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 address
114 * dependency headed by '[L] rq = task_rq()' and the acquire
115 * will pair with the WMB to ensure we then also see migrating.
117 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
121 raw_spin_unlock(&rq->lock);
122 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
124 while (unlikely(task_on_rq_migrating(p)))
130 * RQ-clock updating methods:
133 static void update_rq_clock_task(struct rq *rq, s64 delta)
136 * In theory, the compile should just see 0 here, and optimize out the call
137 * to sched_rt_avg_update. But I don't trust it...
139 s64 __maybe_unused steal = 0, irq_delta = 0;
141 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
142 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
145 * Since irq_time is only updated on {soft,}irq_exit, we might run into
146 * this case when a previous update_rq_clock() happened inside a
149 * When this happens, we stop ->clock_task and only update the
150 * prev_irq_time stamp to account for the part that fit, so that a next
151 * update will consume the rest. This ensures ->clock_task is
154 * It does however cause some slight miss-attribution of {soft,}irq
155 * time, a more accurate solution would be to update the irq_time using
156 * the current rq->clock timestamp, except that would require using
159 if (irq_delta > delta)
162 rq->prev_irq_time += irq_delta;
165 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
166 if (static_key_false((¶virt_steal_rq_enabled))) {
167 steal = paravirt_steal_clock(cpu_of(rq));
168 steal -= rq->prev_steal_time_rq;
170 if (unlikely(steal > delta))
173 rq->prev_steal_time_rq += steal;
178 rq->clock_task += delta;
180 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
181 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
182 update_irq_load_avg(rq, irq_delta + steal);
184 update_rq_clock_pelt(rq, delta);
187 void update_rq_clock(struct rq *rq)
191 lockdep_assert_held(&rq->lock);
193 if (rq->clock_update_flags & RQCF_ACT_SKIP)
196 #ifdef CONFIG_SCHED_DEBUG
197 if (sched_feat(WARN_DOUBLE_CLOCK))
198 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
199 rq->clock_update_flags |= RQCF_UPDATED;
202 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
206 update_rq_clock_task(rq, delta);
210 #ifdef CONFIG_SCHED_HRTICK
212 * Use HR-timers to deliver accurate preemption points.
215 static void hrtick_clear(struct rq *rq)
217 if (hrtimer_active(&rq->hrtick_timer))
218 hrtimer_cancel(&rq->hrtick_timer);
222 * High-resolution timer tick.
223 * Runs from hardirq context with interrupts disabled.
225 static enum hrtimer_restart hrtick(struct hrtimer *timer)
227 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
230 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
234 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
237 return HRTIMER_NORESTART;
242 static void __hrtick_restart(struct rq *rq)
244 struct hrtimer *timer = &rq->hrtick_timer;
246 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
250 * called from hardirq (IPI) context
252 static void __hrtick_start(void *arg)
258 __hrtick_restart(rq);
259 rq->hrtick_csd_pending = 0;
264 * Called to set the hrtick timer state.
266 * called with rq->lock held and irqs disabled
268 void hrtick_start(struct rq *rq, u64 delay)
270 struct hrtimer *timer = &rq->hrtick_timer;
275 * Don't schedule slices shorter than 10000ns, that just
276 * doesn't make sense and can cause timer DoS.
278 delta = max_t(s64, delay, 10000LL);
279 time = ktime_add_ns(timer->base->get_time(), delta);
281 hrtimer_set_expires(timer, time);
283 if (rq == this_rq()) {
284 __hrtick_restart(rq);
285 } else if (!rq->hrtick_csd_pending) {
286 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
287 rq->hrtick_csd_pending = 1;
293 * Called to set the hrtick timer state.
295 * called with rq->lock held and irqs disabled
297 void hrtick_start(struct rq *rq, u64 delay)
300 * Don't schedule slices shorter than 10000ns, that just
301 * doesn't make sense. Rely on vruntime for fairness.
303 delay = max_t(u64, delay, 10000LL);
304 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
305 HRTIMER_MODE_REL_PINNED);
307 #endif /* CONFIG_SMP */
309 static void hrtick_rq_init(struct rq *rq)
312 rq->hrtick_csd_pending = 0;
314 rq->hrtick_csd.flags = 0;
315 rq->hrtick_csd.func = __hrtick_start;
316 rq->hrtick_csd.info = rq;
319 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
320 rq->hrtick_timer.function = hrtick;
322 #else /* CONFIG_SCHED_HRTICK */
323 static inline void hrtick_clear(struct rq *rq)
327 static inline void hrtick_rq_init(struct rq *rq)
330 #endif /* CONFIG_SCHED_HRTICK */
333 * cmpxchg based fetch_or, macro so it works for different integer types
335 #define fetch_or(ptr, mask) \
337 typeof(ptr) _ptr = (ptr); \
338 typeof(mask) _mask = (mask); \
339 typeof(*_ptr) _old, _val = *_ptr; \
342 _old = cmpxchg(_ptr, _val, _val | _mask); \
350 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
352 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
353 * this avoids any races wrt polling state changes and thereby avoids
356 static bool set_nr_and_not_polling(struct task_struct *p)
358 struct thread_info *ti = task_thread_info(p);
359 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
363 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
365 * If this returns true, then the idle task promises to call
366 * sched_ttwu_pending() and reschedule soon.
368 static bool set_nr_if_polling(struct task_struct *p)
370 struct thread_info *ti = task_thread_info(p);
371 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
374 if (!(val & _TIF_POLLING_NRFLAG))
376 if (val & _TIF_NEED_RESCHED)
378 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
387 static bool set_nr_and_not_polling(struct task_struct *p)
389 set_tsk_need_resched(p);
394 static bool set_nr_if_polling(struct task_struct *p)
401 static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
403 struct wake_q_node *node = &task->wake_q;
406 * Atomically grab the task, if ->wake_q is !nil already it means
407 * its already queued (either by us or someone else) and will get the
408 * wakeup due to that.
410 * In order to ensure that a pending wakeup will observe our pending
411 * state, even in the failed case, an explicit smp_mb() must be used.
413 smp_mb__before_atomic();
414 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
418 * The head is context local, there can be no concurrency.
421 head->lastp = &node->next;
426 * wake_q_add() - queue a wakeup for 'later' waking.
427 * @head: the wake_q_head to add @task to
428 * @task: the task to queue for 'later' wakeup
430 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
431 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
434 * This function must be used as-if it were wake_up_process(); IOW the task
435 * must be ready to be woken at this location.
437 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
439 if (__wake_q_add(head, task))
440 get_task_struct(task);
444 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
445 * @head: the wake_q_head to add @task to
446 * @task: the task to queue for 'later' wakeup
448 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
449 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
452 * This function must be used as-if it were wake_up_process(); IOW the task
453 * must be ready to be woken at this location.
455 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
456 * that already hold reference to @task can call the 'safe' version and trust
457 * wake_q to do the right thing depending whether or not the @task is already
460 void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
462 if (!__wake_q_add(head, task))
463 put_task_struct(task);
466 void wake_up_q(struct wake_q_head *head)
468 struct wake_q_node *node = head->first;
470 while (node != WAKE_Q_TAIL) {
471 struct task_struct *task;
473 task = container_of(node, struct task_struct, wake_q);
475 /* Task can safely be re-inserted now: */
477 task->wake_q.next = NULL;
480 * wake_up_process() executes a full barrier, which pairs with
481 * the queueing in wake_q_add() so as not to miss wakeups.
483 wake_up_process(task);
484 put_task_struct(task);
489 * resched_curr - mark rq's current task 'to be rescheduled now'.
491 * On UP this means the setting of the need_resched flag, on SMP it
492 * might also involve a cross-CPU call to trigger the scheduler on
495 void resched_curr(struct rq *rq)
497 struct task_struct *curr = rq->curr;
500 lockdep_assert_held(&rq->lock);
502 if (test_tsk_need_resched(curr))
507 if (cpu == smp_processor_id()) {
508 set_tsk_need_resched(curr);
509 set_preempt_need_resched();
513 if (set_nr_and_not_polling(curr))
514 smp_send_reschedule(cpu);
516 trace_sched_wake_idle_without_ipi(cpu);
519 void resched_cpu(int cpu)
521 struct rq *rq = cpu_rq(cpu);
524 raw_spin_lock_irqsave(&rq->lock, flags);
525 if (cpu_online(cpu) || cpu == smp_processor_id())
527 raw_spin_unlock_irqrestore(&rq->lock, flags);
531 #ifdef CONFIG_NO_HZ_COMMON
533 * In the semi idle case, use the nearest busy CPU for migrating timers
534 * from an idle CPU. This is good for power-savings.
536 * We don't do similar optimization for completely idle system, as
537 * selecting an idle CPU will add more delays to the timers than intended
538 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
540 int get_nohz_timer_target(void)
542 int i, cpu = smp_processor_id();
543 struct sched_domain *sd;
545 if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
549 for_each_domain(cpu, sd) {
550 for_each_cpu(i, sched_domain_span(sd)) {
554 if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
561 if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
562 cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
569 * When add_timer_on() enqueues a timer into the timer wheel of an
570 * idle CPU then this timer might expire before the next timer event
571 * which is scheduled to wake up that CPU. In case of a completely
572 * idle system the next event might even be infinite time into the
573 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
574 * leaves the inner idle loop so the newly added timer is taken into
575 * account when the CPU goes back to idle and evaluates the timer
576 * wheel for the next timer event.
578 static void wake_up_idle_cpu(int cpu)
580 struct rq *rq = cpu_rq(cpu);
582 if (cpu == smp_processor_id())
585 if (set_nr_and_not_polling(rq->idle))
586 smp_send_reschedule(cpu);
588 trace_sched_wake_idle_without_ipi(cpu);
591 static bool wake_up_full_nohz_cpu(int cpu)
594 * We just need the target to call irq_exit() and re-evaluate
595 * the next tick. The nohz full kick at least implies that.
596 * If needed we can still optimize that later with an
599 if (cpu_is_offline(cpu))
600 return true; /* Don't try to wake offline CPUs. */
601 if (tick_nohz_full_cpu(cpu)) {
602 if (cpu != smp_processor_id() ||
603 tick_nohz_tick_stopped())
604 tick_nohz_full_kick_cpu(cpu);
612 * Wake up the specified CPU. If the CPU is going offline, it is the
613 * caller's responsibility to deal with the lost wakeup, for example,
614 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
616 void wake_up_nohz_cpu(int cpu)
618 if (!wake_up_full_nohz_cpu(cpu))
619 wake_up_idle_cpu(cpu);
622 static inline bool got_nohz_idle_kick(void)
624 int cpu = smp_processor_id();
626 if (!(atomic_read(nohz_flags(cpu)) & NOHZ_KICK_MASK))
629 if (idle_cpu(cpu) && !need_resched())
633 * We can't run Idle Load Balance on this CPU for this time so we
634 * cancel it and clear NOHZ_BALANCE_KICK
636 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
640 #else /* CONFIG_NO_HZ_COMMON */
642 static inline bool got_nohz_idle_kick(void)
647 #endif /* CONFIG_NO_HZ_COMMON */
649 #ifdef CONFIG_NO_HZ_FULL
650 bool sched_can_stop_tick(struct rq *rq)
654 /* Deadline tasks, even if single, need the tick */
655 if (rq->dl.dl_nr_running)
659 * If there are more than one RR tasks, we need the tick to effect the
660 * actual RR behaviour.
662 if (rq->rt.rr_nr_running) {
663 if (rq->rt.rr_nr_running == 1)
670 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
671 * forced preemption between FIFO tasks.
673 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
678 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
679 * if there's more than one we need the tick for involuntary
682 if (rq->nr_running > 1)
687 #endif /* CONFIG_NO_HZ_FULL */
688 #endif /* CONFIG_SMP */
690 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
691 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
693 * Iterate task_group tree rooted at *from, calling @down when first entering a
694 * node and @up when leaving it for the final time.
696 * Caller must hold rcu_lock or sufficient equivalent.
698 int walk_tg_tree_from(struct task_group *from,
699 tg_visitor down, tg_visitor up, void *data)
701 struct task_group *parent, *child;
707 ret = (*down)(parent, data);
710 list_for_each_entry_rcu(child, &parent->children, siblings) {
717 ret = (*up)(parent, data);
718 if (ret || parent == from)
722 parent = parent->parent;
729 int tg_nop(struct task_group *tg, void *data)
735 static void set_load_weight(struct task_struct *p, bool update_load)
737 int prio = p->static_prio - MAX_RT_PRIO;
738 struct load_weight *load = &p->se.load;
741 * SCHED_IDLE tasks get minimal weight:
743 if (task_has_idle_policy(p)) {
744 load->weight = scale_load(WEIGHT_IDLEPRIO);
745 load->inv_weight = WMULT_IDLEPRIO;
746 p->se.runnable_weight = load->weight;
751 * SCHED_OTHER tasks have to update their load when changing their
754 if (update_load && p->sched_class == &fair_sched_class) {
755 reweight_task(p, prio);
757 load->weight = scale_load(sched_prio_to_weight[prio]);
758 load->inv_weight = sched_prio_to_wmult[prio];
759 p->se.runnable_weight = load->weight;
763 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
765 if (!(flags & ENQUEUE_NOCLOCK))
768 if (!(flags & ENQUEUE_RESTORE)) {
769 sched_info_queued(rq, p);
770 psi_enqueue(p, flags & ENQUEUE_WAKEUP);
773 p->sched_class->enqueue_task(rq, p, flags);
776 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
778 if (!(flags & DEQUEUE_NOCLOCK))
781 if (!(flags & DEQUEUE_SAVE)) {
782 sched_info_dequeued(rq, p);
783 psi_dequeue(p, flags & DEQUEUE_SLEEP);
786 p->sched_class->dequeue_task(rq, p, flags);
789 void activate_task(struct rq *rq, struct task_struct *p, int flags)
791 if (task_contributes_to_load(p))
792 rq->nr_uninterruptible--;
794 enqueue_task(rq, p, flags);
796 p->on_rq = TASK_ON_RQ_QUEUED;
799 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
801 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
803 if (task_contributes_to_load(p))
804 rq->nr_uninterruptible++;
806 dequeue_task(rq, p, flags);
810 * __normal_prio - return the priority that is based on the static prio
812 static inline int __normal_prio(struct task_struct *p)
814 return p->static_prio;
818 * Calculate the expected normal priority: i.e. priority
819 * without taking RT-inheritance into account. Might be
820 * boosted by interactivity modifiers. Changes upon fork,
821 * setprio syscalls, and whenever the interactivity
822 * estimator recalculates.
824 static inline int normal_prio(struct task_struct *p)
828 if (task_has_dl_policy(p))
829 prio = MAX_DL_PRIO-1;
830 else if (task_has_rt_policy(p))
831 prio = MAX_RT_PRIO-1 - p->rt_priority;
833 prio = __normal_prio(p);
838 * Calculate the current priority, i.e. the priority
839 * taken into account by the scheduler. This value might
840 * be boosted by RT tasks, or might be boosted by
841 * interactivity modifiers. Will be RT if the task got
842 * RT-boosted. If not then it returns p->normal_prio.
844 static int effective_prio(struct task_struct *p)
846 p->normal_prio = normal_prio(p);
848 * If we are RT tasks or we were boosted to RT priority,
849 * keep the priority unchanged. Otherwise, update priority
850 * to the normal priority:
852 if (!rt_prio(p->prio))
853 return p->normal_prio;
858 * task_curr - is this task currently executing on a CPU?
859 * @p: the task in question.
861 * Return: 1 if the task is currently executing. 0 otherwise.
863 inline int task_curr(const struct task_struct *p)
865 return cpu_curr(task_cpu(p)) == p;
869 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
870 * use the balance_callback list if you want balancing.
872 * this means any call to check_class_changed() must be followed by a call to
873 * balance_callback().
875 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
876 const struct sched_class *prev_class,
879 if (prev_class != p->sched_class) {
880 if (prev_class->switched_from)
881 prev_class->switched_from(rq, p);
883 p->sched_class->switched_to(rq, p);
884 } else if (oldprio != p->prio || dl_task(p))
885 p->sched_class->prio_changed(rq, p, oldprio);
888 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
890 const struct sched_class *class;
892 if (p->sched_class == rq->curr->sched_class) {
893 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
895 for_each_class(class) {
896 if (class == rq->curr->sched_class)
898 if (class == p->sched_class) {
906 * A queue event has occurred, and we're going to schedule. In
907 * this case, we can save a useless back to back clock update.
909 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
910 rq_clock_skip_update(rq);
915 static inline bool is_per_cpu_kthread(struct task_struct *p)
917 if (!(p->flags & PF_KTHREAD))
920 if (p->nr_cpus_allowed != 1)
927 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
928 * __set_cpus_allowed_ptr() and select_fallback_rq().
930 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
932 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
935 if (is_per_cpu_kthread(p))
936 return cpu_online(cpu);
938 return cpu_active(cpu);
942 * This is how migration works:
944 * 1) we invoke migration_cpu_stop() on the target CPU using
946 * 2) stopper starts to run (implicitly forcing the migrated thread
948 * 3) it checks whether the migrated task is still in the wrong runqueue.
949 * 4) if it's in the wrong runqueue then the migration thread removes
950 * it and puts it into the right queue.
951 * 5) stopper completes and stop_one_cpu() returns and the migration
956 * move_queued_task - move a queued task to new rq.
958 * Returns (locked) new rq. Old rq's lock is released.
960 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
961 struct task_struct *p, int new_cpu)
963 lockdep_assert_held(&rq->lock);
965 WRITE_ONCE(p->on_rq, TASK_ON_RQ_MIGRATING);
966 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
967 set_task_cpu(p, new_cpu);
970 rq = cpu_rq(new_cpu);
973 BUG_ON(task_cpu(p) != new_cpu);
974 enqueue_task(rq, p, 0);
975 p->on_rq = TASK_ON_RQ_QUEUED;
976 check_preempt_curr(rq, p, 0);
981 struct migration_arg {
982 struct task_struct *task;
987 * Move (not current) task off this CPU, onto the destination CPU. We're doing
988 * this because either it can't run here any more (set_cpus_allowed()
989 * away from this CPU, or CPU going down), or because we're
990 * attempting to rebalance this task on exec (sched_exec).
992 * So we race with normal scheduler movements, but that's OK, as long
993 * as the task is no longer on this CPU.
995 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
996 struct task_struct *p, int dest_cpu)
998 /* Affinity changed (again). */
999 if (!is_cpu_allowed(p, dest_cpu))
1002 update_rq_clock(rq);
1003 rq = move_queued_task(rq, rf, p, dest_cpu);
1009 * migration_cpu_stop - this will be executed by a highprio stopper thread
1010 * and performs thread migration by bumping thread off CPU then
1011 * 'pushing' onto another runqueue.
1013 static int migration_cpu_stop(void *data)
1015 struct migration_arg *arg = data;
1016 struct task_struct *p = arg->task;
1017 struct rq *rq = this_rq();
1021 * The original target CPU might have gone down and we might
1022 * be on another CPU but it doesn't matter.
