUG_PREEMPT
3204 * Spinlock count overflowing soon?
3206 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3209 preempt_latency_start(val);
3211 EXPORT_SYMBOL(preempt_count_add);
3212 NOKPROBE_SYMBOL(preempt_count_add);
3215 * If the value passed in equals to the current preempt count
3216 * then we just enabled preemption. Stop timing the latency.
3218 static inline void preempt_latency_stop(int val)
3220 if (preempt_count() == val)
3221 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3224 void preempt_count_sub(int val)
3226 #ifdef CONFIG_DEBUG_PREEMPT
3230 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3233 * Is the spinlock portion underflowing?
3235 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3236 !(preempt_count() & PREEMPT_MASK)))
3240 preempt_latency_stop(val);
3241 __preempt_count_sub(val);
3243 EXPORT_SYMBOL(preempt_count_sub);
3244 NOKPROBE_SYMBOL(preempt_count_sub);
3247 static inline void preempt_latency_start(int val) { }
3248 static inline void preempt_latency_stop(int val) { }
3251 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3253 #ifdef CONFIG_DEBUG_PREEMPT
3254 return p->preempt_disable_ip;
3261 * Print scheduling while atomic bug:
3263 static noinline void __schedule_bug(struct task_struct *prev)
3265 /* Save this before calling printk(), since that will clobber it */
3266 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3268 if (oops_in_progress)
3271 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3272 prev->comm, prev->pid, preempt_count());
3274 debug_show_held_locks(prev);
3276 if (irqs_disabled())
3277 print_irqtrace_events(prev);
3278 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3279 && in_atomic_preempt_off()) {
3280 pr_err("Preemption disabled at:");
3281 print_ip_sym(preempt_disable_ip);
3285 panic("scheduling while atomic\n");
3288 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3292 * Various schedule()-time debugging checks and statistics:
3294 static inline void schedule_debug(struct task_struct *prev)
3296 #ifdef CONFIG_SCHED_STACK_END_CHECK
3297 if (task_stack_end_corrupted(prev))
3298 panic("corrupted stack end detected inside scheduler\n");
3301 if (unlikely(in_atomic_preempt_off())) {
3302 __schedule_bug(prev);
3303 preempt_count_set(PREEMPT_DISABLED);
3307 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3309 schedstat_inc(this_rq()->sched_count);
3313 * Pick up the highest-prio task:
3315 static inline struct task_struct *
3316 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3318 const struct sched_class *class;
3319 struct task_struct *p;
3322 * Optimization: we know that if all tasks are in the fair class we can
3323 * call that function directly, but only if the @prev task wasn't of a
3324 * higher scheduling class, because otherwise those loose the
3325 * opportunity to pull in more work from other CPUs.
3327 if (likely((prev->sched_class == &idle_sched_class ||
3328 prev->sched_class == &fair_sched_class) &&
3329 rq->nr_running == rq->cfs.h_nr_running)) {
3331 p = fair_sched_class.pick_next_task(rq, prev, rf);
3332 if (unlikely(p == RETRY_TASK))
3335 /* Assumes fair_sched_class->next == idle_sched_class */
3337 p = idle_sched_class.pick_next_task(rq, prev, rf);
3343 for_each_class(class) {
3344 p = class->pick_next_task(rq, prev, rf);
3346 if (unlikely(p == RETRY_TASK))
3352 /* The idle class should always have a runnable task: */
3357 * __schedule() is the main scheduler function.
3359 * The main means of driving the scheduler and thus entering this function are:
3361 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3363 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3364 * paths. For example, see arch/x86/entry_64.S.
3366 * To drive preemption between tasks, the scheduler sets the flag in timer
3367 * interrupt handler scheduler_tick().
3369 * 3. Wakeups don't really cause entry into schedule(). They add a
3370 * task to the run-queue and that's it.
3372 * Now, if the new task added to the run-queue preempts the current
3373 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3374 * called on the nearest possible occasion:
3376 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3378 * - in syscall or exception context, at the next outmost
3379 * preempt_enable(). (this might be as soon as the wake_up()'s
3382 * - in IRQ context, return from interrupt-handler to
3383 * preemptible context
3385 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3388 * - cond_resched() call
3389 * - explicit schedule() call
3390 * - return from syscall or exception to user-space
3391 * - return from interrupt-handler to user-space
3393 * WARNING: must be called with preemption disabled!
3395 static void __sched notrace __schedule(bool preempt)
3397 struct task_struct *prev, *next;
3398 unsigned long *switch_count;
3403 cpu = smp_processor_id();
3407 schedule_debug(prev);
3409 if (sched_feat(HRTICK))
3412 local_irq_disable();
3413 rcu_note_context_switch(preempt);
3416 * Make sure that signal_pending_state()->signal_pending() below
3417 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3418 * done by the caller to avoid the race with signal_wake_up().
3420 * The membarrier system call requires a full memory barrier
3421 * after coming from user-space, before storing to rq->curr.
3424 smp_mb__after_spinlock();
3426 /* Promote REQ to ACT */
3427 rq->clock_update_flags <<= 1;
3428 update_rq_clock(rq);
3430 switch_count = &prev->nivcsw;
3431 if (!preempt && prev->state) {
3432 if (signal_pending_state(prev->state, prev)) {
3433 prev->state = TASK_RUNNING;
3435 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3438 if (prev->in_iowait) {
3439 atomic_inc(&rq->nr_iowait);
3440 delayacct_blkio_start();
3444 * If a worker went to sleep, notify and ask workqueue
3445 * whether it wants to wake up a task to maintain
3448 if (prev->flags & PF_WQ_WORKER) {
3449 struct task_struct *to_wakeup;
3451 to_wakeup = wq_worker_sleeping(prev);
3453 try_to_wake_up_local(to_wakeup, &rf);
3456 switch_count = &prev->nvcsw;
3459 next = pick_next_task(rq, prev, &rf);
3460 clear_tsk_need_resched(prev);
3461 clear_preempt_need_resched();
3463 if (likely(prev != next)) {
3467 * The membarrier system call requires each architecture
3468 * to have a full memory barrier after updating
3469 * rq->curr, before returning to user-space.
3471 * Here are the schemes providing that barrier on the
3472 * various architectures:
3473 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3474 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3475 * - finish_lock_switch() for weakly-ordered
3476 * architectures where spin_unlock is a full barrier,
3477 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3478 * is a RELEASE barrier),
3482 trace_sched_switch(preempt, prev, next);
3484 /* Also unlocks the rq: */
3485 rq = context_switch(rq, prev, next, &rf);
3487 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3488 rq_unlock_irq(rq, &rf);
3491 balance_callback(rq);
3494 void __noreturn do_task_dead(void)
3496 /* Causes final put_task_struct in finish_task_switch(): */
3497 set_special_state(TASK_DEAD);
3499 /* Tell freezer to ignore us: */
3500 current->flags |= PF_NOFREEZE;
3505 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3510 static inline void sched_submit_work(struct task_struct *tsk)
3512 if (!tsk->state || tsk_is_pi_blocked(tsk))
3515 * If we are going to sleep and we have plugged IO queued,
3516 * make sure to submit it to avoid deadlocks.
3518 if (blk_needs_flush_plug(tsk))
3519 blk_schedule_flush_plug(tsk);
3522 asmlinkage __visible void __sched schedule(void)
3524 struct task_struct *tsk = current;
3526 sched_submit_work(tsk);
3530 sched_preempt_enable_no_resched();
3531 } while (need_resched());
3533 EXPORT_SYMBOL(schedule);
3536 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3537 * state (have scheduled out non-voluntarily) by making sure that all
3538 * tasks have either left the run queue or have gone into user space.
3539 * As idle tasks do not do either, they must not ever be preempted
3540 * (schedule out non-voluntarily).
3542 * schedule_idle() is similar to schedule_preempt_disable() except that it
3543 * never enables preemption because it does not call sched_submit_work().
3545 void __sched schedule_idle(void)
3548 * As this skips calling sched_submit_work(), which the idle task does
3549 * regardless because that function is a nop when the task is in a
3550 * TASK_RUNNING state, make sure this isn't used someplace that the
3551 * current task can be in any other state. Note, idle is always in the
3552 * TASK_RUNNING state.
3554 WARN_ON_ONCE(current->state);
3557 } while (need_resched());
3560 #ifdef CONFIG_CONTEXT_TRACKING
3561 asmlinkage __visible void __sched schedule_user(void)
3564 * If we come here after a random call to set_need_resched(),
3565 * or we have been woken up remotely but the IPI has not yet arrived,
3566 * we haven't yet exited the RCU idle mode. Do it here manually until
3567 * we find a better solution.
3569 * NB: There are buggy callers of this function. Ideally we
3570 * should warn if prev_state != CONTEXT_USER, but that will trigger
3571 * too frequently to make sense yet.
3573 enum ctx_state prev_state = exception_enter();
3575 exception_exit(prev_state);
3580 * schedule_preempt_disabled - called with preemption disabled
3582 * Returns with preemption disabled. Note: preempt_count must be 1
3584 void __sched schedule_preempt_disabled(void)
3586 sched_preempt_enable_no_resched();
3591 static void __sched notrace preempt_schedule_common(void)
3595 * Because the function tracer can trace preempt_count_sub()
3596 * and it also uses preempt_enable/disable_notrace(), if
3597 * NEED_RESCHED is set, the preempt_enable_notrace() called
3598 * by the function tracer will call this function again and
3599 * cause infinite recursion.
3601 * Preemption must be disabled here before the function
3602 * tracer can trace. Break up preempt_disable() into two
3603 * calls. One to disable preemption without fear of being
3604 * traced. The other to still record the preemption latency,
3605 * which can also be traced by the function tracer.
3607 preempt_disable_notrace();
3608 preempt_latency_start(1);
3610 preempt_latency_stop(1);
3611 preempt_enable_no_resched_notrace();
3614 * Check again in case we missed a preemption opportunity
3615 * between schedule and now.
3617 } while (need_resched());
3620 #ifdef CONFIG_PREEMPT
3622 * this is the entry point to schedule() from in-kernel preemption
3623 * off of preempt_enable. Kernel preemptions off return from interrupt
3624 * occur there and call schedule directly.
3626 asmlinkage __visible void __sched notrace preempt_schedule(void)
3629 * If there is a non-zero preempt_count or interrupts are disabled,
3630 * we do not want to preempt the current task. Just return..
