4 * Core kernel scheduler code and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
10 #include <linux/nospec.h>
12 #include <linux/kcov.h>
14 #include <asm/switch_to.h>
17 #include "../workqueue_internal.h"
18 #include "../smpboot.h"
22 #define CREATE_TRACE_POINTS
23 #include <trace/events/sched.h>
25 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
27 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
29 * Debugging: various feature bits
31 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
32 * sysctl_sched_features, defined in sched.h, to allow constants propagation
33 * at compile time and compiler optimization based on features default.
35 #define SCHED_FEAT(name, enabled) \
36 (1UL << __SCHED_FEAT_##name) * enabled |
37 const_debug unsigned int sysctl_sched_features =
44 * Number of tasks to iterate in a single balance run.
45 * Limited because this is done with IRQs disabled.
47 const_debug unsigned int sysctl_sched_nr_migrate = 32;
50 * period over which we measure -rt task CPU usage in us.
53 unsigned int sysctl_sched_rt_period = 1000000;
55 __read_mostly int scheduler_running;
58 * part of the period that we allow rt tasks to run in us.
61 int sysctl_sched_rt_runtime = 950000;
64 * __task_rq_lock - lock the rq @p resides on.
66 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
71 lockdep_assert_held(&p->pi_lock);
75 raw_spin_lock(&rq->lock);
76 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
80 raw_spin_unlock(&rq->lock);
82 while (unlikely(task_on_rq_migrating(p)))
88 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
90 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
91 __acquires(p->pi_lock)
97 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
99 raw_spin_lock(&rq->lock);
101 * move_queued_task() task_rq_lock()
104 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
105 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
106 * [S] ->cpu = new_cpu [L] task_rq()
110 * If we observe the old CPU in task_rq_lock, the acquire of
111 * the old rq->lock will fully serialize against the stores.
113 * If we observe the new CPU in task_rq_lock, the acquire will
114 * pair with the WMB to ensure we must then also see migrating.
116 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
120 raw_spin_unlock(&rq->lock);
121 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
123 while (unlikely(task_on_rq_migrating(p)))
129 * RQ-clock updating methods:
132 static void update_rq_clock_task(struct rq *rq, s64 delta)
135 * In theory, the compile should just see 0 here, and optimize out the call
136 * to sched_rt_avg_update. But I don't trust it...
138 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
139 s64 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 HAVE_SCHED_AVG_IRQ
181 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
182 update_irq_load_avg(rq, irq_delta + steal);
186 void update_rq_clock(struct rq *rq)
190 lockdep_assert_held(&rq->lock);
192 if (rq->clock_update_flags & RQCF_ACT_SKIP)
195 #ifdef CONFIG_SCHED_DEBUG
196 if (sched_feat(WARN_DOUBLE_CLOCK))
197 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
198 rq->clock_update_flags |= RQCF_UPDATED;
201 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
205 update_rq_clock_task(rq, delta);
209 #ifdef CONFIG_SCHED_HRTICK
211 * Use HR-timers to deliver accurate preemption points.
214 static void hrtick_clear(struct rq *rq)
216 if (hrtimer_active(&rq->hrtick_timer))
217 hrtimer_cancel(&rq->hrtick_timer);
221 * High-resolution timer tick.
222 * Runs from hardirq context with interrupts disabled.
224 static enum hrtimer_restart hrtick(struct hrtimer *timer)
226 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
229 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
233 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
236 return HRTIMER_NORESTART;
241 static void __hrtick_restart(struct rq *rq)
243 struct hrtimer *timer = &rq->hrtick_timer;
245 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
249 * called from hardirq (IPI) context
251 static void __hrtick_start(void *arg)
257 __hrtick_restart(rq);
258 rq->hrtick_csd_pending = 0;
263 * Called to set the hrtick timer state.
265 * called with rq->lock held and irqs disabled
267 void hrtick_start(struct rq *rq, u64 delay)
269 struct hrtimer *timer = &rq->hrtick_timer;
274 * Don't schedule slices shorter than 10000ns, that just
275 * doesn't make sense and can cause timer DoS.
277 delta = max_t(s64, delay, 10000LL);
278 time = ktime_add_ns(timer->base->get_time(), delta);
280 hrtimer_set_expires(timer, time);
282 if (rq == this_rq()) {
283 __hrtick_restart(rq);
284 } else if (!rq->hrtick_csd_pending) {
285 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
286 rq->hrtick_csd_pending = 1;
292 * Called to set the hrtick timer state.
294 * called with rq->lock held and irqs disabled
296 void hrtick_start(struct rq *rq, u64 delay)
299 * Don't schedule slices shorter than 10000ns, that just
300 * doesn't make sense. Rely on vruntime for fairness.
302 delay = max_t(u64, delay, 10000LL);
303 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
304 HRTIMER_MODE_REL_PINNED);
306 #endif /* CONFIG_SMP */
308 static void hrtick_rq_init(struct rq *rq)
311 rq->hrtick_csd_pending = 0;
313 rq->hrtick_csd.flags = 0;
314 rq->hrtick_csd.func = __hrtick_start;
315 rq->hrtick_csd.info = rq;
318 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
319 rq->hrtick_timer.function = hrtick;
321 #else /* CONFIG_SCHED_HRTICK */
322 static inline void hrtick_clear(struct rq *rq)
326 static inline void hrtick_rq_init(struct rq *rq)
329 #endif /* CONFIG_SCHED_HRTICK */
332 * cmpxchg based fetch_or, macro so it works for different integer types
334 #define fetch_or(ptr, mask) \
336 typeof(ptr) _ptr = (ptr); \
337 typeof(mask) _mask = (mask); \
338 typeof(*_ptr) _old, _val = *_ptr; \
341 _old = cmpxchg(_ptr, _val, _val | _mask); \
349 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
351 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
352 * this avoids any races wrt polling state changes and thereby avoids
355 static bool set_nr_and_not_polling(struct task_struct *p)
357 struct thread_info *ti = task_thread_info(p);
358 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
362 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
364 * If this returns true, then the idle task promises to call
365 * sched_ttwu_pending() and reschedule soon.
367 static bool set_nr_if_polling(struct task_struct *p)
369 struct thread_info *ti = task_thread_info(p);
370 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
373 if (!(val & _TIF_POLLING_NRFLAG))
375 if (val & _TIF_NEED_RESCHED)
377 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
386 static bool set_nr_and_not_polling(struct task_struct *p)
388 set_tsk_need_resched(p);
393 static bool set_nr_if_polling(struct task_struct *p)
400 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
402 struct wake_q_node *node = &task->wake_q;
405 * Atomically grab the task, if ->wake_q is !nil already it means
406 * its already queued (either by us or someone else) and will get the
407 * wakeup due to that.
409 * This cmpxchg() executes a full barrier, which pairs with the full
410 * barrier executed by the wakeup in wake_up_q().
412 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
415 get_task_struct(task);
418 * The head is context local, there can be no concurrency.
421 head->lastp = &node->next;
424 void wake_up_q(struct wake_q_head *head)
426 struct wake_q_node *node = head->first;
428 while (node != WAKE_Q_TAIL) {
429 struct task_struct *task;
431 task = container_of(node, struct task_struct, wake_q);
433 /* Task can safely be re-inserted now: */
435 task->wake_q.next = NULL;
438 * wake_up_process() executes a full barrier, which pairs with
439 * the queueing in wake_q_add() so as not to miss wakeups.
441 wake_up_process(task);
442 put_task_struct(task);
447 * resched_curr - mark rq's current task 'to be rescheduled now'.
449 * On UP this means the setting of the need_resched flag, on SMP it
450 * might also involve a cross-CPU call to trigger the scheduler on
453 void resched_curr(struct rq *rq)
455 struct task_struct *curr = rq->curr;
458 lockdep_assert_held(&rq->lock);
460 if (test_tsk_need_resched(curr))
465 if (cpu == smp_processor_id()) {
466 set_tsk_need_resched(curr);
467 set_preempt_need_resched();
471 if (set_nr_and_not_polling(curr))
472 smp_send_reschedule(cpu);
474 trace_sched_wake_idle_without_ipi(cpu);
477 void resched_cpu(int cpu)
479 struct rq *rq = cpu_rq(cpu);
482 raw_spin_lock_irqsave(&rq->lock, flags);
483 if (cpu_online(cpu) || cpu == smp_processor_id())
485 raw_spin_unlock_irqrestore(&rq->lock, flags);
489 #ifdef CONFIG_NO_HZ_COMMON
491 * In the semi idle case, use the nearest busy CPU for migrating timers
492 * from an idle CPU. This is good for power-savings.
494 * We don't do similar optimization for completely idle system, as
495 * selecting an idle CPU will add more delays to the timers than intended
496 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
498 int get_nohz_timer_target(void)
500 int i, cpu = smp_processor_id();
501 struct sched_domain *sd;
503 if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
507 for_each_domain(cpu, sd) {
508 for_each_cpu(i, sched_domain_span(sd)) {
512 if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
519 if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
520 cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
527 * When add_timer_on() enqueues a timer into the timer wheel of an
528 * idle CPU then this timer might expire before the next timer event
529 * which is scheduled to wake up that CPU. In case of a completely
530 * idle system the next event might even be infinite time into the
531 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
532 * leaves the inner idle loop so the newly added timer is taken into
533 * account when the CPU goes back to idle and evaluates the timer
534 * wheel for the next timer event.
536 static void wake_up_idle_cpu(int cpu)
538 struct rq *rq = cpu_rq(cpu);
540 if (cpu == smp_processor_id())
543 if (set_nr_and_not_polling(rq->idle))
544 smp_send_reschedule(cpu);
546 trace_sched_wake_idle_without_ipi(cpu);
549 static bool wake_up_full_nohz_cpu(int cpu)
552 * We just need the target to call irq_exit() and re-evaluate
553 * the next tick. The nohz full kick at least implies that.
554 * If needed we can still optimize that later with an
557 if (cpu_is_offline(cpu))
558 return true; /* Don't try to wake offline CPUs. */
559 if (tick_nohz_full_cpu(cpu)) {
560 if (cpu != smp_processor_id() ||
561 tick_nohz_tick_stopped())
562 tick_nohz_full_kick_cpu(cpu);
570 * Wake up the specified CPU. If the CPU is going offline, it is the
571 * caller's responsibility to deal with the lost wakeup, for example,
572 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
574 void wake_up_nohz_cpu(int cpu)
576 if (!wake_up_full_nohz_cpu(cpu))
577 wake_up_idle_cpu(cpu);
580 static inline bool got_nohz_idle_kick(void)
582 int cpu = smp_processor_id();
584 if (!(atomic_read(nohz_flags(cpu)) & NOHZ_KICK_MASK))
587 if (idle_cpu(cpu) && !need_resched())
591 * We can't run Idle Load Balance on this CPU for this time so we
592 * cancel it and clear NOHZ_BALANCE_KICK
594 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
598 #else /* CONFIG_NO_HZ_COMMON */
600 static inline bool got_nohz_idle_kick(void)
605 #endif /* CONFIG_NO_HZ_COMMON */
607 #ifdef CONFIG_NO_HZ_FULL
608 bool sched_can_stop_tick(struct rq *rq)
612 /* Deadline tasks, even if single, need the tick */
613 if (rq->dl.dl_nr_running)
617 * If there are more than one RR tasks, we need the tick to effect the
618 * actual RR behaviour.
620 if (rq->rt.rr_nr_running) {
621 if (rq->rt.rr_nr_running == 1)
628 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
629 * forced preemption between FIFO tasks.
631 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
636 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
637 * if there's more than one we need the tick for involuntary
640 if (rq->nr_running > 1)
645 #endif /* CONFIG_NO_HZ_FULL */
646 #endif /* CONFIG_SMP */
648 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
649 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
651 * Iterate task_group tree rooted at *from, calling @down when first entering a
652 * node and @up when leaving it for the final time.
654 * Caller must hold rcu_lock or sufficient equivalent.
656 int walk_tg_tree_from(struct task_group *from,
657 tg_visitor down, tg_visitor up, void *data)
659 struct task_group *parent, *child;
665 ret = (*down)(parent, data);
668 list_for_each_entry_rcu(child, &parent->children, siblings) {
675 ret = (*up)(parent, data);
676 if (ret || parent == from)
680 parent = parent->parent;
687 int tg_nop(struct task_group *tg, void *data)
693 static void set_load_weight(struct task_struct *p, bool update_load)
695 int prio = p->static_prio - MAX_RT_PRIO;
696 struct load_weight *load = &p->se.load;
699 * SCHED_IDLE tasks get minimal weight:
701 if (idle_policy(p->policy)) {
702 load->weight = scale_load(WEIGHT_IDLEPRIO);
703 load->inv_weight = WMULT_IDLEPRIO;
708 * SCHED_OTHER tasks have to update their load when changing their
711 if (update_load && p->sched_class == &fair_sched_class) {
712 reweight_task(p, prio);
714 load->weight = scale_load(sched_prio_to_weight[prio]);
715 load->inv_weight = sched_prio_to_wmult[prio];
719 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
721 if (!(flags & ENQUEUE_NOCLOCK))
724 if (!(flags & ENQUEUE_RESTORE))
725 sched_info_queued(rq, p);
727 p->sched_class->enqueue_task(rq, p, flags);
730 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
732 if (!(flags & DEQUEUE_NOCLOCK))
735 if (!(flags & DEQUEUE_SAVE))
736 sched_info_dequeued(rq, p);
738 p->sched_class->dequeue_task(rq, p, flags);
741 void activate_task(struct rq *rq, struct task_struct *p, int flags)
743 if (task_contributes_to_load(p))
744 rq->nr_uninterruptible--;
746 enqueue_task(rq, p, flags);
749 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
751 if (task_contributes_to_load(p))
752 rq->nr_uninterruptible++;
754 dequeue_task(rq, p, flags);
758 * __normal_prio - return the priority that is based on the static prio
760 static inline int __normal_prio(struct task_struct *p)
762 return p->static_prio;
766 * Calculate the expected normal priority: i.e. priority
767 * without taking RT-inheritance into account. Might be
768 * boosted by interactivity modifiers. Changes upon fork,
769 * setprio syscalls, and whenever the interactivity
770 * estimator recalculates.
772 static inline int normal_prio(struct task_struct *p)
776 if (task_has_dl_policy(p))
777 prio = MAX_DL_PRIO-1;
778 else if (task_has_rt_policy(p))
779 prio = MAX_RT_PRIO-1 - p->rt_priority;
781 prio = __normal_prio(p);
786 * Calculate the current priority, i.e. the priority
787 * taken into account by the scheduler. This value might
788 * be boosted by RT tasks, or might be boosted by
789 * interactivity modifiers. Will be RT if the task got
790 * RT-boosted. If not then it returns p->normal_prio.
792 static int effective_prio(struct task_struct *p)
794 p->normal_prio = normal_prio(p);
796 * If we are RT tasks or we were boosted to RT priority,
797 * keep the priority unchanged. Otherwise, update priority
798 * to the normal priority:
800 if (!rt_prio(p->prio))
801 return p->normal_prio;
806 * task_curr - is this task currently executing on a CPU?
807 * @p: the task in question.
809 * Return: 1 if the task is currently executing. 0 otherwise.
811 inline int task_curr(const struct task_struct *p)
813 return cpu_curr(task_cpu(p)) == p;
817 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
818 * use the balance_callback list if you want balancing.
820 * this means any call to check_class_changed() must be followed by a call to
821 * balance_callback().
823 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
824 const struct sched_class *prev_class,
827 if (prev_class != p->sched_class) {
828 if (prev_class->switched_from)
829 prev_class->switched_from(rq, p);
831 p->sched_class->switched_to(rq, p);
832 } else if (oldprio != p->prio || dl_task(p))
833 p->sched_class->prio_changed(rq, p, oldprio);
836 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
838 const struct sched_class *class;
840 if (p->sched_class == rq->curr->sched_class) {
841 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
843 for_each_class(class) {
844 if (class == rq->curr->sched_class)
846 if (class == p->sched_class) {
854 * A queue event has occurred, and we're going to schedule. In
855 * this case, we can save a useless back to back clock update.
857 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
858 rq_clock_skip_update(rq);
863 static inline bool is_per_cpu_kthread(struct task_struct *p)
865 if (!(p->flags & PF_KTHREAD))
868 if (p->nr_cpus_allowed != 1)
875 * Per-CPU kthreads are allowed to run on !actie && online CPUs, see
876 * __set_cpus_allowed_ptr() and select_fallback_rq().
878 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
880 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
883 if (is_per_cpu_kthread(p))
884 return cpu_online(cpu);
886 return cpu_active(cpu);
890 * This is how migration works:
892 * 1) we invoke migration_cpu_stop() on the target CPU using
894 * 2) stopper starts to run (implicitly forcing the migrated thread
896 * 3) it checks whether the migrated task is still in the wrong runqueue.
897 * 4) if it's in the wrong runqueue then the migration thread removes
898 * it and puts it into the right queue.
899 * 5) stopper completes and stop_one_cpu() returns and the migration
904 * move_queued_task - move a queued task to new rq.
906 * Returns (locked) new rq. Old rq's lock is released.
908 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
909 struct task_struct *p, int new_cpu)
911 lockdep_assert_held(&rq->lock);
913 p->on_rq = TASK_ON_RQ_MIGRATING;
914 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
915 set_task_cpu(p, new_cpu);
918 rq = cpu_rq(new_cpu);
921 BUG_ON(task_cpu(p) != new_cpu);
922 enqueue_task(rq, p, 0);
923 p->on_rq = TASK_ON_RQ_QUEUED;
924 check_preempt_curr(rq, p, 0);
929 struct migration_arg {
930 struct task_struct *task;
935 * Move (not current) task off this CPU, onto the destination CPU. We're doing
936 * this because either it can't run here any more (set_cpus_allowed()
937 * away from this CPU, or CPU going down), or because we're
938 * attempting to rebalance this task on exec (sched_exec).
940 * So we race with normal scheduler movements, but that's OK, as long
941 * as the task is no longer on this CPU.
943 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
944 struct task_struct *p, int dest_cpu)
946 /* Affinity changed (again). */
947 if (!is_cpu_allowed(p, dest_cpu))
951 rq = move_queued_task(rq, rf, p, dest_cpu);
957 * migration_cpu_stop - this will be executed by a highprio stopper thread
958 * and performs thread migration by bumping thread off CPU then
959 * 'pushing' onto another runqueue.
961 static int migration_cpu_stop(void *data)
963 struct migration_arg *arg = data;
964 struct task_struct *p = arg->task;
965 struct rq *rq = this_rq();
969 * The original target CPU might have gone down and we might
970 * be on another CPU but it doesn't matter.
974 * We need to explicitly wake pending tasks before running
975 * __migrate_task() such that we will not miss enforcing cpus_allowed
976 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
978 sched_ttwu_pending();
980 raw_spin_lock(&p->pi_lock);
983 * If task_rq(p) != rq, it cannot be migrated here, because we're
984 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
985 * we're holding p->pi_lock.
987 if (task_rq(p) == rq) {
988 if (task_on_rq_queued(p))
989 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
991 p->wake_cpu = arg->dest_cpu;
994 raw_spin_unlock(&p->pi_lock);
1001 * sched_class::set_cpus_allowed must do the below, but is not required to
1002 * actually call this function.
