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>
14 #include <linux/blkdev.h>
15 #include <linux/kprobes.h>
16 #include <linux/mmu_context.h>
17 #include <linux/module.h>
18 #include <linux/nmi.h>
19 #include <linux/prefetch.h>
20 #include <linux/profile.h>
21 #include <linux/security.h>
22 #include <linux/syscalls.h>
24 #include <asm/switch_to.h>
28 #include "../workqueue_internal.h"
29 #include "../smpboot.h"
31 #define CREATE_TRACE_POINTS
32 #include <trace/events/sched.h>
34 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
37 * Debugging: various feature bits
40 #define SCHED_FEAT(name, enabled) \
41 (1UL << __SCHED_FEAT_##name) * enabled |
43 const_debug
unsigned int sysctl_sched_features
=
50 * Number of tasks to iterate in a single balance run.
51 * Limited because this is done with IRQs disabled.
53 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
56 * period over which we average the RT time consumption, measured
61 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
64 * period over which we measure -rt task CPU usage in us.
67 unsigned int sysctl_sched_rt_period
= 1000000;
69 __read_mostly
int scheduler_running
;
72 * part of the period that we allow rt tasks to run in us.
75 int sysctl_sched_rt_runtime
= 950000;
77 /* CPUs with isolated domains */
78 cpumask_var_t cpu_isolated_map
;
81 * this_rq_lock - lock this runqueue and disable interrupts.
83 static struct rq
*this_rq_lock(void)
90 raw_spin_lock(&rq
->lock
);
96 * __task_rq_lock - lock the rq @p resides on.
98 struct rq
*__task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
103 lockdep_assert_held(&p
->pi_lock
);
107 raw_spin_lock(&rq
->lock
);
108 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
112 raw_spin_unlock(&rq
->lock
);
114 while (unlikely(task_on_rq_migrating(p
)))
120 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
122 struct rq
*task_rq_lock(struct task_struct
*p
, struct rq_flags
*rf
)
123 __acquires(p
->pi_lock
)
129 raw_spin_lock_irqsave(&p
->pi_lock
, rf
->flags
);
131 raw_spin_lock(&rq
->lock
);
133 * move_queued_task() task_rq_lock()
136 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
137 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
138 * [S] ->cpu = new_cpu [L] task_rq()
142 * If we observe the old cpu in task_rq_lock, the acquire of
143 * the old rq->lock will fully serialize against the stores.
145 * If we observe the new CPU in task_rq_lock, the acquire will
146 * pair with the WMB to ensure we must then also see migrating.
148 if (likely(rq
== task_rq(p
) && !task_on_rq_migrating(p
))) {
152 raw_spin_unlock(&rq
->lock
);
153 raw_spin_unlock_irqrestore(&p
->pi_lock
, rf
->flags
);
155 while (unlikely(task_on_rq_migrating(p
)))
161 * RQ-clock updating methods:
164 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
167 * In theory, the compile should just see 0 here, and optimize out the call
168 * to sched_rt_avg_update. But I don't trust it...
170 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
171 s64 steal
= 0, irq_delta
= 0;
173 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
174 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
177 * Since irq_time is only updated on {soft,}irq_exit, we might run into
178 * this case when a previous update_rq_clock() happened inside a
181 * When this happens, we stop ->clock_task and only update the
182 * prev_irq_time stamp to account for the part that fit, so that a next
183 * update will consume the rest. This ensures ->clock_task is
186 * It does however cause some slight miss-attribution of {soft,}irq
187 * time, a more accurate solution would be to update the irq_time using
188 * the current rq->clock timestamp, except that would require using
191 if (irq_delta
> delta
)
194 rq
->prev_irq_time
+= irq_delta
;
197 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
198 if (static_key_false((¶virt_steal_rq_enabled
))) {
199 steal
= paravirt_steal_clock(cpu_of(rq
));
200 steal
-= rq
->prev_steal_time_rq
;
202 if (unlikely(steal
> delta
))
205 rq
->prev_steal_time_rq
+= steal
;
210 rq
->clock_task
+= delta
;
212 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
213 if ((irq_delta
+ steal
) && sched_feat(NONTASK_CAPACITY
))
214 sched_rt_avg_update(rq
, irq_delta
+ steal
);
218 void update_rq_clock(struct rq
*rq
)
222 lockdep_assert_held(&rq
->lock
);
224 if (rq
->clock_update_flags
& RQCF_ACT_SKIP
)
227 #ifdef CONFIG_SCHED_DEBUG
228 rq
->clock_update_flags
|= RQCF_UPDATED
;
230 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
234 update_rq_clock_task(rq
, delta
);
238 #ifdef CONFIG_SCHED_HRTICK
240 * Use HR-timers to deliver accurate preemption points.
243 static void hrtick_clear(struct rq
*rq
)
245 if (hrtimer_active(&rq
->hrtick_timer
))
246 hrtimer_cancel(&rq
->hrtick_timer
);
250 * High-resolution timer tick.
251 * Runs from hardirq context with interrupts disabled.
253 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
255 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
257 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
259 raw_spin_lock(&rq
->lock
);
261 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
262 raw_spin_unlock(&rq
->lock
);
264 return HRTIMER_NORESTART
;
269 static void __hrtick_restart(struct rq
*rq
)
271 struct hrtimer
*timer
= &rq
->hrtick_timer
;
273 hrtimer_start_expires(timer
, HRTIMER_MODE_ABS_PINNED
);
277 * called from hardirq (IPI) context
279 static void __hrtick_start(void *arg
)
283 raw_spin_lock(&rq
->lock
);
284 __hrtick_restart(rq
);
285 rq
->hrtick_csd_pending
= 0;
286 raw_spin_unlock(&rq
->lock
);
290 * Called to set the hrtick timer state.
292 * called with rq->lock held and irqs disabled
294 void hrtick_start(struct rq
*rq
, u64 delay
)
296 struct hrtimer
*timer
= &rq
->hrtick_timer
;
301 * Don't schedule slices shorter than 10000ns, that just
302 * doesn't make sense and can cause timer DoS.
304 delta
= max_t(s64
, delay
, 10000LL);
305 time
= ktime_add_ns(timer
->base
->get_time(), delta
);
307 hrtimer_set_expires(timer
, time
);
309 if (rq
== this_rq()) {
310 __hrtick_restart(rq
);
311 } else if (!rq
->hrtick_csd_pending
) {
312 smp_call_function_single_async(cpu_of(rq
), &rq
->hrtick_csd
);
313 rq
->hrtick_csd_pending
= 1;
319 * Called to set the hrtick timer state.
321 * called with rq->lock held and irqs disabled
323 void hrtick_start(struct rq
*rq
, u64 delay
)
326 * Don't schedule slices shorter than 10000ns, that just
327 * doesn't make sense. Rely on vruntime for fairness.
329 delay
= max_t(u64
, delay
, 10000LL);
330 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
),
331 HRTIMER_MODE_REL_PINNED
);
333 #endif /* CONFIG_SMP */
335 static void init_rq_hrtick(struct rq
*rq
)
338 rq
->hrtick_csd_pending
= 0;
340 rq
->hrtick_csd
.flags
= 0;
341 rq
->hrtick_csd
.func
= __hrtick_start
;
342 rq
->hrtick_csd
.info
= rq
;
345 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
346 rq
->hrtick_timer
.function
= hrtick
;
348 #else /* CONFIG_SCHED_HRTICK */
349 static inline void hrtick_clear(struct rq
*rq
)
353 static inline void init_rq_hrtick(struct rq
*rq
)
356 #endif /* CONFIG_SCHED_HRTICK */
359 * cmpxchg based fetch_or, macro so it works for different integer types
361 #define fetch_or(ptr, mask) \
363 typeof(ptr) _ptr = (ptr); \
364 typeof(mask) _mask = (mask); \
365 typeof(*_ptr) _old, _val = *_ptr; \
368 _old = cmpxchg(_ptr, _val, _val | _mask); \
376 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
378 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
379 * this avoids any races wrt polling state changes and thereby avoids
382 static bool set_nr_and_not_polling(struct task_struct
*p
)
384 struct thread_info
*ti
= task_thread_info(p
);
385 return !(fetch_or(&ti
->flags
, _TIF_NEED_RESCHED
) & _TIF_POLLING_NRFLAG
);
389 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
391 * If this returns true, then the idle task promises to call
392 * sched_ttwu_pending() and reschedule soon.
394 static bool set_nr_if_polling(struct task_struct
*p
)
396 struct thread_info
*ti
= task_thread_info(p
);
397 typeof(ti
->flags
) old
, val
= READ_ONCE(ti
->flags
);
400 if (!(val
& _TIF_POLLING_NRFLAG
))
402 if (val
& _TIF_NEED_RESCHED
)
404 old
= cmpxchg(&ti
->flags
, val
, val
| _TIF_NEED_RESCHED
);
413 static bool set_nr_and_not_polling(struct task_struct
*p
)
415 set_tsk_need_resched(p
);
420 static bool set_nr_if_polling(struct task_struct
*p
)
427 void wake_q_add(struct wake_q_head
*head
, struct task_struct
*task
)
429 struct wake_q_node
*node
= &task
->wake_q
;
432 * Atomically grab the task, if ->wake_q is !nil already it means
433 * its already queued (either by us or someone else) and will get the
434 * wakeup due to that.
436 * This cmpxchg() implies a full barrier, which pairs with the write
437 * barrier implied by the wakeup in wake_up_q().
439 if (cmpxchg(&node
->next
, NULL
, WAKE_Q_TAIL
))
442 get_task_struct(task
);
445 * The head is context local, there can be no concurrency.
448 head
->lastp
= &node
->next
;
451 void wake_up_q(struct wake_q_head
*head
)
453 struct wake_q_node
*node
= head
->first
;
455 while (node
!= WAKE_Q_TAIL
) {
456 struct task_struct
*task
;
458 task
= container_of(node
, struct task_struct
, wake_q
);
460 /* Task can safely be re-inserted now: */
462 task
->wake_q
.next
= NULL
;
465 * wake_up_process() implies a wmb() to pair with the queueing
466 * in wake_q_add() so as not to miss wakeups.
468 wake_up_process(task
);
469 put_task_struct(task
);
474 * resched_curr - mark rq's current task 'to be rescheduled now'.
476 * On UP this means the setting of the need_resched flag, on SMP it
477 * might also involve a cross-CPU call to trigger the scheduler on
480 void resched_curr(struct rq
*rq
)
482 struct task_struct
*curr
= rq
->curr
;
485 lockdep_assert_held(&rq
->lock
);
487 if (test_tsk_need_resched(curr
))
492 if (cpu
== smp_processor_id()) {
493 set_tsk_need_resched(curr
);
494 set_preempt_need_resched();
498 if (set_nr_and_not_polling(curr
))
499 smp_send_reschedule(cpu
);
501 trace_sched_wake_idle_without_ipi(cpu
);
504 void resched_cpu(int cpu
)
506 struct rq
*rq
= cpu_rq(cpu
);
509 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
512 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
516 #ifdef CONFIG_NO_HZ_COMMON
518 * In the semi idle case, use the nearest busy CPU for migrating timers
519 * from an idle CPU. This is good for power-savings.
521 * We don't do similar optimization for completely idle system, as
522 * selecting an idle CPU will add more delays to the timers than intended
523 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
525 int get_nohz_timer_target(void)
527 int i
, cpu
= smp_processor_id();
528 struct sched_domain
*sd
;
530 if (!idle_cpu(cpu
) && is_housekeeping_cpu(cpu
))
534 for_each_domain(cpu
, sd
) {
535 for_each_cpu(i
, sched_domain_span(sd
)) {
539 if (!idle_cpu(i
) && is_housekeeping_cpu(i
)) {
546 if (!is_housekeeping_cpu(cpu
))
547 cpu
= housekeeping_any_cpu();
554 * When add_timer_on() enqueues a timer into the timer wheel of an
555 * idle CPU then this timer might expire before the next timer event
556 * which is scheduled to wake up that CPU. In case of a completely
557 * idle system the next event might even be infinite time into the
558 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
559 * leaves the inner idle loop so the newly added timer is taken into
560 * account when the CPU goes back to idle and evaluates the timer
561 * wheel for the next timer event.
563 static void wake_up_idle_cpu(int cpu
)
565 struct rq
*rq
= cpu_rq(cpu
);
567 if (cpu
== smp_processor_id())
570 if (set_nr_and_not_polling(rq
->idle
))
571 smp_send_reschedule(cpu
);
573 trace_sched_wake_idle_without_ipi(cpu
);
576 static bool wake_up_full_nohz_cpu(int cpu
)
579 * We just need the target to call irq_exit() and re-evaluate
580 * the next tick. The nohz full kick at least implies that.
581 * If needed we can still optimize that later with an
584 if (cpu_is_offline(cpu
))
585 return true; /* Don't try to wake offline CPUs. */
586 if (tick_nohz_full_cpu(cpu
)) {
587 if (cpu
!= smp_processor_id() ||
588 tick_nohz_tick_stopped())
589 tick_nohz_full_kick_cpu(cpu
);
597 * Wake up the specified CPU. If the CPU is going offline, it is the
598 * caller's responsibility to deal with the lost wakeup, for example,
599 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
601 void wake_up_nohz_cpu(int cpu
)
603 if (!wake_up_full_nohz_cpu(cpu
))
604 wake_up_idle_cpu(cpu
);
607 static inline bool got_nohz_idle_kick(void)
609 int cpu
= smp_processor_id();
611 if (!test_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
)))
614 if (idle_cpu(cpu
) && !need_resched())
618 * We can't run Idle Load Balance on this CPU for this time so we
619 * cancel it and clear NOHZ_BALANCE_KICK
621 clear_bit(NOHZ_BALANCE_KICK
, nohz_flags(cpu
));
625 #else /* CONFIG_NO_HZ_COMMON */
627 static inline bool got_nohz_idle_kick(void)
632 #endif /* CONFIG_NO_HZ_COMMON */
634 #ifdef CONFIG_NO_HZ_FULL
635 bool sched_can_stop_tick(struct rq
*rq
)
639 /* Deadline tasks, even if single, need the tick */
640 if (rq
->dl
.dl_nr_running
)
644 * If there are more than one RR tasks, we need the tick to effect the
645 * actual RR behaviour.
647 if (rq
->rt
.rr_nr_running
) {
648 if (rq
->rt
.rr_nr_running
== 1)
655 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
656 * forced preemption between FIFO tasks.
658 fifo_nr_running
= rq
->rt
.rt_nr_running
- rq
->rt
.rr_nr_running
;
663 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
664 * if there's more than one we need the tick for involuntary
667 if (rq
->nr_running
> 1)
672 #endif /* CONFIG_NO_HZ_FULL */
674 void sched_avg_update(struct rq
*rq
)
676 s64 period
= sched_avg_period();
678 while ((s64
)(rq_clock(rq
) - rq
->age_stamp
) > period
) {
680 * Inline assembly required to prevent the compiler
681 * optimising this loop into a divmod call.
682 * See __iter_div_u64_rem() for another example of this.
684 asm("" : "+rm" (rq
->age_stamp
));
685 rq
->age_stamp
+= period
;
690 #endif /* CONFIG_SMP */
692 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
693 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
695 * Iterate task_group tree rooted at *from, calling @down when first entering a
696 * node and @up when leaving it for the final time.
698 * Caller must hold rcu_lock or sufficient equivalent.
700 int walk_tg_tree_from(struct task_group
*from
,
701 tg_visitor down
, tg_visitor up
, void *data
)
703 struct task_group
*parent
, *child
;
709 ret
= (*down
)(parent
, data
);
712 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
719 ret
= (*up
)(parent
, data
);
720 if (ret
|| parent
== from
)
724 parent
= parent
->parent
;
731 int tg_nop(struct task_group
*tg
, void *data
)
737 static void set_load_weight(struct task_struct
*p
)
739 int prio
= p
->static_prio
- MAX_RT_PRIO
;
740 struct load_weight
*load
= &p
->se
.load
;
743 * SCHED_IDLE tasks get minimal weight:
745 if (idle_policy(p
->policy
)) {
746 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
747 load
->inv_weight
= WMULT_IDLEPRIO
;
751 load
->weight
= scale_load(sched_prio_to_weight
[prio
]);
752 load
->inv_weight
= sched_prio_to_wmult
[prio
];
755 static inline void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
758 if (!(flags
& ENQUEUE_RESTORE
))
759 sched_info_queued(rq
, p
);
760 p
->sched_class
->enqueue_task(rq
, p
, flags
);
763 static inline void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
766 if (!(flags
& DEQUEUE_SAVE
))
767 sched_info_dequeued(rq
, p
);
768 p
->sched_class
->dequeue_task(rq
, p
, flags
);
771 void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
773 if (task_contributes_to_load(p
))
774 rq
->nr_uninterruptible
--;
776 enqueue_task(rq
, p
, flags
);
779 void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
781 if (task_contributes_to_load(p
))
782 rq
->nr_uninterruptible
++;
784 dequeue_task(rq
, p
, flags
);
787 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
789 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
790 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
794 * Make it appear like a SCHED_FIFO task, its something
795 * userspace knows about and won't get confused about.
797 * Also, it will make PI more or less work without too
798 * much confusion -- but then, stop work should not
799 * rely on PI working anyway.
801 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
803 stop
->sched_class
= &stop_sched_class
;
806 cpu_rq(cpu
)->stop
= stop
;
810 * Reset it back to a normal scheduling class so that
811 * it can die in pieces.
813 old_stop
->sched_class
= &rt_sched_class
;
818 * __normal_prio - return the priority that is based on the static prio
820 static inline int __normal_prio(struct task_struct
*p
)
822 return p
->static_prio
;
826 * Calculate the expected normal priority: i.e. priority
827 * without taking RT-inheritance into account. Might be
828 * boosted by interactivity modifiers. Changes upon fork,
829 * setprio syscalls, and whenever the interactivity
830 * estimator recalculates.
832 static inline int normal_prio(struct task_struct
*p
)
836 if (task_has_dl_policy(p
))
837 prio
= MAX_DL_PRIO
-1;
838 else if (task_has_rt_policy(p
))
839 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
841 prio
= __normal_prio(p
);
846 * Calculate the current priority, i.e. the priority
847 * taken into account by the scheduler. This value might
848 * be boosted by RT tasks, or might be boosted by
849 * interactivity modifiers. Will be RT if the task got
850 * RT-boosted. If not then it returns p->normal_prio.
852 static int effective_prio(struct task_struct
*p
)
854 p
->normal_prio
= normal_prio(p
);
856 * If we are RT tasks or we were boosted to RT priority,
857 * keep the priority unchanged. Otherwise, update priority
858 * to the normal priority:
860 if (!rt_prio(p
->prio
))
861 return p
->normal_prio
;
866 * task_curr - is this task currently executing on a CPU?
867 * @p: the task in question.
869 * Return: 1 if the task is currently executing. 0 otherwise.
871 inline int task_curr(const struct task_struct
*p
)
873 return cpu_curr(task_cpu(p
)) == p
;
877 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
878 * use the balance_callback list if you want balancing.
880 * this means any call to check_class_changed() must be followed by a call to
881 * balance_callback().
883 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
884 const struct sched_class
*prev_class
,
887 if (prev_class
!= p
->sched_class
) {
888 if (prev_class
->switched_from
)
889 prev_class
->switched_from(rq
, p
);
891 p
->sched_class
->switched_to(rq
, p
);
892 } else if (oldprio
!= p
->prio
|| dl_task(p
))
893 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
896 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
898 const struct sched_class
*class;
900 if (p
->sched_class
== rq
->curr
->sched_class
) {
901 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
903 for_each_class(class) {
904 if (class == rq
->curr
->sched_class
)
906 if (class == p
->sched_class
) {
914 * A queue event has occurred, and we're going to schedule. In
915 * this case, we can save a useless back to back clock update.
917 if (task_on_rq_queued(rq
->curr
) && test_tsk_need_resched(rq
->curr
))
918 rq_clock_skip_update(rq
, true);
923 * This is how migration works:
925 * 1) we invoke migration_cpu_stop() on the target CPU using
927 * 2) stopper starts to run (implicitly forcing the migrated thread
929 * 3) it checks whether the migrated task is still in the wrong runqueue.
930 * 4) if it's in the wrong runqueue then the migration thread removes
931 * it and puts it into the right queue.
932 * 5) stopper completes and stop_one_cpu() returns and the migration
937 * move_queued_task - move a queued task to new rq.
939 * Returns (locked) new rq. Old rq's lock is released.
941 static struct rq
*move_queued_task(struct rq
*rq
, struct task_struct
*p
, int new_cpu
)
943 lockdep_assert_held(&rq
->lock
);
945 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
946 dequeue_task(rq
, p
, 0);
947 set_task_cpu(p
, new_cpu
);
948 raw_spin_unlock(&rq
->lock
);
950 rq
= cpu_rq(new_cpu
);
952 raw_spin_lock(&rq
->lock
);
953 BUG_ON(task_cpu(p
) != new_cpu
);
954 enqueue_task(rq
, p
, 0);
955 p
->on_rq
= TASK_ON_RQ_QUEUED
;
956 check_preempt_curr(rq
, p
, 0);
961 struct migration_arg
{
962 struct task_struct
*task
;
967 * Move (not current) task off this CPU, onto the destination CPU. We're doing
968 * this because either it can't run here any more (set_cpus_allowed()
969 * away from this CPU, or CPU going down), or because we're
970 * attempting to rebalance this task on exec (sched_exec).
972 * So we race with normal scheduler movements, but that's OK, as long
973 * as the task is no longer on this CPU.
975 static struct rq
*__migrate_task(struct rq
*rq
, struct task_struct
*p
, int dest_cpu
)
977 if (unlikely(!cpu_active(dest_cpu
)))
980 /* Affinity changed (again). */
981 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
984 rq
= move_queued_task(rq
, p
, dest_cpu
);
990 * migration_cpu_stop - this will be executed by a highprio stopper thread
991 * and performs thread migration by bumping thread off CPU then
992 * 'pushing' onto another runqueue.
994 static int migration_cpu_stop(void *data
)
996 struct migration_arg
*arg
= data
;
997 struct task_struct
*p
= arg
->task
;
998 struct rq
*rq
= this_rq();
1001 * The original target CPU might have gone down and we might
1002 * be on another CPU but it doesn't matter.
1004 local_irq_disable();
1006 * We need to explicitly wake pending tasks before running
1007 * __migrate_task() such that we will not miss enforcing cpus_allowed
1008 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1010 sched_ttwu_pending();
1012 raw_spin_lock(&p
->pi_lock
);
1013 raw_spin_lock(&rq
->lock
);
1015 * If task_rq(p) != rq, it cannot be migrated here, because we're
1016 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1017 * we're holding p->pi_lock.