1024 local_irq_disable();
1026 * We need to explicitly wake pending tasks before running
1027 * __migrate_task() such that we will not miss enforcing cpus_allowed
1028 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1030 sched_ttwu_pending();
1032 raw_spin_lock(&p->pi_lock);
1035 * If task_rq(p) != rq, it cannot be migrated here, because we're
1036 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1037 * we're holding p->pi_lock.
1039 if (task_rq(p) == rq) {
1040 if (task_on_rq_queued(p))
1041 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1043 p->wake_cpu = arg->dest_cpu;
1046 raw_spin_unlock(&p->pi_lock);
1053 * sched_class::set_cpus_allowed must do the below, but is not required to
1054 * actually call this function.
1056 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1058 cpumask_copy(&p->cpus_allowed, new_mask);
1059 p->nr_cpus_allowed = cpumask_weight(new_mask);
1062 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1064 struct rq *rq = task_rq(p);
1065 bool queued, running;
1067 lockdep_assert_held(&p->pi_lock);
1069 queued = task_on_rq_queued(p);
1070 running = task_current(rq, p);
1074 * Because __kthread_bind() calls this on blocked tasks without
1077 lockdep_assert_held(&rq->lock);
1078 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1081 put_prev_task(rq, p);
1083 p->sched_class->set_cpus_allowed(p, new_mask);
1086 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1088 set_curr_task(rq, p);
1092 * Change a given task's CPU affinity. Migrate the thread to a
1093 * proper CPU and schedule it away if the CPU it's executing on
1094 * is removed from the allowed bitmask.
1096 * NOTE: the caller must have a valid reference to the task, the
1097 * task must not exit() & deallocate itself prematurely. The
1098 * call is not atomic; no spinlocks may be held.
1100 static int __set_cpus_allowed_ptr(struct task_struct *p,
1101 const struct cpumask *new_mask, bool check)
1103 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1104 unsigned int dest_cpu;
1109 rq = task_rq_lock(p, &rf);
1110 update_rq_clock(rq);
1112 if (p->flags & PF_KTHREAD) {
1114 * Kernel threads are allowed on online && !active CPUs
1116 cpu_valid_mask = cpu_online_mask;
1120 * Must re-check here, to close a race against __kthread_bind(),
1121 * sched_setaffinity() is not guaranteed to observe the flag.
1123 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1128 if (cpumask_equal(&p->cpus_allowed, new_mask))
1131 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1136 do_set_cpus_allowed(p, new_mask);
1138 if (p->flags & PF_KTHREAD) {
1140 * For kernel threads that do indeed end up on online &&
1141 * !active we want to ensure they are strict per-CPU threads.
1143 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1144 !cpumask_intersects(new_mask, cpu_active_mask) &&
1145 p->nr_cpus_allowed != 1);
1148 /* Can the task run on the task's current CPU? If so, we're done */
1149 if (cpumask_test_cpu(task_cpu(p), new_mask))
1152 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1153 if (task_running(rq, p) || p->state == TASK_WAKING) {
1154 struct migration_arg arg = { p, dest_cpu };
1155 /* Need help from migration thread: drop lock and wait. */
1156 task_rq_unlock(rq, p, &rf);
1157 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1159 } else if (task_on_rq_queued(p)) {
1161 * OK, since we're going to drop the lock immediately
1162 * afterwards anyway.
1164 rq = move_queued_task(rq, &rf, p, dest_cpu);
1167 task_rq_unlock(rq, p, &rf);
1172 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1174 return __set_cpus_allowed_ptr(p, new_mask, false);
1176 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1178 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1180 #ifdef CONFIG_SCHED_DEBUG
1182 * We should never call set_task_cpu() on a blocked task,
1183 * ttwu() will sort out the placement.
1185 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1189 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1190 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1191 * time relying on p->on_rq.
1193 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1194 p->sched_class == &fair_sched_class &&
1195 (p->on_rq && !task_on_rq_migrating(p)));
1197 #ifdef CONFIG_LOCKDEP
1199 * The caller should hold either p->pi_lock or rq->lock, when changing
1200 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1202 * sched_move_task() holds both and thus holding either pins the cgroup,
1205 * Furthermore, all task_rq users should acquire both locks, see
1208 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1209 lockdep_is_held(&task_rq(p)->lock)));
1212 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1214 WARN_ON_ONCE(!cpu_online(new_cpu));
1217 trace_sched_migrate_task(p, new_cpu);
1219 if (task_cpu(p) != new_cpu) {
1220 if (p->sched_class->migrate_task_rq)
1221 p->sched_class->migrate_task_rq(p, new_cpu);
1222 p->se.nr_migrations++;
1224 perf_event_task_migrate(p);
1227 __set_task_cpu(p, new_cpu);
1230 #ifdef CONFIG_NUMA_BALANCING
1231 static void __migrate_swap_task(struct task_struct *p, int cpu)
1233 if (task_on_rq_queued(p)) {
1234 struct rq *src_rq, *dst_rq;
1235 struct rq_flags srf, drf;
1237 src_rq = task_rq(p);
1238 dst_rq = cpu_rq(cpu);
1240 rq_pin_lock(src_rq, &srf);
1241 rq_pin_lock(dst_rq, &drf);
1243 deactivate_task(src_rq, p, 0);
1244 set_task_cpu(p, cpu);
1245 activate_task(dst_rq, p, 0);
1246 check_preempt_curr(dst_rq, p, 0);
1248 rq_unpin_lock(dst_rq, &drf);
1249 rq_unpin_lock(src_rq, &srf);
1253 * Task isn't running anymore; make it appear like we migrated
1254 * it before it went to sleep. This means on wakeup we make the
1255 * previous CPU our target instead of where it really is.
1261 struct migration_swap_arg {
1262 struct task_struct *src_task, *dst_task;
1263 int src_cpu, dst_cpu;
1266 static int migrate_swap_stop(void *data)
1268 struct migration_swap_arg *arg = data;
1269 struct rq *src_rq, *dst_rq;
1272 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1275 src_rq = cpu_rq(arg->src_cpu);
1276 dst_rq = cpu_rq(arg->dst_cpu);
1278 double_raw_lock(&arg->src_task->pi_lock,
1279 &arg->dst_task->pi_lock);
1280 double_rq_lock(src_rq, dst_rq);
1282 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1285 if (task_cpu(arg->src_task) != arg->src_cpu)
1288 if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1291 if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1294 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1295 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1300 double_rq_unlock(src_rq, dst_rq);
1301 raw_spin_unlock(&arg->dst_task->pi_lock);
1302 raw_spin_unlock(&arg->src_task->pi_lock);
1308 * Cross migrate two tasks
1310 int migrate_swap(struct task_struct *cur, struct task_struct *p,
1311 int target_cpu, int curr_cpu)
1313 struct migration_swap_arg arg;
1316 arg = (struct migration_swap_arg){
1318 .src_cpu = curr_cpu,
1320 .dst_cpu = target_cpu,
1323 if (arg.src_cpu == arg.dst_cpu)
1327 * These three tests are all lockless; this is OK since all of them
1328 * will be re-checked with proper locks held further down the line.
1330 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1333 if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1336 if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1339 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1340 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1345 #endif /* CONFIG_NUMA_BALANCING */
1348 * wait_task_inactive - wait for a thread to unschedule.
1350 * If @match_state is nonzero, it's the @p->state value just checked and
1351 * not expected to change. If it changes, i.e. @p might have woken up,
1352 * then return zero. When we succeed in waiting for @p to be off its CPU,
1353 * we return a positive number (its total switch count). If a second call
1354 * a short while later returns the same number, the caller can be sure that
1355 * @p has remained unscheduled the whole time.
1357 * The caller must ensure that the task *will* unschedule sometime soon,
1358 * else this function might spin for a *long* time. This function can't
1359 * be called with interrupts off, or it may introduce deadlock with
1360 * smp_call_function() if an IPI is sent by the same process we are
1361 * waiting to become inactive.
1363 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1365 int running, queued;
1372 * We do the initial early heuristics without holding
1373 * any task-queue locks at all. We'll only try to get
1374 * the runqueue lock when things look like they will
1380 * If the task is actively running on another CPU
1381 * still, just relax and busy-wait without holding
1384 * NOTE! Since we don't hold any locks, it's not
1385 * even sure that "rq" stays as the right runqueue!
1386 * But we don't care, since "task_running()" will
1387 * return false if the runqueue has changed and p
1388 * is actually now running somewhere else!
1390 while (task_running(rq, p)) {
1391 if (match_state && unlikely(p->state != match_state))
1397 * Ok, time to look more closely! We need the rq
1398 * lock now, to be *sure*. If we're wrong, we'll
1399 * just go back and repeat.
1401 rq = task_rq_lock(p, &rf);
1402 trace_sched_wait_task(p);
1403 running = task_running(rq, p);
1404 queued = task_on_rq_queued(p);
1406 if (!match_state || p->state == match_state)
1407 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1408 task_rq_unlock(rq, p, &rf);
1411 * If it changed from the expected state, bail out now.
1413 if (unlikely(!ncsw))
1417 * Was it really running after all now that we
1418 * checked with the proper locks actually held?
1420 * Oops. Go back and try again..
1422 if (unlikely(running)) {
1428 * It's not enough that it's not actively running,
1429 * it must be off the runqueue _entirely_, and not
1432 * So if it was still runnable (but just not actively
1433 * running right now), it's preempted, and we should
1434 * yield - it could be a while.
1436 if (unlikely(queued)) {
1437 ktime_t to = NSEC_PER_SEC / HZ;
1439 set_current_state(TASK_UNINTERRUPTIBLE);
1440 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1445 * Ahh, all good. It wasn't running, and it wasn't
1446 * runnable, which means that it will never become
1447 * running in the future either. We're all done!
1456 * kick_process - kick a running thread to enter/exit the kernel
1457 * @p: the to-be-kicked thread
1459 * Cause a process which is running on another CPU to enter
1460 * kernel-mode, without any delay. (to get signals handled.)
1462 * NOTE: this function doesn't have to take the runqueue lock,
1463 * because all it wants to ensure is that the remote task enters
1464 * the kernel. If the IPI races and the task has been migrated
1465 * to another CPU then no harm is done and the purpose has been
1468 void kick_process(struct task_struct *p)
1474 if ((cpu != smp_processor_id()) && task_curr(p))
1475 smp_send_reschedule(cpu);
1478 EXPORT_SYMBOL_GPL(kick_process);
1481 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1483 * A few notes on cpu_active vs cpu_online:
1485 * - cpu_active must be a subset of cpu_online
1487 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
1488 * see __set_cpus_allowed_ptr(). At this point the newly online
1489 * CPU isn't yet part of the sched domains, and balancing will not
1492 * - on CPU-down we clear cpu_active() to mask the sched domains and
1493 * avoid the load balancer to place new tasks on the to be removed
1494 * CPU. Existing tasks will remain running there and will be taken
1497 * This means that fallback selection must not select !active CPUs.
1498 * And can assume that any active CPU must be online. Conversely
1499 * select_task_rq() below may allow selection of !active CPUs in order
1500 * to satisfy the above rules.
1502 static int select_fallback_rq(int cpu, struct task_struct *p)
1504 int nid = cpu_to_node(cpu);
1505 const struct cpumask *nodemask = NULL;
1506 enum { cpuset, possible, fail } state = cpuset;
1510 * If the node that the CPU is on has been offlined, cpu_to_node()
1511 * will return -1. There is no CPU on the node, and we should
1512 * select the CPU on the other node.
1515 nodemask = cpumask_of_node(nid);
1517 /* Look for allowed, online CPU in same node. */
1518 for_each_cpu(dest_cpu, nodemask) {
1519 if (!cpu_active(dest_cpu))
1521 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1527 /* Any allowed, online CPU? */
1528 for_each_cpu(dest_cpu, &p->cpus_allowed) {
1529 if (!is_cpu_allowed(p, dest_cpu))
1535 /* No more Mr. Nice Guy. */
1538 if (IS_ENABLED(CONFIG_CPUSETS)) {
1539 cpuset_cpus_allowed_fallback(p);
1545 do_set_cpus_allowed(p, cpu_possible_mask);
1556 if (state != cpuset) {
1558 * Don't tell them about moving exiting tasks or
1559 * kernel threads (both mm NULL), since they never
1562 if (p->mm && printk_ratelimit()) {
1563 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1564 task_pid_nr(p), p->comm, cpu);
1572 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1575 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1577 lockdep_assert_held(&p->pi_lock);
1579 if (p->nr_cpus_allowed > 1)
1580 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1582 cpu = cpumask_any(&p->cpus_allowed);
1585 * In order not to call set_task_cpu() on a blocking task we need
1586 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1589 * Since this is common to all placement strategies, this lives here.
1591 * [ this allows ->select_task() to simply return task_cpu(p) and
1592 * not worry about this generic constraint ]
1594 if (unlikely(!is_cpu_allowed(p, cpu)))
1595 cpu = select_fallback_rq(task_cpu(p), p);
1600 static void update_avg(u64 *avg, u64 sample)
1602 s64 diff = sample - *avg;
1606 void sched_set_stop_task(int cpu, struct task_struct *stop)
1608 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1609 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1613 * Make it appear like a SCHED_FIFO task, its something
1614 * userspace knows about and won't get confused about.
1616 * Also, it will make PI more or less work without too
1617 * much confusion -- but then, stop work should not
1618 * rely on PI working anyway.
1620 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1622 stop->sched_class = &stop_sched_class;
1625 cpu_rq(cpu)->stop = stop;
1629 * Reset it back to a normal scheduling class so that
1630 * it can die in pieces.
1632 old_stop->sched_class = &rt_sched_class;
1638 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1639 const struct cpumask *new_mask, bool check)
1641 return set_cpus_allowed_ptr(p, new_mask);
1644 #endif /* CONFIG_SMP */
1647 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1651 if (!schedstat_enabled())
1657 if (cpu == rq->cpu) {
1658 __schedstat_inc(rq->ttwu_local);
1659 __schedstat_inc(p->se.statistics.nr_wakeups_local);
1661 struct sched_domain *sd;
1663 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
1665 for_each_domain(rq->cpu, sd) {
1666 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1667 __schedstat_inc(sd->ttwu_wake_remote);
1674 if (wake_flags & WF_MIGRATED)
1675 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1676 #endif /* CONFIG_SMP */
1678 __schedstat_inc(rq->ttwu_count);
1679 __schedstat_inc(p->se.statistics.nr_wakeups);
1681 if (wake_flags & WF_SYNC)
1682 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
1686 * Mark the task runnable and perform wakeup-preemption.
1688 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1689 struct rq_flags *rf)
1691 check_preempt_curr(rq, p, wake_flags);
1692 p->state = TASK_RUNNING;
1693 trace_sched_wakeup(p);
1696 if (p->sched_class->task_woken) {
1698 * Our task @p is fully woken up and running; so its safe to
1699 * drop the rq->lock, hereafter rq is only used for statistics.
1701 rq_unpin_lock(rq, rf);
1702 p->sched_class->task_woken(rq, p);
1703 rq_repin_lock(rq, rf);
1706 if (rq->idle_stamp) {
1707 u64 delta = rq_clock(rq) - rq->idle_stamp;
1708 u64 max = 2*rq->max_idle_balance_cost;
1710 update_avg(&rq->avg_idle, delta);
1712 if (rq->avg_idle > max)
1721 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1722 struct rq_flags *rf)
1724 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
1726 lockdep_assert_held(&rq->lock);
1729 if (p->sched_contributes_to_load)
1730 rq->nr_uninterruptible--;
1732 if (wake_flags & WF_MIGRATED)
1733 en_flags |= ENQUEUE_MIGRATED;
1736 activate_task(rq, p, en_flags);
1737 ttwu_do_wakeup(rq, p, wake_flags, rf);
1741 * Called in case the task @p isn't fully descheduled from its runqueue,
1742 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1743 * since all we need to do is flip p->state to TASK_RUNNING, since
1744 * the task is still ->on_rq.
1746 static int ttwu_remote(struct task_struct *p, int wake_flags)
1752 rq = __task_rq_lock(p, &rf);
1753 if (task_on_rq_queued(p)) {
1754 /* check_preempt_curr() may use rq clock */
1755 update_rq_clock(rq);
1756 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1759 __task_rq_unlock(rq, &rf);
1765 void sched_ttwu_pending(void)
1767 struct rq *rq = this_rq();
1768 struct llist_node *llist = llist_del_all(&rq->wake_list);
1769 struct task_struct *p, *t;
1775 rq_lock_irqsave(rq, &rf);
1776 update_rq_clock(rq);
1778 llist_for_each_entry_safe(p, t, llist, wake_entry)
1779 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
1781 rq_unlock_irqrestore(rq, &rf);
1784 void scheduler_ipi(void)
1787 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1788 * TIF_NEED_RESCHED remotely (for the first time) will also send
1791 preempt_fold_need_resched();
1793 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1797 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1798 * traditionally all their work was done from the interrupt return
1799 * path. Now that we actually do some work, we need to make sure
1802 * Some archs already do call them, luckily irq_enter/exit nest
1805 * Arguably we should visit all archs and update all handlers,
1806 * however a fair share of IPIs are still resched only so this would
1807 * somewhat pessimize the simple resched case.
1810 sched_ttwu_pending();
1813 * Check if someone kicked us for doing the nohz idle load balance.
1815 if (unlikely(got_nohz_idle_kick())) {
1816 this_rq()->idle_balance = 1;
1817 raise_softirq_irqoff(SCHED_SOFTIRQ);
1822 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1824 struct rq *rq = cpu_rq(cpu);
1826 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1828 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1829 if (!set_nr_if_polling(rq->idle))
1830 smp_send_reschedule(cpu);
1832 trace_sched_wake_idle_without_ipi(cpu);
1836 void wake_up_if_idle(int cpu)
1838 struct rq *rq = cpu_rq(cpu);
1843 if (!is_idle_task(rcu_dereference(rq->curr)))
1846 if (set_nr_if_polling(rq->idle)) {
1847 trace_sched_wake_idle_without_ipi(cpu);
1849 rq_lock_irqsave(rq, &rf);
1850 if (is_idle_task(rq->curr))
1851 smp_send_reschedule(cpu);
1852 /* Else CPU is not idle, do nothing here: */
1853 rq_unlock_irqrestore(rq, &rf);
1860 bool cpus_share_cache(int this_cpu, int that_cpu)
1862 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1864 #endif /* CONFIG_SMP */
1866 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1868 struct rq *rq = cpu_rq(cpu);
1871 #if defined(CONFIG_SMP)
1872 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1873 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1874 ttwu_queue_remote(p, cpu, wake_flags);
1880 update_rq_clock(rq);
1881 ttwu_do_activate(rq, p, wake_flags, &rf);
1886 * Notes on Program-Order guarantees on SMP systems.
1890 * The basic program-order guarantee on SMP systems is that when a task [t]
1891 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1892 * execution on its new CPU [c1].
1894 * For migration (of runnable tasks) this is provided by the following means:
1896 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1897 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1898 * rq(c1)->lock (if not at the same time, then in that order).
1899 * C) LOCK of the rq(c1)->lock scheduling in task
1901 * Release/acquire chaining guarantees that B happens after A and C after B.