3632 if (likely(!preemptible()))
3635 preempt_schedule_common();
3637 NOKPROBE_SYMBOL(preempt_schedule);
3638 EXPORT_SYMBOL(preempt_schedule);
3641 * preempt_schedule_notrace - preempt_schedule called by tracing
3643 * The tracing infrastructure uses preempt_enable_notrace to prevent
3644 * recursion and tracing preempt enabling caused by the tracing
3645 * infrastructure itself. But as tracing can happen in areas coming
3646 * from userspace or just about to enter userspace, a preempt enable
3647 * can occur before user_exit() is called. This will cause the scheduler
3648 * to be called when the system is still in usermode.
3650 * To prevent this, the preempt_enable_notrace will use this function
3651 * instead of preempt_schedule() to exit user context if needed before
3652 * calling the scheduler.
3654 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3656 enum ctx_state prev_ctx;
3658 if (likely(!preemptible()))
3663 * Because the function tracer can trace preempt_count_sub()
3664 * and it also uses preempt_enable/disable_notrace(), if
3665 * NEED_RESCHED is set, the preempt_enable_notrace() called
3666 * by the function tracer will call this function again and
3667 * cause infinite recursion.
3669 * Preemption must be disabled here before the function
3670 * tracer can trace. Break up preempt_disable() into two
3671 * calls. One to disable preemption without fear of being
3672 * traced. The other to still record the preemption latency,
3673 * which can also be traced by the function tracer.
3675 preempt_disable_notrace();
3676 preempt_latency_start(1);
3678 * Needs preempt disabled in case user_exit() is traced
3679 * and the tracer calls preempt_enable_notrace() causing
3680 * an infinite recursion.
3682 prev_ctx = exception_enter();
3684 exception_exit(prev_ctx);
3686 preempt_latency_stop(1);
3687 preempt_enable_no_resched_notrace();
3688 } while (need_resched());
3690 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3692 #endif /* CONFIG_PREEMPT */
3695 * this is the entry point to schedule() from kernel preemption
3696 * off of irq context.
3697 * Note, that this is called and return with irqs disabled. This will
3698 * protect us against recursive calling from irq.
3700 asmlinkage __visible void __sched preempt_schedule_irq(void)
3702 enum ctx_state prev_state;
3704 /* Catch callers which need to be fixed */
3705 BUG_ON(preempt_count() || !irqs_disabled());
3707 prev_state = exception_enter();
3713 local_irq_disable();
3714 sched_preempt_enable_no_resched();
3715 } while (need_resched());
3717 exception_exit(prev_state);
3720 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
3723 return try_to_wake_up(curr->private, mode, wake_flags);
3725 EXPORT_SYMBOL(default_wake_function);
3727 #ifdef CONFIG_RT_MUTEXES
3729 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
3732 prio = min(prio, pi_task->prio);
3737 static inline int rt_effective_prio(struct task_struct *p, int prio)
3739 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3741 return __rt_effective_prio(pi_task, prio);
3745 * rt_mutex_setprio - set the current priority of a task
3747 * @pi_task: donor task
3749 * This function changes the 'effective' priority of a task. It does
3750 * not touch ->normal_prio like __setscheduler().
3752 * Used by the rt_mutex code to implement priority inheritance
3753 * logic. Call site only calls if the priority of the task changed.
3755 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
3757 int prio, oldprio, queued, running, queue_flag =
3758 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
3759 const struct sched_class *prev_class;
3763 /* XXX used to be waiter->prio, not waiter->task->prio */
3764 prio = __rt_effective_prio(pi_task, p->normal_prio);
3767 * If nothing changed; bail early.
3769 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
3772 rq = __task_rq_lock(p, &rf);
3773 update_rq_clock(rq);
3775 * Set under pi_lock && rq->lock, such that the value can be used under
3778 * Note that there is loads of tricky to make this pointer cache work
3779 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3780 * ensure a task is de-boosted (pi_task is set to NULL) before the
3781 * task is allowed to run again (and can exit). This ensures the pointer
3782 * points to a blocked task -- which guaratees the task is present.
3784 p->pi_top_task = pi_task;
3787 * For FIFO/RR we only need to set prio, if that matches we're done.
3789 if (prio == p->prio && !dl_prio(prio))
3793 * Idle task boosting is a nono in general. There is one
3794 * exception, when PREEMPT_RT and NOHZ is active:
3796 * The idle task calls get_next_timer_interrupt() and holds
3797 * the timer wheel base->lock on the CPU and another CPU wants
3798 * to access the timer (probably to cancel it). We can safely
3799 * ignore the boosting request, as the idle CPU runs this code
3800 * with interrupts disabled and will complete the lock
3801 * protected section without being interrupted. So there is no
3802 * real need to boost.
3804 if (unlikely(p == rq->idle)) {
3805 WARN_ON(p != rq->curr);
3806 WARN_ON(p->pi_blocked_on);
3810 trace_sched_pi_setprio(p, pi_task);
3813 if (oldprio == prio)
3814 queue_flag &= ~DEQUEUE_MOVE;
3816 prev_class = p->sched_class;
3817 queued = task_on_rq_queued(p);
3818 running = task_current(rq, p);
3820 dequeue_task(rq, p, queue_flag);
3822 put_prev_task(rq, p);
3825 * Boosting condition are:
3826 * 1. -rt task is running and holds mutex A
3827 * --> -dl task blocks on mutex A
3829 * 2. -dl task is running and holds mutex A
3830 * --> -dl task blocks on mutex A and could preempt the
3833 if (dl_prio(prio)) {
3834 if (!dl_prio(p->normal_prio) ||
3835 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3836 p->dl.dl_boosted = 1;
3837 queue_flag |= ENQUEUE_REPLENISH;
3839 p->dl.dl_boosted = 0;
3840 p->sched_class = &dl_sched_class;
3841 } else if (rt_prio(prio)) {
3842 if (dl_prio(oldprio))
3843 p->dl.dl_boosted = 0;
3845 queue_flag |= ENQUEUE_HEAD;
3846 p->sched_class = &rt_sched_class;
3848 if (dl_prio(oldprio))
3849 p->dl.dl_boosted = 0;
3850 if (rt_prio(oldprio))
3852 p->sched_class = &fair_sched_class;
3858 enqueue_task(rq, p, queue_flag);
3860 set_curr_task(rq, p);
3862 check_class_changed(rq, p, prev_class, oldprio);
3864 /* Avoid rq from going away on us: */
3866 __task_rq_unlock(rq, &rf);
3868 balance_callback(rq);
3872 static inline int rt_effective_prio(struct task_struct *p, int prio)
3878 void set_user_nice(struct task_struct *p, long nice)
3880 bool queued, running;
3881 int old_prio, delta;
3885 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3888 * We have to be careful, if called from sys_setpriority(),
3889 * the task might be in the middle of scheduling on another CPU.
3891 rq = task_rq_lock(p, &rf);
3892 update_rq_clock(rq);
3895 * The RT priorities are set via sched_setscheduler(), but we still
3896 * allow the 'normal' nice value to be set - but as expected
3897 * it wont have any effect on scheduling until the task is
3898 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3900 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3901 p->static_prio = NICE_TO_PRIO(nice);
3904 queued = task_on_rq_queued(p);
3905 running = task_current(rq, p);
3907 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
3909 put_prev_task(rq, p);
3911 p->static_prio = NICE_TO_PRIO(nice);
3912 set_load_weight(p, true);
3914 p->prio = effective_prio(p);
3915 delta = p->prio - old_prio;
3918 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
3920 * If the task increased its priority or is running and
3921 * lowered its priority, then reschedule its CPU:
3923 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3927 set_curr_task(rq, p);
3929 task_rq_unlock(rq, p, &rf);
3931 EXPORT_SYMBOL(set_user_nice);
3934 * can_nice - check if a task can reduce its nice value
3938 int can_nice(const struct task_struct *p, const int nice)
3940 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3941 int nice_rlim = nice_to_rlimit(nice);
3943 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3944 capable(CAP_SYS_NICE));
3947 #ifdef __ARCH_WANT_SYS_NICE
3950 * sys_nice - change the priority of the current process.
3951 * @increment: priority increment
3953 * sys_setpriority is a more generic, but much slower function that
3954 * does similar things.
3956 SYSCALL_DEFINE1(nice, int, increment)
3961 * Setpriority might change our priority at the same moment.
3962 * We don't have to worry. Conceptually one call occurs first
3963 * and we have a single winner.
3965 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3966 nice = task_nice(current) + increment;
3968 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3969 if (increment < 0 && !can_nice(current, nice))
3972 retval = security_task_setnice(current, nice);
3976 set_user_nice(current, nice);
3983 * task_prio - return the priority value of a given task.
3984 * @p: the task in question.
3986 * Return: The priority value as seen by users in /proc.
3987 * RT tasks are offset by -200. Normal tasks are centered
3988 * around 0, value goes from -16 to +15.
3990 int task_prio(const struct task_struct *p)
3992 return p->prio - MAX_RT_PRIO;
3996 * idle_cpu - is a given CPU idle currently?
3997 * @cpu: the processor in question.
3999 * Return: 1 if the CPU is currently idle. 0 otherwise.
4001 int idle_cpu(int cpu)
4003 struct rq *rq = cpu_rq(cpu);
4005 if (rq->curr != rq->idle)
4012 if (!llist_empty(&rq->wake_list))
4020 * available_idle_cpu - is a given CPU idle for enqueuing work.
4021 * @cpu: the CPU in question.
4023 * Return: 1 if the CPU is currently idle. 0 otherwise.
4025 int available_idle_cpu(int cpu)
4030 if (vcpu_is_preempted(cpu))
4037 * idle_task - return the idle task for a given CPU.
4038 * @cpu: the processor in question.
4040 * Return: The idle task for the CPU @cpu.
4042 struct task_struct *idle_task(int cpu)
4044 return cpu_rq(cpu)->idle;
4048 * find_process_by_pid - find a process with a matching PID value.
4049 * @pid: the pid in question.
4051 * The task of @pid, if found. %NULL otherwise.
4053 static struct task_struct *find_process_by_pid(pid_t pid)
4055 return pid ? find_task_by_vpid(pid) : current;
4059 * sched_setparam() passes in -1 for its policy, to let the functions
4060 * it calls know not to change it.
4062 #define SETPARAM_POLICY -1
4064 static void __setscheduler_params(struct task_struct *p,
4065 const struct sched_attr *attr)
4067 int policy = attr->sched_policy;
4069 if (policy == SETPARAM_POLICY)
4074 if (dl_policy(policy))
4075 __setparam_dl(p, attr);
4076 else if (fair_policy(policy))
4077 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4080 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4081 * !rt_policy. Always setting this ensures that things like
4082 * getparam()/getattr() don't report silly values for !rt tasks.