1019 if (task_rq(p
) == rq
) {
1020 if (task_on_rq_queued(p
))
1021 rq
= __migrate_task(rq
, p
, arg
->dest_cpu
);
1023 p
->wake_cpu
= arg
->dest_cpu
;
1025 raw_spin_unlock(&rq
->lock
);
1026 raw_spin_unlock(&p
->pi_lock
);
1033 * sched_class::set_cpus_allowed must do the below, but is not required to
1034 * actually call this function.
1036 void set_cpus_allowed_common(struct task_struct
*p
, const struct cpumask
*new_mask
)
1038 cpumask_copy(&p
->cpus_allowed
, new_mask
);
1039 p
->nr_cpus_allowed
= cpumask_weight(new_mask
);
1042 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
1044 struct rq
*rq
= task_rq(p
);
1045 bool queued
, running
;
1047 lockdep_assert_held(&p
->pi_lock
);
1049 queued
= task_on_rq_queued(p
);
1050 running
= task_current(rq
, p
);
1054 * Because __kthread_bind() calls this on blocked tasks without
1057 lockdep_assert_held(&rq
->lock
);
1058 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
1061 put_prev_task(rq
, p
);
1063 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
1066 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
1068 set_curr_task(rq
, p
);
1072 * Change a given task's CPU affinity. Migrate the thread to a
1073 * proper CPU and schedule it away if the CPU it's executing on
1074 * is removed from the allowed bitmask.
1076 * NOTE: the caller must have a valid reference to the task, the
1077 * task must not exit() & deallocate itself prematurely. The
1078 * call is not atomic; no spinlocks may be held.
1080 static int __set_cpus_allowed_ptr(struct task_struct
*p
,
1081 const struct cpumask
*new_mask
, bool check
)
1083 const struct cpumask
*cpu_valid_mask
= cpu_active_mask
;
1084 unsigned int dest_cpu
;
1089 rq
= task_rq_lock(p
, &rf
);
1091 if (p
->flags
& PF_KTHREAD
) {
1093 * Kernel threads are allowed on online && !active CPUs
1095 cpu_valid_mask
= cpu_online_mask
;
1099 * Must re-check here, to close a race against __kthread_bind(),
1100 * sched_setaffinity() is not guaranteed to observe the flag.
1102 if (check
&& (p
->flags
& PF_NO_SETAFFINITY
)) {
1107 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
1110 if (!cpumask_intersects(new_mask
, cpu_valid_mask
)) {
1115 do_set_cpus_allowed(p
, new_mask
);
1117 if (p
->flags
& PF_KTHREAD
) {
1119 * For kernel threads that do indeed end up on online &&
1120 * !active we want to ensure they are strict per-CPU threads.
1122 WARN_ON(cpumask_intersects(new_mask
, cpu_online_mask
) &&
1123 !cpumask_intersects(new_mask
, cpu_active_mask
) &&
1124 p
->nr_cpus_allowed
!= 1);
1127 /* Can the task run on the task's current CPU? If so, we're done */
1128 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
1131 dest_cpu
= cpumask_any_and(cpu_valid_mask
, new_mask
);
1132 if (task_running(rq
, p
) || p
->state
== TASK_WAKING
) {
1133 struct migration_arg arg
= { p
, dest_cpu
};
1134 /* Need help from migration thread: drop lock and wait. */
1135 task_rq_unlock(rq
, p
, &rf
);
1136 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
1137 tlb_migrate_finish(p
->mm
);
1139 } else if (task_on_rq_queued(p
)) {
1141 * OK, since we're going to drop the lock immediately
1142 * afterwards anyway.
1144 rq_unpin_lock(rq
, &rf
);
1145 rq
= move_queued_task(rq
, p
, dest_cpu
);
1146 rq_repin_lock(rq
, &rf
);
1149 task_rq_unlock(rq
, p
, &rf
);
1154 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
1156 return __set_cpus_allowed_ptr(p
, new_mask
, false);
1158 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
1160 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1162 #ifdef CONFIG_SCHED_DEBUG
1164 * We should never call set_task_cpu() on a blocked task,
1165 * ttwu() will sort out the placement.
1167 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1171 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1172 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1173 * time relying on p->on_rq.
1175 WARN_ON_ONCE(p
->state
== TASK_RUNNING
&&
1176 p
->sched_class
== &fair_sched_class
&&
1177 (p
->on_rq
&& !task_on_rq_migrating(p
)));
1179 #ifdef CONFIG_LOCKDEP
1181 * The caller should hold either p->pi_lock or rq->lock, when changing
1182 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1184 * sched_move_task() holds both and thus holding either pins the cgroup,
1187 * Furthermore, all task_rq users should acquire both locks, see
1190 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
1191 lockdep_is_held(&task_rq(p
)->lock
)));
1195 trace_sched_migrate_task(p
, new_cpu
);
1197 if (task_cpu(p
) != new_cpu
) {
1198 if (p
->sched_class
->migrate_task_rq
)
1199 p
->sched_class
->migrate_task_rq(p
);
1200 p
->se
.nr_migrations
++;
1201 perf_event_task_migrate(p
);
1204 __set_task_cpu(p
, new_cpu
);
1207 static void __migrate_swap_task(struct task_struct
*p
, int cpu
)
1209 if (task_on_rq_queued(p
)) {
1210 struct rq
*src_rq
, *dst_rq
;
1212 src_rq
= task_rq(p
);
1213 dst_rq
= cpu_rq(cpu
);
1215 p
->on_rq
= TASK_ON_RQ_MIGRATING
;
1216 deactivate_task(src_rq
, p
, 0);
1217 set_task_cpu(p
, cpu
);
1218 activate_task(dst_rq
, p
, 0);
1219 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1220 check_preempt_curr(dst_rq
, p
, 0);
1223 * Task isn't running anymore; make it appear like we migrated
1224 * it before it went to sleep. This means on wakeup we make the
1225 * previous CPU our target instead of where it really is.
1231 struct migration_swap_arg
{
1232 struct task_struct
*src_task
, *dst_task
;
1233 int src_cpu
, dst_cpu
;
1236 static int migrate_swap_stop(void *data
)
1238 struct migration_swap_arg
*arg
= data
;
1239 struct rq
*src_rq
, *dst_rq
;
1242 if (!cpu_active(arg
->src_cpu
) || !cpu_active(arg
->dst_cpu
))
1245 src_rq
= cpu_rq(arg
->src_cpu
);
1246 dst_rq
= cpu_rq(arg
->dst_cpu
);
1248 double_raw_lock(&arg
->src_task
->pi_lock
,
1249 &arg
->dst_task
->pi_lock
);
1250 double_rq_lock(src_rq
, dst_rq
);
1252 if (task_cpu(arg
->dst_task
) != arg
->dst_cpu
)
1255 if (task_cpu(arg
->src_task
) != arg
->src_cpu
)
1258 if (!cpumask_test_cpu(arg
->dst_cpu
, tsk_cpus_allowed(arg
->src_task
)))
1261 if (!cpumask_test_cpu(arg
->src_cpu
, tsk_cpus_allowed(arg
->dst_task
)))
1264 __migrate_swap_task(arg
->src_task
, arg
->dst_cpu
);
1265 __migrate_swap_task(arg
->dst_task
, arg
->src_cpu
);
1270 double_rq_unlock(src_rq
, dst_rq
);
1271 raw_spin_unlock(&arg
->dst_task
->pi_lock
);
1272 raw_spin_unlock(&arg
->src_task
->pi_lock
);
1278 * Cross migrate two tasks
1280 int migrate_swap(struct task_struct
*cur
, struct task_struct
*p
)
1282 struct migration_swap_arg arg
;
1285 arg
= (struct migration_swap_arg
){
1287 .src_cpu
= task_cpu(cur
),
1289 .dst_cpu
= task_cpu(p
),
1292 if (arg
.src_cpu
== arg
.dst_cpu
)
1296 * These three tests are all lockless; this is OK since all of them
1297 * will be re-checked with proper locks held further down the line.
1299 if (!cpu_active(arg
.src_cpu
) || !cpu_active(arg
.dst_cpu
))
1302 if (!cpumask_test_cpu(arg
.dst_cpu
, tsk_cpus_allowed(arg
.src_task
)))
1305 if (!cpumask_test_cpu(arg
.src_cpu
, tsk_cpus_allowed(arg
.dst_task
)))
1308 trace_sched_swap_numa(cur
, arg
.src_cpu
, p
, arg
.dst_cpu
);
1309 ret
= stop_two_cpus(arg
.dst_cpu
, arg
.src_cpu
, migrate_swap_stop
, &arg
);
1316 * wait_task_inactive - wait for a thread to unschedule.
1318 * If @match_state is nonzero, it's the @p->state value just checked and
1319 * not expected to change. If it changes, i.e. @p might have woken up,
1320 * then return zero. When we succeed in waiting for @p to be off its CPU,
1321 * we return a positive number (its total switch count). If a second call
1322 * a short while later returns the same number, the caller can be sure that
1323 * @p has remained unscheduled the whole time.
1325 * The caller must ensure that the task *will* unschedule sometime soon,
1326 * else this function might spin for a *long* time. This function can't
1327 * be called with interrupts off, or it may introduce deadlock with
1328 * smp_call_function() if an IPI is sent by the same process we are
1329 * waiting to become inactive.
1331 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1333 int running
, queued
;
1340 * We do the initial early heuristics without holding
1341 * any task-queue locks at all. We'll only try to get
1342 * the runqueue lock when things look like they will
1348 * If the task is actively running on another CPU
1349 * still, just relax and busy-wait without holding
1352 * NOTE! Since we don't hold any locks, it's not
1353 * even sure that "rq" stays as the right runqueue!
1354 * But we don't care, since "task_running()" will
1355 * return false if the runqueue has changed and p
1356 * is actually now running somewhere else!
1358 while (task_running(rq
, p
)) {
1359 if (match_state
&& unlikely(p
->state
!= match_state
))
1365 * Ok, time to look more closely! We need the rq
1366 * lock now, to be *sure*. If we're wrong, we'll
1367 * just go back and repeat.
1369 rq
= task_rq_lock(p
, &rf
);
1370 trace_sched_wait_task(p
);
1371 running
= task_running(rq
, p
);
1372 queued
= task_on_rq_queued(p
);
1374 if (!match_state
|| p
->state
== match_state
)
1375 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1376 task_rq_unlock(rq
, p
, &rf
);
1379 * If it changed from the expected state, bail out now.
1381 if (unlikely(!ncsw
))
1385 * Was it really running after all now that we
1386 * checked with the proper locks actually held?
1388 * Oops. Go back and try again..
1390 if (unlikely(running
)) {
1396 * It's not enough that it's not actively running,
1397 * it must be off the runqueue _entirely_, and not
1400 * So if it was still runnable (but just not actively
1401 * running right now), it's preempted, and we should
1402 * yield - it could be a while.
1404 if (unlikely(queued
)) {
1405 ktime_t to
= NSEC_PER_SEC
/ HZ
;
1407 set_current_state(TASK_UNINTERRUPTIBLE
);
1408 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
1413 * Ahh, all good. It wasn't running, and it wasn't
1414 * runnable, which means that it will never become
1415 * running in the future either. We're all done!
1424 * kick_process - kick a running thread to enter/exit the kernel
1425 * @p: the to-be-kicked thread
1427 * Cause a process which is running on another CPU to enter
1428 * kernel-mode, without any delay. (to get signals handled.)
1430 * NOTE: this function doesn't have to take the runqueue lock,
1431 * because all it wants to ensure is that the remote task enters
1432 * the kernel. If the IPI races and the task has been migrated
1433 * to another CPU then no harm is done and the purpose has been
1436 void kick_process(struct task_struct
*p
)
1442 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1443 smp_send_reschedule(cpu
);
1446 EXPORT_SYMBOL_GPL(kick_process
);
1449 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1451 * A few notes on cpu_active vs cpu_online:
1453 * - cpu_active must be a subset of cpu_online
1455 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1456 * see __set_cpus_allowed_ptr(). At this point the newly online
1457 * CPU isn't yet part of the sched domains, and balancing will not
1460 * - on CPU-down we clear cpu_active() to mask the sched domains and
1461 * avoid the load balancer to place new tasks on the to be removed
1462 * CPU. Existing tasks will remain running there and will be taken
1465 * This means that fallback selection must not select !active CPUs.
1466 * And can assume that any active CPU must be online. Conversely
1467 * select_task_rq() below may allow selection of !active CPUs in order
1468 * to satisfy the above rules.
1470 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
1472 int nid
= cpu_to_node(cpu
);
1473 const struct cpumask
*nodemask
= NULL
;
1474 enum { cpuset
, possible
, fail
} state
= cpuset
;
1478 * If the node that the CPU is on has been offlined, cpu_to_node()
1479 * will return -1. There is no CPU on the node, and we should
1480 * select the CPU on the other node.
1483 nodemask
= cpumask_of_node(nid
);
1485 /* Look for allowed, online CPU in same node. */
1486 for_each_cpu(dest_cpu
, nodemask
) {
1487 if (!cpu_active(dest_cpu
))
1489 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
1495 /* Any allowed, online CPU? */
1496 for_each_cpu(dest_cpu
, tsk_cpus_allowed(p
)) {
1497 if (!(p
->flags
& PF_KTHREAD
) && !cpu_active(dest_cpu
))
1499 if (!cpu_online(dest_cpu
))
1504 /* No more Mr. Nice Guy. */
1507 if (IS_ENABLED(CONFIG_CPUSETS
)) {
1508 cpuset_cpus_allowed_fallback(p
);
1514 do_set_cpus_allowed(p
, cpu_possible_mask
);
1525 if (state
!= cpuset
) {
1527 * Don't tell them about moving exiting tasks or
1528 * kernel threads (both mm NULL), since they never
1531 if (p
->mm
&& printk_ratelimit()) {
1532 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1533 task_pid_nr(p
), p
->comm
, cpu
);
1541 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1544 int select_task_rq(struct task_struct
*p
, int cpu
, int sd_flags
, int wake_flags
)
1546 lockdep_assert_held(&p
->pi_lock
);
1548 if (tsk_nr_cpus_allowed(p
) > 1)
1549 cpu
= p
->sched_class
->select_task_rq(p
, cpu
, sd_flags
, wake_flags
);
1551 cpu
= cpumask_any(tsk_cpus_allowed(p
));
1554 * In order not to call set_task_cpu() on a blocking task we need
1555 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1558 * Since this is common to all placement strategies, this lives here.
1560 * [ this allows ->select_task() to simply return task_cpu(p) and
1561 * not worry about this generic constraint ]
1563 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
1565 cpu
= select_fallback_rq(task_cpu(p
), p
);
1570 static void update_avg(u64
*avg
, u64 sample
)
1572 s64 diff
= sample
- *avg
;
1578 static inline int __set_cpus_allowed_ptr(struct task_struct
*p
,
1579 const struct cpumask
*new_mask
, bool check
)
1581 return set_cpus_allowed_ptr(p
, new_mask
);
1584 #endif /* CONFIG_SMP */
1587 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
1591 if (!schedstat_enabled())
1597 if (cpu
== rq
->cpu
) {
1598 schedstat_inc(rq
->ttwu_local
);
1599 schedstat_inc(p
->se
.statistics
.nr_wakeups_local
);
1601 struct sched_domain
*sd
;
1603 schedstat_inc(p
->se
.statistics
.nr_wakeups_remote
);
1605 for_each_domain(rq
->cpu
, sd
) {
1606 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
1607 schedstat_inc(sd
->ttwu_wake_remote
);
1614 if (wake_flags
& WF_MIGRATED
)
1615 schedstat_inc(p
->se
.statistics
.nr_wakeups_migrate
);
1616 #endif /* CONFIG_SMP */
1618 schedstat_inc(rq
->ttwu_count
);
1619 schedstat_inc(p
->se
.statistics
.nr_wakeups
);
1621 if (wake_flags
& WF_SYNC
)
1622 schedstat_inc(p
->se
.statistics
.nr_wakeups_sync
);
1625 static inline void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
1627 activate_task(rq
, p
, en_flags
);
1628 p
->on_rq
= TASK_ON_RQ_QUEUED
;
1630 /* If a worker is waking up, notify the workqueue: */
1631 if (p
->flags
& PF_WQ_WORKER
)
1632 wq_worker_waking_up(p
, cpu_of(rq
));
1636 * Mark the task runnable and perform wakeup-preemption.
1638 static void ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1639 struct rq_flags
*rf
)
1641 check_preempt_curr(rq
, p
, wake_flags
);
1642 p
->state
= TASK_RUNNING
;
1643 trace_sched_wakeup(p
);
1646 if (p
->sched_class
->task_woken
) {
1648 * Our task @p is fully woken up and running; so its safe to
1649 * drop the rq->lock, hereafter rq is only used for statistics.
1651 rq_unpin_lock(rq
, rf
);
1652 p
->sched_class
->task_woken(rq
, p
);
1653 rq_repin_lock(rq
, rf
);
1656 if (rq
->idle_stamp
) {
1657 u64 delta
= rq_clock(rq
) - rq
->idle_stamp
;
1658 u64 max
= 2*rq
->max_idle_balance_cost
;
1660 update_avg(&rq
->avg_idle
, delta
);
1662 if (rq
->avg_idle
> max
)
1671 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
,
1672 struct rq_flags
*rf
)
1674 int en_flags
= ENQUEUE_WAKEUP
;
1676 lockdep_assert_held(&rq
->lock
);
1679 if (p
->sched_contributes_to_load
)
1680 rq
->nr_uninterruptible
--;
1682 if (wake_flags
& WF_MIGRATED
)
1683 en_flags
|= ENQUEUE_MIGRATED
;
1686 ttwu_activate(rq
, p
, en_flags
);
1687 ttwu_do_wakeup(rq
, p
, wake_flags
, rf
);
1691 * Called in case the task @p isn't fully descheduled from its runqueue,
1692 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1693 * since all we need to do is flip p->state to TASK_RUNNING, since
1694 * the task is still ->on_rq.
1696 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
1702 rq
= __task_rq_lock(p
, &rf
);
1703 if (task_on_rq_queued(p
)) {
1704 /* check_preempt_curr() may use rq clock */
1705 update_rq_clock(rq
);
1706 ttwu_do_wakeup(rq
, p
, wake_flags
, &rf
);
1709 __task_rq_unlock(rq
, &rf
);
1715 void sched_ttwu_pending(void)
1717 struct rq
*rq
= this_rq();
1718 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
1719 struct task_struct
*p
;
1720 unsigned long flags
;
1726 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1727 rq_pin_lock(rq
, &rf
);
1732 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
1733 llist
= llist_next(llist
);
1735 if (p
->sched_remote_wakeup
)
1736 wake_flags
= WF_MIGRATED
;
1738 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
1741 rq_unpin_lock(rq
, &rf
);
1742 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1745 void scheduler_ipi(void)
1748 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1749 * TIF_NEED_RESCHED remotely (for the first time) will also send
1752 preempt_fold_need_resched();
1754 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
1758 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1759 * traditionally all their work was done from the interrupt return
1760 * path. Now that we actually do some work, we need to make sure
1763 * Some archs already do call them, luckily irq_enter/exit nest
1766 * Arguably we should visit all archs and update all handlers,
1767 * however a fair share of IPIs are still resched only so this would
1768 * somewhat pessimize the simple resched case.
1771 sched_ttwu_pending();
1774 * Check if someone kicked us for doing the nohz idle load balance.
1776 if (unlikely(got_nohz_idle_kick())) {
1777 this_rq()->idle_balance
= 1;
1778 raise_softirq_irqoff(SCHED_SOFTIRQ
);
1783 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
, int wake_flags
)
1785 struct rq
*rq
= cpu_rq(cpu
);
1787 p
->sched_remote_wakeup
= !!(wake_flags
& WF_MIGRATED
);
1789 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
)) {
1790 if (!set_nr_if_polling(rq
->idle
))
1791 smp_send_reschedule(cpu
);
1793 trace_sched_wake_idle_without_ipi(cpu
);
1797 void wake_up_if_idle(int cpu
)
1799 struct rq
*rq
= cpu_rq(cpu
);
1800 unsigned long flags
;
1804 if (!is_idle_task(rcu_dereference(rq
->curr
)))
1807 if (set_nr_if_polling(rq
->idle
)) {
1808 trace_sched_wake_idle_without_ipi(cpu
);
1810 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1811 if (is_idle_task(rq
->curr
))
1812 smp_send_reschedule(cpu
);
1813 /* Else CPU is not idle, do nothing here: */
1814 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1821 bool cpus_share_cache(int this_cpu
, int that_cpu
)
1823 return per_cpu(sd_llc_id
, this_cpu
) == per_cpu(sd_llc_id
, that_cpu
);
1825 #endif /* CONFIG_SMP */
1827 static void ttwu_queue(struct task_struct
*p
, int cpu
, int wake_flags
)
1829 struct rq
*rq
= cpu_rq(cpu
);
1832 #if defined(CONFIG_SMP)
1833 if (sched_feat(TTWU_QUEUE
) && !cpus_share_cache(smp_processor_id(), cpu
)) {
1834 sched_clock_cpu(cpu
); /* Sync clocks across CPUs */
1835 ttwu_queue_remote(p
, cpu
, wake_flags
);
1840 raw_spin_lock(&rq
->lock
);
1841 rq_pin_lock(rq
, &rf
);
1842 ttwu_do_activate(rq
, p
, wake_flags
, &rf
);
1843 rq_unpin_lock(rq
, &rf
);
1844 raw_spin_unlock(&rq
->lock
);
1848 * Notes on Program-Order guarantees on SMP systems.
1852 * The basic program-order guarantee on SMP systems is that when a task [t]
1853 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
1854 * execution on its new CPU [c1].
1856 * For migration (of runnable tasks) this is provided by the following means:
1858 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1859 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1860 * rq(c1)->lock (if not at the same time, then in that order).
1861 * C) LOCK of the rq(c1)->lock scheduling in task
1863 * Transitivity guarantees that B happens after A and C after B.
1864 * Note: we only require RCpc transitivity.