1902 * Note: the CPU doing B need not be c0 or c1
1911 * UNLOCK rq(0)->lock
1913 * LOCK rq(0)->lock // orders against CPU0
1915 * UNLOCK rq(0)->lock
1919 * UNLOCK rq(1)->lock
1921 * LOCK rq(1)->lock // orders against CPU2
1924 * UNLOCK rq(1)->lock
1927 * BLOCKING -- aka. SLEEP + WAKEUP
1929 * For blocking we (obviously) need to provide the same guarantee as for
1930 * migration. However the means are completely different as there is no lock
1931 * chain to provide order. Instead we do:
1933 * 1) smp_store_release(X->on_cpu, 0)
1934 * 2) smp_cond_load_acquire(!X->on_cpu)
1938 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1940 * LOCK rq(0)->lock LOCK X->pi_lock
1943 * smp_store_release(X->on_cpu, 0);
1945 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1951 * X->state = RUNNING
1952 * UNLOCK rq(2)->lock
1954 * LOCK rq(2)->lock // orders against CPU1
1957 * UNLOCK rq(2)->lock
1960 * UNLOCK rq(0)->lock
1963 * However, for wakeups there is a second guarantee we must provide, namely we
1964 * must ensure that CONDITION=1 done by the caller can not be reordered with
1965 * accesses to the task state; see try_to_wake_up() and set_current_state().
1969 * try_to_wake_up - wake up a thread
1970 * @p: the thread to be awakened
1971 * @state: the mask of task states that can be woken
1972 * @wake_flags: wake modifier flags (WF_*)
1974 * If (@state & @p->state) @p->state = TASK_RUNNING.
1976 * If the task was not queued/runnable, also place it back on a runqueue.
1978 * Atomic against schedule() which would dequeue a task, also see
1979 * set_current_state().
1981 * This function executes a full memory barrier before accessing the task
1982 * state; see set_current_state().
1984 * Return: %true if @p->state changes (an actual wakeup was done),
1988 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1990 unsigned long flags;
1991 int cpu, success = 0;
1994 * If we are going to wake up a thread waiting for CONDITION we
1995 * need to ensure that CONDITION=1 done by the caller can not be
1996 * reordered with p->state check below. This pairs with mb() in
1997 * set_current_state() the waiting thread does.
1999 raw_spin_lock_irqsave(&p->pi_lock, flags);
2000 smp_mb__after_spinlock();
2001 if (!(p->state & state))
2004 trace_sched_waking(p);
2006 /* We're going to change ->state: */
2011 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2012 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2013 * in smp_cond_load_acquire() below.
2015 * sched_ttwu_pending() try_to_wake_up()
2016 * STORE p->on_rq = 1 LOAD p->state
2019 * __schedule() (switch to task 'p')
2020 * LOCK rq->lock smp_rmb();
2021 * smp_mb__after_spinlock();
2025 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2027 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2028 * __schedule(). See the comment for smp_mb__after_spinlock().
2031 if (p->on_rq && ttwu_remote(p, wake_flags))
2036 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2037 * possible to, falsely, observe p->on_cpu == 0.
2039 * One must be running (->on_cpu == 1) in order to remove oneself
2040 * from the runqueue.
2042 * __schedule() (switch to task 'p') try_to_wake_up()
2043 * STORE p->on_cpu = 1 LOAD p->on_rq
2046 * __schedule() (put 'p' to sleep)
2047 * LOCK rq->lock smp_rmb();
2048 * smp_mb__after_spinlock();
2049 * STORE p->on_rq = 0 LOAD p->on_cpu
2051 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2052 * __schedule(). See the comment for smp_mb__after_spinlock().
2057 * If the owning (remote) CPU is still in the middle of schedule() with
2058 * this task as prev, wait until its done referencing the task.
2060 * Pairs with the smp_store_release() in finish_task().
2062 * This ensures that tasks getting woken will be fully ordered against
2063 * their previous state and preserve Program Order.
2065 smp_cond_load_acquire(&p->on_cpu, !VAL);
2067 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2068 p->state = TASK_WAKING;
2071 delayacct_blkio_end(p);
2072 atomic_dec(&task_rq(p)->nr_iowait);
2075 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2076 if (task_cpu(p) != cpu) {
2077 wake_flags |= WF_MIGRATED;
2078 psi_ttwu_dequeue(p);
2079 set_task_cpu(p, cpu);
2082 #else /* CONFIG_SMP */
2085 delayacct_blkio_end(p);
2086 atomic_dec(&task_rq(p)->nr_iowait);
2089 #endif /* CONFIG_SMP */
2091 ttwu_queue(p, cpu, wake_flags);
2093 ttwu_stat(p, cpu, wake_flags);
2095 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2101 * wake_up_process - Wake up a specific process
2102 * @p: The process to be woken up.
2104 * Attempt to wake up the nominated process and move it to the set of runnable
2107 * Return: 1 if the process was woken up, 0 if it was already running.
2109 * This function executes a full memory barrier before accessing the task state.
2111 int wake_up_process(struct task_struct *p)
2113 return try_to_wake_up(p, TASK_NORMAL, 0);
2115 EXPORT_SYMBOL(wake_up_process);
2117 int wake_up_state(struct task_struct *p, unsigned int state)
2119 return try_to_wake_up(p, state, 0);
2123 * Perform scheduler related setup for a newly forked process p.
2124 * p is forked by current.
2126 * __sched_fork() is basic setup used by init_idle() too:
2128 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2133 p->se.exec_start = 0;
2134 p->se.sum_exec_runtime = 0;
2135 p->se.prev_sum_exec_runtime = 0;
2136 p->se.nr_migrations = 0;
2138 INIT_LIST_HEAD(&p->se.group_node);
2140 #ifdef CONFIG_FAIR_GROUP_SCHED
2141 p->se.cfs_rq = NULL;
2144 #ifdef CONFIG_SCHEDSTATS
2145 /* Even if schedstat is disabled, there should not be garbage */
2146 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2149 RB_CLEAR_NODE(&p->dl.rb_node);
2150 init_dl_task_timer(&p->dl);
2151 init_dl_inactive_task_timer(&p->dl);
2152 __dl_clear_params(p);
2154 INIT_LIST_HEAD(&p->rt.run_list);
2156 p->rt.time_slice = sched_rr_timeslice;
2160 #ifdef CONFIG_PREEMPT_NOTIFIERS
2161 INIT_HLIST_HEAD(&p->preempt_notifiers);
2164 #ifdef CONFIG_COMPACTION
2165 p->capture_control = NULL;
2167 init_numa_balancing(clone_flags, p);
2170 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2172 #ifdef CONFIG_NUMA_BALANCING
2174 void set_numabalancing_state(bool enabled)
2177 static_branch_enable(&sched_numa_balancing);
2179 static_branch_disable(&sched_numa_balancing);
2182 #ifdef CONFIG_PROC_SYSCTL
2183 int sysctl_numa_balancing(struct ctl_table *table, int write,
2184 void __user *buffer, size_t *lenp, loff_t *ppos)
2188 int state = static_branch_likely(&sched_numa_balancing);
2190 if (write && !capable(CAP_SYS_ADMIN))
2195 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2199 set_numabalancing_state(state);
2205 #ifdef CONFIG_SCHEDSTATS
2207 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2208 static bool __initdata __sched_schedstats = false;
2210 static void set_schedstats(bool enabled)
2213 static_branch_enable(&sched_schedstats);
2215 static_branch_disable(&sched_schedstats);
2218 void force_schedstat_enabled(void)
2220 if (!schedstat_enabled()) {
2221 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2222 static_branch_enable(&sched_schedstats);
2226 static int __init setup_schedstats(char *str)
2233 * This code is called before jump labels have been set up, so we can't
2234 * change the static branch directly just yet. Instead set a temporary
2235 * variable so init_schedstats() can do it later.
2237 if (!strcmp(str, "enable")) {
2238 __sched_schedstats = true;
2240 } else if (!strcmp(str, "disable")) {
2241 __sched_schedstats = false;
2246 pr_warn("Unable to parse schedstats=\n");
2250 __setup("schedstats=", setup_schedstats);
2252 static void __init init_schedstats(void)
2254 set_schedstats(__sched_schedstats);
2257 #ifdef CONFIG_PROC_SYSCTL
2258 int sysctl_schedstats(struct ctl_table *table, int write,
2259 void __user *buffer, size_t *lenp, loff_t *ppos)
2263 int state = static_branch_likely(&sched_schedstats);
2265 if (write && !capable(CAP_SYS_ADMIN))
2270 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2274 set_schedstats(state);
2277 #endif /* CONFIG_PROC_SYSCTL */
2278 #else /* !CONFIG_SCHEDSTATS */
2279 static inline void init_schedstats(void) {}
2280 #endif /* CONFIG_SCHEDSTATS */
2283 * fork()/clone()-time setup:
2285 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2287 unsigned long flags;
2289 __sched_fork(clone_flags, p);
2291 * We mark the process as NEW here. This guarantees that
2292 * nobody will actually run it, and a signal or other external
2293 * event cannot wake it up and insert it on the runqueue either.
2295 p->state = TASK_NEW;
2298 * Make sure we do not leak PI boosting priority to the child.
2300 p->prio = current->normal_prio;
2303 * Revert to default priority/policy on fork if requested.
2305 if (unlikely(p->sched_reset_on_fork)) {
2306 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2307 p->policy = SCHED_NORMAL;
2308 p->static_prio = NICE_TO_PRIO(0);
2310 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2311 p->static_prio = NICE_TO_PRIO(0);
2313 p->prio = p->normal_prio = __normal_prio(p);
2314 set_load_weight(p, false);
2317 * We don't need the reset flag anymore after the fork. It has
2318 * fulfilled its duty:
2320 p->sched_reset_on_fork = 0;
2323 if (dl_prio(p->prio))
2325 else if (rt_prio(p->prio))
2326 p->sched_class = &rt_sched_class;
2328 p->sched_class = &fair_sched_class;
2330 init_entity_runnable_average(&p->se);
2333 * The child is not yet in the pid-hash so no cgroup attach races,
2334 * and the cgroup is pinned to this child due to cgroup_fork()
2335 * is ran before sched_fork().
2337 * Silence PROVE_RCU.
2339 raw_spin_lock_irqsave(&p->pi_lock, flags);
2341 * We're setting the CPU for the first time, we don't migrate,
2342 * so use __set_task_cpu().
2344 __set_task_cpu(p, smp_processor_id());
2345 if (p->sched_class->task_fork)
2346 p->sched_class->task_fork(p);
2347 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2349 #ifdef CONFIG_SCHED_INFO
2350 if (likely(sched_info_on()))
2351 memset(&p->sched_info, 0, sizeof(p->sched_info));
2353 #if defined(CONFIG_SMP)
2356 init_task_preempt_count(p);
2358 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2359 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2364 unsigned long to_ratio(u64 period, u64 runtime)
2366 if (runtime == RUNTIME_INF)
2370 * Doing this here saves a lot of checks in all
2371 * the calling paths, and returning zero seems
2372 * safe for them anyway.
2377 return div64_u64(runtime << BW_SHIFT, period);
2381 * wake_up_new_task - wake up a newly created task for the first time.
2383 * This function will do some initial scheduler statistics housekeeping
2384 * that must be done for every newly created context, then puts the task
2385 * on the runqueue and wakes it.
2387 void wake_up_new_task(struct task_struct *p)
2392 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2393 p->state = TASK_RUNNING;
2396 * Fork balancing, do it here and not earlier because:
2397 * - cpus_allowed can change in the fork path
2398 * - any previously selected CPU might disappear through hotplug
2400 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2401 * as we're not fully set-up yet.
2403 p->recent_used_cpu = task_cpu(p);
2404 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2406 rq = __task_rq_lock(p, &rf);
2407 update_rq_clock(rq);
2408 post_init_entity_util_avg(p);
2410 activate_task(rq, p, ENQUEUE_NOCLOCK);
2411 trace_sched_wakeup_new(p);
2412 check_preempt_curr(rq, p, WF_FORK);
2414 if (p->sched_class->task_woken) {
2416 * Nothing relies on rq->lock after this, so its fine to
2419 rq_unpin_lock(rq, &rf);
2420 p->sched_class->task_woken(rq, p);
2421 rq_repin_lock(rq, &rf);
2424 task_rq_unlock(rq, p, &rf);
2427 #ifdef CONFIG_PREEMPT_NOTIFIERS
2429 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2431 void preempt_notifier_inc(void)
2433 static_branch_inc(&preempt_notifier_key);
2435 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2437 void preempt_notifier_dec(void)
2439 static_branch_dec(&preempt_notifier_key);
2441 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2444 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2445 * @notifier: notifier struct to register
2447 void preempt_notifier_register(struct preempt_notifier *notifier)
2449 if (!static_branch_unlikely(&preempt_notifier_key))
2450 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2452 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2454 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2457 * preempt_notifier_unregister - no longer interested in preemption notifications
2458 * @notifier: notifier struct to unregister
2460 * This is *not* safe to call from within a preemption notifier.
2462 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2464 hlist_del(¬ifier->link);
2466 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2468 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2470 struct preempt_notifier *notifier;
2472 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2473 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2476 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2478 if (static_branch_unlikely(&preempt_notifier_key))
2479 __fire_sched_in_preempt_notifiers(curr);
2483 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2484 struct task_struct *next)
2486 struct preempt_notifier *notifier;
2488 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2489 notifier->ops->sched_out(notifier, next);
2492 static __always_inline void
2493 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2494 struct task_struct *next)
2496 if (static_branch_unlikely(&preempt_notifier_key))
2497 __fire_sched_out_preempt_notifiers(curr, next);
2500 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2502 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2507 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2508 struct task_struct *next)
2512 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2514 static inline void prepare_task(struct task_struct *next)
2518 * Claim the task as running, we do this before switching to it
2519 * such that any running task will have this set.
2525 static inline void finish_task(struct task_struct *prev)
2529 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2530 * We must ensure this doesn't happen until the switch is completely
2533 * In particular, the load of prev->state in finish_task_switch() must
2534 * happen before this.
2536 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2538 smp_store_release(&prev->on_cpu, 0);
2543 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
2546 * Since the runqueue lock will be released by the next
2547 * task (which is an invalid locking op but in the case
2548 * of the scheduler it's an obvious special-case), so we
2549 * do an early lockdep release here:
2551 rq_unpin_lock(rq, rf);
2552 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2553 #ifdef CONFIG_DEBUG_SPINLOCK
2554 /* this is a valid case when another task releases the spinlock */
2555 rq->lock.owner = next;
2559 static inline void finish_lock_switch(struct rq *rq)
2562 * If we are tracking spinlock dependencies then we have to
2563 * fix up the runqueue lock - which gets 'carried over' from
2564 * prev into current:
2566 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
2567 raw_spin_unlock_irq(&rq->lock);
2571 * NOP if the arch has not defined these:
2574 #ifndef prepare_arch_switch
2575 # define prepare_arch_switch(next) do { } while (0)
2578 #ifndef finish_arch_post_lock_switch
2579 # define finish_arch_post_lock_switch() do { } while (0)
2583 * prepare_task_switch - prepare to switch tasks
2584 * @rq: the runqueue preparing to switch
2585 * @prev: the current task that is being switched out
2586 * @next: the task we are going to switch to.
2588 * This is called with the rq lock held and interrupts off. It must
2589 * be paired with a subsequent finish_task_switch after the context
2592 * prepare_task_switch sets up locking and calls architecture specific
2596 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2597 struct task_struct *next)
2599 kcov_prepare_switch(prev);
2600 sched_info_switch(rq, prev, next);
2601 perf_event_task_sched_out(prev, next);
2603 fire_sched_out_preempt_notifiers(prev, next);
2605 prepare_arch_switch(next);
2609 * finish_task_switch - clean up after a task-switch
2610 * @prev: the thread we just switched away from.
2612 * finish_task_switch must be called after the context switch, paired
2613 * with a prepare_task_switch call before the context switch.
2614 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2615 * and do any other architecture-specific cleanup actions.
2617 * Note that we may have delayed dropping an mm in context_switch(). If
2618 * so, we finish that here outside of the runqueue lock. (Doing it
2619 * with the lock held can cause deadlocks; see schedule() for
2622 * The context switch have flipped the stack from under us and restored the
2623 * local variables which were saved when this task called schedule() in the
2624 * past. prev == current is still correct but we need to recalculate this_rq
2625 * because prev may have moved to another CPU.
2627 static struct rq *finish_task_switch(struct task_struct *prev)
2628 __releases(rq->lock)
2630 struct rq *rq = this_rq();
2631 struct mm_struct *mm = rq->prev_mm;
2635 * The previous task will have left us with a preempt_count of 2
2636 * because it left us after:
2639 * preempt_disable(); // 1
2641 * raw_spin_lock_irq(&rq->lock) // 2
2643 * Also, see FORK_PREEMPT_COUNT.
2645 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2646 "corrupted preempt_count: %s/%d/0x%x\n",
2647 current->comm, current->pid, preempt_count()))
2648 preempt_count_set(FORK_PREEMPT_COUNT);
2653 * A task struct has one reference for the use as "current".
2654 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2655 * schedule one last time. The schedule call will never return, and
2656 * the scheduled task must drop that reference.
2658 * We must observe prev->state before clearing prev->on_cpu (in
2659 * finish_task), otherwise a concurrent wakeup can get prev
2660 * running on another CPU and we could rave with its RUNNING -> DEAD
2661 * transition, resulting in a double drop.
2663 prev_state = prev->state;
2664 vtime_task_switch(prev);
2665 perf_event_task_sched_in(prev, current);
2667 finish_lock_switch(rq);
2668 finish_arch_post_lock_switch();
2669 kcov_finish_switch(current);
2671 fire_sched_in_preempt_notifiers(current);
2673 * When switching through a kernel thread, the loop in
2674 * membarrier_{private,global}_expedited() may have observed that
2675 * kernel thread and not issued an IPI. It is therefore possible to
2676 * schedule between user->kernel->user threads without passing though
2677 * switch_mm(). Membarrier requires a barrier after storing to
2678 * rq->curr, before returning to userspace, so provide them here:
2680 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
2681 * provided by mmdrop(),
2682 * - a sync_core for SYNC_CORE.
2685 membarrier_mm_sync_core_before_usermode(mm);
2688 if (unlikely(prev_state == TASK_DEAD)) {
2689 if (prev->sched_class->task_dead)
2690 prev->sched_class->task_dead(prev);
2693 * Remove function-return probe instances associated with this
2694 * task and put them back on the free list.
2696 kprobe_flush_task(prev);
2698 /* Task is done with its stack. */
2699 put_task_stack(prev);
2701 put_task_struct(prev);
2704 tick_nohz_task_switch();
2710 /* rq->lock is NOT held, but preemption is disabled */
2711 static void __balance_callback(struct rq *rq)
2713 struct callback_head *head, *next;
2714 void (*func)(struct rq *rq);
2715 unsigned long flags;
2717 raw_spin_lock_irqsave(&rq->lock, flags);
2718 head = rq->balance_callback;
2719 rq->balance_callback = NULL;
2721 func = (void (*)(struct rq *))head->func;
2728 raw_spin_unlock_irqrestore(&rq->lock, flags);
2731 static inline void balance_callback(struct rq *rq)
2733 if (unlikely(rq->balance_callback))
2734 __balance_callback(rq);
2739 static inline void balance_callback(struct rq *rq)
2746 * schedule_tail - first thing a freshly forked thread must call.
