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);
797 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
799 if (task_contributes_to_load(p))
800 rq->nr_uninterruptible++;
802 dequeue_task(rq, p, flags);
806 * __normal_prio - return the priority that is based on the static prio
808 static inline int __normal_prio(struct task_struct *p)
810 return p->static_prio;
814 * Calculate the expected normal priority: i.e. priority
815 * without taking RT-inheritance into account. Might be
816 * boosted by interactivity modifiers. Changes upon fork,
817 * setprio syscalls, and whenever the interactivity
818 * estimator recalculates.
820 static inline int normal_prio(struct task_struct *p)
824 if (task_has_dl_policy(p))
825 prio = MAX_DL_PRIO-1;
826 else if (task_has_rt_policy(p))
827 prio = MAX_RT_PRIO-1 - p->rt_priority;
829 prio = __normal_prio(p);
834 * Calculate the current priority, i.e. the priority
835 * taken into account by the scheduler. This value might
836 * be boosted by RT tasks, or might be boosted by
837 * interactivity modifiers. Will be RT if the task got
838 * RT-boosted. If not then it returns p->normal_prio.
840 static int effective_prio(struct task_struct *p)
842 p->normal_prio = normal_prio(p);
844 * If we are RT tasks or we were boosted to RT priority,
845 * keep the priority unchanged. Otherwise, update priority
846 * to the normal priority:
848 if (!rt_prio(p->prio))
849 return p->normal_prio;
854 * task_curr - is this task currently executing on a CPU?
855 * @p: the task in question.
857 * Return: 1 if the task is currently executing. 0 otherwise.
859 inline int task_curr(const struct task_struct *p)
861 return cpu_curr(task_cpu(p)) == p;
865 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
866 * use the balance_callback list if you want balancing.
868 * this means any call to check_class_changed() must be followed by a call to
869 * balance_callback().
871 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
872 const struct sched_class *prev_class,
875 if (prev_class != p->sched_class) {
876 if (prev_class->switched_from)
877 prev_class->switched_from(rq, p);
879 p->sched_class->switched_to(rq, p);
880 } else if (oldprio != p->prio || dl_task(p))
881 p->sched_class->prio_changed(rq, p, oldprio);
884 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
886 const struct sched_class *class;
888 if (p->sched_class == rq->curr->sched_class) {
889 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
891 for_each_class(class) {
892 if (class == rq->curr->sched_class)
894 if (class == p->sched_class) {
902 * A queue event has occurred, and we're going to schedule. In
903 * this case, we can save a useless back to back clock update.
905 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
906 rq_clock_skip_update(rq);
911 static inline bool is_per_cpu_kthread(struct task_struct *p)
913 if (!(p->flags & PF_KTHREAD))
916 if (p->nr_cpus_allowed != 1)
923 * Per-CPU kthreads are allowed to run on !actie && online CPUs, see
924 * __set_cpus_allowed_ptr() and select_fallback_rq().
926 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
928 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
931 if (is_per_cpu_kthread(p))
932 return cpu_online(cpu);
934 return cpu_active(cpu);
938 * This is how migration works:
940 * 1) we invoke migration_cpu_stop() on the target CPU using
942 * 2) stopper starts to run (implicitly forcing the migrated thread
944 * 3) it checks whether the migrated task is still in the wrong runqueue.
945 * 4) if it's in the wrong runqueue then the migration thread removes
946 * it and puts it into the right queue.
947 * 5) stopper completes and stop_one_cpu() returns and the migration
952 * move_queued_task - move a queued task to new rq.
954 * Returns (locked) new rq. Old rq's lock is released.
956 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
957 struct task_struct *p, int new_cpu)
959 lockdep_assert_held(&rq->lock);
961 WRITE_ONCE(p->on_rq, TASK_ON_RQ_MIGRATING);
962 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
963 set_task_cpu(p, new_cpu);
966 rq = cpu_rq(new_cpu);
969 BUG_ON(task_cpu(p) != new_cpu);
970 enqueue_task(rq, p, 0);
971 p->on_rq = TASK_ON_RQ_QUEUED;
972 check_preempt_curr(rq, p, 0);
977 struct migration_arg {
978 struct task_struct *task;
983 * Move (not current) task off this CPU, onto the destination CPU. We're doing
984 * this because either it can't run here any more (set_cpus_allowed()
985 * away from this CPU, or CPU going down), or because we're
986 * attempting to rebalance this task on exec (sched_exec).
988 * So we race with normal scheduler movements, but that's OK, as long
989 * as the task is no longer on this CPU.
991 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
992 struct task_struct *p, int dest_cpu)
994 /* Affinity changed (again). */
995 if (!is_cpu_allowed(p, dest_cpu))
999 rq = move_queued_task(rq, rf, p, dest_cpu);
1005 * migration_cpu_stop - this will be executed by a highprio stopper thread
1006 * and performs thread migration by bumping thread off CPU then
1007 * 'pushing' onto another runqueue.
1009 static int migration_cpu_stop(void *data)
1011 struct migration_arg *arg = data;
1012 struct task_struct *p = arg->task;
1013 struct rq *rq = this_rq();
1017 * The original target CPU might have gone down and we might
1018 * be on another CPU but it doesn't matter.
1020 local_irq_disable();
1022 * We need to explicitly wake pending tasks before running
1023 * __migrate_task() such that we will not miss enforcing cpus_allowed
1024 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1026 sched_ttwu_pending();
1028 raw_spin_lock(&p->pi_lock);
1031 * If task_rq(p) != rq, it cannot be migrated here, because we're
1032 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1033 * we're holding p->pi_lock.
1035 if (task_rq(p) == rq) {
1036 if (task_on_rq_queued(p))
1037 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1039 p->wake_cpu = arg->dest_cpu;
1042 raw_spin_unlock(&p->pi_lock);
1049 * sched_class::set_cpus_allowed must do the below, but is not required to
1050 * actually call this function.
1052 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1054 cpumask_copy(&p->cpus_allowed, new_mask);
1055 p->nr_cpus_allowed = cpumask_weight(new_mask);
1058 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1060 struct rq *rq = task_rq(p);
1061 bool queued, running;
1063 lockdep_assert_held(&p->pi_lock);
1065 queued = task_on_rq_queued(p);
1066 running = task_current(rq, p);
1070 * Because __kthread_bind() calls this on blocked tasks without
1073 lockdep_assert_held(&rq->lock);
1074 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1077 put_prev_task(rq, p);
1079 p->sched_class->set_cpus_allowed(p, new_mask);
1082 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1084 set_curr_task(rq, p);
1088 * Change a given task's CPU affinity. Migrate the thread to a
1089 * proper CPU and schedule it away if the CPU it's executing on
1090 * is removed from the allowed bitmask.
1092 * NOTE: the caller must have a valid reference to the task, the
1093 * task must not exit() & deallocate itself prematurely. The
1094 * call is not atomic; no spinlocks may be held.
1096 static int __set_cpus_allowed_ptr(struct task_struct *p,
1097 const struct cpumask *new_mask, bool check)
1099 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1100 unsigned int dest_cpu;
1105 rq = task_rq_lock(p, &rf);
1106 update_rq_clock(rq);
1108 if (p->flags & PF_KTHREAD) {
1110 * Kernel threads are allowed on online && !active CPUs
1112 cpu_valid_mask = cpu_online_mask;
1116 * Must re-check here, to close a race against __kthread_bind(),
1117 * sched_setaffinity() is not guaranteed to observe the flag.
1119 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1124 if (cpumask_equal(&p->cpus_allowed, new_mask))
1127 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1132 do_set_cpus_allowed(p, new_mask);
1134 if (p->flags & PF_KTHREAD) {
1136 * For kernel threads that do indeed end up on online &&
1137 * !active we want to ensure they are strict per-CPU threads.
1139 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1140 !cpumask_intersects(new_mask, cpu_active_mask) &&
1141 p->nr_cpus_allowed != 1);
1144 /* Can the task run on the task's current CPU? If so, we're done */
1145 if (cpumask_test_cpu(task_cpu(p), new_mask))
1148 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1149 if (task_running(rq, p) || p->state == TASK_WAKING) {
1150 struct migration_arg arg = { p, dest_cpu };
1151 /* Need help from migration thread: drop lock and wait. */
1152 task_rq_unlock(rq, p, &rf);
1153 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1154 tlb_migrate_finish(p->mm);
1156 } else if (task_on_rq_queued(p)) {
1158 * OK, since we're going to drop the lock immediately
1159 * afterwards anyway.
1161 rq = move_queued_task(rq, &rf, p, dest_cpu);
1164 task_rq_unlock(rq, p, &rf);
1169 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1171 return __set_cpus_allowed_ptr(p, new_mask, false);
1173 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1175 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1177 #ifdef CONFIG_SCHED_DEBUG
1179 * We should never call set_task_cpu() on a blocked task,
1180 * ttwu() will sort out the placement.
1182 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1186 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1187 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1188 * time relying on p->on_rq.
1190 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1191 p->sched_class == &fair_sched_class &&
1192 (p->on_rq && !task_on_rq_migrating(p)));
1194 #ifdef CONFIG_LOCKDEP
1196 * The caller should hold either p->pi_lock or rq->lock, when changing
1197 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1199 * sched_move_task() holds both and thus holding either pins the cgroup,
1202 * Furthermore, all task_rq users should acquire both locks, see
1205 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1206 lockdep_is_held(&task_rq(p)->lock)));
1209 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1211 WARN_ON_ONCE(!cpu_online(new_cpu));
1214 trace_sched_migrate_task(p, new_cpu);
1216 if (task_cpu(p) != new_cpu) {
1217 if (p->sched_class->migrate_task_rq)
1218 p->sched_class->migrate_task_rq(p, new_cpu);
1219 p->se.nr_migrations++;
1221 perf_event_task_migrate(p);
1224 __set_task_cpu(p, new_cpu);
1227 #ifdef CONFIG_NUMA_BALANCING
1228 static void __migrate_swap_task(struct task_struct *p, int cpu)
1230 if (task_on_rq_queued(p)) {
1231 struct rq *src_rq, *dst_rq;
1232 struct rq_flags srf, drf;
1234 src_rq = task_rq(p);
1235 dst_rq = cpu_rq(cpu);
1237 rq_pin_lock(src_rq, &srf);
1238 rq_pin_lock(dst_rq, &drf);
1240 p->on_rq = TASK_ON_RQ_MIGRATING;
1241 deactivate_task(src_rq, p, 0);
1242 set_task_cpu(p, cpu);
1243 activate_task(dst_rq, p, 0);
1244 p->on_rq = TASK_ON_RQ_QUEUED;
1245 check_preempt_curr(dst_rq, p, 0);
1247 rq_unpin_lock(dst_rq, &drf);
1248 rq_unpin_lock(src_rq, &srf);
1252 * Task isn't running anymore; make it appear like we migrated
1253 * it before it went to sleep. This means on wakeup we make the
1254 * previous CPU our target instead of where it really is.
1260 struct migration_swap_arg {
1261 struct task_struct *src_task, *dst_task;
1262 int src_cpu, dst_cpu;
1265 static int migrate_swap_stop(void *data)
1267 struct migration_swap_arg *arg = data;
1268 struct rq *src_rq, *dst_rq;
1271 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1274 src_rq = cpu_rq(arg->src_cpu);
1275 dst_rq = cpu_rq(arg->dst_cpu);
1277 double_raw_lock(&arg->src_task->pi_lock,
1278 &arg->dst_task->pi_lock);
1279 double_rq_lock(src_rq, dst_rq);
1281 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1284 if (task_cpu(arg->src_task) != arg->src_cpu)
1287 if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1290 if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1293 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1294 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1299 double_rq_unlock(src_rq, dst_rq);
1300 raw_spin_unlock(&arg->dst_task->pi_lock);
1301 raw_spin_unlock(&arg->src_task->pi_lock);
1307 * Cross migrate two tasks
1309 int migrate_swap(struct task_struct *cur, struct task_struct *p,
1310 int target_cpu, int curr_cpu)
1312 struct migration_swap_arg arg;
1315 arg = (struct migration_swap_arg){
1317 .src_cpu = curr_cpu,
1319 .dst_cpu = target_cpu,
1322 if (arg.src_cpu == arg.dst_cpu)
1326 * These three tests are all lockless; this is OK since all of them
1327 * will be re-checked with proper locks held further down the line.
1329 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1332 if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1335 if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1338 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1339 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1344 #endif /* CONFIG_NUMA_BALANCING */
1347 * wait_task_inactive - wait for a thread to unschedule.
1349 * If @match_state is nonzero, it's the @p->state value just checked and
1350 * not expected to change. If it changes, i.e. @p might have woken up,
1351 * then return zero. When we succeed in waiting for @p to be off its CPU,
1352 * we return a positive number (its total switch count). If a second call
1353 * a short while later returns the same number, the caller can be sure that
1354 * @p has remained unscheduled the whole time.
1356 * The caller must ensure that the task *will* unschedule sometime soon,
1357 * else this function might spin for a *long* time. This function can't
1358 * be called with interrupts off, or it may introduce deadlock with
1359 * smp_call_function() if an IPI is sent by the same process we are
1360 * waiting to become inactive.
1362 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1364 int running, queued;
1371 * We do the initial early heuristics without holding
1372 * any task-queue locks at all. We'll only try to get
1373 * the runqueue lock when things look like they will
1379 * If the task is actively running on another CPU
1380 * still, just relax and busy-wait without holding
1383 * NOTE! Since we don't hold any locks, it's not
1384 * even sure that "rq" stays as the right runqueue!
1385 * But we don't care, since "task_running()" will
1386 * return false if the runqueue has changed and p
1387 * is actually now running somewhere else!
1389 while (task_running(rq, p)) {
1390 if (match_state && unlikely(p->state != match_state))
1396 * Ok, time to look more closely! We need the rq
1397 * lock now, to be *sure*. If we're wrong, we'll
1398 * just go back and repeat.
1400 rq = task_rq_lock(p, &rf);
1401 trace_sched_wait_task(p);
1402 running = task_running(rq, p);
1403 queued = task_on_rq_queued(p);
1405 if (!match_state || p->state == match_state)
1406 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1407 task_rq_unlock(rq, p, &rf);
1410 * If it changed from the expected state, bail out now.
1412 if (unlikely(!ncsw))
1416 * Was it really running after all now that we
1417 * checked with the proper locks actually held?
1419 * Oops. Go back and try again..
1421 if (unlikely(running)) {
1427 * It's not enough that it's not actively running,
1428 * it must be off the runqueue _entirely_, and not
1431 * So if it was still runnable (but just not actively
1432 * running right now), it's preempted, and we should
1433 * yield - it could be a while.
1435 if (unlikely(queued)) {
1436 ktime_t to = NSEC_PER_SEC / HZ;
1438 set_current_state(TASK_UNINTERRUPTIBLE);
1439 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1444 * Ahh, all good. It wasn't running, and it wasn't
1445 * runnable, which means that it will never become
1446 * running in the future either. We're all done!
1455 * kick_process - kick a running thread to enter/exit the kernel
1456 * @p: the to-be-kicked thread
1458 * Cause a process which is running on another CPU to enter
1459 * kernel-mode, without any delay. (to get signals handled.)
1461 * NOTE: this function doesn't have to take the runqueue lock,
1462 * because all it wants to ensure is that the remote task enters
1463 * the kernel. If the IPI races and the task has been migrated
1464 * to another CPU then no harm is done and the purpose has been
1467 void kick_process(struct task_struct *p)
1473 if ((cpu != smp_processor_id()) && task_curr(p))
1474 smp_send_reschedule(cpu);
1477 EXPORT_SYMBOL_GPL(kick_process);
1480 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1482 * A few notes on cpu_active vs cpu_online:
1484 * - cpu_active must be a subset of cpu_online
1486 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
1487 * see __set_cpus_allowed_ptr(). At this point the newly online
1488 * CPU isn't yet part of the sched domains, and balancing will not
1491 * - on CPU-down we clear cpu_active() to mask the sched domains and
1492 * avoid the load balancer to place new tasks on the to be removed
1493 * CPU. Existing tasks will remain running there and will be taken
1496 * This means that fallback selection must not select !active CPUs.
1497 * And can assume that any active CPU must be online. Conversely
1498 * select_task_rq() below may allow selection of !active CPUs in order
1499 * to satisfy the above rules.
1501 static int select_fallback_rq(int cpu, struct task_struct *p)
1503 int nid = cpu_to_node(cpu);
1504 const struct cpumask *nodemask = NULL;
1505 enum { cpuset, possible, fail } state = cpuset;
1509 * If the node that the CPU is on has been offlined, cpu_to_node()
1510 * will return -1. There is no CPU on the node, and we should
1511 * select the CPU on the other node.
1514 nodemask = cpumask_of_node(nid);
1516 /* Look for allowed, online CPU in same node. */
1517 for_each_cpu(dest_cpu, nodemask) {
1518 if (!cpu_active(dest_cpu))
1520 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1526 /* Any allowed, online CPU? */
1527 for_each_cpu(dest_cpu, &p->cpus_allowed) {
1528 if (!is_cpu_allowed(p, dest_cpu))
1534 /* No more Mr. Nice Guy. */
1537 if (IS_ENABLED(CONFIG_CPUSETS)) {
1538 cpuset_cpus_allowed_fallback(p);
1544 do_set_cpus_allowed(p, cpu_possible_mask);
1555 if (state != cpuset) {
1557 * Don't tell them about moving exiting tasks or
1558 * kernel threads (both mm NULL), since they never
1561 if (p->mm && printk_ratelimit()) {
1562 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1563 task_pid_nr(p), p->comm, cpu);
1571 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1574 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1576 lockdep_assert_held(&p->pi_lock);
1578 if (p->nr_cpus_allowed > 1)
1579 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1581 cpu = cpumask_any(&p->cpus_allowed);
1584 * In order not to call set_task_cpu() on a blocking task we need
1585 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1588 * Since this is common to all placement strategies, this lives here.