1865 * Note: the CPU doing B need not be c0 or c1
1874 * UNLOCK rq(0)->lock
1876 * LOCK rq(0)->lock // orders against CPU0
1878 * UNLOCK rq(0)->lock
1882 * UNLOCK rq(1)->lock
1884 * LOCK rq(1)->lock // orders against CPU2
1887 * UNLOCK rq(1)->lock
1890 * BLOCKING -- aka. SLEEP + WAKEUP
1892 * For blocking we (obviously) need to provide the same guarantee as for
1893 * migration. However the means are completely different as there is no lock
1894 * chain to provide order. Instead we do:
1896 * 1) smp_store_release(X->on_cpu, 0)
1897 * 2) smp_cond_load_acquire(!X->on_cpu)
1901 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1903 * LOCK rq(0)->lock LOCK X->pi_lock
1906 * smp_store_release(X->on_cpu, 0);
1908 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1914 * X->state = RUNNING
1915 * UNLOCK rq(2)->lock
1917 * LOCK rq(2)->lock // orders against CPU1
1920 * UNLOCK rq(2)->lock
1923 * UNLOCK rq(0)->lock
1926 * However; for wakeups there is a second guarantee we must provide, namely we
1927 * must observe the state that lead to our wakeup. That is, not only must our
1928 * task observe its own prior state, it must also observe the stores prior to
1931 * This means that any means of doing remote wakeups must order the CPU doing
1932 * the wakeup against the CPU the task is going to end up running on. This,
1933 * however, is already required for the regular Program-Order guarantee above,
1934 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1939 * try_to_wake_up - wake up a thread
1940 * @p: the thread to be awakened
1941 * @state: the mask of task states that can be woken
1942 * @wake_flags: wake modifier flags (WF_*)
1944 * If (@state & @p->state) @p->state = TASK_RUNNING.
1946 * If the task was not queued/runnable, also place it back on a runqueue.
1948 * Atomic against schedule() which would dequeue a task, also see
1949 * set_current_state().
1951 * Return: %true if @p->state changes (an actual wakeup was done),
1955 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
1957 unsigned long flags
;
1958 int cpu
, success
= 0;
1961 * If we are going to wake up a thread waiting for CONDITION we
1962 * need to ensure that CONDITION=1 done by the caller can not be
1963 * reordered with p->state check below. This pairs with mb() in
1964 * set_current_state() the waiting thread does.
1966 smp_mb__before_spinlock();
1967 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
1968 if (!(p
->state
& state
))
1971 trace_sched_waking(p
);
1973 /* We're going to change ->state: */
1978 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1979 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1980 * in smp_cond_load_acquire() below.
1982 * sched_ttwu_pending() try_to_wake_up()
1983 * [S] p->on_rq = 1; [L] P->state
1984 * UNLOCK rq->lock -----.
1988 * LOCK rq->lock -----'
1992 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
1994 * Pairs with the UNLOCK+LOCK on rq->lock from the
1995 * last wakeup of our task and the schedule that got our task
1999 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2004 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2005 * possible to, falsely, observe p->on_cpu == 0.
2007 * One must be running (->on_cpu == 1) in order to remove oneself
2008 * from the runqueue.
2010 * [S] ->on_cpu = 1; [L] ->on_rq
2014 * [S] ->on_rq = 0; [L] ->on_cpu
2016 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2017 * from the consecutive calls to schedule(); the first switching to our
2018 * task, the second putting it to sleep.
2023 * If the owning (remote) CPU is still in the middle of schedule() with
2024 * this task as prev, wait until its done referencing the task.
2026 * Pairs with the smp_store_release() in finish_lock_switch().
2028 * This ensures that tasks getting woken will be fully ordered against
2029 * their previous state and preserve Program Order.
2031 smp_cond_load_acquire(&p
->on_cpu
, !VAL
);
2033 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2034 p
->state
= TASK_WAKING
;
2037 delayacct_blkio_end();
2038 atomic_dec(&task_rq(p
)->nr_iowait
);
2041 cpu
= select_task_rq(p
, p
->wake_cpu
, SD_BALANCE_WAKE
, wake_flags
);
2042 if (task_cpu(p
) != cpu
) {
2043 wake_flags
|= WF_MIGRATED
;
2044 set_task_cpu(p
, cpu
);
2047 #else /* CONFIG_SMP */
2050 delayacct_blkio_end();
2051 atomic_dec(&task_rq(p
)->nr_iowait
);
2054 #endif /* CONFIG_SMP */
2056 ttwu_queue(p
, cpu
, wake_flags
);
2058 ttwu_stat(p
, cpu
, wake_flags
);
2060 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2066 * try_to_wake_up_local - try to wake up a local task with rq lock held
2067 * @p: the thread to be awakened
2068 * @cookie: context's cookie for pinning
2070 * Put @p on the run-queue if it's not already there. The caller must
2071 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2074 static void try_to_wake_up_local(struct task_struct
*p
, struct rq_flags
*rf
)
2076 struct rq
*rq
= task_rq(p
);
2078 if (WARN_ON_ONCE(rq
!= this_rq()) ||
2079 WARN_ON_ONCE(p
== current
))
2082 lockdep_assert_held(&rq
->lock
);
2084 if (!raw_spin_trylock(&p
->pi_lock
)) {
2086 * This is OK, because current is on_cpu, which avoids it being
2087 * picked for load-balance and preemption/IRQs are still
2088 * disabled avoiding further scheduler activity on it and we've
2089 * not yet picked a replacement task.
2091 rq_unpin_lock(rq
, rf
);
2092 raw_spin_unlock(&rq
->lock
);
2093 raw_spin_lock(&p
->pi_lock
);
2094 raw_spin_lock(&rq
->lock
);
2095 rq_repin_lock(rq
, rf
);
2098 if (!(p
->state
& TASK_NORMAL
))
2101 trace_sched_waking(p
);
2103 if (!task_on_rq_queued(p
)) {
2105 delayacct_blkio_end();
2106 atomic_dec(&rq
->nr_iowait
);
2108 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2111 ttwu_do_wakeup(rq
, p
, 0, rf
);
2112 ttwu_stat(p
, smp_processor_id(), 0);
2114 raw_spin_unlock(&p
->pi_lock
);
2118 * wake_up_process - Wake up a specific process
2119 * @p: The process to be woken up.
2121 * Attempt to wake up the nominated process and move it to the set of runnable
2124 * Return: 1 if the process was woken up, 0 if it was already running.
2126 * It may be assumed that this function implies a write memory barrier before
2127 * changing the task state if and only if any tasks are woken up.
2129 int wake_up_process(struct task_struct
*p
)
2131 return try_to_wake_up(p
, TASK_NORMAL
, 0);
2133 EXPORT_SYMBOL(wake_up_process
);
2135 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2137 return try_to_wake_up(p
, state
, 0);
2141 * This function clears the sched_dl_entity static params.
2143 void __dl_clear_params(struct task_struct
*p
)
2145 struct sched_dl_entity
*dl_se
= &p
->dl
;
2147 dl_se
->dl_runtime
= 0;
2148 dl_se
->dl_deadline
= 0;
2149 dl_se
->dl_period
= 0;
2153 dl_se
->dl_throttled
= 0;
2154 dl_se
->dl_yielded
= 0;
2158 * Perform scheduler related setup for a newly forked process p.
2159 * p is forked by current.
2161 * __sched_fork() is basic setup used by init_idle() too:
2163 static void __sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2168 p
->se
.exec_start
= 0;
2169 p
->se
.sum_exec_runtime
= 0;
2170 p
->se
.prev_sum_exec_runtime
= 0;
2171 p
->se
.nr_migrations
= 0;
2173 INIT_LIST_HEAD(&p
->se
.group_node
);
2175 #ifdef CONFIG_FAIR_GROUP_SCHED
2176 p
->se
.cfs_rq
= NULL
;
2179 #ifdef CONFIG_SCHEDSTATS
2180 /* Even if schedstat is disabled, there should not be garbage */
2181 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2184 RB_CLEAR_NODE(&p
->dl
.rb_node
);
2185 init_dl_task_timer(&p
->dl
);
2186 __dl_clear_params(p
);
2188 INIT_LIST_HEAD(&p
->rt
.run_list
);
2190 p
->rt
.time_slice
= sched_rr_timeslice
;
2194 #ifdef CONFIG_PREEMPT_NOTIFIERS
2195 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2198 #ifdef CONFIG_NUMA_BALANCING
2199 if (p
->mm
&& atomic_read(&p
->mm
->mm_users
) == 1) {
2200 p
->mm
->numa_next_scan
= jiffies
+ msecs_to_jiffies(sysctl_numa_balancing_scan_delay
);
2201 p
->mm
->numa_scan_seq
= 0;
2204 if (clone_flags
& CLONE_VM
)
2205 p
->numa_preferred_nid
= current
->numa_preferred_nid
;
2207 p
->numa_preferred_nid
= -1;
2209 p
->node_stamp
= 0ULL;
2210 p
->numa_scan_seq
= p
->mm
? p
->mm
->numa_scan_seq
: 0;
2211 p
->numa_scan_period
= sysctl_numa_balancing_scan_delay
;
2212 p
->numa_work
.next
= &p
->numa_work
;
2213 p
->numa_faults
= NULL
;
2214 p
->last_task_numa_placement
= 0;
2215 p
->last_sum_exec_runtime
= 0;
2217 p
->numa_group
= NULL
;
2218 #endif /* CONFIG_NUMA_BALANCING */
2221 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing
);
2223 #ifdef CONFIG_NUMA_BALANCING
2225 void set_numabalancing_state(bool enabled
)
2228 static_branch_enable(&sched_numa_balancing
);
2230 static_branch_disable(&sched_numa_balancing
);
2233 #ifdef CONFIG_PROC_SYSCTL
2234 int sysctl_numa_balancing(struct ctl_table
*table
, int write
,
2235 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2239 int state
= static_branch_likely(&sched_numa_balancing
);
2241 if (write
&& !capable(CAP_SYS_ADMIN
))
2246 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2250 set_numabalancing_state(state
);
2256 #ifdef CONFIG_SCHEDSTATS
2258 DEFINE_STATIC_KEY_FALSE(sched_schedstats
);
2259 static bool __initdata __sched_schedstats
= false;
2261 static void set_schedstats(bool enabled
)
2264 static_branch_enable(&sched_schedstats
);
2266 static_branch_disable(&sched_schedstats
);
2269 void force_schedstat_enabled(void)
2271 if (!schedstat_enabled()) {
2272 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2273 static_branch_enable(&sched_schedstats
);
2277 static int __init
setup_schedstats(char *str
)
2284 * This code is called before jump labels have been set up, so we can't
2285 * change the static branch directly just yet. Instead set a temporary
2286 * variable so init_schedstats() can do it later.
2288 if (!strcmp(str
, "enable")) {
2289 __sched_schedstats
= true;
2291 } else if (!strcmp(str
, "disable")) {
2292 __sched_schedstats
= false;
2297 pr_warn("Unable to parse schedstats=\n");
2301 __setup("schedstats=", setup_schedstats
);
2303 static void __init
init_schedstats(void)
2305 set_schedstats(__sched_schedstats
);
2308 #ifdef CONFIG_PROC_SYSCTL
2309 int sysctl_schedstats(struct ctl_table
*table
, int write
,
2310 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2314 int state
= static_branch_likely(&sched_schedstats
);
2316 if (write
&& !capable(CAP_SYS_ADMIN
))
2321 err
= proc_dointvec_minmax(&t
, write
, buffer
, lenp
, ppos
);
2325 set_schedstats(state
);
2328 #endif /* CONFIG_PROC_SYSCTL */
2329 #else /* !CONFIG_SCHEDSTATS */
2330 static inline void init_schedstats(void) {}
2331 #endif /* CONFIG_SCHEDSTATS */
2334 * fork()/clone()-time setup:
2336 int sched_fork(unsigned long clone_flags
, struct task_struct
*p
)
2338 unsigned long flags
;
2339 int cpu
= get_cpu();
2341 __sched_fork(clone_flags
, p
);
2343 * We mark the process as NEW here. This guarantees that
2344 * nobody will actually run it, and a signal or other external
2345 * event cannot wake it up and insert it on the runqueue either.
2347 p
->state
= TASK_NEW
;
2350 * Make sure we do not leak PI boosting priority to the child.
2352 p
->prio
= current
->normal_prio
;
2355 * Revert to default priority/policy on fork if requested.
2357 if (unlikely(p
->sched_reset_on_fork
)) {
2358 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
2359 p
->policy
= SCHED_NORMAL
;
2360 p
->static_prio
= NICE_TO_PRIO(0);
2362 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2363 p
->static_prio
= NICE_TO_PRIO(0);
2365 p
->prio
= p
->normal_prio
= __normal_prio(p
);
2369 * We don't need the reset flag anymore after the fork. It has
2370 * fulfilled its duty:
2372 p
->sched_reset_on_fork
= 0;
2375 if (dl_prio(p
->prio
)) {
2378 } else if (rt_prio(p
->prio
)) {
2379 p
->sched_class
= &rt_sched_class
;
2381 p
->sched_class
= &fair_sched_class
;
2384 init_entity_runnable_average(&p
->se
);
2387 * The child is not yet in the pid-hash so no cgroup attach races,
2388 * and the cgroup is pinned to this child due to cgroup_fork()
2389 * is ran before sched_fork().
2391 * Silence PROVE_RCU.
2393 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2395 * We're setting the CPU for the first time, we don't migrate,
2396 * so use __set_task_cpu().
2398 __set_task_cpu(p
, cpu
);
2399 if (p
->sched_class
->task_fork
)
2400 p
->sched_class
->task_fork(p
);
2401 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2403 #ifdef CONFIG_SCHED_INFO
2404 if (likely(sched_info_on()))
2405 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2407 #if defined(CONFIG_SMP)
2410 init_task_preempt_count(p
);
2412 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2413 RB_CLEAR_NODE(&p
->pushable_dl_tasks
);
2420 unsigned long to_ratio(u64 period
, u64 runtime
)
2422 if (runtime
== RUNTIME_INF
)
2426 * Doing this here saves a lot of checks in all
2427 * the calling paths, and returning zero seems
2428 * safe for them anyway.
2433 return div64_u64(runtime
<< 20, period
);
2437 inline struct dl_bw
*dl_bw_of(int i
)
2439 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2440 "sched RCU must be held");
2441 return &cpu_rq(i
)->rd
->dl_bw
;
2444 static inline int dl_bw_cpus(int i
)
2446 struct root_domain
*rd
= cpu_rq(i
)->rd
;
2449 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2450 "sched RCU must be held");
2451 for_each_cpu_and(i
, rd
->span
, cpu_active_mask
)
2457 inline struct dl_bw
*dl_bw_of(int i
)
2459 return &cpu_rq(i
)->dl
.dl_bw
;
2462 static inline int dl_bw_cpus(int i
)
2469 * We must be sure that accepting a new task (or allowing changing the
2470 * parameters of an existing one) is consistent with the bandwidth
2471 * constraints. If yes, this function also accordingly updates the currently
2472 * allocated bandwidth to reflect the new situation.
2474 * This function is called while holding p's rq->lock.
2476 * XXX we should delay bw change until the task's 0-lag point, see
2479 static int dl_overflow(struct task_struct
*p
, int policy
,
2480 const struct sched_attr
*attr
)
2483 struct dl_bw
*dl_b
= dl_bw_of(task_cpu(p
));
2484 u64 period
= attr
->sched_period
?: attr
->sched_deadline
;
2485 u64 runtime
= attr
->sched_runtime
;
2486 u64 new_bw
= dl_policy(policy
) ? to_ratio(period
, runtime
) : 0;
2489 /* !deadline task may carry old deadline bandwidth */
2490 if (new_bw
== p
->dl
.dl_bw
&& task_has_dl_policy(p
))
2494 * Either if a task, enters, leave, or stays -deadline but changes
2495 * its parameters, we may need to update accordingly the total
2496 * allocated bandwidth of the container.
2498 raw_spin_lock(&dl_b
->lock
);
2499 cpus
= dl_bw_cpus(task_cpu(p
));
2500 if (dl_policy(policy
) && !task_has_dl_policy(p
) &&
2501 !__dl_overflow(dl_b
, cpus
, 0, new_bw
)) {
2502 __dl_add(dl_b
, new_bw
);
2504 } else if (dl_policy(policy
) && task_has_dl_policy(p
) &&
2505 !__dl_overflow(dl_b
, cpus
, p
->dl
.dl_bw
, new_bw
)) {
2506 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2507 __dl_add(dl_b
, new_bw
);
2509 } else if (!dl_policy(policy
) && task_has_dl_policy(p
)) {
2510 __dl_clear(dl_b
, p
->dl
.dl_bw
);
2513 raw_spin_unlock(&dl_b
->lock
);
2518 extern void init_dl_bw(struct dl_bw
*dl_b
);
2521 * wake_up_new_task - wake up a newly created task for the first time.
2523 * This function will do some initial scheduler statistics housekeeping
2524 * that must be done for every newly created context, then puts the task
2525 * on the runqueue and wakes it.
2527 void wake_up_new_task(struct task_struct
*p
)
2532 raw_spin_lock_irqsave(&p
->pi_lock
, rf
.flags
);
2533 p
->state
= TASK_RUNNING
;
2536 * Fork balancing, do it here and not earlier because:
2537 * - cpus_allowed can change in the fork path
2538 * - any previously selected CPU might disappear through hotplug
2540 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2541 * as we're not fully set-up yet.
2543 __set_task_cpu(p
, select_task_rq(p
, task_cpu(p
), SD_BALANCE_FORK
, 0));
2545 rq
= __task_rq_lock(p
, &rf
);
2546 update_rq_clock(rq
);
2547 post_init_entity_util_avg(&p
->se
);
2549 activate_task(rq
, p
, 0);
2550 p
->on_rq
= TASK_ON_RQ_QUEUED
;
2551 trace_sched_wakeup_new(p
);
2552 check_preempt_curr(rq
, p
, WF_FORK
);
2554 if (p
->sched_class
->task_woken
) {
2556 * Nothing relies on rq->lock after this, so its fine to
2559 rq_unpin_lock(rq
, &rf
);
2560 p
->sched_class
->task_woken(rq
, p
);
2561 rq_repin_lock(rq
, &rf
);
2564 task_rq_unlock(rq
, p
, &rf
);
2567 #ifdef CONFIG_PREEMPT_NOTIFIERS
2569 static struct static_key preempt_notifier_key
= STATIC_KEY_INIT_FALSE
;
2571 void preempt_notifier_inc(void)
2573 static_key_slow_inc(&preempt_notifier_key
);
2575 EXPORT_SYMBOL_GPL(preempt_notifier_inc
);
2577 void preempt_notifier_dec(void)
2579 static_key_slow_dec(&preempt_notifier_key
);
2581 EXPORT_SYMBOL_GPL(preempt_notifier_dec
);
2584 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2585 * @notifier: notifier struct to register
2587 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2589 if (!static_key_false(&preempt_notifier_key
))
2590 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2592 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2594 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2597 * preempt_notifier_unregister - no longer interested in preemption notifications
2598 * @notifier: notifier struct to unregister
2600 * This is *not* safe to call from within a preemption notifier.
2602 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2604 hlist_del(¬ifier
->link
);
2606 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2608 static void __fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2610 struct preempt_notifier
*notifier
;
2612 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2613 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2616 static __always_inline
void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2618 if (static_key_false(&preempt_notifier_key
))
2619 __fire_sched_in_preempt_notifiers(curr
);
2623 __fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2624 struct task_struct
*next
)
2626 struct preempt_notifier
*notifier
;
2628 hlist_for_each_entry(notifier
, &curr
->preempt_notifiers
, link
)
2629 notifier
->ops
->sched_out(notifier
, next
);
2632 static __always_inline
void
2633 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2634 struct task_struct
*next
)
2636 if (static_key_false(&preempt_notifier_key
))
2637 __fire_sched_out_preempt_notifiers(curr
, next
);
2640 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2642 static inline void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2647 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2648 struct task_struct
*next
)
2652 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2655 * prepare_task_switch - prepare to switch tasks
2656 * @rq: the runqueue preparing to switch
2657 * @prev: the current task that is being switched out
2658 * @next: the task we are going to switch to.
2660 * This is called with the rq lock held and interrupts off. It must
2661 * be paired with a subsequent finish_task_switch after the context
2664 * prepare_task_switch sets up locking and calls architecture specific
2668 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2669 struct task_struct
*next
)
2671 sched_info_switch(rq
, prev
, next
);
2672 perf_event_task_sched_out(prev
, next
);
2673 fire_sched_out_preempt_notifiers(prev
, next
);
2674 prepare_lock_switch(rq
, next
);
2675 prepare_arch_switch(next
);
2679 * finish_task_switch - clean up after a task-switch
2680 * @prev: the thread we just switched away from.
2682 * finish_task_switch must be called after the context switch, paired
2683 * with a prepare_task_switch call before the context switch.
2684 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2685 * and do any other architecture-specific cleanup actions.
2687 * Note that we may have delayed dropping an mm in context_switch(). If
2688 * so, we finish that here outside of the runqueue lock. (Doing it
2689 * with the lock held can cause deadlocks; see schedule() for
2692 * The context switch have flipped the stack from under us and restored the
2693 * local variables which were saved when this task called schedule() in the
2694 * past. prev == current is still correct but we need to recalculate this_rq
2695 * because prev may have moved to another CPU.
2697 static struct rq
*finish_task_switch(struct task_struct
*prev
)
2698 __releases(rq
->lock
)
2700 struct rq
*rq
= this_rq();
2701 struct mm_struct
*mm
= rq
->prev_mm
;
2705 * The previous task will have left us with a preempt_count of 2
2706 * because it left us after:
2709 * preempt_disable(); // 1
2711 * raw_spin_lock_irq(&rq->lock) // 2
2713 * Also, see FORK_PREEMPT_COUNT.
2715 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET
,
2716 "corrupted preempt_count: %s/%d/0x%x\n",
2717 current
->comm
, current
->pid
, preempt_count()))
2718 preempt_count_set(FORK_PREEMPT_COUNT
);
2723 * A task struct has one reference for the use as "current".
2724 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2725 * schedule one last time. The schedule call will never return, and
2726 * the scheduled task must drop that reference.
2728 * We must observe prev->state before clearing prev->on_cpu (in
2729 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2730 * running on another CPU and we could rave with its RUNNING -> DEAD
2731 * transition, resulting in a double drop.
2733 prev_state
= prev
->state
;
2734 vtime_task_switch(prev
);
2735 perf_event_task_sched_in(prev
, current
);
2736 finish_lock_switch(rq
, prev
);
2737 finish_arch_post_lock_switch();
2739 fire_sched_in_preempt_notifiers(current
);
2742 if (unlikely(prev_state
== TASK_DEAD
)) {
2743 if (prev
->sched_class
->task_dead
)
2744 prev
->sched_class
->task_dead(prev
);
2747 * Remove function-return probe instances associated with this
2748 * task and put them back on the free list.
2750 kprobe_flush_task(prev
);
2752 /* Task is done with its stack. */
2753 put_task_stack(prev
);
2755 put_task_struct(prev
);
2758 tick_nohz_task_switch();
2764 /* rq->lock is NOT held, but preemption is disabled */
2765 static void __balance_callback(struct rq
*rq
)
2767 struct callback_head
*head
, *next
;
2768 void (*func
)(struct rq
*rq
);
2769 unsigned long flags
;
2771 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2772 head
= rq
->balance_callback
;
2773 rq
->balance_callback
= NULL
;
2775 func
= (void (*)(struct rq
*))head
->func
;
2782 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2785 static inline void balance_callback(struct rq
*rq
)
2787 if (unlikely(rq
->balance_callback
))
2788 __balance_callback(rq
);
2793 static inline void balance_callback(struct rq
*rq
)
2800 * schedule_tail - first thing a freshly forked thread must call.