4084 p->rt_priority = attr->sched_priority;
4085 p->normal_prio = normal_prio(p);
4086 set_load_weight(p, true);
4089 /* Actually do priority change: must hold pi & rq lock. */
4090 static void __setscheduler(struct rq *rq, struct task_struct *p,
4091 const struct sched_attr *attr, bool keep_boost)
4093 __setscheduler_params(p, attr);
4096 * Keep a potential priority boosting if called from
4097 * sched_setscheduler().
4099 p->prio = normal_prio(p);
4101 p->prio = rt_effective_prio(p, p->prio);
4103 if (dl_prio(p->prio))
4104 p->sched_class = &dl_sched_class;
4105 else if (rt_prio(p->prio))
4106 p->sched_class = &rt_sched_class;
4108 p->sched_class = &fair_sched_class;
4112 * Check the target process has a UID that matches the current process's:
4114 static bool check_same_owner(struct task_struct *p)
4116 const struct cred *cred = current_cred(), *pcred;
4120 pcred = __task_cred(p);
4121 match = (uid_eq(cred->euid, pcred->euid) ||
4122 uid_eq(cred->euid, pcred->uid));
4127 static int __sched_setscheduler(struct task_struct *p,
4128 const struct sched_attr *attr,
4131 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4132 MAX_RT_PRIO - 1 - attr->sched_priority;
4133 int retval, oldprio, oldpolicy = -1, queued, running;
4134 int new_effective_prio, policy = attr->sched_policy;
4135 const struct sched_class *prev_class;
4138 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4141 /* The pi code expects interrupts enabled */
4142 BUG_ON(pi && in_interrupt());
4144 /* Double check policy once rq lock held: */
4146 reset_on_fork = p->sched_reset_on_fork;
4147 policy = oldpolicy = p->policy;
4149 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4151 if (!valid_policy(policy))
4155 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4159 * Valid priorities for SCHED_FIFO and SCHED_RR are
4160 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4161 * SCHED_BATCH and SCHED_IDLE is 0.
4163 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4164 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4166 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4167 (rt_policy(policy) != (attr->sched_priority != 0)))
4171 * Allow unprivileged RT tasks to decrease priority:
4173 if (user && !capable(CAP_SYS_NICE)) {
4174 if (fair_policy(policy)) {
4175 if (attr->sched_nice < task_nice(p) &&
4176 !can_nice(p, attr->sched_nice))
4180 if (rt_policy(policy)) {
4181 unsigned long rlim_rtprio =
4182 task_rlimit(p, RLIMIT_RTPRIO);
4184 /* Can't set/change the rt policy: */
4185 if (policy != p->policy && !rlim_rtprio)
4188 /* Can't increase priority: */
4189 if (attr->sched_priority > p->rt_priority &&
4190 attr->sched_priority > rlim_rtprio)
4195 * Can't set/change SCHED_DEADLINE policy at all for now
4196 * (safest behavior); in the future we would like to allow
4197 * unprivileged DL tasks to increase their relative deadline
4198 * or reduce their runtime (both ways reducing utilization)
4200 if (dl_policy(policy))
4204 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4205 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4207 if (task_has_idle_policy(p) && !idle_policy(policy)) {
4208 if (!can_nice(p, task_nice(p)))
4212 /* Can't change other user's priorities: */
4213 if (!check_same_owner(p))
4216 /* Normal users shall not reset the sched_reset_on_fork flag: */
4217 if (p->sched_reset_on_fork && !reset_on_fork)
4222 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4225 retval = security_task_setscheduler(p);
4231 * Make sure no PI-waiters arrive (or leave) while we are
4232 * changing the priority of the task:
4234 * To be able to change p->policy safely, the appropriate
4235 * runqueue lock must be held.
4237 rq = task_rq_lock(p, &rf);
4238 update_rq_clock(rq);
4241 * Changing the policy of the stop threads its a very bad idea:
4243 if (p == rq->stop) {
4244 task_rq_unlock(rq, p, &rf);
4249 * If not changing anything there's no need to proceed further,
4250 * but store a possible modification of reset_on_fork.
4252 if (unlikely(policy == p->policy)) {
4253 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4255 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4257 if (dl_policy(policy) && dl_param_changed(p, attr))
4260 p->sched_reset_on_fork = reset_on_fork;
4261 task_rq_unlock(rq, p, &rf);
4267 #ifdef CONFIG_RT_GROUP_SCHED
4269 * Do not allow realtime tasks into groups that have no runtime
4272 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4273 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4274 !task_group_is_autogroup(task_group(p))) {
4275 task_rq_unlock(rq, p, &rf);
4280 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4281 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4282 cpumask_t *span = rq->rd->span;
4285 * Don't allow tasks with an affinity mask smaller than
4286 * the entire root_domain to become SCHED_DEADLINE. We
4287 * will also fail if there's no bandwidth available.
4289 if (!cpumask_subset(span, &p->cpus_allowed) ||
4290 rq->rd->dl_bw.bw == 0) {
4291 task_rq_unlock(rq, p, &rf);
4298 /* Re-check policy now with rq lock held: */
4299 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4300 policy = oldpolicy = -1;
4301 task_rq_unlock(rq, p, &rf);
4306 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4307 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4310 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4311 task_rq_unlock(rq, p, &rf);
4315 p->sched_reset_on_fork = reset_on_fork;
4320 * Take priority boosted tasks into account. If the new
4321 * effective priority is unchanged, we just store the new
4322 * normal parameters and do not touch the scheduler class and
4323 * the runqueue. This will be done when the task deboost
4326 new_effective_prio = rt_effective_prio(p, newprio);
4327 if (new_effective_prio == oldprio)
4328 queue_flags &= ~DEQUEUE_MOVE;
4331 queued = task_on_rq_queued(p);
4332 running = task_current(rq, p);
4334 dequeue_task(rq, p, queue_flags);
4336 put_prev_task(rq, p);
4338 prev_class = p->sched_class;
4339 __setscheduler(rq, p, attr, pi);
4343 * We enqueue to tail when the priority of a task is
4344 * increased (user space view).
4346 if (oldprio < p->prio)
4347 queue_flags |= ENQUEUE_HEAD;
4349 enqueue_task(rq, p, queue_flags);
4352 set_curr_task(rq, p);
4354 check_class_changed(rq, p, prev_class, oldprio);
4356 /* Avoid rq from going away on us: */
4358 task_rq_unlock(rq, p, &rf);
4361 rt_mutex_adjust_pi(p);
4363 /* Run balance callbacks after we've adjusted the PI chain: */
4364 balance_callback(rq);
4370 static int _sched_setscheduler(struct task_struct *p, int policy,
4371 const struct sched_param *param, bool check)
4373 struct sched_attr attr = {
4374 .sched_policy = policy,
4375 .sched_priority = param->sched_priority,
4376 .sched_nice = PRIO_TO_NICE(p->static_prio),
4379 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4380 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4381 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4382 policy &= ~SCHED_RESET_ON_FORK;
4383 attr.sched_policy = policy;
4386 return __sched_setscheduler(p, &attr, check, true);
4389 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4390 * @p: the task in question.
4391 * @policy: new policy.
4392 * @param: structure containing the new RT priority.
4394 * Return: 0 on success. An error code otherwise.
4396 * NOTE that the task may be already dead.
4398 int sched_setscheduler(struct task_struct *p, int policy,
4399 const struct sched_param *param)
4401 return _sched_setscheduler(p, policy, param, true);
4403 EXPORT_SYMBOL_GPL(sched_setscheduler);
4405 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4407 return __sched_setscheduler(p, attr, true, true);
4409 EXPORT_SYMBOL_GPL(sched_setattr);
4411 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
4413 return __sched_setscheduler(p, attr, false, true);
4417 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4418 * @p: the task in question.
4419 * @policy: new policy.
4420 * @param: structure containing the new RT priority.
4422 * Just like sched_setscheduler, only don't bother checking if the
4423 * current context has permission. For example, this is needed in
4424 * stop_machine(): we create temporary high priority worker threads,
4425 * but our caller might not have that capability.
4427 * Return: 0 on success. An error code otherwise.
4429 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4430 const struct sched_param *param)
4432 return _sched_setscheduler(p, policy, param, false);
4434 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4437 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4439 struct sched_param lparam;
4440 struct task_struct *p;
4443 if (!param || pid < 0)
4445 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4450 p = find_process_by_pid(pid);
4452 retval = sched_setscheduler(p, policy, &lparam);
4459 * Mimics kernel/events/core.c perf_copy_attr().
4461 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4466 if (!access_ok(uattr, SCHED_ATTR_SIZE_VER0))
4469 /* Zero the full structure, so that a short copy will be nice: */
4470 memset(attr, 0, sizeof(*attr));
4472 ret = get_user(size, &uattr->size);
4476 /* Bail out on silly large: */
4477 if (size > PAGE_SIZE)
4480 /* ABI compatibility quirk: */
4482 size = SCHED_ATTR_SIZE_VER0;
4484 if (size < SCHED_ATTR_SIZE_VER0)
4488 * If we're handed a bigger struct than we know of,
4489 * ensure all the unknown bits are 0 - i.e. new
4490 * user-space does not rely on any kernel feature
4491 * extensions we dont know about yet.
4493 if (size > sizeof(*attr)) {
4494 unsigned char __user *addr;
4495 unsigned char __user *end;
4498 addr = (void __user *)uattr + sizeof(*attr);
4499 end = (void __user *)uattr + size;
4501 for (; addr < end; addr++) {
4502 ret = get_user(val, addr);
4508 size = sizeof(*attr);
4511 ret = copy_from_user(attr, uattr, size);
4516 * XXX: Do we want to be lenient like existing syscalls; or do we want
4517 * to be strict and return an error on out-of-bounds values?
4519 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4524 put_user(sizeof(*attr), &uattr->size);
4529 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4530 * @pid: the pid in question.
4531 * @policy: new policy.
4532 * @param: structure containing the new RT priority.
4534 * Return: 0 on success. An error code otherwise.
4536 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4541 return do_sched_setscheduler(pid, policy, param);
4545 * sys_sched_setparam - set/change the RT priority of a thread
4546 * @pid: the pid in question.
4547 * @param: structure containing the new RT priority.
4549 * Return: 0 on success. An error code otherwise.
4551 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4553 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4557 * sys_sched_setattr - same as above, but with extended sched_attr
4558 * @pid: the pid in question.
4559 * @uattr: structure containing the extended parameters.
4560 * @flags: for future extension.
4562 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4563 unsigned int, flags)
4565 struct sched_attr attr;
4566 struct task_struct *p;
4569 if (!uattr || pid < 0 || flags)
4572 retval = sched_copy_attr(uattr, &attr);
4576 if ((int)attr.sched_policy < 0)
4581 p = find_process_by_pid(pid);
4583 retval = sched_setattr(p, &attr);
4590 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4591 * @pid: the pid in question.