2747 * @prev: the thread we just switched away from.
2749 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2750 __releases(rq->lock)
2755 * New tasks start with FORK_PREEMPT_COUNT, see there and
2756 * finish_task_switch() for details.
2758 * finish_task_switch() will drop rq->lock() and lower preempt_count
2759 * and the preempt_enable() will end up enabling preemption (on
2760 * PREEMPT_COUNT kernels).
2763 rq = finish_task_switch(prev);
2764 balance_callback(rq);
2767 if (current->set_child_tid)
2768 put_user(task_pid_vnr(current), current->set_child_tid);
2770 calculate_sigpending();
2774 * context_switch - switch to the new MM and the new thread's register state.
2776 static __always_inline struct rq *
2777 context_switch(struct rq *rq, struct task_struct *prev,
2778 struct task_struct *next, struct rq_flags *rf)
2780 struct mm_struct *mm, *oldmm;
2782 prepare_task_switch(rq, prev, next);
2785 oldmm = prev->active_mm;
2787 * For paravirt, this is coupled with an exit in switch_to to
2788 * combine the page table reload and the switch backend into
2791 arch_start_context_switch(prev);
2794 * If mm is non-NULL, we pass through switch_mm(). If mm is
2795 * NULL, we will pass through mmdrop() in finish_task_switch().
2796 * Both of these contain the full memory barrier required by
2797 * membarrier after storing to rq->curr, before returning to
2801 next->active_mm = oldmm;
2803 enter_lazy_tlb(oldmm, next);
2805 switch_mm_irqs_off(oldmm, mm, next);
2808 prev->active_mm = NULL;
2809 rq->prev_mm = oldmm;
2812 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2814 prepare_lock_switch(rq, next, rf);
2816 /* Here we just switch the register state and the stack. */
2817 switch_to(prev, next, prev);
2820 return finish_task_switch(prev);
2824 * nr_running and nr_context_switches:
2826 * externally visible scheduler statistics: current number of runnable
2827 * threads, total number of context switches performed since bootup.
2829 unsigned long nr_running(void)
2831 unsigned long i, sum = 0;
2833 for_each_online_cpu(i)
2834 sum += cpu_rq(i)->nr_running;
2840 * Check if only the current task is running on the CPU.
2842 * Caution: this function does not check that the caller has disabled
2843 * preemption, thus the result might have a time-of-check-to-time-of-use
2844 * race. The caller is responsible to use it correctly, for example:
2846 * - from a non-preemptible section (of course)
2848 * - from a thread that is bound to a single CPU
2850 * - in a loop with very short iterations (e.g. a polling loop)
2852 bool single_task_running(void)
2854 return raw_rq()->nr_running == 1;
2856 EXPORT_SYMBOL(single_task_running);
2858 unsigned long long nr_context_switches(void)
2861 unsigned long long sum = 0;
2863 for_each_possible_cpu(i)
2864 sum += cpu_rq(i)->nr_switches;
2870 * Consumers of these two interfaces, like for example the cpuidle menu
2871 * governor, are using nonsensical data. Preferring shallow idle state selection
2872 * for a CPU that has IO-wait which might not even end up running the task when
2873 * it does become runnable.
2876 unsigned long nr_iowait_cpu(int cpu)
2878 return atomic_read(&cpu_rq(cpu)->nr_iowait);
2882 * IO-wait accounting, and how its mostly bollocks (on SMP).
2884 * The idea behind IO-wait account is to account the idle time that we could
2885 * have spend running if it were not for IO. That is, if we were to improve the
2886 * storage performance, we'd have a proportional reduction in IO-wait time.
2888 * This all works nicely on UP, where, when a task blocks on IO, we account
2889 * idle time as IO-wait, because if the storage were faster, it could've been
2890 * running and we'd not be idle.
2892 * This has been extended to SMP, by doing the same for each CPU. This however
2895 * Imagine for instance the case where two tasks block on one CPU, only the one
2896 * CPU will have IO-wait accounted, while the other has regular idle. Even
2897 * though, if the storage were faster, both could've ran at the same time,
2898 * utilising both CPUs.
2900 * This means, that when looking globally, the current IO-wait accounting on
2901 * SMP is a lower bound, by reason of under accounting.
2903 * Worse, since the numbers are provided per CPU, they are sometimes
2904 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2905 * associated with any one particular CPU, it can wake to another CPU than it
2906 * blocked on. This means the per CPU IO-wait number is meaningless.
2908 * Task CPU affinities can make all that even more 'interesting'.
2911 unsigned long nr_iowait(void)
2913 unsigned long i, sum = 0;
2915 for_each_possible_cpu(i)
2916 sum += nr_iowait_cpu(i);
2924 * sched_exec - execve() is a valuable balancing opportunity, because at
2925 * this point the task has the smallest effective memory and cache footprint.
2927 void sched_exec(void)
2929 struct task_struct *p = current;
2930 unsigned long flags;
2933 raw_spin_lock_irqsave(&p->pi_lock, flags);
2934 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2935 if (dest_cpu == smp_processor_id())
2938 if (likely(cpu_active(dest_cpu))) {
2939 struct migration_arg arg = { p, dest_cpu };
2941 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2942 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2946 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2951 DEFINE_PER_CPU(struct kernel_stat, kstat);
2952 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2954 EXPORT_PER_CPU_SYMBOL(kstat);
2955 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2958 * The function fair_sched_class.update_curr accesses the struct curr
2959 * and its field curr->exec_start; when called from task_sched_runtime(),
2960 * we observe a high rate of cache misses in practice.
2961 * Prefetching this data results in improved performance.
2963 static inline void prefetch_curr_exec_start(struct task_struct *p)
2965 #ifdef CONFIG_FAIR_GROUP_SCHED
2966 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
2968 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
2971 prefetch(&curr->exec_start);
2975 * Return accounted runtime for the task.
2976 * In case the task is currently running, return the runtime plus current's
2977 * pending runtime that have not been accounted yet.
2979 unsigned long long task_sched_runtime(struct task_struct *p)
2985 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2987 * 64-bit doesn't need locks to atomically read a 64-bit value.
2988 * So we have a optimization chance when the task's delta_exec is 0.
2989 * Reading ->on_cpu is racy, but this is ok.
2991 * If we race with it leaving CPU, we'll take a lock. So we're correct.
2992 * If we race with it entering CPU, unaccounted time is 0. This is
2993 * indistinguishable from the read occurring a few cycles earlier.
2994 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2995 * been accounted, so we're correct here as well.
2997 if (!p->on_cpu || !task_on_rq_queued(p))
2998 return p->se.sum_exec_runtime;
3001 rq = task_rq_lock(p, &rf);
3003 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3004 * project cycles that may never be accounted to this
3005 * thread, breaking clock_gettime().
3007 if (task_current(rq, p) && task_on_rq_queued(p)) {
3008 prefetch_curr_exec_start(p);
3009 update_rq_clock(rq);
3010 p->sched_class->update_curr(rq);
3012 ns = p->se.sum_exec_runtime;
3013 task_rq_unlock(rq, p, &rf);
3019 * This function gets called by the timer code, with HZ frequency.
3020 * We call it with interrupts disabled.
3022 void scheduler_tick(void)
3024 int cpu = smp_processor_id();
3025 struct rq *rq = cpu_rq(cpu);
3026 struct task_struct *curr = rq->curr;
3033 update_rq_clock(rq);
3034 curr->sched_class->task_tick(rq, curr, 0);
3035 cpu_load_update_active(rq);
3036 calc_global_load_tick(rq);
3041 perf_event_task_tick();
3044 rq->idle_balance = idle_cpu(cpu);
3045 trigger_load_balance(rq);
3049 #ifdef CONFIG_NO_HZ_FULL
3053 struct delayed_work work;
3056 static struct tick_work __percpu *tick_work_cpu;
3058 static void sched_tick_remote(struct work_struct *work)
3060 struct delayed_work *dwork = to_delayed_work(work);
3061 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3062 int cpu = twork->cpu;
3063 struct rq *rq = cpu_rq(cpu);
3064 struct task_struct *curr;
3069 * Handle the tick only if it appears the remote CPU is running in full
3070 * dynticks mode. The check is racy by nature, but missing a tick or
3071 * having one too much is no big deal because the scheduler tick updates
3072 * statistics and checks timeslices in a time-independent way, regardless
3073 * of when exactly it is running.
3075 if (idle_cpu(cpu) || !tick_nohz_tick_stopped_cpu(cpu))
3078 rq_lock_irq(rq, &rf);
3080 if (is_idle_task(curr))
3083 update_rq_clock(rq);
3084 delta = rq_clock_task(rq) - curr->se.exec_start;
3087 * Make sure the next tick runs within a reasonable
3090 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3091 curr->sched_class->task_tick(rq, curr, 0);
3094 rq_unlock_irq(rq, &rf);
3098 * Run the remote tick once per second (1Hz). This arbitrary
3099 * frequency is large enough to avoid overload but short enough
3100 * to keep scheduler internal stats reasonably up to date.
3102 queue_delayed_work(system_unbound_wq, dwork, HZ);
3105 static void sched_tick_start(int cpu)
3107 struct tick_work *twork;
3109 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3112 WARN_ON_ONCE(!tick_work_cpu);
3114 twork = per_cpu_ptr(tick_work_cpu, cpu);
3116 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3117 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3120 #ifdef CONFIG_HOTPLUG_CPU
3121 static void sched_tick_stop(int cpu)
3123 struct tick_work *twork;
3125 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3128 WARN_ON_ONCE(!tick_work_cpu);
3130 twork = per_cpu_ptr(tick_work_cpu, cpu);
3131 cancel_delayed_work_sync(&twork->work);
3133 #endif /* CONFIG_HOTPLUG_CPU */
3135 int __init sched_tick_offload_init(void)
3137 tick_work_cpu = alloc_percpu(struct tick_work);
3138 BUG_ON(!tick_work_cpu);
3143 #else /* !CONFIG_NO_HZ_FULL */
3144 static inline void sched_tick_start(int cpu) { }
3145 static inline void sched_tick_stop(int cpu) { }
3148 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3149 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3151 * If the value passed in is equal to the current preempt count
3152 * then we just disabled preemption. Start timing the latency.
3154 static inline void preempt_latency_start(int val)
3156 if (preempt_count() == val) {
3157 unsigned long ip = get_lock_parent_ip();
3158 #ifdef CONFIG_DEBUG_PREEMPT
3159 current->preempt_disable_ip = ip;
3161 trace_preempt_off(CALLER_ADDR0, ip);
3165 void preempt_count_add(int val)
3167 #ifdef CONFIG_DEBUG_PREEMPT
3171 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3174 __preempt_count_add(val);
3175 #ifdef CONFIG_DEBUG_PREEMPT
3177 * Spinlock count overflowing soon?
3179 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3182 preempt_latency_start(val);
3184 EXPORT_SYMBOL(preempt_count_add);
3185 NOKPROBE_SYMBOL(preempt_count_add);
3188 * If the value passed in equals to the current preempt count
3189 * then we just enabled preemption. Stop timing the latency.
3191 static inline void preempt_latency_stop(int val)
3193 if (preempt_count() == val)
3194 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3197 void preempt_count_sub(int val)
3199 #ifdef CONFIG_DEBUG_PREEMPT
3203 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3206 * Is the spinlock portion underflowing?
3208 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3209 !(preempt_count() & PREEMPT_MASK)))
3213 preempt_latency_stop(val);
3214 __preempt_count_sub(val);
3216 EXPORT_SYMBOL(preempt_count_sub);
3217 NOKPROBE_SYMBOL(preempt_count_sub);
3220 static inline void preempt_latency_start(int val) { }
3221 static inline void preempt_latency_stop(int val) { }
3224 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3226 #ifdef CONFIG_DEBUG_PREEMPT
3227 return p->preempt_disable_ip;
3234 * Print scheduling while atomic bug:
3236 static noinline void __schedule_bug(struct task_struct *prev)
3238 /* Save this before calling printk(), since that will clobber it */
3239 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3241 if (oops_in_progress)
3244 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3245 prev->comm, prev->pid, preempt_count());
3247 debug_show_held_locks(prev);
3249 if (irqs_disabled())
3250 print_irqtrace_events(prev);
3251 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3252 && in_atomic_preempt_off()) {
3253 pr_err("Preemption disabled at:");
3254 print_ip_sym(preempt_disable_ip);
3258 panic("scheduling while atomic\n");
3261 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3265 * Various schedule()-time debugging checks and statistics:
3267 static inline void schedule_debug(struct task_struct *prev)
3269 #ifdef CONFIG_SCHED_STACK_END_CHECK
3270 if (task_stack_end_corrupted(prev))
3271 panic("corrupted stack end detected inside scheduler\n");
3274 if (unlikely(in_atomic_preempt_off())) {
3275 __schedule_bug(prev);
3276 preempt_count_set(PREEMPT_DISABLED);
3280 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3282 schedstat_inc(this_rq()->sched_count);
3286 * Pick up the highest-prio task:
3288 static inline struct task_struct *
3289 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3291 const struct sched_class *class;
3292 struct task_struct *p;
3295 * Optimization: we know that if all tasks are in the fair class we can
3296 * call that function directly, but only if the @prev task wasn't of a
3297 * higher scheduling class, because otherwise those loose the
3298 * opportunity to pull in more work from other CPUs.
3300 if (likely((prev->sched_class == &idle_sched_class ||
3301 prev->sched_class == &fair_sched_class) &&
3302 rq->nr_running == rq->cfs.h_nr_running)) {
3304 p = fair_sched_class.pick_next_task(rq, prev, rf);
3305 if (unlikely(p == RETRY_TASK))
3308 /* Assumes fair_sched_class->next == idle_sched_class */
3310 p = idle_sched_class.pick_next_task(rq, prev, rf);
3316 for_each_class(class) {
3317 p = class->pick_next_task(rq, prev, rf);
3319 if (unlikely(p == RETRY_TASK))
3325 /* The idle class should always have a runnable task: */
3330 * __schedule() is the main scheduler function.
3332 * The main means of driving the scheduler and thus entering this function are:
3334 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3336 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3337 * paths. For example, see arch/x86/entry_64.S.
3339 * To drive preemption between tasks, the scheduler sets the flag in timer
3340 * interrupt handler scheduler_tick().
3342 * 3. Wakeups don't really cause entry into schedule(). They add a
3343 * task to the run-queue and that's it.
3345 * Now, if the new task added to the run-queue preempts the current
3346 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3347 * called on the nearest possible occasion:
3349 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3351 * - in syscall or exception context, at the next outmost
3352 * preempt_enable(). (this might be as soon as the wake_up()'s
3355 * - in IRQ context, return from interrupt-handler to
3356 * preemptible context
3358 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3361 * - cond_resched() call
3362 * - explicit schedule() call
3363 * - return from syscall or exception to user-space
3364 * - return from interrupt-handler to user-space
3366 * WARNING: must be called with preemption disabled!
3368 static void __sched notrace __schedule(bool preempt)
3370 struct task_struct *prev, *next;
3371 unsigned long *switch_count;
3376 cpu = smp_processor_id();
3380 schedule_debug(prev);
3382 if (sched_feat(HRTICK))
3385 local_irq_disable();
3386 rcu_note_context_switch(preempt);
3389 * Make sure that signal_pending_state()->signal_pending() below
3390 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3391 * done by the caller to avoid the race with signal_wake_up().
3393 * The membarrier system call requires a full memory barrier
3394 * after coming from user-space, before storing to rq->curr.
3397 smp_mb__after_spinlock();
3399 /* Promote REQ to ACT */
3400 rq->clock_update_flags <<= 1;
3401 update_rq_clock(rq);
3403 switch_count = &prev->nivcsw;
3404 if (!preempt && prev->state) {
3405 if (signal_pending_state(prev->state, prev)) {
3406 prev->state = TASK_RUNNING;
3408 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3410 if (prev->in_iowait) {
3411 atomic_inc(&rq->nr_iowait);
3412 delayacct_blkio_start();
3415 switch_count = &prev->nvcsw;
3418 next = pick_next_task(rq, prev, &rf);
3419 clear_tsk_need_resched(prev);
3420 clear_preempt_need_resched();
3422 if (likely(prev != next)) {
3426 * The membarrier system call requires each architecture
3427 * to have a full memory barrier after updating
3428 * rq->curr, before returning to user-space.
3430 * Here are the schemes providing that barrier on the
3431 * various architectures:
3432 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3433 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3434 * - finish_lock_switch() for weakly-ordered
3435 * architectures where spin_unlock is a full barrier,
3436 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3437 * is a RELEASE barrier),
3441 trace_sched_switch(preempt, prev, next);
3443 /* Also unlocks the rq: */
3444 rq = context_switch(rq, prev, next, &rf);
3446 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3447 rq_unlock_irq(rq, &rf);
3450 balance_callback(rq);
3453 void __noreturn do_task_dead(void)
3455 /* Causes final put_task_struct in finish_task_switch(): */
3456 set_special_state(TASK_DEAD);
3458 /* Tell freezer to ignore us: */
3459 current->flags |= PF_NOFREEZE;
3464 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3469 static inline void sched_submit_work(struct task_struct *tsk)
3471 if (!tsk->state || tsk_is_pi_blocked(tsk))
3475 * If a worker went to sleep, notify and ask workqueue whether
3476 * it wants to wake up a task to maintain concurrency.
3477 * As this function is called inside the schedule() context,
3478 * we disable preemption to avoid it calling schedule() again
3479 * in the possible wakeup of a kworker.
3481 if (tsk->flags & PF_WQ_WORKER) {
3483 wq_worker_sleeping(tsk);
3484 preempt_enable_no_resched();
3488 * If we are going to sleep and we have plugged IO queued,
3489 * make sure to submit it to avoid deadlocks.
3491 if (blk_needs_flush_plug(tsk))
3492 blk_schedule_flush_plug(tsk);
3495 static void sched_update_worker(struct task_struct *tsk)
3497 if (tsk->flags & PF_WQ_WORKER)
3498 wq_worker_running(tsk);
3501 asmlinkage __visible void __sched schedule(void)
3503 struct task_struct *tsk = current;
3505 sched_submit_work(tsk);
3509 sched_preempt_enable_no_resched();
3510 } while (need_resched());
3511 sched_update_worker(tsk);
3513 EXPORT_SYMBOL(schedule);
3516 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3517 * state (have scheduled out non-voluntarily) by making sure that all
3518 * tasks have either left the run queue or have gone into user space.
3519 * As idle tasks do not do either, they must not ever be preempted
3520 * (schedule out non-voluntarily).
3522 * schedule_idle() is similar to schedule_preempt_disable() except that it
3523 * never enables preemption because it does not call sched_submit_work().
3525 void __sched schedule_idle(void)
3528 * As this skips calling sched_submit_work(), which the idle task does
3529 * regardless because that function is a nop when the task is in a
3530 * TASK_RUNNING state, make sure this isn't used someplace that the
3531 * current task can be in any other state. Note, idle is always in the
3532 * TASK_RUNNING state.