2801 * @prev: the thread we just switched away from.
2803 asmlinkage __visible
void schedule_tail(struct task_struct
*prev
)
2804 __releases(rq
->lock
)
2809 * New tasks start with FORK_PREEMPT_COUNT, see there and
2810 * finish_task_switch() for details.
2812 * finish_task_switch() will drop rq->lock() and lower preempt_count
2813 * and the preempt_enable() will end up enabling preemption (on
2814 * PREEMPT_COUNT kernels).
2817 rq
= finish_task_switch(prev
);
2818 balance_callback(rq
);
2821 if (current
->set_child_tid
)
2822 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2826 * context_switch - switch to the new MM and the new thread's register state.
2828 static __always_inline
struct rq
*
2829 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2830 struct task_struct
*next
, struct rq_flags
*rf
)
2832 struct mm_struct
*mm
, *oldmm
;
2834 prepare_task_switch(rq
, prev
, next
);
2837 oldmm
= prev
->active_mm
;
2839 * For paravirt, this is coupled with an exit in switch_to to
2840 * combine the page table reload and the switch backend into
2843 arch_start_context_switch(prev
);
2846 next
->active_mm
= oldmm
;
2847 atomic_inc(&oldmm
->mm_count
);
2848 enter_lazy_tlb(oldmm
, next
);
2850 switch_mm_irqs_off(oldmm
, mm
, next
);
2853 prev
->active_mm
= NULL
;
2854 rq
->prev_mm
= oldmm
;
2857 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
2860 * Since the runqueue lock will be released by the next
2861 * task (which is an invalid locking op but in the case
2862 * of the scheduler it's an obvious special-case), so we
2863 * do an early lockdep release here:
2865 rq_unpin_lock(rq
, rf
);
2866 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2868 /* Here we just switch the register state and the stack. */
2869 switch_to(prev
, next
, prev
);
2872 return finish_task_switch(prev
);
2876 * nr_running and nr_context_switches:
2878 * externally visible scheduler statistics: current number of runnable
2879 * threads, total number of context switches performed since bootup.
2881 unsigned long nr_running(void)
2883 unsigned long i
, sum
= 0;
2885 for_each_online_cpu(i
)
2886 sum
+= cpu_rq(i
)->nr_running
;
2892 * Check if only the current task is running on the CPU.
2894 * Caution: this function does not check that the caller has disabled
2895 * preemption, thus the result might have a time-of-check-to-time-of-use
2896 * race. The caller is responsible to use it correctly, for example:
2898 * - from a non-preemptable section (of course)
2900 * - from a thread that is bound to a single CPU
2902 * - in a loop with very short iterations (e.g. a polling loop)
2904 bool single_task_running(void)
2906 return raw_rq()->nr_running
== 1;
2908 EXPORT_SYMBOL(single_task_running
);
2910 unsigned long long nr_context_switches(void)
2913 unsigned long long sum
= 0;
2915 for_each_possible_cpu(i
)
2916 sum
+= cpu_rq(i
)->nr_switches
;
2922 * IO-wait accounting, and how its mostly bollocks (on SMP).
2924 * The idea behind IO-wait account is to account the idle time that we could
2925 * have spend running if it were not for IO. That is, if we were to improve the
2926 * storage performance, we'd have a proportional reduction in IO-wait time.
2928 * This all works nicely on UP, where, when a task blocks on IO, we account
2929 * idle time as IO-wait, because if the storage were faster, it could've been
2930 * running and we'd not be idle.
2932 * This has been extended to SMP, by doing the same for each CPU. This however
2935 * Imagine for instance the case where two tasks block on one CPU, only the one
2936 * CPU will have IO-wait accounted, while the other has regular idle. Even
2937 * though, if the storage were faster, both could've ran at the same time,
2938 * utilising both CPUs.
2940 * This means, that when looking globally, the current IO-wait accounting on
2941 * SMP is a lower bound, by reason of under accounting.
2943 * Worse, since the numbers are provided per CPU, they are sometimes
2944 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
2945 * associated with any one particular CPU, it can wake to another CPU than it
2946 * blocked on. This means the per CPU IO-wait number is meaningless.
2948 * Task CPU affinities can make all that even more 'interesting'.
2951 unsigned long nr_iowait(void)
2953 unsigned long i
, sum
= 0;
2955 for_each_possible_cpu(i
)
2956 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2962 * Consumers of these two interfaces, like for example the cpufreq menu
2963 * governor are using nonsensical data. Boosting frequency for a CPU that has
2964 * IO-wait which might not even end up running the task when it does become
2968 unsigned long nr_iowait_cpu(int cpu
)
2970 struct rq
*this = cpu_rq(cpu
);
2971 return atomic_read(&this->nr_iowait
);
2974 void get_iowait_load(unsigned long *nr_waiters
, unsigned long *load
)
2976 struct rq
*rq
= this_rq();
2977 *nr_waiters
= atomic_read(&rq
->nr_iowait
);
2978 *load
= rq
->load
.weight
;
2984 * sched_exec - execve() is a valuable balancing opportunity, because at
2985 * this point the task has the smallest effective memory and cache footprint.
2987 void sched_exec(void)
2989 struct task_struct
*p
= current
;
2990 unsigned long flags
;
2993 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2994 dest_cpu
= p
->sched_class
->select_task_rq(p
, task_cpu(p
), SD_BALANCE_EXEC
, 0);
2995 if (dest_cpu
== smp_processor_id())
2998 if (likely(cpu_active(dest_cpu
))) {
2999 struct migration_arg arg
= { p
, dest_cpu
};
3001 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3002 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
3006 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3011 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3012 DEFINE_PER_CPU(struct kernel_cpustat
, kernel_cpustat
);
3014 EXPORT_PER_CPU_SYMBOL(kstat
);
3015 EXPORT_PER_CPU_SYMBOL(kernel_cpustat
);
3018 * The function fair_sched_class.update_curr accesses the struct curr
3019 * and its field curr->exec_start; when called from task_sched_runtime(),
3020 * we observe a high rate of cache misses in practice.
3021 * Prefetching this data results in improved performance.
3023 static inline void prefetch_curr_exec_start(struct task_struct
*p
)
3025 #ifdef CONFIG_FAIR_GROUP_SCHED
3026 struct sched_entity
*curr
= (&p
->se
)->cfs_rq
->curr
;
3028 struct sched_entity
*curr
= (&task_rq(p
)->cfs
)->curr
;
3031 prefetch(&curr
->exec_start
);
3035 * Return accounted runtime for the task.
3036 * In case the task is currently running, return the runtime plus current's
3037 * pending runtime that have not been accounted yet.
3039 unsigned long long task_sched_runtime(struct task_struct
*p
)
3045 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3047 * 64-bit doesn't need locks to atomically read a 64bit value.
3048 * So we have a optimization chance when the task's delta_exec is 0.
3049 * Reading ->on_cpu is racy, but this is ok.
3051 * If we race with it leaving CPU, we'll take a lock. So we're correct.
3052 * If we race with it entering CPU, unaccounted time is 0. This is
3053 * indistinguishable from the read occurring a few cycles earlier.
3054 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3055 * been accounted, so we're correct here as well.
3057 if (!p
->on_cpu
|| !task_on_rq_queued(p
))
3058 return p
->se
.sum_exec_runtime
;
3061 rq
= task_rq_lock(p
, &rf
);
3063 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3064 * project cycles that may never be accounted to this
3065 * thread, breaking clock_gettime().
3067 if (task_current(rq
, p
) && task_on_rq_queued(p
)) {
3068 prefetch_curr_exec_start(p
);
3069 update_rq_clock(rq
);
3070 p
->sched_class
->update_curr(rq
);
3072 ns
= p
->se
.sum_exec_runtime
;
3073 task_rq_unlock(rq
, p
, &rf
);
3079 * This function gets called by the timer code, with HZ frequency.
3080 * We call it with interrupts disabled.
3082 void scheduler_tick(void)
3084 int cpu
= smp_processor_id();
3085 struct rq
*rq
= cpu_rq(cpu
);
3086 struct task_struct
*curr
= rq
->curr
;
3090 raw_spin_lock(&rq
->lock
);
3091 update_rq_clock(rq
);
3092 curr
->sched_class
->task_tick(rq
, curr
, 0);
3093 cpu_load_update_active(rq
);
3094 calc_global_load_tick(rq
);
3095 raw_spin_unlock(&rq
->lock
);
3097 perf_event_task_tick();
3100 rq
->idle_balance
= idle_cpu(cpu
);
3101 trigger_load_balance(rq
);
3103 rq_last_tick_reset(rq
);
3106 #ifdef CONFIG_NO_HZ_FULL
3108 * scheduler_tick_max_deferment
3110 * Keep at least one tick per second when a single
3111 * active task is running because the scheduler doesn't
3112 * yet completely support full dynticks environment.
3114 * This makes sure that uptime, CFS vruntime, load
3115 * balancing, etc... continue to move forward, even
3116 * with a very low granularity.
3118 * Return: Maximum deferment in nanoseconds.
3120 u64
scheduler_tick_max_deferment(void)
3122 struct rq
*rq
= this_rq();
3123 unsigned long next
, now
= READ_ONCE(jiffies
);
3125 next
= rq
->last_sched_tick
+ HZ
;
3127 if (time_before_eq(next
, now
))
3130 return jiffies_to_nsecs(next
- now
);
3134 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3135 defined(CONFIG_PREEMPT_TRACER))
3137 * If the value passed in is equal to the current preempt count
3138 * then we just disabled preemption. Start timing the latency.
3140 static inline void preempt_latency_start(int val
)
3142 if (preempt_count() == val
) {
3143 unsigned long ip
= get_lock_parent_ip();
3144 #ifdef CONFIG_DEBUG_PREEMPT
3145 current
->preempt_disable_ip
= ip
;
3147 trace_preempt_off(CALLER_ADDR0
, ip
);
3151 void preempt_count_add(int val
)
3153 #ifdef CONFIG_DEBUG_PREEMPT
3157 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3160 __preempt_count_add(val
);
3161 #ifdef CONFIG_DEBUG_PREEMPT
3163 * Spinlock count overflowing soon?
3165 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3168 preempt_latency_start(val
);
3170 EXPORT_SYMBOL(preempt_count_add
);
3171 NOKPROBE_SYMBOL(preempt_count_add
);
3174 * If the value passed in equals to the current preempt count
3175 * then we just enabled preemption. Stop timing the latency.
3177 static inline void preempt_latency_stop(int val
)
3179 if (preempt_count() == val
)
3180 trace_preempt_on(CALLER_ADDR0
, get_lock_parent_ip());
3183 void preempt_count_sub(int val
)
3185 #ifdef CONFIG_DEBUG_PREEMPT
3189 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3192 * Is the spinlock portion underflowing?
3194 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3195 !(preempt_count() & PREEMPT_MASK
)))
3199 preempt_latency_stop(val
);
3200 __preempt_count_sub(val
);
3202 EXPORT_SYMBOL(preempt_count_sub
);
3203 NOKPROBE_SYMBOL(preempt_count_sub
);
3206 static inline void preempt_latency_start(int val
) { }
3207 static inline void preempt_latency_stop(int val
) { }
3211 * Print scheduling while atomic bug:
3213 static noinline
void __schedule_bug(struct task_struct
*prev
)
3215 /* Save this before calling printk(), since that will clobber it */
3216 unsigned long preempt_disable_ip
= get_preempt_disable_ip(current
);
3218 if (oops_in_progress
)
3221 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3222 prev
->comm
, prev
->pid
, preempt_count());
3224 debug_show_held_locks(prev
);
3226 if (irqs_disabled())
3227 print_irqtrace_events(prev
);
3228 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
3229 && in_atomic_preempt_off()) {
3230 pr_err("Preemption disabled at:");
3231 print_ip_sym(preempt_disable_ip
);
3235 panic("scheduling while atomic\n");
3238 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
3242 * Various schedule()-time debugging checks and statistics:
3244 static inline void schedule_debug(struct task_struct
*prev
)
3246 #ifdef CONFIG_SCHED_STACK_END_CHECK
3247 if (task_stack_end_corrupted(prev
))
3248 panic("corrupted stack end detected inside scheduler\n");
3251 if (unlikely(in_atomic_preempt_off())) {
3252 __schedule_bug(prev
);
3253 preempt_count_set(PREEMPT_DISABLED
);
3257 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3259 schedstat_inc(this_rq()->sched_count
);
3263 * Pick up the highest-prio task:
3265 static inline struct task_struct
*
3266 pick_next_task(struct rq
*rq
, struct task_struct
*prev
, struct rq_flags
*rf
)
3268 const struct sched_class
*class;
3269 struct task_struct
*p
;
3272 * Optimization: we know that if all tasks are in
3273 * the fair class we can call that function directly:
3275 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
3276 p
= fair_sched_class
.pick_next_task(rq
, prev
, rf
);
3277 if (unlikely(p
== RETRY_TASK
))
3280 /* Assumes fair_sched_class->next == idle_sched_class */
3282 p
= idle_sched_class
.pick_next_task(rq
, prev
, rf
);
3288 for_each_class(class) {
3289 p
= class->pick_next_task(rq
, prev
, rf
);
3291 if (unlikely(p
== RETRY_TASK
))
3297 /* The idle class should always have a runnable task: */
3302 * __schedule() is the main scheduler function.
3304 * The main means of driving the scheduler and thus entering this function are:
3306 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3308 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3309 * paths. For example, see arch/x86/entry_64.S.
3311 * To drive preemption between tasks, the scheduler sets the flag in timer
3312 * interrupt handler scheduler_tick().
3314 * 3. Wakeups don't really cause entry into schedule(). They add a
3315 * task to the run-queue and that's it.
3317 * Now, if the new task added to the run-queue preempts the current
3318 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3319 * called on the nearest possible occasion:
3321 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3323 * - in syscall or exception context, at the next outmost
3324 * preempt_enable(). (this might be as soon as the wake_up()'s
3327 * - in IRQ context, return from interrupt-handler to
3328 * preemptible context
3330 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3333 * - cond_resched() call
3334 * - explicit schedule() call
3335 * - return from syscall or exception to user-space
3336 * - return from interrupt-handler to user-space
3338 * WARNING: must be called with preemption disabled!
3340 static void __sched notrace
__schedule(bool preempt
)
3342 struct task_struct
*prev
, *next
;
3343 unsigned long *switch_count
;
3348 cpu
= smp_processor_id();
3352 schedule_debug(prev
);
3354 if (sched_feat(HRTICK
))
3357 local_irq_disable();
3358 rcu_note_context_switch();
3361 * Make sure that signal_pending_state()->signal_pending() below
3362 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3363 * done by the caller to avoid the race with signal_wake_up().
3365 smp_mb__before_spinlock();
3366 raw_spin_lock(&rq
->lock
);
3367 rq_pin_lock(rq
, &rf
);
3369 /* Promote REQ to ACT */
3370 rq
->clock_update_flags
<<= 1;
3372 switch_count
= &prev
->nivcsw
;
3373 if (!preempt
&& prev
->state
) {
3374 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3375 prev
->state
= TASK_RUNNING
;
3377 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3380 if (prev
->in_iowait
) {
3381 atomic_inc(&rq
->nr_iowait
);
3382 delayacct_blkio_start();
3386 * If a worker went to sleep, notify and ask workqueue
3387 * whether it wants to wake up a task to maintain
3390 if (prev
->flags
& PF_WQ_WORKER
) {
3391 struct task_struct
*to_wakeup
;
3393 to_wakeup
= wq_worker_sleeping(prev
);
3395 try_to_wake_up_local(to_wakeup
, &rf
);
3398 switch_count
= &prev
->nvcsw
;
3401 if (task_on_rq_queued(prev
))
3402 update_rq_clock(rq
);
3404 next
= pick_next_task(rq
, prev
, &rf
);
3405 clear_tsk_need_resched(prev
);
3406 clear_preempt_need_resched();
3408 if (likely(prev
!= next
)) {
3413 trace_sched_switch(preempt
, prev
, next
);
3415 /* Also unlocks the rq: */
3416 rq
= context_switch(rq
, prev
, next
, &rf
);
3418 rq
->clock_update_flags
&= ~(RQCF_ACT_SKIP
|RQCF_REQ_SKIP
);
3419 rq_unpin_lock(rq
, &rf
);
3420 raw_spin_unlock_irq(&rq
->lock
);
3423 balance_callback(rq
);
3426 void __noreturn
do_task_dead(void)
3429 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3430 * when the following two conditions become true.
3431 * - There is race condition of mmap_sem (It is acquired by
3433 * - SMI occurs before setting TASK_RUNINNG.
3434 * (or hypervisor of virtual machine switches to other guest)
3435 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3437 * To avoid it, we have to wait for releasing tsk->pi_lock which
3438 * is held by try_to_wake_up()
3441 raw_spin_unlock_wait(¤t
->pi_lock
);
3443 /* Causes final put_task_struct in finish_task_switch(): */
3444 __set_current_state(TASK_DEAD
);
3446 /* Tell freezer to ignore us: */
3447 current
->flags
|= PF_NOFREEZE
;
3452 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
3457 static inline void sched_submit_work(struct task_struct
*tsk
)
3459 if (!tsk
->state
|| tsk_is_pi_blocked(tsk
))
3462 * If we are going to sleep and we have plugged IO queued,
3463 * make sure to submit it to avoid deadlocks.
3465 if (blk_needs_flush_plug(tsk
))
3466 blk_schedule_flush_plug(tsk
);
3469 asmlinkage __visible
void __sched
schedule(void)
3471 struct task_struct
*tsk
= current
;
3473 sched_submit_work(tsk
);
3477 sched_preempt_enable_no_resched();
3478 } while (need_resched());
3480 EXPORT_SYMBOL(schedule
);
3482 #ifdef CONFIG_CONTEXT_TRACKING
3483 asmlinkage __visible
void __sched
schedule_user(void)
3486 * If we come here after a random call to set_need_resched(),
3487 * or we have been woken up remotely but the IPI has not yet arrived,
3488 * we haven't yet exited the RCU idle mode. Do it here manually until
3489 * we find a better solution.
3491 * NB: There are buggy callers of this function. Ideally we
3492 * should warn if prev_state != CONTEXT_USER, but that will trigger
3493 * too frequently to make sense yet.
3495 enum ctx_state prev_state
= exception_enter();
3497 exception_exit(prev_state
);
3502 * schedule_preempt_disabled - called with preemption disabled
3504 * Returns with preemption disabled. Note: preempt_count must be 1
3506 void __sched
schedule_preempt_disabled(void)
3508 sched_preempt_enable_no_resched();
3513 static void __sched notrace
preempt_schedule_common(void)
3517 * Because the function tracer can trace preempt_count_sub()
3518 * and it also uses preempt_enable/disable_notrace(), if
3519 * NEED_RESCHED is set, the preempt_enable_notrace() called
3520 * by the function tracer will call this function again and
3521 * cause infinite recursion.
3523 * Preemption must be disabled here before the function
3524 * tracer can trace. Break up preempt_disable() into two
3525 * calls. One to disable preemption without fear of being
3526 * traced. The other to still record the preemption latency,
3527 * which can also be traced by the function tracer.
3529 preempt_disable_notrace();
3530 preempt_latency_start(1);
3532 preempt_latency_stop(1);
3533 preempt_enable_no_resched_notrace();
3536 * Check again in case we missed a preemption opportunity
3537 * between schedule and now.
3539 } while (need_resched());
3542 #ifdef CONFIG_PREEMPT
3544 * this is the entry point to schedule() from in-kernel preemption
3545 * off of preempt_enable. Kernel preemptions off return from interrupt
3546 * occur there and call schedule directly.
3548 asmlinkage __visible
void __sched notrace
preempt_schedule(void)
3551 * If there is a non-zero preempt_count or interrupts are disabled,
3552 * we do not want to preempt the current task. Just return..
3554 if (likely(!preemptible()))
3557 preempt_schedule_common();
3559 NOKPROBE_SYMBOL(preempt_schedule
);
3560 EXPORT_SYMBOL(preempt_schedule
);
3563 * preempt_schedule_notrace - preempt_schedule called by tracing
3565 * The tracing infrastructure uses preempt_enable_notrace to prevent
3566 * recursion and tracing preempt enabling caused by the tracing
3567 * infrastructure itself. But as tracing can happen in areas coming
3568 * from userspace or just about to enter userspace, a preempt enable
3569 * can occur before user_exit() is called. This will cause the scheduler
3570 * to be called when the system is still in usermode.
3572 * To prevent this, the preempt_enable_notrace will use this function
3573 * instead of preempt_schedule() to exit user context if needed before
3574 * calling the scheduler.
3576 asmlinkage __visible
void __sched notrace
preempt_schedule_notrace(void)
3578 enum ctx_state prev_ctx
;
3580 if (likely(!preemptible()))
3585 * Because the function tracer can trace preempt_count_sub()
3586 * and it also uses preempt_enable/disable_notrace(), if
3587 * NEED_RESCHED is set, the preempt_enable_notrace() called
3588 * by the function tracer will call this function again and
3589 * cause infinite recursion.
3591 * Preemption must be disabled here before the function
3592 * tracer can trace. Break up preempt_disable() into two
3593 * calls. One to disable preemption without fear of being
3594 * traced. The other to still record the preemption latency,
3595 * which can also be traced by the function tracer.
3597 preempt_disable_notrace();
3598 preempt_latency_start(1);
3600 * Needs preempt disabled in case user_exit() is traced
3601 * and the tracer calls preempt_enable_notrace() causing
3602 * an infinite recursion.
3604 prev_ctx
= exception_enter();
3606 exception_exit(prev_ctx
);
3608 preempt_latency_stop(1);
3609 preempt_enable_no_resched_notrace();
3610 } while (need_resched());
3612 EXPORT_SYMBOL_GPL(preempt_schedule_notrace
);
3614 #endif /* CONFIG_PREEMPT */
3617 * this is the entry point to schedule() from kernel preemption
3618 * off of irq context.
3619 * Note, that this is called and return with irqs disabled. This will
3620 * protect us against recursive calling from irq.
