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 <linux/kcov.h>
15 #include <asm/switch_to.h>
18 #include "../workqueue_internal.h"
19 #include "../smpboot.h"
21 #define CREATE_TRACE_POINTS
22 #include <trace/events/sched.h>
24 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
26 #if defined(CONFIG_SCHED_DEBUG) && defined(HAVE_JUMP_LABEL)
28 * Debugging: various feature bits
30 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
31 * sysctl_sched_features, defined in sched.h, to allow constants propagation
32 * at compile time and compiler optimization based on features default.
34 #define SCHED_FEAT(name, enabled) \
35 (1UL << __SCHED_FEAT_##name) * enabled |
36 const_debug unsigned int sysctl_sched_features =
43 * Number of tasks to iterate in a single balance run.
44 * Limited because this is done with IRQs disabled.
46 const_debug unsigned int sysctl_sched_nr_migrate = 32;
49 * period over which we average the RT time consumption, measured
54 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
57 * period over which we measure -rt task CPU usage in us.
60 unsigned int sysctl_sched_rt_period = 1000000;
62 __read_mostly int scheduler_running;
65 * part of the period that we allow rt tasks to run in us.
68 int sysctl_sched_rt_runtime = 950000;
71 * __task_rq_lock - lock the rq @p resides on.
73 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
78 lockdep_assert_held(&p->pi_lock);
82 raw_spin_lock(&rq->lock);
83 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
87 raw_spin_unlock(&rq->lock);
89 while (unlikely(task_on_rq_migrating(p)))
95 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
97 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
98 __acquires(p->pi_lock)
104 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
106 raw_spin_lock(&rq->lock);
108 * move_queued_task() task_rq_lock()
111 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
112 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
113 * [S] ->cpu = new_cpu [L] task_rq()
117 * If we observe the old CPU in task_rq_lock, the acquire of
118 * the old rq->lock will fully serialize against the stores.
120 * If we observe the new CPU in task_rq_lock, the acquire will
121 * pair with the WMB to ensure we must then also see migrating.
123 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
127 raw_spin_unlock(&rq->lock);
128 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
130 while (unlikely(task_on_rq_migrating(p)))
136 * RQ-clock updating methods:
139 static void update_rq_clock_task(struct rq *rq, s64 delta)
142 * In theory, the compile should just see 0 here, and optimize out the call
143 * to sched_rt_avg_update. But I don't trust it...
145 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
146 s64 steal = 0, irq_delta = 0;
148 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
149 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
152 * Since irq_time is only updated on {soft,}irq_exit, we might run into
153 * this case when a previous update_rq_clock() happened inside a
156 * When this happens, we stop ->clock_task and only update the
157 * prev_irq_time stamp to account for the part that fit, so that a next
158 * update will consume the rest. This ensures ->clock_task is
161 * It does however cause some slight miss-attribution of {soft,}irq
162 * time, a more accurate solution would be to update the irq_time using
163 * the current rq->clock timestamp, except that would require using
166 if (irq_delta > delta)
169 rq->prev_irq_time += irq_delta;
172 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
173 if (static_key_false((¶virt_steal_rq_enabled))) {
174 steal = paravirt_steal_clock(cpu_of(rq));
175 steal -= rq->prev_steal_time_rq;
177 if (unlikely(steal > delta))
180 rq->prev_steal_time_rq += steal;
185 rq->clock_task += delta;
187 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
188 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
189 sched_rt_avg_update(rq, irq_delta + steal);
193 void update_rq_clock(struct rq *rq)
197 lockdep_assert_held(&rq->lock);
199 if (rq->clock_update_flags & RQCF_ACT_SKIP)
202 #ifdef CONFIG_SCHED_DEBUG
203 if (sched_feat(WARN_DOUBLE_CLOCK))
204 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
205 rq->clock_update_flags |= RQCF_UPDATED;
208 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
212 update_rq_clock_task(rq, delta);
216 #ifdef CONFIG_SCHED_HRTICK
218 * Use HR-timers to deliver accurate preemption points.
221 static void hrtick_clear(struct rq *rq)
223 if (hrtimer_active(&rq->hrtick_timer))
224 hrtimer_cancel(&rq->hrtick_timer);
228 * High-resolution timer tick.
229 * Runs from hardirq context with interrupts disabled.
231 static enum hrtimer_restart hrtick(struct hrtimer *timer)
233 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
236 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
240 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
243 return HRTIMER_NORESTART;
248 static void __hrtick_restart(struct rq *rq)
250 struct hrtimer *timer = &rq->hrtick_timer;
252 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
256 * called from hardirq (IPI) context
258 static void __hrtick_start(void *arg)
264 __hrtick_restart(rq);
265 rq->hrtick_csd_pending = 0;
270 * Called to set the hrtick timer state.
272 * called with rq->lock held and irqs disabled
274 void hrtick_start(struct rq *rq, u64 delay)
276 struct hrtimer *timer = &rq->hrtick_timer;
281 * Don't schedule slices shorter than 10000ns, that just
282 * doesn't make sense and can cause timer DoS.
284 delta = max_t(s64, delay, 10000LL);
285 time = ktime_add_ns(timer->base->get_time(), delta);
287 hrtimer_set_expires(timer, time);
289 if (rq == this_rq()) {
290 __hrtick_restart(rq);
291 } else if (!rq->hrtick_csd_pending) {
292 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
293 rq->hrtick_csd_pending = 1;
299 * Called to set the hrtick timer state.
301 * called with rq->lock held and irqs disabled
303 void hrtick_start(struct rq *rq, u64 delay)
306 * Don't schedule slices shorter than 10000ns, that just
307 * doesn't make sense. Rely on vruntime for fairness.
309 delay = max_t(u64, delay, 10000LL);
310 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
311 HRTIMER_MODE_REL_PINNED);
313 #endif /* CONFIG_SMP */
315 static void hrtick_rq_init(struct rq *rq)
318 rq->hrtick_csd_pending = 0;
320 rq->hrtick_csd.flags = 0;
321 rq->hrtick_csd.func = __hrtick_start;
322 rq->hrtick_csd.info = rq;
325 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
326 rq->hrtick_timer.function = hrtick;
328 #else /* CONFIG_SCHED_HRTICK */
329 static inline void hrtick_clear(struct rq *rq)
333 static inline void hrtick_rq_init(struct rq *rq)
336 #endif /* CONFIG_SCHED_HRTICK */
339 * cmpxchg based fetch_or, macro so it works for different integer types
341 #define fetch_or(ptr, mask) \
343 typeof(ptr) _ptr = (ptr); \
344 typeof(mask) _mask = (mask); \
345 typeof(*_ptr) _old, _val = *_ptr; \
348 _old = cmpxchg(_ptr, _val, _val | _mask); \
356 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
358 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
359 * this avoids any races wrt polling state changes and thereby avoids
362 static bool set_nr_and_not_polling(struct task_struct *p)
364 struct thread_info *ti = task_thread_info(p);
365 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
369 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
371 * If this returns true, then the idle task promises to call
372 * sched_ttwu_pending() and reschedule soon.
374 static bool set_nr_if_polling(struct task_struct *p)
376 struct thread_info *ti = task_thread_info(p);
377 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
380 if (!(val & _TIF_POLLING_NRFLAG))
382 if (val & _TIF_NEED_RESCHED)
384 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
393 static bool set_nr_and_not_polling(struct task_struct *p)
395 set_tsk_need_resched(p);
400 static bool set_nr_if_polling(struct task_struct *p)
407 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
409 struct wake_q_node *node = &task->wake_q;
412 * Atomically grab the task, if ->wake_q is !nil already it means
413 * its already queued (either by us or someone else) and will get the
414 * wakeup due to that.
416 * This cmpxchg() implies a full barrier, which pairs with the write
417 * barrier implied by the wakeup in wake_up_q().
419 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
422 get_task_struct(task);
425 * The head is context local, there can be no concurrency.
428 head->lastp = &node->next;
431 void wake_up_q(struct wake_q_head *head)
433 struct wake_q_node *node = head->first;
435 while (node != WAKE_Q_TAIL) {
436 struct task_struct *task;
438 task = container_of(node, struct task_struct, wake_q);
440 /* Task can safely be re-inserted now: */
442 task->wake_q.next = NULL;
445 * wake_up_process() implies a wmb() to pair with the queueing
446 * in wake_q_add() so as not to miss wakeups.
448 wake_up_process(task);
449 put_task_struct(task);
454 * resched_curr - mark rq's current task 'to be rescheduled now'.
456 * On UP this means the setting of the need_resched flag, on SMP it
457 * might also involve a cross-CPU call to trigger the scheduler on
460 void resched_curr(struct rq *rq)
462 struct task_struct *curr = rq->curr;
465 lockdep_assert_held(&rq->lock);
467 if (test_tsk_need_resched(curr))
472 if (cpu == smp_processor_id()) {
473 set_tsk_need_resched(curr);
474 set_preempt_need_resched();
478 if (set_nr_and_not_polling(curr))
479 smp_send_reschedule(cpu);
481 trace_sched_wake_idle_without_ipi(cpu);
484 void resched_cpu(int cpu)
486 struct rq *rq = cpu_rq(cpu);
489 raw_spin_lock_irqsave(&rq->lock, flags);
490 if (cpu_online(cpu) || cpu == smp_processor_id())
492 raw_spin_unlock_irqrestore(&rq->lock, flags);
496 #ifdef CONFIG_NO_HZ_COMMON
498 * In the semi idle case, use the nearest busy CPU for migrating timers
499 * from an idle CPU. This is good for power-savings.
501 * We don't do similar optimization for completely idle system, as
502 * selecting an idle CPU will add more delays to the timers than intended
503 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
505 int get_nohz_timer_target(void)
507 int i, cpu = smp_processor_id();
508 struct sched_domain *sd;
510 if (!idle_cpu(cpu) && housekeeping_cpu(cpu, HK_FLAG_TIMER))
514 for_each_domain(cpu, sd) {
515 for_each_cpu(i, sched_domain_span(sd)) {
519 if (!idle_cpu(i) && housekeeping_cpu(i, HK_FLAG_TIMER)) {
526 if (!housekeeping_cpu(cpu, HK_FLAG_TIMER))
527 cpu = housekeeping_any_cpu(HK_FLAG_TIMER);
534 * When add_timer_on() enqueues a timer into the timer wheel of an
535 * idle CPU then this timer might expire before the next timer event
536 * which is scheduled to wake up that CPU. In case of a completely
537 * idle system the next event might even be infinite time into the
538 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
539 * leaves the inner idle loop so the newly added timer is taken into
540 * account when the CPU goes back to idle and evaluates the timer
541 * wheel for the next timer event.
543 static void wake_up_idle_cpu(int cpu)
545 struct rq *rq = cpu_rq(cpu);
547 if (cpu == smp_processor_id())
550 if (set_nr_and_not_polling(rq->idle))
551 smp_send_reschedule(cpu);
553 trace_sched_wake_idle_without_ipi(cpu);
556 static bool wake_up_full_nohz_cpu(int cpu)
559 * We just need the target to call irq_exit() and re-evaluate
560 * the next tick. The nohz full kick at least implies that.
561 * If needed we can still optimize that later with an
564 if (cpu_is_offline(cpu))
565 return true; /* Don't try to wake offline CPUs. */
566 if (tick_nohz_full_cpu(cpu)) {
567 if (cpu != smp_processor_id() ||
568 tick_nohz_tick_stopped())
569 tick_nohz_full_kick_cpu(cpu);
577 * Wake up the specified CPU. If the CPU is going offline, it is the
578 * caller's responsibility to deal with the lost wakeup, for example,
579 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
581 void wake_up_nohz_cpu(int cpu)
583 if (!wake_up_full_nohz_cpu(cpu))
584 wake_up_idle_cpu(cpu);
587 static inline bool got_nohz_idle_kick(void)
589 int cpu = smp_processor_id();
591 if (!(atomic_read(nohz_flags(cpu)) & NOHZ_KICK_MASK))
594 if (idle_cpu(cpu) && !need_resched())
598 * We can't run Idle Load Balance on this CPU for this time so we
599 * cancel it and clear NOHZ_BALANCE_KICK
601 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(cpu));
605 #else /* CONFIG_NO_HZ_COMMON */
607 static inline bool got_nohz_idle_kick(void)
612 #endif /* CONFIG_NO_HZ_COMMON */
614 #ifdef CONFIG_NO_HZ_FULL
615 bool sched_can_stop_tick(struct rq *rq)
619 /* Deadline tasks, even if single, need the tick */
620 if (rq->dl.dl_nr_running)
624 * If there are more than one RR tasks, we need the tick to effect the
625 * actual RR behaviour.
627 if (rq->rt.rr_nr_running) {
628 if (rq->rt.rr_nr_running == 1)
635 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
636 * forced preemption between FIFO tasks.
638 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
643 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
644 * if there's more than one we need the tick for involuntary
647 if (rq->nr_running > 1)
652 #endif /* CONFIG_NO_HZ_FULL */
654 void sched_avg_update(struct rq *rq)
656 s64 period = sched_avg_period();
658 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
660 * Inline assembly required to prevent the compiler
661 * optimising this loop into a divmod call.
662 * See __iter_div_u64_rem() for another example of this.
664 asm("" : "+rm" (rq->age_stamp));
665 rq->age_stamp += period;
670 #endif /* CONFIG_SMP */
672 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
673 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
675 * Iterate task_group tree rooted at *from, calling @down when first entering a
676 * node and @up when leaving it for the final time.
678 * Caller must hold rcu_lock or sufficient equivalent.
680 int walk_tg_tree_from(struct task_group *from,
681 tg_visitor down, tg_visitor up, void *data)
683 struct task_group *parent, *child;
689 ret = (*down)(parent, data);
692 list_for_each_entry_rcu(child, &parent->children, siblings) {
699 ret = (*up)(parent, data);
700 if (ret || parent == from)
704 parent = parent->parent;
711 int tg_nop(struct task_group *tg, void *data)
717 static void set_load_weight(struct task_struct *p, bool update_load)
719 int prio = p->static_prio - MAX_RT_PRIO;
720 struct load_weight *load = &p->se.load;
723 * SCHED_IDLE tasks get minimal weight:
725 if (idle_policy(p->policy)) {
726 load->weight = scale_load(WEIGHT_IDLEPRIO);
727 load->inv_weight = WMULT_IDLEPRIO;
732 * SCHED_OTHER tasks have to update their load when changing their
735 if (update_load && p->sched_class == &fair_sched_class) {
736 reweight_task(p, prio);
738 load->weight = scale_load(sched_prio_to_weight[prio]);
739 load->inv_weight = sched_prio_to_wmult[prio];
743 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
745 if (!(flags & ENQUEUE_NOCLOCK))
748 if (!(flags & ENQUEUE_RESTORE))
749 sched_info_queued(rq, p);
751 p->sched_class->enqueue_task(rq, p, flags);
754 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
756 if (!(flags & DEQUEUE_NOCLOCK))
759 if (!(flags & DEQUEUE_SAVE))
760 sched_info_dequeued(rq, p);
762 p->sched_class->dequeue_task(rq, p, flags);
765 void activate_task(struct rq *rq, struct task_struct *p, int flags)
767 if (task_contributes_to_load(p))
768 rq->nr_uninterruptible--;
770 enqueue_task(rq, p, flags);
773 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
775 if (task_contributes_to_load(p))
776 rq->nr_uninterruptible++;
778 dequeue_task(rq, p, flags);
782 * __normal_prio - return the priority that is based on the static prio
784 static inline int __normal_prio(struct task_struct *p)
786 return p->static_prio;
790 * Calculate the expected normal priority: i.e. priority
791 * without taking RT-inheritance into account. Might be
792 * boosted by interactivity modifiers. Changes upon fork,
793 * setprio syscalls, and whenever the interactivity
794 * estimator recalculates.
796 static inline int normal_prio(struct task_struct *p)
800 if (task_has_dl_policy(p))
801 prio = MAX_DL_PRIO-1;
802 else if (task_has_rt_policy(p))
803 prio = MAX_RT_PRIO-1 - p->rt_priority;
805 prio = __normal_prio(p);
810 * Calculate the current priority, i.e. the priority
811 * taken into account by the scheduler. This value might
812 * be boosted by RT tasks, or might be boosted by
813 * interactivity modifiers. Will be RT if the task got
814 * RT-boosted. If not then it returns p->normal_prio.
816 static int effective_prio(struct task_struct *p)
818 p->normal_prio = normal_prio(p);
820 * If we are RT tasks or we were boosted to RT priority,
821 * keep the priority unchanged. Otherwise, update priority
822 * to the normal priority:
824 if (!rt_prio(p->prio))
825 return p->normal_prio;
830 * task_curr - is this task currently executing on a CPU?
831 * @p: the task in question.
833 * Return: 1 if the task is currently executing. 0 otherwise.
835 inline int task_curr(const struct task_struct *p)
837 return cpu_curr(task_cpu(p)) == p;
841 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
842 * use the balance_callback list if you want balancing.
844 * this means any call to check_class_changed() must be followed by a call to
845 * balance_callback().
847 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
848 const struct sched_class *prev_class,
851 if (prev_class != p->sched_class) {
852 if (prev_class->switched_from)
853 prev_class->switched_from(rq, p);
855 p->sched_class->switched_to(rq, p);
856 } else if (oldprio != p->prio || dl_task(p))
857 p->sched_class->prio_changed(rq, p, oldprio);
860 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
862 const struct sched_class *class;
864 if (p->sched_class == rq->curr->sched_class) {
865 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
867 for_each_class(class) {
868 if (class == rq->curr->sched_class)
870 if (class == p->sched_class) {
878 * A queue event has occurred, and we're going to schedule. In
879 * this case, we can save a useless back to back clock update.
881 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
882 rq_clock_skip_update(rq);
887 static inline bool is_per_cpu_kthread(struct task_struct *p)
889 if (!(p->flags & PF_KTHREAD))
892 if (p->nr_cpus_allowed != 1)
899 * Per-CPU kthreads are allowed to run on !actie && online CPUs, see
900 * __set_cpus_allowed_ptr() and select_fallback_rq().
902 static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
904 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
907 if (is_per_cpu_kthread(p))
908 return cpu_online(cpu);
910 return cpu_active(cpu);
914 * This is how migration works:
916 * 1) we invoke migration_cpu_stop() on the target CPU using
918 * 2) stopper starts to run (implicitly forcing the migrated thread
920 * 3) it checks whether the migrated task is still in the wrong runqueue.
921 * 4) if it's in the wrong runqueue then the migration thread removes
922 * it and puts it into the right queue.
923 * 5) stopper completes and stop_one_cpu() returns and the migration
928 * move_queued_task - move a queued task to new rq.
930 * Returns (locked) new rq. Old rq's lock is released.
932 static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
933 struct task_struct *p, int new_cpu)
935 lockdep_assert_held(&rq->lock);
937 p->on_rq = TASK_ON_RQ_MIGRATING;
938 dequeue_task(rq, p, DEQUEUE_NOCLOCK);
939 set_task_cpu(p, new_cpu);
942 rq = cpu_rq(new_cpu);
945 BUG_ON(task_cpu(p) != new_cpu);
946 enqueue_task(rq, p, 0);
947 p->on_rq = TASK_ON_RQ_QUEUED;
948 check_preempt_curr(rq, p, 0);
953 struct migration_arg {
954 struct task_struct *task;
959 * Move (not current) task off this CPU, onto the destination CPU. We're doing
960 * this because either it can't run here any more (set_cpus_allowed()
961 * away from this CPU, or CPU going down), or because we're
962 * attempting to rebalance this task on exec (sched_exec).
