4 * Core kernel scheduler code and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 #include <linux/sched.h>
9 #include <linux/cpuset.h>
10 #include <linux/delayacct.h>
11 #include <linux/init_task.h>
12 #include <linux/context_tracking.h>
13 #include <linux/rcupdate_wait.h>
15 #include <linux/blkdev.h>
16 #include <linux/kprobes.h>
17 #include <linux/mmu_context.h>
18 #include <linux/module.h>
19 #include <linux/nmi.h>
20 #include <linux/prefetch.h>
21 #include <linux/profile.h>
22 #include <linux/security.h>
23 #include <linux/syscalls.h>
25 #include <asm/switch_to.h>
27 #ifdef CONFIG_PARAVIRT
28 #include <asm/paravirt.h>
32 #include "../workqueue_internal.h"
33 #include "../smpboot.h"
35 #define CREATE_TRACE_POINTS
36 #include <trace/events/sched.h>
38 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
41 * Debugging: various feature bits
44 #define SCHED_FEAT(name, enabled) \
45 (1UL << __SCHED_FEAT_##name) * enabled |
47 const_debug unsigned int sysctl_sched_features =
54 * Number of tasks to iterate in a single balance run.
55 * Limited because this is done with IRQs disabled.
57 const_debug unsigned int sysctl_sched_nr_migrate = 32;
60 * period over which we average the RT time consumption, measured
65 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
68 * period over which we measure -rt task CPU usage in us.
71 unsigned int sysctl_sched_rt_period = 1000000;
73 __read_mostly int scheduler_running;
76 * part of the period that we allow rt tasks to run in us.
79 int sysctl_sched_rt_runtime = 950000;
81 /* CPUs with isolated domains */
82 cpumask_var_t cpu_isolated_map;
85 * this_rq_lock - lock this runqueue and disable interrupts.
87 static struct rq *this_rq_lock(void)
94 raw_spin_lock(&rq->lock);
100 * __task_rq_lock - lock the rq @p resides on.
102 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
107 lockdep_assert_held(&p->pi_lock);
111 raw_spin_lock(&rq->lock);
112 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
116 raw_spin_unlock(&rq->lock);
118 while (unlikely(task_on_rq_migrating(p)))
124 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
126 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
127 __acquires(p->pi_lock)
133 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
135 raw_spin_lock(&rq->lock);
137 * move_queued_task() task_rq_lock()
140 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
141 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
142 * [S] ->cpu = new_cpu [L] task_rq()
146 * If we observe the old cpu in task_rq_lock, the acquire of
147 * the old rq->lock will fully serialize against the stores.
149 * If we observe the new CPU in task_rq_lock, the acquire will
150 * pair with the WMB to ensure we must then also see migrating.
152 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
156 raw_spin_unlock(&rq->lock);
157 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
159 while (unlikely(task_on_rq_migrating(p)))
165 * RQ-clock updating methods:
168 static void update_rq_clock_task(struct rq *rq, s64 delta)
171 * In theory, the compile should just see 0 here, and optimize out the call
172 * to sched_rt_avg_update. But I don't trust it...
174 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
175 s64 steal = 0, irq_delta = 0;
177 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
178 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
181 * Since irq_time is only updated on {soft,}irq_exit, we might run into
182 * this case when a previous update_rq_clock() happened inside a
185 * When this happens, we stop ->clock_task and only update the
186 * prev_irq_time stamp to account for the part that fit, so that a next
187 * update will consume the rest. This ensures ->clock_task is
190 * It does however cause some slight miss-attribution of {soft,}irq
191 * time, a more accurate solution would be to update the irq_time using
192 * the current rq->clock timestamp, except that would require using
195 if (irq_delta > delta)
198 rq->prev_irq_time += irq_delta;
201 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
202 if (static_key_false((¶virt_steal_rq_enabled))) {
203 steal = paravirt_steal_clock(cpu_of(rq));
204 steal -= rq->prev_steal_time_rq;
206 if (unlikely(steal > delta))
209 rq->prev_steal_time_rq += steal;
214 rq->clock_task += delta;
216 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
217 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
218 sched_rt_avg_update(rq, irq_delta + steal);
222 void update_rq_clock(struct rq *rq)
226 lockdep_assert_held(&rq->lock);
228 if (rq->clock_update_flags & RQCF_ACT_SKIP)
231 #ifdef CONFIG_SCHED_DEBUG
232 rq->clock_update_flags |= RQCF_UPDATED;
234 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
238 update_rq_clock_task(rq, delta);
242 #ifdef CONFIG_SCHED_HRTICK
244 * Use HR-timers to deliver accurate preemption points.
247 static void hrtick_clear(struct rq *rq)
249 if (hrtimer_active(&rq->hrtick_timer))
250 hrtimer_cancel(&rq->hrtick_timer);
254 * High-resolution timer tick.
255 * Runs from hardirq context with interrupts disabled.
257 static enum hrtimer_restart hrtick(struct hrtimer *timer)
259 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
261 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
263 raw_spin_lock(&rq->lock);
265 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
266 raw_spin_unlock(&rq->lock);
268 return HRTIMER_NORESTART;
273 static void __hrtick_restart(struct rq *rq)
275 struct hrtimer *timer = &rq->hrtick_timer;
277 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
281 * called from hardirq (IPI) context
283 static void __hrtick_start(void *arg)
287 raw_spin_lock(&rq->lock);
288 __hrtick_restart(rq);
289 rq->hrtick_csd_pending = 0;
290 raw_spin_unlock(&rq->lock);
294 * Called to set the hrtick timer state.
296 * called with rq->lock held and irqs disabled
298 void hrtick_start(struct rq *rq, u64 delay)
300 struct hrtimer *timer = &rq->hrtick_timer;
305 * Don't schedule slices shorter than 10000ns, that just
306 * doesn't make sense and can cause timer DoS.
308 delta = max_t(s64, delay, 10000LL);
309 time = ktime_add_ns(timer->base->get_time(), delta);
311 hrtimer_set_expires(timer, time);
313 if (rq == this_rq()) {
314 __hrtick_restart(rq);
315 } else if (!rq->hrtick_csd_pending) {
316 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
317 rq->hrtick_csd_pending = 1;
323 * Called to set the hrtick timer state.
325 * called with rq->lock held and irqs disabled
327 void hrtick_start(struct rq *rq, u64 delay)
330 * Don't schedule slices shorter than 10000ns, that just
331 * doesn't make sense. Rely on vruntime for fairness.
333 delay = max_t(u64, delay, 10000LL);
334 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
335 HRTIMER_MODE_REL_PINNED);
337 #endif /* CONFIG_SMP */
339 static void init_rq_hrtick(struct rq *rq)
342 rq->hrtick_csd_pending = 0;
344 rq->hrtick_csd.flags = 0;
345 rq->hrtick_csd.func = __hrtick_start;
346 rq->hrtick_csd.info = rq;
349 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
350 rq->hrtick_timer.function = hrtick;
352 #else /* CONFIG_SCHED_HRTICK */
353 static inline void hrtick_clear(struct rq *rq)
357 static inline void init_rq_hrtick(struct rq *rq)
360 #endif /* CONFIG_SCHED_HRTICK */
363 * cmpxchg based fetch_or, macro so it works for different integer types
365 #define fetch_or(ptr, mask) \
367 typeof(ptr) _ptr = (ptr); \
368 typeof(mask) _mask = (mask); \
369 typeof(*_ptr) _old, _val = *_ptr; \
372 _old = cmpxchg(_ptr, _val, _val | _mask); \
380 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
382 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
383 * this avoids any races wrt polling state changes and thereby avoids
386 static bool set_nr_and_not_polling(struct task_struct *p)
388 struct thread_info *ti = task_thread_info(p);
389 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
393 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
395 * If this returns true, then the idle task promises to call
396 * sched_ttwu_pending() and reschedule soon.
398 static bool set_nr_if_polling(struct task_struct *p)
400 struct thread_info *ti = task_thread_info(p);
401 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
404 if (!(val & _TIF_POLLING_NRFLAG))
406 if (val & _TIF_NEED_RESCHED)
408 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
417 static bool set_nr_and_not_polling(struct task_struct *p)
419 set_tsk_need_resched(p);
424 static bool set_nr_if_polling(struct task_struct *p)
431 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
433 struct wake_q_node *node = &task->wake_q;
436 * Atomically grab the task, if ->wake_q is !nil already it means
437 * its already queued (either by us or someone else) and will get the
438 * wakeup due to that.
440 * This cmpxchg() implies a full barrier, which pairs with the write
441 * barrier implied by the wakeup in wake_up_q().
443 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
446 get_task_struct(task);
449 * The head is context local, there can be no concurrency.
452 head->lastp = &node->next;
455 void wake_up_q(struct wake_q_head *head)
457 struct wake_q_node *node = head->first;
459 while (node != WAKE_Q_TAIL) {
460 struct task_struct *task;
462 task = container_of(node, struct task_struct, wake_q);
464 /* Task can safely be re-inserted now: */
466 task->wake_q.next = NULL;
469 * wake_up_process() implies a wmb() to pair with the queueing
470 * in wake_q_add() so as not to miss wakeups.
472 wake_up_process(task);
473 put_task_struct(task);
478 * resched_curr - mark rq's current task 'to be rescheduled now'.
480 * On UP this means the setting of the need_resched flag, on SMP it
481 * might also involve a cross-CPU call to trigger the scheduler on
484 void resched_curr(struct rq *rq)
486 struct task_struct *curr = rq->curr;
489 lockdep_assert_held(&rq->lock);
491 if (test_tsk_need_resched(curr))
496 if (cpu == smp_processor_id()) {
497 set_tsk_need_resched(curr);
498 set_preempt_need_resched();
502 if (set_nr_and_not_polling(curr))
503 smp_send_reschedule(cpu);
505 trace_sched_wake_idle_without_ipi(cpu);
508 void resched_cpu(int cpu)
510 struct rq *rq = cpu_rq(cpu);
513 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
516 raw_spin_unlock_irqrestore(&rq->lock, flags);
520 #ifdef CONFIG_NO_HZ_COMMON
522 * In the semi idle case, use the nearest busy CPU for migrating timers
523 * from an idle CPU. This is good for power-savings.
525 * We don't do similar optimization for completely idle system, as
526 * selecting an idle CPU will add more delays to the timers than intended
527 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
529 int get_nohz_timer_target(void)
531 int i, cpu = smp_processor_id();
532 struct sched_domain *sd;
534 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
538 for_each_domain(cpu, sd) {
539 for_each_cpu(i, sched_domain_span(sd)) {
543 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
550 if (!is_housekeeping_cpu(cpu))
551 cpu = housekeeping_any_cpu();
558 * When add_timer_on() enqueues a timer into the timer wheel of an
559 * idle CPU then this timer might expire before the next timer event
560 * which is scheduled to wake up that CPU. In case of a completely
561 * idle system the next event might even be infinite time into the
562 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
563 * leaves the inner idle loop so the newly added timer is taken into
564 * account when the CPU goes back to idle and evaluates the timer
565 * wheel for the next timer event.
567 static void wake_up_idle_cpu(int cpu)
569 struct rq *rq = cpu_rq(cpu);
571 if (cpu == smp_processor_id())
574 if (set_nr_and_not_polling(rq->idle))
575 smp_send_reschedule(cpu);
577 trace_sched_wake_idle_without_ipi(cpu);
580 static bool wake_up_full_nohz_cpu(int cpu)
583 * We just need the target to call irq_exit() and re-evaluate
584 * the next tick. The nohz full kick at least implies that.
585 * If needed we can still optimize that later with an
588 if (cpu_is_offline(cpu))
589 return true; /* Don't try to wake offline CPUs. */
590 if (tick_nohz_full_cpu(cpu)) {
591 if (cpu != smp_processor_id() ||
592 tick_nohz_tick_stopped())
593 tick_nohz_full_kick_cpu(cpu);
601 * Wake up the specified CPU. If the CPU is going offline, it is the
602 * caller's responsibility to deal with the lost wakeup, for example,
603 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
605 void wake_up_nohz_cpu(int cpu)
607 if (!wake_up_full_nohz_cpu(cpu))
608 wake_up_idle_cpu(cpu);
611 static inline bool got_nohz_idle_kick(void)
613 int cpu = smp_processor_id();
615 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
618 if (idle_cpu(cpu) && !need_resched())
622 * We can't run Idle Load Balance on this CPU for this time so we
623 * cancel it and clear NOHZ_BALANCE_KICK
625 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
629 #else /* CONFIG_NO_HZ_COMMON */
631 static inline bool got_nohz_idle_kick(void)
636 #endif /* CONFIG_NO_HZ_COMMON */
638 #ifdef CONFIG_NO_HZ_FULL
639 bool sched_can_stop_tick(struct rq *rq)
643 /* Deadline tasks, even if single, need the tick */
644 if (rq->dl.dl_nr_running)
648 * If there are more than one RR tasks, we need the tick to effect the
649 * actual RR behaviour.
651 if (rq->rt.rr_nr_running) {
652 if (rq->rt.rr_nr_running == 1)
659 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
660 * forced preemption between FIFO tasks.
662 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
667 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
668 * if there's more than one we need the tick for involuntary
671 if (rq->nr_running > 1)
676 #endif /* CONFIG_NO_HZ_FULL */
678 void sched_avg_update(struct rq *rq)
680 s64 period = sched_avg_period();
682 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
684 * Inline assembly required to prevent the compiler
685 * optimising this loop into a divmod call.
686 * See __iter_div_u64_rem() for another example of this.
688 asm("" : "+rm" (rq->age_stamp));
689 rq->age_stamp += period;
694 #endif /* CONFIG_SMP */
696 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
697 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
699 * Iterate task_group tree rooted at *from, calling @down when first entering a
700 * node and @up when leaving it for the final time.
702 * Caller must hold rcu_lock or sufficient equivalent.
704 int walk_tg_tree_from(struct task_group *from,
705 tg_visitor down, tg_visitor up, void *data)
707 struct task_group *parent, *child;
713 ret = (*down)(parent, data);
716 list_for_each_entry_rcu(child, &parent->children, siblings) {
723 ret = (*up)(parent, data);
724 if (ret || parent == from)
728 parent = parent->parent;
735 int tg_nop(struct task_group *tg, void *data)
741 static void set_load_weight(struct task_struct *p)
743 int prio = p->static_prio - MAX_RT_PRIO;
744 struct load_weight *load = &p->se.load;
747 * SCHED_IDLE tasks get minimal weight:
749 if (idle_policy(p->policy)) {
750 load->weight = scale_load(WEIGHT_IDLEPRIO);
751 load->inv_weight = WMULT_IDLEPRIO;
755 load->weight = scale_load(sched_prio_to_weight[prio]);
756 load->inv_weight = sched_prio_to_wmult[prio];
759 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
762 if (!(flags & ENQUEUE_RESTORE))
763 sched_info_queued(rq, p);
764 p->sched_class->enqueue_task(rq, p, flags);
767 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
770 if (!(flags & DEQUEUE_SAVE))
771 sched_info_dequeued(rq, p);
772 p->sched_class->dequeue_task(rq, p, flags);
775 void activate_task(struct rq *rq, struct task_struct *p, int flags)
777 if (task_contributes_to_load(p))
778 rq->nr_uninterruptible--;
780 enqueue_task(rq, p, flags);
783 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
785 if (task_contributes_to_load(p))
786 rq->nr_uninterruptible++;
788 dequeue_task(rq, p, flags);
791 void sched_set_stop_task(int cpu, struct task_struct *stop)
793 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
794 struct task_struct *old_stop = cpu_rq(cpu)->stop;
798 * Make it appear like a SCHED_FIFO task, its something
799 * userspace knows about and won't get confused about.
801 * Also, it will make PI more or less work without too
802 * much confusion -- but then, stop work should not
803 * rely on PI working anyway.
805 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
807 stop->sched_class = &stop_sched_class;
810 cpu_rq(cpu)->stop = stop;
814 * Reset it back to a normal scheduling class so that
815 * it can die in pieces.
817 old_stop->sched_class = &rt_sched_class;
822 * __normal_prio - return the priority that is based on the static prio
824 static inline int __normal_prio(struct task_struct *p)
826 return p->static_prio;
830 * Calculate the expected normal priority: i.e. priority
831 * without taking RT-inheritance into account. Might be
832 * boosted by interactivity modifiers. Changes upon fork,
833 * setprio syscalls, and whenever the interactivity
834 * estimator recalculates.
836 static inline int normal_prio(struct task_struct *p)
840 if (task_has_dl_policy(p))
841 prio = MAX_DL_PRIO-1;
842 else if (task_has_rt_policy(p))
843 prio = MAX_RT_PRIO-1 - p->rt_priority;
845 prio = __normal_prio(p);
850 * Calculate the current priority, i.e. the priority
851 * taken into account by the scheduler. This value might
852 * be boosted by RT tasks, or might be boosted by
853 * interactivity modifiers. Will be RT if the task got
854 * RT-boosted. If not then it returns p->normal_prio.
856 static int effective_prio(struct task_struct *p)
858 p->normal_prio = normal_prio(p);
860 * If we are RT tasks or we were boosted to RT priority,
861 * keep the priority unchanged. Otherwise, update priority
862 * to the normal priority:
864 if (!rt_prio(p->prio))
865 return p->normal_prio;
870 * task_curr - is this task currently executing on a CPU?
871 * @p: the task in question.
873 * Return: 1 if the task is currently executing. 0 otherwise.
875 inline int task_curr(const struct task_struct *p)
877 return cpu_curr(task_cpu(p)) == p;
881 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
882 * use the balance_callback list if you want balancing.
884 * this means any call to check_class_changed() must be followed by a call to
885 * balance_callback().
887 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
888 const struct sched_class *prev_class,
891 if (prev_class != p->sched_class) {
892 if (prev_class->switched_from)
893 prev_class->switched_from(rq, p);
895 p->sched_class->switched_to(rq, p);
896 } else if (oldprio != p->prio || dl_task(p))
897 p->sched_class->prio_changed(rq, p, oldprio);
900 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
902 const struct sched_class *class;
904 if (p->sched_class == rq->curr->sched_class) {
905 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
907 for_each_class(class) {
908 if (class == rq->curr->sched_class)
910 if (class == p->sched_class) {
918 * A queue event has occurred, and we're going to schedule. In
919 * this case, we can save a useless back to back clock update.
921 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
922 rq_clock_skip_update(rq, true);
927 * This is how migration works:
929 * 1) we invoke migration_cpu_stop() on the target CPU using
931 * 2) stopper starts to run (implicitly forcing the migrated thread
933 * 3) it checks whether the migrated task is still in the wrong runqueue.
934 * 4) if it's in the wrong runqueue then the migration thread removes
935 * it and puts it into the right queue.
936 * 5) stopper completes and stop_one_cpu() returns and the migration
941 * move_queued_task - move a queued task to new rq.
943 * Returns (locked) new rq. Old rq's lock is released.
945 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
947 lockdep_assert_held(&rq->lock);
949 p->on_rq = TASK_ON_RQ_MIGRATING;
950 dequeue_task(rq, p, 0);
951 set_task_cpu(p, new_cpu);
952 raw_spin_unlock(&rq->lock);
954 rq = cpu_rq(new_cpu);
956 raw_spin_lock(&rq->lock);
957 BUG_ON(task_cpu(p) != new_cpu);
958 enqueue_task(rq, p, 0);
959 p->on_rq = TASK_ON_RQ_QUEUED;
960 check_preempt_curr(rq, p, 0);
965 struct migration_arg {
966 struct task_struct *task;
971 * Move (not current) task off this CPU, onto the destination CPU. We're doing
972 * this because either it can't run here any more (set_cpus_allowed()
973 * away from this CPU, or CPU going down), or because we're
974 * attempting to rebalance this task on exec (sched_exec).
976 * So we race with normal scheduler movements, but that's OK, as long
977 * as the task is no longer on this CPU.
979 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
981 if (unlikely(!cpu_active(dest_cpu)))
984 /* Affinity changed (again). */
985 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
988 rq = move_queued_task(rq, p, dest_cpu);
994 * migration_cpu_stop - this will be executed by a highprio stopper thread
995 * and performs thread migration by bumping thread off CPU then
996 * 'pushing' onto another runqueue.