3622 asmlinkage __visible
void __sched
preempt_schedule_irq(void)
3624 enum ctx_state prev_state
;
3626 /* Catch callers which need to be fixed */
3627 BUG_ON(preempt_count() || !irqs_disabled());
3629 prev_state
= exception_enter();
3635 local_irq_disable();
3636 sched_preempt_enable_no_resched();
3637 } while (need_resched());
3639 exception_exit(prev_state
);
3642 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3645 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3647 EXPORT_SYMBOL(default_wake_function
);
3649 #ifdef CONFIG_RT_MUTEXES
3652 * rt_mutex_setprio - set the current priority of a task
3654 * @prio: prio value (kernel-internal form)
3656 * This function changes the 'effective' priority of a task. It does
3657 * not touch ->normal_prio like __setscheduler().
3659 * Used by the rt_mutex code to implement priority inheritance
3660 * logic. Call site only calls if the priority of the task changed.
3662 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3664 int oldprio
, queued
, running
, queue_flag
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
3665 const struct sched_class
*prev_class
;
3669 BUG_ON(prio
> MAX_PRIO
);
3671 rq
= __task_rq_lock(p
, &rf
);
3672 update_rq_clock(rq
);
3675 * Idle task boosting is a nono in general. There is one
3676 * exception, when PREEMPT_RT and NOHZ is active:
3678 * The idle task calls get_next_timer_interrupt() and holds
3679 * the timer wheel base->lock on the CPU and another CPU wants
3680 * to access the timer (probably to cancel it). We can safely
3681 * ignore the boosting request, as the idle CPU runs this code
3682 * with interrupts disabled and will complete the lock
3683 * protected section without being interrupted. So there is no
3684 * real need to boost.
3686 if (unlikely(p
== rq
->idle
)) {
3687 WARN_ON(p
!= rq
->curr
);
3688 WARN_ON(p
->pi_blocked_on
);
3692 trace_sched_pi_setprio(p
, prio
);
3695 if (oldprio
== prio
)
3696 queue_flag
&= ~DEQUEUE_MOVE
;
3698 prev_class
= p
->sched_class
;
3699 queued
= task_on_rq_queued(p
);
3700 running
= task_current(rq
, p
);
3702 dequeue_task(rq
, p
, queue_flag
);
3704 put_prev_task(rq
, p
);
3707 * Boosting condition are:
3708 * 1. -rt task is running and holds mutex A
3709 * --> -dl task blocks on mutex A
3711 * 2. -dl task is running and holds mutex A
3712 * --> -dl task blocks on mutex A and could preempt the
3715 if (dl_prio(prio
)) {
3716 struct task_struct
*pi_task
= rt_mutex_get_top_task(p
);
3717 if (!dl_prio(p
->normal_prio
) ||
3718 (pi_task
&& dl_entity_preempt(&pi_task
->dl
, &p
->dl
))) {
3719 p
->dl
.dl_boosted
= 1;
3720 queue_flag
|= ENQUEUE_REPLENISH
;
3722 p
->dl
.dl_boosted
= 0;
3723 p
->sched_class
= &dl_sched_class
;
3724 } else if (rt_prio(prio
)) {
3725 if (dl_prio(oldprio
))
3726 p
->dl
.dl_boosted
= 0;
3728 queue_flag
|= ENQUEUE_HEAD
;
3729 p
->sched_class
= &rt_sched_class
;
3731 if (dl_prio(oldprio
))
3732 p
->dl
.dl_boosted
= 0;
3733 if (rt_prio(oldprio
))
3735 p
->sched_class
= &fair_sched_class
;
3741 enqueue_task(rq
, p
, queue_flag
);
3743 set_curr_task(rq
, p
);
3745 check_class_changed(rq
, p
, prev_class
, oldprio
);
3747 /* Avoid rq from going away on us: */
3749 __task_rq_unlock(rq
, &rf
);
3751 balance_callback(rq
);
3756 void set_user_nice(struct task_struct
*p
, long nice
)
3758 bool queued
, running
;
3759 int old_prio
, delta
;
3763 if (task_nice(p
) == nice
|| nice
< MIN_NICE
|| nice
> MAX_NICE
)
3766 * We have to be careful, if called from sys_setpriority(),
3767 * the task might be in the middle of scheduling on another CPU.
3769 rq
= task_rq_lock(p
, &rf
);
3770 update_rq_clock(rq
);
3773 * The RT priorities are set via sched_setscheduler(), but we still
3774 * allow the 'normal' nice value to be set - but as expected
3775 * it wont have any effect on scheduling until the task is
3776 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3778 if (task_has_dl_policy(p
) || task_has_rt_policy(p
)) {
3779 p
->static_prio
= NICE_TO_PRIO(nice
);
3782 queued
= task_on_rq_queued(p
);
3783 running
= task_current(rq
, p
);
3785 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
3787 put_prev_task(rq
, p
);
3789 p
->static_prio
= NICE_TO_PRIO(nice
);
3792 p
->prio
= effective_prio(p
);
3793 delta
= p
->prio
- old_prio
;
3796 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
3798 * If the task increased its priority or is running and
3799 * lowered its priority, then reschedule its CPU:
3801 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
3805 set_curr_task(rq
, p
);
3807 task_rq_unlock(rq
, p
, &rf
);
3809 EXPORT_SYMBOL(set_user_nice
);
3812 * can_nice - check if a task can reduce its nice value
3816 int can_nice(const struct task_struct
*p
, const int nice
)
3818 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
3819 int nice_rlim
= nice_to_rlimit(nice
);
3821 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
3822 capable(CAP_SYS_NICE
));
3825 #ifdef __ARCH_WANT_SYS_NICE
3828 * sys_nice - change the priority of the current process.
3829 * @increment: priority increment
3831 * sys_setpriority is a more generic, but much slower function that
3832 * does similar things.
3834 SYSCALL_DEFINE1(nice
, int, increment
)
3839 * Setpriority might change our priority at the same moment.
3840 * We don't have to worry. Conceptually one call occurs first
3841 * and we have a single winner.
3843 increment
= clamp(increment
, -NICE_WIDTH
, NICE_WIDTH
);
3844 nice
= task_nice(current
) + increment
;
3846 nice
= clamp_val(nice
, MIN_NICE
, MAX_NICE
);
3847 if (increment
< 0 && !can_nice(current
, nice
))
3850 retval
= security_task_setnice(current
, nice
);
3854 set_user_nice(current
, nice
);
3861 * task_prio - return the priority value of a given task.
3862 * @p: the task in question.
3864 * Return: The priority value as seen by users in /proc.
3865 * RT tasks are offset by -200. Normal tasks are centered
3866 * around 0, value goes from -16 to +15.
3868 int task_prio(const struct task_struct
*p
)
3870 return p
->prio
- MAX_RT_PRIO
;
3874 * idle_cpu - is a given CPU idle currently?
3875 * @cpu: the processor in question.
3877 * Return: 1 if the CPU is currently idle. 0 otherwise.
3879 int idle_cpu(int cpu
)
3881 struct rq
*rq
= cpu_rq(cpu
);
3883 if (rq
->curr
!= rq
->idle
)
3890 if (!llist_empty(&rq
->wake_list
))
3898 * idle_task - return the idle task for a given CPU.
3899 * @cpu: the processor in question.
3901 * Return: The idle task for the CPU @cpu.
3903 struct task_struct
*idle_task(int cpu
)
3905 return cpu_rq(cpu
)->idle
;
3909 * find_process_by_pid - find a process with a matching PID value.
3910 * @pid: the pid in question.
3912 * The task of @pid, if found. %NULL otherwise.
3914 static struct task_struct
*find_process_by_pid(pid_t pid
)
3916 return pid
? find_task_by_vpid(pid
) : current
;
3920 * This function initializes the sched_dl_entity of a newly becoming
3921 * SCHED_DEADLINE task.
3923 * Only the static values are considered here, the actual runtime and the
3924 * absolute deadline will be properly calculated when the task is enqueued
3925 * for the first time with its new policy.
3928 __setparam_dl(struct task_struct
*p
, const struct sched_attr
*attr
)
3930 struct sched_dl_entity
*dl_se
= &p
->dl
;
3932 dl_se
->dl_runtime
= attr
->sched_runtime
;
3933 dl_se
->dl_deadline
= attr
->sched_deadline
;
3934 dl_se
->dl_period
= attr
->sched_period
?: dl_se
->dl_deadline
;
3935 dl_se
->flags
= attr
->sched_flags
;
3936 dl_se
->dl_bw
= to_ratio(dl_se
->dl_period
, dl_se
->dl_runtime
);
3939 * Changing the parameters of a task is 'tricky' and we're not doing
3940 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3942 * What we SHOULD do is delay the bandwidth release until the 0-lag
3943 * point. This would include retaining the task_struct until that time
3944 * and change dl_overflow() to not immediately decrement the current
3947 * Instead we retain the current runtime/deadline and let the new
3948 * parameters take effect after the current reservation period lapses.
3949 * This is safe (albeit pessimistic) because the 0-lag point is always
3950 * before the current scheduling deadline.
3952 * We can still have temporary overloads because we do not delay the
3953 * change in bandwidth until that time; so admission control is
3954 * not on the safe side. It does however guarantee tasks will never
3955 * consume more than promised.
3960 * sched_setparam() passes in -1 for its policy, to let the functions
3961 * it calls know not to change it.
3963 #define SETPARAM_POLICY -1
3965 static void __setscheduler_params(struct task_struct
*p
,
3966 const struct sched_attr
*attr
)
3968 int policy
= attr
->sched_policy
;
3970 if (policy
== SETPARAM_POLICY
)
3975 if (dl_policy(policy
))
3976 __setparam_dl(p
, attr
);
3977 else if (fair_policy(policy
))
3978 p
->static_prio
= NICE_TO_PRIO(attr
->sched_nice
);
3981 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3982 * !rt_policy. Always setting this ensures that things like
3983 * getparam()/getattr() don't report silly values for !rt tasks.
3985 p
->rt_priority
= attr
->sched_priority
;
3986 p
->normal_prio
= normal_prio(p
);
3990 /* Actually do priority change: must hold pi & rq lock. */
3991 static void __setscheduler(struct rq
*rq
, struct task_struct
*p
,
3992 const struct sched_attr
*attr
, bool keep_boost
)
3994 __setscheduler_params(p
, attr
);
3997 * Keep a potential priority boosting if called from
3998 * sched_setscheduler().
4001 p
->prio
= rt_mutex_get_effective_prio(p
, normal_prio(p
));
4003 p
->prio
= normal_prio(p
);
4005 if (dl_prio(p
->prio
))
4006 p
->sched_class
= &dl_sched_class
;
4007 else if (rt_prio(p
->prio
))
4008 p
->sched_class
= &rt_sched_class
;
4010 p
->sched_class
= &fair_sched_class
;
4014 __getparam_dl(struct task_struct
*p
, struct sched_attr
*attr
)
4016 struct sched_dl_entity
*dl_se
= &p
->dl
;
4018 attr
->sched_priority
= p
->rt_priority
;
4019 attr
->sched_runtime
= dl_se
->dl_runtime
;
4020 attr
->sched_deadline
= dl_se
->dl_deadline
;
4021 attr
->sched_period
= dl_se
->dl_period
;
4022 attr
->sched_flags
= dl_se
->flags
;
4026 * This function validates the new parameters of a -deadline task.
4027 * We ask for the deadline not being zero, and greater or equal
4028 * than the runtime, as well as the period of being zero or
4029 * greater than deadline. Furthermore, we have to be sure that
4030 * user parameters are above the internal resolution of 1us (we
4031 * check sched_runtime only since it is always the smaller one) and
4032 * below 2^63 ns (we have to check both sched_deadline and
4033 * sched_period, as the latter can be zero).
4036 __checkparam_dl(const struct sched_attr
*attr
)
4039 if (attr
->sched_deadline
== 0)
4043 * Since we truncate DL_SCALE bits, make sure we're at least
4046 if (attr
->sched_runtime
< (1ULL << DL_SCALE
))
4050 * Since we use the MSB for wrap-around and sign issues, make
4051 * sure it's not set (mind that period can be equal to zero).
4053 if (attr
->sched_deadline
& (1ULL << 63) ||
4054 attr
->sched_period
& (1ULL << 63))
4057 /* runtime <= deadline <= period (if period != 0) */
4058 if ((attr
->sched_period
!= 0 &&
4059 attr
->sched_period
< attr
->sched_deadline
) ||
4060 attr
->sched_deadline
< attr
->sched_runtime
)
4067 * Check the target process has a UID that matches the current process's:
4069 static bool check_same_owner(struct task_struct
*p
)
4071 const struct cred
*cred
= current_cred(), *pcred
;
4075 pcred
= __task_cred(p
);
4076 match
= (uid_eq(cred
->euid
, pcred
->euid
) ||
4077 uid_eq(cred
->euid
, pcred
->uid
));
4082 static bool dl_param_changed(struct task_struct
*p
, const struct sched_attr
*attr
)
4084 struct sched_dl_entity
*dl_se
= &p
->dl
;
4086 if (dl_se
->dl_runtime
!= attr
->sched_runtime
||
4087 dl_se
->dl_deadline
!= attr
->sched_deadline
||
4088 dl_se
->dl_period
!= attr
->sched_period
||
4089 dl_se
->flags
!= attr
->sched_flags
)
4095 static int __sched_setscheduler(struct task_struct
*p
,
4096 const struct sched_attr
*attr
,
4099 int newprio
= dl_policy(attr
->sched_policy
) ? MAX_DL_PRIO
- 1 :
4100 MAX_RT_PRIO
- 1 - attr
->sched_priority
;
4101 int retval
, oldprio
, oldpolicy
= -1, queued
, running
;
4102 int new_effective_prio
, policy
= attr
->sched_policy
;
4103 const struct sched_class
*prev_class
;
4106 int queue_flags
= DEQUEUE_SAVE
| DEQUEUE_MOVE
;
4109 /* May grab non-irq protected spin_locks: */
4110 BUG_ON(in_interrupt());
4112 /* Double check policy once rq lock held: */
4114 reset_on_fork
= p
->sched_reset_on_fork
;
4115 policy
= oldpolicy
= p
->policy
;
4117 reset_on_fork
= !!(attr
->sched_flags
& SCHED_FLAG_RESET_ON_FORK
);
4119 if (!valid_policy(policy
))
4123 if (attr
->sched_flags
& ~(SCHED_FLAG_RESET_ON_FORK
))
4127 * Valid priorities for SCHED_FIFO and SCHED_RR are
4128 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4129 * SCHED_BATCH and SCHED_IDLE is 0.
4131 if ((p
->mm
&& attr
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4132 (!p
->mm
&& attr
->sched_priority
> MAX_RT_PRIO
-1))
4134 if ((dl_policy(policy
) && !__checkparam_dl(attr
)) ||
4135 (rt_policy(policy
) != (attr
->sched_priority
!= 0)))
4139 * Allow unprivileged RT tasks to decrease priority:
4141 if (user
&& !capable(CAP_SYS_NICE
)) {
4142 if (fair_policy(policy
)) {
4143 if (attr
->sched_nice
< task_nice(p
) &&
4144 !can_nice(p
, attr
->sched_nice
))
4148 if (rt_policy(policy
)) {
4149 unsigned long rlim_rtprio
=
4150 task_rlimit(p
, RLIMIT_RTPRIO
);
4152 /* Can't set/change the rt policy: */
4153 if (policy
!= p
->policy
&& !rlim_rtprio
)
4156 /* Can't increase priority: */
4157 if (attr
->sched_priority
> p
->rt_priority
&&
4158 attr
->sched_priority
> rlim_rtprio
)
4163 * Can't set/change SCHED_DEADLINE policy at all for now
4164 * (safest behavior); in the future we would like to allow
4165 * unprivileged DL tasks to increase their relative deadline
4166 * or reduce their runtime (both ways reducing utilization)
4168 if (dl_policy(policy
))
4172 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4173 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4175 if (idle_policy(p
->policy
) && !idle_policy(policy
)) {
4176 if (!can_nice(p
, task_nice(p
)))
4180 /* Can't change other user's priorities: */
4181 if (!check_same_owner(p
))
4184 /* Normal users shall not reset the sched_reset_on_fork flag: */
4185 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4190 retval
= security_task_setscheduler(p
);
4196 * Make sure no PI-waiters arrive (or leave) while we are
4197 * changing the priority of the task:
4199 * To be able to change p->policy safely, the appropriate
4200 * runqueue lock must be held.
4202 rq
= task_rq_lock(p
, &rf
);
4203 update_rq_clock(rq
);
4206 * Changing the policy of the stop threads its a very bad idea:
4208 if (p
== rq
->stop
) {
4209 task_rq_unlock(rq
, p
, &rf
);
4214 * If not changing anything there's no need to proceed further,
4215 * but store a possible modification of reset_on_fork.
4217 if (unlikely(policy
== p
->policy
)) {
4218 if (fair_policy(policy
) && attr
->sched_nice
!= task_nice(p
))
4220 if (rt_policy(policy
) && attr
->sched_priority
!= p
->rt_priority
)
4222 if (dl_policy(policy
) && dl_param_changed(p
, attr
))
4225 p
->sched_reset_on_fork
= reset_on_fork
;
4226 task_rq_unlock(rq
, p
, &rf
);
4232 #ifdef CONFIG_RT_GROUP_SCHED
4234 * Do not allow realtime tasks into groups that have no runtime
4237 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4238 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4239 !task_group_is_autogroup(task_group(p
))) {
4240 task_rq_unlock(rq
, p
, &rf
);
4245 if (dl_bandwidth_enabled() && dl_policy(policy
)) {
4246 cpumask_t
*span
= rq
->rd
->span
;
4249 * Don't allow tasks with an affinity mask smaller than
4250 * the entire root_domain to become SCHED_DEADLINE. We
4251 * will also fail if there's no bandwidth available.
4253 if (!cpumask_subset(span
, &p
->cpus_allowed
) ||
4254 rq
->rd
->dl_bw
.bw
== 0) {
4255 task_rq_unlock(rq
, p
, &rf
);
4262 /* Re-check policy now with rq lock held: */
4263 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4264 policy
= oldpolicy
= -1;
4265 task_rq_unlock(rq
, p
, &rf
);
4270 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4271 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4274 if ((dl_policy(policy
) || dl_task(p
)) && dl_overflow(p
, policy
, attr
)) {
4275 task_rq_unlock(rq
, p
, &rf
);
4279 p
->sched_reset_on_fork
= reset_on_fork
;
4284 * Take priority boosted tasks into account. If the new
4285 * effective priority is unchanged, we just store the new
4286 * normal parameters and do not touch the scheduler class and
4287 * the runqueue. This will be done when the task deboost
4290 new_effective_prio
= rt_mutex_get_effective_prio(p
, newprio
);
4291 if (new_effective_prio
== oldprio
)
4292 queue_flags
&= ~DEQUEUE_MOVE
;
4295 queued
= task_on_rq_queued(p
);
4296 running
= task_current(rq
, p
);
4298 dequeue_task(rq
, p
, queue_flags
);
4300 put_prev_task(rq
, p
);
4302 prev_class
= p
->sched_class
;
4303 __setscheduler(rq
, p
, attr
, pi
);
4307 * We enqueue to tail when the priority of a task is
4308 * increased (user space view).
4310 if (oldprio
< p
->prio
)
4311 queue_flags
|= ENQUEUE_HEAD
;
4313 enqueue_task(rq
, p
, queue_flags
);
4316 set_curr_task(rq
, p
);
4318 check_class_changed(rq
, p
, prev_class
, oldprio
);
4320 /* Avoid rq from going away on us: */
4322 task_rq_unlock(rq
, p
, &rf
);
4325 rt_mutex_adjust_pi(p
);
4327 /* Run balance callbacks after we've adjusted the PI chain: */
4328 balance_callback(rq
);
4334 static int _sched_setscheduler(struct task_struct
*p
, int policy
,
4335 const struct sched_param
*param
, bool check
)
4337 struct sched_attr attr
= {
4338 .sched_policy
= policy
,
4339 .sched_priority
= param
->sched_priority
,
4340 .sched_nice
= PRIO_TO_NICE(p
->static_prio
),
4343 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4344 if ((policy
!= SETPARAM_POLICY
) && (policy
& SCHED_RESET_ON_FORK
)) {
4345 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4346 policy
&= ~SCHED_RESET_ON_FORK
;
4347 attr
.sched_policy
= policy
;
4350 return __sched_setscheduler(p
, &attr
, check
, true);
4353 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4354 * @p: the task in question.
4355 * @policy: new policy.
4356 * @param: structure containing the new RT priority.
4358 * Return: 0 on success. An error code otherwise.
4360 * NOTE that the task may be already dead.
4362 int sched_setscheduler(struct task_struct
*p
, int policy
,
4363 const struct sched_param
*param
)
4365 return _sched_setscheduler(p
, policy
, param
, true);
4367 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4369 int sched_setattr(struct task_struct
*p
, const struct sched_attr
*attr
)
4371 return __sched_setscheduler(p
, attr
, true, true);
4373 EXPORT_SYMBOL_GPL(sched_setattr
);
4376 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4377 * @p: the task in question.
4378 * @policy: new policy.
4379 * @param: structure containing the new RT priority.
4381 * Just like sched_setscheduler, only don't bother checking if the
4382 * current context has permission. For example, this is needed in
4383 * stop_machine(): we create temporary high priority worker threads,
4384 * but our caller might not have that capability.
4386 * Return: 0 on success. An error code otherwise.
4388 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4389 const struct sched_param
*param
)
4391 return _sched_setscheduler(p
, policy
, param
, false);
4393 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck
);
4396 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4398 struct sched_param lparam
;
4399 struct task_struct
*p
;
4402 if (!param
|| pid
< 0)
4404 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4409 p
= find_process_by_pid(pid
);
4411 retval
= sched_setscheduler(p
, policy
, &lparam
);
4418 * Mimics kernel/events/core.c perf_copy_attr().
4420 static int sched_copy_attr(struct sched_attr __user
*uattr
, struct sched_attr
*attr
)
4425 if (!access_ok(VERIFY_WRITE
, uattr
, SCHED_ATTR_SIZE_VER0
))
4428 /* Zero the full structure, so that a short copy will be nice: */
4429 memset(attr
, 0, sizeof(*attr
));
4431 ret
= get_user(size
, &uattr
->size
);
4435 /* Bail out on silly large: */
4436 if (size
> PAGE_SIZE
)
4439 /* ABI compatibility quirk: */
4441 size
= SCHED_ATTR_SIZE_VER0
;
4443 if (size
< SCHED_ATTR_SIZE_VER0
)
4447 * If we're handed a bigger struct than we know of,
4448 * ensure all the unknown bits are 0 - i.e. new
4449 * user-space does not rely on any kernel feature
4450 * extensions we dont know about yet.
4452 if (size
> sizeof(*attr
)) {
4453 unsigned char __user
*addr
;
4454 unsigned char __user
*end
;
4457 addr
= (void __user
*)uattr
+ sizeof(*attr
);
4458 end
= (void __user
*)uattr
+ size
;
4460 for (; addr
< end
; addr
++) {
4461 ret
= get_user(val
, addr
);
4467 size
= sizeof(*attr
);
4470 ret
= copy_from_user(attr
, uattr
, size
);
4475 * XXX: Do we want to be lenient like existing syscalls; or do we want
4476 * to be strict and return an error on out-of-bounds values?