964 * So we race with normal scheduler movements, but that's OK, as long
965 * as the task is no longer on this CPU.
967 static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
968 struct task_struct *p, int dest_cpu)
970 /* Affinity changed (again). */
971 if (!is_cpu_allowed(p, dest_cpu))
975 rq = move_queued_task(rq, rf, p, dest_cpu);
981 * migration_cpu_stop - this will be executed by a highprio stopper thread
982 * and performs thread migration by bumping thread off CPU then
983 * 'pushing' onto another runqueue.
985 static int migration_cpu_stop(void *data)
987 struct migration_arg *arg = data;
988 struct task_struct *p = arg->task;
989 struct rq *rq = this_rq();
993 * The original target CPU might have gone down and we might
994 * be on another CPU but it doesn't matter.
998 * We need to explicitly wake pending tasks before running
999 * __migrate_task() such that we will not miss enforcing cpus_allowed
1000 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1002 sched_ttwu_pending();
1004 raw_spin_lock(&p->pi_lock);
1007 * If task_rq(p) != rq, it cannot be migrated here, because we're
1008 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1009 * we're holding p->pi_lock.
1011 if (task_rq(p) == rq) {
1012 if (task_on_rq_queued(p))
1013 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
1015 p->wake_cpu = arg->dest_cpu;
1018 raw_spin_unlock(&p->pi_lock);
1025 * sched_class::set_cpus_allowed must do the below, but is not required to
1026 * actually call this function.
1028 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1030 cpumask_copy(&p->cpus_allowed, new_mask);
1031 p->nr_cpus_allowed = cpumask_weight(new_mask);
1034 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1036 struct rq *rq = task_rq(p);
1037 bool queued, running;
1039 lockdep_assert_held(&p->pi_lock);
1041 queued = task_on_rq_queued(p);
1042 running = task_current(rq, p);
1046 * Because __kthread_bind() calls this on blocked tasks without
1049 lockdep_assert_held(&rq->lock);
1050 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
1053 put_prev_task(rq, p);
1055 p->sched_class->set_cpus_allowed(p, new_mask);
1058 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
1060 set_curr_task(rq, p);
1064 * Change a given task's CPU affinity. Migrate the thread to a
1065 * proper CPU and schedule it away if the CPU it's executing on
1066 * is removed from the allowed bitmask.
1068 * NOTE: the caller must have a valid reference to the task, the
1069 * task must not exit() & deallocate itself prematurely. The
1070 * call is not atomic; no spinlocks may be held.
1072 static int __set_cpus_allowed_ptr(struct task_struct *p,
1073 const struct cpumask *new_mask, bool check)
1075 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1076 unsigned int dest_cpu;
1081 rq = task_rq_lock(p, &rf);
1082 update_rq_clock(rq);
1084 if (p->flags & PF_KTHREAD) {
1086 * Kernel threads are allowed on online && !active CPUs
1088 cpu_valid_mask = cpu_online_mask;
1092 * Must re-check here, to close a race against __kthread_bind(),
1093 * sched_setaffinity() is not guaranteed to observe the flag.
1095 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1100 if (cpumask_equal(&p->cpus_allowed, new_mask))
1103 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1108 do_set_cpus_allowed(p, new_mask);
1110 if (p->flags & PF_KTHREAD) {
1112 * For kernel threads that do indeed end up on online &&
1113 * !active we want to ensure they are strict per-CPU threads.
1115 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1116 !cpumask_intersects(new_mask, cpu_active_mask) &&
1117 p->nr_cpus_allowed != 1);
1120 /* Can the task run on the task's current CPU? If so, we're done */
1121 if (cpumask_test_cpu(task_cpu(p), new_mask))
1124 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1125 if (task_running(rq, p) || p->state == TASK_WAKING) {
1126 struct migration_arg arg = { p, dest_cpu };
1127 /* Need help from migration thread: drop lock and wait. */
1128 task_rq_unlock(rq, p, &rf);
1129 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1130 tlb_migrate_finish(p->mm);
1132 } else if (task_on_rq_queued(p)) {
1134 * OK, since we're going to drop the lock immediately
1135 * afterwards anyway.
1137 rq = move_queued_task(rq, &rf, p, dest_cpu);
1140 task_rq_unlock(rq, p, &rf);
1145 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1147 return __set_cpus_allowed_ptr(p, new_mask, false);
1149 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1151 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1153 #ifdef CONFIG_SCHED_DEBUG
1155 * We should never call set_task_cpu() on a blocked task,
1156 * ttwu() will sort out the placement.
1158 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1162 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1163 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1164 * time relying on p->on_rq.
1166 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1167 p->sched_class == &fair_sched_class &&
1168 (p->on_rq && !task_on_rq_migrating(p)));
1170 #ifdef CONFIG_LOCKDEP
1172 * The caller should hold either p->pi_lock or rq->lock, when changing
1173 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1175 * sched_move_task() holds both and thus holding either pins the cgroup,
1178 * Furthermore, all task_rq users should acquire both locks, see
1181 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1182 lockdep_is_held(&task_rq(p)->lock)));
1185 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
1187 WARN_ON_ONCE(!cpu_online(new_cpu));
1190 trace_sched_migrate_task(p, new_cpu);
1192 if (task_cpu(p) != new_cpu) {
1193 if (p->sched_class->migrate_task_rq)
1194 p->sched_class->migrate_task_rq(p);
1195 p->se.nr_migrations++;
1197 perf_event_task_migrate(p);
1200 __set_task_cpu(p, new_cpu);
1203 static void __migrate_swap_task(struct task_struct *p, int cpu)
1205 if (task_on_rq_queued(p)) {
1206 struct rq *src_rq, *dst_rq;
1207 struct rq_flags srf, drf;
1209 src_rq = task_rq(p);
1210 dst_rq = cpu_rq(cpu);
1212 rq_pin_lock(src_rq, &srf);
1213 rq_pin_lock(dst_rq, &drf);
1215 p->on_rq = TASK_ON_RQ_MIGRATING;
1216 deactivate_task(src_rq, p, 0);
1217 set_task_cpu(p, cpu);
1218 activate_task(dst_rq, p, 0);
1219 p->on_rq = TASK_ON_RQ_QUEUED;
1220 check_preempt_curr(dst_rq, p, 0);
1222 rq_unpin_lock(dst_rq, &drf);
1223 rq_unpin_lock(src_rq, &srf);
1227 * Task isn't running anymore; make it appear like we migrated
1228 * it before it went to sleep. This means on wakeup we make the
1229 * previous CPU our target instead of where it really is.
1235 struct migration_swap_arg {
1236 struct task_struct *src_task, *dst_task;
1237 int src_cpu, dst_cpu;
1240 static int migrate_swap_stop(void *data)
1242 struct migration_swap_arg *arg = data;
1243 struct rq *src_rq, *dst_rq;
1246 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1249 src_rq = cpu_rq(arg->src_cpu);
1250 dst_rq = cpu_rq(arg->dst_cpu);
1252 double_raw_lock(&arg->src_task->pi_lock,
1253 &arg->dst_task->pi_lock);
1254 double_rq_lock(src_rq, dst_rq);
1256 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1259 if (task_cpu(arg->src_task) != arg->src_cpu)
1262 if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1265 if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1268 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1269 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1274 double_rq_unlock(src_rq, dst_rq);
1275 raw_spin_unlock(&arg->dst_task->pi_lock);
1276 raw_spin_unlock(&arg->src_task->pi_lock);
1282 * Cross migrate two tasks
1284 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1286 struct migration_swap_arg arg;
1289 arg = (struct migration_swap_arg){
1291 .src_cpu = task_cpu(cur),
1293 .dst_cpu = task_cpu(p),
1296 if (arg.src_cpu == arg.dst_cpu)
1300 * These three tests are all lockless; this is OK since all of them
1301 * will be re-checked with proper locks held further down the line.
1303 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1306 if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1309 if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1312 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1313 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1320 * wait_task_inactive - wait for a thread to unschedule.
1322 * If @match_state is nonzero, it's the @p->state value just checked and
1323 * not expected to change. If it changes, i.e. @p might have woken up,
1324 * then return zero. When we succeed in waiting for @p to be off its CPU,
1325 * we return a positive number (its total switch count). If a second call
1326 * a short while later returns the same number, the caller can be sure that
1327 * @p has remained unscheduled the whole time.
1329 * The caller must ensure that the task *will* unschedule sometime soon,
1330 * else this function might spin for a *long* time. This function can't
1331 * be called with interrupts off, or it may introduce deadlock with
1332 * smp_call_function() if an IPI is sent by the same process we are
1333 * waiting to become inactive.
1335 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1337 int running, queued;
1344 * We do the initial early heuristics without holding
1345 * any task-queue locks at all. We'll only try to get
1346 * the runqueue lock when things look like they will
1352 * If the task is actively running on another CPU
1353 * still, just relax and busy-wait without holding
1356 * NOTE! Since we don't hold any locks, it's not
1357 * even sure that "rq" stays as the right runqueue!
1358 * But we don't care, since "task_running()" will
1359 * return false if the runqueue has changed and p
1360 * is actually now running somewhere else!
1362 while (task_running(rq, p)) {
1363 if (match_state && unlikely(p->state != match_state))
1369 * Ok, time to look more closely! We need the rq
1370 * lock now, to be *sure*. If we're wrong, we'll
1371 * just go back and repeat.
1373 rq = task_rq_lock(p, &rf);
1374 trace_sched_wait_task(p);
1375 running = task_running(rq, p);
1376 queued = task_on_rq_queued(p);
1378 if (!match_state || p->state == match_state)
1379 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1380 task_rq_unlock(rq, p, &rf);
1383 * If it changed from the expected state, bail out now.
1385 if (unlikely(!ncsw))
1389 * Was it really running after all now that we
1390 * checked with the proper locks actually held?
1392 * Oops. Go back and try again..
1394 if (unlikely(running)) {
1400 * It's not enough that it's not actively running,
1401 * it must be off the runqueue _entirely_, and not
1404 * So if it was still runnable (but just not actively
1405 * running right now), it's preempted, and we should
1406 * yield - it could be a while.
1408 if (unlikely(queued)) {
1409 ktime_t to = NSEC_PER_SEC / HZ;
1411 set_current_state(TASK_UNINTERRUPTIBLE);
1412 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1417 * Ahh, all good. It wasn't running, and it wasn't
1418 * runnable, which means that it will never become
1419 * running in the future either. We're all done!
1428 * kick_process - kick a running thread to enter/exit the kernel
1429 * @p: the to-be-kicked thread
1431 * Cause a process which is running on another CPU to enter
1432 * kernel-mode, without any delay. (to get signals handled.)
1434 * NOTE: this function doesn't have to take the runqueue lock,
1435 * because all it wants to ensure is that the remote task enters
1436 * the kernel. If the IPI races and the task has been migrated
1437 * to another CPU then no harm is done and the purpose has been
1440 void kick_process(struct task_struct *p)
1446 if ((cpu != smp_processor_id()) && task_curr(p))
1447 smp_send_reschedule(cpu);
1450 EXPORT_SYMBOL_GPL(kick_process);
1453 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1455 * A few notes on cpu_active vs cpu_online:
1457 * - cpu_active must be a subset of cpu_online
1459 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
1460 * see __set_cpus_allowed_ptr(). At this point the newly online
1461 * CPU isn't yet part of the sched domains, and balancing will not
1464 * - on CPU-down we clear cpu_active() to mask the sched domains and
1465 * avoid the load balancer to place new tasks on the to be removed
1466 * CPU. Existing tasks will remain running there and will be taken
1469 * This means that fallback selection must not select !active CPUs.
1470 * And can assume that any active CPU must be online. Conversely
1471 * select_task_rq() below may allow selection of !active CPUs in order
1472 * to satisfy the above rules.
1474 static int select_fallback_rq(int cpu, struct task_struct *p)
1476 int nid = cpu_to_node(cpu);
1477 const struct cpumask *nodemask = NULL;
1478 enum { cpuset, possible, fail } state = cpuset;
1482 * If the node that the CPU is on has been offlined, cpu_to_node()
1483 * will return -1. There is no CPU on the node, and we should
1484 * select the CPU on the other node.
1487 nodemask = cpumask_of_node(nid);
1489 /* Look for allowed, online CPU in same node. */
1490 for_each_cpu(dest_cpu, nodemask) {
1491 if (!cpu_active(dest_cpu))
1493 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1499 /* Any allowed, online CPU? */
1500 for_each_cpu(dest_cpu, &p->cpus_allowed) {
1501 if (!is_cpu_allowed(p, dest_cpu))
1507 /* No more Mr. Nice Guy. */
1510 if (IS_ENABLED(CONFIG_CPUSETS)) {
1511 cpuset_cpus_allowed_fallback(p);
1517 do_set_cpus_allowed(p, cpu_possible_mask);
1528 if (state != cpuset) {
1530 * Don't tell them about moving exiting tasks or
1531 * kernel threads (both mm NULL), since they never
1534 if (p->mm && printk_ratelimit()) {
1535 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1536 task_pid_nr(p), p->comm, cpu);
1544 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1547 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1549 lockdep_assert_held(&p->pi_lock);
1551 if (p->nr_cpus_allowed > 1)
1552 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1554 cpu = cpumask_any(&p->cpus_allowed);
1557 * In order not to call set_task_cpu() on a blocking task we need
1558 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1561 * Since this is common to all placement strategies, this lives here.
1563 * [ this allows ->select_task() to simply return task_cpu(p) and
1564 * not worry about this generic constraint ]
1566 if (unlikely(!is_cpu_allowed(p, cpu)))
1567 cpu = select_fallback_rq(task_cpu(p), p);
1572 static void update_avg(u64 *avg, u64 sample)
1574 s64 diff = sample - *avg;
1578 void sched_set_stop_task(int cpu, struct task_struct *stop)
1580 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1581 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1585 * Make it appear like a SCHED_FIFO task, its something
1586 * userspace knows about and won't get confused about.
1588 * Also, it will make PI more or less work without too
1589 * much confusion -- but then, stop work should not
1590 * rely on PI working anyway.
1592 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1594 stop->sched_class = &stop_sched_class;
1597 cpu_rq(cpu)->stop = stop;
1601 * Reset it back to a normal scheduling class so that
1602 * it can die in pieces.
1604 old_stop->sched_class = &rt_sched_class;
1610 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1611 const struct cpumask *new_mask, bool check)
1613 return set_cpus_allowed_ptr(p, new_mask);
1616 #endif /* CONFIG_SMP */
1619 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1623 if (!schedstat_enabled())
1629 if (cpu == rq->cpu) {
1630 __schedstat_inc(rq->ttwu_local);
1631 __schedstat_inc(p->se.statistics.nr_wakeups_local);
1633 struct sched_domain *sd;
1635 __schedstat_inc(p->se.statistics.nr_wakeups_remote);
1637 for_each_domain(rq->cpu, sd) {
1638 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1639 __schedstat_inc(sd->ttwu_wake_remote);
1646 if (wake_flags & WF_MIGRATED)
1647 __schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1648 #endif /* CONFIG_SMP */
1650 __schedstat_inc(rq->ttwu_count);
1651 __schedstat_inc(p->se.statistics.nr_wakeups);
1653 if (wake_flags & WF_SYNC)
1654 __schedstat_inc(p->se.statistics.nr_wakeups_sync);
1657 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1659 activate_task(rq, p, en_flags);
1660 p->on_rq = TASK_ON_RQ_QUEUED;
1662 /* If a worker is waking up, notify the workqueue: */
1663 if (p->flags & PF_WQ_WORKER)
1664 wq_worker_waking_up(p, cpu_of(rq));
1668 * Mark the task runnable and perform wakeup-preemption.
1670 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1671 struct rq_flags *rf)
1673 check_preempt_curr(rq, p, wake_flags);
1674 p->state = TASK_RUNNING;
1675 trace_sched_wakeup(p);
1678 if (p->sched_class->task_woken) {
1680 * Our task @p is fully woken up and running; so its safe to
1681 * drop the rq->lock, hereafter rq is only used for statistics.
1683 rq_unpin_lock(rq, rf);
1684 p->sched_class->task_woken(rq, p);
1685 rq_repin_lock(rq, rf);
1688 if (rq->idle_stamp) {
1689 u64 delta = rq_clock(rq) - rq->idle_stamp;
1690 u64 max = 2*rq->max_idle_balance_cost;
1692 update_avg(&rq->avg_idle, delta);
1694 if (rq->avg_idle > max)
1703 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1704 struct rq_flags *rf)
1706 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
1708 lockdep_assert_held(&rq->lock);
1711 if (p->sched_contributes_to_load)
1712 rq->nr_uninterruptible--;
1714 if (wake_flags & WF_MIGRATED)
1715 en_flags |= ENQUEUE_MIGRATED;
1718 ttwu_activate(rq, p, en_flags);
1719 ttwu_do_wakeup(rq, p, wake_flags, rf);
1723 * Called in case the task @p isn't fully descheduled from its runqueue,
1724 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1725 * since all we need to do is flip p->state to TASK_RUNNING, since
1726 * the task is still ->on_rq.
1728 static int ttwu_remote(struct task_struct *p, int wake_flags)
1734 rq = __task_rq_lock(p, &rf);
1735 if (task_on_rq_queued(p)) {
1736 /* check_preempt_curr() may use rq clock */
1737 update_rq_clock(rq);
1738 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1741 __task_rq_unlock(rq, &rf);
1747 void sched_ttwu_pending(void)
1749 struct rq *rq = this_rq();
1750 struct llist_node *llist = llist_del_all(&rq->wake_list);
1751 struct task_struct *p, *t;
1757 rq_lock_irqsave(rq, &rf);
1758 update_rq_clock(rq);
1760 llist_for_each_entry_safe(p, t, llist, wake_entry)
1761 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
1763 rq_unlock_irqrestore(rq, &rf);
1766 void scheduler_ipi(void)
1769 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1770 * TIF_NEED_RESCHED remotely (for the first time) will also send
1773 preempt_fold_need_resched();
1775 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1779 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1780 * traditionally all their work was done from the interrupt return
1781 * path. Now that we actually do some work, we need to make sure
1784 * Some archs already do call them, luckily irq_enter/exit nest
1787 * Arguably we should visit all archs and update all handlers,
1788 * however a fair share of IPIs are still resched only so this would
1789 * somewhat pessimize the simple resched case.
1792 sched_ttwu_pending();
1795 * Check if someone kicked us for doing the nohz idle load balance.
1797 if (unlikely(got_nohz_idle_kick())) {
1798 this_rq()->idle_balance = 1;
1799 raise_softirq_irqoff(SCHED_SOFTIRQ);
1804 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1806 struct rq *rq = cpu_rq(cpu);
1808 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1810 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1811 if (!set_nr_if_polling(rq->idle))
1812 smp_send_reschedule(cpu);
1814 trace_sched_wake_idle_without_ipi(cpu);
1818 void wake_up_if_idle(int cpu)
1820 struct rq *rq = cpu_rq(cpu);
1825 if (!is_idle_task(rcu_dereference(rq->curr)))
1828 if (set_nr_if_polling(rq->idle)) {
1829 trace_sched_wake_idle_without_ipi(cpu);
1831 rq_lock_irqsave(rq, &rf);
1832 if (is_idle_task(rq->curr))
1833 smp_send_reschedule(cpu);
1834 /* Else CPU is not idle, do nothing here: */
1835 rq_unlock_irqrestore(rq, &rf);
1842 bool cpus_share_cache(int this_cpu, int that_cpu)
1844 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1846 #endif /* CONFIG_SMP */
1848 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1850 struct rq *rq = cpu_rq(cpu);
1853 #if defined(CONFIG_SMP)
1854 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1855 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1856 ttwu_queue_remote(p, cpu, wake_flags);
1862 update_rq_clock(rq);
1863 ttwu_do_activate(rq, p, wake_flags, &rf);
1868 * Notes on Program-Order guarantees on SMP systems.