1004 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1006 cpumask_copy(&p->cpus_allowed, new_mask);
1007 p->nr_cpus_allowed = cpumask_weight(new_mask);
1010 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1012 struct rq *rq = task_rq(p);
1013 bool queued, running;
1015 lockdep_assert_held(&p->pi_lock);
1017 queued = task_on_rq_queued(p);
1018 running = task_current(rq, p);
1022 * Because __kthread_bind() calls this on blocked tasks without
1025 lockdep_assert_held(&rq->lock);
1026 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1029 put_prev_task(rq, p);
1031 p->sched_class->set_cpus_allowed(p, new_mask);
1034 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1036 set_curr_task(rq, p);
1040 * Change a given task's CPU affinity. Migrate the thread to a
1041 * proper CPU and schedule it away if the CPU it's executing on
1042 * is removed from the allowed bitmask.
1044 * NOTE: the caller must have a valid reference to the task, the
1045 * task must not exit() & deallocate itself prematurely. The
1046 * call is not atomic; no spinlocks may be held.
1048 static int __set_cpus_allowed_ptr(struct task_struct *p,
1049 const struct cpumask *new_mask, bool check)
1051 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1052 unsigned int dest_cpu;
1057 rq = task_rq_lock(p, &rf);
1058 update_rq_clock(rq);
1060 if (p->flags & PF_KTHREAD) {
1062 * Kernel threads are allowed on online && !active CPUs
1064 cpu_valid_mask = cpu_online_mask;
1068 * Must re-check here, to close a race against __kthread_bind(),
1069 * sched_setaffinity() is not guaranteed to observe the flag.
1071 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1076 if (cpumask_equal(&p->cpus_allowed, new_mask))
1079 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1084 do_set_cpus_allowed(p, new_mask);
1086 if (p->flags & PF_KTHREAD) {
1088 * For kernel threads that do indeed end up on online &&
1089 * !active we want to ensure they are strict per-CPU threads.
1091 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1092 !cpumask_intersects(new_mask, cpu_active_mask) &&
1093 p->nr_cpus_allowed != 1);
1096 /* Can the task run on the task's current CPU? If so, we're done */
1097 if (cpumask_test_cpu(task_cpu(p), new_mask))
1100 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1101 if (task_running(rq, p) || p->state == TASK_WAKING) {
1102 struct migration_arg arg = { p, dest_cpu };
1103 /* Need help from migration thread: drop lock and wait. */
1104 task_rq_unlock(rq, p, &rf);
1105 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1106 tlb_migrate_finish(p->mm);
1108 } else if (task_on_rq_queued(p)) {
1110 * OK, since we're going to drop the lock immediately
1111 * afterwards anyway.
1113 rq = move_queued_task(rq, &rf, p, dest_cpu);
1116 task_rq_unlock(rq, p, &rf);
1121 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1123 return __set_cpus_allowed_ptr(p, new_mask, false);
1125 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1127 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1129 #ifdef CONFIG_SCHED_DEBUG
1131 * We should never call set_task_cpu() on a blocked task,
1132 * ttwu() will sort out the placement.
1134 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1138 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1139 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1140 * time relying on p->on_rq.
1142 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1143 p->sched_class == &fair_sched_class &&
1144 (p->on_rq && !task_on_rq_migrating(p)));
1146 #ifdef CONFIG_LOCKDEP
1148 * The caller should hold either p->pi_lock or rq->lock, when changing
1149 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1151 * sched_move_task() holds both and thus holding either pins the cgroup,
1154 * Furthermore, all task_rq users should acquire both locks, see
1157 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1158 lockdep_is_held(&task_rq(p)->lock)));
1161 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1163 WARN_ON_ONCE(!cpu_online(new_cpu));
1166 trace_sched_migrate_task(p, new_cpu);
1168 if (task_cpu(p) != new_cpu) {
1169 if (p->sched_class->migrate_task_rq)
1170 p->sched_class->migrate_task_rq(p);
1171 p->se.nr_migrations++;
1173 perf_event_task_migrate(p);
1176 __set_task_cpu(p, new_cpu);
1179 #ifdef CONFIG_NUMA_BALANCING
1180 static void __migrate_swap_task(struct task_struct *p, int cpu)
1182 if (task_on_rq_queued(p)) {
1183 struct rq *src_rq, *dst_rq;
1184 struct rq_flags srf, drf;
1186 src_rq = task_rq(p);
1187 dst_rq = cpu_rq(cpu);
1189 rq_pin_lock(src_rq, &srf);
1190 rq_pin_lock(dst_rq, &drf);
1192 p->on_rq = TASK_ON_RQ_MIGRATING;
1193 deactivate_task(src_rq, p, 0);
1194 set_task_cpu(p, cpu);
1195 activate_task(dst_rq, p, 0);
1196 p->on_rq = TASK_ON_RQ_QUEUED;
1197 check_preempt_curr(dst_rq, p, 0);
1199 rq_unpin_lock(dst_rq, &drf);
1200 rq_unpin_lock(src_rq, &srf);
1204 * Task isn't running anymore; make it appear like we migrated
1205 * it before it went to sleep. This means on wakeup we make the
1206 * previous CPU our target instead of where it really is.
1212 struct migration_swap_arg {
1213 struct task_struct *src_task, *dst_task;
1214 int src_cpu, dst_cpu;
1217 static int migrate_swap_stop(void *data)
1219 struct migration_swap_arg *arg = data;
1220 struct rq *src_rq, *dst_rq;
1223 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1226 src_rq = cpu_rq(arg->src_cpu);
1227 dst_rq = cpu_rq(arg->dst_cpu);
1229 double_raw_lock(&arg->src_task->pi_lock,
1230 &arg->dst_task->pi_lock);
1231 double_rq_lock(src_rq, dst_rq);
1233 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1236 if (task_cpu(arg->src_task) != arg->src_cpu)
1239 if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1242 if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1245 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1246 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1251 double_rq_unlock(src_rq, dst_rq);
1252 raw_spin_unlock(&arg->dst_task->pi_lock);
1253 raw_spin_unlock(&arg->src_task->pi_lock);
1259 * Cross migrate two tasks
1261 int migrate_swap(struct task_struct *cur, struct task_struct *p,
1262 int target_cpu, int curr_cpu)
1264 struct migration_swap_arg arg;
1267 arg = (struct migration_swap_arg){
1269 .src_cpu = curr_cpu,
1271 .dst_cpu = target_cpu,
1274 if (arg.src_cpu == arg.dst_cpu)
1278 * These three tests are all lockless; this is OK since all of them
1279 * will be re-checked with proper locks held further down the line.
1281 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1284 if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1287 if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1290 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1291 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1296 #endif /* CONFIG_NUMA_BALANCING */
1299 * wait_task_inactive - wait for a thread to unschedule.
1301 * If @match_state is nonzero, it's the @p->state value just checked and
1302 * not expected to change. If it changes, i.e. @p might have woken up,
1303 * then return zero. When we succeed in waiting for @p to be off its CPU,
1304 * we return a positive number (its total switch count). If a second call
1305 * a short while later returns the same number, the caller can be sure that
1306 * @p has remained unscheduled the whole time.
1308 * The caller must ensure that the task *will* unschedule sometime soon,
1309 * else this function might spin for a *long* time. This function can't
1310 * be called with interrupts off, or it may introduce deadlock with
1311 * smp_call_function() if an IPI is sent by the same process we are
1312 * waiting to become inactive.
1314 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1316 int running, queued;
1323 * We do the initial early heuristics without holding
1324 * any task-queue locks at all. We'll only try to get
1325 * the runqueue lock when things look like they will
1331 * If the task is actively running on another CPU
1332 * still, just relax and busy-wait without holding
1335 * NOTE! Since we don't hold any locks, it's not
1336 * even sure that "rq" stays as the right runqueue!
1337 * But we don't care, since "task_running()" will
1338 * return false if the runqueue has changed and p
1339 * is actually now running somewhere else!
1341 while (task_running(rq, p)) {
1342 if (match_state && unlikely(p->state != match_state))
1348 * Ok, time to look more closely! We need the rq
1349 * lock now, to be *sure*. If we're wrong, we'll
1350 * just go back and repeat.
1352 rq = task_rq_lock(p, &rf);
1353 trace_sched_wait_task(p);
1354 running = task_running(rq, p);
1355 queued = task_on_rq_queued(p);
1357 if (!match_state || p->state == match_state)
1358 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1359 task_rq_unlock(rq, p, &rf);
1362 * If it changed from the expected state, bail out now.
1364 if (unlikely(!ncsw))
1368 * Was it really running after all now that we
1369 * checked with the proper locks actually held?
1371 * Oops. Go back and try again..
1373 if (unlikely(running)) {
1379 * It's not enough that it's not actively running,
1380 * it must be off the runqueue _entirely_, and not
1383 * So if it was still runnable (but just not actively
1384 * running right now), it's preempted, and we should
1385 * yield - it could be a while.
1387 if (unlikely(queued)) {
1388 ktime_t to = NSEC_PER_SEC / HZ;
1390 set_current_state(TASK_UNINTERRUPTIBLE);
1391 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1396 * Ahh, all good. It wasn't running, and it wasn't
1397 * runnable, which means that it will never become
1398 * running in the future either. We're all done!
1407 * kick_process - kick a running thread to enter/exit the kernel
1408 * @p: the to-be-kicked thread
1410 * Cause a process which is running on another CPU to enter
1411 * kernel-mode, without any delay. (to get signals handled.)
1413 * NOTE: this function doesn't have to take the runqueue lock,
1414 * because all it wants to ensure is that the remote task enters
1415 * the kernel. If the IPI races and the task has been migrated
1416 * to another CPU then no harm is done and the purpose has been
1419 void kick_process(struct task_struct *p)
1425 if ((cpu != smp_processor_id()) && task_curr(p))
1426 smp_send_reschedule(cpu);
1429 EXPORT_SYMBOL_GPL(kick_process);
1432 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1434 * A few notes on cpu_active vs cpu_online:
1436 * - cpu_active must be a subset of cpu_online
1438 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
1439 * see __set_cpus_allowed_ptr(). At this point the newly online
1440 * CPU isn't yet part of the sched domains, and balancing will not
1443 * - on CPU-down we clear cpu_active() to mask the sched domains and
1444 * avoid the load balancer to place new tasks on the to be removed
1445 * CPU. Existing tasks will remain running there and will be taken
1448 * This means that fallback selection must not select !active CPUs.
1449 * And can assume that any active CPU must be online. Conversely
1450 * select_task_rq() below may allow selection of !active CPUs in order
1451 * to satisfy the above rules.
1453 static int select_fallback_rq(int cpu, struct task_struct *p)
1455 int nid = cpu_to_node(cpu);
1456 const struct cpumask *nodemask = NULL;
1457 enum { cpuset, possible, fail } state = cpuset;
1461 * If the node that the CPU is on has been offlined, cpu_to_node()
1462 * will return -1. There is no CPU on the node, and we should
1463 * select the CPU on the other node.
1466 nodemask = cpumask_of_node(nid);
1468 /* Look for allowed, online CPU in same node. */
1469 for_each_cpu(dest_cpu, nodemask) {
1470 if (!cpu_active(dest_cpu))
1472 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1478 /* Any allowed, online CPU? */
1479 for_each_cpu(dest_cpu, &p->cpus_allowed) {
1480 if (!is_cpu_allowed(p, dest_cpu))
1486 /* No more Mr. Nice Guy. */
1489 if (IS_ENABLED(CONFIG_CPUSETS)) {
1490 cpuset_cpus_allowed_fallback(p);
1496 do_set_cpus_allowed(p, cpu_possible_mask);
1507 if (state != cpuset) {
1509 * Don't tell them about moving exiting tasks or
1510 * kernel threads (both mm NULL), since they never
1513 if (p->mm && printk_ratelimit()) {
1514 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1515 task_pid_nr(p), p->comm, cpu);
1523 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1526 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1528 lockdep_assert_held(&p->pi_lock);
1530 if (p->nr_cpus_allowed > 1)
1531 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1533 cpu = cpumask_any(&p->cpus_allowed);
1536 * In order not to call set_task_cpu() on a blocking task we need
1537 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1540 * Since this is common to all placement strategies, this lives here.
1542 * [ this allows ->select_task() to simply return task_cpu(p) and
1543 * not worry about this generic constraint ]
1545 if (unlikely(!is_cpu_allowed(p, cpu)))
1546 cpu = select_fallback_rq(task_cpu(p), p);
1551 static void update_avg(u64 *avg, u64 sample)
1553 s64 diff = sample - *avg;
1557 void sched_set_stop_task(int cpu, struct task_struct *stop)
1559 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1560 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1564 * Make it appear like a SCHED_FIFO task, its something
1565 * userspace knows about and won't get confused about.
1567 * Also, it will make PI more or less work without too
1568 * much confusion -- but then, stop work should not
1569 * rely on PI working anyway.
1571 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1573 stop->sched_class = &stop_sched_class;
1576 cpu_rq(cpu)->stop = stop;
1580 * Reset it back to a normal scheduling class so that
1581 * it can die in pieces.
1583 old_stop->sched_class = &rt_sched_class;
1589 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1590 const struct cpumask *new_mask, bool check)
1592 return set_cpus_allowed_ptr(p, new_mask);
1595 #endif /* CONFIG_SMP */
1598 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1602 if (!schedstat_enabled())
1608 if (cpu == rq->cpu) {
1609 __schedstat_inc(rq->ttwu_local);
1610 __schedstat_inc(p->se.statistics.nr_wakeups_local);
1612 struct sched_domain *sd;
1614 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
1616 for_each_domain(rq->cpu, sd) {
1617 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1618 __schedstat_inc(sd->ttwu_wake_remote);
1625 if (wake_flags & WF_MIGRATED)
1626 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1627 #endif /* CONFIG_SMP */
1629 __schedstat_inc(rq->ttwu_count);
1630 __schedstat_inc(p->se.statistics.nr_wakeups);
1632 if (wake_flags & WF_SYNC)
1633 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
1636 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1638 activate_task(rq, p, en_flags);
1639 p->on_rq = TASK_ON_RQ_QUEUED;
1641 /* If a worker is waking up, notify the workqueue: */
1642 if (p->flags & PF_WQ_WORKER)
1643 wq_worker_waking_up(p, cpu_of(rq));
1647 * Mark the task runnable and perform wakeup-preemption.
1649 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1650 struct rq_flags *rf)
1652 check_preempt_curr(rq, p, wake_flags);
1653 p->state = TASK_RUNNING;
1654 trace_sched_wakeup(p);
1657 if (p->sched_class->task_woken) {
1659 * Our task @p is fully woken up and running; so its safe to
1660 * drop the rq->lock, hereafter rq is only used for statistics.
1662 rq_unpin_lock(rq, rf);
1663 p->sched_class->task_woken(rq, p);
1664 rq_repin_lock(rq, rf);
1667 if (rq->idle_stamp) {
1668 u64 delta = rq_clock(rq) - rq->idle_stamp;
1669 u64 max = 2*rq->max_idle_balance_cost;
1671 update_avg(&rq->avg_idle, delta);
1673 if (rq->avg_idle > max)
1682 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1683 struct rq_flags *rf)
1685 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
1687 lockdep_assert_held(&rq->lock);
1690 if (p->sched_contributes_to_load)
1691 rq->nr_uninterruptible--;
1693 if (wake_flags & WF_MIGRATED)
1694 en_flags |= ENQUEUE_MIGRATED;
1697 ttwu_activate(rq, p, en_flags);
1698 ttwu_do_wakeup(rq, p, wake_flags, rf);
1702 * Called in case the task @p isn't fully descheduled from its runqueue,
1703 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1704 * since all we need to do is flip p->state to TASK_RUNNING, since
1705 * the task is still ->on_rq.
1707 static int ttwu_remote(struct task_struct *p, int wake_flags)
1713 rq = __task_rq_lock(p, &rf);
1714 if (task_on_rq_queued(p)) {
1715 /* check_preempt_curr() may use rq clock */
1716 update_rq_clock(rq);
1717 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1720 __task_rq_unlock(rq, &rf);
1726 void sched_ttwu_pending(void)
1728 struct rq *rq = this_rq();
1729 struct llist_node *llist = llist_del_all(&rq->wake_list);
1730 struct task_struct *p, *t;
1736 rq_lock_irqsave(rq, &rf);
1737 update_rq_clock(rq);
1739 llist_for_each_entry_safe(p, t, llist, wake_entry)
1740 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
1742 rq_unlock_irqrestore(rq, &rf);
1745 void scheduler_ipi(void)
1748 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1749 * TIF_NEED_RESCHED remotely (for the first time) will also send
1752 preempt_fold_need_resched();
1754 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1758 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1759 * traditionally all their work was done from the interrupt return
1760 * path. Now that we actually do some work, we need to make sure
1763 * Some archs already do call them, luckily irq_enter/exit nest
1766 * Arguably we should visit all archs and update all handlers,
1767 * however a fair share of IPIs are still resched only so this would
1768 * somewhat pessimize the simple resched case.
1771 sched_ttwu_pending();
1774 * Check if someone kicked us for doing the nohz idle load balance.
1776 if (unlikely(got_nohz_idle_kick())) {
1777 this_rq()->idle_balance = 1;
1778 raise_softirq_irqoff(SCHED_SOFTIRQ);
1783 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1785 struct rq *rq = cpu_rq(cpu);
1787 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1789 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1790 if (!set_nr_if_polling(rq->idle))
1791 smp_send_reschedule(cpu);
1793 trace_sched_wake_idle_without_ipi(cpu);
1797 void wake_up_if_idle(int cpu)
1799 struct rq *rq = cpu_rq(cpu);
1804 if (!is_idle_task(rcu_dereference(rq->curr)))
1807 if (set_nr_if_polling(rq->idle)) {
1808 trace_sched_wake_idle_without_ipi(cpu);
1810 rq_lock_irqsave(rq, &rf);
1811 if (is_idle_task(rq->curr))
1812 smp_send_reschedule(cpu);
1813 /* Else CPU is not idle, do nothing here: */
1814 rq_unlock_irqrestore(rq, &rf);
1821 bool cpus_share_cache(int this_cpu, int that_cpu)
1823 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1825 #endif /* CONFIG_SMP */
1827 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1829 struct rq *rq = cpu_rq(cpu);
1832 #if defined(CONFIG_SMP)
1833 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1834 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1835 ttwu_queue_remote(p, cpu, wake_flags);
1841 update_rq_clock(rq);
1842 ttwu_do_activate(rq, p, wake_flags, &rf);
1847 * Notes on Program-Order guarantees on SMP systems.
1851 * The basic program-order guarantee on SMP systems is that when a task [t]
1852 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1853 * execution on its new CPU [c1].
1855 * For migration (of runnable tasks) this is provided by the following means:
1857 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1858 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1859 * rq(c1)->lock (if not at the same time, then in that order).
1860 * C) LOCK of the rq(c1)->lock scheduling in task
1862 * Release/acquire chaining guarantees that B happens after A and C after B.