1590 * [ this allows ->select_task() to simply return task_cpu(p) and
1591 * not worry about this generic constraint ]
1593 if (unlikely(!is_cpu_allowed(p, cpu)))
1594 cpu = select_fallback_rq(task_cpu(p), p);
1599 static void update_avg(u64 *avg, u64 sample)
1601 s64 diff = sample - *avg;
1605 void sched_set_stop_task(int cpu, struct task_struct *stop)
1607 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1608 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1612 * Make it appear like a SCHED_FIFO task, its something
1613 * userspace knows about and won't get confused about.
1615 * Also, it will make PI more or less work without too
1616 * much confusion -- but then, stop work should not
1617 * rely on PI working anyway.
1619 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1621 stop->sched_class = &stop_sched_class;
1624 cpu_rq(cpu)->stop = stop;
1628 * Reset it back to a normal scheduling class so that
1629 * it can die in pieces.
1631 old_stop->sched_class = &rt_sched_class;
1637 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1638 const struct cpumask *new_mask, bool check)
1640 return set_cpus_allowed_ptr(p, new_mask);
1643 #endif /* CONFIG_SMP */
1646 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1650 if (!schedstat_enabled())
1656 if (cpu == rq->cpu) {
1657 __schedstat_inc(rq->ttwu_local);
1658 __schedstat_inc(p->se.statistics.nr_wakeups_local);
1660 struct sched_domain *sd;
1662 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
1664 for_each_domain(rq->cpu, sd) {
1665 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1666 __schedstat_inc(sd->ttwu_wake_remote);
1673 if (wake_flags & WF_MIGRATED)
1674 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1675 #endif /* CONFIG_SMP */
1677 __schedstat_inc(rq->ttwu_count);
1678 __schedstat_inc(p->se.statistics.nr_wakeups);
1680 if (wake_flags & WF_SYNC)
1681 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
1684 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1686 activate_task(rq, p, en_flags);
1687 p->on_rq = TASK_ON_RQ_QUEUED;
1689 /* If a worker is waking up, notify the workqueue: */
1690 if (p->flags & PF_WQ_WORKER)
1691 wq_worker_waking_up(p, cpu_of(rq));
1695 * Mark the task runnable and perform wakeup-preemption.
1697 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1698 struct rq_flags *rf)
1700 check_preempt_curr(rq, p, wake_flags);
1701 p->state = TASK_RUNNING;
1702 trace_sched_wakeup(p);
1705 if (p->sched_class->task_woken) {
1707 * Our task @p is fully woken up and running; so its safe to
1708 * drop the rq->lock, hereafter rq is only used for statistics.
1710 rq_unpin_lock(rq, rf);
1711 p->sched_class->task_woken(rq, p);
1712 rq_repin_lock(rq, rf);
1715 if (rq->idle_stamp) {
1716 u64 delta = rq_clock(rq) - rq->idle_stamp;
1717 u64 max = 2*rq->max_idle_balance_cost;
1719 update_avg(&rq->avg_idle, delta);
1721 if (rq->avg_idle > max)
1730 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1731 struct rq_flags *rf)
1733 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
1735 lockdep_assert_held(&rq->lock);
1738 if (p->sched_contributes_to_load)
1739 rq->nr_uninterruptible--;
1741 if (wake_flags & WF_MIGRATED)
1742 en_flags |= ENQUEUE_MIGRATED;
1745 ttwu_activate(rq, p, en_flags);
1746 ttwu_do_wakeup(rq, p, wake_flags, rf);
1750 * Called in case the task @p isn't fully descheduled from its runqueue,
1751 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1752 * since all we need to do is flip p->state to TASK_RUNNING, since
1753 * the task is still ->on_rq.
1755 static int ttwu_remote(struct task_struct *p, int wake_flags)
1761 rq = __task_rq_lock(p, &rf);
1762 if (task_on_rq_queued(p)) {
1763 /* check_preempt_curr() may use rq clock */
1764 update_rq_clock(rq);
1765 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1768 __task_rq_unlock(rq, &rf);
1774 void sched_ttwu_pending(void)
1776 struct rq *rq = this_rq();
1777 struct llist_node *llist = llist_del_all(&rq->wake_list);
1778 struct task_struct *p, *t;
1784 rq_lock_irqsave(rq, &rf);
1785 update_rq_clock(rq);
1787 llist_for_each_entry_safe(p, t, llist, wake_entry)
1788 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
1790 rq_unlock_irqrestore(rq, &rf);
1793 void scheduler_ipi(void)
1796 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1797 * TIF_NEED_RESCHED remotely (for the first time) will also send
1800 preempt_fold_need_resched();
1802 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1806 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1807 * traditionally all their work was done from the interrupt return
1808 * path. Now that we actually do some work, we need to make sure
1811 * Some archs already do call them, luckily irq_enter/exit nest
1814 * Arguably we should visit all archs and update all handlers,
1815 * however a fair share of IPIs are still resched only so this would
1816 * somewhat pessimize the simple resched case.
1819 sched_ttwu_pending();
1822 * Check if someone kicked us for doing the nohz idle load balance.
1824 if (unlikely(got_nohz_idle_kick())) {
1825 this_rq()->idle_balance = 1;
1826 raise_softirq_irqoff(SCHED_SOFTIRQ);
1831 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1833 struct rq *rq = cpu_rq(cpu);
1835 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1837 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1838 if (!set_nr_if_polling(rq->idle))
1839 smp_send_reschedule(cpu);
1841 trace_sched_wake_idle_without_ipi(cpu);
1845 void wake_up_if_idle(int cpu)
1847 struct rq *rq = cpu_rq(cpu);
1852 if (!is_idle_task(rcu_dereference(rq->curr)))
1855 if (set_nr_if_polling(rq->idle)) {
1856 trace_sched_wake_idle_without_ipi(cpu);
1858 rq_lock_irqsave(rq, &rf);
1859 if (is_idle_task(rq->curr))
1860 smp_send_reschedule(cpu);
1861 /* Else CPU is not idle, do nothing here: */
1862 rq_unlock_irqrestore(rq, &rf);
1869 bool cpus_share_cache(int this_cpu, int that_cpu)
1871 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1873 #endif /* CONFIG_SMP */
1875 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1877 struct rq *rq = cpu_rq(cpu);
1880 #if defined(CONFIG_SMP)
1881 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1882 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1883 ttwu_queue_remote(p, cpu, wake_flags);
1889 update_rq_clock(rq);
1890 ttwu_do_activate(rq, p, wake_flags, &rf);
1895 * Notes on Program-Order guarantees on SMP systems.
1899 * The basic program-order guarantee on SMP systems is that when a task [t]
1900 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1901 * execution on its new CPU [c1].
1903 * For migration (of runnable tasks) this is provided by the following means:
1905 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1906 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1907 * rq(c1)->lock (if not at the same time, then in that order).
1908 * C) LOCK of the rq(c1)->lock scheduling in task
1910 * Release/acquire chaining guarantees that B happens after A and C after B.
1911 * Note: the CPU doing B need not be c0 or c1
1920 * UNLOCK rq(0)->lock
1922 * LOCK rq(0)->lock // orders against CPU0
1924 * UNLOCK rq(0)->lock
1928 * UNLOCK rq(1)->lock
1930 * LOCK rq(1)->lock // orders against CPU2
1933 * UNLOCK rq(1)->lock
1936 * BLOCKING -- aka. SLEEP + WAKEUP
1938 * For blocking we (obviously) need to provide the same guarantee as for
1939 * migration. However the means are completely different as there is no lock
1940 * chain to provide order. Instead we do:
1942 * 1) smp_store_release(X->on_cpu, 0)
1943 * 2) smp_cond_load_acquire(!X->on_cpu)
1947 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1949 * LOCK rq(0)->lock LOCK X->pi_lock
1952 * smp_store_release(X->on_cpu, 0);
1954 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1960 * X->state = RUNNING
1961 * UNLOCK rq(2)->lock
1963 * LOCK rq(2)->lock // orders against CPU1
1966 * UNLOCK rq(2)->lock
1969 * UNLOCK rq(0)->lock
1972 * However, for wakeups there is a second guarantee we must provide, namely we
1973 * must ensure that CONDITION=1 done by the caller can not be reordered with
1974 * accesses to the task state; see try_to_wake_up() and set_current_state().
1978 * try_to_wake_up - wake up a thread
1979 * @p: the thread to be awakened
1980 * @state: the mask of task states that can be woken
1981 * @wake_flags: wake modifier flags (WF_*)
1983 * If (@state & @p->state) @p->state = TASK_RUNNING.
1985 * If the task was not queued/runnable, also place it back on a runqueue.
1987 * Atomic against schedule() which would dequeue a task, also see
1988 * set_current_state().
1990 * This function executes a full memory barrier before accessing the task
1991 * state; see set_current_state().
1993 * Return: %true if @p->state changes (an actual wakeup was done),
1997 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1999 unsigned long flags;
2000 int cpu, success = 0;
2003 * If we are going to wake up a thread waiting for CONDITION we
2004 * need to ensure that CONDITION=1 done by the caller can not be
2005 * reordered with p->state check below. This pairs with mb() in
2006 * set_current_state() the waiting thread does.
2008 raw_spin_lock_irqsave(&p->pi_lock, flags);
2009 smp_mb__after_spinlock();
2010 if (!(p->state & state))
2013 trace_sched_waking(p);
2015 /* We're going to change ->state: */
2020 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2021 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2022 * in smp_cond_load_acquire() below.
2024 * sched_ttwu_pending() try_to_wake_up()
2025 * STORE p->on_rq = 1 LOAD p->state
2028 * __schedule() (switch to task 'p')
2029 * LOCK rq->lock smp_rmb();
2030 * smp_mb__after_spinlock();
2034 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
2036 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2037 * __schedule(). See the comment for smp_mb__after_spinlock().
2040 if (p->on_rq && ttwu_remote(p, wake_flags))
2045 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2046 * possible to, falsely, observe p->on_cpu == 0.
2048 * One must be running (->on_cpu == 1) in order to remove oneself
2049 * from the runqueue.
2051 * __schedule() (switch to task 'p') try_to_wake_up()
2052 * STORE p->on_cpu = 1 LOAD p->on_rq
2055 * __schedule() (put 'p' to sleep)
2056 * LOCK rq->lock smp_rmb();
2057 * smp_mb__after_spinlock();
2058 * STORE p->on_rq = 0 LOAD p->on_cpu
2060 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2061 * __schedule(). See the comment for smp_mb__after_spinlock().
2066 * If the owning (remote) CPU is still in the middle of schedule() with
2067 * this task as prev, wait until its done referencing the task.
2069 * Pairs with the smp_store_release() in finish_task().
2071 * This ensures that tasks getting woken will be fully ordered against
2072 * their previous state and preserve Program Order.
2074 smp_cond_load_acquire(&p->on_cpu, !VAL);
2076 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2077 p->state = TASK_WAKING;
2080 delayacct_blkio_end(p);
2081 atomic_dec(&task_rq(p)->nr_iowait);
2084 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2085 if (task_cpu(p) != cpu) {
2086 wake_flags |= WF_MIGRATED;
2087 psi_ttwu_dequeue(p);
2088 set_task_cpu(p, cpu);
2091 #else /* CONFIG_SMP */
2094 delayacct_blkio_end(p);
2095 atomic_dec(&task_rq(p)->nr_iowait);
2098 #endif /* CONFIG_SMP */
2100 ttwu_queue(p, cpu, wake_flags);
2102 ttwu_stat(p, cpu, wake_flags);
2104 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2110 * try_to_wake_up_local - try to wake up a local task with rq lock held
2111 * @p: the thread to be awakened
2112 * @rf: request-queue flags for pinning
2114 * Put @p on the run-queue if it's not already there. The caller must
2115 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2118 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2120 struct rq *rq = task_rq(p);
2122 if (WARN_ON_ONCE(rq != this_rq()) ||
2123 WARN_ON_ONCE(p == current))
2126 lockdep_assert_held(&rq->lock);
2128 if (!raw_spin_trylock(&p->pi_lock)) {
2130 * This is OK, because current is on_cpu, which avoids it being
2131 * picked for load-balance and preemption/IRQs are still
2132 * disabled avoiding further scheduler activity on it and we've
2133 * not yet picked a replacement task.
2136 raw_spin_lock(&p->pi_lock);
2140 if (!(p->state & TASK_NORMAL))
2143 trace_sched_waking(p);
2145 if (!task_on_rq_queued(p)) {
2147 delayacct_blkio_end(p);
2148 atomic_dec(&rq->nr_iowait);
2150 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK);
2153 ttwu_do_wakeup(rq, p, 0, rf);
2154 ttwu_stat(p, smp_processor_id(), 0);
2156 raw_spin_unlock(&p->pi_lock);
2160 * wake_up_process - Wake up a specific process
2161 * @p: The process to be woken up.
2163 * Attempt to wake up the nominated process and move it to the set of runnable
2166 * Return: 1 if the process was woken up, 0 if it was already running.
2168 * This function executes a full memory barrier before accessing the task state.
2170 int wake_up_process(struct task_struct *p)
2172 return try_to_wake_up(p, TASK_NORMAL, 0);
2174 EXPORT_SYMBOL(wake_up_process);
2176 int wake_up_state(struct task_struct *p, unsigned int state)
2178 return try_to_wake_up(p, state, 0);
2182 * Perform scheduler related setup for a newly forked process p.
2183 * p is forked by current.
2185 * __sched_fork() is basic setup used by init_idle() too:
2187 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2192 p->se.exec_start = 0;
2193 p->se.sum_exec_runtime = 0;
2194 p->se.prev_sum_exec_runtime = 0;
2195 p->se.nr_migrations = 0;
2197 INIT_LIST_HEAD(&p->se.group_node);
2199 #ifdef CONFIG_FAIR_GROUP_SCHED
2200 p->se.cfs_rq = NULL;
2203 #ifdef CONFIG_SCHEDSTATS
2204 /* Even if schedstat is disabled, there should not be garbage */
2205 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2208 RB_CLEAR_NODE(&p->dl.rb_node);
2209 init_dl_task_timer(&p->dl);
2210 init_dl_inactive_task_timer(&p->dl);
2211 __dl_clear_params(p);
2213 INIT_LIST_HEAD(&p->rt.run_list);
2215 p->rt.time_slice = sched_rr_timeslice;
2219 #ifdef CONFIG_PREEMPT_NOTIFIERS
2220 INIT_HLIST_HEAD(&p->preempt_notifiers);
2223 init_numa_balancing(clone_flags, p);
2226 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2228 #ifdef CONFIG_NUMA_BALANCING
2230 void set_numabalancing_state(bool enabled)
2233 static_branch_enable(&sched_numa_balancing);
2235 static_branch_disable(&sched_numa_balancing);
2238 #ifdef CONFIG_PROC_SYSCTL
2239 int sysctl_numa_balancing(struct ctl_table *table, int write,
2240 void __user *buffer, size_t *lenp, loff_t *ppos)
2244 int state = static_branch_likely(&sched_numa_balancing);
2246 if (write && !capable(CAP_SYS_ADMIN))
2251 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2255 set_numabalancing_state(state);
2261 #ifdef CONFIG_SCHEDSTATS
2263 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2264 static bool __initdata __sched_schedstats = false;
2266 static void set_schedstats(bool enabled)
2269 static_branch_enable(&sched_schedstats);
2271 static_branch_disable(&sched_schedstats);
2274 void force_schedstat_enabled(void)
2276 if (!schedstat_enabled()) {
2277 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2278 static_branch_enable(&sched_schedstats);
2282 static int __init setup_schedstats(char *str)
2289 * This code is called before jump labels have been set up, so we can't
2290 * change the static branch directly just yet. Instead set a temporary
2291 * variable so init_schedstats() can do it later.
2293 if (!strcmp(str, "enable")) {
2294 __sched_schedstats = true;
2296 } else if (!strcmp(str, "disable")) {
2297 __sched_schedstats = false;
2302 pr_warn("Unable to parse schedstats=\n");
2306 __setup("schedstats=", setup_schedstats);
2308 static void __init init_schedstats(void)
2310 set_schedstats(__sched_schedstats);
2313 #ifdef CONFIG_PROC_SYSCTL
2314 int sysctl_schedstats(struct ctl_table *table, int write,
2315 void __user *buffer, size_t *lenp, loff_t *ppos)
2319 int state = static_branch_likely(&sched_schedstats);
2321 if (write && !capable(CAP_SYS_ADMIN))
2326 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2330 set_schedstats(state);
2333 #endif /* CONFIG_PROC_SYSCTL */
2334 #else /* !CONFIG_SCHEDSTATS */
2335 static inline void init_schedstats(void) {}
2336 #endif /* CONFIG_SCHEDSTATS */
2339 * fork()/clone()-time setup:
2341 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2343 unsigned long flags;
2345 __sched_fork(clone_flags, p);
2347 * We mark the process as NEW here. This guarantees that
2348 * nobody will actually run it, and a signal or other external
2349 * event cannot wake it up and insert it on the runqueue either.
2351 p->state = TASK_NEW;
2354 * Make sure we do not leak PI boosting priority to the child.
2356 p->prio = current->normal_prio;
2359 * Revert to default priority/policy on fork if requested.
