4 * Core kernel scheduler code and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
10 #include <linux/kthread.h>
11 #include <linux/nospec.h>
13 #include <asm/switch_to.h>
16 #include "../workqueue_internal.h"
17 #include "../smpboot.h"
19 #define CREATE_TRACE_POINTS
20 #include <trace/events/sched.h>
22 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
24 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
26 * Debugging: various feature bits
28 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
29 * sysctl_sched_features, defined in sched.h, to allow constants propagation
30 * at compile time and compiler optimization based on features default.
32 #define SCHED_FEAT(name, enabled) \
33 (1UL << __SCHED_FEAT_##name) * enabled |
34 const_debug unsigned int sysctl_sched_features =
41 * Number of tasks to iterate in a single balance run.
42 * Limited because this is done with IRQs disabled.
44 const_debug unsigned int sysctl_sched_nr_migrate = 32;
47 * period over which we average the RT time consumption, measured
52 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
55 * period over which we measure -rt task CPU usage in us.
58 unsigned int sysctl_sched_rt_period = 1000000;
60 __read_mostly int scheduler_running;
63 * part of the period that we allow rt tasks to run in us.
66 int sysctl_sched_rt_runtime = 950000;
69 * __task_rq_lock - lock the rq @p resides on.
71 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
76 lockdep_assert_held(&p->pi_lock);
80 raw_spin_lock(&rq->lock);
81 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
85 raw_spin_unlock(&rq->lock);
87 while (unlikely(task_on_rq_migrating(p)))
93 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
95 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
96 __acquires(p->pi_lock)
102 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
104 raw_spin_lock(&rq->lock);
106 * move_queued_task() task_rq_lock()
109 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
110 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
111 * [S] ->cpu = new_cpu [L] task_rq()
115 * If we observe the old CPU in task_rq_lock, the acquire of
116 * the old rq->lock will fully serialize against the stores.
118 * If we observe the new CPU in task_rq_lock, the acquire will
119 * pair with the WMB to ensure we must then also see migrating.
121 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
125 raw_spin_unlock(&rq->lock);
126 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
128 while (unlikely(task_on_rq_migrating(p)))
134 * RQ-clock updating methods:
137 static void update_rq_clock_task(struct rq *rq, s64 delta)
140 * In theory, the compile should just see 0 here, and optimize out the call
141 * to sched_rt_avg_update. But I don't trust it...
143 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
144 s64 steal = 0, irq_delta = 0;
146 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
147 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
150 * Since irq_time is only updated on {soft,}irq_exit, we might run into
151 * this case when a previous update_rq_clock() happened inside a
154 * When this happens, we stop ->clock_task and only update the
155 * prev_irq_time stamp to account for the part that fit, so that a next
156 * update will consume the rest. This ensures ->clock_task is
159 * It does however cause some slight miss-attribution of {soft,}irq
160 * time, a more accurate solution would be to update the irq_time using
161 * the current rq->clock timestamp, except that would require using
164 if (irq_delta > delta)
167 rq->prev_irq_time += irq_delta;
170 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
171 if (static_key_false((¶virt_steal_rq_enabled))) {
172 steal = paravirt_steal_clock(cpu_of(rq));
173 steal -= rq->prev_steal_time_rq;
175 if (unlikely(steal > delta))
178 rq->prev_steal_time_rq += steal;
183 rq->clock_task += delta;
185 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
186 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
187 sched_rt_avg_update(rq, irq_delta + steal);
191 void update_rq_clock(struct rq *rq)
195 lockdep_assert_held(&rq->lock);
197 if (rq->clock_update_flags & RQCF_ACT_SKIP)
200 #ifdef CONFIG_SCHED_DEBUG
201 if (sched_feat(WARN_DOUBLE_CLOCK))
202 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
203 rq->clock_update_flags |= RQCF_UPDATED;
206 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
210 update_rq_clock_task(rq, delta);
214 #ifdef CONFIG_SCHED_HRTICK
216 * Use HR-timers to deliver accurate preemption points.
219 static void hrtick_clear(struct rq *rq)
221 if (hrtimer_active(&rq->hrtick_timer))
222 hrtimer_cancel(&rq->hrtick_timer);
226 * High-resolution timer tick.
227 * Runs from hardirq context with interrupts disabled.
229 static enum hrtimer_restart hrtick(struct hrtimer *timer)
231 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
234 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
238 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
241 return HRTIMER_NORESTART;
246 static void __hrtick_restart(struct rq *rq)
248 struct hrtimer *timer = &rq->hrtick_timer;
250 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
254 * called from hardirq (IPI) context
256 static void __hrtick_start(void *arg)
262 __hrtick_restart(rq);
263 rq->hrtick_csd_pending = 0;
268 * Called to set the hrtick timer state.
270 * called with rq->lock held and irqs disabled
272 void hrtick_start(struct rq *rq, u64 delay)
274 struct hrtimer *timer = &rq->hrtick_timer;
279 * Don't schedule slices shorter than 10000ns, that just
280 * doesn't make sense and can cause timer DoS.
282 delta = max_t(s64, delay, 10000LL);
283 time = ktime_add_ns(timer->base->get_time(), delta);
285 hrtimer_set_expires(timer, time);
287 if (rq == this_rq()) {
288 __hrtick_restart(rq);
289 } else if (!rq->hrtick_csd_pending) {
290 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
291 rq->hrtick_csd_pending = 1;
297 * Called to set the hrtick timer state.
299 * called with rq->lock held and irqs disabled
301 void hrtick_start(struct rq *rq, u64 delay)
304 * Don't schedule slices shorter than 10000ns, that just
305 * doesn't make sense. Rely on vruntime for fairness.
307 delay = max_t(u64, delay, 10000LL);
308 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
309 HRTIMER_MODE_REL_PINNED);
311 #endif /* CONFIG_SMP */
313 static void hrtick_rq_init(struct rq *rq)
316 rq->hrtick_csd_pending = 0;
318 rq->hrtick_csd.flags = 0;
319 rq->hrtick_csd.func = __hrtick_start;
320 rq->hrtick_csd.info = rq;
323 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
324 rq->hrtick_timer.function = hrtick;
326 #else /* CONFIG_SCHED_HRTICK */
327 static inline void hrtick_clear(struct rq *rq)
331 static inline void hrtick_rq_init(struct rq *rq)
334 #endif /* CONFIG_SCHED_HRTICK */
337 * cmpxchg based fetch_or, macro so it works for different integer types
339 #define fetch_or(ptr, mask) \
341 typeof(ptr) _ptr = (ptr); \
342 typeof(mask) _mask = (mask); \
343 typeof(*_ptr) _old, _val = *_ptr; \
346 _old = cmpxchg(_ptr, _val, _val | _mask); \
354 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
356 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
357 * this avoids any races wrt polling state changes and thereby avoids
360 static bool set_nr_and_not_polling(struct task_struct *p)
362 struct thread_info *ti = task_thread_info(p);
363 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
367 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
369 * If this returns true, then the idle task promises to call
370 * sched_ttwu_pending() and reschedule soon.
372 static bool set_nr_if_polling(struct task_struct *p)
374 struct thread_info *ti = task_thread_info(p);
375 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
378 if (!(val & _TIF_POLLING_NRFLAG))
380 if (val & _TIF_NEED_RESCHED)
382 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
391 static bool set_nr_and_not_polling(struct task_struct *p)
393 set_tsk_need_resched(p);
398 static bool set_nr_if_polling(struct task_struct *p)
405 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
407 struct wake_q_node *node = &task->wake_q;
410 * Atomically grab the task, if ->wake_q is !nil already it means
411 * its already queued (either by us or someone else) and will get the
412 * wakeup due to that.
414 * This cmpxchg() implies a full barrier, which pairs with the write
415 * barrier implied by the wakeup in wake_up_q().
417 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
420 get_task_struct(task);
423 * The head is context local, there can be no concurrency.
426 head->lastp = &node->next;
429 void wake_up_q(struct wake_q_head *head)
431 struct wake_q_node *node = head->first;
433 while (node != WAKE_Q_TAIL) {
434 struct task_struct *task;
436 task = container_of(node, struct task_struct, wake_q);
438 /* Task can safely be re-inserted now: */
440 task->wake_q.next = NULL;
443 * wake_up_process() implies a wmb() to pair with the queueing
444 * in wake_q_add() so as not to miss wakeups.
446 wake_up_process(task);
447 put_task_struct(task);
452 * resched_curr - mark rq's current task 'to be rescheduled now'.
454 * On UP this means the setting of the need_resched flag, on SMP it
455 * might also involve a cross-CPU call to trigger the scheduler on
458 void resched_curr(struct rq *rq)
460 struct task_struct *curr = rq->curr;
463 lockdep_assert_held(&rq->lock);
465 if (test_tsk_need_resched(curr))
470 if (cpu == smp_processor_id()) {
471 set_tsk_need_resched(curr);
472 set_preempt_need_resched();
476 if (set_nr_and_not_polling(curr))
477 smp_send_reschedule(cpu);
479 trace_sched_wake_idle_without_ipi(cpu);
482 void resched_cpu(int cpu)
484 struct rq *rq = cpu_rq(cpu);
487 raw_spin_lock_irqsave(&rq->lock, flags);
488 if (cpu_online(cpu) || cpu == smp_processor_id())
490 raw_spin_unlock_irqrestore(&rq->lock, flags);
494 #ifdef CONFIG_NO_HZ_COMMON
496 * In the semi idle case, use the nearest busy CPU for migrating timers
497 * from an idle CPU. This is good for power-savings.
499 * We don't do similar optimization for completely idle system, as
500 * selecting an idle CPU will add more delays to the timers than intended
501 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
503 int get_nohz_timer_target(void)
505 int i, cpu = smp_processor_id();
506 struct sched_domain *sd;
508 if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
512 for_each_domain(cpu, sd) {
513 for_each_cpu(i, sched_domain_span(sd)) {
517 if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
524 if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
525 cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
532 * When add_timer_on() enqueues a timer into the timer wheel of an
533 * idle CPU then this timer might expire before the next timer event
534 * which is scheduled to wake up that CPU. In case of a completely
535 * idle system the next event might even be infinite time into the
536 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
537 * leaves the inner idle loop so the newly added timer is taken into
538 * account when the CPU goes back to idle and evaluates the timer
539 * wheel for the next timer event.
541 static void wake_up_idle_cpu(int cpu)
543 struct rq *rq = cpu_rq(cpu);
545 if (cpu == smp_processor_id())
548 if (set_nr_and_not_polling(rq->idle))
549 smp_send_reschedule(cpu);
551 trace_sched_wake_idle_without_ipi(cpu);
554 static bool wake_up_full_nohz_cpu(int cpu)
557 * We just need the target to call irq_exit() and re-evaluate
558 * the next tick. The nohz full kick at least implies that.
559 * If needed we can still optimize that later with an
562 if (cpu_is_offline(cpu))
563 return true; /* Don't try to wake offline CPUs. */
564 if (tick_nohz_full_cpu(cpu)) {
565 if (cpu != smp_processor_id() ||
566 tick_nohz_tick_stopped())
567 tick_nohz_full_kick_cpu(cpu);
575 * Wake up the specified CPU. If the CPU is going offline, it is the
576 * caller's responsibility to deal with the lost wakeup, for example,
577 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
579 void wake_up_nohz_cpu(int cpu)
581 if (!wake_up_full_nohz_cpu(cpu))
582 wake_up_idle_cpu(cpu);
585 static inline bool got_nohz_idle_kick(void)
587 int cpu = smp_processor_id();
589 if (!(atomic_read(nohz_flags(cpu)) & NOHZ_KICK_MASK))
592 if (idle_cpu(cpu) && !need_resched())
596 * We can't run Idle Load Balance on this CPU for this time so we
597 * cancel it and clear NOHZ_BALANCE_KICK
599 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
603 #else /* CONFIG_NO_HZ_COMMON */
605 static inline bool got_nohz_idle_kick(void)
610 #endif /* CONFIG_NO_HZ_COMMON */
612 #ifdef CONFIG_NO_HZ_FULL
613 bool sched_can_stop_tick(struct rq *rq)
617 /* Deadline tasks, even if single, need the tick */
618 if (rq->dl.dl_nr_running)
622 * If there are more than one RR tasks, we need the tick to effect the
623 * actual RR behaviour.
625 if (rq->rt.rr_nr_running) {
626 if (rq->rt.rr_nr_running == 1)
633 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
634 * forced preemption between FIFO tasks.
636 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
641 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
642 * if there's more than one we need the tick for involuntary
645 if (rq->nr_running > 1)
650 #endif /* CONFIG_NO_HZ_FULL */
652 void sched_avg_update(struct rq *rq)
654 s64 period = sched_avg_period();
656 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
658 * Inline assembly required to prevent the compiler
659 * optimising this loop into a divmod call.
660 * See __iter_div_u64_rem() for another example of this.
662 asm("" : "+rm" (rq->age_stamp));
663 rq->age_stamp += period;
668 #endif /* CONFIG_SMP */
670 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
671 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
673 * Iterate task_group tree rooted at *from, calling @down when first entering a
674 * node and @up when leaving it for the final time.
676 * Caller must hold rcu_lock or sufficient equivalent.
678 int walk_tg_tree_from(struct task_group *from,
679 tg_visitor down, tg_visitor up, void *data)
681 struct task_group *parent, *child;
687 ret = (*down)(parent, data);
690 list_for_each_entry_rcu(child, &parent->children, siblings) {
697 ret = (*up)(parent, data);
698 if (ret || parent == from)
702 parent = parent->parent;
709 int tg_nop(struct task_group *tg, void *data)
715 static void set_load_weight(struct task_struct *p, bool update_load)
717 int prio = p->static_prio - MAX_RT_PRIO;
718 struct load_weight *load = &p->se.load;
721 * SCHED_IDLE tasks get minimal weight:
723 if (idle_policy(p->policy)) {
724 load->weight = scale_load(WEIGHT_IDLEPRIO);
725 load->inv_weight = WMULT_IDLEPRIO;
730 * SCHED_OTHER tasks have to update their load when changing their
733 if (update_load && p->sched_class == &fair_sched_class) {
734 reweight_task(p, prio);
736 load->weight = scale_load(sched_prio_to_weight[prio]);
737 load->inv_weight = sched_prio_to_wmult[prio];
741 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
743 if (!(flags & ENQUEUE_NOCLOCK))
746 if (!(flags & ENQUEUE_RESTORE))
747 sched_info_queued(rq, p);
749 p->sched_class->enqueue_task(rq, p, flags);
752 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
754 if (!(flags & DEQUEUE_NOCLOCK))
757 if (!(flags & DEQUEUE_SAVE))
758 sched_info_dequeued(rq, p);
760 p->sched_class->dequeue_task(rq, p, flags);
763 void activate_task(struct rq *rq, struct task_struct *p, int flags)
765 if (task_contributes_to_load(p))
766 rq->nr_uninterruptible--;
768 enqueue_task(rq, p, flags);
771 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
773 if (task_contributes_to_load(p))
774 rq->nr_uninterruptible++;
776 dequeue_task(rq, p, flags);
780 * __normal_prio - return the priority that is based on the static prio
782 static inline int __normal_prio(struct task_struct *p)
784 return p->static_prio;
788 * Calculate the expected normal priority: i.e. priority
789 * without taking RT-inheritance into account. Might be
790 * boosted by interactivity modifiers. Changes upon fork,
791 * setprio syscalls, and whenever the interactivity
792 * estimator recalculates.
794 static inline int normal_prio(struct task_struct *p)
798 if (task_has_dl_policy(p))
799 prio = MAX_DL_PRIO-1;
800 else if (task_has_rt_policy(p))
801 prio = MAX_RT_PRIO-1 - p->rt_priority;
803 prio = __normal_prio(p);
808 * Calculate the current priority, i.e. the priority
809 * taken into account by the scheduler. This value might
810 * be boosted by RT tasks, or might be boosted by
811 * interactivity modifiers. Will be RT if the task got
812 * RT-boosted. If not then it returns p->normal_prio.
814 static int effective_prio(struct task_struct *p)
816 p->normal_prio = normal_prio(p);
818 * If we are RT tasks or we were boosted to RT priority,
819 * keep the priority unchanged. Otherwise, update priority
820 * to the normal priority:
822 if (!rt_prio(p->prio))
823 return p->normal_prio;
828 * task_curr - is this task currently executing on a CPU?
829 * @p: the task in question.
831 * Return: 1 if the task is currently executing. 0 otherwise.
833 inline int task_curr(const struct task_struct *p)
835 return cpu_curr(task_cpu(p)) == p;
839 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
840 * use the balance_callback list if you want balancing.
842 * this means any call to check_class_changed() must be followed by a call to
843 * balance_callback().
845 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
846 const struct sched_class *prev_class,
849 if (prev_class != p->sched_class) {
850 if (prev_class->switched_from)
851 prev_class->switched_from(rq, p);
853 p->sched_class->switched_to(rq, p);
854 } else if (oldprio != p->prio || dl_task(p))
855 p->sched_class->prio_changed(rq, p, oldprio);
858 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
860 const struct sched_class *class;
862 if (p->sched_class == rq->curr->sched_class) {
863 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
865 for_each_class(class) {
866 if (class == rq->curr->sched_class)
868 if (class == p->sched_class) {
876 * A queue event has occurred, and we're going to schedule. In
877 * this case, we can save a useless back to back clock update.
879 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
880 rq_clock_skip_update(rq);
885 static inline bool is_per_cpu_kthread(struct task_struct *p)
887 if (!(p->flags & PF_KTHREAD))
890 if (p->nr_cpus_allowed != 1)
897 * Per-CPU kthreads are allowed to run on !actie && online CPUs, see
898 * __set_cpus_allowed_ptr() and select_fallback_rq().
900 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
902 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
905 if (is_per_cpu_kthread(p))
906 return cpu_online(cpu);
908 return cpu_active(cpu);
912 * This is how migration works:
914 * 1) we invoke migration_cpu_stop() on the target CPU using
916 * 2) stopper starts to run (implicitly forcing the migrated thread
918 * 3) it checks whether the migrated task is still in the wrong runqueue.
919 * 4) if it's in the wrong runqueue then the migration thread removes
920 * it and puts it into the right queue.
921 * 5) stopper completes and stop_one_cpu() returns and the migration
926 * move_queued_task - move a queued task to new rq.
928 * Returns (locked) new rq. Old rq's lock is released.
930 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
931 struct task_struct *p, int new_cpu)
933 lockdep_assert_held(&rq->lock);
935 p->on_rq = TASK_ON_RQ_MIGRATING;
936 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
937 set_task_cpu(p, new_cpu);
940 rq = cpu_rq(new_cpu);
943 BUG_ON(task_cpu(p) != new_cpu);
944 enqueue_task(rq, p, 0);
945 p->on_rq = TASK_ON_RQ_QUEUED;
946 check_preempt_curr(rq, p, 0);
951 struct migration_arg {
952 struct task_struct *task;
957 * Move (not current) task off this CPU, onto the destination CPU. We're doing
958 * this because either it can't run here any more (set_cpus_allowed()
959 * away from this CPU, or CPU going down), or because we're
960 * attempting to rebalance this task on exec (sched_exec).
962 * So we race with normal scheduler movements, but that's OK, as long
963 * as the task is no longer on this CPU.