1863 * Note: the CPU doing B need not be c0 or c1
1872 * UNLOCK rq(0)->lock
1874 * LOCK rq(0)->lock // orders against CPU0
1876 * UNLOCK rq(0)->lock
1880 * UNLOCK rq(1)->lock
1882 * LOCK rq(1)->lock // orders against CPU2
1885 * UNLOCK rq(1)->lock
1888 * BLOCKING -- aka. SLEEP + WAKEUP
1890 * For blocking we (obviously) need to provide the same guarantee as for
1891 * migration. However the means are completely different as there is no lock
1892 * chain to provide order. Instead we do:
1894 * 1) smp_store_release(X->on_cpu, 0)
1895 * 2) smp_cond_load_acquire(!X->on_cpu)
1899 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1901 * LOCK rq(0)->lock LOCK X->pi_lock
1904 * smp_store_release(X->on_cpu, 0);
1906 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1912 * X->state = RUNNING
1913 * UNLOCK rq(2)->lock
1915 * LOCK rq(2)->lock // orders against CPU1
1918 * UNLOCK rq(2)->lock
1921 * UNLOCK rq(0)->lock
1924 * However, for wakeups there is a second guarantee we must provide, namely we
1925 * must ensure that CONDITION=1 done by the caller can not be reordered with
1926 * accesses to the task state; see try_to_wake_up() and set_current_state().
1930 * try_to_wake_up - wake up a thread
1931 * @p: the thread to be awakened
1932 * @state: the mask of task states that can be woken
1933 * @wake_flags: wake modifier flags (WF_*)
1935 * If (@state & @p->state) @p->state = TASK_RUNNING.
1937 * If the task was not queued/runnable, also place it back on a runqueue.
1939 * Atomic against schedule() which would dequeue a task, also see
1940 * set_current_state().
1942 * This function executes a full memory barrier before accessing the task
1943 * state; see set_current_state().
1945 * Return: %true if @p->state changes (an actual wakeup was done),
1949 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1951 unsigned long flags;
1952 int cpu, success = 0;
1955 * If we are going to wake up a thread waiting for CONDITION we
1956 * need to ensure that CONDITION=1 done by the caller can not be
1957 * reordered with p->state check below. This pairs with mb() in
1958 * set_current_state() the waiting thread does.
1960 raw_spin_lock_irqsave(&p->pi_lock, flags);
1961 smp_mb__after_spinlock();
1962 if (!(p->state & state))
1965 trace_sched_waking(p);
1967 /* We're going to change ->state: */
1972 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1973 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1974 * in smp_cond_load_acquire() below.
1976 * sched_ttwu_pending() try_to_wake_up()
1977 * STORE p->on_rq = 1 LOAD p->state
1980 * __schedule() (switch to task 'p')
1981 * LOCK rq->lock smp_rmb();
1982 * smp_mb__after_spinlock();
1986 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
1988 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
1989 * __schedule(). See the comment for smp_mb__after_spinlock().
1992 if (p->on_rq && ttwu_remote(p, wake_flags))
1997 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1998 * possible to, falsely, observe p->on_cpu == 0.
2000 * One must be running (->on_cpu == 1) in order to remove oneself
2001 * from the runqueue.
2003 * __schedule() (switch to task 'p') try_to_wake_up()
2004 * STORE p->on_cpu = 1 LOAD p->on_rq
2007 * __schedule() (put 'p' to sleep)
2008 * LOCK rq->lock smp_rmb();
2009 * smp_mb__after_spinlock();
2010 * STORE p->on_rq = 0 LOAD p->on_cpu
2012 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
2013 * __schedule(). See the comment for smp_mb__after_spinlock().
2018 * If the owning (remote) CPU is still in the middle of schedule() with
2019 * this task as prev, wait until its done referencing the task.
2021 * Pairs with the smp_store_release() in finish_task().
2023 * This ensures that tasks getting woken will be fully ordered against
2024 * their previous state and preserve Program Order.
2026 smp_cond_load_acquire(&p->on_cpu, !VAL);
2028 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2029 p->state = TASK_WAKING;
2032 delayacct_blkio_end(p);
2033 atomic_dec(&task_rq(p)->nr_iowait);
2036 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2037 if (task_cpu(p) != cpu) {
2038 wake_flags |= WF_MIGRATED;
2039 set_task_cpu(p, cpu);
2042 #else /* CONFIG_SMP */
2045 delayacct_blkio_end(p);
2046 atomic_dec(&task_rq(p)->nr_iowait);
2049 #endif /* CONFIG_SMP */
2051 ttwu_queue(p, cpu, wake_flags);
2053 ttwu_stat(p, cpu, wake_flags);
2055 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2061 * try_to_wake_up_local - try to wake up a local task with rq lock held
2062 * @p: the thread to be awakened
2063 * @rf: request-queue flags for pinning
2065 * Put @p on the run-queue if it's not already there. The caller must
2066 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2069 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2071 struct rq *rq = task_rq(p);
2073 if (WARN_ON_ONCE(rq != this_rq()) ||
2074 WARN_ON_ONCE(p == current))
2077 lockdep_assert_held(&rq->lock);
2079 if (!raw_spin_trylock(&p->pi_lock)) {
2081 * This is OK, because current is on_cpu, which avoids it being
2082 * picked for load-balance and preemption/IRQs are still
2083 * disabled avoiding further scheduler activity on it and we've
2084 * not yet picked a replacement task.
2087 raw_spin_lock(&p->pi_lock);
2091 if (!(p->state & TASK_NORMAL))
2094 trace_sched_waking(p);
2096 if (!task_on_rq_queued(p)) {
2098 delayacct_blkio_end(p);
2099 atomic_dec(&rq->nr_iowait);
2101 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK);
2104 ttwu_do_wakeup(rq, p, 0, rf);
2105 ttwu_stat(p, smp_processor_id(), 0);
2107 raw_spin_unlock(&p->pi_lock);
2111 * wake_up_process - Wake up a specific process
2112 * @p: The process to be woken up.
2114 * Attempt to wake up the nominated process and move it to the set of runnable
2117 * Return: 1 if the process was woken up, 0 if it was already running.
2119 * This function executes a full memory barrier before accessing the task state.
2121 int wake_up_process(struct task_struct *p)
2123 return try_to_wake_up(p, TASK_NORMAL, 0);
2125 EXPORT_SYMBOL(wake_up_process);
2127 int wake_up_state(struct task_struct *p, unsigned int state)
2129 return try_to_wake_up(p, state, 0);
2133 * Perform scheduler related setup for a newly forked process p.
2134 * p is forked by current.
2136 * __sched_fork() is basic setup used by init_idle() too:
2138 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2143 p->se.exec_start = 0;
2144 p->se.sum_exec_runtime = 0;
2145 p->se.prev_sum_exec_runtime = 0;
2146 p->se.nr_migrations = 0;
2148 INIT_LIST_HEAD(&p->se.group_node);
2150 #ifdef CONFIG_FAIR_GROUP_SCHED
2151 p->se.cfs_rq = NULL;
2154 #ifdef CONFIG_SCHEDSTATS
2155 /* Even if schedstat is disabled, there should not be garbage */
2156 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2159 RB_CLEAR_NODE(&p->dl.rb_node);
2160 init_dl_task_timer(&p->dl);
2161 init_dl_inactive_task_timer(&p->dl);
2162 __dl_clear_params(p);
2164 INIT_LIST_HEAD(&p->rt.run_list);
2166 p->rt.time_slice = sched_rr_timeslice;
2170 #ifdef CONFIG_PREEMPT_NOTIFIERS
2171 INIT_HLIST_HEAD(&p->preempt_notifiers);
2174 init_numa_balancing(clone_flags, p);
2177 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2179 #ifdef CONFIG_NUMA_BALANCING
2181 void set_numabalancing_state(bool enabled)
2184 static_branch_enable(&sched_numa_balancing);
2186 static_branch_disable(&sched_numa_balancing);
2189 #ifdef CONFIG_PROC_SYSCTL
2190 int sysctl_numa_balancing(struct ctl_table *table, int write,
2191 void __user *buffer, size_t *lenp, loff_t *ppos)
2195 int state = static_branch_likely(&sched_numa_balancing);
2197 if (write && !capable(CAP_SYS_ADMIN))
2202 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2206 set_numabalancing_state(state);
2212 #ifdef CONFIG_SCHEDSTATS
2214 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2215 static bool __initdata __sched_schedstats = false;
2217 static void set_schedstats(bool enabled)
2220 static_branch_enable(&sched_schedstats);
2222 static_branch_disable(&sched_schedstats);
2225 void force_schedstat_enabled(void)
2227 if (!schedstat_enabled()) {
2228 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2229 static_branch_enable(&sched_schedstats);
2233 static int __init setup_schedstats(char *str)
2240 * This code is called before jump labels have been set up, so we can't
2241 * change the static branch directly just yet. Instead set a temporary
2242 * variable so init_schedstats() can do it later.
2244 if (!strcmp(str, "enable")) {
2245 __sched_schedstats = true;
2247 } else if (!strcmp(str, "disable")) {
2248 __sched_schedstats = false;
2253 pr_warn("Unable to parse schedstats=\n");
2257 __setup("schedstats=", setup_schedstats);
2259 static void __init init_schedstats(void)
2261 set_schedstats(__sched_schedstats);
2264 #ifdef CONFIG_PROC_SYSCTL
2265 int sysctl_schedstats(struct ctl_table *table, int write,
2266 void __user *buffer, size_t *lenp, loff_t *ppos)
2270 int state = static_branch_likely(&sched_schedstats);
2272 if (write && !capable(CAP_SYS_ADMIN))
2277 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2281 set_schedstats(state);
2284 #endif /* CONFIG_PROC_SYSCTL */
2285 #else /* !CONFIG_SCHEDSTATS */
2286 static inline void init_schedstats(void) {}
2287 #endif /* CONFIG_SCHEDSTATS */
2290 * fork()/clone()-time setup:
2292 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2294 unsigned long flags;
2296 __sched_fork(clone_flags, p);
2298 * We mark the process as NEW here. This guarantees that
2299 * nobody will actually run it, and a signal or other external
2300 * event cannot wake it up and insert it on the runqueue either.
2302 p->state = TASK_NEW;
2305 * Make sure we do not leak PI boosting priority to the child.
2307 p->prio = current->normal_prio;
2310 * Revert to default priority/policy on fork if requested.
2312 if (unlikely(p->sched_reset_on_fork)) {
2313 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2314 p->policy = SCHED_NORMAL;
2315 p->static_prio = NICE_TO_PRIO(0);
2317 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2318 p->static_prio = NICE_TO_PRIO(0);
2320 p->prio = p->normal_prio = __normal_prio(p);
2321 set_load_weight(p, false);
2324 * We don't need the reset flag anymore after the fork. It has
2325 * fulfilled its duty:
2327 p->sched_reset_on_fork = 0;
2330 if (dl_prio(p->prio))
2332 else if (rt_prio(p->prio))
2333 p->sched_class = &rt_sched_class;
2335 p->sched_class = &fair_sched_class;
2337 init_entity_runnable_average(&p->se);
2340 * The child is not yet in the pid-hash so no cgroup attach races,
2341 * and the cgroup is pinned to this child due to cgroup_fork()
2342 * is ran before sched_fork().
2344 * Silence PROVE_RCU.
2346 raw_spin_lock_irqsave(&p->pi_lock, flags);
2348 * We're setting the CPU for the first time, we don't migrate,
2349 * so use __set_task_cpu().
2351 __set_task_cpu(p, smp_processor_id());
2352 if (p->sched_class->task_fork)
2353 p->sched_class->task_fork(p);
2354 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2356 #ifdef CONFIG_SCHED_INFO
2357 if (likely(sched_info_on()))
2358 memset(&p->sched_info, 0, sizeof(p->sched_info));
2360 #if defined(CONFIG_SMP)
2363 init_task_preempt_count(p);
2365 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2366 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2371 unsigned long to_ratio(u64 period, u64 runtime)
2373 if (runtime == RUNTIME_INF)
2377 * Doing this here saves a lot of checks in all
2378 * the calling paths, and returning zero seems
2379 * safe for them anyway.
2384 return div64_u64(runtime << BW_SHIFT, period);
2388 * wake_up_new_task - wake up a newly created task for the first time.
2390 * This function will do some initial scheduler statistics housekeeping
2391 * that must be done for every newly created context, then puts the task
2392 * on the runqueue and wakes it.
2394 void wake_up_new_task(struct task_struct *p)
2399 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2400 p->state = TASK_RUNNING;
2403 * Fork balancing, do it here and not earlier because:
2404 * - cpus_allowed can change in the fork path
2405 * - any previously selected CPU might disappear through hotplug
2407 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2408 * as we're not fully set-up yet.
2410 p->recent_used_cpu = task_cpu(p);
2411 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2413 rq = __task_rq_lock(p, &rf);
2414 update_rq_clock(rq);
2415 post_init_entity_util_avg(&p->se);
2417 activate_task(rq, p, ENQUEUE_NOCLOCK);
2418 p->on_rq = TASK_ON_RQ_QUEUED;
2419 trace_sched_wakeup_new(p);
2420 check_preempt_curr(rq, p, WF_FORK);
2422 if (p->sched_class->task_woken) {
2424 * Nothing relies on rq->lock after this, so its fine to
2427 rq_unpin_lock(rq, &rf);
2428 p->sched_class->task_woken(rq, p);
2429 rq_repin_lock(rq, &rf);
2432 task_rq_unlock(rq, p, &rf);
2435 #ifdef CONFIG_PREEMPT_NOTIFIERS
2437 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2439 void preempt_notifier_inc(void)
2441 static_branch_inc(&preempt_notifier_key);
2443 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2445 void preempt_notifier_dec(void)
2447 static_branch_dec(&preempt_notifier_key);
2449 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2452 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2453 * @notifier: notifier struct to register
2455 void preempt_notifier_register(struct preempt_notifier *notifier)
2457 if (!static_branch_unlikely(&preempt_notifier_key))
2458 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2460 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2462 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2465 * preempt_notifier_unregister - no longer interested in preemption notifications
2466 * @notifier: notifier struct to unregister
2468 * This is *not* safe to call from within a preemption notifier.
2470 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2472 hlist_del(¬ifier->link);
2474 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2476 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2478 struct preempt_notifier *notifier;
2480 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2481 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2484 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2486 if (static_branch_unlikely(&preempt_notifier_key))
2487 __fire_sched_in_preempt_notifiers(curr);
2491 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2492 struct task_struct *next)
2494 struct preempt_notifier *notifier;
2496 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2497 notifier->ops->sched_out(notifier, next);
2500 static __always_inline void
2501 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2502 struct task_struct *next)
2504 if (static_branch_unlikely(&preempt_notifier_key))
2505 __fire_sched_out_preempt_notifiers(curr, next);
2508 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2510 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2515 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2516 struct task_struct *next)
2520 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2522 static inline void prepare_task(struct task_struct *next)
2526 * Claim the task as running, we do this before switching to it
2527 * such that any running task will have this set.
2533 static inline void finish_task(struct task_struct *prev)
2537 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2538 * We must ensure this doesn't happen until the switch is completely
2541 * In particular, the load of prev->state in finish_task_switch() must
2542 * happen before this.
2544 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2546 smp_store_release(&prev->on_cpu, 0);
2551 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
2554 * Since the runqueue lock will be released by the next
2555 * task (which is an invalid locking op but in the case
2556 * of the scheduler it's an obvious special-case), so we
2557 * do an early lockdep release here:
2559 rq_unpin_lock(rq, rf);
2560 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2561 #ifdef CONFIG_DEBUG_SPINLOCK
2562 /* this is a valid case when another task releases the spinlock */
2563 rq->lock.owner = next;
2567 static inline void finish_lock_switch(struct rq *rq)
2570 * If we are tracking spinlock dependencies then we have to
2571 * fix up the runqueue lock - which gets 'carried over' from
2572 * prev into current:
2574 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
2575 raw_spin_unlock_irq(&rq->lock);
2579 * NOP if the arch has not defined these:
2582 #ifndef prepare_arch_switch
2583 # define prepare_arch_switch(next) do { } while (0)
2586 #ifndef finish_arch_post_lock_switch
2587 # define finish_arch_post_lock_switch() do { } while (0)
2591 * prepare_task_switch - prepare to switch tasks
2592 * @rq: the runqueue preparing to switch
2593 * @prev: the current task that is being switched out
2594 * @next: the task we are going to switch to.
2596 * This is called with the rq lock held and interrupts off. It must
2597 * be paired with a subsequent finish_task_switch after the context
2600 * prepare_task_switch sets up locking and calls architecture specific
2604 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2605 struct task_struct *next)
2607 kcov_prepare_switch(prev);
2608 sched_info_switch(rq, prev, next);
2609 perf_event_task_sched_out(prev, next);
2611 fire_sched_out_preempt_notifiers(prev, next);
2613 prepare_arch_switch(next);
2617 * finish_task_switch - clean up after a task-switch
2618 * @prev: the thread we just switched away from.
2620 * finish_task_switch must be called after the context switch, paired
2621 * with a prepare_task_switch call before the context switch.
2622 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2623 * and do any other architecture-specific cleanup actions.
2625 * Note that we may have delayed dropping an mm in context_switch(). If
2626 * so, we finish that here outside of the runqueue lock. (Doing it
2627 * with the lock held can cause deadlocks; see schedule() for
2630 * The context switch have flipped the stack from under us and restored the
2631 * local variables which were saved when this task called schedule() in the
2632 * past. prev == current is still correct but we need to recalculate this_rq
2633 * because prev may have moved to another CPU.
2635 static struct rq *finish_task_switch(struct task_struct *prev)
2636 __releases(rq->lock)
2638 struct rq *rq = this_rq();
2639 struct mm_struct *mm = rq->prev_mm;
2643 * The previous task will have left us with a preempt_count of 2
2644 * because it left us after:
2647 * preempt_disable(); // 1
2649 * raw_spin_lock_irq(&rq->lock) // 2
2651 * Also, see FORK_PREEMPT_COUNT.
2653 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2654 "corrupted preempt_count: %s/%d/0x%x\n",
2655 current->comm, current->pid, preempt_count()))
2656 preempt_count_set(FORK_PREEMPT_COUNT);
2661 * A task struct has one reference for the use as "current".
2662 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2663 * schedule one last time. The schedule call will never return, and
2664 * the scheduled task must drop that reference.
2666 * We must observe prev->state before clearing prev->on_cpu (in
2667 * finish_task), otherwise a concurrent wakeup can get prev
2668 * running on another CPU and we could rave with its RUNNING -> DEAD
2669 * transition, resulting in a double drop.
2671 prev_state = prev->state;
2672 vtime_task_switch(prev);
2673 perf_event_task_sched_in(prev, current);
2675 finish_lock_switch(rq);
2676 finish_arch_post_lock_switch();
2677 kcov_finish_switch(current);
2679 fire_sched_in_preempt_notifiers(current);
2681 * When switching through a kernel thread, the loop in
2682 * membarrier_{private,global}_expedited() may have observed that
2683 * kernel thread and not issued an IPI. It is therefore possible to
2684 * schedule between user->kernel->user threads without passing though
2685 * switch_mm(). Membarrier requires a barrier after storing to
2686 * rq->curr, before returning to userspace, so provide them here:
2688 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
2689 * provided by mmdrop(),
2690 * - a sync_core for SYNC_CORE.