1872 * The basic program-order guarantee on SMP systems is that when a task [t]
1873 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1874 * execution on its new CPU [c1].
1876 * For migration (of runnable tasks) this is provided by the following means:
1878 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1879 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1880 * rq(c1)->lock (if not at the same time, then in that order).
1881 * C) LOCK of the rq(c1)->lock scheduling in task
1883 * Transitivity guarantees that B happens after A and C after B.
1884 * Note: we only require RCpc transitivity.
1885 * Note: the CPU doing B need not be c0 or c1
1894 * UNLOCK rq(0)->lock
1896 * LOCK rq(0)->lock // orders against CPU0
1898 * UNLOCK rq(0)->lock
1902 * UNLOCK rq(1)->lock
1904 * LOCK rq(1)->lock // orders against CPU2
1907 * UNLOCK rq(1)->lock
1910 * BLOCKING -- aka. SLEEP + WAKEUP
1912 * For blocking we (obviously) need to provide the same guarantee as for
1913 * migration. However the means are completely different as there is no lock
1914 * chain to provide order. Instead we do:
1916 * 1) smp_store_release(X->on_cpu, 0)
1917 * 2) smp_cond_load_acquire(!X->on_cpu)
1921 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1923 * LOCK rq(0)->lock LOCK X->pi_lock
1926 * smp_store_release(X->on_cpu, 0);
1928 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1934 * X->state = RUNNING
1935 * UNLOCK rq(2)->lock
1937 * LOCK rq(2)->lock // orders against CPU1
1940 * UNLOCK rq(2)->lock
1943 * UNLOCK rq(0)->lock
1946 * However; for wakeups there is a second guarantee we must provide, namely we
1947 * must observe the state that lead to our wakeup. That is, not only must our
1948 * task observe its own prior state, it must also observe the stores prior to
1951 * This means that any means of doing remote wakeups must order the CPU doing
1952 * the wakeup against the CPU the task is going to end up running on. This,
1953 * however, is already required for the regular Program-Order guarantee above,
1954 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1959 * try_to_wake_up - wake up a thread
1960 * @p: the thread to be awakened
1961 * @state: the mask of task states that can be woken
1962 * @wake_flags: wake modifier flags (WF_*)
1964 * If (@state & @p->state) @p->state = TASK_RUNNING.
1966 * If the task was not queued/runnable, also place it back on a runqueue.
1968 * Atomic against schedule() which would dequeue a task, also see
1969 * set_current_state().
1971 * Return: %true if @p->state changes (an actual wakeup was done),
1975 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1977 unsigned long flags;
1978 int cpu, success = 0;
1981 * If we are going to wake up a thread waiting for CONDITION we
1982 * need to ensure that CONDITION=1 done by the caller can not be
1983 * reordered with p->state check below. This pairs with mb() in
1984 * set_current_state() the waiting thread does.
1986 raw_spin_lock_irqsave(&p->pi_lock, flags);
1987 smp_mb__after_spinlock();
1988 if (!(p->state & state))
1991 trace_sched_waking(p);
1993 /* We're going to change ->state: */
1998 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1999 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2000 * in smp_cond_load_acquire() below.
2002 * sched_ttwu_pending() try_to_wake_up()
2003 * [S] p->on_rq = 1; [L] P->state
2004 * UNLOCK rq->lock -----.
2008 * LOCK rq->lock -----'
2012 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2014 * Pairs with the UNLOCK+LOCK on rq->lock from the
2015 * last wakeup of our task and the schedule that got our task
2019 if (p->on_rq && ttwu_remote(p, wake_flags))
2024 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2025 * possible to, falsely, observe p->on_cpu == 0.
2027 * One must be running (->on_cpu == 1) in order to remove oneself
2028 * from the runqueue.
2030 * [S] ->on_cpu = 1; [L] ->on_rq
2034 * [S] ->on_rq = 0; [L] ->on_cpu
2036 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2037 * from the consecutive calls to schedule(); the first switching to our
2038 * task, the second putting it to sleep.
2043 * If the owning (remote) CPU is still in the middle of schedule() with
2044 * this task as prev, wait until its done referencing the task.
2046 * Pairs with the smp_store_release() in finish_task().
2048 * This ensures that tasks getting woken will be fully ordered against
2049 * their previous state and preserve Program Order.
2051 smp_cond_load_acquire(&p->on_cpu, !VAL);
2053 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2054 p->state = TASK_WAKING;
2057 delayacct_blkio_end(p);
2058 atomic_dec(&task_rq(p)->nr_iowait);
2061 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2062 if (task_cpu(p) != cpu) {
2063 wake_flags |= WF_MIGRATED;
2064 set_task_cpu(p, cpu);
2067 #else /* CONFIG_SMP */
2070 delayacct_blkio_end(p);
2071 atomic_dec(&task_rq(p)->nr_iowait);
2074 #endif /* CONFIG_SMP */
2076 ttwu_queue(p, cpu, wake_flags);
2078 ttwu_stat(p, cpu, wake_flags);
2080 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2086 * try_to_wake_up_local - try to wake up a local task with rq lock held
2087 * @p: the thread to be awakened
2088 * @rf: request-queue flags for pinning
2090 * Put @p on the run-queue if it's not already there. The caller must
2091 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2094 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2096 struct rq *rq = task_rq(p);
2098 if (WARN_ON_ONCE(rq != this_rq()) ||
2099 WARN_ON_ONCE(p == current))
2102 lockdep_assert_held(&rq->lock);
2104 if (!raw_spin_trylock(&p->pi_lock)) {
2106 * This is OK, because current is on_cpu, which avoids it being
2107 * picked for load-balance and preemption/IRQs are still
2108 * disabled avoiding further scheduler activity on it and we've
2109 * not yet picked a replacement task.
2112 raw_spin_lock(&p->pi_lock);
2116 if (!(p->state & TASK_NORMAL))
2119 trace_sched_waking(p);
2121 if (!task_on_rq_queued(p)) {
2123 delayacct_blkio_end(p);
2124 atomic_dec(&rq->nr_iowait);
2126 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK);
2129 ttwu_do_wakeup(rq, p, 0, rf);
2130 ttwu_stat(p, smp_processor_id(), 0);
2132 raw_spin_unlock(&p->pi_lock);
2136 * wake_up_process - Wake up a specific process
2137 * @p: The process to be woken up.
2139 * Attempt to wake up the nominated process and move it to the set of runnable
2142 * Return: 1 if the process was woken up, 0 if it was already running.
2144 * It may be assumed that this function implies a write memory barrier before
2145 * changing the task state if and only if any tasks are woken up.
2147 int wake_up_process(struct task_struct *p)
2149 return try_to_wake_up(p, TASK_NORMAL, 0);
2151 EXPORT_SYMBOL(wake_up_process);
2153 int wake_up_state(struct task_struct *p, unsigned int state)
2155 return try_to_wake_up(p, state, 0);
2159 * Perform scheduler related setup for a newly forked process p.
2160 * p is forked by current.
2162 * __sched_fork() is basic setup used by init_idle() too:
2164 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2169 p->se.exec_start = 0;
2170 p->se.sum_exec_runtime = 0;
2171 p->se.prev_sum_exec_runtime = 0;
2172 p->se.nr_migrations = 0;
2174 INIT_LIST_HEAD(&p->se.group_node);
2176 #ifdef CONFIG_FAIR_GROUP_SCHED
2177 p->se.cfs_rq = NULL;
2180 #ifdef CONFIG_SCHEDSTATS
2181 /* Even if schedstat is disabled, there should not be garbage */
2182 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2185 RB_CLEAR_NODE(&p->dl.rb_node);
2186 init_dl_task_timer(&p->dl);
2187 init_dl_inactive_task_timer(&p->dl);
2188 __dl_clear_params(p);
2190 INIT_LIST_HEAD(&p->rt.run_list);
2192 p->rt.time_slice = sched_rr_timeslice;
2196 #ifdef CONFIG_PREEMPT_NOTIFIERS
2197 INIT_HLIST_HEAD(&p->preempt_notifiers);
2200 init_numa_balancing(clone_flags, p);
2203 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2205 #ifdef CONFIG_NUMA_BALANCING
2207 void set_numabalancing_state(bool enabled)
2210 static_branch_enable(&sched_numa_balancing);
2212 static_branch_disable(&sched_numa_balancing);
2215 #ifdef CONFIG_PROC_SYSCTL
2216 int sysctl_numa_balancing(struct ctl_table *table, int write,
2217 void __user *buffer, size_t *lenp, loff_t *ppos)
2221 int state = static_branch_likely(&sched_numa_balancing);
2223 if (write && !capable(CAP_SYS_ADMIN))
2228 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2232 set_numabalancing_state(state);
2238 #ifdef CONFIG_SCHEDSTATS
2240 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2241 static bool __initdata __sched_schedstats = false;
2243 static void set_schedstats(bool enabled)
2246 static_branch_enable(&sched_schedstats);
2248 static_branch_disable(&sched_schedstats);
2251 void force_schedstat_enabled(void)
2253 if (!schedstat_enabled()) {
2254 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2255 static_branch_enable(&sched_schedstats);
2259 static int __init setup_schedstats(char *str)
2266 * This code is called before jump labels have been set up, so we can't
2267 * change the static branch directly just yet. Instead set a temporary
2268 * variable so init_schedstats() can do it later.
2270 if (!strcmp(str, "enable")) {
2271 __sched_schedstats = true;
2273 } else if (!strcmp(str, "disable")) {
2274 __sched_schedstats = false;
2279 pr_warn("Unable to parse schedstats=\n");
2283 __setup("schedstats=", setup_schedstats);
2285 static void __init init_schedstats(void)
2287 set_schedstats(__sched_schedstats);
2290 #ifdef CONFIG_PROC_SYSCTL
2291 int sysctl_schedstats(struct ctl_table *table, int write,
2292 void __user *buffer, size_t *lenp, loff_t *ppos)
2296 int state = static_branch_likely(&sched_schedstats);
2298 if (write && !capable(CAP_SYS_ADMIN))
2303 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2307 set_schedstats(state);
2310 #endif /* CONFIG_PROC_SYSCTL */
2311 #else /* !CONFIG_SCHEDSTATS */
2312 static inline void init_schedstats(void) {}
2313 #endif /* CONFIG_SCHEDSTATS */
2316 * fork()/clone()-time setup:
2318 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2320 unsigned long flags;
2321 int cpu = get_cpu();
2323 __sched_fork(clone_flags, p);
2325 * We mark the process as NEW here. This guarantees that
2326 * nobody will actually run it, and a signal or other external
2327 * event cannot wake it up and insert it on the runqueue either.
2329 p->state = TASK_NEW;
2332 * Make sure we do not leak PI boosting priority to the child.
2334 p->prio = current->normal_prio;
2337 * Revert to default priority/policy on fork if requested.
2339 if (unlikely(p->sched_reset_on_fork)) {
2340 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2341 p->policy = SCHED_NORMAL;
2342 p->static_prio = NICE_TO_PRIO(0);
2344 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2345 p->static_prio = NICE_TO_PRIO(0);
2347 p->prio = p->normal_prio = __normal_prio(p);
2348 set_load_weight(p, false);
2351 * We don't need the reset flag anymore after the fork. It has
2352 * fulfilled its duty:
2354 p->sched_reset_on_fork = 0;
2357 if (dl_prio(p->prio)) {
2360 } else if (rt_prio(p->prio)) {
2361 p->sched_class = &rt_sched_class;
2363 p->sched_class = &fair_sched_class;
2366 init_entity_runnable_average(&p->se);
2369 * The child is not yet in the pid-hash so no cgroup attach races,
2370 * and the cgroup is pinned to this child due to cgroup_fork()
2371 * is ran before sched_fork().
2373 * Silence PROVE_RCU.
2375 raw_spin_lock_irqsave(&p->pi_lock, flags);
2377 * We're setting the CPU for the first time, we don't migrate,
2378 * so use __set_task_cpu().
2380 __set_task_cpu(p, cpu);
2381 if (p->sched_class->task_fork)
2382 p->sched_class->task_fork(p);
2383 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2385 #ifdef CONFIG_SCHED_INFO
2386 if (likely(sched_info_on()))
2387 memset(&p->sched_info, 0, sizeof(p->sched_info));
2389 #if defined(CONFIG_SMP)
2392 init_task_preempt_count(p);
2394 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2395 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2402 unsigned long to_ratio(u64 period, u64 runtime)
2404 if (runtime == RUNTIME_INF)
2408 * Doing this here saves a lot of checks in all
2409 * the calling paths, and returning zero seems
2410 * safe for them anyway.
2415 return div64_u64(runtime << BW_SHIFT, period);
2419 * wake_up_new_task - wake up a newly created task for the first time.
2421 * This function will do some initial scheduler statistics housekeeping
2422 * that must be done for every newly created context, then puts the task
2423 * on the runqueue and wakes it.
2425 void wake_up_new_task(struct task_struct *p)
2430 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2431 p->state = TASK_RUNNING;
2434 * Fork balancing, do it here and not earlier because:
2435 * - cpus_allowed can change in the fork path
2436 * - any previously selected CPU might disappear through hotplug
2438 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2439 * as we're not fully set-up yet.
2441 p->recent_used_cpu = task_cpu(p);
2442 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2444 rq = __task_rq_lock(p, &rf);
2445 update_rq_clock(rq);
2446 post_init_entity_util_avg(&p->se);
2448 activate_task(rq, p, ENQUEUE_NOCLOCK);
2449 p->on_rq = TASK_ON_RQ_QUEUED;
2450 trace_sched_wakeup_new(p);
2451 check_preempt_curr(rq, p, WF_FORK);
2453 if (p->sched_class->task_woken) {
2455 * Nothing relies on rq->lock after this, so its fine to
2458 rq_unpin_lock(rq, &rf);
2459 p->sched_class->task_woken(rq, p);
2460 rq_repin_lock(rq, &rf);
2463 task_rq_unlock(rq, p, &rf);
2466 #ifdef CONFIG_PREEMPT_NOTIFIERS
2468 static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
2470 void preempt_notifier_inc(void)
2472 static_branch_inc(&preempt_notifier_key);
2474 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2476 void preempt_notifier_dec(void)
2478 static_branch_dec(&preempt_notifier_key);
2480 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2483 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2484 * @notifier: notifier struct to register
2486 void preempt_notifier_register(struct preempt_notifier *notifier)
2488 if (!static_branch_unlikely(&preempt_notifier_key))
2489 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2491 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2493 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2496 * preempt_notifier_unregister - no longer interested in preemption notifications
2497 * @notifier: notifier struct to unregister
2499 * This is *not* safe to call from within a preemption notifier.
2501 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2503 hlist_del(¬ifier->link);
2505 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2507 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2509 struct preempt_notifier *notifier;
2511 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2512 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2515 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2517 if (static_branch_unlikely(&preempt_notifier_key))
2518 __fire_sched_in_preempt_notifiers(curr);
2522 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2523 struct task_struct *next)
2525 struct preempt_notifier *notifier;
2527 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2528 notifier->ops->sched_out(notifier, next);
2531 static __always_inline void
2532 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2533 struct task_struct *next)
2535 if (static_branch_unlikely(&preempt_notifier_key))
2536 __fire_sched_out_preempt_notifiers(curr, next);
2539 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2541 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2546 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2547 struct task_struct *next)
2551 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2553 static inline void prepare_task(struct task_struct *next)
2557 * Claim the task as running, we do this before switching to it
2558 * such that any running task will have this set.
2564 static inline void finish_task(struct task_struct *prev)
2568 * After ->on_cpu is cleared, the task can be moved to a different CPU.
2569 * We must ensure this doesn't happen until the switch is completely
2572 * In particular, the load of prev->state in finish_task_switch() must
2573 * happen before this.
2575 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
2577 smp_store_release(&prev->on_cpu, 0);
2582 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
2585 * Since the runqueue lock will be released by the next
2586 * task (which is an invalid locking op but in the case
2587 * of the scheduler it's an obvious special-case), so we
2588 * do an early lockdep release here:
2590 rq_unpin_lock(rq, rf);
2591 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2592 #ifdef CONFIG_DEBUG_SPINLOCK
2593 /* this is a valid case when another task releases the spinlock */
2594 rq->lock.owner = next;
2598 static inline void finish_lock_switch(struct rq *rq)
2601 * If we are tracking spinlock dependencies then we have to
2602 * fix up the runqueue lock - which gets 'carried over' from
2603 * prev into current:
2605 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
2606 raw_spin_unlock_irq(&rq->lock);
2610 * NOP if the arch has not defined these:
2613 #ifndef prepare_arch_switch
2614 # define prepare_arch_switch(next) do { } while (0)
2617 #ifndef finish_arch_post_lock_switch
2618 # define finish_arch_post_lock_switch() do { } while (0)
2622 * prepare_task_switch - prepare to switch tasks
2623 * @rq: the runqueue preparing to switch
2624 * @prev: the current task that is being switched out
2625 * @next: the task we are going to switch to.
2627 * This is called with the rq lock held and interrupts off. It must
2628 * be paired with a subsequent finish_task_switch after the context
2631 * prepare_task_switch sets up locking and calls architecture specific
2635 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2636 struct task_struct *next)
2638 kcov_prepare_switch(prev);
2639 sched_info_switch(rq, prev, next);
2640 perf_event_task_sched_out(prev, next);
2642 fire_sched_out_preempt_notifiers(prev, next);
2644 prepare_arch_switch(next);
2648 * finish_task_switch - clean up after a task-switch
2649 * @prev: the thread we just switched away from.
2651 * finish_task_switch must be called after the context switch, paired
2652 * with a prepare_task_switch call before the context switch.
2653 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2654 * and do any other architecture-specific cleanup actions.
2656 * Note that we may have delayed dropping an mm in context_switch(). If
2657 * so, we finish that here outside of the runqueue lock. (Doing it
2658 * with the lock held can cause deadlocks; see schedule() for
2661 * The context switch have flipped the stack from under us and restored the
2662 * local variables which were saved when this task called schedule() in the
2663 * past. prev == current is still correct but we need to recalculate this_rq
2664 * because prev may have moved to another CPU.
2666 static struct rq *finish_task_switch(struct task_struct *prev)
2667 __releases(rq->lock)
2669 struct rq *rq = this_rq();
2670 struct mm_struct *mm = rq->prev_mm;
2674 * The previous task will have left us with a preempt_count of 2
2675 * because it left us after:
2678 * preempt_disable(); // 1
2680 * raw_spin_lock_irq(&rq->lock) // 2
2682 * Also, see FORK_PREEMPT_COUNT.
2684 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2685 "corrupted preempt_count: %s/%d/0x%x\n",
2686 current->comm, current->pid, preempt_count()))
2687 preempt_count_set(FORK_PREEMPT_COUNT);
2692 * A task struct has one reference for the use as "current".
2693 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2694 * schedule one last time. The schedule call will never return, and
2695 * the scheduled task must drop that reference.
2697 * We must observe prev->state before clearing prev->on_cpu (in
2698 * finish_task), otherwise a concurrent wakeup can get prev
2699 * running on another CPU and we could rave with its RUNNING -> DEAD
2700 * transition, resulting in a double drop.