3534 WARN_ON_ONCE(current->state);
3537 } while (need_resched());
3540 #ifdef CONFIG_CONTEXT_TRACKING
3541 asmlinkage __visible void __sched schedule_user(void)
3544 * If we come here after a random call to set_need_resched(),
3545 * or we have been woken up remotely but the IPI has not yet arrived,
3546 * we haven't yet exited the RCU idle mode. Do it here manually until
3547 * we find a better solution.
3549 * NB: There are buggy callers of this function. Ideally we
3550 * should warn if prev_state != CONTEXT_USER, but that will trigger
3551 * too frequently to make sense yet.
3553 enum ctx_state prev_state = exception_enter();
3555 exception_exit(prev_state);
3560 * schedule_preempt_disabled - called with preemption disabled
3562 * Returns with preemption disabled. Note: preempt_count must be 1
3564 void __sched schedule_preempt_disabled(void)
3566 sched_preempt_enable_no_resched();
3571 static void __sched notrace preempt_schedule_common(void)
3575 * Because the function tracer can trace preempt_count_sub()
3576 * and it also uses preempt_enable/disable_notrace(), if
3577 * NEED_RESCHED is set, the preempt_enable_notrace() called
3578 * by the function tracer will call this function again and
3579 * cause infinite recursion.
3581 * Preemption must be disabled here before the function
3582 * tracer can trace. Break up preempt_disable() into two
3583 * calls. One to disable preemption without fear of being
3584 * traced. The other to still record the preemption latency,
3585 * which can also be traced by the function tracer.
3587 preempt_disable_notrace();
3588 preempt_latency_start(1);
3590 preempt_latency_stop(1);
3591 preempt_enable_no_resched_notrace();
3594 * Check again in case we missed a preemption opportunity
3595 * between schedule and now.
3597 } while (need_resched());
3600 #ifdef CONFIG_PREEMPT
3602 * this is the entry point to schedule() from in-kernel preemption
3603 * off of preempt_enable. Kernel preemptions off return from interrupt
3604 * occur there and call schedule directly.
3606 asmlinkage __visible void __sched notrace preempt_schedule(void)
3609 * If there is a non-zero preempt_count or interrupts are disabled,
3610 * we do not want to preempt the current task. Just return..
3612 if (likely(!preemptible()))
3615 preempt_schedule_common();
3617 NOKPROBE_SYMBOL(preempt_schedule);
3618 EXPORT_SYMBOL(preempt_schedule);
3621 * preempt_schedule_notrace - preempt_schedule called by tracing
3623 * The tracing infrastructure uses preempt_enable_notrace to prevent
3624 * recursion and tracing preempt enabling caused by the tracing
3625 * infrastructure itself. But as tracing can happen in areas coming
3626 * from userspace or just about to enter userspace, a preempt enable
3627 * can occur before user_exit() is called. This will cause the scheduler
3628 * to be called when the system is still in usermode.
3630 * To prevent this, the preempt_enable_notrace will use this function
3631 * instead of preempt_schedule() to exit user context if needed before
3632 * calling the scheduler.
3634 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3636 enum ctx_state prev_ctx;
3638 if (likely(!preemptible()))
3643 * Because the function tracer can trace preempt_count_sub()
3644 * and it also uses preempt_enable/disable_notrace(), if
3645 * NEED_RESCHED is set, the preempt_enable_notrace() called
3646 * by the function tracer will call this function again and
3647 * cause infinite recursion.
3649 * Preemption must be disabled here before the function
3650 * tracer can trace. Break up preempt_disable() into two
3651 * calls. One to disable preemption without fear of being
3652 * traced. The other to still record the preemption latency,
3653 * which can also be traced by the function tracer.
3655 preempt_disable_notrace();
3656 preempt_latency_start(1);
3658 * Needs preempt disabled in case user_exit() is traced
3659 * and the tracer calls preempt_enable_notrace() causing
3660 * an infinite recursion.
3662 prev_ctx = exception_enter();
3664 exception_exit(prev_ctx);
3666 preempt_latency_stop(1);
3667 preempt_enable_no_resched_notrace();
3668 } while (need_resched());
3670 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3672 #endif /* CONFIG_PREEMPT */
3675 * this is the entry point to schedule() from kernel preemption
3676 * off of irq context.
3677 * Note, that this is called and return with irqs disabled. This will
3678 * protect us against recursive calling from irq.
3680 asmlinkage __visible void __sched preempt_schedule_irq(void)
3682 enum ctx_state prev_state;
3684 /* Catch callers which need to be fixed */
3685 BUG_ON(preempt_count() || !irqs_disabled());
3687 prev_state = exception_enter();
3693 local_irq_disable();
3694 sched_preempt_enable_no_resched();
3695 } while (need_resched());
3697 exception_exit(prev_state);
3700 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
3703 return try_to_wake_up(curr->private, mode, wake_flags);
3705 EXPORT_SYMBOL(default_wake_function);
3707 #ifdef CONFIG_RT_MUTEXES
3709 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
3712 prio = min(prio, pi_task->prio);
3717 static inline int rt_effective_prio(struct task_struct *p, int prio)
3719 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3721 return __rt_effective_prio(pi_task, prio);
3725 * rt_mutex_setprio - set the current priority of a task
3727 * @pi_task: donor task
3729 * This function changes the 'effective' priority of a task. It does
3730 * not touch ->normal_prio like __setscheduler().
3732 * Used by the rt_mutex code to implement priority inheritance
3733 * logic. Call site only calls if the priority of the task changed.
3735 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
3737 int prio, oldprio, queued, running, queue_flag =
3738 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
3739 const struct sched_class *prev_class;
3743 /* XXX used to be waiter->prio, not waiter->task->prio */
3744 prio = __rt_effective_prio(pi_task, p->normal_prio);
3747 * If nothing changed; bail early.
3749 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
3752 rq = __task_rq_lock(p, &rf);
3753 update_rq_clock(rq);
3755 * Set under pi_lock && rq->lock, such that the value can be used under
3758 * Note that there is loads of tricky to make this pointer cache work
3759 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3760 * ensure a task is de-boosted (pi_task is set to NULL) before the
3761 * task is allowed to run again (and can exit). This ensures the pointer
3762 * points to a blocked task -- which guaratees the task is present.
3764 p->pi_top_task = pi_task;
3767 * For FIFO/RR we only need to set prio, if that matches we're done.
3769 if (prio == p->prio && !dl_prio(prio))
3773 * Idle task boosting is a nono in general. There is one
3774 * exception, when PREEMPT_RT and NOHZ is active:
3776 * The idle task calls get_next_timer_interrupt() and holds
3777 * the timer wheel base->lock on the CPU and another CPU wants
3778 * to access the timer (probably to cancel it). We can safely
3779 * ignore the boosting request, as the idle CPU runs this code
3780 * with interrupts disabled and will complete the lock
3781 * protected section without being interrupted. So there is no
3782 * real need to boost.
3784 if (unlikely(p == rq->idle)) {
3785 WARN_ON(p != rq->curr);
3786 WARN_ON(p->pi_blocked_on);
3790 trace_sched_pi_setprio(p, pi_task);
3793 if (oldprio == prio)
3794 queue_flag &= ~DEQUEUE_MOVE;
3796 prev_class = p->sched_class;
3797 queued = task_on_rq_queued(p);
3798 running = task_current(rq, p);
3800 dequeue_task(rq, p, queue_flag);
3802 put_prev_task(rq, p);
3805 * Boosting condition are:
3806 * 1. -rt task is running and holds mutex A
3807 * --> -dl task blocks on mutex A
3809 * 2. -dl task is running and holds mutex A
3810 * --> -dl task blocks on mutex A and could preempt the
3813 if (dl_prio(prio)) {
3814 if (!dl_prio(p->normal_prio) ||
3815 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3816 p->dl.dl_boosted = 1;
3817 queue_flag |= ENQUEUE_REPLENISH;
3819 p->dl.dl_boosted = 0;
3820 p->sched_class = &dl_sched_class;
3821 } else if (rt_prio(prio)) {
3822 if (dl_prio(oldprio))
3823 p->dl.dl_boosted = 0;
3825 queue_flag |= ENQUEUE_HEAD;
3826 p->sched_class = &rt_sched_class;
3828 if (dl_prio(oldprio))
3829 p->dl.dl_boosted = 0;
3830 if (rt_prio(oldprio))
3832 p->sched_class = &fair_sched_class;
3838 enqueue_task(rq, p, queue_flag);
3840 set_curr_task(rq, p);
3842 check_class_changed(rq, p, prev_class, oldprio);
3844 /* Avoid rq from going away on us: */
3846 __task_rq_unlock(rq, &rf);
3848 balance_callback(rq);
3852 static inline int rt_effective_prio(struct task_struct *p, int prio)
3858 void set_user_nice(struct task_struct *p, long nice)
3860 bool queued, running;
3861 int old_prio, delta;
3865 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3868 * We have to be careful, if called from sys_setpriority(),
3869 * the task might be in the middle of scheduling on another CPU.
3871 rq = task_rq_lock(p, &rf);
3872 update_rq_clock(rq);
3875 * The RT priorities are set via sched_setscheduler(), but we still
3876 * allow the 'normal' nice value to be set - but as expected
3877 * it wont have any effect on scheduling until the task is
3878 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3880 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3881 p->static_prio = NICE_TO_PRIO(nice);
3884 queued = task_on_rq_queued(p);
3885 running = task_current(rq, p);
3887 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
3889 put_prev_task(rq, p);
3891 p->static_prio = NICE_TO_PRIO(nice);
3892 set_load_weight(p, true);
3894 p->prio = effective_prio(p);
3895 delta = p->prio - old_prio;
3898 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
3900 * If the task increased its priority or is running and
3901 * lowered its priority, then reschedule its CPU:
3903 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3907 set_curr_task(rq, p);
3909 task_rq_unlock(rq, p, &rf);
3911 EXPORT_SYMBOL(set_user_nice);
3914 * can_nice - check if a task can reduce its nice value
3918 int can_nice(const struct task_struct *p, const int nice)
3920 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3921 int nice_rlim = nice_to_rlimit(nice);
3923 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3924 capable(CAP_SYS_NICE));
3927 #ifdef __ARCH_WANT_SYS_NICE
3930 * sys_nice - change the priority of the current process.
3931 * @increment: priority increment
3933 * sys_setpriority is a more generic, but much slower function that
3934 * does similar things.
3936 SYSCALL_DEFINE1(nice, int, increment)
3941 * Setpriority might change our priority at the same moment.
3942 * We don't have to worry. Conceptually one call occurs first
3943 * and we have a single winner.
3945 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3946 nice = task_nice(current) + increment;
3948 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3949 if (increment < 0 && !can_nice(current, nice))
3952 retval = security_task_setnice(current, nice);
3956 set_user_nice(current, nice);
3963 * task_prio - return the priority value of a given task.
3964 * @p: the task in question.
3966 * Return: The priority value as seen by users in /proc.
3967 * RT tasks are offset by -200. Normal tasks are centered
3968 * around 0, value goes from -16 to +15.
3970 int task_prio(const struct task_struct *p)
3972 return p->prio - MAX_RT_PRIO;
3976 * idle_cpu - is a given CPU idle currently?
3977 * @cpu: the processor in question.
3979 * Return: 1 if the CPU is currently idle. 0 otherwise.
3981 int idle_cpu(int cpu)
3983 struct rq *rq = cpu_rq(cpu);
3985 if (rq->curr != rq->idle)
3992 if (!llist_empty(&rq->wake_list))
4000 * available_idle_cpu - is a given CPU idle for enqueuing work.
4001 * @cpu: the CPU in question.
4003 * Return: 1 if the CPU is currently idle. 0 otherwise.
4005 int available_idle_cpu(int cpu)
4010 if (vcpu_is_preempted(cpu))
4017 * idle_task - return the idle task for a given CPU.
4018 * @cpu: the processor in question.
4020 * Return: The idle task for the CPU @cpu.
4022 struct task_struct *idle_task(int cpu)
4024 return cpu_rq(cpu)->idle;
4028 * find_process_by_pid - find a process with a matching PID value.
4029 * @pid: the pid in question.
4031 * The task of @pid, if found. %NULL otherwise.
4033 static struct task_struct *find_process_by_pid(pid_t pid)
4035 return pid ? find_task_by_vpid(pid) : current;
4039 * sched_setparam() passes in -1 for its policy, to let the functions
4040 * it calls know not to change it.
4042 #define SETPARAM_POLICY -1
4044 static void __setscheduler_params(struct task_struct *p,
4045 const struct sched_attr *attr)
4047 int policy = attr->sched_policy;
4049 if (policy == SETPARAM_POLICY)
4054 if (dl_policy(policy))
4055 __setparam_dl(p, attr);
4056 else if (fair_policy(policy))
4057 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4060 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4061 * !rt_policy. Always setting this ensures that things like
4062 * getparam()/getattr() don't report silly values for !rt tasks.
4064 p->rt_priority = attr->sched_priority;
4065 p->normal_prio = normal_prio(p);
4066 set_load_weight(p, true);
4069 /* Actually do priority change: must hold pi & rq lock. */
4070 static void __setscheduler(struct rq *rq, struct task_struct *p,
4071 const struct sched_attr *attr, bool keep_boost)
4073 __setscheduler_params(p, attr);
4076 * Keep a potential priority boosting if called from
4077 * sched_setscheduler().
4079 p->prio = normal_prio(p);
4081 p->prio = rt_effective_prio(p, p->prio);
4083 if (dl_prio(p->prio))
4084 p->sched_class = &dl_sched_class;
4085 else if (rt_prio(p->prio))
4086 p->sched_class = &rt_sched_class;
4088 p->sched_class = &fair_sched_class;
4092 * Check the target process has a UID that matches the current process's:
4094 static bool check_same_owner(struct task_struct *p)
4096 const struct cred *cred = current_cred(), *pcred;
4100 pcred = __task_cred(p);
4101 match = (uid_eq(cred->euid, pcred->euid) ||
4102 uid_eq(cred->euid, pcred->uid));
4107 static int __sched_setscheduler(struct task_struct *p,
4108 const struct sched_attr *attr,
4111 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4112 MAX_RT_PRIO - 1 - attr->sched_priority;
4113 int retval, oldprio, oldpolicy = -1, queued, running;
4114 int new_effective_prio, policy = attr->sched_policy;
4115 const struct sched_class *prev_class;
4118 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4121 /* The pi code expects interrupts enabled */
4122 BUG_ON(pi && in_interrupt());
4124 /* Double check policy once rq lock held: */
4126 reset_on_fork = p->sched_reset_on_fork;
4127 policy = oldpolicy = p->policy;
4129 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4131 if (!valid_policy(policy))
4135 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4139 * Valid priorities for SCHED_FIFO and SCHED_RR are
4140 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4141 * SCHED_BATCH and SCHED_IDLE is 0.
4143 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4144 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4146 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4147 (rt_policy(policy) != (attr->sched_priority != 0)))
4151 * Allow unprivileged RT tasks to decrease priority:
4153 if (user && !capable(CAP_SYS_NICE)) {
4154 if (fair_policy(policy)) {
4155 if (attr->sched_nice < task_nice(p) &&
4156 !can_nice(p, attr->sched_nice))
4160 if (rt_policy(policy)) {
4161 unsigned long rlim_rtprio =
4162 task_rlimit(p, RLIMIT_RTPRIO);
4164 /* Can't set/change the rt policy: */
4165 if (policy != p->policy && !rlim_rtprio)
4168 /* Can't increase priority: */
4169 if (attr->sched_priority > p->rt_priority &&
4170 attr->sched_priority > rlim_rtprio)
4175 * Can't set/change SCHED_DEADLINE policy at all for now
4176 * (safest behavior); in the future we would like to allow
4177 * unprivileged DL tasks to increase their relative deadline
4178 * or reduce their runtime (both ways reducing utilization)
4180 if (dl_policy(policy))
4184 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4185 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4187 if (task_has_idle_policy(p) && !idle_policy(policy)) {
4188 if (!can_nice(p, task_nice(p)))
4192 /* Can't change other user's priorities: */
4193 if (!check_same_owner(p))
4196 /* Normal users shall not reset the sched_reset_on_fork flag: */
4197 if (p->sched_reset_on_fork && !reset_on_fork)
4202 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4205 retval = security_task_setscheduler(p);
4211 * Make sure no PI-waiters arrive (or leave) while we are
4212 * changing the priority of the task:
4214 * To be able to change p->policy safely, the appropriate
4215 * runqueue lock must be held.
4217 rq = task_rq_lock(p, &rf);
4218 update_rq_clock(rq);
4221 * Changing the policy of the stop threads its a very bad idea:
4223 if (p == rq->stop) {
4224 task_rq_unlock(rq, p, &rf);
4229 * If not changing anything there's no need to proceed further,
4230 * but store a possible modification of reset_on_fork.
4232 if (unlikely(policy == p->policy)) {
4233 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4235 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4237 if (dl_policy(policy) && dl_param_changed(p, attr))
4240 p->sched_reset_on_fork = reset_on_fork;
4241 task_rq_unlock(rq, p, &rf);
4247 #ifdef CONFIG_RT_GROUP_SCHED
4249 * Do not allow realtime tasks into groups that have no runtime
4252 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4253 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4254 !task_group_is_autogroup(task_group(p))) {
4255 task_rq_unlock(rq, p, &rf);
4260 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4261 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4262 cpumask_t *span = rq->rd->span;
4265 * Don't allow tasks with an affinity mask smaller than
4266 * the entire root_domain to become SCHED_DEADLINE. We
4267 * will also fail if there's no bandwidth available.
4269 if (!cpumask_subset(span, &p->cpus_allowed) ||
4270 rq->rd->dl_bw.bw == 0) {
4271 task_rq_unlock(rq, p, &rf);
4278 /* Re-check policy now with rq lock held: */
4279 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4280 policy = oldpolicy = -1;
4281 task_rq_unlock(rq, p, &rf);
4286 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4287 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4290 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4291 task_rq_unlock(rq, p, &rf);
4295 p->sched_reset_on_fork = reset_on_fork;
4300 * Take priority boosted tasks into account. If the new
4301 * effective priority is unchanged, we just store the new
4302 * normal parameters and do not touch the scheduler class and
4303 * the runqueue. This will be done when the task deboost
4306 new_effective_prio = rt_effective_prio(p, newprio);
4307 if (new_effective_prio == oldprio)
4308 queue_flags &= ~DEQUEUE_MOVE;
4311 queued = task_on_rq_queued(p);
4312 running = task_current(rq, p);
4314 dequeue_task(rq, p, queue_flags);
4316 put_prev_task(rq, p);
4318 prev_class = p->sched_class;
4319 __setscheduler(rq, p, attr, pi);
4323 * We enqueue to tail when the priority of a task is
4324 * increased (user space view).