965 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
966 struct task_struct *p, int dest_cpu)
968 /* Affinity changed (again). */
969 if (!is_cpu_allowed(p, dest_cpu))
973 rq = move_queued_task(rq, rf, p, dest_cpu);
979 * migration_cpu_stop - this will be executed by a highprio stopper thread
980 * and performs thread migration by bumping thread off CPU then
981 * 'pushing' onto another runqueue.
983 static int migration_cpu_stop(void *data)
985 struct migration_arg *arg = data;
986 struct task_struct *p = arg->task;
987 struct rq *rq = this_rq();
991 * The original target CPU might have gone down and we might
992 * be on another CPU but it doesn't matter.
996 * We need to explicitly wake pending tasks before running
997 * __migrate_task() such that we will not miss enforcing cpus_allowed
998 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1000 sched_ttwu_pending();
1002 raw_spin_lock(&p->pi_lock);
1005 * If task_rq(p) != rq, it cannot be migrated here, because we're
1006 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1007 * we're holding p->pi_lock.
1009 if (task_rq(p) == rq) {
1010 if (task_on_rq_queued(p))
1011 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1013 p->wake_cpu = arg->dest_cpu;
1016 raw_spin_unlock(&p->pi_lock);
1023 * sched_class::set_cpus_allowed must do the below, but is not required to
1024 * actually call this function.
1026 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1028 cpumask_copy(&p->cpus_allowed, new_mask);
1029 p->nr_cpus_allowed = cpumask_weight(new_mask);
1032 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1034 struct rq *rq = task_rq(p);
1035 bool queued, running;
1037 lockdep_assert_held(&p->pi_lock);
1039 queued = task_on_rq_queued(p);
1040 running = task_current(rq, p);
1044 * Because __kthread_bind() calls this on blocked tasks without
1047 lockdep_assert_held(&rq->lock);
1048 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1051 put_prev_task(rq, p);
1053 p->sched_class->set_cpus_allowed(p, new_mask);
1056 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1058 set_curr_task(rq, p);
1062 * Change a given task's CPU affinity. Migrate the thread to a
1063 * proper CPU and schedule it away if the CPU it's executing on
1064 * is removed from the allowed bitmask.
1066 * NOTE: the caller must have a valid reference to the task, the
1067 * task must not exit() & deallocate itself prematurely. The
1068 * call is not atomic; no spinlocks may be held.
1070 static int __set_cpus_allowed_ptr(struct task_struct *p,
1071 const struct cpumask *new_mask, bool check)
1073 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1074 unsigned int dest_cpu;
1079 rq = task_rq_lock(p, &rf);
1080 update_rq_clock(rq);
1082 if (p->flags & PF_KTHREAD) {
1084 * Kernel threads are allowed on online && !active CPUs
1086 cpu_valid_mask = cpu_online_mask;
1090 * Must re-check here, to close a race against __kthread_bind(),
1091 * sched_setaffinity() is not guaranteed to observe the flag.
1093 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1098 if (cpumask_equal(&p->cpus_allowed, new_mask))
1101 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1106 do_set_cpus_allowed(p, new_mask);
1108 if (p->flags & PF_KTHREAD) {
1110 * For kernel threads that do indeed end up on online &&
1111 * !active we want to ensure they are strict per-CPU threads.
1113 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1114 !cpumask_intersects(new_mask, cpu_active_mask) &&
1115 p->nr_cpus_allowed != 1);
1118 /* Can the task run on the task's current CPU? If so, we're done */
1119 if (cpumask_test_cpu(task_cpu(p), new_mask))
1122 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1123 if (task_running(rq, p) || p->state == TASK_WAKING) {
1124 struct migration_arg arg = { p, dest_cpu };
1125 /* Need help from migration thread: drop lock and wait. */
1126 task_rq_unlock(rq, p, &rf);
1127 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1128 tlb_migrate_finish(p->mm);
1130 } else if (task_on_rq_queued(p)) {
1132 * OK, since we're going to drop the lock immediately
1133 * afterwards anyway.
1135 rq = move_queued_task(rq, &rf, p, dest_cpu);
1138 task_rq_unlock(rq, p, &rf);
1143 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1145 return __set_cpus_allowed_ptr(p, new_mask, false);
1147 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1149 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1151 #ifdef CONFIG_SCHED_DEBUG
1153 * We should never call set_task_cpu() on a blocked task,
1154 * ttwu() will sort out the placement.
1156 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1160 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1161 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1162 * time relying on p->on_rq.
1164 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1165 p->sched_class == &fair_sched_class &&
1166 (p->on_rq && !task_on_rq_migrating(p)));
1168 #ifdef CONFIG_LOCKDEP
1170 * The caller should hold either p->pi_lock or rq->lock, when changing
1171 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1173 * sched_move_task() holds both and thus holding either pins the cgroup,
1176 * Furthermore, all task_rq users should acquire both locks, see
1179 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1180 lockdep_is_held(&task_rq(p)->lock)));
1183 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1185 WARN_ON_ONCE(!cpu_online(new_cpu));
1188 trace_sched_migrate_task(p, new_cpu);
1190 if (task_cpu(p) != new_cpu) {
1191 if (p->sched_class->migrate_task_rq)
1192 p->sched_class->migrate_task_rq(p);
1193 p->se.nr_migrations++;
1195 perf_event_task_migrate(p);
1198 __set_task_cpu(p, new_cpu);
1201 static void __migrate_swap_task(struct task_struct *p, int cpu)
1203 if (task_on_rq_queued(p)) {
1204 struct rq *src_rq, *dst_rq;
1205 struct rq_flags srf, drf;
1207 src_rq = task_rq(p);
1208 dst_rq = cpu_rq(cpu);
1210 rq_pin_lock(src_rq, &srf);
1211 rq_pin_lock(dst_rq, &drf);
1213 p->on_rq = TASK_ON_RQ_MIGRATING;
1214 deactivate_task(src_rq, p, 0);
1215 set_task_cpu(p, cpu);
1216 activate_task(dst_rq, p, 0);
1217 p->on_rq = TASK_ON_RQ_QUEUED;
1218 check_preempt_curr(dst_rq, p, 0);
1220 rq_unpin_lock(dst_rq, &drf);
1221 rq_unpin_lock(src_rq, &srf);
1225 * Task isn't running anymore; make it appear like we migrated
1226 * it before it went to sleep. This means on wakeup we make the
1227 * previous CPU our target instead of where it really is.
1233 struct migration_swap_arg {
1234 struct task_struct *src_task, *dst_task;
1235 int src_cpu, dst_cpu;
1238 static int migrate_swap_stop(void *data)
1240 struct migration_swap_arg *arg = data;
1241 struct rq *src_rq, *dst_rq;
1244 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1247 src_rq = cpu_rq(arg->src_cpu);
1248 dst_rq = cpu_rq(arg->dst_cpu);
1250 double_raw_lock(&arg->src_task->pi_lock,
1251 &arg->dst_task->pi_lock);
1252 double_rq_lock(src_rq, dst_rq);
1254 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1257 if (task_cpu(arg->src_task) != arg->src_cpu)
1260 if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1263 if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1266 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1267 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1272 double_rq_unlock(src_rq, dst_rq);
1273 raw_spin_unlock(&arg->dst_task->pi_lock);
1274 raw_spin_unlock(&arg->src_task->pi_lock);
1280 * Cross migrate two tasks
1282 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1284 struct migration_swap_arg arg;
1287 arg = (struct migration_swap_arg){
1289 .src_cpu = task_cpu(cur),
1291 .dst_cpu = task_cpu(p),
1294 if (arg.src_cpu == arg.dst_cpu)
1298 * These three tests are all lockless; this is OK since all of them
1299 * will be re-checked with proper locks held further down the line.
1301 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1304 if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1307 if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1310 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1311 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1318 * wait_task_inactive - wait for a thread to unschedule.
1320 * If @match_state is nonzero, it's the @p->state value just checked and
1321 * not expected to change. If it changes, i.e. @p might have woken up,
1322 * then return zero. When we succeed in waiting for @p to be off its CPU,
1323 * we return a positive number (its total switch count). If a second call
1324 * a short while later returns the same number, the caller can be sure that
1325 * @p has remained unscheduled the whole time.
1327 * The caller must ensure that the task *will* unschedule sometime soon,
1328 * else this function might spin for a *long* time. This function can't
1329 * be called with interrupts off, or it may introduce deadlock with
1330 * smp_call_function() if an IPI is sent by the same process we are
1331 * waiting to become inactive.
1333 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1335 int running, queued;
1342 * We do the initial early heuristics without holding
1343 * any task-queue locks at all. We'll only try to get
1344 * the runqueue lock when things look like they will
1350 * If the task is actively running on another CPU
1351 * still, just relax and busy-wait without holding
1354 * NOTE! Since we don't hold any locks, it's not
1355 * even sure that "rq" stays as the right runqueue!
1356 * But we don't care, since "task_running()" will
1357 * return false if the runqueue has changed and p
1358 * is actually now running somewhere else!
1360 while (task_running(rq, p)) {
1361 if (match_state && unlikely(p->state != match_state))
1367 * Ok, time to look more closely! We need the rq
1368 * lock now, to be *sure*. If we're wrong, we'll
1369 * just go back and repeat.
1371 rq = task_rq_lock(p, &rf);
1372 trace_sched_wait_task(p);
1373 running = task_running(rq, p);
1374 queued = task_on_rq_queued(p);
1376 if (!match_state || p->state == match_state)
1377 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1378 task_rq_unlock(rq, p, &rf);
1381 * If it changed from the expected state, bail out now.
1383 if (unlikely(!ncsw))
1387 * Was it really running after all now that we
1388 * checked with the proper locks actually held?
1390 * Oops. Go back and try again..
1392 if (unlikely(running)) {
1398 * It's not enough that it's not actively running,
1399 * it must be off the runqueue _entirely_, and not
1402 * So if it was still runnable (but just not actively
1403 * running right now), it's preempted, and we should
1404 * yield - it could be a while.
1406 if (unlikely(queued)) {
1407 ktime_t to = NSEC_PER_SEC / HZ;
1409 set_current_state(TASK_UNINTERRUPTIBLE);
1410 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1415 * Ahh, all good. It wasn't running, and it wasn't
1416 * runnable, which means that it will never become
1417 * running in the future either. We're all done!
1426 * kick_process - kick a running thread to enter/exit the kernel
1427 * @p: the to-be-kicked thread
1429 * Cause a process which is running on another CPU to enter
1430 * kernel-mode, without any delay. (to get signals handled.)
1432 * NOTE: this function doesn't have to take the runqueue lock,
1433 * because all it wants to ensure is that the remote task enters
1434 * the kernel. If the IPI races and the task has been migrated
1435 * to another CPU then no harm is done and the purpose has been
1438 void kick_process(struct task_struct *p)
1444 if ((cpu != smp_processor_id()) && task_curr(p))
1445 smp_send_reschedule(cpu);
1448 EXPORT_SYMBOL_GPL(kick_process);
1451 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1453 * A few notes on cpu_active vs cpu_online:
1455 * - cpu_active must be a subset of cpu_online
1457 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
1458 * see __set_cpus_allowed_ptr(). At this point the newly online
1459 * CPU isn't yet part of the sched domains, and balancing will not
1462 * - on CPU-down we clear cpu_active() to mask the sched domains and
1463 * avoid the load balancer to place new tasks on the to be removed
1464 * CPU. Existing tasks will remain running there and will be taken
1467 * This means that fallback selection must not select !active CPUs.
1468 * And can assume that any active CPU must be online. Conversely
1469 * select_task_rq() below may allow selection of !active CPUs in order
1470 * to satisfy the above rules.
1472 static int select_fallback_rq(int cpu, struct task_struct *p)
1474 int nid = cpu_to_node(cpu);
1475 const struct cpumask *nodemask = NULL;
1476 enum { cpuset, possible, fail } state = cpuset;
1480 * If the node that the CPU is on has been offlined, cpu_to_node()
1481 * will return -1. There is no CPU on the node, and we should
1482 * select the CPU on the other node.
1485 nodemask = cpumask_of_node(nid);
1487 /* Look for allowed, online CPU in same node. */
1488 for_each_cpu(dest_cpu, nodemask) {
1489 if (!cpu_active(dest_cpu))
1491 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1497 /* Any allowed, online CPU? */
1498 for_each_cpu(dest_cpu, &p->cpus_allowed) {
1499 if (!is_cpu_allowed(p, dest_cpu))
1505 /* No more Mr. Nice Guy. */
1508 if (IS_ENABLED(CONFIG_CPUSETS)) {
1509 cpuset_cpus_allowed_fallback(p);
1515 do_set_cpus_allowed(p, cpu_possible_mask);
1526 if (state != cpuset) {
1528 * Don't tell them about moving exiting tasks or
1529 * kernel threads (both mm NULL), since they never
1532 if (p->mm && printk_ratelimit()) {
1533 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1534 task_pid_nr(p), p->comm, cpu);
1542 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1545 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1547 lockdep_assert_held(&p->pi_lock);
1549 if (p->nr_cpus_allowed > 1)
1550 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1552 cpu = cpumask_any(&p->cpus_allowed);
1555 * In order not to call set_task_cpu() on a blocking task we need
1556 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1559 * Since this is common to all placement strategies, this lives here.
1561 * [ this allows ->select_task() to simply return task_cpu(p) and
1562 * not worry about this generic constraint ]
1564 if (unlikely(!is_cpu_allowed(p, cpu)))
1565 cpu = select_fallback_rq(task_cpu(p), p);
1570 static void update_avg(u64 *avg, u64 sample)
1572 s64 diff = sample - *avg;
1576 void sched_set_stop_task(int cpu, struct task_struct *stop)
1578 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1579 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1583 * Make it appear like a SCHED_FIFO task, its something
1584 * userspace knows about and won't get confused about.
1586 * Also, it will make PI more or less work without too
1587 * much confusion -- but then, stop work should not
1588 * rely on PI working anyway.
1590 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1592 stop->sched_class = &stop_sched_class;
1595 cpu_rq(cpu)->stop = stop;
1599 * Reset it back to a normal scheduling class so that
1600 * it can die in pieces.
1602 old_stop->sched_class = &rt_sched_class;
1608 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1609 const struct cpumask *new_mask, bool check)
1611 return set_cpus_allowed_ptr(p, new_mask);
1614 #endif /* CONFIG_SMP */
1617 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1621 if (!schedstat_enabled())
1627 if (cpu == rq->cpu) {
1628 __schedstat_inc(rq->ttwu_local);
1629 __schedstat_inc(p->se.statistics.nr_wakeups_local);
1631 struct sched_domain *sd;
1633 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
1635 for_each_domain(rq->cpu, sd) {
1636 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1637 __schedstat_inc(sd->ttwu_wake_remote);
1644 if (wake_flags & WF_MIGRATED)
1645 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1646 #endif /* CONFIG_SMP */
1648 __schedstat_inc(rq->ttwu_count);
1649 __schedstat_inc(p->se.statistics.nr_wakeups);
1651 if (wake_flags & WF_SYNC)
1652 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
1655 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1657 activate_task(rq, p, en_flags);
1658 p->on_rq = TASK_ON_RQ_QUEUED;
1660 /* If a worker is waking up, notify the workqueue: */
1661 if (p->flags & PF_WQ_WORKER)
1662 wq_worker_waking_up(p, cpu_of(rq));
1666 * Mark the task runnable and perform wakeup-preemption.
1668 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1669 struct rq_flags *rf)
1671 check_preempt_curr(rq, p, wake_flags);
1672 p->state = TASK_RUNNING;
1673 trace_sched_wakeup(p);
1676 if (p->sched_class->task_woken) {
1678 * Our task @p is fully woken up and running; so its safe to
1679 * drop the rq->lock, hereafter rq is only used for statistics.
1681 rq_unpin_lock(rq, rf);
1682 p->sched_class->task_woken(rq, p);
1683 rq_repin_lock(rq, rf);
1686 if (rq->idle_stamp) {
1687 u64 delta = rq_clock(rq) - rq->idle_stamp;
1688 u64 max = 2*rq->max_idle_balance_cost;
1690 update_avg(&rq->avg_idle, delta);
1692 if (rq->avg_idle > max)
1701 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1702 struct rq_flags *rf)
1704 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
1706 lockdep_assert_held(&rq->lock);
1709 if (p->sched_contributes_to_load)
1710 rq->nr_uninterruptible--;
1712 if (wake_flags & WF_MIGRATED)
1713 en_flags |= ENQUEUE_MIGRATED;
1716 ttwu_activate(rq, p, en_flags);
1717 ttwu_do_wakeup(rq, p, wake_flags, rf);
1721 * Called in case the task @p isn't fully descheduled from its runqueue,
1722 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1723 * since all we need to do is flip p->state to TASK_RUNNING, since
1724 * the task is still ->on_rq.
1726 static int ttwu_remote(struct task_struct *p, int wake_flags)
1732 rq = __task_rq_lock(p, &rf);
1733 if (task_on_rq_queued(p)) {
1734 /* check_preempt_curr() may use rq clock */
1735 update_rq_clock(rq);
1736 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1739 __task_rq_unlock(rq, &rf);
1745 void sched_ttwu_pending(void)
1747 struct rq *rq = this_rq();
1748 struct llist_node *llist = llist_del_all(&rq->wake_list);
1749 struct task_struct *p, *t;
1755 rq_lock_irqsave(rq, &rf);
1756 update_rq_clock(rq);
1758 llist_for_each_entry_safe(p, t, llist, wake_entry)
1759 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
1761 rq_unlock_irqrestore(rq, &rf);
1764 void scheduler_ipi(void)
1767 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1768 * TIF_NEED_RESCHED remotely (for the first time) will also send
1771 preempt_fold_need_resched();
1773 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1777 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1778 * traditionally all their work was done from the interrupt return
1779 * path. Now that we actually do some work, we need to make sure
1782 * Some archs already do call them, luckily irq_enter/exit nest
1785 * Arguably we should visit all archs and update all handlers,
1786 * however a fair share of IPIs are still resched only so this would
1787 * somewhat pessimize the simple resched case.
1790 sched_ttwu_pending();
1793 * Check if someone kicked us for doing the nohz idle load balance.
1795 if (unlikely(got_nohz_idle_kick())) {
1796 this_rq()->idle_balance = 1;
1797 raise_softirq_irqoff(SCHED_SOFTIRQ);
1802 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1804 struct rq *rq = cpu_rq(cpu);
1806 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1808 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1809 if (!set_nr_if_polling(rq->idle))
1810 smp_send_reschedule(cpu);
1812 trace_sched_wake_idle_without_ipi(cpu);
1816 void wake_up_if_idle(int cpu)
1818 struct rq *rq = cpu_rq(cpu);
1823 if (!is_idle_task(rcu_dereference(rq->curr)))
1826 if (set_nr_if_polling(rq->idle)) {
1827 trace_sched_wake_idle_without_ipi(cpu);
1829 rq_lock_irqsave(rq, &rf);
1830 if (is_idle_task(rq->curr))
1831 smp_send_reschedule(cpu);
1832 /* Else CPU is not idle, do nothing here: */
1833 rq_unlock_irqrestore(rq, &rf);
1840 bool cpus_share_cache(int this_cpu, int that_cpu)
1842 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1844 #endif /* CONFIG_SMP */
1846 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1848 struct rq *rq = cpu_rq(cpu);
1851 #if defined(CONFIG_SMP)
1852 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1853 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1854 ttwu_queue_remote(p, cpu, wake_flags);
1860 update_rq_clock(rq);
1861 ttwu_do_activate(rq, p, wake_flags, &rf);
1866 * Notes on Program-Order guarantees on SMP systems.