2702 prev_state = prev->state;
2703 vtime_task_switch(prev);
2704 perf_event_task_sched_in(prev, current);
2706 finish_lock_switch(rq);
2707 finish_arch_post_lock_switch();
2708 kcov_finish_switch(current);
2710 fire_sched_in_preempt_notifiers(current);
2712 * When switching through a kernel thread, the loop in
2713 * membarrier_{private,global}_expedited() may have observed that
2714 * kernel thread and not issued an IPI. It is therefore possible to
2715 * schedule between user->kernel->user threads without passing though
2716 * switch_mm(). Membarrier requires a barrier after storing to
2717 * rq->curr, before returning to userspace, so provide them here:
2719 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
2720 * provided by mmdrop(),
2721 * - a sync_core for SYNC_CORE.
2724 membarrier_mm_sync_core_before_usermode(mm);
2727 if (unlikely(prev_state & (TASK_DEAD|TASK_PARKED))) {
2728 switch (prev_state) {
2730 if (prev->sched_class->task_dead)
2731 prev->sched_class->task_dead(prev);
2734 * Remove function-return probe instances associated with this
2735 * task and put them back on the free list.
2737 kprobe_flush_task(prev);
2739 /* Task is done with its stack. */
2740 put_task_stack(prev);
2742 put_task_struct(prev);
2746 kthread_park_complete(prev);
2751 tick_nohz_task_switch();
2757 /* rq->lock is NOT held, but preemption is disabled */
2758 static void __balance_callback(struct rq *rq)
2760 struct callback_head *head, *next;
2761 void (*func)(struct rq *rq);
2762 unsigned long flags;
2764 raw_spin_lock_irqsave(&rq->lock, flags);
2765 head = rq->balance_callback;
2766 rq->balance_callback = NULL;
2768 func = (void (*)(struct rq *))head->func;
2775 raw_spin_unlock_irqrestore(&rq->lock, flags);
2778 static inline void balance_callback(struct rq *rq)
2780 if (unlikely(rq->balance_callback))
2781 __balance_callback(rq);
2786 static inline void balance_callback(struct rq *rq)
2793 * schedule_tail - first thing a freshly forked thread must call.
2794 * @prev: the thread we just switched away from.
2796 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2797 __releases(rq->lock)
2802 * New tasks start with FORK_PREEMPT_COUNT, see there and
2803 * finish_task_switch() for details.
2805 * finish_task_switch() will drop rq->lock() and lower preempt_count
2806 * and the preempt_enable() will end up enabling preemption (on
2807 * PREEMPT_COUNT kernels).
2810 rq = finish_task_switch(prev);
2811 balance_callback(rq);
2814 if (current->set_child_tid)
2815 put_user(task_pid_vnr(current), current->set_child_tid);
2819 * context_switch - switch to the new MM and the new thread's register state.
2821 static __always_inline struct rq *
2822 context_switch(struct rq *rq, struct task_struct *prev,
2823 struct task_struct *next, struct rq_flags *rf)
2825 struct mm_struct *mm, *oldmm;
2827 prepare_task_switch(rq, prev, next);
2830 oldmm = prev->active_mm;
2832 * For paravirt, this is coupled with an exit in switch_to to
2833 * combine the page table reload and the switch backend into
2836 arch_start_context_switch(prev);
2839 * If mm is non-NULL, we pass through switch_mm(). If mm is
2840 * NULL, we will pass through mmdrop() in finish_task_switch().
2841 * Both of these contain the full memory barrier required by
2842 * membarrier after storing to rq->curr, before returning to
2846 next->active_mm = oldmm;
2848 enter_lazy_tlb(oldmm, next);
2850 switch_mm_irqs_off(oldmm, mm, next);
2853 prev->active_mm = NULL;
2854 rq->prev_mm = oldmm;
2857 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2859 prepare_lock_switch(rq, next, rf);
2861 /* Here we just switch the register state and the stack. */
2862 switch_to(prev, next, prev);
2865 return finish_task_switch(prev);
2869 * nr_running and nr_context_switches:
2871 * externally visible scheduler statistics: current number of runnable
2872 * threads, total number of context switches performed since bootup.
2874 unsigned long nr_running(void)
2876 unsigned long i, sum = 0;
2878 for_each_online_cpu(i)
2879 sum += cpu_rq(i)->nr_running;
2885 * Check if only the current task is running on the CPU.
2887 * Caution: this function does not check that the caller has disabled
2888 * preemption, thus the result might have a time-of-check-to-time-of-use
2889 * race. The caller is responsible to use it correctly, for example:
2891 * - from a non-preemptable section (of course)
2893 * - from a thread that is bound to a single CPU
2895 * - in a loop with very short iterations (e.g. a polling loop)
2897 bool single_task_running(void)
2899 return raw_rq()->nr_running == 1;
2901 EXPORT_SYMBOL(single_task_running);
2903 unsigned long long nr_context_switches(void)
2906 unsigned long long sum = 0;
2908 for_each_possible_cpu(i)
2909 sum += cpu_rq(i)->nr_switches;
2915 * IO-wait accounting, and how its mostly bollocks (on SMP).
2917 * The idea behind IO-wait account is to account the idle time that we could
2918 * have spend running if it were not for IO. That is, if we were to improve the
2919 * storage performance, we'd have a proportional reduction in IO-wait time.
2921 * This all works nicely on UP, where, when a task blocks on IO, we account
2922 * idle time as IO-wait, because if the storage were faster, it could've been
2923 * running and we'd not be idle.
2925 * This has been extended to SMP, by doing the same for each CPU. This however
2928 * Imagine for instance the case where two tasks block on one CPU, only the one
2929 * CPU will have IO-wait accounted, while the other has regular idle. Even
2930 * though, if the storage were faster, both could've ran at the same time,
2931 * utilising both CPUs.
2933 * This means, that when looking globally, the current IO-wait accounting on
2934 * SMP is a lower bound, by reason of under accounting.
2936 * Worse, since the numbers are provided per CPU, they are sometimes
2937 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2938 * associated with any one particular CPU, it can wake to another CPU than it
2939 * blocked on. This means the per CPU IO-wait number is meaningless.
2941 * Task CPU affinities can make all that even more 'interesting'.
2944 unsigned long nr_iowait(void)
2946 unsigned long i, sum = 0;
2948 for_each_possible_cpu(i)
2949 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2955 * Consumers of these two interfaces, like for example the cpufreq menu
2956 * governor are using nonsensical data. Boosting frequency for a CPU that has
2957 * IO-wait which might not even end up running the task when it does become
2961 unsigned long nr_iowait_cpu(int cpu)
2963 struct rq *this = cpu_rq(cpu);
2964 return atomic_read(&this->nr_iowait);
2967 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2969 struct rq *rq = this_rq();
2970 *nr_waiters = atomic_read(&rq->nr_iowait);
2971 *load = rq->load.weight;
2977 * sched_exec - execve() is a valuable balancing opportunity, because at
2978 * this point the task has the smallest effective memory and cache footprint.
2980 void sched_exec(void)
2982 struct task_struct *p = current;
2983 unsigned long flags;
2986 raw_spin_lock_irqsave(&p->pi_lock, flags);
2987 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2988 if (dest_cpu == smp_processor_id())
2991 if (likely(cpu_active(dest_cpu))) {
2992 struct migration_arg arg = { p, dest_cpu };
2994 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2995 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2999 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3004 DEFINE_PER_CPU(struct kernel_stat, kstat);
3005 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3007 EXPORT_PER_CPU_SYMBOL(kstat);
3008 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3011 * The function fair_sched_class.update_curr accesses the struct curr
3012 * and its field curr->exec_start; when called from task_sched_runtime(),
3013 * we observe a high rate of cache misses in practice.
3014 * Prefetching this data results in improved performance.
3016 static inline void prefetch_curr_exec_start(struct task_struct *p)
3018 #ifdef CONFIG_FAIR_GROUP_SCHED
3019 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3021 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3024 prefetch(&curr->exec_start);
3028 * Return accounted runtime for the task.
3029 * In case the task is currently running, return the runtime plus current's
3030 * pending runtime that have not been accounted yet.
3032 unsigned long long task_sched_runtime(struct task_struct *p)
3038 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3040 * 64-bit doesn't need locks to atomically read a 64-bit value.
3041 * So we have a optimization chance when the task's delta_exec is 0.
3042 * Reading ->on_cpu is racy, but this is ok.
3044 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3045 * If we race with it entering CPU, unaccounted time is 0. This is
3046 * indistinguishable from the read occurring a few cycles earlier.
3047 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3048 * been accounted, so we're correct here as well.
3050 if (!p->on_cpu || !task_on_rq_queued(p))
3051 return p->se.sum_exec_runtime;
3054 rq = task_rq_lock(p, &rf);
3056 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3057 * project cycles that may never be accounted to this
3058 * thread, breaking clock_gettime().
3060 if (task_current(rq, p) && task_on_rq_queued(p)) {
3061 prefetch_curr_exec_start(p);
3062 update_rq_clock(rq);
3063 p->sched_class->update_curr(rq);
3065 ns = p->se.sum_exec_runtime;
3066 task_rq_unlock(rq, p, &rf);
3072 * This function gets called by the timer code, with HZ frequency.
3073 * We call it with interrupts disabled.
3075 void scheduler_tick(void)
3077 int cpu = smp_processor_id();
3078 struct rq *rq = cpu_rq(cpu);
3079 struct task_struct *curr = rq->curr;
3086 update_rq_clock(rq);
3087 curr->sched_class->task_tick(rq, curr, 0);
3088 cpu_load_update_active(rq);
3089 calc_global_load_tick(rq);
3093 perf_event_task_tick();
3096 rq->idle_balance = idle_cpu(cpu);
3097 trigger_load_balance(rq);
3101 #ifdef CONFIG_NO_HZ_FULL
3105 struct delayed_work work;
3108 static struct tick_work __percpu *tick_work_cpu;
3110 static void sched_tick_remote(struct work_struct *work)
3112 struct delayed_work *dwork = to_delayed_work(work);
3113 struct tick_work *twork = container_of(dwork, struct tick_work, work);
3114 int cpu = twork->cpu;
3115 struct rq *rq = cpu_rq(cpu);
3119 * Handle the tick only if it appears the remote CPU is running in full
3120 * dynticks mode. The check is racy by nature, but missing a tick or
3121 * having one too much is no big deal because the scheduler tick updates
3122 * statistics and checks timeslices in a time-independent way, regardless
3123 * of when exactly it is running.
3125 if (!idle_cpu(cpu) && tick_nohz_tick_stopped_cpu(cpu)) {
3126 struct task_struct *curr;
3129 rq_lock_irq(rq, &rf);
3130 update_rq_clock(rq);
3132 delta = rq_clock_task(rq) - curr->se.exec_start;
3135 * Make sure the next tick runs within a reasonable
3138 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
3139 curr->sched_class->task_tick(rq, curr, 0);
3140 rq_unlock_irq(rq, &rf);
3144 * Run the remote tick once per second (1Hz). This arbitrary
3145 * frequency is large enough to avoid overload but short enough
3146 * to keep scheduler internal stats reasonably up to date.
3148 queue_delayed_work(system_unbound_wq, dwork, HZ);
3151 static void sched_tick_start(int cpu)
3153 struct tick_work *twork;
3155 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3158 WARN_ON_ONCE(!tick_work_cpu);
3160 twork = per_cpu_ptr(tick_work_cpu, cpu);
3162 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
3163 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
3166 #ifdef CONFIG_HOTPLUG_CPU
3167 static void sched_tick_stop(int cpu)
3169 struct tick_work *twork;
3171 if (housekeeping_cpu(cpu, HK_FLAG_TICK))
3174 WARN_ON_ONCE(!tick_work_cpu);
3176 twork = per_cpu_ptr(tick_work_cpu, cpu);
3177 cancel_delayed_work_sync(&twork->work);
3179 #endif /* CONFIG_HOTPLUG_CPU */
3181 int __init sched_tick_offload_init(void)
3183 tick_work_cpu = alloc_percpu(struct tick_work);
3184 BUG_ON(!tick_work_cpu);
3189 #else /* !CONFIG_NO_HZ_FULL */
3190 static inline void sched_tick_start(int cpu) { }
3191 static inline void sched_tick_stop(int cpu) { }
3194 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3195 defined(CONFIG_PREEMPT_TRACER))
3197 * If the value passed in is equal to the current preempt count
3198 * then we just disabled preemption. Start timing the latency.
3200 static inline void preempt_latency_start(int val)
3202 if (preempt_count() == val) {
3203 unsigned long ip = get_lock_parent_ip();
3204 #ifdef CONFIG_DEBUG_PREEMPT
3205 current->preempt_disable_ip = ip;
3207 trace_preempt_off(CALLER_ADDR0, ip);
3211 void preempt_count_add(int val)
3213 #ifdef CONFIG_DEBUG_PREEMPT
3217 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3220 __preempt_count_add(val);
3221 #ifdef CONFIG_DEBUG_PREEMPT
3223 * Spinlock count overflowing soon?
3225 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3228 preempt_latency_start(val);
3230 EXPORT_SYMBOL(preempt_count_add);
3231 NOKPROBE_SYMBOL(preempt_count_add);
3234 * If the value passed in equals to the current preempt count
3235 * then we just enabled preemption. Stop timing the latency.
3237 static inline void preempt_latency_stop(int val)
3239 if (preempt_count() == val)
3240 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3243 void preempt_count_sub(int val)
3245 #ifdef CONFIG_DEBUG_PREEMPT
3249 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3252 * Is the spinlock portion underflowing?
3254 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3255 !(preempt_count() & PREEMPT_MASK)))
3259 preempt_latency_stop(val);
3260 __preempt_count_sub(val);
3262 EXPORT_SYMBOL(preempt_count_sub);
3263 NOKPROBE_SYMBOL(preempt_count_sub);
3266 static inline void preempt_latency_start(int val) { }
3267 static inline void preempt_latency_stop(int val) { }
3270 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3272 #ifdef CONFIG_DEBUG_PREEMPT
3273 return p->preempt_disable_ip;
3280 * Print scheduling while atomic bug:
3282 static noinline void __schedule_bug(struct task_struct *prev)
3284 /* Save this before calling printk(), since that will clobber it */
3285 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3287 if (oops_in_progress)
3290 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3291 prev->comm, prev->pid, preempt_count());
3293 debug_show_held_locks(prev);
3295 if (irqs_disabled())
3296 print_irqtrace_events(prev);
3297 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3298 && in_atomic_preempt_off()) {
3299 pr_err("Preemption disabled at:");
3300 print_ip_sym(preempt_disable_ip);
3304 panic("scheduling while atomic\n");
3307 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3311 * Various schedule()-time debugging checks and statistics:
3313 static inline void schedule_debug(struct task_struct *prev)
3315 #ifdef CONFIG_SCHED_STACK_END_CHECK
3316 if (task_stack_end_corrupted(prev))
3317 panic("corrupted stack end detected inside scheduler\n");
3320 if (unlikely(in_atomic_preempt_off())) {
3321 __schedule_bug(prev);
3322 preempt_count_set(PREEMPT_DISABLED);
3326 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3328 schedstat_inc(this_rq()->sched_count);
3332 * Pick up the highest-prio task:
3334 static inline struct task_struct *
3335 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3337 const struct sched_class *class;
3338 struct task_struct *p;
3341 * Optimization: we know that if all tasks are in the fair class we can
3342 * call that function directly, but only if the @prev task wasn't of a
3343 * higher scheduling class, because otherwise those loose the
3344 * opportunity to pull in more work from other CPUs.
3346 if (likely((prev->sched_class == &idle_sched_class ||
3347 prev->sched_class == &fair_sched_class) &&
3348 rq->nr_running == rq->cfs.h_nr_running)) {
3350 p = fair_sched_class.pick_next_task(rq, prev, rf);
3351 if (unlikely(p == RETRY_TASK))
3354 /* Assumes fair_sched_class->next == idle_sched_class */
3356 p = idle_sched_class.pick_next_task(rq, prev, rf);
3362 for_each_class(class) {
3363 p = class->pick_next_task(rq, prev, rf);
3365 if (unlikely(p == RETRY_TASK))
3371 /* The idle class should always have a runnable task: */
3376 * __schedule() is the main scheduler function.
3378 * The main means of driving the scheduler and thus entering this function are:
3380 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3382 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3383 * paths. For example, see arch/x86/entry_64.S.
3385 * To drive preemption between tasks, the scheduler sets the flag in timer
3386 * interrupt handler scheduler_tick().
3388 * 3. Wakeups don't really cause entry into schedule(). They add a
3389 * task to the run-queue and that's it.
3391 * Now, if the new task added to the run-queue preempts the current
3392 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3393 * called on the nearest possible occasion:
3395 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3397 * - in syscall or exception context, at the next outmost
3398 * preempt_enable(). (this might be as soon as the wake_up()'s
3401 * - in IRQ context, return from interrupt-handler to
3402 * preemptible context
3404 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3407 * - cond_resched() call
3408 * - explicit schedule() call
3409 * - return from syscall or exception to user-space
3410 * - return from interrupt-handler to user-space
3412 * WARNING: must be called with preemption disabled!
3414 static void __sched notrace __schedule(bool preempt)
3416 struct task_struct *prev, *next;
3417 unsigned long *switch_count;
3422 cpu = smp_processor_id();
3426 schedule_debug(prev);
3428 if (sched_feat(HRTICK))
3431 local_irq_disable();
3432 rcu_note_context_switch(preempt);
3435 * Make sure that signal_pending_state()->signal_pending() below
3436 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3437 * done by the caller to avoid the race with signal_wake_up().
3439 * The membarrier system call requires a full memory barrier
3440 * after coming from user-space, before storing to rq->curr.
3443 smp_mb__after_spinlock();
3445 /* Promote REQ to ACT */
3446 rq->clock_update_flags <<= 1;
3447 update_rq_clock(rq);
3449 switch_count = &prev->nivcsw;
3450 if (!preempt && prev->state) {
3451 if (unlikely(signal_pending_state(prev->state, prev))) {
3452 prev->state = TASK_RUNNING;
3454 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
3457 if (prev->in_iowait) {
3458 atomic_inc(&rq->nr_iowait);
3459 delayacct_blkio_start();
3463 * If a worker went to sleep, notify and ask workqueue
3464 * whether it wants to wake up a task to maintain
3467 if (prev->flags & PF_WQ_WORKER) {
3468 struct task_struct *to_wakeup;
3470 to_wakeup = wq_worker_sleeping(prev);
3472 try_to_wake_up_local(to_wakeup, &rf);
3475 switch_count = &prev->nvcsw;
3478 next = pick_next_task(rq, prev, &rf);
3479 clear_tsk_need_resched(prev);
3480 clear_preempt_need_resched();
3482 if (likely(prev != next)) {
3486 * The membarrier system call requires each architecture
3487 * to have a full memory barrier after updating
3488 * rq->curr, before returning to user-space.
3490 * Here are the schemes providing that barrier on the
3491 * various architectures:
3492 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
3493 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
3494 * - finish_lock_switch() for weakly-ordered
3495 * architectures where spin_unlock is a full barrier,
3496 * - switch_to() for arm64 (weakly-ordered, spin_unlock
3497 * is a RELEASE barrier),
3501 trace_sched_switch(preempt, prev, next);
3503 /* Also unlocks the rq: */
3504 rq = context_switch(rq, prev, next, &rf);
3506 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3507 rq_unlock_irq(rq, &rf);
3510 balance_callback(rq);
3513 void __noreturn do_task_dead(void)
3515 /* Causes final put_task_struct in finish_task_switch(): */
3516 set_special_state(TASK_DEAD);
3518 /* Tell freezer to ignore us: */
3519 current->flags |= PF_NOFREEZE;
3524 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3529 static inline void sched_submit_work(struct task_struct *tsk)
3531 if (!tsk->state || tsk_is_pi_blocked(tsk))
3534 * If we are going to sleep and we have plugged IO queued,
3535 * make sure to submit it to avoid deadlocks.