2361 if (unlikely(p->sched_reset_on_fork)) {
2362 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2363 p->policy = SCHED_NORMAL;
2364 p->static_prio = NICE_TO_PRIO(0);
2366 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2367 p->static_prio = NICE_TO_PRIO(0);
2369 p->prio = p->normal_prio = __normal_prio(p);
2370 set_load_weight(p, false);
2373 * We don't need the reset flag anymore after the fork. It has
2374 * fulfilled its duty:
2376 p->sched_reset_on_fork = 0;
2379 if (dl_prio(p->prio))
2381 else if (rt_prio(p->prio))
2382 p->sched_class = &rt_sched_class;
2384 p->sched_class = &fair_sched_class;
2386 init_entity_runnable_average(&p->se);
2389 * The child is not yet in the pid-hash so no cgroup attach races,
2390 * and the cgroup is pinned to this child due to cgroup_fork()
2391 * is ran before sched_fork().
2393 * Silence PROVE_RCU.
2395 raw_spin_lock_irqsave(&p->pi_lock, flags);
2397 * We're setting the CPU for the first time, we don't migrate,
2398 * so use __set_task_cpu().
2400 __set_task_cpu(p, smp_processor_id());
2401 if (p->sched_class->task_fork)
2402 p->sched_class->task_fork(p);
2403 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2405 #ifdef CONFIG_SCHED_INFO
2406 if (likely(sched_info_on()))
2407 memset(&p->sched_info, 0, sizeof(p->sched_info));
2409 #if defined(CONFIG_SMP)
2412 init_task_preempt_count(p);
2414 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2415 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2420 unsigned long to_ratio(u64 period, u64 runtime)
2422 if (runtime == RUNTIME_INF)
2426 * Doing this here saves a lot of checks in all
2427 * the calling paths, and returning zero seems
2428 * safe for them anyway.
2433 return div64_u64(runtime << BW_SHIFT, period);
2437 * wake_up_new_task - wake up a newly created task for the first time.
2439 * This function will do some initial scheduler statistics housekeeping
2440 * that must be done for every newly created context, then puts the task
2441 * on the runqueue and wakes it.
2443 void wake_up_new_task(struct task_struct *p)
2448 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2449 p->state = TASK_RUNNING;
2452 * Fork balancing, do it here and not earlier because:
2453 * - cpus_allowed can change in the fork path
2454 * - any previously selected CPU might disappear through hotplug
2456 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2457 * as we're not fully set-up yet.
2459 p->recent_used_cpu = task_cpu(p);
2460 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2462 rq = __task_rq_lock(p, &rf);
2463 update_rq_clock(rq);
2464 post_init_entity_util_avg(p);
2466 activate_task(rq, p, ENQUEUE_NOCLOCK);
2467 p->on_rq = TASK_ON_RQ_QUEUED;
2468 trace_sched_wakeup_new(p);
2469 check_preempt_curr(rq, p, WF_FORK);
2471 if (p->sched_class->task_woken) {
2473 * Nothing relies on rq->lock after this, so its fine to
2476 rq_unpin_lock(rq, &rf);
2477 p->sched_class->task_woken(rq, p);
2478 rq_repin_lock(rq, &rf);
2481 task_rq_unlock(rq, p, &rf);
2484 #ifdef CONFIG_PREEMPT_NOTIFIERS
2486 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2488 void preempt_notifier_inc(void)
2490 static_branch_inc(&preempt_notifier_key);
2492 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2494 void preempt_notifier_dec(void)
2496 static_branch_dec(&preempt_notifier_key);
2498 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2501 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2502 * @notifier: notifier struct to register
2504 void preempt_notifier_register(struct preempt_notifier *notifier)
2506 if (!static_branch_unlikely(&preempt_notifier_key))
2507 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2509 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2511 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2514 * preempt_notifier_unregister - no longer interested in preemption notifications
2515 * @notifier: notifier struct to unregister
2517 * This is *not* safe to call from within a preemption notifier.
2519 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2521 hlist_del(¬ifier->link);
2523 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2525 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2527 struct preempt_notifier *notifier;
2529 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2530 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2533 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2535 if (static_branch_unlikely(&preempt_notifier_key))
2536 __fire_sched_in_preempt_notifiers(curr);
2540 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2541 struct task_struct *next)
2543 struct preempt_notifier *notifier;
2545 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2546 notifier->ops->sched_out(notifier, next);
2549 static __always_inline void
2550 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2551 struct task_struct *next)
2553 if (static_branch_unlikely(&preempt_notifier_key))
2554 __fire_sched_out_preempt_notifiers(curr, next);
2557 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2559 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2564 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2565 struct task_struct *next)
2569 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2571 static inline void prepare_task(struct task_struct *next)
2575 * Claim the task as running, we do this before switching to it
2576 * such that any running task will have this set.
2582 static inline void finish_task(struct task_struct *prev)
2586 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2587 * We must ensure this doesn't happen until the switch is completely
2590 * In particular, the load of prev->state in finish_task_switch() must
2591 * happen before this.
2593 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2595 smp_store_release(&prev->on_cpu, 0);
2600 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
2603 * Since the runqueue lock will be released by the next
2604 * task (which is an invalid locking op but in the case
2605 * of the scheduler it's an obvious special-case), so we
2606 * do an early lockdep release here:
2608 rq_unpin_lock(rq, rf);
2609 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2610 #ifdef CONFIG_DEBUG_SPINLOCK
2611 /* this is a valid case when another task releases the spinlock */
2612 rq->lock.owner = next;
2616 static inline void finish_lock_switch(struct rq *rq)
2619 * If we are tracking spinlock dependencies then we have to
2620 * fix up the runqueue lock - which gets 'carried over' from
2621 * prev into current:
2623 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
2624 raw_spin_unlock_irq(&rq->lock);
2628 * NOP if the arch has not defined these:
2631 #ifndef prepare_arch_switch
2632 # define prepare_arch_switch(next) do { } while (0)
2635 #ifndef finish_arch_post_lock_switch
2636 # define finish_arch_post_lock_switch() do { } while (0)
2640 * prepare_task_switch - prepare to switch tasks
2641 * @rq: the runqueue preparing to switch
2642 * @prev: the current task that is being switched out
2643 * @next: the task we are going to switch to.
2645 * This is called with the rq lock held and interrupts off. It must
2646 * be paired with a subsequent finish_task_switch after the context
2649 * prepare_task_switch sets up locking and calls architecture specific
2653 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2654 struct task_struct *next)
2656 kcov_prepare_switch(prev);
2657 sched_info_switch(rq, prev, next);
2658 perf_event_task_sched_out(prev, next);
2660 fire_sched_out_preempt_notifiers(prev, next);
2662 prepare_arch_switch(next);
2666 * finish_task_switch - clean up after a task-switch
2667 * @prev: the thread we just switched away from.
2669 * finish_task_switch must be called after the context switch, paired
2670 * with a prepare_task_switch call before the context switch.
2671 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2672 * and do any other architecture-specific cleanup actions.
2674 * Note that we may have delayed dropping an mm in context_switch(). If
2675 * so, we finish that here outside of the runqueue lock. (Doing it
2676 * with the lock held can cause deadlocks; see schedule() for
2679 * The context switch have flipped the stack from under us and restored the
2680 * local variables which were saved when this task called schedule() in the
2681 * past. prev == current is still correct but we need to recalculate this_rq
2682 * because prev may have moved to another CPU.
2684 static struct rq *finish_task_switch(struct task_struct *prev)
2685 __releases(rq->lock)
2687 struct rq *rq = this_rq();
2688 struct mm_struct *mm = rq->prev_mm;
2692 * The previous task will have left us with a preempt_count of 2
2693 * because it left us after:
2696 * preempt_disable(); // 1
2698 * raw_spin_lock_irq(&rq->lock) // 2
2700 * Also, see FORK_PREEMPT_COUNT.
2702 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2703 "corrupted preempt_count: %s/%d/0x%x\n",
2704 current->comm, current->pid, preempt_count()))
2705 preempt_count_set(FORK_PREEMPT_COUNT);
2710 * A task struct has one reference for the use as "current".
2711 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2712 * schedule one last time. The schedule call will never return, and
2713 * the scheduled task must drop that reference.
2715 * We must observe prev->state before clearing prev->on_cpu (in
2716 * finish_task), otherwise a concurrent wakeup can get prev
2717 * running on another CPU and we could rave with its RUNNING -> DEAD
2718 * transition, resulting in a double drop.
2720 prev_state = prev->state;
2721 vtime_task_switch(prev);
2722 perf_event_task_sched_in(prev, current);
2724 finish_lock_switch(rq);
2725 finish_arch_post_lock_switch();
2726 kcov_finish_switch(current);
2728 fire_sched_in_preempt_notifiers(current);
2730 * When switching through a kernel thread, the loop in
2731 * membarrier_{private,global}_expedited() may have observed that
2732 * kernel thread and not issued an IPI. It is therefore possible to
2733 * schedule between user->kernel->user threads without passing though
2734 * switch_mm(). Membarrier requires a barrier after storing to
2735 * rq->curr, before returning to userspace, so provide them here:
2737 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
2738 * provided by mmdrop(),
2739 * - a sync_core for SYNC_CORE.
2742 membarrier_mm_sync_core_before_usermode(mm);
2745 if (unlikely(prev_state == TASK_DEAD)) {
2746 if (prev->sched_class->task_dead)
2747 prev->sched_class->task_dead(prev);
2750 * Remove function-return probe instances associated with this
2751 * task and put them back on the free list.
2753 kprobe_flush_task(prev);
2755 /* Task is done with its stack. */
2756 put_task_stack(prev);
2758 put_task_struct(prev);
2761 tick_nohz_task_switch();
2767 /* rq->lock is NOT held, but preemption is disabled */
2768 static void __balance_callback(struct rq *rq)
2770 struct callback_head *head, *next;
2771 void (*func)(struct rq *rq);
2772 unsigned long flags;
2774 raw_spin_lock_irqsave(&rq->lock, flags);
2775 head = rq->balance_callback;
2776 rq->balance_callback = NULL;
2778 func = (void (*)(struct rq *))head->func;
2785 raw_spin_unlock_irqrestore(&rq->lock, flags);
2788 static inline void balance_callback(struct rq *rq)
2790 if (unlikely(rq->balance_callback))
2791 __balance_callback(rq);
2796 static inline void balance_callback(struct rq *rq)
2803 * schedule_tail - first thing a freshly forked thread must call.
2804 * @prev: the thread we just switched away from.
2806 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2807 __releases(rq->lock)
2812 * New tasks start with FORK_PREEMPT_COUNT, see there and
2813 * finish_task_switch() for details.
2815 * finish_task_switch() will drop rq->lock() and lower preempt_count
2816 * and the preempt_enable() will end up enabling preemption (on
2817 * PREEMPT_COUNT kernels).
2820 rq = finish_task_switch(prev);
2821 balance_callback(rq);
2824 if (current->set_child_tid)
2825 put_user(task_pid_vnr(current), current->set_child_tid);
2827 calculate_sigpending();
2831 * context_switch - switch to the new MM and the new thread's register state.
2833 static __always_inline struct rq *
2834 context_switch(struct rq *rq, struct task_struct *prev,
2835 struct task_struct *next, struct rq_flags *rf)
2837 struct mm_struct *mm, *oldmm;
2839 prepare_task_switch(rq, prev, next);
2842 oldmm = prev->active_mm;
2844 * For paravirt, this is coupled with an exit in switch_to to
2845 * combine the page table reload and the switch backend into
2848 arch_start_context_switch(prev);
2851 * If mm is non-NULL, we pass through switch_mm(). If mm is
2852 * NULL, we will pass through mmdrop() in finish_task_switch().
2853 * Both of these contain the full memory barrier required by
2854 * membarrier after storing to rq->curr, before returning to
2858 next->active_mm = oldmm;
2860 enter_lazy_tlb(oldmm, next);
2862 switch_mm_irqs_off(oldmm, mm, next);
2865 prev->active_mm = NULL;
2866 rq->prev_mm = oldmm;
2869 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2871 prepare_lock_switch(rq, next, rf);
2873 /* Here we just switch the register state and the stack. */
2874 switch_to(prev, next, prev);
2877 return finish_task_switch(prev);
2881 * nr_running and nr_context_switches:
2883 * externally visible scheduler statistics: current number of runnable
2884 * threads, total number of context switches performed since bootup.
2886 unsigned long nr_running(void)
2888 unsigned long i, sum = 0;
2890 for_each_online_cpu(i)
2891 sum += cpu_rq(i)->nr_running;
2897 * Check if only the current task is running on the CPU.
2899 * Caution: this function does not check that the caller has disabled
2900 * preemption, thus the result might have a time-of-check-to-time-of-use
2901 * race. The caller is responsible to use it correctly, for example:
2903 * - from a non-preemptible section (of course)
2905 * - from a thread that is bound to a single CPU
2907 * - in a loop with very short iterations (e.g. a polling loop)
2909 bool single_task_running(void)
2911 return raw_rq()->nr_running == 1;
2913 EXPORT_SYMBOL(single_task_running);
2915 unsigned long long nr_context_switches(void)
2918 unsigned long long sum = 0;
2920 for_each_possible_cpu(i)
2921 sum += cpu_rq(i)->nr_switches;
2927 * Consumers of these two interfaces, like for example the cpuidle menu
2928 * governor, are using nonsensical data. Preferring shallow idle state selection
2929 * for a CPU that has IO-wait which might not even end up running the task when
2930 * it does become runnable.
2933 unsigned long nr_iowait_cpu(int cpu)
2935 return atomic_read(&cpu_rq(cpu)->nr_iowait);
2939 * IO-wait accounting, and how its mostly bollocks (on SMP).
2941 * The idea behind IO-wait account is to account the idle time that we could
2942 * have spend running if it were not for IO. That is, if we were to improve the
2943 * storage performance, we'd have a proportional reduction in IO-wait time.
2945 * This all works nicely on UP, where, when a task blocks on IO, we account
2946 * idle time as IO-wait, because if the storage were faster, it could've been
2947 * running and we'd not be idle.
2949 * This has been extended to SMP, by doing the same for each CPU. This however
2952 * Imagine for instance the case where two tasks block on one CPU, only the one
2953 * CPU will have IO-wait accounted, while the other has regular idle. Even
2954 * though, if the storage were faster, both could've ran at the same time,
2955 * utilising both CPUs.
2957 * This means, that when looking globally, the current IO-wait accounting on
2958 * SMP is a lower bound, by reason of under accounting.
2960 * Worse, since the numbers are provided per CPU, they are sometimes
2961 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2962 * associated with any one particular CPU, it can wake to another CPU than it
2963 * blocked on. This means the per CPU IO-wait number is meaningless.
2965 * Task CPU affinities can make all that even more 'interesting'.
2968 unsigned long nr_iowait(void)
2970 unsigned long i, sum = 0;
2972 for_each_possible_cpu(i)
2973 sum += nr_iowait_cpu(i);
2981 * sched_exec - execve() is a valuable balancing opportunity, because at
2982 * this point the task has the smallest effective memory and cache footprint.
2984 void sched_exec(void)
2986 struct task_struct *p = current;
2987 unsigned long flags;
2990 raw_spin_lock_irqsave(&p->pi_lock, flags);
2991 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2992 if (dest_cpu == smp_processor_id())
2995 if (likely(cpu_active(dest_cpu))) {
2996 struct migration_arg arg = { p, dest_cpu };
2998 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2999 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3003 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3008 DEFINE_PER_CPU(struct kernel_stat, kstat);
3009 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3011 EXPORT_PER_CPU_SYMBOL(kstat);
3012 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3015 * The function fair_sched_class.update_curr accesses the struct curr
3016 * and its field curr->exec_start; when called from task_sched_runtime(),
3017 * we observe a high rate of cache misses in practice.
3018 * Prefetching this data results in improved performance.
3020 static inline void prefetch_curr_exec_start(struct task_struct *p)
3022 #ifdef CONFIG_FAIR_GROUP_SCHED
3023 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3025 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3028 prefetch(&curr->exec_start);
3032 * Return accounted runtime for the task.
3033 * In case the task is currently running, return the runtime plus current's
3034 * pending runtime that have not been accounted yet.
3036 unsigned long long task_sched_runtime(struct task_struct *p)
3042 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3044 * 64-bit doesn't need locks to atomically read a 64-bit value.
3045 * So we have a optimization chance when the task's delta_exec is 0.
3046 * Reading ->on_cpu is racy, but this is ok.
3048 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3049 * If we race with it entering CPU, unaccounted time is 0. This is
3050 * indistinguishable from the read occurring a few cycles earlier.
3051 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3052 * been accounted, so we're correct here as well.
3054 if (!p->on_cpu || !task_on_rq_queued(p))
3055 return p->se.sum_exec_runtime;
3058 rq = task_rq_lock(p, &rf);
3060 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3061 * project cycles that may never be accounted to this
3062 * thread, breaking clock_gettime().
3064 if (task_current(rq, p) && task_on_rq_queued(p)) {
3065 prefetch_curr_exec_start(p);
3066 update_rq_clock(rq);
3067 p->sched_class->update_curr(rq);
3069 ns = p->se.sum_exec_runtime;
3070 task_rq_unlock(rq, p, &rf);
3076 * This function gets called by the timer code, with HZ frequency.
3077 * We call it with interrupts disabled.