4478 attr
->sched_nice
= clamp(attr
->sched_nice
, MIN_NICE
, MAX_NICE
);
4483 put_user(sizeof(*attr
), &uattr
->size
);
4488 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4489 * @pid: the pid in question.
4490 * @policy: new policy.
4491 * @param: structure containing the new RT priority.
4493 * Return: 0 on success. An error code otherwise.
4495 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
, struct sched_param __user
*, param
)
4500 return do_sched_setscheduler(pid
, policy
, param
);
4504 * sys_sched_setparam - set/change the RT priority of a thread
4505 * @pid: the pid in question.
4506 * @param: structure containing the new RT priority.
4508 * Return: 0 on success. An error code otherwise.
4510 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4512 return do_sched_setscheduler(pid
, SETPARAM_POLICY
, param
);
4516 * sys_sched_setattr - same as above, but with extended sched_attr
4517 * @pid: the pid in question.
4518 * @uattr: structure containing the extended parameters.
4519 * @flags: for future extension.
4521 SYSCALL_DEFINE3(sched_setattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4522 unsigned int, flags
)
4524 struct sched_attr attr
;
4525 struct task_struct
*p
;
4528 if (!uattr
|| pid
< 0 || flags
)
4531 retval
= sched_copy_attr(uattr
, &attr
);
4535 if ((int)attr
.sched_policy
< 0)
4540 p
= find_process_by_pid(pid
);
4542 retval
= sched_setattr(p
, &attr
);
4549 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4550 * @pid: the pid in question.
4552 * Return: On success, the policy of the thread. Otherwise, a negative error
4555 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4557 struct task_struct
*p
;
4565 p
= find_process_by_pid(pid
);
4567 retval
= security_task_getscheduler(p
);
4570 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4577 * sys_sched_getparam - get the RT priority of a thread
4578 * @pid: the pid in question.
4579 * @param: structure containing the RT priority.
4581 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4584 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4586 struct sched_param lp
= { .sched_priority
= 0 };
4587 struct task_struct
*p
;
4590 if (!param
|| pid
< 0)
4594 p
= find_process_by_pid(pid
);
4599 retval
= security_task_getscheduler(p
);
4603 if (task_has_rt_policy(p
))
4604 lp
.sched_priority
= p
->rt_priority
;
4608 * This one might sleep, we cannot do it with a spinlock held ...
4610 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4619 static int sched_read_attr(struct sched_attr __user
*uattr
,
4620 struct sched_attr
*attr
,
4625 if (!access_ok(VERIFY_WRITE
, uattr
, usize
))
4629 * If we're handed a smaller struct than we know of,
4630 * ensure all the unknown bits are 0 - i.e. old
4631 * user-space does not get uncomplete information.
4633 if (usize
< sizeof(*attr
)) {
4634 unsigned char *addr
;
4637 addr
= (void *)attr
+ usize
;
4638 end
= (void *)attr
+ sizeof(*attr
);
4640 for (; addr
< end
; addr
++) {
4648 ret
= copy_to_user(uattr
, attr
, attr
->size
);
4656 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4657 * @pid: the pid in question.
4658 * @uattr: structure containing the extended parameters.
4659 * @size: sizeof(attr) for fwd/bwd comp.
4660 * @flags: for future extension.
4662 SYSCALL_DEFINE4(sched_getattr
, pid_t
, pid
, struct sched_attr __user
*, uattr
,
4663 unsigned int, size
, unsigned int, flags
)
4665 struct sched_attr attr
= {
4666 .size
= sizeof(struct sched_attr
),
4668 struct task_struct
*p
;
4671 if (!uattr
|| pid
< 0 || size
> PAGE_SIZE
||
4672 size
< SCHED_ATTR_SIZE_VER0
|| flags
)
4676 p
= find_process_by_pid(pid
);
4681 retval
= security_task_getscheduler(p
);
4685 attr
.sched_policy
= p
->policy
;
4686 if (p
->sched_reset_on_fork
)
4687 attr
.sched_flags
|= SCHED_FLAG_RESET_ON_FORK
;
4688 if (task_has_dl_policy(p
))
4689 __getparam_dl(p
, &attr
);
4690 else if (task_has_rt_policy(p
))
4691 attr
.sched_priority
= p
->rt_priority
;
4693 attr
.sched_nice
= task_nice(p
);
4697 retval
= sched_read_attr(uattr
, &attr
, size
);
4705 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4707 cpumask_var_t cpus_allowed
, new_mask
;
4708 struct task_struct
*p
;
4713 p
= find_process_by_pid(pid
);
4719 /* Prevent p going away */
4723 if (p
->flags
& PF_NO_SETAFFINITY
) {
4727 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4731 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4733 goto out_free_cpus_allowed
;
4736 if (!check_same_owner(p
)) {
4738 if (!ns_capable(__task_cred(p
)->user_ns
, CAP_SYS_NICE
)) {
4740 goto out_free_new_mask
;
4745 retval
= security_task_setscheduler(p
);
4747 goto out_free_new_mask
;
4750 cpuset_cpus_allowed(p
, cpus_allowed
);
4751 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4754 * Since bandwidth control happens on root_domain basis,
4755 * if admission test is enabled, we only admit -deadline
4756 * tasks allowed to run on all the CPUs in the task's
4760 if (task_has_dl_policy(p
) && dl_bandwidth_enabled()) {
4762 if (!cpumask_subset(task_rq(p
)->rd
->span
, new_mask
)) {
4765 goto out_free_new_mask
;
4771 retval
= __set_cpus_allowed_ptr(p
, new_mask
, true);
4774 cpuset_cpus_allowed(p
, cpus_allowed
);
4775 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4777 * We must have raced with a concurrent cpuset
4778 * update. Just reset the cpus_allowed to the
4779 * cpuset's cpus_allowed
4781 cpumask_copy(new_mask
, cpus_allowed
);
4786 free_cpumask_var(new_mask
);
4787 out_free_cpus_allowed
:
4788 free_cpumask_var(cpus_allowed
);
4794 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4795 struct cpumask
*new_mask
)
4797 if (len
< cpumask_size())
4798 cpumask_clear(new_mask
);
4799 else if (len
> cpumask_size())
4800 len
= cpumask_size();
4802 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4806 * sys_sched_setaffinity - set the CPU affinity of a process
4807 * @pid: pid of the process
4808 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4809 * @user_mask_ptr: user-space pointer to the new CPU mask
4811 * Return: 0 on success. An error code otherwise.
4813 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4814 unsigned long __user
*, user_mask_ptr
)
4816 cpumask_var_t new_mask
;
4819 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4822 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4824 retval
= sched_setaffinity(pid
, new_mask
);
4825 free_cpumask_var(new_mask
);
4829 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4831 struct task_struct
*p
;
4832 unsigned long flags
;
4838 p
= find_process_by_pid(pid
);
4842 retval
= security_task_getscheduler(p
);
4846 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4847 cpumask_and(mask
, &p
->cpus_allowed
, cpu_active_mask
);
4848 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4857 * sys_sched_getaffinity - get the CPU affinity of a process
4858 * @pid: pid of the process
4859 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4860 * @user_mask_ptr: user-space pointer to hold the current CPU mask
4862 * Return: size of CPU mask copied to user_mask_ptr on success. An
4863 * error code otherwise.
4865 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4866 unsigned long __user
*, user_mask_ptr
)
4871 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4873 if (len
& (sizeof(unsigned long)-1))
4876 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4879 ret
= sched_getaffinity(pid
, mask
);
4881 size_t retlen
= min_t(size_t, len
, cpumask_size());
4883 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4888 free_cpumask_var(mask
);
4894 * sys_sched_yield - yield the current processor to other threads.
4896 * This function yields the current CPU to other tasks. If there are no
4897 * other threads running on this CPU then this function will return.
4901 SYSCALL_DEFINE0(sched_yield
)
4903 struct rq
*rq
= this_rq_lock();
4905 schedstat_inc(rq
->yld_count
);
4906 current
->sched_class
->yield_task(rq
);
4909 * Since we are going to call schedule() anyway, there's
4910 * no need to preempt or enable interrupts:
4912 __release(rq
->lock
);
4913 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4914 do_raw_spin_unlock(&rq
->lock
);
4915 sched_preempt_enable_no_resched();
4922 #ifndef CONFIG_PREEMPT
4923 int __sched
_cond_resched(void)
4925 if (should_resched(0)) {
4926 preempt_schedule_common();
4931 EXPORT_SYMBOL(_cond_resched
);
4935 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4936 * call schedule, and on return reacquire the lock.
4938 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4939 * operations here to prevent schedule() from being called twice (once via
4940 * spin_unlock(), once by hand).
4942 int __cond_resched_lock(spinlock_t
*lock
)
4944 int resched
= should_resched(PREEMPT_LOCK_OFFSET
);
4947 lockdep_assert_held(lock
);
4949 if (spin_needbreak(lock
) || resched
) {
4952 preempt_schedule_common();
4960 EXPORT_SYMBOL(__cond_resched_lock
);
4962 int __sched
__cond_resched_softirq(void)
4964 BUG_ON(!in_softirq());
4966 if (should_resched(SOFTIRQ_DISABLE_OFFSET
)) {
4968 preempt_schedule_common();
4974 EXPORT_SYMBOL(__cond_resched_softirq
);
4977 * yield - yield the current processor to other threads.
4979 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4981 * The scheduler is at all times free to pick the calling task as the most
4982 * eligible task to run, if removing the yield() call from your code breaks
4983 * it, its already broken.
4985 * Typical broken usage is:
4990 * where one assumes that yield() will let 'the other' process run that will
4991 * make event true. If the current task is a SCHED_FIFO task that will never
4992 * happen. Never use yield() as a progress guarantee!!
4994 * If you want to use yield() to wait for something, use wait_event().
4995 * If you want to use yield() to be 'nice' for others, use cond_resched().
4996 * If you still want to use yield(), do not!
4998 void __sched
yield(void)
5000 set_current_state(TASK_RUNNING
);
5003 EXPORT_SYMBOL(yield
);
5006 * yield_to - yield the current processor to another thread in
5007 * your thread group, or accelerate that thread toward the
5008 * processor it's on.
5010 * @preempt: whether task preemption is allowed or not
5012 * It's the caller's job to ensure that the target task struct
5013 * can't go away on us before we can do any checks.
5016 * true (>0) if we indeed boosted the target task.
5017 * false (0) if we failed to boost the target.
5018 * -ESRCH if there's no task to yield to.
5020 int __sched
yield_to(struct task_struct
*p
, bool preempt
)
5022 struct task_struct
*curr
= current
;
5023 struct rq
*rq
, *p_rq
;
5024 unsigned long flags
;
5027 local_irq_save(flags
);
5033 * If we're the only runnable task on the rq and target rq also
5034 * has only one task, there's absolutely no point in yielding.
5036 if (rq
->nr_running
== 1 && p_rq
->nr_running
== 1) {
5041 double_rq_lock(rq
, p_rq
);
5042 if (task_rq(p
) != p_rq
) {
5043 double_rq_unlock(rq
, p_rq
);
5047 if (!curr
->sched_class
->yield_to_task
)
5050 if (curr
->sched_class
!= p
->sched_class
)
5053 if (task_running(p_rq
, p
) || p
->state
)
5056 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5058 schedstat_inc(rq
->yld_count
);
5060 * Make p's CPU reschedule; pick_next_entity takes care of
5063 if (preempt
&& rq
!= p_rq
)
5068 double_rq_unlock(rq
, p_rq
);
5070 local_irq_restore(flags
);
5077 EXPORT_SYMBOL_GPL(yield_to
);
5079 int io_schedule_prepare(void)
5081 int old_iowait
= current
->in_iowait
;
5083 current
->in_iowait
= 1;
5084 blk_schedule_flush_plug(current
);
5089 void io_schedule_finish(int token
)
5091 current
->in_iowait
= token
;
5095 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5096 * that process accounting knows that this is a task in IO wait state.
5098 long __sched
io_schedule_timeout(long timeout
)
5103 token
= io_schedule_prepare();
5104 ret
= schedule_timeout(timeout
);
5105 io_schedule_finish(token
);
5109 EXPORT_SYMBOL(io_schedule_timeout
);
5111 void io_schedule(void)
5115 token
= io_schedule_prepare();
5117 io_schedule_finish(token
);
5119 EXPORT_SYMBOL(io_schedule
);
5122 * sys_sched_get_priority_max - return maximum RT priority.
5123 * @policy: scheduling class.
5125 * Return: On success, this syscall returns the maximum
5126 * rt_priority that can be used by a given scheduling class.
5127 * On failure, a negative error code is returned.
5129 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5136 ret
= MAX_USER_RT_PRIO
-1;
5138 case SCHED_DEADLINE
:
5149 * sys_sched_get_priority_min - return minimum RT priority.
5150 * @policy: scheduling class.
5152 * Return: On success, this syscall returns the minimum
5153 * rt_priority that can be used by a given scheduling class.
5154 * On failure, a negative error code is returned.
5156 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5165 case SCHED_DEADLINE
:
5175 * sys_sched_rr_get_interval - return the default timeslice of a process.
5176 * @pid: pid of the process.
5177 * @interval: userspace pointer to the timeslice value.
5179 * this syscall writes the default timeslice value of a given process
5180 * into the user-space timespec buffer. A value of '0' means infinity.
5182 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5185 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5186 struct timespec __user
*, interval
)
5188 struct task_struct
*p
;
5189 unsigned int time_slice
;
5200 p
= find_process_by_pid(pid
);
5204 retval
= security_task_getscheduler(p
);
5208 rq
= task_rq_lock(p
, &rf
);
5210 if (p
->sched_class
->get_rr_interval
)
5211 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5212 task_rq_unlock(rq
, p
, &rf
);
5215 jiffies_to_timespec(time_slice
, &t
);
5216 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5224 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5226 void sched_show_task(struct task_struct
*p
)
5228 unsigned long free
= 0;
5230 unsigned long state
= p
->state
;
5232 if (!try_get_task_stack(p
))
5235 state
= __ffs(state
) + 1;
5236 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5237 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5238 if (state
== TASK_RUNNING
)
5239 printk(KERN_CONT
" running task ");
5240 #ifdef CONFIG_DEBUG_STACK_USAGE
5241 free
= stack_not_used(p
);
5246 ppid
= task_pid_nr(rcu_dereference(p
->real_parent
));
5248 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5249 task_pid_nr(p
), ppid
,
5250 (unsigned long)task_thread_info(p
)->flags
);
5252 print_worker_info(KERN_INFO
, p
);
5253 show_stack(p
, NULL
);
5257 void show_state_filter(unsigned long state_filter
)
5259 struct task_struct
*g
, *p
;
5261 #if BITS_PER_LONG == 32
5263 " task PC stack pid father\n");
5266 " task PC stack pid father\n");
5269 for_each_process_thread(g
, p
) {
5271 * reset the NMI-timeout, listing all files on a slow
5272 * console might take a lot of time:
5273 * Also, reset softlockup watchdogs on all CPUs, because
5274 * another CPU might be blocked waiting for us to process
5277 touch_nmi_watchdog();
5278 touch_all_softlockup_watchdogs();
5279 if (!state_filter
|| (p
->state
& state_filter
))
5283 #ifdef CONFIG_SCHED_DEBUG
5285 sysrq_sched_debug_show();
5289 * Only show locks if all tasks are dumped:
5292 debug_show_all_locks();
5295 void init_idle_bootup_task(struct task_struct
*idle
)
5297 idle
->sched_class
= &idle_sched_class
;
5301 * init_idle - set up an idle thread for a given CPU
5302 * @idle: task in question
5303 * @cpu: CPU the idle task belongs to
5305 * NOTE: this function does not set the idle thread's NEED_RESCHED
5306 * flag, to make booting more robust.
5308 void init_idle(struct task_struct
*idle
, int cpu
)
5310 struct rq
*rq
= cpu_rq(cpu
);
5311 unsigned long flags
;
5313 raw_spin_lock_irqsave(&idle
->pi_lock
, flags
);
5314 raw_spin_lock(&rq
->lock
);
5316 __sched_fork(0, idle
);
5317 idle
->state
= TASK_RUNNING
;
5318 idle
->se
.exec_start
= sched_clock();
5319 idle
->flags
|= PF_IDLE
;
5321 kasan_unpoison_task_stack(idle
);
5325 * Its possible that init_idle() gets called multiple times on a task,
5326 * in that case do_set_cpus_allowed() will not do the right thing.
5328 * And since this is boot we can forgo the serialization.
5330 set_cpus_allowed_common(idle
, cpumask_of(cpu
));
5333 * We're having a chicken and egg problem, even though we are
5334 * holding rq->lock, the CPU isn't yet set to this CPU so the
5335 * lockdep check in task_group() will fail.
5337 * Similar case to sched_fork(). / Alternatively we could
5338 * use task_rq_lock() here and obtain the other rq->lock.
5343 __set_task_cpu(idle
, cpu
);
5346 rq
->curr
= rq
->idle
= idle
;
5347 idle
->on_rq
= TASK_ON_RQ_QUEUED
;
5351 raw_spin_unlock(&rq
->lock
);
5352 raw_spin_unlock_irqrestore(&idle
->pi_lock
, flags
);
5354 /* Set the preempt count _outside_ the spinlocks! */
5355 init_idle_preempt_count(idle
, cpu
);
5358 * The idle tasks have their own, simple scheduling class:
5360 idle
->sched_class
= &idle_sched_class
;
5361 ftrace_graph_init_idle_task(idle
, cpu
);
5362 vtime_init_idle(idle
, cpu
);
5364 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
5368 int cpuset_cpumask_can_shrink(const struct cpumask
*cur
,
5369 const struct cpumask
*trial
)
5371 int ret
= 1, trial_cpus
;
5372 struct dl_bw
*cur_dl_b
;
5373 unsigned long flags
;
5375 if (!cpumask_weight(cur
))
5378 rcu_read_lock_sched();
5379 cur_dl_b
= dl_bw_of(cpumask_any(cur
));
5380 trial_cpus
= cpumask_weight(trial
);
5382 raw_spin_lock_irqsave(&cur_dl_b
->lock
, flags
);
5383 if (cur_dl_b
->bw
!= -1 &&
5384 cur_dl_b
->bw
* trial_cpus
< cur_dl_b
->total_bw
)
5386 raw_spin_unlock_irqrestore(&cur_dl_b
->lock
, flags
);
5387 rcu_read_unlock_sched();
5392 int task_can_attach(struct task_struct
*p
,
5393 const struct cpumask
*cs_cpus_allowed
)
5398 * Kthreads which disallow setaffinity shouldn't be moved
5399 * to a new cpuset; we don't want to change their CPU
5400 * affinity and isolating such threads by their set of
5401 * allowed nodes is unnecessary. Thus, cpusets are not
5402 * applicable for such threads. This prevents checking for
5403 * success of set_cpus_allowed_ptr() on all attached tasks
5404 * before cpus_allowed may be changed.
5406 if (p
->flags
& PF_NO_SETAFFINITY
) {
5412 if (dl_task(p
) && !cpumask_intersects(task_rq(p
)->rd
->span
,
5414 unsigned int dest_cpu
= cpumask_any_and(cpu_active_mask
,
5419 unsigned long flags
;
5421 rcu_read_lock_sched();
5422 dl_b
= dl_bw_of(dest_cpu
);
5423 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5424 cpus
= dl_bw_cpus(dest_cpu
);
5425 overflow
= __dl_overflow(dl_b
, cpus
, 0, p
->dl
.dl_bw
);
5430 * We reserve space for this task in the destination
5431 * root_domain, as we can't fail after this point.
5432 * We will free resources in the source root_domain
5433 * later on (see set_cpus_allowed_dl()).
5435 __dl_add(dl_b
, p
->dl
.dl_bw
);
5437 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5438 rcu_read_unlock_sched();
5448 bool sched_smp_initialized __read_mostly
;
5450 #ifdef CONFIG_NUMA_BALANCING
5451 /* Migrate current task p to target_cpu */
5452 int migrate_task_to(struct task_struct
*p
, int target_cpu
)
5454 struct migration_arg arg
= { p
, target_cpu
};
5455 int curr_cpu
= task_cpu(p
);
5457 if (curr_cpu
== target_cpu
)
5460 if (!cpumask_test_cpu(target_cpu
, tsk_cpus_allowed(p
)))
5463 /* TODO: This is not properly updating schedstats */
5465 trace_sched_move_numa(p
, curr_cpu
, target_cpu
);
5466 return stop_one_cpu(curr_cpu
, migration_cpu_stop
, &arg
);
5470 * Requeue a task on a given node and accurately track the number of NUMA
5471 * tasks on the runqueues
5473 void sched_setnuma(struct task_struct
*p
, int nid
)
5475 bool queued
, running
;
5479 rq
= task_rq_lock(p
, &rf
);
5480 queued
= task_on_rq_queued(p
);
5481 running
= task_current(rq
, p
);
5484 dequeue_task(rq
, p
, DEQUEUE_SAVE
);
5486 put_prev_task(rq
, p
);
5488 p
->numa_preferred_nid
= nid
;
5491 enqueue_task(rq
, p
, ENQUEUE_RESTORE
);
5493 set_curr_task(rq
, p
);
5494 task_rq_unlock(rq
, p
, &rf
);
5496 #endif /* CONFIG_NUMA_BALANCING */
5498 #ifdef CONFIG_HOTPLUG_CPU
5500 * Ensure that the idle task is using init_mm right before its CPU goes
5503 void idle_task_exit(void)
5505 struct mm_struct
*mm
= current
->active_mm
;
5507 BUG_ON(cpu_online(smp_processor_id()));
5509 if (mm
!= &init_mm
) {
5510 switch_mm_irqs_off(mm
, &init_mm
, current
);
5511 finish_arch_post_lock_switch();
5517 * Since this CPU is going 'away' for a while, fold any nr_active delta
5518 * we might have. Assumes we're called after migrate_tasks() so that the
5519 * nr_active count is stable. We need to take the teardown thread which
5520 * is calling this into account, so we hand in adjust = 1 to the load
5523 * Also see the comment "Global load-average calculations".
5525 static void calc_load_migrate(struct rq
*rq
)
5527 long delta
= calc_load_fold_active(rq
, 1);
5529 atomic_long_add(delta
, &calc_load_tasks
);
5532 static void put_prev_task_fake(struct rq
*rq
, struct task_struct
*prev
)
5536 static const struct sched_class fake_sched_class
= {
5537 .put_prev_task
= put_prev_task_fake
,
5540 static struct task_struct fake_task
= {
5542 * Avoid pull_{rt,dl}_task()
5544 .prio
= MAX_PRIO
+ 1,
5545 .sched_class
= &fake_sched_class
,
5549 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5550 * try_to_wake_up()->select_task_rq().