4593 * Return: On success, the policy of the thread. Otherwise, a negative error
4596 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4598 struct task_struct *p;
4606 p = find_process_by_pid(pid);
4608 retval = security_task_getscheduler(p);
4611 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4618 * sys_sched_getparam - get the RT priority of a thread
4619 * @pid: the pid in question.
4620 * @param: structure containing the RT priority.
4622 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4625 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4627 struct sched_param lp = { .sched_priority = 0 };
4628 struct task_struct *p;
4631 if (!param || pid < 0)
4635 p = find_process_by_pid(pid);
4640 retval = security_task_getscheduler(p);
4644 if (task_has_rt_policy(p))
4645 lp.sched_priority = p->rt_priority;
4649 * This one might sleep, we cannot do it with a spinlock held ...
4651 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4660 static int sched_read_attr(struct sched_attr __user *uattr,
4661 struct sched_attr *attr,
4666 if (!access_ok(uattr, usize))
4670 * If we're handed a smaller struct than we know of,
4671 * ensure all the unknown bits are 0 - i.e. old
4672 * user-space does not get uncomplete information.
4674 if (usize < sizeof(*attr)) {
4675 unsigned char *addr;
4678 addr = (void *)attr + usize;
4679 end = (void *)attr + sizeof(*attr);
4681 for (; addr < end; addr++) {
4689 ret = copy_to_user(uattr, attr, attr->size);
4697 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4698 * @pid: the pid in question.
4699 * @uattr: structure containing the extended parameters.
4700 * @size: sizeof(attr) for fwd/bwd comp.
4701 * @flags: for future extension.
4703 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4704 unsigned int, size, unsigned int, flags)
4706 struct sched_attr attr = {
4707 .size = sizeof(struct sched_attr),
4709 struct task_struct *p;
4712 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4713 size < SCHED_ATTR_SIZE_VER0 || flags)
4717 p = find_process_by_pid(pid);
4722 retval = security_task_getscheduler(p);
4726 attr.sched_policy = p->policy;
4727 if (p->sched_reset_on_fork)
4728 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4729 if (task_has_dl_policy(p))
4730 __getparam_dl(p, &attr);
4731 else if (task_has_rt_policy(p))
4732 attr.sched_priority = p->rt_priority;
4734 attr.sched_nice = task_nice(p);
4738 retval = sched_read_attr(uattr, &attr, size);
4746 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4748 cpumask_var_t cpus_allowed, new_mask;
4749 struct task_struct *p;
4754 p = find_process_by_pid(pid);
4760 /* Prevent p going away */
4764 if (p->flags & PF_NO_SETAFFINITY) {
4768 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4772 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4774 goto out_free_cpus_allowed;
4777 if (!check_same_owner(p)) {
4779 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4781 goto out_free_new_mask;
4786 retval = security_task_setscheduler(p);
4788 goto out_free_new_mask;
4791 cpuset_cpus_allowed(p, cpus_allowed);
4792 cpumask_and(new_mask, in_mask, cpus_allowed);
4795 * Since bandwidth control happens on root_domain basis,
4796 * if admission test is enabled, we only admit -deadline
4797 * tasks allowed to run on all the CPUs in the task's
4801 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4803 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4806 goto out_free_new_mask;
4812 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4815 cpuset_cpus_allowed(p, cpus_allowed);
4816 if (!cpumask_subset(new_mask, cpus_allowed)) {
4818 * We must have raced with a concurrent cpuset
4819 * update. Just reset the cpus_allowed to the
4820 * cpuset's cpus_allowed
4822 cpumask_copy(new_mask, cpus_allowed);
4827 free_cpumask_var(new_mask);
4828 out_free_cpus_allowed:
4829 free_cpumask_var(cpus_allowed);
4835 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4836 struct cpumask *new_mask)
4838 if (len < cpumask_size())
4839 cpumask_clear(new_mask);
4840 else if (len > cpumask_size())
4841 len = cpumask_size();
4843 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4847 * sys_sched_setaffinity - set the CPU affinity of a process
4848 * @pid: pid of the process
4849 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4850 * @user_mask_ptr: user-space pointer to the new CPU mask
4852 * Return: 0 on success. An error code otherwise.
4854 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4855 unsigned long __user *, user_mask_ptr)
4857 cpumask_var_t new_mask;
4860 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4863 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4865 retval = sched_setaffinity(pid, new_mask);
4866 free_cpumask_var(new_mask);
4870 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4872 struct task_struct *p;
4873 unsigned long flags;
4879 p = find_process_by_pid(pid);
4883 retval = security_task_getscheduler(p);
4887 raw_spin_lock_irqsave(&p->pi_lock, flags);
4888 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4889 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4898 * sys_sched_getaffinity - get the CPU affinity of a process
4899 * @pid: pid of the process
4900 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4901 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4903 * Return: size of CPU mask copied to user_mask_ptr on success. An
4904 * error code otherwise.
4906 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4907 unsigned long __user *, user_mask_ptr)
4912 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4914 if (len & (sizeof(unsigned long)-1))
4917 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4920 ret = sched_getaffinity(pid, mask);
4922 unsigned int retlen = min(len, cpumask_size());
4924 if (copy_to_user(user_mask_ptr, mask, retlen))
4929 free_cpumask_var(mask);
4935 * sys_sched_yield - yield the current processor to other threads.
4937 * This function yields the current CPU to other tasks. If there are no
4938 * other threads running on this CPU then this function will return.
4942 static void do_sched_yield(void)
4947 rq = this_rq_lock_irq(&rf);
4949 schedstat_inc(rq->yld_count);
4950 current->sched_class->yield_task(rq);
4953 * Since we are going to call schedule() anyway, there's
4954 * no need to preempt or enable interrupts:
4958 sched_preempt_enable_no_resched();
4963 SYSCALL_DEFINE0(sched_yield)
4969 #ifndef CONFIG_PREEMPT
4970 int __sched _cond_resched(void)
4972 if (should_resched(0)) {
4973 preempt_schedule_common();
4979 EXPORT_SYMBOL(_cond_resched);
4983 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4984 * call schedule, and on return reacquire the lock.
4986 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4987 * operations here to prevent schedule() from being called twice (once via
4988 * spin_unlock(), once by hand).
4990 int __cond_resched_lock(spinlock_t *lock)
4992 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4995 lockdep_assert_held(lock);
4997 if (spin_needbreak(lock) || resched) {
5000 preempt_schedule_common();
5008 EXPORT_SYMBOL(__cond_resched_lock);
5011 * yield - yield the current processor to other threads.
5013 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5015 * The scheduler is at all times free to pick the calling task as the most
5016 * eligible task to run, if removing the yield() call from your code breaks
5017 * it, its already broken.
5019 * Typical broken usage is:
5024 * where one assumes that yield() will let 'the other' process run that will
5025 * make event true. If the current task is a SCHED_FIFO task that will never
5026 * happen. Never use yield() as a progress guarantee!!
5028 * If you want to use yield() to wait for something, use wait_event().
5029 * If you want to use yield() to be 'nice' for others, use cond_resched().
5030 * If you still want to use yield(), do not!
5032 void __sched yield(void)
5034 set_current_state(TASK_RUNNING);
5037 EXPORT_SYMBOL(yield);
5040 * yield_to - yield the current processor to another thread in
5041 * your thread group, or accelerate that thread toward the
5042 * processor it's on.
5044 * @preempt: whether task preemption is allowed or not
5046 * It's the caller's job to ensure that the target task struct
5047 * can't go away on us before we can do any checks.
5050 * true (>0) if we indeed boosted the target task.
5051 * false (0) if we failed to boost the target.
5052 * -ESRCH if there's no task to yield to.
5054 int __sched yield_to(struct task_struct *p, bool preempt)
5056 struct task_struct *curr = current;
5057 struct rq *rq, *p_rq;
5058 unsigned long flags;
5061 local_irq_save(flags);
5067 * If we're the only runnable task on the rq and target rq also
5068 * has only one task, there's absolutely no point in yielding.
5070 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5075 double_rq_lock(rq, p_rq);
5076 if (task_rq(p) != p_rq) {
5077 double_rq_unlock(rq, p_rq);
5081 if (!curr->sched_class->yield_to_task)
5084 if (curr->sched_class != p->sched_class)
5087 if (task_running(p_rq, p) || p->state)
5090 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5092 schedstat_inc(rq->yld_count);
5094 * Make p's CPU reschedule; pick_next_entity takes care of
5097 if (preempt && rq != p_rq)
5102 double_rq_unlock(rq, p_rq);
5104 local_irq_restore(flags);
5111 EXPORT_SYMBOL_GPL(yield_to);
5113 int io_schedule_prepare(void)
5115 int old_iowait = current->in_iowait;
5117 current->in_iowait = 1;
5118 blk_schedule_flush_plug(current);
5123 void io_schedule_finish(int token)
5125 current->in_iowait = token;
5129 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5130 * that process accounting knows that this is a task in IO wait state.
5132 long __sched io_schedule_timeout(long timeout)
5137 token = io_schedule_prepare();
5138 ret = schedule_timeout(timeout);
5139 io_schedule_finish(token);
5143 EXPORT_SYMBOL(io_schedule_timeout);
5145 void io_schedule(void)
5149 token = io_schedule_prepare();
5151 io_schedule_finish(token);
5153 EXPORT_SYMBOL(io_schedule);
5156 * sys_sched_get_priority_max - return maximum RT priority.
5157 * @policy: scheduling class.
5159 * Return: On success, this syscall returns the maximum
5160 * rt_priority that can be used by a given scheduling class.
5161 * On failure, a negative error code is returned.
5163 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5170 ret = MAX_USER_RT_PRIO-1;
5172 case SCHED_DEADLINE:
5183 * sys_sched_get_priority_min - return minimum RT priority.
5184 * @policy: scheduling class.
5186 * Return: On success, this syscall returns the minimum
5187 * rt_priority that can be used by a given scheduling class.
5188 * On failure, a negative error code is returned.
5190 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5199 case SCHED_DEADLINE:
5208 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5210 struct task_struct *p;
5211 unsigned int time_slice;
5221 p = find_process_by_pid(pid);
5225 retval = security_task_getscheduler(p);
5229 rq = task_rq_lock(p, &rf);
5231 if (p->sched_class->get_rr_interval)
5232 time_slice = p->sched_class->get_rr_interval(rq, p);
5233 task_rq_unlock(rq, p, &rf);
5236 jiffies_to_timespec64(time_slice, t);
5245 * sys_sched_rr_get_interval - return the default timeslice of a process.
5246 * @pid: pid of the process.