3079 void scheduler_tick(void)
3081 int cpu = smp_processor_id();
3082 struct rq *rq = cpu_rq(cpu);
3083 struct task_struct *curr = rq->curr;
3090 update_rq_clock(rq);
3091 curr->sched_class->task_tick(rq, curr, 0);
3092 cpu_load_update_active(rq);
3093 calc_global_load_tick(rq);
3098 perf_event_task_tick();
3101 rq->idle_balance = idle_cpu(cpu);
3102 trigger_load_balance(rq);
3106 #ifdef CONFIG_NO_HZ_FULL
3110 struct delayed_work work;
3113 static struct tick_work __percpu *tick_work_cpu;
3115 static void sched_tick_remote(struct work_struct *work)
3117 struct delayed_work *dwork = to_delayed_work(work);
3118 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3119 int cpu = twork->cpu;
3120 struct rq *rq = cpu_rq(cpu);
3121 struct task_struct *curr;
3126 * Handle the tick only if it appears the remote CPU is running in full
3127 * dynticks mode. The check is racy by nature, but missing a tick or
3128 * having one too much is no big deal because the scheduler tick updates
3129 * statistics and checks timeslices in a time-independent way, regardless
3130 * of when exactly it is running.
3132 if (idle_cpu(cpu) || !tick_nohz_tick_stopped_cpu(cpu))
3135 rq_lock_irq(rq, &rf);
3137 if (is_idle_task(curr))
3140 update_rq_clock(rq);
3141 delta = rq_clock_task(rq) - curr->se.exec_start;
3144 * Make sure the next tick runs within a reasonable
3147 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3148 curr->sched_class->task_tick(rq, curr, 0);
3151 rq_unlock_irq(rq, &rf);
3155 * Run the remote tick once per second (1Hz). This arbitrary
3156 * frequency is large enough to avoid overload but short enough
3157 * to keep scheduler internal stats reasonably up to date.
3159 queue_delayed_work(system_unbound_wq, dwork, HZ);
3162 static void sched_tick_start(int cpu)
3164 struct tick_work *twork;
3166 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3169 WARN_ON_ONCE(!tick_work_cpu);
3171 twork = per_cpu_ptr(tick_work_cpu, cpu);
3173 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3174 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3177 #ifdef CONFIG_HOTPLUG_CPU
3178 static void sched_tick_stop(int cpu)
3180 struct tick_work *twork;
3182 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3185 WARN_ON_ONCE(!tick_work_cpu);
3187 twork = per_cpu_ptr(tick_work_cpu, cpu);
3188 cancel_delayed_work_sync(&twork->work);
3190 #endif /* CONFIG_HOTPLUG_CPU */
3192 int __init sched_tick_offload_init(void)
3194 tick_work_cpu = alloc_percpu(struct tick_work);
3195 BUG_ON(!tick_work_cpu);
3200 #else /* !CONFIG_NO_HZ_FULL */
3201 static inline void sched_tick_start(int cpu) { }
3202 static inline void sched_tick_stop(int cpu) { }
3205 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3206 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
3208 * If the value passed in is equal to the current preempt count
3209 * then we just disabled preemption. Start timing the latency.
3211 static inline void preempt_latency_start(int val)
3213 if (preempt_count() == val) {
3214 unsigned long ip = get_lock_parent_ip();
3215 #ifdef CONFIG_DEBUG_PREEMPT
3216 current->preempt_disable_ip = ip;
3218 trace_preempt_off(CALLER_ADDR0, ip);
3222 void preempt_count_add(int val)
3224 #ifdef CONFIG_DEBUG_PREEMPT
3228 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3231 __preempt_count_add(val);
3232 #ifdef CONFIG_DEBUG_PREEMPT
3234 * Spinlock count overflowing soon?
3236 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3239 preempt_latency_start(val);
3241 EXPORT_SYMBOL(preempt_count_add);
3242 NOKPROBE_SYMBOL(preempt_count_add);
3245 * If the value passed in equals to the current preempt count
3246 * then we just enabled preemption. Stop timing the latency.
3248 static inline void preempt_latency_stop(int val)
3250 if (preempt_count() == val)
3251 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3254 void preempt_count_sub(int val)
3256 #ifdef CONFIG_DEBUG_PREEMPT
3260 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3263 * Is the spinlock portion underflowing?
3265 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3266 !(preempt_count() & PREEMPT_MASK)))
3270 preempt_latency_stop(val);
3271 __preempt_count_sub(val);
3273 EXPORT_SYMBOL(preempt_count_sub);
3274 NOKPROBE_SYMBOL(preempt_count_sub);
3277 static inline void preempt_latency_start(int val) { }
3278 static inline void preempt_latency_stop(int val) { }
3281 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3283 #ifdef CONFIG_DEBUG_PREEMPT
3284 return p->preempt_disable_ip;
3291 * Print scheduling while atomic bug:
3293 static noinline void __schedule_bug(struct task_struct *prev)
3295 /* Save this before calling printk(), since that will clobber it */
3296 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3298 if (oops_in_progress)
3301 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3302 prev->comm, prev->pid, preempt_count());
3304 debug_show_held_locks(prev);
3306 if (irqs_disabled())
3307 print_irqtrace_events(prev);
3308 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3309 && in_atomic_preempt_off()) {
3310 pr_err("Preemption disabled at:");
3311 print_ip_sym(preempt_disable_ip);
3315 panic("scheduling while atomic\n");
3318 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3322 * Various schedule()-time debugging checks and statistics:
3324 static inline void schedule_debug(struct task_struct *prev)
3326 #ifdef CONFIG_SCHED_STACK_END_CHECK
3327 if (task_stack_end_corrupted(prev))
3328 panic("corrupted stack end detected inside scheduler\n");
3331 if (unlikely(in_atomic_preempt_off())) {
3332 __schedule_bug(prev);
3333 preempt_count_set(PREEMPT_DISABLED);
3337 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3339 schedstat_inc(this_rq()->sched_count);
3343 * Pick up the highest-prio task:
3345 static inline struct task_struct *
3346 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3348 const struct sched_class *class;
3349 struct task_struct *p;
3352 * Optimization: we know that if all tasks are in the fair class we can
3353 * call that function directly, but only if the @prev task wasn't of a
3354 * higher scheduling class, because otherwise those loose the
3355 * opportunity to pull in more work from other CPUs.
3357 if (likely((prev->sched_class == &idle_sched_class ||
3358 prev->sched_class == &fair_sched_class) &&
3359 rq->nr_running == rq->cfs.h_nr_running)) {
3361 p = fair_sched_class.pick_next_task(rq, prev, rf);
3362 if (unlikely(p == RETRY_TASK))
3365 /* Assumes fair_sched_class->next == idle_sched_class */
3367 p = idle_sched_class.pick_next_task(rq, prev, rf);
3373 for_each_class(class) {
3374 p = class->pick_next_task(rq, prev, rf);
3376 if (unlikely(p == RETRY_TASK))
3382 /* The idle class should always have a runnable task: */
3387 * __schedule() is the main scheduler function.
3389 * The main means of driving the scheduler and thus entering this function are:
3391 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3393 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3394 * paths. For example, see arch/x86/entry_64.S.
3396 * To drive preemption between tasks, the scheduler sets the flag in timer
3397 * interrupt handler scheduler_tick().
3399 * 3. Wakeups don't really cause entry into schedule(). They add a
3400 * task to the run-queue and that's it.
3402 * Now, if the new task added to the run-queue preempts the current
3403 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3404 * called on the nearest possible occasion:
3406 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3408 * - in syscall or exception context, at the next outmost
3409 * preempt_enable(). (this might be as soon as the wake_up()'s
3412 * - in IRQ context, return from interrupt-handler to
3413 * preemptible context
3415 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3418 * - cond_resched() call
3419 * - explicit schedule() call
3420 * - return from syscall or exception to user-space
3421 * - return from interrupt-handler to user-space
3423 * WARNING: must be called with preemption disabled!
3425 static void __sched notrace __schedule(bool preempt)
3427 struct task_struct *prev, *next;
3428 unsigned long *switch_count;
3433 cpu = smp_processor_id();
3437 schedule_debug(prev);
3439 if (sched_feat(HRTICK))
3442 local_irq_disable();
3443 rcu_note_context_switch(preempt);
3446 * Make sure that signal_pending_state()->signal_pending() below
3447 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3448 * done by the caller to avoid the race with signal_wake_up().
3450 * The membarrier system call requires a full memory barrier
3451 * after coming from user-space, before storing to rq->curr.
3454 smp_mb__after_spinlock();
3456 /* Promote REQ to ACT */
3457 rq->clock_update_flags <<= 1;
3458 update_rq_clock(rq);
3460 switch_count = &prev->nivcsw;
3461 if (!preempt && prev->state) {
3462 if (signal_pending_state(prev->state, prev)) {
3463 prev->state = TASK_RUNNING;
3465 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3468 if (prev->in_iowait) {
3469 atomic_inc(&rq->nr_iowait);
3470 delayacct_blkio_start();
3474 * If a worker went to sleep, notify and ask workqueue
3475 * whether it wants to wake up a task to maintain
3478 if (prev->flags & PF_WQ_WORKER) {
3479 struct task_struct *to_wakeup;
3481 to_wakeup = wq_worker_sleeping(prev);
3483 try_to_wake_up_local(to_wakeup, &rf);
3486 switch_count = &prev->nvcsw;
3489 next = pick_next_task(rq, prev, &rf);
3490 clear_tsk_need_resched(prev);
3491 clear_preempt_need_resched();
3493 if (likely(prev != next)) {
3497 * The membarrier system call requires each architecture
3498 * to have a full memory barrier after updating
3499 * rq->curr, before returning to user-space.
3501 * Here are the schemes providing that barrier on the
3502 * various architectures:
3503 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3504 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3505 * - finish_lock_switch() for weakly-ordered
3506 * architectures where spin_unlock is a full barrier,
3507 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3508 * is a RELEASE barrier),
3512 trace_sched_switch(preempt, prev, next);
3514 /* Also unlocks the rq: */
3515 rq = context_switch(rq, prev, next, &rf);
3517 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3518 rq_unlock_irq(rq, &rf);
3521 balance_callback(rq);
3524 void __noreturn do_task_dead(void)
3526 /* Causes final put_task_struct in finish_task_switch(): */
3527 set_special_state(TASK_DEAD);
3529 /* Tell freezer to ignore us: */
3530 current->flags |= PF_NOFREEZE;
3535 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3540 static inline void sched_submit_work(struct task_struct *tsk)
3542 if (!tsk->state || tsk_is_pi_blocked(tsk))
3545 * If we are going to sleep and we have plugged IO queued,
3546 * make sure to submit it to avoid deadlocks.
3548 if (blk_needs_flush_plug(tsk))
3549 blk_schedule_flush_plug(tsk);
3552 asmlinkage __visible void __sched schedule(void)
3554 struct task_struct *tsk = current;
3556 sched_submit_work(tsk);
3560 sched_preempt_enable_no_resched();
3561 } while (need_resched());
3563 EXPORT_SYMBOL(schedule);
3566 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3567 * state (have scheduled out non-voluntarily) by making sure that all
3568 * tasks have either left the run queue or have gone into user space.
3569 * As idle tasks do not do either, they must not ever be preempted
3570 * (schedule out non-voluntarily).
3572 * schedule_idle() is similar to schedule_preempt_disable() except that it
3573 * never enables preemption because it does not call sched_submit_work().
3575 void __sched schedule_idle(void)
3578 * As this skips calling sched_submit_work(), which the idle task does
3579 * regardless because that function is a nop when the task is in a
3580 * TASK_RUNNING state, make sure this isn't used someplace that the
3581 * current task can be in any other state. Note, idle is always in the
3582 * TASK_RUNNING state.
3584 WARN_ON_ONCE(current->state);
3587 } while (need_resched());
3590 #ifdef CONFIG_CONTEXT_TRACKING
3591 asmlinkage __visible void __sched schedule_user(void)
3594 * If we come here after a random call to set_need_resched(),
3595 * or we have been woken up remotely but the IPI has not yet arrived,
3596 * we haven't yet exited the RCU idle mode. Do it here manually until
3597 * we find a better solution.
3599 * NB: There are buggy callers of this function. Ideally we
3600 * should warn if prev_state != CONTEXT_USER, but that will trigger
3601 * too frequently to make sense yet.
3603 enum ctx_state prev_state = exception_enter();
3605 exception_exit(prev_state);
3610 * schedule_preempt_disabled - called with preemption disabled
3612 * Returns with preemption disabled. Note: preempt_count must be 1
3614 void __sched schedule_preempt_disabled(void)
3616 sched_preempt_enable_no_resched();
3621 static void __sched notrace preempt_schedule_common(void)
3625 * Because the function tracer can trace preempt_count_sub()
3626 * and it also uses preempt_enable/disable_notrace(), if
3627 * NEED_RESCHED is set, the preempt_enable_notrace() called
3628 * by the function tracer will call this function again and
3629 * cause infinite recursion.
3631 * Preemption must be disabled here before the function
3632 * tracer can trace. Break up preempt_disable() into two
3633 * calls. One to disable preemption without fear of being
3634 * traced. The other to still record the preemption latency,
3635 * which can also be traced by the function tracer.
3637 preempt_disable_notrace();
3638 preempt_latency_start(1);
3640 preempt_latency_stop(1);
3641 preempt_enable_no_resched_notrace();
3644 * Check again in case we missed a preemption opportunity
3645 * between schedule and now.
3647 } while (need_resched());
3650 #ifdef CONFIG_PREEMPT
3652 * this is the entry point to schedule() from in-kernel preemption
3653 * off of preempt_enable. Kernel preemptions off return from interrupt
3654 * occur there and call schedule directly.
3656 asmlinkage __visible void __sched notrace preempt_schedule(void)
3659 * If there is a non-zero preempt_count or interrupts are disabled,
3660 * we do not want to preempt the current task. Just return..
3662 if (likely(!preemptible()))
3665 preempt_schedule_common();
3667 NOKPROBE_SYMBOL(preempt_schedule);
3668 EXPORT_SYMBOL(preempt_schedule);
3671 * preempt_schedule_notrace - preempt_schedule called by tracing
3673 * The tracing infrastructure uses preempt_enable_notrace to prevent
3674 * recursion and tracing preempt enabling caused by the tracing
3675 * infrastructure itself. But as tracing can happen in areas coming
3676 * from userspace or just about to enter userspace, a preempt enable
3677 * can occur before user_exit() is called. This will cause the scheduler
3678 * to be called when the system is still in usermode.
3680 * To prevent this, the preempt_enable_notrace will use this function
3681 * instead of preempt_schedule() to exit user context if needed before
3682 * calling the scheduler.
3684 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3686 enum ctx_state prev_ctx;
3688 if (likely(!preemptible()))
3693 * Because the function tracer can trace preempt_count_sub()
3694 * and it also uses preempt_enable/disable_notrace(), if
3695 * NEED_RESCHED is set, the preempt_enable_notrace() called
3696 * by the function tracer will call this function again and
3697 * cause infinite recursion.
3699 * Preemption must be disabled here before the function
3700 * tracer can trace. Break up preempt_disable() into two
3701 * calls. One to disable preemption without fear of being
3702 * traced. The other to still record the preemption latency,
3703 * which can also be traced by the function tracer.
3705 preempt_disable_notrace();
3706 preempt_latency_start(1);
3708 * Needs preempt disabled in case user_exit() is traced
3709 * and the tracer calls preempt_enable_notrace() causing
3710 * an infinite recursion.
3712 prev_ctx = exception_enter();
3714 exception_exit(prev_ctx);
3716 preempt_latency_stop(1);
3717 preempt_enable_no_resched_notrace();
3718 } while (need_resched());
3720 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3722 #endif /* CONFIG_PREEMPT */
3725 * this is the entry point to schedule() from kernel preemption
3726 * off of irq context.
3727 * Note, that this is called and return with irqs disabled. This will
3728 * protect us against recursive calling from irq.
3730 asmlinkage __visible void __sched preempt_schedule_irq(void)
3732 enum ctx_state prev_state;
3734 /* Catch callers which need to be fixed */
3735 BUG_ON(preempt_count() || !irqs_disabled());
3737 prev_state = exception_enter();
3743 local_irq_disable();
3744 sched_preempt_enable_no_resched();
3745 } while (need_resched());
3747 exception_exit(prev_state);
3750 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
3753 return try_to_wake_up(curr->private, mode, wake_flags);
3755 EXPORT_SYMBOL(default_wake_function);
3757 #ifdef CONFIG_RT_MUTEXES
3759 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
3762 prio = min(prio, pi_task->prio);
3767 static inline int rt_effective_prio(struct task_struct *p, int prio)
3769 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3771 return __rt_effective_prio(pi_task, prio);
3775 * rt_mutex_setprio - set the current priority of a task
3777 * @pi_task: donor task
3779 * This function changes the 'effective' priority of a task. It does
3780 * not touch ->normal_prio like __setscheduler().
3782 * Used by the rt_mutex code to implement priority inheritance
3783 * logic. Call site only calls if the priority of the task changed.
3785 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
3787 int prio, oldprio, queued, running, queue_flag =
3788 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
3789 const struct sched_class *prev_class;
3793 /* XXX used to be waiter->prio, not waiter->task->prio */
3794 prio = __rt_effective_prio(pi_task, p->normal_prio);
3797 * If nothing changed; bail early.
3799 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
3802 rq = __task_rq_lock(p, &rf);
3803 update_rq_clock(rq);
3805 * Set under pi_lock && rq->lock, such that the value can be used under
3808 * Note that there is loads of tricky to make this pointer cache work
3809 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3810 * ensure a task is de-boosted (pi_task is set to NULL) before the
3811 * task is allowed to run again (and can exit). This ensures the pointer
3812 * points to a blocked task -- which guaratees the task is present.
3814 p->pi_top_task = pi_task;
3817 * For FIFO/RR we only need to set prio, if that matches we're done.