1870 * The basic program-order guarantee on SMP systems is that when a task [t]
1871 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1872 * execution on its new CPU [c1].
1874 * For migration (of runnable tasks) this is provided by the following means:
1876 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1877 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1878 * rq(c1)->lock (if not at the same time, then in that order).
1879 * C) LOCK of the rq(c1)->lock scheduling in task
1881 * Transitivity guarantees that B happens after A and C after B.
1882 * Note: we only require RCpc transitivity.
1883 * Note: the CPU doing B need not be c0 or c1
1892 * UNLOCK rq(0)->lock
1894 * LOCK rq(0)->lock // orders against CPU0
1896 * UNLOCK rq(0)->lock
1900 * UNLOCK rq(1)->lock
1902 * LOCK rq(1)->lock // orders against CPU2
1905 * UNLOCK rq(1)->lock
1908 * BLOCKING -- aka. SLEEP + WAKEUP
1910 * For blocking we (obviously) need to provide the same guarantee as for
1911 * migration. However the means are completely different as there is no lock
1912 * chain to provide order. Instead we do:
1914 * 1) smp_store_release(X->on_cpu, 0)
1915 * 2) smp_cond_load_acquire(!X->on_cpu)
1919 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1921 * LOCK rq(0)->lock LOCK X->pi_lock
1924 * smp_store_release(X->on_cpu, 0);
1926 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1932 * X->state = RUNNING
1933 * UNLOCK rq(2)->lock
1935 * LOCK rq(2)->lock // orders against CPU1
1938 * UNLOCK rq(2)->lock
1941 * UNLOCK rq(0)->lock
1944 * However; for wakeups there is a second guarantee we must provide, namely we
1945 * must observe the state that lead to our wakeup. That is, not only must our
1946 * task observe its own prior state, it must also observe the stores prior to
1949 * This means that any means of doing remote wakeups must order the CPU doing
1950 * the wakeup against the CPU the task is going to end up running on. This,
1951 * however, is already required for the regular Program-Order guarantee above,
1952 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1957 * try_to_wake_up - wake up a thread
1958 * @p: the thread to be awakened
1959 * @state: the mask of task states that can be woken
1960 * @wake_flags: wake modifier flags (WF_*)
1962 * If (@state & @p->state) @p->state = TASK_RUNNING.
1964 * If the task was not queued/runnable, also place it back on a runqueue.
1966 * Atomic against schedule() which would dequeue a task, also see
1967 * set_current_state().
1969 * Return: %true if @p->state changes (an actual wakeup was done),
1973 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1975 unsigned long flags;
1976 int cpu, success = 0;
1979 * If we are going to wake up a thread waiting for CONDITION we
1980 * need to ensure that CONDITION=1 done by the caller can not be
1981 * reordered with p->state check below. This pairs with mb() in
1982 * set_current_state() the waiting thread does.
1984 raw_spin_lock_irqsave(&p->pi_lock, flags);
1985 smp_mb__after_spinlock();
1986 if (!(p->state & state))
1989 trace_sched_waking(p);
1991 /* We're going to change ->state: */
1996 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1997 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1998 * in smp_cond_load_acquire() below.
2000 * sched_ttwu_pending() try_to_wake_up()
2001 * [S] p->on_rq = 1; [L] P->state
2002 * UNLOCK rq->lock -----.
2006 * LOCK rq->lock -----'
2010 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2012 * Pairs with the UNLOCK+LOCK on rq->lock from the
2013 * last wakeup of our task and the schedule that got our task
2017 if (p->on_rq && ttwu_remote(p, wake_flags))
2022 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2023 * possible to, falsely, observe p->on_cpu == 0.
2025 * One must be running (->on_cpu == 1) in order to remove oneself
2026 * from the runqueue.
2028 * [S] ->on_cpu = 1; [L] ->on_rq
2032 * [S] ->on_rq = 0; [L] ->on_cpu
2034 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2035 * from the consecutive calls to schedule(); the first switching to our
2036 * task, the second putting it to sleep.
2041 * If the owning (remote) CPU is still in the middle of schedule() with
2042 * this task as prev, wait until its done referencing the task.
2044 * Pairs with the smp_store_release() in finish_task().
2046 * This ensures that tasks getting woken will be fully ordered against
2047 * their previous state and preserve Program Order.
2049 smp_cond_load_acquire(&p->on_cpu, !VAL);
2051 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2052 p->state = TASK_WAKING;
2055 delayacct_blkio_end(p);
2056 atomic_dec(&task_rq(p)->nr_iowait);
2059 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2060 if (task_cpu(p) != cpu) {
2061 wake_flags |= WF_MIGRATED;
2062 set_task_cpu(p, cpu);
2065 #else /* CONFIG_SMP */
2068 delayacct_blkio_end(p);
2069 atomic_dec(&task_rq(p)->nr_iowait);
2072 #endif /* CONFIG_SMP */
2074 ttwu_queue(p, cpu, wake_flags);
2076 ttwu_stat(p, cpu, wake_flags);
2078 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2084 * try_to_wake_up_local - try to wake up a local task with rq lock held
2085 * @p: the thread to be awakened
2086 * @rf: request-queue flags for pinning
2088 * Put @p on the run-queue if it's not already there. The caller must
2089 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2092 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2094 struct rq *rq = task_rq(p);
2096 if (WARN_ON_ONCE(rq != this_rq()) ||
2097 WARN_ON_ONCE(p == current))
2100 lockdep_assert_held(&rq->lock);
2102 if (!raw_spin_trylock(&p->pi_lock)) {
2104 * This is OK, because current is on_cpu, which avoids it being
2105 * picked for load-balance and preemption/IRQs are still
2106 * disabled avoiding further scheduler activity on it and we've
2107 * not yet picked a replacement task.
2110 raw_spin_lock(&p->pi_lock);
2114 if (!(p->state & TASK_NORMAL))
2117 trace_sched_waking(p);
2119 if (!task_on_rq_queued(p)) {
2121 delayacct_blkio_end(p);
2122 atomic_dec(&rq->nr_iowait);
2124 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK);
2127 ttwu_do_wakeup(rq, p, 0, rf);
2128 ttwu_stat(p, smp_processor_id(), 0);
2130 raw_spin_unlock(&p->pi_lock);
2134 * wake_up_process - Wake up a specific process
2135 * @p: The process to be woken up.
2137 * Attempt to wake up the nominated process and move it to the set of runnable
2140 * Return: 1 if the process was woken up, 0 if it was already running.
2142 * It may be assumed that this function implies a write memory barrier before
2143 * changing the task state if and only if any tasks are woken up.
2145 int wake_up_process(struct task_struct *p)
2147 return try_to_wake_up(p, TASK_NORMAL, 0);
2149 EXPORT_SYMBOL(wake_up_process);
2151 int wake_up_state(struct task_struct *p, unsigned int state)
2153 return try_to_wake_up(p, state, 0);
2157 * Perform scheduler related setup for a newly forked process p.
2158 * p is forked by current.
2160 * __sched_fork() is basic setup used by init_idle() too:
2162 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2167 p->se.exec_start = 0;
2168 p->se.sum_exec_runtime = 0;
2169 p->se.prev_sum_exec_runtime = 0;
2170 p->se.nr_migrations = 0;
2172 INIT_LIST_HEAD(&p->se.group_node);
2174 #ifdef CONFIG_FAIR_GROUP_SCHED
2175 p->se.cfs_rq = NULL;
2178 #ifdef CONFIG_SCHEDSTATS
2179 /* Even if schedstat is disabled, there should not be garbage */
2180 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2183 RB_CLEAR_NODE(&p->dl.rb_node);
2184 init_dl_task_timer(&p->dl);
2185 init_dl_inactive_task_timer(&p->dl);
2186 __dl_clear_params(p);
2188 INIT_LIST_HEAD(&p->rt.run_list);
2190 p->rt.time_slice = sched_rr_timeslice;
2194 #ifdef CONFIG_PREEMPT_NOTIFIERS
2195 INIT_HLIST_HEAD(&p->preempt_notifiers);
2198 init_numa_balancing(clone_flags, p);
2201 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2203 #ifdef CONFIG_NUMA_BALANCING
2205 void set_numabalancing_state(bool enabled)
2208 static_branch_enable(&sched_numa_balancing);
2210 static_branch_disable(&sched_numa_balancing);
2213 #ifdef CONFIG_PROC_SYSCTL
2214 int sysctl_numa_balancing(struct ctl_table *table, int write,
2215 void __user *buffer, size_t *lenp, loff_t *ppos)
2219 int state = static_branch_likely(&sched_numa_balancing);
2221 if (write && !capable(CAP_SYS_ADMIN))
2226 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2230 set_numabalancing_state(state);
2236 #ifdef CONFIG_SCHEDSTATS
2238 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2239 static bool __initdata __sched_schedstats = false;
2241 static void set_schedstats(bool enabled)
2244 static_branch_enable(&sched_schedstats);
2246 static_branch_disable(&sched_schedstats);
2249 void force_schedstat_enabled(void)
2251 if (!schedstat_enabled()) {
2252 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2253 static_branch_enable(&sched_schedstats);
2257 static int __init setup_schedstats(char *str)
2264 * This code is called before jump labels have been set up, so we can't
2265 * change the static branch directly just yet. Instead set a temporary
2266 * variable so init_schedstats() can do it later.
2268 if (!strcmp(str, "enable")) {
2269 __sched_schedstats = true;
2271 } else if (!strcmp(str, "disable")) {
2272 __sched_schedstats = false;
2277 pr_warn("Unable to parse schedstats=\n");
2281 __setup("schedstats=", setup_schedstats);
2283 static void __init init_schedstats(void)
2285 set_schedstats(__sched_schedstats);
2288 #ifdef CONFIG_PROC_SYSCTL
2289 int sysctl_schedstats(struct ctl_table *table, int write,
2290 void __user *buffer, size_t *lenp, loff_t *ppos)
2294 int state = static_branch_likely(&sched_schedstats);
2296 if (write && !capable(CAP_SYS_ADMIN))
2301 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2305 set_schedstats(state);
2308 #endif /* CONFIG_PROC_SYSCTL */
2309 #else /* !CONFIG_SCHEDSTATS */
2310 static inline void init_schedstats(void) {}
2311 #endif /* CONFIG_SCHEDSTATS */
2314 * fork()/clone()-time setup:
2316 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2318 unsigned long flags;
2319 int cpu = get_cpu();
2321 __sched_fork(clone_flags, p);
2323 * We mark the process as NEW here. This guarantees that
2324 * nobody will actually run it, and a signal or other external
2325 * event cannot wake it up and insert it on the runqueue either.
2327 p->state = TASK_NEW;
2330 * Make sure we do not leak PI boosting priority to the child.
2332 p->prio = current->normal_prio;
2335 * Revert to default priority/policy on fork if requested.
2337 if (unlikely(p->sched_reset_on_fork)) {
2338 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2339 p->policy = SCHED_NORMAL;
2340 p->static_prio = NICE_TO_PRIO(0);
2342 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2343 p->static_prio = NICE_TO_PRIO(0);
2345 p->prio = p->normal_prio = __normal_prio(p);
2346 set_load_weight(p, false);
2349 * We don't need the reset flag anymore after the fork. It has
2350 * fulfilled its duty:
2352 p->sched_reset_on_fork = 0;
2355 if (dl_prio(p->prio)) {
2358 } else if (rt_prio(p->prio)) {
2359 p->sched_class = &rt_sched_class;
2361 p->sched_class = &fair_sched_class;
2364 init_entity_runnable_average(&p->se);
2367 * The child is not yet in the pid-hash so no cgroup attach races,
2368 * and the cgroup is pinned to this child due to cgroup_fork()
2369 * is ran before sched_fork().
2371 * Silence PROVE_RCU.
2373 raw_spin_lock_irqsave(&p->pi_lock, flags);
2375 * We're setting the CPU for the first time, we don't migrate,
2376 * so use __set_task_cpu().
2378 __set_task_cpu(p, cpu);
2379 if (p->sched_class->task_fork)
2380 p->sched_class->task_fork(p);
2381 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2383 #ifdef CONFIG_SCHED_INFO
2384 if (likely(sched_info_on()))
2385 memset(&p->sched_info, 0, sizeof(p->sched_info));
2387 #if defined(CONFIG_SMP)
2390 init_task_preempt_count(p);
2392 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2393 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2400 unsigned long to_ratio(u64 period, u64 runtime)
2402 if (runtime == RUNTIME_INF)
2406 * Doing this here saves a lot of checks in all
2407 * the calling paths, and returning zero seems
2408 * safe for them anyway.
2413 return div64_u64(runtime << BW_SHIFT, period);
2417 * wake_up_new_task - wake up a newly created task for the first time.
2419 * This function will do some initial scheduler statistics housekeeping
2420 * that must be done for every newly created context, then puts the task
2421 * on the runqueue and wakes it.
2423 void wake_up_new_task(struct task_struct *p)
2428 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2429 p->state = TASK_RUNNING;
2432 * Fork balancing, do it here and not earlier because:
2433 * - cpus_allowed can change in the fork path
2434 * - any previously selected CPU might disappear through hotplug
2436 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2437 * as we're not fully set-up yet.
2439 p->recent_used_cpu = task_cpu(p);
2440 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2442 rq = __task_rq_lock(p, &rf);
2443 update_rq_clock(rq);
2444 post_init_entity_util_avg(&p->se);
2446 activate_task(rq, p, ENQUEUE_NOCLOCK);
2447 p->on_rq = TASK_ON_RQ_QUEUED;
2448 trace_sched_wakeup_new(p);
2449 check_preempt_curr(rq, p, WF_FORK);
2451 if (p->sched_class->task_woken) {
2453 * Nothing relies on rq->lock after this, so its fine to
2456 rq_unpin_lock(rq, &rf);
2457 p->sched_class->task_woken(rq, p);
2458 rq_repin_lock(rq, &rf);
2461 task_rq_unlock(rq, p, &rf);
2464 #ifdef CONFIG_PREEMPT_NOTIFIERS
2466 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2468 void preempt_notifier_inc(void)
2470 static_branch_inc(&preempt_notifier_key);
2472 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2474 void preempt_notifier_dec(void)
2476 static_branch_dec(&preempt_notifier_key);
2478 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2481 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2482 * @notifier: notifier struct to register
2484 void preempt_notifier_register(struct preempt_notifier *notifier)
2486 if (!static_branch_unlikely(&preempt_notifier_key))
2487 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2489 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2491 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2494 * preempt_notifier_unregister - no longer interested in preemption notifications
2495 * @notifier: notifier struct to unregister
2497 * This is *not* safe to call from within a preemption notifier.
2499 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2501 hlist_del(¬ifier->link);
2503 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2505 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2507 struct preempt_notifier *notifier;
2509 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2510 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2513 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2515 if (static_branch_unlikely(&preempt_notifier_key))
2516 __fire_sched_in_preempt_notifiers(curr);
2520 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2521 struct task_struct *next)
2523 struct preempt_notifier *notifier;
2525 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2526 notifier->ops->sched_out(notifier, next);
2529 static __always_inline void
2530 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2531 struct task_struct *next)
2533 if (static_branch_unlikely(&preempt_notifier_key))
2534 __fire_sched_out_preempt_notifiers(curr, next);
2537 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2539 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2544 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2545 struct task_struct *next)
2549 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2551 static inline void prepare_task(struct task_struct *next)
2555 * Claim the task as running, we do this before switching to it
2556 * such that any running task will have this set.
2562 static inline void finish_task(struct task_struct *prev)
2566 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2567 * We must ensure this doesn't happen until the switch is completely
2570 * In particular, the load of prev->state in finish_task_switch() must
2571 * happen before this.
2573 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2575 smp_store_release(&prev->on_cpu, 0);
2580 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
2583 * Since the runqueue lock will be released by the next
2584 * task (which is an invalid locking op but in the case
2585 * of the scheduler it's an obvious special-case), so we
2586 * do an early lockdep release here:
2588 rq_unpin_lock(rq, rf);
2589 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2590 #ifdef CONFIG_DEBUG_SPINLOCK
2591 /* this is a valid case when another task releases the spinlock */
2592 rq->lock.owner = next;
2596 static inline void finish_lock_switch(struct rq *rq)
2599 * If we are tracking spinlock dependencies then we have to
2600 * fix up the runqueue lock - which gets 'carried over' from
2601 * prev into current:
2603 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
2604 raw_spin_unlock_irq(&rq->lock);
2608 * NOP if the arch has not defined these:
2611 #ifndef prepare_arch_switch
2612 # define prepare_arch_switch(next) do { } while (0)
2615 #ifndef finish_arch_post_lock_switch
2616 # define finish_arch_post_lock_switch() do { } while (0)
2620 * prepare_task_switch - prepare to switch tasks
2621 * @rq: the runqueue preparing to switch
2622 * @prev: the current task that is being switched out
2623 * @next: the task we are going to switch to.
2625 * This is called with the rq lock held and interrupts off. It must
2626 * be paired with a subsequent finish_task_switch after the context
2629 * prepare_task_switch sets up locking and calls architecture specific
2633 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2634 struct task_struct *next)
2636 sched_info_switch(rq, prev, next);
2637 perf_event_task_sched_out(prev, next);
2639 fire_sched_out_preempt_notifiers(prev, next);
2641 prepare_arch_switch(next);
2645 * finish_task_switch - clean up after a task-switch
2646 * @prev: the thread we just switched away from.
2648 * finish_task_switch must be called after the context switch, paired
2649 * with a prepare_task_switch call before the context switch.
2650 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2651 * and do any other architecture-specific cleanup actions.
2653 * Note that we may have delayed dropping an mm in context_switch(). If
2654 * so, we finish that here outside of the runqueue lock. (Doing it
2655 * with the lock held can cause deadlocks; see schedule() for
2658 * The context switch have flipped the stack from under us and restored the
2659 * local variables which were saved when this task called schedule() in the
2660 * past. prev == current is still correct but we need to recalculate this_rq
2661 * because prev may have moved to another CPU.
2663 static struct rq *finish_task_switch(struct task_struct *prev)
2664 __releases(rq->lock)
2666 struct rq *rq = this_rq();
2667 struct mm_struct *mm = rq->prev_mm;
2671 * The previous task will have left us with a preempt_count of 2
2672 * because it left us after:
2675 * preempt_disable(); // 1
2677 * raw_spin_lock_irq(&rq->lock) // 2
2679 * Also, see FORK_PREEMPT_COUNT.
2681 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2682 "corrupted preempt_count: %s/%d/0x%x\n",
2683 current->comm, current->pid, preempt_count()))
2684 preempt_count_set(FORK_PREEMPT_COUNT);
2689 * A task struct has one reference for the use as "current".
2690 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2691 * schedule one last time. The schedule call will never return, and
2692 * the scheduled task must drop that reference.
2694 * We must observe prev->state before clearing prev->on_cpu (in
2695 * finish_task), otherwise a concurrent wakeup can get prev
2696 * running on another CPU and we could rave with its RUNNING -> DEAD
2697 * transition, resulting in a double drop.