5247 * @interval: userspace pointer to the timeslice value.
5249 * this syscall writes the default timeslice value of a given process
5250 * into the user-space timespec buffer. A value of '0' means infinity.
5252 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5255 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5256 struct __kernel_timespec __user *, interval)
5258 struct timespec64 t;
5259 int retval = sched_rr_get_interval(pid, &t);
5262 retval = put_timespec64(&t, interval);
5267 #ifdef CONFIG_COMPAT_32BIT_TIME
5268 COMPAT_SYSCALL_DEFINE2(sched_rr_get_interval,
5270 struct old_timespec32 __user *, interval)
5272 struct timespec64 t;
5273 int retval = sched_rr_get_interval(pid, &t);
5276 retval = put_old_timespec32(&t, interval);
5281 void sched_show_task(struct task_struct *p)
5283 unsigned long free = 0;
5286 if (!try_get_task_stack(p))
5289 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5291 if (p->state == TASK_RUNNING)
5292 printk(KERN_CONT " running task ");
5293 #ifdef CONFIG_DEBUG_STACK_USAGE
5294 free = stack_not_used(p);
5299 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5301 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5302 task_pid_nr(p), ppid,
5303 (unsigned long)task_thread_info(p)->flags);
5305 print_worker_info(KERN_INFO, p);
5306 show_stack(p, NULL);
5309 EXPORT_SYMBOL_GPL(sched_show_task);
5312 state_filter_match(unsigned long state_filter, struct task_struct *p)
5314 /* no filter, everything matches */
5318 /* filter, but doesn't match */
5319 if (!(p->state & state_filter))
5323 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5326 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5333 void show_state_filter(unsigned long state_filter)
5335 struct task_struct *g, *p;
5337 #if BITS_PER_LONG == 32
5339 " task PC stack pid father\n");
5342 " task PC stack pid father\n");
5345 for_each_process_thread(g, p) {
5347 * reset the NMI-timeout, listing all files on a slow
5348 * console might take a lot of time:
5349 * Also, reset softlockup watchdogs on all CPUs, because
5350 * another CPU might be blocked waiting for us to process
5353 touch_nmi_watchdog();
5354 touch_all_softlockup_watchdogs();
5355 if (state_filter_match(state_filter, p))
5359 #ifdef CONFIG_SCHED_DEBUG
5361 sysrq_sched_debug_show();
5365 * Only show locks if all tasks are dumped:
5368 debug_show_all_locks();
5372 * init_idle - set up an idle thread for a given CPU
5373 * @idle: task in question
5374 * @cpu: CPU the idle task belongs to
5376 * NOTE: this function does not set the idle thread's NEED_RESCHED
5377 * flag, to make booting more robust.
5379 void init_idle(struct task_struct *idle, int cpu)
5381 struct rq *rq = cpu_rq(cpu);
5382 unsigned long flags;
5384 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5385 raw_spin_lock(&rq->lock);
5387 __sched_fork(0, idle);
5388 idle->state = TASK_RUNNING;
5389 idle->se.exec_start = sched_clock();
5390 idle->flags |= PF_IDLE;
5392 kasan_unpoison_task_stack(idle);
5396 * Its possible that init_idle() gets called multiple times on a task,
5397 * in that case do_set_cpus_allowed() will not do the right thing.
5399 * And since this is boot we can forgo the serialization.
5401 set_cpus_allowed_common(idle, cpumask_of(cpu));
5404 * We're having a chicken and egg problem, even though we are
5405 * holding rq->lock, the CPU isn't yet set to this CPU so the
5406 * lockdep check in task_group() will fail.
5408 * Similar case to sched_fork(). / Alternatively we could
5409 * use task_rq_lock() here and obtain the other rq->lock.
5414 __set_task_cpu(idle, cpu);
5417 rq->curr = rq->idle = idle;
5418 idle->on_rq = TASK_ON_RQ_QUEUED;
5422 raw_spin_unlock(&rq->lock);
5423 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5425 /* Set the preempt count _outside_ the spinlocks! */
5426 init_idle_preempt_count(idle, cpu);
5429 * The idle tasks have their own, simple scheduling class:
5431 idle->sched_class = &idle_sched_class;
5432 ftrace_graph_init_idle_task(idle, cpu);
5433 vtime_init_idle(idle, cpu);
5435 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5441 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5442 const struct cpumask *trial)
5446 if (!cpumask_weight(cur))
5449 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5454 int task_can_attach(struct task_struct *p,
5455 const struct cpumask *cs_cpus_allowed)
5460 * Kthreads which disallow setaffinity shouldn't be moved
5461 * to a new cpuset; we don't want to change their CPU
5462 * affinity and isolating such threads by their set of
5463 * allowed nodes is unnecessary. Thus, cpusets are not
5464 * applicable for such threads. This prevents checking for
5465 * success of set_cpus_allowed_ptr() on all attached tasks
5466 * before cpus_allowed may be changed.
5468 if (p->flags & PF_NO_SETAFFINITY) {
5473 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5475 ret = dl_task_can_attach(p, cs_cpus_allowed);
5481 bool sched_smp_initialized __read_mostly;
5483 #ifdef CONFIG_NUMA_BALANCING
5484 /* Migrate current task p to target_cpu */
5485 int migrate_task_to(struct task_struct *p, int target_cpu)
5487 struct migration_arg arg = { p, target_cpu };
5488 int curr_cpu = task_cpu(p);
5490 if (curr_cpu == target_cpu)
5493 if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5496 /* TODO: This is not properly updating schedstats */
5498 trace_sched_move_numa(p, curr_cpu, target_cpu);
5499 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5503 * Requeue a task on a given node and accurately track the number of NUMA
5504 * tasks on the runqueues
5506 void sched_setnuma(struct task_struct *p, int nid)
5508 bool queued, running;
5512 rq = task_rq_lock(p, &rf);
5513 queued = task_on_rq_queued(p);
5514 running = task_current(rq, p);
5517 dequeue_task(rq, p, DEQUEUE_SAVE);
5519 put_prev_task(rq, p);
5521 p->numa_preferred_nid = nid;
5524 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5526 set_curr_task(rq, p);
5527 task_rq_unlock(rq, p, &rf);
5529 #endif /* CONFIG_NUMA_BALANCING */
5531 #ifdef CONFIG_HOTPLUG_CPU
5533 * Ensure that the idle task is using init_mm right before its CPU goes
5536 void idle_task_exit(void)
5538 struct mm_struct *mm = current->active_mm;
5540 BUG_ON(cpu_online(smp_processor_id()));
5542 if (mm != &init_mm) {
5543 switch_mm(mm, &init_mm, current);
5544 current->active_mm = &init_mm;
5545 finish_arch_post_lock_switch();
5551 * Since this CPU is going 'away' for a while, fold any nr_active delta
5552 * we might have. Assumes we're called after migrate_tasks() so that the
5553 * nr_active count is stable. We need to take the teardown thread which
5554 * is calling this into account, so we hand in adjust = 1 to the load
5557 * Also see the comment "Global load-average calculations".
5559 static void calc_load_migrate(struct rq *rq)
5561 long delta = calc_load_fold_active(rq, 1);
5563 atomic_long_add(delta, &calc_load_tasks);
5566 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5570 static const struct sched_class fake_sched_class = {
5571 .put_prev_task = put_prev_task_fake,
5574 static struct task_struct fake_task = {
5576 * Avoid pull_{rt,dl}_task()
5578 .prio = MAX_PRIO + 1,
5579 .sched_class = &fake_sched_class,
5583 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5584 * try_to_wake_up()->select_task_rq().
5586 * Called with rq->lock held even though we'er in stop_machine() and
5587 * there's no concurrency possible, we hold the required locks anyway
5588 * because of lock validation efforts.
5590 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5592 struct rq *rq = dead_rq;
5593 struct task_struct *next, *stop = rq->stop;
5594 struct rq_flags orf = *rf;
5598 * Fudge the rq selection such that the below task selection loop
5599 * doesn't get stuck on the currently eligible stop task.
5601 * We're currently inside stop_machine() and the rq is either stuck
5602 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5603 * either way we should never end up calling schedule() until we're
5609 * put_prev_task() and pick_next_task() sched
5610 * class method both need to have an up-to-date
5611 * value of rq->clock[_task]
5613 update_rq_clock(rq);
5617 * There's this thread running, bail when that's the only
5620 if (rq->nr_running == 1)
5624 * pick_next_task() assumes pinned rq->lock:
5626 next = pick_next_task(rq, &fake_task, rf);
5628 put_prev_task(rq, next);
5631 * Rules for changing task_struct::cpus_allowed are holding
5632 * both pi_lock and rq->lock, such that holding either
5633 * stabilizes the mask.
5635 * Drop rq->lock is not quite as disastrous as it usually is
5636 * because !cpu_active at this point, which means load-balance
5637 * will not interfere. Also, stop-machine.
5640 raw_spin_lock(&next->pi_lock);
5644 * Since we're inside stop-machine, _nothing_ should have
5645 * changed the task, WARN if weird stuff happened, because in
5646 * that case the above rq->lock drop is a fail too.
5648 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5649 raw_spin_unlock(&next->pi_lock);
5653 /* Find suitable destination for @next, with force if needed. */
5654 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5655 rq = __migrate_task(rq, rf, next, dest_cpu);
5656 if (rq != dead_rq) {
5662 raw_spin_unlock(&next->pi_lock);
5667 #endif /* CONFIG_HOTPLUG_CPU */
5669 void set_rq_online(struct rq *rq)
5672 const struct sched_class *class;
5674 cpumask_set_cpu(rq->cpu, rq->rd->online);
5677 for_each_class(class) {
5678 if (class->rq_online)
5679 class->rq_online(rq);
5684 void set_rq_offline(struct rq *rq)
5687 const struct sched_class *class;
5689 for_each_class(class) {
5690 if (class->rq_offline)
5691 class->rq_offline(rq);
5694 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5700 * used to mark begin/end of suspend/resume:
5702 static int num_cpus_frozen;
5705 * Update cpusets according to cpu_active mask. If cpusets are
5706 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5707 * around partition_sched_domains().
5709 * If we come here as part of a suspend/resume, don't touch cpusets because we
5710 * want to restore it back to its original state upon resume anyway.
5712 static void cpuset_cpu_active(void)
5714 if (cpuhp_tasks_frozen) {
5716 * num_cpus_frozen tracks how many CPUs are involved in suspend
5717 * resume sequence. As long as this is not the last online
5718 * operation in the resume sequence, just build a single sched
5719 * domain, ignoring cpusets.