998 static int migration_cpu_stop(void *data)
1000 struct migration_arg *arg = data;
1001 struct task_struct *p = arg->task;
1002 struct rq *rq = this_rq();
1005 * The original target CPU might have gone down and we might
1006 * be on another CPU but it doesn't matter.
1008 local_irq_disable();
1010 * We need to explicitly wake pending tasks before running
1011 * __migrate_task() such that we will not miss enforcing cpus_allowed
1012 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1014 sched_ttwu_pending();
1016 raw_spin_lock(&p->pi_lock);
1017 raw_spin_lock(&rq->lock);
1019 * If task_rq(p) != rq, it cannot be migrated here, because we're
1020 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1021 * we're holding p->pi_lock.
1023 if (task_rq(p) == rq) {
1024 if (task_on_rq_queued(p))
1025 rq = __migrate_task(rq, p, arg->dest_cpu);
1027 p->wake_cpu = arg->dest_cpu;
1029 raw_spin_unlock(&rq->lock);
1030 raw_spin_unlock(&p->pi_lock);
1037 * sched_class::set_cpus_allowed must do the below, but is not required to
1038 * actually call this function.
1040 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1042 cpumask_copy(&p->cpus_allowed, new_mask);
1043 p->nr_cpus_allowed = cpumask_weight(new_mask);
1046 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1048 struct rq *rq = task_rq(p);
1049 bool queued, running;
1051 lockdep_assert_held(&p->pi_lock);
1053 queued = task_on_rq_queued(p);
1054 running = task_current(rq, p);
1058 * Because __kthread_bind() calls this on blocked tasks without
1061 lockdep_assert_held(&rq->lock);
1062 dequeue_task(rq, p, DEQUEUE_SAVE);
1065 put_prev_task(rq, p);
1067 p->sched_class->set_cpus_allowed(p, new_mask);
1070 enqueue_task(rq, p, ENQUEUE_RESTORE);
1072 set_curr_task(rq, p);
1076 * Change a given task's CPU affinity. Migrate the thread to a
1077 * proper CPU and schedule it away if the CPU it's executing on
1078 * is removed from the allowed bitmask.
1080 * NOTE: the caller must have a valid reference to the task, the
1081 * task must not exit() & deallocate itself prematurely. The
1082 * call is not atomic; no spinlocks may be held.
1084 static int __set_cpus_allowed_ptr(struct task_struct *p,
1085 const struct cpumask *new_mask, bool check)
1087 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1088 unsigned int dest_cpu;
1093 rq = task_rq_lock(p, &rf);
1094 update_rq_clock(rq);
1096 if (p->flags & PF_KTHREAD) {
1098 * Kernel threads are allowed on online && !active CPUs
1100 cpu_valid_mask = cpu_online_mask;
1104 * Must re-check here, to close a race against __kthread_bind(),
1105 * sched_setaffinity() is not guaranteed to observe the flag.
1107 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1112 if (cpumask_equal(&p->cpus_allowed, new_mask))
1115 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1120 do_set_cpus_allowed(p, new_mask);
1122 if (p->flags & PF_KTHREAD) {
1124 * For kernel threads that do indeed end up on online &&
1125 * !active we want to ensure they are strict per-CPU threads.
1127 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1128 !cpumask_intersects(new_mask, cpu_active_mask) &&
1129 p->nr_cpus_allowed != 1);
1132 /* Can the task run on the task's current CPU? If so, we're done */
1133 if (cpumask_test_cpu(task_cpu(p), new_mask))
1136 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1137 if (task_running(rq, p) || p->state == TASK_WAKING) {
1138 struct migration_arg arg = { p, dest_cpu };
1139 /* Need help from migration thread: drop lock and wait. */
1140 task_rq_unlock(rq, p, &rf);
1141 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1142 tlb_migrate_finish(p->mm);
1144 } else if (task_on_rq_queued(p)) {
1146 * OK, since we're going to drop the lock immediately
1147 * afterwards anyway.
1149 rq_unpin_lock(rq, &rf);
1150 rq = move_queued_task(rq, p, dest_cpu);
1151 rq_repin_lock(rq, &rf);
1154 task_rq_unlock(rq, p, &rf);
1159 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1161 return __set_cpus_allowed_ptr(p, new_mask, false);
1163 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1165 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1167 #ifdef CONFIG_SCHED_DEBUG
1169 * We should never call set_task_cpu() on a blocked task,
1170 * ttwu() will sort out the placement.
1172 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1176 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1177 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1178 * time relying on p->on_rq.
1180 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1181 p->sched_class == &fair_sched_class &&
1182 (p->on_rq && !task_on_rq_migrating(p)));
1184 #ifdef CONFIG_LOCKDEP
1186 * The caller should hold either p->pi_lock or rq->lock, when changing
1187 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1189 * sched_move_task() holds both and thus holding either pins the cgroup,
1192 * Furthermore, all task_rq users should acquire both locks, see
1195 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1196 lockdep_is_held(&task_rq(p)->lock)));
1200 trace_sched_migrate_task(p, new_cpu);
1202 if (task_cpu(p) != new_cpu) {
1203 if (p->sched_class->migrate_task_rq)
1204 p->sched_class->migrate_task_rq(p);
1205 p->se.nr_migrations++;
1206 perf_event_task_migrate(p);
1209 __set_task_cpu(p, new_cpu);
1212 static void __migrate_swap_task(struct task_struct *p, int cpu)
1214 if (task_on_rq_queued(p)) {
1215 struct rq *src_rq, *dst_rq;
1217 src_rq = task_rq(p);
1218 dst_rq = cpu_rq(cpu);
1220 p->on_rq = TASK_ON_RQ_MIGRATING;
1221 deactivate_task(src_rq, p, 0);
1222 set_task_cpu(p, cpu);
1223 activate_task(dst_rq, p, 0);
1224 p->on_rq = TASK_ON_RQ_QUEUED;
1225 check_preempt_curr(dst_rq, p, 0);
1228 * Task isn't running anymore; make it appear like we migrated
1229 * it before it went to sleep. This means on wakeup we make the
1230 * previous CPU our target instead of where it really is.
1236 struct migration_swap_arg {
1237 struct task_struct *src_task, *dst_task;
1238 int src_cpu, dst_cpu;
1241 static int migrate_swap_stop(void *data)
1243 struct migration_swap_arg *arg = data;
1244 struct rq *src_rq, *dst_rq;
1247 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1250 src_rq = cpu_rq(arg->src_cpu);
1251 dst_rq = cpu_rq(arg->dst_cpu);
1253 double_raw_lock(&arg->src_task->pi_lock,
1254 &arg->dst_task->pi_lock);
1255 double_rq_lock(src_rq, dst_rq);
1257 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1260 if (task_cpu(arg->src_task) != arg->src_cpu)
1263 if (!cpumask_test_cpu(arg->dst_cpu, &arg->src_task->cpus_allowed))
1266 if (!cpumask_test_cpu(arg->src_cpu, &arg->dst_task->cpus_allowed))
1269 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1270 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1275 double_rq_unlock(src_rq, dst_rq);
1276 raw_spin_unlock(&arg->dst_task->pi_lock);
1277 raw_spin_unlock(&arg->src_task->pi_lock);
1283 * Cross migrate two tasks
1285 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1287 struct migration_swap_arg arg;
1290 arg = (struct migration_swap_arg){
1292 .src_cpu = task_cpu(cur),
1294 .dst_cpu = task_cpu(p),
1297 if (arg.src_cpu == arg.dst_cpu)
1301 * These three tests are all lockless; this is OK since all of them
1302 * will be re-checked with proper locks held further down the line.
1304 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1307 if (!cpumask_test_cpu(arg.dst_cpu, &arg.src_task->cpus_allowed))
1310 if (!cpumask_test_cpu(arg.src_cpu, &arg.dst_task->cpus_allowed))
1313 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1314 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1321 * wait_task_inactive - wait for a thread to unschedule.
1323 * If @match_state is nonzero, it's the @p->state value just checked and
1324 * not expected to change. If it changes, i.e. @p might have woken up,
1325 * then return zero. When we succeed in waiting for @p to be off its CPU,
1326 * we return a positive number (its total switch count). If a second call
1327 * a short while later returns the same number, the caller can be sure that
1328 * @p has remained unscheduled the whole time.
1330 * The caller must ensure that the task *will* unschedule sometime soon,
1331 * else this function might spin for a *long* time. This function can't
1332 * be called with interrupts off, or it may introduce deadlock with
1333 * smp_call_function() if an IPI is sent by the same process we are
1334 * waiting to become inactive.
1336 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1338 int running, queued;
1345 * We do the initial early heuristics without holding
1346 * any task-queue locks at all. We'll only try to get
1347 * the runqueue lock when things look like they will
1353 * If the task is actively running on another CPU
1354 * still, just relax and busy-wait without holding
1357 * NOTE! Since we don't hold any locks, it's not
1358 * even sure that "rq" stays as the right runqueue!
1359 * But we don't care, since "task_running()" will
1360 * return false if the runqueue has changed and p
1361 * is actually now running somewhere else!
1363 while (task_running(rq, p)) {
1364 if (match_state && unlikely(p->state != match_state))
1370 * Ok, time to look more closely! We need the rq
1371 * lock now, to be *sure*. If we're wrong, we'll
1372 * just go back and repeat.
1374 rq = task_rq_lock(p, &rf);
1375 trace_sched_wait_task(p);
1376 running = task_running(rq, p);
1377 queued = task_on_rq_queued(p);
1379 if (!match_state || p->state == match_state)
1380 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1381 task_rq_unlock(rq, p, &rf);
1384 * If it changed from the expected state, bail out now.
1386 if (unlikely(!ncsw))
1390 * Was it really running after all now that we
1391 * checked with the proper locks actually held?
1393 * Oops. Go back and try again..
1395 if (unlikely(running)) {
1401 * It's not enough that it's not actively running,
1402 * it must be off the runqueue _entirely_, and not
1405 * So if it was still runnable (but just not actively
1406 * running right now), it's preempted, and we should
1407 * yield - it could be a while.
1409 if (unlikely(queued)) {
1410 ktime_t to = NSEC_PER_SEC / HZ;
1412 set_current_state(TASK_UNINTERRUPTIBLE);
1413 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1418 * Ahh, all good. It wasn't running, and it wasn't
1419 * runnable, which means that it will never become
1420 * running in the future either. We're all done!
1429 * kick_process - kick a running thread to enter/exit the kernel
1430 * @p: the to-be-kicked thread
1432 * Cause a process which is running on another CPU to enter
1433 * kernel-mode, without any delay. (to get signals handled.)
1435 * NOTE: this function doesn't have to take the runqueue lock,
1436 * because all it wants to ensure is that the remote task enters
1437 * the kernel. If the IPI races and the task has been migrated
1438 * to another CPU then no harm is done and the purpose has been
1441 void kick_process(struct task_struct *p)
1447 if ((cpu != smp_processor_id()) && task_curr(p))
1448 smp_send_reschedule(cpu);
1451 EXPORT_SYMBOL_GPL(kick_process);
1454 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1456 * A few notes on cpu_active vs cpu_online:
1458 * - cpu_active must be a subset of cpu_online
1460 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1461 * see __set_cpus_allowed_ptr(). At this point the newly online
1462 * CPU isn't yet part of the sched domains, and balancing will not
1465 * - on CPU-down we clear cpu_active() to mask the sched domains and
1466 * avoid the load balancer to place new tasks on the to be removed
1467 * CPU. Existing tasks will remain running there and will be taken
1470 * This means that fallback selection must not select !active CPUs.
1471 * And can assume that any active CPU must be online. Conversely
1472 * select_task_rq() below may allow selection of !active CPUs in order
1473 * to satisfy the above rules.
1475 static int select_fallback_rq(int cpu, struct task_struct *p)
1477 int nid = cpu_to_node(cpu);
1478 const struct cpumask *nodemask = NULL;
1479 enum { cpuset, possible, fail } state = cpuset;
1483 * If the node that the CPU is on has been offlined, cpu_to_node()
1484 * will return -1. There is no CPU on the node, and we should
1485 * select the CPU on the other node.
1488 nodemask = cpumask_of_node(nid);
1490 /* Look for allowed, online CPU in same node. */
1491 for_each_cpu(dest_cpu, nodemask) {
1492 if (!cpu_active(dest_cpu))
1494 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
1500 /* Any allowed, online CPU? */
1501 for_each_cpu(dest_cpu, &p->cpus_allowed) {
1502 if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1504 if (!cpu_online(dest_cpu))
1509 /* No more Mr. Nice Guy. */
1512 if (IS_ENABLED(CONFIG_CPUSETS)) {
1513 cpuset_cpus_allowed_fallback(p);
1519 do_set_cpus_allowed(p, cpu_possible_mask);
1530 if (state != cpuset) {
1532 * Don't tell them about moving exiting tasks or
1533 * kernel threads (both mm NULL), since they never
1536 if (p->mm && printk_ratelimit()) {
1537 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1538 task_pid_nr(p), p->comm, cpu);
1546 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1549 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1551 lockdep_assert_held(&p->pi_lock);
1553 if (p->nr_cpus_allowed > 1)
1554 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1556 cpu = cpumask_any(&p->cpus_allowed);
1559 * In order not to call set_task_cpu() on a blocking task we need
1560 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1563 * Since this is common to all placement strategies, this lives here.
1565 * [ this allows ->select_task() to simply return task_cpu(p) and
1566 * not worry about this generic constraint ]
1568 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
1570 cpu = select_fallback_rq(task_cpu(p), p);
1575 static void update_avg(u64 *avg, u64 sample)
1577 s64 diff = sample - *avg;
1583 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1584 const struct cpumask *new_mask, bool check)
1586 return set_cpus_allowed_ptr(p, new_mask);
1589 #endif /* CONFIG_SMP */
1592 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1596 if (!schedstat_enabled())
1602 if (cpu == rq->cpu) {
1603 schedstat_inc(rq->ttwu_local);
1604 schedstat_inc(p->se.statistics.nr_wakeups_local);
1606 struct sched_domain *sd;
1608 schedstat_inc(p->se.statistics.nr_wakeups_remote);
1610 for_each_domain(rq->cpu, sd) {
1611 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1612 schedstat_inc(sd->ttwu_wake_remote);
1619 if (wake_flags & WF_MIGRATED)
1620 schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1621 #endif /* CONFIG_SMP */
1623 schedstat_inc(rq->ttwu_count);
1624 schedstat_inc(p->se.statistics.nr_wakeups);
1626 if (wake_flags & WF_SYNC)
1627 schedstat_inc(p->se.statistics.nr_wakeups_sync);
1630 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1632 activate_task(rq, p, en_flags);
1633 p->on_rq = TASK_ON_RQ_QUEUED;
1635 /* If a worker is waking up, notify the workqueue: */
1636 if (p->flags & PF_WQ_WORKER)
1637 wq_worker_waking_up(p, cpu_of(rq));
1641 * Mark the task runnable and perform wakeup-preemption.
1643 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1644 struct rq_flags *rf)
1646 check_preempt_curr(rq, p, wake_flags);
1647 p->state = TASK_RUNNING;
1648 trace_sched_wakeup(p);
1651 if (p->sched_class->task_woken) {
1653 * Our task @p is fully woken up and running; so its safe to
1654 * drop the rq->lock, hereafter rq is only used for statistics.
1656 rq_unpin_lock(rq, rf);
1657 p->sched_class->task_woken(rq, p);
1658 rq_repin_lock(rq, rf);
1661 if (rq->idle_stamp) {
1662 u64 delta = rq_clock(rq) - rq->idle_stamp;
1663 u64 max = 2*rq->max_idle_balance_cost;
1665 update_avg(&rq->avg_idle, delta);
1667 if (rq->avg_idle > max)
1676 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1677 struct rq_flags *rf)
1679 int en_flags = ENQUEUE_WAKEUP;
1681 lockdep_assert_held(&rq->lock);
1684 if (p->sched_contributes_to_load)
1685 rq->nr_uninterruptible--;
1687 if (wake_flags & WF_MIGRATED)
1688 en_flags |= ENQUEUE_MIGRATED;
1691 ttwu_activate(rq, p, en_flags);
1692 ttwu_do_wakeup(rq, p, wake_flags, rf);
1696 * Called in case the task @p isn't fully descheduled from its runqueue,
1697 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1698 * since all we need to do is flip p->state to TASK_RUNNING, since
1699 * the task is still ->on_rq.
1701 static int ttwu_remote(struct task_struct *p, int wake_flags)
1707 rq = __task_rq_lock(p, &rf);
1708 if (task_on_rq_queued(p)) {
1709 /* check_preempt_curr() may use rq clock */
1710 update_rq_clock(rq);
1711 ttwu_do_wakeup(rq, p, wake_flags, &rf);
1714 __task_rq_unlock(rq, &rf);
1720 void sched_ttwu_pending(void)
1722 struct rq *rq = this_rq();
1723 struct llist_node *llist = llist_del_all(&rq->wake_list);
1724 struct task_struct *p;
1725 unsigned long flags;
1731 raw_spin_lock_irqsave(&rq->lock, flags);
1732 rq_pin_lock(rq, &rf);
1737 p = llist_entry(llist, struct task_struct, wake_entry);
1738 llist = llist_next(llist);
1740 if (p->sched_remote_wakeup)
1741 wake_flags = WF_MIGRATED;
1743 ttwu_do_activate(rq, p, wake_flags, &rf);
1746 rq_unpin_lock(rq, &rf);
1747 raw_spin_unlock_irqrestore(&rq->lock, flags);
1750 void scheduler_ipi(void)
1753 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1754 * TIF_NEED_RESCHED remotely (for the first time) will also send
1757 preempt_fold_need_resched();
1759 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1763 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1764 * traditionally all their work was done from the interrupt return
1765 * path. Now that we actually do some work, we need to make sure
1768 * Some archs already do call them, luckily irq_enter/exit nest
1771 * Arguably we should visit all archs and update all handlers,
1772 * however a fair share of IPIs are still resched only so this would
1773 * somewhat pessimize the simple resched case.
1776 sched_ttwu_pending();
1779 * Check if someone kicked us for doing the nohz idle load balance.
1781 if (unlikely(got_nohz_idle_kick())) {
1782 this_rq()->idle_balance = 1;
1783 raise_softirq_irqoff(SCHED_SOFTIRQ);
1788 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1790 struct rq *rq = cpu_rq(cpu);
1792 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1794 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1795 if (!set_nr_if_polling(rq->idle))
1796 smp_send_reschedule(cpu);
1798 trace_sched_wake_idle_without_ipi(cpu);
1802 void wake_up_if_idle(int cpu)
1804 struct rq *rq = cpu_rq(cpu);
1805 unsigned long flags;
1809 if (!is_idle_task(rcu_dereference(rq->curr)))
1812 if (set_nr_if_polling(rq->idle)) {
1813 trace_sched_wake_idle_without_ipi(cpu);
1815 raw_spin_lock_irqsave(&rq->lock, flags);
1816 if (is_idle_task(rq->curr))
1817 smp_send_reschedule(cpu);
1818 /* Else CPU is not idle, do nothing here: */
1819 raw_spin_unlock_irqrestore(&rq->lock, flags);
1826 bool cpus_share_cache(int this_cpu, int that_cpu)
1828 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1830 #endif /* CONFIG_SMP */
1832 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1834 struct rq *rq = cpu_rq(cpu);
1837 #if defined(CONFIG_SMP)
1838 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1839 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
1840 ttwu_queue_remote(p, cpu, wake_flags);
1845 raw_spin_lock(&rq->lock);
1846 rq_pin_lock(rq, &rf);
1847 ttwu_do_activate(rq, p, wake_flags, &rf);
1848 rq_unpin_lock(rq, &rf);
1849 raw_spin_unlock(&rq->lock);
1853 * Notes on Program-Order guarantees on SMP systems.
1857 * The basic program-order guarantee on SMP systems is that when a task [t]
1858 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1859 * execution on its new CPU [c1].
1861 * For migration (of runnable tasks) this is provided by the following means:
1863 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1864 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1865 * rq(c1)->lock (if not at the same time, then in that order).
1866 * C) LOCK of the rq(c1)->lock scheduling in task
1868 * Transitivity guarantees that B happens after A and C after B.
1869 * Note: we only require RCpc transitivity.