2693 membarrier_mm_sync_core_before_usermode(mm);
2696 if (unlikely(prev_state == TASK_DEAD)) {
2697 if (prev->sched_class->task_dead)
2698 prev->sched_class->task_dead(prev);
2701 * Remove function-return probe instances associated with this
2702 * task and put them back on the free list.
2704 kprobe_flush_task(prev);
2706 /* Task is done with its stack. */
2707 put_task_stack(prev);
2709 put_task_struct(prev);
2712 tick_nohz_task_switch();
2718 /* rq->lock is NOT held, but preemption is disabled */
2719 static void __balance_callback(struct rq *rq)
2721 struct callback_head *head, *next;
2722 void (*func)(struct rq *rq);
2723 unsigned long flags;
2725 raw_spin_lock_irqsave(&rq->lock, flags);
2726 head = rq->balance_callback;
2727 rq->balance_callback = NULL;
2729 func = (void (*)(struct rq *))head->func;
2736 raw_spin_unlock_irqrestore(&rq->lock, flags);
2739 static inline void balance_callback(struct rq *rq)
2741 if (unlikely(rq->balance_callback))
2742 __balance_callback(rq);
2747 static inline void balance_callback(struct rq *rq)
2754 * schedule_tail - first thing a freshly forked thread must call.
2755 * @prev: the thread we just switched away from.
2757 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2758 __releases(rq->lock)
2763 * New tasks start with FORK_PREEMPT_COUNT, see there and
2764 * finish_task_switch() for details.
2766 * finish_task_switch() will drop rq->lock() and lower preempt_count
2767 * and the preempt_enable() will end up enabling preemption (on
2768 * PREEMPT_COUNT kernels).
2771 rq = finish_task_switch(prev);
2772 balance_callback(rq);
2775 if (current->set_child_tid)
2776 put_user(task_pid_vnr(current), current->set_child_tid);
2780 * context_switch - switch to the new MM and the new thread's register state.
2782 static __always_inline struct rq *
2783 context_switch(struct rq *rq, struct task_struct *prev,
2784 struct task_struct *next, struct rq_flags *rf)
2786 struct mm_struct *mm, *oldmm;
2788 prepare_task_switch(rq, prev, next);
2791 oldmm = prev->active_mm;
2793 * For paravirt, this is coupled with an exit in switch_to to
2794 * combine the page table reload and the switch backend into
2797 arch_start_context_switch(prev);
2800 * If mm is non-NULL, we pass through switch_mm(). If mm is
2801 * NULL, we will pass through mmdrop() in finish_task_switch().
2802 * Both of these contain the full memory barrier required by
2803 * membarrier after storing to rq->curr, before returning to
2807 next->active_mm = oldmm;
2809 enter_lazy_tlb(oldmm, next);
2811 switch_mm_irqs_off(oldmm, mm, next);
2814 prev->active_mm = NULL;
2815 rq->prev_mm = oldmm;
2818 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2820 prepare_lock_switch(rq, next, rf);
2822 /* Here we just switch the register state and the stack. */
2823 switch_to(prev, next, prev);
2826 return finish_task_switch(prev);
2830 * nr_running and nr_context_switches:
2832 * externally visible scheduler statistics: current number of runnable
2833 * threads, total number of context switches performed since bootup.
2835 unsigned long nr_running(void)
2837 unsigned long i, sum = 0;
2839 for_each_online_cpu(i)
2840 sum += cpu_rq(i)->nr_running;
2846 * Check if only the current task is running on the CPU.
2848 * Caution: this function does not check that the caller has disabled
2849 * preemption, thus the result might have a time-of-check-to-time-of-use
2850 * race. The caller is responsible to use it correctly, for example:
2852 * - from a non-preemptable section (of course)
2854 * - from a thread that is bound to a single CPU
2856 * - in a loop with very short iterations (e.g. a polling loop)
2858 bool single_task_running(void)
2860 return raw_rq()->nr_running == 1;
2862 EXPORT_SYMBOL(single_task_running);
2864 unsigned long long nr_context_switches(void)
2867 unsigned long long sum = 0;
2869 for_each_possible_cpu(i)
2870 sum += cpu_rq(i)->nr_switches;
2876 * IO-wait accounting, and how its mostly bollocks (on SMP).
2878 * The idea behind IO-wait account is to account the idle time that we could
2879 * have spend running if it were not for IO. That is, if we were to improve the
2880 * storage performance, we'd have a proportional reduction in IO-wait time.
2882 * This all works nicely on UP, where, when a task blocks on IO, we account
2883 * idle time as IO-wait, because if the storage were faster, it could've been
2884 * running and we'd not be idle.
2886 * This has been extended to SMP, by doing the same for each CPU. This however
2889 * Imagine for instance the case where two tasks block on one CPU, only the one
2890 * CPU will have IO-wait accounted, while the other has regular idle. Even
2891 * though, if the storage were faster, both could've ran at the same time,
2892 * utilising both CPUs.
2894 * This means, that when looking globally, the current IO-wait accounting on
2895 * SMP is a lower bound, by reason of under accounting.
2897 * Worse, since the numbers are provided per CPU, they are sometimes
2898 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2899 * associated with any one particular CPU, it can wake to another CPU than it
2900 * blocked on. This means the per CPU IO-wait number is meaningless.
2902 * Task CPU affinities can make all that even more 'interesting'.
2905 unsigned long nr_iowait(void)
2907 unsigned long i, sum = 0;
2909 for_each_possible_cpu(i)
2910 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2916 * Consumers of these two interfaces, like for example the cpufreq menu
2917 * governor are using nonsensical data. Boosting frequency for a CPU that has
2918 * IO-wait which might not even end up running the task when it does become
2922 unsigned long nr_iowait_cpu(int cpu)
2924 struct rq *this = cpu_rq(cpu);
2925 return atomic_read(&this->nr_iowait);
2928 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2930 struct rq *rq = this_rq();
2931 *nr_waiters = atomic_read(&rq->nr_iowait);
2932 *load = rq->load.weight;
2938 * sched_exec - execve() is a valuable balancing opportunity, because at
2939 * this point the task has the smallest effective memory and cache footprint.
2941 void sched_exec(void)
2943 struct task_struct *p = current;
2944 unsigned long flags;
2947 raw_spin_lock_irqsave(&p->pi_lock, flags);
2948 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2949 if (dest_cpu == smp_processor_id())
2952 if (likely(cpu_active(dest_cpu))) {
2953 struct migration_arg arg = { p, dest_cpu };
2955 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2956 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2960 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2965 DEFINE_PER_CPU(struct kernel_stat, kstat);
2966 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2968 EXPORT_PER_CPU_SYMBOL(kstat);
2969 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2972 * The function fair_sched_class.update_curr accesses the struct curr
2973 * and its field curr->exec_start; when called from task_sched_runtime(),
2974 * we observe a high rate of cache misses in practice.
2975 * Prefetching this data results in improved performance.
2977 static inline void prefetch_curr_exec_start(struct task_struct *p)
2979 #ifdef CONFIG_FAIR_GROUP_SCHED
2980 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
2982 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
2985 prefetch(&curr->exec_start);
2989 * Return accounted runtime for the task.
2990 * In case the task is currently running, return the runtime plus current's
2991 * pending runtime that have not been accounted yet.
2993 unsigned long long task_sched_runtime(struct task_struct *p)
2999 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3001 * 64-bit doesn't need locks to atomically read a 64-bit value.
3002 * So we have a optimization chance when the task's delta_exec is 0.
3003 * Reading ->on_cpu is racy, but this is ok.
3005 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3006 * If we race with it entering CPU, unaccounted time is 0. This is
3007 * indistinguishable from the read occurring a few cycles earlier.
3008 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3009 * been accounted, so we're correct here as well.
3011 if (!p->on_cpu || !task_on_rq_queued(p))
3012 return p->se.sum_exec_runtime;
3015 rq = task_rq_lock(p, &rf);
3017 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3018 * project cycles that may never be accounted to this
3019 * thread, breaking clock_gettime().
3021 if (task_current(rq, p) && task_on_rq_queued(p)) {
3022 prefetch_curr_exec_start(p);
3023 update_rq_clock(rq);
3024 p->sched_class->update_curr(rq);
3026 ns = p->se.sum_exec_runtime;
3027 task_rq_unlock(rq, p, &rf);
3033 * This function gets called by the timer code, with HZ frequency.
3034 * We call it with interrupts disabled.
3036 void scheduler_tick(void)
3038 int cpu = smp_processor_id();
3039 struct rq *rq = cpu_rq(cpu);
3040 struct task_struct *curr = rq->curr;
3047 update_rq_clock(rq);
3048 curr->sched_class->task_tick(rq, curr, 0);
3049 cpu_load_update_active(rq);
3050 calc_global_load_tick(rq);
3054 perf_event_task_tick();
3057 rq->idle_balance = idle_cpu(cpu);
3058 trigger_load_balance(rq);
3062 #ifdef CONFIG_NO_HZ_FULL
3066 struct delayed_work work;
3069 static struct tick_work __percpu *tick_work_cpu;
3071 static void sched_tick_remote(struct work_struct *work)
3073 struct delayed_work *dwork = to_delayed_work(work);
3074 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3075 int cpu = twork->cpu;
3076 struct rq *rq = cpu_rq(cpu);
3077 struct task_struct *curr;
3082 * Handle the tick only if it appears the remote CPU is running in full
3083 * dynticks mode. The check is racy by nature, but missing a tick or
3084 * having one too much is no big deal because the scheduler tick updates
3085 * statistics and checks timeslices in a time-independent way, regardless
3086 * of when exactly it is running.
3088 if (idle_cpu(cpu) || !tick_nohz_tick_stopped_cpu(cpu))
3091 rq_lock_irq(rq, &rf);
3093 if (is_idle_task(curr))
3096 update_rq_clock(rq);
3097 delta = rq_clock_task(rq) - curr->se.exec_start;
3100 * Make sure the next tick runs within a reasonable
3103 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3104 curr->sched_class->task_tick(rq, curr, 0);
3107 rq_unlock_irq(rq, &rf);
3111 * Run the remote tick once per second (1Hz). This arbitrary
3112 * frequency is large enough to avoid overload but short enough
3113 * to keep scheduler internal stats reasonably up to date.
3115 queue_delayed_work(system_unbound_wq, dwork, HZ);
3118 static void sched_tick_start(int cpu)
3120 struct tick_work *twork;
3122 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3125 WARN_ON_ONCE(!tick_work_cpu);
3127 twork = per_cpu_ptr(tick_work_cpu, cpu);
3129 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3130 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3133 #ifdef CONFIG_HOTPLUG_CPU
3134 static void sched_tick_stop(int cpu)
3136 struct tick_work *twork;
3138 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3141 WARN_ON_ONCE(!tick_work_cpu);
3143 twork = per_cpu_ptr(tick_work_cpu, cpu);
3144 cancel_delayed_work_sync(&twork->work);
3146 #endif /* CONFIG_HOTPLUG_CPU */
3148 int __init sched_tick_offload_init(void)
3150 tick_work_cpu = alloc_percpu(struct tick_work);
3151 BUG_ON(!tick_work_cpu);
3156 #else /* !CONFIG_NO_HZ_FULL */
3157 static inline void sched_tick_start(int cpu) { }
3158 static inline void sched_tick_stop(int cpu) { }
3161 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3162 defined(CONFIG_PREEMPT_TRACER))
3164 * If the value passed in is equal to the current preempt count
3165 * then we just disabled preemption. Start timing the latency.
3167 static inline void preempt_latency_start(int val)
3169 if (preempt_count() == val) {
3170 unsigned long ip = get_lock_parent_ip();
3171 #ifdef CONFIG_DEBUG_PREEMPT
3172 current->preempt_disable_ip = ip;
3174 trace_preempt_off(CALLER_ADDR0, ip);
3178 void preempt_count_add(int val)
3180 #ifdef CONFIG_DEBUG_PREEMPT
3184 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3187 __preempt_count_add(val);
3188 #ifdef CONFIG_DEBUG_PREEMPT
3190 * Spinlock count overflowing soon?
3192 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3195 preempt_latency_start(val);
3197 EXPORT_SYMBOL(preempt_count_add);
3198 NOKPROBE_SYMBOL(preempt_count_add);
3201 * If the value passed in equals to the current preempt count
3202 * then we just enabled preemption. Stop timing the latency.
3204 static inline void preempt_latency_stop(int val)
3206 if (preempt_count() == val)
3207 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3210 void preempt_count_sub(int val)
3212 #ifdef CONFIG_DEBUG_PREEMPT
3216 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3219 * Is the spinlock portion underflowing?
3221 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3222 !(preempt_count() & PREEMPT_MASK)))
3226 preempt_latency_stop(val);
3227 __preempt_count_sub(val);
3229 EXPORT_SYMBOL(preempt_count_sub);
3230 NOKPROBE_SYMBOL(preempt_count_sub);
3233 static inline void preempt_latency_start(int val) { }
3234 static inline void preempt_latency_stop(int val) { }
3237 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3239 #ifdef CONFIG_DEBUG_PREEMPT
3240 return p->preempt_disable_ip;
3247 * Print scheduling while atomic bug:
3249 static noinline void __schedule_bug(struct task_struct *prev)
3251 /* Save this before calling printk(), since that will clobber it */
3252 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3254 if (oops_in_progress)
3257 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3258 prev->comm, prev->pid, preempt_count());
3260 debug_show_held_locks(prev);
3262 if (irqs_disabled())
3263 print_irqtrace_events(prev);
3264 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3265 && in_atomic_preempt_off()) {
3266 pr_err("Preemption disabled at:");
3267 print_ip_sym(preempt_disable_ip);
3271 panic("scheduling while atomic\n");
3274 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3278 * Various schedule()-time debugging checks and statistics:
3280 static inline void schedule_debug(struct task_struct *prev)
3282 #ifdef CONFIG_SCHED_STACK_END_CHECK
3283 if (task_stack_end_corrupted(prev))
3284 panic("corrupted stack end detected inside scheduler\n");
3287 if (unlikely(in_atomic_preempt_off())) {
3288 __schedule_bug(prev);
3289 preempt_count_set(PREEMPT_DISABLED);
3293 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3295 schedstat_inc(this_rq()->sched_count);
3299 * Pick up the highest-prio task:
3301 static inline struct task_struct *
3302 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3304 const struct sched_class *class;
3305 struct task_struct *p;
3308 * Optimization: we know that if all tasks are in the fair class we can
3309 * call that function directly, but only if the @prev task wasn't of a
3310 * higher scheduling class, because otherwise those loose the
3311 * opportunity to pull in more work from other CPUs.
3313 if (likely((prev->sched_class == &idle_sched_class ||
3314 prev->sched_class == &fair_sched_class) &&
3315 rq->nr_running == rq->cfs.h_nr_running)) {
3317 p = fair_sched_class.pick_next_task(rq, prev, rf);
3318 if (unlikely(p == RETRY_TASK))
3321 /* Assumes fair_sched_class->next == idle_sched_class */
3323 p = idle_sched_class.pick_next_task(rq, prev, rf);
3329 for_each_class(class) {
3330 p = class->pick_next_task(rq, prev, rf);
3332 if (unlikely(p == RETRY_TASK))
3338 /* The idle class should always have a runnable task: */
3343 * __schedule() is the main scheduler function.
3345 * The main means of driving the scheduler and thus entering this function are:
3347 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3349 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3350 * paths. For example, see arch/x86/entry_64.S.
3352 * To drive preemption between tasks, the scheduler sets the flag in timer
3353 * interrupt handler scheduler_tick().
3355 * 3. Wakeups don't really cause entry into schedule(). They add a
3356 * task to the run-queue and that's it.
3358 * Now, if the new task added to the run-queue preempts the current
3359 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3360 * called on the nearest possible occasion:
3362 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3364 * - in syscall or exception context, at the next outmost
3365 * preempt_enable(). (this might be as soon as the wake_up()'s
3368 * - in IRQ context, return from interrupt-handler to
3369 * preemptible context
3371 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3374 * - cond_resched() call
3375 * - explicit schedule() call
3376 * - return from syscall or exception to user-space
3377 * - return from interrupt-handler to user-space
3379 * WARNING: must be called with preemption disabled!
3381 static void __sched notrace __schedule(bool preempt)
3383 struct task_struct *prev, *next;
3384 unsigned long *switch_count;
3389 cpu = smp_processor_id();
3393 schedule_debug(prev);
3395 if (sched_feat(HRTICK))
3398 local_irq_disable();
3399 rcu_note_context_switch(preempt);
3402 * Make sure that signal_pending_state()->signal_pending() below
3403 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3404 * done by the caller to avoid the race with signal_wake_up().
3406 * The membarrier system call requires a full memory barrier
3407 * after coming from user-space, before storing to rq->curr.
3410 smp_mb__after_spinlock();
3412 /* Promote REQ to ACT */
3413 rq->clock_update_flags <<= 1;
3414 update_rq_clock(rq);
3416 switch_count = &prev->nivcsw;
3417 if (!preempt && prev->state) {
3418 if (unlikely(signal_pending_state(prev->state, prev))) {
3419 prev->state = TASK_RUNNING;
3421 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3424 if (prev->in_iowait) {
3425 atomic_inc(&rq->nr_iowait);
3426 delayacct_blkio_start();
3430 * If a worker went to sleep, notify and ask workqueue
3431 * whether it wants to wake up a task to maintain
3434 if (prev->flags & PF_WQ_WORKER) {
3435 struct task_struct *to_wakeup;
3437 to_wakeup = wq_worker_sleeping(prev);
3439 try_to_wake_up_local(to_wakeup, &rf);
3442 switch_count = &prev->nvcsw;
3445 next = pick_next_task(rq, prev, &rf);
3446 clear_tsk_need_resched(prev);
3447 clear_preempt_need_resched();
3449 if (likely(prev != next)) {
3453 * The membarrier system call requires each architecture
3454 * to have a full memory barrier after updating
3455 * rq->curr, before returning to user-space.
3457 * Here are the schemes providing that barrier on the
3458 * various architectures:
3459 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3460 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3461 * - finish_lock_switch() for weakly-ordered
3462 * architectures where spin_unlock is a full barrier,
3463 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3464 * is a RELEASE barrier),
3468 trace_sched_switch(preempt, prev, next);
3470 /* Also unlocks the rq: */
3471 rq = context_switch(rq, prev, next, &rf);
3473 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3474 rq_unlock_irq(rq, &rf);
3477 balance_callback(rq);
3480 void __noreturn do_task_dead(void)
3482 /* Causes final put_task_struct in finish_task_switch(): */
3483 set_special_state(TASK_DEAD);
3485 /* Tell freezer to ignore us: */
3486 current->flags |= PF_NOFREEZE;
3491 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3496 static inline void sched_submit_work(struct task_struct *tsk)
3498 if (!tsk->state || tsk_is_pi_blocked(tsk))
3501 * If we are going to sleep and we have plugged IO queued,
3502 * make sure to submit it to avoid deadlocks.
3504 if (blk_needs_flush_plug(tsk))
3505 blk_schedule_flush_plug(tsk);
3508 asmlinkage __visible void __sched schedule(void)
3510 struct task_struct *tsk = current;
3512 sched_submit_work(tsk);
3516 sched_preempt_enable_no_resched();
3517 } while (need_resched());
3519 EXPORT_SYMBOL(schedule);
3522 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3523 * state (have scheduled out non-voluntarily) by making sure that all
3524 * tasks have either left the run queue or have gone into user space.