5721 partition_sched_domains(1, NULL, NULL);
5722 if (--num_cpus_frozen)
5725 * This is the last CPU online operation. So fall through and
5726 * restore the original sched domains by considering the
5727 * cpuset configurations.
5729 cpuset_force_rebuild();
5731 cpuset_update_active_cpus();
5734 static int cpuset_cpu_inactive(unsigned int cpu)
5736 if (!cpuhp_tasks_frozen) {
5737 if (dl_cpu_busy(cpu))
5739 cpuset_update_active_cpus();
5742 partition_sched_domains(1, NULL, NULL);
5747 int sched_cpu_activate(unsigned int cpu)
5749 struct rq *rq = cpu_rq(cpu);
5752 #ifdef CONFIG_SCHED_SMT
5754 * When going up, increment the number of cores with SMT present.
5756 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
5757 static_branch_inc_cpuslocked(&sched_smt_present);
5759 set_cpu_active(cpu, true);
5761 if (sched_smp_initialized) {
5762 sched_domains_numa_masks_set(cpu);
5763 cpuset_cpu_active();
5767 * Put the rq online, if not already. This happens:
5769 * 1) In the early boot process, because we build the real domains
5770 * after all CPUs have been brought up.
5772 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5775 rq_lock_irqsave(rq, &rf);
5777 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5780 rq_unlock_irqrestore(rq, &rf);
5782 update_max_interval();
5787 int sched_cpu_deactivate(unsigned int cpu)
5791 set_cpu_active(cpu, false);
5793 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5794 * users of this state to go away such that all new such users will
5797 * Do sync before park smpboot threads to take care the rcu boost case.
5801 #ifdef CONFIG_SCHED_SMT
5803 * When going down, decrement the number of cores with SMT present.
5805 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
5806 static_branch_dec_cpuslocked(&sched_smt_present);
5809 if (!sched_smp_initialized)
5812 ret = cpuset_cpu_inactive(cpu);
5814 set_cpu_active(cpu, true);
5817 sched_domains_numa_masks_clear(cpu);
5821 static void sched_rq_cpu_starting(unsigned int cpu)
5823 struct rq *rq = cpu_rq(cpu);
5825 rq->calc_load_update = calc_load_update;
5826 update_max_interval();
5829 int sched_cpu_starting(unsigned int cpu)
5831 sched_rq_cpu_starting(cpu);
5832 sched_tick_start(cpu);
5836 #ifdef CONFIG_HOTPLUG_CPU
5837 int sched_cpu_dying(unsigned int cpu)
5839 struct rq *rq = cpu_rq(cpu);
5842 /* Handle pending wakeups and then migrate everything off */
5843 sched_ttwu_pending();
5844 sched_tick_stop(cpu);
5846 rq_lock_irqsave(rq, &rf);
5848 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5851 migrate_tasks(rq, &rf);
5852 BUG_ON(rq->nr_running != 1);
5853 rq_unlock_irqrestore(rq, &rf);
5855 calc_load_migrate(rq);
5856 update_max_interval();
5857 nohz_balance_exit_idle(rq);
5863 void __init sched_init_smp(void)
5868 * There's no userspace yet to cause hotplug operations; hence all the
5869 * CPU masks are stable and all blatant races in the below code cannot
5870 * happen. The hotplug lock is nevertheless taken to satisfy lockdep,
5871 * but there won't be any contention on it.
5874 mutex_lock(&sched_domains_mutex);
5875 sched_init_domains(cpu_active_mask);
5876 mutex_unlock(&sched_domains_mutex);
5879 /* Move init over to a non-isolated CPU */
5880 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
5882 sched_init_granularity();
5884 init_sched_rt_class();
5885 init_sched_dl_class();
5887 sched_smp_initialized = true;
5890 static int __init migration_init(void)
5892 sched_rq_cpu_starting(smp_processor_id());
5895 early_initcall(migration_init);
5898 void __init sched_init_smp(void)
5900 sched_init_granularity();
5902 #endif /* CONFIG_SMP */
5904 int in_sched_functions(unsigned long addr)
5906 return in_lock_functions(addr) ||
5907 (addr >= (unsigned long)__sched_text_start
5908 && addr < (unsigned long)__sched_text_end);
5911 #ifdef CONFIG_CGROUP_SCHED
5913 * Default task group.
5914 * Every task in system belongs to this group at bootup.
5916 struct task_group root_task_group;
5917 LIST_HEAD(task_groups);
5919 /* Cacheline aligned slab cache for task_group */
5920 static struct kmem_cache *task_group_cache __read_mostly;
5923 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5924 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5926 void __init sched_init(void)
5929 unsigned long alloc_size = 0, ptr;
5933 #ifdef CONFIG_FAIR_GROUP_SCHED
5934 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5936 #ifdef CONFIG_RT_GROUP_SCHED
5937 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5940 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5942 #ifdef CONFIG_FAIR_GROUP_SCHED
5943 root_task_group.se = (struct sched_entity **)ptr;
5944 ptr += nr_cpu_ids * sizeof(void **);
5946 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5947 ptr += nr_cpu_ids * sizeof(void **);
5949 #endif /* CONFIG_FAIR_GROUP_SCHED */
5950 #ifdef CONFIG_RT_GROUP_SCHED
5951 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5952 ptr += nr_cpu_ids * sizeof(void **);
5954 root_task_group.rt_rq = (struct rt_rq **)ptr;
5955 ptr += nr_cpu_ids * sizeof(void **);
5957 #endif /* CONFIG_RT_GROUP_SCHED */
5959 #ifdef CONFIG_CPUMASK_OFFSTACK
5960 for_each_possible_cpu(i) {
5961 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5962 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5963 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
5964 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5966 #endif /* CONFIG_CPUMASK_OFFSTACK */
5968 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
5969 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
5972 init_defrootdomain();
5975 #ifdef CONFIG_RT_GROUP_SCHED
5976 init_rt_bandwidth(&root_task_group.rt_bandwidth,
5977 global_rt_period(), global_rt_runtime());
5978 #endif /* CONFIG_RT_GROUP_SCHED */
5980 #ifdef CONFIG_CGROUP_SCHED
5981 task_group_cache = KMEM_CACHE(task_group, 0);
5983 list_add(&root_task_group.list, &task_groups);
5984 INIT_LIST_HEAD(&root_task_group.children);
5985 INIT_LIST_HEAD(&root_task_group.siblings);
5986 autogroup_init(&init_task);
5987 #endif /* CONFIG_CGROUP_SCHED */
5989 for_each_possible_cpu(i) {
5993 raw_spin_lock_init(&rq->lock);
5995 rq->calc_load_active = 0;
5996 rq->calc_load_update = jiffies + LOAD_FREQ;
5997 init_cfs_rq(&rq->cfs);
5998 init_rt_rq(&rq->rt);
5999 init_dl_rq(&rq->dl);
6000 #ifdef CONFIG_FAIR_GROUP_SCHED
6001 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6002 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6003 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6005 * How much CPU bandwidth does root_task_group get?
6007 * In case of task-groups formed thr' the cgroup filesystem, it
6008 * gets 100% of the CPU resources in the system. This overall
6009 * system CPU resource is divided among the tasks of
6010 * root_task_group and its child task-groups in a fair manner,
6011 * based on each entity's (task or task-group's) weight
6012 * (se->load.weight).
6014 * In other words, if root_task_group has 10 tasks of weight
6015 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6016 * then A0's share of the CPU resource is:
6018 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6020 * We achieve this by letting root_task_group's tasks sit
6021 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6023 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6024 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6025 #endif /* CONFIG_FAIR_GROUP_SCHED */
6027 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6028 #ifdef CONFIG_RT_GROUP_SCHED
6029 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6032 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6033 rq->cpu_load[j] = 0;
6038 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6039 rq->balance_callback = NULL;
6040 rq->active_balance = 0;
6041 rq->next_balance = jiffies;
6046 rq->avg_idle = 2*sysctl_sched_migration_cost;
6047 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6049 INIT_LIST_HEAD(&rq->cfs_tasks);
6051 rq_attach_root(rq, &def_root_domain);
6052 #ifdef CONFIG_NO_HZ_COMMON
6053 rq->last_load_update_tick = jiffies;
6054 rq->last_blocked_load_update_tick = jiffies;
6055 atomic_set(&rq->nohz_flags, 0);
6057 #endif /* CONFIG_SMP */
6059 atomic_set(&rq->nr_iowait, 0);
6062 set_load_weight(&init_task, false);
6065 * The boot idle thread does lazy MMU switching as well:
6068 enter_lazy_tlb(&init_mm, current);
6071 * Make us the idle thread. Technically, schedule() should not be
6072 * called from this thread, however somewhere below it might be,
6073 * but because we are the idle thread, we just pick up running again
6074 * when this runqueue becomes "idle".
6076 init_idle(current, smp_processor_id());
6078 calc_load_update = jiffies + LOAD_FREQ;
6081 idle_thread_set_boot_cpu();
6083 init_sched_fair_class();
6089 scheduler_running = 1;
6092 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6093 static inline int preempt_count_equals(int preempt_offset)
6095 int nested = preempt_count() + rcu_preempt_depth();
6097 return (nested == preempt_offset);
6100 void __might_sleep(const char *file, int line, int preempt_offset)
6103 * Blocking primitives will set (and therefore destroy) current->state,
6104 * since we will exit with TASK_RUNNING make sure we enter with it,
6105 * otherwise we will destroy state.
6107 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6108 "do not call blocking ops when !TASK_RUNNING; "
6109 "state=%lx set at [<%p>] %pS\n",
6111 (void *)current->task_state_change,
6112 (void *)current->task_state_change);
6114 ___might_sleep(file, line, preempt_offset);
6116 EXPORT_SYMBOL(__might_sleep);
6118 void ___might_sleep(const char *file, int line, int preempt_offset)
6120 /* Ratelimiting timestamp: */
6121 static unsigned long prev_jiffy;
6123 unsigned long preempt_disable_ip;
6125 /* WARN_ON_ONCE() by default, no rate limit required: */
6128 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6129 !is_idle_task(current)) ||
6130 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6134 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6136 prev_jiffy = jiffies;
6138 /* Save this before calling printk(), since that will clobber it: */
6139 preempt_disable_ip = get_preempt_disable_ip(current);
6142 "BUG: sleeping function called from invalid context at %s:%d\n",
6145 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6146 in_atomic(), irqs_disabled(),
6147 current->pid, current->comm);
6149 if (task_stack_end_corrupted(current))
6150 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6152 debug_show_held_locks(current);
6153 if (irqs_disabled())
6154 print_irqtrace_events(current);
6155 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6156 && !preempt_count_equals(preempt_offset)) {
6157 pr_err("Preemption disabled at:");
6158 print_ip_sym(preempt_disable_ip);
6162 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6164 EXPORT_SYMBOL(___might_sleep);
6167 #ifdef CONFIG_MAGIC_SYSRQ
6168 void normalize_rt_tasks(void)
6170 struct task_struct *g, *p;
6171 struct sched_attr attr = {
6172 .sched_policy = SCHED_NORMAL,
6175 read_lock(&tasklist_lock);
6176 for_each_process_thread(g, p) {
6178 * Only normalize user tasks:
6180 if (p->flags & PF_KTHREAD)
6183 p->se.exec_start = 0;
6184 schedstat_set(p->se.statistics.wait_start, 0);
6185 schedstat_set(p->se.statistics.sleep_start, 0);
6186 schedstat_set(p->se.statistics.block_start, 0);
6188 if (!dl_task(p) && !rt_task(p)) {
6190 * Renice negative nice level userspace
6193 if (task_nice(p) < 0)
6194 set_user_nice(p, 0);
6198 __sched_setscheduler(p, &attr, false, false);
6200 read_unlock(&tasklist_lock);
6203 #endif /* CONFIG_MAGIC_SYSRQ */
6205 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6207 * These functions are only useful for the IA64 MCA handling, or kdb.