1870 * Note: the CPU doing B need not be c0 or c1
1879 * UNLOCK rq(0)->lock
1881 * LOCK rq(0)->lock // orders against CPU0
1883 * UNLOCK rq(0)->lock
1887 * UNLOCK rq(1)->lock
1889 * LOCK rq(1)->lock // orders against CPU2
1892 * UNLOCK rq(1)->lock
1895 * BLOCKING -- aka. SLEEP + WAKEUP
1897 * For blocking we (obviously) need to provide the same guarantee as for
1898 * migration. However the means are completely different as there is no lock
1899 * chain to provide order. Instead we do:
1901 * 1) smp_store_release(X->on_cpu, 0)
1902 * 2) smp_cond_load_acquire(!X->on_cpu)
1906 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1908 * LOCK rq(0)->lock LOCK X->pi_lock
1911 * smp_store_release(X->on_cpu, 0);
1913 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1919 * X->state = RUNNING
1920 * UNLOCK rq(2)->lock
1922 * LOCK rq(2)->lock // orders against CPU1
1925 * UNLOCK rq(2)->lock
1928 * UNLOCK rq(0)->lock
1931 * However; for wakeups there is a second guarantee we must provide, namely we
1932 * must observe the state that lead to our wakeup. That is, not only must our
1933 * task observe its own prior state, it must also observe the stores prior to
1936 * This means that any means of doing remote wakeups must order the CPU doing
1937 * the wakeup against the CPU the task is going to end up running on. This,
1938 * however, is already required for the regular Program-Order guarantee above,
1939 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1944 * try_to_wake_up - wake up a thread
1945 * @p: the thread to be awakened
1946 * @state: the mask of task states that can be woken
1947 * @wake_flags: wake modifier flags (WF_*)
1949 * If (@state & @p->state) @p->state = TASK_RUNNING.
1951 * If the task was not queued/runnable, also place it back on a runqueue.
1953 * Atomic against schedule() which would dequeue a task, also see
1954 * set_current_state().
1956 * Return: %true if @p->state changes (an actual wakeup was done),
1960 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1962 unsigned long flags;
1963 int cpu, success = 0;
1966 * If we are going to wake up a thread waiting for CONDITION we
1967 * need to ensure that CONDITION=1 done by the caller can not be
1968 * reordered with p->state check below. This pairs with mb() in
1969 * set_current_state() the waiting thread does.
1971 smp_mb__before_spinlock();
1972 raw_spin_lock_irqsave(&p->pi_lock, flags);
1973 if (!(p->state & state))
1976 trace_sched_waking(p);
1978 /* We're going to change ->state: */
1983 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1984 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1985 * in smp_cond_load_acquire() below.
1987 * sched_ttwu_pending() try_to_wake_up()
1988 * [S] p->on_rq = 1; [L] P->state
1989 * UNLOCK rq->lock -----.
1993 * LOCK rq->lock -----'
1997 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
1999 * Pairs with the UNLOCK+LOCK on rq->lock from the
2000 * last wakeup of our task and the schedule that got our task
2004 if (p->on_rq && ttwu_remote(p, wake_flags))
2009 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2010 * possible to, falsely, observe p->on_cpu == 0.
2012 * One must be running (->on_cpu == 1) in order to remove oneself
2013 * from the runqueue.
2015 * [S] ->on_cpu = 1; [L] ->on_rq
2019 * [S] ->on_rq = 0; [L] ->on_cpu
2021 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2022 * from the consecutive calls to schedule(); the first switching to our
2023 * task, the second putting it to sleep.
2028 * If the owning (remote) CPU is still in the middle of schedule() with
2029 * this task as prev, wait until its done referencing the task.
2031 * Pairs with the smp_store_release() in finish_lock_switch().
2033 * This ensures that tasks getting woken will be fully ordered against
2034 * their previous state and preserve Program Order.
2036 smp_cond_load_acquire(&p->on_cpu, !VAL);
2038 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2039 p->state = TASK_WAKING;
2042 delayacct_blkio_end();
2043 atomic_dec(&task_rq(p)->nr_iowait);
2046 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2047 if (task_cpu(p) != cpu) {
2048 wake_flags |= WF_MIGRATED;
2049 set_task_cpu(p, cpu);
2052 #else /* CONFIG_SMP */
2055 delayacct_blkio_end();
2056 atomic_dec(&task_rq(p)->nr_iowait);
2059 #endif /* CONFIG_SMP */
2061 ttwu_queue(p, cpu, wake_flags);
2063 ttwu_stat(p, cpu, wake_flags);
2065 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2071 * try_to_wake_up_local - try to wake up a local task with rq lock held
2072 * @p: the thread to be awakened
2073 * @cookie: context's cookie for pinning
2075 * Put @p on the run-queue if it's not already there. The caller must
2076 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2079 static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
2081 struct rq *rq = task_rq(p);
2083 if (WARN_ON_ONCE(rq != this_rq()) ||
2084 WARN_ON_ONCE(p == current))
2087 lockdep_assert_held(&rq->lock);
2089 if (!raw_spin_trylock(&p->pi_lock)) {
2091 * This is OK, because current is on_cpu, which avoids it being
2092 * picked for load-balance and preemption/IRQs are still
2093 * disabled avoiding further scheduler activity on it and we've
2094 * not yet picked a replacement task.
2096 rq_unpin_lock(rq, rf);
2097 raw_spin_unlock(&rq->lock);
2098 raw_spin_lock(&p->pi_lock);
2099 raw_spin_lock(&rq->lock);
2100 rq_repin_lock(rq, rf);
2103 if (!(p->state & TASK_NORMAL))
2106 trace_sched_waking(p);
2108 if (!task_on_rq_queued(p)) {
2110 delayacct_blkio_end();
2111 atomic_dec(&rq->nr_iowait);
2113 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2116 ttwu_do_wakeup(rq, p, 0, rf);
2117 ttwu_stat(p, smp_processor_id(), 0);
2119 raw_spin_unlock(&p->pi_lock);
2123 * wake_up_process - Wake up a specific process
2124 * @p: The process to be woken up.
2126 * Attempt to wake up the nominated process and move it to the set of runnable
2129 * Return: 1 if the process was woken up, 0 if it was already running.
2131 * It may be assumed that this function implies a write memory barrier before
2132 * changing the task state if and only if any tasks are woken up.
2134 int wake_up_process(struct task_struct *p)
2136 return try_to_wake_up(p, TASK_NORMAL, 0);
2138 EXPORT_SYMBOL(wake_up_process);
2140 int wake_up_state(struct task_struct *p, unsigned int state)
2142 return try_to_wake_up(p, state, 0);
2146 * This function clears the sched_dl_entity static params.
2148 void __dl_clear_params(struct task_struct *p)
2150 struct sched_dl_entity *dl_se = &p->dl;
2152 dl_se->dl_runtime = 0;
2153 dl_se->dl_deadline = 0;
2154 dl_se->dl_period = 0;
2158 dl_se->dl_throttled = 0;
2159 dl_se->dl_yielded = 0;
2163 * Perform scheduler related setup for a newly forked process p.
2164 * p is forked by current.
2166 * __sched_fork() is basic setup used by init_idle() too:
2168 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2173 p->se.exec_start = 0;
2174 p->se.sum_exec_runtime = 0;
2175 p->se.prev_sum_exec_runtime = 0;
2176 p->se.nr_migrations = 0;
2178 INIT_LIST_HEAD(&p->se.group_node);
2180 #ifdef CONFIG_FAIR_GROUP_SCHED
2181 p->se.cfs_rq = NULL;
2184 #ifdef CONFIG_SCHEDSTATS
2185 /* Even if schedstat is disabled, there should not be garbage */
2186 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2189 RB_CLEAR_NODE(&p->dl.rb_node);
2190 init_dl_task_timer(&p->dl);
2191 __dl_clear_params(p);
2193 INIT_LIST_HEAD(&p->rt.run_list);
2195 p->rt.time_slice = sched_rr_timeslice;
2199 #ifdef CONFIG_PREEMPT_NOTIFIERS
2200 INIT_HLIST_HEAD(&p->preempt_notifiers);
2203 #ifdef CONFIG_NUMA_BALANCING
2204 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2205 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2206 p->mm->numa_scan_seq = 0;
2209 if (clone_flags & CLONE_VM)
2210 p->numa_preferred_nid = current->numa_preferred_nid;
2212 p->numa_preferred_nid = -1;
2214 p->node_stamp = 0ULL;
2215 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2216 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2217 p->numa_work.next = &p->numa_work;
2218 p->numa_faults = NULL;
2219 p->last_task_numa_placement = 0;
2220 p->last_sum_exec_runtime = 0;
2222 p->numa_group = NULL;
2223 #endif /* CONFIG_NUMA_BALANCING */
2226 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2228 #ifdef CONFIG_NUMA_BALANCING
2230 void set_numabalancing_state(bool enabled)
2233 static_branch_enable(&sched_numa_balancing);
2235 static_branch_disable(&sched_numa_balancing);
2238 #ifdef CONFIG_PROC_SYSCTL
2239 int sysctl_numa_balancing(struct ctl_table *table, int write,
2240 void __user *buffer, size_t *lenp, loff_t *ppos)
2244 int state = static_branch_likely(&sched_numa_balancing);
2246 if (write && !capable(CAP_SYS_ADMIN))
2251 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2255 set_numabalancing_state(state);
2261 #ifdef CONFIG_SCHEDSTATS
2263 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2264 static bool __initdata __sched_schedstats = false;
2266 static void set_schedstats(bool enabled)
2269 static_branch_enable(&sched_schedstats);
2271 static_branch_disable(&sched_schedstats);
2274 void force_schedstat_enabled(void)
2276 if (!schedstat_enabled()) {
2277 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2278 static_branch_enable(&sched_schedstats);
2282 static int __init setup_schedstats(char *str)
2289 * This code is called before jump labels have been set up, so we can't
2290 * change the static branch directly just yet. Instead set a temporary
2291 * variable so init_schedstats() can do it later.
2293 if (!strcmp(str, "enable")) {
2294 __sched_schedstats = true;
2296 } else if (!strcmp(str, "disable")) {
2297 __sched_schedstats = false;
2302 pr_warn("Unable to parse schedstats=\n");
2306 __setup("schedstats=", setup_schedstats);
2308 static void __init init_schedstats(void)
2310 set_schedstats(__sched_schedstats);
2313 #ifdef CONFIG_PROC_SYSCTL
2314 int sysctl_schedstats(struct ctl_table *table, int write,
2315 void __user *buffer, size_t *lenp, loff_t *ppos)
2319 int state = static_branch_likely(&sched_schedstats);
2321 if (write && !capable(CAP_SYS_ADMIN))
2326 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2330 set_schedstats(state);
2333 #endif /* CONFIG_PROC_SYSCTL */
2334 #else /* !CONFIG_SCHEDSTATS */
2335 static inline void init_schedstats(void) {}
2336 #endif /* CONFIG_SCHEDSTATS */
2339 * fork()/clone()-time setup:
2341 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2343 unsigned long flags;
2344 int cpu = get_cpu();
2346 __sched_fork(clone_flags, p);
2348 * We mark the process as NEW here. This guarantees that
2349 * nobody will actually run it, and a signal or other external
2350 * event cannot wake it up and insert it on the runqueue either.
2352 p->state = TASK_NEW;
2355 * Make sure we do not leak PI boosting priority to the child.
2357 p->prio = current->normal_prio;
2360 * Revert to default priority/policy on fork if requested.
2362 if (unlikely(p->sched_reset_on_fork)) {
2363 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2364 p->policy = SCHED_NORMAL;
2365 p->static_prio = NICE_TO_PRIO(0);
2367 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2368 p->static_prio = NICE_TO_PRIO(0);
2370 p->prio = p->normal_prio = __normal_prio(p);
2374 * We don't need the reset flag anymore after the fork. It has
2375 * fulfilled its duty:
2377 p->sched_reset_on_fork = 0;
2380 if (dl_prio(p->prio)) {
2383 } else if (rt_prio(p->prio)) {
2384 p->sched_class = &rt_sched_class;
2386 p->sched_class = &fair_sched_class;
2389 init_entity_runnable_average(&p->se);
2392 * The child is not yet in the pid-hash so no cgroup attach races,
2393 * and the cgroup is pinned to this child due to cgroup_fork()
2394 * is ran before sched_fork().
2396 * Silence PROVE_RCU.
2398 raw_spin_lock_irqsave(&p->pi_lock, flags);
2400 * We're setting the CPU for the first time, we don't migrate,
2401 * so use __set_task_cpu().
2403 __set_task_cpu(p, cpu);
2404 if (p->sched_class->task_fork)
2405 p->sched_class->task_fork(p);
2406 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2408 #ifdef CONFIG_SCHED_INFO
2409 if (likely(sched_info_on()))
2410 memset(&p->sched_info, 0, sizeof(p->sched_info));
2412 #if defined(CONFIG_SMP)
2415 init_task_preempt_count(p);
2417 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2418 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2425 unsigned long to_ratio(u64 period, u64 runtime)
2427 if (runtime == RUNTIME_INF)
2431 * Doing this here saves a lot of checks in all
2432 * the calling paths, and returning zero seems
2433 * safe for them anyway.
2438 return div64_u64(runtime << 20, period);
2442 inline struct dl_bw *dl_bw_of(int i)
2444 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2445 "sched RCU must be held");
2446 return &cpu_rq(i)->rd->dl_bw;
2449 static inline int dl_bw_cpus(int i)
2451 struct root_domain *rd = cpu_rq(i)->rd;
2454 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2455 "sched RCU must be held");
2456 for_each_cpu_and(i, rd->span, cpu_active_mask)
2462 inline struct dl_bw *dl_bw_of(int i)
2464 return &cpu_rq(i)->dl.dl_bw;
2467 static inline int dl_bw_cpus(int i)
2474 * We must be sure that accepting a new task (or allowing changing the
2475 * parameters of an existing one) is consistent with the bandwidth
2476 * constraints. If yes, this function also accordingly updates the currently
2477 * allocated bandwidth to reflect the new situation.
2479 * This function is called while holding p's rq->lock.
2481 * XXX we should delay bw change until the task's 0-lag point, see
2484 static int dl_overflow(struct task_struct *p, int policy,
2485 const struct sched_attr *attr)
2488 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2489 u64 period = attr->sched_period ?: attr->sched_deadline;
2490 u64 runtime = attr->sched_runtime;
2491 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2494 /* !deadline task may carry old deadline bandwidth */
2495 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2499 * Either if a task, enters, leave, or stays -deadline but changes
2500 * its parameters, we may need to update accordingly the total
2501 * allocated bandwidth of the container.
2503 raw_spin_lock(&dl_b->lock);
2504 cpus = dl_bw_cpus(task_cpu(p));
2505 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2506 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2507 __dl_add(dl_b, new_bw);
2509 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2510 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2511 __dl_clear(dl_b, p->dl.dl_bw);
2512 __dl_add(dl_b, new_bw);
2514 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2515 __dl_clear(dl_b, p->dl.dl_bw);
2518 raw_spin_unlock(&dl_b->lock);
2523 extern void init_dl_bw(struct dl_bw *dl_b);
2526 * wake_up_new_task - wake up a newly created task for the first time.
2528 * This function will do some initial scheduler statistics housekeeping
2529 * that must be done for every newly created context, then puts the task
2530 * on the runqueue and wakes it.
2532 void wake_up_new_task(struct task_struct *p)
2537 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2538 p->state = TASK_RUNNING;
2541 * Fork balancing, do it here and not earlier because:
2542 * - cpus_allowed can change in the fork path
2543 * - any previously selected CPU might disappear through hotplug
2545 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2546 * as we're not fully set-up yet.
2548 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2550 rq = __task_rq_lock(p, &rf);
2551 update_rq_clock(rq);
2552 post_init_entity_util_avg(&p->se);
2554 activate_task(rq, p, 0);
2555 p->on_rq = TASK_ON_RQ_QUEUED;
2556 trace_sched_wakeup_new(p);
2557 check_preempt_curr(rq, p, WF_FORK);
2559 if (p->sched_class->task_woken) {
2561 * Nothing relies on rq->lock after this, so its fine to
2564 rq_unpin_lock(rq, &rf);
2565 p->sched_class->task_woken(rq, p);
2566 rq_repin_lock(rq, &rf);
2569 task_rq_unlock(rq, p, &rf);
2572 #ifdef CONFIG_PREEMPT_NOTIFIERS
2574 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2576 void preempt_notifier_inc(void)
2578 static_key_slow_inc(&preempt_notifier_key);
2580 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2582 void preempt_notifier_dec(void)
2584 static_key_slow_dec(&preempt_notifier_key);
2586 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2589 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2590 * @notifier: notifier struct to register
2592 void preempt_notifier_register(struct preempt_notifier *notifier)
2594 if (!static_key_false(&preempt_notifier_key))
2595 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2597 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2599 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2602 * preempt_notifier_unregister - no longer interested in preemption notifications
2603 * @notifier: notifier struct to unregister
2605 * This is *not* safe to call from within a preemption notifier.
2607 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2609 hlist_del(¬ifier->link);
2611 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2613 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2615 struct preempt_notifier *notifier;
2617 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2618 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2621 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2623 if (static_key_false(&preempt_notifier_key))
2624 __fire_sched_in_preempt_notifiers(curr);
2628 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2629 struct task_struct *next)
2631 struct preempt_notifier *notifier;
2633 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2634 notifier->ops->sched_out(notifier, next);
2637 static __always_inline void
2638 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2639 struct task_struct *next)
2641 if (static_key_false(&preempt_notifier_key))
2642 __fire_sched_out_preempt_notifiers(curr, next);
2645 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2647 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2652 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2653 struct task_struct *next)
2657 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2660 * prepare_task_switch - prepare to switch tasks
2661 * @rq: the runqueue preparing to switch
2662 * @prev: the current task that is being switched out
2663 * @next: the task we are going to switch to.
2665 * This is called with the rq lock held and interrupts off. It must
2666 * be paired with a subsequent finish_task_switch after the context
2669 * prepare_task_switch sets up locking and calls architecture specific
2673 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2674 struct task_struct *next)
2676 sched_info_switch(rq, prev, next);
2677 perf_event_task_sched_out(prev, next);
2678 fire_sched_out_preempt_notifiers(prev, next);
2679 prepare_lock_switch(rq, next);
2680 prepare_arch_switch(next);
2684 * finish_task_switch - clean up after a task-switch
2685 * @prev: the thread we just switched away from.
2687 * finish_task_switch must be called after the context switch, paired
2688 * with a prepare_task_switch call before the context switch.
2689 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2690 * and do any other architecture-specific cleanup actions.
2692 * Note that we may have delayed dropping an mm in context_switch(). If
2693 * so, we finish that here outside of the runqueue lock. (Doing it
2694 * with the lock held can cause deadlocks; see schedule() for
2697 * The context switch have flipped the stack from under us and restored the
2698 * local variables which were saved when this task called schedule() in the
2699 * past. prev == current is still correct but we need to recalculate this_rq
2700 * because prev may have moved to another CPU.
2702 static struct rq *finish_task_switch(struct task_struct *prev)
2703 __releases(rq->lock)
2705 struct rq *rq = this_rq();
2706 struct mm_struct *mm = rq->prev_mm;
2710 * The previous task will have left us with a preempt_count of 2
2711 * because it left us after:
2714 * preempt_disable(); // 1
2716 * raw_spin_lock_irq(&rq->lock) // 2
2718 * Also, see FORK_PREEMPT_COUNT.
2720 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2721 "corrupted preempt_count: %s/%d/0x%x\n",
2722 current->comm, current->pid, preempt_count()))
2723 preempt_count_set(FORK_PREEMPT_COUNT);
2728 * A task struct has one reference for the use as "current".
2729 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2730 * schedule one last time. The schedule call will never return, and
2731 * the scheduled task must drop that reference.
2733 * We must observe prev->state before clearing prev->on_cpu (in
2734 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2735 * running on another CPU and we could rave with its RUNNING -> DEAD
2736 * transition, resulting in a double drop.
2738 prev_state = prev->state;
2739 vtime_task_switch(prev);
2740 perf_event_task_sched_in(prev, current);
2741 finish_lock_switch(rq, prev);
2742 finish_arch_post_lock_switch();
2744 fire_sched_in_preempt_notifiers(current);
2747 if (unlikely(prev_state == TASK_DEAD)) {
2748 if (prev->sched_class->task_dead)
2749 prev->sched_class->task_dead(prev);
2752 * Remove function-return probe instances associated with this
2753 * task and put them back on the free list.