4326 if (oldprio < p->prio)
4327 queue_flags |= ENQUEUE_HEAD;
4329 enqueue_task(rq, p, queue_flags);
4332 set_curr_task(rq, p);
4334 check_class_changed(rq, p, prev_class, oldprio);
4336 /* Avoid rq from going away on us: */
4338 task_rq_unlock(rq, p, &rf);
4341 rt_mutex_adjust_pi(p);
4343 /* Run balance callbacks after we've adjusted the PI chain: */
4344 balance_callback(rq);
4350 static int _sched_setscheduler(struct task_struct *p, int policy,
4351 const struct sched_param *param, bool check)
4353 struct sched_attr attr = {
4354 .sched_policy = policy,
4355 .sched_priority = param->sched_priority,
4356 .sched_nice = PRIO_TO_NICE(p->static_prio),
4359 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4360 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4361 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4362 policy &= ~SCHED_RESET_ON_FORK;
4363 attr.sched_policy = policy;
4366 return __sched_setscheduler(p, &attr, check, true);
4369 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4370 * @p: the task in question.
4371 * @policy: new policy.
4372 * @param: structure containing the new RT priority.
4374 * Return: 0 on success. An error code otherwise.
4376 * NOTE that the task may be already dead.
4378 int sched_setscheduler(struct task_struct *p, int policy,
4379 const struct sched_param *param)
4381 return _sched_setscheduler(p, policy, param, true);
4383 EXPORT_SYMBOL_GPL(sched_setscheduler);
4385 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4387 return __sched_setscheduler(p, attr, true, true);
4389 EXPORT_SYMBOL_GPL(sched_setattr);
4391 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
4393 return __sched_setscheduler(p, attr, false, true);
4397 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4398 * @p: the task in question.
4399 * @policy: new policy.
4400 * @param: structure containing the new RT priority.
4402 * Just like sched_setscheduler, only don't bother checking if the
4403 * current context has permission. For example, this is needed in
4404 * stop_machine(): we create temporary high priority worker threads,
4405 * but our caller might not have that capability.
4407 * Return: 0 on success. An error code otherwise.
4409 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4410 const struct sched_param *param)
4412 return _sched_setscheduler(p, policy, param, false);
4414 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4417 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4419 struct sched_param lparam;
4420 struct task_struct *p;
4423 if (!param || pid < 0)
4425 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4430 p = find_process_by_pid(pid);
4432 retval = sched_setscheduler(p, policy, &lparam);
4439 * Mimics kernel/events/core.c perf_copy_attr().
4441 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4446 if (!access_ok(uattr, SCHED_ATTR_SIZE_VER0))
4449 /* Zero the full structure, so that a short copy will be nice: */
4450 memset(attr, 0, sizeof(*attr));
4452 ret = get_user(size, &uattr->size);
4456 /* Bail out on silly large: */
4457 if (size > PAGE_SIZE)
4460 /* ABI compatibility quirk: */
4462 size = SCHED_ATTR_SIZE_VER0;
4464 if (size < SCHED_ATTR_SIZE_VER0)
4468 * If we're handed a bigger struct than we know of,
4469 * ensure all the unknown bits are 0 - i.e. new
4470 * user-space does not rely on any kernel feature
4471 * extensions we dont know about yet.
4473 if (size > sizeof(*attr)) {
4474 unsigned char __user *addr;
4475 unsigned char __user *end;
4478 addr = (void __user *)uattr + sizeof(*attr);
4479 end = (void __user *)uattr + size;
4481 for (; addr < end; addr++) {
4482 ret = get_user(val, addr);
4488 size = sizeof(*attr);
4491 ret = copy_from_user(attr, uattr, size);
4496 * XXX: Do we want to be lenient like existing syscalls; or do we want
4497 * to be strict and return an error on out-of-bounds values?
4499 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4504 put_user(sizeof(*attr), &uattr->size);
4509 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4510 * @pid: the pid in question.
4511 * @policy: new policy.
4512 * @param: structure containing the new RT priority.
4514 * Return: 0 on success. An error code otherwise.
4516 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4521 return do_sched_setscheduler(pid, policy, param);
4525 * sys_sched_setparam - set/change the RT priority of a thread
4526 * @pid: the pid in question.
4527 * @param: structure containing the new RT priority.
4529 * Return: 0 on success. An error code otherwise.
4531 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4533 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4537 * sys_sched_setattr - same as above, but with extended sched_attr
4538 * @pid: the pid in question.
4539 * @uattr: structure containing the extended parameters.
4540 * @flags: for future extension.
4542 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4543 unsigned int, flags)
4545 struct sched_attr attr;
4546 struct task_struct *p;
4549 if (!uattr || pid < 0 || flags)
4552 retval = sched_copy_attr(uattr, &attr);
4556 if ((int)attr.sched_policy < 0)
4561 p = find_process_by_pid(pid);
4563 retval = sched_setattr(p, &attr);
4570 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4571 * @pid: the pid in question.
4573 * Return: On success, the policy of the thread. Otherwise, a negative error
4576 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4578 struct task_struct *p;
4586 p = find_process_by_pid(pid);
4588 retval = security_task_getscheduler(p);
4591 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4598 * sys_sched_getparam - get the RT priority of a thread
4599 * @pid: the pid in question.
4600 * @param: structure containing the RT priority.
4602 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4605 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4607 struct sched_param lp = { .sched_priority = 0 };
4608 struct task_struct *p;
4611 if (!param || pid < 0)
4615 p = find_process_by_pid(pid);
4620 retval = security_task_getscheduler(p);
4624 if (task_has_rt_policy(p))
4625 lp.sched_priority = p->rt_priority;
4629 * This one might sleep, we cannot do it with a spinlock held ...
4631 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4640 static int sched_read_attr(struct sched_attr __user *uattr,
4641 struct sched_attr *attr,
4646 if (!access_ok(uattr, usize))
4650 * If we're handed a smaller struct than we know of,
4651 * ensure all the unknown bits are 0 - i.e. old
4652 * user-space does not get uncomplete information.
4654 if (usize < sizeof(*attr)) {
4655 unsigned char *addr;
4658 addr = (void *)attr + usize;
4659 end = (void *)attr + sizeof(*attr);
4661 for (; addr < end; addr++) {
4669 ret = copy_to_user(uattr, attr, attr->size);
4677 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4678 * @pid: the pid in question.
4679 * @uattr: structure containing the extended parameters.
4680 * @size: sizeof(attr) for fwd/bwd comp.
4681 * @flags: for future extension.
4683 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4684 unsigned int, size, unsigned int, flags)
4686 struct sched_attr attr = {
4687 .size = sizeof(struct sched_attr),
4689 struct task_struct *p;
4692 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4693 size < SCHED_ATTR_SIZE_VER0 || flags)
4697 p = find_process_by_pid(pid);
4702 retval = security_task_getscheduler(p);
4706 attr.sched_policy = p->policy;
4707 if (p->sched_reset_on_fork)
4708 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4709 if (task_has_dl_policy(p))
4710 __getparam_dl(p, &attr);
4711 else if (task_has_rt_policy(p))
4712 attr.sched_priority = p->rt_priority;
4714 attr.sched_nice = task_nice(p);
4718 retval = sched_read_attr(uattr, &attr, size);
4726 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4728 cpumask_var_t cpus_allowed, new_mask;
4729 struct task_struct *p;
4734 p = find_process_by_pid(pid);
4740 /* Prevent p going away */
4744 if (p->flags & PF_NO_SETAFFINITY) {
4748 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4752 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4754 goto out_free_cpus_allowed;
4757 if (!check_same_owner(p)) {
4759 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4761 goto out_free_new_mask;
4766 retval = security_task_setscheduler(p);
4768 goto out_free_new_mask;
4771 cpuset_cpus_allowed(p, cpus_allowed);
4772 cpumask_and(new_mask, in_mask, cpus_allowed);
4775 * Since bandwidth control happens on root_domain basis,
4776 * if admission test is enabled, we only admit -deadline
4777 * tasks allowed to run on all the CPUs in the task's
4781 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4783 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4786 goto out_free_new_mask;
4792 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4795 cpuset_cpus_allowed(p, cpus_allowed);
4796 if (!cpumask_subset(new_mask, cpus_allowed)) {
4798 * We must have raced with a concurrent cpuset
4799 * update. Just reset the cpus_allowed to the
4800 * cpuset's cpus_allowed
4802 cpumask_copy(new_mask, cpus_allowed);
4807 free_cpumask_var(new_mask);
4808 out_free_cpus_allowed:
4809 free_cpumask_var(cpus_allowed);
4815 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4816 struct cpumask *new_mask)
4818 if (len < cpumask_size())
4819 cpumask_clear(new_mask);
4820 else if (len > cpumask_size())
4821 len = cpumask_size();
4823 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4827 * sys_sched_setaffinity - set the CPU affinity of a process
4828 * @pid: pid of the process
4829 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4830 * @user_mask_ptr: user-space pointer to the new CPU mask
4832 * Return: 0 on success. An error code otherwise.
4834 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4835 unsigned long __user *, user_mask_ptr)
4837 cpumask_var_t new_mask;
4840 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4843 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4845 retval = sched_setaffinity(pid, new_mask);
4846 free_cpumask_var(new_mask);
4850 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4852 struct task_struct *p;
4853 unsigned long flags;
4859 p = find_process_by_pid(pid);
4863 retval = security_task_getscheduler(p);
4867 raw_spin_lock_irqsave(&p->pi_lock, flags);
4868 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4869 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4878 * sys_sched_getaffinity - get the CPU affinity of a process
4879 * @pid: pid of the process
4880 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4881 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4883 * Return: size of CPU mask copied to user_mask_ptr on success. An
4884 * error code otherwise.
4886 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4887 unsigned long __user *, user_mask_ptr)
4892 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4894 if (len & (sizeof(unsigned long)-1))
4897 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4900 ret = sched_getaffinity(pid, mask);
4902 unsigned int retlen = min(len, cpumask_size());
4904 if (copy_to_user(user_mask_ptr, mask, retlen))
4909 free_cpumask_var(mask);
4915 * sys_sched_yield - yield the current processor to other threads.
4917 * This function yields the current CPU to other tasks. If there are no
4918 * other threads running on this CPU then this function will return.
4922 static void do_sched_yield(void)
4927 rq = this_rq_lock_irq(&rf);
4929 schedstat_inc(rq->yld_count);
4930 current->sched_class->yield_task(rq);
4933 * Since we are going to call schedule() anyway, there's
4934 * no need to preempt or enable interrupts:
4938 sched_preempt_enable_no_resched();
4943 SYSCALL_DEFINE0(sched_yield)
4949 #ifndef CONFIG_PREEMPT
4950 int __sched _cond_resched(void)
4952 if (should_resched(0)) {
4953 preempt_schedule_common();
4959 EXPORT_SYMBOL(_cond_resched);
4963 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4964 * call schedule, and on return reacquire the lock.
4966 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4967 * operations here to prevent schedule() from being called twice (once via
4968 * spin_unlock(), once by hand).
4970 int __cond_resched_lock(spinlock_t *lock)
4972 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4975 lockdep_assert_held(lock);
4977 if (spin_needbreak(lock) || resched) {
4980 preempt_schedule_common();
4988 EXPORT_SYMBOL(__cond_resched_lock);
4991 * yield - yield the current processor to other threads.
4993 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4995 * The scheduler is at all times free to pick the calling task as the most
4996 * eligible task to run, if removing the yield() call from your code breaks
4997 * it, its already broken.
4999 * Typical broken usage is:
5004 * where one assumes that yield() will let 'the other' process run that will
5005 * make event true. If the current task is a SCHED_FIFO task that will never
5006 * happen. Never use yield() as a progress guarantee!!
5008 * If you want to use yield() to wait for something, use wait_event().
5009 * If you want to use yield() to be 'nice' for others, use cond_resched().
5010 * If you still want to use yield(), do not!
5012 void __sched yield(void)
5014 set_current_state(TASK_RUNNING);
5017 EXPORT_SYMBOL(yield);
5020 * yield_to - yield the current processor to another thread in
5021 * your thread group, or accelerate that thread toward the
5022 * processor it's on.
5024 * @preempt: whether task preemption is allowed or not
5026 * It's the caller's job to ensure that the target task struct
5027 * can't go away on us before we can do any checks.
5030 * true (>0) if we indeed boosted the target task.
5031 * false (0) if we failed to boost the target.
5032 * -ESRCH if there's no task to yield to.
5034 int __sched yield_to(struct task_struct *p, bool preempt)
5036 struct task_struct *curr = current;
5037 struct rq *rq, *p_rq;
5038 unsigned long flags;
5041 local_irq_save(flags);
5047 * If we're the only runnable task on the rq and target rq also
5048 * has only one task, there's absolutely no point in yielding.
5050 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5055 double_rq_lock(rq, p_rq);
5056 if (task_rq(p) != p_rq) {
5057 double_rq_unlock(rq, p_rq);
5061 if (!curr->sched_class->yield_to_task)
5064 if (curr->sched_class != p->sched_class)
5067 if (task_running(p_rq, p) || p->state)
5070 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5072 schedstat_inc(rq->yld_count);
5074 * Make p's CPU reschedule; pick_next_entity takes care of
5077 if (preempt && rq != p_rq)
5082 double_rq_unlock(rq, p_rq);
5084 local_irq_restore(flags);
5091 EXPORT_SYMBOL_GPL(yield_to);
5093 int io_schedule_prepare(void)
5095 int old_iowait = current->in_iowait;
5097 current->in_iowait = 1;
5098 blk_schedule_flush_plug(current);
5103 void io_schedule_finish(int token)
5105 current->in_iowait = token;
5109 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5110 * that process accounting knows that this is a task in IO wait state.
5112 long __sched io_schedule_timeout(long timeout)
5117 token = io_schedule_prepare();
5118 ret = schedule_timeout(timeout);
5119 io_schedule_finish(token);
5123 EXPORT_SYMBOL(io_schedule_timeout);
5125 void io_schedule(void)
5129 token = io_schedule_prepare();
5131 io_schedule_finish(token);
5133 EXPORT_SYMBOL(io_schedule);
5136 * sys_sched_get_priority_max - return maximum RT priority.
5137 * @policy: scheduling class.
5139 * Return: On success, this syscall returns the maximum
5140 * rt_priority that can be used by a given scheduling class.
5141 * On failure, a negative error code is returned.
5143 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5150 ret = MAX_USER_RT_PRIO-1;
5152 case SCHED_DEADLINE:
5163 * sys_sched_get_priority_min - return minimum RT priority.
5164 * @policy: scheduling class.
5166 * Return: On success, this syscall returns the minimum
5167 * rt_priority that can be used by a given scheduling class.
5168 * On failure, a negative error code is returned.
5170 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5179 case SCHED_DEADLINE:
5188 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5190 struct task_struct *p;
5191 unsigned int time_slice;
5201 p = find_process_by_pid(pid);
5205 retval = security_task_getscheduler(p);
5209 rq = task_rq_lock(p, &rf);
5211 if (p->sched_class->get_rr_interval)
5212 time_slice = p->sched_class->get_rr_interval(rq, p);
5213 task_rq_unlock(rq, p, &rf);
5216 jiffies_to_timespec64(time_slice, t);
5225 * sys_sched_rr_get_interval - return the default timeslice of a process.
5226 * @pid: pid of the process.
5227 * @interval: userspace pointer to the timeslice value.
5229 * this syscall writes the default timeslice value of a given process
5230 * into the user-space timespec buffer. A value of '0' means infinity.
5232 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5235 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5236 struct __kernel_timespec __user *, interval)
5238 struct timespec64 t;
5239 int retval = sched_rr_get_interval(pid, &t);
5242 retval = put_timespec64(&t, interval);
5247 #ifdef CONFIG_COMPAT_32BIT_TIME
5248 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
5249 struct old_timespec32 __user *, interval)
5251 struct timespec64 t;
5252 int retval = sched_rr_get_interval(pid, &t);
5255 retval = put_old_timespec32(&t, interval);
5260 void sched_show_task(struct task_struct *p)
5262 unsigned long free = 0;
5265 if (!try_get_task_stack(p))
5268 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5270 if (p->state == TASK_RUNNING)
5271 printk(KERN_CONT " running task ");
5272 #ifdef CONFIG_DEBUG_STACK_USAGE
5273 free = stack_not_used(p);
5278 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5280 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5281 task_pid_nr(p), ppid,
5282 (unsigned long)task_thread_info(p)->flags);
5284 print_worker_info(KERN_INFO, p);
5285 show_stack(p, NULL);
5288 EXPORT_SYMBOL_GPL(sched_show_task);
5291 state_filter_match(unsigned long state_filter, struct task_struct *p)
5293 /* no filter, everything matches */
5297 /* filter, but doesn't match */
5298 if (!(p->state & state_filter))
5302 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5305 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5312 void show_state_filter(unsigned long state_filter)
5314 struct task_struct *g, *p;
5316 #if BITS_PER_LONG == 32
5318 " task PC stack pid father\n");
5321 " task PC stack pid father\n");
5324 for_each_process_thread(g, p) {
5326 * reset the NMI-timeout, listing all files on a slow
5327 * console might take a lot of time:
5328 * Also, reset softlockup watchdogs on all CPUs, because
5329 * another CPU might be blocked waiting for us to process
5332 touch_nmi_watchdog();
5333 touch_all_softlockup_watchdogs();
5334 if (state_filter_match(state_filter, p))
5338 #ifdef CONFIG_SCHED_DEBUG
5340 sysrq_sched_debug_show();
5344 * Only show locks if all tasks are dumped:
5347 debug_show_all_locks();
5351 * init_idle - set up an idle thread for a given CPU
5352 * @idle: task in question
5353 * @cpu: CPU the idle task belongs to
5355 * NOTE: this function does not set the idle thread's NEED_RESCHED
5356 * flag, to make booting more robust.
5358 void init_idle(struct task_struct *idle, int cpu)
5360 struct rq *rq = cpu_rq(cpu);
5361 unsigned long flags;
5363 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5364 raw_spin_lock(&rq->lock);
5366 __sched_fork(0, idle);
5367 idle->state = TASK_RUNNING;
5368 idle->se.exec_start = sched_clock();
5369 idle->flags |= PF_IDLE;
5371 kasan_unpoison_task_stack(idle);
5375 * Its possible that init_idle() gets called multiple times on a task,
5376 * in that case do_set_cpus_allowed() will not do the right thing.
5378 * And since this is boot we can forgo the serialization.
5380 set_cpus_allowed_common(idle, cpumask_of(cpu));
5383 * We're having a chicken and egg problem, even though we are
5384 * holding rq->lock, the CPU isn't yet set to this CPU so the
5385 * lockdep check in task_group() will fail.
5387 * Similar case to sched_fork(). / Alternatively we could
5388 * use task_rq_lock() here and obtain the other rq->lock.
5393 __set_task_cpu(idle, cpu);
5396 rq->curr = rq->idle = idle;
5397 idle->on_rq = TASK_ON_RQ_QUEUED;
5401 raw_spin_unlock(&rq->lock);
5402 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5404 /* Set the preempt count _outside_ the spinlocks! */
5405 init_idle_preempt_count(idle, cpu);
5408 * The idle tasks have their own, simple scheduling class:
5410 idle->sched_class = &idle_sched_class;
5411 ftrace_graph_init_idle_task(idle, cpu);
5412 vtime_init_idle(idle, cpu);
5414 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5420 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5421 const struct cpumask *trial)
5425 if (!cpumask_weight(cur))
5428 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5433 int task_can_attach(struct task_struct *p,
5434 const struct cpumask *cs_cpus_allowed)
5439 * Kthreads which disallow setaffinity shouldn't be moved
5440 * to a new cpuset; we don't want to change their CPU
5441 * affinity and isolating such threads by their set of
5442 * allowed nodes is unnecessary. Thus, cpusets are not
5443 * applicable for such threads. This prevents checking for
5444 * success of set_cpus_allowed_ptr() on all attached tasks
5445 * before cpus_allowed may be changed.