6209 * They can only be called when the whole system has been
6210 * stopped - every CPU needs to be quiescent, and no scheduling
6211 * activity can take place. Using them for anything else would
6212 * be a serious bug, and as a result, they aren't even visible
6213 * under any other configuration.
6217 * curr_task - return the current task for a given CPU.
6218 * @cpu: the processor in question.
6220 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6222 * Return: The current task for @cpu.
6224 struct task_struct *curr_task(int cpu)
6226 return cpu_curr(cpu);
6229 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6233 * set_curr_task - set the current task for a given CPU.
6234 * @cpu: the processor in question.
6235 * @p: the task pointer to set.
6237 * Description: This function must only be used when non-maskable interrupts
6238 * are serviced on a separate stack. It allows the architecture to switch the
6239 * notion of the current task on a CPU in a non-blocking manner. This function
6240 * must be called with all CPU's synchronized, and interrupts disabled, the
6241 * and caller must save the original value of the current task (see
6242 * curr_task() above) and restore that value before reenabling interrupts and
6243 * re-starting the system.
6245 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6247 void ia64_set_curr_task(int cpu, struct task_struct *p)
6254 #ifdef CONFIG_CGROUP_SCHED
6255 /* task_group_lock serializes the addition/removal of task groups */
6256 static DEFINE_SPINLOCK(task_group_lock);
6258 static void sched_free_group(struct task_group *tg)
6260 free_fair_sched_group(tg);
6261 free_rt_sched_group(tg);
6263 kmem_cache_free(task_group_cache, tg);
6266 /* allocate runqueue etc for a new task group */
6267 struct task_group *sched_create_group(struct task_group *parent)
6269 struct task_group *tg;
6271 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6273 return ERR_PTR(-ENOMEM);
6275 if (!alloc_fair_sched_group(tg, parent))
6278 if (!alloc_rt_sched_group(tg, parent))
6284 sched_free_group(tg);
6285 return ERR_PTR(-ENOMEM);
6288 void sched_online_group(struct task_group *tg, struct task_group *parent)
6290 unsigned long flags;
6292 spin_lock_irqsave(&task_group_lock, flags);
6293 list_add_rcu(&tg->list, &task_groups);
6295 /* Root should already exist: */
6298 tg->parent = parent;
6299 INIT_LIST_HEAD(&tg->children);
6300 list_add_rcu(&tg->siblings, &parent->children);
6301 spin_unlock_irqrestore(&task_group_lock, flags);
6303 online_fair_sched_group(tg);
6306 /* rcu callback to free various structures associated with a task group */
6307 static void sched_free_group_rcu(struct rcu_head *rhp)
6309 /* Now it should be safe to free those cfs_rqs: */
6310 sched_free_group(container_of(rhp, struct task_group, rcu));
6313 void sched_destroy_group(struct task_group *tg)
6315 /* Wait for possible concurrent references to cfs_rqs complete: */
6316 call_rcu(&tg->rcu, sched_free_group_rcu);
6319 void sched_offline_group(struct task_group *tg)
6321 unsigned long flags;
6323 /* End participation in shares distribution: */
6324 unregister_fair_sched_group(tg);
6326 spin_lock_irqsave(&task_group_lock, flags);
6327 list_del_rcu(&tg->list);
6328 list_del_rcu(&tg->siblings);
6329 spin_unlock_irqrestore(&task_group_lock, flags);
6332 static void sched_change_group(struct task_struct *tsk, int type)
6334 struct task_group *tg;
6337 * All callers are synchronized by task_rq_lock(); we do not use RCU
6338 * which is pointless here. Thus, we pass "true" to task_css_check()
6339 * to prevent lockdep warnings.
6341 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6342 struct task_group, css);
6343 tg = autogroup_task_group(tsk, tg);
6344 tsk->sched_task_group = tg;
6346 #ifdef CONFIG_FAIR_GROUP_SCHED
6347 if (tsk->sched_class->task_change_group)
6348 tsk->sched_class->task_change_group(tsk, type);
6351 set_task_rq(tsk, task_cpu(tsk));
6355 * Change task's runqueue when it moves between groups.
6357 * The caller of this function should have put the task in its new group by
6358 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6361 void sched_move_task(struct task_struct *tsk)
6363 int queued, running, queue_flags =
6364 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6368 rq = task_rq_lock(tsk, &rf);
6369 update_rq_clock(rq);
6371 running = task_current(rq, tsk);
6372 queued = task_on_rq_queued(tsk);
6375 dequeue_task(rq, tsk, queue_flags);
6377 put_prev_task(rq, tsk);
6379 sched_change_group(tsk, TASK_MOVE_GROUP);
6382 enqueue_task(rq, tsk, queue_flags);
6384 set_curr_task(rq, tsk);
6386 task_rq_unlock(rq, tsk, &rf);
6389 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6391 return css ? container_of(css, struct task_group, css) : NULL;
6394 static struct cgroup_subsys_state *
6395 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6397 struct task_group *parent = css_tg(parent_css);
6398 struct task_group *tg;
6401 /* This is early initialization for the top cgroup */
6402 return &root_task_group.css;
6405 tg = sched_create_group(parent);
6407 return ERR_PTR(-ENOMEM);
6412 /* Expose task group only after completing cgroup initialization */
6413 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6415 struct task_group *tg = css_tg(css);
6416 struct task_group *parent = css_tg(css->parent);
6419 sched_online_group(tg, parent);
6423 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6425 struct task_group *tg = css_tg(css);
6427 sched_offline_group(tg);
6430 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6432 struct task_group *tg = css_tg(css);
6435 * Relies on the RCU grace period between css_released() and this.
6437 sched_free_group(tg);
6441 * This is called before wake_up_new_task(), therefore we really only
6442 * have to set its group bits, all the other stuff does not apply.
6444 static void cpu_cgroup_fork(struct task_struct *task)
6449 rq = task_rq_lock(task, &rf);
6451 update_rq_clock(rq);
6452 sched_change_group(task, TASK_SET_GROUP);
6454 task_rq_unlock(rq, task, &rf);
6457 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6459 struct task_struct *task;
6460 struct cgroup_subsys_state *css;
6463 cgroup_taskset_for_each(task, css, tset) {
6464 #ifdef CONFIG_RT_GROUP_SCHED
6465 if (!sched_rt_can_attach(css_tg(css), task))
6468 /* We don't support RT-tasks being in separate groups */
6469 if (task->sched_class != &fair_sched_class)
6473 * Serialize against wake_up_new_task() such that if its
6474 * running, we're sure to observe its full state.
6476 raw_spin_lock_irq(&task->pi_lock);
6478 * Avoid calling sched_move_task() before wake_up_new_task()
6479 * has happened. This would lead to problems with PELT, due to
6480 * move wanting to detach+attach while we're not attached yet.
6482 if (task->state == TASK_NEW)
6484 raw_spin_unlock_irq(&task->pi_lock);
6492 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6494 struct task_struct *task;
6495 struct cgroup_subsys_state *css;
6497 cgroup_taskset_for_each(task, css, tset)
6498 sched_move_task(task);
6501 #ifdef CONFIG_FAIR_GROUP_SCHED
6502 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6503 struct cftype *cftype, u64 shareval)
6505 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6508 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6511 struct task_group *tg = css_tg(css);
6513 return (u64) scale_load_down(tg->shares);
6516 #ifdef CONFIG_CFS_BANDWIDTH
6517 static DEFINE_MUTEX(cfs_constraints_mutex);
6519 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6520 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6522 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6524 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6526 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6527 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6529 if (tg == &root_task_group)
6533 * Ensure we have at some amount of bandwidth every period. This is
6534 * to prevent reaching a state of large arrears when throttled via
6535 * entity_tick() resulting in prolonged exit starvation.
6537 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6541 * Likewise, bound things on the otherside by preventing insane quota
6542 * periods. This also allows us to normalize in computing quota
6545 if (period > max_cfs_quota_period)
6549 * Prevent race between setting of cfs_rq->runtime_enabled and
6550 * unthrottle_offline_cfs_rqs().