3537 if (blk_needs_flush_plug(tsk))
3538 blk_schedule_flush_plug(tsk);
3541 asmlinkage __visible void __sched schedule(void)
3543 struct task_struct *tsk = current;
3545 sched_submit_work(tsk);
3549 sched_preempt_enable_no_resched();
3550 } while (need_resched());
3552 EXPORT_SYMBOL(schedule);
3555 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
3556 * state (have scheduled out non-voluntarily) by making sure that all
3557 * tasks have either left the run queue or have gone into user space.
3558 * As idle tasks do not do either, they must not ever be preempted
3559 * (schedule out non-voluntarily).
3561 * schedule_idle() is similar to schedule_preempt_disable() except that it
3562 * never enables preemption because it does not call sched_submit_work().
3564 void __sched schedule_idle(void)
3567 * As this skips calling sched_submit_work(), which the idle task does
3568 * regardless because that function is a nop when the task is in a
3569 * TASK_RUNNING state, make sure this isn't used someplace that the
3570 * current task can be in any other state. Note, idle is always in the
3571 * TASK_RUNNING state.
3573 WARN_ON_ONCE(current->state);
3576 } while (need_resched());
3579 #ifdef CONFIG_CONTEXT_TRACKING
3580 asmlinkage __visible void __sched schedule_user(void)
3583 * If we come here after a random call to set_need_resched(),
3584 * or we have been woken up remotely but the IPI has not yet arrived,
3585 * we haven't yet exited the RCU idle mode. Do it here manually until
3586 * we find a better solution.
3588 * NB: There are buggy callers of this function. Ideally we
3589 * should warn if prev_state != CONTEXT_USER, but that will trigger
3590 * too frequently to make sense yet.
3592 enum ctx_state prev_state = exception_enter();
3594 exception_exit(prev_state);
3599 * schedule_preempt_disabled - called with preemption disabled
3601 * Returns with preemption disabled. Note: preempt_count must be 1
3603 void __sched schedule_preempt_disabled(void)
3605 sched_preempt_enable_no_resched();
3610 static void __sched notrace preempt_schedule_common(void)
3614 * Because the function tracer can trace preempt_count_sub()
3615 * and it also uses preempt_enable/disable_notrace(), if
3616 * NEED_RESCHED is set, the preempt_enable_notrace() called
3617 * by the function tracer will call this function again and
3618 * cause infinite recursion.
3620 * Preemption must be disabled here before the function
3621 * tracer can trace. Break up preempt_disable() into two
3622 * calls. One to disable preemption without fear of being
3623 * traced. The other to still record the preemption latency,
3624 * which can also be traced by the function tracer.
3626 preempt_disable_notrace();
3627 preempt_latency_start(1);
3629 preempt_latency_stop(1);
3630 preempt_enable_no_resched_notrace();
3633 * Check again in case we missed a preemption opportunity
3634 * between schedule and now.
3636 } while (need_resched());
3639 #ifdef CONFIG_PREEMPT
3641 * this is the entry point to schedule() from in-kernel preemption
3642 * off of preempt_enable. Kernel preemptions off return from interrupt
3643 * occur there and call schedule directly.
3645 asmlinkage __visible void __sched notrace preempt_schedule(void)
3648 * If there is a non-zero preempt_count or interrupts are disabled,
3649 * we do not want to preempt the current task. Just return..
3651 if (likely(!preemptible()))
3654 preempt_schedule_common();
3656 NOKPROBE_SYMBOL(preempt_schedule);
3657 EXPORT_SYMBOL(preempt_schedule);
3660 * preempt_schedule_notrace - preempt_schedule called by tracing
3662 * The tracing infrastructure uses preempt_enable_notrace to prevent
3663 * recursion and tracing preempt enabling caused by the tracing
3664 * infrastructure itself. But as tracing can happen in areas coming
3665 * from userspace or just about to enter userspace, a preempt enable
3666 * can occur before user_exit() is called. This will cause the scheduler
3667 * to be called when the system is still in usermode.
3669 * To prevent this, the preempt_enable_notrace will use this function
3670 * instead of preempt_schedule() to exit user context if needed before
3671 * calling the scheduler.
3673 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3675 enum ctx_state prev_ctx;
3677 if (likely(!preemptible()))
3682 * Because the function tracer can trace preempt_count_sub()
3683 * and it also uses preempt_enable/disable_notrace(), if
3684 * NEED_RESCHED is set, the preempt_enable_notrace() called
3685 * by the function tracer will call this function again and
3686 * cause infinite recursion.
3688 * Preemption must be disabled here before the function
3689 * tracer can trace. Break up preempt_disable() into two
3690 * calls. One to disable preemption without fear of being
3691 * traced. The other to still record the preemption latency,
3692 * which can also be traced by the function tracer.
3694 preempt_disable_notrace();
3695 preempt_latency_start(1);
3697 * Needs preempt disabled in case user_exit() is traced
3698 * and the tracer calls preempt_enable_notrace() causing
3699 * an infinite recursion.
3701 prev_ctx = exception_enter();
3703 exception_exit(prev_ctx);
3705 preempt_latency_stop(1);
3706 preempt_enable_no_resched_notrace();
3707 } while (need_resched());
3709 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3711 #endif /* CONFIG_PREEMPT */
3714 * this is the entry point to schedule() from kernel preemption
3715 * off of irq context.
3716 * Note, that this is called and return with irqs disabled. This will
3717 * protect us against recursive calling from irq.
3719 asmlinkage __visible void __sched preempt_schedule_irq(void)
3721 enum ctx_state prev_state;
3723 /* Catch callers which need to be fixed */
3724 BUG_ON(preempt_count() || !irqs_disabled());
3726 prev_state = exception_enter();
3732 local_irq_disable();
3733 sched_preempt_enable_no_resched();
3734 } while (need_resched());
3736 exception_exit(prev_state);
3739 int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
3742 return try_to_wake_up(curr->private, mode, wake_flags);
3744 EXPORT_SYMBOL(default_wake_function);
3746 #ifdef CONFIG_RT_MUTEXES
3748 static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
3751 prio = min(prio, pi_task->prio);
3756 static inline int rt_effective_prio(struct task_struct *p, int prio)
3758 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3760 return __rt_effective_prio(pi_task, prio);
3764 * rt_mutex_setprio - set the current priority of a task
3766 * @pi_task: donor task
3768 * This function changes the 'effective' priority of a task. It does
3769 * not touch ->normal_prio like __setscheduler().
3771 * Used by the rt_mutex code to implement priority inheritance
3772 * logic. Call site only calls if the priority of the task changed.
3774 void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
3776 int prio, oldprio, queued, running, queue_flag =
3777 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
3778 const struct sched_class *prev_class;
3782 /* XXX used to be waiter->prio, not waiter->task->prio */
3783 prio = __rt_effective_prio(pi_task, p->normal_prio);
3786 * If nothing changed; bail early.
3788 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
3791 rq = __task_rq_lock(p, &rf);
3792 update_rq_clock(rq);
3794 * Set under pi_lock && rq->lock, such that the value can be used under
3797 * Note that there is loads of tricky to make this pointer cache work
3798 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
3799 * ensure a task is de-boosted (pi_task is set to NULL) before the
3800 * task is allowed to run again (and can exit). This ensures the pointer
3801 * points to a blocked task -- which guaratees the task is present.
3803 p->pi_top_task = pi_task;
3806 * For FIFO/RR we only need to set prio, if that matches we're done.
3808 if (prio == p->prio && !dl_prio(prio))
3812 * Idle task boosting is a nono in general. There is one
3813 * exception, when PREEMPT_RT and NOHZ is active:
3815 * The idle task calls get_next_timer_interrupt() and holds
3816 * the timer wheel base->lock on the CPU and another CPU wants
3817 * to access the timer (probably to cancel it). We can safely
3818 * ignore the boosting request, as the idle CPU runs this code
3819 * with interrupts disabled and will complete the lock
3820 * protected section without being interrupted. So there is no
3821 * real need to boost.
3823 if (unlikely(p == rq->idle)) {
3824 WARN_ON(p != rq->curr);
3825 WARN_ON(p->pi_blocked_on);
3829 trace_sched_pi_setprio(p, pi_task);
3832 if (oldprio == prio)
3833 queue_flag &= ~DEQUEUE_MOVE;
3835 prev_class = p->sched_class;
3836 queued = task_on_rq_queued(p);
3837 running = task_current(rq, p);
3839 dequeue_task(rq, p, queue_flag);
3841 put_prev_task(rq, p);
3844 * Boosting condition are:
3845 * 1. -rt task is running and holds mutex A
3846 * --> -dl task blocks on mutex A
3848 * 2. -dl task is running and holds mutex A
3849 * --> -dl task blocks on mutex A and could preempt the
3852 if (dl_prio(prio)) {
3853 if (!dl_prio(p->normal_prio) ||
3854 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3855 p->dl.dl_boosted = 1;
3856 queue_flag |= ENQUEUE_REPLENISH;
3858 p->dl.dl_boosted = 0;
3859 p->sched_class = &dl_sched_class;
3860 } else if (rt_prio(prio)) {
3861 if (dl_prio(oldprio))
3862 p->dl.dl_boosted = 0;
3864 queue_flag |= ENQUEUE_HEAD;
3865 p->sched_class = &rt_sched_class;
3867 if (dl_prio(oldprio))
3868 p->dl.dl_boosted = 0;
3869 if (rt_prio(oldprio))
3871 p->sched_class = &fair_sched_class;
3877 enqueue_task(rq, p, queue_flag);
3879 set_curr_task(rq, p);
3881 check_class_changed(rq, p, prev_class, oldprio);
3883 /* Avoid rq from going away on us: */
3885 __task_rq_unlock(rq, &rf);
3887 balance_callback(rq);
3891 static inline int rt_effective_prio(struct task_struct *p, int prio)
3897 void set_user_nice(struct task_struct *p, long nice)
3899 bool queued, running;
3900 int old_prio, delta;
3904 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3907 * We have to be careful, if called from sys_setpriority(),
3908 * the task might be in the middle of scheduling on another CPU.
3910 rq = task_rq_lock(p, &rf);
3911 update_rq_clock(rq);
3914 * The RT priorities are set via sched_setscheduler(), but we still
3915 * allow the 'normal' nice value to be set - but as expected
3916 * it wont have any effect on scheduling until the task is
3917 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3919 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3920 p->static_prio = NICE_TO_PRIO(nice);
3923 queued = task_on_rq_queued(p);
3924 running = task_current(rq, p);
3926 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
3928 put_prev_task(rq, p);
3930 p->static_prio = NICE_TO_PRIO(nice);
3931 set_load_weight(p, true);
3933 p->prio = effective_prio(p);
3934 delta = p->prio - old_prio;
3937 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
3939 * If the task increased its priority or is running and
3940 * lowered its priority, then reschedule its CPU:
3942 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3946 set_curr_task(rq, p);
3948 task_rq_unlock(rq, p, &rf);
3950 EXPORT_SYMBOL(set_user_nice);
3953 * can_nice - check if a task can reduce its nice value
3957 int can_nice(const struct task_struct *p, const int nice)
3959 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3960 int nice_rlim = nice_to_rlimit(nice);
3962 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3963 capable(CAP_SYS_NICE));
3966 #ifdef __ARCH_WANT_SYS_NICE
3969 * sys_nice - change the priority of the current process.
3970 * @increment: priority increment
3972 * sys_setpriority is a more generic, but much slower function that
3973 * does similar things.
3975 SYSCALL_DEFINE1(nice, int, increment)
3980 * Setpriority might change our priority at the same moment.
3981 * We don't have to worry. Conceptually one call occurs first
3982 * and we have a single winner.
3984 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3985 nice = task_nice(current) + increment;
3987 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3988 if (increment < 0 && !can_nice(current, nice))
3991 retval = security_task_setnice(current, nice);
3995 set_user_nice(current, nice);
4002 * task_prio - return the priority value of a given task.
4003 * @p: the task in question.
4005 * Return: The priority value as seen by users in /proc.
4006 * RT tasks are offset by -200. Normal tasks are centered
4007 * around 0, value goes from -16 to +15.
4009 int task_prio(const struct task_struct *p)
4011 return p->prio - MAX_RT_PRIO;
4015 * idle_cpu - is a given CPU idle currently?
4016 * @cpu: the processor in question.
4018 * Return: 1 if the CPU is currently idle. 0 otherwise.
4020 int idle_cpu(int cpu)
4022 struct rq *rq = cpu_rq(cpu);
4024 if (rq->curr != rq->idle)
4031 if (!llist_empty(&rq->wake_list))
4039 * available_idle_cpu - is a given CPU idle for enqueuing work.
4040 * @cpu: the CPU in question.
4042 * Return: 1 if the CPU is currently idle. 0 otherwise.
4044 int available_idle_cpu(int cpu)
4049 if (vcpu_is_preempted(cpu))
4056 * idle_task - return the idle task for a given CPU.
4057 * @cpu: the processor in question.
4059 * Return: The idle task for the CPU @cpu.
4061 struct task_struct *idle_task(int cpu)
4063 return cpu_rq(cpu)->idle;
4067 * find_process_by_pid - find a process with a matching PID value.
4068 * @pid: the pid in question.
4070 * The task of @pid, if found. %NULL otherwise.
4072 static struct task_struct *find_process_by_pid(pid_t pid)
4074 return pid ? find_task_by_vpid(pid) : current;
4078 * sched_setparam() passes in -1 for its policy, to let the functions
4079 * it calls know not to change it.
4081 #define SETPARAM_POLICY -1
4083 static void __setscheduler_params(struct task_struct *p,
4084 const struct sched_attr *attr)
4086 int policy = attr->sched_policy;
4088 if (policy == SETPARAM_POLICY)
4093 if (dl_policy(policy))
4094 __setparam_dl(p, attr);
4095 else if (fair_policy(policy))
4096 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
4099 * __sched_setscheduler() ensures attr->sched_priority == 0 when
4100 * !rt_policy. Always setting this ensures that things like
4101 * getparam()/getattr() don't report silly values for !rt tasks.
4103 p->rt_priority = attr->sched_priority;
4104 p->normal_prio = normal_prio(p);
4105 set_load_weight(p, true);
4108 /* Actually do priority change: must hold pi & rq lock. */
4109 static void __setscheduler(struct rq *rq, struct task_struct *p,
4110 const struct sched_attr *attr, bool keep_boost)
4112 __setscheduler_params(p, attr);
4115 * Keep a potential priority boosting if called from
4116 * sched_setscheduler().
4118 p->prio = normal_prio(p);
4120 p->prio = rt_effective_prio(p, p->prio);
4122 if (dl_prio(p->prio))
4123 p->sched_class = &dl_sched_class;
4124 else if (rt_prio(p->prio))
4125 p->sched_class = &rt_sched_class;
4127 p->sched_class = &fair_sched_class;
4131 * Check the target process has a UID that matches the current process's:
4133 static bool check_same_owner(struct task_struct *p)
4135 const struct cred *cred = current_cred(), *pcred;
4139 pcred = __task_cred(p);
4140 match = (uid_eq(cred->euid, pcred->euid) ||
4141 uid_eq(cred->euid, pcred->uid));
4146 static int __sched_setscheduler(struct task_struct *p,
4147 const struct sched_attr *attr,
4150 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4151 MAX_RT_PRIO - 1 - attr->sched_priority;
4152 int retval, oldprio, oldpolicy = -1, queued, running;
4153 int new_effective_prio, policy = attr->sched_policy;
4154 const struct sched_class *prev_class;
4157 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
4160 /* The pi code expects interrupts enabled */
4161 BUG_ON(pi && in_interrupt());
4163 /* Double check policy once rq lock held: */
4165 reset_on_fork = p->sched_reset_on_fork;
4166 policy = oldpolicy = p->policy;
4168 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4170 if (!valid_policy(policy))
4174 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
4178 * Valid priorities for SCHED_FIFO and SCHED_RR are
4179 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4180 * SCHED_BATCH and SCHED_IDLE is 0.
4182 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4183 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4185 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4186 (rt_policy(policy) != (attr->sched_priority != 0)))
4190 * Allow unprivileged RT tasks to decrease priority:
4192 if (user && !capable(CAP_SYS_NICE)) {
4193 if (fair_policy(policy)) {
4194 if (attr->sched_nice < task_nice(p) &&
4195 !can_nice(p, attr->sched_nice))
4199 if (rt_policy(policy)) {
4200 unsigned long rlim_rtprio =
4201 task_rlimit(p, RLIMIT_RTPRIO);
4203 /* Can't set/change the rt policy: */
4204 if (policy != p->policy && !rlim_rtprio)
4207 /* Can't increase priority: */
4208 if (attr->sched_priority > p->rt_priority &&
4209 attr->sched_priority > rlim_rtprio)
4214 * Can't set/change SCHED_DEADLINE policy at all for now
4215 * (safest behavior); in the future we would like to allow
4216 * unprivileged DL tasks to increase their relative deadline
4217 * or reduce their runtime (both ways reducing utilization)
4219 if (dl_policy(policy))
4223 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4224 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4226 if (idle_policy(p->policy) && !idle_policy(policy)) {
4227 if (!can_nice(p, task_nice(p)))
4231 /* Can't change other user's priorities: */
4232 if (!check_same_owner(p))
4235 /* Normal users shall not reset the sched_reset_on_fork flag: */
4236 if (p->sched_reset_on_fork && !reset_on_fork)
4241 if (attr->sched_flags & SCHED_FLAG_SUGOV)
4244 retval = security_task_setscheduler(p);
4250 * Make sure no PI-waiters arrive (or leave) while we are
4251 * changing the priority of the task:
4253 * To be able to change p->policy safely, the appropriate
4254 * runqueue lock must be held.
4256 rq = task_rq_lock(p, &rf);
4257 update_rq_clock(rq);
4260 * Changing the policy of the stop threads its a very bad idea:
4262 if (p == rq->stop) {
4263 task_rq_unlock(rq, p, &rf);
4268 * If not changing anything there's no need to proceed further,
4269 * but store a possible modification of reset_on_fork.
4271 if (unlikely(policy == p->policy)) {
4272 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4274 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4276 if (dl_policy(policy) && dl_param_changed(p, attr))
4279 p->sched_reset_on_fork = reset_on_fork;
4280 task_rq_unlock(rq, p, &rf);
4286 #ifdef CONFIG_RT_GROUP_SCHED
4288 * Do not allow realtime tasks into groups that have no runtime
4291 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4292 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4293 !task_group_is_autogroup(task_group(p))) {
4294 task_rq_unlock(rq, p, &rf);
4299 if (dl_bandwidth_enabled() && dl_policy(policy) &&
4300 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
4301 cpumask_t *span = rq->rd->span;
4304 * Don't allow tasks with an affinity mask smaller than
4305 * the entire root_domain to become SCHED_DEADLINE. We
4306 * will also fail if there's no bandwidth available.