2755 kprobe_flush_task(prev);
2757 /* Task is done with its stack. */
2758 put_task_stack(prev);
2760 put_task_struct(prev);
2763 tick_nohz_task_switch();
2769 /* rq->lock is NOT held, but preemption is disabled */
2770 static void __balance_callback(struct rq *rq)
2772 struct callback_head *head, *next;
2773 void (*func)(struct rq *rq);
2774 unsigned long flags;
2776 raw_spin_lock_irqsave(&rq->lock, flags);
2777 head = rq->balance_callback;
2778 rq->balance_callback = NULL;
2780 func = (void (*)(struct rq *))head->func;
2787 raw_spin_unlock_irqrestore(&rq->lock, flags);
2790 static inline void balance_callback(struct rq *rq)
2792 if (unlikely(rq->balance_callback))
2793 __balance_callback(rq);
2798 static inline void balance_callback(struct rq *rq)
2805 * schedule_tail - first thing a freshly forked thread must call.
2806 * @prev: the thread we just switched away from.
2808 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2809 __releases(rq->lock)
2814 * New tasks start with FORK_PREEMPT_COUNT, see there and
2815 * finish_task_switch() for details.
2817 * finish_task_switch() will drop rq->lock() and lower preempt_count
2818 * and the preempt_enable() will end up enabling preemption (on
2819 * PREEMPT_COUNT kernels).
2822 rq = finish_task_switch(prev);
2823 balance_callback(rq);
2826 if (current->set_child_tid)
2827 put_user(task_pid_vnr(current), current->set_child_tid);
2831 * context_switch - switch to the new MM and the new thread's register state.
2833 static __always_inline struct rq *
2834 context_switch(struct rq *rq, struct task_struct *prev,
2835 struct task_struct *next, struct rq_flags *rf)
2837 struct mm_struct *mm, *oldmm;
2839 prepare_task_switch(rq, prev, next);
2842 oldmm = prev->active_mm;
2844 * For paravirt, this is coupled with an exit in switch_to to
2845 * combine the page table reload and the switch backend into
2848 arch_start_context_switch(prev);
2851 next->active_mm = oldmm;
2853 enter_lazy_tlb(oldmm, next);
2855 switch_mm_irqs_off(oldmm, mm, next);
2858 prev->active_mm = NULL;
2859 rq->prev_mm = oldmm;
2862 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
2865 * Since the runqueue lock will be released by the next
2866 * task (which is an invalid locking op but in the case
2867 * of the scheduler it's an obvious special-case), so we
2868 * do an early lockdep release here:
2870 rq_unpin_lock(rq, rf);
2871 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2873 /* Here we just switch the register state and the stack. */
2874 switch_to(prev, next, prev);
2877 return finish_task_switch(prev);
2881 * nr_running and nr_context_switches:
2883 * externally visible scheduler statistics: current number of runnable
2884 * threads, total number of context switches performed since bootup.
2886 unsigned long nr_running(void)
2888 unsigned long i, sum = 0;
2890 for_each_online_cpu(i)
2891 sum += cpu_rq(i)->nr_running;
2897 * Check if only the current task is running on the CPU.
2899 * Caution: this function does not check that the caller has disabled
2900 * preemption, thus the result might have a time-of-check-to-time-of-use
2901 * race. The caller is responsible to use it correctly, for example:
2903 * - from a non-preemptable section (of course)
2905 * - from a thread that is bound to a single CPU
2907 * - in a loop with very short iterations (e.g. a polling loop)
2909 bool single_task_running(void)
2911 return raw_rq()->nr_running == 1;
2913 EXPORT_SYMBOL(single_task_running);
2915 unsigned long long nr_context_switches(void)
2918 unsigned long long sum = 0;
2920 for_each_possible_cpu(i)
2921 sum += cpu_rq(i)->nr_switches;
2927 * IO-wait accounting, and how its mostly bollocks (on SMP).
2929 * The idea behind IO-wait account is to account the idle time that we could
2930 * have spend running if it were not for IO. That is, if we were to improve the
2931 * storage performance, we'd have a proportional reduction in IO-wait time.
2933 * This all works nicely on UP, where, when a task blocks on IO, we account
2934 * idle time as IO-wait, because if the storage were faster, it could've been
2935 * running and we'd not be idle.
2937 * This has been extended to SMP, by doing the same for each CPU. This however
2940 * Imagine for instance the case where two tasks block on one CPU, only the one
2941 * CPU will have IO-wait accounted, while the other has regular idle. Even
2942 * though, if the storage were faster, both could've ran at the same time,
2943 * utilising both CPUs.
2945 * This means, that when looking globally, the current IO-wait accounting on
2946 * SMP is a lower bound, by reason of under accounting.
2948 * Worse, since the numbers are provided per CPU, they are sometimes
2949 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2950 * associated with any one particular CPU, it can wake to another CPU than it
2951 * blocked on. This means the per CPU IO-wait number is meaningless.
2953 * Task CPU affinities can make all that even more 'interesting'.
2956 unsigned long nr_iowait(void)
2958 unsigned long i, sum = 0;
2960 for_each_possible_cpu(i)
2961 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2967 * Consumers of these two interfaces, like for example the cpufreq menu
2968 * governor are using nonsensical data. Boosting frequency for a CPU that has
2969 * IO-wait which might not even end up running the task when it does become
2973 unsigned long nr_iowait_cpu(int cpu)
2975 struct rq *this = cpu_rq(cpu);
2976 return atomic_read(&this->nr_iowait);
2979 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2981 struct rq *rq = this_rq();
2982 *nr_waiters = atomic_read(&rq->nr_iowait);
2983 *load = rq->load.weight;
2989 * sched_exec - execve() is a valuable balancing opportunity, because at
2990 * this point the task has the smallest effective memory and cache footprint.
2992 void sched_exec(void)
2994 struct task_struct *p = current;
2995 unsigned long flags;
2998 raw_spin_lock_irqsave(&p->pi_lock, flags);
2999 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
3000 if (dest_cpu == smp_processor_id())
3003 if (likely(cpu_active(dest_cpu))) {
3004 struct migration_arg arg = { p, dest_cpu };
3006 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3007 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3011 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3016 DEFINE_PER_CPU(struct kernel_stat, kstat);
3017 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
3019 EXPORT_PER_CPU_SYMBOL(kstat);
3020 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3023 * The function fair_sched_class.update_curr accesses the struct curr
3024 * and its field curr->exec_start; when called from task_sched_runtime(),
3025 * we observe a high rate of cache misses in practice.
3026 * Prefetching this data results in improved performance.
3028 static inline void prefetch_curr_exec_start(struct task_struct *p)
3030 #ifdef CONFIG_FAIR_GROUP_SCHED
3031 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3033 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3036 prefetch(&curr->exec_start);
3040 * Return accounted runtime for the task.
3041 * In case the task is currently running, return the runtime plus current's
3042 * pending runtime that have not been accounted yet.
3044 unsigned long long task_sched_runtime(struct task_struct *p)
3050 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3052 * 64-bit doesn't need locks to atomically read a 64bit value.
3053 * So we have a optimization chance when the task's delta_exec is 0.
3054 * Reading ->on_cpu is racy, but this is ok.
3056 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3057 * If we race with it entering CPU, unaccounted time is 0. This is
3058 * indistinguishable from the read occurring a few cycles earlier.
3059 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3060 * been accounted, so we're correct here as well.
3062 if (!p->on_cpu || !task_on_rq_queued(p))
3063 return p->se.sum_exec_runtime;
3066 rq = task_rq_lock(p, &rf);
3068 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3069 * project cycles that may never be accounted to this
3070 * thread, breaking clock_gettime().
3072 if (task_current(rq, p) && task_on_rq_queued(p)) {
3073 prefetch_curr_exec_start(p);
3074 update_rq_clock(rq);
3075 p->sched_class->update_curr(rq);
3077 ns = p->se.sum_exec_runtime;
3078 task_rq_unlock(rq, p, &rf);
3084 * This function gets called by the timer code, with HZ frequency.
3085 * We call it with interrupts disabled.
3087 void scheduler_tick(void)
3089 int cpu = smp_processor_id();
3090 struct rq *rq = cpu_rq(cpu);
3091 struct task_struct *curr = rq->curr;
3095 raw_spin_lock(&rq->lock);
3096 update_rq_clock(rq);
3097 curr->sched_class->task_tick(rq, curr, 0);
3098 cpu_load_update_active(rq);
3099 calc_global_load_tick(rq);
3100 raw_spin_unlock(&rq->lock);
3102 perf_event_task_tick();
3105 rq->idle_balance = idle_cpu(cpu);
3106 trigger_load_balance(rq);
3108 rq_last_tick_reset(rq);
3111 #ifdef CONFIG_NO_HZ_FULL
3113 * scheduler_tick_max_deferment
3115 * Keep at least one tick per second when a single
3116 * active task is running because the scheduler doesn't
3117 * yet completely support full dynticks environment.
3119 * This makes sure that uptime, CFS vruntime, load
3120 * balancing, etc... continue to move forward, even
3121 * with a very low granularity.
3123 * Return: Maximum deferment in nanoseconds.
3125 u64 scheduler_tick_max_deferment(void)
3127 struct rq *rq = this_rq();
3128 unsigned long next, now = READ_ONCE(jiffies);
3130 next = rq->last_sched_tick + HZ;
3132 if (time_before_eq(next, now))
3135 return jiffies_to_nsecs(next - now);
3139 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3140 defined(CONFIG_PREEMPT_TRACER))
3142 * If the value passed in is equal to the current preempt count
3143 * then we just disabled preemption. Start timing the latency.
3145 static inline void preempt_latency_start(int val)
3147 if (preempt_count() == val) {
3148 unsigned long ip = get_lock_parent_ip();
3149 #ifdef CONFIG_DEBUG_PREEMPT
3150 current->preempt_disable_ip = ip;
3152 trace_preempt_off(CALLER_ADDR0, ip);
3156 void preempt_count_add(int val)
3158 #ifdef CONFIG_DEBUG_PREEMPT
3162 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3165 __preempt_count_add(val);
3166 #ifdef CONFIG_DEBUG_PREEMPT
3168 * Spinlock count overflowing soon?
3170 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3173 preempt_latency_start(val);
3175 EXPORT_SYMBOL(preempt_count_add);
3176 NOKPROBE_SYMBOL(preempt_count_add);
3179 * If the value passed in equals to the current preempt count
3180 * then we just enabled preemption. Stop timing the latency.
3182 static inline void preempt_latency_stop(int val)
3184 if (preempt_count() == val)
3185 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3188 void preempt_count_sub(int val)
3190 #ifdef CONFIG_DEBUG_PREEMPT
3194 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3197 * Is the spinlock portion underflowing?
3199 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3200 !(preempt_count() & PREEMPT_MASK)))
3204 preempt_latency_stop(val);
3205 __preempt_count_sub(val);
3207 EXPORT_SYMBOL(preempt_count_sub);
3208 NOKPROBE_SYMBOL(preempt_count_sub);
3211 static inline void preempt_latency_start(int val) { }
3212 static inline void preempt_latency_stop(int val) { }
3215 static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
3217 #ifdef CONFIG_DEBUG_PREEMPT
3218 return p->preempt_disable_ip;
3225 * Print scheduling while atomic bug:
3227 static noinline void __schedule_bug(struct task_struct *prev)
3229 /* Save this before calling printk(), since that will clobber it */
3230 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3232 if (oops_in_progress)
3235 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3236 prev->comm, prev->pid, preempt_count());
3238 debug_show_held_locks(prev);
3240 if (irqs_disabled())
3241 print_irqtrace_events(prev);
3242 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3243 && in_atomic_preempt_off()) {
3244 pr_err("Preemption disabled at:");
3245 print_ip_sym(preempt_disable_ip);
3249 panic("scheduling while atomic\n");
3252 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3256 * Various schedule()-time debugging checks and statistics:
3258 static inline void schedule_debug(struct task_struct *prev)
3260 #ifdef CONFIG_SCHED_STACK_END_CHECK
3261 if (task_stack_end_corrupted(prev))
3262 panic("corrupted stack end detected inside scheduler\n");
3265 if (unlikely(in_atomic_preempt_off())) {
3266 __schedule_bug(prev);
3267 preempt_count_set(PREEMPT_DISABLED);
3271 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3273 schedstat_inc(this_rq()->sched_count);
3277 * Pick up the highest-prio task:
3279 static inline struct task_struct *
3280 pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
3282 const struct sched_class *class;
3283 struct task_struct *p;
3286 * Optimization: we know that if all tasks are in
3287 * the fair class we can call that function directly:
3289 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3290 p = fair_sched_class.pick_next_task(rq, prev, rf);
3291 if (unlikely(p == RETRY_TASK))
3294 /* Assumes fair_sched_class->next == idle_sched_class */
3296 p = idle_sched_class.pick_next_task(rq, prev, rf);
3302 for_each_class(class) {
3303 p = class->pick_next_task(rq, prev, rf);
3305 if (unlikely(p == RETRY_TASK))
3311 /* The idle class should always have a runnable task: */
3316 * __schedule() is the main scheduler function.
3318 * The main means of driving the scheduler and thus entering this function are:
3320 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3322 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3323 * paths. For example, see arch/x86/entry_64.S.
3325 * To drive preemption between tasks, the scheduler sets the flag in timer
3326 * interrupt handler scheduler_tick().
3328 * 3. Wakeups don't really cause entry into schedule(). They add a
3329 * task to the run-queue and that's it.
3331 * Now, if the new task added to the run-queue preempts the current
3332 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3333 * called on the nearest possible occasion:
3335 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3337 * - in syscall or exception context, at the next outmost
3338 * preempt_enable(). (this might be as soon as the wake_up()'s
3341 * - in IRQ context, return from interrupt-handler to
3342 * preemptible context
3344 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3347 * - cond_resched() call
3348 * - explicit schedule() call
3349 * - return from syscall or exception to user-space
3350 * - return from interrupt-handler to user-space
3352 * WARNING: must be called with preemption disabled!
3354 static void __sched notrace __schedule(bool preempt)
3356 struct task_struct *prev, *next;
3357 unsigned long *switch_count;
3362 cpu = smp_processor_id();
3366 schedule_debug(prev);
3368 if (sched_feat(HRTICK))
3371 local_irq_disable();
3372 rcu_note_context_switch();
3375 * Make sure that signal_pending_state()->signal_pending() below
3376 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3377 * done by the caller to avoid the race with signal_wake_up().
3379 smp_mb__before_spinlock();
3380 raw_spin_lock(&rq->lock);
3381 rq_pin_lock(rq, &rf);
3383 /* Promote REQ to ACT */
3384 rq->clock_update_flags <<= 1;
3386 switch_count = &prev->nivcsw;
3387 if (!preempt && prev->state) {
3388 if (unlikely(signal_pending_state(prev->state, prev))) {
3389 prev->state = TASK_RUNNING;
3391 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3394 if (prev->in_iowait) {
3395 atomic_inc(&rq->nr_iowait);
3396 delayacct_blkio_start();
3400 * If a worker went to sleep, notify and ask workqueue
3401 * whether it wants to wake up a task to maintain
3404 if (prev->flags & PF_WQ_WORKER) {
3405 struct task_struct *to_wakeup;
3407 to_wakeup = wq_worker_sleeping(prev);
3409 try_to_wake_up_local(to_wakeup, &rf);
3412 switch_count = &prev->nvcsw;
3415 if (task_on_rq_queued(prev))
3416 update_rq_clock(rq);
3418 next = pick_next_task(rq, prev, &rf);
3419 clear_tsk_need_resched(prev);
3420 clear_preempt_need_resched();
3422 if (likely(prev != next)) {
3427 trace_sched_switch(preempt, prev, next);
3429 /* Also unlocks the rq: */
3430 rq = context_switch(rq, prev, next, &rf);
3432 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
3433 rq_unpin_lock(rq, &rf);
3434 raw_spin_unlock_irq(&rq->lock);
3437 balance_callback(rq);
3440 void __noreturn do_task_dead(void)
3443 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3444 * when the following two conditions become true.
3445 * - There is race condition of mmap_sem (It is acquired by
3447 * - SMI occurs before setting TASK_RUNINNG.
3448 * (or hypervisor of virtual machine switches to other guest)
3449 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3451 * To avoid it, we have to wait for releasing tsk->pi_lock which
3452 * is held by try_to_wake_up()
3455 raw_spin_unlock_wait(¤t->pi_lock);
3457 /* Causes final put_task_struct in finish_task_switch(): */
3458 __set_current_state(TASK_DEAD);
3460 /* Tell freezer to ignore us: */
3461 current->flags |= PF_NOFREEZE;
3466 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3471 static inline void sched_submit_work(struct task_struct *tsk)
3473 if (!tsk->state || tsk_is_pi_blocked(tsk))
3476 * If we are going to sleep and we have plugged IO queued,
3477 * make sure to submit it to avoid deadlocks.
3479 if (blk_needs_flush_plug(tsk))
3480 blk_schedule_flush_plug(tsk);
3483 asmlinkage __visible void __sched schedule(void)
3485 struct task_struct *tsk = current;
3487 sched_submit_work(tsk);
3491 sched_preempt_enable_no_resched();
3492 } while (need_resched());
3494 EXPORT_SYMBOL(schedule);
3496 #ifdef CONFIG_CONTEXT_TRACKING
3497 asmlinkage __visible void __sched schedule_user(void)
3500 * If we come here after a random call to set_need_resched(),
3501 * or we have been woken up remotely but the IPI has not yet arrived,
3502 * we haven't yet exited the RCU idle mode. Do it here manually until
3503 * we find a better solution.
3505 * NB: There are buggy callers of this function. Ideally we
3506 * should warn if prev_state != CONTEXT_USER, but that will trigger
3507 * too frequently to make sense yet.
3509 enum ctx_state prev_state = exception_enter();
3511 exception_exit(prev_state);
3516 * schedule_preempt_disabled - called with preemption disabled
3518 * Returns with preemption disabled. Note: preempt_count must be 1
3520 void __sched schedule_preempt_disabled(void)
3522 sched_preempt_enable_no_resched();
3527 static void __sched notrace preempt_schedule_common(void)
3531 * Because the function tracer can trace preempt_count_sub()
3532 * and it also uses preempt_enable/disable_notrace(), if
3533 * NEED_RESCHED is set, the preempt_enable_notrace() called
3534 * by the function tracer will call this function again and
3535 * cause infinite recursion.
3537 * Preemption must be disabled here before the function
3538 * tracer can trace. Break up preempt_disable() into two
3539 * calls. One to disable preemption without fear of being
3540 * traced. The other to still record the preemption latency,
3541 * which can also be traced by the function tracer.
3543 preempt_disable_notrace();
3544 preempt_latency_start(1);
3546 preempt_latency_stop(1);
3547 preempt_enable_no_resched_notrace();
3550 * Check again in case we missed a preemption opportunity
3551 * between schedule and now.
3553 } while (need_resched());
3556 #ifdef CONFIG_PREEMPT
3558 * this is the entry point to schedule() from in-kernel preemption
3559 * off of preempt_enable. Kernel preemptions off return from interrupt
3560 * occur there and call schedule directly.
3562 asmlinkage __visible void __sched notrace preempt_schedule(void)
3565 * If there is a non-zero preempt_count or interrupts are disabled,
3566 * we do not want to preempt the current task. Just return..
3568 if (likely(!preemptible()))
3571 preempt_schedule_common();
3573 NOKPROBE_SYMBOL(preempt_schedule);
3574 EXPORT_SYMBOL(preempt_schedule);
3577 * preempt_schedule_notrace - preempt_schedule called by tracing
3579 * The tracing infrastructure uses preempt_enable_notrace to prevent
3580 * recursion and tracing preempt enabling caused by the tracing
3581 * infrastructure itself. But as tracing can happen in areas coming
3582 * from userspace or just about to enter userspace, a preempt enable
3583 * can occur before user_exit() is called. This will cause the scheduler
3584 * to be called when the system is still in usermode.
3586 * To prevent this, the preempt_enable_notrace will use this function
3587 * instead of preempt_schedule() to exit user context if needed before
3588 * calling the scheduler.