3819 if (prio == p->prio && !dl_prio(prio))
3823 * Idle task boosting is a nono in general. There is one
3824 * exception, when PREEMPT_RT and NOHZ is active:
3826 * The idle task calls get_next_timer_interrupt() and holds
3827 * the timer wheel base->lock on the CPU and another CPU wants
3828 * to access the timer (probably to cancel it). We can safely
3829 * ignore the boosting request, as the idle CPU runs this code
3830 * with interrupts disabled and will complete the lock
3831 * protected section without being interrupted. So there is no
3832 * real need to boost.
3834 if (unlikely(p == rq->idle)) {
3835 WARN_ON(p != rq->curr);
3836 WARN_ON(p->pi_blocked_on);
3840 trace_sched_pi_setprio(p, pi_task);
3843 if (oldprio == prio)
3844 queue_flag &= ~DEQUEUE_MOVE;
3846 prev_class = p->sched_class;
3847 queued = task_on_rq_queued(p);
3848 running = task_current(rq, p);
3850 dequeue_task(rq, p, queue_flag);
3852 put_prev_task(rq, p);
3855 * Boosting condition are:
3856 * 1. -rt task is running and holds mutex A
3857 * --> -dl task blocks on mutex A
3859 * 2. -dl task is running and holds mutex A
3860 * --> -dl task blocks on mutex A and could preempt the
3863 if (dl_prio(prio)) {
3864 if (!dl_prio(p->normal_prio) ||
3865 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3866 p->dl.dl_boosted = 1;
3867 queue_flag |= ENQUEUE_REPLENISH;
3869 p->dl.dl_boosted = 0;
3870 p->sched_class = &dl_sched_class;
3871 } else if (rt_prio(prio)) {
3872 if (dl_prio(oldprio))
3873 p->dl.dl_boosted = 0;
3875 queue_flag |= ENQUEUE_HEAD;
3876 p->sched_class = &rt_sched_class;
3878 if (dl_prio(oldprio))
3879 p->dl.dl_boosted = 0;
3880 if (rt_prio(oldprio))
3882 p->sched_class = &fair_sched_class;
3888 enqueue_task(rq, p, queue_flag);
3890 set_curr_task(rq, p);
3892 check_class_changed(rq, p, prev_class, oldprio);
3894 /* Avoid rq from going away on us: */
3896 __task_rq_unlock(rq, &rf);
3898 balance_callback(rq);
3902 static inline int rt_effective_prio(struct task_struct *p, int prio)
3908 void set_user_nice(struct task_struct *p, long nice)
3910 bool queued, running;
3911 int old_prio, delta;
3915 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3918 * We have to be careful, if called from sys_setpriority(),
3919 * the task might be in the middle of scheduling on another CPU.
3921 rq = task_rq_lock(p, &rf);
3922 update_rq_clock(rq);
3925 * The RT priorities are set via sched_setscheduler(), but we still
3926 * allow the 'normal' nice value to be set - but as expected
3927 * it wont have any effect on scheduling until the task is
3928 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3930 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3931 p->static_prio = NICE_TO_PRIO(nice);
3934 queued = task_on_rq_queued(p);
3935 running = task_current(rq, p);
3937 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
3939 put_prev_task(rq, p);
3941 p->static_prio = NICE_TO_PRIO(nice);
3942 set_load_weight(p, true);
3944 p->prio = effective_prio(p);
3945 delta = p->prio - old_prio;
3948 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
3950 * If the task increased its priority or is running and
3951 * lowered its priority, then reschedule its CPU:
3953 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3957 set_curr_task(rq, p);
3959 task_rq_unlock(rq, p, &rf);
3961 EXPORT_SYMBOL(set_user_nice);
3964 * can_nice - check if a task can reduce its nice value
3968 int can_nice(const struct task_struct *p, const int nice)
3970 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3971 int nice_rlim = nice_to_rlimit(nice);
3973 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3974 capable(CAP_SYS_NICE));
3977 #ifdef __ARCH_WANT_SYS_NICE
3980 * sys_nice - change the priority of the current process.
3981 * @increment: priority increment
3983 * sys_setpriority is a more generic, but much slower function that
3984 * does similar things.
3986 SYSCALL_DEFINE1(nice, int, increment)
3991 * Setpriority might change our priority at the same moment.
3992 * We don't have to worry. Conceptually one call occurs first
3993 * and we have a single winner.
3995 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3996 nice = task_nice(current) + increment;
3998 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3999 if (increment < 0 && !can_nice(current, nice))
4002 retval = security_task_setnice(current, nice);
4006 set_user_nice(current, nice);
4013 * task_prio - return the priority value of a given task.
4014 * @p: the task in question.
4016 * Return: The priority value as seen by users in /proc.
4017 * RT tasks are offset by -200. Normal tasks are centered
4018 * around 0, value goes from -16 to +15.
4020 int task_prio(const struct task_struct *p)
4022 return p->prio - MAX_RT_PRIO;
4026 * idle_cpu - is a given CPU idle currently?
4027 * @cpu: the processor in question.
4029 * Return: 1 if the CPU is currently idle. 0 otherwise.
4031 int idle_cpu(int cpu)
4033 struct rq *rq = cpu_rq(cpu);
4035 if (rq->curr != rq->idle)
4042 if (!llist_empty(&rq->wake_list))
4050 * available_idle_cpu - is a given CPU idle for enqueuing work.
4051 * @cpu: the CPU in question.
4053 * Return: 1 if the CPU is currently idle. 0 otherwise.
4055 int available_idle_cpu(int cpu)
4060 if (vcpu_is_preempted(cpu))
4067 * idle_task - return the idle task for a given CPU.
4068 * @cpu: the processor in question.
4070 * Return: The idle task for the CPU @cpu.
4072 struct task_struct *idle_task(int cpu)
4074 return cpu_rq(cpu)->idle;
4078 * find_process_by_pid - find a process with a matching PID value.
4079 * @pid: the pid in question.
4081 * The task of @pid, if found. %NULL otherwise.
4083 static struct task_struct *find_process_by_pid(pid_t pid)
4085 return pid ? find_task_by_vpid(pid) : current;
4089 * sched_setparam() passes in -1 for its policy, to let the functions
4090 * it calls know not to change it.
4092 #define SETPARAM_POLICY -1
4094 static void __setscheduler_params(struct task_struct *p,
4095 const struct sched_attr *attr)
4097 int policy = attr->sched_policy;
4099 if (policy == SETPARAM_POLICY)
4104 if (dl_policy(policy))
4105 __setparam_dl(p, attr);
4106 else if (fair_policy(policy))
4107 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4110 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4111 * !rt_policy. Always setting this ensures that things like
4112 * getparam()/getattr() don't report silly values for !rt tasks.
4114 p->rt_priority = attr->sched_priority;
4115 p->normal_prio = normal_prio(p);
4116 set_load_weight(p, true);
4119 /* Actually do priority change: must hold pi & rq lock. */
4120 static void __setscheduler(struct rq *rq, struct task_struct *p,
4121 const struct sched_attr *attr, bool keep_boost)
4123 __setscheduler_params(p, attr);
4126 * Keep a potential priority boosting if called from
4127 * sched_setscheduler().
4129 p->prio = normal_prio(p);
4131 p->prio = rt_effective_prio(p, p->prio);
4133 if (dl_prio(p->prio))
4134 p->sched_class = &dl_sched_class;
4135 else if (rt_prio(p->prio))
4136 p->sched_class = &rt_sched_class;
4138 p->sched_class = &fair_sched_class;
4142 * Check the target process has a UID that matches the current process's:
4144 static bool check_same_owner(struct task_struct *p)
4146 const struct cred *cred = current_cred(), *pcred;
4150 pcred = __task_cred(p);
4151 match = (uid_eq(cred->euid, pcred->euid) ||
4152 uid_eq(cred->euid, pcred->uid));
4157 static int __sched_setscheduler(struct task_struct *p,
4158 const struct sched_attr *attr,
4161 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4162 MAX_RT_PRIO - 1 - attr->sched_priority;
4163 int retval, oldprio, oldpolicy = -1, queued, running;
4164 int new_effective_prio, policy = attr->sched_policy;
4165 const struct sched_class *prev_class;
4168 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4171 /* The pi code expects interrupts enabled */
4172 BUG_ON(pi && in_interrupt());
4174 /* Double check policy once rq lock held: */
4176 reset_on_fork = p->sched_reset_on_fork;
4177 policy = oldpolicy = p->policy;
4179 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4181 if (!valid_policy(policy))
4185 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4189 * Valid priorities for SCHED_FIFO and SCHED_RR are
4190 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4191 * SCHED_BATCH and SCHED_IDLE is 0.
4193 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4194 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4196 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4197 (rt_policy(policy) != (attr->sched_priority != 0)))
4201 * Allow unprivileged RT tasks to decrease priority:
4203 if (user && !capable(CAP_SYS_NICE)) {
4204 if (fair_policy(policy)) {
4205 if (attr->sched_nice < task_nice(p) &&
4206 !can_nice(p, attr->sched_nice))
4210 if (rt_policy(policy)) {
4211 unsigned long rlim_rtprio =
4212 task_rlimit(p, RLIMIT_RTPRIO);
4214 /* Can't set/change the rt policy: */
4215 if (policy != p->policy && !rlim_rtprio)
4218 /* Can't increase priority: */
4219 if (attr->sched_priority > p->rt_priority &&
4220 attr->sched_priority > rlim_rtprio)
4225 * Can't set/change SCHED_DEADLINE policy at all for now
4226 * (safest behavior); in the future we would like to allow
4227 * unprivileged DL tasks to increase their relative deadline
4228 * or reduce their runtime (both ways reducing utilization)
4230 if (dl_policy(policy))
4234 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4235 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4237 if (task_has_idle_policy(p) && !idle_policy(policy)) {
4238 if (!can_nice(p, task_nice(p)))
4242 /* Can't change other user's priorities: */
4243 if (!check_same_owner(p))
4246 /* Normal users shall not reset the sched_reset_on_fork flag: */
4247 if (p->sched_reset_on_fork && !reset_on_fork)
4252 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4255 retval = security_task_setscheduler(p);
4261 * Make sure no PI-waiters arrive (or leave) while we are
4262 * changing the priority of the task:
4264 * To be able to change p->policy safely, the appropriate
4265 * runqueue lock must be held.
4267 rq = task_rq_lock(p, &rf);
4268 update_rq_clock(rq);
4271 * Changing the policy of the stop threads its a very bad idea:
4273 if (p == rq->stop) {
4274 task_rq_unlock(rq, p, &rf);
4279 * If not changing anything there's no need to proceed further,
4280 * but store a possible modification of reset_on_fork.
4282 if (unlikely(policy == p->policy)) {
4283 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4285 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4287 if (dl_policy(policy) && dl_param_changed(p, attr))
4290 p->sched_reset_on_fork = reset_on_fork;
4291 task_rq_unlock(rq, p, &rf);
4297 #ifdef CONFIG_RT_GROUP_SCHED
4299 * Do not allow realtime tasks into groups that have no runtime
4302 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4303 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4304 !task_group_is_autogroup(task_group(p))) {
4305 task_rq_unlock(rq, p, &rf);
4310 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4311 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4312 cpumask_t *span = rq->rd->span;
4315 * Don't allow tasks with an affinity mask smaller than
4316 * the entire root_domain to become SCHED_DEADLINE. We
4317 * will also fail if there's no bandwidth available.
4319 if (!cpumask_subset(span, &p->cpus_allowed) ||
4320 rq->rd->dl_bw.bw == 0) {
4321 task_rq_unlock(rq, p, &rf);
4328 /* Re-check policy now with rq lock held: */
4329 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4330 policy = oldpolicy = -1;
4331 task_rq_unlock(rq, p, &rf);
4336 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4337 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4340 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4341 task_rq_unlock(rq, p, &rf);
4345 p->sched_reset_on_fork = reset_on_fork;
4350 * Take priority boosted tasks into account. If the new
4351 * effective priority is unchanged, we just store the new
4352 * normal parameters and do not touch the scheduler class and
4353 * the runqueue. This will be done when the task deboost
4356 new_effective_prio = rt_effective_prio(p, newprio);
4357 if (new_effective_prio == oldprio)
4358 queue_flags &= ~DEQUEUE_MOVE;
4361 queued = task_on_rq_queued(p);
4362 running = task_current(rq, p);
4364 dequeue_task(rq, p, queue_flags);
4366 put_prev_task(rq, p);
4368 prev_class = p->sched_class;
4369 __setscheduler(rq, p, attr, pi);
4373 * We enqueue to tail when the priority of a task is
4374 * increased (user space view).
4376 if (oldprio < p->prio)
4377 queue_flags |= ENQUEUE_HEAD;
4379 enqueue_task(rq, p, queue_flags);
4382 set_curr_task(rq, p);
4384 check_class_changed(rq, p, prev_class, oldprio);
4386 /* Avoid rq from going away on us: */
4388 task_rq_unlock(rq, p, &rf);
4391 rt_mutex_adjust_pi(p);
4393 /* Run balance callbacks after we've adjusted the PI chain: */
4394 balance_callback(rq);
4400 static int _sched_setscheduler(struct task_struct *p, int policy,
4401 const struct sched_param *param, bool check)
4403 struct sched_attr attr = {
4404 .sched_policy = policy,
4405 .sched_priority = param->sched_priority,
4406 .sched_nice = PRIO_TO_NICE(p->static_prio),
4409 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4410 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4411 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4412 policy &= ~SCHED_RESET_ON_FORK;
4413 attr.sched_policy = policy;
4416 return __sched_setscheduler(p, &attr, check, true);
4419 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4420 * @p: the task in question.
4421 * @policy: new policy.
4422 * @param: structure containing the new RT priority.
4424 * Return: 0 on success. An error code otherwise.
4426 * NOTE that the task may be already dead.
4428 int sched_setscheduler(struct task_struct *p, int policy,
4429 const struct sched_param *param)
4431 return _sched_setscheduler(p, policy, param, true);
4433 EXPORT_SYMBOL_GPL(sched_setscheduler);
4435 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4437 return __sched_setscheduler(p, attr, true, true);
4439 EXPORT_SYMBOL_GPL(sched_setattr);
4441 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
4443 return __sched_setscheduler(p, attr, false, true);
4447 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4448 * @p: the task in question.
4449 * @policy: new policy.
4450 * @param: structure containing the new RT priority.
4452 * Just like sched_setscheduler, only don't bother checking if the
4453 * current context has permission. For example, this is needed in
4454 * stop_machine(): we create temporary high priority worker threads,
4455 * but our caller might not have that capability.
4457 * Return: 0 on success. An error code otherwise.
4459 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4460 const struct sched_param *param)
4462 return _sched_setscheduler(p, policy, param, false);
4464 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4467 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4469 struct sched_param lparam;
4470 struct task_struct *p;
4473 if (!param || pid < 0)
4475 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4480 p = find_process_by_pid(pid);
4482 retval = sched_setscheduler(p, policy, &lparam);
4489 * Mimics kernel/events/core.c perf_copy_attr().
4491 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4496 if (!access_ok(uattr, SCHED_ATTR_SIZE_VER0))
4499 /* Zero the full structure, so that a short copy will be nice: */
4500 memset(attr, 0, sizeof(*attr));
4502 ret = get_user(size, &uattr->size);
4506 /* Bail out on silly large: */
4507 if (size > PAGE_SIZE)
4510 /* ABI compatibility quirk: */
4512 size = SCHED_ATTR_SIZE_VER0;
4514 if (size < SCHED_ATTR_SIZE_VER0)
4518 * If we're handed a bigger struct than we know of,
4519 * ensure all the unknown bits are 0 - i.e. new
4520 * user-space does not rely on any kernel feature
4521 * extensions we dont know about yet.
4523 if (size > sizeof(*attr)) {
4524 unsigned char __user *addr;
4525 unsigned char __user *end;
4528 addr = (void __user *)uattr + sizeof(*attr);
4529 end = (void __user *)uattr + size;
4531 for (; addr < end; addr++) {
4532 ret = get_user(val, addr);
4538 size = sizeof(*attr);
4541 ret = copy_from_user(attr, uattr, size);
4546 * XXX: Do we want to be lenient like existing syscalls; or do we want
4547 * to be strict and return an error on out-of-bounds values?
4549 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4554 put_user(sizeof(*attr), &uattr->size);
4559 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4560 * @pid: the pid in question.
4561 * @policy: new policy.
4562 * @param: structure containing the new RT priority.
4564 * Return: 0 on success. An error code otherwise.
4566 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4571 return do_sched_setscheduler(pid, policy, param);
4575 * sys_sched_setparam - set/change the RT priority of a thread
4576 * @pid: the pid in question.
4577 * @param: structure containing the new RT priority.
4579 * Return: 0 on success. An error code otherwise.
4581 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4583 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4587 * sys_sched_setattr - same as above, but with extended sched_attr
4588 * @pid: the pid in question.
4589 * @uattr: structure containing the extended parameters.
4590 * @flags: for future extension.
4592 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4593 unsigned int, flags)
4595 struct sched_attr attr;
4596 struct task_struct *p;
4599 if (!uattr || pid < 0 || flags)
4602 retval = sched_copy_attr(uattr, &attr);
4606 if ((int)attr.sched_policy < 0)
4611 p = find_process_by_pid(pid);
4613 retval = sched_setattr(p, &attr);
4620 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4621 * @pid: the pid in question.
4623 * Return: On success, the policy of the thread. Otherwise, a negative error
4626 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4628 struct task_struct *p;
4636 p = find_process_by_pid(pid);
4638 retval = security_task_getscheduler(p);
4641 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4648 * sys_sched_getparam - get the RT priority of a thread
4649 * @pid: the pid in question.