2699 prev_state = prev->state;
2700 vtime_task_switch(prev);
2701 perf_event_task_sched_in(prev, current);
2703 finish_lock_switch(rq);
2704 finish_arch_post_lock_switch();
2706 fire_sched_in_preempt_notifiers(current);
2708 * When switching through a kernel thread, the loop in
2709 * membarrier_{private,global}_expedited() may have observed that
2710 * kernel thread and not issued an IPI. It is therefore possible to
2711 * schedule between user->kernel->user threads without passing though
2712 * switch_mm(). Membarrier requires a barrier after storing to
2713 * rq->curr, before returning to userspace, so provide them here:
2715 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
2716 * provided by mmdrop(),
2717 * - a sync_core for SYNC_CORE.
2720 membarrier_mm_sync_core_before_usermode(mm);
2723 if (unlikely(prev_state & (TASK_DEAD|TASK_PARKED))) {
2724 switch (prev_state) {
2726 if (prev->sched_class->task_dead)
2727 prev->sched_class->task_dead(prev);
2730 * Remove function-return probe instances associated with this
2731 * task and put them back on the free list.
2733 kprobe_flush_task(prev);
2735 /* Task is done with its stack. */
2736 put_task_stack(prev);
2738 put_task_struct(prev);
2742 kthread_park_complete(prev);
2747 tick_nohz_task_switch();
2753 /* rq->lock is NOT held, but preemption is disabled */
2754 static void __balance_callback(struct rq *rq)
2756 struct callback_head *head, *next;
2757 void (*func)(struct rq *rq);
2758 unsigned long flags;
2760 raw_spin_lock_irqsave(&rq->lock, flags);
2761 head = rq->balance_callback;
2762 rq->balance_callback = NULL;
2764 func = (void (*)(struct rq *))head->func;
2771 raw_spin_unlock_irqrestore(&rq->lock, flags);
2774 static inline void balance_callback(struct rq *rq)
2776 if (unlikely(rq->balance_callback))
2777 __balance_callback(rq);
2782 static inline void balance_callback(struct rq *rq)
2789 * schedule_tail - first thing a freshly forked thread must call.
2790 * @prev: the thread we just switched away from.
2792 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2793 __releases(rq->lock)
2798 * New tasks start with FORK_PREEMPT_COUNT, see there and
2799 * finish_task_switch() for details.
2801 * finish_task_switch() will drop rq->lock() and lower preempt_count
2802 * and the preempt_enable() will end up enabling preemption (on
2803 * PREEMPT_COUNT kernels).
2806 rq = finish_task_switch(prev);
2807 balance_callback(rq);
2810 if (current->set_child_tid)
2811 put_user(task_pid_vnr(current), current->set_child_tid);
2815 * context_switch - switch to the new MM and the new thread's register state.
2817 static __always_inline struct rq *
2818 context_switch(struct rq *rq, struct task_struct *prev,
2819 struct task_struct *next, struct rq_flags *rf)
2821 struct mm_struct *mm, *oldmm;
2823 prepare_task_switch(rq, prev, next);
2826 oldmm = prev->active_mm;
2828 * For paravirt, this is coupled with an exit in switch_to to
2829 * combine the page table reload and the switch backend into
2832 arch_start_context_switch(prev);
2835 * If mm is non-NULL, we pass through switch_mm(). If mm is
2836 * NULL, we will pass through mmdrop() in finish_task_switch().
2837 * Both of these contain the full memory barrier required by
2838 * membarrier after storing to rq->curr, before returning to
2842 next->active_mm = oldmm;
2844 enter_lazy_tlb(oldmm, next);
2846 switch_mm_irqs_off(oldmm, mm, next);
2849 prev->active_mm = NULL;
2850 rq->prev_mm = oldmm;
2853 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2855 prepare_lock_switch(rq, next, rf);
2857 /* Here we just switch the register state and the stack. */
2858 switch_to(prev, next, prev);
2861 return finish_task_switch(prev);
2865 * nr_running and nr_context_switches:
2867 * externally visible scheduler statistics: current number of runnable
2868 * threads, total number of context switches performed since bootup.
2870 unsigned long nr_running(void)
2872 unsigned long i, sum = 0;
2874 for_each_online_cpu(i)
2875 sum += cpu_rq(i)->nr_running;
2881 * Check if only the current task is running on the CPU.
2883 * Caution: this function does not check that the caller has disabled
2884 * preemption, thus the result might have a time-of-check-to-time-of-use
2885 * race. The caller is responsible to use it correctly, for example:
2887 * - from a non-preemptable section (of course)
2889 * - from a thread that is bound to a single CPU
2891 * - in a loop with very short iterations (e.g. a polling loop)
2893 bool single_task_running(void)
2895 return raw_rq()->nr_running == 1;
2897 EXPORT_SYMBOL(single_task_running);
2899 unsigned long long nr_context_switches(void)
2902 unsigned long long sum = 0;
2904 for_each_possible_cpu(i)
2905 sum += cpu_rq(i)->nr_switches;
2911 * IO-wait accounting, and how its mostly bollocks (on SMP).
2913 * The idea behind IO-wait account is to account the idle time that we could
2914 * have spend running if it were not for IO. That is, if we were to improve the
2915 * storage performance, we'd have a proportional reduction in IO-wait time.
2917 * This all works nicely on UP, where, when a task blocks on IO, we account
2918 * idle time as IO-wait, because if the storage were faster, it could've been
2919 * running and we'd not be idle.
2921 * This has been extended to SMP, by doing the same for each CPU. This however
2924 * Imagine for instance the case where two tasks block on one CPU, only the one
2925 * CPU will have IO-wait accounted, while the other has regular idle. Even
2926 * though, if the storage were faster, both could've ran at the same time,
2927 * utilising both CPUs.
2929 * This means, that when looking globally, the current IO-wait accounting on
2930 * SMP is a lower bound, by reason of under accounting.
2932 * Worse, since the numbers are provided per CPU, they are sometimes
2933 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2934 * associated with any one particular CPU, it can wake to another CPU than it
2935 * blocked on. This means the per CPU IO-wait number is meaningless.
2937 * Task CPU affinities can make all that even more 'interesting'.
2940 unsigned long nr_iowait(void)
2942 unsigned long i, sum = 0;
2944 for_each_possible_cpu(i)
2945 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2951 * Consumers of these two interfaces, like for example the cpufreq menu
2952 * governor are using nonsensical data. Boosting frequency for a CPU that has
2953 * IO-wait which might not even end up running the task when it does become
2957 unsigned long nr_iowait_cpu(int cpu)
2959 struct rq *this = cpu_rq(cpu);
2960 return atomic_read(&this->nr_iowait);
2963 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2965 struct rq *rq = this_rq();
2966 *nr_waiters = atomic_read(&rq->nr_iowait);
2967 *load = rq->load.weight;
2973 * sched_exec - execve() is a valuable balancing opportunity, because at
2974 * this point the task has the smallest effective memory and cache footprint.
2976 void sched_exec(void)
2978 struct task_struct *p = current;
2979 unsigned long flags;
2982 raw_spin_lock_irqsave(&p->pi_lock, flags);
2983 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2984 if (dest_cpu == smp_processor_id())
2987 if (likely(cpu_active(dest_cpu))) {
2988 struct migration_arg arg = { p, dest_cpu };
2990 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2991 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2995 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3000 DEFINE_PER_CPU(struct kernel_stat, kstat);
3001 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3003 EXPORT_PER_CPU_SYMBOL(kstat);
3004 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3007 * The function fair_sched_class.update_curr accesses the struct curr
3008 * and its field curr->exec_start; when called from task_sched_runtime(),
3009 * we observe a high rate of cache misses in practice.
3010 * Prefetching this data results in improved performance.
3012 static inline void prefetch_curr_exec_start(struct task_struct *p)
3014 #ifdef CONFIG_FAIR_GROUP_SCHED
3015 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3017 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3020 prefetch(&curr->exec_start);
3024 * Return accounted runtime for the task.
3025 * In case the task is currently running, return the runtime plus current's
3026 * pending runtime that have not been accounted yet.
3028 unsigned long long task_sched_runtime(struct task_struct *p)
3034 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3036 * 64-bit doesn't need locks to atomically read a 64-bit value.
3037 * So we have a optimization chance when the task's delta_exec is 0.
3038 * Reading ->on_cpu is racy, but this is ok.
3040 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3041 * If we race with it entering CPU, unaccounted time is 0. This is
3042 * indistinguishable from the read occurring a few cycles earlier.
3043 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3044 * been accounted, so we're correct here as well.
3046 if (!p->on_cpu || !task_on_rq_queued(p))
3047 return p->se.sum_exec_runtime;
3050 rq = task_rq_lock(p, &rf);
3052 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3053 * project cycles that may never be accounted to this
3054 * thread, breaking clock_gettime().
3056 if (task_current(rq, p) && task_on_rq_queued(p)) {
3057 prefetch_curr_exec_start(p);
3058 update_rq_clock(rq);
3059 p->sched_class->update_curr(rq);
3061 ns = p->se.sum_exec_runtime;
3062 task_rq_unlock(rq, p, &rf);
3068 * This function gets called by the timer code, with HZ frequency.
3069 * We call it with interrupts disabled.
3071 void scheduler_tick(void)
3073 int cpu = smp_processor_id();
3074 struct rq *rq = cpu_rq(cpu);
3075 struct task_struct *curr = rq->curr;
3082 update_rq_clock(rq);
3083 curr->sched_class->task_tick(rq, curr, 0);
3084 cpu_load_update_active(rq);
3085 calc_global_load_tick(rq);
3089 perf_event_task_tick();
3092 rq->idle_balance = idle_cpu(cpu);
3093 trigger_load_balance(rq);
3097 #ifdef CONFIG_NO_HZ_FULL
3101 struct delayed_work work;
3104 static struct tick_work __percpu *tick_work_cpu;
3106 static void sched_tick_remote(struct work_struct *work)
3108 struct delayed_work *dwork = to_delayed_work(work);
3109 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3110 int cpu = twork->cpu;
3111 struct rq *rq = cpu_rq(cpu);
3115 * Handle the tick only if it appears the remote CPU is running in full
3116 * dynticks mode. The check is racy by nature, but missing a tick or
3117 * having one too much is no big deal because the scheduler tick updates
3118 * statistics and checks timeslices in a time-independent way, regardless
3119 * of when exactly it is running.
3121 if (!idle_cpu(cpu) && tick_nohz_tick_stopped_cpu(cpu)) {
3122 struct task_struct *curr;
3125 rq_lock_irq(rq, &rf);
3126 update_rq_clock(rq);
3128 delta = rq_clock_task(rq) - curr->se.exec_start;
3131 * Make sure the next tick runs within a reasonable
3134 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3135 curr->sched_class->task_tick(rq, curr, 0);
3136 rq_unlock_irq(rq, &rf);
3140 * Run the remote tick once per second (1Hz). This arbitrary
3141 * frequency is large enough to avoid overload but short enough
3142 * to keep scheduler internal stats reasonably up to date.
3144 queue_delayed_work(system_unbound_wq, dwork, HZ);
3147 static void sched_tick_start(int cpu)
3149 struct tick_work *twork;
3151 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3154 WARN_ON_ONCE(!tick_work_cpu);
3156 twork = per_cpu_ptr(tick_work_cpu, cpu);
3158 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3159 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3162 #ifdef CONFIG_HOTPLUG_CPU
3163 static void sched_tick_stop(int cpu)
3165 struct tick_work *twork;
3167 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3170 WARN_ON_ONCE(!tick_work_cpu);
3172 twork = per_cpu_ptr(tick_work_cpu, cpu);
3173 cancel_delayed_work_sync(&twork->work);
3175 #endif /* CONFIG_HOTPLUG_CPU */
3177 int __init sched_tick_offload_init(void)
3179 tick_work_cpu = alloc_percpu(struct tick_work);
3180 BUG_ON(!tick_work_cpu);
3185 #else /* !CONFIG_NO_HZ_FULL */
3186 static inline void sched_tick_start(int cpu) { }
3187 static inline void sched_tick_stop(int cpu) { }
3190 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3191 defined(CONFIG_PREEMPT_TRACER))
3193 * If the value passed in is equal to the current preempt count
3194 * then we just disabled preemption. Start timing the latency.
3196 static inline void preempt_latency_start(int val)
3198 if (preempt_count() == val) {
3199 unsigned long ip = get_lock_parent_ip();
3200 #ifdef CONFIG_DEBUG_PREEMPT
3201 current->preempt_disable_ip = ip;
3203 trace_preempt_off(CALLER_ADDR0, ip);
3207 void preempt_count_add(int val)
3209 #ifdef CONFIG_DEBUG_PREEMPT
3213 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3216 __preempt_count_add(val);
3217 #ifdef CONFIG_DEBUG_PREEMPT
3219 * Spinlock count overflowing soon?
3221 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3224 preempt_latency_start(val);
3226 EXPORT_SYMBOL(preempt_count_add);
3227 NOKPROBE_SYMBOL(preempt_count_add);
3230 * If the value passed in equals to the current preempt count
3231 * then we just enabled preemption. Stop timing the latency.
3233 static inline void preempt_latency_stop(int val)
3235 if (preempt_count() == val)
3236 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3239 void preempt_count_sub(int val)
3241 #ifdef CONFIG_DEBUG_PREEMPT
3245 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3248 * Is the spinlock portion underflowing?
3250 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3251 !(preempt_count() & PREEMPT_MASK)))
3255 preempt_latency_stop(val);
3256 __preempt_count_sub(val);
3258 EXPORT_SYMBOL(preempt_count_sub);
3259 NOKPROBE_SYMBOL(preempt_count_sub);
3262 static inline void preempt_latency_start(int val) { }
3263 static inline void preempt_latency_stop(int val) { }
3266 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3268 #ifdef CONFIG_DEBUG_PREEMPT
3269 return p->preempt_disable_ip;
3276 * Print scheduling while atomic bug:
3278 static noinline void __schedule_bug(struct task_struct *prev)
3280 /* Save this before calling printk(), since that will clobber it */
3281 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3283 if (oops_in_progress)
3286 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3287 prev->comm, prev->pid, preempt_count());
3289 debug_show_held_locks(prev);
3291 if (irqs_disabled())
3292 print_irqtrace_events(prev);
3293 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3294 && in_atomic_preempt_off()) {
3295 pr_err("Preemption disabled at:");
3296 print_ip_sym(preempt_disable_ip);
3300 panic("scheduling while atomic\n");
3303 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3307 * Various schedule()-time debugging checks and statistics:
3309 static inline void schedule_debug(struct task_struct *prev)
3311 #ifdef CONFIG_SCHED_STACK_END_CHECK
3312 if (task_stack_end_corrupted(prev))
3313 panic("corrupted stack end detected inside scheduler\n");
3316 if (unlikely(in_atomic_preempt_off())) {
3317 __schedule_bug(prev);
3318 preempt_count_set(PREEMPT_DISABLED);
3322 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3324 schedstat_inc(this_rq()->sched_count);
3328 * Pick up the highest-prio task:
3330 static inline struct task_struct *
3331 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3333 const struct sched_class *class;
3334 struct task_struct *p;
3337 * Optimization: we know that if all tasks are in the fair class we can
3338 * call that function directly, but only if the @prev task wasn't of a
3339 * higher scheduling class, because otherwise those loose the
3340 * opportunity to pull in more work from other CPUs.
3342 if (likely((prev->sched_class == &idle_sched_class ||
3343 prev->sched_class == &fair_sched_class) &&
3344 rq->nr_running == rq->cfs.h_nr_running)) {
3346 p = fair_sched_class.pick_next_task(rq, prev, rf);
3347 if (unlikely(p == RETRY_TASK))
3350 /* Assumes fair_sched_class->next == idle_sched_class */
3352 p = idle_sched_class.pick_next_task(rq, prev, rf);
3358 for_each_class(class) {
3359 p = class->pick_next_task(rq, prev, rf);
3361 if (unlikely(p == RETRY_TASK))
3367 /* The idle class should always have a runnable task: */
3372 * __schedule() is the main scheduler function.
3374 * The main means of driving the scheduler and thus entering this function are:
3376 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3378 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3379 * paths. For example, see arch/x86/entry_64.S.
3381 * To drive preemption between tasks, the scheduler sets the flag in timer
3382 * interrupt handler scheduler_tick().
3384 * 3. Wakeups don't really cause entry into schedule(). They add a
3385 * task to the run-queue and that's it.
3387 * Now, if the new task added to the run-queue preempts the current
3388 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3389 * called on the nearest possible occasion:
3391 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3393 * - in syscall or exception context, at the next outmost
3394 * preempt_enable(). (this might be as soon as the wake_up()'s
3397 * - in IRQ context, return from interrupt-handler to
3398 * preemptible context
3400 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3403 * - cond_resched() call
3404 * - explicit schedule() call
3405 * - return from syscall or exception to user-space
3406 * - return from interrupt-handler to user-space
3408 * WARNING: must be called with preemption disabled!
3410 static void __sched notrace __schedule(bool preempt)
3412 struct task_struct *prev, *next;
3413 unsigned long *switch_count;
3418 cpu = smp_processor_id();
3422 schedule_debug(prev);
3424 if (sched_feat(HRTICK))
3427 local_irq_disable();
3428 rcu_note_context_switch(preempt);
3431 * Make sure that signal_pending_state()->signal_pending() below
3432 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3433 * done by the caller to avoid the race with signal_wake_up().
3435 * The membarrier system call requires a full memory barrier
3436 * after coming from user-space, before storing to rq->curr.
3439 smp_mb__after_spinlock();
3441 /* Promote REQ to ACT */
3442 rq->clock_update_flags <<= 1;
3443 update_rq_clock(rq);
3445 switch_count = &prev->nivcsw;
3446 if (!preempt && prev->state) {
3447 if (unlikely(signal_pending_state(prev->state, prev))) {
3448 prev->state = TASK_RUNNING;
3450 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3453 if (prev->in_iowait) {
3454 atomic_inc(&rq->nr_iowait);
3455 delayacct_blkio_start();
3459 * If a worker went to sleep, notify and ask workqueue
3460 * whether it wants to wake up a task to maintain
3463 if (prev->flags & PF_WQ_WORKER) {
3464 struct task_struct *to_wakeup;
3466 to_wakeup = wq_worker_sleeping(prev);
3468 try_to_wake_up_local(to_wakeup, &rf);
3471 switch_count = &prev->nvcsw;
3474 next = pick_next_task(rq, prev, &rf);
3475 clear_tsk_need_resched(prev);
3476 clear_preempt_need_resched();
3478 if (likely(prev != next)) {
3482 * The membarrier system call requires each architecture
3483 * to have a full memory barrier after updating
3484 * rq->curr, before returning to user-space.
3486 * Here are the schemes providing that barrier on the
3487 * various architectures:
3488 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3489 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3490 * - finish_lock_switch() for weakly-ordered
3491 * architectures where spin_unlock is a full barrier,
3492 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3493 * is a RELEASE barrier),
3497 trace_sched_switch(preempt, prev, next);
3499 /* Also unlocks the rq: */
3500 rq = context_switch(rq, prev, next, &rf);
3502 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3503 rq_unlock_irq(rq, &rf);
3506 balance_callback(rq);
3509 void __noreturn do_task_dead(void)
3511 /* Causes final put_task_struct in finish_task_switch(): */
3512 set_special_state(TASK_DEAD);
3514 /* Tell freezer to ignore us: */
3515 current->flags |= PF_NOFREEZE;
3520 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3525 static inline void sched_submit_work(struct task_struct *tsk)
3527 if (!tsk->state || tsk_is_pi_blocked(tsk))
3530 * If we are going to sleep and we have plugged IO queued,
3531 * make sure to submit it to avoid deadlocks.