3590 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3592 enum ctx_state prev_ctx;
3594 if (likely(!preemptible()))
3599 * Because the function tracer can trace preempt_count_sub()
3600 * and it also uses preempt_enable/disable_notrace(), if
3601 * NEED_RESCHED is set, the preempt_enable_notrace() called
3602 * by the function tracer will call this function again and
3603 * cause infinite recursion.
3605 * Preemption must be disabled here before the function
3606 * tracer can trace. Break up preempt_disable() into two
3607 * calls. One to disable preemption without fear of being
3608 * traced. The other to still record the preemption latency,
3609 * which can also be traced by the function tracer.
3611 preempt_disable_notrace();
3612 preempt_latency_start(1);
3614 * Needs preempt disabled in case user_exit() is traced
3615 * and the tracer calls preempt_enable_notrace() causing
3616 * an infinite recursion.
3618 prev_ctx = exception_enter();
3620 exception_exit(prev_ctx);
3622 preempt_latency_stop(1);
3623 preempt_enable_no_resched_notrace();
3624 } while (need_resched());
3626 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3628 #endif /* CONFIG_PREEMPT */
3631 * this is the entry point to schedule() from kernel preemption
3632 * off of irq context.
3633 * Note, that this is called and return with irqs disabled. This will
3634 * protect us against recursive calling from irq.
3636 asmlinkage __visible void __sched preempt_schedule_irq(void)
3638 enum ctx_state prev_state;
3640 /* Catch callers which need to be fixed */
3641 BUG_ON(preempt_count() || !irqs_disabled());
3643 prev_state = exception_enter();
3649 local_irq_disable();
3650 sched_preempt_enable_no_resched();
3651 } while (need_resched());
3653 exception_exit(prev_state);
3656 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3659 return try_to_wake_up(curr->private, mode, wake_flags);
3661 EXPORT_SYMBOL(default_wake_function);
3663 #ifdef CONFIG_RT_MUTEXES
3666 * rt_mutex_setprio - set the current priority of a task
3668 * @prio: prio value (kernel-internal form)
3670 * This function changes the 'effective' priority of a task. It does
3671 * not touch ->normal_prio like __setscheduler().
3673 * Used by the rt_mutex code to implement priority inheritance
3674 * logic. Call site only calls if the priority of the task changed.
3676 void rt_mutex_setprio(struct task_struct *p, int prio)
3678 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3679 const struct sched_class *prev_class;
3683 BUG_ON(prio > MAX_PRIO);
3685 rq = __task_rq_lock(p, &rf);
3686 update_rq_clock(rq);
3689 * Idle task boosting is a nono in general. There is one
3690 * exception, when PREEMPT_RT and NOHZ is active:
3692 * The idle task calls get_next_timer_interrupt() and holds
3693 * the timer wheel base->lock on the CPU and another CPU wants
3694 * to access the timer (probably to cancel it). We can safely
3695 * ignore the boosting request, as the idle CPU runs this code
3696 * with interrupts disabled and will complete the lock
3697 * protected section without being interrupted. So there is no
3698 * real need to boost.
3700 if (unlikely(p == rq->idle)) {
3701 WARN_ON(p != rq->curr);
3702 WARN_ON(p->pi_blocked_on);
3706 trace_sched_pi_setprio(p, prio);
3709 if (oldprio == prio)
3710 queue_flag &= ~DEQUEUE_MOVE;
3712 prev_class = p->sched_class;
3713 queued = task_on_rq_queued(p);
3714 running = task_current(rq, p);
3716 dequeue_task(rq, p, queue_flag);
3718 put_prev_task(rq, p);
3721 * Boosting condition are:
3722 * 1. -rt task is running and holds mutex A
3723 * --> -dl task blocks on mutex A
3725 * 2. -dl task is running and holds mutex A
3726 * --> -dl task blocks on mutex A and could preempt the
3729 if (dl_prio(prio)) {
3730 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3731 if (!dl_prio(p->normal_prio) ||
3732 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3733 p->dl.dl_boosted = 1;
3734 queue_flag |= ENQUEUE_REPLENISH;
3736 p->dl.dl_boosted = 0;
3737 p->sched_class = &dl_sched_class;
3738 } else if (rt_prio(prio)) {
3739 if (dl_prio(oldprio))
3740 p->dl.dl_boosted = 0;
3742 queue_flag |= ENQUEUE_HEAD;
3743 p->sched_class = &rt_sched_class;
3745 if (dl_prio(oldprio))
3746 p->dl.dl_boosted = 0;
3747 if (rt_prio(oldprio))
3749 p->sched_class = &fair_sched_class;
3755 enqueue_task(rq, p, queue_flag);
3757 set_curr_task(rq, p);
3759 check_class_changed(rq, p, prev_class, oldprio);
3761 /* Avoid rq from going away on us: */
3763 __task_rq_unlock(rq, &rf);
3765 balance_callback(rq);
3770 void set_user_nice(struct task_struct *p, long nice)
3772 bool queued, running;
3773 int old_prio, delta;
3777 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3780 * We have to be careful, if called from sys_setpriority(),
3781 * the task might be in the middle of scheduling on another CPU.
3783 rq = task_rq_lock(p, &rf);
3784 update_rq_clock(rq);
3787 * The RT priorities are set via sched_setscheduler(), but we still
3788 * allow the 'normal' nice value to be set - but as expected
3789 * it wont have any effect on scheduling until the task is
3790 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3792 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3793 p->static_prio = NICE_TO_PRIO(nice);
3796 queued = task_on_rq_queued(p);
3797 running = task_current(rq, p);
3799 dequeue_task(rq, p, DEQUEUE_SAVE);
3801 put_prev_task(rq, p);
3803 p->static_prio = NICE_TO_PRIO(nice);
3806 p->prio = effective_prio(p);
3807 delta = p->prio - old_prio;
3810 enqueue_task(rq, p, ENQUEUE_RESTORE);
3812 * If the task increased its priority or is running and
3813 * lowered its priority, then reschedule its CPU:
3815 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3819 set_curr_task(rq, p);
3821 task_rq_unlock(rq, p, &rf);
3823 EXPORT_SYMBOL(set_user_nice);
3826 * can_nice - check if a task can reduce its nice value
3830 int can_nice(const struct task_struct *p, const int nice)
3832 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3833 int nice_rlim = nice_to_rlimit(nice);
3835 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3836 capable(CAP_SYS_NICE));
3839 #ifdef __ARCH_WANT_SYS_NICE
3842 * sys_nice - change the priority of the current process.
3843 * @increment: priority increment
3845 * sys_setpriority is a more generic, but much slower function that
3846 * does similar things.
3848 SYSCALL_DEFINE1(nice, int, increment)
3853 * Setpriority might change our priority at the same moment.
3854 * We don't have to worry. Conceptually one call occurs first
3855 * and we have a single winner.
3857 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3858 nice = task_nice(current) + increment;
3860 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3861 if (increment < 0 && !can_nice(current, nice))
3864 retval = security_task_setnice(current, nice);
3868 set_user_nice(current, nice);
3875 * task_prio - return the priority value of a given task.
3876 * @p: the task in question.
3878 * Return: The priority value as seen by users in /proc.
3879 * RT tasks are offset by -200. Normal tasks are centered
3880 * around 0, value goes from -16 to +15.
3882 int task_prio(const struct task_struct *p)
3884 return p->prio - MAX_RT_PRIO;
3888 * idle_cpu - is a given CPU idle currently?
3889 * @cpu: the processor in question.
3891 * Return: 1 if the CPU is currently idle. 0 otherwise.
3893 int idle_cpu(int cpu)
3895 struct rq *rq = cpu_rq(cpu);
3897 if (rq->curr != rq->idle)
3904 if (!llist_empty(&rq->wake_list))
3912 * idle_task - return the idle task for a given CPU.
3913 * @cpu: the processor in question.
3915 * Return: The idle task for the CPU @cpu.
3917 struct task_struct *idle_task(int cpu)
3919 return cpu_rq(cpu)->idle;
3923 * find_process_by_pid - find a process with a matching PID value.
3924 * @pid: the pid in question.
3926 * The task of @pid, if found. %NULL otherwise.
3928 static struct task_struct *find_process_by_pid(pid_t pid)
3930 return pid ? find_task_by_vpid(pid) : current;
3934 * This function initializes the sched_dl_entity of a newly becoming
3935 * SCHED_DEADLINE task.
3937 * Only the static values are considered here, the actual runtime and the
3938 * absolute deadline will be properly calculated when the task is enqueued
3939 * for the first time with its new policy.
3942 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3944 struct sched_dl_entity *dl_se = &p->dl;
3946 dl_se->dl_runtime = attr->sched_runtime;
3947 dl_se->dl_deadline = attr->sched_deadline;
3948 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3949 dl_se->flags = attr->sched_flags;
3950 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3953 * Changing the parameters of a task is 'tricky' and we're not doing
3954 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3956 * What we SHOULD do is delay the bandwidth release until the 0-lag
3957 * point. This would include retaining the task_struct until that time
3958 * and change dl_overflow() to not immediately decrement the current
3961 * Instead we retain the current runtime/deadline and let the new
3962 * parameters take effect after the current reservation period lapses.
3963 * This is safe (albeit pessimistic) because the 0-lag point is always
3964 * before the current scheduling deadline.
3966 * We can still have temporary overloads because we do not delay the
3967 * change in bandwidth until that time; so admission control is
3968 * not on the safe side. It does however guarantee tasks will never
3969 * consume more than promised.
3974 * sched_setparam() passes in -1 for its policy, to let the functions
3975 * it calls know not to change it.
3977 #define SETPARAM_POLICY -1
3979 static void __setscheduler_params(struct task_struct *p,
3980 const struct sched_attr *attr)
3982 int policy = attr->sched_policy;
3984 if (policy == SETPARAM_POLICY)
3989 if (dl_policy(policy))
3990 __setparam_dl(p, attr);
3991 else if (fair_policy(policy))
3992 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3995 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3996 * !rt_policy. Always setting this ensures that things like
3997 * getparam()/getattr() don't report silly values for !rt tasks.
3999 p->rt_priority = attr->sched_priority;
4000 p->normal_prio = normal_prio(p);
4004 /* Actually do priority change: must hold pi & rq lock. */
4005 static void __setscheduler(struct rq *rq, struct task_struct *p,
4006 const struct sched_attr *attr, bool keep_boost)
4008 __setscheduler_params(p, attr);
4011 * Keep a potential priority boosting if called from
4012 * sched_setscheduler().
4015 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
4017 p->prio = normal_prio(p);
4019 if (dl_prio(p->prio))
4020 p->sched_class = &dl_sched_class;
4021 else if (rt_prio(p->prio))
4022 p->sched_class = &rt_sched_class;
4024 p->sched_class = &fair_sched_class;
4028 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
4030 struct sched_dl_entity *dl_se = &p->dl;
4032 attr->sched_priority = p->rt_priority;
4033 attr->sched_runtime = dl_se->dl_runtime;
4034 attr->sched_deadline = dl_se->dl_deadline;
4035 attr->sched_period = dl_se->dl_period;
4036 attr->sched_flags = dl_se->flags;
4040 * This function validates the new parameters of a -deadline task.
4041 * We ask for the deadline not being zero, and greater or equal
4042 * than the runtime, as well as the period of being zero or
4043 * greater than deadline. Furthermore, we have to be sure that
4044 * user parameters are above the internal resolution of 1us (we
4045 * check sched_runtime only since it is always the smaller one) and
4046 * below 2^63 ns (we have to check both sched_deadline and
4047 * sched_period, as the latter can be zero).
4050 __checkparam_dl(const struct sched_attr *attr)
4053 if (attr->sched_deadline == 0)
4057 * Since we truncate DL_SCALE bits, make sure we're at least
4060 if (attr->sched_runtime < (1ULL << DL_SCALE))
4064 * Since we use the MSB for wrap-around and sign issues, make
4065 * sure it's not set (mind that period can be equal to zero).
4067 if (attr->sched_deadline & (1ULL << 63) ||
4068 attr->sched_period & (1ULL << 63))
4071 /* runtime <= deadline <= period (if period != 0) */
4072 if ((attr->sched_period != 0 &&
4073 attr->sched_period < attr->sched_deadline) ||
4074 attr->sched_deadline < attr->sched_runtime)
4081 * Check the target process has a UID that matches the current process's:
4083 static bool check_same_owner(struct task_struct *p)
4085 const struct cred *cred = current_cred(), *pcred;
4089 pcred = __task_cred(p);
4090 match = (uid_eq(cred->euid, pcred->euid) ||
4091 uid_eq(cred->euid, pcred->uid));
4096 static bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
4098 struct sched_dl_entity *dl_se = &p->dl;
4100 if (dl_se->dl_runtime != attr->sched_runtime ||
4101 dl_se->dl_deadline != attr->sched_deadline ||
4102 dl_se->dl_period != attr->sched_period ||
4103 dl_se->flags != attr->sched_flags)
4109 static int __sched_setscheduler(struct task_struct *p,
4110 const struct sched_attr *attr,
4113 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4114 MAX_RT_PRIO - 1 - attr->sched_priority;
4115 int retval, oldprio, oldpolicy = -1, queued, running;
4116 int new_effective_prio, policy = attr->sched_policy;
4117 const struct sched_class *prev_class;
4120 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
4123 /* May grab non-irq protected spin_locks: */
4124 BUG_ON(in_interrupt());
4126 /* Double check policy once rq lock held: */
4128 reset_on_fork = p->sched_reset_on_fork;
4129 policy = oldpolicy = p->policy;
4131 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4133 if (!valid_policy(policy))
4137 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4141 * Valid priorities for SCHED_FIFO and SCHED_RR are
4142 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4143 * SCHED_BATCH and SCHED_IDLE is 0.
4145 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4146 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4148 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4149 (rt_policy(policy) != (attr->sched_priority != 0)))
4153 * Allow unprivileged RT tasks to decrease priority:
4155 if (user && !capable(CAP_SYS_NICE)) {
4156 if (fair_policy(policy)) {
4157 if (attr->sched_nice < task_nice(p) &&
4158 !can_nice(p, attr->sched_nice))
4162 if (rt_policy(policy)) {
4163 unsigned long rlim_rtprio =
4164 task_rlimit(p, RLIMIT_RTPRIO);
4166 /* Can't set/change the rt policy: */
4167 if (policy != p->policy && !rlim_rtprio)
4170 /* Can't increase priority: */
4171 if (attr->sched_priority > p->rt_priority &&
4172 attr->sched_priority > rlim_rtprio)
4177 * Can't set/change SCHED_DEADLINE policy at all for now
4178 * (safest behavior); in the future we would like to allow
4179 * unprivileged DL tasks to increase their relative deadline
4180 * or reduce their runtime (both ways reducing utilization)
4182 if (dl_policy(policy))
4186 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4187 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4189 if (idle_policy(p->policy) && !idle_policy(policy)) {
4190 if (!can_nice(p, task_nice(p)))
4194 /* Can't change other user's priorities: */
4195 if (!check_same_owner(p))
4198 /* Normal users shall not reset the sched_reset_on_fork flag: */
4199 if (p->sched_reset_on_fork && !reset_on_fork)
4204 retval = security_task_setscheduler(p);
4210 * Make sure no PI-waiters arrive (or leave) while we are
4211 * changing the priority of the task:
4213 * To be able to change p->policy safely, the appropriate
4214 * runqueue lock must be held.
4216 rq = task_rq_lock(p, &rf);
4217 update_rq_clock(rq);
4220 * Changing the policy of the stop threads its a very bad idea:
4222 if (p == rq->stop) {
4223 task_rq_unlock(rq, p, &rf);
4228 * If not changing anything there's no need to proceed further,
4229 * but store a possible modification of reset_on_fork.
4231 if (unlikely(policy == p->policy)) {
4232 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4234 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4236 if (dl_policy(policy) && dl_param_changed(p, attr))
4239 p->sched_reset_on_fork = reset_on_fork;
4240 task_rq_unlock(rq, p, &rf);
4246 #ifdef CONFIG_RT_GROUP_SCHED
4248 * Do not allow realtime tasks into groups that have no runtime
4251 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4252 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4253 !task_group_is_autogroup(task_group(p))) {
4254 task_rq_unlock(rq, p, &rf);
4259 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4260 cpumask_t *span = rq->rd->span;
4263 * Don't allow tasks with an affinity mask smaller than
4264 * the entire root_domain to become SCHED_DEADLINE. We
4265 * will also fail if there's no bandwidth available.
4267 if (!cpumask_subset(span, &p->cpus_allowed) ||
4268 rq->rd->dl_bw.bw == 0) {
4269 task_rq_unlock(rq, p, &rf);
4276 /* Re-check policy now with rq lock held: */
4277 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4278 policy = oldpolicy = -1;
4279 task_rq_unlock(rq, p, &rf);
4284 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4285 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4288 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4289 task_rq_unlock(rq, p, &rf);
4293 p->sched_reset_on_fork = reset_on_fork;
4298 * Take priority boosted tasks into account. If the new
4299 * effective priority is unchanged, we just store the new
4300 * normal parameters and do not touch the scheduler class and
4301 * the runqueue. This will be done when the task deboost
4304 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4305 if (new_effective_prio == oldprio)
4306 queue_flags &= ~DEQUEUE_MOVE;
4309 queued = task_on_rq_queued(p);
4310 running = task_current(rq, p);
4312 dequeue_task(rq, p, queue_flags);
4314 put_prev_task(rq, p);
4316 prev_class = p->sched_class;
4317 __setscheduler(rq, p, attr, pi);
4321 * We enqueue to tail when the priority of a task is
4322 * increased (user space view).
4324 if (oldprio < p->prio)
4325 queue_flags |= ENQUEUE_HEAD;
4327 enqueue_task(rq, p, queue_flags);
4330 set_curr_task(rq, p);
4332 check_class_changed(rq, p, prev_class, oldprio);
4334 /* Avoid rq from going away on us: */
4336 task_rq_unlock(rq, p, &rf);
4339 rt_mutex_adjust_pi(p);
4341 /* Run balance callbacks after we've adjusted the PI chain: */
4342 balance_callback(rq);
4348 static int _sched_setscheduler(struct task_struct *p, int policy,
4349 const struct sched_param *param, bool check)
4351 struct sched_attr attr = {
4352 .sched_policy = policy,
4353 .sched_priority = param->sched_priority,
4354 .sched_nice = PRIO_TO_NICE(p->static_prio),
4357 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4358 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4359 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4360 policy &= ~SCHED_RESET_ON_FORK;
4361 attr.sched_policy = policy;
4364 return __sched_setscheduler(p, &attr, check, true);
4367 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4368 * @p: the task in question.
4369 * @policy: new policy.
4370 * @param: structure containing the new RT priority.
4372 * Return: 0 on success. An error code otherwise.
4374 * NOTE that the task may be already dead.
4376 int sched_setscheduler(struct task_struct *p, int policy,
4377 const struct sched_param *param)
4379 return _sched_setscheduler(p, policy, param, true);
4381 EXPORT_SYMBOL_GPL(sched_setscheduler);
4383 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4385 return __sched_setscheduler(p, attr, true, true);
4387 EXPORT_SYMBOL_GPL(sched_setattr);
4390 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4391 * @p: the task in question.
4392 * @policy: new policy.
4393 * @param: structure containing the new RT priority.
4395 * Just like sched_setscheduler, only don't bother checking if the
4396 * current context has permission. For example, this is needed in
4397 * stop_machine(): we create temporary high priority worker threads,
4398 * but our caller might not have that capability.
4400 * Return: 0 on success. An error code otherwise.
4402 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4403 const struct sched_param *param)
4405 return _sched_setscheduler(p, policy, param, false);
4407 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4410 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4412 struct sched_param lparam;
4413 struct task_struct *p;
4416 if (!param || pid < 0)
4418 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4423 p = find_process_by_pid(pid);
4425 retval = sched_setscheduler(p, policy, &lparam);
4432 * Mimics kernel/events/core.c perf_copy_attr().
4434 static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
4439 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4442 /* Zero the full structure, so that a short copy will be nice: */
4443 memset(attr, 0, sizeof(*attr));
4445 ret = get_user(size, &uattr->size);
4449 /* Bail out on silly large: */
4450 if (size > PAGE_SIZE)
4453 /* ABI compatibility quirk: */
4455 size = SCHED_ATTR_SIZE_VER0;
4457 if (size < SCHED_ATTR_SIZE_VER0)
4461 * If we're handed a bigger struct than we know of,
4462 * ensure all the unknown bits are 0 - i.e. new
4463 * user-space does not rely on any kernel feature
4464 * extensions we dont know about yet.