5447 if (p->flags & PF_NO_SETAFFINITY) {
5452 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5454 ret = dl_task_can_attach(p, cs_cpus_allowed);
5460 bool sched_smp_initialized __read_mostly;
5462 #ifdef CONFIG_NUMA_BALANCING
5463 /* Migrate current task p to target_cpu */
5464 int migrate_task_to(struct task_struct *p, int target_cpu)
5466 struct migration_arg arg = { p, target_cpu };
5467 int curr_cpu = task_cpu(p);
5469 if (curr_cpu == target_cpu)
5472 if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5475 /* TODO: This is not properly updating schedstats */
5477 trace_sched_move_numa(p, curr_cpu, target_cpu);
5478 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5482 * Requeue a task on a given node and accurately track the number of NUMA
5483 * tasks on the runqueues
5485 void sched_setnuma(struct task_struct *p, int nid)
5487 bool queued, running;
5491 rq = task_rq_lock(p, &rf);
5492 queued = task_on_rq_queued(p);
5493 running = task_current(rq, p);
5496 dequeue_task(rq, p, DEQUEUE_SAVE);
5498 put_prev_task(rq, p);
5500 p->numa_preferred_nid = nid;
5503 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5505 set_curr_task(rq, p);
5506 task_rq_unlock(rq, p, &rf);
5508 #endif /* CONFIG_NUMA_BALANCING */
5510 #ifdef CONFIG_HOTPLUG_CPU
5512 * Ensure that the idle task is using init_mm right before its CPU goes
5515 void idle_task_exit(void)
5517 struct mm_struct *mm = current->active_mm;
5519 BUG_ON(cpu_online(smp_processor_id()));
5521 if (mm != &init_mm) {
5522 switch_mm(mm, &init_mm, current);
5523 current->active_mm = &init_mm;
5524 finish_arch_post_lock_switch();
5530 * Since this CPU is going 'away' for a while, fold any nr_active delta
5531 * we might have. Assumes we're called after migrate_tasks() so that the
5532 * nr_active count is stable. We need to take the teardown thread which
5533 * is calling this into account, so we hand in adjust = 1 to the load
5536 * Also see the comment "Global load-average calculations".
5538 static void calc_load_migrate(struct rq *rq)
5540 long delta = calc_load_fold_active(rq, 1);
5542 atomic_long_add(delta, &calc_load_tasks);
5545 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5549 static const struct sched_class fake_sched_class = {
5550 .put_prev_task = put_prev_task_fake,
5553 static struct task_struct fake_task = {
5555 * Avoid pull_{rt,dl}_task()
5557 .prio = MAX_PRIO + 1,
5558 .sched_class = &fake_sched_class,
5562 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5563 * try_to_wake_up()->select_task_rq().
5565 * Called with rq->lock held even though we'er in stop_machine() and
5566 * there's no concurrency possible, we hold the required locks anyway
5567 * because of lock validation efforts.
5569 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5571 struct rq *rq = dead_rq;
5572 struct task_struct *next, *stop = rq->stop;
5573 struct rq_flags orf = *rf;
5577 * Fudge the rq selection such that the below task selection loop
5578 * doesn't get stuck on the currently eligible stop task.
5580 * We're currently inside stop_machine() and the rq is either stuck
5581 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5582 * either way we should never end up calling schedule() until we're
5588 * put_prev_task() and pick_next_task() sched
5589 * class method both need to have an up-to-date
5590 * value of rq->clock[_task]
5592 update_rq_clock(rq);
5596 * There's this thread running, bail when that's the only
5599 if (rq->nr_running == 1)
5603 * pick_next_task() assumes pinned rq->lock:
5605 next = pick_next_task(rq, &fake_task, rf);
5607 put_prev_task(rq, next);
5610 * Rules for changing task_struct::cpus_allowed are holding
5611 * both pi_lock and rq->lock, such that holding either
5612 * stabilizes the mask.
5614 * Drop rq->lock is not quite as disastrous as it usually is
5615 * because !cpu_active at this point, which means load-balance
5616 * will not interfere. Also, stop-machine.
5619 raw_spin_lock(&next->pi_lock);
5623 * Since we're inside stop-machine, _nothing_ should have
5624 * changed the task, WARN if weird stuff happened, because in
5625 * that case the above rq->lock drop is a fail too.
5627 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5628 raw_spin_unlock(&next->pi_lock);
5632 /* Find suitable destination for @next, with force if needed. */
5633 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5634 rq = __migrate_task(rq, rf, next, dest_cpu);
5635 if (rq != dead_rq) {
5641 raw_spin_unlock(&next->pi_lock);
5646 #endif /* CONFIG_HOTPLUG_CPU */
5648 void set_rq_online(struct rq *rq)
5651 const struct sched_class *class;
5653 cpumask_set_cpu(rq->cpu, rq->rd->online);
5656 for_each_class(class) {
5657 if (class->rq_online)
5658 class->rq_online(rq);
5663 void set_rq_offline(struct rq *rq)
5666 const struct sched_class *class;
5668 for_each_class(class) {
5669 if (class->rq_offline)
5670 class->rq_offline(rq);
5673 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5679 * used to mark begin/end of suspend/resume:
5681 static int num_cpus_frozen;
5684 * Update cpusets according to cpu_active mask. If cpusets are
5685 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5686 * around partition_sched_domains().
5688 * If we come here as part of a suspend/resume, don't touch cpusets because we
5689 * want to restore it back to its original state upon resume anyway.
5691 static void cpuset_cpu_active(void)
5693 if (cpuhp_tasks_frozen) {
5695 * num_cpus_frozen tracks how many CPUs are involved in suspend
5696 * resume sequence. As long as this is not the last online
5697 * operation in the resume sequence, just build a single sched
5698 * domain, ignoring cpusets.
5700 partition_sched_domains(1, NULL, NULL);
5701 if (--num_cpus_frozen)
5704 * This is the last CPU online operation. So fall through and
5705 * restore the original sched domains by considering the
5706 * cpuset configurations.
5708 cpuset_force_rebuild();
5710 cpuset_update_active_cpus();
5713 static int cpuset_cpu_inactive(unsigned int cpu)
5715 if (!cpuhp_tasks_frozen) {
5716 if (dl_cpu_busy(cpu))
5718 cpuset_update_active_cpus();
5721 partition_sched_domains(1, NULL, NULL);
5726 int sched_cpu_activate(unsigned int cpu)
5728 struct rq *rq = cpu_rq(cpu);
5731 #ifdef CONFIG_SCHED_SMT
5733 * When going up, increment the number of cores with SMT present.
5735 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
5736 static_branch_inc_cpuslocked(&sched_smt_present);
5738 set_cpu_active(cpu, true);
5740 if (sched_smp_initialized) {
5741 sched_domains_numa_masks_set(cpu);
5742 cpuset_cpu_active();
5746 * Put the rq online, if not already. This happens:
5748 * 1) In the early boot process, because we build the real domains
5749 * after all CPUs have been brought up.
5751 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5754 rq_lock_irqsave(rq, &rf);
5756 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5759 rq_unlock_irqrestore(rq, &rf);
5761 update_max_interval();
5766 int sched_cpu_deactivate(unsigned int cpu)
5770 set_cpu_active(cpu, false);
5772 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5773 * users of this state to go away such that all new such users will
5776 * Do sync before park smpboot threads to take care the rcu boost case.
5780 #ifdef CONFIG_SCHED_SMT
5782 * When going down, decrement the number of cores with SMT present.
5784 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
5785 static_branch_dec_cpuslocked(&sched_smt_present);
5788 if (!sched_smp_initialized)
5791 ret = cpuset_cpu_inactive(cpu);
5793 set_cpu_active(cpu, true);
5796 sched_domains_numa_masks_clear(cpu);
5800 static void sched_rq_cpu_starting(unsigned int cpu)
5802 struct rq *rq = cpu_rq(cpu);
5804 rq->calc_load_update = calc_load_update;
5805 update_max_interval();
5808 int sched_cpu_starting(unsigned int cpu)
5810 sched_rq_cpu_starting(cpu);
5811 sched_tick_start(cpu);
5815 #ifdef CONFIG_HOTPLUG_CPU
5816 int sched_cpu_dying(unsigned int cpu)
5818 struct rq *rq = cpu_rq(cpu);
5821 /* Handle pending wakeups and then migrate everything off */
5822 sched_ttwu_pending();
5823 sched_tick_stop(cpu);
5825 rq_lock_irqsave(rq, &rf);
5827 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5830 migrate_tasks(rq, &rf);
5831 BUG_ON(rq->nr_running != 1);
5832 rq_unlock_irqrestore(rq, &rf);
5834 calc_load_migrate(rq);
5835 update_max_interval();
5836 nohz_balance_exit_idle(rq);
5842 void __init sched_init_smp(void)
5847 * There's no userspace yet to cause hotplug operations; hence all the
5848 * CPU masks are stable and all blatant races in the below code cannot
5851 mutex_lock(&sched_domains_mutex);
5852 sched_init_domains(cpu_active_mask);
5853 mutex_unlock(&sched_domains_mutex);
5855 /* Move init over to a non-isolated CPU */
5856 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
5858 sched_init_granularity();
5860 init_sched_rt_class();
5861 init_sched_dl_class();
5863 sched_smp_initialized = true;
5866 static int __init migration_init(void)
5868 sched_cpu_starting(smp_processor_id());
5871 early_initcall(migration_init);
5874 void __init sched_init_smp(void)
5876 sched_init_granularity();
5878 #endif /* CONFIG_SMP */
5880 int in_sched_functions(unsigned long addr)
5882 return in_lock_functions(addr) ||
5883 (addr >= (unsigned long)__sched_text_start
5884 && addr < (unsigned long)__sched_text_end);
5887 #ifdef CONFIG_CGROUP_SCHED
5889 * Default task group.
5890 * Every task in system belongs to this group at bootup.
5892 struct task_group root_task_group;
5893 LIST_HEAD(task_groups);
5895 /* Cacheline aligned slab cache for task_group */
5896 static struct kmem_cache *task_group_cache __read_mostly;
5899 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5900 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5902 void __init sched_init(void)
5905 unsigned long alloc_size = 0, ptr;
5909 #ifdef CONFIG_FAIR_GROUP_SCHED
5910 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5912 #ifdef CONFIG_RT_GROUP_SCHED
5913 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5916 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5918 #ifdef CONFIG_FAIR_GROUP_SCHED
5919 root_task_group.se = (struct sched_entity **)ptr;
5920 ptr += nr_cpu_ids * sizeof(void **);
5922 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5923 ptr += nr_cpu_ids * sizeof(void **);
5925 #endif /* CONFIG_FAIR_GROUP_SCHED */
5926 #ifdef CONFIG_RT_GROUP_SCHED
5927 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5928 ptr += nr_cpu_ids * sizeof(void **);
5930 root_task_group.rt_rq = (struct rt_rq **)ptr;
5931 ptr += nr_cpu_ids * sizeof(void **);
5933 #endif /* CONFIG_RT_GROUP_SCHED */
5935 #ifdef CONFIG_CPUMASK_OFFSTACK
5936 for_each_possible_cpu(i) {
5937 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5938 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5939 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
5940 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5942 #endif /* CONFIG_CPUMASK_OFFSTACK */
5944 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
5945 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
5948 init_defrootdomain();
5951 #ifdef CONFIG_RT_GROUP_SCHED
5952 init_rt_bandwidth(&root_task_group.rt_bandwidth,
5953 global_rt_period(), global_rt_runtime());
5954 #endif /* CONFIG_RT_GROUP_SCHED */
5956 #ifdef CONFIG_CGROUP_SCHED
5957 task_group_cache = KMEM_CACHE(task_group, 0);
5959 list_add(&root_task_group.list, &task_groups);
5960 INIT_LIST_HEAD(&root_task_group.children);
5961 INIT_LIST_HEAD(&root_task_group.siblings);
5962 autogroup_init(&init_task);
5963 #endif /* CONFIG_CGROUP_SCHED */
5965 for_each_possible_cpu(i) {
5969 raw_spin_lock_init(&rq->lock);
5971 rq->calc_load_active = 0;
5972 rq->calc_load_update = jiffies + LOAD_FREQ;
5973 init_cfs_rq(&rq->cfs);
5974 init_rt_rq(&rq->rt);
5975 init_dl_rq(&rq->dl);
5976 #ifdef CONFIG_FAIR_GROUP_SCHED
5977 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
5978 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
5979 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
5981 * How much CPU bandwidth does root_task_group get?
5983 * In case of task-groups formed thr' the cgroup filesystem, it
5984 * gets 100% of the CPU resources in the system. This overall
5985 * system CPU resource is divided among the tasks of
5986 * root_task_group and its child task-groups in a fair manner,
5987 * based on each entity's (task or task-group's) weight
5988 * (se->load.weight).
5990 * In other words, if root_task_group has 10 tasks of weight
5991 * 1024) and two child groups A0 and A1 (of weight 1024 each),
5992 * then A0's share of the CPU resource is:
5994 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
5996 * We achieve this by letting root_task_group's tasks sit
5997 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
5999 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6000 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6001 #endif /* CONFIG_FAIR_GROUP_SCHED */
6003 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6004 #ifdef CONFIG_RT_GROUP_SCHED
6005 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6008 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6009 rq->cpu_load[j] = 0;
6014 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6015 rq->balance_callback = NULL;
6016 rq->active_balance = 0;
6017 rq->next_balance = jiffies;
6022 rq->avg_idle = 2*sysctl_sched_migration_cost;
6023 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6025 INIT_LIST_HEAD(&rq->cfs_tasks);
6027 rq_attach_root(rq, &def_root_domain);
6028 #ifdef CONFIG_NO_HZ_COMMON
6029 rq->last_load_update_tick = jiffies;
6030 rq->last_blocked_load_update_tick = jiffies;
6031 atomic_set(&rq->nohz_flags, 0);
6033 #endif /* CONFIG_SMP */
6035 atomic_set(&rq->nr_iowait, 0);
6038 set_load_weight(&init_task, false);
6041 * The boot idle thread does lazy MMU switching as well:
6044 enter_lazy_tlb(&init_mm, current);
6047 * Make us the idle thread. Technically, schedule() should not be
6048 * called from this thread, however somewhere below it might be,
6049 * but because we are the idle thread, we just pick up running again
6050 * when this runqueue becomes "idle".
6052 init_idle(current, smp_processor_id());
6054 calc_load_update = jiffies + LOAD_FREQ;
6057 idle_thread_set_boot_cpu();
6059 init_sched_fair_class();
6065 scheduler_running = 1;
6068 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6069 static inline int preempt_count_equals(int preempt_offset)
6071 int nested = preempt_count() + rcu_preempt_depth();
6073 return (nested == preempt_offset);
6076 void __might_sleep(const char *file, int line, int preempt_offset)
6079 * Blocking primitives will set (and therefore destroy) current->state,
6080 * since we will exit with TASK_RUNNING make sure we enter with it,
6081 * otherwise we will destroy state.
6083 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6084 "do not call blocking ops when !TASK_RUNNING; "
6085 "state=%lx set at [<%p>] %pS\n",
6087 (void *)current->task_state_change,
6088 (void *)current->task_state_change);
6090 ___might_sleep(file, line, preempt_offset);
6092 EXPORT_SYMBOL(__might_sleep);
6094 void ___might_sleep(const char *file, int line, int preempt_offset)
6096 /* Ratelimiting timestamp: */
6097 static unsigned long prev_jiffy;
6099 unsigned long preempt_disable_ip;
6101 /* WARN_ON_ONCE() by default, no rate limit required: */
6104 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6105 !is_idle_task(current)) ||
6106 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6110 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6112 prev_jiffy = jiffies;
6114 /* Save this before calling printk(), since that will clobber it: */
6115 preempt_disable_ip = get_preempt_disable_ip(current);
6118 "BUG: sleeping function called from invalid context at %s:%d\n",
6121 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6122 in_atomic(), irqs_disabled(),
6123 current->pid, current->comm);
6125 if (task_stack_end_corrupted(current))
6126 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6128 debug_show_held_locks(current);
6129 if (irqs_disabled())
6130 print_irqtrace_events(current);
6131 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6132 && !preempt_count_equals(preempt_offset)) {
6133 pr_err("Preemption disabled at:");
6134 print_ip_sym(preempt_disable_ip);
6138 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6140 EXPORT_SYMBOL(___might_sleep);
6142 void __cant_sleep(const char *file, int line, int preempt_offset)
6144 static unsigned long prev_jiffy;
6146 if (irqs_disabled())
6149 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
6152 if (preempt_count() > preempt_offset)
6155 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6157 prev_jiffy = jiffies;
6159 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
6160 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6161 in_atomic(), irqs_disabled(),
6162 current->pid, current->comm);
6164 debug_show_held_locks(current);
6166 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6168 EXPORT_SYMBOL_GPL(__cant_sleep);
6171 #ifdef CONFIG_MAGIC_SYSRQ
6172 void normalize_rt_tasks(void)
6174 struct task_struct *g, *p;
6175 struct sched_attr attr = {
6176 .sched_policy = SCHED_NORMAL,
6179 read_lock(&tasklist_lock);
6180 for_each_process_thread(g, p) {
6182 * Only normalize user tasks:
6184 if (p->flags & PF_KTHREAD)
6187 p->se.exec_start = 0;
6188 schedstat_set(p->se.statistics.wait_start, 0);
6189 schedstat_set(p->se.statistics.sleep_start, 0);
6190 schedstat_set(p->se.statistics.block_start, 0);
6192 if (!dl_task(p) && !rt_task(p)) {
6194 * Renice negative nice level userspace
6197 if (task_nice(p) < 0)
6198 set_user_nice(p, 0);
6202 __sched_setscheduler(p, &attr, false, false);
6204 read_unlock(&tasklist_lock);
6207 #endif /* CONFIG_MAGIC_SYSRQ */
6209 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6211 * These functions are only useful for the IA64 MCA handling, or kdb.
6213 * They can only be called when the whole system has been
6214 * stopped - every CPU needs to be quiescent, and no scheduling
6215 * activity can take place. Using them for anything else would
6216 * be a serious bug, and as a result, they aren't even visible
6217 * under any other configuration.
6221 * curr_task - return the current task for a given CPU.
6222 * @cpu: the processor in question.
6224 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6226 * Return: The current task for @cpu.
6228 struct task_struct *curr_task(int cpu)
6230 return cpu_curr(cpu);
6233 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6237 * set_curr_task - set the current task for a given CPU.
6238 * @cpu: the processor in question.
6239 * @p: the task pointer to set.