5552 * Called with rq->lock held even though we'er in stop_machine() and
5553 * there's no concurrency possible, we hold the required locks anyway
5554 * because of lock validation efforts.
5556 static void migrate_tasks(struct rq
*dead_rq
)
5558 struct rq
*rq
= dead_rq
;
5559 struct task_struct
*next
, *stop
= rq
->stop
;
5560 struct rq_flags rf
, old_rf
;
5564 * Fudge the rq selection such that the below task selection loop
5565 * doesn't get stuck on the currently eligible stop task.
5567 * We're currently inside stop_machine() and the rq is either stuck
5568 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5569 * either way we should never end up calling schedule() until we're
5575 * put_prev_task() and pick_next_task() sched
5576 * class method both need to have an up-to-date
5577 * value of rq->clock[_task]
5579 update_rq_clock(rq
);
5583 * There's this thread running, bail when that's the only
5586 if (rq
->nr_running
== 1)
5590 * pick_next_task() assumes pinned rq->lock:
5592 rq_pin_lock(rq
, &rf
);
5593 next
= pick_next_task(rq
, &fake_task
, &rf
);
5595 next
->sched_class
->put_prev_task(rq
, next
);
5598 * Rules for changing task_struct::cpus_allowed are holding
5599 * both pi_lock and rq->lock, such that holding either
5600 * stabilizes the mask.
5602 * Drop rq->lock is not quite as disastrous as it usually is
5603 * because !cpu_active at this point, which means load-balance
5604 * will not interfere. Also, stop-machine.
5606 rq_unpin_lock(rq
, &rf
);
5607 raw_spin_unlock(&rq
->lock
);
5608 raw_spin_lock(&next
->pi_lock
);
5609 raw_spin_lock(&rq
->lock
);
5612 * Since we're inside stop-machine, _nothing_ should have
5613 * changed the task, WARN if weird stuff happened, because in
5614 * that case the above rq->lock drop is a fail too.
5616 if (WARN_ON(task_rq(next
) != rq
|| !task_on_rq_queued(next
))) {
5617 raw_spin_unlock(&next
->pi_lock
);
5622 * __migrate_task() may return with a different
5623 * rq->lock held and a new cookie in 'rf', but we need
5624 * to preserve rf::clock_update_flags for 'dead_rq'.
5628 /* Find suitable destination for @next, with force if needed. */
5629 dest_cpu
= select_fallback_rq(dead_rq
->cpu
, next
);
5631 rq
= __migrate_task(rq
, next
, dest_cpu
);
5632 if (rq
!= dead_rq
) {
5633 raw_spin_unlock(&rq
->lock
);
5635 raw_spin_lock(&rq
->lock
);
5638 raw_spin_unlock(&next
->pi_lock
);
5643 #endif /* CONFIG_HOTPLUG_CPU */
5645 void set_rq_online(struct rq
*rq
)
5648 const struct sched_class
*class;
5650 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5653 for_each_class(class) {
5654 if (class->rq_online
)
5655 class->rq_online(rq
);
5660 void set_rq_offline(struct rq
*rq
)
5663 const struct sched_class
*class;
5665 for_each_class(class) {
5666 if (class->rq_offline
)
5667 class->rq_offline(rq
);
5670 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5675 static void set_cpu_rq_start_time(unsigned int cpu
)
5677 struct rq
*rq
= cpu_rq(cpu
);
5679 rq
->age_stamp
= sched_clock_cpu(cpu
);
5683 * used to mark begin/end of suspend/resume:
5685 static int num_cpus_frozen
;
5688 * Update cpusets according to cpu_active mask. If cpusets are
5689 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
5690 * around partition_sched_domains().
5692 * If we come here as part of a suspend/resume, don't touch cpusets because we
5693 * want to restore it back to its original state upon resume anyway.
5695 static void cpuset_cpu_active(void)
5697 if (cpuhp_tasks_frozen
) {
5699 * num_cpus_frozen tracks how many CPUs are involved in suspend
5700 * resume sequence. As long as this is not the last online
5701 * operation in the resume sequence, just build a single sched
5702 * domain, ignoring cpusets.
5705 if (likely(num_cpus_frozen
)) {
5706 partition_sched_domains(1, NULL
, NULL
);
5710 * This is the last CPU online operation. So fall through and
5711 * restore the original sched domains by considering the
5712 * cpuset configurations.
5715 cpuset_update_active_cpus(true);
5718 static int cpuset_cpu_inactive(unsigned int cpu
)
5720 unsigned long flags
;
5725 if (!cpuhp_tasks_frozen
) {
5726 rcu_read_lock_sched();
5727 dl_b
= dl_bw_of(cpu
);
5729 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
5730 cpus
= dl_bw_cpus(cpu
);
5731 overflow
= __dl_overflow(dl_b
, cpus
, 0, 0);
5732 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
5734 rcu_read_unlock_sched();
5738 cpuset_update_active_cpus(false);
5741 partition_sched_domains(1, NULL
, NULL
);
5746 int sched_cpu_activate(unsigned int cpu
)
5748 struct rq
*rq
= cpu_rq(cpu
);
5749 unsigned long flags
;
5751 set_cpu_active(cpu
, true);
5753 if (sched_smp_initialized
) {
5754 sched_domains_numa_masks_set(cpu
);
5755 cpuset_cpu_active();
5759 * Put the rq online, if not already. This happens:
5761 * 1) In the early boot process, because we build the real domains
5762 * after all CPUs have been brought up.
5764 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
5767 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5769 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5772 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5774 update_max_interval();
5779 int sched_cpu_deactivate(unsigned int cpu
)
5783 set_cpu_active(cpu
, false);
5785 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
5786 * users of this state to go away such that all new such users will
5789 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
5790 * not imply sync_sched(), so wait for both.
5792 * Do sync before park smpboot threads to take care the rcu boost case.
5794 if (IS_ENABLED(CONFIG_PREEMPT
))
5795 synchronize_rcu_mult(call_rcu
, call_rcu_sched
);
5799 if (!sched_smp_initialized
)
5802 ret
= cpuset_cpu_inactive(cpu
);
5804 set_cpu_active(cpu
, true);
5807 sched_domains_numa_masks_clear(cpu
);
5811 static void sched_rq_cpu_starting(unsigned int cpu
)
5813 struct rq
*rq
= cpu_rq(cpu
);
5815 rq
->calc_load_update
= calc_load_update
;
5816 update_max_interval();
5819 int sched_cpu_starting(unsigned int cpu
)
5821 set_cpu_rq_start_time(cpu
);
5822 sched_rq_cpu_starting(cpu
);
5826 #ifdef CONFIG_HOTPLUG_CPU
5827 int sched_cpu_dying(unsigned int cpu
)
5829 struct rq
*rq
= cpu_rq(cpu
);
5830 unsigned long flags
;
5832 /* Handle pending wakeups and then migrate everything off */
5833 sched_ttwu_pending();
5834 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5836 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5840 BUG_ON(rq
->nr_running
!= 1);
5841 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5842 calc_load_migrate(rq
);
5843 update_max_interval();
5844 nohz_balance_exit_idle(cpu
);
5850 #ifdef CONFIG_SCHED_SMT
5851 DEFINE_STATIC_KEY_FALSE(sched_smt_present
);
5853 static void sched_init_smt(void)
5856 * We've enumerated all CPUs and will assume that if any CPU
5857 * has SMT siblings, CPU0 will too.
5859 if (cpumask_weight(cpu_smt_mask(0)) > 1)
5860 static_branch_enable(&sched_smt_present
);
5863 static inline void sched_init_smt(void) { }
5866 void __init
sched_init_smp(void)
5868 cpumask_var_t non_isolated_cpus
;
5870 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
5871 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
5876 * There's no userspace yet to cause hotplug operations; hence all the
5877 * CPU masks are stable and all blatant races in the below code cannot
5880 mutex_lock(&sched_domains_mutex
);
5881 init_sched_domains(cpu_active_mask
);
5882 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
5883 if (cpumask_empty(non_isolated_cpus
))
5884 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
5885 mutex_unlock(&sched_domains_mutex
);
5887 /* Move init over to a non-isolated CPU */
5888 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
5890 sched_init_granularity();
5891 free_cpumask_var(non_isolated_cpus
);
5893 init_sched_rt_class();
5894 init_sched_dl_class();
5897 sched_clock_init_late();
5899 sched_smp_initialized
= true;
5902 static int __init
migration_init(void)
5904 sched_rq_cpu_starting(smp_processor_id());
5907 early_initcall(migration_init
);
5910 void __init
sched_init_smp(void)
5912 sched_init_granularity();
5913 sched_clock_init_late();
5915 #endif /* CONFIG_SMP */
5917 int in_sched_functions(unsigned long addr
)
5919 return in_lock_functions(addr
) ||
5920 (addr
>= (unsigned long)__sched_text_start
5921 && addr
< (unsigned long)__sched_text_end
);
5924 #ifdef CONFIG_CGROUP_SCHED
5926 * Default task group.
5927 * Every task in system belongs to this group at bootup.
5929 struct task_group root_task_group
;
5930 LIST_HEAD(task_groups
);
5932 /* Cacheline aligned slab cache for task_group */
5933 static struct kmem_cache
*task_group_cache __read_mostly
;
5936 DECLARE_PER_CPU(cpumask_var_t
, load_balance_mask
);
5937 DECLARE_PER_CPU(cpumask_var_t
, select_idle_mask
);
5939 #define WAIT_TABLE_BITS 8
5940 #define WAIT_TABLE_SIZE (1 << WAIT_TABLE_BITS)
5941 static wait_queue_head_t bit_wait_table
[WAIT_TABLE_SIZE
] __cacheline_aligned
;
5943 wait_queue_head_t
*bit_waitqueue(void *word
, int bit
)
5945 const int shift
= BITS_PER_LONG
== 32 ? 5 : 6;
5946 unsigned long val
= (unsigned long)word
<< shift
| bit
;
5948 return bit_wait_table
+ hash_long(val
, WAIT_TABLE_BITS
);
5950 EXPORT_SYMBOL(bit_waitqueue
);
5952 void __init
sched_init(void)
5955 unsigned long alloc_size
= 0, ptr
;
5959 for (i
= 0; i
< WAIT_TABLE_SIZE
; i
++)
5960 init_waitqueue_head(bit_wait_table
+ i
);
5962 #ifdef CONFIG_FAIR_GROUP_SCHED
5963 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
5965 #ifdef CONFIG_RT_GROUP_SCHED
5966 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
5969 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
5971 #ifdef CONFIG_FAIR_GROUP_SCHED
5972 root_task_group
.se
= (struct sched_entity
**)ptr
;
5973 ptr
+= nr_cpu_ids
* sizeof(void **);
5975 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
5976 ptr
+= nr_cpu_ids
* sizeof(void **);
5978 #endif /* CONFIG_FAIR_GROUP_SCHED */
5979 #ifdef CONFIG_RT_GROUP_SCHED
5980 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
5981 ptr
+= nr_cpu_ids
* sizeof(void **);
5983 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
5984 ptr
+= nr_cpu_ids
* sizeof(void **);
5986 #endif /* CONFIG_RT_GROUP_SCHED */
5988 #ifdef CONFIG_CPUMASK_OFFSTACK
5989 for_each_possible_cpu(i
) {
5990 per_cpu(load_balance_mask
, i
) = (cpumask_var_t
)kzalloc_node(
5991 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
5992 per_cpu(select_idle_mask
, i
) = (cpumask_var_t
)kzalloc_node(
5993 cpumask_size(), GFP_KERNEL
, cpu_to_node(i
));
5995 #endif /* CONFIG_CPUMASK_OFFSTACK */
5997 init_rt_bandwidth(&def_rt_bandwidth
, global_rt_period(), global_rt_runtime());
5998 init_dl_bandwidth(&def_dl_bandwidth
, global_rt_period(), global_rt_runtime());
6001 init_defrootdomain();
6004 #ifdef CONFIG_RT_GROUP_SCHED
6005 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
6006 global_rt_period(), global_rt_runtime());
6007 #endif /* CONFIG_RT_GROUP_SCHED */
6009 #ifdef CONFIG_CGROUP_SCHED
6010 task_group_cache
= KMEM_CACHE(task_group
, 0);
6012 list_add(&root_task_group
.list
, &task_groups
);
6013 INIT_LIST_HEAD(&root_task_group
.children
);
6014 INIT_LIST_HEAD(&root_task_group
.siblings
);
6015 autogroup_init(&init_task
);
6016 #endif /* CONFIG_CGROUP_SCHED */
6018 for_each_possible_cpu(i
) {
6022 raw_spin_lock_init(&rq
->lock
);
6024 rq
->calc_load_active
= 0;
6025 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
6026 init_cfs_rq(&rq
->cfs
);
6027 init_rt_rq(&rq
->rt
);
6028 init_dl_rq(&rq
->dl
);
6029 #ifdef CONFIG_FAIR_GROUP_SCHED
6030 root_task_group
.shares
= ROOT_TASK_GROUP_LOAD
;
6031 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
6032 rq
->tmp_alone_branch
= &rq
->leaf_cfs_rq_list
;
6034 * How much CPU bandwidth does root_task_group get?
6036 * In case of task-groups formed thr' the cgroup filesystem, it
6037 * gets 100% of the CPU resources in the system. This overall
6038 * system CPU resource is divided among the tasks of
6039 * root_task_group and its child task-groups in a fair manner,
6040 * based on each entity's (task or task-group's) weight
6041 * (se->load.weight).
6043 * In other words, if root_task_group has 10 tasks of weight
6044 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6045 * then A0's share of the CPU resource is:
6047 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6049 * We achieve this by letting root_task_group's tasks sit
6050 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6052 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
6053 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
6054 #endif /* CONFIG_FAIR_GROUP_SCHED */
6056 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
6057 #ifdef CONFIG_RT_GROUP_SCHED
6058 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
6061 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
6062 rq
->cpu_load
[j
] = 0;
6067 rq
->cpu_capacity
= rq
->cpu_capacity_orig
= SCHED_CAPACITY_SCALE
;
6068 rq
->balance_callback
= NULL
;
6069 rq
->active_balance
= 0;
6070 rq
->next_balance
= jiffies
;
6075 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
6076 rq
->max_idle_balance_cost
= sysctl_sched_migration_cost
;
6078 INIT_LIST_HEAD(&rq
->cfs_tasks
);
6080 rq_attach_root(rq
, &def_root_domain
);
6081 #ifdef CONFIG_NO_HZ_COMMON
6082 rq
->last_load_update_tick
= jiffies
;
6085 #ifdef CONFIG_NO_HZ_FULL
6086 rq
->last_sched_tick
= 0;
6088 #endif /* CONFIG_SMP */
6090 atomic_set(&rq
->nr_iowait
, 0);
6093 set_load_weight(&init_task
);
6096 * The boot idle thread does lazy MMU switching as well:
6098 atomic_inc(&init_mm
.mm_count
);
6099 enter_lazy_tlb(&init_mm
, current
);
6102 * Make us the idle thread. Technically, schedule() should not be
6103 * called from this thread, however somewhere below it might be,
6104 * but because we are the idle thread, we just pick up running again
6105 * when this runqueue becomes "idle".
6107 init_idle(current
, smp_processor_id());
6109 calc_load_update
= jiffies
+ LOAD_FREQ
;
6112 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
6113 /* May be allocated at isolcpus cmdline parse time */
6114 if (cpu_isolated_map
== NULL
)
6115 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
6116 idle_thread_set_boot_cpu();
6117 set_cpu_rq_start_time(smp_processor_id());
6119 init_sched_fair_class();
6123 scheduler_running
= 1;
6126 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6127 static inline int preempt_count_equals(int preempt_offset
)
6129 int nested
= preempt_count() + rcu_preempt_depth();
6131 return (nested
== preempt_offset
);
6134 void __might_sleep(const char *file
, int line
, int preempt_offset
)
6137 * Blocking primitives will set (and therefore destroy) current->state,
6138 * since we will exit with TASK_RUNNING make sure we enter with it,
6139 * otherwise we will destroy state.
6141 WARN_ONCE(current
->state
!= TASK_RUNNING
&& current
->task_state_change
,
6142 "do not call blocking ops when !TASK_RUNNING; "
6143 "state=%lx set at [<%p>] %pS\n",
6145 (void *)current
->task_state_change
,
6146 (void *)current
->task_state_change
);
6148 ___might_sleep(file
, line
, preempt_offset
);
6150 EXPORT_SYMBOL(__might_sleep
);
6152 void ___might_sleep(const char *file
, int line
, int preempt_offset
)
6154 /* Ratelimiting timestamp: */
6155 static unsigned long prev_jiffy
;
6157 unsigned long preempt_disable_ip
;
6159 /* WARN_ON_ONCE() by default, no rate limit required: */
6162 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled() &&
6163 !is_idle_task(current
)) ||
6164 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
6166 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6168 prev_jiffy
= jiffies
;
6170 /* Save this before calling printk(), since that will clobber it: */
6171 preempt_disable_ip
= get_preempt_disable_ip(current
);
6174 "BUG: sleeping function called from invalid context at %s:%d\n",
6177 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6178 in_atomic(), irqs_disabled(),
6179 current
->pid
, current
->comm
);
6181 if (task_stack_end_corrupted(current
))
6182 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
6184 debug_show_held_locks(current
);
6185 if (irqs_disabled())
6186 print_irqtrace_events(current
);
6187 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT
)
6188 && !preempt_count_equals(preempt_offset
)) {
6189 pr_err("Preemption disabled at:");
6190 print_ip_sym(preempt_disable_ip
);
6194 add_taint(TAINT_WARN
, LOCKDEP_STILL_OK
);
6196 EXPORT_SYMBOL(___might_sleep
);
6199 #ifdef CONFIG_MAGIC_SYSRQ
6200 void normalize_rt_tasks(void)
6202 struct task_struct
*g
, *p
;
6203 struct sched_attr attr
= {
6204 .sched_policy
= SCHED_NORMAL
,
6207 read_lock(&tasklist_lock
);
6208 for_each_process_thread(g
, p
) {
6210 * Only normalize user tasks:
6212 if (p
->flags
& PF_KTHREAD
)
6215 p
->se
.exec_start
= 0;
6216 schedstat_set(p
->se
.statistics
.wait_start
, 0);
6217 schedstat_set(p
->se
.statistics
.sleep_start
, 0);
6218 schedstat_set(p
->se
.statistics
.block_start
, 0);
6220 if (!dl_task(p
) && !rt_task(p
)) {
6222 * Renice negative nice level userspace
6225 if (task_nice(p
) < 0)
6226 set_user_nice(p
, 0);
6230 __sched_setscheduler(p
, &attr
, false, false);
6232 read_unlock(&tasklist_lock
);
6235 #endif /* CONFIG_MAGIC_SYSRQ */
6237 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6239 * These functions are only useful for the IA64 MCA handling, or kdb.
6241 * They can only be called when the whole system has been
6242 * stopped - every CPU needs to be quiescent, and no scheduling
6243 * activity can take place. Using them for anything else would
6244 * be a serious bug, and as a result, they aren't even visible
6245 * under any other configuration.
6249 * curr_task - return the current task for a given CPU.
6250 * @cpu: the processor in question.
6252 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6254 * Return: The current task for @cpu.
6256 struct task_struct
*curr_task(int cpu
)
6258 return cpu_curr(cpu
);
6261 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6265 * set_curr_task - set the current task for a given CPU.
6266 * @cpu: the processor in question.
6267 * @p: the task pointer to set.
6269 * Description: This function must only be used when non-maskable interrupts
6270 * are serviced on a separate stack. It allows the architecture to switch the
6271 * notion of the current task on a CPU in a non-blocking manner. This function
6272 * must be called with all CPU's synchronized, and interrupts disabled, the
6273 * and caller must save the original value of the current task (see
6274 * curr_task() above) and restore that value before reenabling interrupts and
6275 * re-starting the system.
6277 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6279 void ia64_set_curr_task(int cpu
, struct task_struct
*p
)
6286 #ifdef CONFIG_CGROUP_SCHED
6287 /* task_group_lock serializes the addition/removal of task groups */
6288 static DEFINE_SPINLOCK(task_group_lock
);
6290 static void sched_free_group(struct task_group
*tg
)
6292 free_fair_sched_group(tg
);
6293 free_rt_sched_group(tg
);
6295 kmem_cache_free(task_group_cache
, tg
);
6298 /* allocate runqueue etc for a new task group */
6299 struct task_group
*sched_create_group(struct task_group
*parent
)
6301 struct task_group
*tg
;
6303 tg
= kmem_cache_alloc(task_group_cache
, GFP_KERNEL
| __GFP_ZERO
);
6305 return ERR_PTR(-ENOMEM
);
6307 if (!alloc_fair_sched_group(tg
, parent
))
6310 if (!alloc_rt_sched_group(tg
, parent
))
6316 sched_free_group(tg
);
6317 return ERR_PTR(-ENOMEM
);
6320 void sched_online_group(struct task_group
*tg
, struct task_group
*parent
)
6322 unsigned long flags
;
6324 spin_lock_irqsave(&task_group_lock
, flags
);
6325 list_add_rcu(&tg
->list
, &task_groups
);
6327 /* Root should already exist: */
6330 tg
->parent
= parent
;
6331 INIT_LIST_HEAD(&tg
->children
);
6332 list_add_rcu(&tg
->siblings
, &parent
->children
);
6333 spin_unlock_irqrestore(&task_group_lock
, flags
);
6335 online_fair_sched_group(tg
);
6338 /* rcu callback to free various structures associated with a task group */
6339 static void sched_free_group_rcu(struct rcu_head
*rhp
)
6341 /* Now it should be safe to free those cfs_rqs: */
6342 sched_free_group(container_of(rhp
, struct task_group
, rcu
));
6345 void sched_destroy_group(struct task_group
*tg
)
6347 /* Wait for possible concurrent references to cfs_rqs complete: */
6348 call_rcu(&tg
->rcu
, sched_free_group_rcu
);
6351 void sched_offline_group(struct task_group
*tg
)
6353 unsigned long flags
;
6355 /* End participation in shares distribution: */
6356 unregister_fair_sched_group(tg
);
6358 spin_lock_irqsave(&task_group_lock
, flags
);
6359 list_del_rcu(&tg
->list
);
6360 list_del_rcu(&tg
->siblings
);
6361 spin_unlock_irqrestore(&task_group_lock
, flags
);
6364 static void sched_change_group(struct task_struct
*tsk
, int type
)
6366 struct task_group
*tg
;
6369 * All callers are synchronized by task_rq_lock(); we do not use RCU
6370 * which is pointless here. Thus, we pass "true" to task_css_check()
6371 * to prevent lockdep warnings.