4308 if (!cpumask_subset(span, &p->cpus_allowed) ||
4309 rq->rd->dl_bw.bw == 0) {
4310 task_rq_unlock(rq, p, &rf);
4317 /* Re-check policy now with rq lock held: */
4318 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4319 policy = oldpolicy = -1;
4320 task_rq_unlock(rq, p, &rf);
4325 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4326 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4329 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
4330 task_rq_unlock(rq, p, &rf);
4334 p->sched_reset_on_fork = reset_on_fork;
4339 * Take priority boosted tasks into account. If the new
4340 * effective priority is unchanged, we just store the new
4341 * normal parameters and do not touch the scheduler class and
4342 * the runqueue. This will be done when the task deboost
4345 new_effective_prio = rt_effective_prio(p, newprio);
4346 if (new_effective_prio == oldprio)
4347 queue_flags &= ~DEQUEUE_MOVE;
4350 queued = task_on_rq_queued(p);
4351 running = task_current(rq, p);
4353 dequeue_task(rq, p, queue_flags);
4355 put_prev_task(rq, p);
4357 prev_class = p->sched_class;
4358 __setscheduler(rq, p, attr, pi);
4362 * We enqueue to tail when the priority of a task is
4363 * increased (user space view).
4365 if (oldprio < p->prio)
4366 queue_flags |= ENQUEUE_HEAD;
4368 enqueue_task(rq, p, queue_flags);
4371 set_curr_task(rq, p);
4373 check_class_changed(rq, p, prev_class, oldprio);
4375 /* Avoid rq from going away on us: */
4377 task_rq_unlock(rq, p, &rf);
4380 rt_mutex_adjust_pi(p);
4382 /* Run balance callbacks after we've adjusted the PI chain: */
4383 balance_callback(rq);
4389 static int _sched_setscheduler(struct task_struct *p, int policy,
4390 const struct sched_param *param, bool check)
4392 struct sched_attr attr = {
4393 .sched_policy = policy,
4394 .sched_priority = param->sched_priority,
4395 .sched_nice = PRIO_TO_NICE(p->static_prio),
4398 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4399 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4400 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4401 policy &= ~SCHED_RESET_ON_FORK;
4402 attr.sched_policy = policy;
4405 return __sched_setscheduler(p, &attr, check, true);
4408 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4409 * @p: the task in question.
4410 * @policy: new policy.
4411 * @param: structure containing the new RT priority.
4413 * Return: 0 on success. An error code otherwise.
4415 * NOTE that the task may be already dead.
4417 int sched_setscheduler(struct task_struct *p, int policy,
4418 const struct sched_param *param)
4420 return _sched_setscheduler(p, policy, param, true);
4422 EXPORT_SYMBOL_GPL(sched_setscheduler);
4424 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4426 return __sched_setscheduler(p, attr, true, true);
4428 EXPORT_SYMBOL_GPL(sched_setattr);
4430 int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
4432 return __sched_setscheduler(p, attr, false, true);
4436 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4437 * @p: the task in question.
4438 * @policy: new policy.
4439 * @param: structure containing the new RT priority.
4441 * Just like sched_setscheduler, only don't bother checking if the
4442 * current context has permission. For example, this is needed in
4443 * stop_machine(): we create temporary high priority worker threads,
4444 * but our caller might not have that capability.
4446 * Return: 0 on success. An error code otherwise.
4448 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4449 const struct sched_param *param)
4451 return _sched_setscheduler(p, policy, param, false);
4453 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4456 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4458 struct sched_param lparam;
4459 struct task_struct *p;
4462 if (!param || pid < 0)
4464 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4469 p = find_process_by_pid(pid);
4471 retval = sched_setscheduler(p, policy, &lparam);
4478 * Mimics kernel/events/core.c perf_copy_attr().
4480 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4485 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4488 /* Zero the full structure, so that a short copy will be nice: */
4489 memset(attr, 0, sizeof(*attr));
4491 ret = get_user(size, &uattr->size);
4495 /* Bail out on silly large: */
4496 if (size > PAGE_SIZE)
4499 /* ABI compatibility quirk: */
4501 size = SCHED_ATTR_SIZE_VER0;
4503 if (size < SCHED_ATTR_SIZE_VER0)
4507 * If we're handed a bigger struct than we know of,
4508 * ensure all the unknown bits are 0 - i.e. new
4509 * user-space does not rely on any kernel feature
4510 * extensions we dont know about yet.
4512 if (size > sizeof(*attr)) {
4513 unsigned char __user *addr;
4514 unsigned char __user *end;
4517 addr = (void __user *)uattr + sizeof(*attr);
4518 end = (void __user *)uattr + size;
4520 for (; addr < end; addr++) {
4521 ret = get_user(val, addr);
4527 size = sizeof(*attr);
4530 ret = copy_from_user(attr, uattr, size);
4535 * XXX: Do we want to be lenient like existing syscalls; or do we want
4536 * to be strict and return an error on out-of-bounds values?
4538 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4543 put_user(sizeof(*attr), &uattr->size);
4548 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4549 * @pid: the pid in question.
4550 * @policy: new policy.
4551 * @param: structure containing the new RT priority.
4553 * Return: 0 on success. An error code otherwise.
4555 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4560 return do_sched_setscheduler(pid, policy, param);
4564 * sys_sched_setparam - set/change the RT priority of a thread
4565 * @pid: the pid in question.
4566 * @param: structure containing the new RT priority.
4568 * Return: 0 on success. An error code otherwise.
4570 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4572 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4576 * sys_sched_setattr - same as above, but with extended sched_attr
4577 * @pid: the pid in question.
4578 * @uattr: structure containing the extended parameters.
4579 * @flags: for future extension.
4581 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4582 unsigned int, flags)
4584 struct sched_attr attr;
4585 struct task_struct *p;
4588 if (!uattr || pid < 0 || flags)
4591 retval = sched_copy_attr(uattr, &attr);
4595 if ((int)attr.sched_policy < 0)
4600 p = find_process_by_pid(pid);
4602 retval = sched_setattr(p, &attr);
4609 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4610 * @pid: the pid in question.
4612 * Return: On success, the policy of the thread. Otherwise, a negative error
4615 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4617 struct task_struct *p;
4625 p = find_process_by_pid(pid);
4627 retval = security_task_getscheduler(p);
4630 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4637 * sys_sched_getparam - get the RT priority of a thread
4638 * @pid: the pid in question.
4639 * @param: structure containing the RT priority.
4641 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4644 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4646 struct sched_param lp = { .sched_priority = 0 };
4647 struct task_struct *p;
4650 if (!param || pid < 0)
4654 p = find_process_by_pid(pid);
4659 retval = security_task_getscheduler(p);
4663 if (task_has_rt_policy(p))
4664 lp.sched_priority = p->rt_priority;
4668 * This one might sleep, we cannot do it with a spinlock held ...
4670 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4679 static int sched_read_attr(struct sched_attr __user *uattr,
4680 struct sched_attr *attr,
4685 if (!access_ok(VERIFY_WRITE, uattr, usize))
4689 * If we're handed a smaller struct than we know of,
4690 * ensure all the unknown bits are 0 - i.e. old
4691 * user-space does not get uncomplete information.
4693 if (usize < sizeof(*attr)) {
4694 unsigned char *addr;
4697 addr = (void *)attr + usize;
4698 end = (void *)attr + sizeof(*attr);
4700 for (; addr < end; addr++) {
4708 ret = copy_to_user(uattr, attr, attr->size);
4716 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4717 * @pid: the pid in question.
4718 * @uattr: structure containing the extended parameters.
4719 * @size: sizeof(attr) for fwd/bwd comp.
4720 * @flags: for future extension.
4722 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4723 unsigned int, size, unsigned int, flags)
4725 struct sched_attr attr = {
4726 .size = sizeof(struct sched_attr),
4728 struct task_struct *p;
4731 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4732 size < SCHED_ATTR_SIZE_VER0 || flags)
4736 p = find_process_by_pid(pid);
4741 retval = security_task_getscheduler(p);
4745 attr.sched_policy = p->policy;
4746 if (p->sched_reset_on_fork)
4747 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4748 if (task_has_dl_policy(p))
4749 __getparam_dl(p, &attr);
4750 else if (task_has_rt_policy(p))
4751 attr.sched_priority = p->rt_priority;
4753 attr.sched_nice = task_nice(p);
4757 retval = sched_read_attr(uattr, &attr, size);
4765 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4767 cpumask_var_t cpus_allowed, new_mask;
4768 struct task_struct *p;
4773 p = find_process_by_pid(pid);
4779 /* Prevent p going away */
4783 if (p->flags & PF_NO_SETAFFINITY) {
4787 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4791 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4793 goto out_free_cpus_allowed;
4796 if (!check_same_owner(p)) {
4798 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4800 goto out_free_new_mask;
4805 retval = security_task_setscheduler(p);
4807 goto out_free_new_mask;
4810 cpuset_cpus_allowed(p, cpus_allowed);
4811 cpumask_and(new_mask, in_mask, cpus_allowed);
4814 * Since bandwidth control happens on root_domain basis,
4815 * if admission test is enabled, we only admit -deadline
4816 * tasks allowed to run on all the CPUs in the task's
4820 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4822 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4825 goto out_free_new_mask;
4831 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4834 cpuset_cpus_allowed(p, cpus_allowed);
4835 if (!cpumask_subset(new_mask, cpus_allowed)) {
4837 * We must have raced with a concurrent cpuset
4838 * update. Just reset the cpus_allowed to the
4839 * cpuset's cpus_allowed
4841 cpumask_copy(new_mask, cpus_allowed);
4846 free_cpumask_var(new_mask);
4847 out_free_cpus_allowed:
4848 free_cpumask_var(cpus_allowed);
4854 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4855 struct cpumask *new_mask)
4857 if (len < cpumask_size())
4858 cpumask_clear(new_mask);
4859 else if (len > cpumask_size())
4860 len = cpumask_size();
4862 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4866 * sys_sched_setaffinity - set the CPU affinity of a process
4867 * @pid: pid of the process
4868 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4869 * @user_mask_ptr: user-space pointer to the new CPU mask
4871 * Return: 0 on success. An error code otherwise.
4873 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4874 unsigned long __user *, user_mask_ptr)
4876 cpumask_var_t new_mask;
4879 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4882 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4884 retval = sched_setaffinity(pid, new_mask);
4885 free_cpumask_var(new_mask);
4889 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4891 struct task_struct *p;
4892 unsigned long flags;
4898 p = find_process_by_pid(pid);
4902 retval = security_task_getscheduler(p);
4906 raw_spin_lock_irqsave(&p->pi_lock, flags);
4907 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4908 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4917 * sys_sched_getaffinity - get the CPU affinity of a process
4918 * @pid: pid of the process
4919 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4920 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4922 * Return: size of CPU mask copied to user_mask_ptr on success. An
4923 * error code otherwise.
4925 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4926 unsigned long __user *, user_mask_ptr)
4931 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4933 if (len & (sizeof(unsigned long)-1))
4936 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4939 ret = sched_getaffinity(pid, mask);
4941 unsigned int retlen = min(len, cpumask_size());
4943 if (copy_to_user(user_mask_ptr, mask, retlen))
4948 free_cpumask_var(mask);
4954 * sys_sched_yield - yield the current processor to other threads.
4956 * This function yields the current CPU to other tasks. If there are no
4957 * other threads running on this CPU then this function will return.
4961 static void do_sched_yield(void)
4966 local_irq_disable();
4970 schedstat_inc(rq->yld_count);
4971 current->sched_class->yield_task(rq);
4974 * Since we are going to call schedule() anyway, there's
4975 * no need to preempt or enable interrupts:
4979 sched_preempt_enable_no_resched();
4984 SYSCALL_DEFINE0(sched_yield)
4990 #ifndef CONFIG_PREEMPT
4991 int __sched _cond_resched(void)
4993 if (should_resched(0)) {
4994 preempt_schedule_common();
5000 EXPORT_SYMBOL(_cond_resched);
5004 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5005 * call schedule, and on return reacquire the lock.
5007 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5008 * operations here to prevent schedule() from being called twice (once via
5009 * spin_unlock(), once by hand).
5011 int __cond_resched_lock(spinlock_t *lock)
5013 int resched = should_resched(PREEMPT_LOCK_OFFSET);
5016 lockdep_assert_held(lock);
5018 if (spin_needbreak(lock) || resched) {
5021 preempt_schedule_common();
5029 EXPORT_SYMBOL(__cond_resched_lock);
5032 * yield - yield the current processor to other threads.
5034 * Do not ever use this function, there's a 99% chance you're doing it wrong.
5036 * The scheduler is at all times free to pick the calling task as the most
5037 * eligible task to run, if removing the yield() call from your code breaks
5038 * it, its already broken.
5040 * Typical broken usage is:
5045 * where one assumes that yield() will let 'the other' process run that will
5046 * make event true. If the current task is a SCHED_FIFO task that will never
5047 * happen. Never use yield() as a progress guarantee!!
5049 * If you want to use yield() to wait for something, use wait_event().
5050 * If you want to use yield() to be 'nice' for others, use cond_resched().
5051 * If you still want to use yield(), do not!
5053 void __sched yield(void)
5055 set_current_state(TASK_RUNNING);
5058 EXPORT_SYMBOL(yield);
5061 * yield_to - yield the current processor to another thread in
5062 * your thread group, or accelerate that thread toward the
5063 * processor it's on.
5065 * @preempt: whether task preemption is allowed or not
5067 * It's the caller's job to ensure that the target task struct
5068 * can't go away on us before we can do any checks.
5071 * true (>0) if we indeed boosted the target task.
5072 * false (0) if we failed to boost the target.
5073 * -ESRCH if there's no task to yield to.
5075 int __sched yield_to(struct task_struct *p, bool preempt)
5077 struct task_struct *curr = current;
5078 struct rq *rq, *p_rq;
5079 unsigned long flags;
5082 local_irq_save(flags);
5088 * If we're the only runnable task on the rq and target rq also
5089 * has only one task, there's absolutely no point in yielding.
5091 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5096 double_rq_lock(rq, p_rq);
5097 if (task_rq(p) != p_rq) {
5098 double_rq_unlock(rq, p_rq);
5102 if (!curr->sched_class->yield_to_task)
5105 if (curr->sched_class != p->sched_class)
5108 if (task_running(p_rq, p) || p->state)
5111 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5113 schedstat_inc(rq->yld_count);
5115 * Make p's CPU reschedule; pick_next_entity takes care of
5118 if (preempt && rq != p_rq)
5123 double_rq_unlock(rq, p_rq);
5125 local_irq_restore(flags);
5132 EXPORT_SYMBOL_GPL(yield_to);
5134 int io_schedule_prepare(void)
5136 int old_iowait = current->in_iowait;
5138 current->in_iowait = 1;
5139 blk_schedule_flush_plug(current);
5144 void io_schedule_finish(int token)
5146 current->in_iowait = token;
5150 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5151 * that process accounting knows that this is a task in IO wait state.
5153 long __sched io_schedule_timeout(long timeout)
5158 token = io_schedule_prepare();
5159 ret = schedule_timeout(timeout);
5160 io_schedule_finish(token);
5164 EXPORT_SYMBOL(io_schedule_timeout);
5166 void io_schedule(void)
5170 token = io_schedule_prepare();
5172 io_schedule_finish(token);
5174 EXPORT_SYMBOL(io_schedule);
5177 * sys_sched_get_priority_max - return maximum RT priority.
5178 * @policy: scheduling class.
5180 * Return: On success, this syscall returns the maximum
5181 * rt_priority that can be used by a given scheduling class.
5182 * On failure, a negative error code is returned.
5184 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5191 ret = MAX_USER_RT_PRIO-1;
5193 case SCHED_DEADLINE:
5204 * sys_sched_get_priority_min - return minimum RT priority.
5205 * @policy: scheduling class.
5207 * Return: On success, this syscall returns the minimum
5208 * rt_priority that can be used by a given scheduling class.
5209 * On failure, a negative error code is returned.
5211 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5220 case SCHED_DEADLINE:
5229 static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
5231 struct task_struct *p;
5232 unsigned int time_slice;
5242 p = find_process_by_pid(pid);
5246 retval = security_task_getscheduler(p);
5250 rq = task_rq_lock(p, &rf);
5252 if (p->sched_class->get_rr_interval)
5253 time_slice = p->sched_class->get_rr_interval(rq, p);
5254 task_rq_unlock(rq, p, &rf);
5257 jiffies_to_timespec64(time_slice, t);
5266 * sys_sched_rr_get_interval - return the default timeslice of a process.
5267 * @pid: pid of the process.
5268 * @interval: userspace pointer to the timeslice value.
5270 * this syscall writes the default timeslice value of a given process
5271 * into the user-space timespec buffer. A value of '0' means infinity.
5273 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5276 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5277 struct timespec __user *, interval)
5279 struct timespec64 t;
5280 int retval = sched_rr_get_interval(pid, &t);
5283 retval = put_timespec64(&t, interval);
5288 #ifdef CONFIG_COMPAT
5289 COMPAT_SYSCALL_DEFINE2(sched_rr_get_interval,
5291 struct compat_timespec __user *, interval)
5293 struct timespec64 t;
5294 int retval = sched_rr_get_interval(pid, &t);
5297 retval = compat_put_timespec64(&t, interval);
5302 void sched_show_task(struct task_struct *p)
5304 unsigned long free = 0;
5307 if (!try_get_task_stack(p))
5310 printk(KERN_INFO "%-15.15s %c", p->comm, task_state_to_char(p));
5312 if (p->state == TASK_RUNNING)
5313 printk(KERN_CONT " running task ");
5314 #ifdef CONFIG_DEBUG_STACK_USAGE
5315 free = stack_not_used(p);
5320 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5322 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5323 task_pid_nr(p), ppid,
5324 (unsigned long)task_thread_info(p)->flags);
5326 print_worker_info(KERN_INFO, p);
5327 show_stack(p, NULL);
5330 EXPORT_SYMBOL_GPL(sched_show_task);
5333 state_filter_match(unsigned long state_filter, struct task_struct *p)
5335 /* no filter, everything matches */
5339 /* filter, but doesn't match */
5340 if (!(p->state & state_filter))
5344 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
5347 if (state_filter == TASK_UNINTERRUPTIBLE && p->state == TASK_IDLE)
5354 void show_state_filter(unsigned long state_filter)
5356 struct task_struct *g, *p;
5358 #if BITS_PER_LONG == 32
5360 " task PC stack pid father\n");
5363 " task PC stack pid father\n");
5366 for_each_process_thread(g, p) {
5368 * reset the NMI-timeout, listing all files on a slow
5369 * console might take a lot of time:
5370 * Also, reset softlockup watchdogs on all CPUs, because
5371 * another CPU might be blocked waiting for us to process
5374 touch_nmi_watchdog();
5375 touch_all_softlockup_watchdogs();
5376 if (state_filter_match(state_filter, p))
5380 #ifdef CONFIG_SCHED_DEBUG
5382 sysrq_sched_debug_show();
5386 * Only show locks if all tasks are dumped:
5389 debug_show_all_locks();
5393 * init_idle - set up an idle thread for a given CPU
5394 * @idle: task in question
5395 * @cpu: CPU the idle task belongs to
5397 * NOTE: this function does not set the idle thread's NEED_RESCHED
5398 * flag, to make booting more robust.
5400 void init_idle(struct task_struct *idle, int cpu)
5402 struct rq *rq = cpu_rq(cpu);
5403 unsigned long flags;
5405 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5406 raw_spin_lock(&rq->lock);
5408 __sched_fork(0, idle);
5409 idle->state = TASK_RUNNING;
5410 idle->se.exec_start = sched_clock();
5411 idle->flags |= PF_IDLE;
5413 kasan_unpoison_task_stack(idle);
5417 * Its possible that init_idle() gets called multiple times on a task,
5418 * in that case do_set_cpus_allowed() will not do the right thing.