4466 if (size > sizeof(*attr)) {
4467 unsigned char __user *addr;
4468 unsigned char __user *end;
4471 addr = (void __user *)uattr + sizeof(*attr);
4472 end = (void __user *)uattr + size;
4474 for (; addr < end; addr++) {
4475 ret = get_user(val, addr);
4481 size = sizeof(*attr);
4484 ret = copy_from_user(attr, uattr, size);
4489 * XXX: Do we want to be lenient like existing syscalls; or do we want
4490 * to be strict and return an error on out-of-bounds values?
4492 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4497 put_user(sizeof(*attr), &uattr->size);
4502 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4503 * @pid: the pid in question.
4504 * @policy: new policy.
4505 * @param: structure containing the new RT priority.
4507 * Return: 0 on success. An error code otherwise.
4509 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
4514 return do_sched_setscheduler(pid, policy, param);
4518 * sys_sched_setparam - set/change the RT priority of a thread
4519 * @pid: the pid in question.
4520 * @param: structure containing the new RT priority.
4522 * Return: 0 on success. An error code otherwise.
4524 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4526 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4530 * sys_sched_setattr - same as above, but with extended sched_attr
4531 * @pid: the pid in question.
4532 * @uattr: structure containing the extended parameters.
4533 * @flags: for future extension.
4535 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4536 unsigned int, flags)
4538 struct sched_attr attr;
4539 struct task_struct *p;
4542 if (!uattr || pid < 0 || flags)
4545 retval = sched_copy_attr(uattr, &attr);
4549 if ((int)attr.sched_policy < 0)
4554 p = find_process_by_pid(pid);
4556 retval = sched_setattr(p, &attr);
4563 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4564 * @pid: the pid in question.
4566 * Return: On success, the policy of the thread. Otherwise, a negative error
4569 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4571 struct task_struct *p;
4579 p = find_process_by_pid(pid);
4581 retval = security_task_getscheduler(p);
4584 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4591 * sys_sched_getparam - get the RT priority of a thread
4592 * @pid: the pid in question.
4593 * @param: structure containing the RT priority.
4595 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4598 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4600 struct sched_param lp = { .sched_priority = 0 };
4601 struct task_struct *p;
4604 if (!param || pid < 0)
4608 p = find_process_by_pid(pid);
4613 retval = security_task_getscheduler(p);
4617 if (task_has_rt_policy(p))
4618 lp.sched_priority = p->rt_priority;
4622 * This one might sleep, we cannot do it with a spinlock held ...
4624 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4633 static int sched_read_attr(struct sched_attr __user *uattr,
4634 struct sched_attr *attr,
4639 if (!access_ok(VERIFY_WRITE, uattr, usize))
4643 * If we're handed a smaller struct than we know of,
4644 * ensure all the unknown bits are 0 - i.e. old
4645 * user-space does not get uncomplete information.
4647 if (usize < sizeof(*attr)) {
4648 unsigned char *addr;
4651 addr = (void *)attr + usize;
4652 end = (void *)attr + sizeof(*attr);
4654 for (; addr < end; addr++) {
4662 ret = copy_to_user(uattr, attr, attr->size);
4670 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4671 * @pid: the pid in question.
4672 * @uattr: structure containing the extended parameters.
4673 * @size: sizeof(attr) for fwd/bwd comp.
4674 * @flags: for future extension.
4676 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4677 unsigned int, size, unsigned int, flags)
4679 struct sched_attr attr = {
4680 .size = sizeof(struct sched_attr),
4682 struct task_struct *p;
4685 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4686 size < SCHED_ATTR_SIZE_VER0 || flags)
4690 p = find_process_by_pid(pid);
4695 retval = security_task_getscheduler(p);
4699 attr.sched_policy = p->policy;
4700 if (p->sched_reset_on_fork)
4701 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4702 if (task_has_dl_policy(p))
4703 __getparam_dl(p, &attr);
4704 else if (task_has_rt_policy(p))
4705 attr.sched_priority = p->rt_priority;
4707 attr.sched_nice = task_nice(p);
4711 retval = sched_read_attr(uattr, &attr, size);
4719 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4721 cpumask_var_t cpus_allowed, new_mask;
4722 struct task_struct *p;
4727 p = find_process_by_pid(pid);
4733 /* Prevent p going away */
4737 if (p->flags & PF_NO_SETAFFINITY) {
4741 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4745 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4747 goto out_free_cpus_allowed;
4750 if (!check_same_owner(p)) {
4752 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4754 goto out_free_new_mask;
4759 retval = security_task_setscheduler(p);
4761 goto out_free_new_mask;
4764 cpuset_cpus_allowed(p, cpus_allowed);
4765 cpumask_and(new_mask, in_mask, cpus_allowed);
4768 * Since bandwidth control happens on root_domain basis,
4769 * if admission test is enabled, we only admit -deadline
4770 * tasks allowed to run on all the CPUs in the task's
4774 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4776 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4779 goto out_free_new_mask;
4785 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4788 cpuset_cpus_allowed(p, cpus_allowed);
4789 if (!cpumask_subset(new_mask, cpus_allowed)) {
4791 * We must have raced with a concurrent cpuset
4792 * update. Just reset the cpus_allowed to the
4793 * cpuset's cpus_allowed
4795 cpumask_copy(new_mask, cpus_allowed);
4800 free_cpumask_var(new_mask);
4801 out_free_cpus_allowed:
4802 free_cpumask_var(cpus_allowed);
4808 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4809 struct cpumask *new_mask)
4811 if (len < cpumask_size())
4812 cpumask_clear(new_mask);
4813 else if (len > cpumask_size())
4814 len = cpumask_size();
4816 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4820 * sys_sched_setaffinity - set the CPU affinity of a process
4821 * @pid: pid of the process
4822 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4823 * @user_mask_ptr: user-space pointer to the new CPU mask
4825 * Return: 0 on success. An error code otherwise.
4827 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4828 unsigned long __user *, user_mask_ptr)
4830 cpumask_var_t new_mask;
4833 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4836 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4838 retval = sched_setaffinity(pid, new_mask);
4839 free_cpumask_var(new_mask);
4843 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4845 struct task_struct *p;
4846 unsigned long flags;
4852 p = find_process_by_pid(pid);
4856 retval = security_task_getscheduler(p);
4860 raw_spin_lock_irqsave(&p->pi_lock, flags);
4861 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4862 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4871 * sys_sched_getaffinity - get the CPU affinity of a process
4872 * @pid: pid of the process
4873 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4874 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4876 * Return: size of CPU mask copied to user_mask_ptr on success. An
4877 * error code otherwise.
4879 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4880 unsigned long __user *, user_mask_ptr)
4885 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4887 if (len & (sizeof(unsigned long)-1))
4890 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4893 ret = sched_getaffinity(pid, mask);
4895 size_t retlen = min_t(size_t, len, cpumask_size());
4897 if (copy_to_user(user_mask_ptr, mask, retlen))
4902 free_cpumask_var(mask);
4908 * sys_sched_yield - yield the current processor to other threads.
4910 * This function yields the current CPU to other tasks. If there are no
4911 * other threads running on this CPU then this function will return.
4915 SYSCALL_DEFINE0(sched_yield)
4917 struct rq *rq = this_rq_lock();
4919 schedstat_inc(rq->yld_count);
4920 current->sched_class->yield_task(rq);
4923 * Since we are going to call schedule() anyway, there's
4924 * no need to preempt or enable interrupts:
4926 __release(rq->lock);
4927 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4928 do_raw_spin_unlock(&rq->lock);
4929 sched_preempt_enable_no_resched();
4936 #ifndef CONFIG_PREEMPT
4937 int __sched _cond_resched(void)
4939 if (should_resched(0)) {
4940 preempt_schedule_common();
4945 EXPORT_SYMBOL(_cond_resched);
4949 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4950 * call schedule, and on return reacquire the lock.
4952 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4953 * operations here to prevent schedule() from being called twice (once via
4954 * spin_unlock(), once by hand).
4956 int __cond_resched_lock(spinlock_t *lock)
4958 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4961 lockdep_assert_held(lock);
4963 if (spin_needbreak(lock) || resched) {
4966 preempt_schedule_common();
4974 EXPORT_SYMBOL(__cond_resched_lock);
4976 int __sched __cond_resched_softirq(void)
4978 BUG_ON(!in_softirq());
4980 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4982 preempt_schedule_common();
4988 EXPORT_SYMBOL(__cond_resched_softirq);
4991 * yield - yield the current processor to other threads.
4993 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4995 * The scheduler is at all times free to pick the calling task as the most
4996 * eligible task to run, if removing the yield() call from your code breaks
4997 * it, its already broken.
4999 * Typical broken usage is:
5004 * where one assumes that yield() will let 'the other' process run that will
5005 * make event true. If the current task is a SCHED_FIFO task that will never
5006 * happen. Never use yield() as a progress guarantee!!
5008 * If you want to use yield() to wait for something, use wait_event().
5009 * If you want to use yield() to be 'nice' for others, use cond_resched().
5010 * If you still want to use yield(), do not!
5012 void __sched yield(void)
5014 set_current_state(TASK_RUNNING);
5017 EXPORT_SYMBOL(yield);
5020 * yield_to - yield the current processor to another thread in
5021 * your thread group, or accelerate that thread toward the
5022 * processor it's on.
5024 * @preempt: whether task preemption is allowed or not
5026 * It's the caller's job to ensure that the target task struct
5027 * can't go away on us before we can do any checks.
5030 * true (>0) if we indeed boosted the target task.
5031 * false (0) if we failed to boost the target.
5032 * -ESRCH if there's no task to yield to.
5034 int __sched yield_to(struct task_struct *p, bool preempt)
5036 struct task_struct *curr = current;
5037 struct rq *rq, *p_rq;
5038 unsigned long flags;
5041 local_irq_save(flags);
5047 * If we're the only runnable task on the rq and target rq also
5048 * has only one task, there's absolutely no point in yielding.
5050 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5055 double_rq_lock(rq, p_rq);
5056 if (task_rq(p) != p_rq) {
5057 double_rq_unlock(rq, p_rq);
5061 if (!curr->sched_class->yield_to_task)
5064 if (curr->sched_class != p->sched_class)
5067 if (task_running(p_rq, p) || p->state)
5070 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5072 schedstat_inc(rq->yld_count);
5074 * Make p's CPU reschedule; pick_next_entity takes care of
5077 if (preempt && rq != p_rq)
5082 double_rq_unlock(rq, p_rq);
5084 local_irq_restore(flags);
5091 EXPORT_SYMBOL_GPL(yield_to);
5093 int io_schedule_prepare(void)
5095 int old_iowait = current->in_iowait;
5097 current->in_iowait = 1;
5098 blk_schedule_flush_plug(current);
5103 void io_schedule_finish(int token)
5105 current->in_iowait = token;
5109 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5110 * that process accounting knows that this is a task in IO wait state.
5112 long __sched io_schedule_timeout(long timeout)
5117 token = io_schedule_prepare();
5118 ret = schedule_timeout(timeout);
5119 io_schedule_finish(token);
5123 EXPORT_SYMBOL(io_schedule_timeout);
5125 void io_schedule(void)
5129 token = io_schedule_prepare();
5131 io_schedule_finish(token);
5133 EXPORT_SYMBOL(io_schedule);
5136 * sys_sched_get_priority_max - return maximum RT priority.
5137 * @policy: scheduling class.
5139 * Return: On success, this syscall returns the maximum
5140 * rt_priority that can be used by a given scheduling class.
5141 * On failure, a negative error code is returned.
5143 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5150 ret = MAX_USER_RT_PRIO-1;
5152 case SCHED_DEADLINE:
5163 * sys_sched_get_priority_min - return minimum RT priority.
5164 * @policy: scheduling class.
5166 * Return: On success, this syscall returns the minimum
5167 * rt_priority that can be used by a given scheduling class.
5168 * On failure, a negative error code is returned.
5170 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5179 case SCHED_DEADLINE:
5189 * sys_sched_rr_get_interval - return the default timeslice of a process.
5190 * @pid: pid of the process.
5191 * @interval: userspace pointer to the timeslice value.
5193 * this syscall writes the default timeslice value of a given process
5194 * into the user-space timespec buffer. A value of '0' means infinity.
5196 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5199 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5200 struct timespec __user *, interval)
5202 struct task_struct *p;
5203 unsigned int time_slice;
5214 p = find_process_by_pid(pid);
5218 retval = security_task_getscheduler(p);
5222 rq = task_rq_lock(p, &rf);
5224 if (p->sched_class->get_rr_interval)
5225 time_slice = p->sched_class->get_rr_interval(rq, p);
5226 task_rq_unlock(rq, p, &rf);
5229 jiffies_to_timespec(time_slice, &t);
5230 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5238 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5240 void sched_show_task(struct task_struct *p)
5242 unsigned long free = 0;
5244 unsigned long state = p->state;
5246 /* Make sure the string lines up properly with the number of task states: */
5247 BUILD_BUG_ON(sizeof(TASK_STATE_TO_CHAR_STR)-1 != ilog2(TASK_STATE_MAX)+1);
5249 if (!try_get_task_stack(p))
5252 state = __ffs(state) + 1;
5253 printk(KERN_INFO "%-15.15s %c", p->comm,
5254 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5255 if (state == TASK_RUNNING)
5256 printk(KERN_CONT " running task ");
5257 #ifdef CONFIG_DEBUG_STACK_USAGE
5258 free = stack_not_used(p);
5263 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5265 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5266 task_pid_nr(p), ppid,
5267 (unsigned long)task_thread_info(p)->flags);
5269 print_worker_info(KERN_INFO, p);
5270 show_stack(p, NULL);
5274 void show_state_filter(unsigned long state_filter)
5276 struct task_struct *g, *p;
5278 #if BITS_PER_LONG == 32
5280 " task PC stack pid father\n");
5283 " task PC stack pid father\n");
5286 for_each_process_thread(g, p) {
5288 * reset the NMI-timeout, listing all files on a slow
5289 * console might take a lot of time:
5290 * Also, reset softlockup watchdogs on all CPUs, because
5291 * another CPU might be blocked waiting for us to process
5294 touch_nmi_watchdog();
5295 touch_all_softlockup_watchdogs();
5296 if (!state_filter || (p->state & state_filter))
5300 #ifdef CONFIG_SCHED_DEBUG
5302 sysrq_sched_debug_show();
5306 * Only show locks if all tasks are dumped:
5309 debug_show_all_locks();
5312 void init_idle_bootup_task(struct task_struct *idle)
5314 idle->sched_class = &idle_sched_class;
5318 * init_idle - set up an idle thread for a given CPU
5319 * @idle: task in question
5320 * @cpu: CPU the idle task belongs to
5322 * NOTE: this function does not set the idle thread's NEED_RESCHED
5323 * flag, to make booting more robust.
5325 void init_idle(struct task_struct *idle, int cpu)
5327 struct rq *rq = cpu_rq(cpu);
5328 unsigned long flags;
5330 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5331 raw_spin_lock(&rq->lock);
5333 __sched_fork(0, idle);
5334 idle->state = TASK_RUNNING;
5335 idle->se.exec_start = sched_clock();
5336 idle->flags |= PF_IDLE;
5338 kasan_unpoison_task_stack(idle);
5342 * Its possible that init_idle() gets called multiple times on a task,
5343 * in that case do_set_cpus_allowed() will not do the right thing.
5345 * And since this is boot we can forgo the serialization.
5347 set_cpus_allowed_common(idle, cpumask_of(cpu));
5350 * We're having a chicken and egg problem, even though we are
5351 * holding rq->lock, the CPU isn't yet set to this CPU so the
5352 * lockdep check in task_group() will fail.
5354 * Similar case to sched_fork(). / Alternatively we could
5355 * use task_rq_lock() here and obtain the other rq->lock.
5360 __set_task_cpu(idle, cpu);
5363 rq->curr = rq->idle = idle;
5364 idle->on_rq = TASK_ON_RQ_QUEUED;
5368 raw_spin_unlock(&rq->lock);
5369 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5371 /* Set the preempt count _outside_ the spinlocks! */
5372 init_idle_preempt_count(idle, cpu);
5375 * The idle tasks have their own, simple scheduling class:
5377 idle->sched_class = &idle_sched_class;
5378 ftrace_graph_init_idle_task(idle, cpu);
5379 vtime_init_idle(idle, cpu);
5381 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5385 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5386 const struct cpumask *trial)
5388 int ret = 1, trial_cpus;
5389 struct dl_bw *cur_dl_b;
5390 unsigned long flags;
5392 if (!cpumask_weight(cur))
5395 rcu_read_lock_sched();
5396 cur_dl_b = dl_bw_of(cpumask_any(cur));
5397 trial_cpus = cpumask_weight(trial);
5399 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5400 if (cur_dl_b->bw != -1 &&
5401 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5403 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5404 rcu_read_unlock_sched();
5409 int task_can_attach(struct task_struct *p,
5410 const struct cpumask *cs_cpus_allowed)
5415 * Kthreads which disallow setaffinity shouldn't be moved
5416 * to a new cpuset; we don't want to change their CPU
5417 * affinity and isolating such threads by their set of
5418 * allowed nodes is unnecessary. Thus, cpusets are not
5419 * applicable for such threads. This prevents checking for
5420 * success of set_cpus_allowed_ptr() on all attached tasks
5421 * before cpus_allowed may be changed.
5423 if (p->flags & PF_NO_SETAFFINITY) {
5429 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5431 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5436 unsigned long flags;
5438 rcu_read_lock_sched();
5439 dl_b = dl_bw_of(dest_cpu);
5440 raw_spin_lock_irqsave(&dl_b->lock, flags);
5441 cpus = dl_bw_cpus(dest_cpu);
5442 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5447 * We reserve space for this task in the destination
5448 * root_domain, as we can't fail after this point.
5449 * We will free resources in the source root_domain
5450 * later on (see set_cpus_allowed_dl()).
5452 __dl_add(dl_b, p->dl.dl_bw);
5454 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5455 rcu_read_unlock_sched();
5465 bool sched_smp_initialized __read_mostly;
5467 #ifdef CONFIG_NUMA_BALANCING
5468 /* Migrate current task p to target_cpu */
5469 int migrate_task_to(struct task_struct *p, int target_cpu)
5471 struct migration_arg arg = { p, target_cpu };
5472 int curr_cpu = task_cpu(p);
5474 if (curr_cpu == target_cpu)
5477 if (!cpumask_test_cpu(target_cpu, &p->cpus_allowed))
5480 /* TODO: This is not properly updating schedstats */
5482 trace_sched_move_numa(p, curr_cpu, target_cpu);
5483 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5487 * Requeue a task on a given node and accurately track the number of NUMA
5488 * tasks on the runqueues
5490 void sched_setnuma(struct task_struct *p, int nid)
5492 bool queued, running;
5496 rq = task_rq_lock(p, &rf);
5497 queued = task_on_rq_queued(p);
5498 running = task_current(rq, p);
5501 dequeue_task(rq, p, DEQUEUE_SAVE);
5503 put_prev_task(rq, p);
5505 p->numa_preferred_nid = nid;
5508 enqueue_task(rq, p, ENQUEUE_RESTORE);
5510 set_curr_task(rq, p);
5511 task_rq_unlock(rq, p, &rf);
5513 #endif /* CONFIG_NUMA_BALANCING */
5515 #ifdef CONFIG_HOTPLUG_CPU
5517 * Ensure that the idle task is using init_mm right before its CPU goes
5520 void idle_task_exit(void)
5522 struct mm_struct *mm = current->active_mm;
5524 BUG_ON(cpu_online(smp_processor_id()));
5526 if (mm != &init_mm) {
5527 switch_mm_irqs_off(mm, &init_mm, current);
5528 finish_arch_post_lock_switch();
5534 * Since this CPU is going 'away' for a while, fold any nr_active delta
5535 * we might have. Assumes we're called after migrate_tasks() so that the
5536 * nr_active count is stable. We need to take the teardown thread which
5537 * is calling this into account, so we hand in adjust = 1 to the load
5540 * Also see the comment "Global load-average calculations".