3525 * As idle tasks do not do either, they must not ever be preempted
3526 * (schedule out non-voluntarily).
3528 * schedule_idle() is similar to schedule_preempt_disable() except that it
3529 * never enables preemption because it does not call sched_submit_work().
3531 void __sched schedule_idle(void)
3534 * As this skips calling sched_submit_work(), which the idle task does
3535 * regardless because that function is a nop when the task is in a
3536 * TASK_RUNNING state, make sure this isn't used someplace that the
3537 * current task can be in any other state. Note, idle is always in the
3538 * TASK_RUNNING state.
3540 WARN_ON_ONCE(current->state);
3543 } while (need_resched());
3546 #ifdef CONFIG_CONTEXT_TRACKING
3547 asmlinkage __visible void __sched schedule_user(void)
3550 * If we come here after a random call to set_need_resched(),
3551 * or we have been woken up remotely but the IPI has not yet arrived,
3552 * we haven't yet exited the RCU idle mode. Do it here manually until
3553 * we find a better solution.
3555 * NB: There are buggy callers of this function. Ideally we
3556 * should warn if prev_state != CONTEXT_USER, but that will trigger
3557 * too frequently to make sense yet.
3559 enum ctx_state prev_state = exception_enter();
3561 exception_exit(prev_state);
3566 * schedule_preempt_disabled - called with preemption disabled
3568 * Returns with preemption disabled. Note: preempt_count must be 1
3570 void __sched schedule_preempt_disabled(void)
3572 sched_preempt_enable_no_resched();
3577 static void __sched notrace preempt_schedule_common(void)
3581 * Because the function tracer can trace preempt_count_sub()
3582 * and it also uses preempt_enable/disable_notrace(), if
3583 * NEED_RESCHED is set, the preempt_enable_notrace() called
3584 * by the function tracer will call this function again and
3585 * cause infinite recursion.
3587 * Preemption must be disabled here before the function
3588 * tracer can trace. Break up preempt_disable() into two
3589 * calls. One to disable preemption without fear of being
3590 * traced. The other to still record the preemption latency,
3591 * which can also be traced by the function tracer.
3593 preempt_disable_notrace();
3594 preempt_latency_start(1);
3596 preempt_latency_stop(1);
3597 preempt_enable_no_resched_notrace();
3600 * Check again in case we missed a preemption opportunity
3601 * between schedule and now.
3603 } while (need_resched());
3606 #ifdef CONFIG_PREEMPT
3608 * this is the entry point to schedule() from in-kernel preemption
3609 * off of preempt_enable. Kernel preemptions off return from interrupt
3610 * occur there and call schedule directly.
3612 asmlinkage __visible void __sched notrace preempt_schedule(void)
3615 * If there is a non-zero preempt_count or interrupts are disabled,
3616 * we do not want to preempt the current task. Just return..
3618 if (likely(!preemptible()))
3621 preempt_schedule_common();
3623 NOKPROBE_SYMBOL(preempt_schedule);
3624 EXPORT_SYMBOL(preempt_schedule);
3627 * preempt_schedule_notrace - preempt_schedule called by tracing
3629 * The tracing infrastructure uses preempt_enable_notrace to prevent
3630 * recursion and tracing preempt enabling caused by the tracing
3631 * infrastructure itself. But as tracing can happen in areas coming
3632 * from userspace or just about to enter userspace, a preempt enable
3633 * can occur before user_exit() is called. This will cause the scheduler
3634 * to be called when the system is still in usermode.
3636 * To prevent this, the preempt_enable_notrace will use this function
3637 * instead of preempt_schedule() to exit user context if needed before
3638 * calling the scheduler.
3640 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3642 enum ctx_state prev_ctx;
3644 if (likely(!preemptible()))
3649 * Because the function tracer can trace preempt_count_sub()
3650 * and it also uses preempt_enable/disable_notrace(), if
3651 * NEED_RESCHED is set, the preempt_enable_notrace() called
3652 * by the function tracer will call this function again and
3653 * cause infinite recursion.
3655 * Preemption must be disabled here before the function
3656 * tracer can trace. Break up preempt_disable() into two
3657 * calls. One to disable preemption without fear of being
3658 * traced. The other to still record the preemption latency,
3659 * which can also be traced by the function tracer.
3661 preempt_disable_notrace();
3662 preempt_latency_start(1);
3664 * Needs preempt disabled in case user_exit() is traced
3665 * and the tracer calls preempt_enable_notrace() causing
3666 * an infinite recursion.
3668 prev_ctx = exception_enter();
3670 exception_exit(prev_ctx);
3672 preempt_latency_stop(1);
3673 preempt_enable_no_resched_notrace();
3674 } while (need_resched());
3676 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3678 #endif /* CONFIG_PREEMPT */
3681 * this is the entry point to schedule() from kernel preemption
3682 * off of irq context.
3683 * Note, that this is called and return with irqs disabled. This will
3684 * protect us against recursive calling from irq.
3686 asmlinkage __visible void __sched preempt_schedule_irq(void)
3688 enum ctx_state prev_state;
3690 /* Catch callers which need to be fixed */
3691 BUG_ON(preempt_count() || !irqs_disabled());
3693 prev_state = exception_enter();
3699 local_irq_disable();
3700 sched_preempt_enable_no_resched();
3701 } while (need_resched());
3703 exception_exit(prev_state);
3706 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
3709 return try_to_wake_up(curr->private, mode, wake_flags);
3711 EXPORT_SYMBOL(default_wake_function);
3713 #ifdef CONFIG_RT_MUTEXES
3715 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
3718 prio = min(prio, pi_task->prio);
3723 static inline int rt_effective_prio(struct task_struct *p, int prio)
3725 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3727 return __rt_effective_prio(pi_task, prio);
3731 * rt_mutex_setprio - set the current priority of a task
3733 * @pi_task: donor task
3735 * This function changes the 'effective' priority of a task. It does
3736 * not touch ->normal_prio like __setscheduler().
3738 * Used by the rt_mutex code to implement priority inheritance
3739 * logic. Call site only calls if the priority of the task changed.
3741 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
3743 int prio, oldprio, queued, running, queue_flag =
3744 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
3745 const struct sched_class *prev_class;
3749 /* XXX used to be waiter->prio, not waiter->task->prio */
3750 prio = __rt_effective_prio(pi_task, p->normal_prio);
3753 * If nothing changed; bail early.
3755 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
3758 rq = __task_rq_lock(p, &rf);
3759 update_rq_clock(rq);
3761 * Set under pi_lock && rq->lock, such that the value can be used under
3764 * Note that there is loads of tricky to make this pointer cache work
3765 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3766 * ensure a task is de-boosted (pi_task is set to NULL) before the
3767 * task is allowed to run again (and can exit). This ensures the pointer
3768 * points to a blocked task -- which guaratees the task is present.
3770 p->pi_top_task = pi_task;
3773 * For FIFO/RR we only need to set prio, if that matches we're done.
3775 if (prio == p->prio && !dl_prio(prio))
3779 * Idle task boosting is a nono in general. There is one
3780 * exception, when PREEMPT_RT and NOHZ is active:
3782 * The idle task calls get_next_timer_interrupt() and holds
3783 * the timer wheel base->lock on the CPU and another CPU wants
3784 * to access the timer (probably to cancel it). We can safely
3785 * ignore the boosting request, as the idle CPU runs this code
3786 * with interrupts disabled and will complete the lock
3787 * protected section without being interrupted. So there is no
3788 * real need to boost.
3790 if (unlikely(p == rq->idle)) {
3791 WARN_ON(p != rq->curr);
3792 WARN_ON(p->pi_blocked_on);
3796 trace_sched_pi_setprio(p, pi_task);
3799 if (oldprio == prio)
3800 queue_flag &= ~DEQUEUE_MOVE;
3802 prev_class = p->sched_class;
3803 queued = task_on_rq_queued(p);
3804 running = task_current(rq, p);
3806 dequeue_task(rq, p, queue_flag);
3808 put_prev_task(rq, p);
3811 * Boosting condition are:
3812 * 1. -rt task is running and holds mutex A
3813 * --> -dl task blocks on mutex A
3815 * 2. -dl task is running and holds mutex A
3816 * --> -dl task blocks on mutex A and could preempt the
3819 if (dl_prio(prio)) {
3820 if (!dl_prio(p->normal_prio) ||
3821 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3822 p->dl.dl_boosted = 1;
3823 queue_flag |= ENQUEUE_REPLENISH;
3825 p->dl.dl_boosted = 0;
3826 p->sched_class = &dl_sched_class;
3827 } else if (rt_prio(prio)) {
3828 if (dl_prio(oldprio))
3829 p->dl.dl_boosted = 0;
3831 queue_flag |= ENQUEUE_HEAD;
3832 p->sched_class = &rt_sched_class;
3834 if (dl_prio(oldprio))
3835 p->dl.dl_boosted = 0;
3836 if (rt_prio(oldprio))
3838 p->sched_class = &fair_sched_class;
3844 enqueue_task(rq, p, queue_flag);
3846 set_curr_task(rq, p);
3848 check_class_changed(rq, p, prev_class, oldprio);
3850 /* Avoid rq from going away on us: */
3852 __task_rq_unlock(rq, &rf);
3854 balance_callback(rq);
3858 static inline int rt_effective_prio(struct task_struct *p, int prio)
3864 void set_user_nice(struct task_struct *p, long nice)
3866 bool queued, running;
3867 int old_prio, delta;
3871 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3874 * We have to be careful, if called from sys_setpriority(),
3875 * the task might be in the middle of scheduling on another CPU.
3877 rq = task_rq_lock(p, &rf);
3878 update_rq_clock(rq);
3881 * The RT priorities are set via sched_setscheduler(), but we still
3882 * allow the 'normal' nice value to be set - but as expected
3883 * it wont have any effect on scheduling until the task is
3884 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3886 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3887 p->static_prio = NICE_TO_PRIO(nice);
3890 queued = task_on_rq_queued(p);
3891 running = task_current(rq, p);
3893 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
3895 put_prev_task(rq, p);
3897 p->static_prio = NICE_TO_PRIO(nice);
3898 set_load_weight(p, true);
3900 p->prio = effective_prio(p);
3901 delta = p->prio - old_prio;
3904 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
3906 * If the task increased its priority or is running and
3907 * lowered its priority, then reschedule its CPU:
3909 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3913 set_curr_task(rq, p);
3915 task_rq_unlock(rq, p, &rf);
3917 EXPORT_SYMBOL(set_user_nice);
3920 * can_nice - check if a task can reduce its nice value
3924 int can_nice(const struct task_struct *p, const int nice)
3926 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3927 int nice_rlim = nice_to_rlimit(nice);
3929 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3930 capable(CAP_SYS_NICE));
3933 #ifdef __ARCH_WANT_SYS_NICE
3936 * sys_nice - change the priority of the current process.
3937 * @increment: priority increment
3939 * sys_setpriority is a more generic, but much slower function that
3940 * does similar things.
3942 SYSCALL_DEFINE1(nice, int, increment)
3947 * Setpriority might change our priority at the same moment.
3948 * We don't have to worry. Conceptually one call occurs first
3949 * and we have a single winner.
3951 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3952 nice = task_nice(current) + increment;
3954 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3955 if (increment < 0 && !can_nice(current, nice))
3958 retval = security_task_setnice(current, nice);
3962 set_user_nice(current, nice);
3969 * task_prio - return the priority value of a given task.
3970 * @p: the task in question.
3972 * Return: The priority value as seen by users in /proc.
3973 * RT tasks are offset by -200. Normal tasks are centered
3974 * around 0, value goes from -16 to +15.
3976 int task_prio(const struct task_struct *p)
3978 return p->prio - MAX_RT_PRIO;
3982 * idle_cpu - is a given CPU idle currently?
3983 * @cpu: the processor in question.
3985 * Return: 1 if the CPU is currently idle. 0 otherwise.
3987 int idle_cpu(int cpu)
3989 struct rq *rq = cpu_rq(cpu);
3991 if (rq->curr != rq->idle)
3998 if (!llist_empty(&rq->wake_list))
4006 * available_idle_cpu - is a given CPU idle for enqueuing work.
4007 * @cpu: the CPU in question.
4009 * Return: 1 if the CPU is currently idle. 0 otherwise.
4011 int available_idle_cpu(int cpu)
4016 if (vcpu_is_preempted(cpu))
4023 * idle_task - return the idle task for a given CPU.
4024 * @cpu: the processor in question.
4026 * Return: The idle task for the CPU @cpu.
4028 struct task_struct *idle_task(int cpu)
4030 return cpu_rq(cpu)->idle;
4034 * find_process_by_pid - find a process with a matching PID value.
4035 * @pid: the pid in question.
4037 * The task of @pid, if found. %NULL otherwise.
4039 static struct task_struct *find_process_by_pid(pid_t pid)
4041 return pid ? find_task_by_vpid(pid) : current;
4045 * sched_setparam() passes in -1 for its policy, to let the functions
4046 * it calls know not to change it.
4048 #define SETPARAM_POLICY -1
4050 static void __setscheduler_params(struct task_struct *p,
4051 const struct sched_attr *attr)
4053 int policy = attr->sched_policy;
4055 if (policy == SETPARAM_POLICY)
4060 if (dl_policy(policy))
4061 __setparam_dl(p, attr);
4062 else if (fair_policy(policy))
4063 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4066 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4067 * !rt_policy. Always setting this ensures that things like
4068 * getparam()/getattr() don't report silly values for !rt tasks.
4070 p->rt_priority = attr->sched_priority;
4071 p->normal_prio = normal_prio(p);
4072 set_load_weight(p, true);
4075 /* Actually do priority change: must hold pi & rq lock. */
4076 static void __setscheduler(struct rq *rq, struct task_struct *p,
4077 const struct sched_attr *attr, bool keep_boost)
4079 __setscheduler_params(p, attr);
4082 * Keep a potential priority boosting if called from
4083 * sched_setscheduler().
4085 p->prio = normal_prio(p);
4087 p->prio = rt_effective_prio(p, p->prio);
4089 if (dl_prio(p->prio))
4090 p->sched_class = &dl_sched_class;
4091 else if (rt_prio(p->prio))
4092 p->sched_class = &rt_sched_class;
4094 p->sched_class = &fair_sched_class;
4098 * Check the target process has a UID that matches the current process's:
4100 static bool check_same_owner(struct task_struct *p)
4102 const struct cred *cred = current_cred(), *pcred;
4106 pcred = __task_cred(p);
4107 match = (uid_eq(cred->euid, pcred->euid) ||
4108 uid_eq(cred->euid, pcred->uid));
4113 static int __sched_setscheduler(struct task_struct *p,
4114 const struct sched_attr *attr,
4117 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4118 MAX_RT_PRIO - 1 - attr->sched_priority;
4119 int retval, oldprio, oldpolicy = -1, queued, running;
4120 int new_effective_prio, policy = attr->sched_policy;
4121 const struct sched_class *prev_class;
4124 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4127 /* The pi code expects interrupts enabled */
4128 BUG_ON(pi && in_interrupt());
4130 /* Double check policy once rq lock held: */
4132 reset_on_fork = p->sched_reset_on_fork;
4133 policy = oldpolicy = p->policy;
4135 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4137 if (!valid_policy(policy))
4141 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4145 * Valid priorities for SCHED_FIFO and SCHED_RR are
4146 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4147 * SCHED_BATCH and SCHED_IDLE is 0.
4149 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4150 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4152 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4153 (rt_policy(policy) != (attr->sched_priority != 0)))
4157 * Allow unprivileged RT tasks to decrease priority:
4159 if (user && !capable(CAP_SYS_NICE)) {
4160 if (fair_policy(policy)) {
4161 if (attr->sched_nice < task_nice(p) &&
4162 !can_nice(p, attr->sched_nice))
4166 if (rt_policy(policy)) {
4167 unsigned long rlim_rtprio =
4168 task_rlimit(p, RLIMIT_RTPRIO);
4170 /* Can't set/change the rt policy: */
4171 if (policy != p->policy && !rlim_rtprio)
4174 /* Can't increase priority: */
4175 if (attr->sched_priority > p->rt_priority &&
4176 attr->sched_priority > rlim_rtprio)
4181 * Can't set/change SCHED_DEADLINE policy at all for now
4182 * (safest behavior); in the future we would like to allow
4183 * unprivileged DL tasks to increase their relative deadline
4184 * or reduce their runtime (both ways reducing utilization)
4186 if (dl_policy(policy))
4190 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4191 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4193 if (idle_policy(p->policy) && !idle_policy(policy)) {
4194 if (!can_nice(p, task_nice(p)))
4198 /* Can't change other user's priorities: */
4199 if (!check_same_owner(p))
4202 /* Normal users shall not reset the sched_reset_on_fork flag: */
4203 if (p->sched_reset_on_fork && !reset_on_fork)
4208 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4211 retval = security_task_setscheduler(p);
4217 * Make sure no PI-waiters arrive (or leave) while we are
4218 * changing the priority of the task:
4220 * To be able to change p->policy safely, the appropriate
4221 * runqueue lock must be held.
4223 rq = task_rq_lock(p, &rf);
4224 update_rq_clock(rq);
4227 * Changing the policy of the stop threads its a very bad idea:
4229 if (p == rq->stop) {
4230 task_rq_unlock(rq, p, &rf);
4235 * If not changing anything there's no need to proceed further,
4236 * but store a possible modification of reset_on_fork.
4238 if (unlikely(policy == p->policy)) {
4239 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4241 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4243 if (dl_policy(policy) && dl_param_changed(p, attr))
4246 p->sched_reset_on_fork = reset_on_fork;
4247 task_rq_unlock(rq, p, &rf);
4253 #ifdef CONFIG_RT_GROUP_SCHED
4255 * Do not allow realtime tasks into groups that have no runtime
4258 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4259 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4260 !task_group_is_autogroup(task_group(p))) {
4261 task_rq_unlock(rq, p, &rf);
4266 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4267 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4268 cpumask_t *span = rq->rd->span;
4271 * Don't allow tasks with an affinity mask smaller than
4272 * the entire root_domain to become SCHED_DEADLINE. We
4273 * will also fail if there's no bandwidth available.
4275 if (!cpumask_subset(span, &p->cpus_allowed) ||
4276 rq->rd->dl_bw.bw == 0) {
4277 task_rq_unlock(rq, p, &rf);
4284 /* Re-check policy now with rq lock held: */
4285 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4286 policy = oldpolicy = -1;
4287 task_rq_unlock(rq, p, &rf);
4292 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4293 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4296 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4297 task_rq_unlock(rq, p, &rf);
4301 p->sched_reset_on_fork = reset_on_fork;
4306 * Take priority boosted tasks into account. If the new
4307 * effective priority is unchanged, we just store the new
4308 * normal parameters and do not touch the scheduler class and
4309 * the runqueue. This will be done when the task deboost
4312 new_effective_prio = rt_effective_prio(p, newprio);
4313 if (new_effective_prio == oldprio)
4314 queue_flags &= ~DEQUEUE_MOVE;
4317 queued = task_on_rq_queued(p);
4318 running = task_current(rq, p);
4320 dequeue_task(rq, p, queue_flags);
4322 put_prev_task(rq, p);
4324 prev_class = p->sched_class;
4325 __setscheduler(rq, p, attr, pi);
4329 * We enqueue to tail when the priority of a task is
4330 * increased (user space view).