5420 * And since this is boot we can forgo the serialization.
5422 set_cpus_allowed_common(idle, cpumask_of(cpu));
5425 * We're having a chicken and egg problem, even though we are
5426 * holding rq->lock, the CPU isn't yet set to this CPU so the
5427 * lockdep check in task_group() will fail.
5429 * Similar case to sched_fork(). / Alternatively we could
5430 * use task_rq_lock() here and obtain the other rq->lock.
5435 __set_task_cpu(idle, cpu);
5438 rq->curr = rq->idle = idle;
5439 idle->on_rq = TASK_ON_RQ_QUEUED;
5443 raw_spin_unlock(&rq->lock);
5444 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5446 /* Set the preempt count _outside_ the spinlocks! */
5447 init_idle_preempt_count(idle, cpu);
5450 * The idle tasks have their own, simple scheduling class:
5452 idle->sched_class = &idle_sched_class;
5453 ftrace_graph_init_idle_task(idle, cpu);
5454 vtime_init_idle(idle, cpu);
5456 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5462 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5463 const struct cpumask *trial)
5467 if (!cpumask_weight(cur))
5470 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
5475 int task_can_attach(struct task_struct *p,
5476 const struct cpumask *cs_cpus_allowed)
5481 * Kthreads which disallow setaffinity shouldn't be moved
5482 * to a new cpuset; we don't want to change their CPU
5483 * affinity and isolating such threads by their set of
5484 * allowed nodes is unnecessary. Thus, cpusets are not
5485 * applicable for such threads. This prevents checking for
5486 * success of set_cpus_allowed_ptr() on all attached tasks
5487 * before cpus_allowed may be changed.
5489 if (p->flags & PF_NO_SETAFFINITY) {
5494 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5496 ret = dl_task_can_attach(p, cs_cpus_allowed);
5502 bool sched_smp_initialized __read_mostly;
5504 #ifdef CONFIG_NUMA_BALANCING
5505 /* Migrate current task p to target_cpu */
5506 int migrate_task_to(struct task_struct *p, int target_cpu)
5508 struct migration_arg arg = { p, target_cpu };
5509 int curr_cpu = task_cpu(p);
5511 if (curr_cpu == target_cpu)
5514 if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5517 /* TODO: This is not properly updating schedstats */
5519 trace_sched_move_numa(p, curr_cpu, target_cpu);
5520 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5524 * Requeue a task on a given node and accurately track the number of NUMA
5525 * tasks on the runqueues
5527 void sched_setnuma(struct task_struct *p, int nid)
5529 bool queued, running;
5533 rq = task_rq_lock(p, &rf);
5534 queued = task_on_rq_queued(p);
5535 running = task_current(rq, p);
5538 dequeue_task(rq, p, DEQUEUE_SAVE);
5540 put_prev_task(rq, p);
5542 p->numa_preferred_nid = nid;
5545 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
5547 set_curr_task(rq, p);
5548 task_rq_unlock(rq, p, &rf);
5550 #endif /* CONFIG_NUMA_BALANCING */
5552 #ifdef CONFIG_HOTPLUG_CPU
5554 * Ensure that the idle task is using init_mm right before its CPU goes
5557 void idle_task_exit(void)
5559 struct mm_struct *mm = current->active_mm;
5561 BUG_ON(cpu_online(smp_processor_id()));
5563 if (mm != &init_mm) {
5564 switch_mm(mm, &init_mm, current);
5565 current->active_mm = &init_mm;
5566 finish_arch_post_lock_switch();
5572 * Since this CPU is going 'away' for a while, fold any nr_active delta
5573 * we might have. Assumes we're called after migrate_tasks() so that the
5574 * nr_active count is stable. We need to take the teardown thread which
5575 * is calling this into account, so we hand in adjust = 1 to the load
5578 * Also see the comment "Global load-average calculations".
5580 static void calc_load_migrate(struct rq *rq)
5582 long delta = calc_load_fold_active(rq, 1);
5584 atomic_long_add(delta, &calc_load_tasks);
5587 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5591 static const struct sched_class fake_sched_class = {
5592 .put_prev_task = put_prev_task_fake,
5595 static struct task_struct fake_task = {
5597 * Avoid pull_{rt,dl}_task()
5599 .prio = MAX_PRIO + 1,
5600 .sched_class = &fake_sched_class,
5604 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5605 * try_to_wake_up()->select_task_rq().
5607 * Called with rq->lock held even though we'er in stop_machine() and
5608 * there's no concurrency possible, we hold the required locks anyway
5609 * because of lock validation efforts.
5611 static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
5613 struct rq *rq = dead_rq;
5614 struct task_struct *next, *stop = rq->stop;
5615 struct rq_flags orf = *rf;
5619 * Fudge the rq selection such that the below task selection loop
5620 * doesn't get stuck on the currently eligible stop task.
5622 * We're currently inside stop_machine() and the rq is either stuck
5623 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5624 * either way we should never end up calling schedule() until we're
5630 * put_prev_task() and pick_next_task() sched
5631 * class method both need to have an up-to-date
5632 * value of rq->clock[_task]
5634 update_rq_clock(rq);
5638 * There's this thread running, bail when that's the only
5641 if (rq->nr_running == 1)
5645 * pick_next_task() assumes pinned rq->lock:
5647 next = pick_next_task(rq, &fake_task, rf);
5649 put_prev_task(rq, next);
5652 * Rules for changing task_struct::cpus_allowed are holding
5653 * both pi_lock and rq->lock, such that holding either
5654 * stabilizes the mask.
5656 * Drop rq->lock is not quite as disastrous as it usually is
5657 * because !cpu_active at this point, which means load-balance
5658 * will not interfere. Also, stop-machine.
5661 raw_spin_lock(&next->pi_lock);
5665 * Since we're inside stop-machine, _nothing_ should have
5666 * changed the task, WARN if weird stuff happened, because in
5667 * that case the above rq->lock drop is a fail too.
5669 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5670 raw_spin_unlock(&next->pi_lock);
5674 /* Find suitable destination for @next, with force if needed. */
5675 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5676 rq = __migrate_task(rq, rf, next, dest_cpu);
5677 if (rq != dead_rq) {
5683 raw_spin_unlock(&next->pi_lock);
5688 #endif /* CONFIG_HOTPLUG_CPU */
5690 void set_rq_online(struct rq *rq)
5693 const struct sched_class *class;
5695 cpumask_set_cpu(rq->cpu, rq->rd->online);
5698 for_each_class(class) {
5699 if (class->rq_online)
5700 class->rq_online(rq);
5705 void set_rq_offline(struct rq *rq)
5708 const struct sched_class *class;
5710 for_each_class(class) {
5711 if (class->rq_offline)
5712 class->rq_offline(rq);
5715 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5720 static void set_cpu_rq_start_time(unsigned int cpu)
5722 struct rq *rq = cpu_rq(cpu);
5724 rq->age_stamp = sched_clock_cpu(cpu);
5728 * used to mark begin/end of suspend/resume:
5730 static int num_cpus_frozen;
5733 * Update cpusets according to cpu_active mask. If cpusets are
5734 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5735 * around partition_sched_domains().
5737 * If we come here as part of a suspend/resume, don't touch cpusets because we
5738 * want to restore it back to its original state upon resume anyway.
5740 static void cpuset_cpu_active(void)
5742 if (cpuhp_tasks_frozen) {
5744 * num_cpus_frozen tracks how many CPUs are involved in suspend
5745 * resume sequence. As long as this is not the last online
5746 * operation in the resume sequence, just build a single sched
5747 * domain, ignoring cpusets.
5749 partition_sched_domains(1, NULL, NULL);
5750 if (--num_cpus_frozen)
5753 * This is the last CPU online operation. So fall through and
5754 * restore the original sched domains by considering the
5755 * cpuset configurations.
5757 cpuset_force_rebuild();
5759 cpuset_update_active_cpus();
5762 static int cpuset_cpu_inactive(unsigned int cpu)
5764 if (!cpuhp_tasks_frozen) {
5765 if (dl_cpu_busy(cpu))
5767 cpuset_update_active_cpus();
5770 partition_sched_domains(1, NULL, NULL);
5775 int sched_cpu_activate(unsigned int cpu)
5777 struct rq *rq = cpu_rq(cpu);
5780 set_cpu_active(cpu, true);
5782 if (sched_smp_initialized) {
5783 sched_domains_numa_masks_set(cpu);
5784 cpuset_cpu_active();
5788 * Put the rq online, if not already. This happens:
5790 * 1) In the early boot process, because we build the real domains
5791 * after all CPUs have been brought up.
5793 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5796 rq_lock_irqsave(rq, &rf);
5798 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5801 rq_unlock_irqrestore(rq, &rf);
5803 update_max_interval();
5808 int sched_cpu_deactivate(unsigned int cpu)
5812 set_cpu_active(cpu, false);
5814 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5815 * users of this state to go away such that all new such users will
5818 * Do sync before park smpboot threads to take care the rcu boost case.
5820 synchronize_rcu_mult(call_rcu, call_rcu_sched);
5822 if (!sched_smp_initialized)
5825 ret = cpuset_cpu_inactive(cpu);
5827 set_cpu_active(cpu, true);
5830 sched_domains_numa_masks_clear(cpu);
5834 static void sched_rq_cpu_starting(unsigned int cpu)
5836 struct rq *rq = cpu_rq(cpu);
5838 rq->calc_load_update = calc_load_update;
5839 update_max_interval();
5842 int sched_cpu_starting(unsigned int cpu)
5844 set_cpu_rq_start_time(cpu);
5845 sched_rq_cpu_starting(cpu);
5846 sched_tick_start(cpu);
5850 #ifdef CONFIG_HOTPLUG_CPU
5851 int sched_cpu_dying(unsigned int cpu)
5853 struct rq *rq = cpu_rq(cpu);
5856 /* Handle pending wakeups and then migrate everything off */
5857 sched_ttwu_pending();
5858 sched_tick_stop(cpu);
5860 rq_lock_irqsave(rq, &rf);
5862 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5865 migrate_tasks(rq, &rf);
5866 BUG_ON(rq->nr_running != 1);
5867 rq_unlock_irqrestore(rq, &rf);
5869 calc_load_migrate(rq);
5870 update_max_interval();
5871 nohz_balance_exit_idle(rq);
5877 #ifdef CONFIG_SCHED_SMT
5878 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5880 static void sched_init_smt(void)
5883 * We've enumerated all CPUs and will assume that if any CPU
5884 * has SMT siblings, CPU0 will too.
5886 if (cpumask_weight(cpu_smt_mask(0)) > 1)
5887 static_branch_enable(&sched_smt_present);
5890 static inline void sched_init_smt(void) { }
5893 void __init sched_init_smp(void)
5898 * There's no userspace yet to cause hotplug operations; hence all the
5899 * CPU masks are stable and all blatant races in the below code cannot
5902 mutex_lock(&sched_domains_mutex);
5903 sched_init_domains(cpu_active_mask);
5904 mutex_unlock(&sched_domains_mutex);
5906 /* Move init over to a non-isolated CPU */
5907 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_FLAG_DOMAIN)) < 0)
5909 sched_init_granularity();
5911 init_sched_rt_class();
5912 init_sched_dl_class();
5916 sched_smp_initialized = true;
5919 static int __init migration_init(void)
5921 sched_rq_cpu_starting(smp_processor_id());
5924 early_initcall(migration_init);
5927 void __init sched_init_smp(void)
5929 sched_init_granularity();
5931 #endif /* CONFIG_SMP */
5933 int in_sched_functions(unsigned long addr)
5935 return in_lock_functions(addr) ||
5936 (addr >= (unsigned long)__sched_text_start
5937 && addr < (unsigned long)__sched_text_end);
5940 #ifdef CONFIG_CGROUP_SCHED
5942 * Default task group.
5943 * Every task in system belongs to this group at bootup.
5945 struct task_group root_task_group;
5946 LIST_HEAD(task_groups);
5948 /* Cacheline aligned slab cache for task_group */
5949 static struct kmem_cache *task_group_cache __read_mostly;
5952 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5953 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5955 void __init sched_init(void)
5958 unsigned long alloc_size = 0, ptr;
5963 #ifdef CONFIG_FAIR_GROUP_SCHED
5964 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5966 #ifdef CONFIG_RT_GROUP_SCHED
5967 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5970 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5972 #ifdef CONFIG_FAIR_GROUP_SCHED
5973 root_task_group.se = (struct sched_entity **)ptr;
5974 ptr += nr_cpu_ids * sizeof(void **);
5976 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5977 ptr += nr_cpu_ids * sizeof(void **);
5979 #endif /* CONFIG_FAIR_GROUP_SCHED */
5980 #ifdef CONFIG_RT_GROUP_SCHED
5981 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5982 ptr += nr_cpu_ids * sizeof(void **);
5984 root_task_group.rt_rq = (struct rt_rq **)ptr;
5985 ptr += nr_cpu_ids * sizeof(void **);
5987 #endif /* CONFIG_RT_GROUP_SCHED */
5989 #ifdef CONFIG_CPUMASK_OFFSTACK
5990 for_each_possible_cpu(i) {
5991 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
5992 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5993 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
5994 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
5996 #endif /* CONFIG_CPUMASK_OFFSTACK */
5998 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
5999 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6002 init_defrootdomain();
6005 #ifdef CONFIG_RT_GROUP_SCHED
6006 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6007 global_rt_period(), global_rt_runtime());
6008 #endif /* CONFIG_RT_GROUP_SCHED */
6010 #ifdef CONFIG_CGROUP_SCHED
6011 task_group_cache = KMEM_CACHE(task_group, 0);
6013 list_add(&root_task_group.list, &task_groups);
6014 INIT_LIST_HEAD(&root_task_group.children);
6015 INIT_LIST_HEAD(&root_task_group.siblings);
6016 autogroup_init(&init_task);
6017 #endif /* CONFIG_CGROUP_SCHED */
6019 for_each_possible_cpu(i) {
6023 raw_spin_lock_init(&rq->lock);
6025 rq->calc_load_active = 0;
6026 rq->calc_load_update = jiffies + LOAD_FREQ;
6027 init_cfs_rq(&rq->cfs);
6028 init_rt_rq(&rq->rt);
6029 init_dl_rq(&rq->dl);
6030 #ifdef CONFIG_FAIR_GROUP_SCHED
6031 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6032 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6033 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6035 * How much CPU bandwidth does root_task_group get?
6037 * In case of task-groups formed thr' the cgroup filesystem, it
6038 * gets 100% of the CPU resources in the system. This overall
6039 * system CPU resource is divided among the tasks of
6040 * root_task_group and its child task-groups in a fair manner,
6041 * based on each entity's (task or task-group's) weight
6042 * (se->load.weight).
6044 * In other words, if root_task_group has 10 tasks of weight
6045 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6046 * then A0's share of the CPU resource is:
6048 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6050 * We achieve this by letting root_task_group's tasks sit
6051 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6053 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6054 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6055 #endif /* CONFIG_FAIR_GROUP_SCHED */
6057 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6058 #ifdef CONFIG_RT_GROUP_SCHED
6059 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6062 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6063 rq->cpu_load[j] = 0;
6068 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6069 rq->balance_callback = NULL;
6070 rq->active_balance = 0;
6071 rq->next_balance = jiffies;
6076 rq->avg_idle = 2*sysctl_sched_migration_cost;
6077 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6079 INIT_LIST_HEAD(&rq->cfs_tasks);
6081 rq_attach_root(rq, &def_root_domain);
6082 #ifdef CONFIG_NO_HZ_COMMON
6083 rq->last_load_update_tick = jiffies;
6084 rq->last_blocked_load_update_tick = jiffies;
6085 atomic_set(&rq->nohz_flags, 0);
6087 #endif /* CONFIG_SMP */
6089 atomic_set(&rq->nr_iowait, 0);
6092 set_load_weight(&init_task, false);
6095 * The boot idle thread does lazy MMU switching as well:
6098 enter_lazy_tlb(&init_mm, current);
6101 * Make us the idle thread. Technically, schedule() should not be
6102 * called from this thread, however somewhere below it might be,
6103 * but because we are the idle thread, we just pick up running again
6104 * when this runqueue becomes "idle".
6106 init_idle(current, smp_processor_id());
6108 calc_load_update = jiffies + LOAD_FREQ;
6111 idle_thread_set_boot_cpu();
6112 set_cpu_rq_start_time(smp_processor_id());
6114 init_sched_fair_class();
6118 scheduler_running = 1;
6121 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6122 static inline int preempt_count_equals(int preempt_offset)
6124 int nested = preempt_count() + rcu_preempt_depth();
6126 return (nested == preempt_offset);
6129 void __might_sleep(const char *file, int line, int preempt_offset)
6132 * Blocking primitives will set (and therefore destroy) current->state,
6133 * since we will exit with TASK_RUNNING make sure we enter with it,
6134 * otherwise we will destroy state.
6136 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6137 "do not call blocking ops when !TASK_RUNNING; "
6138 "state=%lx set at [<%p>] %pS\n",
6140 (void *)current->task_state_change,
6141 (void *)current->task_state_change);
6143 ___might_sleep(file, line, preempt_offset);
6145 EXPORT_SYMBOL(__might_sleep);
6147 void ___might_sleep(const char *file, int line, int preempt_offset)
6149 /* Ratelimiting timestamp: */
6150 static unsigned long prev_jiffy;
6152 unsigned long preempt_disable_ip;
6154 /* WARN_ON_ONCE() by default, no rate limit required: */
6157 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6158 !is_idle_task(current)) ||
6159 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
6163 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6165 prev_jiffy = jiffies;
6167 /* Save this before calling printk(), since that will clobber it: */
6168 preempt_disable_ip = get_preempt_disable_ip(current);
6171 "BUG: sleeping function called from invalid context at %s:%d\n",
6174 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6175 in_atomic(), irqs_disabled(),
6176 current->pid, current->comm);
6178 if (task_stack_end_corrupted(current))
6179 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6181 debug_show_held_locks(current);
6182 if (irqs_disabled())
6183 print_irqtrace_events(current);
6184 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6185 && !preempt_count_equals(preempt_offset)) {
6186 pr_err("Preemption disabled at:");
6187 print_ip_sym(preempt_disable_ip);
6191 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6193 EXPORT_SYMBOL(___might_sleep);
6196 #ifdef CONFIG_MAGIC_SYSRQ
6197 void normalize_rt_tasks(void)
6199 struct task_struct *g, *p;
6200 struct sched_attr attr = {
6201 .sched_policy = SCHED_NORMAL,
6204 read_lock(&tasklist_lock);
6205 for_each_process_thread(g, p) {
6207 * Only normalize user tasks:
6209 if (p->flags & PF_KTHREAD)
6212 p->se.exec_start = 0;
6213 schedstat_set(p->se.statistics.wait_start, 0);
6214 schedstat_set(p->se.statistics.sleep_start, 0);
6215 schedstat_set(p->se.statistics.block_start, 0);
6217 if (!dl_task(p) && !rt_task(p)) {
6219 * Renice negative nice level userspace
6222 if (task_nice(p) < 0)
6223 set_user_nice(p, 0);
6227 __sched_setscheduler(p, &attr, false, false);
6229 read_unlock(&tasklist_lock);
6232 #endif /* CONFIG_MAGIC_SYSRQ */
6234 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6236 * These functions are only useful for the IA64 MCA handling, or kdb.
6238 * They can only be called when the whole system has been
6239 * stopped - every CPU needs to be quiescent, and no scheduling
6240 * activity can take place. Using them for anything else would
6241 * be a serious bug, and as a result, they aren't even visible
6242 * under any other configuration.