3533 if (blk_needs_flush_plug(tsk))
3534 blk_schedule_flush_plug(tsk);
3537 asmlinkage __visible void __sched schedule(void)
3539 struct task_struct *tsk = current;
3541 sched_submit_work(tsk);
3545 sched_preempt_enable_no_resched();
3546 } while (need_resched());
3548 EXPORT_SYMBOL(schedule);
3551 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3552 * state (have scheduled out non-voluntarily) by making sure that all
3553 * tasks have either left the run queue or have gone into user space.
3554 * As idle tasks do not do either, they must not ever be preempted
3555 * (schedule out non-voluntarily).
3557 * schedule_idle() is similar to schedule_preempt_disable() except that it
3558 * never enables preemption because it does not call sched_submit_work().
3560 void __sched schedule_idle(void)
3563 * As this skips calling sched_submit_work(), which the idle task does
3564 * regardless because that function is a nop when the task is in a
3565 * TASK_RUNNING state, make sure this isn't used someplace that the
3566 * current task can be in any other state. Note, idle is always in the
3567 * TASK_RUNNING state.
3569 WARN_ON_ONCE(current->state);
3572 } while (need_resched());
3575 #ifdef CONFIG_CONTEXT_TRACKING
3576 asmlinkage __visible void __sched schedule_user(void)
3579 * If we come here after a random call to set_need_resched(),
3580 * or we have been woken up remotely but the IPI has not yet arrived,
3581 * we haven't yet exited the RCU idle mode. Do it here manually until
3582 * we find a better solution.
3584 * NB: There are buggy callers of this function. Ideally we
3585 * should warn if prev_state != CONTEXT_USER, but that will trigger
3586 * too frequently to make sense yet.
3588 enum ctx_state prev_state = exception_enter();
3590 exception_exit(prev_state);
3595 * schedule_preempt_disabled - called with preemption disabled
3597 * Returns with preemption disabled. Note: preempt_count must be 1
3599 void __sched schedule_preempt_disabled(void)
3601 sched_preempt_enable_no_resched();
3606 static void __sched notrace preempt_schedule_common(void)
3610 * Because the function tracer can trace preempt_count_sub()
3611 * and it also uses preempt_enable/disable_notrace(), if
3612 * NEED_RESCHED is set, the preempt_enable_notrace() called
3613 * by the function tracer will call this function again and
3614 * cause infinite recursion.
3616 * Preemption must be disabled here before the function
3617 * tracer can trace. Break up preempt_disable() into two
3618 * calls. One to disable preemption without fear of being
3619 * traced. The other to still record the preemption latency,
3620 * which can also be traced by the function tracer.
3622 preempt_disable_notrace();
3623 preempt_latency_start(1);
3625 preempt_latency_stop(1);
3626 preempt_enable_no_resched_notrace();
3629 * Check again in case we missed a preemption opportunity
3630 * between schedule and now.
3632 } while (need_resched());
3635 #ifdef CONFIG_PREEMPT
3637 * this is the entry point to schedule() from in-kernel preemption
3638 * off of preempt_enable. Kernel preemptions off return from interrupt
3639 * occur there and call schedule directly.
3641 asmlinkage __visible void __sched notrace preempt_schedule(void)
3644 * If there is a non-zero preempt_count or interrupts are disabled,
3645 * we do not want to preempt the current task. Just return..
3647 if (likely(!preemptible()))
3650 preempt_schedule_common();
3652 NOKPROBE_SYMBOL(preempt_schedule);
3653 EXPORT_SYMBOL(preempt_schedule);
3656 * preempt_schedule_notrace - preempt_schedule called by tracing
3658 * The tracing infrastructure uses preempt_enable_notrace to prevent
3659 * recursion and tracing preempt enabling caused by the tracing
3660 * infrastructure itself. But as tracing can happen in areas coming
3661 * from userspace or just about to enter userspace, a preempt enable
3662 * can occur before user_exit() is called. This will cause the scheduler
3663 * to be called when the system is still in usermode.
3665 * To prevent this, the preempt_enable_notrace will use this function
3666 * instead of preempt_schedule() to exit user context if needed before
3667 * calling the scheduler.
3669 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3671 enum ctx_state prev_ctx;
3673 if (likely(!preemptible()))
3678 * Because the function tracer can trace preempt_count_sub()
3679 * and it also uses preempt_enable/disable_notrace(), if
3680 * NEED_RESCHED is set, the preempt_enable_notrace() called
3681 * by the function tracer will call this function again and
3682 * cause infinite recursion.
3684 * Preemption must be disabled here before the function
3685 * tracer can trace. Break up preempt_disable() into two
3686 * calls. One to disable preemption without fear of being
3687 * traced. The other to still record the preemption latency,
3688 * which can also be traced by the function tracer.
3690 preempt_disable_notrace();
3691 preempt_latency_start(1);
3693 * Needs preempt disabled in case user_exit() is traced
3694 * and the tracer calls preempt_enable_notrace() causing
3695 * an infinite recursion.
3697 prev_ctx = exception_enter();
3699 exception_exit(prev_ctx);
3701 preempt_latency_stop(1);
3702 preempt_enable_no_resched_notrace();
3703 } while (need_resched());
3705 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3707 #endif /* CONFIG_PREEMPT */
3710 * this is the entry point to schedule() from kernel preemption
3711 * off of irq context.
3712 * Note, that this is called and return with irqs disabled. This will
3713 * protect us against recursive calling from irq.
3715 asmlinkage __visible void __sched preempt_schedule_irq(void)
3717 enum ctx_state prev_state;
3719 /* Catch callers which need to be fixed */
3720 BUG_ON(preempt_count() || !irqs_disabled());
3722 prev_state = exception_enter();
3728 local_irq_disable();
3729 sched_preempt_enable_no_resched();
3730 } while (need_resched());
3732 exception_exit(prev_state);
3735 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
3738 return try_to_wake_up(curr->private, mode, wake_flags);
3740 EXPORT_SYMBOL(default_wake_function);
3742 #ifdef CONFIG_RT_MUTEXES
3744 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
3747 prio = min(prio, pi_task->prio);
3752 static inline int rt_effective_prio(struct task_struct *p, int prio)
3754 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3756 return __rt_effective_prio(pi_task, prio);
3760 * rt_mutex_setprio - set the current priority of a task
3762 * @pi_task: donor task
3764 * This function changes the 'effective' priority of a task. It does
3765 * not touch ->normal_prio like __setscheduler().
3767 * Used by the rt_mutex code to implement priority inheritance
3768 * logic. Call site only calls if the priority of the task changed.
3770 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
3772 int prio, oldprio, queued, running, queue_flag =
3773 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
3774 const struct sched_class *prev_class;
3778 /* XXX used to be waiter->prio, not waiter->task->prio */
3779 prio = __rt_effective_prio(pi_task, p->normal_prio);
3782 * If nothing changed; bail early.
3784 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
3787 rq = __task_rq_lock(p, &rf);
3788 update_rq_clock(rq);
3790 * Set under pi_lock && rq->lock, such that the value can be used under
3793 * Note that there is loads of tricky to make this pointer cache work
3794 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3795 * ensure a task is de-boosted (pi_task is set to NULL) before the
3796 * task is allowed to run again (and can exit). This ensures the pointer
3797 * points to a blocked task -- which guaratees the task is present.
3799 p->pi_top_task = pi_task;
3802 * For FIFO/RR we only need to set prio, if that matches we're done.
3804 if (prio == p->prio && !dl_prio(prio))
3808 * Idle task boosting is a nono in general. There is one
3809 * exception, when PREEMPT_RT and NOHZ is active:
3811 * The idle task calls get_next_timer_interrupt() and holds
3812 * the timer wheel base->lock on the CPU and another CPU wants
3813 * to access the timer (probably to cancel it). We can safely
3814 * ignore the boosting request, as the idle CPU runs this code
3815 * with interrupts disabled and will complete the lock
3816 * protected section without being interrupted. So there is no
3817 * real need to boost.
3819 if (unlikely(p == rq->idle)) {
3820 WARN_ON(p != rq->curr);
3821 WARN_ON(p->pi_blocked_on);
3825 trace_sched_pi_setprio(p, pi_task);
3828 if (oldprio == prio)
3829 queue_flag &= ~DEQUEUE_MOVE;
3831 prev_class = p->sched_class;
3832 queued = task_on_rq_queued(p);
3833 running = task_current(rq, p);
3835 dequeue_task(rq, p, queue_flag);
3837 put_prev_task(rq, p);
3840 * Boosting condition are:
3841 * 1. -rt task is running and holds mutex A
3842 * --> -dl task blocks on mutex A
3844 * 2. -dl task is running and holds mutex A
3845 * --> -dl task blocks on mutex A and could preempt the
3848 if (dl_prio(prio)) {
3849 if (!dl_prio(p->normal_prio) ||
3850 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3851 p->dl.dl_boosted = 1;
3852 queue_flag |= ENQUEUE_REPLENISH;
3854 p->dl.dl_boosted = 0;
3855 p->sched_class = &dl_sched_class;
3856 } else if (rt_prio(prio)) {
3857 if (dl_prio(oldprio))
3858 p->dl.dl_boosted = 0;
3860 queue_flag |= ENQUEUE_HEAD;
3861 p->sched_class = &rt_sched_class;
3863 if (dl_prio(oldprio))
3864 p->dl.dl_boosted = 0;
3865 if (rt_prio(oldprio))
3867 p->sched_class = &fair_sched_class;
3873 enqueue_task(rq, p, queue_flag);
3875 set_curr_task(rq, p);
3877 check_class_changed(rq, p, prev_class, oldprio);
3879 /* Avoid rq from going away on us: */
3881 __task_rq_unlock(rq, &rf);
3883 balance_callback(rq);
3887 static inline int rt_effective_prio(struct task_struct *p, int prio)
3893 void set_user_nice(struct task_struct *p, long nice)
3895 bool queued, running;
3896 int old_prio, delta;
3900 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3903 * We have to be careful, if called from sys_setpriority(),
3904 * the task might be in the middle of scheduling on another CPU.
3906 rq = task_rq_lock(p, &rf);
3907 update_rq_clock(rq);
3910 * The RT priorities are set via sched_setscheduler(), but we still
3911 * allow the 'normal' nice value to be set - but as expected
3912 * it wont have any effect on scheduling until the task is
3913 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3915 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3916 p->static_prio = NICE_TO_PRIO(nice);
3919 queued = task_on_rq_queued(p);
3920 running = task_current(rq, p);
3922 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
3924 put_prev_task(rq, p);
3926 p->static_prio = NICE_TO_PRIO(nice);
3927 set_load_weight(p, true);
3929 p->prio = effective_prio(p);
3930 delta = p->prio - old_prio;
3933 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
3935 * If the task increased its priority or is running and
3936 * lowered its priority, then reschedule its CPU:
3938 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3942 set_curr_task(rq, p);
3944 task_rq_unlock(rq, p, &rf);
3946 EXPORT_SYMBOL(set_user_nice);
3949 * can_nice - check if a task can reduce its nice value
3953 int can_nice(const struct task_struct *p, const int nice)
3955 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3956 int nice_rlim = nice_to_rlimit(nice);
3958 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3959 capable(CAP_SYS_NICE));
3962 #ifdef __ARCH_WANT_SYS_NICE
3965 * sys_nice - change the priority of the current process.
3966 * @increment: priority increment
3968 * sys_setpriority is a more generic, but much slower function that
3969 * does similar things.
3971 SYSCALL_DEFINE1(nice, int, increment)
3976 * Setpriority might change our priority at the same moment.
3977 * We don't have to worry. Conceptually one call occurs first
3978 * and we have a single winner.
3980 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3981 nice = task_nice(current) + increment;
3983 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3984 if (increment < 0 && !can_nice(current, nice))
3987 retval = security_task_setnice(current, nice);
3991 set_user_nice(current, nice);
3998 * task_prio - return the priority value of a given task.
3999 * @p: the task in question.
4001 * Return: The priority value as seen by users in /proc.
4002 * RT tasks are offset by -200. Normal tasks are centered
4003 * around 0, value goes from -16 to +15.
4005 int task_prio(const struct task_struct *p)
4007 return p->prio - MAX_RT_PRIO;
4011 * idle_cpu - is a given CPU idle currently?
4012 * @cpu: the processor in question.
4014 * Return: 1 if the CPU is currently idle. 0 otherwise.
4016 int idle_cpu(int cpu)
4018 struct rq *rq = cpu_rq(cpu);
4020 if (rq->curr != rq->idle)
4027 if (!llist_empty(&rq->wake_list))
4035 * available_idle_cpu - is a given CPU idle for enqueuing work.
4036 * @cpu: the CPU in question.
4038 * Return: 1 if the CPU is currently idle. 0 otherwise.
4040 int available_idle_cpu(int cpu)
4045 if (vcpu_is_preempted(cpu))
4052 * idle_task - return the idle task for a given CPU.
4053 * @cpu: the processor in question.
4055 * Return: The idle task for the CPU @cpu.
4057 struct task_struct *idle_task(int cpu)
4059 return cpu_rq(cpu)->idle;
4063 * find_process_by_pid - find a process with a matching PID value.
4064 * @pid: the pid in question.
4066 * The task of @pid, if found. %NULL otherwise.
4068 static struct task_struct *find_process_by_pid(pid_t pid)
4070 return pid ? find_task_by_vpid(pid) : current;
4074 * sched_setparam() passes in -1 for its policy, to let the functions
4075 * it calls know not to change it.
4077 #define SETPARAM_POLICY -1
4079 static void __setscheduler_params(struct task_struct *p,
4080 const struct sched_attr *attr)
4082 int policy = attr->sched_policy;
4084 if (policy == SETPARAM_POLICY)
4089 if (dl_policy(policy))
4090 __setparam_dl(p, attr);
4091 else if (fair_policy(policy))
4092 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4095 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4096 * !rt_policy. Always setting this ensures that things like
4097 * getparam()/getattr() don't report silly values for !rt tasks.
4099 p->rt_priority = attr->sched_priority;
4100 p->normal_prio = normal_prio(p);
4101 set_load_weight(p, true);
4104 /* Actually do priority change: must hold pi & rq lock. */
4105 static void __setscheduler(struct rq *rq, struct task_struct *p,
4106 const struct sched_attr *attr, bool keep_boost)
4108 __setscheduler_params(p, attr);
4111 * Keep a potential priority boosting if called from
4112 * sched_setscheduler().
4114 p->prio = normal_prio(p);
4116 p->prio = rt_effective_prio(p, p->prio);
4118 if (dl_prio(p->prio))
4119 p->sched_class = &dl_sched_class;
4120 else if (rt_prio(p->prio))
4121 p->sched_class = &rt_sched_class;
4123 p->sched_class = &fair_sched_class;
4127 * Check the target process has a UID that matches the current process's:
4129 static bool check_same_owner(struct task_struct *p)
4131 const struct cred *cred = current_cred(), *pcred;
4135 pcred = __task_cred(p);
4136 match = (uid_eq(cred->euid, pcred->euid) ||
4137 uid_eq(cred->euid, pcred->uid));
4142 static int __sched_setscheduler(struct task_struct *p,
4143 const struct sched_attr *attr,
4146 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4147 MAX_RT_PRIO - 1 - attr->sched_priority;
4148 int retval, oldprio, oldpolicy = -1, queued, running;
4149 int new_effective_prio, policy = attr->sched_policy;
4150 const struct sched_class *prev_class;
4153 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4156 /* The pi code expects interrupts enabled */
4157 BUG_ON(pi && in_interrupt());
4159 /* Double check policy once rq lock held: */
4161 reset_on_fork = p->sched_reset_on_fork;
4162 policy = oldpolicy = p->policy;
4164 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4166 if (!valid_policy(policy))
4170 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4174 * Valid priorities for SCHED_FIFO and SCHED_RR are
4175 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4176 * SCHED_BATCH and SCHED_IDLE is 0.
4178 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4179 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4181 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4182 (rt_policy(policy) != (attr->sched_priority != 0)))
4186 * Allow unprivileged RT tasks to decrease priority:
4188 if (user && !capable(CAP_SYS_NICE)) {
4189 if (fair_policy(policy)) {
4190 if (attr->sched_nice < task_nice(p) &&
4191 !can_nice(p, attr->sched_nice))
4195 if (rt_policy(policy)) {
4196 unsigned long rlim_rtprio =
4197 task_rlimit(p, RLIMIT_RTPRIO);
4199 /* Can't set/change the rt policy: */
4200 if (policy != p->policy && !rlim_rtprio)
4203 /* Can't increase priority: */
4204 if (attr->sched_priority > p->rt_priority &&
4205 attr->sched_priority > rlim_rtprio)
4210 * Can't set/change SCHED_DEADLINE policy at all for now
4211 * (safest behavior); in the future we would like to allow
4212 * unprivileged DL tasks to increase their relative deadline
4213 * or reduce their runtime (both ways reducing utilization)
4215 if (dl_policy(policy))
4219 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4220 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4222 if (idle_policy(p->policy) && !idle_policy(policy)) {
4223 if (!can_nice(p, task_nice(p)))
4227 /* Can't change other user's priorities: */
4228 if (!check_same_owner(p))
4231 /* Normal users shall not reset the sched_reset_on_fork flag: */
4232 if (p->sched_reset_on_fork && !reset_on_fork)
4237 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4240 retval = security_task_setscheduler(p);
4246 * Make sure no PI-waiters arrive (or leave) while we are
4247 * changing the priority of the task:
4249 * To be able to change p->policy safely, the appropriate
4250 * runqueue lock must be held.
4252 rq = task_rq_lock(p, &rf);
4253 update_rq_clock(rq);
4256 * Changing the policy of the stop threads its a very bad idea:
4258 if (p == rq->stop) {
4259 task_rq_unlock(rq, p, &rf);
4264 * If not changing anything there's no need to proceed further,
4265 * but store a possible modification of reset_on_fork.
4267 if (unlikely(policy == p->policy)) {
4268 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4270 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4272 if (dl_policy(policy) && dl_param_changed(p, attr))
4275 p->sched_reset_on_fork = reset_on_fork;
4276 task_rq_unlock(rq, p, &rf);
4282 #ifdef CONFIG_RT_GROUP_SCHED
4284 * Do not allow realtime tasks into groups that have no runtime
4287 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4288 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4289 !task_group_is_autogroup(task_group(p))) {
4290 task_rq_unlock(rq, p, &rf);
4295 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4296 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4297 cpumask_t *span = rq->rd->span;
4300 * Don't allow tasks with an affinity mask smaller than
4301 * the entire root_domain to become SCHED_DEADLINE. We
4302 * will also fail if there's no bandwidth available.