4650 * @param: structure containing the RT priority.
4652 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4655 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4657 struct sched_param lp = { .sched_priority = 0 };
4658 struct task_struct *p;
4661 if (!param || pid < 0)
4665 p = find_process_by_pid(pid);
4670 retval = security_task_getscheduler(p);
4674 if (task_has_rt_policy(p))
4675 lp.sched_priority = p->rt_priority;
4679 * This one might sleep, we cannot do it with a spinlock held ...
4681 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4690 static int sched_read_attr(struct sched_attr __user *uattr,
4691 struct sched_attr *attr,
4696 if (!access_ok(uattr, usize))
4700 * If we're handed a smaller struct than we know of,
4701 * ensure all the unknown bits are 0 - i.e. old
4702 * user-space does not get uncomplete information.
4704 if (usize < sizeof(*attr)) {
4705 unsigned char *addr;
4708 addr = (void *)attr + usize;
4709 end = (void *)attr + sizeof(*attr);
4711 for (; addr < end; addr++) {
4719 ret = copy_to_user(uattr, attr, attr->size);
4727 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4728 * @pid: the pid in question.
4729 * @uattr: structure containing the extended parameters.
4730 * @size: sizeof(attr) for fwd/bwd comp.
4731 * @flags: for future extension.
4733 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4734 unsigned int, size, unsigned int, flags)
4736 struct sched_attr attr = {
4737 .size = sizeof(struct sched_attr),
4739 struct task_struct *p;
4742 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4743 size < SCHED_ATTR_SIZE_VER0 || flags)
4747 p = find_process_by_pid(pid);
4752 retval = security_task_getscheduler(p);
4756 attr.sched_policy = p->policy;
4757 if (p->sched_reset_on_fork)
4758 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4759 if (task_has_dl_policy(p))
4760 __getparam_dl(p, &attr);
4761 else if (task_has_rt_policy(p))
4762 attr.sched_priority = p->rt_priority;
4764 attr.sched_nice = task_nice(p);
4768 retval = sched_read_attr(uattr, &attr, size);
4776 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4778 cpumask_var_t cpus_allowed, new_mask;
4779 struct task_struct *p;
4784 p = find_process_by_pid(pid);
4790 /* Prevent p going away */
4794 if (p->flags & PF_NO_SETAFFINITY) {
4798 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4802 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4804 goto out_free_cpus_allowed;
4807 if (!check_same_owner(p)) {
4809 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4811 goto out_free_new_mask;
4816 retval = security_task_setscheduler(p);
4818 goto out_free_new_mask;
4821 cpuset_cpus_allowed(p, cpus_allowed);
4822 cpumask_and(new_mask, in_mask, cpus_allowed);
4825 * Since bandwidth control happens on root_domain basis,
4826 * if admission test is enabled, we only admit -deadline
4827 * tasks allowed to run on all the CPUs in the task's
4831 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4833 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4836 goto out_free_new_mask;
4842 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4845 cpuset_cpus_allowed(p, cpus_allowed);
4846 if (!cpumask_subset(new_mask, cpus_allowed)) {
4848 * We must have raced with a concurrent cpuset
4849 * update. Just reset the cpus_allowed to the
4850 * cpuset's cpus_allowed
4852 cpumask_copy(new_mask, cpus_allowed);
4857 free_cpumask_var(new_mask);
4858 out_free_cpus_allowed:
4859 free_cpumask_var(cpus_allowed);
4865 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4866 struct cpumask *new_mask)
4868 if (len < cpumask_size())
4869 cpumask_clear(new_mask);
4870 else if (len > cpumask_size())
4871 len = cpumask_size();
4873 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4877 * sys_sched_setaffinity - set the CPU affinity of a process
4878 * @pid: pid of the process
4879 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4880 * @user_mask_ptr: user-space pointer to the new CPU mask
4882 * Return: 0 on success. An error code otherwise.
4884 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4885 unsigned long __user *, user_mask_ptr)
4887 cpumask_var_t new_mask;
4890 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4893 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4895 retval = sched_setaffinity(pid, new_mask);
4896 free_cpumask_var(new_mask);
4900 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4902 struct task_struct *p;
4903 unsigned long flags;
4909 p = find_process_by_pid(pid);
4913 retval = security_task_getscheduler(p);
4917 raw_spin_lock_irqsave(&p->pi_lock, flags);
4918 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4919 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4928 * sys_sched_getaffinity - get the CPU affinity of a process
4929 * @pid: pid of the process
4930 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4931 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4933 * Return: size of CPU mask copied to user_mask_ptr on success. An
4934 * error code otherwise.
4936 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4937 unsigned long __user *, user_mask_ptr)
4942 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4944 if (len & (sizeof(unsigned long)-1))
4947 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4950 ret = sched_getaffinity(pid, mask);
4952 unsigned int retlen = min(len, cpumask_size());
4954 if (copy_to_user(user_mask_ptr, mask, retlen))
4959 free_cpumask_var(mask);
4965 * sys_sched_yield - yield the current processor to other threads.
4967 * This function yields the current CPU to other tasks. If there are no
4968 * other threads running on this CPU then this function will return.
4972 static void do_sched_yield(void)
4977 rq = this_rq_lock_irq(&rf);
4979 schedstat_inc(rq->yld_count);
4980 current->sched_class->yield_task(rq);
4983 * Since we are going to call schedule() anyway, there's
4984 * no need to preempt or enable interrupts:
4988 sched_preempt_enable_no_resched();
4993 SYSCALL_DEFINE0(sched_yield)
4999 #ifndef CONFIG_PREEMPT
5000 int __sched _cond_resched(void)
5002 if (should_resched(0)) {
5003 preempt_schedule_common();
5009 EXPORT_SYMBOL(_cond_resched);
5013 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5014 * call schedule, and on return reacquire the lock.
5016 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5017 * operations here to prevent schedule() from being called twice (once via
5018 * spin_unlock(), once by hand).
5020 int __cond_resched_lock(spinlock_t *lock)
5022 int resched = should_resched(PREEMPT_LOCK_OFFSET);
5025 lockdep_assert_held(lock);
5027 if (spin_needbreak(lock) || resched) {
5030 preempt_schedule_common();
5038 EXPORT_SYMBOL(__cond_resched_lock);
5041 * yield - yield the current processor to other threads.
5043 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5045 * The scheduler is at all times free to pick the calling task as the most
5046 * eligible task to run, if removing the yield() call from your code breaks
5047 * it, its already broken.
5049 * Typical broken usage is:
5054 * where one assumes that yield() will let 'the other' process run that will
5055 * make event true. If the current task is a SCHED_FIFO task that will never
5056 * happen. Never use yield() as a progress guarantee!!
5058 * If you want to use yield() to wait for something, use wait_event().
5059 * If you want to use yield() to be 'nice' for others, use cond_resched().
5060 * If you still want to use yield(), do not!
5062 void __sched yield(void)
5064 set_current_state(TASK_RUNNING);
5067 EXPORT_SYMBOL(yield);
5070 * yield_to - yield the current processor to another thread in
5071 * your thread group, or accelerate that thread toward the
5072 * processor it's on.
5074 * @preempt: whether task preemption is allowed or not
5076 * It's the caller's job to ensure that the target task struct
5077 * can't go away on us before we can do any checks.
5080 * true (>0) if we indeed boosted the target task.
5081 * false (0) if we failed to boost the target.
5082 * -ESRCH if there's no task to yield to.
5084 int __sched yield_to(struct task_struct *p, bool preempt)
5086 struct task_struct *curr = current;
5087 struct rq *rq, *p_rq;
5088 unsigned long flags;
5091 local_irq_save(flags);
5097 * If we're the only runnable task on the rq and target rq also
5098 * has only one task, there's absolutely no point in yielding.
5100 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5105 double_rq_lock(rq, p_rq);
5106 if (task_rq(p) != p_rq) {
5107 double_rq_unlock(rq, p_rq);
5111 if (!curr->sched_class->yield_to_task)
5114 if (curr->sched_class != p->sched_class)
5117 if (task_running(p_rq, p) || p->state)
5120 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5122 schedstat_inc(rq->yld_count);
5124 * Make p's CPU reschedule; pick_next_entity takes care of
5127 if (preempt && rq != p_rq)
5132 double_rq_unlock(rq, p_rq);
5134 local_irq_restore(flags);
5141 EXPORT_SYMBOL_GPL(yield_to);
5143 int io_schedule_prepare(void)
5145 int old_iowait = current->in_iowait;
5147 current->in_iowait = 1;
5148 blk_schedule_flush_plug(current);
5153 void io_schedule_finish(int token)
5155 current->in_iowait = token;
5159 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5160 * that process accounting knows that this is a task in IO wait state.
5162 long __sched io_schedule_timeout(long timeout)
5167 token = io_schedule_prepare();
5168 ret = schedule_timeout(timeout);
5169 io_schedule_finish(token);
5173 EXPORT_SYMBOL(io_schedule_timeout);
5175 void io_schedule(void)
5179 token = io_schedule_prepare();
5181 io_schedule_finish(token);
5183 EXPORT_SYMBOL(io_schedule);
5186 * sys_sched_get_priority_max - return maximum RT priority.
5187 * @policy: scheduling class.
5189 * Return: On success, this syscall returns the maximum
5190 * rt_priority that can be used by a given scheduling class.
5191 * On failure, a negative error code is returned.
5193 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5200 ret = MAX_USER_RT_PRIO-1;
5202 case SCHED_DEADLINE:
5213 * sys_sched_get_priority_min - return minimum RT priority.
5214 * @policy: scheduling class.
5216 * Return: On success, this syscall returns the minimum
5217 * rt_priority that can be used by a given scheduling class.
5218 * On failure, a negative error code is returned.
5220 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5229 case SCHED_DEADLINE:
5238 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5240 struct task_struct *p;
5241 unsigned int time_slice;
5251 p = find_process_by_pid(pid);
5255 retval = security_task_getscheduler(p);
5259 rq = task_rq_lock(p, &rf);
5261 if (p->sched_class->get_rr_interval)
5262 time_slice = p->sched_class->get_rr_interval(rq, p);
5263 task_rq_unlock(rq, p, &rf);
5266 jiffies_to_timespec64(time_slice, t);
5275 * sys_sched_rr_get_interval - return the default timeslice of a process.
5276 * @pid: pid of the process.
5277 * @interval: userspace pointer to the timeslice value.
5279 * this syscall writes the default timeslice value of a given process
5280 * into the user-space timespec buffer. A value of '0' means infinity.
5282 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5285 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5286 struct __kernel_timespec __user *, interval)
5288 struct timespec64 t;
5289 int retval = sched_rr_get_interval(pid, &t);
5292 retval = put_timespec64(&t, interval);
5297 #ifdef CONFIG_COMPAT_32BIT_TIME
5298 SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
5299 struct old_timespec32 __user *, interval)
5301 struct timespec64 t;
5302 int retval = sched_rr_get_interval(pid, &t);
5305 retval = put_old_timespec32(&t, interval);
5310 void sched_show_task(struct task_struct *p)
5312 unsigned long free = 0;
5315 if (!try_get_task_stack(p))
5318 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5320 if (p->state == TASK_RUNNING)
5321 printk(KERN_CONT " running task ");
5322 #ifdef CONFIG_DEBUG_STACK_USAGE
5323 free = stack_not_used(p);
5328 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5330 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5331 task_pid_nr(p), ppid,
5332 (unsigned long)task_thread_info(p)->flags);
5334 print_worker_info(KERN_INFO, p);
5335 show_stack(p, NULL);
5338 EXPORT_SYMBOL_GPL(sched_show_task);
5341 state_filter_match(unsigned long state_filter, struct task_struct *p)
5343 /* no filter, everything matches */
5347 /* filter, but doesn't match */
5348 if (!(p->state & state_filter))
5352 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5355 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5362 void show_state_filter(unsigned long state_filter)
5364 struct task_struct *g, *p;
5366 #if BITS_PER_LONG == 32
5368 " task PC stack pid father\n");
5371 " task PC stack pid father\n");
5374 for_each_process_thread(g, p) {
5376 * reset the NMI-timeout, listing all files on a slow
5377 * console might take a lot of time:
5378 * Also, reset softlockup watchdogs on all CPUs, because
5379 * another CPU might be blocked waiting for us to process
5382 touch_nmi_watchdog();
5383 touch_all_softlockup_watchdogs();
5384 if (state_filter_match(state_filter, p))
5388 #ifdef CONFIG_SCHED_DEBUG
5390 sysrq_sched_debug_show();
5394 * Only show locks if all tasks are dumped:
5397 debug_show_all_locks();
5401 * init_idle - set up an idle thread for a given CPU
5402 * @idle: task in question
5403 * @cpu: CPU the idle task belongs to
5405 * NOTE: this function does not set the idle thread's NEED_RESCHED
5406 * flag, to make booting more robust.
5408 void init_idle(struct task_struct *idle, int cpu)
5410 struct rq *rq = cpu_rq(cpu);
5411 unsigned long flags;
5413 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5414 raw_spin_lock(&rq->lock);
5416 __sched_fork(0, idle);
5417 idle->state = TASK_RUNNING;
5418 idle->se.exec_start = sched_clock();
5419 idle->flags |= PF_IDLE;
5421 kasan_unpoison_task_stack(idle);
5425 * Its possible that init_idle() gets called multiple times on a task,
5426 * in that case do_set_cpus_allowed() will not do the right thing.
5428 * And since this is boot we can forgo the serialization.
5430 set_cpus_allowed_common(idle, cpumask_of(cpu));
5433 * We're having a chicken and egg problem, even though we are
5434 * holding rq->lock, the CPU isn't yet set to this CPU so the
5435 * lockdep check in task_group() will fail.
5437 * Similar case to sched_fork(). / Alternatively we could
5438 * use task_rq_lock() here and obtain the other rq->lock.
5443 __set_task_cpu(idle, cpu);
5446 rq->curr = rq->idle = idle;
5447 idle->on_rq = TASK_ON_RQ_QUEUED;
5451 raw_spin_unlock(&rq->lock);
5452 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5454 /* Set the preempt count _outside_ the spinlocks! */
5455 init_idle_preempt_count(idle, cpu);
5458 * The idle tasks have their own, simple scheduling class:
5460 idle->sched_class = &idle_sched_class;
5461 ftrace_graph_init_idle_task(idle, cpu);
5462 vtime_init_idle(idle, cpu);
5464 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5470 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5471 const struct cpumask *trial)
5475 if (!cpumask_weight(cur))
5478 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5483 int task_can_attach(struct task_struct *p,
5484 const struct cpumask *cs_cpus_allowed)
5489 * Kthreads which disallow setaffinity shouldn't be moved
5490 * to a new cpuset; we don't want to change their CPU
5491 * affinity and isolating such threads by their set of
5492 * allowed nodes is unnecessary. Thus, cpusets are not
5493 * applicable for such threads. This prevents checking for
5494 * success of set_cpus_allowed_ptr() on all attached tasks
5495 * before cpus_allowed may be changed.
5497 if (p->flags & PF_NO_SETAFFINITY) {
5502 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5504 ret = dl_task_can_attach(p, cs_cpus_allowed);
5510 bool sched_smp_initialized __read_mostly;
5512 #ifdef CONFIG_NUMA_BALANCING
5513 /* Migrate current task p to target_cpu */
5514 int migrate_task_to(struct task_struct *p, int target_cpu)
5516 struct migration_arg arg = { p, target_cpu };
5517 int curr_cpu = task_cpu(p);
5519 if (curr_cpu == target_cpu)
5522 if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5525 /* TODO: This is not properly updating schedstats */
5527 trace_sched_move_numa(p, curr_cpu, target_cpu);
5528 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5532 * Requeue a task on a given node and accurately track the number of NUMA
5533 * tasks on the runqueues
5535 void sched_setnuma(struct task_struct *p, int nid)
5537 bool queued, running;
5541 rq = task_rq_lock(p, &rf);
5542 queued = task_on_rq_queued(p);
5543 running = task_current(rq, p);
5546 dequeue_task(rq, p, DEQUEUE_SAVE);
5548 put_prev_task(rq, p);
5550 p->numa_preferred_nid = nid;
5553 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5555 set_curr_task(rq, p);
5556 task_rq_unlock(rq, p, &rf);
5558 #endif /* CONFIG_NUMA_BALANCING */
5560 #ifdef CONFIG_HOTPLUG_CPU
5562 * Ensure that the idle task is using init_mm right before its CPU goes
5565 void idle_task_exit(void)
5567 struct mm_struct *mm = current->active_mm;
5569 BUG_ON(cpu_online(smp_processor_id()));
5571 if (mm != &init_mm) {
5572 switch_mm(mm, &init_mm, current);
5573 current->active_mm = &init_mm;
5574 finish_arch_post_lock_switch();
5580 * Since this CPU is going 'away' for a while, fold any nr_active delta
5581 * we might have. Assumes we're called after migrate_tasks() so that the
5582 * nr_active count is stable. We need to take the teardown thread which
5583 * is calling this into account, so we hand in adjust = 1 to the load
5586 * Also see the comment "Global load-average calculations".
5588 static void calc_load_migrate(struct rq *rq)
5590 long delta = calc_load_fold_active(rq, 1);
5592 atomic_long_add(delta, &calc_load_tasks);
5595 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5599 static const struct sched_class fake_sched_class = {
5600 .put_prev_task = put_prev_task_fake,
5603 static struct task_struct fake_task = {
5605 * Avoid pull_{rt,dl}_task()
5607 .prio = MAX_PRIO + 1,
5608 .sched_class = &fake_sched_class,
5612 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5613 * try_to_wake_up()->select_task_rq().