6246 * curr_task - return the current task for a given CPU.
6247 * @cpu: the processor in question.
6249 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6251 * Return: The current task for @cpu.
6253 struct task_struct *curr_task(int cpu)
6255 return cpu_curr(cpu);
6258 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6262 * set_curr_task - set the current task for a given CPU.
6263 * @cpu: the processor in question.
6264 * @p: the task pointer to set.
6266 * Description: This function must only be used when non-maskable interrupts
6267 * are serviced on a separate stack. It allows the architecture to switch the
6268 * notion of the current task on a CPU in a non-blocking manner. This function
6269 * must be called with all CPU's synchronized, and interrupts disabled, the
6270 * and caller must save the original value of the current task (see
6271 * curr_task() above) and restore that value before reenabling interrupts and
6272 * re-starting the system.
6274 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6276 void ia64_set_curr_task(int cpu, struct task_struct *p)
6283 #ifdef CONFIG_CGROUP_SCHED
6284 /* task_group_lock serializes the addition/removal of task groups */
6285 static DEFINE_SPINLOCK(task_group_lock);
6287 static void sched_free_group(struct task_group *tg)
6289 free_fair_sched_group(tg);
6290 free_rt_sched_group(tg);
6292 kmem_cache_free(task_group_cache, tg);
6295 /* allocate runqueue etc for a new task group */
6296 struct task_group *sched_create_group(struct task_group *parent)
6298 struct task_group *tg;
6300 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6302 return ERR_PTR(-ENOMEM);
6304 if (!alloc_fair_sched_group(tg, parent))
6307 if (!alloc_rt_sched_group(tg, parent))
6313 sched_free_group(tg);
6314 return ERR_PTR(-ENOMEM);
6317 void sched_online_group(struct task_group *tg, struct task_group *parent)
6319 unsigned long flags;
6321 spin_lock_irqsave(&task_group_lock, flags);
6322 list_add_rcu(&tg->list, &task_groups);
6324 /* Root should already exist: */
6327 tg->parent = parent;
6328 INIT_LIST_HEAD(&tg->children);
6329 list_add_rcu(&tg->siblings, &parent->children);
6330 spin_unlock_irqrestore(&task_group_lock, flags);
6332 online_fair_sched_group(tg);
6335 /* rcu callback to free various structures associated with a task group */
6336 static void sched_free_group_rcu(struct rcu_head *rhp)
6338 /* Now it should be safe to free those cfs_rqs: */
6339 sched_free_group(container_of(rhp, struct task_group, rcu));
6342 void sched_destroy_group(struct task_group *tg)
6344 /* Wait for possible concurrent references to cfs_rqs complete: */
6345 call_rcu(&tg->rcu, sched_free_group_rcu);
6348 void sched_offline_group(struct task_group *tg)
6350 unsigned long flags;
6352 /* End participation in shares distribution: */
6353 unregister_fair_sched_group(tg);
6355 spin_lock_irqsave(&task_group_lock, flags);
6356 list_del_rcu(&tg->list);
6357 list_del_rcu(&tg->siblings);
6358 spin_unlock_irqrestore(&task_group_lock, flags);
6361 static void sched_change_group(struct task_struct *tsk, int type)
6363 struct task_group *tg;
6366 * All callers are synchronized by task_rq_lock(); we do not use RCU
6367 * which is pointless here. Thus, we pass "true" to task_css_check()
6368 * to prevent lockdep warnings.
6370 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6371 struct task_group, css);
6372 tg = autogroup_task_group(tsk, tg);
6373 tsk->sched_task_group = tg;
6375 #ifdef CONFIG_FAIR_GROUP_SCHED
6376 if (tsk->sched_class->task_change_group)
6377 tsk->sched_class->task_change_group(tsk, type);
6380 set_task_rq(tsk, task_cpu(tsk));
6384 * Change task's runqueue when it moves between groups.
6386 * The caller of this function should have put the task in its new group by
6387 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6390 void sched_move_task(struct task_struct *tsk)
6392 int queued, running, queue_flags =
6393 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
6397 rq = task_rq_lock(tsk, &rf);
6398 update_rq_clock(rq);
6400 running = task_current(rq, tsk);
6401 queued = task_on_rq_queued(tsk);
6404 dequeue_task(rq, tsk, queue_flags);
6406 put_prev_task(rq, tsk);
6408 sched_change_group(tsk, TASK_MOVE_GROUP);
6411 enqueue_task(rq, tsk, queue_flags);
6413 set_curr_task(rq, tsk);
6415 task_rq_unlock(rq, tsk, &rf);
6418 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6420 return css ? container_of(css, struct task_group, css) : NULL;
6423 static struct cgroup_subsys_state *
6424 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6426 struct task_group *parent = css_tg(parent_css);
6427 struct task_group *tg;
6430 /* This is early initialization for the top cgroup */
6431 return &root_task_group.css;
6434 tg = sched_create_group(parent);
6436 return ERR_PTR(-ENOMEM);
6441 /* Expose task group only after completing cgroup initialization */
6442 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6444 struct task_group *tg = css_tg(css);
6445 struct task_group *parent = css_tg(css->parent);
6448 sched_online_group(tg, parent);
6452 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6454 struct task_group *tg = css_tg(css);
6456 sched_offline_group(tg);
6459 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6461 struct task_group *tg = css_tg(css);
6464 * Relies on the RCU grace period between css_released() and this.
6466 sched_free_group(tg);
6470 * This is called before wake_up_new_task(), therefore we really only
6471 * have to set its group bits, all the other stuff does not apply.
6473 static void cpu_cgroup_fork(struct task_struct *task)
6478 rq = task_rq_lock(task, &rf);
6480 update_rq_clock(rq);
6481 sched_change_group(task, TASK_SET_GROUP);
6483 task_rq_unlock(rq, task, &rf);
6486 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6488 struct task_struct *task;
6489 struct cgroup_subsys_state *css;
6492 cgroup_taskset_for_each(task, css, tset) {
6493 #ifdef CONFIG_RT_GROUP_SCHED
6494 if (!sched_rt_can_attach(css_tg(css), task))
6497 /* We don't support RT-tasks being in separate groups */
6498 if (task->sched_class != &fair_sched_class)
6502 * Serialize against wake_up_new_task() such that if its
6503 * running, we're sure to observe its full state.
6505 raw_spin_lock_irq(&task->pi_lock);
6507 * Avoid calling sched_move_task() before wake_up_new_task()
6508 * has happened. This would lead to problems with PELT, due to
6509 * move wanting to detach+attach while we're not attached yet.
6511 if (task->state == TASK_NEW)
6513 raw_spin_unlock_irq(&task->pi_lock);
6521 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6523 struct task_struct *task;
6524 struct cgroup_subsys_state *css;
6526 cgroup_taskset_for_each(task, css, tset)
6527 sched_move_task(task);
6530 #ifdef CONFIG_FAIR_GROUP_SCHED
6531 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6532 struct cftype *cftype, u64 shareval)
6534 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6537 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6540 struct task_group *tg = css_tg(css);
6542 return (u64) scale_load_down(tg->shares);
6545 #ifdef CONFIG_CFS_BANDWIDTH
6546 static DEFINE_MUTEX(cfs_constraints_mutex);
6548 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6549 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6551 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6553 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6555 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6556 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6558 if (tg == &root_task_group)
6562 * Ensure we have at some amount of bandwidth every period. This is
6563 * to prevent reaching a state of large arrears when throttled via
6564 * entity_tick() resulting in prolonged exit starvation.
6566 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6570 * Likewise, bound things on the otherside by preventing insane quota
6571 * periods. This also allows us to normalize in computing quota
6574 if (period > max_cfs_quota_period)
6578 * Prevent race between setting of cfs_rq->runtime_enabled and
6579 * unthrottle_offline_cfs_rqs().
6582 mutex_lock(&cfs_constraints_mutex);
6583 ret = __cfs_schedulable(tg, period, quota);
6587 runtime_enabled = quota != RUNTIME_INF;
6588 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6590 * If we need to toggle cfs_bandwidth_used, off->on must occur
6591 * before making related changes, and on->off must occur afterwards
6593 if (runtime_enabled && !runtime_was_enabled)
6594 cfs_bandwidth_usage_inc();
6595 raw_spin_lock_irq(&cfs_b->lock);
6596 cfs_b->period = ns_to_ktime(period);
6597 cfs_b->quota = quota;
6599 __refill_cfs_bandwidth_runtime(cfs_b);
6601 /* Restart the period timer (if active) to handle new period expiry: */
6602 if (runtime_enabled)
6603 start_cfs_bandwidth(cfs_b);
6605 raw_spin_unlock_irq(&cfs_b->lock);
6607 for_each_online_cpu(i) {
6608 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
6609 struct rq *rq = cfs_rq->rq;
6612 rq_lock_irq(rq, &rf);
6613 cfs_rq->runtime_enabled = runtime_enabled;
6614 cfs_rq->runtime_remaining = 0;
6616 if (cfs_rq->throttled)
6617 unthrottle_cfs_rq(cfs_rq);
6618 rq_unlock_irq(rq, &rf);
6620 if (runtime_was_enabled && !runtime_enabled)
6621 cfs_bandwidth_usage_dec();
6623 mutex_unlock(&cfs_constraints_mutex);
6629 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
6633 period = ktime_to_ns(tg->cfs_bandwidth.period);
6634 if (cfs_quota_us < 0)
6635 quota = RUNTIME_INF;
6637 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
6639 return tg_set_cfs_bandwidth(tg, period, quota);
6642 long tg_get_cfs_quota(struct task_group *tg)
6646 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
6649 quota_us = tg->cfs_bandwidth.quota;
6650 do_div(quota_us, NSEC_PER_USEC);
6655 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
6659 period = (u64)cfs_period_us * NSEC_PER_USEC;
6660 quota = tg->cfs_bandwidth.quota;
6662 return tg_set_cfs_bandwidth(tg, period, quota);
6665 long tg_get_cfs_period(struct task_group *tg)
6669 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
6670 do_div(cfs_period_us, NSEC_PER_USEC);
6672 return cfs_period_us;
6675 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
6678 return tg_get_cfs_quota(css_tg(css));
6681 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
6682 struct cftype *cftype, s64 cfs_quota_us)
6684 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
6687 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
6690 return tg_get_cfs_period(css_tg(css));
6693 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
6694 struct cftype *cftype, u64 cfs_period_us)
6696 return tg_set_cfs_period(css_tg(css), cfs_period_us);
6699 struct cfs_schedulable_data {
6700 struct task_group *tg;
6705 * normalize group quota/period to be quota/max_period
6706 * note: units are usecs
6708 static u64 normalize_cfs_quota(struct task_group *tg,
6709 struct cfs_schedulable_data *d)
6717 period = tg_get_cfs_period(tg);
6718 quota = tg_get_cfs_quota(tg);
6721 /* note: these should typically be equivalent */
6722 if (quota == RUNTIME_INF || quota == -1)
6725 return to_ratio(period, quota);
6728 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
6730 struct cfs_schedulable_data *d = data;
6731 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6732 s64 quota = 0, parent_quota = -1;
6735 quota = RUNTIME_INF;
6737 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
6739 quota = normalize_cfs_quota(tg, d);
6740 parent_quota = parent_b->hierarchical_quota;
6743 * Ensure max(child_quota) <= parent_quota. On cgroup2,
6744 * always take the min. On cgroup1, only inherit when no
6747 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
6748 quota = min(quota, parent_quota);
6750 if (quota == RUNTIME_INF)
6751 quota = parent_quota;
6752 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
6756 cfs_b->hierarchical_quota = quota;
6761 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
6764 struct cfs_schedulable_data data = {
6770 if (quota != RUNTIME_INF) {
6771 do_div(data.period, NSEC_PER_USEC);
6772 do_div(data.quota, NSEC_PER_USEC);
6776 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
6782 static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
6784 struct task_group *tg = css_tg(seq_css(sf));
6785 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6787 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
6788 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
6789 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
6793 #endif /* CONFIG_CFS_BANDWIDTH */
6794 #endif /* CONFIG_FAIR_GROUP_SCHED */
6796 #ifdef CONFIG_RT_GROUP_SCHED
6797 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
6798 struct cftype *cft, s64 val)
6800 return sched_group_set_rt_runtime(css_tg(css), val);
6803 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
6806 return sched_group_rt_runtime(css_tg(css));
6809 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
6810 struct cftype *cftype, u64 rt_period_us)
6812 return sched_group_set_rt_period(css_tg(css), rt_period_us);
6815 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
6818 return sched_group_rt_period(css_tg(css));
6820 #endif /* CONFIG_RT_GROUP_SCHED */
6822 static struct cftype cpu_legacy_files[] = {
6823 #ifdef CONFIG_FAIR_GROUP_SCHED
6826 .read_u64 = cpu_shares_read_u64,
6827 .write_u64 = cpu_shares_write_u64,
6830 #ifdef CONFIG_CFS_BANDWIDTH
6832 .name = "cfs_quota_us",
6833 .read_s64 = cpu_cfs_quota_read_s64,
6834 .write_s64 = cpu_cfs_quota_write_s64,
6837 .name = "cfs_period_us",
6838 .read_u64 = cpu_cfs_period_read_u64,
6839 .write_u64 = cpu_cfs_period_write_u64,
6843 .seq_show = cpu_cfs_stat_show,
6846 #ifdef CONFIG_RT_GROUP_SCHED
6848 .name = "rt_runtime_us",
6849 .read_s64 = cpu_rt_runtime_read,
6850 .write_s64 = cpu_rt_runtime_write,
6853 .name = "rt_period_us",
6854 .read_u64 = cpu_rt_period_read_uint,
6855 .write_u64 = cpu_rt_period_write_uint,
6861 static int cpu_extra_stat_show(struct seq_file *sf,
6862 struct cgroup_subsys_state *css)
6864 #ifdef CONFIG_CFS_BANDWIDTH
6866 struct task_group *tg = css_tg(css);
6867 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6870 throttled_usec = cfs_b->throttled_time;
6871 do_div(throttled_usec, NSEC_PER_USEC);
6873 seq_printf(sf, "nr_periods %d\n"
6875 "throttled_usec %llu\n",
6876 cfs_b->nr_periods, cfs_b->nr_throttled,
6883 #ifdef CONFIG_FAIR_GROUP_SCHED
6884 static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
6887 struct task_group *tg = css_tg(css);
6888 u64 weight = scale_load_down(tg->shares);
6890 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
6893 static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
6894 struct cftype *cft, u64 weight)
6897 * cgroup weight knobs should use the common MIN, DFL and MAX
6898 * values which are 1, 100 and 10000 respectively. While it loses
6899 * a bit of range on both ends, it maps pretty well onto the shares
6900 * value used by scheduler and the round-trip conversions preserve
6901 * the original value over the entire range.
6903 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
6906 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
6908 return sched_group_set_shares(css_tg(css), scale_load(weight));
6911 static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
6914 unsigned long weight = scale_load_down(css_tg(css)->shares);
6915 int last_delta = INT_MAX;
6918 /* find the closest nice value to the current weight */
6919 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
6920 delta = abs(sched_prio_to_weight[prio] - weight);
6921 if (delta >= last_delta)
6926 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
6929 static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
6930 struct cftype *cft, s64 nice)
6932 unsigned long weight;
6935 if (nice < MIN_NICE || nice > MAX_NICE)
6938 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
6939 idx = array_index_nospec(idx, 40);
6940 weight = sched_prio_to_weight[idx];
6942 return sched_group_set_shares(css_tg(css), scale_load(weight));
6946 static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
6947 long period, long quota)
6950 seq_puts(sf, "max");
6952 seq_printf(sf, "%ld", quota);
6954 seq_printf(sf, " %ld\n", period);
6957 /* caller should put the current value in *@periodp before calling */
6958 static int __maybe_unused cpu_period_quota_parse(char *buf,
6959 u64 *periodp, u64 *quotap)
6961 char tok[21]; /* U64_MAX */
6963 if (!sscanf(buf, "%s %llu", tok, periodp))
6966 *periodp *= NSEC_PER_USEC;
6968 if (sscanf(tok, "%llu", quotap))
6969 *quotap *= NSEC_PER_USEC;
6970 else if (!strcmp(tok, "max"))
6971 *quotap = RUNTIME_INF;
6978 #ifdef CONFIG_CFS_BANDWIDTH
6979 static int cpu_max_show(struct seq_file *sf, void *v)
6981 struct task_group *tg = css_tg(seq_css(sf));
6983 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
6987 static ssize_t cpu_max_write(struct kernfs_open_file *of,
6988 char *buf, size_t nbytes, loff_t off)
6990 struct task_group *tg = css_tg(of_css(of));
6991 u64 period = tg_get_cfs_period(tg);
6995 ret = cpu_period_quota_parse(buf, &period, "a);
6997 ret = tg_set_cfs_bandwidth(tg, period, quota);
6998 return ret ?: nbytes;
7002 static struct cftype cpu_files[] = {
7003 #ifdef CONFIG_FAIR_GROUP_SCHED
7006 .flags = CFTYPE_NOT_ON_ROOT,
7007 .read_u64 = cpu_weight_read_u64,
7008 .write_u64 = cpu_weight_write_u64,
7011 .name = "weight.nice",
7012 .flags = CFTYPE_NOT_ON_ROOT,
7013 .read_s64 = cpu_weight_nice_read_s64,
7014 .write_s64 = cpu_weight_nice_write_s64,
7017 #ifdef CONFIG_CFS_BANDWIDTH
7020 .flags = CFTYPE_NOT_ON_ROOT,
7021 .seq_show = cpu_max_show,
7022 .write = cpu_max_write,
7028 struct cgroup_subsys cpu_cgrp_subsys = {
7029 .css_alloc = cpu_cgroup_css_alloc,
7030 .css_online = cpu_cgroup_css_online,
7031 .css_released = cpu_cgroup_css_released,
7032 .css_free = cpu_cgroup_css_free,
7033 .css_extra_stat_show = cpu_extra_stat_show,
7034 .fork = cpu_cgroup_fork,
7035 .can_attach = cpu_cgroup_can_attach,
7036 .attach = cpu_cgroup_attach,
7037 .legacy_cftypes = cpu_legacy_files,
7038 .dfl_cftypes = cpu_files,
7043 #endif /* CONFIG_CGROUP_SCHED */
7045 void dump_cpu_task(int cpu)
7047 pr_info("Task dump for CPU %d:\n", cpu);
7048 sched_show_task(cpu_curr(cpu));
7052 * Nice levels are multiplicative, with a gentle 10% change for every
7053 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7054 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7055 * that remained on nice 0.
7057 * The "10% effect" is relative and cumulative: from _any_ nice level,
7058 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7059 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7060 * If a task goes up by ~10% and another task goes down by ~10% then
7061 * the relative distance between them is ~25%.)
7063 const int sched_prio_to_weight[40] = {
7064 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7065 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7066 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7067 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7068 /* 0 */ 1024, 820, 655, 526, 423,
7069 /* 5 */ 335, 272, 215, 172, 137,
7070 /* 10 */ 110, 87, 70, 56, 45,
7071 /* 15 */ 36, 29, 23, 18, 15,
7075 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7077 * In cases where the weight does not change often, we can use the
7078 * precalculated inverse to speed up arithmetics by turning divisions
7079 * into multiplications:
7081 const u32 sched_prio_to_wmult[40] = {
7082 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7083 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7084 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7085 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7086 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7087 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7088 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7089 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
7092 #undef CREATE_TRACE_POINTS