4304 if (!cpumask_subset(span, &p->cpus_allowed) ||
4305 rq->rd->dl_bw.bw == 0) {
4306 task_rq_unlock(rq, p, &rf);
4313 /* Re-check policy now with rq lock held: */
4314 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4315 policy = oldpolicy = -1;
4316 task_rq_unlock(rq, p, &rf);
4321 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4322 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4325 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4326 task_rq_unlock(rq, p, &rf);
4330 p->sched_reset_on_fork = reset_on_fork;
4335 * Take priority boosted tasks into account. If the new
4336 * effective priority is unchanged, we just store the new
4337 * normal parameters and do not touch the scheduler class and
4338 * the runqueue. This will be done when the task deboost
4341 new_effective_prio = rt_effective_prio(p, newprio);
4342 if (new_effective_prio == oldprio)
4343 queue_flags &= ~DEQUEUE_MOVE;
4346 queued = task_on_rq_queued(p);
4347 running = task_current(rq, p);
4349 dequeue_task(rq, p, queue_flags);
4351 put_prev_task(rq, p);
4353 prev_class = p->sched_class;
4354 __setscheduler(rq, p, attr, pi);
4358 * We enqueue to tail when the priority of a task is
4359 * increased (user space view).
4361 if (oldprio < p->prio)
4362 queue_flags |= ENQUEUE_HEAD;
4364 enqueue_task(rq, p, queue_flags);
4367 set_curr_task(rq, p);
4369 check_class_changed(rq, p, prev_class, oldprio);
4371 /* Avoid rq from going away on us: */
4373 task_rq_unlock(rq, p, &rf);
4376 rt_mutex_adjust_pi(p);
4378 /* Run balance callbacks after we've adjusted the PI chain: */
4379 balance_callback(rq);
4385 static int _sched_setscheduler(struct task_struct *p, int policy,
4386 const struct sched_param *param, bool check)
4388 struct sched_attr attr = {
4389 .sched_policy = policy,
4390 .sched_priority = param->sched_priority,
4391 .sched_nice = PRIO_TO_NICE(p->static_prio),
4394 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4395 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4396 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4397 policy &= ~SCHED_RESET_ON_FORK;
4398 attr.sched_policy = policy;
4401 return __sched_setscheduler(p, &attr, check, true);
4404 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4405 * @p: the task in question.
4406 * @policy: new policy.
4407 * @param: structure containing the new RT priority.
4409 * Return: 0 on success. An error code otherwise.
4411 * NOTE that the task may be already dead.
4413 int sched_setscheduler(struct task_struct *p, int policy,
4414 const struct sched_param *param)
4416 return _sched_setscheduler(p, policy, param, true);
4418 EXPORT_SYMBOL_GPL(sched_setscheduler);
4420 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4422 return __sched_setscheduler(p, attr, true, true);
4424 EXPORT_SYMBOL_GPL(sched_setattr);
4426 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
4428 return __sched_setscheduler(p, attr, false, true);
4432 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4433 * @p: the task in question.
4434 * @policy: new policy.
4435 * @param: structure containing the new RT priority.
4437 * Just like sched_setscheduler, only don't bother checking if the
4438 * current context has permission. For example, this is needed in
4439 * stop_machine(): we create temporary high priority worker threads,
4440 * but our caller might not have that capability.
4442 * Return: 0 on success. An error code otherwise.
4444 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4445 const struct sched_param *param)
4447 return _sched_setscheduler(p, policy, param, false);
4449 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4452 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4454 struct sched_param lparam;
4455 struct task_struct *p;
4458 if (!param || pid < 0)
4460 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4465 p = find_process_by_pid(pid);
4467 retval = sched_setscheduler(p, policy, &lparam);
4474 * Mimics kernel/events/core.c perf_copy_attr().
4476 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4481 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4484 /* Zero the full structure, so that a short copy will be nice: */
4485 memset(attr, 0, sizeof(*attr));
4487 ret = get_user(size, &uattr->size);
4491 /* Bail out on silly large: */
4492 if (size > PAGE_SIZE)
4495 /* ABI compatibility quirk: */
4497 size = SCHED_ATTR_SIZE_VER0;
4499 if (size < SCHED_ATTR_SIZE_VER0)
4503 * If we're handed a bigger struct than we know of,
4504 * ensure all the unknown bits are 0 - i.e. new
4505 * user-space does not rely on any kernel feature
4506 * extensions we dont know about yet.
4508 if (size > sizeof(*attr)) {
4509 unsigned char __user *addr;
4510 unsigned char __user *end;
4513 addr = (void __user *)uattr + sizeof(*attr);
4514 end = (void __user *)uattr + size;
4516 for (; addr < end; addr++) {
4517 ret = get_user(val, addr);
4523 size = sizeof(*attr);
4526 ret = copy_from_user(attr, uattr, size);
4531 * XXX: Do we want to be lenient like existing syscalls; or do we want
4532 * to be strict and return an error on out-of-bounds values?
4534 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4539 put_user(sizeof(*attr), &uattr->size);
4544 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4545 * @pid: the pid in question.
4546 * @policy: new policy.
4547 * @param: structure containing the new RT priority.
4549 * Return: 0 on success. An error code otherwise.
4551 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4556 return do_sched_setscheduler(pid, policy, param);
4560 * sys_sched_setparam - set/change the RT priority of a thread
4561 * @pid: the pid in question.
4562 * @param: structure containing the new RT priority.
4564 * Return: 0 on success. An error code otherwise.
4566 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4568 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4572 * sys_sched_setattr - same as above, but with extended sched_attr
4573 * @pid: the pid in question.
4574 * @uattr: structure containing the extended parameters.
4575 * @flags: for future extension.
4577 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4578 unsigned int, flags)
4580 struct sched_attr attr;
4581 struct task_struct *p;
4584 if (!uattr || pid < 0 || flags)
4587 retval = sched_copy_attr(uattr, &attr);
4591 if ((int)attr.sched_policy < 0)
4596 p = find_process_by_pid(pid);
4598 retval = sched_setattr(p, &attr);
4605 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4606 * @pid: the pid in question.
4608 * Return: On success, the policy of the thread. Otherwise, a negative error
4611 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4613 struct task_struct *p;
4621 p = find_process_by_pid(pid);
4623 retval = security_task_getscheduler(p);
4626 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4633 * sys_sched_getparam - get the RT priority of a thread
4634 * @pid: the pid in question.
4635 * @param: structure containing the RT priority.
4637 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4640 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4642 struct sched_param lp = { .sched_priority = 0 };
4643 struct task_struct *p;
4646 if (!param || pid < 0)
4650 p = find_process_by_pid(pid);
4655 retval = security_task_getscheduler(p);
4659 if (task_has_rt_policy(p))
4660 lp.sched_priority = p->rt_priority;
4664 * This one might sleep, we cannot do it with a spinlock held ...
4666 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4675 static int sched_read_attr(struct sched_attr __user *uattr,
4676 struct sched_attr *attr,
4681 if (!access_ok(VERIFY_WRITE, uattr, usize))
4685 * If we're handed a smaller struct than we know of,
4686 * ensure all the unknown bits are 0 - i.e. old
4687 * user-space does not get uncomplete information.
4689 if (usize < sizeof(*attr)) {
4690 unsigned char *addr;
4693 addr = (void *)attr + usize;
4694 end = (void *)attr + sizeof(*attr);
4696 for (; addr < end; addr++) {
4704 ret = copy_to_user(uattr, attr, attr->size);
4712 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4713 * @pid: the pid in question.
4714 * @uattr: structure containing the extended parameters.
4715 * @size: sizeof(attr) for fwd/bwd comp.
4716 * @flags: for future extension.
4718 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4719 unsigned int, size, unsigned int, flags)
4721 struct sched_attr attr = {
4722 .size = sizeof(struct sched_attr),
4724 struct task_struct *p;
4727 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4728 size < SCHED_ATTR_SIZE_VER0 || flags)
4732 p = find_process_by_pid(pid);
4737 retval = security_task_getscheduler(p);
4741 attr.sched_policy = p->policy;
4742 if (p->sched_reset_on_fork)
4743 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4744 if (task_has_dl_policy(p))
4745 __getparam_dl(p, &attr);
4746 else if (task_has_rt_policy(p))
4747 attr.sched_priority = p->rt_priority;
4749 attr.sched_nice = task_nice(p);
4753 retval = sched_read_attr(uattr, &attr, size);
4761 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4763 cpumask_var_t cpus_allowed, new_mask;
4764 struct task_struct *p;
4769 p = find_process_by_pid(pid);
4775 /* Prevent p going away */
4779 if (p->flags & PF_NO_SETAFFINITY) {
4783 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4787 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4789 goto out_free_cpus_allowed;
4792 if (!check_same_owner(p)) {
4794 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4796 goto out_free_new_mask;
4801 retval = security_task_setscheduler(p);
4803 goto out_free_new_mask;
4806 cpuset_cpus_allowed(p, cpus_allowed);
4807 cpumask_and(new_mask, in_mask, cpus_allowed);
4810 * Since bandwidth control happens on root_domain basis,
4811 * if admission test is enabled, we only admit -deadline
4812 * tasks allowed to run on all the CPUs in the task's
4816 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4818 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4821 goto out_free_new_mask;
4827 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4830 cpuset_cpus_allowed(p, cpus_allowed);
4831 if (!cpumask_subset(new_mask, cpus_allowed)) {
4833 * We must have raced with a concurrent cpuset
4834 * update. Just reset the cpus_allowed to the
4835 * cpuset's cpus_allowed
4837 cpumask_copy(new_mask, cpus_allowed);
4842 free_cpumask_var(new_mask);
4843 out_free_cpus_allowed:
4844 free_cpumask_var(cpus_allowed);
4850 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4851 struct cpumask *new_mask)
4853 if (len < cpumask_size())
4854 cpumask_clear(new_mask);
4855 else if (len > cpumask_size())
4856 len = cpumask_size();
4858 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4862 * sys_sched_setaffinity - set the CPU affinity of a process
4863 * @pid: pid of the process
4864 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4865 * @user_mask_ptr: user-space pointer to the new CPU mask
4867 * Return: 0 on success. An error code otherwise.
4869 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4870 unsigned long __user *, user_mask_ptr)
4872 cpumask_var_t new_mask;
4875 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4878 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4880 retval = sched_setaffinity(pid, new_mask);
4881 free_cpumask_var(new_mask);
4885 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4887 struct task_struct *p;
4888 unsigned long flags;
4894 p = find_process_by_pid(pid);
4898 retval = security_task_getscheduler(p);
4902 raw_spin_lock_irqsave(&p->pi_lock, flags);
4903 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4904 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4913 * sys_sched_getaffinity - get the CPU affinity of a process
4914 * @pid: pid of the process
4915 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4916 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4918 * Return: size of CPU mask copied to user_mask_ptr on success. An
4919 * error code otherwise.
4921 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4922 unsigned long __user *, user_mask_ptr)
4927 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4929 if (len & (sizeof(unsigned long)-1))
4932 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4935 ret = sched_getaffinity(pid, mask);
4937 unsigned int retlen = min(len, cpumask_size());
4939 if (copy_to_user(user_mask_ptr, mask, retlen))
4944 free_cpumask_var(mask);
4950 * sys_sched_yield - yield the current processor to other threads.
4952 * This function yields the current CPU to other tasks. If there are no
4953 * other threads running on this CPU then this function will return.
4957 static void do_sched_yield(void)
4962 local_irq_disable();
4966 schedstat_inc(rq->yld_count);
4967 current->sched_class->yield_task(rq);
4970 * Since we are going to call schedule() anyway, there's
4971 * no need to preempt or enable interrupts:
4975 sched_preempt_enable_no_resched();
4980 SYSCALL_DEFINE0(sched_yield)
4986 #ifndef CONFIG_PREEMPT
4987 int __sched _cond_resched(void)
4989 if (should_resched(0)) {
4990 preempt_schedule_common();
4996 EXPORT_SYMBOL(_cond_resched);
5000 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5001 * call schedule, and on return reacquire the lock.
5003 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5004 * operations here to prevent schedule() from being called twice (once via
5005 * spin_unlock(), once by hand).
5007 int __cond_resched_lock(spinlock_t *lock)
5009 int resched = should_resched(PREEMPT_LOCK_OFFSET);
5012 lockdep_assert_held(lock);
5014 if (spin_needbreak(lock) || resched) {
5017 preempt_schedule_common();
5025 EXPORT_SYMBOL(__cond_resched_lock);
5028 * yield - yield the current processor to other threads.
5030 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5032 * The scheduler is at all times free to pick the calling task as the most
5033 * eligible task to run, if removing the yield() call from your code breaks
5034 * it, its already broken.
5036 * Typical broken usage is:
5041 * where one assumes that yield() will let 'the other' process run that will
5042 * make event true. If the current task is a SCHED_FIFO task that will never
5043 * happen. Never use yield() as a progress guarantee!!
5045 * If you want to use yield() to wait for something, use wait_event().
5046 * If you want to use yield() to be 'nice' for others, use cond_resched().
5047 * If you still want to use yield(), do not!
5049 void __sched yield(void)
5051 set_current_state(TASK_RUNNING);
5054 EXPORT_SYMBOL(yield);
5057 * yield_to - yield the current processor to another thread in
5058 * your thread group, or accelerate that thread toward the
5059 * processor it's on.
5061 * @preempt: whether task preemption is allowed or not
5063 * It's the caller's job to ensure that the target task struct
5064 * can't go away on us before we can do any checks.
5067 * true (>0) if we indeed boosted the target task.
5068 * false (0) if we failed to boost the target.
5069 * -ESRCH if there's no task to yield to.
5071 int __sched yield_to(struct task_struct *p, bool preempt)
5073 struct task_struct *curr = current;
5074 struct rq *rq, *p_rq;
5075 unsigned long flags;
5078 local_irq_save(flags);
5084 * If we're the only runnable task on the rq and target rq also
5085 * has only one task, there's absolutely no point in yielding.
5087 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5092 double_rq_lock(rq, p_rq);
5093 if (task_rq(p) != p_rq) {
5094 double_rq_unlock(rq, p_rq);
5098 if (!curr->sched_class->yield_to_task)
5101 if (curr->sched_class != p->sched_class)
5104 if (task_running(p_rq, p) || p->state)
5107 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5109 schedstat_inc(rq->yld_count);
5111 * Make p's CPU reschedule; pick_next_entity takes care of
5114 if (preempt && rq != p_rq)
5119 double_rq_unlock(rq, p_rq);
5121 local_irq_restore(flags);
5128 EXPORT_SYMBOL_GPL(yield_to);
5130 int io_schedule_prepare(void)
5132 int old_iowait = current->in_iowait;
5134 current->in_iowait = 1;
5135 blk_schedule_flush_plug(current);
5140 void io_schedule_finish(int token)
5142 current->in_iowait = token;
5146 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5147 * that process accounting knows that this is a task in IO wait state.
5149 long __sched io_schedule_timeout(long timeout)
5154 token = io_schedule_prepare();
5155 ret = schedule_timeout(timeout);
5156 io_schedule_finish(token);
5160 EXPORT_SYMBOL(io_schedule_timeout);
5162 void io_schedule(void)
5166 token = io_schedule_prepare();
5168 io_schedule_finish(token);
5170 EXPORT_SYMBOL(io_schedule);
5173 * sys_sched_get_priority_max - return maximum RT priority.
5174 * @policy: scheduling class.
5176 * Return: On success, this syscall returns the maximum
5177 * rt_priority that can be used by a given scheduling class.
5178 * On failure, a negative error code is returned.
5180 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5187 ret = MAX_USER_RT_PRIO-1;
5189 case SCHED_DEADLINE:
5200 * sys_sched_get_priority_min - return minimum RT priority.
5201 * @policy: scheduling class.
5203 * Return: On success, this syscall returns the minimum
5204 * rt_priority that can be used by a given scheduling class.
5205 * On failure, a negative error code is returned.
5207 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5216 case SCHED_DEADLINE:
5225 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5227 struct task_struct *p;
5228 unsigned int time_slice;
5238 p = find_process_by_pid(pid);
5242 retval = security_task_getscheduler(p);
5246 rq = task_rq_lock(p, &rf);
5248 if (p->sched_class->get_rr_interval)
5249 time_slice = p->sched_class->get_rr_interval(rq, p);
5250 task_rq_unlock(rq, p, &rf);
5253 jiffies_to_timespec64(time_slice, t);
5262 * sys_sched_rr_get_interval - return the default timeslice of a process.
5263 * @pid: pid of the process.
5264 * @interval: userspace pointer to the timeslice value.
5266 * this syscall writes the default timeslice value of a given process
5267 * into the user-space timespec buffer. A value of '0' means infinity.
5269 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5272 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5273 struct timespec __user *, interval)
5275 struct timespec64 t;
5276 int retval = sched_rr_get_interval(pid, &t);
5279 retval = put_timespec64(&t, interval);
5284 #ifdef CONFIG_COMPAT
5285 COMPAT_SYSCALL_DEFINE2(sched_rr_get_interval,
5287 struct compat_timespec __user *, interval)
5289 struct timespec64 t;
5290 int retval = sched_rr_get_interval(pid, &t);
5293 retval = compat_put_timespec64(&t, interval);
5298 void sched_show_task(struct task_struct *p)
5300 unsigned long free = 0;
5303 if (!try_get_task_stack(p))
5306 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5308 if (p->state == TASK_RUNNING)
5309 printk(KERN_CONT " running task ");
5310 #ifdef CONFIG_DEBUG_STACK_USAGE
5311 free = stack_not_used(p);
5316 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5318 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5319 task_pid_nr(p), ppid,
5320 (unsigned long)task_thread_info(p)->flags);
5322 print_worker_info(KERN_INFO, p);
5323 show_stack(p, NULL);
5326 EXPORT_SYMBOL_GPL(sched_show_task);
5329 state_filter_match(unsigned long state_filter, struct task_struct *p)
5331 /* no filter, everything matches */
5335 /* filter, but doesn't match */
5336 if (!(p->state & state_filter))
5340 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5343 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5350 void show_state_filter(unsigned long state_filter)
5352 struct task_struct *g, *p;
5354 #if BITS_PER_LONG == 32
5356 " task PC stack pid father\n");
5359 " task PC stack pid father\n");
5362 for_each_process_thread(g, p) {
5364 * reset the NMI-timeout, listing all files on a slow
5365 * console might take a lot of time:
5366 * Also, reset softlockup watchdogs on all CPUs, because
5367 * another CPU might be blocked waiting for us to process
5370 touch_nmi_watchdog();
5371 touch_all_softlockup_watchdogs();
5372 if (state_filter_match(state_filter, p))
5376 #ifdef CONFIG_SCHED_DEBUG
5378 sysrq_sched_debug_show();
5382 * Only show locks if all tasks are dumped:
5385 debug_show_all_locks();
5389 * init_idle - set up an idle thread for a given CPU
5390 * @idle: task in question
5391 * @cpu: CPU the idle task belongs to
5393 * NOTE: this function does not set the idle thread's NEED_RESCHED
5394 * flag, to make booting more robust.
5396 void init_idle(struct task_struct *idle, int cpu)
5398 struct rq *rq = cpu_rq(cpu);
5399 unsigned long flags;
5401 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5402 raw_spin_lock(&rq->lock);
5404 __sched_fork(0, idle);
5405 idle->state = TASK_RUNNING;
5406 idle->se.exec_start = sched_clock();
5407 idle->flags |= PF_IDLE;
5409 kasan_unpoison_task_stack(idle);
5413 * Its possible that init_idle() gets called multiple times on a task,
5414 * in that case do_set_cpus_allowed() will not do the right thing.
5416 * And since this is boot we can forgo the serialization.
5418 set_cpus_allowed_common(idle, cpumask_of(cpu));
5421 * We're having a chicken and egg problem, even though we are
5422 * holding rq->lock, the CPU isn't yet set to this CPU so the
5423 * lockdep check in task_group() will fail.