5615 * Called with rq->lock held even though we'er in stop_machine() and
5616 * there's no concurrency possible, we hold the required locks anyway
5617 * because of lock validation efforts.
5619 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5621 struct rq *rq = dead_rq;
5622 struct task_struct *next, *stop = rq->stop;
5623 struct rq_flags orf = *rf;
5627 * Fudge the rq selection such that the below task selection loop
5628 * doesn't get stuck on the currently eligible stop task.
5630 * We're currently inside stop_machine() and the rq is either stuck
5631 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5632 * either way we should never end up calling schedule() until we're
5638 * put_prev_task() and pick_next_task() sched
5639 * class method both need to have an up-to-date
5640 * value of rq->clock[_task]
5642 update_rq_clock(rq);
5646 * There's this thread running, bail when that's the only
5649 if (rq->nr_running == 1)
5653 * pick_next_task() assumes pinned rq->lock:
5655 next = pick_next_task(rq, &fake_task, rf);
5657 put_prev_task(rq, next);
5660 * Rules for changing task_struct::cpus_allowed are holding
5661 * both pi_lock and rq->lock, such that holding either
5662 * stabilizes the mask.
5664 * Drop rq->lock is not quite as disastrous as it usually is
5665 * because !cpu_active at this point, which means load-balance
5666 * will not interfere. Also, stop-machine.
5669 raw_spin_lock(&next->pi_lock);
5673 * Since we're inside stop-machine, _nothing_ should have
5674 * changed the task, WARN if weird stuff happened, because in
5675 * that case the above rq->lock drop is a fail too.
5677 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5678 raw_spin_unlock(&next->pi_lock);
5682 /* Find suitable destination for @next, with force if needed. */
5683 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5684 rq = __migrate_task(rq, rf, next, dest_cpu);
5685 if (rq != dead_rq) {
5691 raw_spin_unlock(&next->pi_lock);
5696 #endif /* CONFIG_HOTPLUG_CPU */
5698 void set_rq_online(struct rq *rq)
5701 const struct sched_class *class;
5703 cpumask_set_cpu(rq->cpu, rq->rd->online);
5706 for_each_class(class) {
5707 if (class->rq_online)
5708 class->rq_online(rq);
5713 void set_rq_offline(struct rq *rq)
5716 const struct sched_class *class;
5718 for_each_class(class) {
5719 if (class->rq_offline)
5720 class->rq_offline(rq);
5723 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5729 * used to mark begin/end of suspend/resume:
5731 static int num_cpus_frozen;
5734 * Update cpusets according to cpu_active mask. If cpusets are
5735 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5736 * around partition_sched_domains().
5738 * If we come here as part of a suspend/resume, don't touch cpusets because we
5739 * want to restore it back to its original state upon resume anyway.
5741 static void cpuset_cpu_active(void)
5743 if (cpuhp_tasks_frozen) {
5745 * num_cpus_frozen tracks how many CPUs are involved in suspend
5746 * resume sequence. As long as this is not the last online
5747 * operation in the resume sequence, just build a single sched
5748 * domain, ignoring cpusets.
5750 partition_sched_domains(1, NULL, NULL);
5751 if (--num_cpus_frozen)
5754 * This is the last CPU online operation. So fall through and
5755 * restore the original sched domains by considering the
5756 * cpuset configurations.
5758 cpuset_force_rebuild();
5760 cpuset_update_active_cpus();
5763 static int cpuset_cpu_inactive(unsigned int cpu)
5765 if (!cpuhp_tasks_frozen) {
5766 if (dl_cpu_busy(cpu))
5768 cpuset_update_active_cpus();
5771 partition_sched_domains(1, NULL, NULL);
5776 int sched_cpu_activate(unsigned int cpu)
5778 struct rq *rq = cpu_rq(cpu);
5781 #ifdef CONFIG_SCHED_SMT
5783 * When going up, increment the number of cores with SMT present.
5785 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
5786 static_branch_inc_cpuslocked(&sched_smt_present);
5788 set_cpu_active(cpu, true);
5790 if (sched_smp_initialized) {
5791 sched_domains_numa_masks_set(cpu);
5792 cpuset_cpu_active();
5796 * Put the rq online, if not already. This happens:
5798 * 1) In the early boot process, because we build the real domains
5799 * after all CPUs have been brought up.
5801 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5804 rq_lock_irqsave(rq, &rf);
5806 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5809 rq_unlock_irqrestore(rq, &rf);
5811 update_max_interval();
5816 int sched_cpu_deactivate(unsigned int cpu)
5820 set_cpu_active(cpu, false);
5822 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5823 * users of this state to go away such that all new such users will
5826 * Do sync before park smpboot threads to take care the rcu boost case.
5830 #ifdef CONFIG_SCHED_SMT
5832 * When going down, decrement the number of cores with SMT present.
5834 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
5835 static_branch_dec_cpuslocked(&sched_smt_present);
5838 if (!sched_smp_initialized)
5841 ret = cpuset_cpu_inactive(cpu);
5843 set_cpu_active(cpu, true);
5846 sched_domains_numa_masks_clear(cpu);
5850 static void sched_rq_cpu_starting(unsigned int cpu)
5852 struct rq *rq = cpu_rq(cpu);
5854 rq->calc_load_update = calc_load_update;
5855 update_max_interval();
5858 int sched_cpu_starting(unsigned int cpu)
5860 sched_rq_cpu_starting(cpu);
5861 sched_tick_start(cpu);
5865 #ifdef CONFIG_HOTPLUG_CPU
5866 int sched_cpu_dying(unsigned int cpu)
5868 struct rq *rq = cpu_rq(cpu);
5871 /* Handle pending wakeups and then migrate everything off */
5872 sched_ttwu_pending();
5873 sched_tick_stop(cpu);
5875 rq_lock_irqsave(rq, &rf);
5877 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5880 migrate_tasks(rq, &rf);
5881 BUG_ON(rq->nr_running != 1);
5882 rq_unlock_irqrestore(rq, &rf);
5884 calc_load_migrate(rq);
5885 update_max_interval();
5886 nohz_balance_exit_idle(rq);
5892 void __init sched_init_smp(void)
5897 * There's no userspace yet to cause hotplug operations; hence all the
5898 * CPU masks are stable and all blatant races in the below code cannot
5901 mutex_lock(&sched_domains_mutex);
5902 sched_init_domains(cpu_active_mask);
5903 mutex_unlock(&sched_domains_mutex);
5905 /* Move init over to a non-isolated CPU */
5906 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
5908 sched_init_granularity();
5910 init_sched_rt_class();
5911 init_sched_dl_class();
5913 sched_smp_initialized = true;
5916 static int __init migration_init(void)
5918 sched_rq_cpu_starting(smp_processor_id());
5921 early_initcall(migration_init);
5924 void __init sched_init_smp(void)
5926 sched_init_granularity();
5928 #endif /* CONFIG_SMP */
5930 int in_sched_functions(unsigned long addr)
5932 return in_lock_functions(addr) ||
5933 (addr >= (unsigned long)__sched_text_start
5934 && addr < (unsigned long)__sched_text_end);
5937 #ifdef CONFIG_CGROUP_SCHED
5939 * Default task group.
5940 * Every task in system belongs to this group at bootup.
5942 struct task_group root_task_group;
5943 LIST_HEAD(task_groups);
5945 /* Cacheline aligned slab cache for task_group */
5946 static struct kmem_cache *task_group_cache __read_mostly;
5949 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5950 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5952 void __init sched_init(void)
5955 unsigned long alloc_size = 0, ptr;
5959 #ifdef CONFIG_FAIR_GROUP_SCHED
5960 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5962 #ifdef CONFIG_RT_GROUP_SCHED
5963 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5966 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5968 #ifdef CONFIG_FAIR_GROUP_SCHED
5969 root_task_group.se = (struct sched_entity **)ptr;
5970 ptr += nr_cpu_ids * sizeof(void **);
5972 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5973 ptr += nr_cpu_ids * sizeof(void **);
5975 #endif /* CONFIG_FAIR_GROUP_SCHED */
5976 #ifdef CONFIG_RT_GROUP_SCHED
5977 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5978 ptr += nr_cpu_ids * sizeof(void **);
5980 root_task_group.rt_rq = (struct rt_rq **)ptr;
5981 ptr += nr_cpu_ids * sizeof(void **);
5983 #endif /* CONFIG_RT_GROUP_SCHED */
5985 #ifdef CONFIG_CPUMASK_OFFSTACK
5986 for_each_possible_cpu(i) {
5987 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5988 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5989 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
5990 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5992 #endif /* CONFIG_CPUMASK_OFFSTACK */
5994 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
5995 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
5998 init_defrootdomain();
6001 #ifdef CONFIG_RT_GROUP_SCHED
6002 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6003 global_rt_period(), global_rt_runtime());
6004 #endif /* CONFIG_RT_GROUP_SCHED */
6006 #ifdef CONFIG_CGROUP_SCHED
6007 task_group_cache = KMEM_CACHE(task_group, 0);
6009 list_add(&root_task_group.list, &task_groups);
6010 INIT_LIST_HEAD(&root_task_group.children);
6011 INIT_LIST_HEAD(&root_task_group.siblings);
6012 autogroup_init(&init_task);
6013 #endif /* CONFIG_CGROUP_SCHED */
6015 for_each_possible_cpu(i) {
6019 raw_spin_lock_init(&rq->lock);
6021 rq->calc_load_active = 0;
6022 rq->calc_load_update = jiffies + LOAD_FREQ;
6023 init_cfs_rq(&rq->cfs);
6024 init_rt_rq(&rq->rt);
6025 init_dl_rq(&rq->dl);
6026 #ifdef CONFIG_FAIR_GROUP_SCHED
6027 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6028 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6029 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6031 * How much CPU bandwidth does root_task_group get?
6033 * In case of task-groups formed thr' the cgroup filesystem, it
6034 * gets 100% of the CPU resources in the system. This overall
6035 * system CPU resource is divided among the tasks of
6036 * root_task_group and its child task-groups in a fair manner,
6037 * based on each entity's (task or task-group's) weight
6038 * (se->load.weight).
6040 * In other words, if root_task_group has 10 tasks of weight
6041 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6042 * then A0's share of the CPU resource is:
6044 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6046 * We achieve this by letting root_task_group's tasks sit
6047 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6049 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6050 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6051 #endif /* CONFIG_FAIR_GROUP_SCHED */
6053 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6054 #ifdef CONFIG_RT_GROUP_SCHED
6055 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6058 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6059 rq->cpu_load[j] = 0;
6064 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6065 rq->balance_callback = NULL;
6066 rq->active_balance = 0;
6067 rq->next_balance = jiffies;
6072 rq->avg_idle = 2*sysctl_sched_migration_cost;
6073 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6075 INIT_LIST_HEAD(&rq->cfs_tasks);
6077 rq_attach_root(rq, &def_root_domain);
6078 #ifdef CONFIG_NO_HZ_COMMON
6079 rq->last_load_update_tick = jiffies;
6080 rq->last_blocked_load_update_tick = jiffies;
6081 atomic_set(&rq->nohz_flags, 0);
6083 #endif /* CONFIG_SMP */
6085 atomic_set(&rq->nr_iowait, 0);
6088 set_load_weight(&init_task, false);
6091 * The boot idle thread does lazy MMU switching as well:
6094 enter_lazy_tlb(&init_mm, current);
6097 * Make us the idle thread. Technically, schedule() should not be
6098 * called from this thread, however somewhere below it might be,
6099 * but because we are the idle thread, we just pick up running again
6100 * when this runqueue becomes "idle".
6102 init_idle(current, smp_processor_id());
6104 calc_load_update = jiffies + LOAD_FREQ;
6107 idle_thread_set_boot_cpu();
6109 init_sched_fair_class();
6115 scheduler_running = 1;
6118 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6119 static inline int preempt_count_equals(int preempt_offset)
6121 int nested = preempt_count() + rcu_preempt_depth();
6123 return (nested == preempt_offset);
6126 void __might_sleep(const char *file, int line, int preempt_offset)
6129 * Blocking primitives will set (and therefore destroy) current->state,
6130 * since we will exit with TASK_RUNNING make sure we enter with it,
6131 * otherwise we will destroy state.
6133 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6134 "do not call blocking ops when !TASK_RUNNING; "
6135 "state=%lx set at [<%p>] %pS\n",
6137 (void *)current->task_state_change,
6138 (void *)current->task_state_change);
6140 ___might_sleep(file, line, preempt_offset);
6142 EXPORT_SYMBOL(__might_sleep);
6144 void ___might_sleep(const char *file, int line, int preempt_offset)
6146 /* Ratelimiting timestamp: */
6147 static unsigned long prev_jiffy;
6149 unsigned long preempt_disable_ip;
6151 /* WARN_ON_ONCE() by default, no rate limit required: */
6154 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6155 !is_idle_task(current)) ||
6156 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6160 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6162 prev_jiffy = jiffies;
6164 /* Save this before calling printk(), since that will clobber it: */
6165 preempt_disable_ip = get_preempt_disable_ip(current);
6168 "BUG: sleeping function called from invalid context at %s:%d\n",
6171 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6172 in_atomic(), irqs_disabled(),
6173 current->pid, current->comm);
6175 if (task_stack_end_corrupted(current))
6176 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6178 debug_show_held_locks(current);
6179 if (irqs_disabled())
6180 print_irqtrace_events(current);
6181 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6182 && !preempt_count_equals(preempt_offset)) {
6183 pr_err("Preemption disabled at:");
6184 print_ip_sym(preempt_disable_ip);
6188 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6190 EXPORT_SYMBOL(___might_sleep);
6192 void __cant_sleep(const char *file, int line, int preempt_offset)
6194 static unsigned long prev_jiffy;
6196 if (irqs_disabled())
6199 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
6202 if (preempt_count() > preempt_offset)
6205 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6207 prev_jiffy = jiffies;
6209 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
6210 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6211 in_atomic(), irqs_disabled(),
6212 current->pid, current->comm);
6214 debug_show_held_locks(current);
6216 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6218 EXPORT_SYMBOL_GPL(__cant_sleep);
6221 #ifdef CONFIG_MAGIC_SYSRQ
6222 void normalize_rt_tasks(void)
6224 struct task_struct *g, *p;
6225 struct sched_attr attr = {
6226 .sched_policy = SCHED_NORMAL,
6229 read_lock(&tasklist_lock);
6230 for_each_process_thread(g, p) {
6232 * Only normalize user tasks:
6234 if (p->flags & PF_KTHREAD)
6237 p->se.exec_start = 0;
6238 schedstat_set(p->se.statistics.wait_start, 0);
6239 schedstat_set(p->se.statistics.sleep_start, 0);
6240 schedstat_set(p->se.statistics.block_start, 0);
6242 if (!dl_task(p) && !rt_task(p)) {
6244 * Renice negative nice level userspace
6247 if (task_nice(p) < 0)
6248 set_user_nice(p, 0);
6252 __sched_setscheduler(p, &attr, false, false);
6254 read_unlock(&tasklist_lock);
6257 #endif /* CONFIG_MAGIC_SYSRQ */
6259 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6261 * These functions are only useful for the IA64 MCA handling, or kdb.
6263 * They can only be called when the whole system has been
6264 * stopped - every CPU needs to be quiescent, and no scheduling
6265 * activity can take place. Using them for anything else would
6266 * be a serious bug, and as a result, they aren't even visible
6267 * under any other configuration.
6271 * curr_task - return the current task for a given CPU.
6272 * @cpu: the processor in question.
6274 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6276 * Return: The current task for @cpu.
6278 struct task_struct *curr_task(int cpu)
6280 return cpu_curr(cpu);
6283 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6287 * set_curr_task - set the current task for a given CPU.
6288 * @cpu: the processor in question.
6289 * @p: the task pointer to set.
6291 * Description: This function must only be used when non-maskable interrupts
6292 * are serviced on a separate stack. It allows the architecture to switch the
6293 * notion of the current task on a CPU in a non-blocking manner. This function
6294 * must be called with all CPU's synchronized, and interrupts disabled, the
6295 * and caller must save the original value of the current task (see
6296 * curr_task() above) and restore that value before reenabling interrupts and
6297 * re-starting the system.
6299 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6301 void ia64_set_curr_task(int cpu, struct task_struct *p)
6308 #ifdef CONFIG_CGROUP_SCHED
6309 /* task_group_lock serializes the addition/removal of task groups */
6310 static DEFINE_SPINLOCK(task_group_lock);
6312 static void sched_free_group(struct task_group *tg)
6314 free_fair_sched_group(tg);
6315 free_rt_sched_group(tg);
6317 kmem_cache_free(task_group_cache, tg);
6320 /* allocate runqueue etc for a new task group */
6321 struct task_group *sched_create_group(struct task_group *parent)
6323 struct task_group *tg;
6325 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6327 return ERR_PTR(-ENOMEM);
6329 if (!alloc_fair_sched_group(tg, parent))
6332 if (!alloc_rt_sched_group(tg, parent))
6338 sched_free_group(tg);
6339 return ERR_PTR(-ENOMEM);
6342 void sched_online_group(struct task_group *tg, struct task_group *parent)
6344 unsigned long flags;
6346 spin_lock_irqsave(&task_group_lock, flags);
6347 list_add_rcu(&tg->list, &task_groups);
6349 /* Root should already exist: */
6352 tg->parent = parent;
6353 INIT_LIST_HEAD(&tg->children);
6354 list_add_rcu(&tg->siblings, &parent->children);
6355 spin_unlock_irqrestore(&task_group_lock, flags);
6357 online_fair_sched_group(tg);
6360 /* rcu callback to free various structures associated with a task group */
6361 static void sched_free_group_rcu(struct rcu_head *rhp)
6363 /* Now it should be safe to free those cfs_rqs: */
6364 sched_free_group(container_of(rhp, struct task_group, rcu));
6367 void sched_destroy_group(struct task_group *tg)
6369 /* Wait for possible concurrent references to cfs_rqs complete: */
6370 call_rcu(&tg->rcu, sched_free_group_rcu);
6373 void sched_offline_group(struct task_group *tg)
6375 unsigned long flags;
6377 /* End participation in shares distribution: */
6378 unregister_fair_sched_group(tg);
6380 spin_lock_irqsave(&task_group_lock, flags);
6381 list_del_rcu(&tg->list);
6382 list_del_rcu(&tg->siblings);
6383 spin_unlock_irqrestore(&task_group_lock, flags);
6386 static void sched_change_group(struct task_struct *tsk, int type)
6388 struct task_group *tg;
6391 * All callers are synchronized by task_rq_lock(); we do not use RCU
6392 * which is pointless here. Thus, we pass "true" to task_css_check()
6393 * to prevent lockdep warnings.