4332 if (oldprio < p->prio)
4333 queue_flags |= ENQUEUE_HEAD;
4335 enqueue_task(rq, p, queue_flags);
4338 set_curr_task(rq, p);
4340 check_class_changed(rq, p, prev_class, oldprio);
4342 /* Avoid rq from going away on us: */
4344 task_rq_unlock(rq, p, &rf);
4347 rt_mutex_adjust_pi(p);
4349 /* Run balance callbacks after we've adjusted the PI chain: */
4350 balance_callback(rq);
4356 static int _sched_setscheduler(struct task_struct *p, int policy,
4357 const struct sched_param *param, bool check)
4359 struct sched_attr attr = {
4360 .sched_policy = policy,
4361 .sched_priority = param->sched_priority,
4362 .sched_nice = PRIO_TO_NICE(p->static_prio),
4365 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4366 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4367 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4368 policy &= ~SCHED_RESET_ON_FORK;
4369 attr.sched_policy = policy;
4372 return __sched_setscheduler(p, &attr, check, true);
4375 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4376 * @p: the task in question.
4377 * @policy: new policy.
4378 * @param: structure containing the new RT priority.
4380 * Return: 0 on success. An error code otherwise.
4382 * NOTE that the task may be already dead.
4384 int sched_setscheduler(struct task_struct *p, int policy,
4385 const struct sched_param *param)
4387 return _sched_setscheduler(p, policy, param, true);
4389 EXPORT_SYMBOL_GPL(sched_setscheduler);
4391 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4393 return __sched_setscheduler(p, attr, true, true);
4395 EXPORT_SYMBOL_GPL(sched_setattr);
4397 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
4399 return __sched_setscheduler(p, attr, false, true);
4403 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4404 * @p: the task in question.
4405 * @policy: new policy.
4406 * @param: structure containing the new RT priority.
4408 * Just like sched_setscheduler, only don't bother checking if the
4409 * current context has permission. For example, this is needed in
4410 * stop_machine(): we create temporary high priority worker threads,
4411 * but our caller might not have that capability.
4413 * Return: 0 on success. An error code otherwise.
4415 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4416 const struct sched_param *param)
4418 return _sched_setscheduler(p, policy, param, false);
4420 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4423 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4425 struct sched_param lparam;
4426 struct task_struct *p;
4429 if (!param || pid < 0)
4431 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4436 p = find_process_by_pid(pid);
4438 retval = sched_setscheduler(p, policy, &lparam);
4445 * Mimics kernel/events/core.c perf_copy_attr().
4447 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4452 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4455 /* Zero the full structure, so that a short copy will be nice: */
4456 memset(attr, 0, sizeof(*attr));
4458 ret = get_user(size, &uattr->size);
4462 /* Bail out on silly large: */
4463 if (size > PAGE_SIZE)
4466 /* ABI compatibility quirk: */
4468 size = SCHED_ATTR_SIZE_VER0;
4470 if (size < SCHED_ATTR_SIZE_VER0)
4474 * If we're handed a bigger struct than we know of,
4475 * ensure all the unknown bits are 0 - i.e. new
4476 * user-space does not rely on any kernel feature
4477 * extensions we dont know about yet.
4479 if (size > sizeof(*attr)) {
4480 unsigned char __user *addr;
4481 unsigned char __user *end;
4484 addr = (void __user *)uattr + sizeof(*attr);
4485 end = (void __user *)uattr + size;
4487 for (; addr < end; addr++) {
4488 ret = get_user(val, addr);
4494 size = sizeof(*attr);
4497 ret = copy_from_user(attr, uattr, size);
4502 * XXX: Do we want to be lenient like existing syscalls; or do we want
4503 * to be strict and return an error on out-of-bounds values?
4505 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4510 put_user(sizeof(*attr), &uattr->size);
4515 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4516 * @pid: the pid in question.
4517 * @policy: new policy.
4518 * @param: structure containing the new RT priority.
4520 * Return: 0 on success. An error code otherwise.
4522 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4527 return do_sched_setscheduler(pid, policy, param);
4531 * sys_sched_setparam - set/change the RT priority of a thread
4532 * @pid: the pid in question.
4533 * @param: structure containing the new RT priority.
4535 * Return: 0 on success. An error code otherwise.
4537 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4539 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4543 * sys_sched_setattr - same as above, but with extended sched_attr
4544 * @pid: the pid in question.
4545 * @uattr: structure containing the extended parameters.
4546 * @flags: for future extension.
4548 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4549 unsigned int, flags)
4551 struct sched_attr attr;
4552 struct task_struct *p;
4555 if (!uattr || pid < 0 || flags)
4558 retval = sched_copy_attr(uattr, &attr);
4562 if ((int)attr.sched_policy < 0)
4567 p = find_process_by_pid(pid);
4569 retval = sched_setattr(p, &attr);
4576 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4577 * @pid: the pid in question.
4579 * Return: On success, the policy of the thread. Otherwise, a negative error
4582 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4584 struct task_struct *p;
4592 p = find_process_by_pid(pid);
4594 retval = security_task_getscheduler(p);
4597 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4604 * sys_sched_getparam - get the RT priority of a thread
4605 * @pid: the pid in question.
4606 * @param: structure containing the RT priority.
4608 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4611 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4613 struct sched_param lp = { .sched_priority = 0 };
4614 struct task_struct *p;
4617 if (!param || pid < 0)
4621 p = find_process_by_pid(pid);
4626 retval = security_task_getscheduler(p);
4630 if (task_has_rt_policy(p))
4631 lp.sched_priority = p->rt_priority;
4635 * This one might sleep, we cannot do it with a spinlock held ...
4637 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4646 static int sched_read_attr(struct sched_attr __user *uattr,
4647 struct sched_attr *attr,
4652 if (!access_ok(VERIFY_WRITE, uattr, usize))
4656 * If we're handed a smaller struct than we know of,
4657 * ensure all the unknown bits are 0 - i.e. old
4658 * user-space does not get uncomplete information.
4660 if (usize < sizeof(*attr)) {
4661 unsigned char *addr;
4664 addr = (void *)attr + usize;
4665 end = (void *)attr + sizeof(*attr);
4667 for (; addr < end; addr++) {
4675 ret = copy_to_user(uattr, attr, attr->size);
4683 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4684 * @pid: the pid in question.
4685 * @uattr: structure containing the extended parameters.
4686 * @size: sizeof(attr) for fwd/bwd comp.
4687 * @flags: for future extension.
4689 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4690 unsigned int, size, unsigned int, flags)
4692 struct sched_attr attr = {
4693 .size = sizeof(struct sched_attr),
4695 struct task_struct *p;
4698 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4699 size < SCHED_ATTR_SIZE_VER0 || flags)
4703 p = find_process_by_pid(pid);
4708 retval = security_task_getscheduler(p);
4712 attr.sched_policy = p->policy;
4713 if (p->sched_reset_on_fork)
4714 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4715 if (task_has_dl_policy(p))
4716 __getparam_dl(p, &attr);
4717 else if (task_has_rt_policy(p))
4718 attr.sched_priority = p->rt_priority;
4720 attr.sched_nice = task_nice(p);
4724 retval = sched_read_attr(uattr, &attr, size);
4732 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4734 cpumask_var_t cpus_allowed, new_mask;
4735 struct task_struct *p;
4740 p = find_process_by_pid(pid);
4746 /* Prevent p going away */
4750 if (p->flags & PF_NO_SETAFFINITY) {
4754 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4758 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4760 goto out_free_cpus_allowed;
4763 if (!check_same_owner(p)) {
4765 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4767 goto out_free_new_mask;
4772 retval = security_task_setscheduler(p);
4774 goto out_free_new_mask;
4777 cpuset_cpus_allowed(p, cpus_allowed);
4778 cpumask_and(new_mask, in_mask, cpus_allowed);
4781 * Since bandwidth control happens on root_domain basis,
4782 * if admission test is enabled, we only admit -deadline
4783 * tasks allowed to run on all the CPUs in the task's
4787 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4789 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4792 goto out_free_new_mask;
4798 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4801 cpuset_cpus_allowed(p, cpus_allowed);
4802 if (!cpumask_subset(new_mask, cpus_allowed)) {
4804 * We must have raced with a concurrent cpuset
4805 * update. Just reset the cpus_allowed to the
4806 * cpuset's cpus_allowed
4808 cpumask_copy(new_mask, cpus_allowed);
4813 free_cpumask_var(new_mask);
4814 out_free_cpus_allowed:
4815 free_cpumask_var(cpus_allowed);
4821 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4822 struct cpumask *new_mask)
4824 if (len < cpumask_size())
4825 cpumask_clear(new_mask);
4826 else if (len > cpumask_size())
4827 len = cpumask_size();
4829 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4833 * sys_sched_setaffinity - set the CPU affinity of a process
4834 * @pid: pid of the process
4835 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4836 * @user_mask_ptr: user-space pointer to the new CPU mask
4838 * Return: 0 on success. An error code otherwise.
4840 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4841 unsigned long __user *, user_mask_ptr)
4843 cpumask_var_t new_mask;
4846 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4849 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4851 retval = sched_setaffinity(pid, new_mask);
4852 free_cpumask_var(new_mask);
4856 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4858 struct task_struct *p;
4859 unsigned long flags;
4865 p = find_process_by_pid(pid);
4869 retval = security_task_getscheduler(p);
4873 raw_spin_lock_irqsave(&p->pi_lock, flags);
4874 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4875 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4884 * sys_sched_getaffinity - get the CPU affinity of a process
4885 * @pid: pid of the process
4886 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4887 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4889 * Return: size of CPU mask copied to user_mask_ptr on success. An
4890 * error code otherwise.
4892 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4893 unsigned long __user *, user_mask_ptr)
4898 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4900 if (len & (sizeof(unsigned long)-1))
4903 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4906 ret = sched_getaffinity(pid, mask);
4908 unsigned int retlen = min(len, cpumask_size());
4910 if (copy_to_user(user_mask_ptr, mask, retlen))
4915 free_cpumask_var(mask);
4921 * sys_sched_yield - yield the current processor to other threads.
4923 * This function yields the current CPU to other tasks. If there are no
4924 * other threads running on this CPU then this function will return.
4928 static void do_sched_yield(void)
4933 local_irq_disable();
4937 schedstat_inc(rq->yld_count);
4938 current->sched_class->yield_task(rq);
4941 * Since we are going to call schedule() anyway, there's
4942 * no need to preempt or enable interrupts:
4946 sched_preempt_enable_no_resched();
4951 SYSCALL_DEFINE0(sched_yield)
4957 #ifndef CONFIG_PREEMPT
4958 int __sched _cond_resched(void)
4960 if (should_resched(0)) {
4961 preempt_schedule_common();
4967 EXPORT_SYMBOL(_cond_resched);
4971 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4972 * call schedule, and on return reacquire the lock.
4974 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4975 * operations here to prevent schedule() from being called twice (once via
4976 * spin_unlock(), once by hand).
4978 int __cond_resched_lock(spinlock_t *lock)
4980 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4983 lockdep_assert_held(lock);
4985 if (spin_needbreak(lock) || resched) {
4988 preempt_schedule_common();
4996 EXPORT_SYMBOL(__cond_resched_lock);
4999 * yield - yield the current processor to other threads.
5001 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5003 * The scheduler is at all times free to pick the calling task as the most
5004 * eligible task to run, if removing the yield() call from your code breaks
5005 * it, its already broken.
5007 * Typical broken usage is:
5012 * where one assumes that yield() will let 'the other' process run that will
5013 * make event true. If the current task is a SCHED_FIFO task that will never
5014 * happen. Never use yield() as a progress guarantee!!
5016 * If you want to use yield() to wait for something, use wait_event().
5017 * If you want to use yield() to be 'nice' for others, use cond_resched().
5018 * If you still want to use yield(), do not!
5020 void __sched yield(void)
5022 set_current_state(TASK_RUNNING);
5025 EXPORT_SYMBOL(yield);
5028 * yield_to - yield the current processor to another thread in
5029 * your thread group, or accelerate that thread toward the
5030 * processor it's on.
5032 * @preempt: whether task preemption is allowed or not
5034 * It's the caller's job to ensure that the target task struct
5035 * can't go away on us before we can do any checks.
5038 * true (>0) if we indeed boosted the target task.
5039 * false (0) if we failed to boost the target.
5040 * -ESRCH if there's no task to yield to.
5042 int __sched yield_to(struct task_struct *p, bool preempt)
5044 struct task_struct *curr = current;
5045 struct rq *rq, *p_rq;
5046 unsigned long flags;
5049 local_irq_save(flags);
5055 * If we're the only runnable task on the rq and target rq also
5056 * has only one task, there's absolutely no point in yielding.
5058 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5063 double_rq_lock(rq, p_rq);
5064 if (task_rq(p) != p_rq) {
5065 double_rq_unlock(rq, p_rq);
5069 if (!curr->sched_class->yield_to_task)
5072 if (curr->sched_class != p->sched_class)
5075 if (task_running(p_rq, p) || p->state)
5078 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5080 schedstat_inc(rq->yld_count);
5082 * Make p's CPU reschedule; pick_next_entity takes care of
5085 if (preempt && rq != p_rq)
5090 double_rq_unlock(rq, p_rq);
5092 local_irq_restore(flags);
5099 EXPORT_SYMBOL_GPL(yield_to);
5101 int io_schedule_prepare(void)
5103 int old_iowait = current->in_iowait;
5105 current->in_iowait = 1;
5106 blk_schedule_flush_plug(current);
5111 void io_schedule_finish(int token)
5113 current->in_iowait = token;
5117 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5118 * that process accounting knows that this is a task in IO wait state.
5120 long __sched io_schedule_timeout(long timeout)
5125 token = io_schedule_prepare();
5126 ret = schedule_timeout(timeout);
5127 io_schedule_finish(token);
5131 EXPORT_SYMBOL(io_schedule_timeout);
5133 void io_schedule(void)
5137 token = io_schedule_prepare();
5139 io_schedule_finish(token);
5141 EXPORT_SYMBOL(io_schedule);
5144 * sys_sched_get_priority_max - return maximum RT priority.
5145 * @policy: scheduling class.
5147 * Return: On success, this syscall returns the maximum
5148 * rt_priority that can be used by a given scheduling class.
5149 * On failure, a negative error code is returned.
5151 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5158 ret = MAX_USER_RT_PRIO-1;
5160 case SCHED_DEADLINE:
5171 * sys_sched_get_priority_min - return minimum RT priority.
5172 * @policy: scheduling class.
5174 * Return: On success, this syscall returns the minimum
5175 * rt_priority that can be used by a given scheduling class.
5176 * On failure, a negative error code is returned.
5178 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5187 case SCHED_DEADLINE:
5196 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5198 struct task_struct *p;
5199 unsigned int time_slice;
5209 p = find_process_by_pid(pid);
5213 retval = security_task_getscheduler(p);
5217 rq = task_rq_lock(p, &rf);
5219 if (p->sched_class->get_rr_interval)
5220 time_slice = p->sched_class->get_rr_interval(rq, p);
5221 task_rq_unlock(rq, p, &rf);
5224 jiffies_to_timespec64(time_slice, t);
5233 * sys_sched_rr_get_interval - return the default timeslice of a process.
5234 * @pid: pid of the process.
5235 * @interval: userspace pointer to the timeslice value.
5237 * this syscall writes the default timeslice value of a given process
5238 * into the user-space timespec buffer. A value of '0' means infinity.
5240 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5243 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5244 struct timespec __user *, interval)
5246 struct timespec64 t;
5247 int retval = sched_rr_get_interval(pid, &t);
5250 retval = put_timespec64(&t, interval);
5255 #ifdef CONFIG_COMPAT
5256 COMPAT_SYSCALL_DEFINE2(sched_rr_get_interval,
5258 struct compat_timespec __user *, interval)
5260 struct timespec64 t;
5261 int retval = sched_rr_get_interval(pid, &t);
5264 retval = compat_put_timespec64(&t, interval);
5269 void sched_show_task(struct task_struct *p)
5271 unsigned long free = 0;
5274 if (!try_get_task_stack(p))
5277 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5279 if (p->state == TASK_RUNNING)
5280 printk(KERN_CONT " running task ");
5281 #ifdef CONFIG_DEBUG_STACK_USAGE
5282 free = stack_not_used(p);
5287 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5289 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5290 task_pid_nr(p), ppid,
5291 (unsigned long)task_thread_info(p)->flags);
5293 print_worker_info(KERN_INFO, p);
5294 show_stack(p, NULL);
5297 EXPORT_SYMBOL_GPL(sched_show_task);
5300 state_filter_match(unsigned long state_filter, struct task_struct *p)
5302 /* no filter, everything matches */
5306 /* filter, but doesn't match */
5307 if (!(p->state & state_filter))
5311 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5314 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5321 void show_state_filter(unsigned long state_filter)
5323 struct task_struct *g, *p;
5325 #if BITS_PER_LONG == 32
5327 " task PC stack pid father\n");
5330 " task PC stack pid father\n");
5333 for_each_process_thread(g, p) {
5335 * reset the NMI-timeout, listing all files on a slow
5336 * console might take a lot of time:
5337 * Also, reset softlockup watchdogs on all CPUs, because
5338 * another CPU might be blocked waiting for us to process
5341 touch_nmi_watchdog();
5342 touch_all_softlockup_watchdogs();
5343 if (state_filter_match(state_filter, p))
5347 #ifdef CONFIG_SCHED_DEBUG
5349 sysrq_sched_debug_show();
5353 * Only show locks if all tasks are dumped:
5356 debug_show_all_locks();
5360 * init_idle - set up an idle thread for a given CPU
5361 * @idle: task in question
5362 * @cpu: CPU the idle task belongs to
5364 * NOTE: this function does not set the idle thread's NEED_RESCHED
5365 * flag, to make booting more robust.
5367 void init_idle(struct task_struct *idle, int cpu)
5369 struct rq *rq = cpu_rq(cpu);
5370 unsigned long flags;
5372 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5373 raw_spin_lock(&rq->lock);
5375 __sched_fork(0, idle);
5376 idle->state = TASK_RUNNING;
5377 idle->se.exec_start = sched_clock();
5378 idle->flags |= PF_IDLE;
5380 kasan_unpoison_task_stack(idle);
5384 * Its possible that init_idle() gets called multiple times on a task,
5385 * in that case do_set_cpus_allowed() will not do the right thing.
5387 * And since this is boot we can forgo the serialization.
5389 set_cpus_allowed_common(idle, cpumask_of(cpu));
5392 * We're having a chicken and egg problem, even though we are
5393 * holding rq->lock, the CPU isn't yet set to this CPU so the
5394 * lockdep check in task_group() will fail.
5396 * Similar case to sched_fork(). / Alternatively we could
5397 * use task_rq_lock() here and obtain the other rq->lock.