6241 * Description: This function must only be used when non-maskable interrupts
6242 * are serviced on a separate stack. It allows the architecture to switch the
6243 * notion of the current task on a CPU in a non-blocking manner. This function
6244 * must be called with all CPU's synchronized, and interrupts disabled, the
6245 * and caller must save the original value of the current task (see
6246 * curr_task() above) and restore that value before reenabling interrupts and
6247 * re-starting the system.
6249 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6251 void ia64_set_curr_task(int cpu, struct task_struct *p)
6258 #ifdef CONFIG_CGROUP_SCHED
6259 /* task_group_lock serializes the addition/removal of task groups */
6260 static DEFINE_SPINLOCK(task_group_lock);
6262 static void sched_free_group(struct task_group *tg)
6264 free_fair_sched_group(tg);
6265 free_rt_sched_group(tg);
6267 kmem_cache_free(task_group_cache, tg);
6270 /* allocate runqueue etc for a new task group */
6271 struct task_group *sched_create_group(struct task_group *parent)
6273 struct task_group *tg;
6275 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6277 return ERR_PTR(-ENOMEM);
6279 if (!alloc_fair_sched_group(tg, parent))
6282 if (!alloc_rt_sched_group(tg, parent))
6288 sched_free_group(tg);
6289 return ERR_PTR(-ENOMEM);
6292 void sched_online_group(struct task_group *tg, struct task_group *parent)
6294 unsigned long flags;
6296 spin_lock_irqsave(&task_group_lock, flags);
6297 list_add_rcu(&tg->list, &task_groups);
6299 /* Root should already exist: */
6302 tg->parent = parent;
6303 INIT_LIST_HEAD(&tg->children);
6304 list_add_rcu(&tg->siblings, &parent->children);
6305 spin_unlock_irqrestore(&task_group_lock, flags);
6307 online_fair_sched_group(tg);
6310 /* rcu callback to free various structures associated with a task group */
6311 static void sched_free_group_rcu(struct rcu_head *rhp)
6313 /* Now it should be safe to free those cfs_rqs: */
6314 sched_free_group(container_of(rhp, struct task_group, rcu));
6317 void sched_destroy_group(struct task_group *tg)
6319 /* Wait for possible concurrent references to cfs_rqs complete: */
6320 call_rcu(&tg->rcu, sched_free_group_rcu);
6323 void sched_offline_group(struct task_group *tg)
6325 unsigned long flags;
6327 /* End participation in shares distribution: */
6328 unregister_fair_sched_group(tg);
6330 spin_lock_irqsave(&task_group_lock, flags);
6331 list_del_rcu(&tg->list);
6332 list_del_rcu(&tg->siblings);
6333 spin_unlock_irqrestore(&task_group_lock, flags);
6336 static void sched_change_group(struct task_struct *tsk, int type)
6338 struct task_group *tg;
6341 * All callers are synchronized by task_rq_lock(); we do not use RCU
6342 * which is pointless here. Thus, we pass "true" to task_css_check()
6343 * to prevent lockdep warnings.
6345 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6346 struct task_group, css);
6347 tg = autogroup_task_group(tsk, tg);
6348 tsk->sched_task_group = tg;
6350 #ifdef CONFIG_FAIR_GROUP_SCHED
6351 if (tsk->sched_class->task_change_group)
6352 tsk->sched_class->task_change_group(tsk, type);
6355 set_task_rq(tsk, task_cpu(tsk));
6359 * Change task's runqueue when it moves between groups.
6361 * The caller of this function should have put the task in its new group by
6362 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6365 void sched_move_task(struct task_struct *tsk)
6367 int queued, running, queue_flags =
6368 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6372 rq = task_rq_lock(tsk, &rf);
6373 update_rq_clock(rq);
6375 running = task_current(rq, tsk);
6376 queued = task_on_rq_queued(tsk);
6379 dequeue_task(rq, tsk, queue_flags);
6381 put_prev_task(rq, tsk);
6383 sched_change_group(tsk, TASK_MOVE_GROUP);
6386 enqueue_task(rq, tsk, queue_flags);
6388 set_curr_task(rq, tsk);
6390 task_rq_unlock(rq, tsk, &rf);
6393 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6395 return css ? container_of(css, struct task_group, css) : NULL;
6398 static struct cgroup_subsys_state *
6399 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6401 struct task_group *parent = css_tg(parent_css);
6402 struct task_group *tg;
6405 /* This is early initialization for the top cgroup */
6406 return &root_task_group.css;
6409 tg = sched_create_group(parent);
6411 return ERR_PTR(-ENOMEM);
6416 /* Expose task group only after completing cgroup initialization */
6417 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6419 struct task_group *tg = css_tg(css);
6420 struct task_group *parent = css_tg(css->parent);
6423 sched_online_group(tg, parent);
6427 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6429 struct task_group *tg = css_tg(css);
6431 sched_offline_group(tg);
6434 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6436 struct task_group *tg = css_tg(css);
6439 * Relies on the RCU grace period between css_released() and this.
6441 sched_free_group(tg);
6445 * This is called before wake_up_new_task(), therefore we really only
6446 * have to set its group bits, all the other stuff does not apply.
6448 static void cpu_cgroup_fork(struct task_struct *task)
6453 rq = task_rq_lock(task, &rf);
6455 update_rq_clock(rq);
6456 sched_change_group(task, TASK_SET_GROUP);
6458 task_rq_unlock(rq, task, &rf);
6461 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6463 struct task_struct *task;
6464 struct cgroup_subsys_state *css;
6467 cgroup_taskset_for_each(task, css, tset) {
6468 #ifdef CONFIG_RT_GROUP_SCHED
6469 if (!sched_rt_can_attach(css_tg(css), task))
6472 /* We don't support RT-tasks being in separate groups */
6473 if (task->sched_class != &fair_sched_class)
6477 * Serialize against wake_up_new_task() such that if its
6478 * running, we're sure to observe its full state.
6480 raw_spin_lock_irq(&task->pi_lock);
6482 * Avoid calling sched_move_task() before wake_up_new_task()
6483 * has happened. This would lead to problems with PELT, due to
6484 * move wanting to detach+attach while we're not attached yet.
6486 if (task->state == TASK_NEW)
6488 raw_spin_unlock_irq(&task->pi_lock);
6496 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6498 struct task_struct *task;
6499 struct cgroup_subsys_state *css;
6501 cgroup_taskset_for_each(task, css, tset)
6502 sched_move_task(task);
6505 #ifdef CONFIG_FAIR_GROUP_SCHED
6506 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6507 struct cftype *cftype, u64 shareval)
6509 if (shareval > scale_load_down(ULONG_MAX))
6510 shareval = MAX_SHARES;
6511 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6514 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6517 struct task_group *tg = css_tg(css);
6519 return (u64) scale_load_down(tg->shares);
6522 #ifdef CONFIG_CFS_BANDWIDTH
6523 static DEFINE_MUTEX(cfs_constraints_mutex);
6525 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6526 static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6528 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6530 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6532 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6533 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6535 if (tg == &root_task_group)
6539 * Ensure we have at some amount of bandwidth every period. This is
6540 * to prevent reaching a state of large arrears when throttled via
6541 * entity_tick() resulting in prolonged exit starvation.
6543 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6547 * Likewise, bound things on the otherside by preventing insane quota
6548 * periods. This also allows us to normalize in computing quota
6551 if (period > max_cfs_quota_period)
6555 * Prevent race between setting of cfs_rq->runtime_enabled and
6556 * unthrottle_offline_cfs_rqs().
6559 mutex_lock(&cfs_constraints_mutex);
6560 ret = __cfs_schedulable(tg, period, quota);
6564 runtime_enabled = quota != RUNTIME_INF;
6565 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6567 * If we need to toggle cfs_bandwidth_used, off->on must occur
6568 * before making related changes, and on->off must occur afterwards
6570 if (runtime_enabled && !runtime_was_enabled)
6571 cfs_bandwidth_usage_inc();
6572 raw_spin_lock_irq(&cfs_b->lock);
6573 cfs_b->period = ns_to_ktime(period);
6574 cfs_b->quota = quota;
6576 __refill_cfs_bandwidth_runtime(cfs_b);
6578 /* Restart the period timer (if active) to handle new period expiry: */
6579 if (runtime_enabled)
6580 start_cfs_bandwidth(cfs_b);
6582 raw_spin_unlock_irq(&cfs_b->lock);
6584 for_each_online_cpu(i) {
6585 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6586 struct rq *rq = cfs_rq->rq;
6589 rq_lock_irq(rq, &rf);
6590 cfs_rq->runtime_enabled = runtime_enabled;
6591 cfs_rq->runtime_remaining = 0;
6593 if (cfs_rq->throttled)
6594 unthrottle_cfs_rq(cfs_rq);
6595 rq_unlock_irq(rq, &rf);
6597 if (runtime_was_enabled && !runtime_enabled)
6598 cfs_bandwidth_usage_dec();
6600 mutex_unlock(&cfs_constraints_mutex);
6606 static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
6610 period = ktime_to_ns(tg->cfs_bandwidth.period);
6611 if (cfs_quota_us < 0)
6612 quota = RUNTIME_INF;
6613 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
6614 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
6618 return tg_set_cfs_bandwidth(tg, period, quota);
6621 static long tg_get_cfs_quota(struct task_group *tg)
6625 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
6628 quota_us = tg->cfs_bandwidth.quota;
6629 do_div(quota_us, NSEC_PER_USEC);
6634 static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
6638 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
6641 period = (u64)cfs_period_us * NSEC_PER_USEC;
6642 quota = tg->cfs_bandwidth.quota;
6644 return tg_set_cfs_bandwidth(tg, period, quota);
6647 static long tg_get_cfs_period(struct task_group *tg)
6651 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
6652 do_div(cfs_period_us, NSEC_PER_USEC);
6654 return cfs_period_us;
6657 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
6660 return tg_get_cfs_quota(css_tg(css));
6663 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
6664 struct cftype *cftype, s64 cfs_quota_us)
6666 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
6669 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
6672 return tg_get_cfs_period(css_tg(css));
6675 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
6676 struct cftype *cftype, u64 cfs_period_us)
6678 return tg_set_cfs_period(css_tg(css), cfs_period_us);
6681 struct cfs_schedulable_data {
6682 struct task_group *tg;
6687 * normalize group quota/period to be quota/max_period
6688 * note: units are usecs
6690 static u64 normalize_cfs_quota(struct task_group *tg,
6691 struct cfs_schedulable_data *d)
6699 period = tg_get_cfs_period(tg);
6700 quota = tg_get_cfs_quota(tg);
6703 /* note: these should typically be equivalent */
6704 if (quota == RUNTIME_INF || quota == -1)
6707 return to_ratio(period, quota);
6710 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
6712 struct cfs_schedulable_data *d = data;
6713 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6714 s64 quota = 0, parent_quota = -1;
6717 quota = RUNTIME_INF;
6719 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
6721 quota = normalize_cfs_quota(tg, d);
6722 parent_quota = parent_b->hierarchical_quota;
6725 * Ensure max(child_quota) <= parent_quota. On cgroup2,
6726 * always take the min. On cgroup1, only inherit when no
6729 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
6730 quota = min(quota, parent_quota);
6732 if (quota == RUNTIME_INF)
6733 quota = parent_quota;
6734 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
6738 cfs_b->hierarchical_quota = quota;
6743 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
6746 struct cfs_schedulable_data data = {
6752 if (quota != RUNTIME_INF) {
6753 do_div(data.period, NSEC_PER_USEC);
6754 do_div(data.quota, NSEC_PER_USEC);
6758 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
6764 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
6766 struct task_group *tg = css_tg(seq_css(sf));
6767 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6769 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
6770 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
6771 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
6773 if (schedstat_enabled() && tg != &root_task_group) {
6777 for_each_possible_cpu(i)
6778 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
6780 seq_printf(sf, "wait_sum %llu\n", ws);
6785 #endif /* CONFIG_CFS_BANDWIDTH */
6786 #endif /* CONFIG_FAIR_GROUP_SCHED */
6788 #ifdef CONFIG_RT_GROUP_SCHED
6789 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
6790 struct cftype *cft, s64 val)
6792 return sched_group_set_rt_runtime(css_tg(css), val);
6795 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
6798 return sched_group_rt_runtime(css_tg(css));
6801 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
6802 struct cftype *cftype, u64 rt_period_us)
6804 return sched_group_set_rt_period(css_tg(css), rt_period_us);
6807 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
6810 return sched_group_rt_period(css_tg(css));
6812 #endif /* CONFIG_RT_GROUP_SCHED */
6814 static struct cftype cpu_legacy_files[] = {
6815 #ifdef CONFIG_FAIR_GROUP_SCHED
6818 .read_u64 = cpu_shares_read_u64,
6819 .write_u64 = cpu_shares_write_u64,
6822 #ifdef CONFIG_CFS_BANDWIDTH
6824 .name = "cfs_quota_us",
6825 .read_s64 = cpu_cfs_quota_read_s64,
6826 .write_s64 = cpu_cfs_quota_write_s64,
6829 .name = "cfs_period_us",
6830 .read_u64 = cpu_cfs_period_read_u64,
6831 .write_u64 = cpu_cfs_period_write_u64,
6835 .seq_show = cpu_cfs_stat_show,
6838 #ifdef CONFIG_RT_GROUP_SCHED
6840 .name = "rt_runtime_us",
6841 .read_s64 = cpu_rt_runtime_read,
6842 .write_s64 = cpu_rt_runtime_write,
6845 .name = "rt_period_us",
6846 .read_u64 = cpu_rt_period_read_uint,
6847 .write_u64 = cpu_rt_period_write_uint,
6853 static int cpu_extra_stat_show(struct seq_file *sf,
6854 struct cgroup_subsys_state *css)
6856 #ifdef CONFIG_CFS_BANDWIDTH
6858 struct task_group *tg = css_tg(css);
6859 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6862 throttled_usec = cfs_b->throttled_time;
6863 do_div(throttled_usec, NSEC_PER_USEC);
6865 seq_printf(sf, "nr_periods %d\n"
6867 "throttled_usec %llu\n",
6868 cfs_b->nr_periods, cfs_b->nr_throttled,
6875 #ifdef CONFIG_FAIR_GROUP_SCHED
6876 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
6879 struct task_group *tg = css_tg(css);
6880 u64 weight = scale_load_down(tg->shares);
6882 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
6885 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
6886 struct cftype *cft, u64 weight)
6889 * cgroup weight knobs should use the common MIN, DFL and MAX
6890 * values which are 1, 100 and 10000 respectively. While it loses
6891 * a bit of range on both ends, it maps pretty well onto the shares
6892 * value used by scheduler and the round-trip conversions preserve
6893 * the original value over the entire range.
6895 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
6898 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
6900 return sched_group_set_shares(css_tg(css), scale_load(weight));
6903 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
6906 unsigned long weight = scale_load_down(css_tg(css)->shares);
6907 int last_delta = INT_MAX;
6910 /* find the closest nice value to the current weight */
6911 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
6912 delta = abs(sched_prio_to_weight[prio] - weight);
6913 if (delta >= last_delta)
6918 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
6921 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
6922 struct cftype *cft, s64 nice)
6924 unsigned long weight;
6927 if (nice < MIN_NICE || nice > MAX_NICE)
6930 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
6931 idx = array_index_nospec(idx, 40);
6932 weight = sched_prio_to_weight[idx];
6934 return sched_group_set_shares(css_tg(css), scale_load(weight));
6938 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
6939 long period, long quota)
6942 seq_puts(sf, "max");
6944 seq_printf(sf, "%ld", quota);
6946 seq_printf(sf, " %ld\n", period);
6949 /* caller should put the current value in *@periodp before calling */
6950 static int __maybe_unused cpu_period_quota_parse(char *buf,
6951 u64 *periodp, u64 *quotap)
6953 char tok[21]; /* U64_MAX */
6955 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
6958 *periodp *= NSEC_PER_USEC;
6960 if (sscanf(tok, "%llu", quotap))
6961 *quotap *= NSEC_PER_USEC;
6962 else if (!strcmp(tok, "max"))
6963 *quotap = RUNTIME_INF;
6970 #ifdef CONFIG_CFS_BANDWIDTH
6971 static int cpu_max_show(struct seq_file *sf, void *v)
6973 struct task_group *tg = css_tg(seq_css(sf));
6975 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
6979 static ssize_t cpu_max_write(struct kernfs_open_file *of,
6980 char *buf, size_t nbytes, loff_t off)
6982 struct task_group *tg = css_tg(of_css(of));
6983 u64 period = tg_get_cfs_period(tg);
6987 ret = cpu_period_quota_parse(buf, &period, "a);
6989 ret = tg_set_cfs_bandwidth(tg, period, quota);
6990 return ret ?: nbytes;
6994 static struct cftype cpu_files[] = {
6995 #ifdef CONFIG_FAIR_GROUP_SCHED
6998 .flags = CFTYPE_NOT_ON_ROOT,
6999 .read_u64 = cpu_weight_read_u64,
7000 .write_u64 = cpu_weight_write_u64,
7003 .name = "weight.nice",
7004 .flags = CFTYPE_NOT_ON_ROOT,
7005 .read_s64 = cpu_weight_nice_read_s64,
7006 .write_s64 = cpu_weight_nice_write_s64,
7009 #ifdef CONFIG_CFS_BANDWIDTH
7012 .flags = CFTYPE_NOT_ON_ROOT,
7013 .seq_show = cpu_max_show,
7014 .write = cpu_max_write,
7020 struct cgroup_subsys cpu_cgrp_subsys = {
7021 .css_alloc = cpu_cgroup_css_alloc,
7022 .css_online = cpu_cgroup_css_online,
7023 .css_released = cpu_cgroup_css_released,
7024 .css_free = cpu_cgroup_css_free,
7025 .css_extra_stat_show = cpu_extra_stat_show,
7026 .fork = cpu_cgroup_fork,
7027 .can_attach = cpu_cgroup_can_attach,
7028 .attach = cpu_cgroup_attach,
7029 .legacy_cftypes = cpu_legacy_files,
7030 .dfl_cftypes = cpu_files,
7035 #endif /* CONFIG_CGROUP_SCHED */
7037 void dump_cpu_task(int cpu)
7039 pr_info("Task dump for CPU %d:\n", cpu);
7040 sched_show_task(cpu_curr(cpu));
7044 * Nice levels are multiplicative, with a gentle 10% change for every
7045 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7046 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7047 * that remained on nice 0.
7049 * The "10% effect" is relative and cumulative: from _any_ nice level,
7050 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7051 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7052 * If a task goes up by ~10% and another task goes down by ~10% then
7053 * the relative distance between them is ~25%.)
7055 const int sched_prio_to_weight[40] = {
7056 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7057 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7058 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7059 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7060 /* 0 */ 1024, 820, 655, 526, 423,
7061 /* 5 */ 335, 272, 215, 172, 137,
7062 /* 10 */ 110, 87, 70, 56, 45,
7063 /* 15 */ 36, 29, 23, 18, 15,
7067 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7069 * In cases where the weight does not change often, we can use the
7070 * precalculated inverse to speed up arithmetics by turning divisions
7071 * into multiplications:
7073 const u32 sched_prio_to_wmult[40] = {
7074 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7075 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7076 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7077 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7078 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7079 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7080 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7081 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7084 #undef CREATE_TRACE_POINTS