6395 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6396 struct task_group, css);
6397 tg = autogroup_task_group(tsk, tg);
6398 tsk->sched_task_group = tg;
6400 #ifdef CONFIG_FAIR_GROUP_SCHED
6401 if (tsk->sched_class->task_change_group)
6402 tsk->sched_class->task_change_group(tsk, type);
6405 set_task_rq(tsk, task_cpu(tsk));
6409 * Change task's runqueue when it moves between groups.
6411 * The caller of this function should have put the task in its new group by
6412 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6415 void sched_move_task(struct task_struct *tsk)
6417 int queued, running, queue_flags =
6418 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6422 rq = task_rq_lock(tsk, &rf);
6423 update_rq_clock(rq);
6425 running = task_current(rq, tsk);
6426 queued = task_on_rq_queued(tsk);
6429 dequeue_task(rq, tsk, queue_flags);
6431 put_prev_task(rq, tsk);
6433 sched_change_group(tsk, TASK_MOVE_GROUP);
6436 enqueue_task(rq, tsk, queue_flags);
6438 set_curr_task(rq, tsk);
6440 task_rq_unlock(rq, tsk, &rf);
6443 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6445 return css ? container_of(css, struct task_group, css) : NULL;
6448 static struct cgroup_subsys_state *
6449 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6451 struct task_group *parent = css_tg(parent_css);
6452 struct task_group *tg;
6455 /* This is early initialization for the top cgroup */
6456 return &root_task_group.css;
6459 tg = sched_create_group(parent);
6461 return ERR_PTR(-ENOMEM);
6466 /* Expose task group only after completing cgroup initialization */
6467 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6469 struct task_group *tg = css_tg(css);
6470 struct task_group *parent = css_tg(css->parent);
6473 sched_online_group(tg, parent);
6477 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6479 struct task_group *tg = css_tg(css);
6481 sched_offline_group(tg);
6484 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6486 struct task_group *tg = css_tg(css);
6489 * Relies on the RCU grace period between css_released() and this.
6491 sched_free_group(tg);
6495 * This is called before wake_up_new_task(), therefore we really only
6496 * have to set its group bits, all the other stuff does not apply.
6498 static void cpu_cgroup_fork(struct task_struct *task)
6503 rq = task_rq_lock(task, &rf);
6505 update_rq_clock(rq);
6506 sched_change_group(task, TASK_SET_GROUP);
6508 task_rq_unlock(rq, task, &rf);
6511 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6513 struct task_struct *task;
6514 struct cgroup_subsys_state *css;
6517 cgroup_taskset_for_each(task, css, tset) {
6518 #ifdef CONFIG_RT_GROUP_SCHED
6519 if (!sched_rt_can_attach(css_tg(css), task))
6522 /* We don't support RT-tasks being in separate groups */
6523 if (task->sched_class != &fair_sched_class)
6527 * Serialize against wake_up_new_task() such that if its
6528 * running, we're sure to observe its full state.
6530 raw_spin_lock_irq(&task->pi_lock);
6532 * Avoid calling sched_move_task() before wake_up_new_task()
6533 * has happened. This would lead to problems with PELT, due to
6534 * move wanting to detach+attach while we're not attached yet.
6536 if (task->state == TASK_NEW)
6538 raw_spin_unlock_irq(&task->pi_lock);
6546 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6548 struct task_struct *task;
6549 struct cgroup_subsys_state *css;
6551 cgroup_taskset_for_each(task, css, tset)
6552 sched_move_task(task);
6555 #ifdef CONFIG_FAIR_GROUP_SCHED
6556 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6557 struct cftype *cftype, u64 shareval)
6559 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6562 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6565 struct task_group *tg = css_tg(css);
6567 return (u64) scale_load_down(tg->shares);
6570 #ifdef CONFIG_CFS_BANDWIDTH
6571 static DEFINE_MUTEX(cfs_constraints_mutex);
6573 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6574 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6576 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6578 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6580 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6581 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6583 if (tg == &root_task_group)
6587 * Ensure we have at some amount of bandwidth every period. This is
6588 * to prevent reaching a state of large arrears when throttled via
6589 * entity_tick() resulting in prolonged exit starvation.
6591 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6595 * Likewise, bound things on the otherside by preventing insane quota
6596 * periods. This also allows us to normalize in computing quota
6599 if (period > max_cfs_quota_period)
6603 * Prevent race between setting of cfs_rq->runtime_enabled and
6604 * unthrottle_offline_cfs_rqs().
6607 mutex_lock(&cfs_constraints_mutex);
6608 ret = __cfs_schedulable(tg, period, quota);
6612 runtime_enabled = quota != RUNTIME_INF;
6613 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6615 * If we need to toggle cfs_bandwidth_used, off->on must occur
6616 * before making related changes, and on->off must occur afterwards
6618 if (runtime_enabled && !runtime_was_enabled)
6619 cfs_bandwidth_usage_inc();
6620 raw_spin_lock_irq(&cfs_b->lock);
6621 cfs_b->period = ns_to_ktime(period);
6622 cfs_b->quota = quota;
6624 __refill_cfs_bandwidth_runtime(cfs_b);
6626 /* Restart the period timer (if active) to handle new period expiry: */
6627 if (runtime_enabled)
6628 start_cfs_bandwidth(cfs_b);
6630 raw_spin_unlock_irq(&cfs_b->lock);
6632 for_each_online_cpu(i) {
6633 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6634 struct rq *rq = cfs_rq->rq;
6637 rq_lock_irq(rq, &rf);
6638 cfs_rq->runtime_enabled = runtime_enabled;
6639 cfs_rq->runtime_remaining = 0;
6641 if (cfs_rq->throttled)
6642 unthrottle_cfs_rq(cfs_rq);
6643 rq_unlock_irq(rq, &rf);
6645 if (runtime_was_enabled && !runtime_enabled)
6646 cfs_bandwidth_usage_dec();
6648 mutex_unlock(&cfs_constraints_mutex);
6654 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
6658 period = ktime_to_ns(tg->cfs_bandwidth.period);
6659 if (cfs_quota_us < 0)
6660 quota = RUNTIME_INF;
6662 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
6664 return tg_set_cfs_bandwidth(tg, period, quota);
6667 long tg_get_cfs_quota(struct task_group *tg)
6671 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
6674 quota_us = tg->cfs_bandwidth.quota;
6675 do_div(quota_us, NSEC_PER_USEC);
6680 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
6684 period = (u64)cfs_period_us * NSEC_PER_USEC;
6685 quota = tg->cfs_bandwidth.quota;
6687 return tg_set_cfs_bandwidth(tg, period, quota);
6690 long tg_get_cfs_period(struct task_group *tg)
6694 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
6695 do_div(cfs_period_us, NSEC_PER_USEC);
6697 return cfs_period_us;
6700 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
6703 return tg_get_cfs_quota(css_tg(css));
6706 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
6707 struct cftype *cftype, s64 cfs_quota_us)
6709 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
6712 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
6715 return tg_get_cfs_period(css_tg(css));
6718 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
6719 struct cftype *cftype, u64 cfs_period_us)
6721 return tg_set_cfs_period(css_tg(css), cfs_period_us);
6724 struct cfs_schedulable_data {
6725 struct task_group *tg;
6730 * normalize group quota/period to be quota/max_period
6731 * note: units are usecs
6733 static u64 normalize_cfs_quota(struct task_group *tg,
6734 struct cfs_schedulable_data *d)
6742 period = tg_get_cfs_period(tg);
6743 quota = tg_get_cfs_quota(tg);
6746 /* note: these should typically be equivalent */
6747 if (quota == RUNTIME_INF || quota == -1)
6750 return to_ratio(period, quota);
6753 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
6755 struct cfs_schedulable_data *d = data;
6756 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6757 s64 quota = 0, parent_quota = -1;
6760 quota = RUNTIME_INF;
6762 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
6764 quota = normalize_cfs_quota(tg, d);
6765 parent_quota = parent_b->hierarchical_quota;
6768 * Ensure max(child_quota) <= parent_quota. On cgroup2,
6769 * always take the min. On cgroup1, only inherit when no
6772 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
6773 quota = min(quota, parent_quota);
6775 if (quota == RUNTIME_INF)
6776 quota = parent_quota;
6777 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
6781 cfs_b->hierarchical_quota = quota;
6786 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
6789 struct cfs_schedulable_data data = {
6795 if (quota != RUNTIME_INF) {
6796 do_div(data.period, NSEC_PER_USEC);
6797 do_div(data.quota, NSEC_PER_USEC);
6801 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
6807 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
6809 struct task_group *tg = css_tg(seq_css(sf));
6810 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6812 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
6813 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
6814 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
6816 if (schedstat_enabled() && tg != &root_task_group) {
6820 for_each_possible_cpu(i)
6821 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
6823 seq_printf(sf, "wait_sum %llu\n", ws);
6828 #endif /* CONFIG_CFS_BANDWIDTH */
6829 #endif /* CONFIG_FAIR_GROUP_SCHED */
6831 #ifdef CONFIG_RT_GROUP_SCHED
6832 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
6833 struct cftype *cft, s64 val)
6835 return sched_group_set_rt_runtime(css_tg(css), val);
6838 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
6841 return sched_group_rt_runtime(css_tg(css));
6844 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
6845 struct cftype *cftype, u64 rt_period_us)
6847 return sched_group_set_rt_period(css_tg(css), rt_period_us);
6850 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
6853 return sched_group_rt_period(css_tg(css));
6855 #endif /* CONFIG_RT_GROUP_SCHED */
6857 static struct cftype cpu_legacy_files[] = {
6858 #ifdef CONFIG_FAIR_GROUP_SCHED
6861 .read_u64 = cpu_shares_read_u64,
6862 .write_u64 = cpu_shares_write_u64,
6865 #ifdef CONFIG_CFS_BANDWIDTH
6867 .name = "cfs_quota_us",
6868 .read_s64 = cpu_cfs_quota_read_s64,
6869 .write_s64 = cpu_cfs_quota_write_s64,
6872 .name = "cfs_period_us",
6873 .read_u64 = cpu_cfs_period_read_u64,
6874 .write_u64 = cpu_cfs_period_write_u64,
6878 .seq_show = cpu_cfs_stat_show,
6881 #ifdef CONFIG_RT_GROUP_SCHED
6883 .name = "rt_runtime_us",
6884 .read_s64 = cpu_rt_runtime_read,
6885 .write_s64 = cpu_rt_runtime_write,
6888 .name = "rt_period_us",
6889 .read_u64 = cpu_rt_period_read_uint,
6890 .write_u64 = cpu_rt_period_write_uint,
6896 static int cpu_extra_stat_show(struct seq_file *sf,
6897 struct cgroup_subsys_state *css)
6899 #ifdef CONFIG_CFS_BANDWIDTH
6901 struct task_group *tg = css_tg(css);
6902 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6905 throttled_usec = cfs_b->throttled_time;
6906 do_div(throttled_usec, NSEC_PER_USEC);
6908 seq_printf(sf, "nr_periods %d\n"
6910 "throttled_usec %llu\n",
6911 cfs_b->nr_periods, cfs_b->nr_throttled,
6918 #ifdef CONFIG_FAIR_GROUP_SCHED
6919 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
6922 struct task_group *tg = css_tg(css);
6923 u64 weight = scale_load_down(tg->shares);
6925 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
6928 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
6929 struct cftype *cft, u64 weight)
6932 * cgroup weight knobs should use the common MIN, DFL and MAX
6933 * values which are 1, 100 and 10000 respectively. While it loses
6934 * a bit of range on both ends, it maps pretty well onto the shares
6935 * value used by scheduler and the round-trip conversions preserve
6936 * the original value over the entire range.
6938 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
6941 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
6943 return sched_group_set_shares(css_tg(css), scale_load(weight));
6946 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
6949 unsigned long weight = scale_load_down(css_tg(css)->shares);
6950 int last_delta = INT_MAX;
6953 /* find the closest nice value to the current weight */
6954 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
6955 delta = abs(sched_prio_to_weight[prio] - weight);
6956 if (delta >= last_delta)
6961 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
6964 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
6965 struct cftype *cft, s64 nice)
6967 unsigned long weight;
6970 if (nice < MIN_NICE || nice > MAX_NICE)
6973 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
6974 idx = array_index_nospec(idx, 40);
6975 weight = sched_prio_to_weight[idx];
6977 return sched_group_set_shares(css_tg(css), scale_load(weight));
6981 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
6982 long period, long quota)
6985 seq_puts(sf, "max");
6987 seq_printf(sf, "%ld", quota);
6989 seq_printf(sf, " %ld\n", period);
6992 /* caller should put the current value in *@periodp before calling */
6993 static int __maybe_unused cpu_period_quota_parse(char *buf,
6994 u64 *periodp, u64 *quotap)
6996 char tok[21]; /* U64_MAX */
6998 if (!sscanf(buf, "%s %llu", tok, periodp))
7001 *periodp *= NSEC_PER_USEC;
7003 if (sscanf(tok, "%llu", quotap))
7004 *quotap *= NSEC_PER_USEC;
7005 else if (!strcmp(tok, "max"))
7006 *quotap = RUNTIME_INF;
7013 #ifdef CONFIG_CFS_BANDWIDTH
7014 static int cpu_max_show(struct seq_file *sf, void *v)
7016 struct task_group *tg = css_tg(seq_css(sf));
7018 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
7022 static ssize_t cpu_max_write(struct kernfs_open_file *of,
7023 char *buf, size_t nbytes, loff_t off)
7025 struct task_group *tg = css_tg(of_css(of));
7026 u64 period = tg_get_cfs_period(tg);
7030 ret = cpu_period_quota_parse(buf, &period, "a);
7032 ret = tg_set_cfs_bandwidth(tg, period, quota);
7033 return ret ?: nbytes;
7037 static struct cftype cpu_files[] = {
7038 #ifdef CONFIG_FAIR_GROUP_SCHED
7041 .flags = CFTYPE_NOT_ON_ROOT,
7042 .read_u64 = cpu_weight_read_u64,
7043 .write_u64 = cpu_weight_write_u64,
7046 .name = "weight.nice",
7047 .flags = CFTYPE_NOT_ON_ROOT,
7048 .read_s64 = cpu_weight_nice_read_s64,
7049 .write_s64 = cpu_weight_nice_write_s64,
7052 #ifdef CONFIG_CFS_BANDWIDTH
7055 .flags = CFTYPE_NOT_ON_ROOT,
7056 .seq_show = cpu_max_show,
7057 .write = cpu_max_write,
7063 struct cgroup_subsys cpu_cgrp_subsys = {
7064 .css_alloc = cpu_cgroup_css_alloc,
7065 .css_online = cpu_cgroup_css_online,
7066 .css_released = cpu_cgroup_css_released,
7067 .css_free = cpu_cgroup_css_free,
7068 .css_extra_stat_show = cpu_extra_stat_show,
7069 .fork = cpu_cgroup_fork,
7070 .can_attach = cpu_cgroup_can_attach,
7071 .attach = cpu_cgroup_attach,
7072 .legacy_cftypes = cpu_legacy_files,
7073 .dfl_cftypes = cpu_files,
7078 #endif /* CONFIG_CGROUP_SCHED */
7080 void dump_cpu_task(int cpu)
7082 pr_info("Task dump for CPU %d:\n", cpu);
7083 sched_show_task(cpu_curr(cpu));
7087 * Nice levels are multiplicative, with a gentle 10% change for every
7088 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7089 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7090 * that remained on nice 0.
7092 * The "10% effect" is relative and cumulative: from _any_ nice level,
7093 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7094 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7095 * If a task goes up by ~10% and another task goes down by ~10% then
7096 * the relative distance between them is ~25%.)
7098 const int sched_prio_to_weight[40] = {
7099 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7100 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7101 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7102 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7103 /* 0 */ 1024, 820, 655, 526, 423,
7104 /* 5 */ 335, 272, 215, 172, 137,
7105 /* 10 */ 110, 87, 70, 56, 45,
7106 /* 15 */ 36, 29, 23, 18, 15,
7110 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7112 * In cases where the weight does not change often, we can use the
7113 * precalculated inverse to speed up arithmetics by turning divisions
7114 * into multiplications:
7116 const u32 sched_prio_to_wmult[40] = {
7117 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7118 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7119 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7120 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7121 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7122 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7123 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7124 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7127 #undef CREATE_TRACE_POINTS