5425 * Similar case to sched_fork(). / Alternatively we could
5426 * use task_rq_lock() here and obtain the other rq->lock.
5431 __set_task_cpu(idle, cpu);
5434 rq->curr = rq->idle = idle;
5435 idle->on_rq = TASK_ON_RQ_QUEUED;
5439 raw_spin_unlock(&rq->lock);
5440 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5442 /* Set the preempt count _outside_ the spinlocks! */
5443 init_idle_preempt_count(idle, cpu);
5446 * The idle tasks have their own, simple scheduling class:
5448 idle->sched_class = &idle_sched_class;
5449 ftrace_graph_init_idle_task(idle, cpu);
5450 vtime_init_idle(idle, cpu);
5452 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5458 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5459 const struct cpumask *trial)
5463 if (!cpumask_weight(cur))
5466 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5471 int task_can_attach(struct task_struct *p,
5472 const struct cpumask *cs_cpus_allowed)
5477 * Kthreads which disallow setaffinity shouldn't be moved
5478 * to a new cpuset; we don't want to change their CPU
5479 * affinity and isolating such threads by their set of
5480 * allowed nodes is unnecessary. Thus, cpusets are not
5481 * applicable for such threads. This prevents checking for
5482 * success of set_cpus_allowed_ptr() on all attached tasks
5483 * before cpus_allowed may be changed.
5485 if (p->flags & PF_NO_SETAFFINITY) {
5490 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5492 ret = dl_task_can_attach(p, cs_cpus_allowed);
5498 bool sched_smp_initialized __read_mostly;
5500 #ifdef CONFIG_NUMA_BALANCING
5501 /* Migrate current task p to target_cpu */
5502 int migrate_task_to(struct task_struct *p, int target_cpu)
5504 struct migration_arg arg = { p, target_cpu };
5505 int curr_cpu = task_cpu(p);
5507 if (curr_cpu == target_cpu)
5510 if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5513 /* TODO: This is not properly updating schedstats */
5515 trace_sched_move_numa(p, curr_cpu, target_cpu);
5516 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5520 * Requeue a task on a given node and accurately track the number of NUMA
5521 * tasks on the runqueues
5523 void sched_setnuma(struct task_struct *p, int nid)
5525 bool queued, running;
5529 rq = task_rq_lock(p, &rf);
5530 queued = task_on_rq_queued(p);
5531 running = task_current(rq, p);
5534 dequeue_task(rq, p, DEQUEUE_SAVE);
5536 put_prev_task(rq, p);
5538 p->numa_preferred_nid = nid;
5541 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5543 set_curr_task(rq, p);
5544 task_rq_unlock(rq, p, &rf);
5546 #endif /* CONFIG_NUMA_BALANCING */
5548 #ifdef CONFIG_HOTPLUG_CPU
5550 * Ensure that the idle task is using init_mm right before its CPU goes
5553 void idle_task_exit(void)
5555 struct mm_struct *mm = current->active_mm;
5557 BUG_ON(cpu_online(smp_processor_id()));
5559 if (mm != &init_mm) {
5560 switch_mm(mm, &init_mm, current);
5561 current->active_mm = &init_mm;
5562 finish_arch_post_lock_switch();
5568 * Since this CPU is going 'away' for a while, fold any nr_active delta
5569 * we might have. Assumes we're called after migrate_tasks() so that the
5570 * nr_active count is stable. We need to take the teardown thread which
5571 * is calling this into account, so we hand in adjust = 1 to the load
5574 * Also see the comment "Global load-average calculations".
5576 static void calc_load_migrate(struct rq *rq)
5578 long delta = calc_load_fold_active(rq, 1);
5580 atomic_long_add(delta, &calc_load_tasks);
5583 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5587 static const struct sched_class fake_sched_class = {
5588 .put_prev_task = put_prev_task_fake,
5591 static struct task_struct fake_task = {
5593 * Avoid pull_{rt,dl}_task()
5595 .prio = MAX_PRIO + 1,
5596 .sched_class = &fake_sched_class,
5600 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5601 * try_to_wake_up()->select_task_rq().
5603 * Called with rq->lock held even though we'er in stop_machine() and
5604 * there's no concurrency possible, we hold the required locks anyway
5605 * because of lock validation efforts.
5607 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5609 struct rq *rq = dead_rq;
5610 struct task_struct *next, *stop = rq->stop;
5611 struct rq_flags orf = *rf;
5615 * Fudge the rq selection such that the below task selection loop
5616 * doesn't get stuck on the currently eligible stop task.
5618 * We're currently inside stop_machine() and the rq is either stuck
5619 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5620 * either way we should never end up calling schedule() until we're
5626 * put_prev_task() and pick_next_task() sched
5627 * class method both need to have an up-to-date
5628 * value of rq->clock[_task]
5630 update_rq_clock(rq);
5634 * There's this thread running, bail when that's the only
5637 if (rq->nr_running == 1)
5641 * pick_next_task() assumes pinned rq->lock:
5643 next = pick_next_task(rq, &fake_task, rf);
5645 put_prev_task(rq, next);
5648 * Rules for changing task_struct::cpus_allowed are holding
5649 * both pi_lock and rq->lock, such that holding either
5650 * stabilizes the mask.
5652 * Drop rq->lock is not quite as disastrous as it usually is
5653 * because !cpu_active at this point, which means load-balance
5654 * will not interfere. Also, stop-machine.
5657 raw_spin_lock(&next->pi_lock);
5661 * Since we're inside stop-machine, _nothing_ should have
5662 * changed the task, WARN if weird stuff happened, because in
5663 * that case the above rq->lock drop is a fail too.
5665 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5666 raw_spin_unlock(&next->pi_lock);
5670 /* Find suitable destination for @next, with force if needed. */
5671 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5672 rq = __migrate_task(rq, rf, next, dest_cpu);
5673 if (rq != dead_rq) {
5679 raw_spin_unlock(&next->pi_lock);
5684 #endif /* CONFIG_HOTPLUG_CPU */
5686 void set_rq_online(struct rq *rq)
5689 const struct sched_class *class;
5691 cpumask_set_cpu(rq->cpu, rq->rd->online);
5694 for_each_class(class) {
5695 if (class->rq_online)
5696 class->rq_online(rq);
5701 void set_rq_offline(struct rq *rq)
5704 const struct sched_class *class;
5706 for_each_class(class) {
5707 if (class->rq_offline)
5708 class->rq_offline(rq);
5711 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5716 static void set_cpu_rq_start_time(unsigned int cpu)
5718 struct rq *rq = cpu_rq(cpu);
5720 rq->age_stamp = sched_clock_cpu(cpu);
5724 * used to mark begin/end of suspend/resume:
5726 static int num_cpus_frozen;
5729 * Update cpusets according to cpu_active mask. If cpusets are
5730 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5731 * around partition_sched_domains().
5733 * If we come here as part of a suspend/resume, don't touch cpusets because we
5734 * want to restore it back to its original state upon resume anyway.
5736 static void cpuset_cpu_active(void)
5738 if (cpuhp_tasks_frozen) {
5740 * num_cpus_frozen tracks how many CPUs are involved in suspend
5741 * resume sequence. As long as this is not the last online
5742 * operation in the resume sequence, just build a single sched
5743 * domain, ignoring cpusets.
5745 partition_sched_domains(1, NULL, NULL);
5746 if (--num_cpus_frozen)
5749 * This is the last CPU online operation. So fall through and
5750 * restore the original sched domains by considering the
5751 * cpuset configurations.
5753 cpuset_force_rebuild();
5755 cpuset_update_active_cpus();
5758 static int cpuset_cpu_inactive(unsigned int cpu)
5760 if (!cpuhp_tasks_frozen) {
5761 if (dl_cpu_busy(cpu))
5763 cpuset_update_active_cpus();
5766 partition_sched_domains(1, NULL, NULL);
5771 int sched_cpu_activate(unsigned int cpu)
5773 struct rq *rq = cpu_rq(cpu);
5776 set_cpu_active(cpu, true);
5778 if (sched_smp_initialized) {
5779 sched_domains_numa_masks_set(cpu);
5780 cpuset_cpu_active();
5784 * Put the rq online, if not already. This happens:
5786 * 1) In the early boot process, because we build the real domains
5787 * after all CPUs have been brought up.
5789 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5792 rq_lock_irqsave(rq, &rf);
5794 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5797 rq_unlock_irqrestore(rq, &rf);
5799 update_max_interval();
5804 int sched_cpu_deactivate(unsigned int cpu)
5808 set_cpu_active(cpu, false);
5810 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5811 * users of this state to go away such that all new such users will
5814 * Do sync before park smpboot threads to take care the rcu boost case.
5816 synchronize_rcu_mult(call_rcu, call_rcu_sched);
5818 if (!sched_smp_initialized)
5821 ret = cpuset_cpu_inactive(cpu);
5823 set_cpu_active(cpu, true);
5826 sched_domains_numa_masks_clear(cpu);
5830 static void sched_rq_cpu_starting(unsigned int cpu)
5832 struct rq *rq = cpu_rq(cpu);
5834 rq->calc_load_update = calc_load_update;
5835 update_max_interval();
5838 int sched_cpu_starting(unsigned int cpu)
5840 set_cpu_rq_start_time(cpu);
5841 sched_rq_cpu_starting(cpu);
5842 sched_tick_start(cpu);
5846 #ifdef CONFIG_HOTPLUG_CPU
5847 int sched_cpu_dying(unsigned int cpu)
5849 struct rq *rq = cpu_rq(cpu);
5852 /* Handle pending wakeups and then migrate everything off */
5853 sched_ttwu_pending();
5854 sched_tick_stop(cpu);
5856 rq_lock_irqsave(rq, &rf);
5858 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5861 migrate_tasks(rq, &rf);
5862 BUG_ON(rq->nr_running != 1);
5863 rq_unlock_irqrestore(rq, &rf);
5865 calc_load_migrate(rq);
5866 update_max_interval();
5867 nohz_balance_exit_idle(rq);
5873 #ifdef CONFIG_SCHED_SMT
5874 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5876 static void sched_init_smt(void)
5879 * We've enumerated all CPUs and will assume that if any CPU
5880 * has SMT siblings, CPU0 will too.
5882 if (cpumask_weight(cpu_smt_mask(0)) > 1)
5883 static_branch_enable(&sched_smt_present);
5886 static inline void sched_init_smt(void) { }
5889 void __init sched_init_smp(void)
5894 * There's no userspace yet to cause hotplug operations; hence all the
5895 * CPU masks are stable and all blatant races in the below code cannot
5898 mutex_lock(&sched_domains_mutex);
5899 sched_init_domains(cpu_active_mask);
5900 mutex_unlock(&sched_domains_mutex);
5902 /* Move init over to a non-isolated CPU */
5903 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
5905 sched_init_granularity();
5907 init_sched_rt_class();
5908 init_sched_dl_class();
5912 sched_smp_initialized = true;
5915 static int __init migration_init(void)
5917 sched_rq_cpu_starting(smp_processor_id());
5920 early_initcall(migration_init);
5923 void __init sched_init_smp(void)
5925 sched_init_granularity();
5927 #endif /* CONFIG_SMP */
5929 int in_sched_functions(unsigned long addr)
5931 return in_lock_functions(addr) ||
5932 (addr >= (unsigned long)__sched_text_start
5933 && addr < (unsigned long)__sched_text_end);
5936 #ifdef CONFIG_CGROUP_SCHED
5938 * Default task group.
5939 * Every task in system belongs to this group at bootup.
5941 struct task_group root_task_group;
5942 LIST_HEAD(task_groups);
5944 /* Cacheline aligned slab cache for task_group */
5945 static struct kmem_cache *task_group_cache __read_mostly;
5948 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5949 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5951 void __init sched_init(void)
5954 unsigned long alloc_size = 0, ptr;
5959 #ifdef CONFIG_FAIR_GROUP_SCHED
5960 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5962 #ifdef CONFIG_RT_GROUP_SCHED
5963 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5966 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5968 #ifdef CONFIG_FAIR_GROUP_SCHED
5969 root_task_group.se = (struct sched_entity **)ptr;
5970 ptr += nr_cpu_ids * sizeof(void **);
5972 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5973 ptr += nr_cpu_ids * sizeof(void **);
5975 #endif /* CONFIG_FAIR_GROUP_SCHED */
5976 #ifdef CONFIG_RT_GROUP_SCHED
5977 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5978 ptr += nr_cpu_ids * sizeof(void **);
5980 root_task_group.rt_rq = (struct rt_rq **)ptr;
5981 ptr += nr_cpu_ids * sizeof(void **);
5983 #endif /* CONFIG_RT_GROUP_SCHED */
5985 #ifdef CONFIG_CPUMASK_OFFSTACK
5986 for_each_possible_cpu(i) {
5987 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5988 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5989 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
5990 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5992 #endif /* CONFIG_CPUMASK_OFFSTACK */
5994 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
5995 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
5998 init_defrootdomain();
6001 #ifdef CONFIG_RT_GROUP_SCHED
6002 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6003 global_rt_period(), global_rt_runtime());
6004 #endif /* CONFIG_RT_GROUP_SCHED */
6006 #ifdef CONFIG_CGROUP_SCHED
6007 task_group_cache = KMEM_CACHE(task_group, 0);
6009 list_add(&root_task_group.list, &task_groups);
6010 INIT_LIST_HEAD(&root_task_group.children);
6011 INIT_LIST_HEAD(&root_task_group.siblings);
6012 autogroup_init(&init_task);
6013 #endif /* CONFIG_CGROUP_SCHED */
6015 for_each_possible_cpu(i) {
6019 raw_spin_lock_init(&rq->lock);
6021 rq->calc_load_active = 0;
6022 rq->calc_load_update = jiffies + LOAD_FREQ;
6023 init_cfs_rq(&rq->cfs);
6024 init_rt_rq(&rq->rt);
6025 init_dl_rq(&rq->dl);
6026 #ifdef CONFIG_FAIR_GROUP_SCHED
6027 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6028 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6029 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6031 * How much CPU bandwidth does root_task_group get?
6033 * In case of task-groups formed thr' the cgroup filesystem, it
6034 * gets 100% of the CPU resources in the system. This overall
6035 * system CPU resource is divided among the tasks of
6036 * root_task_group and its child task-groups in a fair manner,
6037 * based on each entity's (task or task-group's) weight
6038 * (se->load.weight).
6040 * In other words, if root_task_group has 10 tasks of weight
6041 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6042 * then A0's share of the CPU resource is:
6044 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6046 * We achieve this by letting root_task_group's tasks sit
6047 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6049 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6050 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6051 #endif /* CONFIG_FAIR_GROUP_SCHED */
6053 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6054 #ifdef CONFIG_RT_GROUP_SCHED
6055 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6058 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6059 rq->cpu_load[j] = 0;
6064 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6065 rq->balance_callback = NULL;
6066 rq->active_balance = 0;
6067 rq->next_balance = jiffies;
6072 rq->avg_idle = 2*sysctl_sched_migration_cost;
6073 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6075 INIT_LIST_HEAD(&rq->cfs_tasks);
6077 rq_attach_root(rq, &def_root_domain);
6078 #ifdef CONFIG_NO_HZ_COMMON
6079 rq->last_load_update_tick = jiffies;
6080 rq->last_blocked_load_update_tick = jiffies;
6081 atomic_set(&rq->nohz_flags, 0);
6083 #endif /* CONFIG_SMP */
6085 atomic_set(&rq->nr_iowait, 0);
6088 set_load_weight(&init_task, false);
6091 * The boot idle thread does lazy MMU switching as well:
6094 enter_lazy_tlb(&init_mm, current);
6097 * Make us the idle thread. Technically, schedule() should not be
6098 * called from this thread, however somewhere below it might be,
6099 * but because we are the idle thread, we just pick up running again
6100 * when this runqueue becomes "idle".
6102 init_idle(current, smp_processor_id());
6104 calc_load_update = jiffies + LOAD_FREQ;
6107 idle_thread_set_boot_cpu();
6108 set_cpu_rq_start_time(smp_processor_id());
6110 init_sched_fair_class();
6114 scheduler_running = 1;
6117 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6118 static inline int preempt_count_equals(int preempt_offset)
6120 int nested = preempt_count() + rcu_preempt_depth();
6122 return (nested == preempt_offset);
6125 void __might_sleep(const char *file, int line, int preempt_offset)
6128 * Blocking primitives will set (and therefore destroy) current->state,
6129 * since we will exit with TASK_RUNNING make sure we enter with it,
6130 * otherwise we will destroy state.
6132 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6133 "do not call blocking ops when !TASK_RUNNING; "
6134 "state=%lx set at [<%p>] %pS\n",
6136 (void *)current->task_state_change,
6137 (void *)current->task_state_change);
6139 ___might_sleep(file, line, preempt_offset);
6141 EXPORT_SYMBOL(__might_sleep);
6143 void ___might_sleep(const char *file, int line, int preempt_offset)
6145 /* Ratelimiting timestamp: */
6146 static unsigned long prev_jiffy;
6148 unsigned long preempt_disable_ip;
6150 /* WARN_ON_ONCE() by default, no rate limit required: */
6153 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6154 !is_idle_task(current)) ||
6155 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6159 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6161 prev_jiffy = jiffies;
6163 /* Save this before calling printk(), since that will clobber it: */
6164 preempt_disable_ip = get_preempt_disable_ip(current);
6167 "BUG: sleeping function called from invalid context at %s:%d\n",
6170 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6171 in_atomic(), irqs_disabled(),
6172 current->pid, current->comm);
6174 if (task_stack_end_corrupted(current))
6175 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6177 debug_show_held_locks(current);
6178 if (irqs_disabled())
6179 print_irqtrace_events(current);
6180 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6181 && !preempt_count_equals(preempt_offset)) {
6182 pr_err("Preemption disabled at:");
6183 print_ip_sym(preempt_disable_ip);
6187 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6189 EXPORT_SYMBOL(___might_sleep);
6192 #ifdef CONFIG_MAGIC_SYSRQ
6193 void normalize_rt_tasks(void)
6195 struct task_struct *g, *p;
6196 struct sched_attr attr = {
6197 .sched_policy = SCHED_NORMAL,
6200 read_lock(&tasklist_lock);
6201 for_each_process_thread(g, p) {
6203 * Only normalize user tasks:
6205 if (p->flags & PF_KTHREAD)
6208 p->se.exec_start = 0;
6209 schedstat_set(p->se.statistics.wait_start, 0);
6210 schedstat_set(p->se.statistics.sleep_start, 0);
6211 schedstat_set(p->se.statistics.block_start, 0);
6213 if (!dl_task(p) && !rt_task(p)) {
6215 * Renice negative nice level userspace
6218 if (task_nice(p) < 0)
6219 set_user_nice(p, 0);
6223 __sched_setscheduler(p, &attr, false, false);
6225 read_unlock(&tasklist_lock);
6228 #endif /* CONFIG_MAGIC_SYSRQ */
6230 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6232 * These functions are only useful for the IA64 MCA handling, or kdb.
6234 * They can only be called when the whole system has been
6235 * stopped - every CPU needs to be quiescent, and no scheduling
6236 * activity can take place. Using them for anything else would
6237 * be a serious bug, and as a result, they aren't even visible
6238 * under any other configuration.
6242 * curr_task - return the current task for a given CPU.
6243 * @cpu: the processor in question.