5542 static void calc_load_migrate(struct rq *rq)
5544 long delta = calc_load_fold_active(rq, 1);
5546 atomic_long_add(delta, &calc_load_tasks);
5549 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5553 static const struct sched_class fake_sched_class = {
5554 .put_prev_task = put_prev_task_fake,
5557 static struct task_struct fake_task = {
5559 * Avoid pull_{rt,dl}_task()
5561 .prio = MAX_PRIO + 1,
5562 .sched_class = &fake_sched_class,
5566 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5567 * try_to_wake_up()->select_task_rq().
5569 * Called with rq->lock held even though we'er in stop_machine() and
5570 * there's no concurrency possible, we hold the required locks anyway
5571 * because of lock validation efforts.
5573 static void migrate_tasks(struct rq *dead_rq)
5575 struct rq *rq = dead_rq;
5576 struct task_struct *next, *stop = rq->stop;
5581 * Fudge the rq selection such that the below task selection loop
5582 * doesn't get stuck on the currently eligible stop task.
5584 * We're currently inside stop_machine() and the rq is either stuck
5585 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5586 * either way we should never end up calling schedule() until we're
5592 * put_prev_task() and pick_next_task() sched
5593 * class method both need to have an up-to-date
5594 * value of rq->clock[_task]
5596 rq_pin_lock(rq, &rf);
5597 update_rq_clock(rq);
5598 rq_unpin_lock(rq, &rf);
5602 * There's this thread running, bail when that's the only
5605 if (rq->nr_running == 1)
5609 * pick_next_task() assumes pinned rq->lock:
5611 rq_repin_lock(rq, &rf);
5612 next = pick_next_task(rq, &fake_task, &rf);
5614 next->sched_class->put_prev_task(rq, next);
5617 * Rules for changing task_struct::cpus_allowed are holding
5618 * both pi_lock and rq->lock, such that holding either
5619 * stabilizes the mask.
5621 * Drop rq->lock is not quite as disastrous as it usually is
5622 * because !cpu_active at this point, which means load-balance
5623 * will not interfere. Also, stop-machine.
5625 rq_unpin_lock(rq, &rf);
5626 raw_spin_unlock(&rq->lock);
5627 raw_spin_lock(&next->pi_lock);
5628 raw_spin_lock(&rq->lock);
5631 * Since we're inside stop-machine, _nothing_ should have
5632 * changed the task, WARN if weird stuff happened, because in
5633 * that case the above rq->lock drop is a fail too.
5635 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5636 raw_spin_unlock(&next->pi_lock);
5640 /* Find suitable destination for @next, with force if needed. */
5641 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5643 rq = __migrate_task(rq, next, dest_cpu);
5644 if (rq != dead_rq) {
5645 raw_spin_unlock(&rq->lock);
5647 raw_spin_lock(&rq->lock);
5649 raw_spin_unlock(&next->pi_lock);
5654 #endif /* CONFIG_HOTPLUG_CPU */
5656 void set_rq_online(struct rq *rq)
5659 const struct sched_class *class;
5661 cpumask_set_cpu(rq->cpu, rq->rd->online);
5664 for_each_class(class) {
5665 if (class->rq_online)
5666 class->rq_online(rq);
5671 void set_rq_offline(struct rq *rq)
5674 const struct sched_class *class;
5676 for_each_class(class) {
5677 if (class->rq_offline)
5678 class->rq_offline(rq);
5681 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5686 static void set_cpu_rq_start_time(unsigned int cpu)
5688 struct rq *rq = cpu_rq(cpu);
5690 rq->age_stamp = sched_clock_cpu(cpu);
5694 * used to mark begin/end of suspend/resume:
5696 static int num_cpus_frozen;
5699 * Update cpusets according to cpu_active mask. If cpusets are
5700 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5701 * around partition_sched_domains().
5703 * If we come here as part of a suspend/resume, don't touch cpusets because we
5704 * want to restore it back to its original state upon resume anyway.
5706 static void cpuset_cpu_active(void)
5708 if (cpuhp_tasks_frozen) {
5710 * num_cpus_frozen tracks how many CPUs are involved in suspend
5711 * resume sequence. As long as this is not the last online
5712 * operation in the resume sequence, just build a single sched
5713 * domain, ignoring cpusets.
5716 if (likely(num_cpus_frozen)) {
5717 partition_sched_domains(1, NULL, NULL);
5721 * This is the last CPU online operation. So fall through and
5722 * restore the original sched domains by considering the
5723 * cpuset configurations.
5726 cpuset_update_active_cpus(true);
5729 static int cpuset_cpu_inactive(unsigned int cpu)
5731 unsigned long flags;
5736 if (!cpuhp_tasks_frozen) {
5737 rcu_read_lock_sched();
5738 dl_b = dl_bw_of(cpu);
5740 raw_spin_lock_irqsave(&dl_b->lock, flags);
5741 cpus = dl_bw_cpus(cpu);
5742 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5743 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5745 rcu_read_unlock_sched();
5749 cpuset_update_active_cpus(false);
5752 partition_sched_domains(1, NULL, NULL);
5757 int sched_cpu_activate(unsigned int cpu)
5759 struct rq *rq = cpu_rq(cpu);
5760 unsigned long flags;
5762 set_cpu_active(cpu, true);
5764 if (sched_smp_initialized) {
5765 sched_domains_numa_masks_set(cpu);
5766 cpuset_cpu_active();
5770 * Put the rq online, if not already. This happens:
5772 * 1) In the early boot process, because we build the real domains
5773 * after all CPUs have been brought up.
5775 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5778 raw_spin_lock_irqsave(&rq->lock, flags);
5780 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5783 raw_spin_unlock_irqrestore(&rq->lock, flags);
5785 update_max_interval();
5790 int sched_cpu_deactivate(unsigned int cpu)
5794 set_cpu_active(cpu, false);
5796 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5797 * users of this state to go away such that all new such users will
5800 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
5801 * not imply sync_sched(), so wait for both.
5803 * Do sync before park smpboot threads to take care the rcu boost case.
5805 if (IS_ENABLED(CONFIG_PREEMPT))
5806 synchronize_rcu_mult(call_rcu, call_rcu_sched);
5810 if (!sched_smp_initialized)
5813 ret = cpuset_cpu_inactive(cpu);
5815 set_cpu_active(cpu, true);
5818 sched_domains_numa_masks_clear(cpu);
5822 static void sched_rq_cpu_starting(unsigned int cpu)
5824 struct rq *rq = cpu_rq(cpu);
5826 rq->calc_load_update = calc_load_update;
5827 update_max_interval();
5830 int sched_cpu_starting(unsigned int cpu)
5832 set_cpu_rq_start_time(cpu);
5833 sched_rq_cpu_starting(cpu);
5837 #ifdef CONFIG_HOTPLUG_CPU
5838 int sched_cpu_dying(unsigned int cpu)
5840 struct rq *rq = cpu_rq(cpu);
5841 unsigned long flags;
5843 /* Handle pending wakeups and then migrate everything off */
5844 sched_ttwu_pending();
5845 raw_spin_lock_irqsave(&rq->lock, flags);
5847 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5851 BUG_ON(rq->nr_running != 1);
5852 raw_spin_unlock_irqrestore(&rq->lock, flags);
5853 calc_load_migrate(rq);
5854 update_max_interval();
5855 nohz_balance_exit_idle(cpu);
5861 #ifdef CONFIG_SCHED_SMT
5862 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5864 static void sched_init_smt(void)
5867 * We've enumerated all CPUs and will assume that if any CPU
5868 * has SMT siblings, CPU0 will too.
5870 if (cpumask_weight(cpu_smt_mask(0)) > 1)
5871 static_branch_enable(&sched_smt_present);
5874 static inline void sched_init_smt(void) { }
5877 void __init sched_init_smp(void)
5879 cpumask_var_t non_isolated_cpus;
5881 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
5882 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
5887 * There's no userspace yet to cause hotplug operations; hence all the
5888 * CPU masks are stable and all blatant races in the below code cannot
5891 mutex_lock(&sched_domains_mutex);
5892 init_sched_domains(cpu_active_mask);
5893 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
5894 if (cpumask_empty(non_isolated_cpus))
5895 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
5896 mutex_unlock(&sched_domains_mutex);
5898 /* Move init over to a non-isolated CPU */
5899 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
5901 sched_init_granularity();
5902 free_cpumask_var(non_isolated_cpus);
5904 init_sched_rt_class();
5905 init_sched_dl_class();
5908 sched_clock_init_late();
5910 sched_smp_initialized = true;
5913 static int __init migration_init(void)
5915 sched_rq_cpu_starting(smp_processor_id());
5918 early_initcall(migration_init);
5921 void __init sched_init_smp(void)
5923 sched_init_granularity();
5924 sched_clock_init_late();
5926 #endif /* CONFIG_SMP */
5928 int in_sched_functions(unsigned long addr)
5930 return in_lock_functions(addr) ||
5931 (addr >= (unsigned long)__sched_text_start
5932 && addr < (unsigned long)__sched_text_end);
5935 #ifdef CONFIG_CGROUP_SCHED
5937 * Default task group.
5938 * Every task in system belongs to this group at bootup.
5940 struct task_group root_task_group;
5941 LIST_HEAD(task_groups);
5943 /* Cacheline aligned slab cache for task_group */
5944 static struct kmem_cache *task_group_cache __read_mostly;
5947 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
5948 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
5950 #define WAIT_TABLE_BITS 8
5951 #define WAIT_TABLE_SIZE (1 << WAIT_TABLE_BITS)
5952 static wait_queue_head_t bit_wait_table[WAIT_TABLE_SIZE] __cacheline_aligned;
5954 wait_queue_head_t *bit_waitqueue(void *word, int bit)
5956 const int shift = BITS_PER_LONG == 32 ? 5 : 6;
5957 unsigned long val = (unsigned long)word << shift | bit;
5959 return bit_wait_table + hash_long(val, WAIT_TABLE_BITS);
5961 EXPORT_SYMBOL(bit_waitqueue);
5963 void __init sched_init(void)
5966 unsigned long alloc_size = 0, ptr;
5970 for (i = 0; i < WAIT_TABLE_SIZE; i++)
5971 init_waitqueue_head(bit_wait_table + i);
5973 #ifdef CONFIG_FAIR_GROUP_SCHED
5974 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5976 #ifdef CONFIG_RT_GROUP_SCHED
5977 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
5980 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
5982 #ifdef CONFIG_FAIR_GROUP_SCHED
5983 root_task_group.se = (struct sched_entity **)ptr;
5984 ptr += nr_cpu_ids * sizeof(void **);
5986 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
5987 ptr += nr_cpu_ids * sizeof(void **);
5989 #endif /* CONFIG_FAIR_GROUP_SCHED */
5990 #ifdef CONFIG_RT_GROUP_SCHED
5991 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
5992 ptr += nr_cpu_ids * sizeof(void **);
5994 root_task_group.rt_rq = (struct rt_rq **)ptr;
5995 ptr += nr_cpu_ids * sizeof(void **);
5997 #endif /* CONFIG_RT_GROUP_SCHED */
5999 #ifdef CONFIG_CPUMASK_OFFSTACK
6000 for_each_possible_cpu(i) {
6001 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
6002 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6003 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
6004 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
6006 #endif /* CONFIG_CPUMASK_OFFSTACK */
6008 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
6009 init_dl_bandwidth(&def_dl_bandwidth, global_rt_period(), global_rt_runtime());
6012 init_defrootdomain();
6015 #ifdef CONFIG_RT_GROUP_SCHED
6016 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6017 global_rt_period(), global_rt_runtime());
6018 #endif /* CONFIG_RT_GROUP_SCHED */
6020 #ifdef CONFIG_CGROUP_SCHED
6021 task_group_cache = KMEM_CACHE(task_group, 0);
6023 list_add(&root_task_group.list, &task_groups);
6024 INIT_LIST_HEAD(&root_task_group.children);
6025 INIT_LIST_HEAD(&root_task_group.siblings);
6026 autogroup_init(&init_task);
6027 #endif /* CONFIG_CGROUP_SCHED */
6029 for_each_possible_cpu(i) {
6033 raw_spin_lock_init(&rq->lock);
6035 rq->calc_load_active = 0;
6036 rq->calc_load_update = jiffies + LOAD_FREQ;
6037 init_cfs_rq(&rq->cfs);
6038 init_rt_rq(&rq->rt);
6039 init_dl_rq(&rq->dl);
6040 #ifdef CONFIG_FAIR_GROUP_SCHED
6041 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6042 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6043 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
6045 * How much CPU bandwidth does root_task_group get?
6047 * In case of task-groups formed thr' the cgroup filesystem, it
6048 * gets 100% of the CPU resources in the system. This overall
6049 * system CPU resource is divided among the tasks of
6050 * root_task_group and its child task-groups in a fair manner,
6051 * based on each entity's (task or task-group's) weight
6052 * (se->load.weight).
6054 * In other words, if root_task_group has 10 tasks of weight
6055 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6056 * then A0's share of the CPU resource is:
6058 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6060 * We achieve this by letting root_task_group's tasks sit
6061 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6063 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6064 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6065 #endif /* CONFIG_FAIR_GROUP_SCHED */
6067 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6068 #ifdef CONFIG_RT_GROUP_SCHED
6069 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6072 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6073 rq->cpu_load[j] = 0;
6078 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
6079 rq->balance_callback = NULL;
6080 rq->active_balance = 0;
6081 rq->next_balance = jiffies;
6086 rq->avg_idle = 2*sysctl_sched_migration_cost;
6087 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6089 INIT_LIST_HEAD(&rq->cfs_tasks);
6091 rq_attach_root(rq, &def_root_domain);
6092 #ifdef CONFIG_NO_HZ_COMMON
6093 rq->last_load_update_tick = jiffies;
6096 #ifdef CONFIG_NO_HZ_FULL
6097 rq->last_sched_tick = 0;
6099 #endif /* CONFIG_SMP */
6101 atomic_set(&rq->nr_iowait, 0);
6104 set_load_weight(&init_task);
6107 * The boot idle thread does lazy MMU switching as well:
6110 enter_lazy_tlb(&init_mm, current);
6113 * Make us the idle thread. Technically, schedule() should not be
6114 * called from this thread, however somewhere below it might be,
6115 * but because we are the idle thread, we just pick up running again
6116 * when this runqueue becomes "idle".
6118 init_idle(current, smp_processor_id());
6120 calc_load_update = jiffies + LOAD_FREQ;
6123 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6124 /* May be allocated at isolcpus cmdline parse time */
6125 if (cpu_isolated_map == NULL)
6126 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6127 idle_thread_set_boot_cpu();
6128 set_cpu_rq_start_time(smp_processor_id());
6130 init_sched_fair_class();
6134 scheduler_running = 1;
6137 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6138 static inline int preempt_count_equals(int preempt_offset)
6140 int nested = preempt_count() + rcu_preempt_depth();
6142 return (nested == preempt_offset);
6145 void __might_sleep(const char *file, int line, int preempt_offset)
6148 * Blocking primitives will set (and therefore destroy) current->state,
6149 * since we will exit with TASK_RUNNING make sure we enter with it,
6150 * otherwise we will destroy state.
6152 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
6153 "do not call blocking ops when !TASK_RUNNING; "
6154 "state=%lx set at [<%p>] %pS\n",
6156 (void *)current->task_state_change,
6157 (void *)current->task_state_change);
6159 ___might_sleep(file, line, preempt_offset);
6161 EXPORT_SYMBOL(__might_sleep);
6163 void ___might_sleep(const char *file, int line, int preempt_offset)
6165 /* Ratelimiting timestamp: */
6166 static unsigned long prev_jiffy;
6168 unsigned long preempt_disable_ip;
6170 /* WARN_ON_ONCE() by default, no rate limit required: */
6173 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6174 !is_idle_task(current)) ||
6175 system_state != SYSTEM_RUNNING || oops_in_progress)
6177 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6179 prev_jiffy = jiffies;
6181 /* Save this before calling printk(), since that will clobber it: */
6182 preempt_disable_ip = get_preempt_disable_ip(current);
6185 "BUG: sleeping function called from invalid context at %s:%d\n",
6188 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6189 in_atomic(), irqs_disabled(),
6190 current->pid, current->comm);
6192 if (task_stack_end_corrupted(current))
6193 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
6195 debug_show_held_locks(current);
6196 if (irqs_disabled())
6197 print_irqtrace_events(current);
6198 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
6199 && !preempt_count_equals(preempt_offset)) {
6200 pr_err("Preemption disabled at:");
6201 print_ip_sym(preempt_disable_ip);
6205 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
6207 EXPORT_SYMBOL(___might_sleep);
6210 #ifdef CONFIG_MAGIC_SYSRQ
6211 void normalize_rt_tasks(void)
6213 struct task_struct *g, *p;
6214 struct sched_attr attr = {
6215 .sched_policy = SCHED_NORMAL,
6218 read_lock(&tasklist_lock);
6219 for_each_process_thread(g, p) {
6221 * Only normalize user tasks:
6223 if (p->flags & PF_KTHREAD)
6226 p->se.exec_start = 0;
6227 schedstat_set(p->se.statistics.wait_start, 0);
6228 schedstat_set(p->se.statistics.sleep_start, 0);
6229 schedstat_set(p->se.statistics.block_start, 0);
6231 if (!dl_task(p) && !rt_task(p)) {
6233 * Renice negative nice level userspace
6236 if (task_nice(p) < 0)
6237 set_user_nice(p, 0);
6241 __sched_setscheduler(p, &attr, false, false);
6243 read_unlock(&tasklist_lock);
6246 #endif /* CONFIG_MAGIC_SYSRQ */
6248 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6250 * These functions are only useful for the IA64 MCA handling, or kdb.
6252 * They can only be called when the whole system has been
6253 * stopped - every CPU needs to be quiescent, and no scheduling
6254 * activity can take place. Using them for anything else would
6255 * be a serious bug, and as a result, they aren't even visible
6256 * under any other configuration.
6260 * curr_task - return the current task for a given CPU.
6261 * @cpu: the processor in question.
6263 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6265 * Return: The current task for @cpu.
6267 struct task_struct *curr_task(int cpu)
6269 return cpu_curr(cpu);
6272 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6276 * set_curr_task - set the current task for a given CPU.
6277 * @cpu: the processor in question.
6278 * @p: the task pointer to set.
6280 * Description: This function must only be used when non-maskable interrupts
6281 * are serviced on a separate stack. It allows the architecture to switch the
6282 * notion of the current task on a CPU in a non-blocking manner. This function
6283 * must be called with all CPU's synchronized, and interrupts disabled, the
6284 * and caller must save the original value of the current task (see
6285 * curr_task() above) and restore that value before reenabling interrupts and
6286 * re-starting the system.
6288 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6290 void ia64_set_curr_task(int cpu, struct task_struct *p)
6297 #ifdef CONFIG_CGROUP_SCHED
6298 /* task_group_lock serializes the addition/removal of task groups */
6299 static DEFINE_SPINLOCK(task_group_lock);
6301 static void sched_free_group(struct task_group *tg)
6303 free_fair_sched_group(tg);
6304 free_rt_sched_group(tg);
6306 kmem_cache_free(task_group_cache, tg);
6309 /* allocate runqueue etc for a new task group */
6310 struct task_group *sched_create_group(struct task_group *parent)
6312 struct task_group *tg;
6314 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
6316 return ERR_PTR(-ENOMEM);
6318 if (!alloc_fair_sched_group(tg, parent))
6321 if (!alloc_rt_sched_group(tg, parent))
6327 sched_free_group(tg);
6328 return ERR_PTR(-ENOMEM);
6331 void sched_online_group(struct task_group *tg, struct task_group *parent)
6333 unsigned long flags;
6335 spin_lock_irqsave(&task_group_lock, flags);
6336 list_add_rcu(&tg->list, &task_groups);
6338 /* Root should already exist: */
6341 tg->parent = parent;
6342 INIT_LIST_HEAD(&tg->children);
6343 list_add_rcu(&tg->siblings, &parent->children);
6344 spin_unlock_irqrestore(&task_group_lock, flags);
6346 online_fair_sched_group(tg);
6349 /* rcu callback to free various structures associated with a task group */
6350 static void sched_free_group_rcu(struct rcu_head *rhp)
6352 /* Now it should be safe to free those cfs_rqs: */
6353 sched_free_group(container_of(rhp, struct task_group, rcu));
6356 void sched_destroy_group(struct task_group *tg)
6358 /* Wait for possible concurrent references to cfs_rqs complete: */
6359 call_rcu(&tg->rcu, sched_free_group_rcu);
6362 void sched_offline_group(struct task_group *tg)
6364 unsigned long flags;
6366 /* End participation in shares distribution: */
6367 unregister_fair_sched_group(tg);
6369 spin_lock_irqsave(&task_group_lock, flags);
6370 list_del_rcu(&tg->list);
6371 list_del_rcu(&tg->siblings);
6372 spin_unlock_irqrestore(&task_group_lock, flags);
6375 static void sched_change_group(struct task_struct *tsk, int type)
6377 struct task_group *tg;
6380 * All callers are synchronized by task_rq_lock(); we do not use RCU
6381 * which is pointless here. Thus, we pass "true" to task_css_check()
6382 * to prevent lockdep warnings.