5402 __set_task_cpu(idle, cpu);
5405 rq->curr = rq->idle = idle;
5406 idle->on_rq = TASK_ON_RQ_QUEUED;
5410 raw_spin_unlock(&rq->lock);
5411 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5413 /* Set the preempt count _outside_ the spinlocks! */
5414 init_idle_preempt_count(idle, cpu);
5417 * The idle tasks have their own, simple scheduling class:
5419 idle->sched_class = &idle_sched_class;
5420 ftrace_graph_init_idle_task(idle, cpu);
5421 vtime_init_idle(idle, cpu);
5423 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5429 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5430 const struct cpumask *trial)
5434 if (!cpumask_weight(cur))
5437 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5442 int task_can_attach(struct task_struct *p,
5443 const struct cpumask *cs_cpus_allowed)
5448 * Kthreads which disallow setaffinity shouldn't be moved
5449 * to a new cpuset; we don't want to change their CPU
5450 * affinity and isolating such threads by their set of
5451 * allowed nodes is unnecessary. Thus, cpusets are not
5452 * applicable for such threads. This prevents checking for
5453 * success of set_cpus_allowed_ptr() on all attached tasks
5454 * before cpus_allowed may be changed.
5456 if (p->flags & PF_NO_SETAFFINITY) {
5461 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5463 ret = dl_task_can_attach(p, cs_cpus_allowed);
5469 bool sched_smp_initialized __read_mostly;
5471 #ifdef CONFIG_NUMA_BALANCING
5472 /* Migrate current task p to target_cpu */
5473 int migrate_task_to(struct task_struct *p, int target_cpu)
5475 struct migration_arg arg = { p, target_cpu };
5476 int curr_cpu = task_cpu(p);
5478 if (curr_cpu == target_cpu)
5481 if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5484 /* TODO: This is not properly updating schedstats */
5486 trace_sched_move_numa(p, curr_cpu, target_cpu);
5487 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5491 * Requeue a task on a given node and accurately track the number of NUMA
5492 * tasks on the runqueues
5494 void sched_setnuma(struct task_struct *p, int nid)
5496 bool queued, running;
5500 rq = task_rq_lock(p, &rf);
5501 queued = task_on_rq_queued(p);
5502 running = task_current(rq, p);
5505 dequeue_task(rq, p, DEQUEUE_SAVE);
5507 put_prev_task(rq, p);
5509 p->numa_preferred_nid = nid;
5512 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5514 set_curr_task(rq, p);
5515 task_rq_unlock(rq, p, &rf);
5517 #endif /* CONFIG_NUMA_BALANCING */
5519 #ifdef CONFIG_HOTPLUG_CPU
5521 * Ensure that the idle task is using init_mm right before its CPU goes
5524 void idle_task_exit(void)
5526 struct mm_struct *mm = current->active_mm;
5528 BUG_ON(cpu_online(smp_processor_id()));
5530 if (mm != &init_mm) {
5531 switch_mm(mm, &init_mm, current);
5532 current->active_mm = &init_mm;
5533 finish_arch_post_lock_switch();
5539 * Since this CPU is going 'away' for a while, fold any nr_active delta
5540 * we might have. Assumes we're called after migrate_tasks() so that the
5541 * nr_active count is stable. We need to take the teardown thread which
5542 * is calling this into account, so we hand in adjust = 1 to the load
5545 * Also see the comment "Global load-average calculations".
5547 static void calc_load_migrate(struct rq *rq)
5549 long delta = calc_load_fold_active(rq, 1);
5551 atomic_long_add(delta, &calc_load_tasks);
5554 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5558 static const struct sched_class fake_sched_class = {
5559 .put_prev_task = put_prev_task_fake,
5562 static struct task_struct fake_task = {
5564 * Avoid pull_{rt,dl}_task()
5566 .prio = MAX_PRIO + 1,
5567 .sched_class = &fake_sched_class,
5571 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5572 * try_to_wake_up()->select_task_rq().
5574 * Called with rq->lock held even though we'er in stop_machine() and
5575 * there's no concurrency possible, we hold the required locks anyway
5576 * because of lock validation efforts.
5578 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5580 struct rq *rq = dead_rq;
5581 struct task_struct *next, *stop = rq->stop;
5582 struct rq_flags orf = *rf;
5586 * Fudge the rq selection such that the below task selection loop
5587 * doesn't get stuck on the currently eligible stop task.
5589 * We're currently inside stop_machine() and the rq is either stuck
5590 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5591 * either way we should never end up calling schedule() until we're
5597 * put_prev_task() and pick_next_task() sched
5598 * class method both need to have an up-to-date
5599 * value of rq->clock[_task]
5601 update_rq_clock(rq);
5605 * There's this thread running, bail when that's the only
5608 if (rq->nr_running == 1)
5612 * pick_next_task() assumes pinned rq->lock:
5614 next = pick_next_task(rq, &fake_task, rf);
5616 put_prev_task(rq, next);
5619 * Rules for changing task_struct::cpus_allowed are holding
5620 * both pi_lock and rq->lock, such that holding either
5621 * stabilizes the mask.
5623 * Drop rq->lock is not quite as disastrous as it usually is
5624 * because !cpu_active at this point, which means load-balance
5625 * will not interfere. Also, stop-machine.
5628 raw_spin_lock(&next->pi_lock);
5632 * Since we're inside stop-machine, _nothing_ should have
5633 * changed the task, WARN if weird stuff happened, because in
5634 * that case the above rq->lock drop is a fail too.
5636 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5637 raw_spin_unlock(&next->pi_lock);
5641 /* Find suitable destination for @next, with force if needed. */
5642 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5643 rq = __migrate_task(rq, rf, next, dest_cpu);
5644 if (rq != dead_rq) {
5650 raw_spin_unlock(&next->pi_lock);
5655 #endif /* CONFIG_HOTPLUG_CPU */
5657 void set_rq_online(struct rq *rq)
5660 const struct sched_class *class;
5662 cpumask_set_cpu(rq->cpu, rq->rd->online);
5665 for_each_class(class) {
5666 if (class->rq_online)
5667 class->rq_online(rq);
5672 void set_rq_offline(struct rq *rq)
5675 const struct sched_class *class;
5677 for_each_class(class) {
5678 if (class->rq_offline)
5679 class->rq_offline(rq);
5682 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5688 * used to mark begin/end of suspend/resume:
5690 static int num_cpus_frozen;
5693 * Update cpusets according to cpu_active mask. If cpusets are
5694 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5695 * around partition_sched_domains().
5697 * If we come here as part of a suspend/resume, don't touch cpusets because we
5698 * want to restore it back to its original state upon resume anyway.
5700 static void cpuset_cpu_active(void)
5702 if (cpuhp_tasks_frozen) {
5704 * num_cpus_frozen tracks how many CPUs are involved in suspend
5705 * resume sequence. As long as this is not the last online
5706 * operation in the resume sequence, just build a single sched
5707 * domain, ignoring cpusets.
5709 partition_sched_domains(1, NULL, NULL);
5710 if (--num_cpus_frozen)
5713 * This is the last CPU online operation. So fall through and
5714 * restore the original sched domains by considering the
5715 * cpuset configurations.
5717 cpuset_force_rebuild();
5719 cpuset_update_active_cpus();
5722 static int cpuset_cpu_inactive(unsigned int cpu)
5724 if (!cpuhp_tasks_frozen) {
5725 if (dl_cpu_busy(cpu))
5727 cpuset_update_active_cpus();
5730 partition_sched_domains(1, NULL, NULL);
5735 int sched_cpu_activate(unsigned int cpu)
5737 struct rq *rq = cpu_rq(cpu);
5740 set_cpu_active(cpu, true);
5742 if (sched_smp_initialized) {
5743 sched_domains_numa_masks_set(cpu);
5744 cpuset_cpu_active();
5748 * Put the rq online, if not already. This happens:
5750 * 1) In the early boot process, because we build the real domains
5751 * after all CPUs have been brought up.
5753 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5756 rq_lock_irqsave(rq, &rf);
5758 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5761 rq_unlock_irqrestore(rq, &rf);
5763 update_max_interval();
5768 int sched_cpu_deactivate(unsigned int cpu)
5772 set_cpu_active(cpu, false);
5774 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5775 * users of this state to go away such that all new such users will
5778 * Do sync before park smpboot threads to take care the rcu boost case.
5780 synchronize_rcu_mult(call_rcu, call_rcu_sched);
5782 if (!sched_smp_initialized)
5785 ret = cpuset_cpu_inactive(cpu);
5787 set_cpu_active(cpu, true);
5790 sched_domains_numa_masks_clear(cpu);
5794 static void sched_rq_cpu_starting(unsigned int cpu)
5796 struct rq *rq = cpu_rq(cpu);
5798 rq->calc_load_update = calc_load_update;
5799 update_max_interval();
5802 int sched_cpu_starting(unsigned int cpu)
5804 sched_rq_cpu_starting(cpu);
5805 sched_tick_start(cpu);
5809 #ifdef CONFIG_HOTPLUG_CPU
5810 int sched_cpu_dying(unsigned int cpu)
5812 struct rq *rq = cpu_rq(cpu);
5815 /* Handle pending wakeups and then migrate everything off */
5816 sched_ttwu_pending();
5817 sched_tick_stop(cpu);
5819 rq_lock_irqsave(rq, &rf);
5821 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5824 migrate_tasks(rq, &rf);
5825 BUG_ON(rq->nr_running != 1);
5826 rq_unlock_irqrestore(rq, &rf);
5828 calc_load_migrate(rq);
5829 update_max_interval();
5830 nohz_balance_exit_idle(rq);
5836 #ifdef CONFIG_SCHED_SMT
5837 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5839 static void sched_init_smt(void)
5842 * We've enumerated all CPUs and will assume that if any CPU
5843 * has SMT siblings, CPU0 will too.
5845 if (cpumask_weight(cpu_smt_mask(0)) > 1)
5846 static_branch_enable(&sched_smt_present);
5849 static inline void sched_init_smt(void) { }
5852 void __init sched_init_smp(void)
5857 * There's no userspace yet to cause hotplug operations; hence all the
5858 * CPU masks are stable and all blatant races in the below code cannot
5861 mutex_lock(&sched_domains_mutex);
5862 sched_init_domains(cpu_active_mask);
5863 mutex_unlock(&sched_domains_mutex);
5865 /* Move init over to a non-isolated CPU */
5866 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
5868 sched_init_granularity();
5870 init_sched_rt_class();
5871 init_sched_dl_class();
5875 sched_smp_initialized = true;
5878 static int __init migration_init(void)
5880 sched_rq_cpu_starting(smp_processor_id());
5883 early_initcall(migration_init);
5886 void __init sched_init_smp(void)
5888 sched_init_granularity();
5890 #endif /* CONFIG_SMP */
5892 int in_sched_functions(unsigned long addr)
5894 return in_lock_functions(addr) ||
5895 (addr >= (unsigned long)__sched_text_start
5896 && addr < (unsigned long)__sched_text_end);
5899 #ifdef CONFIG_CGROUP_SCHED
5901 * Default task group.
5902 * Every task in system belongs to this group at bootup.
5904 struct task_group root_task_group;
5905 LIST_HEAD(task_groups);
5907 /* Cacheline aligned slab cache for task_group */
5908 static struct kmem_cache *task_group_cache __read_mostly;
5911 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5912 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5914 void __init sched_init(void)
5917 unsigned long alloc_size = 0, ptr;
5921 #ifdef CONFIG_FAIR_GROUP_SCHED
5922 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5924 #ifdef CONFIG_RT_GROUP_SCHED
5925 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5928 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5930 #ifdef CONFIG_FAIR_GROUP_SCHED
5931 root_task_group.se = (struct sched_entity **)ptr;
5932 ptr += nr_cpu_ids * sizeof(void **);
5934 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5935 ptr += nr_cpu_ids * sizeof(void **);
5937 #endif /* CONFIG_FAIR_GROUP_SCHED */
5938 #ifdef CONFIG_RT_GROUP_SCHED
5939 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5940 ptr += nr_cpu_ids * sizeof(void **);
5942 root_task_group.rt_rq = (struct rt_rq **)ptr;
5943 ptr += nr_cpu_ids * sizeof(void **);
5945 #endif /* CONFIG_RT_GROUP_SCHED */
5947 #ifdef CONFIG_CPUMASK_OFFSTACK
5948 for_each_possible_cpu(i) {
5949 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5950 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5951 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
5952 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5954 #endif /* CONFIG_CPUMASK_OFFSTACK */
5956 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
5957 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
5960 init_defrootdomain();
5963 #ifdef CONFIG_RT_GROUP_SCHED
5964 init_rt_bandwidth(&root_task_group.rt_bandwidth,
5965 global_rt_period(), global_rt_runtime());
5966 #endif /* CONFIG_RT_GROUP_SCHED */
5968 #ifdef CONFIG_CGROUP_SCHED
5969 task_group_cache = KMEM_CACHE(task_group, 0);
5971 list_add(&root_task_group.list, &task_groups);
5972 INIT_LIST_HEAD(&root_task_group.children);
5973 INIT_LIST_HEAD(&root_task_group.siblings);
5974 autogroup_init(&init_task);
5975 #endif /* CONFIG_CGROUP_SCHED */
5977 for_each_possible_cpu(i) {
5981 raw_spin_lock_init(&rq->lock);
5983 rq->calc_load_active = 0;
5984 rq->calc_load_update = jiffies + LOAD_FREQ;
5985 init_cfs_rq(&rq->cfs);
5986 init_rt_rq(&rq->rt);
5987 init_dl_rq(&rq->dl);
5988 #ifdef CONFIG_FAIR_GROUP_SCHED
5989 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
5990 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
5991 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
5993 * How much CPU bandwidth does root_task_group get?
5995 * In case of task-groups formed thr' the cgroup filesystem, it
5996 * gets 100% of the CPU resources in the system. This overall
5997 * system CPU resource is divided among the tasks of
5998 * root_task_group and its child task-groups in a fair manner,
5999 * based on each entity's (task or task-group's) weight
6000 * (se->load.weight).
6002 * In other words, if root_task_group has 10 tasks of weight
6003 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6004 * then A0's share of the CPU resource is:
6006 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6008 * We achieve this by letting root_task_group's tasks sit
6009 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6011 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6012 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6013 #endif /* CONFIG_FAIR_GROUP_SCHED */
6015 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6016 #ifdef CONFIG_RT_GROUP_SCHED
6017 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6020 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6021 rq->cpu_load[j] = 0;
6026 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6027 rq->balance_callback = NULL;
6028 rq->active_balance = 0;
6029 rq->next_balance = jiffies;
6034 rq->avg_idle = 2*sysctl_sched_migration_cost;
6035 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6037 INIT_LIST_HEAD(&rq->cfs_tasks);
6039 rq_attach_root(rq, &def_root_domain);
6040 #ifdef CONFIG_NO_HZ_COMMON
6041 rq->last_load_update_tick = jiffies;
6042 rq->last_blocked_load_update_tick = jiffies;
6043 atomic_set(&rq->nohz_flags, 0);
6045 #endif /* CONFIG_SMP */
6047 atomic_set(&rq->nr_iowait, 0);
6050 set_load_weight(&init_task, false);
6053 * The boot idle thread does lazy MMU switching as well:
6056 enter_lazy_tlb(&init_mm, current);
6059 * Make us the idle thread. Technically, schedule() should not be
6060 * called from this thread, however somewhere below it might be,
6061 * but because we are the idle thread, we just pick up running again
6062 * when this runqueue becomes "idle".
6064 init_idle(current, smp_processor_id());
6066 calc_load_update = jiffies + LOAD_FREQ;
6069 idle_thread_set_boot_cpu();
6071 init_sched_fair_class();
6075 scheduler_running = 1;
6078 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6079 static inline int preempt_count_equals(int preempt_offset)
6081 int nested = preempt_count() + rcu_preempt_depth();
6083 return (nested == preempt_offset);
6086 void __might_sleep(const char *file, int line, int preempt_offset)
6089 * Blocking primitives will set (and therefore destroy) current->state,
6090 * since we will exit with TASK_RUNNING make sure we enter with it,
6091 * otherwise we will destroy state.
6093 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6094 "do not call blocking ops when !TASK_RUNNING; "
6095 "state=%lx set at [<%p>] %pS\n",
6097 (void *)current->task_state_change,
6098 (void *)current->task_state_change);
6100 ___might_sleep(file, line, preempt_offset);
6102 EXPORT_SYMBOL(__might_sleep);
6104 void ___might_sleep(const char *file, int line, int preempt_offset)
6106 /* Ratelimiting timestamp: */
6107 static unsigned long prev_jiffy;
6109 unsigned long preempt_disable_ip;
6111 /* WARN_ON_ONCE() by default, no rate limit required: */
6114 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6115 !is_idle_task(current)) ||
6116 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6120 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6122 prev_jiffy = jiffies;
6124 /* Save this before calling printk(), since that will clobber it: */
6125 preempt_disable_ip = get_preempt_disable_ip(current);
6128 "BUG: sleeping function called from invalid context at %s:%d\n",
6131 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6132 in_atomic(), irqs_disabled(),
6133 current->pid, current->comm);
6135 if (task_stack_end_corrupted(current))
6136 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6138 debug_show_held_locks(current);
6139 if (irqs_disabled())
6140 print_irqtrace_events(current);
6141 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6142 && !preempt_count_equals(preempt_offset)) {
6143 pr_err("Preemption disabled at:");
6144 print_ip_sym(preempt_disable_ip);
6148 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6150 EXPORT_SYMBOL(___might_sleep);
6153 #ifdef CONFIG_MAGIC_SYSRQ
6154 void normalize_rt_tasks(void)
6156 struct task_struct *g, *p;
6157 struct sched_attr attr = {
6158 .sched_policy = SCHED_NORMAL,
6161 read_lock(&tasklist_lock);
6162 for_each_process_thread(g, p) {
6164 * Only normalize user tasks:
6166 if (p->flags & PF_KTHREAD)
6169 p->se.exec_start = 0;
6170 schedstat_set(p->se.statistics.wait_start, 0);
6171 schedstat_set(p->se.statistics.sleep_start, 0);
6172 schedstat_set(p->se.statistics.block_start, 0);
6174 if (!dl_task(p) && !rt_task(p)) {
6176 * Renice negative nice level userspace
6179 if (task_nice(p) < 0)
6180 set_user_nice(p, 0);
6184 __sched_setscheduler(p, &attr, false, false);
6186 read_unlock(&tasklist_lock);
6189 #endif /* CONFIG_MAGIC_SYSRQ */
6191 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6193 * These functions are only useful for the IA64 MCA handling, or kdb.
6195 * They can only be called when the whole system has been
6196 * stopped - every CPU needs to be quiescent, and no scheduling
6197 * activity can take place. Using them for anything else would
6198 * be a serious bug, and as a result, they aren't even visible
6199 * under any other configuration.
6203 * curr_task - return the current task for a given CPU.
6204 * @cpu: the processor in question.
6206 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6208 * Return: The current task for @cpu.
6210 struct task_struct *curr_task(int cpu)
6212 return cpu_curr(cpu);
6215 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6219 * set_curr_task - set the current task for a given CPU.
6220 * @cpu: the processor in question.
6221 * @p: the task pointer to set.