6245 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6247 * Return: The current task for @cpu.
6249 struct task_struct *curr_task(int cpu)
6251 return cpu_curr(cpu);
6254 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6258 * set_curr_task - set the current task for a given CPU.
6259 * @cpu: the processor in question.
6260 * @p: the task pointer to set.
6262 * Description: This function must only be used when non-maskable interrupts
6263 * are serviced on a separate stack. It allows the architecture to switch the
6264 * notion of the current task on a CPU in a non-blocking manner. This function
6265 * must be called with all CPU's synchronized, and interrupts disabled, the
6266 * and caller must save the original value of the current task (see
6267 * curr_task() above) and restore that value before reenabling interrupts and
6268 * re-starting the system.
6270 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6272 void ia64_set_curr_task(int cpu, struct task_struct *p)
6279 #ifdef CONFIG_CGROUP_SCHED
6280 /* task_group_lock serializes the addition/removal of task groups */
6281 static DEFINE_SPINLOCK(task_group_lock);
6283 static void sched_free_group(struct task_group *tg)
6285 free_fair_sched_group(tg);
6286 free_rt_sched_group(tg);
6288 kmem_cache_free(task_group_cache, tg);
6291 /* allocate runqueue etc for a new task group */
6292 struct task_group *sched_create_group(struct task_group *parent)
6294 struct task_group *tg;
6296 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6298 return ERR_PTR(-ENOMEM);
6300 if (!alloc_fair_sched_group(tg, parent))
6303 if (!alloc_rt_sched_group(tg, parent))
6309 sched_free_group(tg);
6310 return ERR_PTR(-ENOMEM);
6313 void sched_online_group(struct task_group *tg, struct task_group *parent)
6315 unsigned long flags;
6317 spin_lock_irqsave(&task_group_lock, flags);
6318 list_add_rcu(&tg->list, &task_groups);
6320 /* Root should already exist: */
6323 tg->parent = parent;
6324 INIT_LIST_HEAD(&tg->children);
6325 list_add_rcu(&tg->siblings, &parent->children);
6326 spin_unlock_irqrestore(&task_group_lock, flags);
6328 online_fair_sched_group(tg);
6331 /* rcu callback to free various structures associated with a task group */
6332 static void sched_free_group_rcu(struct rcu_head *rhp)
6334 /* Now it should be safe to free those cfs_rqs: */
6335 sched_free_group(container_of(rhp, struct task_group, rcu));
6338 void sched_destroy_group(struct task_group *tg)
6340 /* Wait for possible concurrent references to cfs_rqs complete: */
6341 call_rcu(&tg->rcu, sched_free_group_rcu);
6344 void sched_offline_group(struct task_group *tg)
6346 unsigned long flags;
6348 /* End participation in shares distribution: */
6349 unregister_fair_sched_group(tg);
6351 spin_lock_irqsave(&task_group_lock, flags);
6352 list_del_rcu(&tg->list);
6353 list_del_rcu(&tg->siblings);
6354 spin_unlock_irqrestore(&task_group_lock, flags);
6357 static void sched_change_group(struct task_struct *tsk, int type)
6359 struct task_group *tg;
6362 * All callers are synchronized by task_rq_lock(); we do not use RCU
6363 * which is pointless here. Thus, we pass "true" to task_css_check()
6364 * to prevent lockdep warnings.
6366 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6367 struct task_group, css);
6368 tg = autogroup_task_group(tsk, tg);
6369 tsk->sched_task_group = tg;
6371 #ifdef CONFIG_FAIR_GROUP_SCHED
6372 if (tsk->sched_class->task_change_group)
6373 tsk->sched_class->task_change_group(tsk, type);
6376 set_task_rq(tsk, task_cpu(tsk));
6380 * Change task's runqueue when it moves between groups.
6382 * The caller of this function should have put the task in its new group by
6383 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6386 void sched_move_task(struct task_struct *tsk)
6388 int queued, running, queue_flags =
6389 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6393 rq = task_rq_lock(tsk, &rf);
6394 update_rq_clock(rq);
6396 running = task_current(rq, tsk);
6397 queued = task_on_rq_queued(tsk);
6400 dequeue_task(rq, tsk, queue_flags);
6402 put_prev_task(rq, tsk);
6404 sched_change_group(tsk, TASK_MOVE_GROUP);
6407 enqueue_task(rq, tsk, queue_flags);
6409 set_curr_task(rq, tsk);
6411 task_rq_unlock(rq, tsk, &rf);
6414 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6416 return css ? container_of(css, struct task_group, css) : NULL;
6419 static struct cgroup_subsys_state *
6420 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6422 struct task_group *parent = css_tg(parent_css);
6423 struct task_group *tg;
6426 /* This is early initialization for the top cgroup */
6427 return &root_task_group.css;
6430 tg = sched_create_group(parent);
6432 return ERR_PTR(-ENOMEM);
6437 /* Expose task group only after completing cgroup initialization */
6438 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6440 struct task_group *tg = css_tg(css);
6441 struct task_group *parent = css_tg(css->parent);
6444 sched_online_group(tg, parent);
6448 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6450 struct task_group *tg = css_tg(css);
6452 sched_offline_group(tg);
6455 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6457 struct task_group *tg = css_tg(css);
6460 * Relies on the RCU grace period between css_released() and this.
6462 sched_free_group(tg);
6466 * This is called before wake_up_new_task(), therefore we really only
6467 * have to set its group bits, all the other stuff does not apply.
6469 static void cpu_cgroup_fork(struct task_struct *task)
6474 rq = task_rq_lock(task, &rf);
6476 update_rq_clock(rq);
6477 sched_change_group(task, TASK_SET_GROUP);
6479 task_rq_unlock(rq, task, &rf);
6482 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6484 struct task_struct *task;
6485 struct cgroup_subsys_state *css;
6488 cgroup_taskset_for_each(task, css, tset) {
6489 #ifdef CONFIG_RT_GROUP_SCHED
6490 if (!sched_rt_can_attach(css_tg(css), task))
6493 /* We don't support RT-tasks being in separate groups */
6494 if (task->sched_class != &fair_sched_class)
6498 * Serialize against wake_up_new_task() such that if its
6499 * running, we're sure to observe its full state.
6501 raw_spin_lock_irq(&task->pi_lock);
6503 * Avoid calling sched_move_task() before wake_up_new_task()
6504 * has happened. This would lead to problems with PELT, due to
6505 * move wanting to detach+attach while we're not attached yet.
6507 if (task->state == TASK_NEW)
6509 raw_spin_unlock_irq(&task->pi_lock);
6517 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6519 struct task_struct *task;
6520 struct cgroup_subsys_state *css;
6522 cgroup_taskset_for_each(task, css, tset)
6523 sched_move_task(task);
6526 #ifdef CONFIG_FAIR_GROUP_SCHED
6527 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6528 struct cftype *cftype, u64 shareval)
6530 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6533 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6536 struct task_group *tg = css_tg(css);
6538 return (u64) scale_load_down(tg->shares);
6541 #ifdef CONFIG_CFS_BANDWIDTH
6542 static DEFINE_MUTEX(cfs_constraints_mutex);
6544 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6545 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6547 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6549 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6551 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6552 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6554 if (tg == &root_task_group)
6558 * Ensure we have at some amount of bandwidth every period. This is
6559 * to prevent reaching a state of large arrears when throttled via
6560 * entity_tick() resulting in prolonged exit starvation.
6562 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6566 * Likewise, bound things on the otherside by preventing insane quota
6567 * periods. This also allows us to normalize in computing quota
6570 if (period > max_cfs_quota_period)
6574 * Prevent race between setting of cfs_rq->runtime_enabled and
6575 * unthrottle_offline_cfs_rqs().
6578 mutex_lock(&cfs_constraints_mutex);
6579 ret = __cfs_schedulable(tg, period, quota);
6583 runtime_enabled = quota != RUNTIME_INF;
6584 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6586 * If we need to toggle cfs_bandwidth_used, off->on must occur
6587 * before making related changes, and on->off must occur afterwards
6589 if (runtime_enabled && !runtime_was_enabled)
6590 cfs_bandwidth_usage_inc();
6591 raw_spin_lock_irq(&cfs_b->lock);
6592 cfs_b->period = ns_to_ktime(period);
6593 cfs_b->quota = quota;
6595 __refill_cfs_bandwidth_runtime(cfs_b);
6597 /* Restart the period timer (if active) to handle new period expiry: */
6598 if (runtime_enabled)
6599 start_cfs_bandwidth(cfs_b);
6601 raw_spin_unlock_irq(&cfs_b->lock);
6603 for_each_online_cpu(i) {
6604 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6605 struct rq *rq = cfs_rq->rq;
6608 rq_lock_irq(rq, &rf);
6609 cfs_rq->runtime_enabled = runtime_enabled;
6610 cfs_rq->runtime_remaining = 0;
6612 if (cfs_rq->throttled)
6613 unthrottle_cfs_rq(cfs_rq);
6614 rq_unlock_irq(rq, &rf);
6616 if (runtime_was_enabled && !runtime_enabled)
6617 cfs_bandwidth_usage_dec();
6619 mutex_unlock(&cfs_constraints_mutex);
6625 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
6629 period = ktime_to_ns(tg->cfs_bandwidth.period);
6630 if (cfs_quota_us < 0)
6631 quota = RUNTIME_INF;
6633 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
6635 return tg_set_cfs_bandwidth(tg, period, quota);
6638 long tg_get_cfs_quota(struct task_group *tg)
6642 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
6645 quota_us = tg->cfs_bandwidth.quota;
6646 do_div(quota_us, NSEC_PER_USEC);
6651 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
6655 period = (u64)cfs_period_us * NSEC_PER_USEC;
6656 quota = tg->cfs_bandwidth.quota;
6658 return tg_set_cfs_bandwidth(tg, period, quota);
6661 long tg_get_cfs_period(struct task_group *tg)
6665 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
6666 do_div(cfs_period_us, NSEC_PER_USEC);
6668 return cfs_period_us;
6671 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
6674 return tg_get_cfs_quota(css_tg(css));
6677 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
6678 struct cftype *cftype, s64 cfs_quota_us)
6680 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
6683 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
6686 return tg_get_cfs_period(css_tg(css));
6689 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
6690 struct cftype *cftype, u64 cfs_period_us)
6692 return tg_set_cfs_period(css_tg(css), cfs_period_us);
6695 struct cfs_schedulable_data {
6696 struct task_group *tg;
6701 * normalize group quota/period to be quota/max_period
6702 * note: units are usecs
6704 static u64 normalize_cfs_quota(struct task_group *tg,
6705 struct cfs_schedulable_data *d)
6713 period = tg_get_cfs_period(tg);
6714 quota = tg_get_cfs_quota(tg);
6717 /* note: these should typically be equivalent */
6718 if (quota == RUNTIME_INF || quota == -1)
6721 return to_ratio(period, quota);
6724 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
6726 struct cfs_schedulable_data *d = data;
6727 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6728 s64 quota = 0, parent_quota = -1;
6731 quota = RUNTIME_INF;
6733 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
6735 quota = normalize_cfs_quota(tg, d);
6736 parent_quota = parent_b->hierarchical_quota;
6739 * Ensure max(child_quota) <= parent_quota. On cgroup2,
6740 * always take the min. On cgroup1, only inherit when no
6743 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
6744 quota = min(quota, parent_quota);
6746 if (quota == RUNTIME_INF)
6747 quota = parent_quota;
6748 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
6752 cfs_b->hierarchical_quota = quota;
6757 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
6760 struct cfs_schedulable_data data = {
6766 if (quota != RUNTIME_INF) {
6767 do_div(data.period, NSEC_PER_USEC);
6768 do_div(data.quota, NSEC_PER_USEC);
6772 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
6778 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
6780 struct task_group *tg = css_tg(seq_css(sf));
6781 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6783 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
6784 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
6785 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
6789 #endif /* CONFIG_CFS_BANDWIDTH */
6790 #endif /* CONFIG_FAIR_GROUP_SCHED */
6792 #ifdef CONFIG_RT_GROUP_SCHED
6793 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
6794 struct cftype *cft, s64 val)
6796 return sched_group_set_rt_runtime(css_tg(css), val);
6799 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
6802 return sched_group_rt_runtime(css_tg(css));
6805 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
6806 struct cftype *cftype, u64 rt_period_us)
6808 return sched_group_set_rt_period(css_tg(css), rt_period_us);
6811 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
6814 return sched_group_rt_period(css_tg(css));
6816 #endif /* CONFIG_RT_GROUP_SCHED */
6818 static struct cftype cpu_legacy_files[] = {
6819 #ifdef CONFIG_FAIR_GROUP_SCHED
6822 .read_u64 = cpu_shares_read_u64,
6823 .write_u64 = cpu_shares_write_u64,
6826 #ifdef CONFIG_CFS_BANDWIDTH
6828 .name = "cfs_quota_us",
6829 .read_s64 = cpu_cfs_quota_read_s64,
6830 .write_s64 = cpu_cfs_quota_write_s64,
6833 .name = "cfs_period_us",
6834 .read_u64 = cpu_cfs_period_read_u64,
6835 .write_u64 = cpu_cfs_period_write_u64,
6839 .seq_show = cpu_cfs_stat_show,
6842 #ifdef CONFIG_RT_GROUP_SCHED
6844 .name = "rt_runtime_us",
6845 .read_s64 = cpu_rt_runtime_read,
6846 .write_s64 = cpu_rt_runtime_write,
6849 .name = "rt_period_us",
6850 .read_u64 = cpu_rt_period_read_uint,
6851 .write_u64 = cpu_rt_period_write_uint,
6857 static int cpu_extra_stat_show(struct seq_file *sf,
6858 struct cgroup_subsys_state *css)
6860 #ifdef CONFIG_CFS_BANDWIDTH
6862 struct task_group *tg = css_tg(css);
6863 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6866 throttled_usec = cfs_b->throttled_time;
6867 do_div(throttled_usec, NSEC_PER_USEC);
6869 seq_printf(sf, "nr_periods %d\n"
6871 "throttled_usec %llu\n",
6872 cfs_b->nr_periods, cfs_b->nr_throttled,
6879 #ifdef CONFIG_FAIR_GROUP_SCHED
6880 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
6883 struct task_group *tg = css_tg(css);
6884 u64 weight = scale_load_down(tg->shares);
6886 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
6889 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
6890 struct cftype *cft, u64 weight)
6893 * cgroup weight knobs should use the common MIN, DFL and MAX
6894 * values which are 1, 100 and 10000 respectively. While it loses
6895 * a bit of range on both ends, it maps pretty well onto the shares
6896 * value used by scheduler and the round-trip conversions preserve
6897 * the original value over the entire range.
6899 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
6902 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
6904 return sched_group_set_shares(css_tg(css), scale_load(weight));
6907 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
6910 unsigned long weight = scale_load_down(css_tg(css)->shares);
6911 int last_delta = INT_MAX;
6914 /* find the closest nice value to the current weight */
6915 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
6916 delta = abs(sched_prio_to_weight[prio] - weight);
6917 if (delta >= last_delta)
6922 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
6925 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
6926 struct cftype *cft, s64 nice)
6928 unsigned long weight;
6931 if (nice < MIN_NICE || nice > MAX_NICE)
6934 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
6935 idx = array_index_nospec(idx, 40);
6936 weight = sched_prio_to_weight[idx];
6938 return sched_group_set_shares(css_tg(css), scale_load(weight));
6942 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
6943 long period, long quota)
6946 seq_puts(sf, "max");
6948 seq_printf(sf, "%ld", quota);
6950 seq_printf(sf, " %ld\n", period);
6953 /* caller should put the current value in *@periodp before calling */
6954 static int __maybe_unused cpu_period_quota_parse(char *buf,
6955 u64 *periodp, u64 *quotap)
6957 char tok[21]; /* U64_MAX */
6959 if (!sscanf(buf, "%s %llu", tok, periodp))
6962 *periodp *= NSEC_PER_USEC;
6964 if (sscanf(tok, "%llu", quotap))
6965 *quotap *= NSEC_PER_USEC;
6966 else if (!strcmp(tok, "max"))
6967 *quotap = RUNTIME_INF;
6974 #ifdef CONFIG_CFS_BANDWIDTH
6975 static int cpu_max_show(struct seq_file *sf, void *v)
6977 struct task_group *tg = css_tg(seq_css(sf));
6979 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
6983 static ssize_t cpu_max_write(struct kernfs_open_file *of,
6984 char *buf, size_t nbytes, loff_t off)
6986 struct task_group *tg = css_tg(of_css(of));
6987 u64 period = tg_get_cfs_period(tg);
6991 ret = cpu_period_quota_parse(buf, &period, "a);
6993 ret = tg_set_cfs_bandwidth(tg, period, quota);
6994 return ret ?: nbytes;
6998 static struct cftype cpu_files[] = {
6999 #ifdef CONFIG_FAIR_GROUP_SCHED
7002 .flags = CFTYPE_NOT_ON_ROOT,
7003 .read_u64 = cpu_weight_read_u64,
7004 .write_u64 = cpu_weight_write_u64,
7007 .name = "weight.nice",
7008 .flags = CFTYPE_NOT_ON_ROOT,
7009 .read_s64 = cpu_weight_nice_read_s64,
7010 .write_s64 = cpu_weight_nice_write_s64,
7013 #ifdef CONFIG_CFS_BANDWIDTH
7016 .flags = CFTYPE_NOT_ON_ROOT,
7017 .seq_show = cpu_max_show,
7018 .write = cpu_max_write,
7024 struct cgroup_subsys cpu_cgrp_subsys = {
7025 .css_alloc = cpu_cgroup_css_alloc,
7026 .css_online = cpu_cgroup_css_online,
7027 .css_released = cpu_cgroup_css_released,
7028 .css_free = cpu_cgroup_css_free,
7029 .css_extra_stat_show = cpu_extra_stat_show,
7030 .fork = cpu_cgroup_fork,
7031 .can_attach = cpu_cgroup_can_attach,
7032 .attach = cpu_cgroup_attach,
7033 .legacy_cftypes = cpu_legacy_files,
7034 .dfl_cftypes = cpu_files,
7039 #endif /* CONFIG_CGROUP_SCHED */
7041 void dump_cpu_task(int cpu)
7043 pr_info("Task dump for CPU %d:\n", cpu);
7044 sched_show_task(cpu_curr(cpu));
7048 * Nice levels are multiplicative, with a gentle 10% change for every
7049 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7050 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7051 * that remained on nice 0.
7053 * The "10% effect" is relative and cumulative: from _any_ nice level,
7054 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7055 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7056 * If a task goes up by ~10% and another task goes down by ~10% then
7057 * the relative distance between them is ~25%.)
7059 const int sched_prio_to_weight[40] = {
7060 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7061 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7062 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7063 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7064 /* 0 */ 1024, 820, 655, 526, 423,
7065 /* 5 */ 335, 272, 215, 172, 137,
7066 /* 10 */ 110, 87, 70, 56, 45,
7067 /* 15 */ 36, 29, 23, 18, 15,
7071 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7073 * In cases where the weight does not change often, we can use the
7074 * precalculated inverse to speed up arithmetics by turning divisions
7075 * into multiplications:
7077 const u32 sched_prio_to_wmult[40] = {
7078 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7079 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7080 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7081 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7082 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7083 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7084 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7085 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7088 #undef CREATE_TRACE_POINTS