6384 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
6385 struct task_group, css);
6386 tg = autogroup_task_group(tsk, tg);
6387 tsk->sched_task_group = tg;
6389 #ifdef CONFIG_FAIR_GROUP_SCHED
6390 if (tsk->sched_class->task_change_group)
6391 tsk->sched_class->task_change_group(tsk, type);
6394 set_task_rq(tsk, task_cpu(tsk));
6398 * Change task's runqueue when it moves between groups.
6400 * The caller of this function should have put the task in its new group by
6401 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6404 void sched_move_task(struct task_struct *tsk)
6406 int queued, running;
6410 rq = task_rq_lock(tsk, &rf);
6411 update_rq_clock(rq);
6413 running = task_current(rq, tsk);
6414 queued = task_on_rq_queued(tsk);
6417 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
6419 put_prev_task(rq, tsk);
6421 sched_change_group(tsk, TASK_MOVE_GROUP);
6424 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
6426 set_curr_task(rq, tsk);
6428 task_rq_unlock(rq, tsk, &rf);
6430 #endif /* CONFIG_CGROUP_SCHED */
6432 #ifdef CONFIG_RT_GROUP_SCHED
6434 * Ensure that the real time constraints are schedulable.
6436 static DEFINE_MUTEX(rt_constraints_mutex);
6438 /* Must be called with tasklist_lock held */
6439 static inline int tg_has_rt_tasks(struct task_group *tg)
6441 struct task_struct *g, *p;
6444 * Autogroups do not have RT tasks; see autogroup_create().
6446 if (task_group_is_autogroup(tg))
6449 for_each_process_thread(g, p) {
6450 if (rt_task(p) && task_group(p) == tg)
6457 struct rt_schedulable_data {
6458 struct task_group *tg;
6463 static int tg_rt_schedulable(struct task_group *tg, void *data)
6465 struct rt_schedulable_data *d = data;
6466 struct task_group *child;
6467 unsigned long total, sum = 0;
6468 u64 period, runtime;
6470 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6471 runtime = tg->rt_bandwidth.rt_runtime;
6474 period = d->rt_period;
6475 runtime = d->rt_runtime;
6479 * Cannot have more runtime than the period.
6481 if (runtime > period && runtime != RUNTIME_INF)
6485 * Ensure we don't starve existing RT tasks.
6487 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
6490 total = to_ratio(period, runtime);
6493 * Nobody can have more than the global setting allows.
6495 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
6499 * The sum of our children's runtime should not exceed our own.
6501 list_for_each_entry_rcu(child, &tg->children, siblings) {
6502 period = ktime_to_ns(child->rt_bandwidth.rt_period);
6503 runtime = child->rt_bandwidth.rt_runtime;
6505 if (child == d->tg) {
6506 period = d->rt_period;
6507 runtime = d->rt_runtime;
6510 sum += to_ratio(period, runtime);
6519 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
6523 struct rt_schedulable_data data = {
6525 .rt_period = period,
6526 .rt_runtime = runtime,
6530 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
6536 static int tg_set_rt_bandwidth(struct task_group *tg,
6537 u64 rt_period, u64 rt_runtime)
6542 * Disallowing the root group RT runtime is BAD, it would disallow the
6543 * kernel creating (and or operating) RT threads.
6545 if (tg == &root_task_group && rt_runtime == 0)
6548 /* No period doesn't make any sense. */
6552 mutex_lock(&rt_constraints_mutex);
6553 read_lock(&tasklist_lock);
6554 err = __rt_schedulable(tg, rt_period, rt_runtime);
6558 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6559 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
6560 tg->rt_bandwidth.rt_runtime = rt_runtime;
6562 for_each_possible_cpu(i) {
6563 struct rt_rq *rt_rq = tg->rt_rq[i];
6565 raw_spin_lock(&rt_rq->rt_runtime_lock);
6566 rt_rq->rt_runtime = rt_runtime;
6567 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6569 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6571 read_unlock(&tasklist_lock);
6572 mutex_unlock(&rt_constraints_mutex);
6577 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
6579 u64 rt_runtime, rt_period;
6581 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6582 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
6583 if (rt_runtime_us < 0)
6584 rt_runtime = RUNTIME_INF;
6586 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6589 static long sched_group_rt_runtime(struct task_group *tg)
6593 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
6596 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
6597 do_div(rt_runtime_us, NSEC_PER_USEC);
6598 return rt_runtime_us;
6601 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
6603 u64 rt_runtime, rt_period;
6605 rt_period = rt_period_us * NSEC_PER_USEC;
6606 rt_runtime = tg->rt_bandwidth.rt_runtime;
6608 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6611 static long sched_group_rt_period(struct task_group *tg)
6615 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
6616 do_div(rt_period_us, NSEC_PER_USEC);
6617 return rt_period_us;
6619 #endif /* CONFIG_RT_GROUP_SCHED */
6621 #ifdef CONFIG_RT_GROUP_SCHED
6622 static int sched_rt_global_constraints(void)
6626 mutex_lock(&rt_constraints_mutex);
6627 read_lock(&tasklist_lock);
6628 ret = __rt_schedulable(NULL, 0, 0);
6629 read_unlock(&tasklist_lock);
6630 mutex_unlock(&rt_constraints_mutex);
6635 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
6637 /* Don't accept realtime tasks when there is no way for them to run */
6638 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
6644 #else /* !CONFIG_RT_GROUP_SCHED */
6645 static int sched_rt_global_constraints(void)
6647 unsigned long flags;
6650 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
6651 for_each_possible_cpu(i) {
6652 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
6654 raw_spin_lock(&rt_rq->rt_runtime_lock);
6655 rt_rq->rt_runtime = global_rt_runtime();
6656 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6658 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
6662 #endif /* CONFIG_RT_GROUP_SCHED */
6664 static int sched_dl_global_validate(void)
6666 u64 runtime = global_rt_runtime();
6667 u64 period = global_rt_period();
6668 u64 new_bw = to_ratio(period, runtime);
6671 unsigned long flags;
6674 * Here we want to check the bandwidth not being set to some
6675 * value smaller than the currently allocated bandwidth in
6676 * any of the root_domains.
6678 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
6679 * cycling on root_domains... Discussion on different/better
6680 * solutions is welcome!
6682 for_each_possible_cpu(cpu) {
6683 rcu_read_lock_sched();
6684 dl_b = dl_bw_of(cpu);
6686 raw_spin_lock_irqsave(&dl_b->lock, flags);
6687 if (new_bw < dl_b->total_bw)
6689 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
6691 rcu_read_unlock_sched();
6700 static void sched_dl_do_global(void)
6705 unsigned long flags;
6707 def_dl_bandwidth.dl_period = global_rt_period();
6708 def_dl_bandwidth.dl_runtime = global_rt_runtime();
6710 if (global_rt_runtime() != RUNTIME_INF)
6711 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
6714 * FIXME: As above...
6716 for_each_possible_cpu(cpu) {
6717 rcu_read_lock_sched();
6718 dl_b = dl_bw_of(cpu);
6720 raw_spin_lock_irqsave(&dl_b->lock, flags);
6722 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
6724 rcu_read_unlock_sched();
6728 static int sched_rt_global_validate(void)
6730 if (sysctl_sched_rt_period <= 0)
6733 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
6734 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
6740 static void sched_rt_do_global(void)
6742 def_rt_bandwidth.rt_runtime = global_rt_runtime();
6743 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
6746 int sched_rt_handler(struct ctl_table *table, int write,
6747 void __user *buffer, size_t *lenp,
6750 int old_period, old_runtime;
6751 static DEFINE_MUTEX(mutex);
6755 old_period = sysctl_sched_rt_period;
6756 old_runtime = sysctl_sched_rt_runtime;
6758 ret = proc_dointvec(table, write, buffer, lenp, ppos);
6760 if (!ret && write) {
6761 ret = sched_rt_global_validate();
6765 ret = sched_dl_global_validate();
6769 ret = sched_rt_global_constraints();
6773 sched_rt_do_global();
6774 sched_dl_do_global();
6778 sysctl_sched_rt_period = old_period;
6779 sysctl_sched_rt_runtime = old_runtime;
6781 mutex_unlock(&mutex);
6786 int sched_rr_handler(struct ctl_table *table, int write,
6787 void __user *buffer, size_t *lenp,
6791 static DEFINE_MUTEX(mutex);
6794 ret = proc_dointvec(table, write, buffer, lenp, ppos);
6796 * Make sure that internally we keep jiffies.
6797 * Also, writing zero resets the timeslice to default:
6799 if (!ret && write) {
6800 sched_rr_timeslice =
6801 sysctl_sched_rr_timeslice <= 0 ? RR_TIMESLICE :
6802 msecs_to_jiffies(sysctl_sched_rr_timeslice);
6804 mutex_unlock(&mutex);
6808 #ifdef CONFIG_CGROUP_SCHED
6810 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6812 return css ? container_of(css, struct task_group, css) : NULL;
6815 static struct cgroup_subsys_state *
6816 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6818 struct task_group *parent = css_tg(parent_css);
6819 struct task_group *tg;
6822 /* This is early initialization for the top cgroup */
6823 return &root_task_group.css;
6826 tg = sched_create_group(parent);
6828 return ERR_PTR(-ENOMEM);
6833 /* Expose task group only after completing cgroup initialization */
6834 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6836 struct task_group *tg = css_tg(css);
6837 struct task_group *parent = css_tg(css->parent);
6840 sched_online_group(tg, parent);
6844 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
6846 struct task_group *tg = css_tg(css);
6848 sched_offline_group(tg);
6851 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6853 struct task_group *tg = css_tg(css);
6856 * Relies on the RCU grace period between css_released() and this.
6858 sched_free_group(tg);
6862 * This is called before wake_up_new_task(), therefore we really only
6863 * have to set its group bits, all the other stuff does not apply.
6865 static void cpu_cgroup_fork(struct task_struct *task)
6870 rq = task_rq_lock(task, &rf);
6872 update_rq_clock(rq);
6873 sched_change_group(task, TASK_SET_GROUP);
6875 task_rq_unlock(rq, task, &rf);
6878 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
6880 struct task_struct *task;
6881 struct cgroup_subsys_state *css;
6884 cgroup_taskset_for_each(task, css, tset) {
6885 #ifdef CONFIG_RT_GROUP_SCHED
6886 if (!sched_rt_can_attach(css_tg(css), task))
6889 /* We don't support RT-tasks being in separate groups */
6890 if (task->sched_class != &fair_sched_class)
6894 * Serialize against wake_up_new_task() such that if its
6895 * running, we're sure to observe its full state.
6897 raw_spin_lock_irq(&task->pi_lock);
6899 * Avoid calling sched_move_task() before wake_up_new_task()
6900 * has happened. This would lead to problems with PELT, due to
6901 * move wanting to detach+attach while we're not attached yet.
6903 if (task->state == TASK_NEW)
6905 raw_spin_unlock_irq(&task->pi_lock);
6913 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
6915 struct task_struct *task;
6916 struct cgroup_subsys_state *css;
6918 cgroup_taskset_for_each(task, css, tset)
6919 sched_move_task(task);
6922 #ifdef CONFIG_FAIR_GROUP_SCHED
6923 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
6924 struct cftype *cftype, u64 shareval)
6926 return sched_group_set_shares(css_tg(css), scale_load(shareval));
6929 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
6932 struct task_group *tg = css_tg(css);
6934 return (u64) scale_load_down(tg->shares);
6937 #ifdef CONFIG_CFS_BANDWIDTH
6938 static DEFINE_MUTEX(cfs_constraints_mutex);
6940 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
6941 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
6943 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
6945 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
6947 int i, ret = 0, runtime_enabled, runtime_was_enabled;
6948 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6950 if (tg == &root_task_group)
6954 * Ensure we have at some amount of bandwidth every period. This is
6955 * to prevent reaching a state of large arrears when throttled via
6956 * entity_tick() resulting in prolonged exit starvation.
6958 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
6962 * Likewise, bound things on the otherside by preventing insane quota
6963 * periods. This also allows us to normalize in computing quota
6966 if (period > max_cfs_quota_period)
6970 * Prevent race between setting of cfs_rq->runtime_enabled and
6971 * unthrottle_offline_cfs_rqs().
6974 mutex_lock(&cfs_constraints_mutex);
6975 ret = __cfs_schedulable(tg, period, quota);
6979 runtime_enabled = quota != RUNTIME_INF;
6980 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
6982 * If we need to toggle cfs_bandwidth_used, off->on must occur
6983 * before making related changes, and on->off must occur afterwards
6985 if (runtime_enabled && !runtime_was_enabled)
6986 cfs_bandwidth_usage_inc();
6987 raw_spin_lock_irq(&cfs_b->lock);
6988 cfs_b->period = ns_to_ktime(period);
6989 cfs_b->quota = quota;
6991 __refill_cfs_bandwidth_runtime(cfs_b);
6993 /* Restart the period timer (if active) to handle new period expiry: */
6994 if (runtime_enabled)
6995 start_cfs_bandwidth(cfs_b);
6997 raw_spin_unlock_irq(&cfs_b->lock);
6999 for_each_online_cpu(i) {
7000 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7001 struct rq *rq = cfs_rq->rq;
7003 raw_spin_lock_irq(&rq->lock);
7004 cfs_rq->runtime_enabled = runtime_enabled;
7005 cfs_rq->runtime_remaining = 0;
7007 if (cfs_rq->throttled)
7008 unthrottle_cfs_rq(cfs_rq);
7009 raw_spin_unlock_irq(&rq->lock);
7011 if (runtime_was_enabled && !runtime_enabled)
7012 cfs_bandwidth_usage_dec();
7014 mutex_unlock(&cfs_constraints_mutex);
7020 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7024 period = ktime_to_ns(tg->cfs_bandwidth.period);
7025 if (cfs_quota_us < 0)
7026 quota = RUNTIME_INF;
7028 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7030 return tg_set_cfs_bandwidth(tg, period, quota);
7033 long tg_get_cfs_quota(struct task_group *tg)
7037 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7040 quota_us = tg->cfs_bandwidth.quota;
7041 do_div(quota_us, NSEC_PER_USEC);
7046 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7050 period = (u64)cfs_period_us * NSEC_PER_USEC;
7051 quota = tg->cfs_bandwidth.quota;
7053 return tg_set_cfs_bandwidth(tg, period, quota);
7056 long tg_get_cfs_period(struct task_group *tg)
7060 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7061 do_div(cfs_period_us, NSEC_PER_USEC);
7063 return cfs_period_us;
7066 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7069 return tg_get_cfs_quota(css_tg(css));
7072 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7073 struct cftype *cftype, s64 cfs_quota_us)
7075 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7078 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7081 return tg_get_cfs_period(css_tg(css));
7084 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7085 struct cftype *cftype, u64 cfs_period_us)
7087 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7090 struct cfs_schedulable_data {
7091 struct task_group *tg;
7096 * normalize group quota/period to be quota/max_period
7097 * note: units are usecs
7099 static u64 normalize_cfs_quota(struct task_group *tg,
7100 struct cfs_schedulable_data *d)
7108 period = tg_get_cfs_period(tg);
7109 quota = tg_get_cfs_quota(tg);
7112 /* note: these should typically be equivalent */
7113 if (quota == RUNTIME_INF || quota == -1)
7116 return to_ratio(period, quota);
7119 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7121 struct cfs_schedulable_data *d = data;
7122 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7123 s64 quota = 0, parent_quota = -1;
7126 quota = RUNTIME_INF;
7128 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7130 quota = normalize_cfs_quota(tg, d);
7131 parent_quota = parent_b->hierarchical_quota;
7134 * Ensure max(child_quota) <= parent_quota, inherit when no
7137 if (quota == RUNTIME_INF)
7138 quota = parent_quota;
7139 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7142 cfs_b->hierarchical_quota = quota;
7147 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7150 struct cfs_schedulable_data data = {
7156 if (quota != RUNTIME_INF) {
7157 do_div(data.period, NSEC_PER_USEC);
7158 do_div(data.quota, NSEC_PER_USEC);
7162 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7168 static int cpu_stats_show(struct seq_file *sf, void *v)
7170 struct task_group *tg = css_tg(seq_css(sf));
7171 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7173 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7174 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7175 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7179 #endif /* CONFIG_CFS_BANDWIDTH */
7180 #endif /* CONFIG_FAIR_GROUP_SCHED */
7182 #ifdef CONFIG_RT_GROUP_SCHED
7183 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7184 struct cftype *cft, s64 val)
7186 return sched_group_set_rt_runtime(css_tg(css), val);
7189 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7192 return sched_group_rt_runtime(css_tg(css));
7195 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7196 struct cftype *cftype, u64 rt_period_us)
7198 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7201 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7204 return sched_group_rt_period(css_tg(css));
7206 #endif /* CONFIG_RT_GROUP_SCHED */
7208 static struct cftype cpu_files[] = {
7209 #ifdef CONFIG_FAIR_GROUP_SCHED
7212 .read_u64 = cpu_shares_read_u64,
7213 .write_u64 = cpu_shares_write_u64,
7216 #ifdef CONFIG_CFS_BANDWIDTH
7218 .name = "cfs_quota_us",
7219 .read_s64 = cpu_cfs_quota_read_s64,
7220 .write_s64 = cpu_cfs_quota_write_s64,
7223 .name = "cfs_period_us",
7224 .read_u64 = cpu_cfs_period_read_u64,
7225 .write_u64 = cpu_cfs_period_write_u64,
7229 .seq_show = cpu_stats_show,
7232 #ifdef CONFIG_RT_GROUP_SCHED
7234 .name = "rt_runtime_us",
7235 .read_s64 = cpu_rt_runtime_read,
7236 .write_s64 = cpu_rt_runtime_write,
7239 .name = "rt_period_us",
7240 .read_u64 = cpu_rt_period_read_uint,
7241 .write_u64 = cpu_rt_period_write_uint,
7247 struct cgroup_subsys cpu_cgrp_subsys = {
7248 .css_alloc = cpu_cgroup_css_alloc,
7249 .css_online = cpu_cgroup_css_online,
7250 .css_released = cpu_cgroup_css_released,
7251 .css_free = cpu_cgroup_css_free,
7252 .fork = cpu_cgroup_fork,
7253 .can_attach = cpu_cgroup_can_attach,
7254 .attach = cpu_cgroup_attach,
7255 .legacy_cftypes = cpu_files,
7259 #endif /* CONFIG_CGROUP_SCHED */
7261 void dump_cpu_task(int cpu)
7263 pr_info("Task dump for CPU %d:\n", cpu);
7264 sched_show_task(cpu_curr(cpu));
7268 * Nice levels are multiplicative, with a gentle 10% change for every
7269 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7270 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7271 * that remained on nice 0.
7273 * The "10% effect" is relative and cumulative: from _any_ nice level,
7274 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7275 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7276 * If a task goes up by ~10% and another task goes down by ~10% then
7277 * the relative distance between them is ~25%.)
7279 const int sched_prio_to_weight[40] = {
7280 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7281 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7282 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7283 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7284 /* 0 */ 1024, 820, 655, 526, 423,
7285 /* 5 */ 335, 272, 215, 172, 137,
7286 /* 10 */ 110, 87, 70, 56, 45,
7287 /* 15 */ 36, 29, 23, 18, 15,
7291 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7293 * In cases where the weight does not change often, we can use the
7294 * precalculated inverse to speed up arithmetics by turning divisions
7295 * into multiplications:
7297 const u32 sched_prio_to_wmult[40] = {
7298 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7299 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7300 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7301 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7302 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7303 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7304 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7305 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,