6223 * Description: This function must only be used when non-maskable interrupts
6224 * are serviced on a separate stack. It allows the architecture to switch the
6225 * notion of the current task on a CPU in a non-blocking manner. This function
6226 * must be called with all CPU's synchronized, and interrupts disabled, the
6227 * and caller must save the original value of the current task (see
6228 * curr_task() above) and restore that value before reenabling interrupts and
6229 * re-starting the system.
6231 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6233 void ia64_set_curr_task(int cpu, struct task_struct *p)
6240 #ifdef CONFIG_CGROUP_SCHED
6241 /* task_group_lock serializes the addition/removal of task groups */
6242 static DEFINE_SPINLOCK(task_group_lock);
6244 static void sched_free_group(struct task_group *tg)
6246 free_fair_sched_group(tg);
6247 free_rt_sched_group(tg);
6249 kmem_cache_free(task_group_cache, tg);
6252 /* allocate runqueue etc for a new task group */
6253 struct task_group *sched_create_group(struct task_group *parent)
6255 struct task_group *tg;
6257 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6259 return ERR_PTR(-ENOMEM);
6261 if (!alloc_fair_sched_group(tg, parent))
6264 if (!alloc_rt_sched_group(tg, parent))
6270 sched_free_group(tg);
6271 return ERR_PTR(-ENOMEM);
6274 void sched_online_group(struct task_group *tg, struct task_group *parent)
6276 unsigned long flags;
6278 spin_lock_irqsave(&task_group_lock, flags);
6279 list_add_rcu(&tg->list, &task_groups);
6281 /* Root should already exist: */
6284 tg->parent = parent;
6285 INIT_LIST_HEAD(&tg->children);
6286 list_add_rcu(&tg->siblings, &parent->children);
6287 spin_unlock_irqrestore(&task_group_lock, flags);
6289 online_fair_sched_group(tg);
6292 /* rcu callback to free various structures associated with a task group */
6293 static void sched_free_group_rcu(struct rcu_head *rhp)
6295 /* Now it should be safe to free those cfs_rqs: */
6296 sched_free_group(container_of(rhp, struct task_group, rcu));
6299 void sched_destroy_group(struct task_group *tg)
6301 /* Wait for possible concurrent references to cfs_rqs complete: */
6302 call_rcu(&tg->rcu, sched_free_group_rcu);
6305 void sched_offline_group(struct task_group *tg)
6307 unsigned long flags;
6309 /* End participation in shares distribution: */
6310 unregister_fair_sched_group(tg);
6312 spin_lock_irqsave(&task_group_lock, flags);
6313 list_del_rcu(&tg->list);
6314 list_del_rcu(&tg->siblings);
6315 spin_unlock_irqrestore(&task_group_lock, flags);
6318 static void sched_change_group(struct task_struct *tsk, int type)
6320 struct task_group *tg;
6323 * All callers are synchronized by task_rq_lock(); we do not use RCU
6324 * which is pointless here. Thus, we pass "true" to task_css_check()
6325 * to prevent lockdep warnings.
6327 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6328 struct task_group, css);
6329 tg = autogroup_task_group(tsk, tg);
6330 tsk->sched_task_group = tg;
6332 #ifdef CONFIG_FAIR_GROUP_SCHED
6333 if (tsk->sched_class->task_change_group)
6334 tsk->sched_class->task_change_group(tsk, type);
6337 set_task_rq(tsk, task_cpu(tsk));
6341 * Change task's runqueue when it moves between groups.
6343 * The caller of this function should have put the task in its new group by
6344 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6347 void sched_move_task(struct task_struct *tsk)
6349 int queued, running, queue_flags =
6350 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6354 rq = task_rq_lock(tsk, &rf);
6355 update_rq_clock(rq);
6357 running = task_current(rq, tsk);
6358 queued = task_on_rq_queued(tsk);
6361 dequeue_task(rq, tsk, queue_flags);
6363 put_prev_task(rq, tsk);
6365 sched_change_group(tsk, TASK_MOVE_GROUP);
6368 enqueue_task(rq, tsk, queue_flags);
6370 set_curr_task(rq, tsk);
6372 task_rq_unlock(rq, tsk, &rf);
6375 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6377 return css ? container_of(css, struct task_group, css) : NULL;
6380 static struct cgroup_subsys_state *
6381 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6383 struct task_group *parent = css_tg(parent_css);
6384 struct task_group *tg;
6387 /* This is early initialization for the top cgroup */
6388 return &root_task_group.css;
6391 tg = sched_create_group(parent);
6393 return ERR_PTR(-ENOMEM);
6398 /* Expose task group only after completing cgroup initialization */
6399 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6401 struct task_group *tg = css_tg(css);
6402 struct task_group *parent = css_tg(css->parent);
6405 sched_online_group(tg, parent);
6409 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6411 struct task_group *tg = css_tg(css);
6413 sched_offline_group(tg);
6416 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6418 struct task_group *tg = css_tg(css);
6421 * Relies on the RCU grace period between css_released() and this.
6423 sched_free_group(tg);
6427 * This is called before wake_up_new_task(), therefore we really only
6428 * have to set its group bits, all the other stuff does not apply.
6430 static void cpu_cgroup_fork(struct task_struct *task)
6435 rq = task_rq_lock(task, &rf);
6437 update_rq_clock(rq);
6438 sched_change_group(task, TASK_SET_GROUP);
6440 task_rq_unlock(rq, task, &rf);
6443 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6445 struct task_struct *task;
6446 struct cgroup_subsys_state *css;
6449 cgroup_taskset_for_each(task, css, tset) {
6450 #ifdef CONFIG_RT_GROUP_SCHED
6451 if (!sched_rt_can_attach(css_tg(css), task))
6454 /* We don't support RT-tasks being in separate groups */
6455 if (task->sched_class != &fair_sched_class)
6459 * Serialize against wake_up_new_task() such that if its
6460 * running, we're sure to observe its full state.
6462 raw_spin_lock_irq(&task->pi_lock);
6464 * Avoid calling sched_move_task() before wake_up_new_task()
6465 * has happened. This would lead to problems with PELT, due to
6466 * move wanting to detach+attach while we're not attached yet.
6468 if (task->state == TASK_NEW)
6470 raw_spin_unlock_irq(&task->pi_lock);
6478 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6480 struct task_struct *task;
6481 struct cgroup_subsys_state *css;
6483 cgroup_taskset_for_each(task, css, tset)
6484 sched_move_task(task);
6487 #ifdef CONFIG_FAIR_GROUP_SCHED
6488 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6489 struct cftype *cftype, u64 shareval)
6491 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6494 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6497 struct task_group *tg = css_tg(css);
6499 return (u64) scale_load_down(tg->shares);
6502 #ifdef CONFIG_CFS_BANDWIDTH
6503 static DEFINE_MUTEX(cfs_constraints_mutex);
6505 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6506 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6508 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6510 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6512 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6513 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6515 if (tg == &root_task_group)
6519 * Ensure we have at some amount of bandwidth every period. This is
6520 * to prevent reaching a state of large arrears when throttled via
6521 * entity_tick() resulting in prolonged exit starvation.
6523 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6527 * Likewise, bound things on the otherside by preventing insane quota
6528 * periods. This also allows us to normalize in computing quota
6531 if (period > max_cfs_quota_period)
6535 * Prevent race between setting of cfs_rq->runtime_enabled and
6536 * unthrottle_offline_cfs_rqs().
6539 mutex_lock(&cfs_constraints_mutex);
6540 ret = __cfs_schedulable(tg, period, quota);
6544 runtime_enabled = quota != RUNTIME_INF;
6545 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6547 * If we need to toggle cfs_bandwidth_used, off->on must occur
6548 * before making related changes, and on->off must occur afterwards
6550 if (runtime_enabled && !runtime_was_enabled)
6551 cfs_bandwidth_usage_inc();
6552 raw_spin_lock_irq(&cfs_b->lock);
6553 cfs_b->period = ns_to_ktime(period);
6554 cfs_b->quota = quota;
6556 __refill_cfs_bandwidth_runtime(cfs_b);
6558 /* Restart the period timer (if active) to handle new period expiry: */
6559 if (runtime_enabled)
6560 start_cfs_bandwidth(cfs_b);
6562 raw_spin_unlock_irq(&cfs_b->lock);
6564 for_each_online_cpu(i) {
6565 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6566 struct rq *rq = cfs_rq->rq;
6569 rq_lock_irq(rq, &rf);
6570 cfs_rq->runtime_enabled = runtime_enabled;
6571 cfs_rq->runtime_remaining = 0;
6573 if (cfs_rq->throttled)
6574 unthrottle_cfs_rq(cfs_rq);
6575 rq_unlock_irq(rq, &rf);
6577 if (runtime_was_enabled && !runtime_enabled)
6578 cfs_bandwidth_usage_dec();
6580 mutex_unlock(&cfs_constraints_mutex);
6586 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
6590 period = ktime_to_ns(tg->cfs_bandwidth.period);
6591 if (cfs_quota_us < 0)
6592 quota = RUNTIME_INF;
6594 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
6596 return tg_set_cfs_bandwidth(tg, period, quota);
6599 long tg_get_cfs_quota(struct task_group *tg)
6603 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
6606 quota_us = tg->cfs_bandwidth.quota;
6607 do_div(quota_us, NSEC_PER_USEC);
6612 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
6616 period = (u64)cfs_period_us * NSEC_PER_USEC;
6617 quota = tg->cfs_bandwidth.quota;
6619 return tg_set_cfs_bandwidth(tg, period, quota);
6622 long tg_get_cfs_period(struct task_group *tg)
6626 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
6627 do_div(cfs_period_us, NSEC_PER_USEC);
6629 return cfs_period_us;
6632 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
6635 return tg_get_cfs_quota(css_tg(css));
6638 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
6639 struct cftype *cftype, s64 cfs_quota_us)
6641 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
6644 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
6647 return tg_get_cfs_period(css_tg(css));
6650 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
6651 struct cftype *cftype, u64 cfs_period_us)
6653 return tg_set_cfs_period(css_tg(css), cfs_period_us);
6656 struct cfs_schedulable_data {
6657 struct task_group *tg;
6662 * normalize group quota/period to be quota/max_period
6663 * note: units are usecs
6665 static u64 normalize_cfs_quota(struct task_group *tg,
6666 struct cfs_schedulable_data *d)
6674 period = tg_get_cfs_period(tg);
6675 quota = tg_get_cfs_quota(tg);
6678 /* note: these should typically be equivalent */
6679 if (quota == RUNTIME_INF || quota == -1)
6682 return to_ratio(period, quota);
6685 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
6687 struct cfs_schedulable_data *d = data;
6688 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6689 s64 quota = 0, parent_quota = -1;
6692 quota = RUNTIME_INF;
6694 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
6696 quota = normalize_cfs_quota(tg, d);
6697 parent_quota = parent_b->hierarchical_quota;
6700 * Ensure max(child_quota) <= parent_quota. On cgroup2,
6701 * always take the min. On cgroup1, only inherit when no
6704 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
6705 quota = min(quota, parent_quota);
6707 if (quota == RUNTIME_INF)
6708 quota = parent_quota;
6709 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
6713 cfs_b->hierarchical_quota = quota;
6718 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
6721 struct cfs_schedulable_data data = {
6727 if (quota != RUNTIME_INF) {
6728 do_div(data.period, NSEC_PER_USEC);
6729 do_div(data.quota, NSEC_PER_USEC);
6733 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
6739 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
6741 struct task_group *tg = css_tg(seq_css(sf));
6742 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6744 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
6745 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
6746 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
6748 if (schedstat_enabled() && tg != &root_task_group) {
6752 for_each_possible_cpu(i)
6753 ws += schedstat_val(tg->se[i]->statistics.wait_sum);
6755 seq_printf(sf, "wait_sum %llu\n", ws);
6760 #endif /* CONFIG_CFS_BANDWIDTH */
6761 #endif /* CONFIG_FAIR_GROUP_SCHED */
6763 #ifdef CONFIG_RT_GROUP_SCHED
6764 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
6765 struct cftype *cft, s64 val)
6767 return sched_group_set_rt_runtime(css_tg(css), val);
6770 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
6773 return sched_group_rt_runtime(css_tg(css));
6776 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
6777 struct cftype *cftype, u64 rt_period_us)
6779 return sched_group_set_rt_period(css_tg(css), rt_period_us);
6782 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
6785 return sched_group_rt_period(css_tg(css));
6787 #endif /* CONFIG_RT_GROUP_SCHED */
6789 static struct cftype cpu_legacy_files[] = {
6790 #ifdef CONFIG_FAIR_GROUP_SCHED
6793 .read_u64 = cpu_shares_read_u64,
6794 .write_u64 = cpu_shares_write_u64,
6797 #ifdef CONFIG_CFS_BANDWIDTH
6799 .name = "cfs_quota_us",
6800 .read_s64 = cpu_cfs_quota_read_s64,
6801 .write_s64 = cpu_cfs_quota_write_s64,
6804 .name = "cfs_period_us",
6805 .read_u64 = cpu_cfs_period_read_u64,
6806 .write_u64 = cpu_cfs_period_write_u64,
6810 .seq_show = cpu_cfs_stat_show,
6813 #ifdef CONFIG_RT_GROUP_SCHED
6815 .name = "rt_runtime_us",
6816 .read_s64 = cpu_rt_runtime_read,
6817 .write_s64 = cpu_rt_runtime_write,
6820 .name = "rt_period_us",
6821 .read_u64 = cpu_rt_period_read_uint,
6822 .write_u64 = cpu_rt_period_write_uint,
6828 static int cpu_extra_stat_show(struct seq_file *sf,
6829 struct cgroup_subsys_state *css)
6831 #ifdef CONFIG_CFS_BANDWIDTH
6833 struct task_group *tg = css_tg(css);
6834 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6837 throttled_usec = cfs_b->throttled_time;
6838 do_div(throttled_usec, NSEC_PER_USEC);
6840 seq_printf(sf, "nr_periods %d\n"
6842 "throttled_usec %llu\n",
6843 cfs_b->nr_periods, cfs_b->nr_throttled,
6850 #ifdef CONFIG_FAIR_GROUP_SCHED
6851 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
6854 struct task_group *tg = css_tg(css);
6855 u64 weight = scale_load_down(tg->shares);
6857 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
6860 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
6861 struct cftype *cft, u64 weight)
6864 * cgroup weight knobs should use the common MIN, DFL and MAX
6865 * values which are 1, 100 and 10000 respectively. While it loses
6866 * a bit of range on both ends, it maps pretty well onto the shares
6867 * value used by scheduler and the round-trip conversions preserve
6868 * the original value over the entire range.
6870 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
6873 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
6875 return sched_group_set_shares(css_tg(css), scale_load(weight));
6878 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
6881 unsigned long weight = scale_load_down(css_tg(css)->shares);
6882 int last_delta = INT_MAX;
6885 /* find the closest nice value to the current weight */
6886 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
6887 delta = abs(sched_prio_to_weight[prio] - weight);
6888 if (delta >= last_delta)
6893 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
6896 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
6897 struct cftype *cft, s64 nice)
6899 unsigned long weight;
6902 if (nice < MIN_NICE || nice > MAX_NICE)
6905 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
6906 idx = array_index_nospec(idx, 40);
6907 weight = sched_prio_to_weight[idx];
6909 return sched_group_set_shares(css_tg(css), scale_load(weight));
6913 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
6914 long period, long quota)
6917 seq_puts(sf, "max");
6919 seq_printf(sf, "%ld", quota);
6921 seq_printf(sf, " %ld\n", period);
6924 /* caller should put the current value in *@periodp before calling */
6925 static int __maybe_unused cpu_period_quota_parse(char *buf,
6926 u64 *periodp, u64 *quotap)
6928 char tok[21]; /* U64_MAX */
6930 if (!sscanf(buf, "%s %llu", tok, periodp))
6933 *periodp *= NSEC_PER_USEC;
6935 if (sscanf(tok, "%llu", quotap))
6936 *quotap *= NSEC_PER_USEC;
6937 else if (!strcmp(tok, "max"))
6938 *quotap = RUNTIME_INF;
6945 #ifdef CONFIG_CFS_BANDWIDTH
6946 static int cpu_max_show(struct seq_file *sf, void *v)
6948 struct task_group *tg = css_tg(seq_css(sf));
6950 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
6954 static ssize_t cpu_max_write(struct kernfs_open_file *of,
6955 char *buf, size_t nbytes, loff_t off)
6957 struct task_group *tg = css_tg(of_css(of));
6958 u64 period = tg_get_cfs_period(tg);
6962 ret = cpu_period_quota_parse(buf, &period, "a);
6964 ret = tg_set_cfs_bandwidth(tg, period, quota);
6965 return ret ?: nbytes;
6969 static struct cftype cpu_files[] = {
6970 #ifdef CONFIG_FAIR_GROUP_SCHED
6973 .flags = CFTYPE_NOT_ON_ROOT,
6974 .read_u64 = cpu_weight_read_u64,
6975 .write_u64 = cpu_weight_write_u64,
6978 .name = "weight.nice",
6979 .flags = CFTYPE_NOT_ON_ROOT,
6980 .read_s64 = cpu_weight_nice_read_s64,
6981 .write_s64 = cpu_weight_nice_write_s64,
6984 #ifdef CONFIG_CFS_BANDWIDTH
6987 .flags = CFTYPE_NOT_ON_ROOT,
6988 .seq_show = cpu_max_show,
6989 .write = cpu_max_write,
6995 struct cgroup_subsys cpu_cgrp_subsys = {
6996 .css_alloc = cpu_cgroup_css_alloc,
6997 .css_online = cpu_cgroup_css_online,
6998 .css_released = cpu_cgroup_css_released,
6999 .css_free = cpu_cgroup_css_free,
7000 .css_extra_stat_show = cpu_extra_stat_show,
7001 .fork = cpu_cgroup_fork,
7002 .can_attach = cpu_cgroup_can_attach,
7003 .attach = cpu_cgroup_attach,
7004 .legacy_cftypes = cpu_legacy_files,
7005 .dfl_cftypes = cpu_files,
7010 #endif /* CONFIG_CGROUP_SCHED */
7012 void dump_cpu_task(int cpu)
7014 pr_info("Task dump for CPU %d:\n", cpu);
7015 sched_show_task(cpu_curr(cpu));
7019 * Nice levels are multiplicative, with a gentle 10% change for every
7020 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7021 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7022 * that remained on nice 0.
7024 * The "10% effect" is relative and cumulative: from _any_ nice level,
7025 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7026 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7027 * If a task goes up by ~10% and another task goes down by ~10% then
7028 * the relative distance between them is ~25%.)
7030 const int sched_prio_to_weight[40] = {
7031 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7032 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7033 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7034 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7035 /* 0 */ 1024, 820, 655, 526, 423,
7036 /* 5 */ 335, 272, 215, 172, 137,
7037 /* 10 */ 110, 87, 70, 56, 45,
7038 /* 15 */ 36, 29, 23, 18, 15,
7042 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7044 * In cases where the weight does not change often, we can use the
7045 * precalculated inverse to speed up arithmetics by turning divisions
7046 * into multiplications:
7048 const u32 sched_prio_to_wmult[40] = {
7049 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7050 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7051 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7052 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7053 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7054 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7055 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7056 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7059 #undef CREATE_TRACE_POINTS