6373 tg
= container_of(task_css_check(tsk
, cpu_cgrp_id
, true),
6374 struct task_group
, css
);
6375 tg
= autogroup_task_group(tsk
, tg
);
6376 tsk
->sched_task_group
= tg
;
6378 #ifdef CONFIG_FAIR_GROUP_SCHED
6379 if (tsk
->sched_class
->task_change_group
)
6380 tsk
->sched_class
->task_change_group(tsk
, type
);
6383 set_task_rq(tsk
, task_cpu(tsk
));
6387 * Change task's runqueue when it moves between groups.
6389 * The caller of this function should have put the task in its new group by
6390 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
6393 void sched_move_task(struct task_struct
*tsk
)
6395 int queued
, running
;
6399 rq
= task_rq_lock(tsk
, &rf
);
6400 update_rq_clock(rq
);
6402 running
= task_current(rq
, tsk
);
6403 queued
= task_on_rq_queued(tsk
);
6406 dequeue_task(rq
, tsk
, DEQUEUE_SAVE
| DEQUEUE_MOVE
);
6408 put_prev_task(rq
, tsk
);
6410 sched_change_group(tsk
, TASK_MOVE_GROUP
);
6413 enqueue_task(rq
, tsk
, ENQUEUE_RESTORE
| ENQUEUE_MOVE
);
6415 set_curr_task(rq
, tsk
);
6417 task_rq_unlock(rq
, tsk
, &rf
);
6419 #endif /* CONFIG_CGROUP_SCHED */
6421 #ifdef CONFIG_RT_GROUP_SCHED
6423 * Ensure that the real time constraints are schedulable.
6425 static DEFINE_MUTEX(rt_constraints_mutex
);
6427 /* Must be called with tasklist_lock held */
6428 static inline int tg_has_rt_tasks(struct task_group
*tg
)
6430 struct task_struct
*g
, *p
;
6433 * Autogroups do not have RT tasks; see autogroup_create().
6435 if (task_group_is_autogroup(tg
))
6438 for_each_process_thread(g
, p
) {
6439 if (rt_task(p
) && task_group(p
) == tg
)
6446 struct rt_schedulable_data
{
6447 struct task_group
*tg
;
6452 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
6454 struct rt_schedulable_data
*d
= data
;
6455 struct task_group
*child
;
6456 unsigned long total
, sum
= 0;
6457 u64 period
, runtime
;
6459 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
6460 runtime
= tg
->rt_bandwidth
.rt_runtime
;
6463 period
= d
->rt_period
;
6464 runtime
= d
->rt_runtime
;
6468 * Cannot have more runtime than the period.
6470 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
6474 * Ensure we don't starve existing RT tasks.
6476 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
6479 total
= to_ratio(period
, runtime
);
6482 * Nobody can have more than the global setting allows.
6484 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
6488 * The sum of our children's runtime should not exceed our own.
6490 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
6491 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
6492 runtime
= child
->rt_bandwidth
.rt_runtime
;
6494 if (child
== d
->tg
) {
6495 period
= d
->rt_period
;
6496 runtime
= d
->rt_runtime
;
6499 sum
+= to_ratio(period
, runtime
);
6508 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
6512 struct rt_schedulable_data data
= {
6514 .rt_period
= period
,
6515 .rt_runtime
= runtime
,
6519 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
6525 static int tg_set_rt_bandwidth(struct task_group
*tg
,
6526 u64 rt_period
, u64 rt_runtime
)
6531 * Disallowing the root group RT runtime is BAD, it would disallow the
6532 * kernel creating (and or operating) RT threads.
6534 if (tg
== &root_task_group
&& rt_runtime
== 0)
6537 /* No period doesn't make any sense. */
6541 mutex_lock(&rt_constraints_mutex
);
6542 read_lock(&tasklist_lock
);
6543 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
6547 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
6548 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
6549 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
6551 for_each_possible_cpu(i
) {
6552 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
6554 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
6555 rt_rq
->rt_runtime
= rt_runtime
;
6556 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
6558 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
6560 read_unlock(&tasklist_lock
);
6561 mutex_unlock(&rt_constraints_mutex
);
6566 static int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
6568 u64 rt_runtime
, rt_period
;
6570 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
6571 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
6572 if (rt_runtime_us
< 0)
6573 rt_runtime
= RUNTIME_INF
;
6575 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
6578 static long sched_group_rt_runtime(struct task_group
*tg
)
6582 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
6585 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
6586 do_div(rt_runtime_us
, NSEC_PER_USEC
);
6587 return rt_runtime_us
;
6590 static int sched_group_set_rt_period(struct task_group
*tg
, u64 rt_period_us
)
6592 u64 rt_runtime
, rt_period
;
6594 rt_period
= rt_period_us
* NSEC_PER_USEC
;
6595 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
6597 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
6600 static long sched_group_rt_period(struct task_group
*tg
)
6604 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
6605 do_div(rt_period_us
, NSEC_PER_USEC
);
6606 return rt_period_us
;
6608 #endif /* CONFIG_RT_GROUP_SCHED */
6610 #ifdef CONFIG_RT_GROUP_SCHED
6611 static int sched_rt_global_constraints(void)
6615 mutex_lock(&rt_constraints_mutex
);
6616 read_lock(&tasklist_lock
);
6617 ret
= __rt_schedulable(NULL
, 0, 0);
6618 read_unlock(&tasklist_lock
);
6619 mutex_unlock(&rt_constraints_mutex
);
6624 static int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
6626 /* Don't accept realtime tasks when there is no way for them to run */
6627 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
6633 #else /* !CONFIG_RT_GROUP_SCHED */
6634 static int sched_rt_global_constraints(void)
6636 unsigned long flags
;
6639 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
6640 for_each_possible_cpu(i
) {
6641 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
6643 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
6644 rt_rq
->rt_runtime
= global_rt_runtime();
6645 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
6647 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
6651 #endif /* CONFIG_RT_GROUP_SCHED */
6653 static int sched_dl_global_validate(void)
6655 u64 runtime
= global_rt_runtime();
6656 u64 period
= global_rt_period();
6657 u64 new_bw
= to_ratio(period
, runtime
);
6660 unsigned long flags
;
6663 * Here we want to check the bandwidth not being set to some
6664 * value smaller than the currently allocated bandwidth in
6665 * any of the root_domains.
6667 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
6668 * cycling on root_domains... Discussion on different/better
6669 * solutions is welcome!
6671 for_each_possible_cpu(cpu
) {
6672 rcu_read_lock_sched();
6673 dl_b
= dl_bw_of(cpu
);
6675 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
6676 if (new_bw
< dl_b
->total_bw
)
6678 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
6680 rcu_read_unlock_sched();
6689 static void sched_dl_do_global(void)
6694 unsigned long flags
;
6696 def_dl_bandwidth
.dl_period
= global_rt_period();
6697 def_dl_bandwidth
.dl_runtime
= global_rt_runtime();
6699 if (global_rt_runtime() != RUNTIME_INF
)
6700 new_bw
= to_ratio(global_rt_period(), global_rt_runtime());
6703 * FIXME: As above...
6705 for_each_possible_cpu(cpu
) {
6706 rcu_read_lock_sched();
6707 dl_b
= dl_bw_of(cpu
);
6709 raw_spin_lock_irqsave(&dl_b
->lock
, flags
);
6711 raw_spin_unlock_irqrestore(&dl_b
->lock
, flags
);
6713 rcu_read_unlock_sched();
6717 static int sched_rt_global_validate(void)
6719 if (sysctl_sched_rt_period
<= 0)
6722 if ((sysctl_sched_rt_runtime
!= RUNTIME_INF
) &&
6723 (sysctl_sched_rt_runtime
> sysctl_sched_rt_period
))
6729 static void sched_rt_do_global(void)
6731 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
6732 def_rt_bandwidth
.rt_period
= ns_to_ktime(global_rt_period());
6735 int sched_rt_handler(struct ctl_table
*table
, int write
,
6736 void __user
*buffer
, size_t *lenp
,
6739 int old_period
, old_runtime
;
6740 static DEFINE_MUTEX(mutex
);
6744 old_period
= sysctl_sched_rt_period
;
6745 old_runtime
= sysctl_sched_rt_runtime
;
6747 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
6749 if (!ret
&& write
) {
6750 ret
= sched_rt_global_validate();
6754 ret
= sched_dl_global_validate();
6758 ret
= sched_rt_global_constraints();
6762 sched_rt_do_global();
6763 sched_dl_do_global();
6767 sysctl_sched_rt_period
= old_period
;
6768 sysctl_sched_rt_runtime
= old_runtime
;
6770 mutex_unlock(&mutex
);
6775 int sched_rr_handler(struct ctl_table
*table
, int write
,
6776 void __user
*buffer
, size_t *lenp
,
6780 static DEFINE_MUTEX(mutex
);
6783 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
6785 * Make sure that internally we keep jiffies.
6786 * Also, writing zero resets the timeslice to default:
6788 if (!ret
&& write
) {
6789 sched_rr_timeslice
=
6790 sysctl_sched_rr_timeslice
<= 0 ? RR_TIMESLICE
:
6791 msecs_to_jiffies(sysctl_sched_rr_timeslice
);
6793 mutex_unlock(&mutex
);
6797 #ifdef CONFIG_CGROUP_SCHED
6799 static inline struct task_group
*css_tg(struct cgroup_subsys_state
*css
)
6801 return css
? container_of(css
, struct task_group
, css
) : NULL
;
6804 static struct cgroup_subsys_state
*
6805 cpu_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6807 struct task_group
*parent
= css_tg(parent_css
);
6808 struct task_group
*tg
;
6811 /* This is early initialization for the top cgroup */
6812 return &root_task_group
.css
;
6815 tg
= sched_create_group(parent
);
6817 return ERR_PTR(-ENOMEM
);
6819 sched_online_group(tg
, parent
);
6824 static void cpu_cgroup_css_released(struct cgroup_subsys_state
*css
)
6826 struct task_group
*tg
= css_tg(css
);
6828 sched_offline_group(tg
);
6831 static void cpu_cgroup_css_free(struct cgroup_subsys_state
*css
)
6833 struct task_group
*tg
= css_tg(css
);
6836 * Relies on the RCU grace period between css_released() and this.
6838 sched_free_group(tg
);
6842 * This is called before wake_up_new_task(), therefore we really only
6843 * have to set its group bits, all the other stuff does not apply.
6845 static void cpu_cgroup_fork(struct task_struct
*task
)
6850 rq
= task_rq_lock(task
, &rf
);
6852 update_rq_clock(rq
);
6853 sched_change_group(task
, TASK_SET_GROUP
);
6855 task_rq_unlock(rq
, task
, &rf
);
6858 static int cpu_cgroup_can_attach(struct cgroup_taskset
*tset
)
6860 struct task_struct
*task
;
6861 struct cgroup_subsys_state
*css
;
6864 cgroup_taskset_for_each(task
, css
, tset
) {
6865 #ifdef CONFIG_RT_GROUP_SCHED
6866 if (!sched_rt_can_attach(css_tg(css
), task
))
6869 /* We don't support RT-tasks being in separate groups */
6870 if (task
->sched_class
!= &fair_sched_class
)
6874 * Serialize against wake_up_new_task() such that if its
6875 * running, we're sure to observe its full state.
6877 raw_spin_lock_irq(&task
->pi_lock
);
6879 * Avoid calling sched_move_task() before wake_up_new_task()
6880 * has happened. This would lead to problems with PELT, due to
6881 * move wanting to detach+attach while we're not attached yet.
6883 if (task
->state
== TASK_NEW
)
6885 raw_spin_unlock_irq(&task
->pi_lock
);
6893 static void cpu_cgroup_attach(struct cgroup_taskset
*tset
)
6895 struct task_struct
*task
;
6896 struct cgroup_subsys_state
*css
;
6898 cgroup_taskset_for_each(task
, css
, tset
)
6899 sched_move_task(task
);
6902 #ifdef CONFIG_FAIR_GROUP_SCHED
6903 static int cpu_shares_write_u64(struct cgroup_subsys_state
*css
,
6904 struct cftype
*cftype
, u64 shareval
)
6906 return sched_group_set_shares(css_tg(css
), scale_load(shareval
));
6909 static u64
cpu_shares_read_u64(struct cgroup_subsys_state
*css
,
6912 struct task_group
*tg
= css_tg(css
);
6914 return (u64
) scale_load_down(tg
->shares
);
6917 #ifdef CONFIG_CFS_BANDWIDTH
6918 static DEFINE_MUTEX(cfs_constraints_mutex
);
6920 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
6921 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
6923 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
6925 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
6927 int i
, ret
= 0, runtime_enabled
, runtime_was_enabled
;
6928 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
6930 if (tg
== &root_task_group
)
6934 * Ensure we have at some amount of bandwidth every period. This is
6935 * to prevent reaching a state of large arrears when throttled via
6936 * entity_tick() resulting in prolonged exit starvation.
6938 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
6942 * Likewise, bound things on the otherside by preventing insane quota
6943 * periods. This also allows us to normalize in computing quota
6946 if (period
> max_cfs_quota_period
)
6950 * Prevent race between setting of cfs_rq->runtime_enabled and
6951 * unthrottle_offline_cfs_rqs().
6954 mutex_lock(&cfs_constraints_mutex
);
6955 ret
= __cfs_schedulable(tg
, period
, quota
);
6959 runtime_enabled
= quota
!= RUNTIME_INF
;
6960 runtime_was_enabled
= cfs_b
->quota
!= RUNTIME_INF
;
6962 * If we need to toggle cfs_bandwidth_used, off->on must occur
6963 * before making related changes, and on->off must occur afterwards
6965 if (runtime_enabled
&& !runtime_was_enabled
)
6966 cfs_bandwidth_usage_inc();
6967 raw_spin_lock_irq(&cfs_b
->lock
);
6968 cfs_b
->period
= ns_to_ktime(period
);
6969 cfs_b
->quota
= quota
;
6971 __refill_cfs_bandwidth_runtime(cfs_b
);
6973 /* Restart the period timer (if active) to handle new period expiry: */
6974 if (runtime_enabled
)
6975 start_cfs_bandwidth(cfs_b
);
6977 raw_spin_unlock_irq(&cfs_b
->lock
);
6979 for_each_online_cpu(i
) {
6980 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
6981 struct rq
*rq
= cfs_rq
->rq
;
6983 raw_spin_lock_irq(&rq
->lock
);
6984 cfs_rq
->runtime_enabled
= runtime_enabled
;
6985 cfs_rq
->runtime_remaining
= 0;
6987 if (cfs_rq
->throttled
)
6988 unthrottle_cfs_rq(cfs_rq
);
6989 raw_spin_unlock_irq(&rq
->lock
);
6991 if (runtime_was_enabled
&& !runtime_enabled
)
6992 cfs_bandwidth_usage_dec();
6994 mutex_unlock(&cfs_constraints_mutex
);
7000 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
7004 period
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7005 if (cfs_quota_us
< 0)
7006 quota
= RUNTIME_INF
;
7008 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
7010 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7013 long tg_get_cfs_quota(struct task_group
*tg
)
7017 if (tg
->cfs_bandwidth
.quota
== RUNTIME_INF
)
7020 quota_us
= tg
->cfs_bandwidth
.quota
;
7021 do_div(quota_us
, NSEC_PER_USEC
);
7026 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
7030 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
7031 quota
= tg
->cfs_bandwidth
.quota
;
7033 return tg_set_cfs_bandwidth(tg
, period
, quota
);
7036 long tg_get_cfs_period(struct task_group
*tg
)
7040 cfs_period_us
= ktime_to_ns(tg
->cfs_bandwidth
.period
);
7041 do_div(cfs_period_us
, NSEC_PER_USEC
);
7043 return cfs_period_us
;
7046 static s64
cpu_cfs_quota_read_s64(struct cgroup_subsys_state
*css
,
7049 return tg_get_cfs_quota(css_tg(css
));
7052 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state
*css
,
7053 struct cftype
*cftype
, s64 cfs_quota_us
)
7055 return tg_set_cfs_quota(css_tg(css
), cfs_quota_us
);
7058 static u64
cpu_cfs_period_read_u64(struct cgroup_subsys_state
*css
,
7061 return tg_get_cfs_period(css_tg(css
));
7064 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state
*css
,
7065 struct cftype
*cftype
, u64 cfs_period_us
)
7067 return tg_set_cfs_period(css_tg(css
), cfs_period_us
);
7070 struct cfs_schedulable_data
{
7071 struct task_group
*tg
;
7076 * normalize group quota/period to be quota/max_period
7077 * note: units are usecs
7079 static u64
normalize_cfs_quota(struct task_group
*tg
,
7080 struct cfs_schedulable_data
*d
)
7088 period
= tg_get_cfs_period(tg
);
7089 quota
= tg_get_cfs_quota(tg
);
7092 /* note: these should typically be equivalent */
7093 if (quota
== RUNTIME_INF
|| quota
== -1)
7096 return to_ratio(period
, quota
);
7099 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
7101 struct cfs_schedulable_data
*d
= data
;
7102 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7103 s64 quota
= 0, parent_quota
= -1;
7106 quota
= RUNTIME_INF
;
7108 struct cfs_bandwidth
*parent_b
= &tg
->parent
->cfs_bandwidth
;
7110 quota
= normalize_cfs_quota(tg
, d
);
7111 parent_quota
= parent_b
->hierarchical_quota
;
7114 * Ensure max(child_quota) <= parent_quota, inherit when no
7117 if (quota
== RUNTIME_INF
)
7118 quota
= parent_quota
;
7119 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
7122 cfs_b
->hierarchical_quota
= quota
;
7127 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
7130 struct cfs_schedulable_data data
= {
7136 if (quota
!= RUNTIME_INF
) {
7137 do_div(data
.period
, NSEC_PER_USEC
);
7138 do_div(data
.quota
, NSEC_PER_USEC
);
7142 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
7148 static int cpu_stats_show(struct seq_file
*sf
, void *v
)
7150 struct task_group
*tg
= css_tg(seq_css(sf
));
7151 struct cfs_bandwidth
*cfs_b
= &tg
->cfs_bandwidth
;
7153 seq_printf(sf
, "nr_periods %d\n", cfs_b
->nr_periods
);
7154 seq_printf(sf
, "nr_throttled %d\n", cfs_b
->nr_throttled
);
7155 seq_printf(sf
, "throttled_time %llu\n", cfs_b
->throttled_time
);
7159 #endif /* CONFIG_CFS_BANDWIDTH */
7160 #endif /* CONFIG_FAIR_GROUP_SCHED */
7162 #ifdef CONFIG_RT_GROUP_SCHED
7163 static int cpu_rt_runtime_write(struct cgroup_subsys_state
*css
,
7164 struct cftype
*cft
, s64 val
)
7166 return sched_group_set_rt_runtime(css_tg(css
), val
);
7169 static s64
cpu_rt_runtime_read(struct cgroup_subsys_state
*css
,
7172 return sched_group_rt_runtime(css_tg(css
));
7175 static int cpu_rt_period_write_uint(struct cgroup_subsys_state
*css
,
7176 struct cftype
*cftype
, u64 rt_period_us
)
7178 return sched_group_set_rt_period(css_tg(css
), rt_period_us
);
7181 static u64
cpu_rt_period_read_uint(struct cgroup_subsys_state
*css
,
7184 return sched_group_rt_period(css_tg(css
));
7186 #endif /* CONFIG_RT_GROUP_SCHED */
7188 static struct cftype cpu_files
[] = {
7189 #ifdef CONFIG_FAIR_GROUP_SCHED
7192 .read_u64
= cpu_shares_read_u64
,
7193 .write_u64
= cpu_shares_write_u64
,
7196 #ifdef CONFIG_CFS_BANDWIDTH
7198 .name
= "cfs_quota_us",
7199 .read_s64
= cpu_cfs_quota_read_s64
,
7200 .write_s64
= cpu_cfs_quota_write_s64
,
7203 .name
= "cfs_period_us",
7204 .read_u64
= cpu_cfs_period_read_u64
,
7205 .write_u64
= cpu_cfs_period_write_u64
,
7209 .seq_show
= cpu_stats_show
,
7212 #ifdef CONFIG_RT_GROUP_SCHED
7214 .name
= "rt_runtime_us",
7215 .read_s64
= cpu_rt_runtime_read
,
7216 .write_s64
= cpu_rt_runtime_write
,
7219 .name
= "rt_period_us",
7220 .read_u64
= cpu_rt_period_read_uint
,
7221 .write_u64
= cpu_rt_period_write_uint
,
7227 struct cgroup_subsys cpu_cgrp_subsys
= {
7228 .css_alloc
= cpu_cgroup_css_alloc
,
7229 .css_released
= cpu_cgroup_css_released
,
7230 .css_free
= cpu_cgroup_css_free
,
7231 .fork
= cpu_cgroup_fork
,
7232 .can_attach
= cpu_cgroup_can_attach
,
7233 .attach
= cpu_cgroup_attach
,
7234 .legacy_cftypes
= cpu_files
,
7238 #endif /* CONFIG_CGROUP_SCHED */
7240 void dump_cpu_task(int cpu
)
7242 pr_info("Task dump for CPU %d:\n", cpu
);
7243 sched_show_task(cpu_curr(cpu
));
7247 * Nice levels are multiplicative, with a gentle 10% change for every
7248 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
7249 * nice 1, it will get ~10% less CPU time than another CPU-bound task
7250 * that remained on nice 0.
7252 * The "10% effect" is relative and cumulative: from _any_ nice level,
7253 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
7254 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
7255 * If a task goes up by ~10% and another task goes down by ~10% then
7256 * the relative distance between them is ~25%.)
7258 const int sched_prio_to_weight
[40] = {
7259 /* -20 */ 88761, 71755, 56483, 46273, 36291,
7260 /* -15 */ 29154, 23254, 18705, 14949, 11916,
7261 /* -10 */ 9548, 7620, 6100, 4904, 3906,
7262 /* -5 */ 3121, 2501, 1991, 1586, 1277,
7263 /* 0 */ 1024, 820, 655, 526, 423,
7264 /* 5 */ 335, 272, 215, 172, 137,
7265 /* 10 */ 110, 87, 70, 56, 45,
7266 /* 15 */ 36, 29, 23, 18, 15,
7270 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
7272 * In cases where the weight does not change often, we can use the
7273 * precalculated inverse to speed up arithmetics by turning divisions
7274 * into multiplications:
7276 const u32 sched_prio_to_wmult
[40] = {
7277 /* -20 */ 48388, 59856, 76040, 92818, 118348,
7278 /* -15 */ 147320, 184698, 229616, 287308, 360437,
7279 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
7280 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
7281 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
7282 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
7283 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
7284 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,