6553 mutex_lock(&cfs_constraints_mutex);
6554 ret = __cfs_schedulable(tg, period, quota);
6558 runtime_enabled = quota != RUNTIME_INF;
6559 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6561 * If we need to toggle cfs_bandwidth_used, off->on must occur
6562 * before making related changes, and on->off must occur afterwards
6564 if (runtime_enabled && !runtime_was_enabled)
6565 cfs_bandwidth_usage_inc();
6566 raw_spin_lock_irq(&cfs_b->lock);
6567 cfs_b->period = ns_to_ktime(period);
6568 cfs_b->quota = quota;
6570 __refill_cfs_bandwidth_runtime(cfs_b);
6572 /* Restart the period timer (if active) to handle new period expiry: */
6573 if (runtime_enabled)
6574 start_cfs_bandwidth(cfs_b);
6576 raw_spin_unlock_irq(&cfs_b->lock);
6578 for_each_online_cpu(i) {
6579 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6580 struct rq *rq = cfs_rq->rq;
6583 rq_lock_irq(rq, &rf);
6584 cfs_rq->runtime_enabled = runtime_enabled;
6585 cfs_rq->runtime_remaining = 0;
6587 if (cfs_rq->throttled)
6588 unthrottle_cfs_rq(cfs_rq);
6589 rq_unlock_irq(rq, &rf);
6591 if (runtime_was_enabled && !runtime_enabled)
6592 cfs_bandwidth_usage_dec();
6594 mutex_unlock(&cfs_constraints_mutex);
6600 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
6604 period = ktime_to_ns(tg->cfs_bandwidth.period);
6605 if (cfs_quota_us < 0)
6606 quota = RUNTIME_INF;
6608 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
6610 return tg_set_cfs_bandwidth(tg, period, quota);
6613 long tg_get_cfs_quota(struct task_group *tg)
6617 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
6620 quota_us = tg->cfs_bandwidth.quota;
6621 do_div(quota_us, NSEC_PER_USEC);
6626 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
6630 period = (u64)cfs_period_us * NSEC_PER_USEC;
6631 quota = tg->cfs_bandwidth.quota;
6633 return tg_set_cfs_bandwidth(tg, period, quota);
6636 long tg_get_cfs_period(struct task_group *tg)
6640 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
6641 do_div(cfs_period_us, NSEC_PER_USEC);
6643 return cfs_period_us;
6646 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
6649 return tg_get_cfs_quota(css_tg(css));
6652 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
6653 struct cftype *cftype, s64 cfs_quota_us)
6655 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
6658 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
6661 return tg_get_cfs_period(css_tg(css));
6664 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
6665 struct cftype *cftype, u64 cfs_period_us)
6667 return tg_set_cfs_period(css_tg(css), cfs_period_us);
6670 struct cfs_schedulable_data {
6671 struct task_group *tg;
6676 * normalize group quota/period to be quota/max_period
6677 * note: units are usecs
6679 static u64 normalize_cfs_quota(struct task_group *tg,
6680 struct cfs_schedulable_data *d)
6688 period = tg_get_cfs_period(tg);
6689 quota = tg_get_cfs_quota(tg);
6692 /* note: these should typically be equivalent */
6693 if (quota == RUNTIME_INF || quota == -1)
6696 return to_ratio(period, quota);
6699 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
6701 struct cfs_schedulable_data *d = data;
6702 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6703 s64 quota = 0, parent_quota = -1;
6706 quota = RUNTIME_INF;
6708 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
6710 quota = normalize_cfs_quota(tg, d);
6711 parent_quota = parent_b->hierarchical_quota;
6714 * Ensure max(child_quota) <= parent_quota. On cgroup2,
6715 * always take the min. On cgroup1, only inherit when no
6718 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
6719 quota = min(quota, parent_quota);
6721 if (quota == RUNTIME_INF)
6722 quota = parent_quota;
6723 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
6727 cfs_b->hierarchical_quota = quota;
6732 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
6735 struct cfs_schedulable_data data = {
6741 if (quota != RUNTIME_INF) {
6742 do_div(data.period, NSEC_PER_USEC);
6743 do_div(data.quota, NSEC_PER_USEC);
6747 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
6753 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
6755 struct task_group *tg = css_tg(seq_css(sf));
6756 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6758 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
6759 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
6760 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
6762 if (schedstat_enabled() && tg != &root_task_group) {
6766 for_each_possible_cpu(i)
6767 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
6769 seq_printf(sf, "wait_sum %llu\n", ws);
6774 #endif /* CONFIG_CFS_BANDWIDTH */
6775 #endif /* CONFIG_FAIR_GROUP_SCHED */
6777 #ifdef CONFIG_RT_GROUP_SCHED
6778 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
6779 struct cftype *cft, s64 val)
6781 return sched_group_set_rt_runtime(css_tg(css), val);
6784 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
6787 return sched_group_rt_runtime(css_tg(css));
6790 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
6791 struct cftype *cftype, u64 rt_period_us)
6793 return sched_group_set_rt_period(css_tg(css), rt_period_us);
6796 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
6799 return sched_group_rt_period(css_tg(css));
6801 #endif /* CONFIG_RT_GROUP_SCHED */
6803 static struct cftype cpu_legacy_files[] = {
6804 #ifdef CONFIG_FAIR_GROUP_SCHED
6807 .read_u64 = cpu_shares_read_u64,
6808 .write_u64 = cpu_shares_write_u64,
6811 #ifdef CONFIG_CFS_BANDWIDTH
6813 .name = "cfs_quota_us",
6814 .read_s64 = cpu_cfs_quota_read_s64,
6815 .write_s64 = cpu_cfs_quota_write_s64,
6818 .name = "cfs_period_us",
6819 .read_u64 = cpu_cfs_period_read_u64,
6820 .write_u64 = cpu_cfs_period_write_u64,
6824 .seq_show = cpu_cfs_stat_show,
6827 #ifdef CONFIG_RT_GROUP_SCHED
6829 .name = "rt_runtime_us",
6830 .read_s64 = cpu_rt_runtime_read,
6831 .write_s64 = cpu_rt_runtime_write,
6834 .name = "rt_period_us",
6835 .read_u64 = cpu_rt_period_read_uint,
6836 .write_u64 = cpu_rt_period_write_uint,
6842 static int cpu_extra_stat_show(struct seq_file *sf,
6843 struct cgroup_subsys_state *css)
6845 #ifdef CONFIG_CFS_BANDWIDTH
6847 struct task_group *tg = css_tg(css);
6848 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6851 throttled_usec = cfs_b->throttled_time;
6852 do_div(throttled_usec, NSEC_PER_USEC);
6854 seq_printf(sf, "nr_periods %d\n"
6856 "throttled_usec %llu\n",
6857 cfs_b->nr_periods, cfs_b->nr_throttled,
6864 #ifdef CONFIG_FAIR_GROUP_SCHED
6865 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
6868 struct task_group *tg = css_tg(css);
6869 u64 weight = scale_load_down(tg->shares);
6871 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
6874 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
6875 struct cftype *cft, u64 weight)
6878 * cgroup weight knobs should use the common MIN, DFL and MAX
6879 * values which are 1, 100 and 10000 respectively. While it loses
6880 * a bit of range on both ends, it maps pretty well onto the shares
6881 * value used by scheduler and the round-trip conversions preserve
6882 * the original value over the entire range.
6884 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
6887 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
6889 return sched_group_set_shares(css_tg(css), scale_load(weight));
6892 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
6895 unsigned long weight = scale_load_down(css_tg(css)->shares);
6896 int last_delta = INT_MAX;
6899 /* find the closest nice value to the current weight */
6900 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
6901 delta = abs(sched_prio_to_weight[prio] - weight);
6902 if (delta >= last_delta)
6907 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
6910 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
6911 struct cftype *cft, s64 nice)
6913 unsigned long weight;
6916 if (nice < MIN_NICE || nice > MAX_NICE)
6919 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
6920 idx = array_index_nospec(idx, 40);
6921 weight = sched_prio_to_weight[idx];
6923 return sched_group_set_shares(css_tg(css), scale_load(weight));
6927 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
6928 long period, long quota)
6931 seq_puts(sf, "max");
6933 seq_printf(sf, "%ld", quota);
6935 seq_printf(sf, " %ld\n", period);
6938 /* caller should put the current value in *@periodp before calling */
6939 static int __maybe_unused cpu_period_quota_parse(char *buf,
6940 u64 *periodp, u64 *quotap)
6942 char tok[21]; /* U64_MAX */
6944 if (!sscanf(buf, "%s %llu", tok, periodp))
6947 *periodp *= NSEC_PER_USEC;
6949 if (sscanf(tok, "%llu", quotap))
6950 *quotap *= NSEC_PER_USEC;
6951 else if (!strcmp(tok, "max"))
6952 *quotap = RUNTIME_INF;
6959 #ifdef CONFIG_CFS_BANDWIDTH
6960 static int cpu_max_show(struct seq_file *sf, void *v)
6962 struct task_group *tg = css_tg(seq_css(sf));
6964 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
6968 static ssize_t cpu_max_write(struct kernfs_open_file *of,
6969 char *buf, size_t nbytes, loff_t off)
6971 struct task_group *tg = css_tg(of_css(of));
6972 u64 period = tg_get_cfs_period(tg);
6976 ret = cpu_period_quota_parse(buf, &period, "a);
6978 ret = tg_set_cfs_bandwidth(tg, period, quota);
6979 return ret ?: nbytes;
6983 static struct cftype cpu_files[] = {
6984 #ifdef CONFIG_FAIR_GROUP_SCHED
6987 .flags = CFTYPE_NOT_ON_ROOT,
6988 .read_u64 = cpu_weight_read_u64,
6989 .write_u64 = cpu_weight_write_u64,
6992 .name = "weight.nice",
6993 .flags = CFTYPE_NOT_ON_ROOT,
6994 .read_s64 = cpu_weight_nice_read_s64,
6995 .write_s64 = cpu_weight_nice_write_s64,
6998 #ifdef CONFIG_CFS_BANDWIDTH
7001 .flags = CFTYPE_NOT_ON_ROOT,
7002 .seq_show = cpu_max_show,
7003 .write = cpu_max_write,
7009 struct cgroup_subsys cpu_cgrp_subsys = {
7010 .css_alloc = cpu_cgroup_css_alloc,
7011 .css_online = cpu_cgroup_css_online,
7012 .css_released = cpu_cgroup_css_released,
7013 .css_free = cpu_cgroup_css_free,
7014 .css_extra_stat_show = cpu_extra_stat_show,
7015 .fork = cpu_cgroup_fork,
7016 .can_attach = cpu_cgroup_can_attach,
7017 .attach = cpu_cgroup_attach,
7018 .legacy_cftypes = cpu_legacy_files,
7019 .dfl_cftypes = cpu_files,
7024 #endif /* CONFIG_CGROUP_SCHED */
7026 void dump_cpu_task(int cpu)
7028 pr_info("Task dump for CPU %d:\n", cpu);
7029 sched_show_task(cpu_curr(cpu));
7033 * Nice levels are multiplicative, with a gentle 10% change for every
7034 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7035 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7036 * that remained on nice 0.
7038 * The "10% effect" is relative and cumulative: from _any_ nice level,
7039 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7040 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7041 * If a task goes up by ~10% and another task goes down by ~10% then
7042 * the relative distance between them is ~25%.)
7044 const int sched_prio_to_weight[40] = {
7045 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7046 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7047 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7048 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7049 /* 0 */ 1024, 820, 655, 526, 423,
7050 /* 5 */ 335, 272, 215, 172, 137,
7051 /* 10 */ 110, 87, 70, 56, 45,
7052 /* 15 */ 36, 29, 23, 18, 15,
7056 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7058 * In cases where the weight does not change often, we can use the
7059 * precalculated inverse to speed up arithmetics by turning divisions
7060 * into multiplications:
7062 const u32 sched_prio_to_wmult[40] = {
7063 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7064 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7065 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7066 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7067 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7068 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7069 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7070 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7073 #undef CREATE_TRACE_POINTS